From f8364997c0d736a35e884d609e254e05c78b1b2c Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Sat, 24 Aug 2019 09:20:57 +0300
Subject: [PATCH 01/56] [WIP] maxpool2d_bp fix (#160)

* one test for maxpool2d_bp

Signed-off-by: raver119 <raver119@gmail.com>

* - maxpool2d_bp cuda fix for NaNs
- streamSync after each custom op execution

Signed-off-by: raver119 <raver119@gmail.com>
---
 libnd4j/blas/cuda/NativeOps.cu                | 74 +++++++++----------
 .../declarable/helpers/cuda/convolutions.cu   |  2 +
 .../nd4j/linalg/custom/CustomOpsTests.java    | 16 ++++
 3 files changed, 54 insertions(+), 38 deletions(-)

diff --git a/libnd4j/blas/cuda/NativeOps.cu b/libnd4j/blas/cuda/NativeOps.cu
index 5c6dadbaf..e75aa422c 100755
--- a/libnd4j/blas/cuda/NativeOps.cu
+++ b/libnd4j/blas/cuda/NativeOps.cu
@@ -866,9 +866,10 @@ void initializeFunctions(Nd4jPointer *functions) {
 Nd4jPointer mallocHost(Nd4jLong memorySize, int flags) {
 	Nd4jPointer pointer;
 	// cudaHostAllocMapped |cudaHostAllocPortable
-	cudaError_t res = cudaHostAlloc(reinterpret_cast<void **>(&pointer), memorySize, cudaHostAllocDefault);
+	auto res = cudaHostAlloc(reinterpret_cast<void **>(&pointer), memorySize, cudaHostAllocDefault);
 	if (res != 0)
-		pointer = 0L;
+		throw nd4j::cuda_exception::build("cudaHostAlloc(...) failed", res);
+
 	return pointer;
 }
 
@@ -884,7 +885,7 @@ Nd4jPointer mallocDevice(Nd4jLong memorySize, int deviceId, int flags) {
 	Nd4jPointer pointer;
 	auto res = cudaMalloc(reinterpret_cast<void **>(&pointer), memorySize);
 	if (res != 0)
-		pointer = 0L;
+		throw nd4j::cuda_exception::build("cudaMalloc(...) failed", res);
 	return pointer;
 }
 
@@ -894,9 +895,9 @@ Nd4jPointer mallocDevice(Nd4jLong memorySize, int deviceId, int flags) {
  * @param pointer pointer that'll be freed
  */
 int freeHost(Nd4jPointer pointer) {
-	cudaError_t res = cudaFreeHost(reinterpret_cast<void *>(pointer));
+	auto res = cudaFreeHost(reinterpret_cast<void *>(pointer));
 	if (res != 0)
-		pointer = 0L;
+		throw nd4j::cuda_exception::build("cudaFreeHost(...) failed", res);
 	return 1L;
 }
 
@@ -907,9 +908,10 @@ int freeHost(Nd4jPointer pointer) {
  * @param ptrToDeviceId pointer to deviceId.
  */
 int freeDevice(Nd4jPointer pointer, int deviceId) {
-	cudaError_t res = cudaFree(reinterpret_cast<void *>(pointer));
+	auto res = cudaFree(reinterpret_cast<void *>(pointer));
 	if (res != 0)
-		pointer = 0L;
+		throw nd4j::cuda_exception::build("cudaFree(...) failed", res);
+
 	return 1L;
 }
 
@@ -934,7 +936,7 @@ Nd4jPointer createStream() {
     auto stream = new cudaStream_t();
     auto dZ = cudaStreamCreate(stream);
     if (dZ != 0)
-        throw std::runtime_error("cudaStreamCreate(...) failed");
+        throw nd4j::cuda_exception::build("cudaStreamCreate(...) failed", dZ);
 
     return stream;
 }
@@ -944,23 +946,21 @@ Nd4jPointer createEvent() {
 
 	CHECK_ALLOC(nativeEvent, "Failed to allocate new CUDA event buffer", sizeof(cudaEvent_t));
 
-	cudaError_t dZ = cudaEventCreateWithFlags(reinterpret_cast<cudaEvent_t *>(&nativeEvent), cudaEventDisableTiming);
-	checkCudaErrors(dZ);
+	auto dZ = cudaEventCreateWithFlags(reinterpret_cast<cudaEvent_t *>(&nativeEvent), cudaEventDisableTiming);
 	if (dZ != 0)
-		throw std::runtime_error("cudaEventCreateWithFlags(...) failed");
+		throw nd4j::cuda_exception::build("cudaEventCreateWithFlags(...) failed", dZ);
 
 
 	return nativeEvent;
 }
 
 int registerEvent(Nd4jPointer event, Nd4jPointer stream) {
-	cudaEvent_t *pEvent = reinterpret_cast<cudaEvent_t *>(&event);
-	cudaStream_t *pStream = reinterpret_cast<cudaStream_t *>(stream);
+	auto pEvent = reinterpret_cast<cudaEvent_t *>(&event);
+	auto pStream = reinterpret_cast<cudaStream_t *>(stream);
 
-	cudaError_t dZ = cudaEventRecord(*pEvent, *pStream);
-	checkCudaErrors(dZ);
+    auto dZ = cudaEventRecord(*pEvent, *pStream);
 	if (dZ != 0)
-		throw std::runtime_error("cudaEventRecord(...) failed");
+		throw nd4j::cuda_exception::build("cudaEventRecord(...) failed", dZ);
 
 	return 1;
 }
@@ -1065,53 +1065,48 @@ int memcpyAsync(Nd4jPointer dst, Nd4jPointer src, Nd4jLong size, int flags, Nd4j
 }
 
 int memsetSync(Nd4jPointer dst, int value, Nd4jLong size, int flags, Nd4jPointer reserved) {
-	cudaError_t dZ = cudaMemset(reinterpret_cast<void *>(dst), value, static_cast<size_t>(size));
-	checkCudaErrors(dZ);
+	auto dZ = cudaMemset(reinterpret_cast<void *>(dst), value, static_cast<size_t>(size));
 	if (dZ != 0)
-		throw std::runtime_error("cudaMemset(...) failed");
+		throw nd4j::cuda_exception::build("cudaMemset(...) failed", dZ);
 
 	return 1;
 }
 
 int memsetAsync(Nd4jPointer dst, int value, Nd4jLong size, int flags, Nd4jPointer reserved) {
-	cudaStream_t *pStream = reinterpret_cast<cudaStream_t *>(reserved);
+	auto pStream = reinterpret_cast<cudaStream_t *>(reserved);
 
-	cudaError_t dZ = cudaMemsetAsync(reinterpret_cast<void *>(dst), value, static_cast<size_t>(size), *pStream);
-	checkCudaErrors(dZ);
+	auto dZ = cudaMemsetAsync(reinterpret_cast<void *>(dst), value, static_cast<size_t>(size), *pStream);
 	if (dZ != 0)
-		throw std::runtime_error("cudaMemsetAsync(...) failed");
+		throw nd4j::cuda_exception::build("cudaMemsetAsync(...) failed", dZ);
 
 	return 1;
 }
 
 int destroyEvent(Nd4jPointer event) {
-	cudaEvent_t *pEvent = reinterpret_cast<cudaEvent_t *>(&event);
-	cudaError_t dZ = cudaEventDestroy(*pEvent);
-	checkCudaErrors(dZ);
+	auto pEvent = reinterpret_cast<cudaEvent_t *>(&event);
+	auto dZ = cudaEventDestroy(*pEvent);
 	if (dZ != 0)
-		throw std::runtime_error("cudaEvenDestroy(...) failed");
+		throw nd4j::cuda_exception::build("cudaEvenDestroy(...) failed", dZ);
 
 	return 1;
 }
 
 int streamSynchronize(Nd4jPointer stream) {
-	cudaStream_t *pStream = reinterpret_cast<cudaStream_t *>(stream);
+	auto pStream = reinterpret_cast<cudaStream_t *>(stream);
 
-	cudaError_t dZ = cudaStreamSynchronize(*pStream);
-	checkCudaErrors(dZ);
+	auto dZ = cudaStreamSynchronize(*pStream);
 	if (dZ != 0)
-        throw std::runtime_error("cudaStreamSynchronize(...) failed");
+        throw nd4j::cuda_exception::build("cudaStreamSynchronize(...) failed", dZ);
 
 	return 1L;
 }
 
 int eventSynchronize(Nd4jPointer event) {
-	cudaEvent_t *pEvent = reinterpret_cast<cudaEvent_t *>(&event);
+	auto pEvent = reinterpret_cast<cudaEvent_t *>(&event);
 
-	cudaError_t dZ = cudaEventSynchronize(*pEvent);
-	checkCudaErrors(dZ);
+	auto  dZ = cudaEventSynchronize(*pEvent);
 	if (dZ != 0)
-        throw std::runtime_error("cudaEventSynchronize(...) failed");
+        throw nd4j::cuda_exception::build("cudaEventSynchronize(...) failed", dZ);
 
 	return 1L;
 }
@@ -2697,13 +2692,16 @@ int execCustomOp2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer opConte
 
     auto result = op->execute(context);
 
-    // FIXME: remove once CUDA backend is 100% ready
+    auto res = cudaStreamSynchronize(*context->launchContext()->getCudaStream());
+    if (res != 0)
+        throw nd4j::cuda_exception::build("customOp execution failed", res);
+
     for (auto v:context->fastpath_in()) {
-        v->makeBothActual();
+        v->syncToDevice();
     }
 
     for (auto v:context->fastpath_out()) {
-        v->makeBothActual();
+        v->syncToDevice();
     }
 
     return result;
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/convolutions.cu b/libnd4j/include/ops/declarable/helpers/cuda/convolutions.cu
index e224329f0..98ab86dec 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/convolutions.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/convolutions.cu
@@ -907,6 +907,8 @@ __global__ static void pooling2dBPCuda(const void* vx, const Nd4jLong* xShapeInf
 
         /*** max ***/
         case 0: {
+            coord2 = hstart;
+            coord3 = hend;
 
             T max = -DataTypeUtils::max<T>();
             for (coords[2] = hstart; coords[2] < hend; coords[2] += dH) {
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/custom/CustomOpsTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/custom/CustomOpsTests.java
index c2f5dedc5..6c4595a0c 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/custom/CustomOpsTests.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/custom/CustomOpsTests.java
@@ -732,4 +732,20 @@ public class CustomOpsTests extends BaseNd4jTest {
             fail("Failed datatypes: " + failed.toString());
         }
     }
+
+    @Test
+    public void testMaxPool2Dbp_1() {
+        val x = Nd4j.create(DataType.HALF, 2,3,16,16).assign(Double.NaN);
+        val y = Nd4j.create(DataType.HALF, 2,3,8,8).assign(Double.NaN);
+        val z = Nd4j.create(DataType.HALF, 2,3,16,16);
+
+        val op = DynamicCustomOp.builder("maxpool2d_bp")
+                .addInputs(x, y)
+                .addOutputs(z)
+                .addIntegerArguments(2, 2, 2, 2, 8,8, 1,1,1, 0,0)
+                .build();
+
+        Nd4j.exec(op);
+        Nd4j.getExecutioner().commit();
+    }
 }

From b85238a6df8ce72021b33f40e9beb99d9d3661a0 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Sat, 24 Aug 2019 17:33:11 +1000
Subject: [PATCH 02/56] MLN/CG: Don't swallow exceptions if a second exception
 occurs during workspace closing (#161)

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 .../nn/graph/ComputationGraph.java            | 93 ++++++++++++-------
 .../nn/multilayer/MultiLayerNetwork.java      | 91 ++++++++++++------
 2 files changed, 121 insertions(+), 63 deletions(-)

diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/ComputationGraph.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/ComputationGraph.java
index 99f8aeff0..00c0cf7d6 100755
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/ComputationGraph.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/ComputationGraph.java
@@ -2278,6 +2278,7 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
         LayerWorkspaceMgr allNone = noWS ? LayerWorkspaceMgr.noWorkspaces(helperWorkspaces) : null;
         List<MemoryWorkspace>[] closeAtEndIteraton = (List<MemoryWorkspace>[])new List[topologicalOrder.length];
         MemoryWorkspace initialWorkspace = Nd4j.getMemoryManager().getCurrentWorkspace();
+        Throwable t = null;
         try {
             for (int i = 0; i <= stopIndex; i++) {
                 GraphVertex current = vertices[topologicalOrder[i]];
@@ -2302,14 +2303,14 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                                 .with(ArrayType.RNN_FF_LOOP_WORKING_MEM, WS_RNN_LOOP_WORKING_MEM, WS_RNN_LOOP_WORKING_MEM_CONFIG)
                                 .build();
 
-                        if(detachedInputs){
+                        if (detachedInputs) {
                             //Sometimes (like: external errors use cases) we don't want the activations/inputs to be
                             // in a workspace
                             workspaceMgr.setScopedOutFor(ArrayType.INPUT);
                             workspaceMgr.setScopedOutFor(ArrayType.ACTIVATIONS);
                         } else {
                             //Don't leverage out of async MultiDataSetIterator workspaces
-                            if(features[0].isAttached()){
+                            if (features[0].isAttached()) {
                                 workspaceMgr.setNoLeverageOverride(features[0].data().getParentWorkspace().getId());
                             }
                         }
@@ -2326,7 +2327,7 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                 if (ArrayUtils.contains(layerIndexes, vIdx)) {
                     isRequiredOutput = true;
 
-                    if(outputWorkspace != null && !(outputWorkspace instanceof DummyWorkspace)){
+                    if (outputWorkspace != null && !(outputWorkspace instanceof DummyWorkspace)) {
                         //Place activations in user-specified workspace
                         origWSAct = workspaceMgr.getWorkspaceName(ArrayType.ACTIVATIONS);
                         origWSActConf = workspaceMgr.getConfiguration(ArrayType.ACTIVATIONS);
@@ -2345,7 +2346,7 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                 //Open the relevant workspace for the activations.
                 //Note that this will be closed only once the current vertex's activations have been consumed
                 MemoryWorkspace wsActivations = null;
-                if(outputWorkspace == null || outputWorkspace instanceof DummyWorkspace || !isRequiredOutput ){    //Open WS if (a) no external/output WS (if present, it's already open), or (b) not being placed in external/output WS
+                if (outputWorkspace == null || outputWorkspace instanceof DummyWorkspace || !isRequiredOutput) {    //Open WS if (a) no external/output WS (if present, it's already open), or (b) not being placed in external/output WS
                     wsActivations = workspaceMgr.notifyScopeEntered(ArrayType.ACTIVATIONS);
                     openActivationsWorkspaces.put(wsActivations, workspaceMgr);
                 }
@@ -2353,11 +2354,11 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                 //Note that because we're opening activation workspaces not in any defined order (i.e., workspace
                 // use isn't simply nested), we'll manually override the previous workspace setting. Otherwise, when we
                 // close these workspaces, the "current" workspace may be set to the incorrect one
-                if(wsActivations != null )
+                if (wsActivations != null)
                     wsActivations.setPreviousWorkspace(initialWorkspace);
 
                 int closeableAt = vertexOutputsFullyConsumedByStep[vIdx];
-                if(outputWorkspace == null || outputWorkspace instanceof DummyWorkspace || (wsActivations != null && !outputWorkspace.getId().equals(wsActivations.getId()))) {
+                if (outputWorkspace == null || outputWorkspace instanceof DummyWorkspace || (wsActivations != null && !outputWorkspace.getId().equals(wsActivations.getId()))) {
                     if (closeAtEndIteraton[closeableAt] == null) {
                         closeAtEndIteraton[closeableAt] = new ArrayList<>();
                     }
@@ -2373,18 +2374,18 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                         out = features[vIdx];
                     } else {
 
-                        if(fwdPassType == FwdPassType.STANDARD){
+                        if (fwdPassType == FwdPassType.STANDARD) {
                             //Standard feed-forward case
                             out = current.doForward(train, workspaceMgr);
-                        } else if(fwdPassType == FwdPassType.RNN_TIMESTEP){
+                        } else if (fwdPassType == FwdPassType.RNN_TIMESTEP) {
                             if (current.hasLayer()) {
                                 //Layer
                                 INDArray input = current.getInputs()[0];
                                 Layer l = current.getLayer();
                                 if (l instanceof RecurrentLayer) {
                                     out = ((RecurrentLayer) l).rnnTimeStep(reshapeTimeStepInput(input), workspaceMgr);
-                                } else if(l instanceof org.deeplearning4j.nn.layers.wrapper.BaseWrapperLayer && ((org.deeplearning4j.nn.layers.wrapper.BaseWrapperLayer)l).getUnderlying() instanceof RecurrentLayer){
-                                    RecurrentLayer rl = ((RecurrentLayer) ((org.deeplearning4j.nn.layers.wrapper.BaseWrapperLayer)l).getUnderlying());
+                                } else if (l instanceof org.deeplearning4j.nn.layers.wrapper.BaseWrapperLayer && ((org.deeplearning4j.nn.layers.wrapper.BaseWrapperLayer) l).getUnderlying() instanceof RecurrentLayer) {
+                                    RecurrentLayer rl = ((RecurrentLayer) ((org.deeplearning4j.nn.layers.wrapper.BaseWrapperLayer) l).getUnderlying());
                                     out = rl.rnnTimeStep(reshapeTimeStepInput(input), workspaceMgr);
                                 } else if (l instanceof MultiLayerNetwork) {
                                     out = ((MultiLayerNetwork) l).rnnTimeStep(reshapeTimeStepInput(input));
@@ -2402,7 +2403,7 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                         validateArrayWorkspaces(workspaceMgr, out, ArrayType.ACTIVATIONS, vName, false, "Feed forward (inference)");
                     }
 
-                    if(inputsTo != null) {  //Output vertices may not input to any other vertices
+                    if (inputsTo != null) {  //Output vertices may not input to any other vertices
                         for (VertexIndices v : inputsTo) {
                             //Note that we don't have to do anything special here: the activations are always detached in
                             // this method
@@ -2412,13 +2413,13 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                         }
                     }
 
-                    if(clearLayerInputs) {
+                    if (clearLayerInputs) {
                         current.clear();
                     }
 
-                    if(isRequiredOutput){
+                    if (isRequiredOutput) {
                         outputs[ArrayUtils.indexOf(layerIndexes, vIdx)] = out;
-                        if(origWSAct != null){
+                        if (origWSAct != null) {
                             //Reset the configuration, as we may reuse this workspace manager...
                             workspaceMgr.setWorkspace(ArrayType.ACTIVATIONS, origWSAct, origWSActConf);
                         }
@@ -2428,14 +2429,16 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                 //Close any activations workspaces that we no longer require
                 //Note that activations workspaces can be closed only once the corresponding output activations have
                 // been fully consumed
-                if(closeAtEndIteraton[i] != null){
-                    for(MemoryWorkspace wsAct : closeAtEndIteraton[i]){
+                if (closeAtEndIteraton[i] != null) {
+                    for (MemoryWorkspace wsAct : closeAtEndIteraton[i]) {
                         wsAct.close();
                         LayerWorkspaceMgr canNowReuse = openActivationsWorkspaces.remove(wsAct);
                         freeWorkspaceManagers.add(canNowReuse);
                     }
                 }
             }
+        } catch (Throwable t2){
+            t = t2;
         } finally {
             //Close all open workspaces... usually this list will be empty, but not if an exception is thrown
             //Though if stopIndex < numLayers, some might still be open
@@ -2444,7 +2447,15 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                     //Edge case here: seems that scoping out can increase the tagScope of the current WS
                     //and if we hit an exception during forward pass, we aren't guaranteed to call close a sufficient
                     // number of times to actually close it, in all cases
-                    ws.close();
+                    try{
+                        ws.close();
+                    } catch (Throwable t2){
+                        if(t != null){
+                            log.error("Encountered second exception while trying to close workspace after initial exception");
+                            log.error("Original exception:", t);
+                            throw t2;
+                        }
+                    }
                 }
             }
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
@@ -2581,28 +2592,29 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
 
         boolean traceLog = log.isTraceEnabled();
 
-        try{
-            for(int i=topologicalOrder.length-1; i>= 0; i--){
+        Throwable t = null;
+        try {
+            for (int i = topologicalOrder.length - 1; i >= 0; i--) {
                 boolean hitFrozen = false;
                 GraphVertex current = vertices[topologicalOrder[i]];
                 int vIdx = current.getVertexIndex();
                 String vertexName = current.getVertexName();
 
-                if(traceLog){
+                if (traceLog) {
                     log.trace("About backprop: {} (\"{}\") - {}", i, vertexName, current.getClass().getSimpleName());
                 }
 
                 //FIXME: make the frozen vertex feature extraction more flexible
-                if (current.hasLayer() && current.getLayer() instanceof FrozenLayer || current instanceof FrozenVertex){
+                if (current.hasLayer() && current.getLayer() instanceof FrozenLayer || current instanceof FrozenVertex) {
                     hitFrozen = true;
                 }
 
-                if (current.isInputVertex() || hitFrozen){
+                if (current.isInputVertex() || hitFrozen) {
                     //Close any activation gradient workspaces that we no longer require
                     //Note that activation gradient workspaces can be closed only once the corresponding activations
                     // gradients have been fully consumed
-                    if(closeAtEndIteraton[i] != null){
-                        for(MemoryWorkspace wsAct : closeAtEndIteraton[i]){
+                    if (closeAtEndIteraton[i] != null) {
+                        for (MemoryWorkspace wsAct : closeAtEndIteraton[i]) {
                             wsAct.close();
                             LayerWorkspaceMgr canNowReuse = openActivationsWorkspaces.remove(wsAct);
                             freeWorkspaceManagers.add(canNowReuse);
@@ -2680,7 +2692,7 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                 wsActivationGrads.setPreviousWorkspace(initialWorkspace);
 
                 int closeableAt = vertexActGradsFullyConsumedByStep[vIdx];
-                if(closeableAt >= 0) {
+                if (closeableAt >= 0) {
                     if (closeAtEndIteraton[closeableAt] == null) {
                         closeAtEndIteraton[closeableAt] = new ArrayList<>();
                     }
@@ -2689,14 +2701,14 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
 
                 Pair<Gradient, INDArray[]> pair;
                 INDArray[] epsilons;
-                try(MemoryWorkspace wsWorkingMem = workspaceMgr.notifyScopeEntered(ArrayType.BP_WORKING_MEM)){
+                try (MemoryWorkspace wsWorkingMem = workspaceMgr.notifyScopeEntered(ArrayType.BP_WORKING_MEM)) {
                     pair = current.doBackward(truncatedBPTT, workspaceMgr);
                     epsilons = pair.getSecond();
 
                     //Validate workspace location for the activation gradients:
                     //validateArrayWorkspaces(LayerWorkspaceMgr mgr, INDArray array, ArrayType arrayType, String vertexName, boolean isInputVertex, String op){
                     for (INDArray epsilon : epsilons) {
-                        if(epsilon != null) {
+                        if (epsilon != null) {
                             //May be null for EmbeddingLayer, etc
                             validateArrayWorkspaces(workspaceMgr, epsilon, ArrayType.ACTIVATION_GRAD, vertexName, false, "Backprop");
                         }
@@ -2732,15 +2744,15 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                         tempList.addFirst(new Triple<>(newName, entry.getValue(),
                                 g.flatteningOrderForVariable(origName)));
                     }
-                    for (Triple<String, INDArray, Character> t : tempList)
-                        gradients.addFirst(t);
+                    for (Triple<String, INDArray, Character> triple : tempList)
+                        gradients.addFirst(triple);
                 }
 
                 //Close any activation gradient workspaces that we no longer require
                 //Note that activation gradient workspaces can be closed only once the corresponding activations
                 // gradients have been fully consumed
-                if(closeAtEndIteraton[i] != null){
-                    for(MemoryWorkspace wsAct : closeAtEndIteraton[i]){
+                if (closeAtEndIteraton[i] != null) {
+                    for (MemoryWorkspace wsAct : closeAtEndIteraton[i]) {
                         wsAct.close();
                         LayerWorkspaceMgr canNowReuse = openActivationsWorkspaces.remove(wsAct);
                         freeWorkspaceManagers.add(canNowReuse);
@@ -2748,23 +2760,32 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                     closeAtEndIteraton[i] = null;
                 }
 
-                if(traceLog){
+                if (traceLog) {
                     log.trace("Completed backprop: {} (\"{}\") - {}", i, vertexName, current.getClass().getSimpleName());
                 }
             }
-
+        } catch (Throwable t2){
+            t = t2;
         } finally {
             //Close all open workspaces... usually this list will be empty, but not if an exception is thrown
             for(MemoryWorkspace ws : openActivationsWorkspaces.keySet()){
-                ws.close();
+                try{
+                    ws.close();
+                } catch (Throwable t2){
+                    if(t != null){
+                        log.error("Encountered second exception while trying to close workspace after initial exception");
+                        log.error("Original exception:", t);
+                        throw t2;
+                    }
+                }
             }
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
         }
 
         //Now, add the gradients in the order we need them in for flattening (same as params order)
         Gradient gradient = new DefaultGradient(flattenedGradients);
-        for (Triple<String, INDArray, Character> t : gradients) {
-            gradient.setGradientFor(t.getFirst(), t.getSecond(), t.getThird());
+        for (Triple<String, INDArray, Character> tr : gradients) {
+            gradient.setGradientFor(tr.getFirst(), tr.getSecond(), tr.getThird());
         }
 
         this.gradient = gradient;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/multilayer/MultiLayerNetwork.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/multilayer/MultiLayerNetwork.java
index 731ca398b..dd495a620 100755
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/multilayer/MultiLayerNetwork.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/multilayer/MultiLayerNetwork.java
@@ -1242,17 +1242,18 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
 
         boolean traceLog = log.isTraceEnabled();
 
+        Throwable t = null;
         try {
             for (int i = 0; i <= layerIndex; i++) {
                 LayerWorkspaceMgr mgr = (i % 2 == 0 ? mgrEven : mgrOdd);
 
-                if(traceLog){
+                if (traceLog) {
                     log.trace("About to forward pass: {} - {}", i, layers[i].getClass().getSimpleName());
                 }
 
                 //Edge case: for first layer with dropout, inputs can't be in previous workspace (as it hasn't been opened yet)
                 //Hence: put inputs in working memory
-                if(i == 0 && wsm != WorkspaceMode.NONE){
+                if (i == 0 && wsm != WorkspaceMode.NONE) {
                     mgr.setWorkspace(ArrayType.INPUT, WS_LAYER_WORKING_MEM, WS_LAYER_WORKING_MEM_CONFIG);
                 }
 
@@ -1268,7 +1269,7 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
                     temp.setPreviousWorkspace(initialWorkspace);
 
 
-                    if(i == 0 && input.isAttached()){
+                    if (i == 0 && input.isAttached()) {
                         //Don't leverage out of async DataSetIterator workspaces
                         mgr.setNoLeverageOverride(input.data().getParentWorkspace().getId());
                     }
@@ -1279,8 +1280,8 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
                         validateArrayWorkspaces(mgr, input, ArrayType.ACTIVATIONS, i, true, "Output of layer (inference)");
                     }
 
-                    if ( i == layerIndex ) {
-                        if(outputWorkspace != null && !(outputWorkspace instanceof DummyWorkspace)){
+                    if (i == layerIndex) {
+                        if (outputWorkspace != null && !(outputWorkspace instanceof DummyWorkspace)) {
                             //Place activations in user-specified workspace
                             mgr.setWorkspace(ArrayType.ACTIVATIONS, outputWorkspace.getId(), outputWorkspace.getWorkspaceConfiguration());
                         } else {
@@ -1289,15 +1290,15 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
                         }
                     }
 
-                    if(fwdPassType == FwdPassType.STANDARD){
+                    if (fwdPassType == FwdPassType.STANDARD) {
                         //Standard feed-forward case
                         input = layers[i].activate(input, train, mgr);
-                    } else if(fwdPassType == FwdPassType.RNN_TIMESTEP){
+                    } else if (fwdPassType == FwdPassType.RNN_TIMESTEP) {
                         //rnnTimeStep case
                         if (layers[i] instanceof RecurrentLayer) {
                             input = ((RecurrentLayer) layers[i]).rnnTimeStep(reshapeTimeStepInput(input), mgr);
-                        } else if(layers[i] instanceof BaseWrapperLayer && ((BaseWrapperLayer)layers[i]).getUnderlying() instanceof RecurrentLayer){
-                            RecurrentLayer rl = ((RecurrentLayer) ((BaseWrapperLayer)layers[i]).getUnderlying());
+                        } else if (layers[i] instanceof BaseWrapperLayer && ((BaseWrapperLayer) layers[i]).getUnderlying() instanceof RecurrentLayer) {
+                            RecurrentLayer rl = ((RecurrentLayer) ((BaseWrapperLayer) layers[i]).getUnderlying());
                             input = rl.rnnTimeStep(reshapeTimeStepInput(input), mgr);
                         } else if (layers[i] instanceof MultiLayerNetwork) {
                             input = ((MultiLayerNetwork) layers[i]).rnnTimeStep(reshapeTimeStepInput(input));
@@ -1311,34 +1312,51 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
                     //Validation: Exception if invalid (bad layer implementation)
                     validateArrayWorkspaces(mgr, input, ArrayType.ACTIVATIONS, i, false, "Output of layer (inference)");
 
-                    if(wsActCloseNext != null){
+                    if (wsActCloseNext != null) {
                         wsActCloseNext.close();
                     }
                     wsActCloseNext = temp;
                     temp = null;
                 }
 
-                if(traceLog){
+                if (traceLog) {
                     log.trace("Completed forward pass: {} - {}", i, layers[i].getClass().getSimpleName());
                 }
 
                 //Edge case: for first layer with dropout, inputs can't be in previous workspace (as it hasn't been opened yet)
                 //Hence: put inputs in working memory -> set back to default for next use of workspace mgr
-                if(i == 0 && wsm != WorkspaceMode.NONE){
+                if (i == 0 && wsm != WorkspaceMode.NONE) {
                     mgr.setWorkspace(ArrayType.INPUT, WS_LAYER_ACT_2, WS_LAYER_ACT_X_CONFIG);            //Inputs should always be in the previous WS
                 }
             }
-
+        } catch (Throwable t2){
+            t = t2;
         } finally {
             if(wsActCloseNext != null){
-                wsActCloseNext.close();
+                try {
+                    wsActCloseNext.close();
+                } catch (Throwable t2){
+                    if(t != null){
+                        log.error("Encountered second exception while trying to close workspace after initial exception");
+                        log.error("Original exception:", t);
+                        throw t2;
+                    }
+                }
             }
             if(temp != null){
                 //Should only be non-null on exception
                 while(temp.isScopeActive()){
                     //For safety, should never occur in theory: a single close() call may not be sufficient, if
                     // workspace scope was borrowed and not properly closed when exception occurred
-                    temp.close();
+                    try{
+                        temp.close();
+                    } catch (Throwable t2){
+                        if(t != null){
+                            log.error("Encountered second exception while trying to close workspace after initial exception");
+                            log.error("Original exception:", t);
+                            throw t2;
+                        }
+                    }
                 }
             }
 
@@ -1871,13 +1889,14 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
 
         boolean traceLog = log.isTraceEnabled();
 
+        Throwable t = null;
         try {
             for (int i = layers.length - 1; i >= 0; i--) {
                 if (layers[i] instanceof FrozenLayer) {
                     break;
                 }
 
-                if(traceLog){
+                if (traceLog) {
                     log.trace("About to backprop: {} - {}", i, layers[i].getClass().getSimpleName());
                 }
 
@@ -1897,7 +1916,7 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
 
                 //Open activation gradients WS *then* BP working memory, so BP working memory is opened last for use in layers
                 wsActGradTemp = workspaceMgr.notifyScopeEntered(ArrayType.ACTIVATION_GRAD);
-                try(MemoryWorkspace wsBPWorking = workspaceMgr.notifyScopeEntered(ArrayType.BP_WORKING_MEM)){
+                try (MemoryWorkspace wsBPWorking = workspaceMgr.notifyScopeEntered(ArrayType.BP_WORKING_MEM)) {
 
                     //Note that because we're opening activation workspaces not in a simple nested order, we'll manually
                     // override the previous workspace setting. Otherwise, when we close these workspaces, the "current"
@@ -1907,7 +1926,7 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
 
                     INDArray eps = (i == layers.length - 1 ? epsilon : currPair.getRight());  //eps is null for OutputLayer
 
-                    if(!tbptt){
+                    if (!tbptt) {
                         //Standard case
                         currPair = layers[i].backpropGradient(eps, workspaceMgr);
                     } else {
@@ -1920,7 +1939,7 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
                         }
                     }
 
-                    if(currPair.getSecond() != null) {
+                    if (currPair.getSecond() != null) {
                         //Edge case: may be null for Embedding layer, for example
                         validateArrayWorkspaces(workspaceMgr, currPair.getSecond(), ArrayType.ACTIVATION_GRAD, i,
                                 false, "Backprop");
@@ -1936,38 +1955,56 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
                         currPair = new Pair<>(currPair.getFirst(),
                                 this.layerWiseConfigurations.getInputPreProcess(i)
                                         .backprop(currPair.getSecond(), getInputMiniBatchSize(), workspaceMgr));
-                        if (i > 0 && currPair.getSecond() != null){
+                        if (i > 0 && currPair.getSecond() != null) {
                             validateArrayWorkspaces(workspaceMgr, currPair.getSecond(), ArrayType.ACTIVATION_GRAD, i,
                                     true, "Backprop");
                         }
                     }
 
-                    if(i == 0 ){
-                        if(returnInputActGrad && currPair.getSecond() != null){
+                    if (i == 0) {
+                        if (returnInputActGrad && currPair.getSecond() != null) {
                             currPair.setSecond(currPair.getSecond().detach());
                         } else {
                             currPair.setSecond(null);
                         }
                     }
 
-                    if(wsActGradCloseNext != null){
+                    if (wsActGradCloseNext != null) {
                         wsActGradCloseNext.close();
                     }
                     wsActGradCloseNext = wsActGradTemp;
                     wsActGradTemp = null;
                 }
 
-                if(traceLog){
+                if (traceLog) {
                     log.trace("Completed backprop: {} - {}", i, layers[i].getClass().getSimpleName());
                 }
             }
+        } catch (Throwable thr ){
+            t = thr;
         } finally {
             if(wsActGradCloseNext != null){
-                wsActGradCloseNext.close();
+                try {
+                    wsActGradCloseNext.close();
+                } catch (Throwable t2){
+                    if(t != null){
+                        log.error("Encountered second exception while trying to close workspace after initial exception");
+                        log.error("Original exception:", t);
+                        throw t2;
+                    }
+                }
             }
-            if(wsActGradTemp != null){
+            if(wsActGradTemp != null) {
                 //Should only be non-null on exception
-                wsActGradTemp.close();
+                try {
+                    wsActGradTemp.close();
+                } catch (Throwable t2) {
+                    if (t != null) {
+                        log.error("Encountered second exception while trying to close workspace after initial exception");
+                        log.error("Original exception:", t);
+                        throw t2;
+                    }
+                }
             }
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
         }

From a9b08cc163847eda270b940056035d08311215f0 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Sat, 24 Aug 2019 19:22:36 +1000
Subject: [PATCH 03/56] Upgrade protobuf version (#162)

* First steps for protobuf version upgrade

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Phase 2

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Update imports to shaded protobuf

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Version fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Switch to single execution for protobuf codegen to work around plugin bug

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Automatically delete old PB generated files after name change

Signed-off-by: Alex Black <blacka101@gmail.com>
---
 .../nd4j-api-parent/nd4j-api/pom.xml          |  92 ++++---
 .../functions/DifferentialFunction.java       |  10 +-
 .../nd4j/autodiff/samediff/SDVariable.java    |   2 +-
 .../TensorflowDescriptorParser.java           |   2 +-
 .../imports/graphmapper/BaseGraphMapper.java  |   4 +-
 .../nd4j/imports/graphmapper/GraphMapper.java |   2 +-
 .../graphmapper/onnx/OnnxGraphMapper.java     | 120 ++++-----
 .../imports/graphmapper/tf/TFGraphMapper.java |   2 +-
 .../tf/tensors/TFTensorMappers.java           |   2 +-
 .../linalg/api/ops/BaseBroadcastBoolOp.java   |   4 +-
 .../nd4j/linalg/api/ops/BaseBroadcastOp.java  |   4 +-
 .../java/org/nd4j/linalg/api/ops/BaseOp.java  |   4 +-
 .../org/nd4j/linalg/api/ops/BaseReduceOp.java |   4 +-
 .../nd4j/linalg/api/ops/DynamicCustomOp.java  |   4 +-
 .../java/org/nd4j/linalg/api/ops/NoOp.java    |   4 +-
 .../linalg/api/ops/impl/controlflow/If.java   |   4 +-
 .../api/ops/impl/controlflow/While.java       |   4 +-
 .../impl/layers/ExternalErrorsFunction.java   |   4 +-
 .../linalg/api/ops/impl/layers/Linear.java    |   4 +-
 .../impl/layers/convolution/AvgPooling2D.java |   4 +-
 .../impl/layers/convolution/AvgPooling3D.java |   4 +-
 .../impl/layers/convolution/BatchNorm.java    |   4 +-
 .../ops/impl/layers/convolution/Conv1D.java   |   2 +-
 .../ops/impl/layers/convolution/Conv2D.java   |   4 +-
 .../ops/impl/layers/convolution/DeConv2D.java |   4 +-
 .../layers/convolution/DepthwiseConv2D.java   |   4 +-
 .../LocalResponseNormalization.java           |   4 +-
 .../impl/layers/convolution/MaxPooling2D.java |   4 +-
 .../impl/layers/convolution/MaxPooling3D.java |   4 +-
 .../impl/layers/convolution/Pooling2D.java    |   4 +-
 .../ops/impl/layers/recurrent/GRUCell.java    |   2 +-
 .../ops/impl/layers/recurrent/LSTMCell.java   |   4 +-
 .../api/ops/impl/layers/recurrent/SRU.java    |   4 +-
 .../ops/impl/layers/recurrent/SRUCell.java    |   4 +-
 .../nd4j/linalg/api/ops/impl/reduce/Mmul.java |   4 +-
 .../api/ops/impl/reduce/TensorMmul.java       |   4 +-
 .../linalg/api/ops/impl/shape/Concat.java     |   4 +-
 .../nd4j/linalg/api/ops/impl/shape/Diag.java  |   4 +-
 .../linalg/api/ops/impl/shape/DiagPart.java   |   4 +-
 .../linalg/api/ops/impl/shape/Gather.java     |   4 +-
 .../linalg/api/ops/impl/shape/GatherNd.java   |   2 +-
 .../linalg/api/ops/impl/shape/MergeAvg.java   |   4 +-
 .../linalg/api/ops/impl/shape/MergeMax.java   |   4 +-
 .../linalg/api/ops/impl/shape/MergeSum.java   |   4 +-
 .../api/ops/impl/shape/ParallelStack.java     |   4 +-
 .../nd4j/linalg/api/ops/impl/shape/Rank.java  |   4 +-
 .../linalg/api/ops/impl/shape/Repeat.java     |   4 +-
 .../linalg/api/ops/impl/shape/Reshape.java    |   4 +-
 .../api/ops/impl/shape/SequenceMask.java      |   2 +-
 .../nd4j/linalg/api/ops/impl/shape/Shape.java |   4 +-
 .../linalg/api/ops/impl/shape/ShapeN.java     |   2 +-
 .../nd4j/linalg/api/ops/impl/shape/Size.java  |   2 +-
 .../nd4j/linalg/api/ops/impl/shape/Stack.java |   4 +-
 .../linalg/api/ops/impl/shape/Transpose.java  |   4 +-
 .../linalg/api/ops/impl/shape/Unstack.java    |   4 +-
 .../api/ops/impl/shape/bp/ConcatBp.java       |   4 +-
 .../shape/tensorops/TensorArrayConcat.java    |   4 +-
 .../shape/tensorops/TensorArrayGather.java    |   4 +-
 .../impl/shape/tensorops/TensorArrayRead.java |   4 +-
 .../shape/tensorops/TensorArrayScatter.java   |   4 +-
 .../impl/shape/tensorops/TensorArraySize.java |   4 +-
 .../shape/tensorops/TensorArraySplit.java     |   4 +-
 .../ops/impl/transforms/clip/ClipByNorm.java  |   4 +-
 .../ops/impl/transforms/clip/ClipByValue.java |   4 +-
 .../ops/impl/transforms/custom/Assign.java    |   4 +-
 .../ops/impl/transforms/custom/CumProd.java   |   4 +-
 .../ops/impl/transforms/custom/CumSum.java    |   4 +-
 .../api/ops/impl/transforms/custom/Fill.java  |   4 +-
 .../impl/transforms/strict/RectifiedTanh.java |   4 +-
 .../api/ops/random/impl/DropOutInverted.java  |   4 +-
 .../linalg/api/ops/random/impl/Range.java     |   2 +-
 .../onnx/{onnx-ml.proto3 => onnx-ml.proto}    |   0
 ...-operators.proto3 => onnx-operators.proto} |   2 +-
 .../protobuf/onnx/{onnx.proto3 => onnx.proto} |   0
 .../conversion/GraphRunnerTest.java           |   2 +-
 .../conversion/GpuGraphRunnerTest.java        |   2 +-
 nd4j/nd4j-shade/pom.xml                       |   1 +
 nd4j/nd4j-shade/protobuf/pom.xml              | 228 ++++++++++++++++++
 .../conversion/TensorflowConversion.java      |   2 +-
 .../conversion/graphrunner/GraphRunner.java   |   8 +-
 80 files changed, 487 insertions(+), 228 deletions(-)
 rename nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/{onnx-ml.proto3 => onnx-ml.proto} (100%)
 rename nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/{onnx-operators.proto3 => onnx-operators.proto} (99%)
 rename nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/{onnx.proto3 => onnx.proto} (100%)
 create mode 100644 nd4j/nd4j-shade/protobuf/pom.xml

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/pom.xml b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/pom.xml
index 18680e699..21924f80a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/pom.xml
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/pom.xml
@@ -31,10 +31,35 @@
 
     <build>
         <plugins>
+
+            <!-- AB 2019/08/24 This plugin is to be added TEMPORARILY due to a change in the filenames of the generated ONNX -->
+            <!-- Normal "mvn clean" etc won't delete these files, and any users who have built ND4J even once before the
+                 change will run into a compilation error. This can be removed after a few weeks.-->
+            <plugin>
+                <groupId>org.apache.maven.plugins</groupId>
+                <artifactId>maven-antrun-plugin</artifactId>
+                <version>1.8</version>
+                <executions>
+                    <execution>
+                        <phase>generate-sources</phase>
+                        <goals>
+                            <goal>run</goal>
+                        </goals>
+                        <configuration>
+                            <target>
+                                <delete file="${project.build.sourceDirectory}/onnx/OnnxMlProto3.java" />
+                                <delete file="${project.build.sourceDirectory}/onnx/OnnxOperatorsProto3.java" />
+                                <delete file="${project.build.sourceDirectory}/onnx/OnnxProto3.java" />
+                            </target>
+                        </configuration>
+                    </execution>
+                </executions>
+            </plugin>
+
             <plugin>
                 <groupId>com.github.os72</groupId>
                 <artifactId>protoc-jar-maven-plugin</artifactId>
-                <version>3.5.1.1</version>
+                <version>3.8.0</version>
                 <executions>
                     <execution>
                         <id>tensorflow</id>
@@ -43,30 +68,14 @@
                             <goal>run</goal>
                         </goals>
                         <configuration>
-                            <type>java-shaded</type>
-                            <protocVersion>3.5.1</protocVersion>
+                            <protocVersion>3.8.0</protocVersion>
+                            <extension>.proto</extension>
                             <includeDirectories>
                                 <include>src/main/protobuf/tf</include>
+                                <include>src/main/protobuf/onnx</include>
                             </includeDirectories>
                             <inputDirectories>
                                 <include>src/main/protobuf/tf/tensorflow</include>
-                            </inputDirectories>
-                            <addSources>main</addSources>
-                            <cleanOutputFolder>false</cleanOutputFolder>
-                            <outputDirectory>src/main/java/</outputDirectory>
-                        </configuration>
-                    </execution>
-                    <execution>
-                        <id>onnx</id>
-                        <phase>generate-sources</phase>
-                        <goals>
-                            <goal>run</goal>
-                        </goals>
-                        <configuration>
-                            <type>java-shaded</type>
-                            <extension>.proto3</extension>
-                            <protocVersion>3.5.1</protocVersion>
-                            <inputDirectories>
                                 <include>src/main/protobuf/onnx</include>
                             </inputDirectories>
                             <addSources>main</addSources>
@@ -76,6 +85,32 @@
                     </execution>
                 </executions>
             </plugin>
+
+            <plugin>
+                <groupId>com.google.code.maven-replacer-plugin</groupId>
+                <artifactId>replacer</artifactId>
+                <version>1.5.3</version>
+                <configuration>
+                    <includes>
+                        <include>${project.build.sourceDirectory}/org/tensorflow/**</include>
+                        <include>${project.build.sourceDirectory}/tensorflow/**</include>
+                        <include>${project.build.sourceDirectory}/onnx/**</include>
+                    </includes>
+                    <token>com.google.protobuf.</token>
+                    <value>org.nd4j.shade.protobuf.</value>
+                </configuration>
+                <executions>
+                    <execution>
+                        <id>replace-imports</id>
+                        <phase>generate-sources</phase>
+                        <goals>
+                            <goal>replace</goal>
+                        </goals>
+                    </execution>
+                </executions>
+            </plugin>
+
+
             <plugin>
                 <groupId>org.apache.maven.plugins</groupId>
                 <artifactId>maven-compiler-plugin</artifactId>
@@ -148,20 +183,15 @@
             <version>${flatbuffers.version}</version>
         </dependency>
 
-        <!-- Note that this is shaded flatbuffers, see the protoc declaration above
-        mentioning java-shaded as the type for why we use this instead of google's (mainly due ot other systems packaging
-        their own older protobuf versions-->
+        <!-- Note that this is shaded protobuf. We use this instead of google's version mainly due ot other systems packaging
+        their own older (incompatible) protobuf versions-->
         <dependency>
-            <groupId>com.github.os72</groupId>
-            <artifactId>protobuf-java-shaded-351</artifactId>
-            <version>0.9</version>
-        </dependency>
-        <dependency>
-            <groupId>com.github.os72</groupId>
-            <artifactId>protobuf-java-util-shaded-351</artifactId>
-            <version>0.9</version>
+            <groupId>org.nd4j</groupId>
+            <artifactId>protobuf</artifactId>
+            <version>${project.version}</version>
         </dependency>
 
+
         <dependency>
             <groupId>org.objenesis</groupId>
             <artifactId>objenesis</artifactId>
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java
index 2d49ce56f..71bbd26ee 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java
@@ -21,7 +21,7 @@ import lombok.Getter;
 import lombok.Setter;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.autodiff.samediff.serde.FlatBuffersMapper;
@@ -101,10 +101,10 @@ public abstract class DifferentialFunction {
 
     /**
      * Initialize the function from the given
-     * {@link onnx.OnnxProto3.NodeProto}
+     * {@link onnx.Onnx.NodeProto}
      * @param node
      */
-    public DifferentialFunction(SameDiff sameDiff,onnx.OnnxProto3.NodeProto node,Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public DifferentialFunction(SameDiff sameDiff,onnx.Onnx.NodeProto node,Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         this.sameDiff = sameDiff;
         setInstanceId();
         initFromOnnx(node, sameDiff, attributesForNode, graph);
@@ -731,13 +731,13 @@ public abstract class DifferentialFunction {
 
     /**
      * Iniitialize the function from the given
-     * {@link onnx.OnnxProto3.NodeProto}
+     * {@link onnx.Onnx.NodeProto}
      * @param node
      * @param initWith
      * @param attributesForNode
      * @param graph
      */
-    public abstract void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph);
+    public abstract void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph);
 
 
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SDVariable.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SDVariable.java
index a7fb35520..430b4d83a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SDVariable.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SDVariable.java
@@ -19,7 +19,7 @@ package org.nd4j.autodiff.samediff;
 import java.util.Objects;
 import lombok.*;
 import lombok.extern.slf4j.Slf4j;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.internal.Variable;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/descriptors/tensorflow/TensorflowDescriptorParser.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/descriptors/tensorflow/TensorflowDescriptorParser.java
index fad55d101..3d0464782 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/descriptors/tensorflow/TensorflowDescriptorParser.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/descriptors/tensorflow/TensorflowDescriptorParser.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.imports.descriptors.tensorflow;
 
-import com.github.os72.protobuf351.TextFormat;
+import org.nd4j.shade.protobuf.TextFormat;
 import org.nd4j.linalg.exception.ND4JIllegalStateException;
 import org.nd4j.linalg.io.ClassPathResource;
 import org.tensorflow.framework.OpDef;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/BaseGraphMapper.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/BaseGraphMapper.java
index 92c888e0c..fe252aeeb 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/BaseGraphMapper.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/BaseGraphMapper.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.imports.graphmapper;
 
-import com.github.os72.protobuf351.Message;
-import com.github.os72.protobuf351.TextFormat;
+import org.nd4j.shade.protobuf.Message;
+import org.nd4j.shade.protobuf.TextFormat;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
 import org.apache.commons.io.IOUtils;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/GraphMapper.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/GraphMapper.java
index 8aad0f4d9..2d89a2b07 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/GraphMapper.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/GraphMapper.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.imports.graphmapper;
 
-import com.github.os72.protobuf351.Message;
+import org.nd4j.shade.protobuf.Message;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.imports.descriptors.properties.PropertyMapping;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/onnx/OnnxGraphMapper.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/onnx/OnnxGraphMapper.java
index 0bbfece6f..719ac792d 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/onnx/OnnxGraphMapper.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/onnx/OnnxGraphMapper.java
@@ -16,13 +16,13 @@
 
 package org.nd4j.imports.graphmapper.onnx;
 
-import com.github.os72.protobuf351.ByteString;
-import com.github.os72.protobuf351.Message;
+import org.nd4j.shade.protobuf.ByteString;
+import org.nd4j.shade.protobuf.Message;
 import com.google.common.primitives.Floats;
 import com.google.common.primitives.Ints;
 import com.google.common.primitives.Longs;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
@@ -52,7 +52,7 @@ import java.util.*;
  *
  * @author Adam Gibson
  */
-public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, OnnxProto3.NodeProto, OnnxProto3.AttributeProto,  onnx.OnnxProto3.TypeProto.Tensor> {
+public class OnnxGraphMapper extends BaseGraphMapper<Onnx.GraphProto, Onnx.NodeProto, Onnx.AttributeProto,  onnx.Onnx.TypeProto.Tensor> {
     private static OnnxGraphMapper INSTANCE = new OnnxGraphMapper();
 
 
@@ -64,9 +64,9 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
     @Override
     public void dumpBinaryProtoAsText(InputStream inputFile, File outputFile) {
         try {
-            OnnxProto3.ModelProto graphDef = OnnxProto3.ModelProto.parseFrom(inputFile);
+            Onnx.ModelProto graphDef = Onnx.ModelProto.parseFrom(inputFile);
             BufferedWriter bufferedWriter = new BufferedWriter(new FileWriter(outputFile,true));
-            for(OnnxProto3.NodeProto node : graphDef.getGraph().getNodeList()) {
+            for(Onnx.NodeProto node : graphDef.getGraph().getNodeList()) {
                 bufferedWriter.write(node.toString() + "\n");
             }
 
@@ -88,7 +88,7 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
      * @param node
      * @param graph
      */
-    public void initFunctionFromProperties(String mappedTfName, DifferentialFunction on, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.NodeProto node, OnnxProto3.GraphProto graph) {
+    public void initFunctionFromProperties(String mappedTfName, DifferentialFunction on, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.NodeProto node, Onnx.GraphProto graph) {
         val properties = on.mappingsForFunction();
         val tfProperties = properties.get(mappedTfName);
         val fields = DifferentialFunctionClassHolder.getInstance().getFieldsForFunction(on);
@@ -170,18 +170,18 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
     }
 
     @Override
-    public boolean isOpIgnoreException(OnnxProto3.NodeProto node) {
+    public boolean isOpIgnoreException(Onnx.NodeProto node) {
         return false;
     }
 
     @Override
-    public String getTargetMappingForOp(DifferentialFunction function, OnnxProto3.NodeProto node) {
+    public String getTargetMappingForOp(DifferentialFunction function, Onnx.NodeProto node) {
         return function.opName();
     }
 
 
     @Override
-    public void mapProperty(String name, DifferentialFunction on, OnnxProto3.NodeProto node, OnnxProto3.GraphProto graph, SameDiff sameDiff, Map<String, Map<String, PropertyMapping>> propertyMappingsForFunction) {
+    public void mapProperty(String name, DifferentialFunction on, Onnx.NodeProto node, Onnx.GraphProto graph, SameDiff sameDiff, Map<String, Map<String, PropertyMapping>> propertyMappingsForFunction) {
         val mapping = propertyMappingsForFunction.get(name).get(getTargetMappingForOp(on, node));
         val fields = DifferentialFunctionClassHolder.getInstance().getFieldsForFunction(on);
         /**
@@ -263,7 +263,7 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
 
 
     @Override
-    public OnnxProto3.NodeProto getNodeWithNameFromGraph(OnnxProto3.GraphProto graph, String name) {
+    public Onnx.NodeProto getNodeWithNameFromGraph(Onnx.GraphProto graph, String name) {
         for(int i = 0; i < graph.getNodeCount(); i++) {
             val node = graph.getNode(i);
             if(node.getName().equals(name))
@@ -274,21 +274,21 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
     }
 
     @Override
-    public boolean isPlaceHolderNode(OnnxProto3.TypeProto.Tensor node) {
+    public boolean isPlaceHolderNode(Onnx.TypeProto.Tensor node) {
         return false;
     }
 
     @Override
-    public List<String> getControlDependencies(OnnxProto3.NodeProto node) {
+    public List<String> getControlDependencies(Onnx.NodeProto node) {
         throw new UnsupportedOperationException("Not yet implemented");
     }
 
     @Override
     public void dumpBinaryProtoAsText(File inputFile, File outputFile) {
         try {
-            OnnxProto3.ModelProto graphDef = OnnxProto3.ModelProto.parseFrom(new BufferedInputStream(new FileInputStream(inputFile)));
+            Onnx.ModelProto graphDef = Onnx.ModelProto.parseFrom(new BufferedInputStream(new FileInputStream(inputFile)));
             BufferedWriter bufferedWriter = new BufferedWriter(new FileWriter(outputFile,true));
-            for(OnnxProto3.NodeProto node : graphDef.getGraph().getNodeList()) {
+            for(Onnx.NodeProto node : graphDef.getGraph().getNodeList()) {
                 bufferedWriter.write(node.toString());
             }
 
@@ -316,12 +316,12 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
 
 
     @Override
-    public Map<String,onnx.OnnxProto3.TypeProto.Tensor> variablesForGraph(OnnxProto3.GraphProto graphProto) {
+    public Map<String,onnx.Onnx.TypeProto.Tensor> variablesForGraph(Onnx.GraphProto graphProto) {
         /**
          * Need to figure out why
          * gpu_0/conv1_1 isn't present in VGG
          */
-        Map<String,onnx.OnnxProto3.TypeProto.Tensor> ret = new HashMap<>();
+        Map<String,onnx.Onnx.TypeProto.Tensor> ret = new HashMap<>();
         for(int i = 0; i < graphProto.getInputCount(); i++) {
             ret.put(graphProto.getInput(i).getName(),graphProto.getInput(i).getType().getTensorType());
         }
@@ -356,19 +356,19 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
     }
 
     @Override
-    public String translateToSameDiffName(String name, OnnxProto3.NodeProto node) {
+    public String translateToSameDiffName(String name, Onnx.NodeProto node) {
         return null;
     }
 
 
-    protected void addDummyTensor(String name, Map<String, OnnxProto3.TypeProto.Tensor> to) {
-        OnnxProto3.TensorShapeProto.Dimension dim =  OnnxProto3.TensorShapeProto.Dimension.
+    protected void addDummyTensor(String name, Map<String, Onnx.TypeProto.Tensor> to) {
+        Onnx.TensorShapeProto.Dimension dim =  Onnx.TensorShapeProto.Dimension.
                 newBuilder()
                 .setDimValue(-1)
                 .build();
-        OnnxProto3.TypeProto.Tensor typeProto = OnnxProto3.TypeProto.Tensor.newBuilder()
+        Onnx.TypeProto.Tensor typeProto = Onnx.TypeProto.Tensor.newBuilder()
                 .setShape(
-                        OnnxProto3.TensorShapeProto.newBuilder()
+                        Onnx.TensorShapeProto.newBuilder()
                                 .addDim(dim)
                                 .addDim(dim).build())
                 .build();
@@ -377,23 +377,23 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
 
     @Override
     public Message.Builder getNewGraphBuilder() {
-        return OnnxProto3.GraphProto.newBuilder();
+        return Onnx.GraphProto.newBuilder();
     }
 
     @Override
-    public OnnxProto3.GraphProto parseGraphFrom(byte[] inputStream) throws IOException {
-        return OnnxProto3.ModelProto.parseFrom(inputStream).getGraph();
+    public Onnx.GraphProto parseGraphFrom(byte[] inputStream) throws IOException {
+        return Onnx.ModelProto.parseFrom(inputStream).getGraph();
     }
 
     @Override
-    public OnnxProto3.GraphProto parseGraphFrom(InputStream inputStream) throws IOException {
-        return OnnxProto3.ModelProto.parseFrom(inputStream).getGraph();
+    public Onnx.GraphProto parseGraphFrom(InputStream inputStream) throws IOException {
+        return Onnx.ModelProto.parseFrom(inputStream).getGraph();
     }
 
     @Override
-    public void mapNodeType(OnnxProto3.NodeProto tfNode, ImportState<OnnxProto3.GraphProto, OnnxProto3.TypeProto.Tensor> importState,
-                            OpImportOverride<OnnxProto3.GraphProto, OnnxProto3.NodeProto, OnnxProto3.AttributeProto> opImportOverride,
-                            OpImportFilter<OnnxProto3.GraphProto, OnnxProto3.NodeProto, OnnxProto3.AttributeProto> opFilter) {
+    public void mapNodeType(Onnx.NodeProto tfNode, ImportState<Onnx.GraphProto, Onnx.TypeProto.Tensor> importState,
+                            OpImportOverride<Onnx.GraphProto, Onnx.NodeProto, Onnx.AttributeProto> opImportOverride,
+                            OpImportFilter<Onnx.GraphProto, Onnx.NodeProto, Onnx.AttributeProto> opFilter) {
         val differentialFunction = DifferentialFunctionClassHolder.getInstance().getOpWithOnnxName(tfNode.getOpType());
         if(differentialFunction == null) {
             throw new NoOpNameFoundException("No op name found " + tfNode.getOpType());
@@ -425,13 +425,13 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
 
 
     @Override
-    public DataType dataTypeForTensor(OnnxProto3.TypeProto.Tensor tensorProto, int outputNum) {
+    public DataType dataTypeForTensor(Onnx.TypeProto.Tensor tensorProto, int outputNum) {
        return nd4jTypeFromOnnxType(tensorProto.getElemType());
     }
 
     @Override
-    public boolean isStringType(OnnxProto3.TypeProto.Tensor tensor) {
-        return tensor.getElemType() == OnnxProto3.TensorProto.DataType.STRING;
+    public boolean isStringType(Onnx.TypeProto.Tensor tensor) {
+        return tensor.getElemType() == Onnx.TensorProto.DataType.STRING;
     }
 
 
@@ -440,7 +440,7 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
      * @param dataType the data type to convert
      * @return the nd4j type for the onnx type
      */
-    public DataType nd4jTypeFromOnnxType(OnnxProto3.TensorProto.DataType dataType) {
+    public DataType nd4jTypeFromOnnxType(Onnx.TensorProto.DataType dataType) {
         switch (dataType) {
             case DOUBLE: return DataType.DOUBLE;
             case FLOAT: return DataType.FLOAT;
@@ -452,8 +452,8 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
     }
 
     @Override
-    public String getAttrValueFromNode(OnnxProto3.NodeProto nodeProto, String key) {
-        for(OnnxProto3.AttributeProto attributeProto : nodeProto.getAttributeList()) {
+    public String getAttrValueFromNode(Onnx.NodeProto nodeProto, String key) {
+        for(Onnx.AttributeProto attributeProto : nodeProto.getAttributeList()) {
             if(attributeProto.getName().equals(key)) {
                 return attributeProto.getS().toString();
             }
@@ -463,29 +463,29 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
     }
 
     @Override
-    public long[] getShapeFromAttribute(OnnxProto3.AttributeProto attributeProto) {
+    public long[] getShapeFromAttribute(Onnx.AttributeProto attributeProto) {
         return Longs.toArray(attributeProto.getT().getDimsList());
     }
 
     @Override
-    public boolean isPlaceHolder(OnnxProto3.TypeProto.Tensor nodeType) {
+    public boolean isPlaceHolder(Onnx.TypeProto.Tensor nodeType) {
         return false;
     }
 
     @Override
-    public boolean isConstant(OnnxProto3.TypeProto.Tensor nodeType) {
+    public boolean isConstant(Onnx.TypeProto.Tensor nodeType) {
         return false;
     }
 
 
     @Override
-    public INDArray getNDArrayFromTensor(String tensorName, OnnxProto3.TypeProto.Tensor tensorProto, OnnxProto3.GraphProto graph) {
+    public INDArray getNDArrayFromTensor(String tensorName, Onnx.TypeProto.Tensor tensorProto, Onnx.GraphProto graph) {
         DataType type = dataTypeForTensor(tensorProto, 0);
         if(!tensorProto.isInitialized()) {
             throw new ND4JIllegalStateException("Unable to retrieve ndarray. Tensor was not initialized");
         }
 
-        OnnxProto3.TensorProto tensor = null;
+        Onnx.TensorProto tensor = null;
         for(int i = 0; i < graph.getInitializerCount(); i++) {
             val initializer = graph.getInitializer(i);
             if(initializer.getName().equals(tensorName)) {
@@ -508,7 +508,7 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
         return arr;
     }
 
-    public INDArray mapTensorProto(OnnxProto3.TensorProto tensor) {
+    public INDArray mapTensorProto(Onnx.TensorProto tensor) {
         if(tensor == null)
             return null;
 
@@ -527,7 +527,7 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
     }
 
     @Override
-    public long[] getShapeFromTensor(onnx.OnnxProto3.TypeProto.Tensor tensorProto) {
+    public long[] getShapeFromTensor(onnx.Onnx.TypeProto.Tensor tensorProto) {
         val ret = new long[Math.max(2,tensorProto.getShape().getDimCount())];
         int dimCount = tensorProto.getShape().getDimCount();
         if(dimCount >= 2)
@@ -548,11 +548,11 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
 
     /**
      * Get the shape from a tensor proto.
-     * Note that this is different from {@link #getShapeFromTensor(OnnxProto3.TensorProto)}
+     * Note that this is different from {@link #getShapeFromTensor(Onnx.TensorProto)}
      * @param tensorProto the tensor to get the shape from
      * @return
      */
-    public long[] getShapeFromTensor(OnnxProto3.TensorProto tensorProto) {
+    public long[] getShapeFromTensor(Onnx.TensorProto tensorProto) {
         val ret = new long[Math.max(2,tensorProto.getDimsCount())];
         int dimCount = tensorProto.getDimsCount();
         if(dimCount >= 2)
@@ -577,74 +577,74 @@ public class OnnxGraphMapper extends BaseGraphMapper<OnnxProto3.GraphProto, Onnx
 
 
     @Override
-    public String getInputFromNode(OnnxProto3.NodeProto node, int index) {
+    public String getInputFromNode(Onnx.NodeProto node, int index) {
         return node.getInput(index);
     }
 
     @Override
-    public int numInputsFor(OnnxProto3.NodeProto nodeProto) {
+    public int numInputsFor(Onnx.NodeProto nodeProto) {
         return nodeProto.getInputCount();
     }
 
 
     @Override
-    public long[] getShapeFromAttr(OnnxProto3.AttributeProto attr) {
+    public long[] getShapeFromAttr(Onnx.AttributeProto attr) {
         return Longs.toArray(attr.getT().getDimsList());
     }
 
     @Override
-    public Map<String, OnnxProto3.AttributeProto> getAttrMap(OnnxProto3.NodeProto nodeProto) {
-        Map<String,OnnxProto3.AttributeProto> proto = new HashMap<>();
+    public Map<String, Onnx.AttributeProto> getAttrMap(Onnx.NodeProto nodeProto) {
+        Map<String,Onnx.AttributeProto> proto = new HashMap<>();
         for(int i = 0; i < nodeProto.getAttributeCount(); i++) {
-            OnnxProto3.AttributeProto attributeProto = nodeProto.getAttribute(i);
+            Onnx.AttributeProto attributeProto = nodeProto.getAttribute(i);
             proto.put(attributeProto.getName(),attributeProto);
         }
         return proto;
     }
 
     @Override
-    public String getName(OnnxProto3.NodeProto nodeProto) {
+    public String getName(Onnx.NodeProto nodeProto) {
         return nodeProto.getName();
     }
 
     @Override
-    public boolean alreadySeen(OnnxProto3.NodeProto nodeProto) {
+    public boolean alreadySeen(Onnx.NodeProto nodeProto) {
         return false;
     }
 
     @Override
-    public boolean isVariableNode(OnnxProto3.NodeProto nodeProto) {
+    public boolean isVariableNode(Onnx.NodeProto nodeProto) {
         return nodeProto.getOpType().contains("Var");
     }
 
     @Override
-    public boolean shouldSkip(OnnxProto3.NodeProto opType) {
+    public boolean shouldSkip(Onnx.NodeProto opType) {
         return false;
     }
 
     @Override
-    public boolean hasShape(OnnxProto3.NodeProto nodeProto) {
+    public boolean hasShape(Onnx.NodeProto nodeProto) {
         return false;
     }
 
     @Override
-    public long[] getShape(OnnxProto3.NodeProto nodeProto) {
+    public long[] getShape(Onnx.NodeProto nodeProto) {
         return null;
     }
 
     @Override
-    public INDArray getArrayFrom(OnnxProto3.NodeProto nodeProto, OnnxProto3.GraphProto graph) {
+    public INDArray getArrayFrom(Onnx.NodeProto nodeProto, Onnx.GraphProto graph) {
 
         return null;
     }
 
     @Override
-    public String getOpType(OnnxProto3.NodeProto nodeProto) {
+    public String getOpType(Onnx.NodeProto nodeProto) {
         return nodeProto.getOpType();
     }
 
     @Override
-    public List<OnnxProto3.NodeProto> getNodeList(OnnxProto3.GraphProto graphProto) {
+    public List<Onnx.NodeProto> getNodeList(Onnx.GraphProto graphProto) {
         return graphProto.getNodeList();
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/TFGraphMapper.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/TFGraphMapper.java
index 5579569c3..f57fef4c7 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/TFGraphMapper.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/TFGraphMapper.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.imports.graphmapper.tf;
 
-import com.github.os72.protobuf351.Message;
+import org.nd4j.shade.protobuf.Message;
 import com.google.common.primitives.Floats;
 import com.google.common.primitives.Ints;
 import lombok.extern.slf4j.Slf4j;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/tensors/TFTensorMappers.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/tensors/TFTensorMappers.java
index 722168541..e9a99c6c0 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/tensors/TFTensorMappers.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/tensors/TFTensorMappers.java
@@ -1,6 +1,6 @@
 package org.nd4j.imports.graphmapper.tf.tensors;
 
-import com.github.os72.protobuf351.Descriptors;
+import org.nd4j.shade.protobuf.Descriptors;
 import org.bytedeco.javacpp.indexer.Bfloat16ArrayIndexer;
 import org.bytedeco.javacpp.indexer.HalfIndexer;
 import org.nd4j.linalg.api.buffer.DataType;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseBroadcastBoolOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseBroadcastBoolOp.java
index a844b04c7..a41dc8790 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseBroadcastBoolOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseBroadcastBoolOp.java
@@ -19,7 +19,7 @@ package org.nd4j.linalg.api.ops;
 import lombok.NoArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -205,7 +205,7 @@ public abstract class BaseBroadcastBoolOp extends BaseOp implements BroadcastOp
 
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
 
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseBroadcastOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseBroadcastOp.java
index d65ff377e..7f0d7e40c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseBroadcastOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseBroadcastOp.java
@@ -19,7 +19,7 @@ package org.nd4j.linalg.api.ops;
 import lombok.NoArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -200,7 +200,7 @@ public abstract class BaseBroadcastOp extends BaseOp implements BroadcastOp {
 
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
 
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseOp.java
index 925a5924f..8c9cdf4e0 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseOp.java
@@ -20,7 +20,7 @@ import lombok.Data;
 import lombok.Getter;
 import lombok.Setter;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
@@ -134,7 +134,7 @@ public abstract class BaseOp extends DifferentialFunction implements Op {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
     }
 
     @Override
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseReduceOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseReduceOp.java
index ebf9b9c18..7fc0679db 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseReduceOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseReduceOp.java
@@ -21,7 +21,7 @@ import lombok.Getter;
 import lombok.Setter;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.imports.graphmapper.onnx.OnnxGraphMapper;
@@ -218,7 +218,7 @@ public abstract class BaseReduceOp extends BaseOp implements ReduceOp {
 
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         if (!attributesForNode.containsKey("axes")) {
             this.dimensions = new int[] { Integer.MAX_VALUE };
         }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java
index 9b5b190c1..f52450eee 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java
@@ -21,7 +21,7 @@ import com.google.common.primitives.Doubles;
 import com.google.common.primitives.Longs;
 import lombok.*;
 import lombok.extern.slf4j.Slf4j;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
@@ -603,7 +603,7 @@ public class DynamicCustomOp extends DifferentialFunction implements CustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
 
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/NoOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/NoOp.java
index 19d8fe987..6b174bd07 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/NoOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/NoOp.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.linalg.api.buffer.DataType;
@@ -61,7 +61,7 @@ public class NoOp extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
 
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/If.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/If.java
index 6e0db97a5..03dc26313 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/If.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/If.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.controlflow;
 
 import lombok.*;
 import lombok.extern.slf4j.Slf4j;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
@@ -367,7 +367,7 @@ public class If extends DifferentialFunction implements CustomOp {
 
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
 
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/While.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/While.java
index eba0e1145..e26b0ea5f 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/While.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/While.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.controlflow;
 
 import lombok.*;
 import lombok.extern.slf4j.Slf4j;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
@@ -468,7 +468,7 @@ public class While extends DifferentialFunction implements CustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
 
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/ExternalErrorsFunction.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/ExternalErrorsFunction.java
index 378fbb06b..fd2134aad 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/ExternalErrorsFunction.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/ExternalErrorsFunction.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.layers;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
@@ -122,7 +122,7 @@ public class ExternalErrorsFunction extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
 
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/Linear.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/Linear.java
index da2c26f54..27f357b4b 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/Linear.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/Linear.java
@@ -19,7 +19,7 @@ package org.nd4j.linalg.api.ops.impl.layers;
 import lombok.Builder;
 import lombok.NoArgsConstructor;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
@@ -96,7 +96,7 @@ public class Linear extends BaseModule {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
 
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/AvgPooling2D.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/AvgPooling2D.java
index 3198a6a56..ac13c6224 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/AvgPooling2D.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/AvgPooling2D.java
@@ -21,7 +21,7 @@ import lombok.Getter;
 import lombok.NoArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -260,7 +260,7 @@ public class AvgPooling2D extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         val paddingVal = !attributesForNode.containsKey("auto_pad") ? "VALID" : attributesForNode.get("auto_pad").getS().toStringUtf8();
         val kernelShape = attributesForNode.get("kernel_shape").getIntsList();
         val padding = !attributesForNode.containsKey("pads") ? Arrays.asList(1L) : attributesForNode.get("pads").getIntsList();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/AvgPooling3D.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/AvgPooling3D.java
index 2c57c68de..6f58884f0 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/AvgPooling3D.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/AvgPooling3D.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.layers.convolution;
 
 import lombok.Getter;
 import lombok.extern.slf4j.Slf4j;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -78,7 +78,7 @@ public class AvgPooling3D extends Pooling3D {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         throw new UnsupportedOperationException("Not yet implemented");
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/BatchNorm.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/BatchNorm.java
index 67fc9f3a5..bad975cb5 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/BatchNorm.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/BatchNorm.java
@@ -21,7 +21,7 @@ import lombok.Getter;
 import lombok.NoArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.autodiff.samediff.internal.SameDiffOp;
@@ -139,7 +139,7 @@ public class BatchNorm extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         OnnxGraphMapper.getInstance().initFunctionFromProperties(node.getOpType(), this, attributesForNode, node, graph);
         addArgs();
     }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv1D.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv1D.java
index 3d61de716..5ae2ac144 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv1D.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv1D.java
@@ -21,7 +21,7 @@ import lombok.Getter;
 import lombok.NoArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv2D.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv2D.java
index 4335f4561..04db5874c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv2D.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv2D.java
@@ -21,7 +21,7 @@ import lombok.Getter;
 import lombok.NoArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -127,7 +127,7 @@ public class Conv2D extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         OnnxGraphMapper.getInstance().initFunctionFromProperties(node.getOpType(), this, attributesForNode, node, graph);
         addArgs();
     }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/DeConv2D.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/DeConv2D.java
index 6cba853d0..65c0fccc3 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/DeConv2D.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/DeConv2D.java
@@ -21,7 +21,7 @@ import lombok.Getter;
 import lombok.NoArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -247,7 +247,7 @@ public class DeConv2D extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         val autoPad = !attributesForNode.containsKey("auto_pad") ? "VALID" : attributesForNode.get("auto_pad").getS().toStringUtf8();
         val dilations = attributesForNode.get("dilations");
         val dilationY = dilations == null ? 1 : dilations.getIntsList().get(0).intValue();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/DepthwiseConv2D.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/DepthwiseConv2D.java
index 0ea84e081..92a39f188 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/DepthwiseConv2D.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/DepthwiseConv2D.java
@@ -20,7 +20,7 @@ import lombok.Builder;
 import lombok.Getter;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -151,7 +151,7 @@ public class DepthwiseConv2D extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         OnnxGraphMapper.getInstance().initFunctionFromProperties(node.getOpType(), this, attributesForNode, node, graph);
         addArgs();
     }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/LocalResponseNormalization.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/LocalResponseNormalization.java
index de4e763bc..421598d13 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/LocalResponseNormalization.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/LocalResponseNormalization.java
@@ -21,7 +21,7 @@ import lombok.Getter;
 import lombok.NoArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -115,7 +115,7 @@ public class LocalResponseNormalization extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         val aAlpha = attributesForNode.get("alpha");
         val aBeta = attributesForNode.get("beta");
         val aBias = attributesForNode.get("bias");
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/MaxPooling2D.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/MaxPooling2D.java
index f996fc29f..b321334a5 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/MaxPooling2D.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/MaxPooling2D.java
@@ -21,7 +21,7 @@ import lombok.Getter;
 import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -221,7 +221,7 @@ public class MaxPooling2D extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         val paddingVal = !attributesForNode.containsKey("auto_pad") ? "VALID" : attributesForNode.get("auto_pad").getS().toStringUtf8();
         val isSameNode = paddingVal.equals("SAME");
         val kernelShape = attributesForNode.get("kernel_shape").getIntsList();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/MaxPooling3D.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/MaxPooling3D.java
index a243dec9b..99d73d2af 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/MaxPooling3D.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/MaxPooling3D.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.layers.convolution;
 
 import lombok.Getter;
 import lombok.extern.slf4j.Slf4j;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -78,7 +78,7 @@ public class MaxPooling3D extends Pooling3D {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         throw new UnsupportedOperationException("Not yet implemented");
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Pooling2D.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Pooling2D.java
index f7f21e78d..c45d106e7 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Pooling2D.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Pooling2D.java
@@ -20,7 +20,7 @@ import lombok.Builder;
 import lombok.Getter;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -183,7 +183,7 @@ public class Pooling2D extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         val isSameNode = attributesForNode.get("auto_pad").getS().equals("SAME");
         val kernelShape = attributesForNode.get("kernel_shape").getIntsList();
         val padding = attributesForNode.get("pads").getIntsList();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/GRUCell.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/GRUCell.java
index 678a4afef..6c7daca69 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/GRUCell.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/GRUCell.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.layers.recurrent;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/LSTMCell.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/LSTMCell.java
index 1fdd6b191..e9d2ffd3b 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/LSTMCell.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/LSTMCell.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.layers.recurrent;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.linalg.api.ops.DynamicCustomOp;
 import org.nd4j.linalg.api.ops.impl.layers.recurrent.config.LSTMCellConfiguration;
@@ -73,7 +73,7 @@ public class LSTMCell extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
 
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/SRU.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/SRU.java
index aaac14131..b916d4961 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/SRU.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/SRU.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.layers.recurrent;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.imports.NoOpNameFoundException;
 import org.nd4j.linalg.api.ops.DynamicCustomOp;
@@ -65,7 +65,7 @@ public class SRU extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/SRUCell.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/SRUCell.java
index 625e09e91..4880b90fe 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/SRUCell.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/recurrent/SRUCell.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.layers.recurrent;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.imports.NoOpNameFoundException;
 import org.nd4j.linalg.api.ops.DynamicCustomOp;
@@ -66,7 +66,7 @@ public class SRUCell extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/Mmul.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/Mmul.java
index 00cad1f88..7d711ca58 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/Mmul.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/Mmul.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.reduce;
 
 import lombok.EqualsAndHashCode;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -204,7 +204,7 @@ public class Mmul extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         val isTransposeA = !attributesForNode.containsKey("transA") ? false : attributesForNode.get("transA").getI() > 0;
         val isTransposeB = !attributesForNode.containsKey("transB") ? false : attributesForNode.get("transB").getI() > 0;
         MMulTranspose mMulTranspose = MMulTranspose.builder()
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/TensorMmul.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/TensorMmul.java
index 62e373832..3de44537a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/TensorMmul.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/TensorMmul.java
@@ -20,7 +20,7 @@ import com.google.common.primitives.Ints;
 import com.google.common.primitives.Longs;
 import lombok.NoArgsConstructor;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.linalg.api.blas.params.MMulTranspose;
@@ -283,7 +283,7 @@ public class TensorMmul extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         val isTransposeA = !attributesForNode.containsKey("transA") ? false : attributesForNode.get("transA").getI() > 0;
         val isTransposeB = !attributesForNode.containsKey("transB") ? false : attributesForNode.get("transB").getI() > 0;
         MMulTranspose mMulTranspose = MMulTranspose.builder()
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Concat.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Concat.java
index e0b0450d3..5c6beb945 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Concat.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Concat.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -163,7 +163,7 @@ public class Concat extends DynamicCustomOp {
 
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Diag.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Diag.java
index 90aed14bf..b6d08784b 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Diag.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Diag.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -77,7 +77,7 @@ public class Diag extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/DiagPart.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/DiagPart.java
index d2807e36d..6b1688602 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/DiagPart.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/DiagPart.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -79,7 +79,7 @@ public class DiagPart extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Gather.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Gather.java
index 1782f75df..31718d337 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Gather.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Gather.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.NoArgsConstructor;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.imports.descriptors.properties.PropertyMapping;
@@ -78,7 +78,7 @@ public class Gather extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         OnnxGraphMapper.getInstance().initFunctionFromProperties(node.getOpType(), this, attributesForNode, node, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/GatherNd.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/GatherNd.java
index 1be7c56bf..cfe4fe8be 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/GatherNd.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/GatherNd.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.NoArgsConstructor;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeAvg.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeAvg.java
index d8319cab2..ec86c6553 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeAvg.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeAvg.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.extern.slf4j.Slf4j;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -65,7 +65,7 @@ public class MergeAvg extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeMax.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeMax.java
index c9118990c..046f06c3c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeMax.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeMax.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -64,7 +64,7 @@ public class MergeMax extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeSum.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeSum.java
index b7c370615..6b87ca5c8 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeSum.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/MergeSum.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -66,7 +66,7 @@ public class MergeSum extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ParallelStack.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ParallelStack.java
index 8d7dcf6a6..1856e6804 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ParallelStack.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ParallelStack.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -68,7 +68,7 @@ public class ParallelStack extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         throw new UnsupportedOperationException("No analog found for onnx for " + opName());
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Rank.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Rank.java
index 96f28dbf1..aacfa19e1 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Rank.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Rank.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -66,7 +66,7 @@ public class Rank extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
 
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Repeat.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Repeat.java
index 14d67d912..02f8f9445 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Repeat.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Repeat.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.NoArgsConstructor;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -106,7 +106,7 @@ public class Repeat extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Reshape.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Reshape.java
index 42e401859..b30bacc22 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Reshape.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Reshape.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -126,7 +126,7 @@ public class Reshape extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         val shape = new OnnxGraphMapper().getShape(node);
         this.shape = shape;
     }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/SequenceMask.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/SequenceMask.java
index 5faa82609..a2f6bd208 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/SequenceMask.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/SequenceMask.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.NoArgsConstructor;
 import lombok.val;
-import onnx.OnnxMlProto3;
+import onnx.OnnxMl;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Shape.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Shape.java
index a1133ee82..6cd2eec06 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Shape.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Shape.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.autodiff.samediff.serde.FlatBuffersMapper;
@@ -87,7 +87,7 @@ public class Shape extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         throw new NoOpNameFoundException("No onnx name found for shape " + opName());
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ShapeN.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ShapeN.java
index 241cc950f..55d9dd806 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ShapeN.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ShapeN.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Size.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Size.java
index 1ba9156bc..71b52a92a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Size.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Size.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Stack.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Stack.java
index 44cb0539c..6cd09f9bd 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Stack.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Stack.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -93,7 +93,7 @@ public class Stack extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         throw new UnsupportedOperationException("No analog found for onnx for " + opName());
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java
index 965d071c3..2de0a29c5 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape;
 
 import com.google.common.primitives.Ints;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.autodiff.samediff.VariableType;
@@ -156,7 +156,7 @@ public class Transpose extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         if (!attributesForNode.containsKey("perm")) {
 
         } else
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Unstack.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Unstack.java
index 3d7e07a72..9dd6b6338 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Unstack.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Unstack.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.api.ops.impl.shape;
 
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -127,7 +127,7 @@ public class Unstack extends DynamicCustomOp {
 
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         throw new UnsupportedOperationException("No analog found for onnx for " + opName());
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/bp/ConcatBp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/bp/ConcatBp.java
index ead0f2747..70bc1b087 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/bp/ConcatBp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/bp/ConcatBp.java
@@ -18,7 +18,7 @@ package org.nd4j.linalg.api.ops.impl.shape.bp;
 
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -71,7 +71,7 @@ public class ConcatBp extends DynamicCustomOp {
 
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         //No op
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayConcat.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayConcat.java
index 07bdab586..7759f96dd 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayConcat.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayConcat.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape.tensorops;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.imports.NoOpNameFoundException;
@@ -59,7 +59,7 @@ public class TensorArrayConcat extends BaseTensorOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         throw new UnsupportedOperationException();
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayGather.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayGather.java
index 3ab0d91c9..9e7669725 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayGather.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayGather.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape.tensorops;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.imports.NoOpNameFoundException;
@@ -59,7 +59,7 @@ public class TensorArrayGather extends BaseTensorOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         throw new UnsupportedOperationException();
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayRead.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayRead.java
index 619216813..6d8cff91c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayRead.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayRead.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape.tensorops;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
@@ -54,7 +54,7 @@ public class TensorArrayRead extends BaseTensorOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
     }
 
     @Override
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayScatter.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayScatter.java
index add288d89..9e1d93e2f 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayScatter.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArrayScatter.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape.tensorops;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.linalg.api.buffer.DataType;
@@ -52,7 +52,7 @@ public class TensorArrayScatter extends BaseTensorOp {
 
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
     }
 
     @Override
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySize.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySize.java
index 9734515d3..276dadcab 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySize.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySize.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape.tensorops;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.imports.descriptors.properties.PropertyMapping;
 import org.nd4j.linalg.api.buffer.DataType;
@@ -58,7 +58,7 @@ public class TensorArraySize extends BaseTensorOp {
    }
 
    @Override
-   public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+   public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
    }
 
    @Override
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySplit.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySplit.java
index fb52c78a7..589805641 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySplit.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySplit.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape.tensorops;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.linalg.api.buffer.DataType;
@@ -52,7 +52,7 @@ public class TensorArraySplit extends BaseTensorOp {
 
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
     }
 
     @Override
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/clip/ClipByNorm.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/clip/ClipByNorm.java
index 1cda0257d..59b7ec2f5 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/clip/ClipByNorm.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/clip/ClipByNorm.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.transforms.clip;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -64,7 +64,7 @@ public class ClipByNorm extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         throw new UnsupportedOperationException("Not yet implemented");
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/clip/ClipByValue.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/clip/ClipByValue.java
index d25b0df62..11d3e9004 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/clip/ClipByValue.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/clip/ClipByValue.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.transforms.clip;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -77,7 +77,7 @@ public class ClipByValue extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         throw new UnsupportedOperationException("Not yet implemented");
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Assign.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Assign.java
index 7ca0b342a..35c209870 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Assign.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Assign.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.transforms.custom;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -62,7 +62,7 @@ public class Assign extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CumProd.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CumProd.java
index d1d0176ef..9c04aeb12 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CumProd.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CumProd.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.api.ops.impl.transforms.custom;
 
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -132,7 +132,7 @@ public class CumProd extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CumSum.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CumSum.java
index 2b62b73cf..b8c7d5c51 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CumSum.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CumSum.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.api.ops.impl.transforms.custom;
 
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 
@@ -133,7 +133,7 @@ public class CumSum extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Fill.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Fill.java
index db95ee728..af4097870 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Fill.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Fill.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.api.ops.impl.transforms.custom;
 
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -80,7 +80,7 @@ public class Fill extends DynamicCustomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/strict/RectifiedTanh.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/strict/RectifiedTanh.java
index 4bd56ea4d..da439cec7 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/strict/RectifiedTanh.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/strict/RectifiedTanh.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.transforms.strict;
 
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
@@ -81,7 +81,7 @@ public class RectifiedTanh extends BaseTransformStrictOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/impl/DropOutInverted.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/impl/DropOutInverted.java
index bb4b86f12..6b174ae63 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/impl/DropOutInverted.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/impl/DropOutInverted.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.api.ops.random.impl;
 
 import lombok.NonNull;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.linalg.api.ndarray.INDArray;
@@ -75,7 +75,7 @@ public class DropOutInverted extends BaseRandomOp {
     }
 
     @Override
-    public void initFromOnnx(OnnxProto3.NodeProto node, SameDiff initWith, Map<String, OnnxProto3.AttributeProto> attributesForNode, OnnxProto3.GraphProto graph) {
+    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/impl/Range.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/impl/Range.java
index 27e9d9f3c..c3670b52f 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/impl/Range.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/impl/Range.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.api.ops.random.impl;
 
 import lombok.val;
-import onnx.OnnxProto3;
+import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx-ml.proto3 b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx-ml.proto
similarity index 100%
rename from nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx-ml.proto3
rename to nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx-ml.proto
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx-operators.proto3 b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx-operators.proto
similarity index 99%
rename from nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx-operators.proto3
rename to nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx-operators.proto
index a8db3ca23..48890a516 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx-operators.proto3
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx-operators.proto
@@ -9,7 +9,7 @@
 syntax = "proto3";
 
 package onnx;
-import "onnx.proto3";
+import "onnx.proto";
 
 //
 // This file contains the proto definitions for OperatorSetProto and
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx.proto3 b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx.proto
similarity index 100%
rename from nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx.proto3
rename to nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/protobuf/onnx/onnx.proto
diff --git a/nd4j/nd4j-backends/nd4j-tests-tensorflow/src/test/cpujava/org/nd4j/tensorflow/conversion/GraphRunnerTest.java b/nd4j/nd4j-backends/nd4j-tests-tensorflow/src/test/cpujava/org/nd4j/tensorflow/conversion/GraphRunnerTest.java
index af1af3a75..ee188605d 100644
--- a/nd4j/nd4j-backends/nd4j-tests-tensorflow/src/test/cpujava/org/nd4j/tensorflow/conversion/GraphRunnerTest.java
+++ b/nd4j/nd4j-backends/nd4j-tests-tensorflow/src/test/cpujava/org/nd4j/tensorflow/conversion/GraphRunnerTest.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.tensorflow.conversion;
 
-import com.github.os72.protobuf351.util.JsonFormat;
+import org.nd4j.shade.protobuf.util.JsonFormat;
 import org.apache.commons.io.IOUtils;
 import org.junit.Ignore;
 import org.junit.Rule;
diff --git a/nd4j/nd4j-backends/nd4j-tests-tensorflow/src/test/gpujava/org/nd4j/tensorflow/conversion/GpuGraphRunnerTest.java b/nd4j/nd4j-backends/nd4j-tests-tensorflow/src/test/gpujava/org/nd4j/tensorflow/conversion/GpuGraphRunnerTest.java
index accde5b1b..1ecc0e39a 100644
--- a/nd4j/nd4j-backends/nd4j-tests-tensorflow/src/test/gpujava/org/nd4j/tensorflow/conversion/GpuGraphRunnerTest.java
+++ b/nd4j/nd4j-backends/nd4j-tests-tensorflow/src/test/gpujava/org/nd4j/tensorflow/conversion/GpuGraphRunnerTest.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.tensorflow.conversion;
 
-import com.github.os72.protobuf351.util.JsonFormat;
+import org.nd4j.shade.protobuf.util.JsonFormat;
 import org.apache.commons.io.IOUtils;
 import org.junit.Ignore;
 import org.junit.Test;
diff --git a/nd4j/nd4j-shade/pom.xml b/nd4j/nd4j-shade/pom.xml
index 4a2c4ca1b..36b58087b 100644
--- a/nd4j/nd4j-shade/pom.xml
+++ b/nd4j/nd4j-shade/pom.xml
@@ -29,6 +29,7 @@
     <packaging>pom</packaging>
     <modules>
         <module>jackson</module>
+        <module>protobuf</module>
     </modules>
 
     <properties>
diff --git a/nd4j/nd4j-shade/protobuf/pom.xml b/nd4j/nd4j-shade/protobuf/pom.xml
new file mode 100644
index 000000000..1cbd7d5a8
--- /dev/null
+++ b/nd4j/nd4j-shade/protobuf/pom.xml
@@ -0,0 +1,228 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project xmlns="http://maven.apache.org/POM/4.0.0"
+         xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
+         xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
+    <parent>
+        <artifactId>nd4j-shade</artifactId>
+        <groupId>org.nd4j</groupId>
+        <version>1.0.0-SNAPSHOT</version>
+    </parent>
+    <modelVersion>4.0.0</modelVersion>
+
+    <artifactId>protobuf</artifactId>
+
+    <properties>
+        <skipTestResourceEnforcement>true</skipTestResourceEnforcement>
+    </properties>
+
+    <dependencies>
+        <dependency>
+            <groupId>com.google.protobuf</groupId>
+            <artifactId>protobuf-java</artifactId>
+            <version>3.8.0</version>
+        </dependency>
+        <dependency>
+            <groupId>com.google.protobuf</groupId>
+            <artifactId>protobuf-java-util</artifactId>
+            <version>3.8.0</version>
+        </dependency>
+    </dependencies>
+
+
+    <profiles>
+        <profile>
+            <id>custom-lifecycle</id>
+
+            <activation>
+                <property><name>!skip.custom.lifecycle</name></property>
+            </activation>
+            <build>
+                <plugins>
+
+                    <plugin>
+                        <groupId>org.apache.portals.jetspeed-2</groupId>
+                        <artifactId>jetspeed-mvn-maven-plugin</artifactId>
+                        <version>2.3.1</version>
+                        <executions>
+                            <execution>
+                                <id>compile-and-pack</id>
+                                <phase>compile</phase>
+                                <goals>
+                                    <goal>mvn</goal>
+                                </goals>
+                            </execution>
+                        </executions>
+                        <dependencies>
+                            <dependency>
+                                <groupId>org.apache.maven.shared</groupId>
+                                <artifactId>maven-invoker</artifactId>
+                                <version>2.2</version>
+                            </dependency>
+                        </dependencies>
+                        <configuration>
+                            <targets combine.children="merge">
+
+                                <target>
+                                    <id>create-shaded-jars</id>
+                                    <dir>@rootdir@/nd4j/nd4j-shade/protobuf/</dir>
+                                    <goals>clean,compile,package</goals>
+                                    <properties>
+                                        <skip.custom.lifecycle>true</skip.custom.lifecycle>
+                                    </properties>
+                                </target>
+
+                            </targets>
+                            <defaultTarget>create-shaded-jars</defaultTarget>
+                        </configuration>
+                    </plugin>
+
+                </plugins>
+            </build>
+        </profile>
+    </profiles>
+
+    <build>
+        <plugins>
+
+            <!-- Disable Maven Lint plugin in this module. For some reason it chokes on this module (internal NPE) and we don't need it anyway here -->
+            <plugin>
+                <groupId>com.lewisd</groupId>
+                <artifactId>lint-maven-plugin</artifactId>
+                <version>0.0.11</version>
+                <executions>
+                    <execution>
+                        <id>pom-lint</id>
+                        <phase>none</phase>
+                    </execution>
+                </executions>
+            </plugin>
+
+
+            <!--
+            Use Maven Shade plugin to add a shaded version of the Protobuf dependencies, that can be imported by
+            including this module (org.nd4j.protobuf) as a dependency.
+            The standard com.google.protobuf dependencies will be provided, though are prefixed by org.nd4j.shade.protobuf
+            -->
+            <plugin>
+                <groupId>org.apache.maven.plugins</groupId>
+                <artifactId>maven-shade-plugin</artifactId>
+                <version>${maven-shade-plugin.version}</version>
+                <executions>
+                    <execution>
+                        <phase>package</phase>
+                        <goals>
+                            <goal>shade</goal>
+                        </goals>
+                        <configuration>
+                            <transformers>
+                                <transformer implementation="org.apache.maven.plugins.shade.resource.AppendingTransformer">
+                                    <resource>reference.conf</resource>
+                                </transformer>
+                                <transformer implementation="org.apache.maven.plugins.shade.resource.ServicesResourceTransformer"/>
+                                <transformer implementation="org.apache.maven.plugins.shade.resource.ManifestResourceTransformer">
+                                </transformer>
+                            </transformers>
+                        </configuration>
+                    </execution>
+                </executions>
+
+                <configuration>
+                    <!--
+                    Important configuration options here:
+                    createDependencyReducedPom: remove the shaded artifacts from the module dependencies. Without this, the
+                                                original dependencies will be shaded, AND still included as transitive deps
+                                                in the final POM. This is not what we want.
+                    shadedArtifactAttached:     If true, the shaded artifact will be a separate JAR file for install, with
+                                                the original un-shaded JAR being separate. With this being set to false,
+                                                the original JAR will be modified, and no extra jar will be produced.
+                    promoteTransitiveDependencies:  This will promote the transitive dependencies of the shaded dependencies
+                                                to direct dependencies. Without this, we need to manually manage the transitive
+                                                dependencies of the shaded artifacts.
+
+                    Note that using <optional>true</optional> in the dependencies also allows the deps to be shaded (and
+                    original dependencies to not be included), but does NOT work with promoteTransitiveDependencies
+                    -->
+                    <shadedArtifactAttached>false</shadedArtifactAttached>
+                    <createDependencyReducedPom>true</createDependencyReducedPom>
+                    <promoteTransitiveDependencies>true</promoteTransitiveDependencies>
+
+                    <artifactSet>
+                        <includes>
+                            <include>com.google.protobuf:*</include>
+                            <include>com.google.protobuf.*:*</include>
+                        </includes>
+                    </artifactSet>
+
+                    <relocations>
+                        <!-- Protobuf dependencies -->
+                        <relocation>
+                            <pattern>com.google.protobuf</pattern>
+                            <shadedPattern>org.nd4j.shade.protobuf</shadedPattern>
+                        </relocation>
+                    </relocations>
+
+                </configuration>
+            </plugin>
+            <plugin>
+                <groupId>org.apache.maven.plugins</groupId>
+                <artifactId>maven-jar-plugin</artifactId>
+                <configuration>
+                    <forceCreation>true</forceCreation>
+                </configuration>
+                <executions>
+                    <execution>
+                        <id>empty-javadoc-jar</id>
+                        <phase>package</phase>
+                        <goals>
+                            <goal>jar</goal>
+                        </goals>
+                        <configuration>
+                            <classifier>javadoc</classifier>
+                            <classesDirectory>${basedir}/javadoc</classesDirectory>
+                        </configuration>
+                    </execution>
+                    <execution>
+                        <id>empty-sources-jar</id>
+                        <phase>package</phase>
+                        <goals>
+                            <goal>jar</goal>
+                        </goals>
+                        <configuration>
+                            <classifier>sources</classifier>
+                            <classesDirectory>${basedir}/src</classesDirectory>
+                        </configuration>
+                    </execution>
+                </executions>
+            </plugin>
+
+            <plugin>
+                <groupId>org.apache.maven.plugins</groupId>
+                <artifactId>maven-dependency-plugin</artifactId>
+                <version>3.0.0</version>
+                <executions>
+                    <execution>
+                        <id>unpack</id>
+                        <phase>package</phase>
+                        <goals>
+                            <goal>unpack</goal>
+                        </goals>
+                        <configuration>
+                            <artifactItems>
+                                <artifactItem>
+                                    <groupId>org.nd4j</groupId>
+                                    <artifactId>protobuf</artifactId>
+                                    <version>${project.version}</version>
+                                    <type>jar</type>
+                                    <overWrite>false</overWrite>
+                                    <outputDirectory>${project.build.directory}/classes/</outputDirectory>
+                                    <includes>**/*.class,**/*.xml</includes>
+                                </artifactItem>
+                            </artifactItems>
+                        </configuration>
+                    </execution>
+                </executions>
+            </plugin>
+        </plugins>
+    </build>
+
+</project>
\ No newline at end of file
diff --git a/nd4j/nd4j-tensorflow/src/main/java/org/nd4j/tensorflow/conversion/TensorflowConversion.java b/nd4j/nd4j-tensorflow/src/main/java/org/nd4j/tensorflow/conversion/TensorflowConversion.java
index b47cd30d1..6eff18ecc 100644
--- a/nd4j/nd4j-tensorflow/src/main/java/org/nd4j/tensorflow/conversion/TensorflowConversion.java
+++ b/nd4j/nd4j-tensorflow/src/main/java/org/nd4j/tensorflow/conversion/TensorflowConversion.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.tensorflow.conversion;
 
-import com.github.os72.protobuf351.InvalidProtocolBufferException;
+import org.nd4j.shade.protobuf.InvalidProtocolBufferException;
 import org.bytedeco.javacpp.*;
 import org.bytedeco.javacpp.indexer.*;
 import org.nd4j.linalg.api.buffer.DataBuffer;
diff --git a/nd4j/nd4j-tensorflow/src/main/java/org/nd4j/tensorflow/conversion/graphrunner/GraphRunner.java b/nd4j/nd4j-tensorflow/src/main/java/org/nd4j/tensorflow/conversion/graphrunner/GraphRunner.java
index 633535197..79d45f781 100644
--- a/nd4j/nd4j-tensorflow/src/main/java/org/nd4j/tensorflow/conversion/graphrunner/GraphRunner.java
+++ b/nd4j/nd4j-tensorflow/src/main/java/org/nd4j/tensorflow/conversion/graphrunner/GraphRunner.java
@@ -16,9 +16,9 @@
 
 package org.nd4j.tensorflow.conversion.graphrunner;
 
-import com.github.os72.protobuf351.ByteString;
-import com.github.os72.protobuf351.InvalidProtocolBufferException;
-import com.github.os72.protobuf351.util.JsonFormat;
+import org.nd4j.shade.protobuf.ByteString;
+import org.nd4j.shade.protobuf.InvalidProtocolBufferException;
+import org.nd4j.shade.protobuf.util.JsonFormat;
 import lombok.Getter;
 import lombok.Setter;
 import lombok.extern.slf4j.Slf4j;
@@ -638,7 +638,7 @@ public class GraphRunner implements Closeable {
 
     /**
      * Convert a json string written out
-     * by {@link com.github.os72.protobuf351.util.JsonFormat}
+     * by {@link org.nd4j.shade.protobuf.util.JsonFormat}
      * to a {@link org.bytedeco.tensorflow.ConfigProto}
      * @param json the json to read
      * @return the config proto to use

From b091e972ef48affb5efd5337963f68f4788db592 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Sat, 24 Aug 2019 14:16:34 +0300
Subject: [PATCH 04/56] - string NDArray flat serde impl + tests (#163)

- string NDArray equalsTo impl

Signed-off-by: raver119 <raver119@gmail.com>
---
 libnd4j/blas/NDArray.hpp                      |  76 +++++++++----
 libnd4j/blas/cpu/GraphExecutioner.cpp         |  11 +-
 libnd4j/include/graph/FlatUtils.h             |   2 +
 libnd4j/include/graph/impl/FlatUtils.cpp      |  11 ++
 .../tests_cpu/layers_tests/FlatUtilsTests.cpp | 100 ++++++++++++++++++
 .../tests_cpu/layers_tests/StringTests.cpp    |   3 +-
 6 files changed, 172 insertions(+), 31 deletions(-)
 create mode 100644 libnd4j/tests_cpu/layers_tests/FlatUtilsTests.cpp

diff --git a/libnd4j/blas/NDArray.hpp b/libnd4j/blas/NDArray.hpp
index 1404afc96..fdbcae49f 100644
--- a/libnd4j/blas/NDArray.hpp
+++ b/libnd4j/blas/NDArray.hpp
@@ -476,19 +476,36 @@ std::vector<Nd4jLong> NDArray::getShapeInfoAsVector() {
 ////////////////////////////////////////////////////////////////////////
 std::vector<int8_t> NDArray::asByteVector() {
 
-    std::vector<int8_t> result((unsigned long long) this->lengthOf() * sizeOfT());
 
-    if (this->isView()) {
-        auto tmp = this->dup(this->ordering());
 
-        memcpy(result.data(), tmp->getBuffer(), (unsigned long long) lengthOf() * sizeOfT());
+    if (isS()) {
+        // string data type requires special treatment
+        syncToHost();
+        auto numWords = this->lengthOf();
+        auto offsetsBuffer = this->bufferAsT<Nd4jLong>();
+        auto headerLength = ShapeUtils::stringBufferHeaderRequirements(numWords);
+        auto dataLength = offsetsBuffer[numWords];
+        std::vector<int8_t> result(headerLength + dataLength);
 
-        delete tmp;
+        memcpy(result.data(), getBuffer(), headerLength + dataLength);
+
+        return result;
+    } else {
+        // all other types are linear
+        std::vector<int8_t> result((unsigned long long) this->lengthOf() * sizeOfT());
+
+        if (this->isView()) {
+            auto tmp = this->dup(this->ordering());
+            syncToHost();
+            memcpy(result.data(), tmp->getBuffer(), (unsigned long long) lengthOf() * sizeOfT());
+
+            delete tmp;
+        } else {
+            syncToHost();
+            memcpy(result.data(), getBuffer(), (unsigned long long) lengthOf() * sizeOfT());
+        }
+        return result;
     }
-    else {
-        memcpy(result.data(), getBuffer(), (unsigned long long) lengthOf() * sizeOfT());
-    }
-    return result;
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -1584,9 +1601,7 @@ std::string* NDArray::bufferAsT() const {
 //////////////////////////////////////////////////////////////////////////
 template <typename T>
 T* NDArray::bufferAsT() const {
-    if (isS())
-        throw std::runtime_error("You can't use this method on String array");
-
+    // FIXME: do we REALLY want sync here?
     syncToHost();
 
     return reinterpret_cast<T*>(getBuffer());
@@ -3202,20 +3217,39 @@ bool NDArray::equalsTo(const NDArray *other, double eps) const {
     } else if (!shape::equalsSoft(getShapeInfo(), other->getShapeInfo()))
         return false;
 
-    NDArray tmp(nd4j::DataType::FLOAT32, getContext()); // scalar = 0
+    if (isS()) {
+        // string is special case, we'll compare them one by one, considering both arrays are guaranteed to have the same length
+        for (int e = 0; e < this->lengthOf(); e++) {
+            auto s1 = this->e<std::string>(e);
+            auto s2 = other->e<std::string>(e);
 
-    ExtraArguments extras({eps});
+            if (s1 != s2)
+                return false;
+        }
 
-    NDArray::prepareSpecialUse({&tmp}, {this, other});
-    NativeOpExecutioner::execReduce3Scalar(getContext(), reduce3::EqualsWithEps, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), extras.argumentsAsT(DataType::FLOAT32), other->getBuffer(), other->getShapeInfo(), other->getSpecialBuffer(), other->getSpecialShapeInfo(), tmp.buffer(), tmp.shapeInfo(), tmp.specialBuffer(), tmp.specialShapeInfo());
-    NDArray::registerSpecialUse({&tmp}, {this, other});
+        return true;
+    } else {
+        // regular numeric types
+        NDArray tmp(nd4j::DataType::FLOAT32, getContext()); // scalar = 0
 
-    synchronize("NDArray::equalsTo");
+        ExtraArguments extras({eps});
 
-    if (tmp.e<int>(0) > 0)
-        return false;
+        NDArray::prepareSpecialUse({&tmp}, {this, other});
+        NativeOpExecutioner::execReduce3Scalar(getContext(), reduce3::EqualsWithEps, getBuffer(), getShapeInfo(),
+                                               getSpecialBuffer(), getSpecialShapeInfo(),
+                                               extras.argumentsAsT(DataType::FLOAT32), other->getBuffer(),
+                                               other->getShapeInfo(), other->getSpecialBuffer(),
+                                               other->getSpecialShapeInfo(), tmp.buffer(), tmp.shapeInfo(),
+                                               tmp.specialBuffer(), tmp.specialShapeInfo());
+        NDArray::registerSpecialUse({&tmp}, {this, other});
 
-    return true;
+        synchronize("NDArray::equalsTo");
+
+        if (tmp.e<int>(0) > 0)
+            return false;
+
+        return true;
+    }
 }
 
 //////////////////////////////////////////////////////////////////////////
diff --git a/libnd4j/blas/cpu/GraphExecutioner.cpp b/libnd4j/blas/cpu/GraphExecutioner.cpp
index b5e7d9bf2..6f97bc024 100644
--- a/libnd4j/blas/cpu/GraphExecutioner.cpp
+++ b/libnd4j/blas/cpu/GraphExecutioner.cpp
@@ -54,6 +54,7 @@
 #include <graph/ExecutionResult.h>
 #include <exceptions/graph_execution_exception.h>
 #include <exceptions/no_results_exception.h>
+#include <graph/FlatUtils.h>
 
 namespace nd4j{
 namespace graph {
@@ -575,15 +576,9 @@ Nd4jStatus GraphExecutioner::execute(Graph *graph, VariableSpace* variableSpace)
             continue;
 
 
-        NDArray* array = var->getNDArray();
-        auto byteVector = array->asByteVector();
+        auto array = var->getNDArray();
 
-        auto fBuffer = builder.CreateVector(byteVector);
-        auto fShape = builder.CreateVector(array->getShapeInfoAsFlatVector());
-
-        auto bo = static_cast<nd4j::graph::ByteOrder>(BitwiseUtils::asByteOrder());
-
-        auto fArray = CreateFlatArray(builder, fShape, fBuffer, static_cast<nd4j::graph::DataType>(array->dataType()), bo);
+        auto fArray = FlatUtils::toFlatArray(builder, *array);
 
         auto fName = builder.CreateString(*(var->getName()));
         auto id = CreateIntPair(builder, var->id(), var->index());
diff --git a/libnd4j/include/graph/FlatUtils.h b/libnd4j/include/graph/FlatUtils.h
index abfff5915..939db1fb7 100644
--- a/libnd4j/include/graph/FlatUtils.h
+++ b/libnd4j/include/graph/FlatUtils.h
@@ -36,6 +36,8 @@ namespace nd4j {
             static std::pair<Nd4jLong, Nd4jLong> fromLongPair(LongPair* pair);
 
             static NDArray* fromFlatArray(const nd4j::graph::FlatArray* flatArray);
+
+            static flatbuffers::Offset<FlatArray> toFlatArray(flatbuffers::FlatBufferBuilder &builder, NDArray &array);
         };
     }
 }
diff --git a/libnd4j/include/graph/impl/FlatUtils.cpp b/libnd4j/include/graph/impl/FlatUtils.cpp
index ad0c5112d..bc8ff7e33 100644
--- a/libnd4j/include/graph/impl/FlatUtils.cpp
+++ b/libnd4j/include/graph/impl/FlatUtils.cpp
@@ -102,5 +102,16 @@ namespace nd4j {
             delete[] newShape;
             return array;
         }
+
+        flatbuffers::Offset<FlatArray> FlatUtils::toFlatArray(flatbuffers::FlatBufferBuilder &builder, NDArray &array) {
+            auto byteVector = array.asByteVector();
+
+            auto fBuffer = builder.CreateVector(byteVector);
+            auto fShape = builder.CreateVector(array.getShapeInfoAsFlatVector());
+
+            auto bo = static_cast<nd4j::graph::ByteOrder>(BitwiseUtils::asByteOrder());
+
+            return CreateFlatArray(builder, fShape, fBuffer, static_cast<nd4j::graph::DataType>(array.dataType()), bo);
+        }
     }
 }
\ No newline at end of file
diff --git a/libnd4j/tests_cpu/layers_tests/FlatUtilsTests.cpp b/libnd4j/tests_cpu/layers_tests/FlatUtilsTests.cpp
new file mode 100644
index 000000000..bf428b833
--- /dev/null
+++ b/libnd4j/tests_cpu/layers_tests/FlatUtilsTests.cpp
@@ -0,0 +1,100 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2019 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+
+#include <NDArray.h>
+#include <NDArrayFactory.h>
+#include "testlayers.h"
+#include <graph/Stash.h>
+#include <FlatUtils.h>
+
+using namespace nd4j;
+
+class FlatUtilsTests : public testing::Test {
+public:
+
+};
+
+TEST_F(FlatUtilsTests, flat_float_serde_1) {
+    auto array = NDArrayFactory::create<float>('c', {4}, {1.f, 2.f, 3.f, 4.f});
+
+    flatbuffers::FlatBufferBuilder builder(1024);
+    auto flatArray = FlatUtils::toFlatArray(builder, array);
+    builder.Finish(flatArray);
+
+
+    auto pfArray = GetFlatArray(builder.GetBufferPointer());
+
+    auto restored = FlatUtils::fromFlatArray(pfArray);
+
+    ASSERT_EQ(array, *restored);
+
+    delete restored;
+}
+
+TEST_F(FlatUtilsTests, flat_int_serde_1) {
+    auto array = NDArrayFactory::create<int>('c', {4}, {1, 2, 3, 4});
+
+    flatbuffers::FlatBufferBuilder builder(1024);
+    auto flatArray = FlatUtils::toFlatArray(builder, array);
+    builder.Finish(flatArray);
+
+
+    auto pfArray = GetFlatArray(builder.GetBufferPointer());
+
+    auto restored = FlatUtils::fromFlatArray(pfArray);
+
+    ASSERT_EQ(array, *restored);
+
+    delete restored;
+}
+
+TEST_F(FlatUtilsTests, flat_bool_serde_1) {
+    auto array = NDArrayFactory::create<bool>('c', {4}, {true, false, true, false});
+
+    flatbuffers::FlatBufferBuilder builder(1024);
+    auto flatArray = FlatUtils::toFlatArray(builder, array);
+    builder.Finish(flatArray);
+
+
+    auto pfArray = GetFlatArray(builder.GetBufferPointer());
+
+    auto restored = FlatUtils::fromFlatArray(pfArray);
+
+    ASSERT_EQ(array, *restored);
+
+    delete restored;
+}
+
+TEST_F(FlatUtilsTests, flat_string_serde_1) {
+    auto array = NDArrayFactory::string('c', {3}, {"alpha", "beta", "gamma"});
+
+    flatbuffers::FlatBufferBuilder builder(1024);
+    auto flatArray = FlatUtils::toFlatArray(builder, array);
+    builder.Finish(flatArray);
+
+
+    auto pfArray = GetFlatArray(builder.GetBufferPointer());
+
+    auto restored = FlatUtils::fromFlatArray(pfArray);
+
+    ASSERT_EQ(array, *restored);
+
+    delete restored;
+}
\ No newline at end of file
diff --git a/libnd4j/tests_cpu/layers_tests/StringTests.cpp b/libnd4j/tests_cpu/layers_tests/StringTests.cpp
index a023dcdd3..2ae236210 100644
--- a/libnd4j/tests_cpu/layers_tests/StringTests.cpp
+++ b/libnd4j/tests_cpu/layers_tests/StringTests.cpp
@@ -24,7 +24,6 @@
 #include "testlayers.h"
 #include <graph/Stash.h>
 
-using namespace nd4j;
 using namespace nd4j;
 
 class StringTests : public testing::Test {
@@ -91,4 +90,4 @@ TEST_F(StringTests, Basic_dup_1) {
     ASSERT_EQ(f, z1);
 
     delete dup;
-}
+}
\ No newline at end of file

From 841eeb56c5d8d78d3df6804024393a8022e5f09d Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Sat, 24 Aug 2019 16:18:38 +0300
Subject: [PATCH 05/56] get rid of context variable

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../ops/declarable/helpers/cuda/lup.cu        | 34 ++++++-------------
 1 file changed, 10 insertions(+), 24 deletions(-)

diff --git a/libnd4j/include/ops/declarable/helpers/cuda/lup.cu b/libnd4j/include/ops/declarable/helpers/cuda/lup.cu
index bf9c73e7c..97d47079b 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/lup.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/lup.cu
@@ -31,8 +31,6 @@
 namespace nd4j {
 namespace ops {
 namespace helpers {
-    nd4j::LaunchContext* defaultContext = nd4j::LaunchContext::defaultContext();
-
 //    template <typename T>
 //    static __device__ void swapRows_(T* matrix, Nd4jLong* shape, int theFirst, int theSecond, Nd4jLong N) {
 //        if (theFirst != theSecond) {
@@ -204,7 +202,7 @@ namespace helpers {
 
         if (inputMatrix->isIdentityMatrix()) return;
 
-        auto stream = defaultContext->getCudaStream();
+        auto stream = LaunchContext::defaultContext()->getCudaStream();
 
         // invert main diagonal
         upvertKernel<T> << < 1, n, 512, *stream >> >
@@ -227,7 +225,7 @@ namespace helpers {
     static void invertUpperMatrix_(NDArray* inputMatrix, NDArray* invertedMatrix) {
         int n = inputMatrix->rows();
         invertedMatrix->setIdentity();
-        auto stream = defaultContext->getCudaStream();
+        auto stream = LaunchContext::defaultContext()->getCudaStream();
         if (inputMatrix->isIdentityMatrix()) { // the inverse for I is I
             return;
         }
@@ -392,7 +390,6 @@ namespace helpers {
             auto n = input->rows();
             cusolverDnHandle_t cusolverH = nullptr;
             cusolverStatus_t status = cusolverDnCreate(&cusolverH);
-            defaultContext = context;
             if (CUSOLVER_STATUS_SUCCESS != status) {
                 throw cuda_exception::build("Cannot create cuSolver handle", status);
             }
@@ -543,9 +540,8 @@ namespace helpers {
 //        DataType dtype = input->dataType();
 //        if (dtype != DataType::DOUBLE)
 //            dtype = DataType::FLOAT32;
-            defaultContext = context;
             auto matrix = NDArrayFactory::create(input->ordering(), {n, n}, DataTypeUtils::fromT<T>(),
-                                                 defaultContext); //, block.getWorkspace());
+                                                 LaunchContext::defaultContext()); //, block.getWorkspace());
             auto det = NDArrayFactory::create<T>(1);
             auto stream = context->getCudaStream();
             NDArray::prepareSpecialUse({output}, {input});
@@ -578,7 +574,6 @@ namespace helpers {
         }
 
         int determinant(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
-            defaultContext = context;
             NDArray::prepareSpecialUse({output}, {input});
             BUILD_SINGLE_SELECTOR(input->dataType(), return determinant_, (context, input, output), FLOAT_NATIVE);
             NDArray::registerSpecialUse({output}, {input});
@@ -586,7 +581,6 @@ namespace helpers {
 
         template<typename T>
         int logAbsDeterminant_(LaunchContext *context, NDArray *input, NDArray *output) {
-            defaultContext = context;
             Nd4jLong n = input->sizeAt(-1);
             Nd4jLong n2 = n * n;
             std::vector<int> dims();
@@ -598,7 +592,7 @@ namespace helpers {
                 dtype = DataType::FLOAT32;
 
             auto matrix = NDArrayFactory::create(input->ordering(), {n, n}, dtype,
-                                                 defaultContext); //, block.getWorkspace());
+                                                 LaunchContext::defaultContext()); //, block.getWorkspace());
             auto det = NDArrayFactory::create<T>(1);
             auto stream = context->getCudaStream();
             NDArray::prepareSpecialUse({output}, {input});
@@ -633,7 +627,6 @@ namespace helpers {
         }
 
         int logAbsDeterminant(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
-            defaultContext = context;
             NDArray::prepareSpecialUse({output}, {input});
             BUILD_SINGLE_SELECTOR(input->dataType(), return logAbsDeterminant_, (context, input, output), FLOAT_NATIVE);
             NDArray::registerSpecialUse({output}, {input});
@@ -696,17 +689,16 @@ namespace helpers {
 
         template<typename T>
         static int inverse_(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
-            defaultContext = context;
             auto n = input->sizeAt(-1);
             auto n2 = n * n;
             auto dtype = DataTypeUtils::fromT<T>(); //input->dataType();
 //            if (dtype != DataType::DOUBLE)
 //                dtype = DataType::FLOAT32;
-            NDArray matrix = NDArrayFactory::create('c', {n, n}, dtype, defaultContext);
-            NDArray upper = NDArrayFactory::create('c', {n, n}, dtype, defaultContext);
-            NDArray lower = NDArrayFactory::create('c', {n, n}, dtype, defaultContext);
-            NDArray compound = NDArrayFactory::create('c', {n, n}, dtype, defaultContext);
-            NDArray permutation = NDArrayFactory::create('c', {n, n}, dtype, defaultContext);
+            NDArray matrix = NDArrayFactory::create('c', {n, n}, dtype, LaunchContext::defaultContext());
+            NDArray upper = NDArrayFactory::create('c', {n, n}, dtype, LaunchContext::defaultContext());
+            NDArray lower = NDArrayFactory::create('c', {n, n}, dtype, LaunchContext::defaultContext());
+            NDArray compound = NDArrayFactory::create('c', {n, n}, dtype, LaunchContext::defaultContext());
+            NDArray permutation = NDArrayFactory::create('c', {n, n}, dtype, LaunchContext::defaultContext());
             auto packX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(),
                                                                                   {input->rankOf() - 2,
                                                                                    input->rankOf() - 1});
@@ -745,7 +737,6 @@ namespace helpers {
         }
 
         int inverse(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
-            defaultContext = context;
             NDArray::prepareSpecialUse({output}, {input});
             BUILD_SINGLE_SELECTOR(input->dataType(), return inverse_, (context, input, output), FLOAT_NATIVE);
             NDArray::registerSpecialUse({output}, {input});
@@ -788,7 +779,6 @@ namespace helpers {
         int cholesky__(LaunchContext *context, NDArray *input, NDArray *output, bool inplace) {
             if (!inplace)
                 output->assign(input);
-            defaultContext = context;
             std::unique_ptr<NDArray> tempOutput(output->dup());
             cusolverDnHandle_t handle = nullptr;
             auto n = input->sizeAt(-1);
@@ -868,7 +858,6 @@ namespace helpers {
 
 //    template <typename T>
         int cholesky_(LaunchContext *context, NDArray *input, NDArray *output, bool inplace) {
-            defaultContext = context;
             NDArray::prepareSpecialUse({output}, {input});
             if (input->dataType() == DataType::DOUBLE)
                 cholesky__<double>(context, input, output, inplace);
@@ -877,7 +866,7 @@ namespace helpers {
             else {
                 std::unique_ptr<NDArray> tempOutput(
                         NDArrayFactory::create_('c', input->getShapeAsVector(), DataType::FLOAT32,
-                                                defaultContext));
+                                                LaunchContext::defaultContext()));
                 tempOutput->assign(input);
                 cholesky__<float>(context, tempOutput.get(), tempOutput.get(), true);
                 output->assign(tempOutput.get());
@@ -888,7 +877,6 @@ namespace helpers {
 
         int cholesky(nd4j::LaunchContext *context, NDArray *input, NDArray *output, bool inplace) {
 //        BUILD_SINGLE_SELECTOR(input->dataType(), return cholesky_, (context, input, output, inplace), FLOAT_TYPES);
-            defaultContext = context;
             return cholesky_(context, input, output, inplace);
         }
 //    BUILD_SINGLE_TEMPLATE(template int cholesky_, (LaunchContext* context, NDArray* input, NDArray* output, bool inplace), FLOAT_TYPES);
@@ -927,7 +915,6 @@ namespace helpers {
 
         template<typename T>
         int logdetFunctor_(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
-            defaultContext = context;
             NDArray::prepareSpecialUse({output}, {input});
             auto n2 = input->sizeAt(-1) * input->sizeAt(-2);
             auto stream = context->getCudaStream();
@@ -957,7 +944,6 @@ namespace helpers {
         }
 
         int logdetFunctor(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
-            defaultContext = context;
             BUILD_SINGLE_SELECTOR(output->dataType(), logdetFunctor_, (context, input, output), FLOAT_NATIVE);
         }
 

From ece6a17b1197d8ee57487b57253c92906b1bf2f5 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Sat, 24 Aug 2019 16:57:48 +0300
Subject: [PATCH 06/56] lup context fix (#164)

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../ops/declarable/helpers/cpu/lup.cpp        | 61 ++++++--------
 .../ops/declarable/helpers/cuda/lup.cu        | 81 +++++++------------
 2 files changed, 57 insertions(+), 85 deletions(-)

diff --git a/libnd4j/include/ops/declarable/helpers/cpu/lup.cpp b/libnd4j/include/ops/declarable/helpers/cpu/lup.cpp
index ee9a78cee..1e3c798e2 100644
--- a/libnd4j/include/ops/declarable/helpers/cpu/lup.cpp
+++ b/libnd4j/include/ops/declarable/helpers/cpu/lup.cpp
@@ -26,7 +26,6 @@
 namespace nd4j {
 namespace ops {
 namespace helpers {
-    nd4j::LaunchContext* defaultContext = nd4j::LaunchContext::defaultContext();
 
     template <typename T>
     static void swapRows_(NDArray* matrix, int theFirst, int theSecond) {
@@ -108,14 +107,14 @@ namespace helpers {
 
 
     template <typename T>
-    static NDArray lup_(NDArray* input, NDArray* compound, NDArray* permutation) {
+    static NDArray lup_(LaunchContext *context, NDArray* input, NDArray* compound, NDArray* permutation) {
 
         const int rowNum = input->rows();
         const int columnNum = input->columns();
 
         NDArray determinant = NDArrayFactory::create<T>(1.f);
         NDArray compoundMatrix = *input; // copy
-        NDArray permutationMatrix(input, false, defaultContext); // has same shape as input and contiguous strides
+        NDArray permutationMatrix(input, false, context); // has same shape as input and contiguous strides
         permutationMatrix.setIdentity();
 
         T pivotValue; // = T(0.0);
@@ -161,46 +160,43 @@ namespace helpers {
         return determinant;
     }
 
-    BUILD_SINGLE_TEMPLATE(template NDArray lup_, (NDArray* input, NDArray* output, NDArray* permutation), FLOAT_TYPES);
+    BUILD_SINGLE_TEMPLATE(template NDArray lup_, (LaunchContext *context, NDArray* input, NDArray* output, NDArray* permutation), FLOAT_TYPES);
 
 
 
     template <typename T>
-    static int determinant_(NDArray* input, NDArray* output) {
+    static int determinant_(LaunchContext *context, NDArray* input, NDArray* output) {
 
         Nd4jLong n = input->sizeAt(-1);
         Nd4jLong n2 = n * n;
 
-        auto matrix = NDArrayFactory::create(input->ordering(), {n, n}, input->dataType(), defaultContext); //, block.getWorkspace());
+        auto matrix = NDArrayFactory::create(input->ordering(), {n, n}, input->dataType(), context); //, block.getWorkspace());
 
         for (int e = 0; e < output->lengthOf(); e++) {
             for (int k = e * n2, row = 0; k < (e + 1) * n2; ++k, ++row)
                 matrix.p(row, input->e<T>(k));
-            output->p(e, lup_<T>(&matrix, (NDArray*)nullptr, (NDArray*)nullptr));
+            output->p(e, lup_<T>(context, &matrix, (NDArray*)nullptr, (NDArray*)nullptr));
         }
 
         return Status::OK();
     }
 
-    BUILD_SINGLE_TEMPLATE(template int determinant_, (NDArray* input, NDArray* output), FLOAT_TYPES);
-
     int determinant(nd4j::LaunchContext * context, NDArray* input, NDArray* output) {
-        defaultContext = context;
-        BUILD_SINGLE_SELECTOR(input->dataType(), return determinant_, (input, output), FLOAT_TYPES);
+        BUILD_SINGLE_SELECTOR(input->dataType(), return determinant_, (context, input, output), FLOAT_TYPES);
     }
 
 template <typename T>
-    int logAbsDeterminant_(NDArray* input, NDArray* output) {
+    int logAbsDeterminant_(LaunchContext *context, NDArray* input, NDArray* output) {
 
         Nd4jLong n = input->sizeAt(-1);
         Nd4jLong n2 = n * n;
 
-        NDArray matrix = NDArrayFactory::create(input->ordering(), {n, n}, input->dataType(), defaultContext); //, block.getWorkspace());
+        NDArray matrix = NDArrayFactory::create(input->ordering(), {n, n}, input->dataType(), context); //, block.getWorkspace());
         for (int e = 0; e < output->lengthOf(); e++) {
             for (int k = e * n2, row = 0; k < (e + 1) * n2; ++k, ++row) {
                 matrix.p(row, input->e<T>(k));
             }
-	    NDArray det = lup_<T>(&matrix, (NDArray*)nullptr, (NDArray*)nullptr);
+	    NDArray det = lup_<T>(context, &matrix, (NDArray*)nullptr, (NDArray*)nullptr);
 	    if (det.e<T>(0) != 0.f)
              	output->p(e, nd4j::math::nd4j_log<T,T>(nd4j::math::nd4j_abs(det.t<T>(0))));
         }
@@ -208,25 +204,23 @@ template <typename T>
         return ND4J_STATUS_OK;
     }
 
-    BUILD_SINGLE_TEMPLATE(template int logAbsDeterminant_, (NDArray* input, NDArray* output), FLOAT_TYPES);
-
     int logAbsDeterminant(nd4j::LaunchContext * context, NDArray* input, NDArray* output) {
-        BUILD_SINGLE_SELECTOR(input->dataType(), return logAbsDeterminant_, (input, output), FLOAT_TYPES);
+        BUILD_SINGLE_SELECTOR(input->dataType(), return logAbsDeterminant_, (context, input, output), FLOAT_TYPES);
     }
 
     template <typename T>
-    static int inverse_(NDArray* input, NDArray* output) {
+    static int inverse_(LaunchContext *context, NDArray* input, NDArray* output) {
 
         auto n = input->sizeAt(-1);
         auto n2 = n * n;
         auto totalCount = output->lengthOf() / n2;
 
         output->assign(0.f); // fill up output tensor with zeros
-        auto matrix = NDArrayFactory::create('c', {n, n}, DataTypeUtils::fromT<T>(), defaultContext); //, block.getWorkspace());
-        auto compound = NDArrayFactory::create('c', {n, n}, DataTypeUtils::fromT<T>(), defaultContext); //, block.getWorkspace());
-        auto permutation = NDArrayFactory::create('c', {n, n}, DataTypeUtils::fromT<T>(), defaultContext);
-        auto lowerMatrix = NDArrayFactory::create('c', {n, n}, DataTypeUtils::fromT<T>(), defaultContext);
-        auto upperMatrix = NDArrayFactory::create('c', {n, n}, DataTypeUtils::fromT<T>(), defaultContext);
+        auto matrix = NDArrayFactory::create('c', {n, n}, DataTypeUtils::fromT<T>(), context); //, block.getWorkspace());
+        auto compound = NDArrayFactory::create('c', {n, n}, DataTypeUtils::fromT<T>(), context); //, block.getWorkspace());
+        auto permutation = NDArrayFactory::create('c', {n, n}, DataTypeUtils::fromT<T>(), context);
+        auto lowerMatrix = NDArrayFactory::create('c', {n, n}, DataTypeUtils::fromT<T>(), context);
+        auto upperMatrix = NDArrayFactory::create('c', {n, n}, DataTypeUtils::fromT<T>(), context);
 
         for (int e = 0; e < totalCount; e++) {
             if (e)
@@ -235,7 +229,7 @@ template <typename T>
             for (int k = e * n2, row = 0; k < (e + 1) * n2; k++) {
                 matrix.p(row++, input->e<T>(k));
             }
-            T det = lup_<T>(&matrix, &compound, &permutation).template e<T>(0);
+            T det = lup_<T>(context, &matrix, &compound, &permutation).template e<T>(0);
 
             // FIXME: and how this is going to work on float16?
             if (nd4j::math::nd4j_abs<T>(det) < T(0.000001)) {
@@ -268,8 +262,7 @@ template <typename T>
     }
 
     int inverse(nd4j::LaunchContext * context, NDArray* input, NDArray* output) {
-        defaultContext = context;
-        BUILD_SINGLE_SELECTOR(input->dataType(), return inverse_, (input, output), FLOAT_TYPES);
+        BUILD_SINGLE_SELECTOR(input->dataType(), return inverse_, (context, input, output), FLOAT_TYPES);
     }
 
     template <typename T>
@@ -296,14 +289,13 @@ template <typename T>
 
         return true;
     }
-    BUILD_SINGLE_TEMPLATE(template bool checkCholeskyInput_, (nd4j::LaunchContext * context, NDArray const* input), FLOAT_TYPES);
 
     bool checkCholeskyInput(nd4j::LaunchContext * context, NDArray const* input) {
         BUILD_SINGLE_SELECTOR(input->dataType(), return checkCholeskyInput_, (context, input), FLOAT_TYPES);
     }
 
     template <typename T>
-    int cholesky_(NDArray* input, NDArray* output, bool inplace) {
+    int cholesky_(LaunchContext *context, NDArray* input, NDArray* output, bool inplace) {
 
         auto n = input->sizeAt(-1);
         auto n2 = n * n;
@@ -311,8 +303,8 @@ template <typename T>
         if (!inplace)
              output->assign(0.f); // fill up output tensor with zeros only inplace=false
 
-        std::unique_ptr<NDArray> matrix(NDArrayFactory::create_('c', {n, n}, input->dataType(), defaultContext)); //, block.getWorkspace());
-        std::unique_ptr<NDArray> lowerMatrix(NDArrayFactory::create_('c',{n, n}, input->dataType(), defaultContext));
+        std::unique_ptr<NDArray> matrix(NDArrayFactory::create_('c', {n, n}, input->dataType(), context)); //, block.getWorkspace());
+        std::unique_ptr<NDArray> lowerMatrix(NDArrayFactory::create_('c',{n, n}, input->dataType(), context));
 
         for (int e = 0; e < totalCount; e++) {
 
@@ -346,14 +338,13 @@ template <typename T>
     }
 
     int cholesky(nd4j::LaunchContext * context, NDArray* input, NDArray* output, bool inplace) {
-        defaultContext = context;
-        BUILD_SINGLE_SELECTOR(input->dataType(), return cholesky_, (input, output, inplace), FLOAT_TYPES);
+        BUILD_SINGLE_SELECTOR(input->dataType(), return cholesky_, (context, input, output, inplace), FLOAT_TYPES);
     }
 
     template <typename T>
-    int logdetFunctor_(NDArray* input, NDArray* output) {
+    int logdetFunctor_(LaunchContext *context, NDArray* input, NDArray* output) {
         std::unique_ptr<NDArray> tempOutput(input->dup());
-        int res = cholesky_<T>(input, tempOutput.get(), false);
+        int res = cholesky_<T>(context, input, tempOutput.get(), false);
         if (res != ND4J_STATUS_OK)
             return res;
         auto n = input->sizeAt(-1);
@@ -372,7 +363,7 @@ template <typename T>
     }
 
     int logdetFunctor(nd4j::LaunchContext * context, NDArray* input, NDArray* output) {
-        BUILD_SINGLE_SELECTOR(input->dataType(), return logdetFunctor_, (input, output), FLOAT_TYPES);
+        BUILD_SINGLE_SELECTOR(input->dataType(), return logdetFunctor_, (context, input, output), FLOAT_TYPES);
     }
 
 }
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/lup.cu b/libnd4j/include/ops/declarable/helpers/cuda/lup.cu
index 97d47079b..f11b56745 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/lup.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/lup.cu
@@ -196,36 +196,33 @@ namespace helpers {
     }
 
     template<typename T>
-    static void invertLowerMatrix_(NDArray *inputMatrix, NDArray *invertedMatrix) {
+    static void invertLowerMatrix_(LaunchContext *context, NDArray *inputMatrix, NDArray *invertedMatrix) {
         int n = inputMatrix->rows();
         invertedMatrix->setIdentity();
 
         if (inputMatrix->isIdentityMatrix()) return;
 
-        auto stream = LaunchContext::defaultContext()->getCudaStream();
+        auto stream = context->getCudaStream();
 
         // invert main diagonal
-        upvertKernel<T> << < 1, n, 512, *stream >> >
-                                        (invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(), inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
+        upvertKernel<T><<<1, n, 512, *stream>>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(), inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
         // invert the second diagonal
-        invertKernelLow<T> << < 1, n, 512, *stream >> >
-                                           (invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(), inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
+        invertKernelLow<T><<<1, n, 512, *stream>>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(), inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
 //        invertKernelLow<T><<<1, n, 128, *stream>>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(), inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
-        invertLowKernel<T><<< n, n, 512, *stream >> >
-                                           (invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(), inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
+        invertLowKernel<T><<<n, n, 512, *stream>>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(), inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
     }
 
-    void invertLowerMatrix(NDArray *inputMatrix, NDArray *invertedMatrix) {
+    void invertLowerMatrix(LaunchContext *context, NDArray *inputMatrix, NDArray *invertedMatrix) {
         NDArray::prepareSpecialUse({invertedMatrix}, {inputMatrix});
-        BUILD_SINGLE_SELECTOR(inputMatrix->dataType(), invertLowerMatrix_, (inputMatrix, invertedMatrix), FLOAT_NATIVE);
+        BUILD_SINGLE_SELECTOR(inputMatrix->dataType(), invertLowerMatrix_, (context, inputMatrix, invertedMatrix), FLOAT_NATIVE);
         NDArray::registerSpecialUse({invertedMatrix}, {inputMatrix});
     }
 
     template<typename T>
-    static void invertUpperMatrix_(NDArray* inputMatrix, NDArray* invertedMatrix) {
+    static void invertUpperMatrix_(LaunchContext *context, NDArray* inputMatrix, NDArray* invertedMatrix) {
         int n = inputMatrix->rows();
         invertedMatrix->setIdentity();
-        auto stream = LaunchContext::defaultContext()->getCudaStream();
+        auto stream = context->getCudaStream();
         if (inputMatrix->isIdentityMatrix()) { // the inverse for I is I
             return;
         }
@@ -235,13 +232,12 @@ namespace helpers {
                 inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
         invertedMatrix->tickWriteDevice();
         invertedMatrix->printIndexedBuffer("Step1 UP inversion");
-        invertUpKernel<T><<<n, n, 512, *stream >>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(),
-                inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
+        invertUpKernel<T><<<n, n, 512, *stream >>>(invertedMatrix->specialBuffer(), invertedMatrix->specialShapeInfo(),inputMatrix->specialBuffer(), inputMatrix->specialShapeInfo(), n);
     }
 
-    void invertUpperMatrix(NDArray *inputMatrix, NDArray *invertedMatrix) {
+    void invertUpperMatrix(LaunchContext *context, NDArray *inputMatrix, NDArray *invertedMatrix) {
         NDArray::prepareSpecialUse({invertedMatrix}, {inputMatrix});
-        BUILD_SINGLE_SELECTOR(invertedMatrix->dataType(), invertUpperMatrix_, (inputMatrix, invertedMatrix), FLOAT_NATIVE);
+        BUILD_SINGLE_SELECTOR(invertedMatrix->dataType(), invertUpperMatrix_, (context, inputMatrix, invertedMatrix), FLOAT_NATIVE);
         NDArray::prepareSpecialUse({invertedMatrix}, {inputMatrix});
     }
 
@@ -525,23 +521,19 @@ namespace helpers {
             input->tickWriteDevice();
         }
 
-        BUILD_SINGLE_TEMPLATE(template void lup_,
-                              (LaunchContext * context, NDArray * input, NDArray * output, NDArray * permutation),
-                              FLOAT_NATIVE);
+        BUILD_SINGLE_TEMPLATE(template void lup_,(LaunchContext * context, NDArray * input, NDArray * output, NDArray * permutation), FLOAT_NATIVE);
 
         template<typename T>
         static int determinant_(nd4j::LaunchContext *context, NDArray *input, NDArray *output) {
             Nd4jLong n = input->sizeAt(-1);
             Nd4jLong n2 = n * n;
             std::vector<int> dims();
-            auto packX = ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(),
-                                                                            {input->rankOf() - 2, input->rankOf() - 1});
+            auto packX = ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(), {input->rankOf() - 2, input->rankOf() - 1});
             //auto packZ = ConstantTadHelper::getInstance()->tadForDimensions(output->shapeInfo(), {output->rankOf() - 1});
 //        DataType dtype = input->dataType();
 //        if (dtype != DataType::DOUBLE)
 //            dtype = DataType::FLOAT32;
-            auto matrix = NDArrayFactory::create(input->ordering(), {n, n}, DataTypeUtils::fromT<T>(),
-                                                 LaunchContext::defaultContext()); //, block.getWorkspace());
+            auto matrix = NDArrayFactory::create(input->ordering(), {n, n}, DataTypeUtils::fromT<T>(), context); //, block.getWorkspace());
             auto det = NDArrayFactory::create<T>(1);
             auto stream = context->getCudaStream();
             NDArray::prepareSpecialUse({output}, {input});
@@ -550,8 +542,7 @@ namespace helpers {
             for (int e = 0; e < output->lengthOf(); e++) {
                 Nd4jLong pos = e * n2;
 //            if (matrix.dataType() == input->dataType())
-                fillMatrix<T, T> << < launchDims.x, launchDims.y, launchDims.z, *stream >> >
-                                                                                (matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), pos, n);
+                fillMatrix<T, T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), pos, n);
 //            else
 //                fillMatrix<T, float><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), pos, n);
 
@@ -584,15 +575,13 @@ namespace helpers {
             Nd4jLong n = input->sizeAt(-1);
             Nd4jLong n2 = n * n;
             std::vector<int> dims();
-            auto packX = ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(),
-                                                                            {input->rankOf() - 2, input->rankOf() - 1});
+            auto packX = ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(), {input->rankOf() - 2, input->rankOf() - 1});
             //auto packZ = ConstantTadHelper::getInstance()->tadForDimensions(output->shapeInfo(), {output->rankOf() - 1});
             DataType dtype = input->dataType();
             if (dtype != DataType::DOUBLE)
                 dtype = DataType::FLOAT32;
 
-            auto matrix = NDArrayFactory::create(input->ordering(), {n, n}, dtype,
-                                                 LaunchContext::defaultContext()); //, block.getWorkspace());
+            auto matrix = NDArrayFactory::create(input->ordering(), {n, n}, dtype, context); //, block.getWorkspace());
             auto det = NDArrayFactory::create<T>(1);
             auto stream = context->getCudaStream();
             NDArray::prepareSpecialUse({output}, {input});
@@ -601,8 +590,7 @@ namespace helpers {
             for (int e = 0; e < output->lengthOf(); e++) {
                 Nd4jLong pos = e * n2;
 //            if (matrix.dataType() == input->dataType())
-                fillMatrix<T, T> << < launchDims.x, launchDims.y, launchDims.z, *stream >> >
-                                                                                (matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), pos, n);
+                fillMatrix<T, T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), pos, n);
 //            else
 //                fillMatrix<T, float><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), pos, n);
 
@@ -614,8 +602,7 @@ namespace helpers {
                 auto inputBuf = reinterpret_cast<T *>(matrix.specialBuffer());
                 auto outputBuf = reinterpret_cast<T *>(output->specialBuffer()) + offset;
 //            if (matrix.dataType() == input->dataType())
-                determinantLogKernel<T> << < launchDims.x, launchDims.y, launchDims.z, *stream >> >
-                                                                                       (inputBuf, outputBuf, n);
+                determinantLogKernel<T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(inputBuf, outputBuf, n);
 //            else
 //                determinantLogKernel<T, float><<<launchDims.x, launchDims.y, launchDims.z, *stream >>> (inputBuf, outputBuf, n);
             }
@@ -694,11 +681,11 @@ namespace helpers {
             auto dtype = DataTypeUtils::fromT<T>(); //input->dataType();
 //            if (dtype != DataType::DOUBLE)
 //                dtype = DataType::FLOAT32;
-            NDArray matrix = NDArrayFactory::create('c', {n, n}, dtype, LaunchContext::defaultContext());
-            NDArray upper = NDArrayFactory::create('c', {n, n}, dtype, LaunchContext::defaultContext());
-            NDArray lower = NDArrayFactory::create('c', {n, n}, dtype, LaunchContext::defaultContext());
-            NDArray compound = NDArrayFactory::create('c', {n, n}, dtype, LaunchContext::defaultContext());
-            NDArray permutation = NDArrayFactory::create('c', {n, n}, dtype, LaunchContext::defaultContext());
+            NDArray matrix = NDArrayFactory::create('c', {n, n}, dtype, context);
+            NDArray upper = NDArrayFactory::create('c', {n, n}, dtype, context);
+            NDArray lower = NDArrayFactory::create('c', {n, n}, dtype, context);
+            NDArray compound = NDArrayFactory::create('c', {n, n}, dtype, context);
+            NDArray permutation = NDArrayFactory::create('c', {n, n}, dtype, context);
             auto packX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(),
                                                                                   {input->rankOf() - 2,
                                                                                    input->rankOf() - 1});
@@ -708,20 +695,17 @@ namespace helpers {
             auto stream = context->getCudaStream();
 
             for (auto i = 0LL; i < packX.numberOfTads(); i++) {
-                fillMatrix<T, T> << < 1, n2, 1024, *stream >> >
-                                                   (matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(),
-                                                           i * n2, n);
+                fillMatrix<T, T><<<1, n2, 1024, *stream>>>(matrix.specialBuffer(), matrix.specialShapeInfo(), input->specialBuffer(), input->specialShapeInfo(), i * n2, n);
                 matrix.tickWriteDevice();
                 compound.assign(matrix);
                 lup_<T>(context, &compound, nullptr, nullptr);
-                fillLowerUpperKernel<T> << < n, n, 1024, *stream >> >
-                                                         (lower.specialBuffer(), lower.specialShapeInfo(), upper.specialBuffer(), upper.specialShapeInfo(), compound.specialBuffer(), compound.specialShapeInfo(), n);
+                fillLowerUpperKernel<T><<<n, n, 1024, *stream>>>(lower.specialBuffer(), lower.specialShapeInfo(), upper.specialBuffer(), upper.specialShapeInfo(), compound.specialBuffer(), compound.specialShapeInfo(), n);
                 matrix.assign(0);
-                invertUpperMatrix(&upper, &matrix); // U^{-1}
+                invertUpperMatrix(context, &upper, &matrix); // U^{-1}
                 matrix.tickWriteDevice();
 //                matrix.printIndexedBuffer("Upper Inverted");
                 compound.assign(0);
-                invertLowerMatrix(&lower, &compound); // L{-1}
+                invertLowerMatrix(context, &lower, &compound); // L{-1}
                 compound.tickWriteDevice();
 //                compound.printIndexedBuffer("Lower Inverted");
 //                matrix.tickWriteDevice();
@@ -729,9 +713,7 @@ namespace helpers {
                 nd4j::MmulHelper::mmul(&matrix, &compound, &upper, 1.0, 0.0);
                 upper.tickWriteDevice();
 //                upper.printIndexedBuffer("Full inverted");
-                returnMatrix<T> << < 1, n2, 1024, *stream >> >
-                                                     (output->specialBuffer(), output->specialShapeInfo(), upper.specialBuffer(), upper.specialShapeInfo(),
-                                                             i * n2, n);
+                returnMatrix<T> <<<1, n2, 1024, *stream>>>(output->specialBuffer(), output->specialShapeInfo(), upper.specialBuffer(), upper.specialShapeInfo(), i * n2, n);
             }
             return Status::OK();
         }
@@ -865,8 +847,7 @@ namespace helpers {
                 cholesky__<float>(context, input, output, inplace);
             else {
                 std::unique_ptr<NDArray> tempOutput(
-                        NDArrayFactory::create_('c', input->getShapeAsVector(), DataType::FLOAT32,
-                                                LaunchContext::defaultContext()));
+                        NDArrayFactory::create_('c', input->getShapeAsVector(), DataType::FLOAT32, context));
                 tempOutput->assign(input);
                 cholesky__<float>(context, tempOutput.get(), tempOutput.get(), true);
                 output->assign(tempOutput.get());

From d607bec6f928c666368b1c6e832537c0910f32d1 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Mon, 26 Aug 2019 11:45:33 +1000
Subject: [PATCH 07/56] Small test fixes (#165)

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 .../org/nd4j/evaluation/RegressionEvalTest.java  |  4 ++--
 .../java/org/nd4j/linalg/crash/SpecialTests.java | 16 ++++++----------
 2 files changed, 8 insertions(+), 12 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/evaluation/RegressionEvalTest.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/evaluation/RegressionEvalTest.java
index d182377fe..1bd6fd22c 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/evaluation/RegressionEvalTest.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/evaluation/RegressionEvalTest.java
@@ -291,7 +291,7 @@ public class RegressionEvalTest  extends BaseNd4jTest {
         for (Metric m : Metric.values()) {
             double d1 = e4d.scoreForMetric(m);
             double d2 = e2d.scoreForMetric(m);
-            assertEquals(m.toString(), d2, d1, 1e-6);
+            assertEquals(m.toString(), d2, d1, 1e-5);
         }
     }
 
@@ -385,7 +385,7 @@ public class RegressionEvalTest  extends BaseNd4jTest {
         for(Metric m : Metric.values()){
             double d1 = e4d_m1.scoreForMetric(m);
             double d2 = e2d_m1.scoreForMetric(m);
-            assertEquals(m.toString(), d2, d1, 1e-6);
+            assertEquals(m.toString(), d2, d1, 1e-5);
         }
 
         //Check per-output masking:
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/crash/SpecialTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/crash/SpecialTests.java
index 75a91263d..cfff870a6 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/crash/SpecialTests.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/crash/SpecialTests.java
@@ -551,15 +551,13 @@ public class SpecialTests extends BaseNd4jTest {
         int[] inputShape = new int[]{1, 2, 2, 1};
 
         int M = 2;
-        int[] blockShape = new int[]{M, 1};
-        int[] paddingShape = new int[]{M, 2};
 
         INDArray input = Nd4j.randn(inputShape).castTo(DataType.DOUBLE);
-        INDArray blocks = Nd4j.create(new float[]{2, 2}, blockShape).castTo(DataType.INT);
-        INDArray padding = Nd4j.create(new float[]{0, 0, 0, 0}, paddingShape).castTo(DataType.INT);
+        INDArray blocks = Nd4j.createFromArray(2, 2);
+        INDArray padding = Nd4j.createFromArray(0, 0, 0, 0).reshape(2,2);
 
         INDArray expOut = Nd4j.create(DataType.DOUBLE, miniBatch, 1, 1, 1);
-        val op = DynamicCustomOp.builder("space_to_batch")
+        val op = DynamicCustomOp.builder("space_to_batch_nd")
                 .addInputs(input, blocks, padding)
                 .addOutputs(expOut).build();
         Nd4j.getExecutioner().execAndReturn(op);
@@ -573,15 +571,13 @@ public class SpecialTests extends BaseNd4jTest {
         int[] inputShape = new int[]{miniBatch, 1, 1, 1};
 
         int M = 2;
-        int[] blockShape = new int[]{M, 1};
-        int[] cropShape = new int[]{M, 2};
 
         INDArray input = Nd4j.randn(inputShape).castTo(DataType.DOUBLE);
-        INDArray blocks = Nd4j.create(new float[]{2, 2}, blockShape).castTo(DataType.INT);
-        INDArray crops = Nd4j.create(new float[]{0, 0, 0, 0}, cropShape).castTo(DataType.INT);
+        INDArray blocks = Nd4j.createFromArray(2, 2);
+        INDArray crops = Nd4j.createFromArray(0, 0, 0, 0).reshape(2,2);
 
         INDArray expOut = Nd4j.create(DataType.DOUBLE, 1, 2, 2, 1);
-        DynamicCustomOp op = DynamicCustomOp.builder("batch_to_space")
+        DynamicCustomOp op = DynamicCustomOp.builder("batch_to_space_nd")
                 .addInputs(input, blocks, crops)
                 .addOutputs(expOut).build();
         Nd4j.getExecutioner().execAndReturn(op);

From daf5420f4c2b47ec0e615a04b34d11880c4a86f2 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Mon, 26 Aug 2019 08:13:22 +0300
Subject: [PATCH 08/56] cmake fix for windows debug build

Signed-off-by: raver119 <raver119@gmail.com>
---
 libnd4j/blas/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/libnd4j/blas/CMakeLists.txt b/libnd4j/blas/CMakeLists.txt
index 8e940bedb..add8960a3 100755
--- a/libnd4j/blas/CMakeLists.txt
+++ b/libnd4j/blas/CMakeLists.txt
@@ -341,7 +341,7 @@ elseif(CPU_BLAS)
         endif()
     endif()
 
-    if (CMAKE_BUILD_TYPE STREQUAL "Debug" AND NOT(APPLE))
+    if (CMAKE_BUILD_TYPE STREQUAL "Debug" AND NOT(APPLE) AND NOT(WIN32))
         SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -rdynamic -Wl,-export-dynamic")
         SET(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -export-dynamic")
     endif()

From 9a513a9aa66a9ad8efb3a498640963309512213d Mon Sep 17 00:00:00 2001
From: Robert Altena <Rob@ra-ai.com>
Date: Mon, 26 Aug 2019 16:01:20 +0900
Subject: [PATCH 09/56]  INDArray javadocs (#166)

* small fix of compiler warnings in nd4j.

Signed-off-by: Robert Altena <Rob@Ra-ai.com>

* indarray javadoc start.

Signed-off-by: Robert Altena <Rob@Ra-ai.com>
---
 .../nd4j/linalg/api/ndarray/BaseNDArray.java  | 27 -------------------
 .../linalg/api/ndarray/BaseSparseNDArray.java |  3 ---
 .../org/nd4j/linalg/api/ndarray/INDArray.java | 26 +++++++++---------
 .../java/org/nd4j/linalg/factory/Nd4j.java    |  4 +--
 4 files changed, 13 insertions(+), 47 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
index 121ca2b43..c73a56c61 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
@@ -142,21 +142,11 @@ public abstract class BaseNDArray implements INDArray, Iterable {
 
     }
 
-    /**
-     * Returns true if this array is compressed, and false otherwise
-     * @return
-     */
     @Override
     public boolean isCompressed() {
         return compressed;
     }
 
-    /**
-     * This method marks INDArray instance as compressed
-     * PLEASE NOTE: Do not use this method unless you 100% have to
-     *
-     * @param reallyCompressed
-     */
     @Override
     public void markAsCompressed(boolean reallyCompressed) {
         this.compressed = reallyCompressed;
@@ -949,17 +939,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
 
     @Override
     public int elementWiseStride() {
-        /*
-        if(Shape.elementWiseStride(shapeInfo()) < 0 && !attemptedToFindElementWiseStride) {
-            INDArray reshapeAttempt = Shape.newShapeNoCopy(this,new int[]{1,length()}, ordering() == 'f');
-            if(reshapeAttempt != null && reshapeAttempt.elementWiseStride() > 0) {
-               Shape.setElementWiseStride(shapeInfo(), reshapeAttempt.stride(-1));
-               this.shapeInformation = Nd4j.getShapeInfoProvider().createShapeInformation(shape(), stride(), offset(),reshapeAttempt.stride(-1), ordering());
-            }
-            attemptedToFindElementWiseStride = true;
-        
-        }
-        */
         return Shape.elementWiseStride(shapeInfoDataBuffer());
     }
 
@@ -5349,12 +5328,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return jvmShapeInfo.shape;
     }
 
-    /**
-     * Returns the shape information debugging
-     * information
-     *
-     * @return the shape information debugging information
-     */
     @Override
     public String shapeInfoToString() {
         return Shape.shapeToString(this);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
index 3570ed7ad..5257be12e 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
@@ -328,13 +328,11 @@ public abstract class BaseSparseNDArray implements ISparseNDArray {
         return sparseInformation;
     }
 
-
     @Override
     public LongBuffer shapeInfo() {
         return null;
     }
 
-
     @Override
     public boolean isCompressed() {
         return false;
@@ -364,7 +362,6 @@ public abstract class BaseSparseNDArray implements ISparseNDArray {
         return Shape.sparseOffsets(sparseInformation);
     }
 
-
     @Override
     public int stride(int dimension) {
         int rank = Shape.rank(shapeInformation);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java
index 727b5db6d..88cda5b4f 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java
@@ -41,45 +41,44 @@ import org.nd4j.linalg.string.NDArrayStrings;
  */
 public interface INDArray extends Serializable, AutoCloseable {
     /**
-     * Returns the shape information debugging
-     * information
-     * @return the shape information debugging information
+     * Returns the shape information debugging information
+     * @return the shape information.
      */
     String shapeInfoToString();
 
     /**
      * Shape info
-     * @return
+     * @return Shape info
      */
     DataBuffer shapeInfoDataBuffer();
 
     /**
      * Sparse info
-     * @return
+     * @return Sparse info.
      */
     DataBuffer sparseInfoDataBuffer();
 
     /**
      * Shape info
-     * @return
+     * @return Shape info
      */
     LongBuffer shapeInfo();
 
     /**
-     * Returns true if this array is a view or not
-     * @return
+     * Check if this array is a view or not.
+     * @return true if array is a view.
      */
     boolean isView();
 
     /**
-     * Returns true if this array is sparse
-     * @return
+     * Check if this array is sparse
+     * @return true if this array is sparse.
      */
     boolean isSparse();
 
     /**
-     * Returns true if this array is compressed, and false otherwise
-     * @return
+     * Check if this array is compressed.
+     * @return true if this array is compressed.
      */
     boolean isCompressed();
 
@@ -87,11 +86,10 @@ public interface INDArray extends Serializable, AutoCloseable {
      * This method marks INDArray instance as compressed
      * PLEASE NOTE: Do not use this method unless you 100% have to
      *
-     * @param reallyCompressed
+     * @param reallyCompressed new value for compressed.
      */
     void markAsCompressed(boolean reallyCompressed);
 
-
     /**
      * Returns the rank of the ndarray (the number of dimensions).
      *
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
index eb3250e3c..d2a4f94a4 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
@@ -103,7 +103,6 @@ import java.text.NumberFormat;
 import java.text.ParseException;
 import java.util.*;
 import java.util.concurrent.atomic.AtomicBoolean;
-import java.util.concurrent.atomic.AtomicInteger;
 import java.util.concurrent.atomic.AtomicReference;
 import java.util.logging.Logger;
 
@@ -153,7 +152,6 @@ public class Nd4j {
     public static RandomFactory randomFactory;
     private static MemoryWorkspaceManager workspaceManager;
     private static DeallocatorService deallocatorService;
-    private static final AtomicInteger numThreads = new AtomicInteger(-1);
     private static AtomicReference<DataType> defaultFloatingPointDataType;
 
     private static DataBufferFactory DATA_BUFFER_FACTORY_INSTANCE;
@@ -4755,7 +4753,7 @@ public class Nd4j {
      * @param toStrip the ndarray to newShapeNoCopy
      * @return the reshaped ndarray
      */
-    @SuppressWarnings("WeakerAccess") // Needs tests if part of public API.
+    @SuppressWarnings({"unused"}) // Needs tests if part of public API.
     public static INDArray stripOnes(INDArray toStrip) {
         if (toStrip.isVector())
             return toStrip;

From b417ca21bfb00f24e100d5308ae499021463498f Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Mon, 26 Aug 2019 23:10:28 +1000
Subject: [PATCH 10/56] Fix for concat op shape function (empty shapes) (#167)

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 .../declarable/generic/transforms/concat.cpp  |  4 +--
 .../opvalidation/ShapeOpValidation.java       | 33 +++++++++++++++++++
 2 files changed, 34 insertions(+), 3 deletions(-)

diff --git a/libnd4j/include/ops/declarable/generic/transforms/concat.cpp b/libnd4j/include/ops/declarable/generic/transforms/concat.cpp
index 5249758bf..3c165f64f 100644
--- a/libnd4j/include/ops/declarable/generic/transforms/concat.cpp
+++ b/libnd4j/include/ops/declarable/generic/transforms/concat.cpp
@@ -167,9 +167,7 @@ DECLARE_SHAPE_FN(concat) {
     }
 
     for(int i = 1; i < numOfArrs; ++i)
-        if (!shape::isEmpty(arrShapes[i])) {
-            outShapeInfo[axis + 1] += arrShapes[i][axis + 1];
-        }
+        outShapeInfo[axis + 1] += arrShapes[i][axis + 1];
 
     ShapeUtils::updateStridesAndType(outShapeInfo, arrShapes[0], shape::order(arrShapes[0]));
 
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java
index e53cfa5ff..2965f367f 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java
@@ -2109,6 +2109,38 @@ public class ShapeOpValidation extends BaseOpValidation {
         Nd4j.exec(op);
     }
 
+    @Test
+    public void testConcatEmpty2(){
+        INDArray empty10a = Nd4j.create(DataType.INT, 1, 0);
+        INDArray empty10b = Nd4j.create(DataType.INT, 1, 0);
+
+        DynamicCustomOp op = DynamicCustomOp.builder("concat")
+                .addInputs(empty10a, empty10b)
+                .addIntegerArguments(0) //axis = 0
+                .build();
+
+        List<LongShapeDescriptor> l = op.calculateOutputShape();
+        assertEquals(1, l.size());
+        assertTrue(l.get(0).isEmpty());
+        assertArrayEquals(new long[]{2, 0}, l.get(0).getShape());
+        assertEquals(DataType.INT, l.get(0).dataType());
+
+        op.addOutputArgument(Nd4j.create(DataType.INT, 2, 0));
+        Nd4j.exec(op);
+
+
+        op = DynamicCustomOp.builder("concat")
+                .addInputs(empty10a, empty10b)
+                .addIntegerArguments(1) //axis = 1
+                .build();
+        l = op.calculateOutputShape();
+        assertEquals(1, l.size());
+        assertTrue(l.get(0).isEmpty());
+        assertArrayEquals(new long[]{1, 0}, l.get(0).getShape());
+        op.addOutputArgument(Nd4j.create(DataType.INT, 1, 0));
+        Nd4j.exec(op);
+    }
+
     @Test
     public void testEmptyGather(){
         /*
@@ -2434,4 +2466,5 @@ public class ShapeOpValidation extends BaseOpValidation {
                 .addInputs(Nd4j.createFromArray(1, 0))
                 .build();
     }
+
 }

From bb5fc36e5e4d446d1f60fa949c2780e9cba7f75d Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Mon, 26 Aug 2019 19:37:05 +0300
Subject: [PATCH 11/56] [WIP] ops fixes (#168)

* - correct layer_norm

Signed-off-by: Yurii <yurii@skymind.io>

* - further fix of layer norm

Signed-off-by: Yurii <yurii@skymind.io>

* - correct scatter_upd op

Signed-off-by: Yurii <yurii@skymind.io>

* - correct cuda kernel for histogram_fixed_width op

Signed-off-by: Yurii <yurii@skymind.io>

* - delete comments

Signed-off-by: Yurii <yurii@skymind.io>

* enabled one ignored test

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../generic/transforms/layer_norm.cpp         |  28 +-
 .../helpers/cpu/histogramFixedWidth.cpp       |   9 +-
 .../ops/declarable/helpers/cpu/scatter.cpp    |   2 +
 .../helpers/cuda/histogramFixedWidth.cu       | 245 +++++++++++-------
 .../ops/declarable/helpers/cuda/scatter.cu    |  11 +-
 .../layers_tests/DeclarableOpsTests10.cpp     |  26 +-
 .../layers_tests/DeclarableOpsTests15.cpp     |   4 +-
 .../layers_tests/DeclarableOpsTests16.cpp     |  23 +-
 .../opvalidation/LayerOpValidation.java       |   1 -
 9 files changed, 240 insertions(+), 109 deletions(-)

diff --git a/libnd4j/include/ops/declarable/generic/transforms/layer_norm.cpp b/libnd4j/include/ops/declarable/generic/transforms/layer_norm.cpp
index 4e612565e..684d98d6d 100644
--- a/libnd4j/include/ops/declarable/generic/transforms/layer_norm.cpp
+++ b/libnd4j/include/ops/declarable/generic/transforms/layer_norm.cpp
@@ -32,9 +32,12 @@ namespace ops  {
         auto input = INPUT_VARIABLE(0);
         auto gain = INPUT_VARIABLE(1);
         auto output = OUTPUT_VARIABLE(0);
-        
+
         std::vector<int> axis = *block.getIArguments();
 
+        const bool isNCHW = block.getBArguments()->size() > 0 ? B_ARG(0) : true;       // INT_ARG(9): 0-NCHW,  1-NHWC
+        const int dimC = isNCHW ? 1 : input->rankOf() - 1;
+
         NDArray* bias = nullptr;
         if (block.width() > 2)
             bias = INPUT_VARIABLE(2);
@@ -48,9 +51,12 @@ namespace ops  {
         std::vector<bool> bargs = {};
         standardizeOp.execute(inputs, outputs, targs, longAxis, bargs);
 
-        output->applyTrueBroadcast(nd4j::BroadcastOpsTuple::Multiply(), gain, output);
-        if(bias != nullptr)
-            output->applyTrueBroadcast(nd4j::BroadcastOpsTuple::Add(), bias, output);
+        // output->applyTrueBroadcast(nd4j::BroadcastOpsTuple::Multiply(), gain, output);
+        output->applyBroadcast(nd4j::broadcast::Multiply, {dimC}, gain);
+        if(bias != nullptr) {
+            // output->applyTrueBroadcast(nd4j::BroadcastOpsTuple::Add(), bias, output);
+            output->applyBroadcast(nd4j::broadcast::Add, {dimC}, bias);
+        }
 
         return Status::OK();
     }
@@ -71,12 +77,17 @@ namespace ops  {
         auto dLdg = OUTPUT_VARIABLE(1);
         auto dLdb = block.width() == 4 ? OUTPUT_VARIABLE(2) : nullptr;
 
+        const bool isNCHW = block.getBArguments()->size() > 0 ? B_ARG(0) : true;       // INT_ARG(9): 0-NCHW,  1-NHWC
+        const int dimC = isNCHW ? 1 : input->rankOf() - 1;
+
         std::vector<int> axis = *block.getIArguments();
 
         std::vector<Nd4jLong> longAxis = ArrayUtils::toLongVector(axis);
 
-        if(bias != nullptr)
-            eps->reduceAlongDimension(nd4j::reduce::Sum, dLdb, {0}, true);
+        if(bias != nullptr) {
+            // eps->reduceAlongDimension(nd4j::reduce::Sum, dLdb, {0}, true);
+            eps->reduceAlongDimension(nd4j::reduce::Sum, dLdb, ShapeUtils::evalDimsToExclude(input->rankOf(), {dimC}), true);
+        }
 
         NDArray standardized(input->shapeInfo(), false, block.launchContext());
 
@@ -88,10 +99,11 @@ namespace ops  {
 
         standardizeOp.execute(inputs, outputs, targs, longAxis, bargs);
         standardized.applyPairwiseTransform(nd4j::pairwise::Multiply, eps, &standardized, nullptr);
-        standardized.reduceAlongDimension(nd4j::reduce::Sum, dLdg, {0}, true);
+        standardized.reduceAlongDimension(nd4j::reduce::Sum, dLdg, ShapeUtils::evalDimsToExclude(input->rankOf(), {dimC}), true);
 
         nd4j::ops::standardize_bp standardizeBp;
-        eps->applyTrueBroadcast(nd4j::BroadcastOpsTuple::Multiply(), gain, dLdx);
+        // eps->applyTrueBroadcast(nd4j::BroadcastOpsTuple::Multiply(), gain, dLdx);
+        eps->applyBroadcast(nd4j::broadcast::Multiply, {dimC}, gain, dLdx);
 
         auto dLdx_tmp = dLdx->dup();
         std::vector<NDArray *> standardizeBpArgs = {input, dLdx_tmp};
diff --git a/libnd4j/include/ops/declarable/helpers/cpu/histogramFixedWidth.cpp b/libnd4j/include/ops/declarable/helpers/cpu/histogramFixedWidth.cpp
index a3f3e2f9e..349d0381a 100644
--- a/libnd4j/include/ops/declarable/helpers/cpu/histogramFixedWidth.cpp
+++ b/libnd4j/include/ops/declarable/helpers/cpu/histogramFixedWidth.cpp
@@ -28,13 +28,10 @@ namespace helpers {
 template <typename T>
 void histogramFixedWidth_(const NDArray& input, const NDArray& range, NDArray& output) {
 
-     const int nbins = output.lengthOf();
+    const int nbins = output.lengthOf();
 
-    // firstly initialize output with zeros 
-    if(output.ews() == 1)
-        memset(output.buffer(), 0, nbins * output.sizeOfT());
-    else
-        output = 0;
+    // firstly initialize output with zeros
+    output.nullify();
 
     const T leftEdge  = range.e<double>(0);
     const T rightEdge = range.e<double>(1);
diff --git a/libnd4j/include/ops/declarable/helpers/cpu/scatter.cpp b/libnd4j/include/ops/declarable/helpers/cpu/scatter.cpp
index c1d01930c..0b16ac989 100644
--- a/libnd4j/include/ops/declarable/helpers/cpu/scatter.cpp
+++ b/libnd4j/include/ops/declarable/helpers/cpu/scatter.cpp
@@ -54,6 +54,8 @@ PRAGMA_OMP_PARALLEL_FOR_ARGS(OMP_IF(!lock) schedule(guided))
         std::vector<int> dimsToExcludeUpd(sizeOfDims);
         std::iota(dimsToExcludeUpd.begin(), dimsToExcludeUpd.end(), 0);
 
+        shape::printIntArray(dimsToExcludeUpd.data(),dimsToExcludeUpd.size());
+
 // PRAGMA_OMP_PARALLEL_FOR_ARGS(OMP_IF(indLen > Environment::getInstance()->elementwiseThreshold()) schedule(guided)) // causes known openMP asan bug !
 PRAGMA_OMP_PARALLEL_FOR_ARGS(OMP_IF(!lock) schedule(guided))
         for(Nd4jLong i = 0; i < indLen; ++i) {
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/histogramFixedWidth.cu b/libnd4j/include/ops/declarable/helpers/cuda/histogramFixedWidth.cu
index 2c46210cf..ebde4909c 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/histogramFixedWidth.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/histogramFixedWidth.cu
@@ -20,110 +20,181 @@
 
 #include <ops/declarable/helpers/histogramFixedWidth.h>
 #include <cuda_exception.h>
+#include <PointersManager.h>
 
-namespace nd4j {
-namespace ops {
+namespace nd4j    {
+namespace ops     {
 namespace helpers {
 
-    template <typename T>
-    __global__ static void copyBuffers(Nd4jLong* destination, void const* source, Nd4jLong* sourceShape, Nd4jLong bufferLength) {
-        const auto tid = blockIdx.x * gridDim.x + threadIdx.x;
-        const auto step = gridDim.x * blockDim.x;
-        for (int t = tid; t < bufferLength; t += step) {
-            destination[t] = reinterpret_cast<T const*>(source)[shape::getIndexOffset(t, sourceShape, bufferLength)];
-        }
+///////////////////////////////////////////////////////////////////
+template<typename T>
+__global__ static void histogramFixedWidthCuda( const void* vx, const Nd4jLong* xShapeInfo,
+                                                      void* vz, const Nd4jLong* zShapeInfo,
+                                                const T leftEdge, const T rightEdge) {
+
+    const T* x  = reinterpret_cast<const T*>(vx);
+    Nd4jLong* z = reinterpret_cast<Nd4jLong*>(vz);
+
+    __shared__ Nd4jLong xLen, zLen, totalThreads, nbins;
+    __shared__ T binWidth, secondEdge, lastButOneEdge;
+
+    if (threadIdx.x == 0) {
+
+        xLen  = shape::length(xShapeInfo);
+        nbins = shape::length(zShapeInfo);          // nbins = zLen
+        totalThreads = gridDim.x * blockDim.x;
+
+        binWidth       = (rightEdge - leftEdge ) / nbins;
+        secondEdge     = leftEdge + binWidth;
+        lastButOneEdge = rightEdge - binWidth;
     }
 
-    template <typename T>
-    __global__ static void returnBuffers(void* destination, Nd4jLong const* source, Nd4jLong* destinationShape, Nd4jLong bufferLength) {
-        const auto tid = blockIdx.x * gridDim.x + threadIdx.x;
-        const auto step = gridDim.x * blockDim.x;
-        for (int t = tid; t < bufferLength; t += step) {
-            reinterpret_cast<T*>(destination)[shape::getIndexOffset(t, destinationShape, bufferLength)] = source[t];
-        }
+    __syncthreads();
+
+    const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+
+    for (Nd4jLong i = tid; i < xLen; i += totalThreads) {
+
+        const T value = x[shape::getIndexOffset(i, xShapeInfo, xLen)];
+
+        Nd4jLong zIndex;
+
+        if(value < secondEdge)
+            zIndex = 0;
+        else if(value >= lastButOneEdge)
+            zIndex = nbins - 1;
+        else
+            zIndex = static_cast<Nd4jLong>((value - leftEdge) / binWidth);
+
+        nd4j::math::atomics::nd4j_atomicAdd(&z[shape::getIndexOffset(zIndex, zShapeInfo, nbins)], 1LL);
     }
+}
 
-    template <typename T>
-    static __global__ void histogramFixedWidthKernel(void* outputBuffer, Nd4jLong outputLength, void const* inputBuffer, Nd4jLong* inputShape, Nd4jLong inputLength, double const leftEdge, double binWidth, double secondEdge, double lastButOneEdge) {
+///////////////////////////////////////////////////////////////////
+template<typename T>
+__host__ static void histogramFixedWidthCudaLauncher(const cudaStream_t *stream, const NDArray& input, const NDArray& range, NDArray& output) {
 
-        __shared__ T const* x;
-        __shared__ Nd4jLong* z; // output buffer
+    const T leftEdge  = range.e<T>(0);
+    const T rightEdge = range.e<T>(1);
 
-        if (threadIdx.x == 0) {
-            z = reinterpret_cast<Nd4jLong*>(outputBuffer);
-            x = reinterpret_cast<T const*>(inputBuffer);
-        }
-        __syncthreads();
-        auto tid = blockIdx.x * gridDim.x + threadIdx.x;
-        auto step = blockDim.x * gridDim.x;
+    histogramFixedWidthCuda<T><<<512, MAX_NUM_THREADS / 2, 512, *stream>>>(input.getSpecialBuffer(), input.getSpecialShapeInfo(), output.specialBuffer(), output.specialShapeInfo(), leftEdge, rightEdge);
+}
 
-        for(auto i = tid; i < inputLength; i += step) {
+////////////////////////////////////////////////////////////////////////
+void histogramFixedWidth(nd4j::LaunchContext* context, const NDArray& input, const NDArray& range, NDArray& output) {
 
-            const T value = x[shape::getIndexOffset(i, inputShape, inputLength)];
-            Nd4jLong currInd = static_cast<Nd4jLong>((value - leftEdge) / binWidth);
+    // firstly initialize output with zeros
+    output.nullify();
 
-            if(value < secondEdge)
-                currInd = 0;
-            else if(value >= lastButOneEdge)
-                currInd = outputLength - 1;
-            nd4j::math::atomics::nd4j_atomicAdd(&z[currInd], 1LL);
-        }
-    }
+    PointersManager manager(context, "histogramFixedWidth");
+
+    NDArray::prepareSpecialUse({&output}, {&input});
+    BUILD_SINGLE_SELECTOR(input.dataType(), histogramFixedWidthCudaLauncher, (context->getCudaStream(), input, range, output), LIBND4J_TYPES);
+    NDArray::registerSpecialUse({&output}, {&input});
+
+    manager.synchronize();
+}
 
 
-    template <typename T>
-    void histogramFixedWidth_(nd4j::LaunchContext * context, const NDArray& input, const NDArray& range, NDArray& output) {
-        const int nbins = output.lengthOf();
-        auto stream = context->getCudaStream();
-        // firstly initialize output with zeros
-        //if(output.ews() == 1)
-        //    memset(output.buffer(), 0, nbins * output.sizeOfT());
-        //else
-        output.assign(0);
-        if (!input.isActualOnDeviceSide())
-            input.syncToDevice();
+//     template <typename T>
+//     __global__ static void copyBuffers(Nd4jLong* destination, void const* source, Nd4jLong* sourceShape, Nd4jLong bufferLength) {
+//         const auto tid = blockIdx.x * gridDim.x + threadIdx.x;
+//         const auto step = gridDim.x * blockDim.x;
+//         for (int t = tid; t < bufferLength; t += step) {
+//             destination[t] = reinterpret_cast<T const*>(source)[shape::getIndexOffset(t, sourceShape, bufferLength)];
+//         }
+//     }
 
-        const double leftEdge  = range.e<double>(0);
-        const double rightEdge = range.e<double>(1);
+//     template <typename T>
+//     __global__ static void returnBuffers(void* destination, Nd4jLong const* source, Nd4jLong* destinationShape, Nd4jLong bufferLength) {
+//         const auto tid = blockIdx.x * gridDim.x + threadIdx.x;
+//         const auto step = gridDim.x * blockDim.x;
+//         for (int t = tid; t < bufferLength; t += step) {
+//             reinterpret_cast<T*>(destination)[shape::getIndexOffset(t, destinationShape, bufferLength)] = source[t];
+//         }
+//     }
 
-        const double binWidth       = (rightEdge - leftEdge ) / nbins;
-        const double secondEdge     = leftEdge + binWidth;
-        double lastButOneEdge = rightEdge - binWidth;
-        Nd4jLong* outputBuffer;
-        cudaError_t err = cudaMalloc(&outputBuffer, output.lengthOf() * sizeof(Nd4jLong));
-        if (err != 0)
-            throw cuda_exception::build("helpers::histogramFixedWidth: Cannot allocate memory for output", err);
-        copyBuffers<Nd4jLong ><<<256, 512, 8192, *stream>>>(outputBuffer, output.getSpecialBuffer(), output.getSpecialShapeInfo(), output.lengthOf());
-        histogramFixedWidthKernel<T><<<256, 512, 8192, *stream>>>(outputBuffer, output.lengthOf(), input.getSpecialBuffer(), input.getSpecialShapeInfo(), input.lengthOf(), leftEdge, binWidth, secondEdge, lastButOneEdge);
-        returnBuffers<Nd4jLong><<<256, 512, 8192, *stream>>>(output.specialBuffer(), outputBuffer, output.specialShapeInfo(), output.lengthOf());
-        //cudaSyncStream(*stream);
-        err = cudaFree(outputBuffer);
-        if (err != 0)
-            throw cuda_exception::build("helpers::histogramFixedWidth: Cannot deallocate memory for output buffer", err);
-        output.tickWriteDevice();
-//#pragma omp parallel for schedule(guided)
-//        for(Nd4jLong i = 0; i < input.lengthOf(); ++i) {
-//
-//            const T value = input.e<T>(i);
-//
-//            if(value < secondEdge)
-//#pragma omp critical
-//                output.p<Nd4jLong>(0, output.e<Nd4jLong>(0) + 1);
-//            else if(value >= lastButOneEdge)
-//#pragma omp critical
-//                output.p<Nd4jLong>(nbins-1, output.e<Nd4jLong>(nbins-1) + 1);
-//            else {
-//                Nd4jLong currInd = static_cast<Nd4jLong>((value - leftEdge) / binWidth);
-//#pragma omp critical
-//                output.p<Nd4jLong>(currInd, output.e<Nd4jLong>(currInd) + 1);
-//            }
-//        }
-    }
+//     template <typename T>
+//     static __global__ void histogramFixedWidthKernel(void* outputBuffer, Nd4jLong outputLength, void const* inputBuffer, Nd4jLong* inputShape, Nd4jLong inputLength, double const leftEdge, double binWidth, double secondEdge, double lastButOneEdge) {
 
-    void histogramFixedWidth(nd4j::LaunchContext * context, const NDArray& input, const NDArray& range, NDArray& output) {
-        BUILD_SINGLE_SELECTOR(input.dataType(), histogramFixedWidth_, (context, input, range, output), LIBND4J_TYPES);
-    }
-    BUILD_SINGLE_TEMPLATE(template void histogramFixedWidth_, (nd4j::LaunchContext * context, const NDArray& input, const NDArray& range, NDArray& output), LIBND4J_TYPES);
+//         __shared__ T const* x;
+//         __shared__ Nd4jLong* z; // output buffer
+
+//         if (threadIdx.x == 0) {
+//             z = reinterpret_cast<Nd4jLong*>(outputBuffer);
+//             x = reinterpret_cast<T const*>(inputBuffer);
+//         }
+//         __syncthreads();
+//         auto tid = blockIdx.x * gridDim.x + threadIdx.x;
+//         auto step = blockDim.x * gridDim.x;
+
+//         for(auto i = tid; i < inputLength; i += step) {
+
+//             const T value = x[shape::getIndexOffset(i, inputShape, inputLength)];
+//             Nd4jLong currInd = static_cast<Nd4jLong>((value - leftEdge) / binWidth);
+
+//             if(value < secondEdge)
+//                 currInd = 0;
+//             else if(value >= lastButOneEdge)
+//                 currInd = outputLength - 1;
+//             nd4j::math::atomics::nd4j_atomicAdd(&z[currInd], 1LL);
+//         }
+//     }
+
+
+//     template <typename T>
+//     void histogramFixedWidth_(nd4j::LaunchContext * context, const NDArray& input, const NDArray& range, NDArray& output) {
+//         const int nbins = output.lengthOf();
+//         auto stream = context->getCudaStream();
+//         // firstly initialize output with zeros
+//         //if(output.ews() == 1)
+//         //    memset(output.buffer(), 0, nbins * output.sizeOfT());
+//         //else
+//         output.assign(0);
+//         if (!input.isActualOnDeviceSide())
+//             input.syncToDevice();
+
+//         const double leftEdge  = range.e<double>(0);
+//         const double rightEdge = range.e<double>(1);
+
+//         const double binWidth       = (rightEdge - leftEdge ) / nbins;
+//         const double secondEdge     = leftEdge + binWidth;
+//         double lastButOneEdge = rightEdge - binWidth;
+//         Nd4jLong* outputBuffer;
+//         cudaError_t err = cudaMalloc(&outputBuffer, output.lengthOf() * sizeof(Nd4jLong));
+//         if (err != 0)
+//             throw cuda_exception::build("helpers::histogramFixedWidth: Cannot allocate memory for output", err);
+//         copyBuffers<Nd4jLong ><<<256, 512, 8192, *stream>>>(outputBuffer, output.getSpecialBuffer(), output.getSpecialShapeInfo(), output.lengthOf());
+//         histogramFixedWidthKernel<T><<<256, 512, 8192, *stream>>>(outputBuffer, output.lengthOf(), input.getSpecialBuffer(), input.getSpecialShapeInfo(), input.lengthOf(), leftEdge, binWidth, secondEdge, lastButOneEdge);
+//         returnBuffers<Nd4jLong><<<256, 512, 8192, *stream>>>(output.specialBuffer(), outputBuffer, output.specialShapeInfo(), output.lengthOf());
+//         //cudaSyncStream(*stream);
+//         err = cudaFree(outputBuffer);
+//         if (err != 0)
+//             throw cuda_exception::build("helpers::histogramFixedWidth: Cannot deallocate memory for output buffer", err);
+//         output.tickWriteDevice();
+// //#pragma omp parallel for schedule(guided)
+// //        for(Nd4jLong i = 0; i < input.lengthOf(); ++i) {
+// //
+// //            const T value = input.e<T>(i);
+// //
+// //            if(value < secondEdge)
+// //#pragma omp critical
+// //                output.p<Nd4jLong>(0, output.e<Nd4jLong>(0) + 1);
+// //            else if(value >= lastButOneEdge)
+// //#pragma omp critical
+// //                output.p<Nd4jLong>(nbins-1, output.e<Nd4jLong>(nbins-1) + 1);
+// //            else {
+// //                Nd4jLong currInd = static_cast<Nd4jLong>((value - leftEdge) / binWidth);
+// //#pragma omp critical
+// //                output.p<Nd4jLong>(currInd, output.e<Nd4jLong>(currInd) + 1);
+// //            }
+// //        }
+//     }
+
+//     void histogramFixedWidth(nd4j::LaunchContext * context, const NDArray& input, const NDArray& range, NDArray& output) {
+//         BUILD_SINGLE_SELECTOR(input.dataType(), histogramFixedWidth_, (context, input, range, output), LIBND4J_TYPES);
+//     }
+//     BUILD_SINGLE_TEMPLATE(template void histogramFixedWidth_, (nd4j::LaunchContext * context, const NDArray& input, const NDArray& range, NDArray& output), LIBND4J_TYPES);
 
 }
 }
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/scatter.cu b/libnd4j/include/ops/declarable/helpers/cuda/scatter.cu
index ec0d304df..54d350f47 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/scatter.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/scatter.cu
@@ -398,10 +398,15 @@ void scatter(nd4j::LaunchContext  *context, pairwise::Ops op, const NDArray& ind
         const int xRank = indices.rankOf();
 
         std::vector<int> zTadDims = ShapeUtils::evalDimsToExclude(output.rankOf(), {0});
-        std::vector<int> yTadDims(xRank);
-        std::iota(yTadDims.begin(), yTadDims.end(), xRank == 1 ? 0 : xRank);
 
-        auto packY = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(updates.getShapeInfo(), yTadDims);
+        int sizeOfUpdDims = xRank;
+        if(output.rankOf() == updates.rankOf() && indices.isVector())
+            sizeOfUpdDims = 1;
+
+        std::vector<int> yTadDims(sizeOfUpdDims);
+        std::iota(yTadDims.begin(), yTadDims.end(), 0);
+
+        auto packY = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(updates.getShapeInfo(), ShapeUtils::evalDimsToExclude(updates.rankOf(), yTadDims));
         auto packZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(output.getShapeInfo(), zTadDims);
 
         const Nd4jLong zTadLen = shape::length(packZ.primaryShapeInfo());
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests10.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests10.cpp
index 1d4bf7338..82ed21709 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests10.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests10.cpp
@@ -910,7 +910,31 @@ TEST_F(DeclarableOpsTests10, histogram_fixed_width_test5) {
     auto *out = results->at(0);
 
     ASSERT_TRUE(exp.isSameShape(out));
-    out->printBuffer("5HIST");
+    // out->printBuffer("5HIST");
+    ASSERT_TRUE(exp.equalsTo(out));
+
+    delete results;
+}
+
+///////////////////////////////////////////////////////////////////
+TEST_F(DeclarableOpsTests10, histogram_fixed_width_test6) {
+
+    auto input = NDArrayFactory::create<double>('c', {7},{0.0, 0.1, 0.1, 0.3, 0.5, 0.5, 0.9});
+    auto range = NDArrayFactory::create<double>('c', {2}, {0, 1});
+    auto bins  = NDArrayFactory::create<int>(5);
+
+    auto exp = NDArrayFactory::create<Nd4jLong>('c', {5}, {3, 1, 2, 0, 1});
+
+    nd4j::ops::histogram_fixed_width op;
+    auto results = op.execute({&input, &range, &bins}, {}, {}, {});
+
+    ASSERT_EQ(ND4J_STATUS_OK, results->status());
+
+    auto out = results->at(0);
+    // out->printShapeInfo();
+    // out->printIndexedBuffer();
+
+    ASSERT_TRUE(exp.isSameShape(out));
     ASSERT_TRUE(exp.equalsTo(out));
 
     delete results;
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests15.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests15.cpp
index df1421d71..21a0381e9 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests15.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests15.cpp
@@ -249,7 +249,7 @@ TEST_F(DeclarableOpsTests15, Test_layer_norm_1) {
     auto b = NDArrayFactory::create<float>('c', {1, 5}, {1., 2., 3., 4., 5.});
 
     nd4j::ops::layer_norm op;
-    auto result = op.execute({&x, &g, &b}, {}, {0}, {});
+    auto result = op.execute({&x, &g, &b}, {}, {0}, {false});
     ASSERT_EQ(Status::OK(), result->status());
     delete result;
 }
@@ -261,7 +261,7 @@ TEST_F(DeclarableOpsTests15, Test_layer_norm_bp_1) {
     auto eps = NDArrayFactory::create<float>('c', {1, 5}, {0., 0., 0., 0., 0.});
 
     nd4j::ops::layer_norm_bp op;
-    auto result = op.execute({&x, &g, &b, &eps}, {}, {0}, {});
+    auto result = op.execute({&x, &g, &b, &eps}, {}, {0}, {false});
     ASSERT_EQ(Status::OK(), result->status());
     delete result;
 }
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
index a23d5421e..992b21c0f 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
@@ -39,7 +39,7 @@ public:
     }
 };
 
-TEST_F(DeclarableOpsTests16, test_scatter_update_119) {
+TEST_F(DeclarableOpsTests16, scatter_upd_1) {
     auto x = NDArrayFactory::create<float>('c', {3}, {1, 1, 1});
     auto y = NDArrayFactory::create<int>(0);
     auto w = NDArrayFactory::create<float>(3.0f);
@@ -56,6 +56,27 @@ TEST_F(DeclarableOpsTests16, test_scatter_update_119) {
     delete result;
 }
 
+TEST_F(DeclarableOpsTests16, scatter_upd_2) {
+
+    NDArray x('c', {10, 3}, nd4j::DataType::FLOAT32);
+    NDArray indices('c', {2}, {2,5}, nd4j::DataType::INT32);
+    NDArray updates('c', {2, 3}, {100,101,102,  200,201,202}, nd4j::DataType::FLOAT32);
+    NDArray e('c', {10, 3}, {1,2,3, 4,5,6, 100,101,102, 10,11,12, 13,14,15, 200,201,202, 19,20,21, 22,23,24, 25,26,27, 28,29,30}, nd4j::DataType::FLOAT32);
+
+    x.linspace(1);
+
+    nd4j::ops::scatter_upd op;
+    auto result = op.execute({&x, &indices, &updates}, {}, {});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    auto z = result->at(0);
+
+    ASSERT_EQ(e, *z);
+
+    delete result;
+}
+
+
 TEST_F(DeclarableOpsTests16, test_size_dtype_1) {
     auto x = NDArrayFactory::create<float>('c', {3}, {1, 1, 1});
     auto z = NDArrayFactory::create<float>(0.0f);
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
index 9437ad7b2..84bd96ad6 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
@@ -1297,7 +1297,6 @@ public class LayerOpValidation extends BaseOpValidation {
     }
 
     @Test
-    @Ignore("AB 2019/06/24 - Failing: Ignored to get to all passing baseline to prevent regressions via CI - see issue #7912")
     public void testLayerNormMixedOrders(){
         Nd4j.getRandom().setSeed(12345);
         INDArray input = Nd4j.rand(DataType.DOUBLE, 3, 8).dup('f');

From 25e5c23eae642c21c0bf43fc0df16113ea6dea29 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Mon, 26 Aug 2019 19:57:51 +0300
Subject: [PATCH 12/56] [WIP] Error handling (#169)

* CUDA reverse rewrite + couple of tests

Signed-off-by: raver119 <raver119@gmail.com>

* don't throw exception on invalid pointer

Signed-off-by: raver119 <raver119@gmail.com>

* data types validation for fastpath exec mode + 2 tests

Signed-off-by: raver119 <raver119@gmail.com>

* data types validation for fastpath exec mode + 2 tests

Signed-off-by: raver119 <raver119@gmail.com>

* ismax allowed dtypes tweak

Signed-off-by: raver119 <raver119@gmail.com>

* lastErrorCode + lastErrorMessage for native exceptions handling

Signed-off-by: raver119 <raver119@gmail.com>

* exportable ErrorReference

Signed-off-by: raver119 <raver119@gmail.com>

* check error codes in java

Signed-off-by: raver119 <raver119@gmail.com>

* - consume lastErrorCode
- fast_in dtype validation fix

Signed-off-by: raver119 <raver119@gmail.com>

* - sg/cb allowed output type change
- minor logging fix for data type validation

Signed-off-by: raver119 <raver119@gmail.com>
---
 libnd4j/blas/NativeOps.h                      |   44 +-
 libnd4j/blas/cpu/NativeOps.cpp                | 1798 ++++++------
 libnd4j/blas/cuda/NativeOps.cu                | 2417 +++++++++--------
 libnd4j/include/execution/ContextBuffers.h    |    4 +
 .../ErrorReference.h}                         |   40 +-
 libnd4j/include/execution/LaunchContext.h     |    4 +
 .../include/execution/cpu/ContextBuffers.cpp  |    4 +
 .../include/execution/cpu/LaunchContext.cpp   |    8 +
 .../include/execution/cuda/ContextBuffers.cu  |    4 +
 .../include/execution/cuda/LaunchContext.cu   |    4 +
 .../impl/ErrorReference.cpp}                  |   51 +-
 .../ops/declarable/generic/convo/ismax.cpp    |    2 +-
 .../ops/declarable/generic/nlp/cbow.cpp       |    3 +-
 .../ops/declarable/generic/nlp/skipgram.cpp   |    2 +-
 .../ops/declarable/generic/nn/softmax.cpp     |    2 +-
 .../ops/declarable/helpers/cuda/reverse.cu    |  114 +-
 .../ops/declarable/impl/DeclarableOp.cpp      |  199 +-
 .../layers_tests/JavaInteropTests.cpp         |   26 +
 .../tests_cpu/layers_tests/NativeOpsTests.cpp |   75 -
 .../java/org/nd4j/nativeblas/NativeOps.java   |   40 +-
 .../allocator/pointers/cuda/cudaEvent_t.java  |    9 +-
 .../allocator/pointers/cuda/cudaStream_t.java |    5 +-
 .../linalg/jcublas/JCublasNDArrayFactory.java |  261 +-
 .../ops/executioner/CudaExecutioner.java      |  184 +-
 .../java/org/nd4j/nativeblas/Nd4jCuda.java    |  139 +-
 .../org/nd4j/nativeblas/Nd4jCudaPresets.java  |    1 +
 .../cpu/nativecpu/CpuNDArrayFactory.java      |  187 +-
 .../nativecpu/ops/NativeOpExecutioner.java    |  209 +-
 .../java/org/nd4j/nativeblas/Nd4jCpu.java     |  139 +-
 .../org/nd4j/nativeblas/Nd4jCpuPresets.java   |    1 +
 .../test/java/org/nd4j/linalg/Nd4jTestsC.java |   98 +-
 31 files changed, 3002 insertions(+), 3072 deletions(-)
 rename libnd4j/include/{helpers/ProviderRNG.h => execution/ErrorReference.h} (57%)
 rename libnd4j/include/{helpers/impl/ProviderRNG.cpp => execution/impl/ErrorReference.cpp} (52%)

diff --git a/libnd4j/blas/NativeOps.h b/libnd4j/blas/NativeOps.h
index 9ce90176f..9bca7bb10 100755
--- a/libnd4j/blas/NativeOps.h
+++ b/libnd4j/blas/NativeOps.h
@@ -79,6 +79,18 @@ bool verbose = false;
 
 extern "C" {
 
+/**
+ * This function returns last error code stored,
+ * @return non-zero if something bad happened
+ */
+ND4J_EXPORT int lastErrorCode();
+
+/**
+ * This function returns last error message, if last error code > 0
+ * @return
+ */
+ND4J_EXPORT const char* lastErrorMessage();
+
 /**
  *
  * @param p
@@ -557,38 +569,6 @@ ND4J_EXPORT void execScalarBoolTad(Nd4jPointer *extraPointers,
                 Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
                 Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ);
 
-
-/**
-* Append an input array
-* to the end of a flat array
-* in a particular order
-* @param offset the offset of the array to start at
-* @param order the order
-* @param result the result array
-* @param resultShapeInfo the shape info for te array
-* @param input the input for the array
-* @param inputShapeInfo the shape information for that array
-*/
-ND4J_EXPORT void flatten(
-        Nd4jPointer *extraPointers,
-        int offset,
-        char order,
-        void *result, Nd4jLong *resultShapeInfo,
-        void *dresult, Nd4jLong *dresultShapeInfo,
-        void *input, Nd4jLong *inputShapeInfo,
-        void *dinput, Nd4jLong *dinputShapeInfo);
-
-ND4J_EXPORT void concat(
-        Nd4jPointer *extraPointers,
-        int dimension,
-        int numArrays,
-        Nd4jPointer *data, Nd4jPointer *inputShapeInfo,
-        Nd4jPointer *ddata, Nd4jPointer *dinputShapeInfo,
-        void *result, Nd4jLong *resultShapeInfo,
-        void *dresult, Nd4jLong *dresultShapeInfo,
-        Nd4jPointer *tadPointers, Nd4jPointer *offsetPointers);
-
-
 ND4J_EXPORT void specialConcat (
         Nd4jPointer *extraPointers,
         int dimension,
diff --git a/libnd4j/blas/cpu/NativeOps.cpp b/libnd4j/blas/cpu/NativeOps.cpp
index f5d4996e4..86bc04fc4 100644
--- a/libnd4j/blas/cpu/NativeOps.cpp
+++ b/libnd4j/blas/cpu/NativeOps.cpp
@@ -102,8 +102,12 @@ void execIndexReduceScalar(Nd4jPointer *extraPointers,
                                                 void *extraParams,
                                                 void *hZ, Nd4jLong *hZShapeInfo,
                                                 void *dZ, Nd4jLong *dZShapeInfo) {
-
-    NativeOpExecutioner::execIndexReduceScalar(nullptr, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    try {
+        NativeOpExecutioner::execIndexReduceScalar(nullptr, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -125,31 +129,36 @@ void  execIndexReduce(Nd4jPointer *extraPointers,int opNum,
                                         void *dZ, Nd4jLong *dZShapeInfo,
                                         void *hDimension, Nd4jLong *hDimensionShape,
                                         void *dDimension, Nd4jLong *dDimensionShape) {
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension,
+                                                                                dimensionLength);
 
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension, dimensionLength);
+        auto hTADShapeInfo = tadPack.primaryShapeInfo();
+        auto hTADOffsets = tadPack.primaryOffsets();
 
-    auto hTADShapeInfo = tadPack.primaryShapeInfo();
-    auto hTADOffsets = tadPack.primaryOffsets();
+        auto hz = reinterpret_cast<Nd4jLong *>(hZ);
 
-    auto hz = reinterpret_cast<Nd4jLong*>(hZ);
-
-    NativeOpExecutioner::execIndexReduce(nullptr, opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            extraParams,
-            hz,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            dimension,
-            dimensionLength,
-            hTADShapeInfo,
-            hTADOffsets);
+        NativeOpExecutioner::execIndexReduce(nullptr, opNum,
+                                             hX,
+                                             hXShapeInfo,
+                                             dX,
+                                             dXShapeInfo,
+                                             extraParams,
+                                             hz,
+                                             hZShapeInfo,
+                                             dZ,
+                                             dZShapeInfo,
+                                             dimension,
+                                             dimensionLength,
+                                             hTADShapeInfo,
+                                             hTADOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -175,31 +184,38 @@ void execBroadcast(Nd4jPointer *extraPointers,
                                       void *dZ, Nd4jLong *dZShapeInfo,
                                       void *hDimension, Nd4jLong *hDimensionShape,
                                       void *dDimension, Nd4jLong *dDimensionShape) {
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto tadPackX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension, dimensionLength);
-    auto tadPackZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hZShapeInfo, dimension, dimensionLength);
+        auto tadPackX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension,
+                                                                                 dimensionLength);
+        auto tadPackZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hZShapeInfo, dimension,
+                                                                                 dimensionLength);
 
-    auto hTADShapeInfo = tadPackX.primaryShapeInfo();
-    auto hTADOffsets = tadPackX.primaryOffsets();
-    auto hTADShapeInfoZ = tadPackZ.primaryShapeInfo();
-    auto hTADOffsetsZ = tadPackZ.primaryOffsets();
+        auto hTADShapeInfo = tadPackX.primaryShapeInfo();
+        auto hTADOffsets = tadPackX.primaryOffsets();
+        auto hTADShapeInfoZ = tadPackZ.primaryShapeInfo();
+        auto hTADOffsetsZ = tadPackZ.primaryOffsets();
 
-    NativeOpExecutioner::execBroadcast(nullptr,
-                                      opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            hY,
-            hYShapeInfo,
-            dY,
-            dYShapeInfo,
-            hZ, hZShapeInfo,
-            dZ, dZShapeInfo,
-            dimension,
-            dimensionLength, hTADShapeInfo, hTADOffsets, hTADShapeInfoZ, hTADOffsetsZ);
+        NativeOpExecutioner::execBroadcast(nullptr,
+                                           opNum,
+                                           hX,
+                                           hXShapeInfo,
+                                           dX,
+                                           dXShapeInfo,
+                                           hY,
+                                           hYShapeInfo,
+                                           dY,
+                                           dYShapeInfo,
+                                           hZ, hZShapeInfo,
+                                           dZ, dZShapeInfo,
+                                           dimension,
+                                           dimensionLength, hTADShapeInfo, hTADOffsets, hTADShapeInfoZ, hTADOffsetsZ);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execBroadcastBool(Nd4jPointer *extraPointers,
@@ -212,31 +228,39 @@ void execBroadcastBool(Nd4jPointer *extraPointers,
                               void *dZ, Nd4jLong *dZShapeInfo,
                                   void *hDimension, Nd4jLong *hDimensionShape,
                                   void *dDimension, Nd4jLong *dDimensionShape) {
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto tadPackX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension, dimensionLength);
-    auto tadPackZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hZShapeInfo, dimension, dimensionLength);
+        auto tadPackX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension,
+                                                                                 dimensionLength);
+        auto tadPackZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hZShapeInfo, dimension,
+                                                                                 dimensionLength);
 
-    auto hTADShapeInfo = tadPackX.primaryShapeInfo();
-    auto hTADOffsets = tadPackX.primaryOffsets();
-    auto hTADShapeInfoZ = tadPackZ.primaryShapeInfo();
-    auto hTADOffsetsZ = tadPackZ.primaryOffsets();
+        auto hTADShapeInfo = tadPackX.primaryShapeInfo();
+        auto hTADOffsets = tadPackX.primaryOffsets();
+        auto hTADShapeInfoZ = tadPackZ.primaryShapeInfo();
+        auto hTADOffsetsZ = tadPackZ.primaryOffsets();
 
-    NativeOpExecutioner::execBroadcastBool(nullptr,
-                                          opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            hY,
-            hYShapeInfo,
-            dY,
-            dYShapeInfo,
-            hZ, hZShapeInfo,
-            dZ, dZShapeInfo,
-            dimension,
-            dimensionLength, hTADShapeInfo, hTADOffsets, hTADShapeInfoZ, hTADOffsetsZ);
+        NativeOpExecutioner::execBroadcastBool(nullptr,
+                                               opNum,
+                                               hX,
+                                               hXShapeInfo,
+                                               dX,
+                                               dXShapeInfo,
+                                               hY,
+                                               hYShapeInfo,
+                                               dY,
+                                               dYShapeInfo,
+                                               hZ, hZShapeInfo,
+                                               dZ, dZShapeInfo,
+                                               dimension,
+                                               dimensionLength, hTADShapeInfo, hTADOffsets, hTADShapeInfoZ,
+                                               hTADOffsetsZ);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -261,21 +285,26 @@ void execPairwiseTransform(
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo,
         void *extraParams) {
-    NativeOpExecutioner::execPairwiseTransform(nullptr,
-                                              opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            hY,
-            hYShapeInfo,
-            dY,
-            dYShapeInfo,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            extraParams);
+    try {
+        NativeOpExecutioner::execPairwiseTransform(nullptr,
+                                                   opNum,
+                                                   hX,
+                                                   hXShapeInfo,
+                                                   dX,
+                                                   dXShapeInfo,
+                                                   hY,
+                                                   hYShapeInfo,
+                                                   dY,
+                                                   dYShapeInfo,
+                                                   hZ,
+                                                   hZShapeInfo,
+                                                   dZ,
+                                                   dZShapeInfo,
+                                                   extraParams);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execPairwiseTransformBool(
@@ -288,21 +317,27 @@ void execPairwiseTransformBool(
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo,
         void *extraParams) {
-    NativeOpExecutioner::execPairwiseBoolTransform(nullptr,
-                                                  opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            hY,
-            hYShapeInfo,
-            dY,
-            dYShapeInfo,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            extraParams);
+
+    try {
+        NativeOpExecutioner::execPairwiseBoolTransform(nullptr,
+                                                       opNum,
+                                                       hX,
+                                                       hXShapeInfo,
+                                                       dX,
+                                                       dXShapeInfo,
+                                                       hY,
+                                                       hYShapeInfo,
+                                                       dY,
+                                                       dYShapeInfo,
+                                                       hZ,
+                                                       hZShapeInfo,
+                                                       dZ,
+                                                       dZShapeInfo,
+                                                       extraParams);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -323,18 +358,22 @@ void execReduceFloat(
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo) {
 
-    NativeOpExecutioner::execReduceFloatScalar(nullptr,
-                                              opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            extraParams,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo);
-
+    try {
+        NativeOpExecutioner::execReduceFloatScalar(nullptr,
+                                                   opNum,
+                                                   hX,
+                                                   hXShapeInfo,
+                                                   dX,
+                                                   dXShapeInfo,
+                                                   extraParams,
+                                                   hZ,
+                                                   hZShapeInfo,
+                                                   dZ,
+                                                   dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execReduceSame(
@@ -346,18 +385,22 @@ void execReduceSame(
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo) {
 
-    NativeOpExecutioner::execReduceSameScalar(nullptr,
-                                             opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            extraParams,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo);
-
+    try {
+        NativeOpExecutioner::execReduceSameScalar(nullptr,
+                                                  opNum,
+                                                  hX,
+                                                  hXShapeInfo,
+                                                  dX,
+                                                  dXShapeInfo,
+                                                  extraParams,
+                                                  hZ,
+                                                  hZShapeInfo,
+                                                  dZ,
+                                                  dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execReduceBool(
@@ -368,19 +411,22 @@ void execReduceBool(
         void *extraParams,
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo) {
-
-    NativeOpExecutioner::execReduceBoolScalar(nullptr,
-                                             opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            extraParams,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo);
-
+    try {
+        NativeOpExecutioner::execReduceBoolScalar(nullptr,
+                                                  opNum,
+                                                  hX,
+                                                  hXShapeInfo,
+                                                  dX,
+                                                  dXShapeInfo,
+                                                  extraParams,
+                                                  hZ,
+                                                  hZShapeInfo,
+                                                  dZ,
+                                                  dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execReduceLong(
@@ -391,19 +437,22 @@ void execReduceLong(
         void *extraParams,
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo) {
-
-    NativeOpExecutioner::execReduceLongScalar(nullptr,
-                                             opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            extraParams,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo);
-
+    try {
+        NativeOpExecutioner::execReduceLongScalar(nullptr,
+                                                  opNum,
+                                                  hX,
+                                                  hXShapeInfo,
+                                                  dX,
+                                                  dXShapeInfo,
+                                                  extraParams,
+                                                  hZ,
+                                                  hZShapeInfo,
+                                                  dZ,
+                                                  dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -424,28 +473,34 @@ void execReduceFloat2(Nd4jPointer *extraPointers,
                                    void *dZ, Nd4jLong *dZShapeInfo,
                                 void *hDimension, Nd4jLong *hDimensionShape,
                                 void *dDimension, Nd4jLong *dDimensionShape) {
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto tadPackX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension, dimensionLength);
+        auto tadPackX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension,
+                                                                                 dimensionLength);
 
-    auto hTADShapeInfo = tadPackX.primaryShapeInfo();
-    auto hTADOffsets = tadPackX.primaryOffsets();
+        auto hTADShapeInfo = tadPackX.primaryShapeInfo();
+        auto hTADOffsets = tadPackX.primaryOffsets();
 
-    NativeOpExecutioner::execReduceFloat(nullptr, opNum,
-                                           hX,
-                                           hXShapeInfo,
-                                           dX,
-                                           dXShapeInfo,
-                                           extraParams,
-                                           hZ,
-                                           hZShapeInfo,
-                                           dZ,
-                                           dZShapeInfo,
-                                           dimension,
-                                           dimensionLength,
-                                           hTADShapeInfo,
-                                           hTADOffsets);
+        NativeOpExecutioner::execReduceFloat(nullptr, opNum,
+                                             hX,
+                                             hXShapeInfo,
+                                             dX,
+                                             dXShapeInfo,
+                                             extraParams,
+                                             hZ,
+                                             hZShapeInfo,
+                                             dZ,
+                                             dZShapeInfo,
+                                             dimension,
+                                             dimensionLength,
+                                             hTADShapeInfo,
+                                             hTADOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execReduceBool2(Nd4jPointer *extraPointers,
@@ -457,28 +512,34 @@ void execReduceBool2(Nd4jPointer *extraPointers,
                                 void *dZ, Nd4jLong *dZShapeInfo,
                                void *hDimension, Nd4jLong *hDimensionShape,
                                void *dDimension, Nd4jLong *dDimensionShape) {
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension, dimensionLength);
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension,
+                                                                                dimensionLength);
 
-    auto hTADShapeInfo = tadPack.primaryShapeInfo();
-    auto hTADOffsets = tadPack.primaryOffsets();
+        auto hTADShapeInfo = tadPack.primaryShapeInfo();
+        auto hTADOffsets = tadPack.primaryOffsets();
 
-    NativeOpExecutioner::execReduceBool(nullptr, opNum,
-                                        hX,
-                                        hXShapeInfo,
-                                        dX,
-                                        dXShapeInfo,
-                                        extraParams,
-                                        hZ,
-                                        hZShapeInfo,
-                                        dZ,
-                                        dZShapeInfo,
-                                        dimension,
-                                        dimensionLength,
-                                        hTADShapeInfo,
-                                        hTADOffsets);
+        NativeOpExecutioner::execReduceBool(nullptr, opNum,
+                                            hX,
+                                            hXShapeInfo,
+                                            dX,
+                                            dXShapeInfo,
+                                            extraParams,
+                                            hZ,
+                                            hZShapeInfo,
+                                            dZ,
+                                            dZShapeInfo,
+                                            dimension,
+                                            dimensionLength,
+                                            hTADShapeInfo,
+                                            hTADOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execReduceSame2(Nd4jPointer *extraPointers,
@@ -490,28 +551,34 @@ void execReduceSame2(Nd4jPointer *extraPointers,
                                 void *dZ, Nd4jLong *dZShapeInfo,
                                void *hDimension, Nd4jLong *hDimensionShape,
                                void *dDimension, Nd4jLong *dDimensionShape) {
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension, dimensionLength);
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension,
+                                                                                dimensionLength);
 
-    auto hTADShapeInfo = tadPack.primaryShapeInfo();
-    auto hTADOffsets = tadPack.primaryOffsets();
+        auto hTADShapeInfo = tadPack.primaryShapeInfo();
+        auto hTADOffsets = tadPack.primaryOffsets();
 
-    NativeOpExecutioner::execReduceSame(nullptr, opNum,
-                                        hX,
-                                        hXShapeInfo,
-                                        dX,
-                                        dXShapeInfo,
-                                        extraParams,
-                                        hZ,
-                                        hZShapeInfo,
-                                        dZ,
-                                        dZShapeInfo,
-                                        dimension,
-                                        dimensionLength,
-                                        hTADShapeInfo,
-                                        hTADOffsets);
+        NativeOpExecutioner::execReduceSame(nullptr, opNum,
+                                            hX,
+                                            hXShapeInfo,
+                                            dX,
+                                            dXShapeInfo,
+                                            extraParams,
+                                            hZ,
+                                            hZShapeInfo,
+                                            dZ,
+                                            dZShapeInfo,
+                                            dimension,
+                                            dimensionLength,
+                                            hTADShapeInfo,
+                                            hTADOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execReduceLong2(Nd4jPointer *extraPointers,
@@ -523,28 +590,34 @@ void execReduceLong2(Nd4jPointer *extraPointers,
                                 void *dZ, Nd4jLong *dZShapeInfo,
                                void *hDimension, Nd4jLong *hDimensionShape,
                                void *dDimension, Nd4jLong *dDimensionShape) {
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension, dimensionLength);
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension,
+                                                                                dimensionLength);
 
-    auto hTADShapeInfo = tadPack.primaryShapeInfo();
-    auto hTADOffsets = tadPack.primaryOffsets();
+        auto hTADShapeInfo = tadPack.primaryShapeInfo();
+        auto hTADOffsets = tadPack.primaryOffsets();
 
-    NativeOpExecutioner::execReduceLong(nullptr, opNum,
-                                        hX,
-                                        hXShapeInfo,
-                                        dX,
-                                        dXShapeInfo,
-                                        extraParams,
-                                        hZ,
-                                        hZShapeInfo,
-                                        dZ,
-                                        dZShapeInfo,
-                                        dimension,
-                                        dimensionLength,
-                                        hTADShapeInfo,
-                                        hTADOffsets);
+        NativeOpExecutioner::execReduceLong(nullptr, opNum,
+                                            hX,
+                                            hXShapeInfo,
+                                            dX,
+                                            dXShapeInfo,
+                                            extraParams,
+                                            hZ,
+                                            hZShapeInfo,
+                                            dZ,
+                                            dZShapeInfo,
+                                            dimension,
+                                            dimensionLength,
+                                            hTADShapeInfo,
+                                            hTADOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -567,8 +640,13 @@ void execReduce3(Nd4jPointer *extraPointers,
                                     void *dY, Nd4jLong *dYShapeInfo,
                                     void *hZ, Nd4jLong *hZShapeInfo,
                                     void *dZ, Nd4jLong *dZShapeInfo) {
-
-    NativeOpExecutioner::execReduce3(nullptr, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    try {
+        NativeOpExecutioner::execReduce3(nullptr, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo,
+                                         dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -588,8 +666,13 @@ void execReduce3Scalar(Nd4jPointer *extraPointers,int opNum,
                                             void *dY, Nd4jLong *dYShapeInfo,
                                             void *hZ, Nd4jLong *hZShapeInfo,
                                             void *dZ, Nd4jLong *dZShapeInfo) {
-
-    NativeOpExecutioner::execReduce3Scalar(nullptr, opNum,hX,hXShapeInfo,dX, dXShapeInfo,extraParams,hY,hYShapeInfo,dY,dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    try {
+        NativeOpExecutioner::execReduce3Scalar(nullptr, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY,
+                                               hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 /**
  *
@@ -617,19 +700,31 @@ void execReduce3Tad(Nd4jPointer *extraPointers,
                                     void *dDimension, Nd4jLong *dDimensionShape,
                                     Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets,
                                     Nd4jLong *yTadOnlyShapeInfo, Nd4jLong *yTadOffsets) {
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    if (extraPointers == nullptr || extraPointers[2] == 0) {
-        NativeOpExecutioner::execReduce3(LaunchContext::defaultContext(), opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, yTadOnlyShapeInfo, yTadOffsets);
-    } else {
-        // going tad-way
-        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension, dimensionLength);
+        if (extraPointers == nullptr || extraPointers[2] == 0) {
+            NativeOpExecutioner::execReduce3(LaunchContext::defaultContext(), opNum, hX, hXShapeInfo, dX, dXShapeInfo,
+                                             extraParams, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ,
+                                             dZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets,
+                                             yTadOnlyShapeInfo, yTadOffsets);
+        } else {
+            // going tad-way
+            auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension,
+                                                                                    dimensionLength);
 
-        auto hTADShapeInfo = tadPack.primaryShapeInfo();
-        auto hTADOffsets = tadPack.primaryOffsets();
+            auto hTADShapeInfo = tadPack.primaryShapeInfo();
+            auto hTADOffsets = tadPack.primaryOffsets();
 
-        NativeOpExecutioner::execReduce3TAD(LaunchContext::defaultContext(), opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, hTADShapeInfo, hTADOffsets, nullptr, nullptr);
+            NativeOpExecutioner::execReduce3TAD(LaunchContext::defaultContext(), opNum, hX, hXShapeInfo, dX,
+                                                dXShapeInfo, extraParams, hY, hYShapeInfo, dY, dYShapeInfo, hZ,
+                                                hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, hTADShapeInfo,
+                                                hTADOffsets, nullptr, nullptr);
+        }
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
     }
 }
 
@@ -654,36 +749,9 @@ void execScalar(
         void *hScalar, Nd4jLong *hScalarShapeInfo,
         void *dScalar, Nd4jLong *dScalarShapeInfo,
         void *extraParams) {
-    NativeOpExecutioner::execScalar(nullptr,
-                                   opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            hScalar,
-            hScalarShapeInfo,
-            dScalar,
-            dScalarShapeInfo,
-            extraParams);
-}
-
-void execScalarBool(
-        Nd4jPointer *extraPointers,
-        int opNum,
-        void *hX, Nd4jLong *hXShapeInfo,
-        void *dX, Nd4jLong *dXShapeInfo,
-        void *hZ, Nd4jLong *hZShapeInfo,
-        void *dZ, Nd4jLong *dZShapeInfo,
-        void *hScalar, Nd4jLong *hScalarShapeInfo,
-        void *dScalar, Nd4jLong *dScalarShapeInfo,
-        void *extraParams) {
-
-    NativeOpExecutioner::execScalarBool(nullptr,
-                                       opNum,
+    try {
+        NativeOpExecutioner::execScalar(nullptr,
+                                        opNum,
                                         hX,
                                         hXShapeInfo,
                                         dX,
@@ -696,7 +764,43 @@ void execScalarBool(
                                         hScalarShapeInfo,
                                         dScalar,
                                         dScalarShapeInfo,
-                                       extraParams);
+                                        extraParams);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
+}
+
+void execScalarBool(
+        Nd4jPointer *extraPointers,
+        int opNum,
+        void *hX, Nd4jLong *hXShapeInfo,
+        void *dX, Nd4jLong *dXShapeInfo,
+        void *hZ, Nd4jLong *hZShapeInfo,
+        void *dZ, Nd4jLong *dZShapeInfo,
+        void *hScalar, Nd4jLong *hScalarShapeInfo,
+        void *dScalar, Nd4jLong *dScalarShapeInfo,
+        void *extraParams) {
+    try {
+        NativeOpExecutioner::execScalarBool(nullptr,
+                                            opNum,
+                                            hX,
+                                            hXShapeInfo,
+                                            dX,
+                                            dXShapeInfo,
+                                            hZ,
+                                            hZShapeInfo,
+                                            dZ,
+                                            dZShapeInfo,
+                                            hScalar,
+                                            hScalarShapeInfo,
+                                            dScalar,
+                                            dScalarShapeInfo,
+                                            extraParams);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -714,18 +818,23 @@ void execSummaryStatsScalar(Nd4jPointer *extraPointers,
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo,
         bool biasCorrected) {
-    NativeOpExecutioner::execSummaryStatsScalar(nullptr,
-                                               opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            extraParams,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            biasCorrected);
+    try {
+        NativeOpExecutioner::execSummaryStatsScalar(nullptr,
+                                                    opNum,
+                                                    hX,
+                                                    hXShapeInfo,
+                                                    dX,
+                                                    dXShapeInfo,
+                                                    extraParams,
+                                                    hZ,
+                                                    hZShapeInfo,
+                                                    dZ,
+                                                    dZShapeInfo,
+                                                    biasCorrected);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 /**
  *
@@ -744,18 +853,23 @@ void execSummaryStats(Nd4jPointer *extraPointers,
                                          void *hZ, Nd4jLong *hZShapeInfo,
                                          void *dZ, Nd4jLong *dZShapeInfo,
                                          bool biasCorrected) {
-    NativeOpExecutioner::execSummaryStats(nullptr,
-                                         opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            extraParams,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            biasCorrected);
+    try {
+        NativeOpExecutioner::execSummaryStats(nullptr,
+                                              opNum,
+                                              hX,
+                                              hXShapeInfo,
+                                              dX,
+                                              dXShapeInfo,
+                                              extraParams,
+                                              hZ,
+                                              hZShapeInfo,
+                                              dZ,
+                                              dZShapeInfo,
+                                              biasCorrected);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 /**
  *
@@ -779,27 +893,31 @@ void execSummaryStatsTad(Nd4jPointer *extraPointers,
                                          void *dDimension, Nd4jLong *dDimensionShape,
                                          bool biasCorrected,
                                          Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
 
-    NativeOpExecutioner::execSummaryStats(nullptr,
-                                         opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            extraParams,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            dimension,
-            dimensionLength,
-            tadShapeInfo,
-            tadOffsets,
-            biasCorrected);
-
+        NativeOpExecutioner::execSummaryStats(nullptr,
+                                              opNum,
+                                              hX,
+                                              hXShapeInfo,
+                                              dX,
+                                              dXShapeInfo,
+                                              extraParams,
+                                              hZ,
+                                              hZShapeInfo,
+                                              dZ,
+                                              dZShapeInfo,
+                                              dimension,
+                                              dimensionLength,
+                                              tadShapeInfo,
+                                              tadOffsets,
+                                              biasCorrected);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -820,20 +938,24 @@ void execTransformFloat(
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo,
         void *extraParams) {
-
-    NativeOpExecutioner::execTransformFloat(nullptr,
-                                           opNum,
-            hX,
-            hXShapeInfo,
-            dZ,
-            dXShapeInfo,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            extraParams,
-            nullptr,
-            nullptr);
+    try {
+        NativeOpExecutioner::execTransformFloat(nullptr,
+                                                opNum,
+                                                hX,
+                                                hXShapeInfo,
+                                                dZ,
+                                                dXShapeInfo,
+                                                hZ,
+                                                hZShapeInfo,
+                                                dZ,
+                                                dZShapeInfo,
+                                                extraParams,
+                                                nullptr,
+                                                nullptr);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execTransformSame(
@@ -844,20 +966,24 @@ void execTransformSame(
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo,
         void *extraParams) {
-
-    NativeOpExecutioner::execTransformSame(nullptr,
-                                          opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            extraParams,
-            nullptr,
-            nullptr);
+    try {
+        NativeOpExecutioner::execTransformSame(nullptr,
+                                               opNum,
+                                               hX,
+                                               hXShapeInfo,
+                                               dX,
+                                               dXShapeInfo,
+                                               hZ,
+                                               hZShapeInfo,
+                                               dZ,
+                                               dZShapeInfo,
+                                               extraParams,
+                                               nullptr,
+                                               nullptr);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execTransformBool(
@@ -868,20 +994,24 @@ void execTransformBool(
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo,
         void *extraParams) {
-
-    NativeOpExecutioner::execTransformBool(nullptr,
-                                          opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            extraParams,
-            nullptr,
-            nullptr);
+    try {
+        NativeOpExecutioner::execTransformBool(nullptr,
+                                               opNum,
+                                               hX,
+                                               hXShapeInfo,
+                                               dX,
+                                               dXShapeInfo,
+                                               hZ,
+                                               hZShapeInfo,
+                                               dZ,
+                                               dZShapeInfo,
+                                               extraParams,
+                                               nullptr,
+                                               nullptr);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execTransformAny(
@@ -892,20 +1022,24 @@ void execTransformAny(
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo,
         void *extraParams) {
-
-    NativeOpExecutioner::execTransformAny(nullptr,
-                                         opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            extraParams,
-            nullptr,
-            nullptr);
+    try {
+        NativeOpExecutioner::execTransformAny(nullptr,
+                                              opNum,
+                                              hX,
+                                              hXShapeInfo,
+                                              dX,
+                                              dXShapeInfo,
+                                              hZ,
+                                              hZShapeInfo,
+                                              dZ,
+                                              dZShapeInfo,
+                                              extraParams,
+                                              nullptr,
+                                              nullptr);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execTransformStrict(
@@ -916,20 +1050,24 @@ void execTransformStrict(
         void *hZ, Nd4jLong *hZShapeInfo,
         void *dZ, Nd4jLong *dZShapeInfo,
         void *extraParams) {
-
-    NativeOpExecutioner::execTransformStrict(nullptr,
-                                            opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            extraParams,
-            nullptr,
-            nullptr);
+    try {
+        NativeOpExecutioner::execTransformStrict(nullptr,
+                                                 opNum,
+                                                 hX,
+                                                 hXShapeInfo,
+                                                 dX,
+                                                 dXShapeInfo,
+                                                 hZ,
+                                                 hZShapeInfo,
+                                                 dZ,
+                                                 dZShapeInfo,
+                                                 extraParams,
+                                                 nullptr,
+                                                 nullptr);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execReduce3All(Nd4jPointer *extraPointers,
@@ -948,158 +1086,18 @@ void execReduce3All(Nd4jPointer *extraPointers,
                                      Nd4jLong *yTadShapeInfo,
                                      Nd4jLong *yOffsets) {
 
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
 
-    NativeOpExecutioner::execReduce3All(nullptr, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParamsVals, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, xTadShapeInfo, xOffsets, yTadShapeInfo, yOffsets);
-}
-
-
-template <typename T>
-void flattenGeneric(Nd4jPointer *extraPointers,
-                    int offset,
-                    char order,
-                    void *vresult,
-                    Nd4jLong *hZShapeInfo,
-                    void *vinput,
-                    Nd4jLong *inputShapeInfo) {
-
-    auto hZ = reinterpret_cast<T *>(vresult);
-    auto input = reinterpret_cast<T *>(vinput);
-
-    int numOnes = 0;
-    auto shape = shape::shapeOf(inputShapeInfo);
-    int wholeRank = shape::rank(inputShapeInfo);
-    for(int i = 0; i < wholeRank; i++) {
-        if(shape[i] == 1)
-            numOnes++;
+        NativeOpExecutioner::execReduce3All(nullptr, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParamsVals, hY,
+                                            hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension,
+                                            dimensionLength, xTadShapeInfo, xOffsets, yTadShapeInfo, yOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
     }
-
-
-
-    //start at the given offset
-    hZ += offset;
-    char inputOrder = shape::order(inputShapeInfo);
-    auto len = shape::length(inputShapeInfo);
-    auto resultEleStride = shape::elementWiseStride(hZShapeInfo);
-    auto inputEleStride = shape::elementWiseStride(inputShapeInfo);
-    Nd4jLong numTads, stride;
-    int dimension, dimensionLength;
-    int rank = shape::rank(inputShapeInfo);
-    auto xStride = shape::stride(inputShapeInfo);
-    auto xShape = shape::shapeOf(inputShapeInfo);
-
-    dimensionLength = 1;
-    if(order == 'f') {
-        dimension = 0;
-    }
-    else {
-        dimension = rank - 1;
-    }
-    stride  = xStride[dimension];
-    // numTads is product of length of all dimensions excluding
-    // the one we do the tad on
-    numTads = 1;
-    for (int i = 0; i < rank; i++) {
-        if (i != dimension)
-            numTads *= xShape[i];
-    }
-
-    if (inputOrder == order) {
-        if (resultEleStride == 1 && inputEleStride == 1) {
-            memcpy(hZ, input, len* sizeof(T));
-        }
-        else if (resultEleStride >= 1 && inputEleStride >= 1) {
-            if (len < ELEMENT_THRESHOLD) {
-
-                PRAGMA_OMP_SIMD
-                for (Nd4jLong i = 0; i < len; i++) {
-                    hZ[i * resultEleStride] = input[i * inputEleStride];
-                }
-            }
-            else {
-
-                PRAGMA_OMP_PARALLEL_FOR_SIMD
-                for (Nd4jLong i = 0; i < len; i++) {
-                    hZ[i * resultEleStride] = input[i * inputEleStride];
-                }
-            }
-        }
-        else {
-            int idx = 0;
-            for(Nd4jLong i = 0; i < len; i++)
-                    hZ[idx++] = input[shape::getIndexOffset(i, inputShapeInfo, len)];
-        }
-    }
-    else {
-        int rank = shape::rank(inputShapeInfo);
-        auto xShape = shape::shapeOf(inputShapeInfo);
-        auto tadShape = xShape[dimension];
-
-        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(inputShapeInfo, dimension);
-
-        PRAGMA_OMP_PARALLEL_FOR
-        for(int i = 0; i < numTads; i++) {
-
-            Nd4jLong resultOffset;
-
-            if (order == 'f') {
-                // 1. get c ordering coordinates
-                auto cIndexCoordinates = new Nd4jLong[rank - 1];
-                Nd4jLong divisor = 1;
-                for (int dim = rank - 1; dim > 0; dim--) {
-                    cIndexCoordinates[dim - 1] = (i / divisor) % xShape[dim];
-                    divisor *= xShape[dim];
-                }
-
-
-                // 2. convert to f ordering index
-                int fIndex = 0;
-                Nd4jLong multiplier = 1;
-                for (int dim = 1; dim <= rank - 1; dim++) {
-                    fIndex += cIndexCoordinates[dim - 1] * multiplier;
-                    multiplier *= xShape[dim];
-                }
-
-                resultOffset = fIndex * tadShape;
-                delete[] cIndexCoordinates;
-
-            }
-            else {
-                resultOffset = i *  tadShape;
-            }
-
-            auto tadOffset = tadPack.primaryOffsets()[i];
-            for( int j = 0; j < tadShape; j++) {
-
-                // TAD are returned in C ordering always
-                hZ[resultOffset + j] = input[tadOffset + j * stride];
-
-            }
-        }
-    }
-}
-
-
-/**
-    * Concatneate multi array of the same shape together
-    * along a particular dimension
-    */
-void concat(
-        Nd4jPointer *extraPointers,
-        int dimension,
-        int numArrays,
-        Nd4jPointer *data, Nd4jPointer *inputShapeInfo,
-        Nd4jPointer *ddata, Nd4jPointer *dinputShapeInfo,
-        void *hZ, Nd4jLong *hZShapeInfo,
-        void *dZ, Nd4jLong *dZShapeInfo,
-        Nd4jPointer *tadPointers,
-        Nd4jPointer *offsetPointers) {
-
-    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
-
-    BUILD_SINGLE_SELECTOR(zType, nd4j::SpecialMethods, ::concatCpuGeneric(dimension, numArrays, data, inputShapeInfo, hZ, hZShapeInfo), LIBND4J_TYPES);
 }
 
 /**
@@ -1116,39 +1114,14 @@ void specialConcat(
         Nd4jLong *hZShapeInfo,
         Nd4jPointer *tadPointers,
         Nd4jPointer *offsetPointers) {
+    try {
+        auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
 
-    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
-
-    BUILD_SINGLE_SELECTOR(zType, nd4j::SpecialMethods, ::concatCpuGeneric(dimension, numArrays, data, inputShapeInfo, hZ, hZShapeInfo), LIBND4J_TYPES);
-}
-
-/**
-* Append an input array
-* to the end of a flat array
-* in a particular order
-* @param offset the offset of the array to start at
-* @param order the order
-* @param hZ the hZ array
-* @param hZShapeInfo the shape info for te array
-* @param input the input for the array
-* @param inputShapeInfo the shape information for that array
-*/
-void flatten(
-        Nd4jPointer *extraPointers,
-        int offset,
-        char order,
-        void *hZ, Nd4jLong *hZShapeInfo,
-        void *dZ, Nd4jLong *dZShapeInfo,
-        void *input, Nd4jLong *inputShapeInfo,
-        void *dinput, Nd4jLong *dinputShapeInfo) {
-
-    auto xType = nd4j::ArrayOptions::dataType(inputShapeInfo);
-    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
-
-    if (xType != zType)
-        throw std::runtime_error("NativeOps::flatten requires all operands to have same data type");
-
-    BUILD_SINGLE_SELECTOR(xType, flattenGeneric, (extraPointers, offset, order, hZ, hZShapeInfo, input, inputShapeInfo), LIBND4J_TYPES);
+        BUILD_SINGLE_SELECTOR(zType, nd4j::SpecialMethods,::concatCpuGeneric(dimension, numArrays, data, inputShapeInfo, hZ, hZShapeInfo), LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -1324,7 +1297,13 @@ void setGridLimit(int gridSize) {
 
 nd4j::TadPack* tadOnlyShapeInfo(Nd4jLong *hXShapeInfo, int *dimension, int dimensionLength) {
     auto pack = new TadPack();
-    *pack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension, dimensionLength);
+    try {
+        *pack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, dimension, dimensionLength);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
+
     return pack;
 }
 
@@ -1421,9 +1400,14 @@ void pullRows(Nd4jPointer *extraPointers,
         Nd4jLong *tadOffsets,
         Nd4jLong *zTadShapeInfo,
         Nd4jLong *zTadOffsets) {
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    try {
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
 
-    BUILD_SINGLE_SELECTOR(xType, pullRowsGeneric, (hX, hXShapeInfo, hZ, hZShapeInfo, n, indexes, tadShapeInfo, tadOffsets, zTadShapeInfo, zTadOffsets), LIBND4J_TYPES);
+        BUILD_SINGLE_SELECTOR(xType, pullRowsGeneric, (hX, hXShapeInfo, hZ, hZShapeInfo, n, indexes, tadShapeInfo, tadOffsets, zTadShapeInfo, zTadOffsets), LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 template<typename T>
@@ -1474,9 +1458,14 @@ void tear(Nd4jPointer *extraPointers,
         Nd4jLong *hZShapeInfo,
         Nd4jLong *tadShapeInfo,
         Nd4jLong *tadOffsets) {
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    try {
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
 
-    BUILD_SINGLE_SELECTOR(xType, tearGeneric, (hX, hXShapeInfo, targets, hZShapeInfo, tadShapeInfo, tadOffsets), LIBND4J_TYPES);
+        BUILD_SINGLE_SELECTOR(xType, tearGeneric, (hX, hXShapeInfo, targets, hZShapeInfo, tadShapeInfo, tadOffsets), LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -1488,9 +1477,14 @@ void average(Nd4jPointer *extras,
         int n,
         Nd4jLong length,
         bool propagate) {
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    try {
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
 
-    BUILD_SINGLE_SELECTOR(xType, nd4j::SpecialMethods, ::averageGeneric(hX, z, hZShapeInfo, n, length, propagate), LIBND4J_TYPES);
+        BUILD_SINGLE_SELECTOR(xType, nd4j::SpecialMethods, ::averageGeneric(hX, z, hZShapeInfo, n, length, propagate), LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void accumulate(Nd4jPointer *extras,
@@ -1500,10 +1494,14 @@ void accumulate(Nd4jPointer *extras,
         void *dz, Nd4jLong *dZShapeInfo,
         int n,
         Nd4jLong length) {
+    try {
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
 
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
-
-    BUILD_SINGLE_SELECTOR(xType, nd4j::SpecialMethods, ::accumulateGeneric(hX, hz, hZShapeInfo, n, length), LIBND4J_TYPES);
+        BUILD_SINGLE_SELECTOR(xType, nd4j::SpecialMethods, ::accumulateGeneric(hX, hz, hZShapeInfo, n, length), LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void enableP2P(bool enable) {
@@ -1613,14 +1611,20 @@ void shuffle(Nd4jPointer *extras,
                               int *shuffleMap,
                               Nd4jPointer *tadShapeInfo,
                               Nd4jPointer *tadOffsets) {
-    auto xShape = reinterpret_cast<Nd4jLong **>(hXShapeInfo);
-    auto zShape = reinterpret_cast<Nd4jLong **>(hZShapeInfo);
-    auto tadOnlyShapeInfo = reinterpret_cast<Nd4jLong **>(tadShapeInfo);
-    auto tadOffset = reinterpret_cast<Nd4jLong **>(tadOffsets);
+    try {
+        auto xShape = reinterpret_cast<Nd4jLong **>(hXShapeInfo);
+        auto zShape = reinterpret_cast<Nd4jLong **>(hZShapeInfo);
+        auto tadOnlyShapeInfo = reinterpret_cast<Nd4jLong **>(tadShapeInfo);
+        auto tadOffset = reinterpret_cast<Nd4jLong **>(tadOffsets);
 
-    auto xType = nd4j::ArrayOptions::dataType(xShape[0]);
+        auto xType = nd4j::ArrayOptions::dataType(xShape[0]);
 
-    BUILD_SINGLE_SELECTOR(xType, shuffleGeneric, (hX, xShape, hz, zShape, N, shuffleMap, tadOnlyShapeInfo, tadOffset), LIBND4J_TYPES);
+        BUILD_SINGLE_SELECTOR(xType, shuffleGeneric,
+                              (hX, xShape, hz, zShape, N, shuffleMap, tadOnlyShapeInfo, tadOffset), LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -1633,27 +1637,6 @@ void setOmpMinThreads(int threads) {
     // TODO: to be implemented
 }
 
-/*
-void execMetaPredicateShape(Nd4jPointer *extras,
-                                        const int opTypeA,
-                                        const int opNumA,
-                                        const int opTypeB,
-                                        const int opNumB,
-                                        Nd4jLong N,
-                                        void *hX, Nd4jLong *hXShapeInfo,
-                                        void *dX, Nd4jLong *dXShapeInfo,
-                                        void *hY, Nd4jLong *hYShapeInfo,
-                                        void *dY, Nd4jLong *dYShapeInfo,
-                                        void *hZ, Nd4jLong *hZShapeInfo,
-                                        void *dZ, Nd4jLong *dZShapeInfo,
-                                        void *extraA,
-                                        void *extraB,
-                                        double scalarA,
-                                        double scalarB) {
-    // no-op;
-}
-*/
-
 int getDevice() {
     return 0;
 }
@@ -1671,31 +1654,35 @@ void execScalarTad(Nd4jPointer *extraPointers,
                                  void *dDimension, Nd4jLong *dDimensionShape,
                                  Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
                                  Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
-
-    NativeOpExecutioner::execScalar(nullptr,
-            opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            extraParams,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            hScalars,
-            hScalarShapeInfo,
-            dScalars,
-            dScalarShapeInfo,
-            dimension,
-            shape::length(hDimensionShape),
-            tadShapeInfo,
-            tadOffsets,
-            tadShapeInfoZ,
-            tadOffsetsZ);
+        NativeOpExecutioner::execScalar(nullptr,
+                                        opNum,
+                                        hX,
+                                        hXShapeInfo,
+                                        dX,
+                                        dXShapeInfo,
+                                        extraParams,
+                                        hZ,
+                                        hZShapeInfo,
+                                        dZ,
+                                        dZShapeInfo,
+                                        hScalars,
+                                        hScalarShapeInfo,
+                                        dScalars,
+                                        dScalarShapeInfo,
+                                        dimension,
+                                        shape::length(hDimensionShape),
+                                        tadShapeInfo,
+                                        tadOffsets,
+                                        tadShapeInfoZ,
+                                        tadOffsetsZ);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execScalarBoolTad(Nd4jPointer *extraPointers,
@@ -1711,44 +1698,53 @@ void execScalarBoolTad(Nd4jPointer *extraPointers,
                            void *dDimension, Nd4jLong *dDimensionShape,
                            Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
                            Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+    try {
+        auto dimension = reinterpret_cast<int *>(hDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto dimension = reinterpret_cast<int *>(hDimension);
-    int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
-
-    NativeOpExecutioner::execScalarBool(nullptr,
-                                       opNum,
-            hX,
-            hXShapeInfo,
-            dX,
-            dXShapeInfo,
-            extraParams,
-            hZ,
-            hZShapeInfo,
-            dZ,
-            dZShapeInfo,
-            hScalars,
-            hScalarShapeInfo,
-            dScalars,
-            dScalarShapeInfo,
-            dimension,
-            dimensionLength,
-            tadShapeInfo,
-            tadOffsets,
-            tadShapeInfoZ,
-            tadOffsetsZ);
+        NativeOpExecutioner::execScalarBool(nullptr,
+                                            opNum,
+                                            hX,
+                                            hXShapeInfo,
+                                            dX,
+                                            dXShapeInfo,
+                                            extraParams,
+                                            hZ,
+                                            hZShapeInfo,
+                                            dZ,
+                                            dZShapeInfo,
+                                            hScalars,
+                                            hScalarShapeInfo,
+                                            dScalars,
+                                            dScalarShapeInfo,
+                                            dimension,
+                                            dimensionLength,
+                                            tadShapeInfo,
+                                            tadOffsets,
+                                            tadShapeInfoZ,
+                                            tadOffsetsZ);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 const char * getDeviceName(int deviceId) {
-    if (!nameSet) {
-        name = reinterpret_cast<char *>(malloc(256 * sizeof(char)));
+    try {
+        if (!nameSet) {
+            name = reinterpret_cast<char *>(malloc(256 * sizeof(char)));
 
-        CHECK_ALLOC(name, "Failed to allocate new string buffer", 256);
+            CHECK_ALLOC(name, "Failed to allocate new string buffer", 256);
 
-        std::memset(name, 0, 256 * sizeof(char));
-        nameSet = true;
+            std::memset(name, 0, 256 * sizeof(char));
+            nameSet = true;
 
-        // TODO: provide proper CPU model name here
-        sprintf(name, "x86-compatible CPU");
+            // TODO: provide proper CPU model name here
+            sprintf(name, "x86-compatible CPU");
+        }
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
     }
 
 
@@ -1768,8 +1764,12 @@ void execAggregate(Nd4jPointer *extraPointers,int opNum,
                                     void *realArguments,
                                     int numRealArguments,
                                     nd4j::DataType dtype) {
-
-    BUILD_SINGLE_SELECTOR(dtype, NativeOpExecutioner::execAggregate, (nullptr, opNum, arguments, numArguments, shapeArguments, numShapeArguments, indexArguments, numIndexArguments, intArrays, numIntArrays, realArguments, numRealArguments), FLOAT_TYPES);
+    try {
+        BUILD_SINGLE_SELECTOR(dtype, NativeOpExecutioner::execAggregate, (nullptr, opNum, arguments, numArguments, shapeArguments, numShapeArguments, indexArguments, numIndexArguments, intArrays, numIntArrays, realArguments, numRealArguments), FLOAT_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 
 }
 
@@ -1841,7 +1841,12 @@ void batchExecutor(Nd4jPointer *extraPointers,
                                int maxReals,
                                void *ptrToArguments,
                                nd4j::DataType dtype) {
-    BUILD_SINGLE_SELECTOR(dtype, _batchExecutor, (extraPointers, numAggregates, opNum, maxArgs, maxShapes, maxIntArrays, maxIntArraySize, maxIdx, maxReals, ptrToArguments, dtype), FLOAT_TYPES);
+    try {
+        BUILD_SINGLE_SELECTOR(dtype, _batchExecutor, (extraPointers, numAggregates, opNum, maxArgs, maxShapes, maxIntArrays, maxIntArraySize, maxIdx, maxReals, ptrToArguments, dtype), FLOAT_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execAggregateBatch(Nd4jPointer *extraPointers,
@@ -1855,7 +1860,12 @@ void execAggregateBatch(Nd4jPointer *extraPointers,
                                          int maxReals,
                                          void *ptrToArguments,
                                          nd4j::DataType dtype) {
-    BUILD_SINGLE_SELECTOR(dtype, _batchExecutor, (extraPointers, numAggregates, opNum, maxArgs, maxShapes, maxIntArrays, maxIntArraySize, maxIdx, maxReals, ptrToArguments, dtype), FLOAT_TYPES);
+    try {
+        BUILD_SINGLE_SELECTOR(dtype, _batchExecutor, (extraPointers, numAggregates, opNum, maxArgs, maxShapes, maxIntArrays, maxIntArraySize, maxIdx, maxReals, ptrToArguments, dtype), FLOAT_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -1865,7 +1875,12 @@ void execRandom(Nd4jPointer *extraPointers,
                                  void *hZ, Nd4jLong *hZShapeInfo,
                                  void *dZ, Nd4jLong *dZShapeInfo,
                                  void *extraArguments) {
-    NativeOpExecutioner::execRandom(nullptr, opNum, state, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    try {
+        NativeOpExecutioner::execRandom(nullptr, opNum, state, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execRandom3(Nd4jPointer *extraPointers,
@@ -1878,8 +1893,12 @@ void execRandom3(Nd4jPointer *extraPointers,
                                  void *hZ, Nd4jLong *hZShapeInfo,
                                  void *dZ, Nd4jLong *dZShapeInfo,
                                  void *extraArguments) {
-
-    NativeOpExecutioner::execRandom(nullptr, opNum, state, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    try {
+        NativeOpExecutioner::execRandom(nullptr, opNum, state, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execRandom2(Nd4jPointer *extraPointers,
@@ -1890,19 +1909,25 @@ void execRandom2(Nd4jPointer *extraPointers,
                                  void *hZ, Nd4jLong *hZShapeInfo,
                                  void *dZ, Nd4jLong *dZShapeInfo,
                                  void *extraArguments) {
-
-    NativeOpExecutioner::execRandom(nullptr, opNum, state, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    try {
+        NativeOpExecutioner::execRandom(nullptr, opNum, state, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 Nd4jPointer initRandom(Nd4jPointer *extraPointers, long seed, long bufferSize, Nd4jPointer ptrToBuffer) {
-    graph::RandomGenerator* generator = new graph::RandomGenerator(seed, seed);
-//    auto ptrBuf = reinterpret_cast<long *>(ptrToBuffer);
-//    auto buffer = new nd4j::random::RandomBuffer(seed, bufferSize, reinterpret_cast<uint64_t *>(ptrBuf));
-//
-//    nd4j::random::Xoroshiro128 generator(buffer);
-//    generator.refreshBuffer();
-//
-    return (Nd4jPointer) generator;
+    try {
+        auto generator = new graph::RandomGenerator(seed, seed);
+
+        return (Nd4jPointer) generator;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+
+        return nullptr;
+    }
 }
 
 void refreshBuffer(Nd4jPointer *extraPointers, long seed, Nd4jPointer ptrRandom) {
@@ -1953,7 +1978,12 @@ void sort(Nd4jPointer *extraPointers,
         void *hX, Nd4jLong *hXShapeInfo,
         void *dX, Nd4jLong *dXShapeInfo,
         bool descending) {
-    NativeOpExecutioner::execSort(hX, hXShapeInfo, descending);
+    try {
+        NativeOpExecutioner::execSort(hX, hXShapeInfo, descending);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void sortTad(Nd4jPointer *extraPointers,
@@ -1964,7 +1994,12 @@ void sortTad(Nd4jPointer *extraPointers,
             Nd4jLong *tadShapeInfo,
             Nd4jLong *tadOffsets,
             bool descending) {
-    NativeOpExecutioner::execSort(hX, hXShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, descending);
+    try {
+        NativeOpExecutioner::execSort(hX, hXShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, descending);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void sortCooIndices(Nd4jPointer *extraPointers,
@@ -1972,7 +2007,12 @@ void sortCooIndices(Nd4jPointer *extraPointers,
         void *values,
         Nd4jLong length,
         int rank) {
-    NativeOpExecutioner::execSortCooIndices(indices, values, length, rank);
+    try {
+        NativeOpExecutioner::execSortCooIndices(indices, values, length, rank);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 Nd4jLong encodeBitmap(Nd4jPointer *extraPointers, void *hX, Nd4jLong *hXShapeInfo, Nd4jLong N, int *dz, float threshold) {
@@ -1983,7 +2023,7 @@ Nd4jLong encodeBitmap(Nd4jPointer *extraPointers, void *hX, Nd4jLong *hXShapeInf
 
 Nd4jLong* mmapFile(Nd4jPointer *extraPointers, const char *fileName, Nd4jLong length) {
     auto hZ = new Nd4jLong[2];errno = 0;
-
+try {
 #if defined(_WIN32) || defined(_WIN64)
     _mmap(hZ, static_cast<size_t>(length), fileName);
 #else
@@ -1992,7 +2032,7 @@ Nd4jLong* mmapFile(Nd4jPointer *extraPointers, const char *fileName, Nd4jLong le
         nd4j_printf("Errno: %i\n", errno);
         throw std::runtime_error("Failed to open file for MMAP");
     }
-    void * ptr = mmap(NULL, length, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+    void *ptr = mmap(NULL, length, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 
 // check for failed allocation
     if (ptr == MAP_FAILED)
@@ -2004,7 +2044,11 @@ Nd4jLong* mmapFile(Nd4jPointer *extraPointers, const char *fileName, Nd4jLong le
 #endif
 
     return hZ;
-
+} catch (std::exception &e) {
+    nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+    nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    return nullptr;
+}
 }
 
 void munmapFile(Nd4jPointer *extraPointers, Nd4jLong *ptrMap, Nd4jLong length) {
@@ -2019,7 +2063,13 @@ void munmapFile(Nd4jPointer *extraPointers, Nd4jLong *ptrMap, Nd4jLong length) {
 }
 
 nd4j::graph::ResultWrapper* executeFlatGraph(Nd4jPointer *extraPointers, Nd4jPointer flatBufferPointer) {
-    return nd4j::graph::GraphExecutioner::executeFlatBuffer(flatBufferPointer);
+    try {
+        return nd4j::graph::GraphExecutioner::executeFlatBuffer(flatBufferPointer);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 Nd4jLong getResultWrapperSize(nd4j::graph::ResultWrapper* ptr) {
@@ -2061,8 +2111,14 @@ FORCEINLINE int estimateThresholdGeneric(Nd4jPointer *extraPointers, Nd4jPointer
 
 
 int estimateThreshold(Nd4jPointer *extraPointers, Nd4jPointer hX, Nd4jLong *hXShapeInfo, int N, float threshold) {
-    auto xType = ArrayOptions::dataType(hXShapeInfo);
-    BUILD_SINGLE_SELECTOR(xType, return estimateThresholdGeneric, (extraPointers, hX, N, threshold), FLOAT_TYPES);
+    try {
+        auto xType = ArrayOptions::dataType(hXShapeInfo);
+        BUILD_SINGLE_SELECTOR(xType, return estimateThresholdGeneric, (extraPointers, hX, N, threshold), FLOAT_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 0;
+    }
 }
 
 Nd4jLong getShapeListSize(nd4j::ShapeList* list) {
@@ -2122,9 +2178,15 @@ nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::D
 }
 
 nd4j::ShapeList* calculateOutputShapes2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs) {
-    auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
+    try {
+        auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
 
-    return _calculateOutputShapes(extraPointers, op, inputBuffers, inputShapes, numInputShapes, tArgs, numTArgs, iArgs, numIArgs, bArgs, numBArgs);
+        return _calculateOutputShapes(extraPointers, op, inputBuffers, inputShapes, numInputShapes, tArgs, numTArgs, iArgs, numIArgs, bArgs, numBArgs);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::DeclarableOp *op, Nd4jPointer* inputShapes, int numInputShapes, double *tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs) {
@@ -2147,16 +2209,28 @@ nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::D
 }
 
 nd4j::ShapeList* calculateOutputShapes(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs) {
-    auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
+    try {
+        auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
 
-    return _calculateOutputShapes(extraPointers, op, inputShapes, numInputShapes, tArgs, numTArgs, iArgs, numIArgs);
+        return _calculateOutputShapes(extraPointers, op, inputShapes, numInputShapes, tArgs, numTArgs, iArgs, numIArgs);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 int execCustomOp2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer opContext) {
-    auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
-    auto context = reinterpret_cast<Context*>(opContext);
+    try {
+        auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
+        auto context = reinterpret_cast<Context *>(opContext);
 
-    return op->execute(context);
+        return op->execute(context);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 20;
+    }
 }
 
 Nd4jStatus realExec(nd4j::ops::DeclarableOp* op, Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputs, Nd4jPointer* outputBuffers, Nd4jPointer* outputShapes, int numOutputs, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool* bArgs, int numBArgs, bool isInplace) {
@@ -2234,34 +2308,6 @@ Nd4jStatus realExec(nd4j::ops::DeclarableOp* op, Nd4jPointer* extraPointers, Nd4
                 outputs[e]->streamline(shape::order(reinterpret_cast<Nd4jLong *>(outputShapes[e])));
         }
 
-/*
-    if (!isInplace) {
-        if (hZ->size() != numOutputs) {
-            return ND4J_STATUS_BAD_OUTPUT;
-        }
-
-        for (int e = 0; e < numOutputs; e++) {
-            auto buffer = (T *) outputBuffers[e];
-            auto shape = (int *) outputShapes[e];
-            nd4j::NDArray<T> tmp(buffer, shape);
-
-            if (tmp.lengthOf() != hZ->at(e)->lengthOf()) {
-                nd4j_printf("Provided output array for [%s] has length of %i, but actual hZ has length of %i\n", op->getOpName()->c_str(), tmp.lengthOf(), hZ->at(e)->lengthOf());
-                return ND4J_STATUS_BAD_OUTPUT;
-            }
-
-            tmp.assign(hZ->at(e));
-        }
-    } else {
-        // if op is inplace, our ResultSet holds pointers
-        hZ->purge();
-    }
-
-
-    delete hZ;
-
-*/
-
     for (auto v: inputs)
         delete v;
 
@@ -2273,16 +2319,28 @@ Nd4jStatus realExec(nd4j::ops::DeclarableOp* op, Nd4jPointer* extraPointers, Nd4
 
 
 int execCustomOp(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputs, Nd4jPointer* outputBuffers, Nd4jPointer* outputShapes, int numOutputs, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool* bArgs, int numBArgs, bool isInplace) {
-    auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
-    return realExec(op, extraPointers, hash, inputBuffers, inputShapes, numInputs, outputBuffers, outputShapes, numOutputs, tArgs, numTArgs, iArgs, numIArgs, bArgs, numBArgs, isInplace);
+    try {
+        auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
+        return realExec(op, extraPointers, hash, inputBuffers, inputShapes, numInputs, outputBuffers, outputShapes, numOutputs, tArgs, numTArgs, iArgs, numIArgs, bArgs, numBArgs, isInplace);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 1;
+    }
 }
 
 int registerGraph(Nd4jPointer *extraPointers, Nd4jLong graphId, Nd4jPointer flatBufferPointer) {
-    auto graph = nd4j::graph::GraphExecutioner::importFromFlatPointer(flatBufferPointer);
+    try {
+        auto graph = nd4j::graph::GraphExecutioner::importFromFlatPointer(flatBufferPointer);
 
-    nd4j::graph::GraphHolder::getInstance()->registerGraph(graphId, graph);
+        nd4j::graph::GraphHolder::getInstance()->registerGraph(graphId, graph);
 
-    return ND4J_STATUS_OK;
+        return ND4J_STATUS_OK;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 1;
+    }
 }
 
 static VariablesSet* executeStoredGraphT(Nd4jPointer *extraPointers, Nd4jLong graphId, Nd4jPointer *inputBuffers, Nd4jPointer *inputShapes, int* inputIndices, int numInputs) {
@@ -2478,7 +2536,13 @@ Nd4jStatus execCustomOpWithScope_(Nd4jPointer *extraPointers, nd4j::graph::Graph
 }
 
 Nd4jStatus execCustomOpWithScope(Nd4jPointer *extraPointers, Nd4jPointer state, Nd4jLong opHash, Nd4jLong *scopes, int numScopes, Nd4jPointer *inputBuffers, Nd4jPointer *inputShapes, int numInputs, Nd4jPointer *outputBuffers, Nd4jPointer *outputShapes, int numOutputs) {
-    return execCustomOpWithScope_(extraPointers, reinterpret_cast<nd4j::graph::GraphState*>(state), opHash, scopes, numScopes, inputBuffers, inputShapes, numInputs, outputBuffers, outputShapes, numOutputs);
+    try {
+        return execCustomOpWithScope_(extraPointers, reinterpret_cast<nd4j::graph::GraphState *>(state), opHash, scopes, numScopes, inputBuffers, inputShapes, numInputs, outputBuffers, outputShapes, numOutputs);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 1;
+    }
 }
 
 void deleteResultWrapper(Nd4jPointer ptr) {
@@ -2704,73 +2768,98 @@ void scatterUpdate(Nd4jPointer *extraPointers, int opCode, int numOfSubArrs,
                       void* dY, Nd4jLong* dYShapeInfo, Nd4jLong* dYOffsets,
                       int* hIindexes, int* dIindexes) {
 
+    try {
 
-    int numThreads = omp_get_max_threads();
+        int numThreads = omp_get_max_threads();
 
-    PRAGMA_OMP_PARALLEL_THREADS(numThreads)
-    {
-        for (int i = 0; i < numOfSubArrs; ++i) {
+        PRAGMA_OMP_PARALLEL_THREADS(numThreads)
+        {
+            for (int i = 0; i < numOfSubArrs; ++i) {
 
-            int threadIndex = omp_get_thread_num();
-            const auto xIndex = hIindexes[i];
-            const bool isOwner = xIndex < numThreads ? threadIndex == xIndex : threadIndex == xIndex % numThreads;
+                int threadIndex = omp_get_thread_num();
+                const auto xIndex = hIindexes[i];
+                const bool isOwner = xIndex < numThreads ? threadIndex == xIndex : threadIndex == xIndex % numThreads;
 
-            if (!isOwner)
-                continue;
-
-            NDArray inSubArr(reinterpret_cast<int8_t *>(hX) + (hXOffsets[hIindexes[i]] * DataTypeUtils::sizeOf(hXShapeInfo)), hXShapeInfo);
-            NDArray updSubArr(reinterpret_cast<int8_t *>(hY) + (hYOffsets[i] * DataTypeUtils::sizeOf(hXShapeInfo)), hYShapeInfo);
-
-            if (inSubArr.lengthOf() != updSubArr.lengthOf()) {
-                continue;
-            }
-
-            switch (opCode) {
-                case 0:
-                    inSubArr.applyPairwiseTransform(pairwise::Add, &updSubArr, &inSubArr, nullptr);
-                    break;
-                case 1:
-                    inSubArr.applyPairwiseTransform(pairwise::Subtract, &updSubArr, &inSubArr, nullptr);
-                    break;
-                case 2:
-                    inSubArr.applyPairwiseTransform(pairwise::Multiply, &updSubArr, &inSubArr, nullptr);
-                    break;
-                case 3:
-                    inSubArr.applyPairwiseTransform(pairwise::Divide, &updSubArr, &inSubArr, nullptr);
-                    break;
-                case 4:
-                    inSubArr.applyPairwiseTransform(pairwise::ReverseSubtract, &updSubArr, &inSubArr, nullptr);
-                    break;
-                case 5:
-                    inSubArr.applyPairwiseTransform(pairwise::ReverseDivide, &updSubArr, &inSubArr, nullptr);
-                    break;
-                case 6:
-                    inSubArr.applyPairwiseTransform(pairwise::CopyPws, &updSubArr, &inSubArr, nullptr);
-                    break;
-                default:
+                if (!isOwner)
                     continue;
+
+                NDArray inSubArr(
+                        reinterpret_cast<int8_t *>(hX) + (hXOffsets[hIindexes[i]] * DataTypeUtils::sizeOf(hXShapeInfo)),
+                        hXShapeInfo);
+                NDArray updSubArr(reinterpret_cast<int8_t *>(hY) + (hYOffsets[i] * DataTypeUtils::sizeOf(hXShapeInfo)),
+                                  hYShapeInfo);
+
+                if (inSubArr.lengthOf() != updSubArr.lengthOf()) {
+                    continue;
+                }
+
+                switch (opCode) {
+                    case 0:
+                        inSubArr.applyPairwiseTransform(pairwise::Add, &updSubArr, &inSubArr, nullptr);
+                        break;
+                    case 1:
+                        inSubArr.applyPairwiseTransform(pairwise::Subtract, &updSubArr, &inSubArr, nullptr);
+                        break;
+                    case 2:
+                        inSubArr.applyPairwiseTransform(pairwise::Multiply, &updSubArr, &inSubArr, nullptr);
+                        break;
+                    case 3:
+                        inSubArr.applyPairwiseTransform(pairwise::Divide, &updSubArr, &inSubArr, nullptr);
+                        break;
+                    case 4:
+                        inSubArr.applyPairwiseTransform(pairwise::ReverseSubtract, &updSubArr, &inSubArr, nullptr);
+                        break;
+                    case 5:
+                        inSubArr.applyPairwiseTransform(pairwise::ReverseDivide, &updSubArr, &inSubArr, nullptr);
+                        break;
+                    case 6:
+                        inSubArr.applyPairwiseTransform(pairwise::CopyPws, &updSubArr, &inSubArr, nullptr);
+                        break;
+                    default:
+                        continue;
+                }
             }
         }
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
     }
 }
 
 void inspectArray(Nd4jPointer *extraPointers, Nd4jPointer buffer, Nd4jLong *shapeInfo, Nd4jPointer specialBuffer, Nd4jLong *specialShapeInfo, Nd4jPointer debugInfo) {
-    auto p = reinterpret_cast<nd4j::DebugInfo*>(debugInfo);
-    NDArray array(buffer, shapeInfo);
-    nd4j::DebugHelper::retrieveDebugStatistics(p, &array);
+    try {
+        auto p = reinterpret_cast<nd4j::DebugInfo *>(debugInfo);
+        NDArray array(buffer, shapeInfo);
+        nd4j::DebugHelper::retrieveDebugStatistics(p, &array);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void tryPointer(Nd4jPointer extra, Nd4jPointer p, int len) {
-    auto buf = reinterpret_cast<int8_t*>(p);
-    int cnt = 0;
-    for (int i = 0; i < len; i++)
-        cnt += buf[cnt];
+    try {
+        auto buf = reinterpret_cast<int8_t *>(p);
+        int cnt = 0;
+        for (int i = 0; i < len; i++)
+            cnt += buf[cnt];
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 nd4j::ConstantDataBuffer* shapeBuffer(int rank, Nd4jLong *shape, Nd4jLong *strides, nd4j::DataType dtype, char order, Nd4jLong ews, bool empty) {
-    auto buffer = new ConstantDataBuffer();
-    *buffer = nd4j::ConstantShapeHelper::getInstance()->bufferForShapeInfo(ShapeDescriptor(dtype, order, shape, strides, rank, ews, empty));
-    return buffer;
+    try {
+        auto buffer = new ConstantDataBuffer();
+        *buffer = nd4j::ConstantShapeHelper::getInstance()->bufferForShapeInfo(
+                ShapeDescriptor(dtype, order, shape, strides, rank, ews, empty));
+        return buffer;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 void deleteShapeBuffer(nd4j::ConstantDataBuffer* ptr) {
@@ -2790,7 +2879,13 @@ nd4j::ConstantDataBuffer* constantBufferDouble(nd4j::DataType dtype, double *dat
 }
 
 nd4j::ConstantDataBuffer* constantBuffer(nd4j::DataType dtype, nd4j::ConstantDescriptor *descriptor) {
-    return nd4j::ConstantHelper::getInstance()->constantBuffer(*descriptor, dtype);
+    try {
+        return nd4j::ConstantHelper::getInstance()->constantBuffer(*descriptor, dtype);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 Nd4jPointer getConstantDataBufferPrimary(nd4j::ConstantDataBuffer* dbf) {
@@ -2808,7 +2903,13 @@ Nd4jLong getConstantDataBufferSizeOf(nd4j::ConstantDataBuffer* dbf) {
 
 
 nd4j::graph::Context* createGraphContext(int nodeId) {
-    return new nd4j::graph::Context(nodeId);
+    try {
+        return new nd4j::graph::Context(nodeId);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 nd4j::graph::RandomGenerator* getGraphContextRandomGenerator(nd4j::graph::Context* ptr) {
     return &ptr->randomGenerator();
@@ -2872,32 +2973,38 @@ int dataTypeFromNpyHeader(void *header) {
 }
 
 Nd4jPointer shapeBufferForNumpy(Nd4jPointer npyArray) {
-    cnpy::NpyArray arr = cnpy::loadNpyFromPointer(reinterpret_cast<char *>(npyArray));
-    unsigned int shapeSize = arr.shape.size();
-    std::vector<Nd4jLong> shape(shapeSize);
-    bool _empty = false;
-    for(unsigned int i = 0; i < shapeSize; i++) {
-        shape[i] = arr.shape[i];
+    try {
+        cnpy::NpyArray arr = cnpy::loadNpyFromPointer(reinterpret_cast<char *>(npyArray));
+        unsigned int shapeSize = arr.shape.size();
+        std::vector<Nd4jLong> shape(shapeSize);
+        bool _empty = false;
+        for (unsigned int i = 0; i < shapeSize; i++) {
+            shape[i] = arr.shape[i];
 
-        if (arr.shape[i] == 0)
-            _empty = true;
+            if (arr.shape[i] == 0)
+                _empty = true;
+        }
+
+        auto dtype = cnpy::dataTypeFromHeader(reinterpret_cast<char *>(npyArray));
+
+        Nd4jLong *shapeBuffer;
+        if (shape.size() == 1 && shape[0] == 0) {
+            // scalar case
+            shapeBuffer = nd4j::ShapeBuilders::createScalarShapeInfo(dtype);
+        } else if (_empty) {
+            if (shapeSize > 0)
+                shapeBuffer = nd4j::ShapeBuilders::emptyShapeInfo(dtype, arr.fortranOrder ? 'f' : 'c', shape);
+            else
+                shapeBuffer = nd4j::ShapeBuilders::emptyShapeInfo(dtype);
+        } else {
+            shapeBuffer = nd4j::ShapeBuilders::createShapeInfo(dtype, arr.fortranOrder ? 'f' : 'c', shape);
+        }
+        return reinterpret_cast<Nd4jPointer>(nd4j::ConstantShapeHelper::getInstance()->createFromExisting(shapeBuffer, true));
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
     }
-
-    auto dtype = cnpy::dataTypeFromHeader(reinterpret_cast<char *>(npyArray));
-
-    Nd4jLong *shapeBuffer;
-    if (shape.size() == 1 && shape[0] == 0) {
-    // scalar case
-        shapeBuffer = nd4j::ShapeBuilders::createScalarShapeInfo(dtype);
-    } else if (_empty) {
-        if (shapeSize > 0)
-            shapeBuffer = nd4j::ShapeBuilders::emptyShapeInfo(dtype, arr.fortranOrder ? 'f' : 'c', shape);
-        else
-            shapeBuffer = nd4j::ShapeBuilders::emptyShapeInfo(dtype);
-    } else {
-        shapeBuffer = nd4j::ShapeBuilders::createShapeInfo(dtype, arr.fortranOrder ? 'f' : 'c', shape);
-    }
-    return reinterpret_cast<Nd4jPointer>(nd4j::ConstantShapeHelper::getInstance()->createFromExisting(shapeBuffer, true));
 }
 
 void sortByKey(Nd4jPointer *extraPointers,
@@ -2906,10 +3013,15 @@ void sortByKey(Nd4jPointer *extraPointers,
                           void *y, Nd4jLong *yShapeInfo,
                           void *dy, Nd4jLong *dyShapeInfo,
                           bool descending) {
-    auto xType = ArrayOptions::dataType(xShapeInfo);
-    auto yType = ArrayOptions::dataType(yShapeInfo);
+    try {
+        auto xType = ArrayOptions::dataType(xShapeInfo);
+        auto yType = ArrayOptions::dataType(yShapeInfo);
 
-    BUILD_DOUBLE_SELECTOR(xType, yType, nd4j::DoubleMethods, ::sortByKey(x, xShapeInfo, y, yShapeInfo, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+        BUILD_DOUBLE_SELECTOR(xType, yType, nd4j::DoubleMethods, ::sortByKey(x, xShapeInfo, y, yShapeInfo, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void sortByValue(Nd4jPointer *extraPointers,
@@ -2918,11 +3030,15 @@ void sortByValue(Nd4jPointer *extraPointers,
                             void *y, Nd4jLong *yShapeInfo,
                             void *dy, Nd4jLong *dyShapeInfo,
                             bool descending) {
+    try {
+        auto xType = ArrayOptions::dataType(xShapeInfo);
+        auto yType = ArrayOptions::dataType(yShapeInfo);
 
-    auto xType = ArrayOptions::dataType(xShapeInfo);
-    auto yType = ArrayOptions::dataType(yShapeInfo);
-
-    BUILD_DOUBLE_SELECTOR(xType, yType, nd4j::DoubleMethods, ::sortByValue(x, xShapeInfo, y, yShapeInfo, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+        BUILD_DOUBLE_SELECTOR(xType, yType, nd4j::DoubleMethods, ::sortByValue(x, xShapeInfo, y, yShapeInfo, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void sortTadByKey(Nd4jPointer *extraPointers,
@@ -2933,10 +3049,15 @@ void sortTadByKey(Nd4jPointer *extraPointers,
                   int *dimension,
                   int dimensionLength,
                   bool descending) {
-    auto xType = ArrayOptions::dataType(xShapeInfo);
-    auto yType = ArrayOptions::dataType(yShapeInfo);
+    try {
+        auto xType = ArrayOptions::dataType(xShapeInfo);
+        auto yType = ArrayOptions::dataType(yShapeInfo);
 
-    BUILD_DOUBLE_SELECTOR(xType, yType, nd4j::DoubleMethods, ::sortTadByKey(x, xShapeInfo, y, yShapeInfo, dimension, dimensionLength, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+        BUILD_DOUBLE_SELECTOR(xType, yType, nd4j::DoubleMethods, ::sortTadByKey(x, xShapeInfo, y, yShapeInfo, dimension, dimensionLength, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void sortTadByValue(Nd4jPointer *extraPointers,
@@ -2947,24 +3068,35 @@ void sortTadByValue(Nd4jPointer *extraPointers,
                     int *dimension,
                     int dimensionLength,
                     bool descending) {
-    auto xType = ArrayOptions::dataType(xShapeInfo);
-    auto yType = ArrayOptions::dataType(yShapeInfo);
+    try {
+        auto xType = ArrayOptions::dataType(xShapeInfo);
+        auto yType = ArrayOptions::dataType(yShapeInfo);
 
-    BUILD_DOUBLE_SELECTOR(xType, yType, nd4j::DoubleMethods, ::sortTadByValue(x, xShapeInfo, y, yShapeInfo, dimension, dimensionLength, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+        BUILD_DOUBLE_SELECTOR(xType, yType, nd4j::DoubleMethods, ::sortTadByValue(x, xShapeInfo, y, yShapeInfo, dimension, dimensionLength, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 const char* runLightBenchmarkSuit(bool printOut) {
-    nd4j::LightBenchmarkSuit suit;
-    auto result = suit.runSuit();
+    try {
+        nd4j::LightBenchmarkSuit suit;
+        auto result = suit.runSuit();
 
-    if (printOut)
-        nd4j_printf("%s\n", result.data());
+        if (printOut)
+            nd4j_printf("%s\n", result.data());
 
-    auto chars = new char[result.length()+1];
-    std::memcpy(chars, result.data(), result.length());
-    chars[result.length()] = (char) 0x0;
+        auto chars = new char[result.length() + 1];
+        std::memcpy(chars, result.data(), result.length());
+        chars[result.length()] = (char) 0x0;
 
-    return chars;
+        return chars;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 Nd4jLong getCachedMemory(int deviceId) {
@@ -2972,17 +3104,23 @@ Nd4jLong getCachedMemory(int deviceId) {
 }
 
 const char* runFullBenchmarkSuit(bool printOut) {
-    nd4j::FullBenchmarkSuit suit;
-    auto result = suit.runSuit();
+    try {
+        nd4j::FullBenchmarkSuit suit;
+        auto result = suit.runSuit();
 
-    if (printOut)
-        nd4j_printf("%s\n", result.data());
+        if (printOut)
+            nd4j_printf("%s\n", result.data());
 
-    auto chars = new char[result.length()+1];
-    std::memcpy(chars, result.data(), result.length());
-    chars[result.length()] = (char) 0x0;
+        auto chars = new char[result.length() + 1];
+        std::memcpy(chars, result.data(), result.length());
+        chars[result.length()] = (char) 0x0;
 
-    return chars;
+        return chars;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 nd4j::LaunchContext* defaultLaunchContext() {
@@ -3017,8 +3155,14 @@ Nd4jPointer lcSolverHandle(OpaqueLaunchContext* lc) {
     return nullptr;
 }
 
+int lastErrorCode() {
+    return nd4j::LaunchContext::defaultContext()->errorReference()->errorCode();
+}
+
+const char* lastErrorMessage() {
+    return nd4j::LaunchContext::defaultContext()->errorReference()->errorMessage();
+}
 
-BUILD_SINGLE_TEMPLATE(template void flattenGeneric,(Nd4jPointer*, int, char, void*, Nd4jLong*, void*, Nd4jLong*), LIBND4J_TYPES);
 BUILD_SINGLE_TEMPLATE(template void pullRowsGeneric, (void *, Nd4jLong*, void*, Nd4jLong*, const int, Nd4jLong*, Nd4jLong*, Nd4jLong*, Nd4jLong*, Nd4jLong*), LIBND4J_TYPES);
 BUILD_SINGLE_TEMPLATE(template void tearGeneric, (void *, Nd4jLong*, Nd4jPointer*, Nd4jLong*, Nd4jLong*, Nd4jLong*), LIBND4J_TYPES);
 BUILD_SINGLE_TEMPLATE(template void shuffleGeneric, (void**, Nd4jLong**, void**, Nd4jLong**, int, int*, Nd4jLong**, Nd4jLong**), LIBND4J_TYPES);
diff --git a/libnd4j/blas/cuda/NativeOps.cu b/libnd4j/blas/cuda/NativeOps.cu
index e75aa422c..626b0ea26 100755
--- a/libnd4j/blas/cuda/NativeOps.cu
+++ b/libnd4j/blas/cuda/NativeOps.cu
@@ -68,21 +68,6 @@ int minThreads = 32;
 
 __constant__ char deviceConstantMemory[49152];
 
-typedef struct {
-    long streamId;
-    long callId;
-} __syncInfo;
-
-typedef __syncInfo SyncInfo;
-
-
-// this method isn't used, left here for legacy and caution purposes
-// TLDR: don't use this way, it sucks
-void CUDART_CB syncCallback(cudaStream_t stream, cudaError_t status, void *data){
-    SyncInfo *sync = reinterpret_cast<SyncInfo *>(data);
-
-    //printf("Finished stream: [%i], kernel call: [%i]\n", sync->streamId, sync->callId);
-}
 
 // this method just does type conversion in fancy way
 int getDeviceId(Nd4jPointer ptrToDeviceId) {
@@ -250,9 +235,14 @@ void execPairwiseTransform( Nd4jPointer *extraPointers,
         								void *hZ, Nd4jLong *hZShapeInfo,
         								void *dZ, Nd4jLong *dZShapeInfo,
         								void *extraParams) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execPairwiseTransform(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraParams);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execPairwiseTransform(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY,
+                                                   dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraParams);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -265,9 +255,14 @@ void execPairwiseTransformBool(Nd4jPointer *extraPointers,
         								void *hZ, Nd4jLong *hZShapeInfo,
         								void *dZ, Nd4jLong *dZShapeInfo,
         								void *extraParams) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execPairwiseBoolTransform(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraParams);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execPairwiseBoolTransform(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo,
+                                                       dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraParams);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -279,9 +274,14 @@ void execSummaryStatsScalar(Nd4jPointer *extraPointers,
                                        void *hZ, Nd4jLong *hZShapeInfo,
                                        void *dZ, Nd4jLong *dZShapeInfo,
                                        bool biasCorrected) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execSummaryStatsScalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo, biasCorrected);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execSummaryStatsScalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ,
+                                                    hZShapeInfo, dZ, dZShapeInfo, biasCorrected);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -295,24 +295,30 @@ void execBroadcastBool(Nd4jPointer *extraPointers,
         						void *dZ, Nd4jLong *dZShapeInfo,
         						void *hDimension, Nd4jLong *hDimensionShape,
 		void *dDimension, Nd4jLong *dDimensionShape) {
+    try {
+        //Nd4jLong *tadOnlyShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[0]);
+        //Nd4jLong *tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers[1]);
+        //Nd4jLong *tadOnlyShapeInfoZ = reinterpret_cast<Nd4jLong *>(extraPointers[2]);
+        //Nd4jLong *tadOffsetsZ = reinterpret_cast<Nd4jLong *>(extraPointers[3]);
 
-	//Nd4jLong *tadOnlyShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[0]);
-    //Nd4jLong *tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers[1]);
-    //Nd4jLong *tadOnlyShapeInfoZ = reinterpret_cast<Nd4jLong *>(extraPointers[2]);
-    //Nd4jLong *tadOffsetsZ = reinterpret_cast<Nd4jLong *>(extraPointers[3]);
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
-
-	auto hTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[9]);
-	auto tadOnlyShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[10]);
-	auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers[11]);
-	auto tadOnlyShapeInfoZ = reinterpret_cast<Nd4jLong *>(extraPointers[12]);
-	auto tadOffsetsZ = reinterpret_cast<Nd4jLong *>(extraPointers[13]);
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execBroadcastBool(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ);
+        auto hTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[9]);
+        auto tadOnlyShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[10]);
+        auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers[11]);
+        auto tadOnlyShapeInfoZ = reinterpret_cast<Nd4jLong *>(extraPointers[12]);
+        auto tadOffsetsZ = reinterpret_cast<Nd4jLong *>(extraPointers[13]);
 
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execBroadcastBool(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY,
+                                               dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension,
+                                               dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ,
+                                               tadOffsetsZ);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -338,38 +344,33 @@ void   execBroadcast(
 		void *dZ, Nd4jLong *dZShapeInfo,
 		void *hDimension, Nd4jLong *hDimensionShape,
 		void *dDimension, Nd4jLong *dDimensionShape) {
-/*
-    cudaEvent_t start;
-    cudaEventCreateWithFlags(&start, cudaEventDisableTiming);
-    timespec tsX;
-    timespec tsY;
-    clock_gettime(CLOCK_REALTIME, &tsX);
-*/
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-	cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
 
-	auto hTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[9]);
-	auto tadOnlyShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[10]);
-	auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers[11]);
-	auto tadOnlyShapeInfoZ = reinterpret_cast<Nd4jLong *>(extraPointers[12]);
-	auto tadOffsetsZ = reinterpret_cast<Nd4jLong *>(extraPointers[13]);
+        auto hTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[9]);
+        auto tadOnlyShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[10]);
+        auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers[11]);
+        auto tadOnlyShapeInfoZ = reinterpret_cast<Nd4jLong *>(extraPointers[12]);
+        auto tadOffsetsZ = reinterpret_cast<Nd4jLong *>(extraPointers[13]);
 
-	auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
-	auto yType = nd4j::ArrayOptions::dataType(hYShapeInfo);
-    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+        auto yType = nd4j::ArrayOptions::dataType(hYShapeInfo);
+        auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
 
-	if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-		printf("F3 opNum:[%i]\n", opNum);
+        if (nd4j::Environment::getInstance()->isDebugAndVerbose())
+            printf("F3 opNum:[%i]\n", opNum);
 
-	//Nd4jLong *tadOnlyShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[0]);
-    //Nd4jLong *tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers[1]);
-    //Nd4jLong *tadOnlyShapeInfoZ = reinterpret_cast<Nd4jLong *>(extraPointers[2]);
-    //Nd4jLong *tadOffsetsZ = reinterpret_cast<Nd4jLong *>(extraPointers[3]);
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execBroadcast(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ);
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execBroadcast(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY,
+                                           dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength,
+                                           tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -390,9 +391,14 @@ void execReduceFloat(Nd4jPointer *extraPointers,
 							void *extraParams,
 							void *hZ, Nd4jLong *hZShapeInfo,
 							void *dZ, Nd4jLong *dZShapeInfo) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execReduceFloatScalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execReduceFloatScalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ,
+                                                   hZShapeInfo, dZ, dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -403,9 +409,14 @@ void execReduceSame(Nd4jPointer *extraPointers,
                                 void *extraParams,
                                 void *hZ, Nd4jLong *hZShapeInfo,
                                 void *dZ, Nd4jLong *dZShapeInfo) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execReduceSameScalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execReduceSameScalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ,
+                                                  hZShapeInfo, dZ, dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -418,13 +429,22 @@ void execReduceSame2(Nd4jPointer *extraPointers,
                             void *dZ, Nd4jLong *dZShapeInfo,
 							   void *hDimension, Nd4jLong *hDimensionShape,
 							   void *dDimension, Nd4jLong *dDimensionShape) {
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-	auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, reinterpret_cast<int*>(hDimension), shape::length(hDimensionShape));
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo,
+                                                                                reinterpret_cast<int *>(hDimension),
+                                                                                shape::length(hDimensionShape));
 
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execReduceSame(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadPack.specialShapeInfo(), tadPack.specialOffsets());
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execReduceSame(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo,
+                                            dZ, dZShapeInfo, dimension, dimensionLength, tadPack.specialShapeInfo(),
+                                            tadPack.specialOffsets());
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -437,13 +457,22 @@ void execReduceLong2(Nd4jPointer *extraPointers,
                             void *dZ, Nd4jLong *dZShapeInfo,
 							   void *hDimension, Nd4jLong *hDimensionShape,
 							   void *dDimension, Nd4jLong *dDimensionShape) {
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, reinterpret_cast<int*>(hDimension), shape::length(hDimensionShape));
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo,
+                                                                                reinterpret_cast<int *>(hDimension),
+                                                                                shape::length(hDimensionShape));
 
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execReduceLong(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadPack.specialShapeInfo(), tadPack.specialOffsets());
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execReduceLong(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo,
+                                            dZ, dZShapeInfo, dimension, dimensionLength, tadPack.specialShapeInfo(),
+                                            tadPack.specialOffsets());
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -454,30 +483,37 @@ void   execReduceLong(Nd4jPointer *extraPointers,
                                 void *extraParams,
                                 void *hZ, Nd4jLong *hZShapeInfo,
                                 void *dZ, Nd4jLong *dZShapeInfo) {
+    try {
+        auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        auto hTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[9]);
+        auto dTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[10]);
 
-    auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-    auto hTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[9]);
-	auto dTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[10]);
+        if (nd4j::Environment::getInstance()->isDebugAndVerbose())
+            printf("LF7 opNum:[%i]\n", opNum);
 
-    if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-        printf("LF7 opNum:[%i]\n", opNum);
+        auto reductionPointer = reinterpret_cast<void *>(extraPointers[4]);
 
-    auto reductionPointer = reinterpret_cast<void *>(extraPointers[4]);
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+        auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
 
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
-    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+        if (zType != nd4j::DataType::INT64)
+            throw datatype_exception::build("execReduceLong wrong Z data type", nd4j::DataType::INT64, zType);
 
-    if (zType != nd4j::DataType::INT64)
-        throw datatype_exception::build("execReduceLong wrong Z data type", nd4j::DataType::INT64, zType);
+        auto xLength = shape::length(hXShapeInfo);
+        auto blockWidth = 256;
+        auto numBlocks = CudaLaunchHelper::getReductionBlocks(xLength, blockWidth);
+        dim3 launchDims(numBlocks, blockWidth, 32768);
 
-    auto xLength = shape::length(hXShapeInfo);
-    auto blockWidth = 256;
-    auto numBlocks = CudaLaunchHelper::getReductionBlocks(xLength, blockWidth);
-    dim3 launchDims(numBlocks, blockWidth, 32768);
+        BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceLongFunction,
+                              ::execReduceScalar(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams,
+                                                 dZ, dZShapeInfo, hXShapeInfo, nullptr, 0, reductionPointer,
+                                                 dTADShapeInfo), LIBND4J_TYPES, LONG_TYPES);
 
-    BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceLongFunction, ::execReduceScalar(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams, dZ, dZShapeInfo, hXShapeInfo, nullptr, 0, reductionPointer, dTADShapeInfo), LIBND4J_TYPES, LONG_TYPES);
-
-    nd4j::DebugHelper::checkErrorCode(stream, "execReduceLong(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "execReduceLong(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -490,13 +526,22 @@ void execReduceBool2(Nd4jPointer *extraPointers,
                             void *dZ, Nd4jLong *dZShapeInfo,
 							   void *hDimension, Nd4jLong *hDimensionShape,
 							   void *dDimension, Nd4jLong *dDimensionShape) {
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, reinterpret_cast<int*>(hDimension), shape::length(hDimensionShape));
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo,
+                                                                                reinterpret_cast<int *>(hDimension),
+                                                                                shape::length(hDimensionShape));
 
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execReduceBool(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadPack.specialShapeInfo(), tadPack.specialOffsets());
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execReduceBool(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo,
+                                            dZ, dZShapeInfo, dimension, dimensionLength, tadPack.specialShapeInfo(),
+                                            tadPack.specialOffsets());
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -507,30 +552,37 @@ void   execReduceBool(Nd4jPointer *extraPointers,
                                 void *extraParams,
                                 void *hZ, Nd4jLong *hZShapeInfo,
                                 void *dZ, Nd4jLong *dZShapeInfo) {
+    try {
+        auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        auto hTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[9]);
+        auto dTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[10]);
 
-    auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-    auto hTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[9]);
-	auto dTADShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[10]);
+        if (nd4j::Environment::getInstance()->isDebugAndVerbose())
+            printf("BF7 opNum:[%i]\n", opNum);
 
-    if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-        printf("BF7 opNum:[%i]\n", opNum);
+        auto reductionPointer = reinterpret_cast<void *>(extraPointers[4]);
 
-    auto reductionPointer = reinterpret_cast<void *>(extraPointers[4]);
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+        auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
 
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
-    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+        if (zType != nd4j::DataType::BOOL)
+            throw std::runtime_error("execReduceBool requires Z operand to have BOOL type");
 
-    if (zType != nd4j::DataType::BOOL)
-        throw std::runtime_error("execReduceBool requires Z operand to have BOOL type");
+        auto xLength = shape::length(hXShapeInfo);
+        auto blockWidth = 256;
+        auto numBlocks = CudaLaunchHelper::getReductionBlocks(xLength, blockWidth);
+        dim3 launchDims(numBlocks, blockWidth, 32768);
 
-    auto xLength = shape::length(hXShapeInfo);
-    auto blockWidth = 256;
-    auto numBlocks = CudaLaunchHelper::getReductionBlocks(xLength, blockWidth);
-    dim3 launchDims(numBlocks, blockWidth, 32768);
+        BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceBoolFunction,
+                              ::execReduceScalar(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams,
+                                                 dZ, dZShapeInfo, hZShapeInfo, nullptr, 0, reductionPointer,
+                                                 dTADShapeInfo), LIBND4J_TYPES, BOOL_TYPES);
 
-    BUILD_DOUBLE_SELECTOR(xType, zType, functions::reduce::ReduceBoolFunction, ::execReduceScalar(launchDims, stream, opNum, dX, dXShapeInfo, hXShapeInfo, extraParams, dZ, dZShapeInfo, hZShapeInfo, nullptr, 0, reductionPointer, dTADShapeInfo), LIBND4J_TYPES, BOOL_TYPES);
-
-    nd4j::DebugHelper::checkErrorCode(stream, "execReduceBool(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "execReduceBool(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -554,13 +606,22 @@ void execIndexReduce(Nd4jPointer *extraPointers,
         						 void *dZ, Nd4jLong *dZShapeInfo,
 								 void *hDimension, Nd4jLong *hDimensionShape,
 		void *dDimension, Nd4jLong *dDimensionShape) {
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, reinterpret_cast<int*>(hDimension), shape::length(hDimensionShape));
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo,
+                                                                                reinterpret_cast<int *>(hDimension),
+                                                                                shape::length(hDimensionShape));
 
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execIndexReduce(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadPack.specialShapeInfo(), tadPack.specialOffsets());
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execIndexReduce(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo,
+                                             dZ, dZShapeInfo, dimension, dimensionLength, tadPack.specialShapeInfo(),
+                                             tadPack.specialOffsets());
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -582,13 +643,22 @@ void execReduceFloat2(Nd4jPointer *extraPointers,
 								void *dZ, Nd4jLong *dZShapeInfo,
 								void *hDimension, Nd4jLong *hDimensionShape,
 		void *dDimension, Nd4jLong *dDimensionShape) {
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, reinterpret_cast<int*>(hDimension), shape::length(hDimensionShape));
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo,
+                                                                                reinterpret_cast<int *>(hDimension),
+                                                                                shape::length(hDimensionShape));
 
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execReduceFloat(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadPack.specialShapeInfo(), tadPack.specialOffsets());
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execReduceFloat(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo,
+                                             dZ, dZShapeInfo, dimension, dimensionLength, tadPack.specialShapeInfo(),
+                                             tadPack.specialOffsets());
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 /**
@@ -607,9 +677,14 @@ void execIndexReduceScalar(
         void *extraParams,
         void *hZ, Nd4jLong *hZShapeInfo,
 		void *dZ, Nd4jLong *dZShapeInfo){
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execIndexReduceScalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execIndexReduceScalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ,
+                                                   hZShapeInfo, dZ, dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -619,12 +694,17 @@ void execTransformSame(Nd4jPointer *extraPointers,int opNum,
                                    void *hZ, Nd4jLong *hZShapeInfo,
                                    void *dZ, Nd4jLong *dZShapeInfo,
                                    void *extraParams) {
+    try {
+        auto tadShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[0] : nullptr);
+        auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[1] : nullptr);
 
-    auto tadShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[0] : nullptr);
-    auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[1] : nullptr);
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execTransformSame(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraParams, tadShapeInfo, tadOffsets);
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execTransformSame(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ,
+                                               dZShapeInfo, extraParams, tadShapeInfo, tadOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -634,12 +714,17 @@ void execTransformBool(Nd4jPointer *extraPointers,int opNum,
 								  void *hZ, Nd4jLong *hZShapeInfo,
 								  void *dZ, Nd4jLong *dZShapeInfo,
 								  void *extraParams) {
+    try {
+        auto tadShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[0] : nullptr);
+        auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[1] : nullptr);
 
-	auto tadShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[0] : nullptr);
-    auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[1] : nullptr);
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execTransformBool(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraParams, tadShapeInfo, tadOffsets);
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execTransformBool(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ,
+                                               dZShapeInfo, extraParams, tadShapeInfo, tadOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -649,12 +734,18 @@ void execTransformAny(Nd4jPointer *extraPointers,int opNum,
 								  void *hZ, Nd4jLong *hZShapeInfo,
 								  void *dZ, Nd4jLong *dZShapeInfo,
 								  void *extraParams) {
+    try {
+        auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        auto streamSpecial = reinterpret_cast<cudaStream_t &>(extraPointers[4]);
+        LaunchContext lc(stream, streamSpecial, extraPointers[5], extraPointers[3],
+                         reinterpret_cast<int *>(extraPointers[6]));
 
-    auto stream = reinterpret_cast<cudaStream_t*>(extraPointers[1]);
-    auto streamSpecial = reinterpret_cast<cudaStream_t&>(extraPointers[4]);
-    LaunchContext lc(stream, streamSpecial, extraPointers[5], extraPointers[3], reinterpret_cast<int*>(extraPointers[6]));
-
-    NativeOpExecutioner::execTransformAny(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraParams, nullptr, nullptr);
+        NativeOpExecutioner::execTransformAny(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ,
+                                              dZShapeInfo, extraParams, nullptr, nullptr);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -664,12 +755,17 @@ void execTransformStrict(Nd4jPointer *extraPointers,int opNum,
                                   void *hZ, Nd4jLong *hZShapeInfo,
                                   void *dZ, Nd4jLong *dZShapeInfo,
                                   void *extraParams) {
+    try {
+        auto tadShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[10] : nullptr);
+        auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[11] : nullptr);
 
-    auto tadShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[10] : nullptr);
-    auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[11] : nullptr);
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execTransformStrict(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraParams, tadShapeInfo, tadOffsets);
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execTransformStrict(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ,
+                                                 dZShapeInfo, extraParams, tadShapeInfo, tadOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -679,55 +775,19 @@ void execTransformFloat(Nd4jPointer *extraPointers,int opNum,
                                     void *hZ, Nd4jLong *hZShapeInfo,
                                     void *dZ, Nd4jLong *dZShapeInfo,
                                     void *extraParams) {
+    try {
+        auto tadShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[10] : nullptr);
+        auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[11] : nullptr);
 
-    auto tadShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[10] : nullptr);
-    auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers != nullptr ? extraPointers[11] : nullptr);
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execTransformFloat(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraParams, tadShapeInfo, tadOffsets);
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execTransformFloat(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ,
+                                                dZShapeInfo, extraParams, tadShapeInfo, tadOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
-
-/**
- * Append an input array
- * to the end of a flat array
- * in a particular order
- * @param offset the offset of the array to start at
- * @param order the order
- * @param dZ the dZ array
- * @param dZShapeInfo the shape info for te array
- * @param input the input for the array
- * @param inputShapeInfo the shape information for that array
- */
-void flatten(Nd4jPointer *extraPointers,
-						int offset,
-						char order,
-						void *hZ, Nd4jLong *hZShapeInfo,
-						void *dZ, Nd4jLong *dZShapeInfo,
-						void *hInput, Nd4jLong *hInputShapeInfo,
-						void *dInput, Nd4jLong *dInputShapeInfo) {
-
-	cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-	auto hYShapeInfo = reinterpret_cast<Nd4jLong *>(extraPointers[7]);
-
-	if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-		printf("F22 opNum:[7]\n");
-
-	// int *allocPointer = reinterpret_cast<int *>(extraPointers[3]);
-
-	dim3 launchDims(256, 256, 2048);
-
-	if (nd4j::Environment::getInstance()->isVerbose() && launchDims.x == 1)
-		printf("AF222 opNum:[7]\n");
-
-	auto type = nd4j::ArrayOptions::dataType(hInputShapeInfo);
-    BUILD_SINGLE_SELECTOR(type, flattenKernelGeneric, (launchDims, stream, extraPointers, offset, order, dZ, dZShapeInfo, dInput, dInputShapeInfo), LIBND4J_TYPES);
-
-	DEBUG_KERNEL(stream, -1);
-}
-
-
-
 void checkP2P() {
 	int curDevice = 0;
 
@@ -821,23 +881,28 @@ bool isP2PAvailable() {
 
 
 void initializeDevicesAndFunctions() {
-	int devCnt = 0;
-	cudaGetDeviceCount(&devCnt);
-	deviceProperties = new cudaDeviceProp[devCnt];
-	for (int i = 0; i < devCnt; i++) {
-		cudaSetDevice(i);
-		cudaGetDeviceProperties(&deviceProperties[i], i);
+    try {
+        int devCnt = 0;
+        cudaGetDeviceCount(&devCnt);
+        deviceProperties = new cudaDeviceProp[devCnt];
+        for (int i = 0; i < devCnt; i++) {
+            cudaSetDevice(i);
+            cudaGetDeviceProperties(&deviceProperties[i], i);
 
-		cudaDeviceSetLimit(cudaLimitStackSize, 4096);
-	}
+            cudaDeviceSetLimit(cudaLimitStackSize, 4096);
+        }
 
-	cudaSetDevice(0);
+        cudaSetDevice(0);
 
-	checkP2P();
+        checkP2P();
 
-	// enabling p2p gpu access if it's supported
-	if (supportedP2P && devCnt > 1)
-    	enableP2P(allowedP2P);
+        // enabling p2p gpu access if it's supported
+        if (supportedP2P && devCnt > 1)
+            enableP2P(allowedP2P);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void initializeFunctions(Nd4jPointer *functions) {
@@ -867,8 +932,10 @@ Nd4jPointer mallocHost(Nd4jLong memorySize, int flags) {
 	Nd4jPointer pointer;
 	// cudaHostAllocMapped |cudaHostAllocPortable
 	auto res = cudaHostAlloc(reinterpret_cast<void **>(&pointer), memorySize, cudaHostAllocDefault);
-	if (res != 0)
-		throw nd4j::cuda_exception::build("cudaHostAlloc(...) failed", res);
+	if (res != 0) {
+	    nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(res);
+	    nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaHostAlloc failed");
+    }
 
 	return pointer;
 }
@@ -884,8 +951,11 @@ Nd4jPointer mallocHost(Nd4jLong memorySize, int flags) {
 Nd4jPointer mallocDevice(Nd4jLong memorySize, int deviceId, int flags) {
 	Nd4jPointer pointer;
 	auto res = cudaMalloc(reinterpret_cast<void **>(&pointer), memorySize);
-	if (res != 0)
-		throw nd4j::cuda_exception::build("cudaMalloc(...) failed", res);
+	if (res != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(res);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaMalloc failed");
+	}
+
 	return pointer;
 }
 
@@ -896,8 +966,11 @@ Nd4jPointer mallocDevice(Nd4jLong memorySize, int deviceId, int flags) {
  */
 int freeHost(Nd4jPointer pointer) {
 	auto res = cudaFreeHost(reinterpret_cast<void *>(pointer));
-	if (res != 0)
-		throw nd4j::cuda_exception::build("cudaFreeHost(...) failed", res);
+	if (res != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(res);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaFreeHost failed");
+    }
+
 	return 1L;
 }
 
@@ -909,10 +982,14 @@ int freeHost(Nd4jPointer pointer) {
  */
 int freeDevice(Nd4jPointer pointer, int deviceId) {
 	auto res = cudaFree(reinterpret_cast<void *>(pointer));
-	if (res != 0)
-		throw nd4j::cuda_exception::build("cudaFree(...) failed", res);
 
-	return 1L;
+	// we're intentionally skipping
+	if (res != 0 && res != 1) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(res);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaFree failed");
+	}
+
+	return res == 0 ? 1L : 0L;
 }
 
 
@@ -921,22 +998,13 @@ Nd4jPointer createContext() {
 }
 
 Nd4jPointer createStream() {
-    /*
-	Nd4jPointer nativeStream = (Nd4jPointer) malloc(sizeof(cudaStream_t));
 
-	CHECK_ALLOC(nativeStream, "Failed to allocate memory for new CUDA stream", sizeof(cudaStream_t));
-
-	cudaError_t dZ = cudaStreamCreate(reinterpret_cast<cudaStream_t *>(&nativeStream));
-	checkCudaErrors(dZ);
-	if (dZ != 0)
-		throw std::runtime_error("cudaStreamCreate(...) failed");
-
-	return nativeStream;
-     */
     auto stream = new cudaStream_t();
     auto dZ = cudaStreamCreate(stream);
-    if (dZ != 0)
-        throw nd4j::cuda_exception::build("cudaStreamCreate(...) failed", dZ);
+    if (dZ != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaStreamCreate failed");
+    }
 
     return stream;
 }
@@ -947,9 +1015,10 @@ Nd4jPointer createEvent() {
 	CHECK_ALLOC(nativeEvent, "Failed to allocate new CUDA event buffer", sizeof(cudaEvent_t));
 
 	auto dZ = cudaEventCreateWithFlags(reinterpret_cast<cudaEvent_t *>(&nativeEvent), cudaEventDisableTiming);
-	if (dZ != 0)
-		throw nd4j::cuda_exception::build("cudaEventCreateWithFlags(...) failed", dZ);
-
+	if (dZ != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaEventCreateWithFlags failed");
+	}
 
 	return nativeEvent;
 }
@@ -959,8 +1028,10 @@ int registerEvent(Nd4jPointer event, Nd4jPointer stream) {
 	auto pStream = reinterpret_cast<cudaStream_t *>(stream);
 
     auto dZ = cudaEventRecord(*pEvent, *pStream);
-	if (dZ != 0)
-		throw nd4j::cuda_exception::build("cudaEventRecord(...) failed", dZ);
+	if (dZ != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaEventRecord failed");
+	}
 
 	return 1;
 }
@@ -1048,8 +1119,11 @@ int memcpyAsync(Nd4jPointer dst, Nd4jPointer src, Nd4jLong size, int flags, Nd4j
 			    kind = cudaMemcpyDeviceToDevice;
 		    }
 			break;
-		default:
-			throw nd4j::cuda_exception::build("UNDEFINED MEMCPY!\n", 119);
+		default: {
+            nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+            nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("UNDEFNED MEMCPY");
+            return 0;
+		}
 	}
 
 	auto dZ = cudaMemcpyAsync(reinterpret_cast<void *>(dst), const_cast<const void *>(reinterpret_cast<void *>(src)), static_cast<size_t>(size), kind, *pStream);
@@ -1058,7 +1132,8 @@ int memcpyAsync(Nd4jPointer dst, Nd4jPointer src, Nd4jLong size, int flags, Nd4j
         printf("Failed on [%lu] -> [%lu], size: [%i], direction: [%i], dZ: [%i]\n", src, dst, size, flags, static_cast<int>(dZ));
         fflush(stdout);
         fflush(stderr);
-        throw nd4j::cuda_exception::build("cudaMemcpyAsync(...) failed", dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaMemcpyAsync failed");
 	}
 
 	return 1;
@@ -1066,8 +1141,10 @@ int memcpyAsync(Nd4jPointer dst, Nd4jPointer src, Nd4jLong size, int flags, Nd4j
 
 int memsetSync(Nd4jPointer dst, int value, Nd4jLong size, int flags, Nd4jPointer reserved) {
 	auto dZ = cudaMemset(reinterpret_cast<void *>(dst), value, static_cast<size_t>(size));
-	if (dZ != 0)
-		throw nd4j::cuda_exception::build("cudaMemset(...) failed", dZ);
+	if (dZ != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaMemset failed");
+	}
 
 	return 1;
 }
@@ -1076,8 +1153,10 @@ int memsetAsync(Nd4jPointer dst, int value, Nd4jLong size, int flags, Nd4jPointe
 	auto pStream = reinterpret_cast<cudaStream_t *>(reserved);
 
 	auto dZ = cudaMemsetAsync(reinterpret_cast<void *>(dst), value, static_cast<size_t>(size), *pStream);
-	if (dZ != 0)
-		throw nd4j::cuda_exception::build("cudaMemsetAsync(...) failed", dZ);
+	if (dZ != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaMemsetAsync failed");
+	}
 
 	return 1;
 }
@@ -1085,8 +1164,10 @@ int memsetAsync(Nd4jPointer dst, int value, Nd4jLong size, int flags, Nd4jPointe
 int destroyEvent(Nd4jPointer event) {
 	auto pEvent = reinterpret_cast<cudaEvent_t *>(&event);
 	auto dZ = cudaEventDestroy(*pEvent);
-	if (dZ != 0)
-		throw nd4j::cuda_exception::build("cudaEvenDestroy(...) failed", dZ);
+	if (dZ != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaEventDestroy failed");
+	}
 
 	return 1;
 }
@@ -1095,8 +1176,10 @@ int streamSynchronize(Nd4jPointer stream) {
 	auto pStream = reinterpret_cast<cudaStream_t *>(stream);
 
 	auto dZ = cudaStreamSynchronize(*pStream);
-	if (dZ != 0)
-        throw nd4j::cuda_exception::build("cudaStreamSynchronize(...) failed", dZ);
+	if (dZ != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaStreamSynchronize failed");
+	}
 
 	return 1L;
 }
@@ -1105,8 +1188,10 @@ int eventSynchronize(Nd4jPointer event) {
 	auto pEvent = reinterpret_cast<cudaEvent_t *>(&event);
 
 	auto  dZ = cudaEventSynchronize(*pEvent);
-	if (dZ != 0)
-        throw nd4j::cuda_exception::build("cudaEventSynchronize(...) failed", dZ);
+	if (dZ != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaEventSynchronize failed");
+	}
 
 	return 1L;
 }
@@ -1162,268 +1247,6 @@ const char * getDeviceName(int device) {
     return deviceProperties[device].name;
 }
 
-///////////////////////////////////////////////////////////////////
-template<typename T>
-__global__ static void concatCuda(const int numOfArrs, void* pVx,  void* pxShapeInfo, void* pVz, void* pzShapeInfo) {
-
-    __shared__ int arrIdx, blocksPerArr;
-    __shared__ T *x, *z;
-    __shared__ Nd4jLong *zShapeInfo, *xShapeInfo, arrLen, arrLenZ, arrLenPerBlock, start, end;
-
-    if (threadIdx.x == 0) {
-        blocksPerArr = (gridDim.x - gridDim.x % numOfArrs) / numOfArrs;     // floor
-        arrIdx = blockIdx.x / blocksPerArr;
-        if (arrIdx >= numOfArrs)
-            arrIdx = numOfArrs - 1;
-        x = reinterpret_cast<T*>(reinterpret_cast<void**>(pVx)[arrIdx]);
-        z = reinterpret_cast<T*>(reinterpret_cast<void**>(pVz)[arrIdx]);
-        xShapeInfo = reinterpret_cast<Nd4jLong**>(pxShapeInfo)[arrIdx];
-        zShapeInfo = reinterpret_cast<Nd4jLong**>(pzShapeInfo)[arrIdx];
-
-        arrLen = shape::length(xShapeInfo);
-        arrLenZ = shape::length(zShapeInfo);
-        arrLenPerBlock = (arrLen + blocksPerArr - arrLen % blocksPerArr) / blocksPerArr;  // ceil
-
-        start = arrLenPerBlock * (blockIdx.x % blocksPerArr);
-        end   = (start + arrLenPerBlock) > arrLen ? arrLen : (start + arrLenPerBlock);
-    }
-    __syncthreads();
-
-    for (Nd4jLong i = threadIdx.x + start; i < end; i += blockDim.x) {
-        auto zOffset = shape::getIndexOffset(i, zShapeInfo, arrLenZ);
-        auto xOffset = shape::getIndexOffset(i, xShapeInfo, arrLen);
-        //printf("z[%i][%lld] = x[%i][%lld]\n", arrIdx, zOffset, arrIdx, xOffset);
-        z[zOffset] = x[xOffset];
-    }
-}
-template<typename T>
-__host__ static void concatCudaLauncher(const int numOfArrs, cudaStream_t *stream,  void* pVx, void* pxShapeInfo, void* pVz, void* pzShapeInfo) {
-    //int blocks = numOfArrs * 16; // >> 1 << 2);
-    //nd4j_printf("gridDim.x is %i\n", blocks);
-    //if (blocks > 8192)
-    //    blocks = 8192; // restrict grid dims to 8K max
-    concatCuda<T><<<numOfArrs, 128, 512, *stream>>>(numOfArrs, pVx, pxShapeInfo, pVz, pzShapeInfo);
-    nd4j::DebugHelper::checkErrorCode(stream, "concat(...) failed");
-}
-BUILD_SINGLE_TEMPLATE(template void concatCudaLauncher, (const int numOfArrs, cudaStream_t *stream,  void* pVx, void* pxShapeInfo, void* pVz, void* pzShapeInfo), LIBND4J_TYPES);
-
-static void
-specialBufferAndShapeWithOffset(void* vZ, Nd4jLong* hZShapeInfo, Nd4jLong* dZShapeInfo, std::vector<Nd4jLong> const& idx, void*& outBuffer, Nd4jLong*& outShape) {
-    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
-    const int rank = shape::rank(hZShapeInfo);
-    Nd4jLong* newShape = new Nd4jLong[shape::shapeInfoLength(rank)];
-    //ALLOCATE(newShape, nullptr, , Nd4jLong)
-    auto shapeSize = shape::shapeInfoByteLength(rank);
-    memcpy(newShape, hZShapeInfo, shapeSize);
-
-    auto shapeOf = shape::shapeOf(newShape);
-    auto stridesOf = shape::stride(newShape);
-
-    Nd4jLong offset(0), subArrLen(1);
-    int n(2), first, last, stride;
-
-    for (int d = rank - 1; d >= 0; --d) {
-
-        if (idx[n * d] != idx[n * d + 1]) {
-            auto axeDim = shape::sizeAt(hZShapeInfo, d);
-            first  = idx[n * d]     >= 0 ? idx[n * d]     : idx[n * d]     + axeDim + 1;
-            last   = idx[n * d + 1] >= 0 ? idx[n * d + 1] : idx[n * d + 1] + axeDim + 1;
-            stride = 1;
-
-            shapeOf[d] = (last - first + stride - 1) / stride;      // ceil (last - first) / stride;
-            offset += first * stridesOf[d];
-
-            if(shapeOf[d] != 1)
-                stridesOf[d] *= stride;
-        }
-
-        subArrLen *= shapeOf[d];
-    }
-
-    // check if there is possibility to set ews = 1
-    //shape::setEws(newShape, subArrLen);
-
-    //makeBothBuffersActual();
-    outBuffer = (void*)((int8_t*)vZ + offset * DataTypeUtils::sizeOfElement(zType));
-    cudaError_t err = cudaMalloc(&outShape, shapeSize);
-    if (err != 0) {
-        printf("Cannot allocate memory with error %d\n", err);
-        throw std::runtime_error("Cannot allocate memory for shape");
-    }
-    cudaMemcpy(outShape, newShape, shapeSize, cudaMemcpyHostToDevice);
-    delete [] newShape;
-}
-
-/**
-  * Concatneate multi array of the same shape together
-  * along a particular dimension
-  */
-void concat(
-        Nd4jPointer *extraPointers,
-        int dimension,
-        int numArrays,
-        Nd4jPointer *data, Nd4jPointer *inputShapeInfo,
-        Nd4jPointer *ddata, Nd4jPointer *dinputShapeInfo,
-        void *hZ, Nd4jLong *hZShapeInfo,
-        void *dZ, Nd4jLong *dZShapeInfo,
-        Nd4jPointer *tadPointers, Nd4jPointer *offsetPointers) {
-
-    auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-
-    auto hXShapeInfo = hZShapeInfo;
-    auto hShapePointers = reinterpret_cast<Nd4jLong **>(inputShapeInfo);
-    auto dShapePointers = reinterpret_cast<Nd4jLong **>(dinputShapeInfo);
-    // numArrays will be used as number of TADs, so each block process 1 input
-    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
-    auto axis = dimension;
-
-    const int rank  = shape::rank(hZShapeInfo); //reinterpret_cast<Nd4jLong*>(inputShapeInfo[0]));
-    const int rank2 = 2 * rank;
-    std::vector<std::vector<Nd4jLong>> indices(numArrays, std::vector<Nd4jLong>(rank2,0));
-
-    // take into account indices for first array
-    auto axisSize = shape::sizeAt(reinterpret_cast<Nd4jLong*>(inputShapeInfo[0]), axis);
-    indices[0][2 * axis + 1] = axisSize;
-
-    for(int i = 1; i < numArrays; ++i) {
-        indices[i][2 * axis]     = indices[i-1][2 * axis + 1];                                // index start from
-        indices[i][2 * axis + 1] = indices[i-1][2 * axis + 1] + shape::sizeAt(reinterpret_cast<Nd4jLong*>(inputShapeInfo[i]), axis);      // index end with (excluding)
-    }
-
-    std::vector<void*> outSubArrsBuffs(numArrays);
-    std::vector<Nd4jLong*> outSubArrsShapes(numArrays);
-    for(int i = 0; i < numArrays; ++i) {
-        specialBufferAndShapeWithOffset(dZ, hZShapeInfo, dZShapeInfo, indices[i], outSubArrsBuffs[i], outSubArrsShapes[i]);
-    }
-
-    LaunchContext context(stream);
-    PointersManager manager(&context, "concat");
-    void* dOutBuffers	= manager.replicatePointer(outSubArrsBuffs.data(),  outSubArrsBuffs.size() * sizeof(void*));
-    void* dInBuffers	= manager.replicatePointer(ddata,    numArrays * sizeof(void*));
-    void* dInShapeInfo  = manager.replicatePointer(dShapePointers,  numArrays * sizeof(Nd4jLong*));
-    void* dOutShapeInfo = manager.replicatePointer(outSubArrsShapes.data(), outSubArrsShapes.size() * sizeof(Nd4jLong*));
-
-    BUILD_SINGLE_SELECTOR(zType, concatCudaLauncher, (numArrays, stream, dInBuffers, dInShapeInfo, dOutBuffers, dOutShapeInfo), LIBND4J_TYPES);
-    manager.synchronize();
-
-    cudaError_t err;
-    for(int i = 0; i < numArrays; ++i) {
-        err = cudaFree(outSubArrsShapes[i]);
-        if (err != 0) {
-            printf("Error %d occured when shape %i was deallocating.\n", err, i);
-            throw std::runtime_error("Cannot deallocate memory for shapes.");
-        }
-    }
-}
-
-/**
-  * Concatneate multi array of the same shape together
-  * along a particular dimension
-  */
-// void concat(
-//		Nd4jPointer *extraPointers,
-//        int dimension,
-//        int numArrays,
-//        Nd4jPointer *data, Nd4jPointer *inputShapeInfo,
-//		Nd4jPointer *ddata, Nd4jPointer *dinputShapeInfo,
-//		void *hZ, Nd4jLong *hZShapeInfo,
-//        void *dZ, Nd4jLong *dZShapeInfo,
-//		Nd4jPointer *tadPointers, Nd4jPointer *offsetPointers) {
-//
-//	cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-//	auto hXShapeInfo = hZShapeInfo;
-//	auto hShapePointers = reinterpret_cast<Nd4jLong **>(inputShapeInfo);
-//	// numArrays will be used as number of TADs, so each block process 1 input
-//
-//	int smem = 8192;
-//	bool isVstack = false;
-//	bool isScalar = true;
-//	bool isHstack = false;
-//
-//	for (int i = 0; i < numArrays; i++) {
-//		if (!shape::isScalar(hShapePointers[i])) {
-//			isScalar = false;
-//			break;
-//		}
-//	}
-//
-//	if (!isScalar && dimension == 0 && shape::rank(hZShapeInfo) == 2 && shape::order(hZShapeInfo) == 'c' ) {
-//		isVstack = true;
-//        for (int i = 0; i < numArrays; i++) {
-//			if (!shape::isVector(hShapePointers[i]) || shape::elementWiseStride(hShapePointers[i]) <= 0 ||
-//				shape::order(hShapePointers[i]) != 'c') {
-//				isVstack = false;
-//				break;
-//			}
-//		}
-//	}
-//
-//    // let's try to fit N-dimensional vstack
-//    if (!isVstack && !isScalar && dimension == 0 && shape::order(hXShapeInfo) == 'c') {
-//		auto length0 = shape::length(hShapePointers[0]);
-//        isVstack = true;
-//        for (int i = 0; i < numArrays; i++) {
-//            if (shape::elementWiseStride(hShapePointers[i]) <= 0 || shape::order(hShapePointers[i]) != 'c' || length0 != shape::length(hShapePointers[i])) {
-//                isVstack = false;
-//                break;
-//            }
-//        }
-//    }
-//
-//	if (!isScalar && !isVstack && dimension == 1 && shape::isVector(hZShapeInfo)) {
-//		isHstack = true;
-//		for (int i = 0; i < numArrays; i++) {
-//			if (!shape::isVector(hShapePointers[i]) || shape::elementWiseStride(hShapePointers[i]) <= 0) {
-//				isHstack = false;
-//				break;
-//			}
-//		}
-//	}
-//
-//	if (isScalar) {
-//		if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-//			printf("Going scalar concat\n");
-//
-//		dim3 launchDims(128, 128, 16384);
-//		auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
-//		BUILD_SINGLE_SELECTOR(zType, concatKernelScalarGeneric, (launchDims, stream, numArrays, reinterpret_cast<Nd4jPointer *>(ddata[0]), dZ), LIBND4J_TYPES);
-//
-//	} else if (isVstack) {
-//		if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-//			printf("Going VStack concat\n");
-//
-//		dim3 launchDims(128, 512, 16384);
-//		auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
-//		BUILD_SINGLE_SELECTOR(zType, concatKernelVStackGeneric, (launchDims, stream, numArrays, reinterpret_cast<Nd4jPointer *>(ddata[0]), reinterpret_cast<Nd4jPointer *>(dinputShapeInfo[0]), dZ, dZShapeInfo), LIBND4J_TYPES);
-//
-//	} else if (isHstack) {
-//		if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-//			printf("Going HStack concat\n");
-//
-//		dim3 launchDims(128, 128, 16384);
-//		auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
-//		BUILD_SINGLE_SELECTOR(zType, concatKernelHStackGeneric, (launchDims, stream, numArrays, reinterpret_cast<Nd4jPointer *>(ddata[0]), reinterpret_cast<Nd4jPointer *>(dinputShapeInfo[0]), dZ, dZShapeInfo), LIBND4J_TYPES);
-//	} else {
-//		if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-//			printf("Going generic concat\n");
-//
-//        auto devZTadShape = reinterpret_cast<Nd4jLong *>(extraPointers[10]);
-//		auto devZOffsets = reinterpret_cast<Nd4jLong *>(extraPointers[11]);
-//
-//		dim3 launchDims(128, 128, 8192);
-//		auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
-//		BUILD_SINGLE_SELECTOR(zType, concatKernelGeneric, (launchDims, stream, numArrays, reinterpret_cast<Nd4jPointer *>(ddata[0]), reinterpret_cast<Nd4jPointer *>(dinputShapeInfo[0]), dZ, dZShapeInfo,  reinterpret_cast<Nd4jPointer *>(tadPointers[0]), reinterpret_cast<Nd4jPointer *>(offsetPointers[0]), devZTadShape, devZOffsets), LIBND4J_TYPES);
-//	}
-//	if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-//		printf("sharedMemory requested for concatFloat: [%i], registers: [%i]\n", smem, funcAttributes[31].numRegs);
-//
-//    cudaError_t res = cudaStreamSynchronize(*stream);
-//    checkCudaErrors(res);
-//    nd4j::DebugHelper::checkErrorCode(stream, "Legacy ConcatFloat(...) failed");
-//}
-
-
-
 void specialConcat(
         Nd4jPointer *extraPointers,
         int dimension,
@@ -1432,8 +1255,14 @@ void specialConcat(
         Nd4jPointer *inputShapeInfo,
         void *dZ,
         Nd4jLong *dZShapeInfo, Nd4jPointer *tadPointers, Nd4jPointer *offsetPointers) {
-
-    BUILD_SINGLE_SELECTOR(ArrayOptions::dataType(dZShapeInfo), nd4j::SpecialMethods ,::concatCpuGeneric(dimension, numArrays, data, inputShapeInfo, dZ, dZShapeInfo), LIBND4J_TYPES);
+    try {
+        BUILD_SINGLE_SELECTOR(ArrayOptions::dataType(dZShapeInfo), nd4j::SpecialMethods,
+                              ::concatCpuGeneric(dimension, numArrays, data, inputShapeInfo, dZ, dZShapeInfo),
+                              LIBND4J_TYPES);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -1441,9 +1270,15 @@ void specialConcat(
  * This method saves
  */
 nd4j::TadPack* tadOnlyShapeInfo(Nd4jLong *dXShapeInfo, int *dimension, int dimensionLength) {
-	auto pack = new TadPack();
-	*pack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(dXShapeInfo, dimension, dimensionLength);
-    return pack;
+    try {
+        auto pack = new TadPack();
+        *pack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(dXShapeInfo, dimension, dimensionLength);
+        return pack;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 Nd4jLong* getPrimaryShapeInfo(nd4j::TadPack* pack) {
@@ -1489,11 +1324,11 @@ int memcpyConstantAsync(Nd4jLong dst, Nd4jPointer src, Nd4jLong size, int flags,
 		}
 			break;
 	}
-	//cudaError_t dZ = cudaMemcpyAsync((void *) dst, (const void *) src, (size_t) size, kind, *pStream);
-	cudaError_t dZ = cudaMemcpyToSymbolAsync(deviceConstantMemory, const_cast<const void *>(src), size, dst, kind, *pStream);
-	checkCudaErrors(dZ);
-	if (dZ != 0)
-        throw std::runtime_error("cudaMemcpyToSymbolAsync(...) failed");
+	auto dZ = cudaMemcpyToSymbolAsync(deviceConstantMemory, const_cast<const void *>(src), size, dst, kind, *pStream);
+	if (dZ != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaMemcpyToSymbolAsync failed");
+	}
 
 	return 1;
 }
@@ -1502,8 +1337,10 @@ Nd4jPointer getConstantSpace() {
 	Nd4jPointer dConstAddr;
 	cudaError_t dZ = cudaGetSymbolAddress(reinterpret_cast<void **>(&dConstAddr), deviceConstantMemory);
 
-	if (dZ != 0)
-        throw std::runtime_error("cudaGetSymbolAddress(...) failed");
+	if (dZ != 0) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(dZ);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage("cudaGetSymbolAddress failed");
+	}
 
 	return dConstAddr;
 }
@@ -1519,13 +1356,19 @@ void pullRows(Nd4jPointer *extraPointers,
 						 Nd4jLong *tadOffsets,
 						 Nd4jLong *zTadShapeInfo,
 						 Nd4jLong *zTadOffsets) {
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        dim3 launchDims(64, 256, 1024);
+        auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
+        BUILD_SINGLE_SELECTOR(xType, pullRowsKernelGeneric,
+                              (launchDims, stream, dX, dZ, n, indexes, tadShapeInfo, tadOffsets, zTadShapeInfo, zTadOffsets),
+                              LIBND4J_TYPES);
 
-	cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-	dim3 launchDims(64, 256, 1024);
-	auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
-    BUILD_SINGLE_SELECTOR(xType, pullRowsKernelGeneric, (launchDims, stream, dX, dZ, n, indexes, tadShapeInfo, tadOffsets,  zTadShapeInfo,  zTadOffsets), LIBND4J_TYPES);
-
-	DEBUG_KERNEL(stream, -1);
+        DEBUG_KERNEL(stream, -1);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -1537,25 +1380,31 @@ void average(Nd4jPointer *extras,
 						int n,
 						Nd4jLong length,
 						bool propagate) {
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extras[1]);
+        int mode = getDeviceId(extras[3]);
 
-	cudaStream_t * stream = reinterpret_cast<cudaStream_t *>(extras[1]);
-	int mode = getDeviceId(extras[3]);
+        auto dX = reinterpret_cast<void **>(dx);
 
-	auto dX = reinterpret_cast<void **>(dx);
+        if (nd4j::Environment::getInstance()->isDebugAndVerbose())
+            printf("averageFloat called\n");
 
-	if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-		printf("averageFloat called\n");
-
-	auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
-	// launching on gpu
-	if (mode == 0) {
-		dim3 launchDims(256, 256, 4096);
-    	BUILD_SINGLE_SELECTOR(xType, averagingKernelGeneric, (launchDims, stream, dX, dz, n, length, propagate), LIBND4J_TYPES);
-        nd4j::DebugHelper::checkErrorCode(stream, "AverageFloat(...) failed");
-	} else {
-		// launching on host memory
-        BUILD_SINGLE_SELECTOR(xType, nd4j::SpecialMethods, ::averageGeneric(x, z, zShapeInfo, n, length, propagate), LIBND4J_TYPES);
-	}
+        auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
+        // launching on gpu
+        if (mode == 0) {
+            dim3 launchDims(256, 256, 4096);
+            BUILD_SINGLE_SELECTOR(xType, averagingKernelGeneric, (launchDims, stream, dX, dz, n, length, propagate),
+                                  LIBND4J_TYPES);
+            nd4j::DebugHelper::checkErrorCode(stream, "AverageFloat(...) failed");
+        } else {
+            // launching on host memory
+            BUILD_SINGLE_SELECTOR(xType, nd4j::SpecialMethods, ::averageGeneric(x, z, zShapeInfo, n, length, propagate),
+                                  LIBND4J_TYPES);
+        }
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void accumulate(Nd4jPointer *extras,
@@ -1565,25 +1414,31 @@ void accumulate(Nd4jPointer *extras,
 						   void *dz, Nd4jLong *dzShapeInfo,
 						   int n,
 						   Nd4jLong length) {
+    try {
+        auto stream = reinterpret_cast<cudaStream_t *>(extras[1]);
+        int mode = getDeviceId(extras[3]);
 
-	auto stream = reinterpret_cast<cudaStream_t *>(extras[1]);
-	int mode = getDeviceId(extras[3]);
+        auto dX = reinterpret_cast<void **>(dx);
 
-	auto dX = reinterpret_cast<void **>(dx);
+        if (nd4j::Environment::getInstance()->isDebugAndVerbose())
+            printf("accumulateFloat called\n");
+        auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
 
-	if (nd4j::Environment::getInstance()->isDebugAndVerbose())
-		printf("accumulateFloat called\n");
-	auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
-
-	// launching on gpu
-	if (mode == 0) {
-		dim3 launchDims(n, 256, 16384);
-        BUILD_SINGLE_SELECTOR(xType, accumulateKernelGeneric, (launchDims, stream, dX, dz, n,length), LIBND4J_TYPES);
-        nd4j::DebugHelper::checkErrorCode(stream, "AccumulateFloat(...) failed");
-	} else {
-		// launching on host memory
-        BUILD_SINGLE_SELECTOR(xType, nd4j::SpecialMethods, ::accumulateGeneric(x, z, zShapeInfo, n, length), LIBND4J_TYPES);
-	}
+        // launching on gpu
+        if (mode == 0) {
+            dim3 launchDims(n, 256, 16384);
+            BUILD_SINGLE_SELECTOR(xType, accumulateKernelGeneric, (launchDims, stream, dX, dz, n, length),
+                                  LIBND4J_TYPES);
+            nd4j::DebugHelper::checkErrorCode(stream, "AccumulateFloat(...) failed");
+        } else {
+            // launching on host memory
+            BUILD_SINGLE_SELECTOR(xType, nd4j::SpecialMethods, ::accumulateGeneric(x, z, zShapeInfo, n, length),
+                                  LIBND4J_TYPES);
+        }
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -1596,50 +1451,29 @@ void shuffle(Nd4jPointer *extras,
 						int *shuffleMap,
 						Nd4jPointer *tadShapeInfo,
 						Nd4jPointer *tadOffsets) {
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extras[1]);
 
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extras[1]);
+        auto dX = reinterpret_cast<void **>(dx);
+        auto dZ = reinterpret_cast<void **>(dz);
+        auto xShape = reinterpret_cast<Nd4jLong **>(xShapeInfo);
+        auto dxShape = reinterpret_cast<Nd4jLong **>(dXShapeInfo);
+        auto tadOnlyShapeInfo = reinterpret_cast<Nd4jLong **>(tadShapeInfo);
+        auto tadOffset = reinterpret_cast<Nd4jLong **>(tadOffsets);
 
-    auto dX = reinterpret_cast<void **>(dx);
-    auto dZ = reinterpret_cast<void **>(dz);
-    auto xShape = reinterpret_cast<Nd4jLong **>(xShapeInfo);
-    auto dxShape = reinterpret_cast<Nd4jLong **>(dXShapeInfo);
-    auto tadOnlyShapeInfo = reinterpret_cast<Nd4jLong **>(tadShapeInfo);
-    auto tadOffset = reinterpret_cast<Nd4jLong **>(tadOffsets);
+        auto xType = nd4j::ArrayOptions::dataType(xShape[0]);
+        dim3 launchDims(256, 512, 8192);
+        BUILD_SINGLE_SELECTOR(xType, shuffleKernelGeneric,
+                              (launchDims, stream, dX, dxShape, dZ, N, shuffleMap, tadOnlyShapeInfo, tadOffset),
+                              LIBND4J_TYPES);
 
-    auto xType = nd4j::ArrayOptions::dataType(xShape[0]);
-    dim3 launchDims(256, 512, 8192);
-    BUILD_SINGLE_SELECTOR(xType, shuffleKernelGeneric, (launchDims, stream, dX, dxShape, dZ, N, shuffleMap,  tadOnlyShapeInfo, tadOffset), LIBND4J_TYPES);
-
-    nd4j::DebugHelper::checkErrorCode(stream, "shuffle(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "shuffle(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
-/*
-void execMetaPredicateShape(Nd4jPointer *extras,
-	                                  const int opTypeA,
-	                                  const int opNumA,
-	                                  const int opTypeB,
-	                                  const int opNumB,
-	                                  Nd4jLong N,
-	                                  void *hX, Nd4jLong *hXShapeInfo,
-                                      void *dX, Nd4jLong *dXShapeInfo,
-                                      void *hY, Nd4jLong *hYShapeInfo,
-                                      void *dY, Nd4jLong *dYShapeInfo,
-                                      void *hZ, Nd4jLong *hZShapeInfo,
-                                      void *dZ, Nd4jLong *dZShapeInfo,
-	                                  void *extraA,
-	                                  void *extraB,
-	                                  double scalarA,
-	                                  double scalarB) {
-
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extras[1]);
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
-    BUILD_SINGLE_SELECTOR(xType, functions::grid::GRIDShaped, ::execMetaPredicateShaped(stream, extras, opTypeA, opNumA, opTypeB, opNumB, N, dX, dXShapeInfo, dY, dYShapeInfo, dZ, dZShapeInfo, extraA, extraB, scalarA, scalarB), LIBND4J_TYPES);
-    // functions::grid::GRIDShaped<float>::execMetaPredicateShaped(stream, extras, opTypeA, opNumA, opTypeB, opNumB, N, dX, dXShapeInfo, dy, dYShapeInfo, dz, zShapeInfo, extraA, extraB, scalarA, scalarB);
-
-	DEBUG_KERNEL(stream, opNumA);
-}
-*/
-
 bool isExperimentalEnabled() {
     return nd4j::Environment::getInstance()->isExperimentalBuild();
 }
@@ -1670,9 +1504,14 @@ void execSummaryStats(Nd4jPointer *extraPointers,
                                  void *hZ, Nd4jLong *hZShapeInfo,
                                  void *dZ, Nd4jLong *dZShapeInfo,
                                  bool biasCorrected) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execSummaryStats(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo, biasCorrected);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execSummaryStats(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ,
+                                              hZShapeInfo, dZ, dZShapeInfo, biasCorrected);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1686,11 +1525,18 @@ void execSummaryStatsTad(Nd4jPointer *extraPointers,
 								 void *hDimension, Nd4jLong *hDimensionShape, void *dDimension, Nd4jLong *dDimensionShape,
                                  bool biasCorrected,
 								 Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execSummaryStats(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, biasCorrected);
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execSummaryStats(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ,
+                                              hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadShapeInfo,
+                                              tadOffsets, biasCorrected);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1703,9 +1549,14 @@ void execReduce3(Nd4jPointer *extraPointers,
                             void *dY, Nd4jLong *dYShapeInfo,
                             void *hZ, Nd4jLong *hZShapeInfo,
                             void *dZ, Nd4jLong *dZShapeInfo) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execReduce3(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execReduce3(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo, dY,
+                                         dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1721,35 +1572,35 @@ void execReduce3Tad(Nd4jPointer *extraPointers,
 							void *hDimension, Nd4jLong *hDimensionShape, void *dDimension, Nd4jLong *dDimensionShape,
                             Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets,
                             Nd4jLong *yTadOnlyShapeInfo, Nd4jLong *yTadOffsets) {
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-	// if (extraPointers == nullptr || extraPointers[2] == 0)
- //        NativeOpExecutioner::execReduce3(nullptr, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, yTadOnlyShapeInfo, yTadOffsets);
- //    else {
- //        // going tad-ways
- //        auto tadShapeInfo = reinterpret_cast<Nd4jLong *> (extraPointers[0]);
- //        auto tadOffsets = reinterpret_cast<Nd4jLong *>(extraPointers[1]);
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo,
+                                                                                reinterpret_cast<int *>(hDimension),
+                                                                                shape::length(hDimensionShape));
+        auto tadLength = shape::length(tadPack.primaryShapeInfo());
+        auto yLength = shape::length(hYShapeInfo);
+        auto xLength = shape::length(hXShapeInfo);
 
- //        NativeOpExecutioner::execReduce3TAD(nullptr, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets);
- //    }
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
 
-    // nd4j_printf("Starting...\n","");
+        if (tadLength == yLength || tadLength == xLength) {
+            // nd4j_printf("== way\n","");
+            NativeOpExecutioner::execReduce3(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo,
+                                             dY,
+                                             dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength,
+                                             tadOnlyShapeInfo, tadOffsets, yTadOnlyShapeInfo, yTadOffsets);
+        } else
+            NativeOpExecutioner::execReduce3TAD(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY,
+                                                hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo,
+                                                dimension, dimensionLength, tadOnlyShapeInfo, yTadOffsets,
+                                                yTadOnlyShapeInfo, yTadOffsets);
 
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(hXShapeInfo, reinterpret_cast<int*>(hDimension), shape::length(hDimensionShape));
-    auto tadLength = shape::length(tadPack.primaryShapeInfo());
-    auto yLength = shape::length(hYShapeInfo);
-    auto xLength = shape::length(hXShapeInfo);
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-
-    if (tadLength == yLength || tadLength == xLength) {
-        // nd4j_printf("== way\n","");
-        NativeOpExecutioner::execReduce3(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo, dY,
-                                         dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength,
-                                         tadOnlyShapeInfo, tadOffsets, yTadOnlyShapeInfo, yTadOffsets);
-    } else
-        NativeOpExecutioner::execReduce3TAD(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, yTadOffsets, yTadOnlyShapeInfo, yTadOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1761,9 +1612,14 @@ void execReduce3Scalar(Nd4jPointer *extraPointers,int opNum,
                                   void *dY, Nd4jLong *dYShapeInfo,
                                   void *hZ, Nd4jLong *hZShapeInfo,
                                   void *dZ, Nd4jLong *dZShapeInfo) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execReduce3Scalar(&lc, opNum,hX,hXShapeInfo,dX, dXShapeInfo,extraParams,hY,hYShapeInfo,dY,dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execReduce3Scalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hY,
+                                               hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1776,9 +1632,15 @@ void execScalarBool(Nd4jPointer *extraPointers,
 							void *hScalar, Nd4jLong *hScalarShapeInfo,
 							void *dScalar, Nd4jLong *dScalarShapeInfo,
 							void *extraParams) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execScalarBool(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, hScalar, hScalarShapeInfo, dScalar, dScalarShapeInfo, extraParams);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execScalarBool(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ,
+                                            dZShapeInfo, hScalar, hScalarShapeInfo, dScalar, dScalarShapeInfo,
+                                            extraParams);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1794,11 +1656,19 @@ void execScalarBoolTad(Nd4jPointer *extraPointers,
 							   void *hDimension, Nd4jLong *hDimensionShape, void *dDimension, Nd4jLong *dDimensionShape,
                            Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
                            Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execScalarBool(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo, hScalars, hScalarShapeInfo, dScalars, dScalarShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ, tadOffsetsZ);
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execScalarBool(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo,
+                                            dZ, dZShapeInfo, hScalars, hScalarShapeInfo, dScalars, dScalarShapeInfo,
+                                            dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ,
+                                            tadOffsetsZ);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1811,9 +1681,14 @@ void execScalar(Nd4jPointer *extraPointers,
 						void *hScalar, Nd4jLong *hScalarShapeInfo,
 						void *dScalar, Nd4jLong *dScalarShapeInfo,
 						void *extraParams) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-	NativeOpExecutioner::execScalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, hScalar, hScalarShapeInfo, dScalar, dScalarShapeInfo, extraParams);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execScalar(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo,
+                                        hScalar, hScalarShapeInfo, dScalar, dScalarShapeInfo, extraParams);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1830,27 +1705,36 @@ void execScalarTad(Nd4jPointer *extraPointers,
 						   void *dDimension, Nd4jLong *dDimensionShape,
                      Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
                      Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
 
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
-    auto yType = nd4j::ArrayOptions::dataType(hScalarShapeInfo);
-    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+        auto yType = nd4j::ArrayOptions::dataType(hScalarShapeInfo);
+        auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
 
-	if (yType != xType && yType != nd4j::DataType::BOOL && !isExperimentalEnabled())
-		throw nd4j::datatype_exception::build("execScalar both operands must have same data type", xType, yType);
+        if (yType != xType && yType != nd4j::DataType::BOOL && !isExperimentalEnabled())
+            throw nd4j::datatype_exception::build("execScalar both operands must have same data type", xType, yType);
 
-	dim3 launchDims(256, 256, 16384);
+        dim3 launchDims(256, 256, 16384);
 
 #ifdef __ND4J_EXPERIMENTAL__
-    BUILD_PAIRWISE_SELECTOR(xType, yType, zType, functions::scalar::ScalarTransform, ::executeCudaAlongDimension(launchDims, stream, opNum, dX, dXShapeInfo, dZ, dZShapeInfo, dScalars, extraParams, dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ, tadOffsetsZ), LIBND4J_TYPES, LIBND4J_TYPES);
+        BUILD_PAIRWISE_SELECTOR(xType, yType, zType, functions::scalar::ScalarTransform, ::executeCudaAlongDimension(launchDims, stream, opNum, dX, dXShapeInfo, dZ, dZShapeInfo, dScalars, extraParams, dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ, tadOffsetsZ), LIBND4J_TYPES, LIBND4J_TYPES);
 #else
-	BUILD_SINGLE_SELECTOR_THRICE(xType, functions::scalar::ScalarTransform, ::executeCudaAlongDimension(launchDims, stream, opNum, dX, dXShapeInfo, dZ, dZShapeInfo, dScalars, extraParams, dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ, tadOffsetsZ), LIBND4J_TYPES);
+        BUILD_SINGLE_SELECTOR_THRICE(xType, functions::scalar::ScalarTransform,
+                                     ::executeCudaAlongDimension(launchDims, stream, opNum, dX, dXShapeInfo, dZ,
+                                                                 dZShapeInfo, dScalars, extraParams, dimension,
+                                                                 dimensionLength, tadShapeInfo, tadOffsets,
+                                                                 tadShapeInfoZ, tadOffsetsZ), LIBND4J_TYPES);
 #endif
 
-	DEBUG_KERNEL(stream, opNum);
+        DEBUG_KERNEL(stream, opNum);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void execAggregate(Nd4jPointer *extraPointers,
@@ -1866,16 +1750,23 @@ void execAggregate(Nd4jPointer *extraPointers,
                                    void *realArguments,
                                    int numRealArguments,
                                    nd4j::DataType dtype) {
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        int numBlocks = getDeviceId(extraPointers[2]);
+        int numThreads = getDeviceId(extraPointers[3]);
+        int shmem = getDeviceId(extraPointers[4]);
 
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-    int numBlocks = getDeviceId(extraPointers[2]);
-    int numThreads = getDeviceId(extraPointers[3]);
-    int shmem = getDeviceId(extraPointers[4]);
+        dim3 launchDims = dim3(numBlocks, numThreads, shmem);
 
-    dim3 launchDims = dim3(numBlocks, numThreads, shmem);
-
-    BUILD_SINGLE_SELECTOR(dtype, functions::aggregate::AggregatedFunction, ::aggregateKernelGeneric(launchDims, stream, opNum, arguments, numArguments, shapes, numShapes, indexArguments, numIndexArguments, intArrays, numIntArrays, realArguments, numRealArguments), FLOAT_TYPES);
-    nd4j::DebugHelper::checkErrorCode(stream, "execAggregateFloat(...) failed");
+        BUILD_SINGLE_SELECTOR(dtype, functions::aggregate::AggregatedFunction,
+                              ::aggregateKernelGeneric(launchDims, stream, opNum, arguments, numArguments, shapes,
+                                                       numShapes, indexArguments, numIndexArguments, intArrays,
+                                                       numIntArrays, realArguments, numRealArguments), FLOAT_TYPES);
+        nd4j::DebugHelper::checkErrorCode(stream, "execAggregateFloat(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void batchExecutor(Nd4jPointer *extraPointers,
@@ -1897,17 +1788,25 @@ void execAggregateBatch(Nd4jPointer *extraPointers,
 									int maxIntArrays, int maxIntArraySize,
 									int maxIdx, int maxReals,
 									void *ptrToArguments, nd4j::DataType dtype) {
-    // not implemented yet
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-    int numBlocks = getDeviceId(extraPointers[2]);
-    int numThreads = getDeviceId(extraPointers[3]);
-    int shmem = getDeviceId(extraPointers[4]);
+    try {
+        // not implemented yet
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        int numBlocks = getDeviceId(extraPointers[2]);
+        int numThreads = getDeviceId(extraPointers[3]);
+        int shmem = getDeviceId(extraPointers[4]);
 
-    dim3 launchDims = dim3(numAggregates, numThreads, shmem);
+        dim3 launchDims = dim3(numAggregates, numThreads, shmem);
 
-	BUILD_SINGLE_SELECTOR(dtype, functions::aggregate::AggregatedFunction, ::aggregateBatchKernelGeneric(launchDims, stream, opNum, numAggregates, maxArgs, maxShapes, maxIntArrays, maxIntArraySize, maxIdx, maxReals, ptrToArguments), FLOAT_TYPES);
+        BUILD_SINGLE_SELECTOR(dtype, functions::aggregate::AggregatedFunction,
+                              ::aggregateBatchKernelGeneric(launchDims, stream, opNum, numAggregates, maxArgs,
+                                                            maxShapes, maxIntArrays, maxIntArraySize, maxIdx, maxReals,
+                                                            ptrToArguments), FLOAT_TYPES);
 
-	DEBUG_KERNEL(stream, opNum);
+        DEBUG_KERNEL(stream, opNum);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1917,9 +1816,13 @@ void execRandom(Nd4jPointer *extraPointers,
                           void *hZ, Nd4jLong *hZShapeInfo,
                           void *dZ, Nd4jLong *dZShapeInfo,
                           void *extraArguments) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execRandom(&lc, opNum, stateHost, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execRandom(&lc, opNum, stateHost, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1929,9 +1832,14 @@ void execRandom2(Nd4jPointer *extraPointers, int opNum, Nd4jPointer stateHost,
 						   void *hZ, Nd4jLong *hZShapeInfo,
 						   void *dZ, Nd4jLong *dZShapeInfo,
 						   void *extraArguments) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execRandom(&lc, opNum, stateHost, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execRandom(&lc, opNum, stateHost, hX, hXShapeInfo, dX, dXShapeInfo, hZ, hZShapeInfo, dZ,
+                                        dZShapeInfo, extraArguments);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -1943,9 +1851,14 @@ void execRandom3(Nd4jPointer *extraPointers, int opNum, Nd4jPointer stateHost,
 							void *hZ, Nd4jLong *hZShapeInfo,
 							void *dZ, Nd4jLong *dZShapeInfo,
 							void *extraArguments) {
-
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execRandom(&lc, opNum, stateHost, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execRandom(&lc, opNum, stateHost, hX, hXShapeInfo, dX, dXShapeInfo, hY, hYShapeInfo, dY,
+                                        dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, extraArguments);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -2053,13 +1966,19 @@ void tear(Nd4jPointer *extras,
 					 Nd4jLong *zShapeInfo,
 					 Nd4jLong *tadShapeInfo,
 					 Nd4jLong *tadOffsets) {
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extras[1]);
+        dim3 launchDims(512, 512, 512);
+        auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
+        BUILD_SINGLE_SELECTOR(xType, tearKernelGeneric,
+                              (launchDims, stream, dX, dXShapeInfo, targets, zShapeInfo, tadShapeInfo, tadOffsets),
+                              LIBND4J_TYPES);
 
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extras[1]);
-    dim3 launchDims(512, 512, 512);
-    auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
-    BUILD_SINGLE_SELECTOR(xType, tearKernelGeneric, (launchDims, stream, dX, dXShapeInfo, targets, zShapeInfo, tadShapeInfo, tadOffsets), LIBND4J_TYPES);
-
-    nd4j::DebugHelper::checkErrorCode(stream, "tearFloat(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "tearFloat(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -2146,56 +2065,72 @@ void prescanArrayRecursive(Nd4jPointer *extras, int *dZ, int *dX, int numElement
 
 
 void encodeThresholdP1(Nd4jPointer *extras, void *dx, Nd4jLong *hXShapeInfo, Nd4jLong N, int *dz, float threshold) {
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extras[1]);
 
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extras[1]);
+        int blockSize = 1024;
+        int numBlocks = N / blockSize + (N % blockSize ? 1 : 0);
 
-    int blockSize = 1024;
-    int numBlocks = N / blockSize + (N % blockSize ? 1 : 0);
+        dim3 launchDims(numBlocks, blockSize, 1024);
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+        BUILD_SINGLE_SELECTOR(xType, encoderKernelP1Generic, (launchDims, stream, dx, N, dz, threshold), LIBND4J_TYPES);
 
-    dim3 launchDims(numBlocks, blockSize, 1024);
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
-    BUILD_SINGLE_SELECTOR(xType, encoderKernelP1Generic, (launchDims, stream, dx, N, dz, threshold), LIBND4J_TYPES);
-
-    nd4j::DebugHelper::checkErrorCode(stream, "encodeThresholdP1Float(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "encodeThresholdP1Float(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
 
 void encodeThresholdP2Int(Nd4jPointer *extraPointers, int *dx, Nd4jLong N, int *dz) {
-
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-    //encoderKernelP2Float<<<numBlocks, blockSize , 1024 * sizeof(float), *stream>>>(dx, N, dz);
-    prescanArrayRecursive(extraPointers, dz, dx + 1, (int) N, 0);
-    nd4j::DebugHelper::checkErrorCode(stream, "encodeThresholdP2Int(...) failed");
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        //encoderKernelP2Float<<<numBlocks, blockSize , 1024 * sizeof(float), *stream>>>(dx, N, dz);
+        prescanArrayRecursive(extraPointers, dz, dx + 1, (int) N, 0);
+        nd4j::DebugHelper::checkErrorCode(stream, "encodeThresholdP2Int(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void encodeThresholdP3(Nd4jPointer *extraPointers, void *dx, Nd4jLong *hXShapeInfo, int *offsets, Nd4jLong N, int *dz){
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
 
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        int blockSize = 1024;
+        int numBlocks = N / blockSize + (N % blockSize ? 1 : 0);
 
-    int blockSize = 1024;
-    int numBlocks = N / blockSize + (N % blockSize ? 1 : 0);
+        dim3 launchDims(numBlocks, blockSize, 4096);
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+        BUILD_SINGLE_SELECTOR(xType, encoderKernelP3Generic, (launchDims, stream, dx, offsets, N, dz), LIBND4J_TYPES);
 
-    dim3 launchDims(numBlocks, blockSize, 4096);
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
-    BUILD_SINGLE_SELECTOR(xType, encoderKernelP3Generic, (launchDims, stream, dx, offsets, N, dz), LIBND4J_TYPES);
-
-    nd4j::DebugHelper::checkErrorCode(stream, "encodeThresholdP3Float(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "encodeThresholdP3Float(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void decodeThreshold(Nd4jPointer *extraPointers, void *dx, Nd4jLong N, void *dz, Nd4jLong *zShapeInfo){
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
 
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        // we probably want to have smaller blocks here, memory writes are misaligned anyway
+        int blockSize = 128;
+        int numBlocks = N / blockSize + (N % blockSize ? 1 : 0);
 
-    // we probably want to have smaller blocks here, memory writes are misaligned anyway
-    int blockSize = 128;
-    int numBlocks = N / blockSize + (N % blockSize ? 1 : 0);
+        dim3 launchDims(numBlocks, blockSize, 1024);
+        auto zType = nd4j::ArrayOptions::dataType(zShapeInfo);
+        BUILD_SINGLE_SELECTOR(zType, decoderKernelGeneric, (launchDims, stream, dx, N, dz), LIBND4J_TYPES);
 
-    dim3 launchDims(numBlocks, blockSize, 1024);
-    auto zType = nd4j::ArrayOptions::dataType(zShapeInfo);
-    BUILD_SINGLE_SELECTOR(zType, decoderKernelGeneric, (launchDims, stream, dx, N, dz), LIBND4J_TYPES);
-
-    nd4j::DebugHelper::checkErrorCode(stream, "decodeThresholdFloat(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "decodeThresholdFloat(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -2212,11 +2147,18 @@ void execReduce3All(Nd4jPointer *extraPointers,
 							   		void *dDimension, Nd4jLong *dDimensionShape,
 									Nd4jLong *xTadShapeInfo, Nd4jLong *xOffsets,
 									Nd4jLong *yTadShapeInfo, Nd4jLong *yOffsets) {
-	auto dimension = reinterpret_cast<int *>(dDimension);
-	int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
+    try {
+        auto dimension = reinterpret_cast<int *>(dDimension);
+        int dimensionLength = static_cast<int>(shape::length(hDimensionShape));
 
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    NativeOpExecutioner::execReduce3All(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParamsVals, hY, hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, xTadShapeInfo, xOffsets, yTadShapeInfo, yOffsets);
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        NativeOpExecutioner::execReduce3All(&lc, opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParamsVals, hY,
+                                            hYShapeInfo, dY, dYShapeInfo, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension,
+                                            dimensionLength, xTadShapeInfo, xOffsets, yTadShapeInfo, yOffsets);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -2224,57 +2166,65 @@ void sort(Nd4jPointer *extraPointers,
 					 void *x, Nd4jLong *xShapeInfo,
 					 void *dX, Nd4jLong *dXShapeInfo,
 					 bool descending) {
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
 
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-
-    auto xLength = shape::length(xShapeInfo);
-    auto xEWS = shape::elementWiseStride(xShapeInfo);
-    auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
+        auto xLength = shape::length(xShapeInfo);
+        auto xEWS = shape::elementWiseStride(xShapeInfo);
+        auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
 
 
-    // check if xLength is a power of 2, and use bitonic sort, if that's the case
-    if ((xLength != 0) && ((xLength & (xLength - 1)) == 0) && (xLength <= 1024 * 1024 * 10)) {
-        int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
-        int numBlocks = xLength / numThreads;
-        if (xLength % numThreads > 0 || numBlocks == 0)
-            numBlocks++;
+        // check if xLength is a power of 2, and use bitonic sort, if that's the case
+        if ((xLength != 0) && ((xLength & (xLength - 1)) == 0) && (xLength <= 1024 * 1024 * 10)) {
+            int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
+            int numBlocks = xLength / numThreads;
+            if (xLength % numThreads > 0 || numBlocks == 0)
+                numBlocks++;
 
-        dim3 launchDims(numBlocks, numThreads, 32768);
+            dim3 launchDims(numBlocks, numThreads, 32768);
 
-        for (int k = 2; k <= xLength; k = 2*k) {
-            for (int j = k >> 1; j > 0; j = j >> 1) {
-				BUILD_SINGLE_SELECTOR(xType, bitonicSortStepGeneric, (launchDims, stream, dX, dXShapeInfo, j, k, xLength, descending), LIBND4J_TYPES);
-			}
+            for (int k = 2; k <= xLength; k = 2 * k) {
+                for (int j = k >> 1; j > 0; j = j >> 1) {
+                    BUILD_SINGLE_SELECTOR(xType, bitonicSortStepGeneric,
+                                          (launchDims, stream, dX, dXShapeInfo, j, k, xLength, descending),
+                                          LIBND4J_TYPES);
+                }
+            }
+        } else {
+            int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
+            int numBlocks = xLength / numThreads;
+            if (xLength % numThreads > 0 || numBlocks == 0)
+                numBlocks++;
+
+            numBlocks = nd4j::math::nd4j_min<int>(512, numBlocks);
+            dim3 launchDims(numBlocks, numThreads, 32768);
+
+            int max = 2, dg = 0;
+            while (max < xLength) {
+                max <<= 1;
+                dg++;
+            }
+            max <<= 1;
+
+            for (int window = 2; window < max; window <<= 1) {
+                int n = window;
+                int rev = 0;
+                do {
+                    int half = n >> 1;
+                    BUILD_SINGLE_SELECTOR(xType, bitonicArbitraryStepGeneric,
+                                          (launchDims, stream, dX, dXShapeInfo, n, xLength, rev, descending),
+                                          LIBND4J_TYPES);
+                    n >>= 1;
+                    rev = 1;
+                } while (n > 1);
+            }
         }
-    } else {
-    	int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
-    	int numBlocks = xLength / numThreads;
-    	if (xLength % numThreads > 0 || numBlocks == 0)
-    		numBlocks++;
 
-    	numBlocks = nd4j::math::nd4j_min<int>(512, numBlocks);
-    	dim3 launchDims(numBlocks, numThreads, 32768);
-
-    	int max = 2, dg = 0;
-    	while (max < xLength) {
-    		max <<= 1;
-    		dg++;
-    	}
-    	max <<= 1;
-
-    	for (int window = 2; window < max; window<<=1) {
-    		int n = window;
-    		int rev = 0;
-    		do{
-    			int half = n >> 1;
-    			BUILD_SINGLE_SELECTOR(xType, bitonicArbitraryStepGeneric, (launchDims, stream, dX, dXShapeInfo, n, xLength, rev, descending), LIBND4J_TYPES);
-    			n>>=1;
-    			rev = 1;
-    		} while(n > 1);
-    	}
+        nd4j::DebugHelper::checkErrorCode(stream, "sort(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
     }
-
-    nd4j::DebugHelper::checkErrorCode(stream, "sort(...) failed");
 }
 
 
@@ -2284,55 +2234,64 @@ void sortByKey(Nd4jPointer *extraPointers,
                void *y, Nd4jLong *yShapeInfo,
                void *dy, Nd4jLong *dyShapeInfo,
                bool descending) {
+    try {
+        auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
 
-    auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-
-    auto xLength = shape::length(xShapeInfo);
-    auto xEWS = shape::elementWiseStride(xShapeInfo);
-    auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
-    auto yType = nd4j::ArrayOptions::dataType(yShapeInfo);
+        auto xLength = shape::length(xShapeInfo);
+        auto xEWS = shape::elementWiseStride(xShapeInfo);
+        auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
+        auto yType = nd4j::ArrayOptions::dataType(yShapeInfo);
 
 
-    // check if xLength is a power of 2, and use bitonic sort, if that's the case
-    if ((xLength != 0) && ((xLength & (xLength - 1)) == 0) && (xLength <= 1024 * 1024 * 10)) {
-        int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
-        int numBlocks = xLength / numThreads;
-        if (xLength % numThreads > 0 || numBlocks == 0)
-            numBlocks++;
+        // check if xLength is a power of 2, and use bitonic sort, if that's the case
+        if ((xLength != 0) && ((xLength & (xLength - 1)) == 0) && (xLength <= 1024 * 1024 * 10)) {
+            int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
+            int numBlocks = xLength / numThreads;
+            if (xLength % numThreads > 0 || numBlocks == 0)
+                numBlocks++;
 
-        dim3 launchDims(numBlocks, numThreads, 32768);
+            dim3 launchDims(numBlocks, numThreads, 32768);
 
-        for (int k = 2; k <= xLength; k = 2*k) {
-            for (int j = k >> 1; j > 0; j = j >> 1) {
-                BUILD_DOUBLE_SELECTOR(xType, yType, bitonicSortStepGenericKey, (launchDims, stream, dX, dXShapeInfo, dy, dyShapeInfo, j, k, xLength, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+            for (int k = 2; k <= xLength; k = 2 * k) {
+                for (int j = k >> 1; j > 0; j = j >> 1) {
+                    BUILD_DOUBLE_SELECTOR(xType, yType, bitonicSortStepGenericKey,
+                                          (launchDims, stream, dX, dXShapeInfo, dy, dyShapeInfo, j, k, xLength, descending),
+                                          LIBND4J_TYPES, LIBND4J_TYPES);
+                }
+            }
+        } else {
+            int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
+            int numBlocks = xLength / numThreads;
+            if (xLength % numThreads > 0 || numBlocks == 0)
+                numBlocks++;
+
+            numBlocks = nd4j::math::nd4j_min<int>(512, numBlocks);
+            dim3 launchDims(numBlocks, numThreads, 32768);
+
+            int max = 2, dg = 0;
+            while (max < xLength) {
+                max <<= 1;
+                dg++;
+            }
+            max <<= 1;
+
+            for (int window = 2; window < max; window <<= 1) {
+                int n = window;
+                int rev = 0;
+                do {
+                    int half = n >> 1;
+                    BUILD_DOUBLE_SELECTOR(xType, yType, bitonicArbitraryStepGenericKey,
+                                          (launchDims, stream, dX, dXShapeInfo, dy, dyShapeInfo, n, xLength, rev, descending),
+                                          LIBND4J_TYPES, LIBND4J_TYPES);
+                    n >>= 1;
+                    rev = 1;
+                } while (n > 1);
             }
         }
-    } else {
-        int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
-        int numBlocks = xLength / numThreads;
-        if (xLength % numThreads > 0 || numBlocks == 0)
-            numBlocks++;
 
-        numBlocks = nd4j::math::nd4j_min<int>(512, numBlocks);
-        dim3 launchDims(numBlocks, numThreads, 32768);
-
-        int max = 2, dg = 0;
-        while (max < xLength) {
-            max <<= 1;
-            dg++;
-        }
-        max <<= 1;
-
-        for (int window = 2; window < max; window<<=1) {
-            int n = window;
-            int rev = 0;
-            do{
-                int half = n >> 1;
-                BUILD_DOUBLE_SELECTOR(xType, yType, bitonicArbitraryStepGenericKey, (launchDims, stream, dX, dXShapeInfo, dy, dyShapeInfo, n, xLength, rev, descending), LIBND4J_TYPES, LIBND4J_TYPES);
-                n>>=1;
-                rev = 1;
-            } while(n > 1);
-        }
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
     }
 }
 
@@ -2342,54 +2301,63 @@ void sortByValue(Nd4jPointer *extraPointers,
                  void *y, Nd4jLong *yShapeInfo,
                  void *dy, Nd4jLong *dyShapeInfo,
                  bool descending) {
-    auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+    try {
+        auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
 
-    auto xLength = shape::length(xShapeInfo);
-    auto xEWS = shape::elementWiseStride(xShapeInfo);
-    auto xType = nd4j::ArrayOptions::dataType(yShapeInfo);
-    auto yType = nd4j::ArrayOptions::dataType(xShapeInfo);
+        auto xLength = shape::length(xShapeInfo);
+        auto xEWS = shape::elementWiseStride(xShapeInfo);
+        auto xType = nd4j::ArrayOptions::dataType(yShapeInfo);
+        auto yType = nd4j::ArrayOptions::dataType(xShapeInfo);
 
 
-    // check if xLength is a power of 2, and use bitonic sort, if that's the case
-    if ((xLength != 0) && ((xLength & (xLength - 1)) == 0) && (xLength <= 1024 * 1024 * 10)) {
-        int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
-        int numBlocks = xLength / numThreads;
-        if (xLength % numThreads > 0 || numBlocks == 0)
-            numBlocks++;
+        // check if xLength is a power of 2, and use bitonic sort, if that's the case
+        if ((xLength != 0) && ((xLength & (xLength - 1)) == 0) && (xLength <= 1024 * 1024 * 10)) {
+            int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
+            int numBlocks = xLength / numThreads;
+            if (xLength % numThreads > 0 || numBlocks == 0)
+                numBlocks++;
 
-        dim3 launchDims(numBlocks, numThreads, 32768);
+            dim3 launchDims(numBlocks, numThreads, 32768);
 
-        for (int k = 2; k <= xLength; k = 2*k) {
-            for (int j = k >> 1; j > 0; j = j >> 1) {
-                BUILD_DOUBLE_SELECTOR(xType, yType, bitonicSortStepGenericKey, (launchDims, stream, dy, dyShapeInfo, dX, dXShapeInfo, j, k, xLength, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+            for (int k = 2; k <= xLength; k = 2 * k) {
+                for (int j = k >> 1; j > 0; j = j >> 1) {
+                    BUILD_DOUBLE_SELECTOR(xType, yType, bitonicSortStepGenericKey,
+                                          (launchDims, stream, dy, dyShapeInfo, dX, dXShapeInfo, j, k, xLength, descending),
+                                          LIBND4J_TYPES, LIBND4J_TYPES);
+                }
+            }
+        } else {
+            int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
+            int numBlocks = xLength / numThreads;
+            if (xLength % numThreads > 0 || numBlocks == 0)
+                numBlocks++;
+
+            numBlocks = nd4j::math::nd4j_min<int>(512, numBlocks);
+            dim3 launchDims(numBlocks, numThreads, 32768);
+
+            int max = 2, dg = 0;
+            while (max < xLength) {
+                max <<= 1;
+                dg++;
+            }
+            max <<= 1;
+
+            for (int window = 2; window < max; window <<= 1) {
+                int n = window;
+                int rev = 0;
+                do {
+                    int half = n >> 1;
+                    BUILD_DOUBLE_SELECTOR(xType, yType, bitonicArbitraryStepGenericKey,
+                                          (launchDims, stream, dy, dyShapeInfo, dX, dXShapeInfo, n, xLength, rev, descending),
+                                          LIBND4J_TYPES, LIBND4J_TYPES);
+                    n >>= 1;
+                    rev = 1;
+                } while (n > 1);
             }
         }
-    } else {
-        int numThreads = nd4j::math::nd4j_min<int>(512, xLength);
-        int numBlocks = xLength / numThreads;
-        if (xLength % numThreads > 0 || numBlocks == 0)
-            numBlocks++;
-
-        numBlocks = nd4j::math::nd4j_min<int>(512, numBlocks);
-        dim3 launchDims(numBlocks, numThreads, 32768);
-
-        int max = 2, dg = 0;
-        while (max < xLength) {
-            max <<= 1;
-            dg++;
-        }
-        max <<= 1;
-
-        for (int window = 2; window < max; window<<=1) {
-            int n = window;
-            int rev = 0;
-            do{
-                int half = n >> 1;
-                BUILD_DOUBLE_SELECTOR(xType, yType, bitonicArbitraryStepGenericKey, (launchDims, stream, dy, dyShapeInfo, dX, dXShapeInfo, n, xLength, rev, descending), LIBND4J_TYPES, LIBND4J_TYPES);
-                n>>=1;
-                rev = 1;
-            } while(n > 1);
-        }
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
     }
 }
 
@@ -2403,15 +2371,23 @@ void sortTadByKey(Nd4jPointer *extraPointers,
                              int *dimension,
                              int dimensionLength,
                              bool descending) {
-    auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-    auto context = extraPointers[0] == 0 ? LaunchContext::defaultContext(): reinterpret_cast<LaunchContext*>(extraPointers[0]);
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dimension, dimensionLength);
-    dim3 launchDims((int) tadPack.numberOfTads(), 256, 2048);
-    auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
-    auto yType = nd4j::ArrayOptions::dataType(yShapeInfo);
-    BUILD_DOUBLE_SELECTOR(xType, yType, oesTadGenericKey, (launchDims, stream, dX, dXShapeInfo, dy, dyShapeInfo, nullptr, dimensionLength, tadPack.platformShapeInfo(), tadPack.platformOffsets(), descending), LIBND4J_TYPES, LIBND4J_TYPES);
+    try {
+        auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        auto context = extraPointers[0] == 0 ? LaunchContext::defaultContext()
+                                             : reinterpret_cast<LaunchContext *>(extraPointers[0]);
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dimension, dimensionLength);
+        dim3 launchDims((int) tadPack.numberOfTads(), 256, 2048);
+        auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
+        auto yType = nd4j::ArrayOptions::dataType(yShapeInfo);
+        BUILD_DOUBLE_SELECTOR(xType, yType, oesTadGenericKey,
+                              (launchDims, stream, dX, dXShapeInfo, dy, dyShapeInfo, nullptr, dimensionLength, tadPack.platformShapeInfo(), tadPack.platformOffsets(), descending),
+                              LIBND4J_TYPES, LIBND4J_TYPES);
 
-    nd4j::DebugHelper::checkErrorCode(stream, "sortTadKey(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "sortTadKey(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void sortTadByValue(Nd4jPointer *extraPointers,
@@ -2422,16 +2398,24 @@ void sortTadByValue(Nd4jPointer *extraPointers,
                                int *dimension,
                                int dimensionLength,
                                bool descending) {
-    auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-    auto context = extraPointers[0] == 0 ? LaunchContext::defaultContext(): reinterpret_cast<LaunchContext*>(extraPointers[0]);
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dimension, dimensionLength);
-    dim3 launchDims((int) tadPack.numberOfTads(), 256, 2048);
-    auto xType = nd4j::ArrayOptions::dataType(yShapeInfo);
-    auto yType = nd4j::ArrayOptions::dataType(xShapeInfo);
+    try {
+        auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        auto context = extraPointers[0] == 0 ? LaunchContext::defaultContext()
+                                             : reinterpret_cast<LaunchContext *>(extraPointers[0]);
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dimension, dimensionLength);
+        dim3 launchDims((int) tadPack.numberOfTads(), 256, 2048);
+        auto xType = nd4j::ArrayOptions::dataType(yShapeInfo);
+        auto yType = nd4j::ArrayOptions::dataType(xShapeInfo);
 
-    BUILD_DOUBLE_SELECTOR(xType, yType, oesTadGenericKey, (launchDims, stream, dy, dyShapeInfo, dX, dXShapeInfo, nullptr, dimensionLength, tadPack.platformShapeInfo(), tadPack.platformOffsets(), descending), LIBND4J_TYPES, LIBND4J_TYPES);
+        BUILD_DOUBLE_SELECTOR(xType, yType, oesTadGenericKey,
+                              (launchDims, stream, dy, dyShapeInfo, dX, dXShapeInfo, nullptr, dimensionLength, tadPack.platformShapeInfo(), tadPack.platformOffsets(), descending),
+                              LIBND4J_TYPES, LIBND4J_TYPES);
 
-    nd4j::DebugHelper::checkErrorCode(stream, "sortTadValue(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "sortTadValue(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 
@@ -2443,15 +2427,23 @@ void sortTad(Nd4jPointer *extraPointers,
 						Nd4jLong *tadShapeInfo,
 						Nd4jLong *tadOffsets,
 						bool descending) {
-    // to be implemented
-    auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-    auto context = extraPointers[0] == 0 ? LaunchContext::defaultContext(): reinterpret_cast<LaunchContext*>(extraPointers[0]);
-    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dimension, dimensionLength);
-    dim3 launchDims((int) tadPack.numberOfTads(), 512, 33768);
-	auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
-    BUILD_SINGLE_SELECTOR(xType, oesTadGeneric, (launchDims, stream, dX, dXShapeInfo, nullptr, dimensionLength, tadShapeInfo, tadOffsets, descending), LIBND4J_TYPES);
+    try {
+        // to be implemented
+        auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        auto context = extraPointers[0] == 0 ? LaunchContext::defaultContext()
+                                             : reinterpret_cast<LaunchContext *>(extraPointers[0]);
+        auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dimension, dimensionLength);
+        dim3 launchDims((int) tadPack.numberOfTads(), 512, 33768);
+        auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
+        BUILD_SINGLE_SELECTOR(xType, oesTadGeneric,
+                              (launchDims, stream, dX, dXShapeInfo, nullptr, dimensionLength, tadShapeInfo, tadOffsets, descending),
+                              LIBND4J_TYPES);
 
-    nd4j::DebugHelper::checkErrorCode(stream, "sortTad(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "sortTad(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void sortCooIndices(Nd4jPointer *extraPointers, Nd4jLong *indices, void *values, Nd4jLong length, int rank) {
@@ -2464,21 +2456,29 @@ Nd4jLong encodeBitmap(Nd4jPointer *extraPointers,
 								Nd4jLong N,
 								int *dz,
 								float threshold) {
+    try {
 
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-    int *resultPointer = reinterpret_cast<int *>(extraPointers[2]);
-    int *reductionPointer = reinterpret_cast<int *>(extraPointers[3]);
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        int *resultPointer = reinterpret_cast<int *>(extraPointers[2]);
+        int *reductionPointer = reinterpret_cast<int *>(extraPointers[3]);
 
-    dim3 launchDims(512, 512, 32768);
-    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
-    BUILD_SINGLE_SELECTOR(xType, cudaEncodeBitmapGeneric, (launchDims, stream, dx, N, dz, resultPointer, reductionPointer, threshold), LIBND4J_TYPES);
+        dim3 launchDims(512, 512, 32768);
+        auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+        BUILD_SINGLE_SELECTOR(xType, cudaEncodeBitmapGeneric,
+                              (launchDims, stream, dx, N, dz, resultPointer, reductionPointer, threshold),
+                              LIBND4J_TYPES);
 
-    nd4j::DebugHelper::checkErrorCode(stream, "encodeBitmapFloat(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "encodeBitmapFloat(...) failed");
 
-    Nd4jLong dZ = (Nd4jLong) resultPointer[0];
-    resultPointer[0] = 0;
+        Nd4jLong dZ = (Nd4jLong) resultPointer[0];
+        resultPointer[0] = 0;
 
-    return dZ;
+        return dZ;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 0;
+    }
 }
 
 
@@ -2486,13 +2486,17 @@ void decodeBitmap(Nd4jPointer *extraPointers,
 							void *dx,
 							Nd4jLong N,
 							void *dz, Nd4jLong *zShapeInfo) {
+    try {
+        cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        dim3 launchDims(512, 512, 16384);
+        auto xType = nd4j::ArrayOptions::dataType(zShapeInfo);
+        BUILD_SINGLE_SELECTOR(xType, cudaDecodeBitmapGeneric, (launchDims, stream, dx, N, dz), LIBND4J_TYPES);
 
-    cudaStream_t *stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
-    dim3 launchDims(512, 512, 16384);
-    auto xType = nd4j::ArrayOptions::dataType(zShapeInfo);
-    BUILD_SINGLE_SELECTOR(xType, cudaDecodeBitmapGeneric, (launchDims, stream, dx, N, dz), LIBND4J_TYPES);
-
-    nd4j::DebugHelper::checkErrorCode(stream, "decodeBitmapFloat(...) failed");
+        nd4j::DebugHelper::checkErrorCode(stream, "decodeBitmapFloat(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 Nd4jLong* mmapFile(Nd4jPointer *extraPointers, const char *fileName, Nd4jLong length) {
@@ -2505,7 +2509,13 @@ void munmapFile(Nd4jPointer *extraPointers, Nd4jLong* ptrMap, Nd4jLong length) {
 
 
 nd4j::graph::ResultWrapper* executeFlatGraph(Nd4jPointer *extraPointers, Nd4jPointer flatBufferPointer) {
-    return nd4j::graph::GraphExecutioner::executeFlatBuffer(flatBufferPointer);
+    try {
+        return nd4j::graph::GraphExecutioner::executeFlatBuffer(flatBufferPointer);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 Nd4jLong getResultWrapperSize(nd4j::graph::ResultWrapper* ptr) {
@@ -2560,9 +2570,16 @@ nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::D
 }
 
 nd4j::ShapeList* calculateOutputShapes2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool *bArgs, int numBArgs) {
-    auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
+    try {
+        auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
 
-    return _calculateOutputShapes(extraPointers, op, inputBuffers, inputShapes, numInputShapes, tArgs, numTArgs, iArgs, numIArgs, bArgs, numBArgs);
+        return _calculateOutputShapes(extraPointers, op, inputBuffers, inputShapes, numInputShapes, tArgs, numTArgs,
+                                      iArgs, numIArgs, bArgs, numBArgs);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::DeclarableOp* op, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs) {
@@ -2584,9 +2601,15 @@ nd4j::ShapeList* _calculateOutputShapes(Nd4jPointer* extraPointers, nd4j::ops::D
 }
 
 nd4j::ShapeList* calculateOutputShapes(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputShapes, int numInputShapes, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs) {
-	auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
+    try {
+        auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
 
-	return _calculateOutputShapes(extraPointers, op, inputShapes, numInputShapes, tArgs, numTArgs, iArgs, numIArgs);
+        return _calculateOutputShapes(extraPointers, op, inputShapes, numInputShapes, tArgs, numTArgs, iArgs, numIArgs);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 Nd4jLong getShapeListSize(nd4j::ShapeList* list) {
@@ -2681,39 +2704,57 @@ static FORCEINLINE Nd4jStatus realExec(nd4j::ops::DeclarableOp* op, Nd4jPointer*
 
 
 int execCustomOp(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer* inputBuffers, Nd4jPointer* inputShapes, int numInputs, Nd4jPointer* outputBuffers, Nd4jPointer* outputShapes, int numOutputs, double* tArgs, int numTArgs, Nd4jLong *iArgs, int numIArgs, bool* bArgs, int numBArgs, bool isInplace) {
-	auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
+    try {
+        auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
 
-	return realExec(op, extraPointers, hash, inputBuffers, inputShapes, numInputs, outputBuffers, outputShapes, numOutputs, tArgs, numTArgs, iArgs, numIArgs, bArgs, numBArgs, isInplace);
+        return realExec(op, extraPointers, hash, inputBuffers, inputShapes, numInputs, outputBuffers, outputShapes,
+                        numOutputs, tArgs, numTArgs, iArgs, numIArgs, bArgs, numBArgs, isInplace);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 1;
+    }
 }
 
 int execCustomOp2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer opContext) {
-    auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
-    auto context = reinterpret_cast<Context*>(opContext);
+    try {
+        auto op = nd4j::ops::OpRegistrator::getInstance()->getOperation(hash);
+        auto context = reinterpret_cast<Context *>(opContext);
 
-    auto result = op->execute(context);
+        auto result = op->execute(context);
 
-    auto res = cudaStreamSynchronize(*context->launchContext()->getCudaStream());
-    if (res != 0)
-        throw nd4j::cuda_exception::build("customOp execution failed", res);
+        auto res = cudaStreamSynchronize(*context->launchContext()->getCudaStream());
+        if (res != 0)
+            throw nd4j::cuda_exception::build("customOp execution failed", res);
 
-    for (auto v:context->fastpath_in()) {
-        v->syncToDevice();
+        for (auto v:context->fastpath_in()) {
+            v->syncToDevice();
+        }
+
+        for (auto v:context->fastpath_out()) {
+            v->syncToDevice();
+        }
+
+        return result;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 1;
     }
-
-    for (auto v:context->fastpath_out()) {
-        v->syncToDevice();
-    }
-
-    return result;
 }
 
 int registerGraph(Nd4jPointer *extraPointers, Nd4jLong graphId, Nd4jPointer flatBufferPointer) {
+    try {
+        auto graph = nd4j::graph::GraphExecutioner::importFromFlatPointer(flatBufferPointer);
 
-	auto graph = nd4j::graph::GraphExecutioner::importFromFlatPointer(flatBufferPointer);
+        nd4j::graph::GraphHolder::getInstance()->registerGraph(graphId, graph);
 
-	nd4j::graph::GraphHolder::getInstance()->registerGraph(graphId, graph);
-
-	return ND4J_STATUS_OK;
+        return ND4J_STATUS_OK;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 1;
+    }
 }
 
 
@@ -2764,7 +2805,13 @@ static VariablesSet* executeStoredGraphT(Nd4jPointer *extraPointers, Nd4jLong gr
 }
 
 VariablesSet* executeStoredGraph(Nd4jPointer *extraPointers, Nd4jLong graphId, Nd4jPointer *inputBuffers, Nd4jPointer *inputShapes, int* inputIndices, int numInputs) {
-	return executeStoredGraphT(extraPointers, graphId, inputBuffers, inputShapes, inputIndices, numInputs);
+    try {
+        return executeStoredGraphT(extraPointers, graphId, inputBuffers, inputShapes, inputIndices, numInputs);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 Nd4jLong getVariablesSetSize(nd4j::graph::VariablesSet* set) {
@@ -2800,10 +2847,15 @@ void* getVariableBuffer(nd4j::graph::Variable* variable) {
 }
 
 int unregisterGraph(Nd4jPointer *extraPointers, Nd4jLong graphId) {
+    try {
+        nd4j::graph::GraphHolder::getInstance()->dropGraphAny(graphId);
 
-	nd4j::graph::GraphHolder::getInstance()->dropGraphAny(graphId);
-
-	return ND4J_STATUS_OK;
+        return ND4J_STATUS_OK;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 1;
+    }
 }
 
 void deletePointerArray(Nd4jPointer pointer) {
@@ -2918,8 +2970,15 @@ Nd4jStatus execCustomOpWithScope(Nd4jPointer *extraPointers, nd4j::graph::GraphS
 
 
 Nd4jStatus execCustomOpWithScope(Nd4jPointer *extraPointers, Nd4jPointer state, Nd4jLong opHash, Nd4jLong *scopes, int numScopes, Nd4jPointer *inputBuffers, Nd4jPointer *inputShapes, int numInputs, Nd4jPointer *outputBuffers, Nd4jPointer *outputShapes, int numOutputs) {
-
-    return execCustomOpWithScope(extraPointers, reinterpret_cast<nd4j::graph::GraphState*>(state), opHash, scopes, numScopes, inputBuffers, inputShapes, numInputs, outputBuffers, outputShapes, numOutputs);
+    try {
+        return execCustomOpWithScope(extraPointers, reinterpret_cast<nd4j::graph::GraphState *>(state), opHash, scopes,
+                                     numScopes, inputBuffers, inputShapes, numInputs, outputBuffers, outputShapes,
+                                     numOutputs);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return 1;
+    }
 }
 
 void deleteResultWrapper(Nd4jPointer ptr) {
@@ -2937,181 +2996,186 @@ int estimateThreshold(Nd4jPointer *extraPointers, Nd4jPointer dX, Nd4jLong *dXSh
  *     void convertTypes(Nd4jPointer *extras, int srcType, Nd4jPointer dX, long N, int dstType, Nd4jPointer dZ);
  */
 void convertTypes(Nd4jPointer *extras, int srcType, Nd4jPointer dX, Nd4jLong N, int dstType, Nd4jPointer dZ) {
- 	auto dx = reinterpret_cast<void *>(dX);
-	auto dz = reinterpret_cast<void *>(dZ);
+    try {
+        auto dx = reinterpret_cast<void *>(dX);
+        auto dz = reinterpret_cast<void *>(dZ);
 
-    if (srcType == ND4J_FLOAT8) {
-        if (dstType == ND4J_FLOAT8) {
-            // convertKernel<double, nd4j::float8>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT8) {
-            //nd4j::TypeCast::convertGenericCuda<nd4j::float8, nd4j::int8>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT8) {
-            //nd4j::TypeCast::convertGenericCuda<nd4j::float8, nd4j::uint8>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT16) {
-            //nd4j::TypeCast::convertGenericCuda<nd4j::float8, float16>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT16) {
-            //nd4j::TypeCast::convertGenericCuda<nd4j::float8, nd4j::int16>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT16) {
-            //nd4j::TypeCast::convertGenericCuda<nd4j::float8, nd4j::uint16>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT24) {
+        if (srcType == ND4J_FLOAT8) {
+            if (dstType == ND4J_FLOAT8) {
+                // convertKernel<double, nd4j::float8>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT8) {
+                //nd4j::TypeCast::convertGenericCuda<nd4j::float8, nd4j::int8>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT8) {
+                //nd4j::TypeCast::convertGenericCuda<nd4j::float8, nd4j::uint8>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT16) {
+                //nd4j::TypeCast::convertGenericCuda<nd4j::float8, float16>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT16) {
+                //nd4j::TypeCast::convertGenericCuda<nd4j::float8, nd4j::int16>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT16) {
+                //nd4j::TypeCast::convertGenericCuda<nd4j::float8, nd4j::uint16>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT24) {
 
-        } else if (dstType == ND4J_FLOAT32) {
-            //nd4j::TypeCast::convertGenericCuda<nd4j::float8, float>(extras, dx, N, dz);
-        } else if (dstType == ND4J_DOUBLE) {
-            //nd4j::TypeCast::convertGenericCuda<nd4j::float8, double>(extras, dx, N, dz);
-        } else {
-            nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
-        }
-    } else if (srcType == ND4J_INT8) {
-        if (dstType == ND4J_FLOAT8) {
-            //nd4j::TypeCast::convertGenericCuda<nd4j::int8, nd4j::float8>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT8) {
-            //convertKernel<nd4j::int8, nd4j::int8>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT8) {
-            nd4j::TypeCast::convertGenericCuda<int8_t, uint8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT16) {
-            nd4j::TypeCast::convertGenericCuda<int8_t, float16>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT16) {
-            nd4j::TypeCast::convertGenericCuda<int8_t, int16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT16) {
-            nd4j::TypeCast::convertGenericCuda<int8_t, uint16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT24) {
-            // TODO: eventually we might want to add it
-        } else if (dstType == ND4J_FLOAT32) {
-            nd4j::TypeCast::convertGenericCuda<int8_t, float>(extras, dx, N, dz);
-        } else if (dstType == ND4J_DOUBLE) {
-            nd4j::TypeCast::convertGenericCuda<int8_t, double>(extras, dx, N, dz);
-        } else {
-            nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
-        }
-    } else if (srcType == ND4J_UINT8) {
-        if (dstType == ND4J_FLOAT8) {
-            //nd4j::TypeCast::convertGenericCuda<uint8_t, nd4j::float8>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT8) {
-            nd4j::TypeCast::convertGenericCuda<uint8_t, int8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT8) {
-            nd4j::TypeCast::convertGenericCuda<uint8_t, uint8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT16) {
-            nd4j::TypeCast::convertGenericCuda<uint8_t, float16>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT16) {
-            nd4j::TypeCast::convertGenericCuda<uint8_t, int16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT16) {
-            nd4j::TypeCast::convertGenericCuda<uint8_t, uint16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT24) {
-            // TODO: still might want to add
-        } else if (dstType == ND4J_FLOAT32) {
-            nd4j::TypeCast::convertGenericCuda<uint8_t, float>(extras, dx, N, dz);
-        } else if (dstType == ND4J_DOUBLE) {
-            nd4j::TypeCast::convertGenericCuda<uint8_t, double>(extras, dx, N, dz);
-        } else {
-            nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
-        }
-    } else if (srcType == ND4J_FLOAT16) {
-        if (dstType == ND4J_FLOAT8) {
-            //nd4j::TypeCast::convertGenericCuda<float16, nd4j::float8>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT8) {
-            nd4j::TypeCast::convertGenericCuda<float16, int8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT8) {
-            nd4j::TypeCast::convertGenericCuda<float16, uint8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT16) {
-            nd4j::TypeCast::convertGenericCuda<float16, float16>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT16) {
-            nd4j::TypeCast::convertGenericCuda<float16, int16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT16) {
-            nd4j::TypeCast::convertGenericCuda<float16, uint16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT24) {
-            // TODO: .... ^^^
-        } else if (dstType == ND4J_FLOAT32) {
-            nd4j::TypeCast::convertGenericCuda<float16, float>(extras, dx, N, dz);
-        } else if (dstType == ND4J_DOUBLE) {
-            nd4j::TypeCast::convertGenericCuda<float16, double>(extras, dx, N, dz);
-        } else if (dstType == ND4J_THRESHOLD) {
-            //nd4j::convertToThreshold<float16>(nullptr, dx, N, dz);
-        } else {
-            nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
-        }
-    } else if (srcType == ND4J_INT16) {
-        if (dstType == ND4J_FLOAT8) {
-            //nd4j::TypeCast::convertGenericCuda<int16_t, nd4j::float8>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT8) {
-            nd4j::TypeCast::convertGenericCuda<int16_t, int8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT8) {
-            nd4j::TypeCast::convertGenericCuda<int16_t, uint8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT16) {
-            nd4j::TypeCast::convertGenericCuda<int16_t, float16>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT16) {
-            nd4j::TypeCast::convertGenericCuda<int16_t, int16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT16) {
-            nd4j::TypeCast::convertGenericCuda<int16_t, uint16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT24) {
-            // TODO...
-        } else if (dstType == ND4J_FLOAT32) {
-            nd4j::TypeCast::convertGenericCuda<int16_t, float>(extras, dx, N, dz);
-        } else if (dstType == ND4J_DOUBLE) {
-            nd4j::TypeCast::convertGenericCuda<int16_t, double>(extras, dx, N, dz);
-        } else {
-            printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
-        }
-    } else if (srcType == ND4J_FLOAT24) {
+            } else if (dstType == ND4J_FLOAT32) {
+                //nd4j::TypeCast::convertGenericCuda<nd4j::float8, float>(extras, dx, N, dz);
+            } else if (dstType == ND4J_DOUBLE) {
+                //nd4j::TypeCast::convertGenericCuda<nd4j::float8, double>(extras, dx, N, dz);
+            } else {
+                nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
+            }
+        } else if (srcType == ND4J_INT8) {
+            if (dstType == ND4J_FLOAT8) {
+                //nd4j::TypeCast::convertGenericCuda<nd4j::int8, nd4j::float8>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT8) {
+                //convertKernel<nd4j::int8, nd4j::int8>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT8) {
+                nd4j::TypeCast::convertGenericCuda<int8_t, uint8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT16) {
+                nd4j::TypeCast::convertGenericCuda<int8_t, float16>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT16) {
+                nd4j::TypeCast::convertGenericCuda<int8_t, int16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT16) {
+                nd4j::TypeCast::convertGenericCuda<int8_t, uint16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT24) {
+                // TODO: eventually we might want to add it
+            } else if (dstType == ND4J_FLOAT32) {
+                nd4j::TypeCast::convertGenericCuda<int8_t, float>(extras, dx, N, dz);
+            } else if (dstType == ND4J_DOUBLE) {
+                nd4j::TypeCast::convertGenericCuda<int8_t, double>(extras, dx, N, dz);
+            } else {
+                nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
+            }
+        } else if (srcType == ND4J_UINT8) {
+            if (dstType == ND4J_FLOAT8) {
+                //nd4j::TypeCast::convertGenericCuda<uint8_t, nd4j::float8>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT8) {
+                nd4j::TypeCast::convertGenericCuda<uint8_t, int8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT8) {
+                nd4j::TypeCast::convertGenericCuda<uint8_t, uint8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT16) {
+                nd4j::TypeCast::convertGenericCuda<uint8_t, float16>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT16) {
+                nd4j::TypeCast::convertGenericCuda<uint8_t, int16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT16) {
+                nd4j::TypeCast::convertGenericCuda<uint8_t, uint16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT24) {
+                // TODO: still might want to add
+            } else if (dstType == ND4J_FLOAT32) {
+                nd4j::TypeCast::convertGenericCuda<uint8_t, float>(extras, dx, N, dz);
+            } else if (dstType == ND4J_DOUBLE) {
+                nd4j::TypeCast::convertGenericCuda<uint8_t, double>(extras, dx, N, dz);
+            } else {
+                nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
+            }
+        } else if (srcType == ND4J_FLOAT16) {
+            if (dstType == ND4J_FLOAT8) {
+                //nd4j::TypeCast::convertGenericCuda<float16, nd4j::float8>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT8) {
+                nd4j::TypeCast::convertGenericCuda<float16, int8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT8) {
+                nd4j::TypeCast::convertGenericCuda<float16, uint8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT16) {
+                nd4j::TypeCast::convertGenericCuda<float16, float16>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT16) {
+                nd4j::TypeCast::convertGenericCuda<float16, int16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT16) {
+                nd4j::TypeCast::convertGenericCuda<float16, uint16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT24) {
+                // TODO: .... ^^^
+            } else if (dstType == ND4J_FLOAT32) {
+                nd4j::TypeCast::convertGenericCuda<float16, float>(extras, dx, N, dz);
+            } else if (dstType == ND4J_DOUBLE) {
+                nd4j::TypeCast::convertGenericCuda<float16, double>(extras, dx, N, dz);
+            } else if (dstType == ND4J_THRESHOLD) {
+                //nd4j::convertToThreshold<float16>(nullptr, dx, N, dz);
+            } else {
+                nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
+            }
+        } else if (srcType == ND4J_INT16) {
+            if (dstType == ND4J_FLOAT8) {
+                //nd4j::TypeCast::convertGenericCuda<int16_t, nd4j::float8>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT8) {
+                nd4j::TypeCast::convertGenericCuda<int16_t, int8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT8) {
+                nd4j::TypeCast::convertGenericCuda<int16_t, uint8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT16) {
+                nd4j::TypeCast::convertGenericCuda<int16_t, float16>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT16) {
+                nd4j::TypeCast::convertGenericCuda<int16_t, int16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT16) {
+                nd4j::TypeCast::convertGenericCuda<int16_t, uint16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT24) {
+                // TODO...
+            } else if (dstType == ND4J_FLOAT32) {
+                nd4j::TypeCast::convertGenericCuda<int16_t, float>(extras, dx, N, dz);
+            } else if (dstType == ND4J_DOUBLE) {
+                nd4j::TypeCast::convertGenericCuda<int16_t, double>(extras, dx, N, dz);
+            } else {
+                printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
+            }
+        } else if (srcType == ND4J_FLOAT24) {
 
-    } else if (srcType == ND4J_FLOAT32) {
-        if (dstType == ND4J_FLOAT8) {
-            //nd4j::TypeCast::convertGenericCuda<float, nd4j::float8>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT8) {
-            nd4j::TypeCast::convertGenericCuda<float, int8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT8) {
-            nd4j::TypeCast::convertGenericCuda<float, uint8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT16) {
-            nd4j::TypeCast::convertGenericCuda<float, float16>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT16) {
-            nd4j::TypeCast::convertGenericCuda<float, int16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT16) {
-            nd4j::TypeCast::convertGenericCuda<float, uint16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT24) {
+        } else if (srcType == ND4J_FLOAT32) {
+            if (dstType == ND4J_FLOAT8) {
+                //nd4j::TypeCast::convertGenericCuda<float, nd4j::float8>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT8) {
+                nd4j::TypeCast::convertGenericCuda<float, int8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT8) {
+                nd4j::TypeCast::convertGenericCuda<float, uint8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT16) {
+                nd4j::TypeCast::convertGenericCuda<float, float16>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT16) {
+                nd4j::TypeCast::convertGenericCuda<float, int16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT16) {
+                nd4j::TypeCast::convertGenericCuda<float, uint16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT24) {
 
-        } else if (dstType == ND4J_DOUBLE) {
-            nd4j::TypeCast::convertGenericCuda<float, double>(extras, dx, N, dz);
-        } else if (dstType == ND4J_THRESHOLD) {
-            //nd4j::convertToThreshold<float>(nullptr, dx, N, dz);
-        } else {
-            nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
-        }
-    } else if (srcType == ND4J_DOUBLE) {
-        if (dstType == ND4J_FLOAT8) {
-            //nd4j::TypeCast::convertGenericCuda<double, nd4j::float8>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT8) {
-            nd4j::TypeCast::convertGenericCuda<double, int8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT8) {
-            nd4j::TypeCast::convertGenericCuda<double, uint8_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT16) {
-            nd4j::TypeCast::convertGenericCuda<double, float16>(extras, dx, N, dz);
-        } else if (dstType == ND4J_INT16) {
-            nd4j::TypeCast::convertGenericCuda<double, int16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_UINT16) {
-            nd4j::TypeCast::convertGenericCuda<double, uint16_t>(extras, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT24) {
+            } else if (dstType == ND4J_DOUBLE) {
+                nd4j::TypeCast::convertGenericCuda<float, double>(extras, dx, N, dz);
+            } else if (dstType == ND4J_THRESHOLD) {
+                //nd4j::convertToThreshold<float>(nullptr, dx, N, dz);
+            } else {
+                nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
+            }
+        } else if (srcType == ND4J_DOUBLE) {
+            if (dstType == ND4J_FLOAT8) {
+                //nd4j::TypeCast::convertGenericCuda<double, nd4j::float8>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT8) {
+                nd4j::TypeCast::convertGenericCuda<double, int8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT8) {
+                nd4j::TypeCast::convertGenericCuda<double, uint8_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT16) {
+                nd4j::TypeCast::convertGenericCuda<double, float16>(extras, dx, N, dz);
+            } else if (dstType == ND4J_INT16) {
+                nd4j::TypeCast::convertGenericCuda<double, int16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_UINT16) {
+                nd4j::TypeCast::convertGenericCuda<double, uint16_t>(extras, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT24) {
 
-        } else if (dstType == ND4J_FLOAT32) {
-            nd4j::TypeCast::convertGenericCuda<double, float>(extras, dx, N, dz);
-        } else if (dstType == ND4J_DOUBLE) {
-            //
-        } else if (dstType == ND4J_THRESHOLD) {
-            //nd4j::convertToThreshold<double>(nullptr, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT32) {
+                nd4j::TypeCast::convertGenericCuda<double, float>(extras, dx, N, dz);
+            } else if (dstType == ND4J_DOUBLE) {
+                //
+            } else if (dstType == ND4J_THRESHOLD) {
+                //nd4j::convertToThreshold<double>(nullptr, dx, N, dz);
+            } else {
+                nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
+            }
+        } else if (srcType == ND4J_THRESHOLD) {
+            if (dstType == ND4J_FLOAT16) {
+                //nd4j::convertFromThreshold<float16>(nullptr, dx, N, dz);
+            } else if (dstType == ND4J_FLOAT32) {
+                //nd4j::convertFromThreshold<float>(nullptr, dx, N, dz);
+            } else if (dstType == ND4J_DOUBLE) {
+                //nd4j::convertFromThreshold<double>(nullptr, dx, N, dz);
+            } else {
+                nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
+            }
         } else {
             nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
         }
-    } else if (srcType == ND4J_THRESHOLD) {
-        if (dstType == ND4J_FLOAT16) {
-            //nd4j::convertFromThreshold<float16>(nullptr, dx, N, dz);
-        } else if (dstType == ND4J_FLOAT32) {
-            //nd4j::convertFromThreshold<float>(nullptr, dx, N, dz);
-        } else if (dstType == ND4J_DOUBLE) {
-            //nd4j::convertFromThreshold<double>(nullptr, dx, N, dz);
-        } else {
-            nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
-        }
-    } else {
-        nd4j_printf("Unsupported types conversion: [%i] -> [%i]\n", srcType, dstType);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
     }
 }
 
@@ -3209,20 +3273,31 @@ void scatterUpdate(Nd4jPointer *extraPointers, int opCode, int numOfSubArrs,
                       			void* hY, Nd4jLong* hYShapeInfo, Nd4jLong* hYOffsets,
                       			void* dY, Nd4jLong* dYShapeInfo, Nd4jLong* dYOffsets,
                       			int* hIindexes, int* dIndexes) {
+    try {
+        auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
 
-	auto stream = reinterpret_cast<cudaStream_t *>(extraPointers[1]);
+        nd4j::DataType type = ArrayOptions::dataType(hXShapeInfo);
 
-	nd4j::DataType type = ArrayOptions::dataType(hXShapeInfo);
-
-    BUILD_SINGLE_SELECTOR(type, scatterUpdateCudaLauncher, (stream, opCode, numOfSubArrs, dX, dXShapeInfo, dXOffsets, dY, dYShapeInfo, dYOffsets, dIndexes), LIBND4J_TYPES);
-    nd4j::DebugHelper::checkErrorCode(stream, "scatterUpdate(...) failed");
+        BUILD_SINGLE_SELECTOR(type, scatterUpdateCudaLauncher,
+                              (stream, opCode, numOfSubArrs, dX, dXShapeInfo, dXOffsets, dY, dYShapeInfo, dYOffsets, dIndexes),
+                              LIBND4J_TYPES);
+        nd4j::DebugHelper::checkErrorCode(stream, "scatterUpdate(...) failed");
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void inspectArray(Nd4jPointer *extraPointers, Nd4jPointer buffer, Nd4jLong *shapeInfo, Nd4jPointer specialBuffer, Nd4jLong *specialShapeInfo, Nd4jPointer debugInfo) {
-    LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
-    auto p = reinterpret_cast<nd4j::DebugInfo*>(debugInfo);
-    NDArray array(buffer, specialBuffer, shapeInfo, &lc);
-    nd4j::DebugHelper::retrieveDebugStatistics(p, &array);
+    try {
+        LaunchContext lc(extraPointers[1], extraPointers[4], extraPointers[5], extraPointers[3]);
+        auto p = reinterpret_cast<nd4j::DebugInfo *>(debugInfo);
+        NDArray array(buffer, specialBuffer, shapeInfo, &lc);
+        nd4j::DebugHelper::retrieveDebugStatistics(p, &array);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 void __global__ tryPointerKernel(void* p, int len) {
@@ -3239,26 +3314,37 @@ void __global__ tryPointerKernel(void* p, int len) {
 }
 
 void tryPointer(Nd4jPointer extra, Nd4jPointer p, int len) {
+    try {
+        cudaStream_t stream;
+        cudaStreamCreate(&stream);
 
-    cudaStream_t stream;
-    cudaStreamCreate(&stream);
+        tryPointerKernel << < 256, 512, len + 64, stream >> > (p, len);
+        auto e = cudaStreamSynchronize(stream);
 
-    tryPointerKernel<<<256, 512, len+64, stream>>>(p, len);
-    auto e = cudaStreamSynchronize(stream);
+        if (e != 0)
+            throw nd4j::cuda_exception::build("tryPointer failed", e);
 
-    if (e != 0)
-        throw nd4j::cuda_exception::build("tryPointer failed", e);
-
-    cudaStreamDestroy(stream);
+        cudaStreamDestroy(stream);
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+    }
 }
 
 int dataTypeFromNpyHeader(void *header) {
     return (int) cnpy::dataTypeFromHeader(reinterpret_cast<char *>(header));
 }
 nd4j::ConstantDataBuffer* shapeBuffer(int rank, Nd4jLong *shape, Nd4jLong *strides, nd4j::DataType dtype, char order, Nd4jLong ews, bool empty) {
-    auto buffer = new ConstantDataBuffer();
-    *buffer = nd4j::ConstantShapeHelper::getInstance()->bufferForShapeInfo(ShapeDescriptor(dtype, order, shape, strides, rank, ews, empty));
-    return buffer;
+    try {
+        auto buffer = new ConstantDataBuffer();
+        *buffer = nd4j::ConstantShapeHelper::getInstance()->bufferForShapeInfo(
+                ShapeDescriptor(dtype, order, shape, strides, rank, ews, empty));
+        return buffer;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 void deleteShapeBuffer(nd4j::ConstantDataBuffer* ptr) {
@@ -3359,60 +3445,79 @@ void deleteRandomGenerator(nd4j::graph::RandomGenerator* ptr) {
 
 
 Nd4jPointer shapeBufferForNumpy(Nd4jPointer npyArray) {
-    cnpy::NpyArray arr = cnpy::loadNpyFromPointer(reinterpret_cast<char *>(npyArray));
-    unsigned int shapeSize = arr.shape.size();
-    std::vector<Nd4jLong> shape(shapeSize);
-    bool _empty = false;
-    for(unsigned int i = 0; i < shapeSize; i++) {
-        shape[i] = arr.shape[i];
+    try {
+        cnpy::NpyArray arr = cnpy::loadNpyFromPointer(reinterpret_cast<char *>(npyArray));
+        unsigned int shapeSize = arr.shape.size();
+        std::vector<Nd4jLong> shape(shapeSize);
+        bool _empty = false;
+        for (unsigned int i = 0; i < shapeSize; i++) {
+            shape[i] = arr.shape[i];
 
-        if (arr.shape[i] == 0)
-            _empty = true;
+            if (arr.shape[i] == 0)
+                _empty = true;
+        }
+
+        auto dtype = cnpy::dataTypeFromHeader(reinterpret_cast<char *>(npyArray));
+
+        Nd4jLong *shapeBuffer;
+        if (shape.size() == 1 && shape[0] == 0) {
+            // scalar case
+            shapeBuffer = nd4j::ShapeBuilders::createScalarShapeInfo(dtype);
+        } else if (_empty) {
+            if (shapeSize > 0)
+                shapeBuffer = nd4j::ShapeBuilders::emptyShapeInfo(dtype, arr.fortranOrder ? 'f' : 'c', shape);
+            else
+                shapeBuffer = nd4j::ShapeBuilders::emptyShapeInfo(dtype);
+        } else {
+            shapeBuffer = nd4j::ShapeBuilders::createShapeInfo(dtype, arr.fortranOrder ? 'f' : 'c', shape);
+        }
+        return reinterpret_cast<Nd4jPointer>(nd4j::ConstantShapeHelper::getInstance()->createFromExisting(shapeBuffer,
+                                                                                                          true));
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
     }
-
-    auto dtype = cnpy::dataTypeFromHeader(reinterpret_cast<char *>(npyArray));
-
-    Nd4jLong *shapeBuffer;
-    if (shape.size() == 1 && shape[0] == 0) {
-        // scalar case
-        shapeBuffer = nd4j::ShapeBuilders::createScalarShapeInfo(dtype);
-    } else if (_empty) {
-        if (shapeSize > 0)
-            shapeBuffer = nd4j::ShapeBuilders::emptyShapeInfo(dtype, arr.fortranOrder ? 'f' : 'c', shape);
-        else
-            shapeBuffer = nd4j::ShapeBuilders::emptyShapeInfo(dtype);
-    } else {
-        shapeBuffer = nd4j::ShapeBuilders::createShapeInfo(dtype, arr.fortranOrder ? 'f' : 'c', shape);
-    }
-    return reinterpret_cast<Nd4jPointer>(nd4j::ConstantShapeHelper::getInstance()->createFromExisting(shapeBuffer, true));
 }
 
 const char* runLightBenchmarkSuit(bool printOut) {
-    nd4j::LightBenchmarkSuit suit;
-    auto result = suit.runSuit();
+    try {
+        nd4j::LightBenchmarkSuit suit;
+        auto result = suit.runSuit();
 
-    if (printOut)
-        nd4j_printf("%s\n", result.data());
+        if (printOut)
+            nd4j_printf("%s\n", result.data());
 
-    auto chars = new char[result.length()+1];
-    std::memcpy(chars, result.data(), result.length());
-    chars[result.length()] = (char) 0x0;
+        auto chars = new char[result.length() + 1];
+        std::memcpy(chars, result.data(), result.length());
+        chars[result.length()] = (char) 0x0;
 
-    return chars;
+        return chars;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 const char* runFullBenchmarkSuit(bool printOut) {
-    nd4j::FullBenchmarkSuit suit;
-    auto result = suit.runSuit();
+    try {
+        nd4j::FullBenchmarkSuit suit;
+        auto result = suit.runSuit();
 
-    if (printOut)
-        nd4j_printf("%s\n", result.data());
+        if (printOut)
+            nd4j_printf("%s\n", result.data());
 
-    auto chars = new char[result.length()+1];
-    std::memcpy(chars, result.data(), result.length());
-    chars[result.length()] = (char) 0x0;
+        auto chars = new char[result.length() + 1];
+        std::memcpy(chars, result.data(), result.length());
+        chars[result.length()] = (char) 0x0;
 
-    return chars;
+        return chars;
+    } catch (std::exception &e) {
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorCode(1);
+        nd4j::LaunchContext::defaultContext()->errorReference()->setErrorMessage(e.what());
+        return nullptr;
+    }
 }
 
 Nd4jLong getCachedMemory(int deviceId) {
@@ -3449,4 +3554,12 @@ Nd4jPointer lcBlasHandle(OpaqueLaunchContext* lc) {
 
 Nd4jPointer lcSolverHandle(OpaqueLaunchContext* lc) {
     return lc->getCusolverHandle();
+}
+
+int lastErrorCode() {
+    return nd4j::LaunchContext::defaultContext()->errorReference()->errorCode();
+}
+
+const char* lastErrorMessage() {
+    return nd4j::LaunchContext::defaultContext()->errorReference()->errorMessage();
 }
\ No newline at end of file
diff --git a/libnd4j/include/execution/ContextBuffers.h b/libnd4j/include/execution/ContextBuffers.h
index 130354070..67c428d27 100644
--- a/libnd4j/include/execution/ContextBuffers.h
+++ b/libnd4j/include/execution/ContextBuffers.h
@@ -23,6 +23,7 @@
 
 #include <dll.h>
 #include <pointercast.h>
+#include <execution/ErrorReference.h>
 
 namespace nd4j {
     class ND4J_EXPORT ContextBuffers {
@@ -32,6 +33,7 @@ namespace nd4j {
         void* _allocationPointer = nullptr;
         void* _execStream = nullptr;
         void* _specialStream = nullptr;
+        sd::ErrorReference _errorReference;
         bool _allocated = false;
         bool _initialized = false;
 
@@ -60,6 +62,8 @@ namespace nd4j {
         void setScalarBuffer(void* pointer);
         void setAllocationBuffer(void* pointer);
 
+        sd::ErrorReference* errorReference();
+
         void triggerOwnership(bool isOwner);
 
         int deviceId();
diff --git a/libnd4j/include/helpers/ProviderRNG.h b/libnd4j/include/execution/ErrorReference.h
similarity index 57%
rename from libnd4j/include/helpers/ProviderRNG.h
rename to libnd4j/include/execution/ErrorReference.h
index e82f6ac98..2b68d5855 100644
--- a/libnd4j/include/helpers/ProviderRNG.h
+++ b/libnd4j/include/execution/ErrorReference.h
@@ -15,32 +15,32 @@
  ******************************************************************************/
 
 //
-// Created by Yurii Shyrma on 27.01.2018
+// @author raver119@gmail.com
 //
 
-#ifndef LIBND4J_PROVIDERRNG_H
-#define LIBND4J_PROVIDERRNG_H
+#ifndef DEV_TESTS_ERRORREFERENCE_H
+#define DEV_TESTS_ERRORREFERENCE_H
 
-#include <helpers/helper_random.h>
-#include <mutex>
-
-namespace nd4j {
-    
-class ProviderRNG {
-        
-    protected:
-        random::RandomBuffer* _rng;
-        static std::mutex _mutex;
-        ProviderRNG();
+#include <string>
+#include <dll.h>
 
+namespace sd {
+    class ND4J_EXPORT ErrorReference {
+    private:
+        int _errorCode = 0;
+        std::string _errorMessage;
     public:
-        ProviderRNG(const ProviderRNG&)    = delete;
-        void operator=(const ProviderRNG&) = delete;   
-        random::RandomBuffer* getRNG() const;
-        static ProviderRNG& getInstance();        
-};
+        ErrorReference() = default;
+        ~ErrorReference() = default;
 
+        int errorCode();
+        const char* errorMessage();
 
+        void setErrorCode(int errorCode);
+        void setErrorMessage(std::string message);
+        void setErrorMessage(const char* message);
+    };
 }
 
-#endif //LIBND4J_PROVIDERRNG_H
+
+#endif //DEV_TESTS_ERRORREFERENCE_H
diff --git a/libnd4j/include/execution/LaunchContext.h b/libnd4j/include/execution/LaunchContext.h
index 23165fa0e..5fae2162c 100644
--- a/libnd4j/include/execution/LaunchContext.h
+++ b/libnd4j/include/execution/LaunchContext.h
@@ -37,6 +37,7 @@
 #include <vector>
 #include <mutex>
 #include <execution/ContextBuffers.h>
+#include <execution/ErrorReference.h>
 
 
 
@@ -97,9 +98,12 @@ class ND4J_EXPORT LaunchContext {
 
     	int getDeviceID() const {return _deviceID;}
     	void setDeviceID(int deviceID) { _deviceID = deviceID; }
+        sd::ErrorReference* errorReference();
 
     	static bool isInitialized();
     	static void releaseBuffers();
+
+
 	    static LaunchContext* defaultContext();
 
 
diff --git a/libnd4j/include/execution/cpu/ContextBuffers.cpp b/libnd4j/include/execution/cpu/ContextBuffers.cpp
index 3bf0a01eb..0038990c2 100644
--- a/libnd4j/include/execution/cpu/ContextBuffers.cpp
+++ b/libnd4j/include/execution/cpu/ContextBuffers.cpp
@@ -99,4 +99,8 @@ namespace nd4j {
     ContextBuffers& ContextBuffers::operator=(ContextBuffers&& other) {
         return *this;
     }
+
+    sd::ErrorReference* ContextBuffers::errorReference() {
+        return &_errorReference;
+    }
 }
\ No newline at end of file
diff --git a/libnd4j/include/execution/cpu/LaunchContext.cpp b/libnd4j/include/execution/cpu/LaunchContext.cpp
index 3ee460350..60e29c7ca 100644
--- a/libnd4j/include/execution/cpu/LaunchContext.cpp
+++ b/libnd4j/include/execution/cpu/LaunchContext.cpp
@@ -23,7 +23,11 @@
 #include <exceptions/cuda_exception.h>
 #include <thread>
 
+#ifdef IOS_BUILD
 nd4j::ContextBuffers contextBuffers = nd4j::ContextBuffers();
+#else
+thread_local nd4j::ContextBuffers contextBuffers = nd4j::ContextBuffers();
+#endif
 
 namespace nd4j {
 
@@ -65,4 +69,8 @@ namespace nd4j {
     void LaunchContext::releaseBuffers() {
         //
     }
+
+    sd::ErrorReference* LaunchContext::errorReference() {
+        return contextBuffers.errorReference();
+    }
 }
\ No newline at end of file
diff --git a/libnd4j/include/execution/cuda/ContextBuffers.cu b/libnd4j/include/execution/cuda/ContextBuffers.cu
index 84db0c284..895bb6623 100644
--- a/libnd4j/include/execution/cuda/ContextBuffers.cu
+++ b/libnd4j/include/execution/cuda/ContextBuffers.cu
@@ -220,5 +220,9 @@ namespace nd4j {
     bool ContextBuffers::isInitialized() {
         return _initialized;
     }
+
+    sd::ErrorReference* ContextBuffers::errorReference() {
+        return &_errorReference;
+    }
 }
 
diff --git a/libnd4j/include/execution/cuda/LaunchContext.cu b/libnd4j/include/execution/cuda/LaunchContext.cu
index 1292f756c..9d9f2c506 100644
--- a/libnd4j/include/execution/cuda/LaunchContext.cu
+++ b/libnd4j/include/execution/cuda/LaunchContext.cu
@@ -168,4 +168,8 @@ LaunchContext::LaunchContext() {
     bool LaunchContext::isInitialized() {
         return contextBuffers.isInitialized();
     }
+
+    sd::ErrorReference* LaunchContext::errorReference() {
+        return contextBuffers.errorReference();
+    }
 }
\ No newline at end of file
diff --git a/libnd4j/include/helpers/impl/ProviderRNG.cpp b/libnd4j/include/execution/impl/ErrorReference.cpp
similarity index 52%
rename from libnd4j/include/helpers/impl/ProviderRNG.cpp
rename to libnd4j/include/execution/impl/ErrorReference.cpp
index 216aa3a32..7b3409aa1 100644
--- a/libnd4j/include/helpers/impl/ProviderRNG.cpp
+++ b/libnd4j/include/execution/impl/ErrorReference.cpp
@@ -15,37 +15,32 @@
  ******************************************************************************/
 
 //
-// Created by Yurii Shyrma on 27.01.2018
+// @author raver119@gmail.com
 //
 
-#include <helpers/ProviderRNG.h>
-#include <NativeOps.h>
 
+#include <execution/ErrorReference.h>
 
-namespace nd4j {
-    
-ProviderRNG::ProviderRNG() {
+namespace sd {
+    int ErrorReference::errorCode() {
+        return _errorCode;
+    }
 
-    Nd4jLong *buffer = new Nd4jLong[100000];
-    std::lock_guard<std::mutex> lock(_mutex);
-    #ifndef __CUDABLAS__
-    // at this moment we don't have streams etc, so let's just skip this for now
-    _rng = (nd4j::random::RandomBuffer *) initRandom(nullptr, 123, 100000, (Nd4jPointer) buffer);    
-    #endif
-    // if(_rng != nullptr)        
-}
-
-ProviderRNG& ProviderRNG::getInstance() {     
-    
-    static ProviderRNG instance; 
-    return instance;
-}
-
-random::RandomBuffer* ProviderRNG::getRNG() const {
-
-    return _rng;
-}
-
-std::mutex ProviderRNG::_mutex;
-    
+    const char* ErrorReference::errorMessage() {
+        // since we're fetching error message - error code will be assumed consumed & nullified
+        _errorCode = 0;
+        return _errorMessage.c_str();
+    }
+
+    void ErrorReference::setErrorCode(int errorCode) {
+        _errorCode = errorCode;
+    }
+
+    void ErrorReference::setErrorMessage(std::string message) {
+        _errorMessage = message;
+    }
+
+    void ErrorReference::setErrorMessage(const char* message) {
+        _errorMessage = std::string(message);
+    }
 }
diff --git a/libnd4j/include/ops/declarable/generic/convo/ismax.cpp b/libnd4j/include/ops/declarable/generic/convo/ismax.cpp
index ad5a485e1..13de73e81 100644
--- a/libnd4j/include/ops/declarable/generic/convo/ismax.cpp
+++ b/libnd4j/include/ops/declarable/generic/convo/ismax.cpp
@@ -45,7 +45,7 @@ DECLARE_SYN(IsMax, ismax);
     DECLARE_TYPES(ismax) {
         getOpDescriptor()
                 ->setAllowedInputTypes(0, DataType::ANY)
-                ->setAllowedOutputTypes(0, DataType::BOOL);
+                ->setAllowedOutputTypes(0, DataType::ANY);
 
     }
 
diff --git a/libnd4j/include/ops/declarable/generic/nlp/cbow.cpp b/libnd4j/include/ops/declarable/generic/nlp/cbow.cpp
index 2ae69e296..21906f4eb 100644
--- a/libnd4j/include/ops/declarable/generic/nlp/cbow.cpp
+++ b/libnd4j/include/ops/declarable/generic/nlp/cbow.cpp
@@ -84,7 +84,8 @@ namespace nd4j {
                     ->setAllowedInputTypes(11, nd4j::DataType::INT64)
                     ->setAllowedInputTypes(12, nd4j::DataType::INT32)
                     ->setAllowedInputTypes(13, nd4j::DataType::INT32)
-                    ->setAllowedInputTypes(14, {ALL_FLOATS});
+                    ->setAllowedInputTypes(14, {ALL_FLOATS})
+                    ->setAllowedOutputTypes(nd4j::DataType::ANY);
         }
     }
 }
diff --git a/libnd4j/include/ops/declarable/generic/nlp/skipgram.cpp b/libnd4j/include/ops/declarable/generic/nlp/skipgram.cpp
index 78c6e3818..a97e1a79e 100644
--- a/libnd4j/include/ops/declarable/generic/nlp/skipgram.cpp
+++ b/libnd4j/include/ops/declarable/generic/nlp/skipgram.cpp
@@ -79,7 +79,7 @@ namespace nd4j {
                     ->setAllowedInputTypes(9, {ALL_FLOATS})
                     ->setAllowedInputTypes(10, nd4j::DataType::INT64)
                     ->setAllowedInputTypes(11, {ALL_FLOATS})
-                    ->setAllowedOutputTypes(nd4j::DataType::INT8);
+                    ->setAllowedOutputTypes(nd4j::DataType::ANY);
         }
 
         /*
diff --git a/libnd4j/include/ops/declarable/generic/nn/softmax.cpp b/libnd4j/include/ops/declarable/generic/nn/softmax.cpp
index 08dba09f2..d96f97c10 100644
--- a/libnd4j/include/ops/declarable/generic/nn/softmax.cpp
+++ b/libnd4j/include/ops/declarable/generic/nn/softmax.cpp
@@ -70,7 +70,7 @@ CONFIGURABLE_OP_IMPL(softmax_bp, 2, 1, true, 0, 0) {
 
     DECLARE_TYPES(softmax_bp) {
         getOpDescriptor()
-                ->setAllowedInputTypes(DataType::ANY)
+                ->setAllowedInputTypes({ALL_FLOATS})
                 ->setAllowedOutputTypes({ALL_FLOATS});
     }
 
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/reverse.cu b/libnd4j/include/ops/declarable/helpers/cuda/reverse.cu
index fab9577d6..cf891feab 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/reverse.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/reverse.cu
@@ -30,51 +30,9 @@ namespace nd4j    {
 namespace ops     {
 namespace helpers {
 
-    template <typename T>
-    inline void __device__ indexSwap(T* arr, Nd4jLong idx1, Nd4jLong idx2) {
-        T tmp = arr[idx1];
-        arr[idx1] = arr[idx2];
-        arr[idx2] = tmp;
-    }
-//    template <typename T>
-//    void reverseArray(nd4j::LaunchContext * context, void* inArr, Nd4jLong *inShapeBuffer, void *result, Nd4jLong *zShapeBuffer, int numOfElemsToReverse = 0);
-
-    /////////////////////////////////////////////////////////////////////////////////////
-    template <typename T>
-    static __global__ void reverseArrayInplaceKernel(void *input, Nd4jLong *inputShape, Nd4jLong numOfElemsToReverse) {
-        const auto tid = blockIdx.x * gridDim.x + threadIdx.x;
-        const auto step = gridDim.x * blockDim.x;
-        __shared__ Nd4jLong length;
-        __shared__ int linearStatus;
-        __shared__ T* inputArr;
-        if (threadIdx.x == 0) {
-            length = shape::length(inputShape);
-            linearStatus = shape::elementWiseStride(inputShape);
-            inputArr = reinterpret_cast<T*>(input);
-        }
-        __syncthreads();
-
-        for (Nd4jLong e = tid; e < numOfElemsToReverse / 2; e += step) {
-            if (linearStatus == 1) {
-                auto idx = numOfElemsToReverse - e - 1;
-                indexSwap(inputArr, e, idx);
-            }
-            else if (linearStatus > 1) {
-                auto idx1 = (numOfElemsToReverse - e - 1) * linearStatus;
-                Nd4jLong idx2 =  e * linearStatus;
-                indexSwap(inputArr, idx1, idx2);
-            }
-            else {
-                auto inOffset  = shape::getIndexOffset(e, inputShape, length);
-                auto outOffset = shape::getIndexOffset(numOfElemsToReverse - e - 1, inputShape, length);
-                indexSwap(inputArr, inOffset, outOffset);
-            }
-        }
-    }
-
     template <typename T>
     static __global__ void reverseArrayKernel(void* input, Nd4jLong *inputShape, void* output, Nd4jLong *outputShape, Nd4jLong numOfElemsToReverse) {
-        const auto tid = blockIdx.x * gridDim.x + threadIdx.x;
+        const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
         const auto step = gridDim.x * blockDim.x;
         __shared__ Nd4jLong length;
         __shared__ int linearStatus;
@@ -93,51 +51,47 @@ namespace helpers {
         }
         __syncthreads();
 
-        for (Nd4jLong e = tid; e < length; e += step) {
-            if (e < numOfElemsToReverse ) {
-                if (linearStatus == 1) {
-                    auto idx = numOfElemsToReverse - e - 1;
-                    outputArr[idx] = inputArr[e];
-                } else if (linearStatus > 1) {
-                    auto idx1 = (numOfElemsToReverse - e - 1) * linearStatus;
-                    Nd4jLong idx2 = e * linearStatus;
-                    outputArr[idx1] = inputArr[idx2];
-                } else {
-                    auto inOffset = shape::getIndexOffset(e, inputShape, length);
-                    auto outOffset = shape::getIndexOffset(numOfElemsToReverse - e - 1, outputShape, length);
-                    outputArr[outOffset] = inputArr[inOffset];
-                }
-            }
-            else {
-                if (linearStatus == 1) {
-                    outputArr[e] = inputArr[e];
-                } else if (linearStatus > 1) {
-                    auto idx1 = e * linearStatus;
-                    Nd4jLong idx2 = e * linearStatus;
-                    outputArr[idx1] = inputArr[idx2];
-                } else {
-                    auto inOffset = shape::getIndexOffset(e, inputShape, length);
-                    auto outOffset = shape::getIndexOffset(e, outputShape, length);
-                    outputArr[outOffset] = inputArr[inOffset];
-                }
-            }
+        auto odd = length % 2 != 0;
+        auto limit = length / 2;
+
+        for (Nd4jLong e = tid; e < limit; e += step) {
+            // we're calculating offsets within input array
+            auto fOffset = shape::getIndexOffset(e, inputShape, length);
+            auto lOffset = shape::getIndexOffset(numOfElemsToReverse - e - 1, inputShape, length);
+
+            // now we're storing input values
+            auto v1 = inputArr[fOffset];
+            auto v2 = inputArr[lOffset];
+
+            // now we're calculating offsets within output array
+            auto zfOffset = shape::getIndexOffset(e, outputShape, length);
+            auto zlOffset = shape::getIndexOffset(numOfElemsToReverse - e - 1, outputShape, length);
+
+            // and saving values to output arrays
+            outputArr[zfOffset] = v2;
+            outputArr[zlOffset] = v1;
+
+            //printf("TID: %i; E: %lld; z[%lld], z[%lld] = x[%lld], x[%lld];\n", tid, e, zfOffset, zlOffset, lOffset, fOffset);
         }
 
-        //printf("\n");
+        // in case of odd array we'll have to move middle value
+        if (odd && tid == 0) {
+            auto xOffset = shape::getIndexOffset(limit, inputShape, length);
+            auto zOffset = shape::getIndexOffset(limit, outputShape, length);
+
+            outputArr[zOffset] = inputArr[xOffset];
+            //printf("TID: %i; E: %lld; z[%lld] = x[%lld];\n", tid, limit, zOffset, xOffset);
+        }
     }
 
     template<typename T>
-    static void reverseArray(nd4j::LaunchContext * context, NDArray* input, NDArray* output, int numOfElemsToReverse) {
+    static void reverseArray(nd4j::LaunchContext * context, NDArray* input, NDArray* output, Nd4jLong numOfElemsToReverse) {
         auto stream = context->getCudaStream();
         Nd4jLong numOfReverse = numOfElemsToReverse;
         if (numOfElemsToReverse == 0)
             numOfReverse = input->lengthOf();
-        if (input == output) {
-            reverseArrayInplaceKernel<T><<<256, 512, 8192, *stream>>>(input->specialBuffer(), input->specialShapeInfo(), numOfReverse);
-        }
-        else {
-            reverseArrayKernel<T><<<256, 512, 8192, *stream>>>(input->specialBuffer(), input->specialShapeInfo(), output->specialBuffer(), output->specialShapeInfo(), numOfReverse);
-        }
+
+        reverseArrayKernel<T><<<256, 512, 8192, *stream>>>(input->specialBuffer(), input->specialShapeInfo(), output->specialBuffer(), output->specialShapeInfo(), numOfReverse);
     }
 
 
@@ -221,7 +175,7 @@ namespace helpers {
         delete listIn;
     }
 
-BUILD_SINGLE_TEMPLATE(template void reverseArray, (nd4j::LaunchContext * context, NDArray *inArr, NDArray *outArr, int numOfElemsToReverse), LIBND4J_TYPES);
+BUILD_SINGLE_TEMPLATE(template void reverseArray, (nd4j::LaunchContext * context, NDArray *inArr, NDArray *outArr, Nd4jLong numOfElemsToReverse), LIBND4J_TYPES);
 
 }
 }
diff --git a/libnd4j/include/ops/declarable/impl/DeclarableOp.cpp b/libnd4j/include/ops/declarable/impl/DeclarableOp.cpp
index 5d29ed826..4fe28df8c 100644
--- a/libnd4j/include/ops/declarable/impl/DeclarableOp.cpp
+++ b/libnd4j/include/ops/declarable/impl/DeclarableOp.cpp
@@ -19,7 +19,6 @@
 //
 
 #include <ops/declarable/DeclarableOp.h>
-#include <helpers/ProviderRNG.h>
 #include <Status.h>
 #include <helpers/ShapeUtils.h>
 #include <NDArrayFactory.h>
@@ -190,32 +189,6 @@ namespace nd4j {
                 auto outSha = this->calculateOutputShape(&inSha, ctx);
                 results = outSha->size();
 
-                // we must "validate" our output shapes
-                /*
-                for (int e = 0; e < results; e++) {
-                    auto ptr = outSha->at(e);
-
-                    // checking for the same pointer used twice
-                    for (int i = 0; i < results; i++){
-                        if (i == e)
-                            continue;
-
-                        auto com = outSha->at(i);
-
-                        if (ptr == com)
-                            throw std::runtime_error("ShapeFunction returned same shape instance twice [" + *_descriptor->getOpName() + "]");
-                    }
-
-                    // checking for input pointer returned back
-                    for (int i = 0; i < inSha.size(); i++){
-                        auto com = inSha.at(i);
-
-                        if (ptr == com)
-                            throw std::runtime_error("ShapeFunction returned input shape instance as output [" + *_descriptor->getOpName() + "]");
-                    }
-                }
-                */
-
                 // optionally saving shapeTime
                 if (Environment::getInstance()->isProfiling() && node != nullptr) {
                     shapeEnd = std::chrono::system_clock::now();
@@ -355,75 +328,139 @@ namespace nd4j {
             // rolling over inputs first
             int cnt = 0, inT = 0;
             std::vector<nd4j::DataType> inputTypes(block.width());
-            for (auto &p: *(block.inputs())) {
-                auto var = block.variable(p);
-
-                // we're not checking validity, if ANY types were explicitly allowed
-                //if (block.dataType(cnt) == nd4j::DataType::ANY)
-                //    continue;
-
-                // only validating non-null variables
-                if (var != nullptr && var->hasNDArray()) {
-                    auto array = var->getNDArray();
-
+            if (block.isFastPath()) {
+                for (auto array: block.fastpath_in()) {
                     inputTypes[inT++] = array->dataType();
                     if (!_descriptor->checkInputMatch(cnt, array->dataType())) {
                         auto ctype = DataTypeUtils::asString(array->dataType());
-                        nd4j_printf("Op [%s] failed check for input [%i], DataType: [%s]\n", _descriptor->getOpName()->data(), cnt, ctype.c_str());
+                        nd4j_printf("Op [%s] failed check for input [%i], DataType: [%s]\n",
+                                    _descriptor->getOpName()->data(), cnt, ctype.c_str());
                         return ND4J_STATUS_BAD_ARGUMENTS;
                     }
+                    cnt++;
                 }
+            } else {
+                for (auto &p: *(block.inputs())) {
+                    auto var = block.variable(p);
 
-                cnt++;
-            }
-
-            // checking optionally available outputs
-            auto varSpace = block.getVariableSpace();
-            for (int index = 0; index < DataTypeUtils::max<int>(); index++) {
-                if (varSpace != nullptr && varSpace->hasVariable(block.nodeId(), index)) {
-                    auto var = block.variable(block.nodeId(), index);
+                    // we're not checking validity, if ANY types were explicitly allowed
+                    //if (block.dataType(cnt) == nd4j::DataType::ANY)
+                    //    continue;
 
                     // only validating non-null variables
                     if (var != nullptr && var->hasNDArray()) {
                         auto array = var->getNDArray();
-                        auto cType = array->dataType();
 
-                        if (_descriptor->isSameMode()) {
-
-                            if (index >= block.width()) {
-                                auto iv = block.variable(0);
-
-                                if (iv->getNDArray()->dataType() != cType) {
-                                    auto t = DataTypeUtils::asString(cType);
-                                    nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s]\n", _descriptor->getOpName()->data(), index, t.c_str());
-                                    return ND4J_STATUS_BAD_ARGUMENTS;
-                                }
-                            } else {
-                                // for same mode, output type must be the same as input type
-                                auto iv = block.variable(index);
-
-                                if (iv->getNDArray()->dataType() != cType) {
-                                    auto t = DataTypeUtils::asString(cType);
-                                    nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s]\n", _descriptor->getOpName()->data(), index, t.c_str());
-                                    return ND4J_STATUS_BAD_ARGUMENTS;
-                                }
-                            }
-                        } else if (_descriptor->isInherit(index)) {
-                            // in inherit mode, output type must be the same as one of input types
-                            if (std::find(inputTypes.begin(), inputTypes.end(), cType) == inputTypes.end()) {
-                                auto t = DataTypeUtils::asString(cType);
-                                nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s].\n", _descriptor->getOpName()->data(), index, t.c_str());
-                                return ND4J_STATUS_BAD_ARGUMENTS;
-                            }
-
-                        } else if (!_descriptor->checkOutputMatch(index, cType)) {
-                            auto t = DataTypeUtils::asString(cType);
-                            nd4j_printf("Op [%s] failed check for output [%i], DataType: [%i];\n", _descriptor->getOpName()->data(), index, t.c_str());
+                        inputTypes[inT++] = array->dataType();
+                        if (!_descriptor->checkInputMatch(cnt, array->dataType())) {
+                            auto ctype = DataTypeUtils::asString(array->dataType());
+                            nd4j_printf("Op [%s] failed check for input [%i], DataType: [%s]\n",
+                                        _descriptor->getOpName()->data(), cnt, ctype.c_str());
                             return ND4J_STATUS_BAD_ARGUMENTS;
                         }
                     }
-                } else
-                    break;
+
+                    cnt++;
+                }
+            }
+
+            if (block.isFastPath()) {
+                int index = 0;
+                for (auto array: block.fastpath_out()) {
+                    auto cType = array->dataType();
+
+                    if (_descriptor->isSameMode()) {
+
+                        if (index >= block.width()) {
+                            auto ia = block.fastpath_in()[0];
+
+                            if (ia->dataType() != cType) {
+                                auto t = DataTypeUtils::asString(cType);
+                                nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s]\n",
+                                            _descriptor->getOpName()->data(), index, t.c_str());
+                                return ND4J_STATUS_BAD_ARGUMENTS;
+                            }
+                        } else {
+                            // for same mode, output type must be the same as input type
+                            auto ia = block.fastpath_in()[index];
+
+                            if (ia->dataType() != cType) {
+                                auto t = DataTypeUtils::asString(cType);
+                                nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s]\n",
+                                            _descriptor->getOpName()->data(), index, t.c_str());
+                                return ND4J_STATUS_BAD_ARGUMENTS;
+                            }
+                        }
+                    } else if (_descriptor->isInherit(index)) {
+                        // in inherit mode, output type must be the same as one of input types
+                        if (std::find(inputTypes.begin(), inputTypes.end(), cType) == inputTypes.end()) {
+                            auto t = DataTypeUtils::asString(cType);
+                            nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s].\n",
+                                        _descriptor->getOpName()->data(), index, t.c_str());
+                            return ND4J_STATUS_BAD_ARGUMENTS;
+                        }
+
+                    } else if (!_descriptor->checkOutputMatch(index, cType)) {
+                        auto t = DataTypeUtils::asString(cType);
+                        nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s];\n",
+                                    _descriptor->getOpName()->data(), index, t.c_str());
+                        return ND4J_STATUS_BAD_ARGUMENTS;
+                    }
+                    index++;
+                }
+            } else {
+                // checking optionally available outputs
+                auto varSpace = block.getVariableSpace();
+                for (int index = 0; index < DataTypeUtils::max<int>(); index++) {
+                    if (varSpace != nullptr && varSpace->hasVariable(block.nodeId(), index)) {
+                        auto var = block.variable(block.nodeId(), index);
+
+                        // only validating non-null variables
+                        if (var != nullptr && var->hasNDArray()) {
+                            auto array = var->getNDArray();
+                            auto cType = array->dataType();
+
+                            if (_descriptor->isSameMode()) {
+
+                                if (index >= block.width()) {
+                                    auto iv = block.variable(0);
+
+                                    if (iv->getNDArray()->dataType() != cType) {
+                                        auto t = DataTypeUtils::asString(cType);
+                                        nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s]\n",
+                                                    _descriptor->getOpName()->data(), index, t.c_str());
+                                        return ND4J_STATUS_BAD_ARGUMENTS;
+                                    }
+                                } else {
+                                    // for same mode, output type must be the same as input type
+                                    auto iv = block.variable(index);
+
+                                    if (iv->getNDArray()->dataType() != cType) {
+                                        auto t = DataTypeUtils::asString(cType);
+                                        nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s]\n",
+                                                    _descriptor->getOpName()->data(), index, t.c_str());
+                                        return ND4J_STATUS_BAD_ARGUMENTS;
+                                    }
+                                }
+                            } else if (_descriptor->isInherit(index)) {
+                                // in inherit mode, output type must be the same as one of input types
+                                if (std::find(inputTypes.begin(), inputTypes.end(), cType) == inputTypes.end()) {
+                                    auto t = DataTypeUtils::asString(cType);
+                                    nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s].\n",
+                                                _descriptor->getOpName()->data(), index, t.c_str());
+                                    return ND4J_STATUS_BAD_ARGUMENTS;
+                                }
+
+                            } else if (!_descriptor->checkOutputMatch(index, cType)) {
+                                auto t = DataTypeUtils::asString(cType);
+                                nd4j_printf("Op [%s] failed check for output [%i], DataType: [%s];\n",
+                                            _descriptor->getOpName()->data(), index, t.c_str());
+                                return ND4J_STATUS_BAD_ARGUMENTS;
+                            }
+                        }
+                    } else
+                        break;
+                }
             }
 
 
diff --git a/libnd4j/tests_cpu/layers_tests/JavaInteropTests.cpp b/libnd4j/tests_cpu/layers_tests/JavaInteropTests.cpp
index ef3710371..d0d67000b 100644
--- a/libnd4j/tests_cpu/layers_tests/JavaInteropTests.cpp
+++ b/libnd4j/tests_cpu/layers_tests/JavaInteropTests.cpp
@@ -400,6 +400,32 @@ TEST_F(JavaInteropTests, Test_Synonyms_3) {
     ASSERT_EQ(nameRef, name);
 }
 
+TEST_F(JavaInteropTests, Test_FastPath_Validation_1) {
+    auto x = NDArrayFactory::create<int>('c', {4}, {1, 2, 3, 4});
+    auto z = NDArrayFactory::create<int>('c', {4}, {1, 2, 3, 4});
+
+    Context ctx(1);
+    ctx.setInputArray(0, x.buffer(), x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo());
+    ctx.setOutputArray(0, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo());
+
+    nd4j::ops::softmax op;
+    auto status = op.execute(&ctx);
+    ASSERT_NE(Status::OK(), status);
+}
+
+TEST_F(JavaInteropTests, Test_FastPath_Validation_2) {
+    auto x = NDArrayFactory::create<float>('c', {4}, {1.f, 2.f, 3.f, 4.f});
+    auto z = NDArrayFactory::create<int>('c', {4}, {1, 2, 3, 4});
+
+    Context ctx(1);
+    ctx.setInputArray(0, x.buffer(), x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo());
+    ctx.setOutputArray(0, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo());
+
+    nd4j::ops::softmax op;
+    auto status = op.execute(&ctx);
+    ASSERT_NE(Status::OK(), status);
+}
+
 /*
 TEST_F(JavaInteropTests, test_avgpooling_edge_1) {
     int inOutH = 35;
diff --git a/libnd4j/tests_cpu/layers_tests/NativeOpsTests.cpp b/libnd4j/tests_cpu/layers_tests/NativeOpsTests.cpp
index fe190d9bb..9aac42ddf 100644
--- a/libnd4j/tests_cpu/layers_tests/NativeOpsTests.cpp
+++ b/libnd4j/tests_cpu/layers_tests/NativeOpsTests.cpp
@@ -992,81 +992,6 @@ TEST_F(NativeOpsTests, ScalarTadTest_2) {
     ASSERT_TRUE(exp.e<bool>(5) == z.e<bool>(5) && exp.e<bool>(15));
 }
 
-TEST_F(NativeOpsTests, FlattenTest_1) {
-    auto x = NDArrayFactory::create<float>('c', {5, 5});
-    auto y = NDArrayFactory::create<float>('c', {5, 5});
-    auto exp = NDArrayFactory::create<float>('c', {2, 5,5});
-    auto z = NDArrayFactory::create<float>('c', {2, 5,5});
-
-    Nd4jPointer extra[6];
-#ifdef __CUDABLAS__
-    extra[1] = x.getContext()->getCudaStream();
-    extra[0] = extra[2] = extra[3] = extra[4] = extra[5] = nullptr;
-    x.syncToHost();
-    y.syncToHost();
-    printf("Unsupported for CUDA platform yet.\n");
-    return;
-#endif
-    x.linspace(1.0,2);
-    y.linspace(2,2);
-
-    //y.assign(2.);
-    x.syncToDevice();
-    z.syncToDevice();
-    auto dimension = NDArrayFactory::create<int>({0, 1});
-    auto dimensions = reinterpret_cast<int*>(dimension.buffer());
-    auto tadPackX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(x.shapeInfo(), dimensions, dimension.lengthOf());
-    auto tadPackZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(z.shapeInfo(), dimensions, dimension.lengthOf());
-    exp(1, {0}).linspace(1,2);
-    ::flatten(extra,
-                        25, 'c', z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(),
-                        x.buffer(), x.shapeInfo(),
-                        x.specialBuffer(), x.specialShapeInfo());
-
-//    exp.printIndexedBuffer("Exp");
-//    z.printIndexedBuffer("Flatten");
-    ASSERT_TRUE(exp.equalsTo(z));
-}
-
-TEST_F(NativeOpsTests, ConcatTest_1) {
-    auto x = NDArrayFactory::create<float>('c', {5, 5});
-    auto y = NDArrayFactory::create<float>('c', {5, 5});
-    auto exp = NDArrayFactory::create<float>('c', {10,5});
-    auto z = NDArrayFactory::create<float>('c', {10,5});
-
-    Nd4jPointer extra[6];
-#ifdef __CUDABLAS__
-    extra[1] = x.getContext()->getCudaStream();
-    extra[0] = extra[2] = extra[3] = extra[4] = extra[5] = nullptr;
-    x.syncToHost();
-    y.syncToHost();
-    printf("Unsupported for CUDA platform yet.\n");
-    return;
-#endif
-    x.linspace(1.0);
-    y.linspace(26);
-
-    //y.assign(2.);
-    x.syncToDevice();
-    z.syncToDevice();
-    int d = 0;
-    auto dimension = NDArrayFactory::create<int>('c', {1}, {d});
-    auto dimensions = reinterpret_cast<int*>(dimension.buffer());
-    //auto tadPackX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(x.shapeInfo(), dimensions, dimension.lengthOf());
-    auto tadPackZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(z.shapeInfo(), dimensions, dimension.lengthOf());
-    exp.linspace(1);
-    Nd4jPointer datas[] = {x.buffer(), y.buffer()};
-    Nd4jPointer shapes[] = {x.shapeInfo(), y.shapeInfo()};
-
-    ::concat(extra,
-                     0, 2, datas, shapes, nullptr, nullptr, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(),
-                     nullptr, nullptr);
-
-//    exp.printIndexedBuffer("Exp");
-//    z.printIndexedBuffer("Concat");
-    ASSERT_TRUE(exp.equalsTo(z));
-}
-
 TEST_F(NativeOpsTests, ConcatTest_2) {
     auto x = NDArrayFactory::create<float>('c', {5, 5});
     auto y = NDArrayFactory::create<float>('c', {5, 5});
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-native-api/src/main/java/org/nd4j/nativeblas/NativeOps.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-native-api/src/main/java/org/nd4j/nativeblas/NativeOps.java
index 174be9a7d..576cea78a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-native-api/src/main/java/org/nd4j/nativeblas/NativeOps.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-native-api/src/main/java/org/nd4j/nativeblas/NativeOps.java
@@ -557,43 +557,6 @@ public interface NativeOps {
                                         @Cast("Nd4jLong *") LongPointer tadShapeInfo, @Cast("Nd4jLong *") LongPointer tadOffsets,
                                         @Cast("Nd4jLong *") LongPointer tadShapeInfoZ, @Cast("Nd4jLong *") LongPointer tadOffsetsZ);
 
-    /**
-     * @param extraPointers
-     * @param offset
-     * @param order
-     * @param results
-     * @param resultShapeInfo
-     * @param input
-     * @param inputShapeInfo
-     */
-    void flatten(PointerPointer extraPointers,
-                                 int offset,
-                                 char order,
-                                 Pointer results, @Cast("Nd4jLong *") LongPointer resultShapeInfo,
-                                 Pointer dresults, @Cast("Nd4jLong *") LongPointer dresultShapeInfo,
-                                 Pointer input, @Cast("Nd4jLong *") LongPointer inputShapeInfo,
-                                 Pointer dinput, @Cast("Nd4jLong *") LongPointer dinputShapeInfo);
-
-    /**
-     * @param extraPointers
-     * @param dimension
-     * @param numArrays
-     * @param data
-     * @param inputShapeInfo
-     * @param results
-     * @param resultShapeInfo
-     * @param tadPointers
-     * @param tadOffsets
-     */
-    void concat(PointerPointer extraPointers,
-                                int dimension,
-                                int numArrays,
-                                PointerPointer data, PointerPointer inputShapeInfo,
-                                PointerPointer ddata, PointerPointer dinputShapeInfo,
-                                Pointer results, @Cast("Nd4jLong *") LongPointer resultShapeInfo,
-                                Pointer dresults, @Cast("Nd4jLong *") LongPointer dresultShapeInfo,
-                                PointerPointer tadPointers,
-                                PointerPointer tadOffsets);
 
     void specialConcat(PointerPointer extraPointers,
                                        int dimension,
@@ -1185,4 +1148,7 @@ public interface NativeOps {
     Pointer lcCopyStream(OpaqueLaunchContext lc);
     Pointer lcBlasHandle(OpaqueLaunchContext lc);
     Pointer lcSolverHandle(OpaqueLaunchContext lc);
+
+    int lastErrorCode();
+    String lastErrorMessage();
 }
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/pointers/cuda/cudaEvent_t.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/pointers/cuda/cudaEvent_t.java
index 1650e08ac..52b7d7332 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/pointers/cuda/cudaEvent_t.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/pointers/cuda/cudaEvent_t.java
@@ -22,6 +22,7 @@ import org.bytedeco.javacpp.Pointer;
 import org.nd4j.jita.allocator.pointers.CudaPointer;
 import org.nd4j.linalg.exception.ND4JException;
 import org.nd4j.linalg.factory.Nd4j;
+import org.nd4j.nativeblas.NativeOps;
 import org.nd4j.nativeblas.NativeOpsHolder;
 
 /**
@@ -67,14 +68,18 @@ public class cudaEvent_t extends CudaPointer {
             int res = NativeOpsHolder.getInstance().getDeviceNativeOps().eventSynchronize(this);
             if (res == 0)
                 throw new ND4JException("CUDA exception happened. Terminating. Last op: [" + Nd4j.getExecutioner().getLastOp() +"]");
+
+            if (NativeOpsHolder.getInstance().getDeviceNativeOps().lastErrorCode() != 0)
+                throw new RuntimeException(NativeOpsHolder.getInstance().getDeviceNativeOps().lastErrorMessage());
         }
     }
 
     public void register(cudaStream_t stream) {
         if (!isDestroyed()) {
             int res = NativeOpsHolder.getInstance().getDeviceNativeOps().registerEvent(this, stream);
-            if (res == 0)
-                throw new ND4JException("CUDA exception happened. Terminating. Last op: [" + Nd4j.getExecutioner().getLastOp() +"]");
+
+            if (NativeOpsHolder.getInstance().getDeviceNativeOps().lastErrorCode() != 0)
+                throw new RuntimeException(NativeOpsHolder.getInstance().getDeviceNativeOps().lastErrorMessage());
         }
     }
 }
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/pointers/cuda/cudaStream_t.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/pointers/cuda/cudaStream_t.java
index b18ceb2fa..8d78ee950 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/pointers/cuda/cudaStream_t.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/pointers/cuda/cudaStream_t.java
@@ -36,8 +36,9 @@ public class cudaStream_t extends CudaPointer {
     public int synchronize() {
         NativeOps nativeOps = NativeOpsHolder.getInstance().getDeviceNativeOps();
         int res = nativeOps.streamSynchronize(this);
-        if (res == 0)
-            throw new ND4JException("CUDA exception happened. Terminating. Last op: [" + Nd4j.getExecutioner().getLastOp() +"]");
+
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         return res;
     }
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/JCublasNDArrayFactory.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/JCublasNDArrayFactory.java
index 44c361d87..9e9dc34b2 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/JCublasNDArrayFactory.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/JCublasNDArrayFactory.java
@@ -24,6 +24,7 @@ import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.buffer.DataTypeEx;
 import org.nd4j.linalg.api.buffer.Utf8Buffer;
 import org.nd4j.linalg.api.memory.enums.MemoryKind;
+import org.nd4j.linalg.api.ops.custom.Flatten;
 import org.nd4j.linalg.api.ops.impl.shape.Concat;
 import org.nd4j.linalg.api.ops.performance.PerformanceTracker;
 import org.nd4j.linalg.api.shape.options.ArrayOptionsHelper;
@@ -104,6 +105,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
         functions.put(11, Loader.addressof("cusolverDnSgesvd"));
         functions.put(12, Loader.addressof("cusolverDnDgesvd"));
         nativeOps.initializeFunctions(functions);
+
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
     }
 
     @Override
@@ -335,75 +339,7 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
         if (Nd4j.getExecutioner() instanceof GridExecutioner)
             ((GridExecutioner) Nd4j.getExecutioner()).flushQueue();
 
-        int length = 0;
-        DataType t = null;
-        for (INDArray m : matrices) {
-            length += m.length();
-            if (t == null)
-                t = m.dataType();
-
-            Preconditions.checkArgument(t == m.dataType(), "Arrays must have same data type");
-        }
-
-        INDArray ret = Nd4j.create(t, new long[] {length}, order);
-        int linearIndex = 0;
-
-        AtomicAllocator allocator = AtomicAllocator.getInstance();
-
-
-        for (INDArray m : matrices) {
-            if (m.isEmpty())
-                continue;
-
-            CudaContext context = allocator.getFlowController().prepareAction(ret, m);
-
-            if (m.ordering() == order && ret.elementWiseStride() == m.elementWiseStride()
-                            && ret.elementWiseStride() == 1) {
-                // do memcpy in proper direction and forget about that
-                // FIXME: get rid of this
-                ((BaseCudaDataBuffer) m.data()).lazyAllocateHostPointer();
-                allocator.memcpyAsync(ret.data(), new CudaPointer(allocator.getHostPointer(m).address()),
-                                AllocationUtils.getRequiredMemory(AllocationUtils.buildAllocationShape(m)),
-                                linearIndex * (m.data().dataType() == DataType.DOUBLE ? 8
-                                                : m.data().dataType() == DataType.FLOAT ? 4 : 2));
-                linearIndex += m.length();
-            } else {
-                Pointer hostYShapeInfo = AddressRetriever.retrieveHostPointer(m.shapeInfoDataBuffer());
-
-                PointerPointer extras = new PointerPointer(
-                                AddressRetriever.retrieveHostPointer(ret.shapeInfoDataBuffer()), context.getOldStream(),
-                                allocator.getDeviceIdPointer(), null,
-                                context.getBufferReduction(), context.getBufferScalar(), null,
-                                hostYShapeInfo, AddressRetriever.retrieveHostPointer(ret.shapeInfoDataBuffer()));
-
-
-                nativeOps.flatten(extras, linearIndex, order,
-                                     null,
-                                    (LongPointer) allocator.getHostPointer(ret.shapeInfoDataBuffer()),
-                                     allocator.getPointer(ret, context),
-                                    (LongPointer) allocator.getPointer(ret.shapeInfoDataBuffer(), context),
-                                     null,
-                                    (LongPointer) allocator.getHostPointer(m.shapeInfoDataBuffer()),
-                                     allocator.getPointer(m, context),
-                                    (LongPointer) allocator.getPointer(m.shapeInfoDataBuffer(), context));
-
-
-
-
-                //Works for all cases...
-
-                /* NdIndexIterator iter = new NdIndexIterator(order, m.shape());
-                while (iter.hasNext()) {
-                    ret.putScalar(linearIndex++, m.getDouble(iter.next()));
-                }*/
-
-                linearIndex += m.length();
-            }
-
-            if (ret != null)
-                allocator.registerAction(context, ret, m);
-        }
-        return ret;
+        return Nd4j.exec(new Flatten(order, matrices.toArray(new INDArray[0])))[0];
     }
 
     @Override
@@ -412,131 +348,6 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
             ((GridExecutioner) Nd4j.getExecutioner()).flushQueue();
 
         return Nd4j.exec(new Concat(dimension, toConcat))[0];
-
-        // legacy implementation
-/*
-        boolean allScalars = true;
-
-        var outputShape = ArrayUtil.copy(toConcat[0].shape());
-
-        if (toConcat.length == 1)
-            return toConcat[0];
-
-        int sumAlongDim = 0;
-        for (int i = 0; i < toConcat.length; i++) {
-            if (toConcat[i].isCompressed())
-                Nd4j.getCompressor().decompressi(toConcat[i]);
-
-            allScalars &= toConcat[i].rank() == 0;
-
-            sumAlongDim += toConcat[i].size(dimension);
-        }
-
-        if (allScalars) {
-            outputShape = new long[]{sumAlongDim};
-        } else {
-            outputShape[dimension] = sumAlongDim;
-        }
-
-        INDArray ret = Nd4j.createUninitialized(toConcat[0].dataType(), outputShape, Nd4j.order());
-
-        AtomicAllocator allocator = AtomicAllocator.getInstance();
-
-        CudaContext context = allocator.getFlowController().prepareAction(ret, toConcat);
-
-        val shapeInfoPointers = new long[toConcat.length];
-        val dataPointers = new long[toConcat.length];
-        val tadPointers = new long[toConcat.length];
-        val offsetsPointers = new long[toConcat.length];
-        val hostShapeInfoPointers = new long[toConcat.length];
-
-        TADManager tadManager = Nd4j.getExecutioner().getTADManager();
-        for (int i = 0; i < toConcat.length; i++) {
-            shapeInfoPointers[i] = AddressRetriever.retrieveDeviceAddress(toConcat[i].shapeInfoDataBuffer(), context);
-            dataPointers[i] = AtomicAllocator.getInstance().getPointer(toConcat[i], context).address();
-            hostShapeInfoPointers[i] = AtomicAllocator.getInstance().getHostPointer(toConcat[i].shapeInfoDataBuffer()).address();
-
-            sumAlongDim += toConcat[i].size(dimension);
-            for (int j = 0; j < toConcat[i].rank(); j++)
-                if (j != dimension && toConcat[i].size(j) != outputShape[j]) {
-                    throw new IllegalArgumentException(
-                                    "Illegal concatenation at array " + i + " and shape element " + j);
-                }
-
-            if (!allScalars) {
-                val tadBuffers = tadManager.getTADOnlyShapeInfo(toConcat[i], new int[]{dimension});
-
-                long devTadShapeInfo = AtomicAllocator.getInstance().getPointer(tadBuffers.getFirst(), context).address();
-
-                val offsets = tadBuffers.getSecond();
-                long devTadOffsets = AtomicAllocator.getInstance().getPointer(offsets, context).address();
-
-                tadPointers[i] = devTadShapeInfo;
-                offsetsPointers[i] = devTadOffsets;
-            }
-        }
-
-        // getting tadOnlyShape for result
-        val zBuffers = tadManager.getTADOnlyShapeInfo(ret, new int[] {dimension});
-        val hostPointers = new LongPointer(hostShapeInfoPointers);
-        val hosthost = new PointerPointerWrapper(hostPointers);
-
-        //System.out.println("shapePointers: " + Arrays.toString(shapeInfoPointers));
-
-        val dZ = AtomicAllocator.getInstance().getPointer(ret, context);
-        val dZShapeInfo = AddressRetriever.retrieveDevicePointer(ret.shapeInfoDataBuffer(), context);
-
-
-
-        //val tempData = new CudaDoubleDataBuffer(toConcat.length);
-        //val tempShapes = new CudaDoubleDataBuffer(toConcat.length);
-        //val tempTAD = new CudaDoubleDataBuffer(toConcat.length);
-        //val tempOffsets = new CudaDoubleDataBuffer(toConcat.length);
-
-        //AtomicAllocator.getInstance().memcpyBlocking(tempData, new LongPointer(dataPointers), dataPointers.length * 8,0);
-        //AtomicAllocator.getInstance().memcpyBlocking(tempShapes, new LongPointer(shapeInfoPointers), shapeInfoPointers.length * 8, 0);
-        //AtomicAllocator.getInstance().memcpyBlocking(tempTAD, new LongPointer(tadPointers), tadPointers.length * 8, 0);
-        //AtomicAllocator.getInstance().memcpyBlocking(tempOffsets, new LongPointer(offsetsPointers), offsetsPointers.length * 8, 0);
-
-        val dataPointer = new PointerPointerWrapper(new LongPointer(dataPointers)); //AtomicAllocator.getInstance().getPointer(tempData, context);
-        val shapesPointer = new PointerPointerWrapper(new LongPointer(shapeInfoPointers));//AtomicAllocator.getInstance().getPointer(tempShapes, context);
-        //val tadPointer = AtomicAllocator.getInstance().getPointer(tempTAD, context);
-        //val offsetPointer = AtomicAllocator.getInstance().getPointer(tempOffsets, context);
-
-
-        // System.out.println("ShapesPointer after conversion: " + shapesPointer);
-
-        val extras = new PointerPointer(AddressRetriever.retrieveHostPointer(ret.shapeInfoDataBuffer()),
-                        context.getOldStream(), allocator.getDeviceIdPointer(), null,
-                        context.getBufferReduction(), context.getBufferScalar(), null,
-                        AddressRetriever.retrieveHostPointer(toConcat[0].shapeInfoDataBuffer()),
-                        AddressRetriever.retrieveHostPointer(ret.shapeInfoDataBuffer()),
-                        new LongPointer(hostShapeInfoPointers),
-                        AtomicAllocator.getInstance().getPointer(zBuffers.getFirst(), context), // getting zTADShape
-                        AtomicAllocator.getInstance().getPointer(zBuffers.getSecond(), context) // getting zOffset
-        );
-
-
-        nativeOps.concat(extras,
-                dimension,
-                toConcat.length,
-                null,
-                hosthost,
-                dataPointer,
-                shapesPointer,
-                null,
-                (LongPointer) ret.shapeInfoDataBuffer().addressPointer(),
-                dZ,
-                (LongPointer) dZShapeInfo,
-                 null,
-                null);
-
-
-        allocator.registerAction(context, ret, toConcat);
-
-        return ret;
-        //return super.concat(dimension, toConcat);
-        */
     }
 
 
@@ -590,6 +401,8 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                     (LongPointer) ret.shapeInfoDataBuffer().addressPointer(),
                     null, null);
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         AllocationPoint point = allocator.getAllocationPoint(ret);
 
@@ -598,6 +411,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
         nativeOps.memcpyAsync(point.getDevicePointer(), point.getHostPointer(), ret.lengthLong() * Nd4j.sizeOfDataType(ret.data().dataType()), CudaConstants.cudaMemcpyHostToDevice, context.getSpecialStream());
         context.getSpecialStream().synchronize();
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         PerformanceTracker.getInstance().helperRegisterTransaction(point.getDeviceId(), perfD, point.getNumberOfBytes(), MemcpyDirection.HOST_TO_DEVICE);
 
         point.tickHostRead();
@@ -729,6 +545,8 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                 (LongPointer) zTadShapeInfo,
                 new LongPointerWrapper(zTadOffsets));
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         allocator.registerAction(context, ret, source);
 
@@ -743,7 +561,6 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
             return target.assign(arrays[0]);
 
         // we do averaging on GPU only if ALL devices have p2p links
-        //if (CudaEnvironment.getInstance().getConfiguration().isCrossDeviceAccessAllowed() && nativeOps.isP2PAvailable()) {
         if (true) {
             Nd4j.getExecutioner().push();
 
@@ -781,6 +598,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
 
             nativeOps.accumulate(extras, null, (LongPointer) arrays[0].shapeInfoDataBuffer().addressPointer(), x, null, null, (LongPointer)  allocator.getHostPointer(target.shapeInfoDataBuffer()) , z, (LongPointer)  allocator.getPointer(target.shapeInfoDataBuffer()), arrays.length, len);
 
+            if (nativeOps.lastErrorCode() != 0)
+                throw new RuntimeException(nativeOps.lastErrorMessage());
+
             allocator.getFlowController().registerAction(context, target, arrays);
 
             return target;
@@ -824,6 +644,8 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                     arrays.length,
                     len);
 
+            if (nativeOps.lastErrorCode() != 0)
+                throw new RuntimeException(nativeOps.lastErrorMessage());
 
             AtomicAllocator.getInstance().getAllocationPoint(target).tickHostWrite();
 
@@ -895,6 +717,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                     arrays.length,
                     len, true);
 
+            if (nativeOps.lastErrorCode() != 0)
+                throw new RuntimeException(nativeOps.lastErrorMessage());
+
             allocator.getFlowController().registerAction(context, target, arrays);
 
             return target;
@@ -940,6 +765,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                     arrays.length,
                     len, true);
 
+            if (nativeOps.lastErrorCode() != 0)
+                throw new RuntimeException(nativeOps.lastErrorMessage());
+
             if (target != null)
                 AtomicAllocator.getInstance().getAllocationPoint(target).tickHostWrite();
 
@@ -1115,6 +943,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                             (IntPointer) shuffleMap, new PointerPointer(allocator.getPointer(tempTAD, context)),
                             new PointerPointer(allocator.getPointer(tempOffsets, context)));
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         for (int f = 0; f < arrays.size(); f++) {
             allocator.getFlowController().registerAction(context, arrays.get(f));
         }
@@ -1260,6 +1091,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
         val p = new PointerPointer<>(new Pointer[]{null, stream});
 
         nativeOps.convertTypes(p, typeSrc.ordinal(), source, length, typeDst.ordinal(), target);
+
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
     }
 
     @Override
@@ -1277,7 +1111,13 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
             srcPtr = nativeOps.mallocDevice(ssize, 0, 0);
             dstPtr = nativeOps.mallocDevice(size, 0, 0);
 
+            if (nativeOps.lastErrorCode() != 0)
+                throw new RuntimeException(nativeOps.lastErrorMessage());
+
             nativeOps.memcpyAsync(srcPtr, source, ssize, CudaConstants.cudaMemcpyHostToDevice, stream);
+
+            if (nativeOps.lastErrorCode() != 0)
+                throw new RuntimeException(nativeOps.lastErrorMessage());
         } else {
             // decompressing
             throw new UnsupportedOperationException();
@@ -1288,9 +1128,15 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
 
         stream.synchronize();
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         if (buffer instanceof CompressedDataBuffer) {
             nativeOps.freeDevice(srcPtr, 0);
             nativeOps.freeDevice(dstPtr, 0);
+
+            if (nativeOps.lastErrorCode() != 0)
+                throw new RuntimeException(nativeOps.lastErrorMessage());
         }
     }
 
@@ -1309,13 +1155,15 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                 val size = ((CompressedDataBuffer) source).getCompressionDescriptor().getCompressedLength();
                 srcPtr = ws.alloc(size, MemoryKind.DEVICE, DataType.HALF, false);
                 nativeOps.memcpyAsync(srcPtr, source.addressPointer(), size, CudaConstants.cudaMemcpyHostToHost, stream);
+
+                if (nativeOps.lastErrorCode() != 0)
+                    throw new RuntimeException(nativeOps.lastErrorMessage());
             }
 
             // if true - we're compressing into host memory
             if (target instanceof CompressedDataBuffer) {
                 val size = ((CompressedDataBuffer) target).getCompressionDescriptor().getCompressedLength();
                 dstPtr = ws.alloc(size, MemoryKind.DEVICE, DataType.HALF, false);
-                //nativeOps.memcpyAsync(dstPtr, target.addressPointer(), size, CudaConstants.cudaMemcpyHostToHost, stream);
             }
         } else {
             // if true - we're decompressing from host memory
@@ -1325,6 +1173,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                 srcPtr = nativeOps.mallocDevice(size, 0, 0);
                 nativeOps.memcpyAsync(srcPtr, source.addressPointer(), size, CudaConstants.cudaMemcpyHostToHost, stream);
                 stream.synchronize();
+
+                if (nativeOps.lastErrorCode() != 0)
+                    throw new RuntimeException(nativeOps.lastErrorMessage());
             } else
                 srcPtr = AtomicAllocator.getInstance().getPointer(source);
 
@@ -1333,8 +1184,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                 log.info("Replacing target ptr");
                 val size = ((CompressedDataBuffer) target).getCompressionDescriptor().getCompressedLength();
                 dstPtr = nativeOps.mallocDevice(size, 0, 0);
-                //nativeOps.memcpyAsync(dstPtr, source.addressPointer(), size, CudaConstants.cudaMemcpyHostToHost, stream);
-                //stream.synchronize();
+
+                if (nativeOps.lastErrorCode() != 0)
+                    throw new RuntimeException(nativeOps.lastErrorMessage());
             } else
                 dstPtr = AtomicAllocator.getInstance().getPointer(target);
         }
@@ -1342,6 +1194,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
 
         convertDataEx(typeSrc, srcPtr, typeDst, dstPtr, target.length());
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         Nd4j.getExecutioner().commit();
 
 
@@ -1364,6 +1219,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
 
         }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         Nd4j.getExecutioner().commit();
     }
 
@@ -1462,6 +1320,9 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                     new LongPointerWrapper(AtomicAllocator.getInstance().getPointer(tadBuffers.getSecond(), context))
             );
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         AtomicAllocator.getInstance().getFlowController().registerActionAllWrite(context, result);
         AtomicAllocator.getInstance().getFlowController().registerAction(context,null, result);
 
@@ -1517,6 +1378,8 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                     descending
             );
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         AtomicAllocator.getInstance().getFlowController().registerAction(context, x);
 
@@ -1565,6 +1428,8 @@ public class JCublasNDArrayFactory extends BaseNativeNDArrayFactory {
                     descending
             );
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         AtomicAllocator.getInstance().getFlowController().registerAction(context, x);
 
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
index 789f0f1a3..38a1ba382 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
@@ -207,6 +207,10 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                 throw new UnsupportedOperationException("Unknown op type: " + op.getOpType());
         }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
+
         AtomicAllocator.getInstance().registerAction(context, op.z(), op.x(), op.y());
 
         profilingConfigurableHookOut(op, st);
@@ -461,6 +465,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                 }
             }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         profilingConfigurableHookOut(op, st);
 
         return op.z();
@@ -619,7 +626,8 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                 AtomicAllocator.getInstance().getPointer(op.dimensions(), context),
                 null);
 
-
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         AtomicAllocator.getInstance().registerAction(context, op.z(), op.x(), op.y());
 
@@ -777,6 +785,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                 throw new UnsupportedOperationException("Unknown opType: " + op.getOpType());
         }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         AtomicAllocator.getInstance().registerAction(context, op.z(), op.x(), op.y());
 
         profilingConfigurableHookOut(op, st);
@@ -868,6 +879,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
             AtomicAllocator.getInstance().registerAction(context, null, op.x(), op.y());
         }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         profilingConfigurableHookOut(op, st);
 
         return null;
@@ -1105,6 +1119,8 @@ public class CudaExecutioner extends DefaultOpExecutioner {
             AtomicAllocator.getInstance().registerAction(context, op.z(), op.x(), op.y());
         }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         profilingConfigurableHookOut(op, st);
 
@@ -1194,6 +1210,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                 throw new UnsupportedOperationException();
         }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         AtomicAllocator.getInstance().getFlowController().registerAction(context, op.z(), op.x(), op.y());
 
         profilingConfigurableHookOut(op, st);
@@ -1268,6 +1287,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                 throw new UnsupportedOperationException("Unknown op type: " + op.getOpType());
         }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         AtomicAllocator.getInstance().registerAction(context, op.z(), op.x(), op.scalar());
 
         profilingConfigurableHookOut(op, st);
@@ -1423,6 +1445,8 @@ public class CudaExecutioner extends DefaultOpExecutioner {
             }
         }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         AtomicAllocator.getInstance().registerAction(context, op.z(), op.x(), op.y());
 
@@ -1582,6 +1606,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                     batch.getSample().maxIndexArguments(), batch.getSample().maxRealArguments(),
                     AtomicAllocator.getInstance().getPointer(surfaceBuffer, context), FlatBuffersMapper.getDataTypeAsByte(dataType));
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         surfacePoint.tickHostWrite();
     }
 
@@ -1676,6 +1703,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                     numIndexArguments, iPtr, numIntArrays,
                     AtomicAllocator.getInstance().getPointer(realsBuffer.data(), context),
                     numRealArguments, FlatBuffersMapper.getDataTypeAsByte(dataType));
+
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
     }
 
     /**
@@ -1739,6 +1769,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                          AtomicAllocator.getInstance().getPointer(op.extraArgsDataBuff(op.z().dataType()), context));
         }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         AtomicAllocator.getInstance().getFlowController().registerAction(context, op.z(), op.x(), op.y());
 
         profilingConfigurableHookOut(op, st);
@@ -1969,6 +2002,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
 
         nativeOps.decodeThreshold(extras, AtomicAllocator.getInstance().getPointer(buffer), compressedLength, AtomicAllocator.getInstance().getPointer(result), (LongPointer) AtomicAllocator.getInstance().getHostPointer(target.shapeInfoDataBuffer()));
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         AtomicAllocator.getInstance().getAllocationPoint(result).tickDeviceWrite();
 
         return target;
@@ -2013,7 +2049,8 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                     (IntPointer) AtomicAllocator.getInstance().getPointer(buffer, context),
                     (float) threshold);
 
-
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         AtomicAllocator.getInstance().getFlowController().registerAction(context, indArray);
 
@@ -2039,6 +2076,8 @@ public class CudaExecutioner extends DefaultOpExecutioner {
 
         nativeOps.decodeBitmap(extras, AtomicAllocator.getInstance().getPointer(encoded.data(), context), target.lengthLong(), AtomicAllocator.getInstance().getPointer(target, context), (LongPointer) AtomicAllocator.getInstance().getHostPointer(target.shapeInfoDataBuffer()));
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         AtomicAllocator.getInstance().getFlowController().registerAction(context, target);
 
@@ -2151,6 +2190,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
 
         OpaqueShapeList ptrptr = nativeOps.calculateOutputShapes2(null, hash, inputBuffers, inputShapes, op.inputArguments().length, tArgs, op.tArgs().length, iArgs, op.iArgs().length, bArgs, op.numBArguments());
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         if (ptrptr == null)
             throw new RuntimeException();
 
@@ -2221,109 +2263,6 @@ public class CudaExecutioner extends DefaultOpExecutioner {
         } catch (Exception e) {
             throw new RuntimeException("Op [" + name + "] execution failed", e);
         }
-
-        /*
-        long st = profilingConfigurableHookIn(op);
-
-        CudaContext context =(CudaContext) AtomicAllocator.getInstance().getDeviceContext().getContext();
-        //AtomicAllocator.getInstance().getFlowController().prepareActionAllWrite(op.outputArguments());
-
-        if (extraz.get() == null)
-            extraz.set(new PointerPointer(32));
-
-
-        PointerPointer extras = extraz.get().put(
-                new CudaPointer(1),
-                context.getOldStream(),
-                context.getBufferScalar(),
-                context.getBufferReduction());
-
-        val outputArgs = op.outputArguments();
-        val inputArgs = op.inputArguments();
-
-        if (outputArgs.length == 0 && !op.isInplaceCall())
-            throw new ND4JIllegalStateException("You can't execute non-inplace CustomOp without outputs being specified");
-
-        val lc = op.opName().toLowerCase();
-        val hash = op.opHash();
-
-
-        val inputShapes = new PointerPointer<>(inputArgs.length * 2);
-        val inputBuffers = new PointerPointer<>(inputArgs.length * 2);
-
-        int cnt= 0;
-        for (val in: inputArgs) {
-            val hp = AtomicAllocator.getInstance().getHostPointer(in.shapeInfoDataBuffer());
-            inputBuffers.put(cnt,  AtomicAllocator.getInstance().getHostPointer(in));
-            inputShapes.put(cnt, hp);
-
-
-            val dp = AtomicAllocator.getInstance().getPointer(in.shapeInfoDataBuffer(), context);
-
-            inputBuffers.put(cnt + inputArgs.length, AtomicAllocator.getInstance().getPointer(in, context));
-            inputShapes.put(cnt+ inputArgs.length, dp);
-
-            if (op.isInplaceCall()) {
-                val ap = AtomicAllocator.getInstance().getAllocationPoint(in);
-                if (ap != null)
-                    ap.tickHostWrite();
-            }
-
-            cnt++;
-        }
-
-
-        val outputShapes = new PointerPointer<>(outputArgs.length * 2);
-        val outputBuffers = new PointerPointer<>(outputArgs.length * 2);
-
-        cnt= 0;
-        for (val out: outputArgs) {
-            outputBuffers.put(cnt,  AtomicAllocator.getInstance().getHostPointer(out));
-            outputShapes.put(cnt,  AtomicAllocator.getInstance().getHostPointer(out.shapeInfoDataBuffer()));
-
-            outputBuffers.put(cnt + outputArgs.length,  AtomicAllocator.getInstance().getPointer(out, context));
-            outputShapes.put(cnt + outputArgs.length,  AtomicAllocator.getInstance().getPointer(out.shapeInfoDataBuffer(), context));
-
-            val ap = AtomicAllocator.getInstance().getAllocationPoint(out);
-
-            if (ap != null)
-                ap.tickHostWrite();
-
-            cnt++;
-        }
-
-        val iArgs = op.iArgs().length > 0 ? new LongPointer(op.iArgs().length) : null;
-
-        cnt = 0;
-        for (val i: op.iArgs())
-            iArgs.put(cnt++, i);
-
-
-        val tArgs = op.tArgs().length > 0 ? new DoublePointer(op.tArgs().length) : null;
-
-        val bArgs = op.bArgs().length > 0 ? new BooleanPointer(op.numBArguments()) : null;
-
-        cnt = 0;
-        for (val t: op.tArgs())
-            tArgs.put(cnt++, t);
-
-        cnt = 0;
-        for (val b: op.bArgs())
-            bArgs.put(cnt++, b);
-
-        try {
-            val status = OpStatus.byNumber(nativeOps.execCustomOp(extras, hash, inputBuffers, inputShapes, inputArgs.length, outputBuffers, outputShapes, outputArgs.length, tArgs, op.tArgs().length, iArgs, op.iArgs().length, bArgs, op.numBArguments(), op.isInplaceCall()));
-            if (status != OpStatus.ND4J_STATUS_OK)
-                throw new ND4JIllegalStateException("Op execution failed: " + status);
-        } catch (Exception e) {
-            throw new RuntimeException("Op [" + op.opName() + "] execution failed");
-        }
-
-        //AtomicAllocator.getInstance().getFlowController().prepareActionAllWrite(op.outputArguments());
-
-        profilingConfigurableHookOut(op, st);
-        return op.outputArguments();
-         */
     }
 
     @Override
@@ -2341,6 +2280,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
     @Override
     public void registerGraph(long id, Pointer graph) {
         nativeOps.registerGraph(null, id, graph);
+
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
     }
 
     @Override
@@ -2368,6 +2310,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
 
         OpaqueVariablesSet result = nativeOps.executeStoredGraph(null, id, ptrBuffers, ptrShapes, ptrIndices, map.size());
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         OpStatus status = OpStatus.byNumber(nativeOps.getVariablesSetStatus(result));
 
         if (status != OpStatus.ND4J_STATUS_OK)
@@ -2398,6 +2343,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
             newMap.put(nodeName, array);
         }
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         nativeOps.deleteVariablesSet(result);
 
         return newMap;
@@ -2406,6 +2354,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
     @Override
     public void forgetGraph(long id) {
         nativeOps.unregisterGraph(null, id);
+
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
     }
 
     /**
@@ -2474,6 +2425,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                 null, (LongPointer) AtomicAllocator.getInstance().getHostPointer(tadY.getFirst()), null, AtomicAllocator.getInstance().getPointer(updates, context), (LongPointer) AtomicAllocator.getInstance().getPointer(tadY.getFirst()), (LongPointer) AtomicAllocator.getInstance().getPointer(tadY.getSecond()),
                 null, (IntPointer) AtomicAllocator.getInstance().getPointer(indices, context));
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         AtomicAllocator.getInstance().getFlowController().registerAction(context, array, indices, updates);
     }
 
@@ -2490,9 +2444,14 @@ public class CudaExecutioner extends DefaultOpExecutioner {
         ((CudaOpContext) context).setCudaStream(ctx.getOldStream(), ctx.getBufferReduction(), ctx.getBufferAllocation());
 
         val status = nativeOps.execCustomOp2(null, op.opHash(), context.contextPointer());
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         if (status != 0)
             throw new RuntimeException("Op [" + op.opName() + "] execution failed");
 
+
+
         for (val arr:op.outputArguments())
             AtomicAllocator.getInstance().registerAction(ctx, arr);
 
@@ -2527,6 +2486,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
 
         nativeOps.inspectArray(extras, AtomicAllocator.getInstance().getHostPointer(array), (LongPointer) AtomicAllocator.getInstance().getHostPointer(array.shapeInfoDataBuffer()), AtomicAllocator.getInstance().getPointer(array, ctx), (LongPointer) AtomicAllocator.getInstance().getPointer(array.shapeInfoDataBuffer()), debugInfo);
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         return INDArrayStatistics.builder()
                 .minValue(debugInfo._minValue())
                 .maxValue(debugInfo._maxValue())
@@ -2545,6 +2507,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
     public DataBuffer createShapeInfo(long[] shape, long[] stride, long elementWiseStride, char order, DataType dtype, boolean empty) {
         OpaqueConstantDataBuffer dbf = nativeOps.shapeBuffer(shape.length, new LongPointer(shape), new LongPointer(stride), dtype.toInt(), order, elementWiseStride, empty);
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         val result = new CudaLongDataBuffer(nativeOps.getConstantDataBufferPrimary(dbf), nativeOps.getConstantDataBufferSpecial(dbf), Shape.shapeInfoLength(shape.length));
 
         nativeOps.deleteShapeBuffer(dbf);
@@ -2556,6 +2521,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
     public TadPack tadShapeInfoAndOffsets(INDArray array, int[] dimension) {
         OpaqueTadPack pack = nativeOps.tadOnlyShapeInfo((LongPointer) array.shapeInfoDataBuffer().addressPointer(), new IntPointer(dimension), dimension.length);
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         val tadShape = new CudaLongDataBuffer(nativeOps.getPrimaryShapeInfo(pack), nativeOps.getSpecialShapeInfo(pack), nativeOps.getShapeInfoLength(pack));
         val tadOffsets = new CudaLongDataBuffer(nativeOps.getPrimaryOffsets(pack), nativeOps.getSpecialOffsets(pack), nativeOps.getNumberOfTads(pack));
 
@@ -2568,6 +2536,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
     public DataBuffer createConstantBuffer(long[] values, DataType desiredType) {
         OpaqueConstantDataBuffer dbf = nativeOps.constantBufferLong(desiredType.toInt(), new LongPointer(values), values.length);
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         val buffer = Nd4j.createBuffer(nativeOps.getConstantDataBufferPrimary(dbf), nativeOps.getConstantDataBufferSpecial(dbf), values.length, desiredType);
         buffer.setConstant(true);
 
@@ -2578,6 +2549,9 @@ public class CudaExecutioner extends DefaultOpExecutioner {
     public DataBuffer createConstantBuffer(double[] values, DataType desiredType)  {
         OpaqueConstantDataBuffer dbf = nativeOps.constantBufferDouble(desiredType.toInt(), new DoublePointer(values), values.length);
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         val buffer = Nd4j.createBuffer(nativeOps.getConstantDataBufferPrimary(dbf), nativeOps.getConstantDataBufferSpecial(dbf), values.length, desiredType);
         buffer.setConstant(true);
 
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java
index 47cfa2584..603413fd6 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java
@@ -449,6 +449,60 @@ public class Nd4jCuda extends org.nd4j.nativeblas.Nd4jCudaHelper {
 // #endif //DEV_TESTS_TADPACK_H
 
 
+// Parsed from execution/ErrorReference.h
+
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+
+// #ifndef DEV_TESTS_ERRORREFERENCE_H
+// #define DEV_TESTS_ERRORREFERENCE_H
+
+// #include <string>
+// #include <dll.h>
+    @Namespace("sd") @NoOffset public static class ErrorReference extends Pointer {
+        static { Loader.load(); }
+        /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */
+        public ErrorReference(Pointer p) { super(p); }
+        /** Native array allocator. Access with {@link Pointer#position(long)}. */
+        public ErrorReference(long size) { super((Pointer)null); allocateArray(size); }
+        private native void allocateArray(long size);
+        @Override public ErrorReference position(long position) {
+            return (ErrorReference)super.position(position);
+        }
+    
+        public ErrorReference() { super((Pointer)null); allocate(); }
+        private native void allocate();
+
+        public native int errorCode();
+        public native @Cast("char*") String errorMessage();
+
+        public native void setErrorCode(int errorCode);
+        public native void setErrorMessage(@StdString BytePointer message);
+        public native void setErrorMessage(@StdString String message);
+    }
+
+
+
+// #endif //DEV_TESTS_ERRORREFERENCE_H
+
+
 // Parsed from memory/MemoryType.h
 
 //
@@ -688,6 +742,18 @@ bool verbose = false;
 // #include <graph/ResultWrapper.h>
 // #include <DebugInfo.h>
 
+/**
+ * This function returns last error code stored,
+ * @return non-zero if something bad happened
+ */
+public native int lastErrorCode();
+
+/**
+ * This function returns last error message, if last error code > 0
+ * @return
+ */
+public native @Cast("char*") String lastErrorMessage();
+
 /**
  *
  * @param p
@@ -1710,72 +1776,6 @@ public native void execScalarBoolTad(@Cast("Nd4jPointer*") PointerPointer extraP
                 @Cast("Nd4jLong*") long[] tadShapeInfo, @Cast("Nd4jLong*") long[] tadOffsets,
                 @Cast("Nd4jLong*") long[] tadShapeInfoZ, @Cast("Nd4jLong*") long[] tadOffsetsZ);
 
-
-/**
-* Append an input array
-* to the end of a flat array
-* in a particular order
-* @param offset the offset of the array to start at
-* @param order the order
-* @param result the result array
-* @param resultShapeInfo the shape info for te array
-* @param input the input for the array
-* @param inputShapeInfo the shape information for that array
-*/
-public native void flatten(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int offset,
-        char order,
-        Pointer result, @Cast("Nd4jLong*") LongPointer resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") LongPointer dresultShapeInfo,
-        Pointer input, @Cast("Nd4jLong*") LongPointer inputShapeInfo,
-        Pointer dinput, @Cast("Nd4jLong*") LongPointer dinputShapeInfo);
-public native void flatten(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int offset,
-        char order,
-        Pointer result, @Cast("Nd4jLong*") LongBuffer resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") LongBuffer dresultShapeInfo,
-        Pointer input, @Cast("Nd4jLong*") LongBuffer inputShapeInfo,
-        Pointer dinput, @Cast("Nd4jLong*") LongBuffer dinputShapeInfo);
-public native void flatten(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int offset,
-        char order,
-        Pointer result, @Cast("Nd4jLong*") long[] resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") long[] dresultShapeInfo,
-        Pointer input, @Cast("Nd4jLong*") long[] inputShapeInfo,
-        Pointer dinput, @Cast("Nd4jLong*") long[] dinputShapeInfo);
-
-public native void concat(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int dimension,
-        int numArrays,
-        @Cast("Nd4jPointer*") PointerPointer data, @Cast("Nd4jPointer*") PointerPointer inputShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer ddata, @Cast("Nd4jPointer*") PointerPointer dinputShapeInfo,
-        Pointer result, @Cast("Nd4jLong*") LongPointer resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") LongPointer dresultShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer tadPointers, @Cast("Nd4jPointer*") PointerPointer offsetPointers);
-public native void concat(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int dimension,
-        int numArrays,
-        @Cast("Nd4jPointer*") PointerPointer data, @Cast("Nd4jPointer*") PointerPointer inputShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer ddata, @Cast("Nd4jPointer*") PointerPointer dinputShapeInfo,
-        Pointer result, @Cast("Nd4jLong*") LongBuffer resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") LongBuffer dresultShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer tadPointers, @Cast("Nd4jPointer*") PointerPointer offsetPointers);
-public native void concat(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int dimension,
-        int numArrays,
-        @Cast("Nd4jPointer*") PointerPointer data, @Cast("Nd4jPointer*") PointerPointer inputShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer ddata, @Cast("Nd4jPointer*") PointerPointer dinputShapeInfo,
-        Pointer result, @Cast("Nd4jLong*") long[] resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") long[] dresultShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer tadPointers, @Cast("Nd4jPointer*") PointerPointer offsetPointers);
-
-
 public native void specialConcat(
         @Cast("Nd4jPointer*") PointerPointer extraPointers,
         int dimension,
@@ -9950,6 +9950,7 @@ public static final int PREALLOC_SIZE = 33554432;
 
 // #include <dll.h>
 // #include <pointercast.h>
+// #include <execution/ErrorReference.h>
     @Namespace("nd4j") @NoOffset public static class ContextBuffers extends Pointer {
         static { Loader.load(); }
         /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */
@@ -9985,6 +9986,8 @@ public static final int PREALLOC_SIZE = 33554432;
         public native void setScalarBuffer(Pointer pointer);
         public native void setAllocationBuffer(Pointer pointer);
 
+        public native ErrorReference errorReference();
+
         public native void triggerOwnership(@Cast("bool") boolean isOwner);
 
         public native int deviceId();
@@ -10038,6 +10041,7 @@ public static final int PREALLOC_SIZE = 33554432;
 // #include <vector>
 // #include <mutex>
 // #include <execution/ContextBuffers.h>
+// #include <execution/ErrorReference.h>
 
 @Namespace("nd4j") @NoOffset public static class LaunchContext extends Pointer {
     static { Loader.load(); }
@@ -10067,9 +10071,12 @@ public static final int PREALLOC_SIZE = 33554432;
 
     	public native int getDeviceID();
     	public native void setDeviceID(int deviceID);
+        public native ErrorReference errorReference();
 
     	public static native @Cast("bool") boolean isInitialized();
     	public static native void releaseBuffers();
+
+
 	    public static native LaunchContext defaultContext();
 
 
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCudaPresets.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCudaPresets.java
index 4466cf4b5..8f95fe5cb 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCudaPresets.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCudaPresets.java
@@ -32,6 +32,7 @@ import org.bytedeco.javacpp.tools.InfoMapper;
                         "array/ConstantDescriptor.h",
                         "array/ConstantDataBuffer.h",
                         "array/TadPack.h",
+                        "execution/ErrorReference.h",
                         "memory/MemoryType.h",
                         "Environment.h",
                         "types/utf8string.h",
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/linalg/cpu/nativecpu/CpuNDArrayFactory.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/linalg/cpu/nativecpu/CpuNDArrayFactory.java
index 2b47103c3..cacf32b38 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/linalg/cpu/nativecpu/CpuNDArrayFactory.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/linalg/cpu/nativecpu/CpuNDArrayFactory.java
@@ -106,6 +106,9 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
         functions.put(8, Loader.addressof("LAPACKE_sgesdd"));
         functions.put(9, Loader.addressof("LAPACKE_dgesdd"));
         nativeOps.initializeFunctions(functions);
+
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
     }
 
     @Override
@@ -489,32 +492,7 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
     @Override
     public INDArray toFlattened(char order, Collection<INDArray> matrices) {
         Preconditions.checkArgument(matrices.size() > 0, "toFlattened expects > 0 operands");
-/*
-        int length = 0;
-        val list = new ArrayList<INDArray>(matrices);
-        val t = list.get(0).dataType();
-        for (INDArray m : matrices) {
-            length += m.length();
-            Preconditions.checkArgument(m.dataType() == t, "All operands must have same data type");
-        }
 
-        INDArray ret = Nd4j.create(t, new long[] {length}, order);
-        int linearIndex = 0;
-        PointerPointer dummy = new PointerPointer(new Pointer[] {null});
-        for (INDArray m : matrices) {
-            Nd4j.getCompressor().autoDecompress(m);
-
-            nativeOps.flatten(dummy, linearIndex, order,
-                            ret.data().addressPointer(), (LongPointer) ret.shapeInfoDataBuffer().addressPointer(),
-                            null, null,
-                            m.data().addressPointer(),
-                            (LongPointer) m.shapeInfoDataBuffer().addressPointer(),
-                            null, null);
-
-            linearIndex += m.length();
-        }
-        return ret;
-        */
         return Nd4j.exec(new Flatten(order, matrices.toArray(new INDArray[matrices.size()])))[0];
     }
 
@@ -555,6 +533,9 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
                     new LongPointerWrapper(tadBuffers.getSecond().pointer())
             );
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         return result;
     }
 
@@ -574,65 +555,6 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
             return toConcat[0];
 
         return Nd4j.exec(new Concat(dimension, toConcat))[0];
-
-        // legacy implementation
-/*
-        // if reusable var wasn't created for this thread, or is smaller then needed - set it to new value
-        if (extrazA.get() == null || extrazB.get() == null || extrazSize.get() == null || extrazSize.get() < toConcat.length) {
-            extrazA.set(new PointerPointer(toConcat.length));
-            extrazB.set(new PointerPointer(toConcat.length));
-            extrazSize.set(toConcat.length);
-        }
-
-        PointerPointer shapeInfoPointers = extrazA.get();
-        PointerPointer dataPointers = extrazB.get();
-        int sumAlongDim = 0;
-
-        long[] outputShape = ArrayUtil.copy(toConcat[0].shape());
-
-        boolean allScalars = true;
-
-        for (int i = 0; i < toConcat.length; i++) {
-            Preconditions.checkState(toConcat[i].rank() == outputShape.length, "Encountered different array ranks for concat: input[0].shape()=%ndShape, input[%s].shape()=%ndShape",
-                    toConcat[0], i, toConcat[i]);
-
-            if (toConcat[i].isCompressed())
-                Nd4j.getCompressor().decompressi(toConcat[i]);
-
-            Preconditions.checkArgument(toConcat[i].dataType() == toConcat[0].dataType(), "All operands must have same data type: input 0 has type %s, input %s has type %s",
-                    toConcat[0].dataType(), i, toConcat[i].dataType());
-
-            allScalars &= toConcat[i].rank() == 0;
-
-            shapeInfoPointers.put(i, toConcat[i].shapeInfoDataBuffer().addressPointer());
-            dataPointers.put(i, toConcat[i].data().addressPointer());
-            sumAlongDim += toConcat[i].size(dimension);
-            for (int j = 0; j < toConcat[i].rank(); j++) {
-
-                if (j != dimension && toConcat[i].size(j) != outputShape[j]) {
-                    throw new IllegalArgumentException(
-                            "Illegal concatenation at array " + i + " and shape element " + j);
-                }
-            }
-        }
-
-        if (allScalars) {
-            outputShape = new long[]{sumAlongDim};
-        } else {
-            outputShape[dimension] = sumAlongDim;
-        }
-
-        INDArray ret = Nd4j.createUninitialized(toConcat[0].dataType(), outputShape, Nd4j.order());
-
-        nativeOps.concat(null, dimension, toConcat.length,
-                    dataPointers, shapeInfoPointers,
-                null, null,
-                    ret.data().addressPointer(), (LongPointer) ret.shapeInfoDataBuffer().addressPointer(),
-                    null, null,
-                    null, null);
-
-        return ret;
-        */
     }
 
 
@@ -757,6 +679,8 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
                     (LongPointer) zTadShapeInfo,
                     new LongPointerWrapper(zTadOffsets));
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         return ret;
     }
@@ -794,6 +718,9 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
                 arrays.length,
                 len);
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         return target;
     }
 
@@ -846,6 +773,9 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
                     len,
                 true);
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         return target;
     }
 
@@ -983,6 +913,8 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
                 arrays.size(),
                     ptrMap, tadPointers, offsetPointers);
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
 
         dataPointers.address();
         shapePointers.address();
@@ -990,84 +922,6 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
         offsetPointers.address();
     }
 
-
-    /**
-     * This method converts Half-precision databuffer to current dType buffer.
-     *
-     * @param buffer
-     * @return
-     */
-    /*
-    @Override
-    public DataBuffer restoreFromHalfs(DataBuffer buffer) {
-        if (buffer.dataType() != DataType.COMPRESSED)
-            throw new IllegalStateException("DataBuffer contains wrong data: " + buffer.dataType());
-    
-        CompressedDataBuffer comp = (CompressedDataBuffer) buffer;
-        CompressionDescriptor descriptor = comp.getCompressionDescriptor();
-    
-        DataBuffer targetBuffer = Nd4j.createBuffer(descriptor.getCompressedLength() / 2);
-    
-        if (Nd4j.dataType() == DataType.DOUBLE) {
-            nativeOps.convertHalfsToDoubles(
-                    null,
-                    comp.addressPointer(),
-                    (int) descriptor.getCompressedLength() / 2,
-                    targetBuffer.addressPointer()
-            );
-        } else if (Nd4j.dataType() == DataType.FLOAT) {
-            nativeOps.convertHalfsToFloats(
-                    null,
-                    comp.addressPointer(),
-                    (int) descriptor.getCompressedLength() / 2,
-                    targetBuffer.addressPointer()
-            );
-        } else {
-            throw new UnsupportedOperationException("Target dtype isn't supported: " + Nd4j.dataType());
-        }
-    
-        return targetBuffer;
-    }
-    */
-
-    /**
-     * This method converts Single/Double precision databuffer to Half-precision databuffer
-     *
-     * @param buffer
-     * @return
-     */
-    /*@Override
-    public DataBuffer convertToHalfs(DataBuffer buffer) {
-        // we allocate pointer
-        ShortPointer pointer = new ShortPointer(buffer.length());
-    
-        if (buffer.dataType() == DataType.DOUBLE) {
-            nativeOps.convertDoublesToHalfs(
-                    null,
-                    buffer.addressPointer(),
-                    (int) buffer.length(),
-                    pointer
-            );
-        } else if (buffer.dataType() == DataType.FLOAT) {
-            nativeOps.convertFloatsToHalfs(
-                    null,
-                    buffer.addressPointer(),
-                    (int) buffer.length(),
-                    pointer
-            );
-        } else {
-            throw new UnsupportedOperationException("Source dtype isn't supported: " + buffer.dataType());
-        }
-    
-        CompressionDescriptor descriptor = new CompressionDescriptor(buffer, new Float16());
-        descriptor.setCompressedLength(buffer.length() * 2);
-    
-    
-        CompressedDataBuffer result = new CompressedDataBuffer(pointer, descriptor);
-        return result;
-    }
-    */
-
     /**
      * This method converts Single/Double precision databuffer to Half-precision databuffer
      *
@@ -1081,6 +935,9 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
             throw new UnsupportedOperationException("Impossible to compress View. Consider using dup() before. ");
 
         DataBuffer buffer = convertDataEx(typeSrc, source.data(), typeDst);
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         source.setData(buffer);
 
         if (buffer instanceof CompressedDataBuffer)
@@ -1125,6 +982,9 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
 
         convertDataEx(typeSrc, source, typeDst, buffer);
 
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
+
         return buffer;
     }
 
@@ -1132,6 +992,9 @@ public class CpuNDArrayFactory extends BaseNativeNDArrayFactory {
     public void convertDataEx(DataTypeEx typeSrc, Pointer source, DataTypeEx typeDst, Pointer target,
                               long length) {
         nativeOps.convertTypes(null, typeSrc.ordinal(), source, length, typeDst.ordinal(), target);
+
+        if (nativeOps.lastErrorCode() != 0)
+            throw new RuntimeException(nativeOps.lastErrorMessage());
     }
 
     @Override
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/linalg/cpu/nativecpu/ops/NativeOpExecutioner.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/linalg/cpu/nativecpu/ops/NativeOpExecutioner.java
index 11373c440..e79c21feb 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/linalg/cpu/nativecpu/ops/NativeOpExecutioner.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/linalg/cpu/nativecpu/ops/NativeOpExecutioner.java
@@ -234,6 +234,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                         null);
             }
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         profilingConfigurableHookOut(op, st);
         return op.z();
     }
@@ -563,6 +566,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                 }
             }
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         return ret;
     }
 
@@ -644,6 +650,8 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                 throw new UnsupportedOperationException();
         }
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
     }
 
     public INDArray exec(ScalarOp op) {
@@ -690,6 +698,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                 throw new ND4JIllegalStateException("Unknown op type: [" + op.getOpType() +"]");
         }
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         profilingConfigurableHookOut(op, st);
 
         return op.z();
@@ -886,6 +897,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
 
             }
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         profilingConfigurableHookOut(op, st);
     }
 
@@ -962,6 +976,8 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                 throw new UnsupportedOperationException("Unknown operation type: [" + op.getOpType() + "]");
         }
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
 
         return op.z();
     }
@@ -1091,6 +1107,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                     batch.getSample().maxIntArrays(), batch.getSample().maxIntArraySize(),
                     batch.getSample().maxIndexArguments(), batch.getSample().maxRealArguments(), pointer, FlatBuffersMapper.getDataTypeAsByte(dataType));
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
     }
 
     /**
@@ -1197,6 +1216,8 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                     numIndexArguments, intArrays, numIntArrays, block.getRealArgumentsPointer(),
                     numRealArguments, FlatBuffersMapper.getDataTypeAsByte(dataType));
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
     }
 
     /**
@@ -1284,6 +1305,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                         op.extraArgsDataBuff(op.z().dataType()).addressPointer());
         }
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         profilingConfigurableHookOut(op, st);
 
         return op.z();
@@ -1370,6 +1394,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                 (float) threshold);
         //long t2 = System.currentTimeMillis();
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         if (cntAbs < 2)
             return null;
 
@@ -1429,6 +1456,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
 
         loop.convertTypes(null, DataTypeEx.THRESHOLD.ordinal(), buffer.addressPointer(), target.length(), typeDst.ordinal(), target.data().addressPointer());
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         return target;
     }
 
@@ -1460,6 +1490,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                 (IntPointer) buffer.addressPointer(),
                 (float) threshold);
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         return affected;
     }
 
@@ -1473,6 +1506,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                 (LongPointer) target.shapeInfoDataBuffer().addressPointer()
         );
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         return target;
     }
 
@@ -1673,136 +1709,6 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
         } catch (Exception e) {
             throw new RuntimeException("Op [" + name + "] execution failed", e);
         }
-/*
-        val name = op.opName().toLowerCase();
-        val hash = op.opHash();
-
-        if (name.equals("noop")) {
-            return op.outputArguments();
-        }
-
-        val inputShapes = getInputShapes(op.numInputArguments());
-        val inputBuffers = getInputBuffers(op.numInputArguments());
-
-        int cnt= 0;
-        val inputArgs = op.inputArguments();
-        for (val in: inputArgs) {
-            if(in == null)
-                throw new NullPointerException("Input argument is null for op " + op.getClass().getName());
-
-            if (!in.isEmpty())
-                inputBuffers.put(cnt, in.data().addressPointer());
-
-            inputShapes.put(cnt++, in.shapeInfoDataBuffer().addressPointer());
-        }
-
-        val outputArgs = op.outputArguments();
-        for(int i = 0; i < outputArgs.length; i++) {
-            if(outputArgs[i] == null)
-                throw new ND4JIllegalStateException("Op output arguments must not be null! Op " + op.getClass().getName());
-        }
-
-
-        val outputShapes = getOutputShapes(op.numOutputArguments());
-        val outputBuffers = getOutputBuffers(op.numOutputArguments());
-
-        cnt= 0;
-        for (val out: outputArgs) {
-            if(out.isEmpty()){
-                outputBuffers.put(cnt, null);
-            } else {
-                outputBuffers.put(cnt, out.data().addressPointer());
-            }
-            outputShapes.put(cnt++, out.shapeInfoDataBuffer().addressPointer());
-        }
-
-        val iArgs = op.numIArguments() > 0 ? getLongPointerFrom(iArgsPointer,op.numIArguments()) : null;
-        val tArgs = op.numTArguments() > 0 ? getDoublePointerFrom(tArgsPointer,op.numTArguments()) : null;
-        val bArgs = op.numBArguments() > 0 ? getBooleanPointerFrom(bArgsPointer,op.numBArguments()) : null;
-
-        cnt = 0;
-        val iArgs1 = op.iArgs();
-        for (val i: iArgs1)
-            iArgs.put(cnt++, i);
-
-        cnt = 0;
-        val bArgs1 = op.bArgs();
-        for (val b: bArgs1)
-            bArgs.put(cnt++, b);
-
-        cnt = 0;
-        val tArgs1 = op.tArgs();
-        for (val t: tArgs1)
-            tArgs.put(cnt++, t);
-
-        val t = op.numInputArguments();
-
-        OpStatus status = OpStatus.ND4J_STATUS_OK;
-        try {
-            val code = loop.execCustomOp(
-                        null,
-                        hash,
-                        inputBuffers,
-                        inputShapes,
-                        op.numInputArguments(),
-                        outputBuffers,
-                        outputShapes,
-                        op.numOutputArguments(),
-                        tArgs, op.numTArguments(),
-                        iArgs, op.numIArguments(),
-                        bArgs, op.numBArguments(),
-                        op.isInplaceCall());
-
-                status = OpStatus.byNumber(code);
-
-                if (status != OpStatus.ND4J_STATUS_OK)
-                    throw new ND4JIllegalStateException("Failed to execute op [" + name + "] with error code [" + status +"]");
-            }catch(Exception e) {
-                val sb = new StringBuilder();
-                sb.append("Inputs: [(");
-                for( int i=0; i<inputArgs.length; i++ ){
-                    if(i > 0)
-                        sb.append("), (");
-                    sb.append(Shape.shapeToStringShort(inputArgs[i]));
-                }
-                sb.append(")]. Outputs: [(");
-                for( int i=0; i<outputArgs.length; i++){
-                    if(i > 0)
-                        sb.append("), (");
-                    sb.append(Shape.shapeToStringShort(outputArgs[i]));
-                }
-                sb.append(")]. tArgs: ");
-                if(op.numTArguments() > 0){
-                    sb.append(Arrays.toString(op.tArgs()));
-                } else {
-                    sb.append("-");
-                }
-                sb.append(". iArgs: ");
-                if(op.numIArguments() > 0){
-                    sb.append(Arrays.toString(op.iArgs()));
-                } else {
-                    sb.append("-");
-                }
-                if(op instanceof DifferentialFunction){
-                    String n = ((DifferentialFunction) op).getOwnName();
-                    if(n != null && !n.equals(op.opName())){
-                        sb.append(". Op own name: \"").append(n).append("\"");
-                    }
-                }
-                log.error("Failed to execute op " + op.opName() + ". Attempted to execute with " +
-                                String.valueOf(op.numInputArguments()) + " inputs, " +
-                                String.valueOf(op.numOutputArguments()) + " outputs, "+
-                                String.valueOf(op.numTArguments()) + " targs and " +
-                                String.valueOf(op.numIArguments()) + " iargs. " +
-                                sb.toString() +
-                " - Please see above message (printed out from c++) for a possible cause of error.");
-                throw e;
-        }
-
-        profilingConfigurableHookOut(op, st);
-
-        return op.outputArguments();
- */
     }
 
     protected LongShapeDescriptor getShapeFromPointer(LongPointer ptr) {
@@ -1870,6 +1776,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                 ptrptr = loop.calculateOutputShapes2(null,
                         hash, inputBuffers, inputShapes, op.numInputArguments(), tArgs,
                         op.numTArguments(), iArgs, op.numIArguments(), bArgs, op.numBArguments());
+
+                if (loop.lastErrorCode() != 0)
+                    throw new RuntimeException(loop.lastErrorMessage());
             } catch (Throwable t){
                 StringBuilder sb = new StringBuilder();
                 sb.append("Inputs: [(");
@@ -1893,6 +1802,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                 throw t;
             }
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
             if (ptrptr == null)
                 throw new RuntimeException();
 
@@ -1929,6 +1841,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
     @Override
     public void registerGraph(long id, Pointer graph) {
          loop.registerGraph(null, id, graph);
+
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
     }
 
     @Override
@@ -1952,7 +1867,10 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
 
         val newMap = new LinkedHashMap<String, INDArray>();
 
-            OpaqueVariablesSet result = loop.executeStoredGraph(null, id, ptrBuffers, ptrShapes, ptrIndices, map.size());
+        OpaqueVariablesSet result = loop.executeStoredGraph(null, id, ptrBuffers, ptrShapes, ptrIndices, map.size());
+
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
 
             OpStatus status = OpStatus.byNumber(loop.getVariablesSetStatus(result));
 
@@ -1996,6 +1914,8 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
     @Override
     public void forgetGraph(long id) {
         loop.unregisterGraph(null, id);
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
     }
 
     /**
@@ -2055,6 +1975,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
                 array.data().addressPointer(), (LongPointer) tadX.getFirst().addressPointer(), (LongPointer) tadX.getSecond().addressPointer(), null, null, null,
                 updates.data().addressPointer(), (LongPointer) tadY.getFirst().addressPointer(), (LongPointer) tadY.getSecond().addressPointer(), null, null, null,
                 (IntPointer) indices.data().addressPointer(), null);
+
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
     }
 
     @Override
@@ -2078,6 +2001,10 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
 
 
             val status = loop.execCustomOp2(null, op.opHash(), context.contextPointer());
+
+            if (loop.lastErrorCode() != 0)
+                throw new RuntimeException(loop.lastErrorMessage());
+
             if (status != 0)
                 throw new RuntimeException("Op [" + op.opName() + "] execution failed");
 
@@ -2155,6 +2082,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
 
         loop.inspectArray(null, array.data().addressPointer(), (LongPointer) array.shapeInfoDataBuffer().addressPointer(), null, null, debugInfo);
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         return INDArrayStatistics.builder()
                 .minValue(debugInfo._minValue())
                 .maxValue(debugInfo._maxValue())
@@ -2171,6 +2101,8 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
     @Override
     public DataBuffer createShapeInfo(long[] shape, long[] stride, long elementWiseStride, char order, DataType dtype, boolean empty) {
         OpaqueConstantDataBuffer dbf = loop.shapeBuffer(shape.length, new LongPointer(shape), new LongPointer(stride), dtype.toInt(), order, elementWiseStride, empty);
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
 
         val result = new LongBuffer(loop.getConstantDataBufferPrimary(dbf), Shape.shapeInfoLength(shape.length));
 
@@ -2183,6 +2115,9 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
     public TadPack tadShapeInfoAndOffsets(INDArray array, int[] dimension) {
         OpaqueTadPack pack = loop.tadOnlyShapeInfo((LongPointer) array.shapeInfoDataBuffer().addressPointer(), new IntPointer(dimension), dimension.length);
 
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
         val tadShape = new LongBuffer(loop.getPrimaryShapeInfo(pack), loop.getShapeInfoLength(pack));
         val tadOffsets = new LongBuffer(loop.getPrimaryOffsets(pack), loop.getNumberOfTads(pack));
 
@@ -2205,11 +2140,19 @@ public class NativeOpExecutioner extends DefaultOpExecutioner {
 
     @Override
     public String runLightBenchmarkSuit(boolean printOut) {
-        return loop.runLightBenchmarkSuit(printOut);
+        val s = loop.runLightBenchmarkSuit(printOut);
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
+        return s;
     }
 
     @Override
     public String runFullBenchmarkSuit(boolean printOut) {
-        return loop.runFullBenchmarkSuit(printOut);
+        val s = loop.runFullBenchmarkSuit(printOut);
+        if (loop.lastErrorCode() != 0)
+            throw new RuntimeException(loop.lastErrorMessage());
+
+        return s;
     }
 }
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java
index 8e71816f8..38c0cb8c4 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java
@@ -467,6 +467,60 @@ public class Nd4jCpu extends org.nd4j.nativeblas.Nd4jCpuHelper {
 // #endif //DEV_TESTS_TADPACK_H
 
 
+// Parsed from execution/ErrorReference.h
+
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+
+// #ifndef DEV_TESTS_ERRORREFERENCE_H
+// #define DEV_TESTS_ERRORREFERENCE_H
+
+// #include <string>
+// #include <dll.h>
+    @Namespace("sd") @NoOffset public static class ErrorReference extends Pointer {
+        static { Loader.load(); }
+        /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */
+        public ErrorReference(Pointer p) { super(p); }
+        /** Native array allocator. Access with {@link Pointer#position(long)}. */
+        public ErrorReference(long size) { super((Pointer)null); allocateArray(size); }
+        private native void allocateArray(long size);
+        @Override public ErrorReference position(long position) {
+            return (ErrorReference)super.position(position);
+        }
+    
+        public ErrorReference() { super((Pointer)null); allocate(); }
+        private native void allocate();
+
+        public native int errorCode();
+        public native @Cast("char*") String errorMessage();
+
+        public native void setErrorCode(int errorCode);
+        public native void setErrorMessage(@StdString BytePointer message);
+        public native void setErrorMessage(@StdString String message);
+    }
+
+
+
+// #endif //DEV_TESTS_ERRORREFERENCE_H
+
+
 // Parsed from Environment.h
 
 /*******************************************************************************
@@ -688,6 +742,18 @@ bool verbose = false;
 // #include <graph/ResultWrapper.h>
 // #include <DebugInfo.h>
 
+/**
+ * This function returns last error code stored,
+ * @return non-zero if something bad happened
+ */
+public native int lastErrorCode();
+
+/**
+ * This function returns last error message, if last error code > 0
+ * @return
+ */
+public native @Cast("char*") String lastErrorMessage();
+
 /**
  *
  * @param p
@@ -1710,72 +1776,6 @@ public native void execScalarBoolTad(@Cast("Nd4jPointer*") PointerPointer extraP
                 @Cast("Nd4jLong*") long[] tadShapeInfo, @Cast("Nd4jLong*") long[] tadOffsets,
                 @Cast("Nd4jLong*") long[] tadShapeInfoZ, @Cast("Nd4jLong*") long[] tadOffsetsZ);
 
-
-/**
-* Append an input array
-* to the end of a flat array
-* in a particular order
-* @param offset the offset of the array to start at
-* @param order the order
-* @param result the result array
-* @param resultShapeInfo the shape info for te array
-* @param input the input for the array
-* @param inputShapeInfo the shape information for that array
-*/
-public native void flatten(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int offset,
-        char order,
-        Pointer result, @Cast("Nd4jLong*") LongPointer resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") LongPointer dresultShapeInfo,
-        Pointer input, @Cast("Nd4jLong*") LongPointer inputShapeInfo,
-        Pointer dinput, @Cast("Nd4jLong*") LongPointer dinputShapeInfo);
-public native void flatten(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int offset,
-        char order,
-        Pointer result, @Cast("Nd4jLong*") LongBuffer resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") LongBuffer dresultShapeInfo,
-        Pointer input, @Cast("Nd4jLong*") LongBuffer inputShapeInfo,
-        Pointer dinput, @Cast("Nd4jLong*") LongBuffer dinputShapeInfo);
-public native void flatten(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int offset,
-        char order,
-        Pointer result, @Cast("Nd4jLong*") long[] resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") long[] dresultShapeInfo,
-        Pointer input, @Cast("Nd4jLong*") long[] inputShapeInfo,
-        Pointer dinput, @Cast("Nd4jLong*") long[] dinputShapeInfo);
-
-public native void concat(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int dimension,
-        int numArrays,
-        @Cast("Nd4jPointer*") PointerPointer data, @Cast("Nd4jPointer*") PointerPointer inputShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer ddata, @Cast("Nd4jPointer*") PointerPointer dinputShapeInfo,
-        Pointer result, @Cast("Nd4jLong*") LongPointer resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") LongPointer dresultShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer tadPointers, @Cast("Nd4jPointer*") PointerPointer offsetPointers);
-public native void concat(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int dimension,
-        int numArrays,
-        @Cast("Nd4jPointer*") PointerPointer data, @Cast("Nd4jPointer*") PointerPointer inputShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer ddata, @Cast("Nd4jPointer*") PointerPointer dinputShapeInfo,
-        Pointer result, @Cast("Nd4jLong*") LongBuffer resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") LongBuffer dresultShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer tadPointers, @Cast("Nd4jPointer*") PointerPointer offsetPointers);
-public native void concat(
-        @Cast("Nd4jPointer*") PointerPointer extraPointers,
-        int dimension,
-        int numArrays,
-        @Cast("Nd4jPointer*") PointerPointer data, @Cast("Nd4jPointer*") PointerPointer inputShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer ddata, @Cast("Nd4jPointer*") PointerPointer dinputShapeInfo,
-        Pointer result, @Cast("Nd4jLong*") long[] resultShapeInfo,
-        Pointer dresult, @Cast("Nd4jLong*") long[] dresultShapeInfo,
-        @Cast("Nd4jPointer*") PointerPointer tadPointers, @Cast("Nd4jPointer*") PointerPointer offsetPointers);
-
-
 public native void specialConcat(
         @Cast("Nd4jPointer*") PointerPointer extraPointers,
         int dimension,
@@ -22877,6 +22877,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
 
 // #include <dll.h>
 // #include <pointercast.h>
+// #include <execution/ErrorReference.h>
     @Namespace("nd4j") @NoOffset public static class ContextBuffers extends Pointer {
         static { Loader.load(); }
         /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */
@@ -22912,6 +22913,8 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
         public native void setScalarBuffer(Pointer pointer);
         public native void setAllocationBuffer(Pointer pointer);
 
+        public native ErrorReference errorReference();
+
         public native void triggerOwnership(@Cast("bool") boolean isOwner);
 
         public native int deviceId();
@@ -22961,6 +22964,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
 // #include <vector>
 // #include <mutex>
 // #include <execution/ContextBuffers.h>
+// #include <execution/ErrorReference.h>
 
 @Namespace("nd4j") @NoOffset public static class LaunchContext extends Pointer {
     static { Loader.load(); }
@@ -22985,9 +22989,12 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
 
     	public native int getDeviceID();
     	public native void setDeviceID(int deviceID);
+        public native ErrorReference errorReference();
 
     	public static native @Cast("bool") boolean isInitialized();
     	public static native void releaseBuffers();
+
+
 	    public static native LaunchContext defaultContext();
 
 
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpuPresets.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpuPresets.java
index 58a2a7d02..554016686 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpuPresets.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpuPresets.java
@@ -38,6 +38,7 @@ import java.util.Scanner;
                                               "array/ConstantDataBuffer.h",
                                               "array/ConstantDescriptor.h",
                                               "array/TadPack.h",
+                                              "execution/ErrorReference.h",
                                               "Environment.h",
                                               "types/utf8string.h",
                                               "NativeOps.h",
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/Nd4jTestsC.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/Nd4jTestsC.java
index b302c8c0f..915d6f650 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/Nd4jTestsC.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/Nd4jTestsC.java
@@ -5216,6 +5216,8 @@ public class Nd4jTestsC extends BaseNd4jTest {
         INDArray array = Nd4j.create(new double[] {10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0});
         INDArray exp = Nd4j.create(new double[] {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
 
+        log.info("Array shapeInfo: {}", array.shapeInfoJava());
+
         INDArray rev = Nd4j.reverse(array);
 
         assertEquals(exp, rev);
@@ -5226,7 +5228,7 @@ public class Nd4jTestsC extends BaseNd4jTest {
         INDArray array = Nd4j.create(new double[] {9, 8, 7, 6, 5, 4, 3, 2, 1, 0});
         INDArray exp = Nd4j.create(new double[] {0, 1, 2, 3, 4, 5, 6, 7, 8, 9});
 
-        INDArray rev = Nd4j.getExecutioner().exec(new Reverse(array, Nd4j.createUninitialized(array.length())))[0];
+        INDArray rev = Nd4j.getExecutioner().exec(new Reverse(array, array.ulike()))[0];
 
         assertEquals(exp, rev);
     }
@@ -5236,7 +5238,7 @@ public class Nd4jTestsC extends BaseNd4jTest {
         INDArray array = Nd4j.create(new double[] {10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0});
         INDArray exp = Nd4j.create(new double[] {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10});
 
-        INDArray rev = Nd4j.getExecutioner().exec(new Reverse(array, Nd4j.createUninitialized(array.length())))[0];
+        INDArray rev = Nd4j.getExecutioner().exec(new Reverse(array,array.ulike()))[0];
 
         assertEquals(exp, rev);
     }
@@ -5335,11 +5337,103 @@ public class Nd4jTestsC extends BaseNd4jTest {
         assertNotNull(lsd);     //Fails here on CUDA, OK on native/cpu
     }
 
+    @Test
+    public void testReverseSmall_1() {
+        val array = Nd4j.linspace(1, 10, 10, DataType.INT);
+        val exp = array.dup(array.ordering());
+
+        Transforms.reverse(array, false);
+        Transforms.reverse(array, false);
+
+        val jexp = exp.data().asInt();
+        val jarr = array.data().asInt();
+        assertArrayEquals(jexp, jarr);
+        assertEquals(exp, array);
+    }
+
+    @Test
+    public void testReverseSmall_2() {
+        val array = Nd4j.linspace(1, 10, 10, DataType.INT);
+        val exp = array.dup(array.ordering());
+
+        val reversed = Transforms.reverse(array, true);
+        val rereversed = Transforms.reverse(reversed, true);
+
+        val jexp = exp.data().asInt();
+        val jarr = rereversed.data().asInt();
+        assertArrayEquals(jexp, jarr);
+        assertEquals(exp, rereversed);
+    }
+
+    @Test
+    public void testReverseSmall_3() {
+        val array = Nd4j.linspace(1, 11, 11, DataType.INT);
+        val exp = array.dup(array.ordering());
+
+        Transforms.reverse(array, false);
+
+        log.info("Reversed shapeInfo: {}", array.shapeInfoJava());
+        log.info("Reversed: {}", array);
+
+        Transforms.reverse(array, false);
+
+        val jexp = exp.data().asInt();
+        val jarr = array.data().asInt();
+        assertArrayEquals(jexp, jarr);
+        assertEquals(exp, array);
+    }
+
+    @Test
+    public void testReverseSmall_4() {
+        val array = Nd4j.linspace(1, 11, 11, DataType.INT);
+        val exp = array.dup(array.ordering());
+
+        val reversed = Transforms.reverse(array, true);
+
+        log.info("Reversed: {}", reversed);
+
+        val rereversed = Transforms.reverse(reversed, true);
+
+        val jexp = exp.data().asInt();
+        val jarr = rereversed.data().asInt();
+        assertArrayEquals(jexp, jarr);
+        assertEquals(exp, rereversed);
+    }
+
+    @Test
+    public void testReverse_1() {
+        val array = Nd4j.linspace(1, 2017152, 2017152, DataType.INT);
+        val exp = array.dup(array.ordering());
+
+        Transforms.reverse(array, false);
+        Transforms.reverse(array, false);
+
+        val jexp = exp.data().asInt();
+        val jarr = array.data().asInt();
+        assertArrayEquals(jexp, jarr);
+        assertEquals(exp, array);
+    }
+
+    @Test
+    public void testReverse_2() {
+        val array = Nd4j.linspace(1, 2017152, 2017152, DataType.INT);
+        val exp = array.dup(array.ordering());
+
+        val reversed = Transforms.reverse(array, true);
+        val rereversed = Transforms.reverse(reversed, true);
+
+        val jexp = exp.data().asInt();
+        val jarr = rereversed.data().asInt();
+        assertArrayEquals(jexp, jarr);
+        assertEquals(exp, rereversed);
+    }
+
     @Test
     public void testNativeSort3_1() {
         INDArray array = Nd4j.linspace(1, 2017152, 2017152, DataType.DOUBLE).reshape(1, -1);
         INDArray exp = array.dup();
         Transforms.reverse(array, false);
+        log.info("Reverse: {}", array);
 
 
         long time1 = System.currentTimeMillis();

From e92f7218f3bfe5833faa00670a5144903e552d95 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Tue, 27 Aug 2019 12:15:56 +1000
Subject: [PATCH 13/56] Add new tests (#171)

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 .../opvalidation/MiscOpValidation.java        | 58 +++++++++++++++++++
 .../opvalidation/RandomOpValidation.java      |  9 +++
 .../opvalidation/ShapeOpValidation.java       | 20 +++++++
 .../opvalidation/TransformOpValidation.java   | 39 +++++++++++++
 4 files changed, 126 insertions(+)

diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/MiscOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/MiscOpValidation.java
index 2a4b032b5..99a2f57ac 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/MiscOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/MiscOpValidation.java
@@ -438,6 +438,27 @@ public class MiscOpValidation extends BaseOpValidation {
         assertEquals(failed.toString(), 0, failed.size());
     }
 
+    @Test
+    public void testScatterUpdate(){
+        INDArray x = Nd4j.linspace(DataType.FLOAT, 1, 30, 1).reshape(10, 3);
+        INDArray updates = Nd4j.create(new float[][]{
+                {100, 101, 102},
+                {200, 201, 202}});
+        INDArray indices = Nd4j.createFromArray(2, 5);
+
+        INDArray exp = x.dup();
+        exp.putRow(2, updates.getRow(0));
+        exp.putRow(5, updates.getRow(1));
+
+        INDArray out = exp.ulike();
+        Nd4j.exec(DynamicCustomOp.builder("scatter_upd")
+                .addInputs(x, indices, updates)
+                .addOutputs(out)
+                .build());
+
+        assertEquals(exp, out);
+    }
+
     @Test
     public void testGatherGradient() {
         Nd4j.getRandom().setSeed(12345);
@@ -1688,4 +1709,41 @@ public class MiscOpValidation extends BaseOpValidation {
 
         Nd4j.getExecutioner().exec(op);
     }
+
+    @Test
+    public void testHistogramFixedWidth(){
+        //Bins: [-inf, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), [0.8, inf]
+        INDArray in = Nd4j.createFromArray(0.0, 0.1, 0.1, 0.3, 0.5, 0.5, 0.9);
+        INDArray range = Nd4j.createFromArray(0.0, 1.0);
+        INDArray n = Nd4j.scalar(5);
+
+        INDArray out = Nd4j.create(DataType.INT, 5);
+
+        Nd4j.exec(DynamicCustomOp.builder("histogram_fixed_width")
+                .addInputs(in, range, n)
+                .addOutputs(out)
+                .build());
+
+        INDArray exp = Nd4j.createFromArray(3, 1, 2, 0, 1);
+        assertEquals(exp, out);
+    }
+
+    @Test
+    public void testListDiff(){
+        INDArray x = Nd4j.createFromArray(0, 1, 2, 3);
+        INDArray y = Nd4j.createFromArray(3, 1);
+
+        INDArray out = Nd4j.create(DataType.INT, 2);
+        INDArray outIdx = Nd4j.create(DataType.INT, 2);
+
+        Nd4j.exec(DynamicCustomOp.builder("listdiff")
+                .addInputs(x, y)
+                .addOutputs(out, outIdx)
+                .build());
+
+        INDArray exp = Nd4j.createFromArray(0, 2);
+
+        assertEquals(exp, out);         //Values in x not in y
+        assertEquals(exp, outIdx);      //Indices of the values in x not in y
+    }
 }
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/RandomOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/RandomOpValidation.java
index 646cae454..8d64f6404 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/RandomOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/RandomOpValidation.java
@@ -27,6 +27,7 @@ import org.nd4j.autodiff.validation.TestCase;
 import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.iter.NdIndexIterator;
 import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.api.ops.impl.reduce.bool.All;
 import org.nd4j.linalg.api.ops.random.custom.RandomBernoulli;
 import org.nd4j.linalg.api.ops.random.custom.RandomExponential;
 import org.nd4j.linalg.api.ops.random.impl.BinomialDistribution;
@@ -371,6 +372,14 @@ public class RandomOpValidation extends BaseOpValidation {
 
             assertNull(OpValidation.validate(tc));
         }
+    }
 
+    @Test
+    public void testAllEmptyReduce(){
+        INDArray x = Nd4j.createFromArray(true, true, true);
+        All all = new All(x);
+        all.setEmptyReduce(true);   //For TF compatibility - empty array for axis (which means no-op - and NOT all array reduction)
+        INDArray out = Nd4j.exec(all);
+        assertEquals(x, out);
     }
 }
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java
index 2965f367f..ffb585183 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java
@@ -1342,6 +1342,26 @@ public class ShapeOpValidation extends BaseOpValidation {
         assertEquals(failed.toString(), 0, failed.size());
     }
 
+    @Test
+    public void testSegmentMean(){
+        INDArray x = Nd4j.linspace(DataType.FLOAT, 1, 18, 1).reshape(6, 3);
+        INDArray segmentIds = Nd4j.createFromArray(0, 0, 1, 1, 2, 2);
+
+        INDArray out = Nd4j.create(DataType.FLOAT, 3, 3);
+
+        Nd4j.exec(DynamicCustomOp.builder("segment_mean")
+                .addInputs(x, segmentIds)
+                .addOutputs(out)
+                .build());
+
+        INDArray exp = out.like();
+        exp.putRow(0, x.getRow(0).add(x.getRow(1)).muli(0.5));
+        exp.putRow(1, x.getRow(2).add(x.getRow(3)).muli(0.5));
+        exp.putRow(2, x.getRow(4).add(x.getRow(5)).muli(0.5));
+
+        assertEquals(exp, out);
+    }
+
     @Test
     public void testSequenceMask() {
         OpValidationSuite.ignoreFailing();  //2018-01-09: output datatype issue?
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java
index 0d177027d..9183a0884 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java
@@ -415,6 +415,24 @@ public class TransformOpValidation extends BaseOpValidation {
         assertNull(err, err);
     }
 
+    @Test
+    public void testDynamicPartition2(){
+        INDArray data = Nd4j.createFromArray(2, 1, 2, 0);
+        INDArray partitions = Nd4j.createFromArray(0, 2, 1, 0);
+        INDArray[] out = Nd4j.exec(DynamicCustomOp.builder("dynamic_partition")
+                .addOutputs(Nd4j.createUninitialized(DataType.INT, 2), Nd4j.createUninitialized(DataType.INT, 1), Nd4j.createUninitialized(DataType.INT, 1))
+                .addIntegerArguments(3) //3 partitions
+                .addInputs(data, partitions).build());
+
+        INDArray exp0 = Nd4j.createFromArray(2, 0);
+        INDArray exp1 = Nd4j.createFromArray(2);
+        INDArray exp2 = Nd4j.createFromArray(1);
+
+        assertEquals(exp0, out[0]);     //Usually just gives [0,0]
+        assertEquals(exp1, out[1]);
+        assertEquals(exp2, out[2]);
+    }
+
     @Test
     public void testDynamicStitch() {
         SameDiff sd = SameDiff.create();
@@ -1612,6 +1630,27 @@ public class TransformOpValidation extends BaseOpValidation {
         }
     }
 
+    @Test
+    public void testTopK1(){
+        INDArray x = Nd4j.createFromArray(0.0, 0.0, 0.0, 10.0, 0.0);
+        INDArray k = Nd4j.scalar(1);
+        INDArray outValue = Nd4j.create(DataType.DOUBLE, 1);
+        INDArray outIdx = Nd4j.create(DataType.INT, 1);
+
+        Nd4j.exec(DynamicCustomOp.builder("top_k")
+                .addInputs(x, k)
+                .addOutputs(outValue, outIdx)
+                .addBooleanArguments(false) //not sorted
+                .addIntegerArguments(1)
+                .build());
+
+        INDArray expValue = Nd4j.createFromArray(10.0);
+        INDArray expIdx = Nd4j.createFromArray(3);
+
+        assertEquals(expValue, outValue);
+        assertEquals(expIdx, outIdx);
+    }
+
     @Test
     public void testInTopK() {
         for( int k=4; k>= 1; k--){

From df84bc7255b53c359ab1e2284eab73de4d390a79 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Tue, 27 Aug 2019 10:37:10 +0300
Subject: [PATCH 14/56] [WIP] More tweaks (#173)

* CUDA empty reduction

Signed-off-by: raver119 <raver119@gmail.com>

* - listdiff synchronization fix for CUDA
- listdiff test

Signed-off-by: raver119 <raver119@gmail.com>

* - IndexReduce ops now allow INDEXING_TYPES output
- topK op accepts only INDEXING_TYPES as output

Signed-off-by: raver119 <raver119@gmail.com>
---
 libnd4j/blas/NDArray.hpp                      |   4 +-
 libnd4j/blas/cpu/NativeOpExecutioner.cpp      |   6 +-
 libnd4j/blas/cuda/NativeOpExecutioner.cu      |  14 +--
 libnd4j/include/helpers/Loops.h               |   6 +-
 .../helpers/cpu/loops/IndexReductionLoops.cpp |  35 +++---
 libnd4j/include/loops/cpu/indexreduce.cpp     |  32 +++---
 libnd4j/include/loops/cuda/indexreduce.cu     | 105 +++++++++---------
 libnd4j/include/loops/indexreduce.h           |  16 +--
 .../declarable/generic/parity_ops/top_k.cpp   |   2 +-
 .../ops/declarable/helpers/impl/listdiff.cpp  |  25 +----
 libnd4j/include/ops/ops.h                     |  12 +-
 .../ops/executioner/CudaExecutioner.java      |  47 +++++++-
 .../nd4j/linalg/custom/CustomOpsTests.java    |  40 +++++++
 .../org/nd4j/linalg/shape/EmptyTests.java     |  10 ++
 14 files changed, 217 insertions(+), 137 deletions(-)

diff --git a/libnd4j/blas/NDArray.hpp b/libnd4j/blas/NDArray.hpp
index fdbcae49f..72b029c0b 100644
--- a/libnd4j/blas/NDArray.hpp
+++ b/libnd4j/blas/NDArray.hpp
@@ -3590,8 +3590,8 @@ void NDArray::applyIndexReduce(nd4j::indexreduce::Ops op, NDArray* target, const
     if (isS())
         throw std::runtime_error("NDArray::applyIndexReduce: you can't use this method on String array!");
 
-    if (target->dataType() != nd4j::DataType::INT64)
-        throw std::runtime_error("NDArray::applyIndexReduce operations return INT64");
+    if (target->dataType() != nd4j::DataType::INT64 && target->dataType() != nd4j::DataType::INT32)
+        throw std::runtime_error("NDArray::applyIndexReduce operations return INT32/INT64");
 
     void* params = extraParams != nullptr ? const_cast<ExtraArguments*>(extraParams)->argumentsAsT(this->dataType()) : nullptr;
 
diff --git a/libnd4j/blas/cpu/NativeOpExecutioner.cpp b/libnd4j/blas/cpu/NativeOpExecutioner.cpp
index e320b4f57..b2ce7846a 100644
--- a/libnd4j/blas/cpu/NativeOpExecutioner.cpp
+++ b/libnd4j/blas/cpu/NativeOpExecutioner.cpp
@@ -79,9 +79,10 @@ void NativeOpExecutioner::execIndexReduceScalar(nd4j::LaunchContext  *lc, int op
 #endif
 
     auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
     auto hz = reinterpret_cast<Nd4jLong*>(hZ);
 
-    BUILD_SINGLE_SELECTOR(xType, hz[0] = functions::indexreduce::IndexReduce, ::execScalar(opNum,hX,hXShapeInfo,extraParams), LIBND4J_TYPES);
+    BUILD_DOUBLE_SELECTOR(xType, zType, hz[0] = functions::indexreduce::IndexReduce, ::execScalar(opNum,hX,hXShapeInfo,extraParams), LIBND4J_TYPES, INDEXING_TYPES);
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -111,9 +112,10 @@ void NativeOpExecutioner::execIndexReduce(nd4j::LaunchContext  *lc,
 #endif
 
     auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
     Nd4jLong* hz = reinterpret_cast<Nd4jLong*>(hZ);
 
-    BUILD_SINGLE_SELECTOR(xType, functions::indexreduce::IndexReduce, ::exec(opNum, hX, hXShapeInfo, extraParams, hz, hZShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets), LIBND4J_TYPES);
+    BUILD_DOUBLE_SELECTOR(xType, zType, functions::indexreduce::IndexReduce, ::exec(opNum, hX, hXShapeInfo, extraParams, hz, hZShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets), LIBND4J_TYPES, INDEXING_TYPES);
 //    BUILD_SINGLE_SELECTOR(xType, functions::indexreduce::IndexReduce, ::exec(opNum, hX, hXShapeInfo, dX, dXShapeInfo, extraParams, hZ, hZShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets), LIBND4J_TYPES);
 }
 
diff --git a/libnd4j/blas/cuda/NativeOpExecutioner.cu b/libnd4j/blas/cuda/NativeOpExecutioner.cu
index b3573c7ab..8c4f1d3fa 100644
--- a/libnd4j/blas/cuda/NativeOpExecutioner.cu
+++ b/libnd4j/blas/cuda/NativeOpExecutioner.cu
@@ -475,12 +475,12 @@ void NativeOpExecutioner::execIndexReduce(nd4j::LaunchContext  *lc,
 	auto numBlocks = shape::length(hZShapeInfo);
     dim3 launchDims(numBlocks, 256, 32768);
 
-    if (zType != nd4j::DataType::INT64)
-        throw datatype_exception::build("NativeOpExecutioner::execIndexReduce requires Z operand to have INT64 type", zType);
+    if (zType != nd4j::DataType::INT64 && zType != nd4j::DataType::INT32)
+        throw datatype_exception::build("NativeOpExecutioner::execIndexReduce requires Z operand to have INT32/INT64 type", zType);
 
 	auto dz = reinterpret_cast<Nd4jLong*>(dZ);
 
-	BUILD_SINGLE_SELECTOR(xType, functions::indexreduce::IndexReduce,  ::executeIndexReduce(launchDims, stream, opNum, dX, dXShapeInfo, shape::rank(hXShapeInfo), extraParams, dz, dZShapeInfo, shape::rank(hZShapeInfo), dimension, dimensionLength, 1, allocationPointer, reductionPointer, tadShapeInfo, tadOffsets), LIBND4J_TYPES);
+	BUILD_DOUBLE_SELECTOR(xType, zType, functions::indexreduce::IndexReduce,  ::executeIndexReduce(launchDims, stream, opNum, dX, dXShapeInfo, shape::rank(hXShapeInfo), extraParams, dz, dZShapeInfo, shape::rank(hZShapeInfo), dimension, dimensionLength, 1, allocationPointer, reductionPointer, tadShapeInfo, tadOffsets), LIBND4J_TYPES, INDEXING_TYPES);
 
     // TODO: remove after the release
     auto res = cudaStreamSynchronize(*stream);
@@ -567,12 +567,12 @@ void NativeOpExecutioner::execIndexReduceScalar(nd4j::LaunchContext  *lc,
     // FIXME: we want Z to be one of integer types
 	//if (!DataTypeUtils::isZ(zType))
 	//    throw nd4j::datatype_exception("NativeOpExecutioner::execIndexReduceScalar requires Z operand to have one of integer types")
-	if (zType != nd4j::DataType::INT64)
-        throw nd4j::datatype_exception::build("NativeOpExecutioner::execIndexReduceScalar requires Z operand to have INT64 data type", zType);
+	if (zType != nd4j::DataType::INT64 && zType != nd4j::DataType::INT32)
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execIndexReduceScalar requires Z operand to have INT32/INT64 data type", zType);
 
     auto dz = reinterpret_cast<Nd4jLong*>(dZ);
 
-    BUILD_SINGLE_SELECTOR(xType, functions::indexreduce::IndexReduce, ::executeIndexReduceScalar(launchDims, stream,
+    BUILD_DOUBLE_SELECTOR(xType, zType, functions::indexreduce::IndexReduce, ::executeIndexReduceScalar(launchDims, stream,
                                                                                                 opNum,
                                                                                                 dX, dXShapeInfo, shape::rank(hXShapeInfo),
                                                                                                 extraParams,
@@ -580,7 +580,7 @@ void NativeOpExecutioner::execIndexReduceScalar(nd4j::LaunchContext  *lc,
                                                                                                 nullptr, 0,
                                                                                                 1,
                                                                                                 allocationPointer, reductionPointer,
-                                                                                                nullptr, nullptr), LIBND4J_TYPES);
+                                                                                                nullptr, nullptr), LIBND4J_TYPES, INDEXING_TYPES);
     // TODO: remove after the release
     auto res = cudaStreamSynchronize(*stream);
     if (res != 0)
diff --git a/libnd4j/include/helpers/Loops.h b/libnd4j/include/helpers/Loops.h
index bda04414f..d04d3315d 100644
--- a/libnd4j/include/helpers/Loops.h
+++ b/libnd4j/include/helpers/Loops.h
@@ -80,14 +80,14 @@ namespace nd4j {
     };
 
 
-    template <typename X>
+    template <typename X, typename Z>
     class ND4J_EXPORT IndexReductionLoops {
     private:
     public:
-        static void wrapIndexReduce(const int opNum, void* x, Nd4jLong* xShapeInfo, Nd4jLong* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, void* extraParams);
+        static void wrapIndexReduce(const int opNum, void* x, Nd4jLong* xShapeInfo, void* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, void* extraParams);
 
         template <typename OpType>
-        static void loopIndexReduce(X* x, Nd4jLong* xShapeInfo, Nd4jLong* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams);
+        static void loopIndexReduce(X* x, Nd4jLong* xShapeInfo, Z* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, X* extraParams);
     };
 
 
diff --git a/libnd4j/include/helpers/cpu/loops/IndexReductionLoops.cpp b/libnd4j/include/helpers/cpu/loops/IndexReductionLoops.cpp
index 33e230bd5..0a096b65f 100644
--- a/libnd4j/include/helpers/cpu/loops/IndexReductionLoops.cpp
+++ b/libnd4j/include/helpers/cpu/loops/IndexReductionLoops.cpp
@@ -24,10 +24,10 @@ using namespace simdOps;
 
 
 //////////////////////////////////////////////////////////////////////////////
-template <typename X>
+template <typename X, typename Z>
 template <typename OpType>
-void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
-                           Nd4jLong* z, Nd4jLong* zShapeInfo,
+void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
+                           Z* z, Nd4jLong* zShapeInfo,
                            Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets,
                            X* extraParams) {
 
@@ -62,7 +62,7 @@ void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
                     indexValue = OpType::update(indexValue, comp, extraParams);
                 }
 
-                z[i] = indexValue.index;
+                z[i] = (Z) indexValue.index;
             }
         }
             break;
@@ -80,7 +80,7 @@ void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
                     indexValue = OpType::update(indexValue, comp, extraParams);
                 }
 
-                z[i * zEws] = indexValue.index;
+                z[i * zEws] = (Z) indexValue.index;
             }
         }
             break;
@@ -98,7 +98,7 @@ void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
                     indexValue = OpType::update(indexValue, comp, extraParams);
                 }
 
-                z[i] = indexValue.index;
+                z[i] = (Z) indexValue.index;
             }
         }
             break;
@@ -122,7 +122,7 @@ void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
                     }
                 }
 
-                z[i] = indexValue.index;
+                z[i] = (Z) indexValue.index;
             }
         }
             break;
@@ -148,7 +148,7 @@ void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
                     }
                 }
 
-                z[i] = indexValue.index;
+                z[i] = (Z) indexValue.index;
             }
         }
             break;
@@ -176,7 +176,7 @@ void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
                     }
                 }
 
-                z[i] = indexValue.index;
+                z[i] = (Z) indexValue.index;
             }
         }
             break;
@@ -206,7 +206,7 @@ void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
                     }
                 }
 
-                z[i] = indexValue.index;
+                z[i] = (Z) indexValue.index;
             }
         }
             break;
@@ -227,7 +227,7 @@ void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
                 }
 
                 auto zOffset = shape::indexOffset(i, zShapeInfo, castZShapeInfo, zLen, canCastZ);
-                z[zOffset] = indexValue.index;
+                z[zOffset] = (Z) indexValue.index;
             }
         }
             break;
@@ -248,7 +248,7 @@ void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
                     indexValue = OpType::update(indexValue, comp, extraParams);
                 }
 
-                z[i * zEws] = indexValue.index;
+                z[i * zEws] = (Z) indexValue.index;
             }
         }
             break;
@@ -272,18 +272,19 @@ void nd4j::IndexReductionLoops<X>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
                 }
 
                 auto zOffset = shape::indexOffset(i, zShapeInfo, castZShapeInfo, zLen, canCastZ);
-                z[zOffset] = indexValue.index;
+                z[zOffset] = (Z) indexValue.index;
             }
         }
     }
 }
 
-template <typename X>
-void nd4j::IndexReductionLoops<X>::wrapIndexReduce(const int opNum, void* vx, Nd4jLong* xShapeInfo, Nd4jLong* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, void* vextraParams) {
+template <typename X, typename Y>
+void nd4j::IndexReductionLoops<X, Y>::wrapIndexReduce(const int opNum, void* vx, Nd4jLong* xShapeInfo, void* vz, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, void* vextraParams) {
     auto x = reinterpret_cast<X *>(vx);
+    auto z = reinterpret_cast<Y *>(vz);
     auto extraParams = reinterpret_cast<X *>(vextraParams);
 
-    DISPATCH_BY_OPNUM_T(loopIndexReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams), INDEX_REDUCE_OPS);
+    DISPATCH_BY_OPNUM_TT(loopIndexReduce, PARAMS(x, xShapeInfo, z, zShapeInfo, tadShapeInfo, tadOffsets, extraParams), INDEX_REDUCE_OPS);
 }
 
-BUILD_SINGLE_TEMPLATE(template void nd4j::IndexReductionLoops, ::wrapIndexReduce(const int opNum, void* vx, Nd4jLong* xShapeInfo, Nd4jLong* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, void* vextraParams), LIBND4J_TYPES);
\ No newline at end of file
+BUILD_DOUBLE_TEMPLATE(template void nd4j::IndexReductionLoops, ::wrapIndexReduce(const int opNum, void* vx, Nd4jLong* xShapeInfo, void* z, Nd4jLong* zShapeInfo, Nd4jLong* tadShapeInfo, Nd4jLong* tadOffsets, void* vextraParams), LIBND4J_TYPES, INDEXING_TYPES);
\ No newline at end of file
diff --git a/libnd4j/include/loops/cpu/indexreduce.cpp b/libnd4j/include/loops/cpu/indexreduce.cpp
index 951ac287b..5a7beee24 100644
--- a/libnd4j/include/loops/cpu/indexreduce.cpp
+++ b/libnd4j/include/loops/cpu/indexreduce.cpp
@@ -31,26 +31,27 @@ namespace functions   {
 namespace indexreduce {
 
 ////////////////////////////////////////////////////////////////////////
-template <typename X> Nd4jLong IndexReduce<X>::execScalar( const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams) {
-    RETURNING_DISPATCH_BY_OPNUM_T(execScalar, PARAMS(x, xShapeInfo, extraParams), INDEX_REDUCE_OPS);
+template <typename X, typename Y>
+Nd4jLong IndexReduce<X,Y>::execScalar( const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams) {
+    RETURNING_DISPATCH_BY_OPNUM_TT(execScalar, PARAMS(x, xShapeInfo, extraParams), INDEX_REDUCE_OPS);
 }
 
 ////////////////////////////////////////////////////////////////////////
-template <typename X>
-void IndexReduce<X>::exec(const int opNum,
+template <typename X, typename Y>
+void IndexReduce<X,Y>::exec(const int opNum,
                         void *x,  Nd4jLong *xShapeInfo,
                         void *extraParams,
-                        Nd4jLong *z, Nd4jLong *zShapeInfo,
+                        void *z, Nd4jLong *zShapeInfo,
                         int *dimension, int dimensionLength,
                         Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset) {
 
-DISPATCH_BY_OPNUM_T(exec, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffset), INDEX_REDUCE_OPS);
+DISPATCH_BY_OPNUM_TT(exec, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffset), INDEX_REDUCE_OPS);
 }
 
 ////////////////////////////////////////////////////////////////////////
-template <typename X>
+template <typename X, typename Y>
 template<typename OpType>
-Nd4jLong IndexReduce<X>::execScalar(void *vx, Nd4jLong *xShapeInfo, void *vextraParams) {
+Nd4jLong IndexReduce<X, Y>::execScalar(void *vx, Nd4jLong *xShapeInfo, void *vextraParams) {
 
     auto x = reinterpret_cast<X *>(vx);
     auto extraParams = reinterpret_cast<X *>(vextraParams);
@@ -105,15 +106,16 @@ Nd4jLong IndexReduce<X>::execScalar(void *vx, Nd4jLong *xShapeInfo, void *vextra
 
 
 ////////////////////////////////////////////////////////////////////////
-template <typename X>
+template <typename X, typename Z>
 template<typename OpType>
-void IndexReduce<X>::exec(void *vx, Nd4jLong *xShapeInfo,
+void IndexReduce<X, Z>::exec(void *vx, Nd4jLong *xShapeInfo,
                         void *vextraParams,
-                        Nd4jLong *z, Nd4jLong *zShapeInfo,
+                        void *vz, Nd4jLong *zShapeInfo,
                         int *dimension, int dimensionLength,
                         Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset) {
 
     auto x = reinterpret_cast<X *>(vx);
+    auto z = reinterpret_cast<Z *>(vz);
     auto extraParams = reinterpret_cast<X *>(vextraParams);
 
     const Nd4jLong zLen = shape::length(zShapeInfo);
@@ -124,12 +126,12 @@ void IndexReduce<X>::exec(void *vx, Nd4jLong *xShapeInfo,
         const auto indexValue = OpType::startingIndexValue(x);
         PRAGMA_OMP_PARALLEL_FOR_IF(zLen > nd4j::Environment::getInstance()->elementwiseThreshold())
         for (uint i = 0; i < zLen; i++)
-            z[i] = indexValue.index;;
+            z[i] = (Z) indexValue.index;;
         return;
     }
 
     if(shape::isScalar(zShapeInfo)) {
-        z[0] = execScalar<OpType>(x,xShapeInfo,extraParams);
+        z[0] = (Z) execScalar<OpType>(x,xShapeInfo,extraParams);
         return;
     }
 
@@ -146,11 +148,11 @@ void IndexReduce<X>::exec(void *vx, Nd4jLong *xShapeInfo,
         tadOffsets = tadPack.primaryOffsets();
     }
 
-    nd4j::IndexReductionLoops<X>::template loopIndexReduce<OpType>(x, xShapeInfo, z, zShapeInfo,  tadOnlyShapeInfo, tadOffsets, extraParams);
+    nd4j::IndexReductionLoops<X,Z>::template loopIndexReduce<OpType>(x, xShapeInfo, z, zShapeInfo,  tadOnlyShapeInfo, tadOffsets, extraParams);
 }
 
 
-BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT IndexReduce, , LIBND4J_TYPES);
+BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT IndexReduce, , LIBND4J_TYPES, INDEXING_TYPES);
 
 }
 }
\ No newline at end of file
diff --git a/libnd4j/include/loops/cuda/indexreduce.cu b/libnd4j/include/loops/cuda/indexreduce.cu
index 18e5b1432..5f0cf07ae 100644
--- a/libnd4j/include/loops/cuda/indexreduce.cu
+++ b/libnd4j/include/loops/cuda/indexreduce.cu
@@ -29,37 +29,37 @@
 using namespace simdOps;
 
 
-template <typename T>
+template <typename X, typename Z>
 static __global__ void simpleIndexReduceGeneric(const int op,
                                            void *dx,
                                            Nd4jLong *xShapeInfo, int xRank,
                                            void *extraParams,
-                                           Nd4jLong *result,
+                                           void *result,
                                            Nd4jLong *resultShapeInfo, int zRank,
                                            int *dimension,
                                            int dimensionLength,
                                            int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets) {
 
-     functions::indexreduce::IndexReduce<T>::transform(op,dx,xShapeInfo,extraParams,result,resultShapeInfo,dimension,dimensionLength,postProcessOrNot,allocationBuffer,reductionBuffer,tadOnlyShapeInfo,tadOffsets);
+     functions::indexreduce::IndexReduce<X, Z>::transform(op,dx,xShapeInfo,extraParams,result,resultShapeInfo,dimension,dimensionLength,postProcessOrNot,allocationBuffer,reductionBuffer,tadOnlyShapeInfo,tadOffsets);
 }
 
 namespace functions {
     namespace indexreduce {
 
-        template <typename T>
-        _CUDA_H void IndexReduce<T>::executeIndexReduceScalar(dim3 launchDims, cudaStream_t *stream,
+        template <typename X, typename Z>
+        _CUDA_H void IndexReduce<X,Z>::executeIndexReduceScalar(dim3 launchDims, cudaStream_t *stream,
                                                                 const int opNum,
                                                                 void *dx, Nd4jLong *xShapeInfo,
                                                                 int xRank,
                                                                 void *extraParams,
-                                                                Nd4jLong *result, Nd4jLong *resultShapeInfo,
+                                                                void *result, Nd4jLong *resultShapeInfo,
                                                                 int zRank,
                                                                 int *dimension, int dimensionLength,
                                                                 int postProcessOrNot,
                                                                 int *allocationBuffer, void *reductionBuffer,
                                                                 Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets) {
 
-            simpleIndexReduceGeneric<T><<<launchDims.x,launchDims.y,launchDims.z, *stream>>>(opNum,
+            simpleIndexReduceGeneric<X, Z><<<launchDims.x,launchDims.y,launchDims.z, *stream>>>(opNum,
                                                                                             dx, xShapeInfo, xRank,
                                                                                             extraParams,
                                                                                             result, resultShapeInfo, 0,
@@ -67,13 +67,11 @@ namespace functions {
                                                                                             1,
                                                                                             allocationBuffer, reductionBuffer,
                                                                                             tadOnlyShapeInfo, tadOffsets);
-
-            nd4j::DebugHelper::checkErrorCode(stream, "execIndexReduceScalar(...) failed");
         }
 
-        template <typename T>
-        _CUDA_H void IndexReduce<T>::executeIndexReduce(dim3 launchDims, cudaStream_t *stream, const int opNum, void *dx, Nd4jLong *xShapeInfo, int xRank, void *extraParams, Nd4jLong *result, Nd4jLong *resultShapeInfo, int zRank, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets) {
-            simpleIndexReduceGeneric<T><<<launchDims.x,launchDims.y,launchDims.z, *stream>>>(
+        template <typename X, typename Z>
+        _CUDA_H void IndexReduce<X, Z>::executeIndexReduce(dim3 launchDims, cudaStream_t *stream, const int opNum, void *dx, Nd4jLong *xShapeInfo, int xRank, void *extraParams, void *result, Nd4jLong *resultShapeInfo, int zRank, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets) {
+            simpleIndexReduceGeneric<X, Z><<<launchDims.x,launchDims.y,launchDims.z, *stream>>>(
 			 opNum,
 			 dx,
 			 xShapeInfo, xRank,
@@ -83,8 +81,6 @@ namespace functions {
 			 dimension,
 			 dimensionLength,
 			 1, allocationBuffer, reductionBuffer, tadOnlyShapeInfo, tadOffsets);
-
-            DEBUG_KERNEL(stream, opNum);
         }
 
         // This is the un-specialized struct.  Note that we prevent instantiation of this
@@ -122,14 +118,14 @@ namespace functions {
             }
         };
 
-        template <typename T>
+        template <typename X, typename Z>
         template <typename OpType>
-        __device__ void IndexReduce<T>::aggregatePartials(IndexValue<T> **sPartialsRef, Nd4jLong tid, Nd4jLong numElements, void *vextraParams) {
+        __device__ void IndexReduce<X, Z>::aggregatePartials(IndexValue<X> **sPartialsRef, Nd4jLong tid, Nd4jLong numElements, void *vextraParams) {
             // start the shared memory loop on the next power of 2 less
             // than the block size.  If block size is not a power of 2,
             // accumulate the intermediate sums in the remainder range.
-            auto extraParams = static_cast<T*>(vextraParams);
-            IndexValue<T> *sPartials = *sPartialsRef;
+            auto extraParams = static_cast<X*>(vextraParams);
+            IndexValue<X> *sPartials = *sPartialsRef;
             Nd4jLong floorPow2 = blockDim.x;
 
             if (floorPow2 & (floorPow2 - 1)) {
@@ -138,8 +134,8 @@ namespace functions {
                 }
 
                 if (tid >= floorPow2) {
-                    IndexValue<T> prev = sPartials[tid - floorPow2];
-                    IndexValue<T> curr = sPartials[tid];
+                    IndexValue<X> prev = sPartials[tid - floorPow2];
+                    IndexValue<X> curr = sPartials[tid];
                     sPartials[tid - floorPow2] = OpType::update(prev,curr,extraParams);
                 }
                 __syncthreads();
@@ -147,21 +143,21 @@ namespace functions {
 
             for (int activeThreads = floorPow2 >> 1;activeThreads; activeThreads >>= 1) {
                 if (tid < activeThreads && tid + activeThreads < numElements) {
-                    IndexValue<T> curr = sPartials[tid];
-                    IndexValue<T> next = sPartials[tid + activeThreads];
+                    IndexValue<X> curr = sPartials[tid];
+                    IndexValue<X> next = sPartials[tid + activeThreads];
                     sPartials[tid] = OpType::update(curr,next,extraParams);
                 }
                 __syncthreads();
             }
         }
 
-        template <typename X>
-        __device__ void IndexReduce<X>::transform(
+        template <typename X, typename Y>
+        __device__ void IndexReduce<X, Y>::transform(
                 const int opNum,
                 void *x,
                 Nd4jLong *xShapeInfo,
                 void *extraParams,
-                Nd4jLong *result,
+                void *result,
                 Nd4jLong *resultShapeInfo,
                 int *dimension,
                 int dimensionLength,
@@ -170,15 +166,15 @@ namespace functions {
                 void *reductionBuffer,
                 Nd4jLong *tadShapeInfo,
                 Nd4jLong *tadOffset) {
-             DISPATCH_BY_OPNUM_T(transform, PARAMS(x, xShapeInfo, extraParams, result, resultShapeInfo, dimension, dimensionLength, postProcessOrNot, allocationBuffer, reductionBuffer, tadShapeInfo, tadOffset), INDEX_REDUCE_OPS);
+             DISPATCH_BY_OPNUM_TT(transform, PARAMS(x, xShapeInfo, extraParams, result, resultShapeInfo, dimension, dimensionLength, postProcessOrNot, allocationBuffer, reductionBuffer, tadShapeInfo, tadOffset), INDEX_REDUCE_OPS);
         }
 
 
-        template <typename T>
+        template <typename X, typename Z>
         template <typename OpType>
-        __device__ void IndexReduce<T>::transform(void *vdx, Nd4jLong *xShapeInfo,
+        __device__ void IndexReduce<X, Z>::transform(void *vdx, Nd4jLong *xShapeInfo,
                                                 void *vextraParams,
-                                                Nd4jLong *result, Nd4jLong *resultShapeInfo,
+                                                void *vresult, Nd4jLong *resultShapeInfo,
                                                 int *dimension, int dimensionLength,
                                                 int postProcessOrNot,
                                                 int *allocationBuffer, void *vreductionBuffer,
@@ -186,18 +182,19 @@ namespace functions {
             /**int
              * Gpu information for the problem
              */
-            auto dx = static_cast<T*>(vdx);
-            auto extraParams = static_cast<T*>(vextraParams);
-            auto reductionBuffer = static_cast<T*>(vreductionBuffer);
+            auto dx = reinterpret_cast<X*>(vdx);
+            auto result = reinterpret_cast<Z*>(vresult);
+            auto extraParams = static_cast<X*>(vextraParams);
+            auto reductionBuffer = static_cast<X*>(vreductionBuffer);
             auto order = shape::order(xShapeInfo);
             int tid = blockIdx.x * blockDim.x + threadIdx.x;
             __shared__ volatile int resultScalar;
 
             //shared memory space for storing intermediate results
-            __shared__ IndexValue<T>* sPartials;
+            __shared__ IndexValue<X>* sPartials;
             if(threadIdx.x == 0) {
                 extern __shared__ unsigned char shmem[];
-                sPartials = reinterpret_cast<IndexValue<T>*>(shmem);
+                sPartials = reinterpret_cast<IndexValue<X>*>(shmem);
             }
             __syncthreads();
 
@@ -210,7 +207,7 @@ namespace functions {
 
 
             //only compute the tad indexes once
-            IndexValue <T> reduction = OpType::startingIndexValue(dx);
+            IndexValue<X> reduction = OpType::startingIndexValue(dx);
 
             if (threadIdx.x == 0) {
                 if (resultShapeInfo != nullptr)
@@ -255,7 +252,7 @@ namespace functions {
 
                         for(int i = threadIdx.x;i < tadLength; i += blockDim.x) {                            
                             auto xOffset = tadOffsetForBlock + shape::getIndexOffset(i, tadOnlyShapeInfo, tadLength);
-                            IndexValue<T> comp {dx[xOffset], i};
+                            IndexValue<X> comp {dx[xOffset], i};
                             sPartials[threadIdx.x] = OpType::update(sPartials[threadIdx.x], comp, extraParams);
                         }
 
@@ -264,7 +261,7 @@ namespace functions {
 
                         __syncthreads();
                         if (threadIdx.x == 0) {
-                            result[r] = sPartials[threadIdx.x].index;
+                            result[r] = (Z) sPartials[threadIdx.x].index;
                         }
                         __syncthreads();
                     }
@@ -276,7 +273,7 @@ namespace functions {
                         sPartials[threadIdx.x] = OpType::startingIndexValue(dx);
 
                         for (int x = threadIdx.x; x < tadLength; x+= blockDim.x) {
-                            IndexValue<T> comp {dx[tadOffsetForBlock + x * tadEWS], x};
+                            IndexValue<X> comp {dx[tadOffsetForBlock + x * tadEWS], x};
                             sPartials[threadIdx.x] =  OpType::update(sPartials[threadIdx.x], comp, extraParams);
                         }
 
@@ -285,7 +282,7 @@ namespace functions {
 
                         __syncthreads();
                         if (threadIdx.x == 0) {
-                            result[i] = sPartials[threadIdx.x].index; //postProcess(sPartials[0],tadLength ,extraParams);
+                            result[i] = (Z) sPartials[threadIdx.x].index; //postProcess(sPartials[0],tadLength ,extraParams);
                         }
                         __syncthreads();
                     }
@@ -296,14 +293,14 @@ namespace functions {
 
                 if(xElementWiseStride >= 1 && order == 'c') {
                     for(Nd4jLong i = tid;i < n; i += (blockDim.x * gridDim.x)) {
-                        IndexValue <T> indexVal = {dx[i * xElementWiseStride], i};
+                        IndexValue<X> indexVal = {dx[i * xElementWiseStride], i};
                         reduction = OpType::update(reduction, indexVal, extraParams);
                     }
                 } else {
                                         
                     for(Nd4jLong i = tid;i < n; i += blockDim.x * gridDim.x) {                                                
                         auto offset = shape::getIndexOffset(i, xShapeInfo, n);
-                        IndexValue <T> indexVal = {dx[offset], i};
+                        IndexValue<X> indexVal = {dx[offset], i};
                         reduction = OpType::update(reduction, indexVal, extraParams);
                     }
                 }
@@ -320,7 +317,7 @@ namespace functions {
                     unsigned int *tc = (unsigned int *) reductionBuffer;
                     tid = threadIdx.x;
                     if (threadIdx.x == 0) {
-                        auto pBuffer = reinterpret_cast<IndexValue<T> *>(reductionBuffer);
+                        auto pBuffer = reinterpret_cast<IndexValue<X> *>(reductionBuffer);
                         pBuffer[blockIdx.x] = {sPartials[0].value, sPartials[0].index};
                     }
                     __threadfence();
@@ -335,7 +332,7 @@ namespace functions {
 
                     if (amLast) {
                         tc[16384] = 0;
-                        IndexValue<T> *pBuffer = (IndexValue<T> *) reductionBuffer;
+                        IndexValue<X> *pBuffer = (IndexValue<X> *) reductionBuffer;
 
                         sPartials[threadIdx.x] = OpType::startingIndexValue(dx);
 
@@ -348,14 +345,14 @@ namespace functions {
 
                         __syncthreads();
                         if (tid == 0) {
-                            result[0] = sPartials[0].index;
+                            result[0] = (Z) sPartials[0].index;
                         }
                     }
                 } else {
                     if (tid == 0) {
                         auto tc = reinterpret_cast<unsigned int *>(reductionBuffer);
                         tc[16384] = 0;
-                        result[0] = sPartials[0].index;
+                        result[0] = (Z) sPartials[0].index;
                     }
                 }
 
@@ -365,30 +362,30 @@ namespace functions {
 
 
 
-        template <typename T>
-        Nd4jLong IndexReduce<T>::execScalar(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams) {
+        template <typename X, typename Z>
+        Nd4jLong IndexReduce<X,Z>::execScalar(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams) {
             return 0;
         }
 
-        template <typename T>
-        void IndexReduce<T>::exec(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams, Nd4jLong *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset) {
+        template <typename X, typename Z>
+        void IndexReduce<X,Z>::exec(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams, void *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset) {
 
         }
 
-        template <typename T>
+        template <typename X, typename Z>
         template<typename OpType>
-        Nd4jLong IndexReduce<T>:: execScalar(void *x, Nd4jLong *xShapeInfo, void *extraParams) {
+        Nd4jLong IndexReduce<X,Z>:: execScalar(void *x, Nd4jLong *xShapeInfo, void *extraParams) {
             return 0;
         }
 
-        template <typename T>
+        template <typename X, typename Z>
         template<typename OpType>
-        _CUDA_H void IndexReduce<T>::exec(void *x, Nd4jLong *xShapeInfo, void *extraParams, Nd4jLong *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset) {
+        _CUDA_H void IndexReduce<X,Z>::exec(void *x, Nd4jLong *xShapeInfo, void *extraParams, void *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset) {
 
         }
 
 
-        BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT IndexReduce, , LIBND4J_TYPES);
+        BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT IndexReduce, , LIBND4J_TYPES, INDEXING_TYPES);
     }
 }
 
diff --git a/libnd4j/include/loops/indexreduce.h b/libnd4j/include/loops/indexreduce.h
index 40f98c692..792ed16a9 100755
--- a/libnd4j/include/loops/indexreduce.h
+++ b/libnd4j/include/loops/indexreduce.h
@@ -52,35 +52,35 @@
 namespace functions {
 	namespace indexreduce {
 
-		template<typename T>
+		template<typename X, typename Z>
 		class IndexReduce {
 		public:
 #ifdef __CUDACC__
 
-	static __device__ void transform(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams, Nd4jLong *result, Nd4jLong *resultShapeInfo, int *dimension,int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset);
+	static __device__ void transform(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams, void *result, Nd4jLong *resultShapeInfo, int *dimension,int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset);
 
     template<typename OpType>
-	static __device__ void aggregatePartials(IndexValue<T> **sPartialsRef, Nd4jLong tid, Nd4jLong numElements,void *extraParams);
+	static __device__ void aggregatePartials(IndexValue<X> **sPartialsRef, Nd4jLong tid, Nd4jLong numElements,void *extraParams);
 
 
     template<typename OpType>
-	static __device__ void transform(void *dx, Nd4jLong *xShapeInfo, void *extraParams, Nd4jLong *result, Nd4jLong *resultShapeInfo, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets);
+	static __device__ void transform(void *dx, Nd4jLong *xShapeInfo, void *extraParams, void *result, Nd4jLong *resultShapeInfo, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets);
 
 
-    static _CUDA_H void executeIndexReduceScalar(dim3 launchDims, cudaStream_t *stream, const int op, void *dx, Nd4jLong *xShapeInfo, int xRank, void *extraParams, Nd4jLong *result, Nd4jLong *resultShapeInfo, int zRank, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets);
+    static _CUDA_H void executeIndexReduceScalar(dim3 launchDims, cudaStream_t *stream, const int op, void *dx, Nd4jLong *xShapeInfo, int xRank, void *extraParams, void *result, Nd4jLong *resultShapeInfo, int zRank, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets);
 
-    static _CUDA_H void executeIndexReduce(dim3 launchDims, cudaStream_t *stream, const int op, void *dx, Nd4jLong *xShapeInfo, int xRank, void *extraParams, Nd4jLong *result, Nd4jLong *resultShapeInfo, int zRank, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets);
+    static _CUDA_H void executeIndexReduce(dim3 launchDims, cudaStream_t *stream, const int op, void *dx, Nd4jLong *xShapeInfo, int xRank, void *extraParams, void *result, Nd4jLong *resultShapeInfo, int zRank, int *dimension, int dimensionLength, int postProcessOrNot, int *allocationBuffer, void *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets);
 #endif
 
 		static Nd4jLong execScalar(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams);
 
-		static void exec(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams, Nd4jLong *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset);
+		static void exec(const int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams, void *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset);
 
 		template<typename OpType>
 		static _CUDA_H Nd4jLong execScalar(void *x, Nd4jLong *xShapeInfo, void *extraParams);
 
 		template<typename OpType>
-		static _CUDA_H void exec(void *x, Nd4jLong *xShapeInfo, void *extraParams, Nd4jLong *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset);
+		static _CUDA_H void exec(void *x, Nd4jLong *xShapeInfo, void *extraParams, void *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset);
 		};
 	}
 }
diff --git a/libnd4j/include/ops/declarable/generic/parity_ops/top_k.cpp b/libnd4j/include/ops/declarable/generic/parity_ops/top_k.cpp
index ea2e3330a..bd16cdd79 100644
--- a/libnd4j/include/ops/declarable/generic/parity_ops/top_k.cpp
+++ b/libnd4j/include/ops/declarable/generic/parity_ops/top_k.cpp
@@ -87,7 +87,7 @@ namespace nd4j {
             getOpDescriptor()
                     ->setAllowedInputTypes(nd4j::DataType::ANY)
                     ->setAllowedOutputTypes(0, nd4j::DataType::ANY)
-                    ->setAllowedOutputTypes(1, {ALL_INTS});
+                    ->setAllowedOutputTypes(1, {ALL_INDICES});
         }
     }
 }
diff --git a/libnd4j/include/ops/declarable/helpers/impl/listdiff.cpp b/libnd4j/include/ops/declarable/helpers/impl/listdiff.cpp
index baa08dad9..c840f6960 100644
--- a/libnd4j/include/ops/declarable/helpers/impl/listdiff.cpp
+++ b/libnd4j/include/ops/declarable/helpers/impl/listdiff.cpp
@@ -42,10 +42,11 @@ namespace helpers {
     Nd4jLong listDiffCount(nd4j::LaunchContext * context, NDArray* values, NDArray* keep) {
         auto xType = values->dataType();
 
-        values->syncToHost();
-        keep->syncToHost();
+        NDArray::preparePrimaryUse({},{values, keep});
 
         BUILD_SINGLE_SELECTOR(xType, return listDiffCount_, (values, keep), LIBND4J_TYPES);
+
+        NDArray::registerPrimaryUse({},{values, keep});
     }
 
     BUILD_SINGLE_TEMPLATE(template Nd4jLong listDiffCount_, (NDArray* values, NDArray* keep);, LIBND4J_TYPES);
@@ -97,16 +98,7 @@ namespace helpers {
     int listDiffFunctor(nd4j::LaunchContext * context, NDArray* values, NDArray* keep, NDArray* output1, NDArray* output2) {
         auto xType = values->dataType();
 
-        values->syncToHost();
-
-        if (keep != nullptr)
-            keep->syncToHost();
-
-        if (output1 != nullptr)
-            output1->syncToHost();
-
-        if (output2 != nullptr)
-            output2->syncToHost();
+        NDArray::preparePrimaryUse({output1, output2}, {values, keep});
 
         int result = 0;
 
@@ -118,14 +110,7 @@ namespace helpers {
             throw std::runtime_error("ListDiff: Only integer and floating point data types are supported");
         }
 
-        if (keep != nullptr)
-            keep->syncToDevice();
-
-        if (output1 != nullptr)
-            output1->syncToDevice();
-
-        if (output2 != nullptr)
-            output2->syncToDevice();
+        NDArray::registerPrimaryUse({output1, output2}, {values, keep});
 
         return result;
     }
diff --git a/libnd4j/include/ops/ops.h b/libnd4j/include/ops/ops.h
index 38122f985..fe6bfae81 100644
--- a/libnd4j/include/ops/ops.h
+++ b/libnd4j/include/ops/ops.h
@@ -3746,7 +3746,7 @@ namespace simdOps {
 	};
 
 
-	template <typename X>
+	template <typename X, typename Z>
 	class IndexAbsoluteMax  {
 	public:
 		static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> val, X *extraParams) {
@@ -3799,7 +3799,7 @@ namespace simdOps {
 		}
 	};
 
-    template <typename X>
+    template <typename X, typename Z>
     class FirstIndex {
     public:
         static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> val, X *extraParams) {
@@ -3861,7 +3861,7 @@ namespace simdOps {
     };
 
 
-    template <typename X>
+    template <typename X, typename Z>
     class LastIndex {
     public:
         static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> val, X *extraParams) {
@@ -3920,7 +3920,7 @@ namespace simdOps {
     };
 
 
-	template <typename X>
+	template <typename X, typename Z>
 	class IndexMax  {
 	public:
 
@@ -3974,7 +3974,7 @@ namespace simdOps {
 	};
 
 
-	template <typename X>
+	template <typename X, typename Z>
 	class IndexAbsoluteMin {
 	public:
 		static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(
@@ -4030,7 +4030,7 @@ namespace simdOps {
 	};
 
 
-	template <typename X>
+	template <typename X, typename Z>
 	class IndexMin {
 	public:
         static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
index 38a1ba382..54649692c 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
@@ -226,6 +226,21 @@ public class CudaExecutioner extends DefaultOpExecutioner {
      */
     protected INDArray naiveExec(ReduceOp op, int... dimension) {
         long st = profilingConfigurableHookIn(op);
+
+        if(op instanceof BaseReduceOp && ((BaseReduceOp)op).isEmptyReduce()){
+            //Edge case for TF import compatibility: [x,y].reduce(empty) = [x,y]
+            //Note that "empty" axis is NOT the same as length 0, as in INDArray.sum(new int[0]), which means "all dimensions"
+            if(op.z() != null){
+                Preconditions.checkState(op.x().equalShapes(op.z()), "For empty reductions, result (z) array must have same shape as x shape." +
+                        " Got: x=%ndShape, z=%ndShape", op.x(), op.z());
+                op.z().assign(op.x());
+                return op.z();
+            } else {
+                op.setZ(op.x().dup());
+                return op.z();
+            }
+        }
+
         INDArray ret = op.z();
 
         checkForCompression(op);
@@ -482,6 +497,20 @@ public class CudaExecutioner extends DefaultOpExecutioner {
     public INDArray exec(ReduceOp op) {
         checkForCompression(op);
 
+        if(op instanceof BaseReduceOp && ((BaseReduceOp)op).isEmptyReduce()){
+            //Edge case for TF import compatibility: [x,y].reduce(empty) = [x,y]
+            //Note that "empty" axis is NOT the same as length 0, as in INDArray.sum(new int[0]), which means "all dimensions"
+            if(op.z() != null){
+                Preconditions.checkState(op.x().equalShapes(op.z()), "For empty reductions, result (z) array must have same shape as x shape." +
+                        " Got: x=%ndShape, z=%ndShape", op.x(), op.z());
+                op.z().assign(op.x());
+                return op.z();
+            } else {
+                op.setZ(op.x().dup());
+                return op.z();
+            }
+        }
+
         val dimension = op.dimensions().toIntVector();
 
         if (extraz.get() == null)
@@ -890,6 +919,22 @@ public class CudaExecutioner extends DefaultOpExecutioner {
 
 
     protected CudaContext invoke(ReduceOp op, int[] dimension) {
+        CudaContext context = AtomicAllocator.getInstance().getFlowController().prepareAction(op.z(), op.x(), op.y());
+
+        if(op instanceof BaseReduceOp && ((BaseReduceOp)op).isEmptyReduce()){
+            //Edge case for TF import compatibility: [x,y].reduce(empty) = [x,y]
+            //Note that "empty" axis is NOT the same as length 0, as in INDArray.sum(new int[0]), which means "all dimensions"
+            if(op.z() != null){
+                Preconditions.checkState(op.x().equalShapes(op.z()), "For empty reductions, result (z) array must have same shape as x shape." +
+                        " Got: x=%ndShape, z=%ndShape", op.x(), op.z());
+                op.z().assign(op.x());
+                return context;
+            } else {
+                op.setZ(op.x().dup());
+                return context;
+            }
+        }
+
         long st = profilingConfigurableHookIn(op);
 
         checkForCompression(op);
@@ -913,8 +958,6 @@ public class CudaExecutioner extends DefaultOpExecutioner {
                 throw new ND4JIllegalStateException("Op target dimension " + Arrays.toString(dimension)
                         + " contains element that higher then rank of op.X: [" + op.x().rank() + "]");
 
-        CudaContext context = AtomicAllocator.getInstance().getFlowController().prepareAction(op.z(), op.x(), op.y());
-
         if (CudaEnvironment.getInstance().getConfiguration().isDebug())
             lastOp.set(op.opName());
 
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/custom/CustomOpsTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/custom/CustomOpsTests.java
index 6c4595a0c..f325348fb 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/custom/CustomOpsTests.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/custom/CustomOpsTests.java
@@ -733,6 +733,46 @@ public class CustomOpsTests extends BaseNd4jTest {
         }
     }
 
+    @Test
+    public void testListDiff(){
+        INDArray x = Nd4j.createFromArray(0, 1, 2, 3);
+        INDArray y = Nd4j.createFromArray(3, 1);
+
+        INDArray out = Nd4j.create(DataType.INT, 2);
+        INDArray outIdx = Nd4j.create(DataType.INT, 2);
+
+        Nd4j.exec(DynamicCustomOp.builder("listdiff")
+                .addInputs(x, y)
+                .addOutputs(out, outIdx)
+                .build());
+
+        INDArray exp = Nd4j.createFromArray(0, 2);
+
+        assertEquals(exp, out);         //Values in x not in y
+        assertEquals(exp, outIdx);      //Indices of the values in x not in y
+    }
+
+    @Test
+    public void testTopK1(){
+        INDArray x = Nd4j.createFromArray(0.0, 0.0, 0.0, 10.0, 0.0);
+        INDArray k = Nd4j.scalar(1);
+        INDArray outValue = Nd4j.create(DataType.DOUBLE, 1);
+        INDArray outIdx = Nd4j.create(DataType.INT, 1);
+
+        Nd4j.exec(DynamicCustomOp.builder("top_k")
+                .addInputs(x, k)
+                .addOutputs(outValue, outIdx)
+                .addBooleanArguments(false) //not sorted
+                .addIntegerArguments(1)
+                .build());
+
+        INDArray expValue = Nd4j.createFromArray(10.0);
+        INDArray expIdx = Nd4j.createFromArray(3);
+
+        assertEquals(expValue, outValue);
+        assertEquals(expIdx, outIdx);
+    }
+
     @Test
     public void testMaxPool2Dbp_1() {
         val x = Nd4j.create(DataType.HALF, 2,3,16,16).assign(Double.NaN);
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/EmptyTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/EmptyTests.java
index 261e1e300..e7e8f8288 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/EmptyTests.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/EmptyTests.java
@@ -25,6 +25,7 @@ import org.nd4j.linalg.BaseNd4jTest;
 import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.api.ops.DynamicCustomOp;
+import org.nd4j.linalg.api.ops.impl.reduce.bool.All;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.factory.Nd4jBackend;
 
@@ -299,6 +300,15 @@ public class EmptyTests extends BaseNd4jTest {
         assertNotNull(result[0].shapeInfoDataBuffer().asLong());
     }
 
+    @Test
+    public void testAllEmptyReduce(){
+        INDArray x = Nd4j.createFromArray(true, true, true);
+        val all = new All(x);
+        all.setEmptyReduce(true);   //For TF compatibility - empty array for axis (which means no-op - and NOT all array reduction)
+        INDArray out = Nd4j.exec(all);
+        assertEquals(x, out);
+    }
+
     @Override
     public char ordering() {
         return 'c';

From 05d45ec0508def2034510b373a0f61d1abe8bf0e Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Tue, 27 Aug 2019 11:31:59 +0300
Subject: [PATCH 15/56] IndexReduce along dim CUDA fix

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java  | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
index 54649692c..afeff4d8b 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
@@ -894,13 +894,13 @@ public class CudaExecutioner extends DefaultOpExecutioner {
 
             //long dimensionPointer = AtomicAllocator.getInstance().getPointer(Nd4j.createBuffer(dimension), context);
             Pointer dimensionPointer = AtomicAllocator.getInstance()
-                    .getPointer(AtomicAllocator.getInstance().getConstantBuffer(dimension), context);
+                    .getHostPointer(AtomicAllocator.getInstance().getConstantBuffer(dimension));
 
             nativeOps.execIndexReduce(xShapeInfoHostPointer, op.opNum(),
                     null, (LongPointer) hostXShapeInfo, x, (LongPointer) xShapeInfo,
                      extraArgs,
                     null, (LongPointer) hostZShapeInfo, z, (LongPointer) zShapeInfo,
-                    null,
+                    dimensionPointer,
                     (LongPointer) op.dimensions().shapeInfoDataBuffer().addressPointer(),
                     AtomicAllocator.getInstance().getPointer(op.dimensions(), context),
                     null);

From dce4751fc12065a285af294dda299e54a6468add Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Tue, 27 Aug 2019 18:34:53 +1000
Subject: [PATCH 16/56] Layer norm 4d case fixes (#174)

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 .../DifferentialFunctionFactory.java          | 16 ++++----
 .../org/nd4j/autodiff/samediff/ops/SDNN.java  | 18 +++++----
 .../nd4j/linalg/api/ops/DynamicCustomOp.java  | 16 ++++++++
 .../ops/impl/transforms/custom/LayerNorm.java | 21 +++++-----
 .../impl/transforms/custom/LayerNormBp.java   | 31 ++++++++-------
 .../opvalidation/LayerOpValidation.java       | 38 +++++++++++++++++--
 6 files changed, 96 insertions(+), 44 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
index 3bf1754db..e16bd3dc2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
@@ -790,20 +790,20 @@ public class DifferentialFunctionFactory {
         return new StandardizeBp(sameDiff(), stdInput, gradient, dimensions).outputVariable();
     }
 
-    public SDVariable layerNorm(SDVariable input, SDVariable gain, SDVariable bias, int... dimensions) {
-        return new LayerNorm(sameDiff(), input, gain, bias, dimensions).outputVariable();
+    public SDVariable layerNorm(SDVariable input, SDVariable gain, SDVariable bias, boolean channelsFirst, int... dimensions) {
+        return new LayerNorm(sameDiff(), input, gain, bias, channelsFirst, dimensions).outputVariable();
     }
 
-    public SDVariable[] layerNormBp(SDVariable input, SDVariable gain, SDVariable bias, SDVariable gradient, int... dimensions) {
-        return new LayerNormBp(sameDiff(), input, gain, bias, gradient, dimensions).outputVariables();
+    public SDVariable[] layerNormBp(SDVariable input, SDVariable gain, SDVariable bias, SDVariable gradient, boolean channelsFirst, int... dimensions) {
+        return new LayerNormBp(sameDiff(), input, gain, bias, gradient, channelsFirst, dimensions).outputVariables();
     }
 
-    public SDVariable layerNorm(SDVariable input, SDVariable gain, int... dimensions) {
-        return new LayerNorm(sameDiff(), input, gain, dimensions).outputVariable();
+    public SDVariable layerNorm(SDVariable input, SDVariable gain, boolean channelsFirst, int... dimensions) {
+        return new LayerNorm(sameDiff(), input, gain, channelsFirst, dimensions).outputVariable();
     }
 
-    public SDVariable[] layerNormBp(SDVariable input, SDVariable gain, SDVariable gradient, int... dimensions) {
-        return new LayerNormBp(sameDiff(), input, gain, gradient, dimensions).outputVariables();
+    public SDVariable[] layerNormBp(SDVariable input, SDVariable gain, SDVariable gradient, boolean channelsFirst, int... dimensions) {
+        return new LayerNormBp(sameDiff(), input, gain, gradient, channelsFirst, dimensions).outputVariables();
     }
 
     public SDVariable squaredNorm(SDVariable input, boolean keepDims, int... dimensions) {
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDNN.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDNN.java
index 928bf3e6e..eb89a0f3a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDNN.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDNN.java
@@ -759,8 +759,8 @@ public class SDNN extends SDOps {
      *
      * @return Output variable
      */
-    public SDVariable layerNorm(SDVariable input, SDVariable gain, SDVariable bias, int... dimensions) {
-        return layerNorm(null, input, gain, bias, dimensions);
+    public SDVariable layerNorm(SDVariable input, SDVariable gain, SDVariable bias, boolean channelsFirst, int... dimensions) {
+        return layerNorm(null, input, gain, bias, channelsFirst, dimensions);
     }
 
     /**
@@ -772,13 +772,15 @@ public class SDNN extends SDOps {
      * @param input Input variable
      * @param gain gain
      * @param bias bias
+     * @param channelsFirst For 2D input - unused. True for NCHW (minibatch, channels, height, width), false for NHWC data
+     * @param dimensions Dimensions to perform layer norm over - dimension=1 for 2d/MLP data, dimension=1,2,3 for CNNs
      * @return Output variable
      */
-    public SDVariable layerNorm(String name, SDVariable input, SDVariable gain, SDVariable bias, int... dimensions) {
+    public SDVariable layerNorm(String name, SDVariable input, SDVariable gain, SDVariable bias, boolean channelsFirst, int... dimensions) {
         validateFloatingPoint("layerNorm", "input", input);
         validateFloatingPoint("layerNorm", "gain", gain);
         validateFloatingPoint("layerNorm", "bias", bias);
-        SDVariable result = f().layerNorm(input, gain, bias, dimensions);
+        SDVariable result = f().layerNorm(input, gain, bias, channelsFirst, dimensions);
         return updateVariableNameAndReference(result, name);
     }
 
@@ -789,8 +791,8 @@ public class SDNN extends SDOps {
      *
      * @return Output variable
      */
-    public SDVariable layerNorm(SDVariable input, SDVariable gain, int... dimensions) {
-        return layerNorm((String)null, input, gain, dimensions);
+    public SDVariable layerNorm(SDVariable input, SDVariable gain, boolean channelsFirst, int... dimensions) {
+        return layerNorm((String)null, input, gain, channelsFirst, dimensions);
     }
 
     /**
@@ -803,10 +805,10 @@ public class SDNN extends SDOps {
      * @param gain gain
      * @return Output variable
      */
-    public SDVariable layerNorm(String name, SDVariable input, SDVariable gain, int... dimensions) {
+    public SDVariable layerNorm(String name, SDVariable input, SDVariable gain, boolean channelsFirst, int... dimensions) {
         validateFloatingPoint("layerNorm", "input", input);
         validateFloatingPoint("layerNorm", "gain", gain);
-        SDVariable result = f().layerNorm(input, gain, dimensions);
+        SDVariable result = f().layerNorm(input, gain, channelsFirst, dimensions);
         return updateVariableNameAndReference(result, name);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java
index f52450eee..27e8ae281 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java
@@ -35,6 +35,7 @@ import org.tensorflow.framework.AttrValue;
 import org.tensorflow.framework.GraphDef;
 import org.tensorflow.framework.NodeDef;
 
+import java.lang.reflect.Array;
 import java.util.*;
 
 /**
@@ -611,6 +612,21 @@ public class DynamicCustomOp extends DifferentialFunction implements CustomOp {
         return in == null ? null : new INDArray[]{in};
     }
 
+    protected static <T> T[] wrapFilterNull(T... in){
+        int count = 0;
+        for( int i=0; i<in.length; i++ ) {
+            if (in[i] != null) count++;
+        }
+        T[] out = (T[]) Array.newInstance(in.getClass().getComponentType(), count);
+        int j=0;
+        for( int i=0; i<in.length; i++ ){
+            if(in[i] != null){
+                out[j++] = in[i];
+            }
+        }
+        return out;
+    }
+
     public static class DynamicCustomOpsBuilder {
         protected String opName;
         protected int numInputs;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java
index dde92a947..65222fe98 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java
@@ -17,6 +17,7 @@
 package org.nd4j.linalg.api.ops.impl.transforms.custom;
 
 import lombok.NoArgsConstructor;
+import lombok.NonNull;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -41,17 +42,16 @@ import java.util.List;
 public class LayerNorm extends DynamicCustomOp {
 
     private boolean noBias = false;
+    private boolean channelsFirst;
 
-    public LayerNorm(SameDiff sameDiff, SDVariable input, SDVariable gain, SDVariable bias, int... dimensions) {
-        super(null, sameDiff, new SDVariable[] {input, gain, bias}, false);
-        Preconditions.checkArgument(bias != null, "LayerNorm: Use constructor without bias argument if bias is null / not available.");
+    public LayerNorm(@NonNull SameDiff sameDiff, @NonNull SDVariable input, @NonNull SDVariable gain, SDVariable bias, boolean channelsFirst, int... dimensions) {
+        super(null, sameDiff, wrapFilterNull(input, gain, bias), false);
+        this.channelsFirst = channelsFirst;
         setDimensions(dimensions);
     }
 
-    public LayerNorm(SameDiff sameDiff, SDVariable input, SDVariable gain, int... dimensions) {
-        super(null, sameDiff, new SDVariable[] {input, gain}, false);
-        noBias = true;
-        setDimensions(dimensions);
+    public LayerNorm(SameDiff sameDiff, SDVariable input, SDVariable gain, boolean channelsFirst, int... dimensions) {
+        this(sameDiff, input, gain, null, channelsFirst, dimensions);
     }
 
     public LayerNorm(INDArray input, INDArray gain, INDArray bias, INDArray result, int... dimensions) {
@@ -73,7 +73,10 @@ public class LayerNorm extends DynamicCustomOp {
         Preconditions.checkArgument(dimensions.length > 0, "LayerNorm: You have to provide dimensions");
 
         this.dimensions = dimensions;
+        this.iArguments.clear();
         addIArgument(dimensions);
+        this.bArguments.clear();
+        this.bArguments.add(channelsFirst);
     }
 
     @Override
@@ -96,9 +99,9 @@ public class LayerNorm extends DynamicCustomOp {
     public List<SDVariable> doDiff(List<SDVariable> gradient) {
         SDVariable[] ret;
         if(noBias){
-            ret = f().layerNormBp(arg(0), arg(1), gradient.get(0), dimensions);
+            ret = f().layerNormBp(arg(0), arg(1), gradient.get(0), channelsFirst, dimensions);
         }else{
-            ret = f().layerNormBp(arg(0), arg(1), arg(2), gradient.get(0), dimensions);
+            ret = f().layerNormBp(arg(0), arg(1), arg(2), gradient.get(0), channelsFirst, dimensions);
         }
         return Arrays.asList(ret);
     }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNormBp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNormBp.java
index cfd4fff65..2168fd165 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNormBp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNormBp.java
@@ -17,6 +17,7 @@
 package org.nd4j.linalg.api.ops.impl.transforms.custom;
 
 import lombok.NoArgsConstructor;
+import lombok.NonNull;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
@@ -39,33 +40,28 @@ import java.util.List;
 public class LayerNormBp extends DynamicCustomOp {
 
     private boolean noBias = false;
+    private boolean channelsFirst;
 
 
-    public LayerNormBp(SameDiff sameDiff, SDVariable input, SDVariable gain, SDVariable bias, SDVariable gradient, int... dimensions) {
-        super(null, sameDiff, new SDVariable[] {input, gain, bias, gradient}, false);
-        Preconditions.checkArgument(bias != null, "LayerNormBp: Use constructor without bias argument if bias is null / not available.");
-
+    public LayerNormBp(@NonNull SameDiff sameDiff, @NonNull SDVariable input, @NonNull SDVariable gain, SDVariable bias, @NonNull SDVariable gradient, boolean channelsFirst, int... dimensions) {
+        super(null, sameDiff, wrapFilterNull(input, gain, bias, gradient), false);
+        this.channelsFirst = channelsFirst;
         setDimensions(dimensions);
     }
 
-    public LayerNormBp(INDArray input, INDArray gain, INDArray bias, INDArray grad, INDArray dLdx, INDArray dLdg, INDArray dLdb, int... dimensions) {
-        super("layer_norm_bp", new INDArray[]{input, gain, bias, grad}, new INDArray[]{dLdx, dLdg, dLdb});
-        Preconditions.checkArgument(bias != null, "LayerNormBp: Use constructor without bias argument if bias is null / not available.");
-
+    public LayerNormBp(@NonNull INDArray input, @NonNull INDArray gain, INDArray bias, @NonNull INDArray grad, @NonNull INDArray dLdx, @NonNull INDArray dLdg, INDArray dLdb, boolean channelsFirst, int... dimensions) {
+        super("layer_norm_bp", wrapFilterNull(input, gain, bias, grad), wrapFilterNull(dLdx, dLdg, dLdb));
+        this.channelsFirst = channelsFirst;
         setDimensions(dimensions);
     }
 
 
-    public LayerNormBp(SameDiff sameDiff, SDVariable input, SDVariable gain, SDVariable gradient, int... dimensions) {
-        super(null, sameDiff, new SDVariable[] {input, gain, gradient}, false);
-        noBias = true;
-        setDimensions(dimensions);
+    public LayerNormBp(SameDiff sameDiff, SDVariable input, SDVariable gain, SDVariable gradient, boolean channelsFirst, int... dimensions) {
+        this(sameDiff, input, gain, null, gradient, channelsFirst, dimensions);
     }
 
-    public LayerNormBp(INDArray input, INDArray gain, INDArray grad, INDArray dLdx, INDArray dLdg, int... dimensions) {
-        super("layer_norm_bp", new INDArray[]{input, gain, grad}, new INDArray[]{dLdx, dLdg});
-        noBias = true;
-        setDimensions(dimensions);
+    public LayerNormBp(INDArray input, INDArray gain, INDArray grad, INDArray dLdx, INDArray dLdg, boolean channelsFirst, int... dimensions) {
+        this(input, gain, null, grad, dLdx, dLdg, null, channelsFirst, dimensions);
     }
 
     @Override
@@ -74,7 +70,10 @@ public class LayerNormBp extends DynamicCustomOp {
         Preconditions.checkArgument(dimensions.length > 0, "LayerNormBp: You have to provide dimensions");
 
         this.dimensions = dimensions;
+        this.iArguments.clear();
         addIArgument(dimensions);
+        this.bArguments.clear();
+        addBArgument(channelsFirst);
     }
 
     @Override
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
index 84bd96ad6..fde2170a6 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
@@ -1126,7 +1126,7 @@ public class LayerOpValidation extends BaseOpValidation {
         SDVariable sdInput = sd.var("input", standardized);
         SDVariable sdGain = sd.var("gain", gain);
         SDVariable sdBias = sd.var("bias", bias);
-        SDVariable out = sd.nn.layerNorm(sdInput, sdGain, sdBias, axis);
+        SDVariable out = sd.nn.layerNorm(sdInput, sdGain, sdBias, true, axis);
         out.norm1("out");
 
         String err = OpValidation.validate(new TestCase(sd)
@@ -1135,6 +1135,38 @@ public class LayerOpValidation extends BaseOpValidation {
         assertNull(err, err);
     }
 
+    @Test
+    public void testLayerNorm4d() {
+        int mb = 3;
+        int ch = 4;
+        for(boolean nchw : new boolean[]{true, false}) {
+            double eps = 0.0;
+            INDArray x = Nd4j.rand(DataType.FLOAT, nchw ? new long[]{mb, ch, 8, 8} : new long[]{mb, 8, 8, ch});
+            INDArray gain4d = Nd4j.rand(DataType.FLOAT, nchw ? new long[]{1, ch, 1, 1} : new long[]{1, 1, 1, ch});
+            INDArray bias4d = Nd4j.rand(DataType.FLOAT, nchw ? new long[]{1, ch, 1, 1} : new long[]{1, 1, 1, ch});
+            INDArray mean = x.mean(true, 1, 2, 3);
+            INDArray std = Transforms.sqrt(x.var(false,1,2,3).addi(eps)).reshape(mb, 1, 1, 1);
+
+            INDArray standardized = x.sub(mean).div(std);
+            INDArray exp = standardized.mul(gain4d).add(bias4d);
+
+            final int[] axis = new int[]{1, 2, 3};
+            SameDiff sd = SameDiff.create();
+            SDVariable sdInput = sd.var("input", x);
+            SDVariable sdGain = sd.var("gain", gain4d.reshape(ch));
+            SDVariable sdBias = sd.var("bias", bias4d.reshape(ch));
+            SDVariable out = sd.nn.layerNorm("layernorm", sdInput, sdGain, sdBias, nchw, axis);
+
+            SDVariable loss = sd.loss.l2Loss(out);
+
+            String err = OpValidation.validate(new TestCase(sd)
+                    .expectedOutput("layernorm", exp)
+                    .gradientCheck(true));
+            assertNull(err);
+        }
+    }
+
+
     @Test
     public void testLayerNormOP() {
         final INDArray random = Nd4j.rand(new int[]{10, 4});
@@ -1165,7 +1197,7 @@ public class LayerOpValidation extends BaseOpValidation {
         SameDiff sd = SameDiff.create();
         SDVariable sdInput = sd.var("input", standardized);
         SDVariable sdGain = sd.var("gain", gain);
-        SDVariable out = sd.nn.layerNorm(sdInput, sdGain, axis);
+        SDVariable out = sd.nn.layerNorm(sdInput, sdGain, true, axis);
         out.norm1("out");
 
         String err = OpValidation.validate(new TestCase(sd)
@@ -1209,7 +1241,7 @@ public class LayerOpValidation extends BaseOpValidation {
         SDVariable sdInput = sd.var("input", standardized);
         SDVariable sdGain = sd.var("gain", gain);
         SDVariable sdBias = sd.var("bias", bias);
-        SDVariable out = sd.nn.layerNorm(sdInput, sdGain, sdBias, axis);
+        SDVariable out = sd.nn.layerNorm(sdInput, sdGain, sdBias, true, axis);
         out.norm1("out");
 
         String err = OpValidation.validate(new TestCase(sd)

From 5cfbeb64ac3d0e84a885719f146f1162d5a9de48 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Tue, 27 Aug 2019 19:10:31 +1000
Subject: [PATCH 17/56] Another small fix (#175)

* Layer norm 4d case fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 .../main/java/org/nd4j/autodiff/validation/OpValidation.java  | 4 ++--
 .../imports/converters/DifferentialFunctionClassHolder.java   | 1 +
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/validation/OpValidation.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/validation/OpValidation.java
index e9ad61c04..3e329ad13 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/validation/OpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/validation/OpValidation.java
@@ -341,8 +341,8 @@ public class OpValidation {
 
 
         //Finally: check execution/output
-        Map<String,INDArray> outOrig = original.execAll(tc.placeholderValues());
-        Map<String,INDArray> outDe = deserialized.execAll(tc.placeholderValues());
+        Map<String,INDArray> outOrig = original.outputAll(tc.placeholderValues());
+        Map<String,INDArray> outDe = deserialized.outputAll(tc.placeholderValues());
         Preconditions.checkState(outOrig.keySet().equals(outDe.keySet()), "Keysets for execution after deserialization does not match key set for original model");
 
         for(String s : outOrig.keySet()){
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/DifferentialFunctionClassHolder.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/DifferentialFunctionClassHolder.java
index a92066d7a..82bfdc843 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/DifferentialFunctionClassHolder.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/DifferentialFunctionClassHolder.java
@@ -64,6 +64,7 @@ public class DifferentialFunctionClassHolder {
         add("outputVariables");
         add("tArguments");
         add("iArguments");
+        add("bArguments");
         add("hash");
         add("opName");
         add("sameDiff");

From efbfafe3f7ebd0ef593b1a88a398e1fc4e3d9958 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Tue, 27 Aug 2019 12:35:14 +0300
Subject: [PATCH 18/56] [WIP] gatherND fix (#176)

* one test for gather_nd

Signed-off-by: raver119 <raver119@gmail.com>

* get rid of old concat tests

Signed-off-by: raver119 <raver119@gmail.com>

* one printf

Signed-off-by: raver119 <raver119@gmail.com>

* one more legacy test removed

Signed-off-by: raver119 <raver119@gmail.com>

* gatherNd launch params fix

Signed-off-by: raver119 <raver119@gmail.com>

* gatherNd launch params fix

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../ops/declarable/helpers/cuda/gather_nd.cu  |   3 +-
 .../layers_tests/DeclarableOpsTests5.cpp      |  17 +
 .../layers_tests/NDArrayCudaBasicsTests.cu    | 302 +-----------------
 3 files changed, 20 insertions(+), 302 deletions(-)

diff --git a/libnd4j/include/ops/declarable/helpers/cuda/gather_nd.cu b/libnd4j/include/ops/declarable/helpers/cuda/gather_nd.cu
index 709f0ed2c..6587b4ca7 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/gather_nd.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/gather_nd.cu
@@ -106,6 +106,7 @@ namespace nd4j {
                     const auto xOffset = shape::getOffset(0, xShapeInfo + 1, xShapeInfo + xRank + 1, xCoordStart, xRank);
 
                     z[zOffset] = x[xOffset];
+                    printf("z[%lld] = x[%lld] = %f\n", zOffset, xOffset, (float) z[zOffset]);
                 }
             }
 
@@ -124,7 +125,7 @@ namespace nd4j {
 
                 const int maxRank = nd4j::math::nd4j_max<int>(indices.rankOf(), nd4j::math::nd4j_max<int>(input.rankOf(), output.rankOf()));
 
-                const int threadsPerBlock = MAX_NUM_THREADS;
+                const int threadsPerBlock = 256;
                 const int blocksPerGrid = (output.lengthOf() + threadsPerBlock - 1) / threadsPerBlock;
                 const int sharedMem = 8 * threadsPerBlock * maxRank + 128;
 
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
index b596ebcd5..1fbe81046 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
@@ -815,6 +815,23 @@ TEST_F(DeclarableOpsTests5, gatherNd_test7) {
     delete results;
 }
 
+//////////////////////////////////////////////////////////////////////
+TEST_F(DeclarableOpsTests5, gatherNd_test8) {
+    auto x = NDArrayFactory::create<double>('c', {2, 2}, {1., 2., 3., 4.});
+    auto y = NDArrayFactory::create<int>('c', {2, 2}, {0, 0, 1, 1});
+    auto e = NDArrayFactory::create<double>('c', {2}, {1., 4.});
+
+    nd4j::ops::gather_nd op;
+    auto result = op.execute({&x, &y}, {}, {});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    auto z = result->at(0);
+
+    ASSERT_EQ(e, *z);
+
+    delete result;
+}
+
 //////////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests5, reverse_sequense_test1) {
 
diff --git a/libnd4j/tests_cpu/layers_tests/NDArrayCudaBasicsTests.cu b/libnd4j/tests_cpu/layers_tests/NDArrayCudaBasicsTests.cu
index 7b9e788f7..4ab884d28 100644
--- a/libnd4j/tests_cpu/layers_tests/NDArrayCudaBasicsTests.cu
+++ b/libnd4j/tests_cpu/layers_tests/NDArrayCudaBasicsTests.cu
@@ -2261,304 +2261,4 @@ TEST_F(NDArrayCudaBasicsTests, Test_Empty_4) {
 
     ASSERT_TRUE(x->isEmpty());
     delete x;
-}
-
-// printCudaGlobal<double><<<1,1,0,*stream>>>(dX, 6);
-//     printCudaGlobal<Nd4jLong><<<1,1,0,*stream>>>(dXShapeInfo, 8);
-//     printCudaGlobal<double><<<1,1,0,*stream>>>(dZ, 2);
-//     printCudaGlobal<Nd4jLong><<<1,1,0,*stream>>>(dZShapeInfo, 6);
-//     printCudaGlobal<int><<<1,1,0,*stream>>>(dimension, 1);
-//     printCudaGlobal<Nd4jLong><<<1,1,0,*stream>>>(tadShapeInfo, 6);
-//     printCudaGlobal<Nd4jLong><<<1,1,0,*stream>>>(tadOffsets, 2);
-//     cudaStreamSynchronize(*stream);
-
-TEST_F(NDArrayCudaBasicsTests, Test_ConcatNative_1) {
-
-    auto x = NDArrayFactory::create<float>('c', {5,2}, {0,1,2,3,4,5,6,7,8,9});
-    x.syncToHost();
-    auto z = NDArrayFactory::create<float>('c', {5, 8});
-    z.syncToHost();
-
-    std::vector<void*> buffers(4);
-    std::vector<Nd4jLong*> shapes(4);
-    std::vector<Nd4jLong*> hostShapes(4);
-
-    for (size_t i = 0; i < buffers.size(); i++) {
-        buffers[i] = x.specialBuffer();
-        shapes[i] = x.specialShapeInfo();
-        hostShapes[i] = x.shapeInfo();
-    }
-    Nd4jPointer extra[2];
-    extra[1] = x.getContext()->getCudaStream();
-    ::concat(extra, 1, 4, nullptr, (Nd4jPointer*)hostShapes.data(), (Nd4jPointer*)buffers.data(), (Nd4jPointer*)shapes.data(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, nullptr);
-}
-
-TEST_F(NDArrayCudaBasicsTests, Test_ConcatNative_2) {
-
-    auto x = NDArrayFactory::create<float>('c', {5,2}, {0,1,2,3,4,5,6,7,8,9});
-    auto z = NDArrayFactory::create<float>('f', {5, 8});
-
-    std::vector<void*> buffers(4);
-    std::vector<Nd4jLong*> shapes(4);
-    std::vector<Nd4jLong*> hostShapes(4);
-
-    x.syncToHost();
-    z.syncToHost();
-
-    for (size_t i = 0; i < buffers.size(); i++) {
-        buffers[i] = x.specialBuffer();
-        shapes[i] = x.specialShapeInfo();
-        hostShapes[i] = x.shapeInfo();
-    }
-
-    Nd4jPointer extra[2];
-    extra[1] = x.getContext()->getCudaStream();
-
-    ::concat(extra, 1, 4, nullptr, (Nd4jPointer*)hostShapes.data(), (Nd4jPointer*)buffers.data(), (Nd4jPointer*)shapes.data(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, nullptr);
-}
-
-TEST_F(NDArrayCudaBasicsTests, Test_ConcatNative_3) {
-
-    auto x = NDArrayFactory::create<float>('c', {2,3}, {1,2,3,4,5,6});
-    auto y = NDArrayFactory::create<float>('c', {1,3}, {7,8,9});
-    auto z = NDArrayFactory::create<float>('f', {3, 3});
-
-
-    std::vector<void*> buffers(2);
-    std::vector<Nd4jLong*> shapes(2);
-    std::vector<Nd4jLong*> hostShapes(2);
-
-    x.syncToHost();
-    y.syncToHost();
-    z.syncToHost();
-
-    buffers[0] = x.specialBuffer(); shapes[0] = x.specialShapeInfo(); hostShapes[0] = x.shapeInfo();
-    buffers[1] = y.specialBuffer(); shapes[1] = y.specialShapeInfo(); hostShapes[1] = y.shapeInfo();
-
-    Nd4jPointer extra[2];
-    extra[1] = x.getContext()->getCudaStream();
-
-    ::concat(extra, 0, 2, nullptr, (Nd4jPointer*)hostShapes.data(), (Nd4jPointer*)buffers.data(), (Nd4jPointer*)shapes.data(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, nullptr);
-}
-
-TEST_F(NDArrayCudaBasicsTests, Test_ConcatNative_4) {
-
-    auto x = NDArrayFactory::create<float>('c', {2,3}, {1,2,3,4,5,6});
-    auto y = NDArrayFactory::create<float>('c', {1,3}, {7,8,9});
-    auto z = NDArrayFactory::create<float>('c', {3, 3});
-
-    x.syncToHost();
-    y.syncToHost();
-    z.syncToHost();
-
-    std::vector<void*> buffers(2);
-    std::vector<Nd4jLong*> shapes(2);
-    std::vector<Nd4jLong*> hostShapes(2);
-
-    buffers[0] = x.specialBuffer(); shapes[0] = x.specialShapeInfo(); hostShapes[0] = x.shapeInfo();
-    buffers[1] = y.specialBuffer(); shapes[1] = y.specialShapeInfo(); hostShapes[1] = y.shapeInfo();
-
-    Nd4jPointer extra[2];
-    extra[1] = x.getContext()->getCudaStream();
-
-    ::concat(extra, 0, 2, nullptr, (Nd4jPointer*)hostShapes.data(), (Nd4jPointer*)buffers.data(), (Nd4jPointer*)shapes.data(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, nullptr);
-}
-
-TEST_F(NDArrayCudaBasicsTests, Test_ConcatNative_5) {
-
-    auto x = NDArrayFactory::create<float>('c', {1,2,3}, {1,2,3,4,5,6});
-    auto y = NDArrayFactory::create<float>('c', {1,2,3}, {7,8,9,10,11, 12});
-
-    auto z = NDArrayFactory::create<float>('c', {2, 2, 3});
-    auto stream = x.getContext()->getCudaStream();//reinterpret_cast<cudaStream_t *>(&nativeStream);
-    std::vector<void*> buffers(2);
-    std::vector<Nd4jLong*> shapes(2);
-    std::vector<Nd4jLong*> hostShapes(2);
-
-    buffers[0] = x.specialBuffer(); shapes[0] = x.specialShapeInfo(); hostShapes[0] = x.shapeInfo();
-    buffers[1] = y.specialBuffer(); shapes[1] = y.specialShapeInfo(); hostShapes[1] = y.shapeInfo();
-
-    Nd4jPointer extra[2];
-    extra[1] = x.getContext()->getCudaStream();
-
-    ::concat(extra, 0, 2, nullptr, (Nd4jPointer*)hostShapes.data(), (Nd4jPointer*)buffers.data(), (Nd4jPointer*)shapes.data(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, nullptr);
-}
-
-TEST_F(NDArrayCudaBasicsTests, Test_ConcatNative_6) {
-
-    auto x1 = NDArrayFactory::create<float>('c', {2,2,3}, {1,2,3,4,5,6,7,8, 9, 10,11,12});
-    auto x2 = NDArrayFactory::create<float>('c', {1,2,3}, {13,14,15,16,17, 18});
-    auto x3 = NDArrayFactory::create<float>('c', {1,2,3}, {19,20,21,22,23, 24});
-
-    x1.syncToHost();
-    x2.syncToHost();
-    x3.syncToHost();
-
-    auto z = NDArrayFactory::create<float>('c', {4, 2, 3});
-
-    std::vector<void*> buffers(3);
-    std::vector<Nd4jLong*> shapes(3);
-    std::vector<Nd4jLong*> hostShapes(3);
-
-    buffers[0] = x1.specialBuffer(); shapes[0] = x1.specialShapeInfo(); hostShapes[0] = x1.shapeInfo();
-    buffers[1] = x2.specialBuffer(); shapes[1] = x2.specialShapeInfo(); hostShapes[1] = x2.shapeInfo();
-    buffers[2] = x3.specialBuffer(); shapes[2] = x3.specialShapeInfo(); hostShapes[2] = x3.shapeInfo();
-
-    Nd4jPointer extra[2];
-    extra[1] = x1.getContext()->getCudaStream();
-
-    ::concat(extra, 0, 3, nullptr, (Nd4jPointer*)hostShapes.data(), (Nd4jPointer*)buffers.data(), (Nd4jPointer*)shapes.data(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, nullptr);
-}
-
-TEST_F(NDArrayCudaBasicsTests, Test_ConcatNative_7) {
-
-    auto x1 = NDArrayFactory::create<float>(1);
-    auto x2 = NDArrayFactory::create<float>(2);
-    auto x3 = NDArrayFactory::create<float>(3);
-
-    auto z = NDArrayFactory::create<float>('c', {3}, {1,2,3});
-
-    x1.syncToHost();
-    x2.syncToHost();
-    x3.syncToHost();
-
-    std::vector<void*> buffers(3);
-    std::vector<Nd4jLong*> shapes(3);
-    std::vector<Nd4jLong*> hostShapes(3);
-
-    buffers[0] = x1.specialBuffer(); shapes[0] = x1.specialShapeInfo(); hostShapes[0] = x1.shapeInfo();
-    buffers[1] = x2.specialBuffer(); shapes[1] = x2.specialShapeInfo(); hostShapes[1] = x2.shapeInfo();
-    buffers[2] = x3.specialBuffer(); shapes[2] = x3.specialShapeInfo(); hostShapes[2] = x3.shapeInfo();
-
-    Nd4jPointer extra[2];
-    extra[1] = x1.getContext()->getCudaStream();
-
-    ::concat(extra, 0, 3, nullptr, (Nd4jPointer*)hostShapes.data(), (Nd4jPointer*)buffers.data(), (Nd4jPointer*)shapes.data(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, nullptr);
-}
-
-TEST_F(NDArrayCudaBasicsTests, Test_ConcatNative_8) {
-
-    auto totalCount = 1000;
-    auto width = 300;
-    std::vector<NDArray> lx(totalCount);
-    for (int i = 0; i < totalCount; i++) {
-        lx[i] = NDArrayFactory::create<float>('c', {1, width});
-        lx[i].assign(i);
-        lx[i].syncToHost();
-    }
-
-    auto z = NDArrayFactory::create<float>('c', {totalCount, width});
-
-    std::vector<void*> buffers(totalCount);
-    std::vector<Nd4jLong*> shapes(totalCount);
-    std::vector<Nd4jLong*> hostShapes(totalCount);
-
-    for (size_t i = 0; i < lx.size(); i++) {
-        buffers[i] = lx[i].specialBuffer();
-        shapes[i] = lx[i].specialShapeInfo();
-        hostShapes[i] = lx[i].shapeInfo();
-    }
-
-    Nd4jPointer extra[2];
-    extra[1] = nd4j::LaunchContext::defaultContext()->getCudaStream();
-
-    ::concat(extra, 0, totalCount, nullptr, (Nd4jPointer*)hostShapes.data(), (Nd4jPointer*)buffers.data(), (Nd4jPointer*)shapes.data(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, nullptr);
-}
-
-TEST_F(NDArrayCudaBasicsTests, TestTear_1) {
-    auto input = NDArrayFactory::create<float>('c', {1, 10, 10});
-    std::vector<NDArray> arrays; // = {NDArrayFactory::create<float>('c', {1, 10, 10}), NDArrayFactory::create<float>('c', {1, 10, 10}), NDArrayFactory::create<float>('c', {1, 10, 10}), NDArrayFactory::create<float>('c', {1, 10, 10}), NDArrayFactory::create<float>('c', {1, 10, 10})};
-    int total = 151;
-    for (int e = 0; e < total; e++) {
-        input.assign(e);
-        arrays.emplace_back(input);
-    }
-    auto z = NDArrayFactory::create<float>('c', {total, 10, 10});
-
-    Nd4jPointer extra[1];
-    extra[1] = input.getContext()->getCudaStream();
-
-    std::vector<void*> buffers(total);
-    std::vector<Nd4jLong*> shapes(total);
-    std::vector<Nd4jLong*> hostShapes(total);
-
-    for (size_t i = 0; i < buffers.size(); i++) {
-        buffers[i] = arrays[i].specialBuffer();
-        shapes[i] = arrays[i].specialShapeInfo();
-        hostShapes[i] = arrays[i].shapeInfo();
-    }
-
-    ::concat(extra, 0, total, nullptr, (Nd4jPointer*)hostShapes.data(), (Nd4jPointer*)buffers.data(), (Nd4jPointer*)shapes.data(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, nullptr);
-    nd4j::ops::tear op;
-
-    auto result = op.execute({&z}, {}, {1, 2});
-    //ASSERT_EQ(10, result->size());
-    auto e = result->size() - 1;
-    //for (size_t e = 0; e < result->size(); e++) {
-//        arrays[e].printIndexedBuffer("Input list at 40");
-//        result->at(e)->printIndexedBuffer("OUtput TEAR at 40");
-    //}
-//        ASSERT_TRUE(tads->at(e)->equalsTo(result->at(e)));
-
-    delete result;
-//    delete tads;
-}
-
-TEST_F(NDArrayCudaBasicsTests, TestTear_2) {
-
-    auto input = NDArrayFactory::create<float>('c', {1, 10, 10});
-
-    std::vector<NDArray> arrays; // = {NDArrayFactory::create<float>('c', {1, 10, 10}), NDArrayFactory::create<float>('c', {1, 10, 10}), NDArrayFactory::create<float>('c', {1, 10, 10}), NDArrayFactory::create<float>('c', {1, 10, 10}), NDArrayFactory::create<float>('c', {1, 10, 10})};
-    for (int e = 0; e < 10; e++) {
-        input.assign(e);
-        arrays.emplace_back(input);
-        arrays[e].syncToHost();
-    }
-
-    auto z = NDArrayFactory::create<float>('c', {10, 10, 10});
-
-    Nd4jPointer extra[2];
-    extra[1] = input.getContext()->getCudaStream();
-
-    std::vector<void*> buffers(10);
-    std::vector<Nd4jLong*> shapes(10);
-    std::vector<Nd4jLong*> hostShapes(10);
-
-    for (size_t i = 0; i < buffers.size(); i++) {
-        buffers[i] = arrays[i].specialBuffer();
-        shapes[i] = arrays[i].specialShapeInfo();
-        hostShapes[i] = arrays[i].shapeInfo();
-    }
-
-    std::vector<int> dimsToExclude({1,2});
-
-
-    ::concat(extra, 0, 10, nullptr, (Nd4jPointer*)hostShapes.data(), (Nd4jPointer*)buffers.data(), (Nd4jPointer*)shapes.data(), nullptr, z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), nullptr, nullptr);
-
-    auto packX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(input.getShapeInfo(), dimsToExclude);
-    //std::vector<void*> arraysData(arrays.size());
-    Nd4jPointer* arraysData;
-    cudaError_t err = cudaMalloc(&arraysData, arrays.size() * sizeof(void*));
-    if (err != 0) {
-        printf("Cannot allocate device memory for targets due error %d\n", err);
-        ASSERT_TRUE(false);
-    }
-    for (size_t i = 0; i < arrays.size(); i++) {
-        Nd4jPointer target = arrays[i].specialBuffer();
-        cudaMemcpy(&arraysData[i], &target, sizeof(Nd4jPointer), cudaMemcpyHostToDevice);
-    }
-    ::tear(extra, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo(), arraysData, input.specialShapeInfo(), packX.specialShapeInfo(), packX.specialOffsets());
-//    auto result = op.execute({&z}, {}, {1, 2});
-
-    //ASSERT_EQ(10, result->size());
-    err = cudaFree(arraysData);
-    if (err != 0) {
-        printf("Cannot deallocate device memory for targets due error %d\n", err);
-        ASSERT_TRUE(false);
-    }
-
-//        ASSERT_TRUE(tads->at(e)->equalsTo(result->at(e)));
-
-//    delete result;
-//    delete tads;
-}
+}
\ No newline at end of file

From fd22a8ecc753b4cbddc89e8eb8260fc439bbda90 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Tue, 27 Aug 2019 19:46:26 +1000
Subject: [PATCH 19/56] Small build fix, after last PR (#177)

Signed-off-by: Alex Black <blacka101@gmail.com>
---
 .../org/deeplearning4j/nn/layers/BaseLayer.java     |  4 ++--
 .../nn/layers/recurrent/SimpleRnn.java              |  8 ++++----
 .../api/ops/impl/transforms/custom/LayerNorm.java   | 13 +++++--------
 3 files changed, 11 insertions(+), 14 deletions(-)

diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/BaseLayer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/BaseLayer.java
index 405889f0e..00ca7e7c4 100755
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/BaseLayer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/BaseLayer.java
@@ -93,7 +93,7 @@ public abstract class BaseLayer<LayerConfT extends org.deeplearning4j.nn.conf.la
             INDArray g = getParam(DefaultParamInitializer.GAIN_KEY);
 
             INDArray dldg = gradientViews.get(DefaultParamInitializer.GAIN_KEY);
-            Nd4j.getExecutioner().exec(new LayerNormBp(preNorm, g, delta, delta, dldg, 1));
+            Nd4j.getExecutioner().exec(new LayerNormBp(preNorm, g, delta, delta, dldg, true, 1));
             ret.gradientForVariable().put(DefaultParamInitializer.GAIN_KEY, dldg);
 
         }
@@ -318,7 +318,7 @@ public abstract class BaseLayer<LayerConfT extends org.deeplearning4j.nn.conf.la
         INDArray preNorm = ret;
         if(hasLayerNorm()){
             preNorm = (forBackprop ? ret.dup(ret.ordering()) : ret);
-            Nd4j.getExecutioner().exec(new LayerNorm(preNorm, g, ret, 1));
+            Nd4j.getExecutioner().exec(new LayerNorm(preNorm, g, ret, true, 1));
         }
 
         if(hasBias()){
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/SimpleRnn.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/SimpleRnn.java
index 5704e57d3..e4bdcf56f 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/SimpleRnn.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/SimpleRnn.java
@@ -158,7 +158,7 @@ public class SimpleRnn extends BaseRecurrentLayer<org.deeplearning4j.nn.conf.lay
                 dldnCurrent = workspaceMgr.createUninitialized(ArrayType.BP_WORKING_MEM, dldzCurrent.dataType(), dldzCurrent.shape());
                 INDArray ggCur = workspaceMgr.createUninitialized(ArrayType.BP_WORKING_MEM, gg.dataType(), gxg.shape());
                 INDArray bgCur = workspaceMgr.createUninitialized(ArrayType.BP_WORKING_MEM, bg.dataType(), bg.shape());
-                Nd4j.getExecutioner().exec(new LayerNormBp(nCurrent, gx, b, dldzCurrent, dldnCurrent, ggCur, bgCur, 1));
+                Nd4j.getExecutioner().exec(new LayerNormBp(nCurrent, gx, b, dldzCurrent, dldnCurrent, ggCur, bgCur, true, 1));
                 gxg.addi(ggCur);
                 bg.addi(bgCur);
             }else{
@@ -177,7 +177,7 @@ public class SimpleRnn extends BaseRecurrentLayer<org.deeplearning4j.nn.conf.lay
             if(hasLayerNorm() && i > end){
                 dldzNext = workspaceMgr.createUninitialized(ArrayType.BP_WORKING_MEM, dldzCurrent.dataType(), dldzCurrent.shape());
                 INDArray ggCur = workspaceMgr.createUninitialized(ArrayType.BP_WORKING_MEM, gg.dataType(), grg.shape());
-                Nd4j.getExecutioner().exec(new LayerNormBp(rCurrent, gr, dldzCurrent, dldzNext, ggCur, 1));
+                Nd4j.getExecutioner().exec(new LayerNormBp(rCurrent, gr, dldzCurrent, dldzNext, ggCur, true, 1));
                 grg.addi(ggCur);
             }else{
                 dldzNext = dldzCurrent;
@@ -256,7 +256,7 @@ public class SimpleRnn extends BaseRecurrentLayer<org.deeplearning4j.nn.conf.lay
             if(hasLayerNorm()){
                 INDArray currOutPreNorm = (forBackprop ? outPreNorm : out).get(all(), all(), point(i));
                 Nd4j.gemm(currIn, w, currOutPreNorm, false, false, 1.0, 0.0);
-                Nd4j.getExecutioner().exec(new LayerNorm(currOutPreNorm, gx, b, currOut, 1));
+                Nd4j.getExecutioner().exec(new LayerNorm(currOutPreNorm, gx, b, currOut, true, 1));
             }else{
                 Nd4j.gemm(currIn, w, currOut, false, false, 1.0, 1.0);  //beta = 1.0 to keep previous contents (bias)
             }
@@ -266,7 +266,7 @@ public class SimpleRnn extends BaseRecurrentLayer<org.deeplearning4j.nn.conf.lay
                     INDArray currRecPreNorm = forBackprop ? recPreNorm.get(all(), all(), point(i)) : workspaceMgr.createUninitialized(ArrayType.FF_WORKING_MEM, currOut.dataType(), currOut.shape(), 'f');;
                     Nd4j.gemm(prevStepOut, rw, currRecPreNorm, false, false, 1.0, 0.0);
                     INDArray recNorm = workspaceMgr.createUninitialized(ArrayType.FF_WORKING_MEM, currOut.dataType(), currOut.shape(), 'f');
-                    Nd4j.getExecutioner().exec(new LayerNorm(currRecPreNorm, gr, recNorm, 1));
+                    Nd4j.getExecutioner().exec(new LayerNorm(currRecPreNorm, gr, recNorm, true, 1));
                     currOut.addi(recNorm);
                 }else {
                     Nd4j.gemm(prevStepOut, rw, currOut, false, false, 1.0, 1.0);    //beta = 1.0 to keep previous contents
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java
index 65222fe98..137bc6693 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java
@@ -54,17 +54,14 @@ public class LayerNorm extends DynamicCustomOp {
         this(sameDiff, input, gain, null, channelsFirst, dimensions);
     }
 
-    public LayerNorm(INDArray input, INDArray gain, INDArray bias, INDArray result, int... dimensions) {
-        super("layer_norm", new INDArray[]{input, gain, bias}, new INDArray[]{result});
-        Preconditions.checkArgument(bias != null, "LayerNorm: Use different constructor if bias is null.");
-
+    public LayerNorm(INDArray input, INDArray gain, INDArray bias, INDArray result, boolean channelsFirst, int... dimensions) {
+        super("layer_norm", wrapFilterNull(input, gain, bias), wrapOrNull(result));
+        this.channelsFirst = channelsFirst;
         setDimensions(dimensions);
     }
 
-    public LayerNorm(INDArray input, INDArray gain, INDArray result, int... dimensions) {
-        super("layer_norm", new INDArray[]{input, gain}, new INDArray[]{result});
-        noBias = true;
-        setDimensions(dimensions);
+    public LayerNorm(INDArray input, INDArray gain, INDArray result, boolean channelsFirst, int... dimensions) {
+        this(input, gain, null, result, channelsFirst, dimensions);
     }
 
     @Override

From a49f7c908b0118368aacedd04d3b5a0859f249e2 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Tue, 27 Aug 2019 13:21:01 +0300
Subject: [PATCH 20/56] [WIP] More fixes (#178)

* skip string arrays for device validation

Signed-off-by: raver119 <raver119@gmail.com>

* histogram_fixed_width now really supports indexing types

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../transforms/histogram_fixed_width.cpp      |  2 +-
 .../helpers/cuda/histogramFixedWidth.cu       | 24 +++++++++----------
 .../jita/allocator/impl/AtomicAllocator.java  |  2 +-
 .../flow/impl/SynchronousFlowController.java  |  5 ++--
 .../ops/executioner/CudaOpContext.java        |  4 ++--
 5 files changed, 19 insertions(+), 18 deletions(-)

diff --git a/libnd4j/include/ops/declarable/generic/transforms/histogram_fixed_width.cpp b/libnd4j/include/ops/declarable/generic/transforms/histogram_fixed_width.cpp
index b3063c75d..529446e12 100644
--- a/libnd4j/include/ops/declarable/generic/transforms/histogram_fixed_width.cpp
+++ b/libnd4j/include/ops/declarable/generic/transforms/histogram_fixed_width.cpp
@@ -49,7 +49,7 @@ CUSTOM_OP_IMPL(histogram_fixed_width, 2, 1, false, 0, 0) {
 DECLARE_TYPES(histogram_fixed_width) {
     getOpDescriptor()
         ->setAllowedInputTypes(nd4j::DataType::ANY)
-        ->setAllowedOutputTypes({ALL_INTS});
+        ->setAllowedOutputTypes({ALL_INDICES});
 }
 
 
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/histogramFixedWidth.cu b/libnd4j/include/ops/declarable/helpers/cuda/histogramFixedWidth.cu
index ebde4909c..317f1d857 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/histogramFixedWidth.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/histogramFixedWidth.cu
@@ -27,16 +27,16 @@ namespace ops     {
 namespace helpers {
 
 ///////////////////////////////////////////////////////////////////
-template<typename T>
+template<typename X, typename Z>
 __global__ static void histogramFixedWidthCuda( const void* vx, const Nd4jLong* xShapeInfo,
                                                       void* vz, const Nd4jLong* zShapeInfo,
-                                                const T leftEdge, const T rightEdge) {
+                                                const X leftEdge, const X rightEdge) {
 
-    const T* x  = reinterpret_cast<const T*>(vx);
-    Nd4jLong* z = reinterpret_cast<Nd4jLong*>(vz);
+    const auto x  = reinterpret_cast<const X*>(vx);
+    auto z = reinterpret_cast<Z*>(vz);
 
     __shared__ Nd4jLong xLen, zLen, totalThreads, nbins;
-    __shared__ T binWidth, secondEdge, lastButOneEdge;
+    __shared__ X binWidth, secondEdge, lastButOneEdge;
 
     if (threadIdx.x == 0) {
 
@@ -55,7 +55,7 @@ __global__ static void histogramFixedWidthCuda( const void* vx, const Nd4jLong*
 
     for (Nd4jLong i = tid; i < xLen; i += totalThreads) {
 
-        const T value = x[shape::getIndexOffset(i, xShapeInfo, xLen)];
+        const X value = x[shape::getIndexOffset(i, xShapeInfo, xLen)];
 
         Nd4jLong zIndex;
 
@@ -66,18 +66,18 @@ __global__ static void histogramFixedWidthCuda( const void* vx, const Nd4jLong*
         else
             zIndex = static_cast<Nd4jLong>((value - leftEdge) / binWidth);
 
-        nd4j::math::atomics::nd4j_atomicAdd(&z[shape::getIndexOffset(zIndex, zShapeInfo, nbins)], 1LL);
+        nd4j::math::atomics::nd4j_atomicAdd<Z>(&z[shape::getIndexOffset(zIndex, zShapeInfo, nbins)], 1);
     }
 }
 
 ///////////////////////////////////////////////////////////////////
-template<typename T>
+template<typename X, typename Z>
 __host__ static void histogramFixedWidthCudaLauncher(const cudaStream_t *stream, const NDArray& input, const NDArray& range, NDArray& output) {
 
-    const T leftEdge  = range.e<T>(0);
-    const T rightEdge = range.e<T>(1);
+    const X leftEdge  = range.e<X>(0);
+    const X rightEdge = range.e<X>(1);
 
-    histogramFixedWidthCuda<T><<<512, MAX_NUM_THREADS / 2, 512, *stream>>>(input.getSpecialBuffer(), input.getSpecialShapeInfo(), output.specialBuffer(), output.specialShapeInfo(), leftEdge, rightEdge);
+    histogramFixedWidthCuda<X, Z><<<256, 256, 1024, *stream>>>(input.getSpecialBuffer(), input.getSpecialShapeInfo(), output.specialBuffer(), output.specialShapeInfo(), leftEdge, rightEdge);
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -89,7 +89,7 @@ void histogramFixedWidth(nd4j::LaunchContext* context, const NDArray& input, con
     PointersManager manager(context, "histogramFixedWidth");
 
     NDArray::prepareSpecialUse({&output}, {&input});
-    BUILD_SINGLE_SELECTOR(input.dataType(), histogramFixedWidthCudaLauncher, (context->getCudaStream(), input, range, output), LIBND4J_TYPES);
+    BUILD_DOUBLE_SELECTOR(input.dataType(), output.dataType(), histogramFixedWidthCudaLauncher, (context->getCudaStream(), input, range, output), LIBND4J_TYPES, INDEXING_TYPES);
     NDArray::registerSpecialUse({&output}, {&input});
 
     manager.synchronize();
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/impl/AtomicAllocator.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/impl/AtomicAllocator.java
index 8fbf0a000..0ec1876ca 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/impl/AtomicAllocator.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/impl/AtomicAllocator.java
@@ -312,7 +312,7 @@ public class AtomicAllocator implements Allocator {
     @Override
     public Pointer getPointer(INDArray array, CudaContext context) {
         //    DataBuffer buffer = array.data().originalDataBuffer() == null ? array.data() : array.data().originalDataBuffer();
-        if (array.isEmpty())
+        if (array.isEmpty() || array.isS())
             return null;
 
         return memoryHandler.getDevicePointer(array.data(), context);
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/flow/impl/SynchronousFlowController.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/flow/impl/SynchronousFlowController.java
index fb4510f1b..d81de381a 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/flow/impl/SynchronousFlowController.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/flow/impl/SynchronousFlowController.java
@@ -172,7 +172,7 @@ public class SynchronousFlowController implements FlowController {
         val cId = allocator.getDeviceId();
 
 
-        if (result != null && !result.isEmpty()) {
+        if (result != null && !result.isEmpty() && !result.isS()) {
             Nd4j.getCompressor().autoDecompress(result);
             prepareDelayedMemory(result);
             val pointData = allocator.getAllocationPoint(result);
@@ -198,7 +198,8 @@ public class SynchronousFlowController implements FlowController {
             return context;
 
         for (INDArray operand : operands) {
-            if (operand == null || operand.isEmpty())
+            // empty or String arrays can be skipped
+            if (operand == null || operand.isEmpty() || operand.isS())
                 continue;
 
             Nd4j.getCompressor().autoDecompress(operand);
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaOpContext.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaOpContext.java
index 26d363f32..cf779f537 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaOpContext.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaOpContext.java
@@ -100,7 +100,7 @@ public class CudaOpContext extends BaseOpContext implements OpContext {
     @Override
     public Pointer contextPointer() {
         for (val v:fastpath_in.values()) {
-            if (v.isEmpty())
+            if (v.isEmpty() || v.isS())
                 continue;
 
             AtomicAllocator.getInstance().getAllocationPoint(v).tickHostRead();
@@ -111,7 +111,7 @@ public class CudaOpContext extends BaseOpContext implements OpContext {
         }
 
         for (val v:fastpath_out.values()) {
-            if (v.isEmpty())
+            if (v.isEmpty() || v.isS())
                 continue;
 
             AtomicAllocator.getInstance().getAllocationPoint(v).tickHostRead();

From dff599aa8fbbaa45765bb4ea82d494082c6300ac Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Tue, 27 Aug 2019 20:43:36 +1000
Subject: [PATCH 21/56] Test fix (#179)

Signed-off-by: Alex Black <blacka101@gmail.com>
---
 .../org/nd4j/autodiff/opvalidation/LayerOpValidation.java     | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
index fde2170a6..760165b3b 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
@@ -1178,7 +1178,7 @@ public class LayerOpValidation extends BaseOpValidation {
         final INDArray res = standardized.mulRowVector(gain).addRowVector(bias);
 
         final INDArray output = Nd4j.zerosLike(res);
-        Nd4j.getExecutioner().exec(new LayerNorm(standardized, gain, bias, output, 1));
+        Nd4j.getExecutioner().exec(new LayerNorm(standardized, gain, bias, output, true, 1));
 
         assertEquals(res, output);
     }
@@ -1216,7 +1216,7 @@ public class LayerOpValidation extends BaseOpValidation {
         final INDArray res = standardized.mulRowVector(gain);
 
         final INDArray output = Nd4j.zerosLike(res);
-        Nd4j.getExecutioner().exec(new LayerNorm(standardized, gain, output, 1));
+        Nd4j.getExecutioner().exec(new LayerNorm(standardized, gain, output, true, 1));
 
         assertEquals(res, output);
     }

From 0e523490e93d1b2be71a238edbf3ec84cfa8f988 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Tue, 27 Aug 2019 14:30:37 +0300
Subject: [PATCH 22/56] [WIP] confusion (#180)

* skip string arrays for device validation

Signed-off-by: raver119 <raver119@gmail.com>

* confusion_matrix fix

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../ops/declarable/helpers/cuda/confusion.cu  | 56 +++++++------------
 1 file changed, 20 insertions(+), 36 deletions(-)

diff --git a/libnd4j/include/ops/declarable/helpers/cuda/confusion.cu b/libnd4j/include/ops/declarable/helpers/cuda/confusion.cu
index 513911f97..12f14b20b 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/confusion.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/confusion.cu
@@ -30,10 +30,10 @@ namespace helpers {
 
     template <typename T>
     __global__ static void copyBuffers(Nd4jLong* destination, void const* source, Nd4jLong bufferLength) {
-        const auto tid = blockIdx.x * gridDim.x + threadIdx.x;
+        const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
         const auto step = gridDim.x * blockDim.x;
         for (int t = tid; t < bufferLength; t += step) {
-            destination[t] = reinterpret_cast<T const*>(source)[t];
+            destination[t] = static_cast<Nd4jLong>(reinterpret_cast<T const*>(source)[t]);
         }
     }
 
@@ -51,38 +51,24 @@ namespace helpers {
         }
         __syncthreads();
 
-        const auto tid = blockIdx.x * gridDim.x + threadIdx.x;
+        const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
         const auto step = gridDim.x * blockDim.x;
         for (int t = tid; t < bufferLength; t += step) {
-            //auto tX = reinterpret_cast<T*>(inputList[t]);
-            //auto xShape = reinterpret_cast<Nd4jLong*>(inputShapeList[t]);
             auto label = labelsBuffer[t]; //->e<Nd4jLong>(j);
             auto pred = predictionBuffer[t]; //->e<Nd4jLong>(j);
             auto tZ = z + tadOffsets[label];
             T val = (weightsBuffer == nullptr ? (T)1.0f : w[t]);
 
-            //for (int e = threadIdx.x; e < arrLen; e += blockDim.x) {
-
-            tZ[shape::getIndexOffset(pred, tadShape, arrLen)] = val; //tX[shape::getIndexOffset(e, , arrLen)];
+            auto idx = shape::getIndexOffset(pred, tadShape, arrLen);
+            tZ[idx] = val;
         }
     }
 
-    template <typename T>
+    template <typename X, typename Z>
     void _confusionFunctor(nd4j::LaunchContext * context, NDArray* labels, NDArray* predictions, NDArray* weights, NDArray* output) {
-//        std::unique_ptr<ResultSet> arrs(output->allTensorsAlongDimension({1}));
-//
-//#pragma omp parallel for if(labels->lengthOf() > Environment::getInstance()->elementwiseThreshold()) schedule(static)
-//        for (int j = 0; j < labels->lengthOf(); ++j){
-//            auto label = labels->e<Nd4jLong>(j);
-//            auto pred = predictions->e<Nd4jLong>(j);
-//            T value = (weights == nullptr ? (T)1.0f : weights->e<T>(j));
-//            (*arrs->at(label)).p<T>(pred, value);
-//        }
-
-        int dimension = 1;
-
-        auto pack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(output->shapeInfo(), dimension);
+        auto stream = context->getCudaStream();
 
+        auto pack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(output->shapeInfo(), 1);
 
         PointersManager manager(context, "helpers::confusion");
 
@@ -90,26 +76,26 @@ namespace helpers {
         Nd4jLong* predictionLongBuffer = predictions->dataType() == nd4j::DataType::INT64?(Nd4jLong*)predictions->specialBuffer():nullptr;
 
         if (labelsLongBuffer == nullptr) {
-            cudaError_t err = cudaMalloc(&labelsLongBuffer, labels->lengthOf() * sizeof(Nd4jLong));
+            auto err = cudaMalloc(&labelsLongBuffer, labels->lengthOf() * sizeof(Nd4jLong));
             if (err != 0)
                 throw nd4j::cuda_exception::build("Cannot allocate memory for labels long buffer", err);
             // copy with type conversion
-            copyBuffers<T><<<256, 512, 8192>>>(labelsLongBuffer, labels->getSpecialBuffer(), labels->lengthOf());
+            copyBuffers<X><<<256, 512, 1024, *stream>>>(labelsLongBuffer, labels->getSpecialBuffer(), labels->lengthOf());
         }
 
         if (predictionLongBuffer == nullptr) {
-            cudaError_t err = cudaMalloc(&predictionLongBuffer, predictions->lengthOf() * sizeof(Nd4jLong));
+            auto err = cudaMalloc(&predictionLongBuffer, predictions->lengthOf() * sizeof(Nd4jLong));
             if (err != 0)
                 throw nd4j::cuda_exception::build("Cannot allocate memory for predictions long buffer", err);
             // copy with type conversion
-            copyBuffers<T><<<256, 512, 8192>>>(predictionLongBuffer, predictions->getSpecialBuffer(), predictions->lengthOf());
+            copyBuffers<X><<<256, 512, 1024, *stream>>>(predictionLongBuffer, predictions->getSpecialBuffer(), predictions->lengthOf());
         }
 
         auto bufferLength = labels->lengthOf();
         dim3 launchDims(32, 32, 1024);
-        auto stream = context->getCudaStream();
-        confusionFunctorKernel<T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(labelsLongBuffer, predictionLongBuffer,
-                bufferLength, weights != nullptr? weights->getSpecialBuffer():nullptr, output->specialBuffer(), pack.specialShapeInfo(), pack.specialOffsets());
+        confusionFunctorKernel<Z><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(labelsLongBuffer, predictionLongBuffer, bufferLength, weights != nullptr? weights->getSpecialBuffer():nullptr, output->specialBuffer(), pack.specialShapeInfo(), pack.specialOffsets());
+
+        manager.synchronize();
 
         if (predictionLongBuffer != predictions->getSpecialBuffer()) {
             cudaError_t err = cudaFree(predictionLongBuffer);
@@ -122,17 +108,15 @@ namespace helpers {
             if (err != 0)
                 throw nd4j::cuda_exception::build("Cannot deallocate memory for labels long buffer", err);
         }
-        manager.synchronize();
     }
 
     void confusionFunctor(nd4j::LaunchContext * context, NDArray* labels, NDArray* predictions, NDArray* weights, NDArray* output) {
-        auto xType = output->dataType(); // weights can be null
-
-        BUILD_SINGLE_SELECTOR(xType, _confusionFunctor, (context, labels, predictions, weights, output), NUMERIC_TYPES);
+        auto xType = predictions->dataType();
+        auto zType = output->dataType(); // weights can be null
+        NDArray::prepareSpecialUse({output}, {labels, predictions, weights});
+        BUILD_DOUBLE_SELECTOR(xType, zType, _confusionFunctor, (context, labels, predictions, weights, output), INDEXING_TYPES, NUMERIC_TYPES);
+        NDArray::registerSpecialUse({output}, {labels, predictions, weights});
     }
-
-    BUILD_SINGLE_TEMPLATE(template void _confusionFunctor, (nd4j::LaunchContext * context, NDArray* labels, NDArray* predictions, NDArray* weights, NDArray* output);, NUMERIC_TYPES);
-
 }
 }
 }
\ No newline at end of file

From 7f0c660d8b7ad69bcf4f5f733ea130ac1e1bc619 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Tue, 27 Aug 2019 15:05:43 +0300
Subject: [PATCH 23/56] [WIP] HGemm (#181)

* skip string arrays for device validation

Signed-off-by: raver119 <raver119@gmail.com>

* confusion_matrix fix

Signed-off-by: raver119 <raver119@gmail.com>

* exclude cublasHGemm from archs < 530

Signed-off-by: raver119 <raver119@gmail.com>
---
 libnd4j/include/helpers/cuda_off/MmulHelper.cu | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/libnd4j/include/helpers/cuda_off/MmulHelper.cu b/libnd4j/include/helpers/cuda_off/MmulHelper.cu
index dda709545..19e0d5baf 100644
--- a/libnd4j/include/helpers/cuda_off/MmulHelper.cu
+++ b/libnd4j/include/helpers/cuda_off/MmulHelper.cu
@@ -228,6 +228,7 @@ NDArray* MmulHelper::mmulMxM(const NDArray* A, const NDArray* B, NDArray* C, dou
         float alphaF(alpha), betaF(beta);
         status = cublasSgemm(*handle, transAblas, transBblas, M, N, K, &alphaF, (float*)pA->getSpecialBuffer(), lda, (float*)pB->getSpecialBuffer(), ldb, &betaF, (float*)pC->getSpecialBuffer(), ldc);
     }
+#if __CUDA_ARCH__ >= 530
     else if(ABC && aType == DataType::HALF) {
         float16 alphaH(alpha), betaH(beta);
         status = cublasHgemm(*handle, transAblas, transBblas, M, N, K, &alphaH.data, (__half*)pA->getSpecialBuffer(), lda, (__half*)pB->getSpecialBuffer(), ldb, &betaH.data, (__half*)pC->getSpecialBuffer(), ldc);
@@ -240,6 +241,7 @@ NDArray* MmulHelper::mmulMxM(const NDArray* A, const NDArray* B, NDArray* C, dou
         float alphaF(alpha), betaF(beta);
         status = cublasSgemmEx(*handle, transAblas, transBblas, M, N, K, &alphaF, pA->getSpecialBuffer(), CUDA_R_16F, lda, pB->getSpecialBuffer(), CUDA_R_16F, ldb, &betaF, pC->getSpecialBuffer(), CUDA_R_32F, ldc);
     }
+#endif
     else {
         dim3 threadsPerBlock(N, M);
         dim3 blocksPerGrid(1, 1);

From 9d325ad0700f23754b54405241a2aa5f9d253219 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Tue, 27 Aug 2019 23:27:41 +1000
Subject: [PATCH 24/56] Small optimization to Nd4j.readNumpy (#183)

Signed-off-by: Alex Black <blacka101@gmail.com>
---
 .../src/main/java/org/nd4j/linalg/factory/Nd4j.java   | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
index d2a4f94a4..c8baedfa5 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
@@ -2265,15 +2265,12 @@ public class Nd4j {
                 Preconditions.checkState(data.length == numColumns,
                         "Data has inconsistent number of columns: data length %s, numColumns %s", data.length, numColumns);
             data2.add(readSplit(data));
-
-
         }
-        ret = Nd4j.create(dataType, data2.size(), numColumns);
-        for (int i = 0; i < data2.size(); i++) {
-            float[] row = data2.get(i);
-            INDArray arr = Nd4j.create(row, new long[]{1, row.length}, dataType);
-            ret.putRow(i, arr);
+        float[][] fArr = new float[data2.size()][0];
+        for(int i=0; i<data2.size(); i++ ){
+            fArr[i] = data2.get(i);
         }
+        ret = Nd4j.createFromArray(fArr).castTo(dataType);
         return ret;
     }
 

From 0849b3c1a4e7fd746a5812b4ac0526aa22b4e83e Mon Sep 17 00:00:00 2001
From: shugeo <sgazeos@gmail.com>
Date: Tue, 27 Aug 2019 18:25:39 +0300
Subject: [PATCH 25/56] Shugeo segment fix2 (#185)

* Added test for segment_mean.

* Added another test for segment_mean.

* Fixed segment_* ops helpers for cuda to proper use external data.
---
 .../declarable/helpers/cuda/segment_max.cu    |  8 +++++
 .../declarable/helpers/cuda/segment_mean.cu   | 14 +++++++-
 .../declarable/helpers/cuda/segment_min.cu    |  8 +++++
 .../declarable/helpers/cuda/segment_prod.cu   |  8 +++++
 .../declarable/helpers/cuda/segment_sqrtn.cu  |  4 +++
 .../declarable/helpers/cuda/segment_sum.cu    |  8 +++++
 .../layers_tests/DeclarableOpsTests7.cpp      | 36 +++++++++++++++++++
 7 files changed, 85 insertions(+), 1 deletion(-)

diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_max.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_max.cu
index 180af41e1..dc91a2704 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_max.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_max.cu
@@ -201,7 +201,9 @@ namespace nd4j {
             }
             // -------------------------------------------------------------------------------------------------------------- //
             void segmentMaxFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* output) {
+                NDArray::prepareSpecialUse({output}, {input, indices});
                 BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), segmentMaxFunctor_, (context, input, indices, output), NUMERIC_TYPES, INDEXING_TYPES);
+                NDArray::registerSpecialUse({output}, {input, indices});
             }
             // -------------------------------------------------------------------------------------------------------------- //
 
@@ -240,7 +242,9 @@ namespace nd4j {
             }
             // -------------------------------------------------------------------------------------------------------------- //
             void unsortedSegmentMaxFunctor(nd4j::LaunchContext* context, NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
+                NDArray::prepareSpecialUse({output}, {input, indices});
                 BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), unsortedSegmentMaxFunctor_, (context, input, indices, numOfClasses, output), NUMERIC_TYPES, INDEXING_TYPES);
+                NDArray::registerSpecialUse({output}, {input, indices});
             }
 
             // -------------------------------------------------------------------------------------------------------------- //
@@ -370,8 +374,10 @@ namespace nd4j {
             }
             // -------------------------------------------------------------------------------------------------------------- //
             int segmentMaxFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, NDArray* output) {
+                NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
                 BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return segmentMaxFunctorBP_, (context, input,
                         indices, gradOut, output), FLOAT_TYPES, INDEXING_TYPES);
+                NDArray::registerSpecialUse({output}, {input, indices, gradOut});
             }
 
             // -------------------------------------------------------------------------------------------------------------- //
@@ -416,7 +422,9 @@ namespace nd4j {
             }
             // -------------------------------------------------------------------------------------------------------------- //
             int unsortedSegmentMaxFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, Nd4jLong numOfClasses, NDArray* output) {
+                NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
                 BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return unsortedSegmentMaxFunctorBP_, (context, input, indices, gradOut, numOfClasses, output), FLOAT_TYPES, INDEXING_TYPES);
+                NDArray::registerSpecialUse({output}, {input, indices, gradOut});
             }
         }
     }
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_mean.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_mean.cu
index 3f2168da4..fbb45a375 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_mean.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_mean.cu
@@ -163,7 +163,7 @@ namespace helpers {
 
         classesRangesBegs.assign(indices->lengthOf());
         classesRangesLens.assign(0);
-
+        NDArray::prepareSpecialUse({output}, {input, indices});
         dim3 dims(numClasses, indices->lengthOf(), numClasses * 32 + 32);
         int* begins = reinterpret_cast<int*>(classesRangesBegs.specialBuffer());
         int* lengths = reinterpret_cast<int*>(classesRangesLens.specialBuffer());
@@ -182,11 +182,14 @@ namespace helpers {
             Nd4jLong* outputTadOffsets = packZ.specialOffsets();
             segmentMeanTadKernel<T,I><<<input->sizeAt(0), 512, 2048, *stream>>>(input->specialBuffer(), input->specialShapeInfo(), inputTads, inputTadOffsets, reinterpret_cast<I*>(indices->specialBuffer()), begins, lengths, numClasses, output->specialBuffer(), output->specialShapeInfo(), outputTads, outputTadOffsets);
         }
+        NDArray::registerSpecialUse({output}, {input, indices});
 
     }
     // -------------------------------------------------------------------------------------------------------------- //
     void segmentMeanFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), segmentMeanFunctor_, (context, input, indices, output), NUMERIC_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices});
     }
 
     // -------------------------------------------------------------------------------------------------------------- //
@@ -194,6 +197,8 @@ namespace helpers {
     static void unsortedSegmentMeanFunctor_(nd4j::LaunchContext* context, NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
         auto stream = context->getCudaStream();
 //        NDArray classes = NDArrayFactory::create<int>('c', {numOfClasses, 2});
+        NDArray::prepareSpecialUse({output}, {input, indices});
+
         NDArray classesRangesBegs = NDArrayFactory::create<int>('c', {numOfClasses});
         NDArray classesRangesLens = NDArrayFactory::create<int>('c', {numOfClasses});
 //        NDArray row = NDArrayFactory::create<int>('c', {1, 2}, {(int)indices->lengthOf(), (int)0});
@@ -221,12 +226,15 @@ namespace helpers {
             dims.x = input->sizeAt(0);
             segmentMeanTadKernel<T,I><<<dims.x, dims.y, dims.z, *stream>>>(input->specialBuffer(), input->specialShapeInfo(), inputTads, inputTadOffsets, reinterpret_cast<I*>(indices->specialBuffer()), begins, lengths, numOfClasses, output->specialBuffer(), output->specialShapeInfo(), outputTads, outputTadOffsets);
         }
+        NDArray::registerSpecialUse({output}, {input, indices});
 
     }
     // -------------------------------------------------------------------------------------------------------------- //
     void unsortedSegmentMeanFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices});
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), unsortedSegmentMeanFunctor_, (context, input, indices, numOfClasses, output),
                               NUMERIC_TYPES, INDEXING_TYPES);
+        NDArray::prepareSpecialUse({output}, {input, indices});
     }
 
     // -------------------------------------------------------------------------------------------------------------- //
@@ -349,8 +357,10 @@ namespace helpers {
     // -------------------------------------------------------------------------------------------------------------- //
     // segmen mean bp main
     int segmentMeanFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return segmentMeanFunctorBP_, (context, input,
                 indices, gradOut, output), FLOAT_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices, gradOut});
     }
     // -------------------------------------------------------------------------------------------------------------- //
 
@@ -399,7 +409,9 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     int unsortedSegmentMeanFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, Nd4jLong numOfClasses, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return unsortedSegmentMeanFunctorBP_, (context, input, indices, gradOut, numOfClasses, output), FLOAT_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices, gradOut});
     }
 
 }
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_min.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_min.cu
index 0c67b41d5..950abde67 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_min.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_min.cu
@@ -192,7 +192,9 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     void segmentMinFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices});
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), segmentMinFunctor_, (context, input, indices, output), NUMERIC_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices});
     }
 
     // -------------------------------------------------------------------------------------------------------------- //
@@ -233,8 +235,10 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     void unsortedSegmentMinFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices});
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), unsortedSegmentMinFunctor_, (context, input, indices, numOfClasses, output),
                               NUMERIC_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices});
     }
 
     template <typename T, typename I>
@@ -364,8 +368,10 @@ namespace helpers {
     // -------------------------------------------------------------------------------------------------------------- //
     // segmen min
     int segmentMinFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return segmentMinFunctorBP_, (context, input,
                 indices, gradOut, output), FLOAT_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices, gradOut});
     }
 
     template <typename T, typename I>
@@ -409,7 +415,9 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     int unsortedSegmentMinFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, Nd4jLong numOfClasses, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return unsortedSegmentMinFunctorBP_, (context, input, indices, gradOut, numOfClasses, output), FLOAT_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices, gradOut});
     }
 }
 }
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_prod.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_prod.cu
index 78f21916d..3ae4ebcb8 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_prod.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_prod.cu
@@ -192,7 +192,9 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     void segmentProdFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), segmentProdFunctor_, (context, input, indices, output), NUMERIC_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices});
     }
 
     // -------------------------------------------------------------------------------------------------------------- //
@@ -231,8 +233,10 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     void unsortedSegmentProdFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices});
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), unsortedSegmentProdFunctor_, (context, input, indices, numOfClasses, output),
                               NUMERIC_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices});
     }
 
     // -------------------------------------------------------------------------------------------------------------- //
@@ -358,8 +362,10 @@ namespace helpers {
     // -------------------------------------------------------------------------------------------------------------- //
 
     int segmentProdFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return segmentProdFunctorBP_, (context, input,
                 indices, gradOut, output), FLOAT_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices, gradOut});
     }
 
     // -------------------------------------------------------------------------------------------------------------- //
@@ -404,7 +410,9 @@ namespace helpers {
 
     // -------------------------------------------------------------------------------------------------------------- //
     int unsortedSegmentProdFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, Nd4jLong numOfClasses, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return unsortedSegmentProdFunctorBP_, (context, input, indices, gradOut, numOfClasses, output), FLOAT_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices, gradOut});
     }
 
     // -------------------------------------------------------------------------------------------------------------- //
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_sqrtn.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_sqrtn.cu
index 4141cefba..229d41cc9 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_sqrtn.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_sqrtn.cu
@@ -146,8 +146,10 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     void unsortedSegmentSqrtNFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices});
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), unsortedSegmentSqrtNFunctor_, (context, input, indices, numOfClasses, output),
                               FLOAT_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices});
     }
     // -------------------------------------------------------------------------------------------------------------- //
     template <typename T, typename I>
@@ -270,7 +272,9 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     int unsortedSegmentSqrtNFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, Nd4jLong numOfClasses, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return unsortedSegmentSqrtNFunctorBP_, (context, input, indices, gradOut, numOfClasses, output), FLOAT_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices, gradOut});
     }
 }
 }
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_sum.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_sum.cu
index 37dacee09..08b36253a 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_sum.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_sum.cu
@@ -190,7 +190,9 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     void segmentSumFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices});
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), segmentSumFunctor_, (context, input, indices, output), NUMERIC_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices});
     }
 
     // -------------------------------------------------------------------------------------------------------------- //
@@ -229,8 +231,10 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     void unsortedSegmentSumFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices});
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), unsortedSegmentSumFunctor_, (context, input, indices, numOfClasses, output),
                               NUMERIC_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices});
 
     }
 
@@ -343,8 +347,10 @@ namespace helpers {
     // -------------------------------------------------------------------------------------------------------------- //
 
     int segmentSumFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return segmentSumFunctorBP_, (context, input,
                 indices, gradOut, output), FLOAT_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices, gradOut});
     }
 
     template <typename T, typename I>
@@ -381,7 +387,9 @@ namespace helpers {
     }
     // -------------------------------------------------------------------------------------------------------------- //
     int unsortedSegmentSumFunctorBP(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, Nd4jLong numOfClasses, NDArray* output) {
+        NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
         BUILD_DOUBLE_SELECTOR(output->dataType(), indices->dataType(), return unsortedSegmentSumFunctorBP_, (context, input, indices, gradOut, numOfClasses, output), FLOAT_TYPES, INDEXING_TYPES);
+        NDArray::registerSpecialUse({output}, {input, indices, gradOut});
     }
 
 }
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests7.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests7.cpp
index 2e1dab1a3..b0488c23a 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests7.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests7.cpp
@@ -1423,6 +1423,42 @@ TEST_F(DeclarableOpsTests7, TestSegmentMean_2) {
     delete result;
 }
 
+TEST_F(DeclarableOpsTests7, TestSegmentMean_02) {
+    auto x = NDArrayFactory::create<double>('c', {6, 3}, {1, 2,  3., 4., 5., 6.,  7., 8., 9., 10., 11., 12., 13., 14., 15., 16., 17., 18.});
+    auto idx = NDArrayFactory::create<int>({0, 0, 1, 1, 2,2});
+    auto exp = NDArrayFactory::create<double>('c', {3, 3}, {    2.5, 3.5, 4.5,      8.5, 9.5, 10.5,   14.5, 15.5,  16.5});
+
+    nd4j::ops::segment_mean op;
+
+    auto result = op.execute({&x, &idx}, {}, {});
+    ASSERT_EQ(result->status(), Status::OK());
+    ASSERT_EQ(result->size(), 1);
+    exp.printIndexedBuffer("Expect Mean");
+    result->at(0)->printIndexedBuffer("Output Mean");
+//    exp.printShapeInfo("Exp Shape");
+    ASSERT_TRUE(exp.equalsTo(result->at(0)));
+
+    delete result;
+}
+
+TEST_F(DeclarableOpsTests7, TestSegmentMean_021) {
+    auto x = NDArrayFactory::create<float>('c', {6, 3});//, {1, 2,  3., 4., 5., 6.,  7., 8., 9., 10., 11., 12., 13., 14., 15., 16., 17., 18.});
+    auto idx = NDArrayFactory::create<int>({0, 0, 1, 1, 2,2});
+    auto exp = NDArrayFactory::create<float>('c', {3, 3}, {    2.5, 3.5, 4.5,      8.5, 9.5, 10.5,   14.5, 15.5,  16.5});
+
+    nd4j::ops::segment_mean op;
+    x.linspace(1.);
+    auto result = op.execute({&x, &idx}, {}, {});
+    ASSERT_EQ(result->status(), Status::OK());
+    ASSERT_EQ(result->size(), 1);
+    exp.printIndexedBuffer("Expect Mean");
+    result->at(0)->printIndexedBuffer("Output Mean");
+//    exp.printShapeInfo("Exp Shape");
+    ASSERT_TRUE(exp.equalsTo(result->at(0)));
+
+    delete result;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests7, TestSegmentMeanBP_2) {
     auto x = NDArrayFactory::create<double>('c', {4, 4}, {1.8, 2.5,  4.,  9.,2.1, 2.4,  3.,  9.,2.1, 2.1, 0.7, 0.1,3., 4.2, 2.2, 1.});

From 650539528cc7e78aebd76e29d1d8cbb264407bf4 Mon Sep 17 00:00:00 2001
From: Alexander Stoyakin <alexander.stoyakin@gmail.com>
Date: Tue, 27 Aug 2019 19:56:04 +0300
Subject: [PATCH 26/56] [WIP] Updating failed tests to reflect code changes
 (#184)

* Test updated to reflect changes in random generation

* Invalid resource removed and fixed tests

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
---
 .../clustering/kmeans/KMeansTest.java           |   6 +++---
 .../models/fasttext/FastTextTest.java           |   2 --
 .../models/fasttext/supervised.model.bin        | Bin 99274 -> 0 bytes
 3 files changed, 3 insertions(+), 5 deletions(-)
 delete mode 100644 deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/resources/models/fasttext/supervised.model.bin

diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/kmeans/KMeansTest.java b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/kmeans/KMeansTest.java
index fe4fac1b7..c9140942d 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/kmeans/KMeansTest.java
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/kmeans/KMeansTest.java
@@ -269,9 +269,9 @@ public class KMeansTest extends BaseDL4JTest {
 
             double[] centroid1 = {2.44e8,    2.71e8,    2.98e8,    3.25e8};
             double[] centroid2 = {5.14e8,    5.41e8,    5.68e8,    5.95e8};
-            double[] centroid3 = {1.63e8,     1.9e8,    2.17e8,    2.44e8};
-            double[] centroid4 = {6.76e8,    7.03e8,     7.3e8,    7.57e8};
-            double[] centroid5 = {4.06e8,    4.33e8,     4.6e8,    4.87e8};
+            double[] centroid3 = {1000000.0, 2.8E7, 5.5E7, 8.2E7};
+            double[] centroid4 = {7.03E8, 7.3E8, 7.57E8, 7.84E8};
+            double[] centroid5 = {3.79E8, 4.06E8, 4.33E8, 4.6E8};
 
             assertArrayEquals(centroid1, clusterSet.getClusters().get(0).getCenter().getArray().toDoubleVector(), 1e-4);
             assertArrayEquals(centroid2, clusterSet.getClusters().get(1).getCenter().getArray().toDoubleVector(), 1e-4);
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/fasttext/FastTextTest.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/fasttext/FastTextTest.java
index cd3737e04..b60af71d5 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/fasttext/FastTextTest.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/fasttext/FastTextTest.java
@@ -103,7 +103,6 @@ public class FastTextTest extends BaseDL4JTest {
     }
 
     @Test
-    @Ignore("AB 2019/06/24 - Failing: Ignored to get to all passing baseline to prevent regressions via CI - see issue #7912")
     public void testPredict() throws IOException {
             String text = "I like soccer";
 
@@ -119,7 +118,6 @@ public class FastTextTest extends BaseDL4JTest {
     }
 
     @Test
-    @Ignore("AB 2019/06/24 - Failing: Ignored to get to all passing baseline to prevent regressions via CI - see issue #7912")
     public void testPredictProbability() throws IOException {
         String text = "I like soccer";
 
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/resources/models/fasttext/supervised.model.bin b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/resources/models/fasttext/supervised.model.bin
deleted file mode 100644
index dcf5646840872ba7d78b30769b65a84542ba67bc..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 99274
zcmZ6y2Rv5q8$WLEz4vZYQL5)$cT-DCODa*6N)bgvQ%Z@9tVoeUHc5ESbw?V=2$7Mj
zLZzhAF!Ou9-{0^5{d`{k^LpLSbFXuq`&{Q9?`yFtRx_md`1tnllJNga{QvqF;N}05
zB6&%Wmqd8!KYdLsv_8L^qp{G8m!HW?|LL=wm;Q5&|G$gO@L2HQ8~$56oloe$|GXb(
zkN?v8Z*iZ4tJ6Ov&;3sSCH3DX54YXBosRrdUieRTubZ2vql1gfKQ*)X{xj_VUZ!UM
z<UJh@I(ht4JnC_jcbC^c6^sALT5tL%?|kT=yt|wGQ5T0J|5UiTd3gR)+V6Js$iJT8
zaA?mzZM$8LI{vH8)%nm-&wmeRsoOt&z1)ueQ*d!U==4v?@?SJ~-TpO{y$;9RjyQY%
ztM7mK`LFKlk2twH^O*QQ=4|@EO#NTk;h!dV7Y8q=e~sSpUwraB{;#2WIyt!h>lto)
z|DoaJ;&jaE-^}|D&42S}Gw%!k@8Qe;zt8x;nl(-i`;Pvrhr_>UEjzT&#o1&3KexLd
zaog+U;o<Cd$id~GT9^MHiM?+BdVN=?LzDhBbO#UKBy)D~<dOJ?z|Ng64vtPPJ9qxS
z3HYCR|DTKhe+~bee*bIYEBU{D{<}E-n;jOVbR!=*Xjg#`GllVHWE}im7t5q}l*6j{
zqnrnf0h!gKMm*jWvv01fgxN%uSl#S}jL}OdUEIz#Ou5M^{qzV1wD&U=lZ9z(*nM`M
z^gEEK$$>N9Um$&y4~4I5@b|H5#xTwU_1PXaY=b^Iw&*QBIldL0S|#Y~wXabvR+Sny
zPGx=P?15_|(zNB69G0(a!K6Ap^0ns&+*soXF1_8%v%uT<=+gir920@<2O_b4pAfxJ
zXaf2EUASIZkr;L=)4lRHK)ri5Yt--%7wthNT~-`sK0OaxT5LeE<29DmCV;@VdN|US
z2qVU>=$kT?i5c#}Hzh?7V<Jj^eR&Koi}h&JtBsudzkKnt!YxKzAZy-`pcf|T>yRy`
zYQ&-868qtqH-0-PPh(r3gTWOYazj&*KJrb)Iez>!z-<om_IL;9K(jsih`3=|gccc+
zy8-$MN@Q!iB6Xcq$o%S&rIAb}C<tqlw#)};f8-Z?`<ND-?@0oVniyH$tV;K0H-V?6
z5Lt7(9@X-)nPZ0>aOOpEyxGUFQb((C`;Az1mXe2GSb`VjKclIxJ(T1=hlNs|cw<1B
zO$?4hxoUlq=;Q-dG4H|1Qkotgd4UU!^y#AoajbKg9NLwAMGLnC6gj+#9klX+$r6uo
z{?-~q4+(5{>qU!>t1xk$4EtE#9B*APq~ZY@q{%rM(^@UT#8-|yUOd7)I4VSn#%(cX
zi8sE;-V1{_FR<igKbtI~3EAszL(y&}dMVKf1T=N%R}XW>I7F5hFVMvcs};DHU@z1}
zv(PZeN91mW!}j~msO9ky6O7W)`Y*+-LoA-YRmmJLPK5l>LdaY24IiG-!XLXQ;QX2A
zP`PG4+=!`zKa;qiF2l!MW()Dr$PAE_nghc(ZbCzj3k3fX#OPpsS|%5R*Hym4$93~z
z#SH_<>ze?%#-A9e8){_funH{moXE-)=zu=A9UCY0;`KTG;Hf7>)g|58txkJz$F)e#
zTGel8HY!E$PSKzvM|8-sbt>eOi6NaUtb@7gf7qtMr?}xF!TY7rP^)VMUmgO!>KjJc
z)f-`>-Um=NZ9v(ES_IP>@W8a1YpjvWHXKcYI$I5T{ihh+5FrhBVL4<;eT3=TLb0fR
z8Dt%mqEq+u;=>ugaDl7>xgjA-%5}wwNXICKn@7QAubUt;Q>kEL`!SIC5DA87gh`Cu
z2)y}G3gp~3uxYGet&UW1y_@cX%vCO~snDS}+yzMQ=ATSwttjnMZ^5&rS8!^sE}6JN
zlE}<ygQnHZ&^DzS{|>B$qq2F(CoGS>cP@b3S9$t-!5665vlsC5PLQ+LqrzoMbf*3m
zxNt;{EIM)qM`e}B6v9Wl1!M>}zz7db=|Rsq$rzd<i35{5(I`a~m5MY;Qu=E+rQ(X4
z0$!r*3<Wk<s}tNM*5J12tuXi_jV-B@#8BOIs10}wq5g72@>wqD_Agapu;&Hlt=x&`
zrZ&uN9YMmedWIfuRot>kCD5|09s=C`v7IUq{qiq3%P0uW?8<`s!TGpFs|~-E|Hc01
zap>8%0<P|sAhM}aBv<D;_p?zhGs=F%@3(IPXDuH&+p2~?->A@p`*%>Sy$ST2GvNJJ
zMYj2;1Q~N2WuH6qk<oH#SfF+hwMV<zIo1KNa=I+IE^Fsrv-L(P>o&Y}dovr*9)d-=
z{h)E~CL1gq33FT;Vf`viny~l@PB{OVc^mr{ZwLKh!~bNURCyA2#=_Zfwm2B3ESSUe
z>=vfWZ<_$0#9Fu$BTg^9c!%zXHJ~hU9`iOh6W!wSSr_-qs1YYkOfF1;2yJC1_VYy$
zeA`{{=b<=ZKHdcDR#j4L*9;bQ!|0!Wg>kT52aa2Xh^>1Dqz!42%ag<AmI}wR(}#Y8
z{8m+3_(hIRu!>?H?|qKPWWAV!xvL@IL>X@U?1P`<i#czjg22I47u`On0(<Hzh8f=h
z5_FRD!t4bTv0j5Hx2uwozhPhytU(v#WrJ-@0Mul@<ci$7fh+FEpw97`AeJOclsz{w
zNsaSB=!*g_no)>gs7UN0MTzt~3#?TS!;M0!^qQ~**p+#~j_ayqyO}$fAN~d@rSC!N
zZ#FC!k*5iF_-RDj7uL+;G$SkA0~rsMX#OKLBDeez1bZxIs@5DuiB3I~PaTAX9qL4~
zAsSpy8Pcj)M>aiMmJ~XSvB@(nk(-%}o2)W`Rs?~+PA2D8mIj$Ivmd<WQ!pSh5mtQA
zq?+R@WNr2mXd3>&G&+SKZjWI-EWX0E-Yh8p7=edR$kF-&E;_FCgRT>oQEJ{Xw&VB@
z99rK6Mg?~;vhfPa3UwkW)}^QPB4Bfa9&Vq=N7H@AnHR^FG3&<cq5iiS+ZwtIldUGe
ztT^7rVH=LiuK$G6jC_b~U&&nSR3zcus%%4x8id^&25sBVTrQELyN}nPPVO@t;a4Zc
z&k~@?>IJBiF#NnH2_}a+!`nuFT7Pj4O5T#B$6M5ig3_UR@kZyMG9wLjlAp8n3+hn+
z&3Tw-`x-6UL}`S}cc|txV(ghu_#%9m(Jik9k98KP%G)pNGG2jCvuy!m){GarM5(S$
z6V9CW4vm%C0airgpRZx;mg(M5q*(>Azdxhww?wqMB}Q)rDv&vUD^bAdIrrvJ5T3B!
z3_ov<LwVmS++DdC6NmKavJI&i%nFf92^2>*O3=boJ(}tL3MRM*!0s25Kz-sU!_@!b
z#%1s^*WYTQD@an`%oCV)Rg!!=?}58_#=(<O6L_#mm+r5w!P!D-Xe>Mep+%AymeB<F
z-s+IDyZW?C)EetjG%&(zIXrS4LpGuepFG)#gPCbyDw_r^N2kDut|2Mf+5|xz_t}tQ
zHM*Re%Q!!jr2`Xn$nqK2*_oGH*akZhnlB{*#Il*Q$agYl$`2`8`AUzq;D3sC%ouWY
z`ay!D&pvr0P3~HYQ8g_u439ViM=P$v;JQY3s^S4=Pq`Rlc1nZkD|`y)tn}$1Z{1v(
zm4Z7P#EF4#Cu?t@geh|s8ELz_VBkLlO`&Qe)Zi;)8F`&|KJmkl^aH40l8+njEMXtV
z3~;{RRwXY*ny}H^khWed!!^NQna<y5V7%%%=thUZ<JbG(m5DG{@BSrbY~Mv3ywd~|
zm!(6;!8;f#I1Dcf%Fro6iwLY5M$>cqxOdu0(0w!smSxsL?e#~vedBS?lkr<nl4p;r
zA0@Jem)F2qVPPV9c#O5X_L=+O$t#dDQln>cmFZ)f=Q!W`DFjS2pfl!Ngik#uSiTE}
zq`#7%4ysFm@0ASvJ9|F*{>ov3%T6-&Yc;r48wBXerF_)NeI+^y)x+VHW3Z<8GCZ1J
zi=yFOaA;Z@20!@=!B4XwFi4U7bPeLH8-K<w$i9KrxvQ~_9%uF9n=$j0AbmX~h>I8b
z;LRjqQn6B<p6hJF*)jWZLW?L>PisQ`?~-J}HxnHG?2DaNq0HpyGfc_Z7If~~4a;Uq
z(~oZjIo`o)WV2s2O!m0T)-6oKl{M#>3x{MuHZ2|oPUJAA9}l66;afc6I2%S*7NBif
zA+EBH!erl6n6WAfHIoz}*k=t}VelJ_PO6csmRI0gggTurl?WSW36jnWigb&pFp2Wj
zgqHouuvoJICh%EsdMn>Ref}M=h~k~u!9$S89f8cc3JCq9O_o*E;bVCn*6>Fzc!o$2
zkzO5`u&Ia5df39*DelL-UHu4>I=<l8{;O<(zaYJ1u1-!3%F!&#*SPM*cXZ3)z~3)M
z7=BcmzCIj_l@p)im4k(l6d^-@h6XYUXT(Suk2hiq$}#2o8xS-O$AyWDu(nYHolX6)
zt$7#@H4BoH-Hwc-i!#}HeFr;5_8RO6CCn*Uk1^Y|$mD@t_*O-Qp8Fetul)_E&%9#v
z{%Vb8mU=XJdN)k8<DlZ=638SEvG9R7Nu7ECY2h~rd!a`Uy#LF7Go1<fC_yd$?#GjV
zWoY47W5iSEv2T|isXKQW?z_Fm&HKxldnzgLBvyj_nZ1bBnj}lhOJ-qj{w|nYQH<_2
zrQCTEPB_-0L|Tt1(bAc6<mb;cAd3f>&yIQ|cwszbomRvY=PvvyX8^0VNE6kNi*V4p
zj>$QqO$0I}sCK6|6~C`VZiIXQ=`TM)f(X(ptB%5zH^~qOD^Ob9fW*8zgeUEzq22X0
zzES!CQFYC%^Tw~>XS$6INK~aZy;pGPcOslUH50o8s+gPI&%v-ppX?qs!E<}t@$f|%
z^2aO+Wq5df_gP!qvvV)#K9ZwGE9KzGA2DLAt4ik07=WJ+b*LCq0yzefWa~>Ua(`Mi
zdlD?MN^+1r@@p8YhvZ1a{7&fYOoj7qI#6}qkF&~50Yc?&vrE6nVwk2RnZUdR$t%&=
z$fttU5A@Ocvj&;O{R6Mu4`5<u5Bh4aU>|sWXA)<3L7zf3>$;)~j76hh;!$tLrP2?}
z#NOlD*&ng*`EAS@dkqs@v}oeFKR7F-3}q89;?pns^zw-b@UZU$9^Vy;iRTueY_m3e
zK5_}S<Ve$3^Dl63NcVu;(J?Hamj{dQcVP3>sbD3_quCPzpEuk`zBztaA@GKs*Y^;W
z=SULYNgU=?fgOiDvw~;08{rTug&+J+!V9l(*2}+<qw*z&xAw&GiSt^vtgV?nvNV9r
zlM|$!l}{M04dPgP_92X|3WG0lTcEp3nH0BZ6TZ*+P_+Lh<}BBRBR^&pj0E0*273>@
z=iP~uidJC47bUW1zboFhkA&;s4mT&H;{B+4#{T+K@K~t9(JLy1d8^mrrtOi;KH3Zc
zPxD#FS}9Wa<pr94$Op#*YSdb)6{CPBx1Kko13q_P_of`ySARu;i?B8X8IM4QxIX!w
zdk;-4ROsf3Rv0TMP0465Gjb#rLze2(2)jmzJtYI*yDL!HSb*@qQYCE-cbTn9L+muu
zM{Ke4JT_xc0}pgd5Q)f@Oy?;9PVuT4Ou$(M%$l_tEtiY3UQau*GE0H1z1jv|0p^g>
z!m_?oHHluaA>Em99fvxGuvz3c)V_7VV|&k_ylw|`Fj<#$d5+=5#3*juTr1r2tN{wm
z1gL+6ZNVWqYqtJy3M*-$MDM+i!;<f*?3qvU)F8<MV^rE;i@PBaADxNI&!z)kLKCW<
z(qe`k_2|&6Ky*LG<CE+KcHw4QC{OLfr?U!~z&H(J^u8UEto#_Yh0j4JwGf2Po`w);
zfUfP{*q)_BCUUNV-?o*^i{rO&+G-7A#wd|rb}D2Yw*(IR{KChNCu6#P59<<8$kFT`
zh6jaL!DJ`{f5n^O?d^q7cF7s#M%Ljz&?df7^%$n-joYglITdR@q1|T<ymu%Wi#~3G
zAm?saGa^POJ=LSK$tv{z=0bR0G6SR9GH?N(H>^mIB-c2#@akg%W|hm3F6)a}pG2WV
z!;s{6?q<S`gy@;F7+j~@jW#Q@8M%Ufz<Y~W<L+`;mSV$}R_SxjY#YX?@><S~X=0#|
zX$4ZDWpLWfhjE))1J+(0n0&Peto_d7pl3WV!@5*3$sJ;xPr+e7St4^>2*1S}(yOy0
zn9m<%nJb6q&b^&D4WEY1WOw)sF_yD;fzI(b2<Y#ISv$pv@cS5;bTXfzcU@UCyQhrt
znvW<eHxBWp%~<=UlrwtC2UN#|=rVr=G7u07{)=?sONtYobL)W4rv~)<>Su7RLJy{=
z{DxoYV&wXMVZvAKkE+xaH?9u?EAKwMRUiU)f84|Kl^o_#mm+yteH&U_t8jnvT3CES
z77G^|p~&q&FyWjWZ53?;S21t6FItVk76|^l{r}mA0yti#Kt>a5n9><FjQu7x+9Nj=
zqdjCvZI?6G;*&OEJ2D|&&zybMqCjH<wD8OyH9EDW6hFXMMpXViKCk$IFG6yV%aJ4A
z@im;KP0BP($O+etdO~*FR;<}9gCeH#g#D;Tif#&!^{zJ1`QkboYvjtF`?L*rEgJ>#
zdr|mfg*4GSQ_4MP83q2QzcaITug48j2XJjg9S9v3BXWX#<dZ=N-tB(QBrIG7sl0v6
z)<Fci3Y2I`xhvoa8RDnjfOV&MGTn|gHd){)wC;X@r`D+v@g=`-hvzT2yqS;A@s(!J
z+>;=_&;By*hf|q}@1x+7ln&jq=m@JOcprO*tWh(Wi^Jxln6c9xy*q|ceQ78jIky!w
zH&#PeVG1ahyR(6hh3JWWA3;S+j`sD86s&KIh8;E8_@mn&dcV|xt!xoYjlYGb7ge**
zn}p~l+qE$HjuQF9x0P9<+Ydj*6PXDc-Y}7U97voOfo7YsK-lFxb4F$jld@kE-^wQA
zctjY}pHjl8<tx!=<*(VK9x3|LXBVDX(}Irk@4?~as?^ATBX_yqFFf1t2*<_`;3(gH
zsQ=UhE(O`>vHJ-`*zbn8;HS9ooe&Y$H-x?BdccXk1xD@C^yZWl80wjh<;pK$zg{-<
z=82LU4%+nU3TxE<v<SlkgV3b$77jeT%`A-1fQ+dJp~F~&gvcgAhgKVCa?6?iree@A
z{=nf*n}lM0cQE7wKWPj83N?))^xO?O+WsUG-WKfP=>vIZi9|iq{&pXG*!mU(Y$=By
z&JXeIZ+U8dDTFDTash_y9dS&-giT-i3vA3~sp<3cV0J=_rfBP9rNtLaDZT;=Ul`I|
zC&MA~hb)!Mxr=(MW`lFIK6#NWk8ADoaEnYF_UTzOYJYA)uq>e32UWUrT?s6b4u@a%
z>)Cs;_t3Ox2p<>=lcPtKDWBO}+`Xuq9WJ{M4dTIYil3i`&+LXleP#T;;S)^#_zM(g
z@lm5yOEBSG4qW}r!<!@lZZ^b2k4X)dtrkYN2ZhWQ2Q_TmtcX%>+T{2gD|8j&hZIJV
zj!jkr!<<n_Ppp7NE41kyBVD*x?}{g0ePeF5?gpc%K-{!co)BdnI+uE)_UhYMracRv
z=U(H^E*Byb&L*N?u>rX>zZtKFmP7VqFXmTL4PG1<hJ9i8;JN8f5d5W0j|@JBpYnI%
zOHDA&S@ai*Is~ZKf=&#LQ6#6F;&F;pJx8K29n4J&u`ybKrkw3SrQKEV@W5?+wdp3e
zefmfC`KciG+EN`NnUaro$<jo(a6VY+4k9zL71*J*nCJ14{ln9--A$^+jIhg`+zbYb
zUrz)vs}NXO*NQ)<OoY*4O`53R3?e_|$d;%!%wPYIyHtM;BONuyIxJWXg=a;maLioH
zkM)O}5nA-vZ$-++s*=B~IME*CC)(r$`bSEU^ZDrz^h$^xS;4aU`hr9<TLUXiZ(!u{
z0;nq$f&R+P7&cFy?%|3Mw;49D<xvI(i(dq<3xZVJO^t3DUyKVb>|*AeMrLTvMRxuv
z1~xx`4I_JeaIup#{qg+(SWc^f!JZo2`}+$PCwk-hXK&!c2^RkRkf-66-OLYz$uJ|P
z0`yc3$)=05Fl$N`hxqm56zdPTdqkG#Su0ZVR*jxJph?1#FLRs?9O2}oY+RMBPTomr
zQjhO>xGmEce#P^VGmq7YVs<T7c64Iqq#{&Xph0(er(xXmD-8FB4hcMdjXT(r0`eo8
zq@v&z1i!qB+gA!xHSTH{h%g|Z%F3B-o@20{c;m)29rEQG!<cv-z)}k?Y?<{76qwl<
zmeq;w8~fngg~PbtEgZJmzeL|3ZMbEdKkQob3`&1j;HF(FMEgt$Bk}$P_DeQH&gGAA
z?}IS$Q8u9Sym@k2q9Cy)PAIU=8qeO5rBN2j&^19Fp7Zr^*38u>MOj<H;K>5mlc9)$
z55D5)yHI$ct3vhsvYC($J$kEKliI${f<JBX@N3fnG+S82MsXg&O7BqYn>Zi0oMk~H
z)(i~$&cNyzQ8?%7jlP#y=6-<-c)x3g42^0Wt~-x!D<z2$xd%<83HU!hhBKAP_^KlW
z9*!J?$#YBbjq7*jRDdJotCV7Vk|wV6^<%}iP~fxIB`K?4Lbvc1CeCXHjE>d8`W2q+
z-p;AGcXb?8YU-06mMv`3&XbI4#R$GUphJs)y~U)YgRI4=6Uad&vfgYG8)(vxcDohG
z@@bEtDzyumm-#{5DM@lnP>5};;)3e*kMM5$Id-e=bBMiQ02LLvOhaic23%2LOGzNR
z@45zA@_Z&%KC5S6t-Z`RkDh_3HXZsdNtxVxEe)=oiXitb0O)*KQs>sjsN1B#wXh3V
z{>u+d)?J3JM-++WhAL?GDS=O)Pje^wDF7c2+u26*lf3c**d<bk-iddaxp~D*+evZ4
z-6jGrwv~W=usYpiBu9U`y@9D$Gr>hcfIOSf1-{M-<WQw9c_`rw<s<s^>aid;_Gdmg
zI6i@*v&MM0;tSl0A4EmvZ46^B0o=^{?4cQoG-@^<6&0b-+M59+PM$8mq=vUo7=vkf
zHAdO3<BIF)&~JUZMD5LQ2!j`#8B%9i*CVR5voQjOl?S*ho~qE_`B%}wxBxjnzQHTT
z5Ki=s-~ly0JlPuo(OZMj!|njO3WqW8hzs6(^pZ1Jvm5g>TjBX`Kc+!I6<P|bQ9R^3
z7?}QLN?{SMcyR(ugRSBC^Dazq&I8a%#%qNc$mtKlp`kIHRWcVO&F|sK25;QFOMuL6
zZ^K&=!BF>nl!^ExiOSt!Y|~N|dO*Jx?@90zjmyg9!)JROdioQ3t8#$9K%LmNDPrW2
zFHq~#j#n-{hY7{WjNoP)Ca#fT&UDJtOZHk=H|a0*zW)pZE#hR*RhLvw(80i0LYSDL
zP3uig!|7s4R&!zsGx@e4w&iw#owWqbb(SN^3mWk1ytiO8)fW1)D>*z%H6F7tMYEN1
z7||1lg=Sn>yQ3aHt`Q(tUVlKLaegY2RK<8~<fG1pE;#HZN`B%Q*5Ketm|j+b@jSdR
z_{tZixkwO?j5Y{8SkKdQXCiZKD^{=;5E9{ywwYRB*woBKhdkg)IxK=GjZ4ul&>UU5
zf5F6)I@CtCk@>PghMt`7i62Ax=?88ZG#|N*j}4tso2L^itCE56VSZ{;DMYp1uS4{~
zhhQn8&F*X8#~hThL;rm*n9#54G%)EG)=ZQqV^U>o>(A>LKPE+s(v*m)zZl&-s{j(K
z$1!m5C0jI_g}b8_$<%g1lJ!oJl>C$-`?7ktpGw5YvsY3?Xu$;b!qRzA`KS)`g^r=)
z@gcPTwFH)BhhwALMb>e<7Z`mEWBrQub1&41Ku^b8@W`6Qh%dRoT-Ta}u9MT@#|%yS
zebHUGEF{FZr%Mp8T5V#o=r-rfjXL<Kbsuu9=W(6)<S?Ri7#2@bCNp?gMAu~i2Sg`f
z=CNXIWi~+giq&|!={%Gt9S7O{SMkOAKOkuE1}gi7iRVmd(or77oV{Yjp1VW27pGr=
z&HaN2c3RZ$-7KiTVGr3G2N`ZzG}9swg>Ok2o_Q-nt*dypya{<QWz-o4W=WFvHQ^wi
zoQ;i!19-0aDB`y|*bt)xfyst&{kA)_8fvfuvIre>-@>FV`!M{ICOi9*BwaA@114TG
z$MjFqbWck=n|?S9PHY&#R8a}CBP<d4pJ@<#CCj;M1!bs8?N<(;nl!ce7>Z|3#(;yj
z9t@}5=ae)jAyrBNy!#HeRWhg@m;rcLn0k)N(d@I%46{y&O1xy5k(EI(F0={xwuVB~
z&T$BwycnikS0>y1{-X4AK@5$K!L55skSt5b(JkZbwmpp?oKpg$nJhfsTFB;WN)b69
zWgz21%m&jUPD!dF%}jg(ojjf6!j&Oxv%@2ZJEu;!EAiyM_b=F5r)sds6r|Z~5Ss7&
z&S}ypL{(uv;=RKRerMH!UAPRW*(uN;u6JQkm=@8q8AJ;kL&$GLi0M24dyi?7q3LqC
zu_**=k7P1GCdYv0XgRp;wug@{W#E}R3oeb_#3Pw~@KT5OT@?&zteP<H8fu21)uGJI
zC?8N0xPhminP9v9bPyK1jB($aP<o()^^>!~gv{&cZNC^o=E=~CYxj`ZuSmA@@Zr)W
z9nf<}nFtgKP;T#6CNOIfQ@c%qG&C1OWo9)@-5ZVFPrVp}X(zxiT$~671wf2X1J-t^
z(N~Ne@45M?+;$MHv;@e$^eH&e_!{c1sR9XC4Z8IJAKdQu!!2c*5SAxSB_g(9bFU5b
zKh=hf;`+EN_cgwX)1e_dmB<9c_wb}ggQN*Jpp?ZhjP#d)&-7VrfY%|eyZtz}OnL*t
zP2;dY%oo!Jf8m`q*I>GSG1G8r8Om|w$*Yobc8R(&Z6#Tdl&wSr{-!~-F+W%*$kM&%
z@=#85CYpmT4N&Vxx0GwRV#Nx0Z^2JCwQa@&=S8Til>+QkFMvJ$x-`Q<57yYWB35Uk
z;gVK%Taz?tGpS+!%JB4<68@Makp|CF;^FS+H_YM@ee$g<53{aGku&Pjbjnn1GNfF=
zJy|!x4LlTs^@0?ddk<sO>L0k_X(7|(t4SgS2H}sPHB%ERPM&=J#bIY_(AR-cnBg-U
zU;4_?>zP;KeRvA&2~{MAdI`qJW^i}NY=Uf~aXh7Y16nOsfR{3G%7Ub6_)<Ph-Srx_
zBpd<t>xFm(^Uxva5N7mF#=SrOpxFyP*6E8qYf@2)>3<(`ll1~Xb|io?ACaTUC-tbt
ziU+Xx@bP()oM(^_;RvuplUmkIWkca&KFKm-f*T@GuKYIS$7<2aEIsHk^JVPkE`dn*
z37moGTiEO^3nrhxp{d{j(DaidsXq>KEOOK5)yxjS^2l}!i5vpCGdJ;%U@OW@PDi)H
zDx~YHJ~dC`GBriFnPM>!;^174ckJh5)ZEuF|AZj<$nfk#_TqHupB7MBWXgFbN}y0l
zgyfvM0wU2N+`*CqxM8mfM{{W_Q@H#qT2ToSR^J59HDBQQ#tvL-@dg}2pRgzDv`AKp
z7M*kQC-3|`${aKGg&QABVQr@nDgQ1+>$ZBMQbjc^s<mb9vUQ1_x&lmjeGL}!eSvM^
zAD}=-lIR(~#18^-pte+o`cJfCUhmYSPnXI;+RvAe6EDk^Lmg7n7>aMI%J8d~C|%{H
zOjM_-5=)&R7|f9&1p@l`?x-R)2#Lb}Zc)<Y#!qaE#h_VY7v33{oEPa+hEHW5fE^PI
z?`3#v>Z>k!D(%f|g168!R{?}h`taoBHR#5Z?{^&)pfgWqGiSG+!DiJU?kl23UzF+K
z+S4g)%&+;l{oGMd3^&J&00DT#ld(Rmn#Y|Wr%zfN*CCnpw!nP`Z$8}d=jmT$=<*NG
z*|D7hRINV&I%3?xIM;~DfBBmeO_Sl|c`f>POBxDozXKBz6>-fo38K{~hUd;GlGzz7
zPUy=)Y&Z{=w|`>ehB)@}$0}TY&I;Z%=n%`r!Zc_=fGivP3A56L>8ITSbZqApFj&*e
z3Z>3~3AWYv_1J92TtJ!5UzCWh30kz^p&jI}l!q(QfjGMFA+-I~p>JIJ;nR!?$Qm4k
zD^pZx#4MgJO70WX`eot5P6M!@A{0YKXy-H?G|A%Wkv7c&k*~sJeMK1<uUigd0eocc
z21DAnP=)MUau>`Jt#QQrJo+a8KoN@!)YZC-=S+j3KI=F8ef%>l-oQr!PRSEmbrHXH
zUO^50K-BzghAVb_#MjFt@bR@zIC@r}nm)Y>33Fqh@R&9p29`-_FlSV5gs?#itJvl_
zn)JbFAhzg~f%)dM1^kk^SbI7fe@!2Q_ha>tv0s_&^K)dN{tKMB!$nUXULVTmA^3g1
zc<G=taad)FSN?8>2!V1iUTi?-W<7-J0~FSu`VAld7Jz<;0Z3b^Q|V2juxEQLtWm6G
zj)z2|ll=ocS||_G1HVF}%rN>^YJ+Kf4XgJmk<ncy1(o*%$s{p%?D9;5!xMsGwRr@i
zT4oQQ#%?m&kC!v-j8j-4_8w-OiDc!rJcVe%9oT4Ak7<<}r0v@uHu<O#n>Tqs#9I6T
zkFi5oQ<?-HxqJ}I=`C=QG{?cIQ{ah=3_h}YkFkcU!MU{zqYhla-&ZuKfOt829NB^)
zUOhPHxB&AlOMvX07>Vw)OW?sgNgARP0G4Y7h+>HzQI$Q14m@3{ap_)+YB>h{{VVXM
zlNCC+36r91f9xAKfL-%jV1B6?&Y`NrK~I|m?r%WJXZi49#2gbUE`we)A61#9Lzg~4
zY#J{GWzICtjl>JEa_T|6wrdg=J>#Ryv{xuHZIm^T_yjwTZ-j808?eIX20Hx-ge_=5
zvgX}ne1h}}Ja%nI9jg*(($S{E`}g6glaf?9|32qfU<!Emg@Si+Dkf~-47U&bgg-GI
z=&mGBCFhM}{JjZyc%LYd3Vww{)3eZ{TAmyz;G<v1rRaXHB<%~4A@9d#g3#-q=(1ao
zdVlN#rIpQ0w{ae$x0;7hmDZxk<4xf6QUEHqIlyBPeY&gjJwB3+c&HU6K^E-y!FxQt
zM1;q4Y~r0)OYMSD-dLRW<g3xRD_i0C^eX1hv@TRlihu)d>ww*Vo2_@zp!>7h;q=sN
zSaeB*J`qpFX-&0wYhD!os@a1p1r><3u^1`6X$f8j#mEJ*77V|d$6cah%lO#jF|WN^
zU_efb^7|EV%EDCPXs7{=6Vs-xZM?O9Y$t5pVgmg6m$61$j0#%>!S&6TI9A)g!M=Ca
zu%vn-Kw&#4y6YY!y)?#F?Otw(;w60S^czJb=CfzsWELD9qiA~bK8$YlL%Kr{`4@gg
zsYGd-*)L1|+q*DTLW<gMIRRcuM$C2I8t|&Rhsh>Wx$dg{sJ$eRwa@TpUfG>w>raS5
zm8Ak1k30y%jltmTu#nBMtwM=Yk({(*14!1|3k756naZzA@pZB%3e-k3zm58E_K{?K
zmZVMEpUc706|T6zd;$)OU&LMT3TLOtk@C17sAI$8tmq=-hU(FW?g?z2*JqqrsX{+B
zEMhjR^~3nR8fbRT#aBHp%+v#U9QW>g_IFhk_kn6PZ0PyHzV&a1GF^YnIb97w!&>xV
zKV{o5`Z2Y`X)qjQNS<C-AbVQX=&cVcVNF{AquEglWX(%B@Q|NQoBWIu@kEGf@V^F+
zTqCx4(F+*Tj>WyN<;m$oVXS7vZ&;l&7nd)LXR8=7QW2knk~*b~gNZubHARa?JrE^B
zujE+i&#{c0)GD~pE=G=QRHt#px@4ot68M(c%-&fw#2j=NC&rE=OyAZm@NLIa?q5D(
z@?-ch^zrz!_?!~<6$F6uhysngt4Px1M97WwRjgI76m6Il#pv`EFlVO*qR8Ynu&XS<
z#deZ({SH~|cofMv4#*M+7bN3d!!T!H12eh*H`B0BgZ$jP6>HC=!N{jySSBorXsAXN
zEG6jaK5Zthgy#$R`8+c}`Yw8Q&qufLLvYXhA{Zw4ad%qr_HIKi$lP7S)1}li>DDUf
zuvebc2Ha&MB!viTs7!-0Q*c?2D7g<ocp>ovt{<fMp|%~CeEx&gRgyI3&I~q0dM3sM
z4uiA2EyMUL;-Rz8(N3})^TzsM`}Q~xj!+;CJl~Z4dxc=;_i&g$Nt({!8^ec=V$_RX
zXCjlj@X7`q(krA(?wpe+_vfs~rcVkq(D@;|Y|s)G=iO%&`(C5a1U0&!l_8sLwaD=J
zQ(QhKL-%@oV@o=B;n8clq_|Lw#Lub4yf3PB;<R>%%vB_+%N5CL2R?EwR*3}h<mIiM
zL30I)#3;8&5!*MbkiSO!JfGNDlq@S``IR`(@?4qRPkjVls#37)JI}ABZwVMLvV^|R
zZ9IFeF>Y+tCk3+wNq}HFoL{pU4_<nLIvYi>A+{NcyMr;tN($_pzq0<)%YZbbGn?sE
z^xb(6q#6StL+C4{{VHM|*O+saelyHOXCW|hz6@Fg!MI$nmn~VbhIx~bgWpmU;fMHK
zSZyW9z2|lwZoW#NXEeBu*|=JPY_zxvUmTqwcV;JQcgW#(qa?hTxsd6V*Q2lW_=#iv
z8#v$ll-;__0z00B;gaFxf=T`(FoAy;`>6N=>@qZl$h9NTdh#o}@UX~ai`TGk!ZkG7
z-vi<91(?2Gl-{$L0moc?86W*{IQc;abcD31!CQ6G-!4FuTE3%x%6;51DHRsFX5-BM
zDs*Xzg1U(YB&c7Ujz`CU>bXeFbBkb0y;Ml%-|w(Ys~WQ^A3=BQXEdKyh2eE=+zZ~9
zfPdOyDE>2xow>*qEW2NG3{(_oScN7L_cr9dKC*%f4<t#3*j3m)LxlQ1>4xodRjJ)Z
zO{&=N77{yE>BnUP_{vd%6liy&EZ+~b*}NU?cC~`Z*I$gfe*~kaXg2pCZ~hE?y$Jhx
zcAJYv0<@)SHKbd90IQrrJe_a}*{-?lc{~RBYbLQJB78J}QGmv)q42&VA6RjB&enw!
zaBrp}>8jt%*`m9!pz8Y=9De#0ukq}zcE?1S3ZZ=bF-eki52?VmNDD0bbg1Cx&x;^;
zVlBGAOM>p2+weVG6r8<8h>b!GQ*<X2?T*Eu%CAV+^yV_2x7~<3vOJt?90(6=Zr}li
zAE|jS-YdPq4HA;3Otcofb@s%$>hIxLhYd&_evjo|S72Or6Fc6oPWP#>;1<rqB+Gwe
z>ar8;y(TZX80Z3bRw>f#NER6zZ8)0b2`45-;o0#tv{g|i7t;CY#YYPCr(z@;_uhb(
z(ihC9kY~*B^fxf2!+|?ELz}4iJ%I%agFz}pm-45}lYS*bwjrdJJvpfbobJV<{7VfI
zaN#MKG^o*)Kl4Ch(?ys%=7uuEdbC-n3yMM$FnY&Re6(MQPHA|GmfxMRL0y7gc^n7k
z+ht(ZXIc9Dl>|9HZzin2^91(1RHGIfSu`!nhog~_)Z~yD$=w@MV0Y>l#JAqSKdC(!
ze{C+@5zl4<S8T)+OU)o<ayQ6o3UQzF?2r8~THut^5jYX^3r2tHlJ)l&GNZ!kV6;9S
z1TK$)*)85a>v|L)Y_4V2dXAvB(n)+aYd_oLACIOu6M~yAfs2U>8PfMb|D|KF@>C95
zXn1ovdoN*&=N*jx69UsqR%1)A4j69#0b{3G$T8s~(pCM8=S^dLz9SenIhDXQl|nG=
z6eQ&ZL7Y~;54d%1B4&A>g0TzrY&y>$*r)p&r*}3m8*OHA<eI0zy_ti!Xq7hEVyj5g
zlDcux>rPzu=`iTHpTx8SLZoV~0XC>fQMo=ru#>1lokK4%4z8iRT^(+GRs~ZAglW^e
zARHN2C+(5mcx{CU{ctl3LpJKu7k5i>g|<5V5q1Sf!$nY`QHib-??#2)nb4l)1NXh9
z$v3$R*mrjaDr~#XU6YoDC%Q^-HV@ms{?0|1DNp+<Lh^Ub>cnRI5NvSzjFwM6;4{A8
z__8buAW^*FMnV#5-Yx*49qJg7polj{<;e*fK~f)f9oLL8tbbkwdc^zT`4UBVX;FsF
z^LkJs#}D$<<jLQm2h0dhMtFZ&lUVgs<E0`8*uGd5=h*H6d0R~y;2nq;O2sHkl*l8w
z2TXvU9$WNIf<Er*V;}Q9V=tD5aFquAK<z|19BMnxo;|2X@1)LS^+prWOvVBPc<*j(
zU<v0f642%P4+u?Ep^DS^sg6P!+oGpOiW8;Cx8*|IU_EmP7<+>{KW?#KvJ2RO%R(R$
z@(uZnM2U6z0hlzk6IFUe$&zX3ph)={=wv0}^BE7I`KtlFH2fN3PTht5mL=#cR?IRV
z4d@St3%E;Nm~`I=#O>#`h~PFgx=CD}KwUFz<@xn4SNMsBvdwUL<S(SS?*jR=o8iPL
zqTGZo6ySe>6$2NU79Q@dSt3a%wrxShR8cyp(hS!=DA78>^T6fifuOxQ;j?gns#(Is
zX6z(Fmpl!=-iDJt+k@IFSyF+ipqryc)K9fC@-9lmetikl9EoP~>;upzZYt;V8*ebX
z#Iv;)6wK4lUJUJz-r{HD21YBJXP-0;XZJYQL-A5+>JuwORfcz?vDht4o#%n3$K|Qq
z7e1!iaWC_hXHPmN?SP@CHDF(CKyPbpM%zCIWMytJyP&}ozFcdCu!~pm*>_L4BKw|A
z-+KpqJ~6OAxdbP!@?-8@9szT&6}ZuT9I`(ig19P4Qrw;d7ag_fLJdXQzE_!8T3&@7
zzaGeZ&C@-n&Bysu9hm%_M_}mR3za<o-<P*yL0Xp&;+oT053LqR{BWIFa7%|aoL8dT
zhC47ZPmz9&-3sa-qp@Pq49=f{d<+Xa53aKnaRKvuo_5(hEV$qY(qW63XK_ie_do^1
z`A`dezI^n~{!Uc?uni1z`N&#+4OS%E71sntGao>QtV-i>0_ylE-pR)JKehPH?-qVp
z@&=N0Lg2_TYYe&n0NKTH5H>dvdanne`g%cnc;OxBJ$DF%i$8(DI&V09eH7cXdAha*
zk8#d>AzEqmg(qJckoN|2LC#a3E<bo5H?65bADJhN$z(hDQW=k7vnN5$_8@2-`~=ko
z{LK6Z{`0=K=o87*)kwW<aD|X3H*5B1UMx@xSSd&1<)Z-@6VeU_ZqlTwM2gN@V2v*=
z&Di+_V43g*bn}W~y@nqGe^w#W_(GFbdfUVKDetgh!9{dlra{6d&V%Qh-{JPw=U8KJ
z2t5--sgH&p>EY>4os(mj-hl?3Gc1R_)v8qUQ7NjI{=t;<`%ocdfN3%Q4F_MuVg-Cc
zBa=uxZT%erm#=}Cg+3Tkoq#!mcQBz;j>OmwW8g9uv>O&Bemh%mqlq?q*77Q99F!u{
zL~U@AhDpJ5zcHR+e>y9Yeh%jZsSs6r0dllp6gJKYL|>7+%(g%gB2ptk(#LP1s_!dM
zO$f!sd_135b91(LM;BvumLK+>;>lGrxs0Rv9=6gljy>d{L|!;vfJ&Qq+`G*b>V0&m
z)6x;(7Dq88ZYD!kd%*C5Czvwcg+-~_WXh!gnArcC8Jar+&p?neP7T<;r3_$g6XfaI
zFdO^I(WbNvmK>_(%%5k4b>_MFV7el-UXZ3H$9yq<@>Ou(ElNUb<l$o0JE#~DL4V;J
zaBa;l>|1n+JwG)H8WWnCWutYdVJ1Yzw+WNZy-&EgI>VTHPKQ1)35THFD{%YlD)w-M
zK3#g9=WF_2jr5-#grFDh@XcU4qj-X!{#Cq=sm}mqgNotPcZBEN@!ax`Ga$6l6<4Vh
zb2VD@Xq$2{$~1;Czt8xyiZ`ahe3M%^t0faV`Gugz`6J#d_GfSKa8}EwOuYC#1P0Ri
zF!KCbsH&7F0grmv7<vM_XP;tiy$(UmeV!^yZx568=qLp8ix3_ESh#N`Lqe-^VWm7j
z^>FgR;gJULdmG0J`zX=0lu%3`kA!zpJRCO|gFDRXaM*GMzT@|WhBwLVw74u-6Y~|@
z2d2PafH3jjU5si`RXiVZJ`$oW4}8*DXuz{E9o#BTm+^~|<(mSrgn59EkJLjbPoMRo
zM}j&?#o?V}GDLb^JI@b07c4i(62A|!Fn!-LTva1V9}DDw<)$9I%$EjdBYM%J^$7cW
z#EBjF>W??$!(rvCepW;L1qO`E(AgeB<b<6N*_3$!S4ex~0h>e|c%(^<&#eW{?P|8p
z+ki-C7?PW94bW7)7fqDnV46xdV^Vhq9{%Q|kH1x;@D+cMdDww7<YS=Zlq?Ai-44sE
zSHYW?SJ{Wp9-y409{QW9lY2}1+4^nwVa8W3Ol+^f<TeVcCU@Yxm?7-8eE`YPLezJf
zAo+4R6%}&Ci1<B4ax&~1>KHde<hnE{+RU?co8&^=YJI$ID!`O^E2B?m1Ws^{gYdJj
zz(!t;cEpR(M6WiGR=tmFd9g+hSG|Na*%dg=QlI_Yl7!6fF=qI2GaDu+L<e%LV5?0M
zMEgvEx07<QJ6;46*%;=hf*DBUeaGXIrD)Q@C~R!xhrGAF?2|LSsC~1Sb$Ip+i;w7`
zXPpKS7CHsY-~+au7jw3~Nr@C~G$7-pL(u#(8$Y$Zhj+{WGS+o~+P!Vqw0`0|mD2T?
zuBA)y>jik}xB%vElp}*<JXvfO9}!65r?Z0P>F~WE5PVcRZ_ZSCdTc-iG(L%=hLZ$s
z$}fh2Fk8m^%V+pnI2rZAf8k{hSBN%KA?`jR<g14O5n-mYl0OybUx`56VkSjj#i^5O
z4Pk1SBgI?mYjIaw5L0<VoF31M!o}wgz*z4T@JkzJXIt`-f_z^bVO8j~yU)Ns^bKQK
ze+TBCx(D(zMX@-o41-RqlXx2oIP_#OY@g_fA+_<mcoPwpHR%N3dPAxywhG84P3mt@
z3qw*r;Px3895ULAXPe4!Y1a@O(tilwY_f4=y&5F>NfJ8k*xXe&q^Nve1QzRVo|k=i
z02?QpGha_s;D@C9vm~N?uzX=Uepyw=HZ+uj=7GEL$@?ur*%<fk7v#mgw8Eg*Q?|O7
zXQwev1EsY!FfK1ePYzTf%X`ONKj$+Obkg8Ul>x=lX5>tEWvpl9Fmnn{Vzp2e+ANfY
z3!()OCBchjh{*wwOG>2c(@HpBYt78Ed<0_7(>SZW{lLi29qYGM!a9euc+hGFm+wF_
z4vN3WWvaEX)%zsNDR>L|d{4pPU<;m}8x9FU0nC(2Mfz}5fUan+!6Vyh@p$Nc2snG6
z%|0$g!rouQp@&5n#`S}Da-!7pV8Gn72MQSTOS4#G$z||y@m|!5H9}i+Wg<S6CzA;-
z1?ll}PP#=n<3Ce`cD<FN%O*QBPo(5|@?R~Nem{!ocY=XE8;6;XAHYn3Z}3`TC0m$(
zA9klI(P#n0y(+Sl(NLftC&_TF7QKNTGX%)CySo^x3~@4M=ZKNYzgf?4-nqAxA0^i3
zGER%E*;xHTP;38<26Ftw<hleMo_GL`FR;V~?a{dL8UuH``kC^3d^C1z0xYb*#OBSF
zq!zQ%VDoZy>~GLSowO8Gkn#n4>kt+<Okmo1akWc$m`Q%|JzRWLg(z3))An=vbiZaI
zZj>8hBGmK2;E@W=*fbj#4FrK(#7#J5D#;!y_Qf3`N0IXMY?WJv*=+e+tmyYM(0s6p
zE$n^(T9rD)CI2(mVrCLlncv1kkEh|pnmOQj!ww%!xr3|E>(CL`Mo52S171A8@Xh9r
zaB+SlaE}Z#LS1=`eX|N>mP(*TB*2+vcOiAxE}XPXo6I-3%$D)a8eejmS;+HHI2&jT
zPN7Ti;w@?FpA(9;5x-bfdp{6!d50#_e<4*U4NAU8V^E(ATKNPr6D7sT{BcPVe_V#F
zQXWVBX`aY8`V}sPCgZvZwrt(06HsW9h!<*d@xXCeVlXclrBVgpr12=q3T|g6BqhM`
z<9e79twT0BEWktE286sh2sb44h?QsoE}5oGQcSF%+Kd-F@=J;MXgq;)?L8>4NuAop
z?#9y6@8Gfa6PhISvx0|z!=2sYtk4xLxculcT(c2p=D6(TJ}PfStz$`$w`MvnR9T6W
zc`-fi{a0YSsR4;xnSoa;^~hMR9A<vhr3QC1;liybsL7fI+iSnTx~OOPlZSn$tjz^|
zMM2UrLzI5})WUJNRfJi|r*NcC4^O+M<K#DrG&VwxUUuZ+MP)(eVOlyI_pt#hkvy>Z
zG9N1<{$hi(2pNicg3~p{hyvup*o#qY&Y2Ibj})m~zZo>=uS3n=KXC9|0_Hx^q1Ipc
z@!ZjLc8P*Axp>hL_Bm-2&3Zwc?{2`In-K%4r&}45<-rhXZwx2T`G9jvKL|w~!nKK~
zA?Rim*sXkz@n=~0;PDQ_-iCry!94tUj)lsX1&rKYO>*mmIQ%uQgafyH8S4cOm?|Ph
zJqO>w*PRCRZJr?9_@f5*^K2o1%O}jfra*cfX5x&2SzPHmpIMVH$+$*jGeC|G&G?wa
zRl5BHrqSnUef0+_HY$*>HCv&-UY`o(xS>UUE(SfE#E5<pqPoTMpyqiC9(##U$DmY@
zu(5@A{Az4mtv;18+K=HYw?SqWFAm!bfIn3NHtaYAlSE?B-9mu6*a=e2MUxSDdUKs4
zez0BlHT-RT539D7GMwr{IMurXE;-+V6PkLom{F!H;ta{hb_KF9F^HM4c$69QSEVY#
z30S+V2t##t!vCS@Jp6k6+Bn|cd+%*$ufF$n$o|=zjEpEND<eflv<IT8v^3Gy_}<rv
zq*B_0rj&+4A`K!v_wyI@dR6y1=lYEI)v~vmb#v7xVm=knIQSQ(rpGWv3kuOSCW7^?
z8o=iv=h1$98rE)!g!n;W`nn?xFNCJ>mAlPBk~a!PXAbjvr!35_db*Mox#5UA3&qI2
z>W^5$-4Ue|1jy_;71%X)99Q?K)65CdBrK+iVGn50(^V>@Swo#_O?!cq%U+)RR097=
zWPy^^Cul$J$b4J46OyMYlI0f7OlR6@ESow5wlA>6V~<-v*IR^sT4_Xfr9XrVO>sGf
zZcbqI&=m{fCt~OJIs6H=E+|rv$n+JO!qLb|X3)+8zuiN2fl(tGE84=gJ>JZNnFauZ
zO~_3Zq}f)0)C-7HquJq1^`3Z4`rU=`6%-cTHm3LH2!ZJ5Ay}vRj&Hk3i}f$b;$3C_
z;r`QxB&xU%SBWoXw6<J_cZO!LW~3hvx^+Q~&=)M?F0QE{N~fHDn&ZXg$8Q!H(T#82
zna~yEtp0k0JG%<-@`E>^mhu7=7M0?~*rT{xoX202nt`>(UFfbChc_HJ9slSb%s=u3
zj(kmJRtWWB!=_Fkt47f@(gVkL&V>$_HMp-vg8q4-M1NGqv&x+Q7}|G&Ikt5OQqC$u
z;;1BeD#&oXp-^<JJI7_-4M_YM4R%F!D0C@Wuw%mrK?<wb97`SAmYN4=&3>}ET(-5T
zumO70r!dkBHSo8z8VQ}z26Og?Vq~=p{mOZFuarbd#q34UFQiR{-|J#V*a)0Z$OOgs
zmuPAtL&b-;Gs5nI3%uldV2c@LP3ND1kgdX0Y33Dn&jC05??^2E!$M5ryk(V~AGrU7
zF)8iw2bGC3^g>Jldu&-2uIP`!sF%^~>g~!T_S9zn(Gdmsx<itlu9l{^V*Y|*nGu!X
zdgKuvb{IA(9-C`iV0WB2wfeV)cdD7o;UAfd+oHo@@#Io`I7NZ{2;au&e>w|7Eek>H
zju`D~P@*xiD!9L1mbQOxg82)UgL2bv7~ZEy@?vXY6ElV)PW@<;{F7NzAWoys2trYc
zubKY@V{(tX)6**RF{JK4ytP1wHcUSW#Z^2|x?sbiTOACZ{tT(-55ea;U7Q>5%eE){
zWj1lWht6ZsJmG2<W?YFw?>}19`hX<;USh=#MxA6n+N;pXZF^byx!*A*X)ZhWt})$o
zq#Ymc$zU#hjRc`B3n6J+IlH>g6N6^HfKi?cbCAoezrXbsms;eaYNIk;8Qcg*%e3es
z&+GW_?^pajSDG5gzQKM2LHeO-J`>SD&KMt)q;32Jv<y{cT6KzW^AlS<<7P~k)o77%
z-}C7FmGd@A49K$?+t^j#%%C}Fj0u(*1eb~?ymC>3ylD8qF5x<D&uKSW_uha@OGQY6
zo;Wpa6~;|T#<cr9cfPqr@*V0*V9VM;G|@FAKRHeQPjD4%6wbj0j<Haex(tUWY~nm?
zGuV5h2@_Sfusf%I#{#YcaCC1wH<u;inge67PdJagCt8T}uCBo)ae8zvTZysDrKmmk
zJw=Fl!t+!$V!o*j*Kf`T$Il{EAz7asawx!)52|2t)py*m`UAe*twt@6<X}s4F#hZ<
zLcR40G^9`jI(zuo8}kPJW|ZKsab3ne>l?mKm;|Hdg5>Y7C9vv(J3bp1CkjGc@HxE`
zclSu6i_$X~slA8}_j5pENdwM}&?oQI<=~U40IGULfZVJ1_*Rw${V9`~;=X1^eWxJJ
zdtid2$4218{*SowNk2+n41_onDU$CdOH*>W+|1fbaLd4$K1vzF@A~?5+*}-w#=eAu
zd#A(qK3y8wDn#=J<FNXE3iEcc1bKAO1rqmkak_sZjOs+NEAHqMyAeN7m?%Z38CT)r
zBegvD`px{5#S-*cbub$kYX!obu2rI(3Wct>ajwx5{``vuL{?=ko0Jj`$G4@ROtB7K
zo0<z0ufX3>N7QjvppVyYMClGMjK8oC)U%vXzP}DtwRSRbl5NlhH*vygYl!^p#;Rp!
zq3YRnxKpVbr*E$WH`AHuJIfiYXA2Xxoh~rVS&!5#^WYVj?8J@r!8zhfWQa@he)#;~
zS(x>piU~KZ!PM{3_+r5f=r8<%f2VUlr)4vntX#%S)l7u+`?FxDrzV+K5zAMck&ee+
zh!TmvA0WoM6knAc$9)YSP;&oEtW(mVL28FEsyP5stcB^C=sd=I{0bVYC*ayrZ-mNP
zyp!^ktsmAVJ-0uAG3SHXIZD%ZJ9{)1u|%6ladwmaYV^Iq$F&=yVTQp6X6hzSoN`W@
zcHX>&Mb9`NwWkWwMEJ0OQ$5!kj=(ORU0_urMvu-mpr2j8fJFNkzP9(rhtK87rc1*}
zw)CKXlN<l5>Q~0c<UPxr@C8dgg9$SKK)}I><lY?x$Ai((yEPGRJ<ntZms>I$`%QRp
z6^Ge4Z4p9LQZb7s&SdeWsPpyvoX+wBcE4zVxbSwE8@!u2XkdkE%^#U=n`f}#{Q%b$
zQYDpN_36dY{aDRwM%(o!C^UWtWsfQ{;gW}QV!aN*YR+dKc_~K1I`r9n<BBxcFa&PT
zpNaM6nr1=X!X$X86Azl-#*>jDZ2phiczwDqO&n9FMp+SX-1jaTK7WK0Vx?&H=lyuZ
zvL1R()8W4|b=u~Y3SI|R;7Y?@w3}7H<fXplW~)c+oNIq!>24!(zWpy`>@}hvoL)26
zW)H#SOV1%?fjc-gNs&<7N_1c7ir=Pg<L_O15I(K=&dR>kAYG}UxMG?xJ$o<$duHE7
ztxH_rWJ^ElyQ@HS@-Aj*v;_ao7ox78Gnu}#(V(?ffcy}aBVUBu+2s{~vAF#*8Vr@g
zy_ig>GEpUV?z*JjI0|OT7BUJ4ow=UbZT@q;DBfAy#rR`IFgq_ziX7kU41EWr$X5?R
z@VL8!8B)<Fvo>pDW`jJTw_5Srq)3=t;)yZ5C`LJT3S4;n5$)ZZAzM{}UOXm=>3hRK
z_)HcrUU3n+g~*a)No6p(*cJ+8_CwGoQLqn_<1#Dr*o4;veI1SH6aQA2);$w)cHV&U
zCT;rmya+qUn&70RlHluFgVp(?kZ4;3Gqu9dAg&kBW{u&_uxu=lmZW35{h9P%t~}Lg
z?hqjM71kbZW+KkWk=@x*>^m!Gys)DJZ+?#ged8qlRTC{Le0UmqIT_H-5DQ3^UW)IZ
z%TimdZOqsyU-)mDGJRVjN7Bp(plHSbe(!Pwg~VrgB{CR-TIwMr@-h?|e}OakT#qLE
z4|_*F7&G{9aIx4~Tx7L|*<ijIJFiJlIGx0s+LVe9i9ShR>;N5EvQ+BePK=4`<ndJ`
z>7E6ubgQlqT)#d9`$wu_QK>SCtI;Plu0@dKHiA!A7MuC6Y{3JL{b+Q|hzLFkU^m_S
z#4kMc8wI&skbFmV_Jd!~V5<&yE;8AWBVh$iYzuGU0v+P<)&%D%t-ul-LKs?zYYo~k
zY9N@OEb|p|3;&__Ks8VHYCB4g?tq6uLiD!&zN`TCVA$|@GxR<VhN0zy7^t-dQbQC;
zy<r<XEFS?Q_7!8lu9DLb_>lIv5(h3B5PsY<^eE!Ps(UQV@vB9ZNO$mx>BZr7YOH$r
zRye8Ui#(qtn0X`zj+VbcZG$AdCV3vBCEhYqat-nIEGN8q={=4eSEPY(873Oa<IJNP
zD08d|M%ht(+H;S&pe{phsG4Hn94Vq2rov?cOHt#&1UMn-&RjohL@h!G@T!yo!S6%3
z>)Z|~c7Mz`cU$4E!MmX1S_hSP%E8w#g|XoBbN=T)L6R8a2}L8EmVF!TZ_0qmMIJjQ
z<ifn=vXbSYc{rb&#ZwBpFteZnyaMG(;*3Ukzp;o>8g_xF-I6o~wdoz6CjIv(l({}-
z1Q~}am~~f?8Y>IZ<*Axf^rk4?Jzx*Dhrgmmn*@Cy=mm~fbm)!b8m9Z5BH3qZ1{$9;
z@tnv#oN~_}1r6Q#n-8eatLih5m-`aNdoDBOJzVb5N{ifg>VtoIJ7K2t99X~TF8)3s
zL#$Q`k@zpmpvgd-4!x7ZKY#atO`Zn4x)g&qE(F55ZWmS)4Vj6JNjPiKKCCTOAeE74
zS=Tww_$q?U80j^Oz4mz>YEI5(7TGA!4DSeV%QMIIk6Jhki3+aQ`i>1o55Ty!9$yXV
z!#$r!m{nqpYlJG%Q_=^Cmkvq(?++}_J%$;vKjD~20)Eu`&DYq~&MrHe0pCw^os8o;
z<n_lVcvAlXJ6~}np6UJx!JJ=WfA%2e+*PJxRt(#-S_`_4|A7U|&$I2Z<)CC1j`o$D
ze#2#qW?$Y3nd7NAnf2oJ6n`}nSXl${+3#_NsTPSC(4@!vgdoye0p~s#gAH4Zh*PE@
z7&TPDhMFw2UQ&fyzPN+W%O<#c^(f@6Zf9*1M{wxwSvc=?9=jOwe;y8Y8`t4u109-H
zeFG&qzC`5nRHpRtd${ze9Q?R$VG-tG;B8&p>*dDOKWha0Oif}ZB1Y#7hCxwSEM&Uq
zkSAM(i9(+~jXQf1lm3kJW;#_M?=6>$Xi8xeKi1<OsZ*e8-Vd{bpTfc3cd&Y9JU-5F
z#?}~9sO*h^%@s-zb07-Wb3OaI!}DO}8!1$malvRIBVxHu6`Ev|QNh88YL(tZ4ciP{
zWW;I0*9!2xq!HJjn+n2i-%;7Y3I`UPf_Zsjq~T%+=S8Vgr?N=c{d*ZqxrO+)`voM}
zuVj5&lxcOT1gY%!%m_bmgxZVquu(V?-d#{316)SJcAYU59@3*~$r`MFbro(>6d{{4
zN104@DXRKW0#BThCF0FraG!tzc{SY`RSo3nsSVjI`}!yR3fCr@Ypd|IRvRX3v@kKE
z#f<%lS$I%SpI-MD2kE4rsL9u&vU}WF!%GU_^f4Yi2hw2GtW@mLl_9nFuE6x#7QEh)
z!0y%-reZyZ*i*)L_#Gn0AfR814j<E@P5Hri_q6~$I6n^pWkqqapD<5bc@-`w-^40E
zYQ~In+?z<anQ^+didnbiA*X%!;8_b}qA56^9dF))E|X%QG4~ZydPtm(a5;OC7h@2s
z6aZ&41<5z5OmIIMhke4$py_@UH^|ttM(x_T{E0lNeDn}!)d|r$&j0xGyMgKcb{t~F
z|KiU3;#B8g9{4Zk7<PV<U@IMuZAMMlwM~Y6HJ2iqRRUllcM)ra*WvCh^LSSl%93p+
z(YWd3FUBXq9k#zjkeDh;`rq{7Mgdcp_*;?uJ(_{vF88x-&nLr)Lu0rmIui9eR)A_H
z!ZDLUXxl1|mzM~Ga-}-QdsqfN7Pp~HI2TTSQlOad3*Xpw;fb>bG`my=T$T$^dnF}u
zMNF8=uiK8NJUZB|Cz5dVsw*^GhG5L#FbbArKqfc0#LIn!z=jqGJgQ1f6?I9T(0+dO
zqbH#CUl6PMK#02B`^|pbIK-MBn+&JVB!EeLA_n9LktQV-;`~k@`%aa!1>TxuxqCid
z7QTbAq9g3Vw^C&8(s@jv9UpAoN?;OSjhy&gff^gf;76Jw7$w-kWak#nbH0n?$6c7b
zIt4a$e>t3gRRl((Qy^i&kDRqJRru%2Huj2SHMq>mhoxLDO2Va*XPqO632lbd+$|V&
zCz&(Niw|S&OAjUt2I0@fYp}|*k-40G2L?4AnejjcQnh&zIt~9poy{(==+_<8+H(i4
zay*@_>Z6!+A%&Gp6eUx)$dLVJ#aORzimg{&m`?-$(BxYqES)=y{Sw+FZ$=}>V5z}@
zp)T-oufejqLiCciVdko2;b=+^@Tn`zOxGg2S_iP;^)d**--IQ%vN1^N9@JXt5Q#Sr
zxW2Oo{GI&|TiQ*a&0#ZB=aPjnnk<w=ijwTQRy04r!e0$3x+1@Xc|x`6#4I)H-g1cL
zU(SQ(T<)&)dkdRBZp3Q0BMbL#o`LVCkMkY4e6d{0F4Wh24-@YOp+HzTe!eD1U5r)m
z`bGsj5`PS}qJo&gZ7oddiRXCr38z7u=%HR@5kxPzg%YDY=2V>nE-~i(fSyd=>3I_{
zTw0C#8Vgf;Ll;e-@!{C=r(jT;%tn`bz<wPT-n&S`CBb$^O#COuQ7XZ+oR99(Y{VAv
zxa?ZkALgvQ3wA8;W#7EJfHBJ}pw^tv+Bsf?rgL#H@lF6VjwG|LE2e_7-#8mNE=HnX
zy+Cm}1sYMCiZWNlsOH%Wh{&A9E?zLgF$cq-=F~VYPg#YV1{6{BtP-_z^x!|#mmmQq
zL$F2Lkgn7DfDeKlacpzj!k1cxxLZb!Fw@_oZrWcM&5y(n_X{DQXcZQ4XXGkN1A2w?
z9~(d1hd+x`AeU#srmEfnV>M2<7F4ID#YSZI9|bbu$A7#E{q5*EU5V@;*CUT_|Kz2u
zk|7yBHn2SHIWKO}V^rQ9hmZY4X~APzYO#JQ1UL`k>bF|tqmvXh(p!i&+ss*|i_h_w
z$4Pj4hRb0MzJbFZZ{X807Cp>(tP}eHI_rgq*Nb>Iew_||zO$3<{-O`pRvlqYK9sP7
zA3E5`sebHil}X^4t`D;YRp{X;H~hzS%f^RNz<tM8s4%F7k83(XcHcA1x%vqA+6`m7
zo(hpnybV*{|H8xEy?JHv8TOshOnk&KFnV)7qY>A;DN2&0OLCORVOED(`lA+B6w5KO
z7kqes*M5bSy?ppP`8))s36go)Cd>xCFK~Fj0~C&_lZ~}1B=X1#_-|B=*qv(Sy4=4|
z$Mg<k|IC+N5xO2}N;`Y-iV>NE`Xs>Dhs*kHLhX`JcF<-Y=chEnj*I%FE!o>_BtV14
zm?#q0KhmTm<qqguS;3KWYruS$00}LyHama2ft@5)1_p6DDD5!{{}rsj;9ddpqHGv<
zp$ew~zk=#qMf!7QEw8^I9Wy<Q>D#w5^u3iR{C9!}1LxH-bg>rsWuZ(Kp6I}w>8|`q
zffE>MT>-M{v=)u(cn?p{N5R>RV)T;D2-YZ1zy*W%k>5a3dhR?%cBK!)vL4{8{Tky$
zzhLx&Ly+dC!D`xmhl`UFSk)v2I>kbP8fL$OQPHEQdRUGw2-G28@+HW_?iB3(qe$+Y
z)~B+M`Y`8-6yp<r2O4AqS;hT=L|rfzLk*Rwms~g7we~OAMDJzKrK{0Y*$I4haW7g;
zUx#&5joIy89sG$W_2>tiUI>pl&Im3$j0KGX#KJcfOU@0$GMQ>#-gR#nD>a}`m`Eso
zQwH}x*P>NjBNMq=il$dIL$7ES+L;N{o?$8S`a%(NGu8tq$L)bniVGmpaw0fIu&^T3
zfKJ+|Oe{8pv#+kku)Wu_nTf*P?B;^&{N4%$wqif`PBFO#^Ha9~$D0D(%onisK@q#J
z<~`W0$bkDqnOrcBXV1TLhI*L~%$fbSL66IdTs_gt2*)Jhxq)Kl)T~kn*&|Gzz8-;B
zTz8_@p_~8hL?<o@tzbJ9p2Nd!Zm8C#LZW3VuqTC^G56izMH`DzJBK)!FeC-KbpRFX
zOE}J*Iz9IBDNG3zg82AAta0o@am5Vojp+mR(aYKBkT8C4RSWj%*0Tq}07BLY5W63y
z5Pf3?d=nXECdpUg(wV{RU4LPyh|(eRJU63?(iB{FRfl$1c7pK`pwNa@D3NMJ-|QM@
z?+yKjk2&tD>y>6W(w_u7Rqujf$TZer+j7Q_V^{rS)8NpghfwMzLbNX!(~<%)YQkR#
zs`aWQpJQk&v&{j$ugl<I=}Xw&SO8VKxqgp)6Br38lcaMuaE0<)v<*;XW_4?lTNP#O
zgSZWl%=yq~dc}!|e=Y2tZ4J}yINnBmJ@{W~1OdZDwr`0P^mOZi0WZ(&b)O{jeWDQ6
z<9)}2$0Q)y_c~r|Qo(7Mf=Uy)_rSAGzN_t57+N8M?U!_^?>A{W>z5V$nCHQ;g58W@
z-w-b3vgPtF(d@HV@vvfdH7uDWMqYk4XL>G~<9Ue#IWwaaiME6^RczmmyZl7xDTli-
zvMwEe^1OhT6$-n;^vH~Dv3T;-0}ve&BeO-*;nlP&xUbF!vdsD*n&-#3-2cXCxp-h~
zu|D~GCJR<@IV7Kx@?_S7myFVZ=NPUdLKd8P&F1V{02k(b14xsjh933s_-+I*_SkM<
zJRf7;<wx+(Bo_GfzJLqA;C7K1p0JT78<dg`uK9Z8{H+I|{-P9Yp0JGU)~)FHvKreA
zicsU@vz*<3Mala$=FmGLMW>f_1Cx-3ot!^?sJ#j97(Il^g0Gny91Bo)_$svMDnqnE
z4!oSjhsqCPRH}^Y{kPpjzfmnLcWlK~GA~(08*y5{WEg(E(_`WnxZv}%qGWh~9d0h1
z##DSNX6A8xiRiUiFnwhZ_S^OHt<Ol%05%_f{#lRv=7ce~vrOUIS8?db?!(1`A_P6u
zsN&X1+<ljYr8~dE5neBD(FkVfBnDjsZBazI3^k_xV&yLMVZ!xvs5Ef^1@5kz9-0PA
zJ<GA?b_SZ>Ol7<jU-OfcB+$-8o;dxwix=!Q@gdi5>FUV=QO`mQ>(M5Ygxp}Wkql+q
zY#}xx9);Yb=n3`)^T~E09$Zz0VjNe<XGns6A6NzU`xMAo*od|3T2XTX$Ml`n3r?rs
z!d(xJ2XKE8xUA8qySV=6wD@y)u2qJZ%lD(Zyb*P27R0!c5BMwL34bV68jobG19nY0
z^2Yzbyq;HZ<d`)5;Uz>XtpA}|#vl}J*CYRC{e(c3OgweN8<*iBJSZwm1H}f+UKHF$
z#e;@K{!KQvaO}j+ogYE)84p_2AEC<$8G7_o3rL=iz{M8?=`tpnv35{e=yz}cACJCf
zSczM>DsVQ$zBI${W9gV6*oQZMJcK6|Us2CPl@4U2^A0VQB9mhi@C?`Ey|XP2rf@tY
zbvJ#wJEDvIGb4#NC;0{}RgQ*q4?(irzL3e9*os?2SMkgK#Xy7uH|x&ixO)~0@#l?)
z;C51y$@KGL-#s)Si}wFU`Tf1@bYU&{tR+t-PnIBZPg~&Ipdcx1R-hYOB(h^xrs2oG
z4RE9UByUadUJP8H1Jx-ylo>E2Gn6Odt-bdkYri7$oPROq&y~op@)K<PJ_)K5%=K8Z
z7Bh2nm8ew4Lp&Gw4svJ2WAbGsVrCHlX@b=_Q<7sxHK~$Xn^-I(GGyqDDy$CIWggGB
zMvWjz>T4-M8{+FA%Re`}Vs#}vZR|mH|1cCh_8bGjiM>1F2!Hyu2)LtSK=hhEGP~wQ
zvpw56e6i1RGf^vbqOS52nxwrM-F@?M(pMIu>$z_HoCi?-{t1}QH6-nvX7KHh2pr7b
zh22eDAHcx^E#^Dm##a*LCT!sM4v3Q8EK|%g<g?F=%Ha4}H|PjfBr@i*#CTo|Hca}!
zW-cCFm~|*O2e?ea>AHEi;t^%sg1(!5oU2KqO_AMjKnaYW8qo<0OxUi&($uJm%kVda
z^7NWj$?0SHOwk=>=v#3HUhX!e3v&jcBT|eyPFM#1V&!NRMZj}gH41*vrE`C9OrZt8
z;j@E3Ji0Rk^Y3e7%yR?oZBU5UY=AwY@r<3$gtNzQq$BTxEIFy5fteXjQ0H+UUPS5B
zFSeR+K+K3JiiYA+oQ?sr-r;a@4^HxQfs5N6aIM)3*uJp_##ctb>DQbVWp@MoGV^h|
zxdQnZwGr<Ki;>5_IF|VgQTlRj9Yl_;hKog4*`)?LMB1|yMxAD0euX=$n2>|90r{Zv
zL7s~5c0kPw^5jLA0Lh9{q4SPPl6bL?;Ii-*|N4S>*7?3O!+UxPZtn<Z73`$w(gaoN
zV7?Ah{x-1#yC1SEkD8lRgsj3!6eb%Di*QM^Fr94d3<_NTV~)ZOIQ>waEJ>{d3!Qmz
z!(k47+G_yX1stn@V>AD*Q^tYuK=jHQfO9LA$=<z2cw%!f&OaRsrkY!!GgzLiGcSdy
zciZ4u)mr#q@(oLeSnQB}4JRwS(LGU?ik^?>x4z4OyT3c2qq-5ZeD3gn#k^tvdl=4J
zn4QMjB<FHWRRvZ_IfQ*I^9f>ftDt}KE0}u!AsC<aK=qg*%$DTtzKJ^IYuPDCn>m0y
zM|&)vR|nC0Ms(lGZtUbTQKl2UnB1gxP>g+mXO2JTC(f})t>neb&mJ}Ey5<#r(euVD
zhW|K@zbN^1PMMWUK8?+lhP2^u6*7_`%xf+`9Jy15>^5@eM_*Q;%hwx&xp+P7v-^!F
zdYWK7GKI<Im;vv8%7DRDAyOfA2&=Sfvwdt`pgUKK$afjiwKxV$WFv%FKLzjA;b_U}
zp}%j+lahJHFqz{(hN!5}{|>%nZq{u9su&2J95-;+#3kSnv=$9Y8*m`ri?JNyvcjXP
zxZkuGtoAlA55G@<orRn~VU)r@nxISMHX}^JllZs04x6(9ZuoA&TC*Ae<zX0z*2a;x
zr(mxwN;)#iV9Ad*T(h$UyskaO)dj`y-k$qj7tX*HBjNDD&J|WmacA;-JzBE!A&BNG
zkh+D<d{9h-vG<P1vk=ECcT+I(t_JxzmIfiidzgq<HB3;G25DR62nk<r!Scl%&t`u%
z9G$t9c_F$8_NM%1X3YN1KJ-%~vK^u{ZQdhrNL-4WO#PU>LZ47Q*_&~@VuDZW+_BfX
z5H^H#;Q7#TDBaWqL!lq}*E@B|**?N(G|9u^_^-IV?j*?WX=Wpee?iTe5G;v}$~i5}
z^?Aa);21rEhWAP_)JGA|9I=K9?_jL?pihQB&Sw6t4aIUH4WgPDj|012;U|6Wtuq*i
z!@=fobwvU=mT|gx)?;XVs7c=QWvTD3B6w!C2=2NH5WQRdd=E~Har>-Krt3I^zmg`E
z3-p8EdO2wC>jgT01o$sU?t`YU64=&i!4x+^__9Wb)=UwkI{9%B+^oiUDslSUPZqCj
zRHT!I|6y7{DSu0@Fts#PB)c**aL%|a{aC9;R+&vhsfWfK<JJpHevRYwW-0PL;TQ^v
z8q$Ya;>2)?8vXNK2u*`sAyDojG+gI{p`#Nzr&)sM+Z;B6Xwess-eL3B27cL!Y}{cM
z1KlxOc=l7{aFXkErt?w<?y_AC(Ok!{&axDLB+9_$#pgh~JOu9eeqi-e8d;mndZhP@
z(I<Hv7hhfk2cDNR8B>2TTlM|G`b$1A(l=55l?%_~@)#?b?gzuu?}FJ38~6;~F!|kK
z_$&Mbvj-JO?`Ls(P3H!*oQi~V2UcOB|3fTX*@WgRdojg!4JxS!(6jdE(dCE*?As|s
z4*&6ECKrY=LTv@;mu1B)Jg-JxJ1oJV?=HAQ#)vEy;`+nu{84jKKOQOVWH<dM0d4Uw
zaP@-+%xx`8xMX&NZSplBGYl1})E!IAC@#c=Dbuior$cn_8&ZM)67c%=YIOQ}94!x7
z;<P1sc(_NLw9i|{KeRg#Dy`kY)moLzJHUAs;S#vC><XOS)(rWRm4OESf`U^K%qF9+
zc<DzNYY?GM?@alMzoX(Y+T$OD9ZEy#HjW9l(wI8fDnnyz5$I>HgwVcD{CM1%-OGK?
z8glBiGV22bXNTk0OImQib{~G>m~O>x;$(BJAz3D&!Y1gnfScuC_@btdmT??EbCNb$
z7uo}x8XvL+Y3U&HT8^H{)g;a>nOvtshg!My!flC>oP<wz(YRWP4Ckwp5}SwUt{(zD
z9+Fg}F9SWd2XLIwCN78X1GupaJ|9?(%@V5Yd=YWHx#A04T%bVDI9o!jk1;*!oy2_K
zk<1)1(;!pBdLjCw8sVS&g3??r`-16oNE9(3-?`3yx5qf<Hfxi4C`5Np?Bng~sD`j&
z7bbbYm6@q`2}&d1q22laVEKJP;(16IB?34N`%5Ks709FZSp|AKu^intNK)s+-!mn5
z+Okdp;_P{94G#o-Ku}naB<y#C9HB!vbKeP+Sau#$N;7cut|gKlC8{8xPTuYoqZ@nt
z@NTXib6ttWk6~SSlSxELS62{mYBCEP3&NLzt`L<xow3ze1uAn4i6fWAu+o?V2dm{s
zaB3E;`YA@HY}2B}ZRgO}_#vCe?IoG$QVR~7A7%3nZG>x<*Ky}&1G1S{gP-~(iA<m|
zB(^Vv;e|;kIYXS9FpKa|YcwzFfipW|rGuIIt(fv`C$r~X5NyA$N8-1-VYI#^d7?Lt
zpSf;Cd+0)VXfKDilmoENEDYUJdzgfgL^z<tWm{|3vwFFH3(kJYWdhzP(uq#ZP${3l
zdM9s0-y_DfNb?h?G4F@zDl<VKv7A|<+64gzxGZ$xC}{1e0q5)e5RorHd*>Zy|7l1Q
zv7(hZZ&I#<{|yD|ZC#1SxjE(Y+BmjMnfn{9<FTq%I&?%Z9O`eIFp2YCfxNICF1LNm
z7g+ofp4$0?+f*YGyi<gJ<T@FX+*7fb<G3e?ZNSz~+t~R96{w`&%2<eLk|2kFY)V@I
z>~y`3tg$PbY^TA@JvPEf)_(@s+Z@kon-lC_?*pGxKkzH9M9|Pa6`8d4pf$V(iiWdr
z*en!Ol(pI2DTbt8<_Vr&z`gzFb#nbEKQ_iUgbh)SK>;Jmp3a(=t*4s-7i4F_r(j>m
zU)%?02d%K==4%{UH^iDN=;97N3uu&YXCKJ_gso;`C^jz;?`=8+zGf~UO{16<x0N{#
zrUtRwt3l?~aA!hkGaBy+Vvj$TMR>A=i5b=hwW3}Swk^V2``&=YUI#RNUJUu1pQ3q8
zgN8bou+;n=Th%W_;_m8_O&285@ioUDIG_dBR)0aw@-E!F{}FZ=RAP^lDjrT%!Jrc_
zKv30*t=yFlIft$BkS%wIwfuwe`XX4)WigHh+(Dk;YtVi34wqjELWd11Br(VtMg&`M
z((X4{uiXW%90$WO<}*Yje!~XKC2(?s9e<yKC?T5?;6FAG<t#<%ta2MBcBl~-y63Q_
z@?vD=_e_|avmft29K=)kz3hUPT6km9#B6DJ1vk2MsbUSM*<6lg@`aRW)j<)80THM+
zAjW8(TnU5CA|!aZ26-9q5~`D>iItZmH4|?}y_XEmJ{*FJf_+%^SbbXQXw4plb`<Yv
zU`BcG*>}<sv{k(U*F8@I*cQz0?7RodBm@X=kt)3Z8w7jx%TYt0W6Nex2u?o6Hiync
zl|UB5(~Q`W3Ps}QX2IK;KgP{y>O?9f4tE?GWscl0XS05BxuUDxY$})2s>*AHsHk@=
zJ1ZO$Ui5?U(kBpm?J3@q@aH%O%|IinaK#2!tlD@QcbHV7@$@2g`O#5y=XO)b3MKMh
zgg%C=uS`I{@eBK3O$ct<rpG8vz09P%>BoUB;k-#B-Eb!ADe$6Svx~0?;q~<#`^2o6
z|BCC-n)uh?%#$Hl(>?`EFB&3LS2NC9Wl$mJ3M+ZWr18^cytw8Z)|b0u+Iu51^+i2)
zO-#i)=jk9C)Q66CD<J7g6>NU5Mb@7>jg~U4oCe?x9e(BTRIr8p&1Jk?X5GUprNeMp
zNtqaER=~5?B`|0ziAnpvp<;LyGs<Nr?+16lrx<D2G20stOz{NGBokbma|-uY5?*cO
zH#ACkjCbXSFyKHard|_4iF#)Yu!~^E#}jb+&hIQ+9}ZW~%g|5`VFOz@c50OlDgRps
zh1Qz%-1C(=$7AcEZQV`YeJ3mCO1TSCx8=<3M=k7eacw$(xhNgBkf8HJG>EmdI&b#R
z1js!jMaR-!^K9lmz;DBKs5yHUK+FJh(@CAYxZ(ma4ff0y0|h!&@-Q}W8uj0tTOj-V
z3mdSVo3H!j$S*%p5_(RPn4FU!8|<5y^3icPTNZ}ljjANzC#UP~&cZ~+dwBSl9_don
z!0(kCnX75;Oq{n7S!C#sC({(j1I-4USUrYeoVOn@T!&>s5$vZarJ%dn6An8nW7<Ul
z8q*|5OK+yZiF58;u45DiuOwjR<mq6WxCr@$_26DH22NWmAhz%x1nD8#OqmCp&OT+%
z#00Z%XDVZQc?7=wsYdw)y6D0&Ctr*O@Ftn(@t<(r<ZEXu@Xb;~?DrF=hG(C!l93ve
z_cj*WtdGJrcW-=Wy^Xn4eiLrZG9ZlPQMhg_!01j}466BItVI#`20T3#d&v^!#TPX?
za;%Z5-@<XcxO`Sb{9>kTg(5p}a~QrGe}{(?A2ZcYxLrvz`KVs<6wfu3W0SBv?Uxgw
zl^u<kXWhtl328%M^bK4zn8mtvAAs`{6!EdKAsG({VO6>Rm-CHFa=6MV&-63*zU(vt
z@6=ipo$~@L21S`7hi~wEP>Ys*E{7R{olyTb7mZUmFShs>yMXKBj2*lSjAa0beZGNV
z=Y{FJ*uP*ku?F4<`oM_xREYKw!hpZ}WJ-WE=nH0pjx@)$JZMB?KbfNEUZ0$U{6o;c
z(g|{fb=ffPj&t}SjX^)#QKYz=X`dVkXDn~Bk5HeI>IB?v_X&<K7-ki@yx)&6)fk?0
zn^6pl!V3mQ#GsWA#hf-}|K$hPuKb55gg)S{$G(ij%s_s3LLWZ07KRV@^Rc-$jBPk{
z6D?UAh^f=&jwM;DZsdy1!)wqeYc|LEIm}MmE<sN27{YWpDcnE6G0m1u#d4(_-rdIm
zSV8^J#zC6;J4d0#%f0aO%rIEZea>D|_knZ5+-@^*e-QjBj>cE)U@n*K%bhoX(r?`{
zZ-OG-bX13^rdGkZwhzpO6aRQqI4(+<+X1+$DFH?2w!mPOCgZgF0B9O^LIRgT$kJMZ
zzt&6AExQh*^AbN?y-|bag++pTe+|>VJt%uTsg=o%76Um$?v2<UhEXR5v7p)(x(8BV
z->Ol#sGSLaD<zRlSO+m@ZsF@^eK<?p8*3K$ao_m}2ssp?ohRqH9&hJ5`AO{G*6p}5
zUYjmk#_eSKH;y%pT2%h}RqVQW1Ai1M;pxlC_^!E#wMhSj>F4${KG}H?+o4J-JbKt&
z-lx!g>RUEYT8TQgdE(q{@-$9Klh__Nfio9s*uME8xU$C_FZXw1#JSbr9r+H#c=5~}
z8)Ncq+7dRY1E4Q<6a02eW}VcdFt5*(Rn30`Ti*zh(>^TvYCQq*Yn$Li^Ai*};==CZ
zW~0!)ZZ4NCNOL$JXh*dq*>4`eWUO^VtD13an<7Z+6F=hf<;qN>*aeKtHwO<QPTzD&
zQTe2oFi+T@nQxhln`Z9d)kH;t*xp4<^Vlm$tP=z``NtSN`WvUL*Q0Bl<cYgrmf6%q
zP4Z>;Ka9MoLdJy~p!;$+1VtzDi=P|P`_pdowh7qqZ4CE-X+$v;Eb76>Ki*-{^+C8W
zwS~25k)~JPOOlEvZ<x8%9F-N7m=f#FP`E;!Xh=VTMZ1qdty2?Dl$wtvXMeCWxjhy)
zmi4p0k8kC6vQ6R($K1qg$+zKIN;_Op?ZU3jz7Q^6!_>Ov;;(P2^yyy82z1;5=Xwgh
zpSXQ})|!mvS~=1f)r#KtE6|wp)0V3RWB3MLD(;gGv#efW@Z8^w5yxi8CBN~HrUZTN
zDGhRZAt0N=F&9_YpulzkvP`>(mHAqNj|sQ4M{x|MD$0{V_8eaJ-iBiqoX0<}ld0h5
z2x~1VnlSY@?CO&QxvL^HATt0AtGOK2C4ExzLV+~=xq~N7S7Xl66wHrrM5&NVD3|X9
z-S6We=!76~bWo)}N*@`GJ=VDGk2LkXI|7hZfIF5p;^dbiu$0RMJM%i3TZ_g)D@>cT
zZb%3HlR4PDRG$APZXS%S?!+q#*1$q9F(gi#S%<)T(CWu!?`JaDk)Q=fGWKGFyC`W{
z&<HVeJ>X}hEM}ChVb;da0Q(){L~3Fqls>D#d}lE-dfyzsq~3!8+d8wqhhMYK)~fu}
zt&OPQ-^K9}w1^{5i*_EZU^XW`MVNjQi>8eMRP{l#S`yn(e+>dgF8~N}yFFz7;M;R$
zIb|yH^tlk1YkgG@MSCvic;8Qh3EQur@a;}4y)8h`zvko8H`*}nbqf}*`HPn`Rf%$^
zAc;HG&x^X=1CjS{py!_NFe&yqe7NI<qaJ-s*F-VWd_Vy<-)w-50+CodO_W(Ym}FKX
zp+b9F5NB<v0|%QSR*Tycmh{U3t~|a4+9pkyQ;>!Kl}plJD-ZLZN;qTljVcUFx51s}
z90#9Y14pddK)=ijex4dbi!foPPFj>oDXP<5`lrBT%Om{DH(~zy8<V<^oNpj`A6?zM
zS<g+a+)e@tz5DLth4)-{=BqyW^LrAyaeTZdqB^uN(E;rHI4;dvAtdx29)6`j)=R3=
z7kT**wX_|+I7<=#E$hHyqXiCdy@W@n4`N)sCauv@M~ONIC^nprYjrcxy0i_<O=NM?
z(`&4a{a1)PIL5k#wqRiYN4E6!cJ^?LDDIw70%9DG^P1T^UVgMXO&$$LxlU)M?Pe-k
zeBq<-#CDur8wnm4Epb8kLli9i37wW#Kz%|WT1bhL{gGE8@S-fS37^Tv$`vr>3u7VY
zo&f#6P?V%uOHxC(Zai<4165TIpzwGHyiN(mM}i8Z|MMNzw<iXE`|9Js@c`)BEkJyw
zXX8_Ej$3Tr2PTq7AZl9x<W<RnRiP|Nnzshq;vAVv2TtIA?GkL`DblV}3MAVm3AbS|
z-kI_n@(RSswI@E9vbU2h9u}s0HiFcomYel&UVx*UpW)fw4d5TDNhX~f;yOCPxTIVd
zyUp}LeO@l=1i!)sUk<>u+uVDr$QytE<Cq3(SHjJd>v#fG$;mMpSik=#ssyFMjuY2W
zw%?Q8I#V2%hAEL-Q<GsZKasCj-2igElH^bMeLQ6&Lw~Ff=T*2rXRc*!WtI-A()hw#
zyu|DE$jk6F`<1Oq!Zhyk-s)Th<GiiRyK5J5@E6CY@K!Tx33!OYCzj(v>t?8X?FCWl
z_t>lD-{9RhBVuPf6^`+g>56S-+_%>S?WYT%$<7<ZyDf0>yJ58XI*0<U(=q&jGR-(K
z#ALs{f{{EUG9%(0TwE+q#N6^&dtVdu$oUSxI`v4Ok|{{bPXI4ZAxNM0hS6x)igs@s
z(WFqH<4x{j-X0dAW_LV5zi0qI=s(8!xEdGzNrFt>XZU1aE4y)qCY9j$$A??>=&=0_
zW_EQ5Oh5Y`81C$U^I`{DUut8fXR6XY;}xKIelv4)dpk_lb_cocy=e961&G(><8L1c
z;<rwj%N5*$Uj=?pu}Y2feapeiccM8~zzf*t?aziz_haJ!sglW@Up?(*6H|6l5}sd)
zz{{tSP+^52`F_!lHC#FaW^kFN#7!UZuFX$GQ(d}qRVn_bxf?roHo%V3L{6{m0@WAM
zoc<kz9cg~-IgzXEdAAzWxGO;t%lq)_Bv1HVlYmFJXj2nTx8`|`K<{n^sIb@tXG&gX
z)_5;vhVHarmLjK%85qza&#UO`BTYA6eTgTBY9T%B1jnNKg1@F25cz~pFs%6(Z<KO-
zY@R96O-rmHrO6%7IalT^^%#Id8~3t~kCuWDmlM78#GSA5r~$1_&%u@oQBuCV54>-1
z*}7%=WJ!5H|F41!$QZms$w6nf$=QJ27tV1VF8$<1vY%M-m?S1RY&ZBDMS;u6S}sHX
z7{{_kaPP4*pkKEbmUE1&BbU>e-|s&`n2tX2+onyz=GNhfN?oFLQ-PLj4FW-30Xnhm
z0bI<{;&yom(9wUbU@Nr?*P4ox%&Kx`dAS666)#1{mA_$7yAkmacEmdpk70L>0(EL?
zVag<w=)p-6ROXNpon(`RF9l+-AbbOiI>zAGy*KQVOJRJO+X~cZJEym~%8~AaACW!n
z#9y>Qgg7-TQt$JeK=9cD%evCxhUR~$C^&$Hd()XIt%A@uC{M3;`NC&@7${y;A(`f9
z`DYf^VcXiDV0XZqpQcv_%bhttXy*pTSGx}54T|BvL#;UTa4O#FSH(|<G)RT<1+?+D
zKqao<Jnef9t~A<+A8e1pn8q#G(0>RwhsU7Id5U+}mEvZhBg}-zLKyz{8v^sYV6XT>
zY|~XjnJ!Pr@+igJXdk?0rifQ9w}a2SGZ<VIgvBEn=-?<%CQ5fQ_Xh33f5UT(_U?lw
zy=169BuSL4OQ5|~gIs=T2JPY*F#UEmE`8YuvY)@9T!cL9`@Nnq2|5f0`%{_1oJ_pP
z?Jp7|!MsHwIjquLJ!Z|090XH6vTTMc6T3{04m?!H8A12ajR<4*)F^1}j^TOQ=@Wk!
zBXkX41wnU;Ij*4{o^BF=y}sgvdCcI8bF*;j#&FiaQ;Lyu3P=5iXTZ@l9$GeyLQ9(t
zj>{;Zmq#^54ckKbe}XhNypm^fe*+{$zvR8@c+QMkgo0<WI@xymDy+8U7!LXen9S?)
z<a3885sQw1ubtDNE;)&H8YzLNcb}pB;pH%{=Z{x!_(5agXAFLQ7YxSkpngFQ^VaYf
zuK4~5%>A>#{o{S`TE+d{)c3*B6mMW|nsIzRDUxt8ncYruAzx!I$3M^qk7PaixLb|(
zL_GqJJ|(JQc@xDue#5s7`MB@&2~^!6jh!Vr<fn!l+`GRB4{ub%7sY=duvL?SwI%3;
zk8=By#mKq!>g<t|G2kV74_2m&5PM};uw7jU3w9`z#^59H<J%BW-gAC$hB2Kk9S-+{
zyP)`JJpBH8fz6wohik7lupL$LO!VdqP~B-vtds99oOXfZ>=$!)XXpv2%^L*on<v5U
zxi2qfnA3xYKZEotJD_u<;ny~Qet4G<$@MCM?ENy}FRepr(_FF7#gHC|Q=?Zb9Wi!8
zH1nrNo82My3#D$&!MWq1e53O-*ntdb%q{u?=ZA7|UixEx=&u%jv_b}JF;||;)f-S{
zu8X4-DoT$|vjY7D5ol^Yh!WjvF^*$WAKIG+2Vaas`lKO%GI!kWD?-0E9p`*pEuyeA
z98dR^K;?)bk)sW$lXx4w@6^KNiSNK?sXE!8DM2izBtpo63T979BTDK>(nd!suvU|V
zR?ZJgkKD`QUw)z-_ZIZnCx9O-vYFGHqL|*DT4dwVTDTDMfThKTL}xIWZMO``iI_6N
z-?C_#S-Zs;9Qdb7U!o7r<GgcC6LDH2C4g4$ELKK%fvs{sJo#`6=Z4)xo5}T{YGVt#
zG@Rjl{t9j<q!d+CS%h|ZZZJHQizb!9xJH2M4jUSh^j|T|8%aCNPJ9K9U&}D#NIzq{
z#1D$81?X@Ln^lF4T#w=fr)y2XDXF6nH|-KXj+kP_PEQ;tr~GK%3!G;Z#!h;D0rWHc
zu`ynm3MAx#&6L;dgU|AGXp0K1J6Z*@yKX~8K`XvKBSt0T#khS-M)dK{6;L|};1c)=
zu1)#@o*toK+r{mQZu|jzC-d37ox{xah5E$QP?DbePz5TM+H~^MRd`57fMyDB0j0rN
zcu~&-@5^#*juZ`YO~8@=vaA6T0zJUy>Ns8-;eI#sbD56o9Zd5IaS}V>Dl6vY0t?U$
z=+zAP8&H7n{?p)3Abcj^s3&e1tIIyjF;$W!6WIRs)sT2!oy2lyuingb_^I&?RBmRW
zve6+7p5TQBiQIfCBMhw}2iV$9E(0mnfwDWEL+Zz;c>8+|w?ASVHuoefbXlAV>#lOX
zg-s_Yr@AngFK~?LD30~B!hrHe5pHph+arGeq1xKuoHq%xvoD<SV&m#XX?V3hvnTcy
z&e*aZU4J~oL+vcvm$wod)XZ4jD<(`tvLH_7Aug~Jq;I+`p!JXj<-|d{eX~D2^AM*`
zPJ4k<II#crtmMaSm16AcjLGeYiO~Ep1FXw}@f>p;gG{B#qCZ^UK<zz@MoKYpFQwrU
zktCk?kFya^1M&Acj>q7af~l!Ow1;D>oUdJGwkzQ|KDNJz0%|{?clb8DV)JR(S(ya3
zb-GkT<ty&~zLVW#y9UO@m2p<`0-V472dX#<kZA_4xIMuI=D&IY?Y$Ht1Nz|RW+f81
zRGAj?teNUhf7yl)!Kkhf462J};2||`j!syB8{aKt2Yz?p{<RvU#b1rsK8wXHr&z2!
zpi3t^hA*&pDZ?GcO1xbY<cL&~4JvYP<u`mDcHch3-w?hCEq#kPZN3kEo+#08y#{ci
zLJK2BY9MlbF*~-;56@a|g7xR@A@SJ&N_w_4(VfzC)9&N!+rV!0yJ5*jzrTqhvl77N
zBi93x6`+e^zeDsk15#cnLjoJV!1eSVU`xem)}lkW>_<7MzpI37TaLT8Z4rh|Q>3CB
za&V#5186lAAX-<BfmYHDzW6Oga_|P%Y1~l3mTXagU`s8+_Zh<LUn<#UH<D09;~f|*
zeF@rv%1o<gEW2=)1i2h;!`=5u_)E)|A?~*^=if<8%<AEI2>m=cjti@O|2?D|@G<@9
zO-L?(jYh+BpkAp6w=)fpAD~Z#97^!Cls{Nq$>m=O55NHfIe6)q1@~SFfZ~!@P(MYC
zdJD)<Q~P2#mpg!Y5uxCJNSTr0_EJ3(n}T!8kHRJHJ_&iKg=@HchhhJ5{L<3R{N^$o
z*#d(&I(h=f#pa;h=T*2*%!9qlF^Y3#;!)yYFE$R0KzQp-cr*GEW|Swxxx#FCZ^rFa
zUAdH*IjIS@Y<v&fT)7>IBHFa~*-vK0=yZ-@ww(z+tHPdrZ%A+ErSn?i<}eS;H3&QH
zDO}-pkoKcL8t#@Q7g8el1I1tPbyET^U!=~y;%5GwkJ{8?z9b!4@Eokpdqaw0B+ll$
z;^P=Ol9iH&62lUtb&oFBk-iUmT(05m8Jt$L=qrfRP_{%S4T`nInU&E&sA3WUTe)n`
z$q#x|%Gnh{yjZr(=qalj-V05lo0vy$>fw3LX>8~E3?|QBVaK2}0iPmP{Gb5EpjVK!
zcMtj=OvJ$ZE3xE{H(sw8g*r7=I?HPXJLZEZRKazuOclXn6Q7OT<BA`T%9F6iCtx^i
z7!Mw@gli2xtf|v}+{(OVT6#q3>3z>})qhQV3*Y@9;kOB9d-lWiz*dZ06wGBK#<`A<
z4w={~NmYY5uPy5iTACDadVC9L9224Pvx`t!N0O}4QKL_fDw2RQVD}!cXISF~bh~2(
z(X2GN`azAlgv@4{(!!jfEBU<DnQDw$)J-O9kw1GYSDJ>sF2o9nCiIVcjW5Jic&nGk
zq33)Jyjyn`{zuVy$5Z*faopZ}@7d6xQr~i(>uzdDrM=VC-aBas4YHM$>=hzXoaed|
zDOn*DMMXv-T2vB#_wR2n{y5G#&wXFl=ktF5RX)z!cvhq6QfG|6A0kY5{e$*DlwnN1
zJT-p!2TkAi;{p3Y_}9gLw|gRNbH0LETFhDd{sqQglceu67-v;nhScZOK&QVb+5Gby
zip*r^#Z(14CME~n_icjFcY3j4p#rh%_rp1MDrD|p41D~P$$v{^pNlKyP!#YKF7%YM
zGinj<eK{W|#>>$0n-!=Y>(t`eY}j^8mx|o~30?U&c*k3b;Ksaw`4?Sas7e9$)imH_
zQ&W1)Uzx5<H{lOF_e5#tavnM`3TA%$4q->dsZX^Q)f+6ru1^QJf|^QhnOPBUnPp0M
zxQlYmemtc67oacOS4v!Mg^oD`{14{lS^mC;m$+`qYrIJ3JOa$gKcPMI6iaYL``%++
z?nxXum+jnj)9_qhFh19d#gnP#Fuw5#xU1&Dqe(4TmG~0#y#{bvyEL`g@{_r(@<At1
zg3Lej4mSQw5!?x?;RobI(b;+eWK|CGmfu^!^(gyH%NK%a>@7_1Z)Z*lQKAyIl@A&5
z1h=kD!j406@cv01M9i9vb&I9xD&saBw?cs&Gczaib_c_%j7A*!Cl$od500m+P^H{(
zEHTp~b!$|J*=B3>^fRJQ@vz`jz+})mqeh)B>d<wD^6+rcdEURy3>9Xrh86GcfvN0F
zfoM$|7VOJGWxXa&Y@sx@+GUC7w|8TlyfRguRR&U9USr*^G;q@kVRy$Ibe-}PpPPCM
z;$7;H%aa5<H+}9+bUDs8kf*{sD?y88Lw7Ar;dV&W!!hY3=8izP+CB$1Wv;_{TIwY3
zR5q7+L6RPCmZzibJh^;@esJ1whF2?Z28|Ntw=EClit88gUQ=#ky0#8Y&3wg=V||yk
zCfY<(d>O`HnSt_oO5Ddq<|q+hPUF*M>Bl}5yricJ6_KCOdXx!t$EU&S2c^8N!xG+L
zU>%OTYf9!tox<5_0bJcNDlE1~?m+5O`0-Vh+|RfQOM7ML^~xpuuWi>cZNgbJ`0^Mw
z2mZ!)SIY51;|yFJ`V)6my1=O&3jE4R{#>C-CBMC?1?GRwMxkR2>p7JQz8X1lsB{Ng
zzklO1PankmIlA=q(G*n9T8nM}gkW~ulrkm~I7Wm_^PIAayARAKd&H2O2vs4GCIy1Z
zvRJO{{E%R($wK^U^+1?WFOM;|^dQzkirltTArE!L=<#tf^x9k%QnTd}^C;_+0~0J+
zjv|uhcfSD%n?ty`c?L{M*QWKowFps)WK{TLtbR&i@QD{+kof^3JXySWyC%KyN1fy(
zB|xd21Tp-U4Mi(8s8L87S2p)I#+*0@o3H;yk?<6L%uNaU;^J%mlaDq8%W9K9_Y}!+
zPBezCG$0=*X_EPSgAnGIa-GJqaQV7A={s-7=HhK&)N6wyo!dAqqiujYgfJ<@g?W#H
zpnHx9k$%+<kCn6uq=#VR)0;Tfq8t)Z^tqTN5g=)-%NTTK^paM+AUH>j{`vBsK&X|0
zWws^Q=}-YyLPz|xM})ecE98nD4XN!W)`x#wD;!ccrBfG&V*B-XSUcWdXt=l(7VWBl
zV+S-rZqu+}c~dD&ijX2+QyQS`LIB(~Vf;Y5be542B{`C7U{CRG*k!FsrphmXY6r$D
z)h@wyrPso*^7nA-#akG+Eds2@1JeEFsC>+epR0BP)ci9A#s#7zdtf@Z;gJDNX?O{v
z*=O5-sXAWDrot&hG2GZJ9^@PJNP5>dY+K!lji0=8$9~*{K^J@l+O9_Quu(5sRjlAM
z_l(9UODS?HO_SbqkivKVx|BFx<KAgW(5h1^Wa+j(Zfe^YsL5jaidJ_tD9;xzFFu7;
zbIrKNpP%A_LyELD%Zwx>X5*bVH*nM@IWkT7nDZ<w<`yI|fUa%|d#wRabw}`T&dbu-
zOV{$_vZA@UMIv-|?FW=Rq)Cpy4aTA=v%sje0!DTxz(U1hv>z5BzPEh2IXhZFHPr)h
zvIn49$r;+pEb*|n98Q@lO^2gp>A9O7p!q_dh`ul97QHYd3I~=z>6CSNL^c)wvkc~&
zKX7n&p(^$5Sc=EUPq-E*P0KI118GNmE)c`j5o>W7_l)hT?xFgXRov&9slbnWiE}J&
zgLl*dm|rbIH(xiQ16&sDnOBd!ZV#CU@hJEDE!*J=8sV@%6@I_I1TU+Z(8(cf5WI6K
zZ{U9ruFi6Vt<1%I<%b!a?Nf&y7srE=f+@`m)g;mRKj4I~5%F~&#<1`le$m-Vh+^!1
zuP+8Pu1k$~HFChj3@xhLy#*_0YLeW5>G(qW3inX)rf{2=7O9T3h7i_&AGPf@H{;7B
zaF}CGS05|~{r+z-$(83P&AgB1g_rRj>d+l3mKZ$$DX!YY+z6XQ$)jIdIFy@%i6=f_
z-=+^B7OqF_-B0l2JcZos!*9XDp$<c4enS<;w!C<IJ~z(06n92?gVm)^xOM4aYz%dV
zSt_bDqt=N(Z)yffWzk?0)`jG!0>!cI{AYc4fuKd3=+7^Nh114iO3*>}Y@G!G1^IAX
z|2#VWk>R~17~8HgOJE#v8?>t<Fpj*yq6>a-*xHagUS<GaKN`}r-=wK<dl8g`8B#A>
z<_<I{$Dv}z*csCY-rC);xGzR<!LW+c4GKi=J7aL@s#DEfW5GX00mfW3AqrwWpe%m?
zRIjr0?ZzUgcJJqRZ14r&1{oUv=RK<NrsVT#5%|wjgD#ddp&Pv%;QI9UQ1&~GeaGlQ
zgcozTGal{Sr<&yRkry~^mL&5}MZh{efZ2<>G4W<4rZN_q1NF&mw`c72P<67iP@H!A
z#KW!(ZD>#%;2Teu@xxB0Sm_##a^c?aD{}#?R%VWmB@#G3$qRn0ThH%(<`2w*lk;-Q
zZ0=v+d;CD2<NcMU!sUTt)K)wm#@}Jj!#z8MFSvQ|sjCf&4t~d&%>h`}RD~P4tGHry
zBs}=d+>t6O>=}Ft<Tu)Z*nGBwYz>3FSD*NXBb{(FU<0?i{4$1ERN%?o!FU->iCon}
zZ1*^XCR@JXr-@qN?5IvH_rHJ_FJ#GG(^v3#svMEZG^Lx)^@6gMCBX9eHbbvd@K{4M
zEU0B{^Mg~N_u(own`5lN94lVuz7Z+3>%;B01Nmo5w8(MBhg!jD)4qu^bi$={n6jl9
z9_fh?m79!99{3wbA=}C5Z{^~<NAp8TeekR_4KuT5qjl;gKI=*wH{A6UR(94xNa#b}
z_OStV@;0RJZ*~Z!XEwoC!*aj^aaw&{mu3#1!Q*by#Mf^JP7}Iw@~MmQy6;<f=erSK
zD0~(++0H_j@Q>i$d<uTfN$1P;;(05%0a*3#8hq<n!_EKRAk<yH8}INzuvV=Mmqf3G
z3G0}vB_R})S;p4nWFu@IrB5DyIEv=QpP^{yEL_SQNt>RS;{Fxt^rW6Ip3Gnlq$daQ
zva1GuC^Uxgle|%T!zqxS%NX2CY9Mq_38Y)Tp!PmreB$sOg5ONQew}N$VQLy|lPUs@
zglAZ^A{25XI<ai(AqX0-fgiUesoGFJluP+wpyo(e{8Wq=x#Er1%(JTJ{{_c?nU7<C
z2;e=Nb0^=_rMkDO;ofdJ5@KgWGX9L=-QNF0)mciku(29+*#A`P?Et)0@x=dfmB`Bl
zL9qQrCjJmf$G!qZDswRohQC|mW$?yBWj|o+Z7WDsQ>H!pG-x;D%H$^&L+F9|_?FMY
z7>6hDW~C^d`@RQSRW5_dhC5tNysKc!A2Iy7cOeEgs*#JS1~{!zjd~ilgL<k7UDI}$
z<v2}gU5F|DXvTahz8drru=l!!Hu&Dz$ys4K9Fl1h1P$x(_h&wU^G*FIUlWFF_spBt
zBs>b6E?R=?%^Ik)e46{BAQYV&1NdDX9sGgAVq|~!aXzyn4POt*;_f(Osy!nBzn>Lg
z!FGL8eP%DVe7pqrt_*T7Bs3_S`hoF+yKvsv1t+r{U}WzTKEwYzsAWr#RckK8*Ke*E
z=5Yu8F6-m-94*>s@e=pw#lehy;`F6SANE9d;Ha(N(CJkNY@K%#F3wLt?b<LrcvFmw
zKF8c~ySC%Z_j&kz)iTKMH>Bw|38=6vv(KIybZqGqHb_l|Z_G>mBlH8$7q1XzL<T_k
zr4)GmWfK~SE#cMf9>+18zM_kp9tmBv91jdTqG$Oa_9sk$MR5h3vdVW}VVw-UbT~#J
zy4jc*j(v{7m6tKORSJzH#K@*(=C;WHjb}u53q#qg=D5NETv_0bp`EWVM^l=l$0?E6
zf_OZ0;fnA~$2nLvc`i0MB*C=q<#=g>7ox&EK4x2mP{vx0&U6sP@|o?7-Db?auF%4#
zoAt?``X%7AQJ-9nG!@29mLicJ%;Qr<q2ZT0l+R|o3e81$q5lzYpdJLPrEY?c$PD<l
zEd+}mEX64slKI{133g5^2enIXaBtpcXmOkj_KQDrlUWXX@T!3OC8&o7cDVx8jSWzv
zYQ?V&wi1S)yTdOqlqU@bvIN!gA7M*=1n&8m#5*VKg?2$U%HBT@>wDecNyT&Y4lpIM
zr==m`{sQQp_!h_2%ma%Tk6>Sn04oy*;p%q-y1BUz41HxuHS0kQXTQRoi}g9xiy83b
zh#gE(tL7G`x?+>tEtue{$xAO*AT6O9%;{1u*dC=rmv88S>cCz+HoXZ$cX;BIlhfcV
z=LX9ylyDTE$~pdc#nrIf#ulp{7;9IFpB8Vzswv%QJ&)b-cX8aX=_7nG--H-^5TT_n
z_2`;yQT(@k7U0_2$D1x^&f@r0d{wO{4AisTdO(w)&@~S4rZeW~WR{ux*T?;MAwqho
z9Tw&cVeXs)VR%>qY^~gnqtD0?P1YOzpx?@A7pjsRogq|tx}Gb_Ga+AIB!jQA0-h~%
z%1u`t;)fD%f#oSB&e>uu^g0jnCDV0jo!nDYx0e8~G(8%rP>1*4YmvbnHn2iF5h8pg
zVSlGSu_@icJ$-l%Rm;q%fnpW^&Fcz|^)1BV?_VKfL^@3TuLzDuX2Lpl?~ssW*2C3j
zk-CVHC#Q62?5kW@eY=luofwOLzJB~(Gja0bkt0;Ic^b}qgmT*{ufZ67!}f1*<4ti|
z5WJC>vk~$Ke!sxrK@N{4EQ7}{oA{aggK`#EY7+hCJ^b1E#`MUuK~B0(6REyG+}YL0
zm26~AmIL>})<Bf%nxt~q>y+`$X<1sZOo@hyRH4i9OtkNpM$e;HVEL$TIKjpW*M^1(
zOO}h!88rp?E65HFMpt6Ho)9l}=i%AM@p$Z_IpL$+a7>00O;LWp-8WE&ysiHKomjq<
z51jV~_9}GZmkZN*S;h?XWL?rR%__Xp{z`7`iv%EL+T5(G$$WNAsPLNoC!zGV!|-QI
zG(HOr74CDZ<3}BA=knK&Lx=wu$LP8$&AAr>#wVYl&bu=7b3cU#`nx$3VIIi&Tk=L4
z3Ghid0@d=>$enL)==V^BSg*UmkIGL+;}y?2vyD%2hKM3gAO8t_t+a@fz6CZ}v%Ex;
z3O#YloQ#=U2th*`IJ<Q*KH47-N1Lng?W6Um7XA+=51bU9+aX2_&h??D0nfEbGk#|i
zV3>LUM%1|Sw>qX{*7k7R;BQX0{njQTZ<p~iTTMvbR>|D7z8_#TV9cdaRp?r5O4lV9
zqkf7ZcGecd!boi@@$m)Dnl=|w&c6Vo+vdbP-xr<_45M<a7s!rZ4jt_mv94k|YHbiB
zJ8pP!u3^1!`#bB9cxw{PUK^p<`EKq(%SJ4Dqz`?s-C?(5E@!dz322uX!puqovea7^
z)mE0E;w*X8J>7*Ratb6b(~qwlqW~65>TO<@F;B<w22i!BfxnMKq4pr#!T34zwu}*1
zbYTyFN_{`bLObJ|zvIX5%>bt@j5Cm{2AZ)6kgg(4e%>j^Nh!_v;VI$;uN(Mbn<i11
zAi=q1U%_tIY>ZaljR*83Nn}VVHl6{#vat+jCbz*(K@v3FSinK9DOtWlg-$tL3^g9%
zLSPKLP?ixDRU9jDim5~6S=(`Gt{>d5NrnCZYtWjIf;N>Y5dQueD&LC)-{*{>7V(9j
zGf|RMRzC!nxz+r^U9#+AEea`&;Y$)+Ve`Amyht*06+7~{e@rSK&$<FBGpE5ru`N)3
zaVDNR_6ucJi@`oG4}RV1gxuJx-*^LM1v)ZX5_R9Q9S-B9*!7pfSmzjl8DE3C*V%LK
z<~OXN6JT1fD2ZtO%O_Rs<@|44hnfG?Lco7naD?S(muUXvPYNgu$uV}{OcUypVZ=K*
zYrvFIzwmzYc*f=0z|U|qp^iJN7;7RI<mwd2=-7n<@ykC^WOfm}Ib;RLQb%KzS{CeX
z+=La3Np>dWH*_iYgX4rjJRdJd>IT@k%jF$6$xVV<NwOUC#Bm^-s7f|B8}g0;Rd7cC
z7)UT~ZRM0g{&a8#>}9)|vYeMNeup?Ykt0sR4hS*BN{?({S*yU3&zLZ)0Mh4$a#FAj
z=Wo)W{=z4a3SO`%A)8ML2nDf+pYZqW4!qnRj^EdefEa5P>gXPa1!I_(#{L(-Vdfym
zqy_x0c{R9mr~q$hCgFv#o!Bd_O-;s{l9i1kF{$wbo~V3>C!hVqONni~TGMl0{bdd+
z7HQElnnQg2)t|8Stuepyi#F%XS@O4YnlR}vpxmM-*f;7cYMqJ2Px?=BPFXUVoBxK7
zH&{nt1Y=96i_q>F#du-jZycq)3AM+E!)32hu*(ZUl?UxO&t4q^(^4VKP@TH{{VVjI
z6b~WsKhW}rA+b6#0OJ;AqE9>Pt2|sOTy><7*JPQlFW;x(c@dWHjzZ9L(xLX#wqrv}
zhVajsaCE#N<f~T6k&&-o!mx}GK0encg#mRSEKTDbXI{r;U1oIf?i?uB4MW5KH0V0L
zy$|Oc8palR#sphzNXtD<iTao*wAYBiZZ|a&zVZa;aC-$ligFg-b-ab-{#)Vi7JE2x
z`77wPcET4&TW(dn6O6e05;nvsPz|n{TV!#ZapJeY!y+mA@c1Gw)#6Fcg(yRax}Z(B
zs%r~V^aLP`)yCDPpU~yi8?bK(!edv&=&OnT;9j#5?gq$`J=n#V&LSjyNRGs|$Z(&%
z=fbJ|pLpxODwx2Tk+hTHU^c1_VmkA<tpQ!U^L=9)e(ETysEk8LUXLWSC{pils$^pI
z7v!{+;b7S@IGW=QKYxpG8L!;1rQ#oNE1iNLlPoYa+LRi6Z{n19sBr5?DG~jsGECbj
zLH=`p3S%ut@gHSO>3E?O@%;9fJKGV6NuM9%@Av0erpt_I+9ctTv&Lj!fhoRiJHrdM
zYZ23(l92p09u}%7(-%R%vBkgzEhlwgtYi(WEy;m~Pp`NoTJxE|crMlnQ(^mm{y3j?
zVWy00f|~gnWEAtc+_iX!=0TJBxCzX08$AfCM9;HK-#eT;tO1V4t8jkg2QF?P8$Dm;
zL2`CK%QAX#Z<`hATmNZzY@08)YSSS+tbH3LL#IL2gKXA6+05I^+~=+Z407u)t%sdw
ze?Z?d6XA?SSD^c96Z9Nh$)8SYM1{p##Ojq2$8BkcZOa(r)9*T;d1EseH?T~5wgFke
zdKjiPMG&XZEgbc*nsfLp2gi*fq0MmxUy!89I+|T@W~>eDYKh|R-FOMVmgtZ-f_%^%
zd76{h9f*scKg0<JE_{(t5*M75g9BV3em;2_#$;Z?`m*JaXxxc5Zm&S(i4y&3oQ+>M
zOOY3Ae#6?D8c1-ug<ECvQR;*{_j9!!7}lr}uW6Jk*~@rVFDGGr^$XSo$wdkOzdTgA
zGN%EXW&66~lGN2O;$j##>&r)&{J;-1x`Z&#`Y(vQN`$!@e*AVPe~b-y4;fL5AZ$1V
zg9f&6pXa~j`dLQr>Vsxf-lI*N!|kAJ?iyHO>xHl}7rc5l;lO1D7#tA>6EZGAmxMWK
z9={oO^^Rb!p-e&f$n$u(LYX{0|Bsup^e1M$9)=5zPdLxoXZZ7`ITgMf0fr?b_^ACY
z&}`KVi38G9XR0|Vk!P$5*Lqy8s)Mrz7!%dV$HvCn0(T6mkyBR#xJcC&%**Y^q;eZv
zWN$*h<bK8N2b&PbnBkmgU1G?-+q&OQg~<&D@Zftagp85ESRD!e`_HX{&$Av0$0c46
zUfM+YwBK);v*I^*aKj)-nh!&ZiVj&g=^AP*TLcEICwJY|gVUa@2+2d4FxvA8Z#Jt7
z&!5&NzduR*e@BO@zg6ga{up03Ns~GZ_A>XICN?jO;YDo^^5v{s{`1^*jIyvp?+#6J
z$?-70?O_fC4Og4Xc2D5fdMU2Dl`$D}#c0Q@V3;>Q6RXZ@kbPmb=$)fZ#8sG={gwhg
zzV!k$iX%`y+7^D_`HT7Om%&_WE$b0Khev9nq{D71JddkHe_sulr+o}U<F12Q{ukIX
z+Lh0)mmrQF)|hGA2M3*_vD)MZJPi@0LC3_%UNsZ&bDsg#XFu?3)@YLF!fznGNs1<C
zMd0vsDGZv^21g1F$iI<m@Yib@j4QpwPr7G-PAwl%R!oQf{?gC8ooDy5i;U~E%^PK^
z^I_lMM7(4-G%aIL2+M4YAm^eyR8FwPHipu?a9n}ph@XVqjj3GmnrLoFgt7G(ionV@
zKY1n5Y@yy)1*-Vn1@&*l!ryBl9CvO#mMoJHo=cZtEOIBDE!_)+@_*p5=PSO=+ZuwF
zcJXoSySnAKI=9)|KBsn;1sDk&@hG&wij97}a$+JDYFvl9Q8&2Y*-OA`z8Am#pea3g
z@h)WV(4bqVCgome`vSRC86+29MUA<4cy56Uw&-~0UMgtDG?AaM>a-3`e0CjYwl?u%
zhg$Gl$PgAx?S=7@3S`s=)|a_0@xZit2Rg=ba7*qFw${6Fo|Wma@lH0Z+)@L@lU2ys
zlL>4l%l1XAZ{ZQ~1m_8L=;Z$UEY}>$8$Jr+hNZ$Farilo2o8mkUny`l^fT(&sZw6g
z9gZFL;hge<@M4+~`sA~&(Ia_Mk^2Glj=w?GKoc@|x&iqwMhP>+3?OjsSTt1kLvLDz
zB61@kzha1g7XA`-Gu?3SNE=u;{~I=qPk}uL3{j-th>nkcg3cb5{4cvS^w*Jsoimtc
zVAlp{XB|RXGt70YJt>&#5yVaWbPIzXUjyBrjBBzc7F|YhsN1^;b1F1w9_vmnyLgpz
z81Y_^`sp5s#<jyC&u9Fo(_bK}yPkI$|5Ugd^I-4l+Zg@2l|NGv2ziCyQM}54M(oQ3
z&wnkLwd@02y=qP>&z^yGfu*oPBMjkW7`HA|0-welhqs=|xKdG?h!bnh;z9<B7n$)3
z?Ek^^)^2E3u*O|cmw7u0X_}bN{5<A{bb)CqOumZz!o9lm_(^XlENF%9r5$|W>R@=l
zIJ?^xi_wgunZhWQPQ2)r1lJNbaRa;=?X7r+hfbCAkoyHYd^T|(-e$tj9c*sPxXio1
zX%XL;g&?G6<fX3?JuzRFXe@1m5i0(C%RDVC-8L4N3zx#lGda-yM-(0Avh)4cRyL1b
zj9WT>LPBy0PMP=)MeI5R??VDOxjP)k?Lp4<TRfC*um{H!dk~*yLesrl_@Ur*j*By)
zM$^M>tX=NH@a-aQlBg8*`KU*otA64YZVkxePgoq;h+5f7^wP<8v~3qcXK*yPUiB!}
zQZX|9#C05ed=wV%6~b!fE3(=3947va!vk`0;4}3cP9BlX;h`dQV|U`ke;ZJ;sU8o^
z+<}^RJHY_uiQO*>mtKp4n)?xu+c{gfyRDk*fAgKU>-vgIJb&POYgsy@R{^vCedV>z
zDAOZfZgIBiH?j24aooA!AG{Sr!6+F$>e~H>n{_D=j(LbtO-BLm-Q*AB*Jx4E%}#iz
ze=6*gGQda?H+XK;3y%eEoap(foKszzFnn?lOw<lzE`~uq{Awm2)HxYs^yTQSS2AR$
ziYw~=)a26*<;al<W+ZD>0q<{_fq&C<>GuhLar?7Jpk?L?SzAnKe1MRvA5fujOCP{*
zwsUcEmFJ&rNCG}%7>o)G$)pxVbfgk6w|@uc@Y|EIzO+bOgci-OIKUW<;zYcV&HG1)
zllI{<xK-kVPNQ_m0%Mjf$`U7q|Jv|A-WIkz)TH{)4)UsMb1-G?DZXx0AE(eQNpz;a
z=XGmSP=n?1JXG#+J=WE19xwt)M>^DRvVhpD5~TZJF@MKT8M7)Eqew~+w2W`X2op_A
z&e5dDmMnk`|77U!tW*q7=w&?sOK97F0@qe4QISDsenWXHzMh$ZZzILYit((=^?>6$
z4xPr@dQrMaT$_GUqIm68JlbqwZnZTFz-V_7*Y3U@WaVG*CTm4dXzdQ4A5`&{uT$Z<
zjw|fHVTE<^%UA?=A6{xP;5VZV-k6qzD=z%u+d6(gE9)JrxI}ULqFeczzm+gy=SmzN
zRL3xf0C@21GgrUL9ELpAiRN>8oVQVqHut51@jwW!ddJukr+?$kFDr19rap+*o`<7-
zL7*1-0Agej`yTZRE*~_b;jTfz9DF!Yp`UwP(uKVp9bCdaNv<@08MvgfEXwyJ*2mB#
z*2TlT^kwET9x$W(ti}l!m3~9_hziVjs6&^?731x#YPixp2L`xQZpPh8C_dqdLCRtH
zw}UZ?{>{U2@t1IP`Xuh|-9Y@7dP8`snz2<EpM;6H3QKLwS=Z+r{Hw0zRCfgD=9HS#
z)2wUR^JoHGn*IepKk~wY&^z30*0qw_6o88D$z0K)0MzVi=6q75XzE5S)(g39^Ov!x
z_4VZGu9_u~m?cKWYp3y-4^4$~<zQTTtO#68dtjefJ>2ip;;l|Pp{)nY1;-eY72&;b
z_x686M=Oq7A0S7gZzglp=nk$sq5}^{ufWMOFK_{wGT7BC!Y92f!T#MvaP^KVd0lUb
zV>V{N_l&QgxcmU%3LPv@&>>5^e_}<O9+BwE#Lwz#7(S{3k1vd6o~j0Hd~e4+XuZlc
z+`5V9b2W(T83{5%X99NSjDP{HI7n*JB=oW>$?Bhtd251UYDy+PTdU7?kDiFii?X<l
zXIpT)t`0H%`jvAU-T<fPegMZR6=*(hO7dPxqtAFL40s!l1I?^wwbv2~6Uvbe*5S30
z8YD!)PPi+i7QH%E$&<U{@ZfA0>bSA)<j+Z1;jBVM%&PIxY9EL-xW?~076X|d;^2$^
zJIw0Rq_&?#S&nck)Jtg5z3&L;R(BPHcdo~`b3THFzYGnp_#u2Ec@AYk3Kd5kL$jV(
zm@N~^*B8HpQKwo#EYyrD%+`d8F+H%!$r1*PMTz1Caca5jAeTYn`5rL`+^2O*C}eB|
zT>)csRAxf=I2AJ6yAXeEE8(jgL}|_x6(aUE8FucS1ZG!yp|sL~F1_c;_HXKR`O88m
zxYG{<Iv)5nLq)Kx#Z%B_;LXwJ-*BR52KM~*$8j02;U3g-8ngdk@0x3v)b|cN41&?k
zVGHgLQlb)h`=D^rHyrOJMH95A!RFb|v9wEqJgX`I&${h!irpbaAQPQ>N^w(Q7~XeD
zhioNzm?AKN%3r>@ql>i2x806h2YZcd3+G_=c^7V0$VN^vnavHlYEdWE7<N7Qj$5kP
zdFGBOEvS;>(tZf>csGaAtM8(T;ukKjxgV4lcww(b2w3-aKxal9jxS?5z&A>QH&a|;
zhG7xvrG@eGD=k1`>M*Z*yoVpH_KM?uPIBWu*MhR}J6t$tK#$h-fvkBizwoCDtumd4
z37d84+bkv0c=9hwDn@fJiVNYwo;NtfQ3DSDiNM!t7hzAJ2fnm5AqU%;<2CIb9Q>w7
z)^zJb51V6=vN>GKm(e)Q_CNlmTniRHmLrjDr!jQo76iq&bC2d3(U-fvapUsdf-@+Q
z^ly34%^d39P9e~Fl<g=t>tfq88w^oYCWYEP+$>8ctofqHIvA61Y(uM{={<*YW;){d
z=_bVQ>R6%Jk90U-(FFdRO~^^cG?thygw#h8)al0`yngK!j`rTm1udEY1}2$|ZSTt6
zOEV@3No6>&hV{lq`Jn5FGw3%q9DVxq$XE4c;HaR4$90&a_NN3nU>1haHznx?5)1Qw
z$8!5NK7&0EbjS&xD7+mjN1gK1peWP{N>ZQOc<opZ<Ig?8)82o$87}}N-Hhm8)3^92
z#}k{DUIJewaZ=Lo9ro6!QsqOo+&2RuM(@4D9ID5-6H|5YKexyHH}h10F&fmsr5TN8
z)I&m!9Dn(eIW3pJ4ih*0!>-GH7@Z(bs)mj-PQiV&WV5?!B@`|&`p4D#?uB8KXL!hc
z5=3bjLABQj3>Le?XYNcw8;2CwvdRzqI(4YtR#kK!iiN}DRzpF*0u`^#LvMF|vd&+N
z2FRyEbDlKp^w%c~2h=ciiVWk!GWKC`KjRjazz&zsV3i+>kv7K!F&pm--_MK0-GO>U
zN7Easvzb$8+CNBQ{)J7J;~AT&Q21A?A4X_@f;2W49y3ymERB$%=}Pf%{A)h0voWJL
z4jHrViV{(^eFzgKZpVNw4RZfMfG~AqDr&Nw{Q3blAGxPPRcGA7<E(G&l9VU7uN#Sf
zyyx)O>_<SlMJ^^UpUXunyyJsnUh#Dan=n1_9Nuzeej`J|T%74PlZ>TtrpY^$|HoK~
zS{vbffHKt>+0EOY$b<22^>Au)G|KZkvAbmj9A451PYeRty;FrQZM%n7tDLd8)0<nX
z`3@d`Y(WhVFMK;x0WFC&*jJ{9e-B6Fh9pb=oZcd6$(AE;BxSjZ!|U+G`gWZ2<s`gu
zSpj8JJ-MIOjJ@Zx0TryXVB^?y?2N2~cJWf~8S{eTOi9|A?T212?XcEJoes{@A|t&Q
zVx!hquDI2JLWKlL8RvoZuX1ss9?PPpJ9FogtvU0)A>8yv4x^s@pH+ac!EXlh!u<yM
z01dL1?UIyPE=W+(!j0;YAia-ofx;nGYTFtJjvnd=Wr4WQt{;}p4rl+@C;XU8k=%{P
zH{k&9iVJ@9;NHnAVY@}2&9vTPn00$2U(%^ZZ+=e0Yg4yC#=AJ!#df%(3Kb!K?>IED
zzJWtWWGS)K;x~JDpiG<(JQE6Wc&P?I+qnm`e_LZ<&l1e{JBUl`$3i&slS`e6;U^vq
zz@efmnBS5kEFK|2(yuGvvQ<26X&&U>tdO9a1HG~AN-?g}sTOXZVuEgcfAB$~gD|;t
zE3PKY%{{Aw`*rgj?w4V1Ut=@ky*d>dRW-@qwG=!HJ;2H65iHVt%r)1#bGLte#n(5*
ziFsBxT>O*()u+1Q`_d-3HZLD{uly{K=nlcx=mU?B*Fmz!LvGc4BXUpu5U>BH8(-=L
zVaS_nV6|F?n2l-&ZTC;$`FAWYIX8)q7kz@Fic_&iPmhQ^^5tiRYoUx}6`T+$$FTS7
zA?k4u1ZO8edrK*=zxE5HkIBT$_|<|HF4e+2(<TXm3Rj@j;3xQ&rB6SFDPkkzH2sNj
z0@=Oy!9M0GYBDEhVs{*_OPY_Wm$wOSSlHre-Fj|dsVNzIU7EOdros#Pr6@hN2Y=4;
z<^<|)!UmSx|F`)E{*$PK^^BABV%-2gehdd)W0@P#>KfQ_sSs~qPDWP7!)2cVY?~9v
zFHTb@(I>KTy<sni$KA)x2@+Iujx_DMn~Je239wW;9JeX9;gq%8a6yM6u{Tzu)6F))
z&$DV&-r9|qh^XRr&XlMBNqoeZ)jlYy@fsCb4sXVb51jwzZ)g+L1Cmmvgge*`**!j-
z)a?N@@zAG_nI|>mnGwA-?mGI6d<1uPR6xUn=b$gaVi;7Hwr{9I|Dz(*p!Ewrw3*HA
za*9HCmrL-{s|FjbRH*%1DJn0P!6&tifV2CjVvz1knB^o+#IA(%CK0c&Jey^-95ktd
zJKG1o_lLvti*bJEK1gaXrXNSw3p{P=(brvqxB4=`?M%`oo7i5_<(UXQBW^~{8s=~-
z@3B0TlO%Jh)bfiqg~R=FDYC-f27{SD9<OWD#ar0>Le2$jf1A>k^a8(N&L_CB=_Z;O
zF2s*{%tbia5}sQOfn&5kXKxz=X7j)E9rAzBuh9VChGlWL79d`c_>Qhw7vQVT4UkG~
z$2}%epnkrYcL`m~FWE0o53SOsnUXSOw}&V(Z&rXwnnkeVP7Pz7n3K)YlC+%hT{Dgw
z(~hBT(E8FU2wawl`Wn|z5K@d!4E1SNeJB3Rnass~kfA@Ujp%<@bm$n%F<fS6FoxY_
zZj;}PL&whm`K719)<T2J-SYulB|^DBgs~ghGpkvl1|l>s;@YFvI8z-Nn)&n*PK--u
zp3kkq%Z0bO@tp|DRp!*;KQZzEf|0wbPc}tVquL%NQanE#z28*hDX%-6mu@jkzl3n<
z)g3f(Nq{w}?`+J!8&a(bFHrLk!}C$;@Yhp<+(>%}CzE7J&Q>qs#;4(Me0>}!RJ|3f
zIqih^^0MHq^h;j*Xdnb!smB|QSMX8237%EbCw{v-u%Rb~KhkGL-86P$YJnb^bmIxu
zd#~me9{Gid^Cjto87k<z<{yf%`-S}B?{NO~jhy}WM%v6e*?{~SQKC7*N+5PX8vT?r
zLG8^ha5!pArTW~!ri5i2R#;$Kyajr!Fein*5g;FP8Gd*hQ59J`=vn#?3qROlRD={*
zY8%q*`^?$0lfwD~UvbS{2{N`yla?$<11rzJ_=mISt!uA5oFcI0wAGuq(H<jllOwX9
z6{F>w9{99ImBc)J5B?W?uw<($Jk~CTG3zoo^SwPNttmvVsuC5nZgUfr-(cZV6XIbz
z6~g~b19k1wT;RI^-atj0wC~d=GrLOJXIYnSw>-=LPPq)+PcKk0e+?-cGoYExQFb-x
zklKF{@J+i5woD5IR~a$V(dZ3*r!7&Nu0f+-BWf0&$VFD!;))z6%>LWOYpSTwG7&E@
zGzkK!b*n*nbdXK8+%MsZ<DdA&R@a%2<ALyVuK>#b-o`!UPw??O_BT5JL9x^${IdH1
zPF>}Je++kUUzlUH<%bq#tX3htDk8Y1S`C!4AL13IV02;K)Kt3)zAfAYCEy7+dT5lO
z;e7ynIwpbL*7mrxPZdr-I0>2(ci5iPjD~IfjtTA0`O_kWSeFN|G+LQnh-^Tm69zQ0
zA_bl^{&U?rIhwEe2H$kPgsZRAsd%?IwGSqEW3@bWRF|WFCQd@B?TX~F>SJiq7Kh@F
zdWc0a(sTO&=au#Vi`Vy}?x<c&RT{uA7Jp&U5@`}uJB-oBlR4t2M)37ferf#ybTZT=
z)7#r1#qB!iev}~2aj)RN%2qV{auJtz-Uaj1tlKCtfaV}U)t~3{@i$e7>@I0Kt#mrN
z&r5@(BIe0k{21j_lfY}&O+2!2JjR`6{pMXS1#70qW4LD~H2s}~?=NcN7PAgqzbl%j
zc{NZqp50-eC&I~!=lq-7oAI0nJ0{2!qYEZO{b5nz9_Fv!Gj<DqEyELB)@E?q0+^qq
z^Dn#-*M^;cHRxge3b4(ugX6nc2RE`0H=b3(fF6Hrk5Q(dI6v^*x(BY#y9|?sx8d{V
z4ER>9096v7akl0x^s4!c`RS|C-C37bvgcHovl115W-SOyssyDbdHh$C0h5o3lS#dA
z_;a(L!aw>8(mOlgANc~aPpMPMGw<O=cs8!sDMu`qu0`)2Hrw_%i1)hRa5Gi?Ia#S0
z5Z_n|CAWKc>1&s9Dq|ZRsZ}OUdlcx*@K#=0(+G^lw&JawlQ@mhPG~5Xh+^v(akZ;d
zV8N#^VEs~x2>+9&jir*H;Px30RIyn?XEoF*-s4i4pHI0Ufw|S>=|x{TG%Ly$O63@m
zRxSxn@B9o0>T)1&e<W{zLk5q`7J|&wYM7D{im@*xNqlQRMw*R-;*u!AOuIrpd73_n
zyui4MleTl@;Sk2J%R}GNX5L6u6xMft1(mf&xScx(@oUvPTyclp8wHEu;+wtDa3}=y
z45Rq>Nxv}U=sV1-xPX^j%}J=p0kE$Rh7rN<P}nsU>>i(lUN#F+P-gEJ)$3g3g=qZU
z$m5_C^Xz`S1wL<kF~Bqrqt;CU@nBb+6Cy)Te)td0S%~l(q(;JtS$<IaNu6X@-GG4C
zF<`YsflN5Ji=Xv;5})J1_}4e|$U>z&aJC-;yAidhvBr>)iSu#yMJZ|$I>cBziLl@M
zGPf|;l`Bf0&51;fMwg!Vf}Z^WSpV%1G>Npr=}{8I{H7|rZ5OARao1qg19>V|9|#F{
z_XMtN4*&SB1a+=(#X{zY@iO*;37Wxhiqj|F`kq{4_k5g1y1;Hgj{6!n7Y`ii;p}>b
z_#-!cAdJ31{rxNxY3dKoA13hus+ZW@pdY8sZ{p+b%FuaJCqRm57%JR2$XS0K!l|=9
z^9OCk2ye%jff9#Mq}x|ezL{nA6BTKmg+0VKrSrWro`RL_FkIgoFG!X11F?jC=)Gwl
zm!chv=iFF*ZVJcWk`ZzRGqXAWMU^<RNS!oHlb|;gBC+n$I@Ad$0)_H__%d0Kt1OQM
zg{`kKLpy`t7~KO7jcO2hSB_RpzXI3(GbS5!)QIZ5J<vXUm{U^Bz=@xYG4--Go-Cv|
z<$fXVf22T7CVd2Z54Oj<b{raNnZNkYHu!s@Q1DSZmjCLv1E#Y)Ox1*8Z18<72xs5X
zk2g&M|2{jkbugoQPCkXUkFglYvI#qx2fSena~7@A6gr2EMGIYacGs-nJ_SYzzOQ@4
zl9(HDI6{;r?qug<)h{4BVIS+kvA!P5!c})0ko;}2czurwS$|BQ&Nz?=&Y%6!W?Knz
zECcVhY$|xVy#?c#x%ixW$=S9y;D3vsGj#4{?z^}mog-5TZ%><(Iqs)8-Pt=KbEF;k
z6dTc%m276a@;mk@pT=&_eukYJ!epF?Oa3(T-fL?4J0FU0TO;M?VGD3;%Y<eVuX1l5
zDw8g~Lb&<IgqmFX2YLnXaTDtfoGG{gRQfreGQSJTo;(C8kE^ht-WWsvMc}4NFWh-r
z3x0i=g+)z^@WIO=!8FX`_U~AaGiI_Lz+`b+nlDHE*BIi>{x%$O)Dv&Yy@M%DufY<c
z@#j-Bx=Wj7Ry;;=t%G8;{}9{xxh%k9>lBD<9E+t|h2UY=$Op_z!MdUyn7Z^j{B1VG
zue<N_>h4-p*0qa2kXHoP#N%N1tyr9OQkOhA@rtXPEJqxs<U_-;Mfh)yJXv{LoYoe$
zfoYE$-k}+wsAPwUnjRQ)dKO&GQ6#44a$w2cVf6C<jfQ^SI5#61XGPdzq#=9$9$*}L
zu~0Bu>;Q!iqEYDM$BBiQlWYB_az9=jgCj5I;lLnsh&O)4(Q~@_HTIeOYI!GEXQoN#
z%DQ7n(_u($mY|Q9e}m4za11(h4N^uO#_l82z}eFstQQA!^&U}z6|;|m@>5lMXG#VZ
zJgVmk(~4oTtv$p}Q>KoMnP~kl9)%;EV6E;`e!{$KkS`g8V@D~{*NX4(Ms+9@2UcQ0
zSxoMpRVTRX-=0HV-aELv;x^-du7ha-C7{KRMu8;TkM-<_X|}bTj)w=D_l)CAUWt;D
zUE*{Re;j8<)?npIFIc!Z5&q3M3z2Jn@bOx_P^RoSSoS1ggpDzcDon)CCHEoVM<XAz
zwir^ctb;2ne!w=^!aptw$9D^_3K9}pVSIZyTsS=mrOUS9nf3a#?P9qgg>Qs+l7_VV
zqdx5UX+Q=~r@=JFpV{Wn2qSiQLbHzyS#`k=xBq<yTV_0lhBp;p7^OugraZ*+&*$S@
zzY^A!)`3IC0k{qa@lKa9+oKHQ6vO3ECzl7JUnYas?O4pdUX4BRlhKIXV^&_M#cTG<
z&|u9UoVCvn*Oy*{agEBv@Lt!nva!=}+KP)fX6|fINtVa2t{Oy1NtXD<Si*+3NVrhA
zkgF-sC)pbf$z7$nj2kx!i<$eRc-a&5WW2CDpJO5I$V;%9a1fkEgoFJZX>zGK5nY~X
zlS6iG_^LVrCR`PzO35kUvQL{N#<1C3nhbT4C>305x{j3-Oo+)CdAi_OD|bj~k1*%i
zbhLXM0fUUwzH(DB7pPtiGv@0u_lq?3d-fSzdNM(4Ukc2dT4=NRw2-?s)C8t{HMB)Y
z)2b>h+It`te9p$=_YI>2)R1{xtm4qOcngY^4dAJekFdv8m26q5j1jtZpxgNeW<6O6
z3V#jA$d@8?xx6?jdi`DSyG9nR0xKZhEe5}yq}-9HSzK27dp>RZ7Vvuh0OqR6vfjmf
z5KMglCw4kw_!L>9>vt(9I&~^W&6|oRoOk3}Evx1_lKY@Q@d_{b^a1#2{l&%)5;UXp
z4g_=R#6=$=`F#=`b1egt)>C}AO*V@9WT0q&63awPg<DPus1<fIH@^N1?w0RH;k_i-
zEo|cUyh;V@r2pXU0(1J=XAvK}zCcj!AOhzP1;X}wPr%O*c@g!i+(!#}Qr@dbV@f5+
z`HvR-GB+;>td*zU7qoHY{AO68@fh6Mxi77u2M5=$$7v>Cg^VVR+8_6V*>r7`J)Q&8
zBk#h`rQK++D1n+I?qOz76pA=foY8EG8?!#Z^INLynW{tYvmIYDZ^|o-9mW!y0l~xX
z3-QV3Yl7hLChq1@5pwR<c95Pv4qHE@V@bIV{P*<?uf^C-bDi~I&F`&HxAG<2+j$4&
znr?$ny(#Pt>4R*k_popH6MUMJ$n89@L7la`@w|&C##Nldg7?K3_HQA#v-u~u^RIAn
zktdv2`HV`xB#GhvQaEZp2@@(=_q>y0o0u9IA1_KKO|OPcD}RDZxH+C%piBB4mFa6}
zf*%XwaKr>PPCq3K-tbA-9Q^{9nLUOFInyy$!43K^#>3(L;c)v-C+~2%9<^6>!#bnC
z+-bvqSh>3gx73%zs9|k#Wb$uV^EU#|`PJe1xvJEO@d+PXT#N6*Vj=v_RbFOj5IScX
zlU>mwbam(>$ogap@$-Jej*2CGWtgYU;>rOWTJ{vnj{o9UjoHFG%wGeGX0VQ$148Tv
zar#_Giw=b^f%|3KP|fl%EVw1h{kJe0?x)rB(;DL7&)Pk(p7jBD?A9c5=WEgWl>|hw
z8L2H_0A^ojquZgmFuUgsepIf9+*e98_NqAZge%hrOVntI%w5P+>jj7MP%v+N#Kj%c
z<b99I5XZ<sEZJj1j;_^*#}nScllSj(=H3t?<Ik#771I#D;_@3jS<(Tk^1neQFG-_i
zzjJrZWT>6EF??+Qz}s(3<Bp~qF!oLdSJpJlC*^u@N`(ciAEk(Io-_W9pd2>VFlIgT
z8mpF{LHWm8u=tz>TE{j(zcyp4NP41`{u5X_Gle^!rvfn(5#T>la-U0v#Imh$rs)8l
zA6fz*Dp>D!%pkwZ;Vop_SitkE&%kT98@mHF<2c;`D0w{+o5z0RO;$N#+>}n?vZDJ?
zXrl<)t_HMld^l%moey$3fpBv=%UoRy$JU+~P~u}oQnMOhd;EMn^yfRcOs|K`nUPp~
z^&VXF)uzi&`$0^e9C_xP4Te_kV8~um^C&g?tBCn7%!YYmr63r6;{eo%-{iZh<KR+H
z6i(SPfbwF>bidwN?p?uKl;cal;?4tbxcwHQx{^U4p29tGo{n>;=#j%4PYbHT-r-2*
zNR`}?0ts3_apK`dyxVqK5M7%CgPkldJ2e<b>PPdxMHojT!4A$&k{}}sCCKGG9V-3F
zn3k+chUP#j{IimAMBDRV)TMC#gYpP&{xJ{!gOdyqY;omef0&V;yE^n-$6-Fy@fW_c
z?B|tj8u6rA3|7k6z{wA4r0k&qiI8M7_L-@CxXE5TCos<yVf~Ab_bV_#Ap@5$G=Qz4
zPhkq3gz`bU<aZ<pXZWtfGj-bJ>MSD~T$+Vl3svaX?|)HfDnSRvHACR8chF+*1g-4^
zZ>7J5aC0#-#?FV6+#!akwe6fBtQ+W*HfXxk$z7bMM<<W-heOUsxV_Gv+$AgK2U%Z?
z8-%B9fOVI<S|-4tgCh7Ocj48YLFjX&5qHY#(qQp^ob0a4xA{ra6E~C5eTxRsbr2&(
zrR{im>tYC7t3{es2KghenU8{g;ceTtaenK!@->^2v92zFi)AigtK4`vm#0p$T$T&|
z-Y(_FC3x6eUS0xGCwIW8m>cM3c?muTh||vfk~Aa16s|2&B))gsp}BHB?B`$eC-5wP
z`G^vYOw5Lh>+&GKwHT(%)<BV#=K|k{!!VeX1&5~xAq`Ft1WaPxt0X<Lrz`=667@Kd
zLVxb#bYrZ(Ab`~PE0}P;7sPJ2;Popt;Gu0wFWAiEJzHz=mD)epD)9gWs*1F8oDLZw
z{uV5aFLUP@d&Mx;0g{ZQsaB*MG4}fcFHdx^j+Qd5x0I#!IwB<C^=i-wl0dW$#jc;V
z!Ytbz{OmptVRMBJ-9IiI7Jn3{S>wgY=E!Ftw#@}UPAY=5<eN~iz#T`#s8CDmJ1~dM
z{e8>c@QZq7gyUYQkhL#f;Ptufu;9KpJ&>Ly*s<FkB~Ok6uMyVVvT<z2!sa4I-d1qe
zc`WSu>k5LyQIL2#4exHe2Id>3@%Ex3K5^tjI32GBRu4x&@`hT_8~y_@R+@kJrV+NX
z9PE<oO7yRO5+?T<kvFU}Ur)ucRNqH9sCG#B!0tXSRyTlC`#$2W>#Q#@Iu54Z`YtRu
zSOR;rozYZy9v7^l@b{Y>4O>O|OxBhDa3caerhkNhxFWP*tR4x)0h}6V0rxExNcs^o
zGVV$hFI`&)*B>@;ce+CavDX$0r!%f@kGVPNnhfxErzlr@a++|q-a}j%Rx1>SBy*{g
z4M|VM5Z<j^z?~HulDVhaU`BU2NQ9fwCoXk-&^^X;+!2oBpomvGmJbE@mI>`QKF5HC
zhP)Mdh)=gE5r^Dc=;?HZ8~SSjT9f6eSJy%8Zu|!!znt;b6iGUy{uASViBSiSN=X0w
z73SHA(bVdE9B&rL2cA$M3+%t)5{ou&rE3+Ky*Hs#b^I}N`zZW8(2s*#en1^a<H~C%
zpgH*(?h>5`%84>W-n0^jU-*LY!>#<}akkj!JBkAZW%By76nWa(%g=21$~n7@gM8It
zKBm}&?#p|~7_MQEvO^uxMIYh)=m>6YR2Fkz<zugH2A+9ylq-GHh>x{QXz9%m?pApz
zs_r$RcdHKL?VQDenCy0#@jr^rI}pq74deFSTQ=>j>3!}~BJHJU50&=Np5G=yDneu=
z6lEu6yw82I5@n05D2h-hqoIu7`TgZ@UhngqbKlqX`B+PntzE~t5ypYc=hlc8UD~r(
zyBN?zyW~j1;u_q!{VjJmK9+0QJ_m#CD|kBk7#4iI#pcE`^dn<UzI$(vXB!xE=tZzV
zK3j)eKXM7g6UuqNi*Yy{u0n>=^vHtdc2o=eCh%F&%l>C6XViBU0#^9r*niI4m+!7P
zz5f;$F*gZ4lExtefIv%G3+VgY=kx}d&wLi+D2K@qt!Y8v8KprKA1uWn+xrkRRLx86
zw!q+{57~Dj3c^`7-0nsZtiP;9y>xxSDb5e>T#SH&$&7m!d=Gtp*zqm?lpE1E2j9o4
z@eyOv@n~oXF0l?0O1l<tugxBTu3S5qA4=q|MClWok!Iju7msJp-o*Js(j*~Gg#3(R
zdy;rI%Z*i_;W5mmQfETiKe4}GuRf-QI--4_IH~&27wZ4?<NCFZ>2nSlL(_^()H=%C
zz+0T5Z_fyt9n#D<MzOu{g${^Z*n{@7vV=KRQZUkI5g(tOB*d{dP)ep6Y}UA<adt06
zEJ)#>Jl_jyqsm}yqZ=%)w8uzEc~E)p2E8gdyyNI{XgI4$auajG-QfpJ{v!_#r*D9|
z(^o98%!d`rLohyk0t`4#1?w6QbT(<^=FRQ{%9w9uy3hHbYiyq}Q3h4|7D7&sKG}2n
zBkoh3jE7UxP<(X=icU8pGqqen>V`iW97yFRbN4ti>vU*s)x+=WU-G6yA~-@NQ8;$@
z9607AMZ))wg7i(_@YL69ki@bF7xh}PLCTO{ZT|<4EsVi!sgL0|cLoCTu7FG2emIwG
zNR}CQ!l5EV;>ox|8oSi!V&zbN+^tsR?Cm(s##d0czMiw6W(!vCFEID`esI4gNxDiu
zf!mFbpy`>*Es5{uIXxB9yfKmQu+o9d{Xuvt@*KbO2XnRLv9qzD7or^<P_?Qa&KS=J
z!#DNZlatmM$$8)<_ebc$xZ@UHQh4rk3b^i0=N54pIP>-`!BM$>uoAtAD;OKW;mS3f
zueKPbsPbT~^$V8wh*L|!574%=MXk@lz^84%bIm`wX`6qeky9_k4rf_-Oh1967#FIz
zu?~xcF>oS1SZJDAg{K*V+{E<{m%7j!w&>i2tnW7P^E<nv&A))*w^;8vB@bIJ-QzuV
z<cOk7Hte@*fb7&pFx?vqt+s!_EJ}hXj|zsdybZi!oF-_?f_Y6ISb+0Tw&P{`uTNhg
zZ_HB|u+wIGho`XWK{+%{Xa>#j_k61KHApNk=RH!EU=MrWo_O*cEOvE4SD`NP_bS4u
z18-1lN+z#sVZaz_QZ#U8KELIv2|f0kW$OmQG4BAo_nOI22N^kH&v+<G1#e*%|CsNZ
z;UhF|V$ZsV@^t!ENfIBo4&9>4p*hb95AVH=U0fXu$V-sw`O&P0FM!lgWqK-q5PvPr
zgvzczpb;oV%Ra~u%WW^=0%N4b?lGdG!%?8HNQ}0w{0mza7^N4;Rw4~8!pQfw5dSC%
zW`06A_&O4-rp;xX>M*z~=gEKGw;Oz9*Q4HJ74SYMOR`$y*|X#^&X!Uk*C+jfahb-X
z;^HyJa#bQic1P;h=zu_lN>m=X6O<Y6_IN0Bi2l8V9ertxnW9N|vvYHWdJEWoF2tB4
z`>`<YC)Tn1_L_&on7LS%9Q?78KR3mM-19C*W9!cXPit{n+AM&>UrPCNv9(Z={}@tc
zM5FmvEoyMhANOiTv3_(77v5hFA3{H1I(ye^OUsjj!X9iYspZaOYy)e3HF`N%p4J}S
z4_`GTXtJv+jZi&;;sM5FUdIf+-~1$gu`k5cEH~NLpTRx3?T%+}%VYm(U!k+kMA$b)
zAXs^*6jB%df}=0GvD>i-Q~2$$yH<yMSzH4fN0x%wsYh78YY<LId2ma|9K*og-~4DR
z37k2mUZ8c`4K^(Qz`1HE(dDiSxoaEF3vYLILj9F;uzqDgxMgopc7`7;yCeduc1MAe
z7!Os8HSq<@Kc<x|<>t?RiqmB>*gGhIJDJEl(@DEw4&$spSKW`3cAka5`XcmF#0NYh
zE5kUY`gC+`2Rxl7f)US@;8fyzY*XxkJemuO*KGoKu}Ez4p2eqi+~kU`GN(lCHvBS0
z3$-_XWV@_Ou>YPY%Sbn&neHzrsP)9@a2@8a3&i+Yzu|C@K1~-bL+9La(A4!ELk4?r
z59=d3o_5FYR!<<|Q6{f0Z9-j(MfqenfL-}2<j5p-vNLP|_nX{@sHcjgH?a*o9}>WW
z%;TY!%8~=0QR3(-uwSZ9b7Hel&sCS#ZB!)_Yz{#Gb2n&w8IMZdJ-EiG7?KQCh}x7u
z&hJDA_wchSXfyYWOjr;1`QIx35#s`n-mlF+U7<v$$;^fFHQqQEym%?&R@n7QoNnxx
z1y}A)L$mEdO!}t52l&d;K!acW>h^QOJF16q@4V;eeyf6UOC))h<uTBxk&D;Y_`}Uv
zj9s?xITzf1j5{dh4u(?<a61%o3m&>*<si@5J<ugn{bWgEY$o=N$$$~X6`(P|GC`BR
zanqO|TG@ehk58y_<}+-$u$zowlO;iVuDQaUvwlJo)sI+M!JMn=V$?3|9*jxJ#e-4y
zkQk{+FR(kY{{d~9wD=9!vpepcUu*}S*n{hm>-qCtH=yzBXJJZi5cZ22&=aF{VVhs-
z?1aDqESh=(8}-ysa$g)yT^og8ciBO+bQ8X{s^I%Fl;}Z!Kal>XN!6BphlS=+bj!sl
z+)&m--so*2HXMz_AIqOW%U?rk8`R013!nM6<FC05g3+LOH4wyAGGJfx2rTmY0CpEN
zz*XV~R!n@voG&>L%se(t6Lg7!vnY}4n~bBDTXK!>BVgp&0NALL1e+N@Zu6gQd`Ev0
z1|}CnaJd%g;Kk`Ub9u;KBTeR7?BY$D#HhLTL0CTf6CSrY&FAio#=?K&xTrb#e2{Pu
zBiK%VUuhwC-mMABzDSU_quSxwAr*-93gGO-YPh~`X?l%i;~grM=$@^>8_v1S2g*tk
ztNjE|ye@&PH6AEs5snW&xWMdJqPXLS232j{4;_z8$ji4UP_DHSg}<F~O`iejNM($(
zYgd?4W4xeR%^#PnzkwMlXVG($2zy3Nfy$BZ@PUXjS+pq{FYVLAg==~REo~k+zv?DV
z^am{e6N%yfOv%m_&p~P2K4{2d-Gi`7jE^1-Z5<kDll~R=g{;F%MO7H^LX++{P=Y-5
zMS_-3#_%(88z|pOg*odfzOk`~w;Kvz@7b9+bC*6fR$@Gqs64oLY6wRA=;FT`NovEK
zvf^Q;L_EoiteAgW*c4+8H=i>12g@!sGWT7n%VpSl_$hC%>HxX#8AC8qhh9^crM-@d
zxa5-|$?q#bwKMC4mJ>~&UcMi4M?dFgCp_Y9o5zDs-frG0SDMZoI|c*F%JJW;Do}ub
zyyK|n{PS!pZu`BRP^~V`-s7*}=+z{g^G6A`giYbxM=t`Au>GLx^&eJ!Z-&~(k-|93
z3`k2$71&*zgW=<D2<|zn63t6{S?6jhU*$7~v6Q^v&VT3eyv=qnUu1z5yYw;7I0@XG
z#HoD4EFRB(!gp&<<H-|Zu+3?V#oeJIylv-#g8{yvl2pOzF7{y0UzW>SJ`*<YZ4lhN
zbRDKz{NYEO6rrlaeZmHj|1dRvKBVgnq2xG!h|x-5{V&SqAZ)HRnMoyFZ*zr99nmH=
z1t&$dfvagx+IEZsi5f-HwDB|MxoeP-wc^-v*$pyG6o`XD2=ucvY>7z(UU$2MtB#72
zxM?!v_ISow-w*@I8PUQGD?M@fo)2tqB~O|I<w<yxE<LHR95+0QfW0&4L(C#seA?#3
zSM5#`COnZMT)ha@Ue^m7M@z%vmzs36eW-9$zYs&^^7xPDclp<a-vR3-iT<l<v|PvB
zB9R6((n5qZY3f7W^5t-2Wgsv4yc;EUHDYJ!ZVX{v%8Id5Ks@IktXNPDj*e%!zUg_`
zD6<cG?DXjFIXw_HRT4y^QsCe^cieH;m$?$X@uB4i8v1-5Y-v+L8&Of>m~scEAIst!
zJSzCr19#B!Wggt|Ex~VM+VlmRwex=hFnLu5jGbQ!R?)1d`0)>v?+8VW7xS@6#TznC
zc%b}YO)^(imB!C?fukOVM1G$M3~dL9jtqd;s-`slbPQJ^#xgOChqCUSF<M96#$DaA
z;5diP*>7!tF8e0DYQ+4AHKTAj{^A$zs7K3bcR22Q81@fN!V42S(?jm>1l`5s1=IX$
zarV0e{GcljAAkJ=S|tQ}F9be+b;WO4pRwe<D;Ju58Fm^<VC#gX!t+UA(^g#-;Vfqc
z;O+Uk<gJuCnX|$Kf>s%k4Z)EpS*Sy*Os7CX0-M9<y#nP)2E^-<DpBoa`x2r}?VhL*
z-+Ss%9h`-8a}cXeYSO+l$*8Za5A!YWbK}kzpn0Vvc{8&D<UhUxD~}*NG^9jxk`uW}
zZNXq{6b47ap5XO`=fKId9Mc<2sO*8ycyD|stWZhC4e3|7xL#{`TC$U~l8*sP#g}+<
z=q+5{dx0|=V+b+{x0wsx9+lqt3RXzO3!85r#3e_{K$x){UUrD`>K5CD84YXsqu+c{
zU=s#UzSOaM;sMw^8BrrA6#Yftpin!Pzau3@Rg@IT&DYBOlpc3}W0^McQv-OzsWEt>
z#|^ZWEW#-hWXbKfngqTd;i}KKa69~mV1H5wIFVndKjj6cH2&aB3ie|byatuWPr*_8
z4Lc7$1heA{u{Hb-hCTTKxHldCq)tFDF-x3aC66jcbcs<$9xr%cEKv0JL#s{$5_a|{
zKfG9+UUX&m#XvQ(u452=o;}AWGg3fuhAzvatI(On%n6jpdgUwRso8@@$TSM#7c=L?
zo2}Uxt<#6^+23MiSTL-RyaONB7elYB5BHU^V&Xn0!n5J){O1`7+&&LGw3CYBgV^50
z#)*B#C-ZTM<ZINu$<BJND`A1hFf^)7M~BK1h<~9%m19d$b6gaDNN22W*?;K1R);2>
zRK!_x6+rK&CUyKb8`d0_;6_zN!?zXf+z#K#@axt_aIv_BrN{qqUke!L^!ax<xzU(R
z(Ts!fiJ#ylJ;7<1Y0(2G#=zL%2WY?A6<cTze>=zrRu)!pEqhBL<Z3)hosJWfpZ<q(
z?m5_HCCR%xHt}%}9`SRt8(>${Q&dnL#v0Ru_&3i9y2e+-mihe{ytV{WW`{z7AeiOJ
z<WT>MIyy>R1)-=kO1FOE*TxQDbDlQ2>fi&HN1wr!xt+LOSryckC*r|YP3o0CA4Bzz
z0?96d@j(aih4Oejb!-VIv41~o7V!bEA5vKNz!FnA3F3aek2z@AJ2K)vj5%jOy5l1t
zSA)&FD>TXRB*t+5XF!8&G^oO6#_S#+0MglZ;1m=IesQJThl&y8?!6tzbE|~!_j0hZ
z`WoH|i^71aC&G5~16*;{56FJk!*Vk^<WcK8P<*NanUUe3c(@uzE$G6PlOs5jGm<1i
zr5a^duj37_tQNS{n~?XJ3j8_4Ziw>LBC1Ufaa)ZV-S?>umM?n`QCpSh_H=(BU9z+(
z)D7wAL=1lO4vu)J)2{O}R9aMxlqzY__JRuVm}J9++$)AzkFTL??F~Hc^*})O6mk7W
zqi{pLJOq+#7$?%sU3+*Ii`)vh$z4bJ8<Ssx*@17^*6o7XwOuH>HWi9b3pp3L4d93s
zFmDOP1NCBbZ$qE(0^6y~N-?4D)!cB!SrZbuu?~fadN?ISll}^D!|u7JaHiElSiy2Z
zB{J{$`Rj(6ci9KyV;QnVQk1lXW#R^QFUsDnN!$*2!NksLJi+n-uSP_{eG@q%$(Z4~
zr4f9^?ji8lqR%HXwtD+&MOt66DXm^!m7ckJ6}z1ufa$D27*0yX-i%-TsHkKR(>{Q!
z9@^nfaW(3F*a2(0ZU~#BJfMA-7Go4wa!$-uVsp6!B;OZ9+`i2yZG8vqXGCD_++o}t
z`~+WRoxvU3--7m0W6s&h8o!*=q2|Ap@TG(ziF(F%d#5L1Fw19no9=;%7p2(#-v{XW
zCP}#rWhydZ3%ZTi#O<ryC1^R>46>IhVexcL>i(@6W=wm|kNy^im-okW@k3|8VbwBW
zKt6?QjK?x*=^}h9@dT^nOz8SsHG-kJBQdClb$X2V!1HY9*%30#cXdU`^}Sn<i}DLG
zYJnQOQu_*mz+t{{n=c;P(vInj38f&fO{)z&;VSEbRavd##WDslpFfr^yz)tK>be@N
zb+F}UJ@|~%V*0@HWg&P^Rl&_oj2qau8MvvUe1<6j-DB;HZLdt{)=uTRUrmLRji%&O
z>RQ~<bQTByji4H@9l51Dnjl{y3sZW<@n@Y7o_?1n>xU+CX)a}WG-d!R9T#AoaSUz@
ztb`OdFJayLckq&Zzpn1_<IERU!Ae6v{zYXBev2ODkBE!1PVoTNu$kk<IikcOyi(vT
zYAqOkpp5wowds`FMKG#uBz9UH7FwGzZku@vOkmD?`jmBcPItgzdjOqD%V6bGNhoVH
zp-MUvvA%8u(MWp^pEdv-uaKkZ1wwv)v@AX4e*z0wUeKyJ8O7?@l5ejPvAQ8b#U0qs
zNfaJ?SD@nCK`dE%5gS`W(_|X{?~SeI9v(>qdoqzTeV{|`pEIBVzd0DI-^g1nEa%1)
zi;}Gd>(F_?3-0lHoQZA{7kxGc@&@GTo1xdR^?WC48GL~+o3>%+uXfN=UjphE9^zs{
z1yc4q8K&*AMOw?US-uK%L#qgBmuo}G!~eLNeX-o$Pl-^!nK_@OmV->Z50<p;#RE>|
zKm$Lc-1M`!@?Zsw`C9^^A7#nxta1$7Cjpy#hI!4#XP6x92iMzg^I>cc_2#ZD(L7QO
zeH{T@hjkP#u8<>_XX;U$<AMu9p5WE>?_sfg4Nlzejg>4v=QnB#HVAI>Vs;#e|9#IJ
zs(W)slPhrL{hMfd!Hfn^cgCW|OF}~v*84tJCscZp0so!I!2G#!f)|WGBxAP%B+rLn
zYSI(_{OAR^ik&YMZ9b#u*=mS4;JLBURou!OU(jHbBz@zvGTqkkHb2n*8>Xw(3VMbM
zaqbQoV%*6(I`yHre}gA3?0ANfi31qWw*@|JbApZoqF_691jw;0@P8Y2q@1kI!@ol>
z@T`Y1H8OmLi=J3;Q(w73rmY`NA0028o>j;g4n<&ppqLXlyYQz}n_zY4Rs4PxQM2O+
z$OJzY>{s;#)gMOu@!uEVC`wVEM^oUr+;4t!=Ud+DxdK&tDN5$1$&s<kHQ@Xy9-VuB
z^E*_P=&f&6aMr;cNthh}&X@ArialZ7?kAx0V*q`myTJF57ee(9e7?^bHdt^lJGK(+
zmS`b)ph{*BeZ#L6V&q7)Aq+CMtj@Qky!lWX-s{f788?lH(iuHasVjkRubKZwKajt{
zJM#aod%y-qHk+JOjyA6^ap%NGL*%%8!MI1xxH-^(oM<{P_!6u|ypwNnj~B<lcRP6!
zle8R8X|wNpOAAg`RVQ1kZ^F8zt?+L8BeugAp#g*Xz|Ap-TGnlk7xbb}`FkvW_8x|F
zi}4-NC+!+KFhO<{#ATV%KV=Au4VB4?CtpF5eV4A<9D&W6Ujge<QOQyk^Md<uS^gdj
zevc@x_=feqy^+|4@KOJSFm?Se+!Z&P-#K>~uD<dBCLcHhrsD@O{PYOYdxiCYC#llI
zUG8A{Ar#E5BH+o~?Jz%~ijPG(*xv9)SiB+?{HqP=k{cfU&mE0|1!1Y^G{*<VnYY1V
zamFg$EC4G*#{WBZ2NumNgxFK+<P~EAu`&;)oeIOU3pYUPf;N`tiJ*9SJ`99>hN?Pm
zXga?J3k`FiLG%khQki826vU{;DGh49X&9Y+qoD200raW3iXB?&H2kpzEXyq8LZiC5
zK^H~HJu(Btid3m<R}bu-CQD*&-$KRBc6{dx6VkFyz}Zt(I^OX(>|W+=(b)SAw)X6Y
zX@>VX?a}E%!%}5>MI;xF%zDms-A_rc-!#bY>-5JDQ~JPrdoMU0c!=#~zwui5ANI~M
zz-Ph~j5E0}e4#22uLVuGPg@i&Pd<wJq32N{VhvoBdI#55Gfw#-CAzSd;_rqz0-aL!
z?6BVeQ*&zYV2L5U#bx34d5HqAKdbn8Ry=kKwD8V`DBSTtkywn9<O^#I&>&hI?y*jY
z^SdwzPMw2lUbS$lwhvFXEyJPT+wsJy7Gd=yZS)z>e6{7({MgAlbVm+_o986x(jWsm
zaP9}}iu#T&cho@A|1+LD@&~f@PvOzM9egvJT^=2_Vg85zxQ|co@op>XAZ?l)8FD{?
z8iy6g)8qM&ozRT_Nq&r3Aqg!ts=}QYB}sU;D9idi<?ffP!m;E3a6!+*v2wd6O%rE(
zxIf>xXul*pG|e0Dr3z7O&vTr9$_>jHSG?fYT>RMinz<<=(OK07wndxa>Q#mq;Sq+y
zn>o1Ax(Oy;?1m}QI_&$SKogq$us}+hytC1vG6$8&%WMPMA+Ja)?oWZ>Fba2zrD%Z4
z0#pQD`g8GZjxdJ5{+)}sY>hN{rH4V}VKWl2;4j#=D+&r(mgd13Ihes5%x`C?5_L&w
z5_It<SCtk4UgRP8TnmBwBVXa`@DXt1R5@&B&S?q84V{o%iNO~{iTF}u3~w$I?%3G|
z5!-UoC*1Xel%p50X^|dTZGUd|9LvkNJggAzTZvGWg&i<6D;(FqUBW+mu@^*s|Avb*
zbTRhWPP{=v1R@s{saZG0zmx9cjEmp+B}HPSzu_DhFKH0y+S>40X;Z*q$O^4R8}VVR
zE-cO3foUgWF_Y}$JU+>grkBhw*dhrLvvP2>VGdfp+6tO}{e07{{m|L#0!bfhaQL7o
zPV#$;6J|W(Y>MOH(z_f?JD^Vvf9l44+vMq!9T!+v#Q>%@mE*Xee#Godn9}?Jw<<n|
z=6gQSa_B2Klz)du`)_k)+KOq%7P7ggtqvD5Zlmzp=Uc2NEJk;qlOYCj2l;frNm!B5
zjrxu+V68~GAj7kjyP2TCGO_OHo$&$#l!~C;K%b7gYy|1+l;Nw+Fq-H6<XlF-1F5}+
zAXSpgo>jYW-%FM!dAk|sv1du#!EjWJ>}K7TD;PRs5bvyb11}Vx;@i5Vm}hc}&2dxV
z0)${=`UmXIj)qUq3UF7;6VTnxoHm#J;YUdwZg-jl{#QoOuI%U75f;nu=`^O}rhkE1
zM>L50Luc;RX%n=od<GT5Yq<4(DKuo-bC1vKQdj1|GHqZPoNHH5lCj^gK?oy1Gk%$C
z7fxTugbtBY@XQ1~i2LxDj}eUqrH9YBDy1;|$>vaLCr%67cin?lgEZ7O%E!NFYdPQW
zFA!E8&N=Sc4udn2vE3*Il|IIDA(dP3<mdUQEHuY8hI=i(&5HuL?ilbJV9y7w?_kQu
z@<z%mA9=DD1XTtgEY0WMopJ-yrk#*C%9zYzobNK{=ir)BgKs`|aFYFuqdUz3eQ+Lh
zG<f6Ifydy}%)!li6FOO1f$Yy%f@W8=$eC0bvhjK@6jz(j>3yO2e^zZzsWyG%*39WJ
zMs{@A0Q|TWl;)vSCaAbS5!7XP$kJ1yWqT@N<+NUy{?ZGbZnJr4>N0F9Rwuzqico6S
zh}wZtq(y2eG{>3H)+7a5kn08BlhpC+-LY6LTaK%4Plv56JLKtaz@55!52aR~!Dj<6
zxZd7>sH^J_X>!$kU{MoqoqnD3k+Q+2Nvk-6t-dHG<H}`3U&m+aTJ+l1-`qUiN5U`f
zMX2A@V|+xG8MJ4$Sggo);8*-GA~W?5aeDPF+~zdKmWloaa!Qh9QiD7>mGKE0V?@ZS
zmC5*i?heNJ>w&Yn)6vqW9DR>5A1QM!pI^)v5nHUmesc(%5!d7=Ya5{Ere?nTP73$$
z?Qw8_Bty+v?{ViZci8?e0k-5_!WADr;E}yQ_!f2^wLK?751&ot);jpZ!^L7GGkv}A
zJL`Dtvm63V{Y!8=QwpwYNR!r|ov_u-8LiiULGyj~_^Uz+yt1?j?=4B<P9dUbB4~JA
z2jA-DxG~d^yjfv|yT|)-+iV}9<R&e0_g@>zPQ1;Xi)P%fycmAnhgqEBoX6azhgFcd
zZUi>fCJD|rH(>W_UXZe(7Ne{rh!%5jsc!iRqW^Bd+f{O8n(8yuA2o$fo~%Zq#Pnh8
zw+6g#CP{tS{AbL@-Pm5JLYgL+lBe(VNszT7m3=2mJk8HQVTmmYw9PsHug+L6F&13M
zGWPkn5p=JWJ`rol1wZP7J5`ENs>&km*T=W8#&00~=Bi=1y443|lH{pcG0WFJO~JaJ
zKhO~$k0s|tsiN*GK0@6A4aO-DjC1D~jp)Z?lE&nwM>)1us0w53nX^__oqRlJghG{i
zTz^~@9hK+dGnV<De$Rx&pL@n;UpKi=PK@K^9FD#hfAVExuD~vjZ@3^y9B)k5pav2i
z{I{zQxwkf!SSuw7%Aa3zD)S3DZjljDOuUC9HRZT%LLGqoKV0L%xscnl6?Togfxq5e
zhkHFf%=_JlO}~9{)!fH|`6`~A$&5IZNmM7>SD$4p9#s-P6oB2&bZLXR733^v!<Vnh
zV0B%+aHCN)m^LTDN&QBw@i(Q}dz1Oq;Z=}$)Caa5-+(97RN+!yFDS51kEOZ*7o8i%
zmFzPvQymZL<3eD(QVUO+k9;QMZdi{~#=M?l9GK;b^VzQUzY&u3qpBtj3%Ynk-hkTL
zNYk*VkGahk2Epu+C^0E~kEPUyA70PA^M`F__qq(gYX2mhc*hkJ?irE=^#`Hn<Wnf$
z`iyU@VS7RO?>MaX6Z_nZNl4{CET|EIJ)vqed2%TG-Y>x|nXkBrOTO@Sei*O0E<@lK
zHyIY~y}|unJrnQcg~7+&U*YCqO=?6`iQHu`&>EB?569||whsc_aAg&?j_`v*fe_qQ
zeS_x{^wBWjABx<#1iYz$4^h7Y*F|oz-j^P|Eha*5jgCtnlWL9T+KGaHH?MH9qNY$0
zPzqy1kKqiBE<7`D74ID&LI+KiaeuV|-TXQXEbJdb$RTTZ_OXWTLI&~q@DAZwJ{=d&
zFGQQI@4-85m^;a4OgAs*!1z6;WX6%>T*9|l=01_4m$#UaC*$+sz}*TIp0|Y6${nyW
za5c^>D8-i6ox%pM5!5|44~BdhuQH#<4UNn##kkwYx4hxMk5MAwlG5}XyI#2MFs0o?
zy;v0N0<xYtn44lm@>##>?<E}wlaAsS?3D&%X&n;2#+a<S!d%cU(GWvD;ME#qdgq@j
zzb^44|MQvyxIHi-3lGVoZH56&iufb^BYziU-2xad{ee(hN1ANXz6a^m6Ywc}&R0sv
zkjG8N)bYj*oUV2g0<B`8I3^O*zJG$n<L5(2?ns=`6%O*b-#N`Hd8)6W3VL3yP???v
zrPZH7bLkQE7`Gdy-oFPyf$P~B=qML5a1LmiC!A9;Bm19*!<3{EWU2-06(zo73=Dn9
z*{F@}ng8(a&Px!*zRTzA?{l^Po<O0LJ~hiUq`QC4g+oeJNS7IK=S?KYf055nyGMjB
z73ve0=wYxhG3NaivUiH7CKa3+WOL3GtQ1NRvDtUfm1QN94h_NH$z{xO?~f-W3@t>r
zF-F6s0n{lop_>nfL3U3Qrkr^WlErS^di!ArPItpAhjXEf55(*pYzF?bQQ$5T!x@fi
zzzrQPa9i^tykT4e4K-C-bw~qeejCT%w94Xwhr0P)LnF{&k~J<55)rJMBH(>=jmeA)
zTBQBmZCLUC7MQ!d!jU$y(6IX{ZjQQx^H{I<_t$PbeN~6p>@!P0uWm{M{EE21Cuum#
zAOwFYFB6E361A9jv;x9r8xzeX>O^%y5*{~H$Fm#ch}mu-mXwTwfY^@$`(uT$c9A8f
zT$dr!cS=%z_c(O=@IB4Ya6UMy{RhXc5Uzx|1!GrB;IZr{e8WBkZta9h&SMYrQM^?l
zF+5|LT6fM`6*2-g-ZZB2jEgnz<XJw^qZlPLrgBY;v7DM0giFRm;`XyV+%s4KZs$a4
z)87hicicQ~^YKhU)QMlH74Q^oqnmJ-`A2x#9|g_@?-<`fpX9XUqsxQS+@7WvXcnwS
zqQ`%NyhZB#hm02PyG}ZYPLIR<;tgEiO&MzP%N88Bn$k024VWMs3p<&QqHd)IP9mxB
zIyRi~i7T*JZwMsP^kDA?6SDTy8m#fUiK8Z&(S-IAIKNq*XgR+{y}WNAb+rtlMt{ZA
z5n^;jzCC9hHD2)7#~JPW=W*ZLrm|g#Gj`ekfY6<PVb<nNFy=Bl)BQb+2g<&Hb!9Df
z>y4m-5x*fTtr&mWn&PBytZO_b4mW9-(9vzGv@xkzu)HW2-n>32^qf$_ulh5BdRYC$
z0SzZEb5|Otv`LPZtzmwu!)5%jXP$WUz81Y%zZ3UO(;|`zMUZJ-4bI+$aMU3J-}wze
z%xxuJmice4da`%R_;k+l_Z0Zr=8GY=Nm%lw0P1R&!QQ|1*uRNATiKpq`7;B^c#{IX
zk*}c1Pk>90F}C((QF7~a6lXP_<=!4_=09oJ;nmY(d{&(?slVDNuy+cCuQ&WT%Wx&~
zCsrC(Yi0=lTcVHCB6NscM+&a8F2=-wEcCeX7os|iNp$K$IPj(ki#nL!aHlEW$}izF
zZx~>RKI3PH*uc}^dMI>~r;fu3(Aip!eXef!=1V4;L`CAv2oaLOSmeKVeq@fi%g`t8
z!~eGV$z2V$!=id`(4W9&!F#mu@)LcUz15gpKkp1$@#0h^h2?Q=O3>5&7hlC#anh3S
zu<pDhS)iE!hNB~K{@52_`RvU9&na}y1U2%{O`n>E%z+1SW6{VXhac4=M-rdjf}u^_
zP*ie=_h`<BTi)$hd}TU*F|x&nc03gBWgId#LmSicgPp?$g-%`iba-7T1|=)-dB4ly
znw>Sif7A!Q&e~8kwu9el#&Qw<&G2CE4)7LZyBirL8tR+Cx(;oSbSD!c|B7+T*Xz^4
zxi9%O<;ryaLMNR0bq2p~3=01!_kcnBKIl5u&Zj+|0Cyr|`6{nc*brrpC)+Rar^Y-+
zvk9hjmBB4=xT^%q1Un&X-XI8tWBAk0*5dE+J=|G|aPVon%=3S{VRU9UI+==*+i(yn
zs#&jct}Z!vp@ef`pF{OOW6t=51fB2C!*eV4cSuqtPmuK_c0Pf)TvNy|mZXPv)&R2J
zLi5C6kTMMeeWl0zvfem&`B#N*;NCJ8-c3Gn-6i;9eH!!kY7^(>E-)DR9s_-cA&lkq
zOQuC|DlY=j=d%WpYEr|@8z*s!<^c2T#)71@DV1!}r?u7kWGt@7fKg$1GmkkC9=4<F
z)?mSaw=7MtTZJ_SSE1DB0B=zwM=V))#c)?6Ke^{2_xg_`-VbfV*mQeN(ftNQODu;~
zd2g|*BZv=<_5sB=$q-)Ch9ZCOgY`HuG9WEM98xo&Vn-RYOiP0%pI+Qzq(ySo!f{K<
zWd7a646w+ofn%ftjH(`@1Dmm`%sm7Du{&Dz&oH=k>k`cJ>){IPDupR3e+4BfhNMbW
z32HB1V$KQ;a%pP?rYd`5>Y{Y4W-Q-3MK+*$Cof&H{~)yVtHGO`F9O$Ib-MOdEZ@)W
zf!ZgO=(sNh_+Ow7c{jOi_M+pF>4z38&|~Nh%`BIfC}%`Y?R^P#9EaYt5)HMasJfFh
ziMaa~B1<l!XKy(s51R6h(PgN>&XS+>bcOp(|6)q(Tkt6}qA9OLNz<9@Fvji(XU69X
z>^5BijYBTL_YOj7ZU;A_M2}n_uTDkZuENl)s~Bo`1#Ja77K3%;Ky(yxr_)Yz=hzG}
zwX~P}{b3z^QSm@&=OFafVr)L;B&a*8jL%rcyZo*l*c}^!Wd$L4?^&5(>1aLt7IPcY
z<ac92Mk}Zk#KKSJ*9v63oy6d5)L&c*J5)4@Sp5SOeL?YD&kY>=UX~mV79qn=J7H##
zGL+>gQ2o6x(Bh6f-aEIOzxjpjn=($qRhdN8yiv%<CjG;<`c53_$<9GxUff(;IsSaJ
z8R}Ha(F2xdoXBqGoR`T(vFoq7ZStKE&=H1G5m&fv7e9l_P7|~d`^KfZOoBCDdZbM(
z39K5u(JNybth`){w!tyjFKECQ`dhJMK%AR#Ck4(sh!BJI@|feMOB2T@V91^*z>FwZ
zRA5LoYjsJG+Z_zAXvYWx#6XuKxZ)B6U-oJ7G@t{Ti`2mUMhphceuH^dmtf9G0}`Cz
zj3-0btTLwy`ju9~^6j~3Q#KjbI*!1ln?B>k|N3A9<APcD_i(N6^~ikh037(>&3yEM
z=%CpN2@Qv#YlH$Ve<(&K?e4?;9p328vOmM=&Tvi175t|%Pvo?>&=w*^r<+A{{fxmf
zL0ppfofM-NJB*2Yz6fp0OM~D3m!WXpHc(VghMmF|`1t5HdKMSLEelDQC^(6ptV3p*
zHHZfDw&LB<l5|vnBn>^HOyqxxl11|rsa~ZFu8vTnDh+Q0f{A}%Jj>{)k7sVUSH;|z
z+(w++X8?Lps${eEfMBIj4>!q62L21(fZ2y{;M_BApnmlc_wtJr&A+o5mc3G^$<zQ#
zCK-|K3opT~&&$APQVt%d*#j>fG-&e1TkP{1<cyw<<FYkH>6!j)VIZ41U2&SpMVc=J
zxASG7?|+k%DN&&Z4@=WvyF%QfXaL_^hVe_5CdhWEptMp67D~CnbVV~^<=H*_MIy!A
zfrDt7BnLfDB^dL_nHx1xpC9i#4?F%n!m!kfpj`i*TX1V0%zg3+w>2A(N!|)1>6bgM
z&}zVv4-sfAcn1eZ_j3ETTmhNpCOEi-!#4_|WDny{sN1;XHdzNu{&SRN3Jl1p=r!=<
zg*v8e(jyJanFl912uBZv!&s&H@bu9ynA3J#5Vr6Hx75fNZm$2y9TwYwTdrmZM(MrA
zpT}Ha{Z|p9_Ck~FUl5LGUBt+<G06z8mhobv8S7hLnHVOqJzC%gh)vaD_l&O)Ay$OJ
z(&}_+q8hb(n9g_qSBHj&wa6taX}GUwM58?l8Luo`kY9R|J)<<qqxy%OYUmF9TWCZ)
zb-cK3M3Pi=sgZ-<I4s})4}6>5;oPJ({PbsExk)S=<GFYuHgU|weDD!2xHJn6KUacM
zn`UsTP$E;`t%mK)|MBOGJ%q1Kf%NZM<h<%9+;C2xDqoT&nc5npQ*<6U=sM$SCnYXx
z>umIzu0xv6MnQ0>4wv$t7nX1nVaNOgbYu7NrwKvOd8`e@eY27Gx{c~lB>l?9RFF)+
z#xui-u$k@Rqc^%DjwuIQ!4lSeRwmT$I^L9v!^h{lV7iDhUAbP0NSrVwFWcvGGZS;*
zxC_ftbe_j(Hj}e<SPUKPJk_8V2Mb<3g-Tz5_fpO%ycUD6$$U6AQJ&7&ua|z&ESrnJ
z>xhYeEn)rlLvX_H3O_s(;ZwyojQ8?{)S5*8P_F^KYubb|k51rt*7vGAKN|0B9|Ov>
z9^t_9<2a$m0*g*u6L`Pq!*n+Dc-@yTFt2Ah=+n#%w@ZdfHFkja3se51wGOSA-HFp5
z>Jsx=-f(Nk6W^p65$6DLa>ghWY#$nv`=X5PRHQ|>Gq#zD+d9sCw*#lgX1-~h8ExqC
zfH$KX;QX^4{E4S6Xjm4>uT)gwh7QdX{QZ)HbG<m2B>M^ke^l^uvnEZ+ZQy*el@PB~
zLiY4z{&W5b{=dyKyh_|J3%Pm9v>{!TJHnVI;;B7wsOmBv|4)a?9+?aMzkjehktxkc
zFo3^7T4dryaZ(?nO=H%ZlBKhMLVVb1ICMsujC_>{3kD77Sb2M1L0g5UG}%J%;!b{W
z%w0USNslg{6C&t~X@az!fhaRqhb!$L!k3*<knhxHk?cQ{AG!Pwh#gvt2OA3^sP+MG
zeIUxhW+B@FYFa|6%mBR2oC&oPy?OU<p%8jUlBU<^Ly^7<s+cC=zXiki{F5qaZ@dGk
zB#5)RDnXJ$(=c#oEz3=B!6kP0+3%4I66RAtv~O5gb~1;1{5}+dr%u3A-;Ak$>QO;^
z!6~Rr3jsxTK6eWH$r;Rl2hJ09;O$OTx~gUl4oDGv9x(}Btc?Ih8<R1%<?!8R9lU<a
zd^XpHF|R|Hx8VmsWRwA|*l5SPN}mM2QxkCWo_@F%m=ELZH}h*Yu<pbZF9_VNLLQ&{
z2--*b@RHpRKD_TM8cdW+PdpHf$piH;`|(Qnc4ZOURl37F+gC96yB<v!=#gnlC4@UV
z9l&m}2AHf6C4FZ`(91iIVA#$|#$}=WAQ2^}rt%zXR|{=eCr79}4St1s@Dk?YRHo7e
zyC&RXjH43Jiqhg{PHjQu?h;ropNo&b`r<_9ZdE-a2``OSfYRX;*!@_ITK%#>wex3$
z3)l96+_({>%jXizmde9nrFHzIJtBO<Z6WSRt%CzcqT#;JcV591gtOeFX@zheJh`06
zxUr$|IMt9?-`Av~S#mUGuL7Zq`jMA^&e<eCgyf2TezBDuR+}5~S4>QWh3VD!>scMP
zKl%V^vKt|MxEI<3MTp*2Q}Sj;Ic|G!6HM=TV!w7NW=+?p3!k}R*O?A({G&$vyfK6I
z*$0GiXY#pQ8Sz|_+XUQ39}2T_L$Uq60@3=f6joe{#x=In`H-C1Y3KJ?^4$w(;g?)B
zi`U;x=}Qd-cDHB&Ur?rTEuv(M41n~WxhVek4urayl1D3iKr~mCxJKNB^kYr%spvU6
zUJl@QF001}hvp)$*TwPWlJtSII?;&#CD=Lg1XwN+A@kbL31S(GwO3uAEE{^w<@JVg
zqvAJ1(GF>xwZ9r?d$kG{?pq3_(l-G2)x%v|6;e0-7tUX31~2zy^V`|I>w1d_y`9_#
zq+S*3NeCt#inidd=?QMEUCzI|Siqm(_7T&AFJtUMLo#`?0a0%_%>7s-Mm~S>7gSjn
zb2Y5LVflv_=+?{Ne~d%5Po*8f&=u3KDA3q|``n~mwQy+kbJ%OIL&SH;Vd{xe!ISga
z<lcoN{LY0U^jLu^5&p{ok2B@GmeYMubdSY}7oI`Qg8R^u@DPqpP2mqsxeV*)*n)zg
z5fy(K#h-*gh<TNR+74^@s+qUpXO1p?BxOKF+<Lea{0nz}gu<#j>crfv65`=6-2B6w
zQ4be_if+81!j8>}7k+{WMJZb8;>#zRKfqG45#Z-B3K$6#{*fY7T^$VK_w9wXf$DUj
zPcO`x6Ala49EH-U$IvlEjA-ragX=*?gzIFT#ep`?@WC=}XURl(CG!pTJ1LPj;xXK<
zEgHmg&wJ>c=*V)Cid3@2mV3ZhiFvRVG8J^F-aB@VDoWtIE`;Ff4Lw{RXG9CU{_^$t
zZNk%x9c6R%H$O986i=wK9OQ3hy8cZZoPANlAN^gy52RS)6}pL&`1BvRO}GwvW4&P3
z$HQ<{Tn-*AI*4+aA0VPYi5lfAaHoZysM?b$tiDEh^iU#&BXY1J_98dlT^ZYYO5xDy
zCf>#7E5v{Gg7~F>u=cVhwRrj(M4InGZ-OqBo05k%4Xz-%$%MMyu0g-7Og?TyI8-T*
z=8ruNMMo|fKCu7&zFfw{_^82lq@Lh#EFC{FMy>whD0sZ449^HuXh(c87!fnt9b*a$
zytFvaoN?%@TaWp<7a%NElw97eOIpV+!(4MSYTJQ=&!Wcs_YJkYtL!Z<tSyS2!z#EB
zdGBG0AQ>i$m4I1x7S`%{;^zshqZ1~?r&p!Pzol2v$;6*8tjI`TmtaJ#jg_fV)dwyi
z(+(xCGj~qhQ7lZzghP69ywr9Ts2Lc~<vHB7sE9Fxr15%0!K+W$K357#*S+MnPZOg(
zrfPISWid_;H>EYQ&rs}=8-H|ZKQ3st;|=|FXzqX-&XD>m6g_>IvkH-bF7uTzp<}I}
z-taE>!+bXweNE-!gQdx+Aq~253hSI5%;oon&f#)8W6;-BAKLN*QN8xD@bmO?Zgj!~
z{8$jgwV4JW|2qv%2ed+B>1V<6TdC;1;St|*b_BiPQj1E5wE6CV64?JLowsw<1IZHw
zyh@iBy633U0;5RG?oy!lKFYD~wJHvMuM+GXtOZ(D3vQYLm>K*CmayORjd>yFk6~v@
zFH?+j2*EwQ%+0&@5EhNdhs|4z$^OUA@bB_P-nBx7tjR@Qu5c>oKJ~}c>x}K}a~7WO
zmn26uc5ri@ck()W@1-v|P{DoDO2ZNA3HYuwp36TL3^P~;Fn8uYaJwIYu>%*lO2&Rr
z4toWe2|Q+xF+eHPLA>ir_yGqUK5&vEtt~u><NVISm~D;NM=KD<E0C|h8CP}TDcspD
z59-&v*&N?R=rK`&oNT<!O<wB8&s5cdY;Tr@UaC(Va}4SIU8*$mY9yw1j^KwfmheWO
zo(a0-O1a!WJ@{>7Fi7t;qiU8F{4R+Gye{Jn_w66xm7Yd^%7ack)vU?SOufKcnwZfL
z1-APVpACDSN5am*NPJdzi?jUD15f75km41RXmd{+VjiF30xZ&SYT6<&SoI!5Or^-K
z@`G5)W@XEN74uCu7_a($AS4~xjZ;_61jVoz5L6h^C8ZuHnec^|J+4If*OA=MxoY4#
zoC3=K#RwnoS;Ln+W-iCP#R6d{+Y|27K!^BeIMrVjj`JSEQteK#5YFQLBhA56NfuPT
zsl)7fdbFsc2)sle^Do4I^RZ#O@cNsxjQeUtu7>(T)8Qx_T%b!7Q+R0E-Ui)8qU2D|
z171w@Eox+pAbFNGFlu%`Zr<s}iA8uITCu)JU60_wcXtfhE<;un9EH&e>XcO?;L{Na
zx;r8RC#o!fg&lzyV&R1z&mY6SjE`yO$FaHJDl_4Ou2l4#BS%~HUf}Gzef-zJVo+<7
z<jid*LcdcL9PkbT;k@UZlIupkP2~}|y8VXcmuj@(i8c;gY=wzRPe9pXG~OLQ5-VGk
z$at5h+(tPqdah;=F1{?}@43u^u_?ajKH~`IBHY3IWnSb$;`B*^geh&&IfZ8S*M*m^
zE5N($Qy|>q93H(H1|kj((7&}Bw`-i^zs4AG`MRr7O}mZFc$G=xlMY_?{8>1Hy0jq7
z8DtnYL%COiI%<@H{lqLBuh5E%)s0D!oHOq`+mJV69e|^bdm&?S2yfXjnO~E{_`x}=
zx#0coSi62Au6}$G+X~p(tm8U*-Q10b{v|=p5^>_2^pyK4!8muJQ~BWAbMVY56IdM{
ziOKg~aRxJr@Rc7Ak3_2Z3o|Ey<K%Mq)765_VQsK)^>!?MF&P}%f-r-1g0{sv3F+=E
zer4f$sN55Yx7FSt+3$!;GA^R&q7bkn(vZkz?lqOWEbh(p$KYG8VD*S1b}$dGX&~M>
z!`Q4RrlVw6AM9H-g?R>caH?A(@v>SvUh&Vtm`|HAxmBLqt!O|B<XUlIf*#$c8^vwe
z^N2sV^&QmK)PRhg65Da;lDol4{EU$Yz;pj<{If!rdi!^y)$thEy}TIomo33pbLuhf
zhzwP6QK!FEU2(WFo!eS^9k;A|2^ak?;l&spPlcWXnc-3xZ}t$!O^b(`3`eXUQX!`e
z^=Q-yb(Hjb07-r~aP2#mHM_hQBd#l;a_bf5^j2bXLL(eA)&q7;k|W24AEEz+{fw=y
zKp*_;hq&XT`TL`9Ls^6jEHSykbv_Q}R>tjuk6a7vsd5J&&0BD()f8rltmc&q-r(%x
zzUa2;1)l*|Ae7xnJZ|j3tY~p+wFe=%W;1TSqJ{36UC`35MRYtZ;7p#AAigXD4;=c0
z$@aTo<3fGfc858+ZDq)m@rK0BZaVjVMKNr77mcIh2Qf$91~<};ymzn-q%HHn-w~a7
zJRnn0XB#Bw+j|RoGUMU<WEqnELY7#DmO;_f5yZ5VaYWU-@Z6kASgIhKK2bgpRvXFC
zn{u-0OL|nemsZbl%3zws3qd-B{`$wgtTdv-{d3_-?i3h$+lT}67{{tR5e#Ef!9w*b
z&azaczkeu`$A3KVqOmpOBhJKPDdytPR3S@F<Z@*{jc8Nl2Y%K}E%GNP8!Q8K$)HvM
zeBJBBZ)#owKl_yN^D2E(Q!PX8?pp-^k>xn=c{eXEZ%8UEW<&2ZBYMUoA0~dU;H$?r
z@#<~YFjsUXpRr^&+;Q3hpZ70-r&Stc_TRfuaCRlOg&5FFo7CvXZ9e>!+Nb>NEn4Jp
zKrlZFKjFz*DeT!j4GJCQ!6QbEYVP;~Bd1QmUk|=;zuyjnJkMDDF%R&UM+|3i?IqYR
z*QJi86h1A8;nf|};o;geJd)A@!WjmFiCOj7MQWh>;y(2Jg}7vr0FG*kQQ6iAcwSNo
z*ISH8#5yHj<*Oz&Y15=nRP&H;SkG_HI*qkj%kab4X?Ut32K8+<$R;%-a=oDn%uZI~
zwuRI1?C@_C*8hW#)^b!WvXMWyI0);!Uhuw*b1T*sK4T5*emdUw=0-J2!@<XHunRP>
zslr6~xO$_Y>x?n<nH+<L7zx_kDn+k6%)#jeskljt?X}-KL*?`;++ODbYMCAQz3>sP
z0|hG5@5kvDg~Q1HE{NUv0i%+K(Byn8b~zk^IT6g0Y|#OWK3>LSMIvNoD9fm+jDZVS
z3lX;-@(YWfF}I8+z2xVG4jU)nfA2HFX4XwEj=4a^Gp~Sym;y~6;e?F_-5}6j#k?h*
za7;^-Zhr2G_Qyw1&b|W`PZ)EulM5jE_bUi69zwU8Fw8!66*Zg2!@ZoBIOz=KY$BL%
zHC&!PKB_<ijwo2nGf<*}j2E2$yxUOG9fX#{DV%<qJDjRs&r5Ccg0`j7<nW_wpr9y1
zJu?z<I_o91IUfbJBrV9@k}4eiONO|bTVd0YSFkvlId(iv$PdXN*fi%4XR@;%H%_X6
z_yzLRd`%jT96ZDKv~)w`%wRbF!UGq~mmtmeKk?E<nQ;7YBL)n+b1x<+lHaGU;lbVJ
zD6H|uqu-BV)CO(hHt-N{OL|~R#0yYlXY_}bqjAB>TC5wTNo=)3VFP2NXhJlG-Zde0
z&Ykc+(i49+F_*({QL=sD0k_J~h;(OC+;OoANA`E4V~i*-{=Sv#N)jWFZy93Y-BwJg
zZiQ&|LhPSoLdPZqVd{Zk2wC9_=KHnj{rp^LwyNV~1dLh6cFVTcqItDmX=1Qtxu9*|
zayZsy19zp1P)tIWly-K*2t7smZCM0kntS39<IcQ!=m$l8H*nsgc&_xH4;;EM8iG}l
zF-=tae=MDMTu*P<$I~7f+FOH^k&OD@=iJ8(NkW;~lu<-v{me8eO)IHHgOo_)+~-^;
zqKqV>C=FR9Dx<U%p7Xq($3OnzkFS^R?|t3Z=ktE^_m7p)0h<!Ba*61q2x|n{Wi<ap
z^vzWLX#$7VfB4FrWPFMXrG}UWQIEQ2;G5!^I>K=c&Tv>j=T|1e*5EGqimv19V()0M
zpcWf#pUT~hamEH)%^8yCfV1fe-tez1^Qcor?CF>bF4C@0eo{`b+}8>IzM^+Q%mvph
ziN}XGS74tDJ0Up#3Eo*{3agXEoXp4%AU~=dLN*`5(koeRPVxYD+jI)o+MEYbKn;s}
z#N3hUDv0;Fja&LnvCE=!_(6gS^I-Ks$Xmg|yI>XO$Q1#v{m0`2--if~ceK()ihH<I
zq91nWu@*QoZwoew@5KiP-eJ?%YC=@Kyx<l+8)vU5fupnDiad}V*kz0^vrzjGtUnsb
zyX-Jwrkyhqwl$du@gsimy1BsZYV*hMCnobPQCh;UeSJJLIR(z1cjFu94P}aDFMuCN
zF?)75i}_az?(vv%+_W_oyGJzB-d|tf-DmpnVCFJd9&sHyZfFWzm?=}({Sv~Tw?SJN
z2WLNB#14@n3p7OJwv?t}<5|YS7rAmgTGttzH80~&_JLsZ=>o5o4WPB~4*D`P2tRnW
zP;{x;;#mVFu;x?%q~&yg<BHX=D<g*gibd{J-ebOGMG}r*ugBc{dj~&iF=h<ky}^SQ
z9^fk@h6-ubc65M$8$Oo*3C1P;#eDb!TvKxoXY8zm7pKqR`uYf*VV(|Nwn6aW?owP>
zD9^k$JcB<5YB2PHp^Qg*0<3tJ#_OEQ=08U$!RY+!{9TViKJ$b-Txhro74F-i@{=Yx
zIx-2%xD-Q&|1$75<Y7YT1Nhv(0rm%&LhRE&Twja0lkE!RJ>yqVrz0QZ*xF#cL06Tz
zr<Dmp@;taapv9QfX)zss<<MFE6+9aY`T4t-aP+k;FjUNPAF2w$rm9QfL!ccV+3<^=
zy7Ct{LbV;Zo1u86#zp?bnrV=qK8R1wJO#39>9}m<0oumPh*8%x6Q&)w45PQa24kxb
zxa2>K=~*Ijt_s&fV90Hpx88&?3D;zb*Riz!6$9AicMlGih&^^=V<sW(GhXd*5cjsI
zF`=R(MJ_Q6Z0=ex-Dh9IoV{7_@dp9#{Y7}QauuBDG-Xu%w(>cFA$abAVa(1wSFm;A
zWoQ$7l|}qJxSt`*w2Z63(|TnfVE0^nx<<TL&)N_D7PUCyztf-{KLVz{`$?BNw?T96
zP~noRlCZ|dOqiANjpUo?G0*R)F!_`L<7>GW%!aY}tgX5**S;Rs{^LZ??tVP>f%rW^
zx8fsf;;>KOLmazZ7Fr@S1ovutTzJS>&|YpXM6dsaGt_*bYWOqgejv}d6&GTs|D0*$
z)rOnzXTzn42|x(XH!Rm<+GqXZ2lty`nUZR3yKyR?Dfx?^**(R<FSEhl+yREIJcsKp
z^LVM)gG-i(r{%V1*iCK|yuL|;^tvECZCf9$yJs2p_rHdv+TYQ8Bd_5xnV0zR9%;rc
zycHjN_69f2_TdlxI>xVA`he3E+4{~C#ID_2F;lLlBrMer#LG|+=&O|RGgM0W`(-);
zeKH0!LFPi^ZJG~B0!aBc3Jd+RjFH0-z$48l+fX@Xgr}S!B@@p5*6zjx-cv~TC%hlH
z1BcvyLaEtg`k%%`2$9^wL050$rmCNi`y&eUVr%(v37-6tjz4%$vJ;H0?*q&I=fP;j
zFyUnJUp#Bx6C51!9FAAC;hM3ZAidxp#0{!3?K<-zWk(x4Z9EG`jkmyP*;$I!m!_k)
zq(iacHt;_j3&(#BfsOL*`0?K1xJzUom41<7ax*%>HSrzYBXYlzi=>&~BGW!R^B(S7
zUdPXiQpK`o1iJ6RJ{ZtGLchOL!+&e2!L{uXp!+Bk;;Qe!i!O6!_<><!hxsc0a#uxI
zZM6Zf6yFV_tx~W!MZwgFFK`QvU=z(-n4R_&Hfw&sDtd!_fA%(!$*#uu<l4g>uhI1K
zSMmHcu@n5NQyGqSt_Nc=TUp@n1D0iefVs|JV82y7{`f>ma2#nQaBAXgn)#16^tp_O
zy0r1rr=5WGjbevMJ{Z64T0qY>Er48`&(Iut8+NpbJ?t*P$J(VBFY*(d?VG`UOeLm1
zstXBgrG#er0b15429_0^gxzu~LgL}$IO$OmtWxn6J9mSSBN#)|`X%%Pzb~}@`Tua<
zj3EBeBm-gAPANF5_XFS0zKWL*cfnJsAx!VQX8cLk1Djh~F#qB$1#R*G;;&rdifCP-
z?|m=$`mceXcRgX~1uvZ19Se2oa<I=(OSsth1}|kd;A2w_@%NaUBI9ci=hzuAzocYg
zro{^QEZ59mie#Zzun>wEkuf@<4O_2u#M0d&pHMRi56zE<#moQTGo=~~4yfiGf?L5q
zH6N=4(sb}}BVonJO*HfQBpmQp6>_fi!uxg|n3&7rovAnRv3YO#dNn<vCA9~)oRJkO
z{YAI8(k-Z-F$m$6m*~PqJz;Z74)*L=iC>Ak?~CUH@y!ccVeh0?t_lrjD%+!RNSQW6
zJ_JLAj<G;Bn=wr<bQq_MAbzb}CVx>yyeAK63LjsD@jw4<fQic#A;#?&JU`!r>-NpV
zuKm}zTRq?Lt2rX?VkzLyzjcIxV~@D4o{8MrS{i@(JQfPgR>DIkec_L?h3LZoShe>(
zm~T>I_H6tCj}H~#!v`#d`2H^H#&{{gFGlP~WOmZ_OFx0yzQeexF#`u3`UOi@9)@d)
z<zS_2B8<g~OnqEA9{1B+xV%?OPy-s<9So-}FI>O|-r1mdMGE(dy;mFmhq$8lGB~c*
z6wGfMFqtRRgiEXB`LPL`;BQJLt$0nA9;Q-5Io~kE3qSsZD@pzMR62*Jwk?N$L+#;P
zRyp>a;E11%S7E-qpzy!NrQlg;$t1#1=09GES=8eJrBCjIa_U@u`xZ6)K;ME1)m3MD
zOw5_k(;7nYTU%;exHNX(B#Z5)-sVfEB%#@K1gL%Z2Zuh%3PI8;%&|iv7x4NPTv(?n
z)U(|%ciu04-t0Xh2k0RFc_b3lx9SLZ=UrMSD1Z*X9R?w5+OU$kqOi<vDc&S<PSpbb
z;Gy;gLTgMnuF?1e!>=XsJ{hS{;-Jmgq#lMR`z(Ysx2I75Oc|DIJJA~t9^+@Z)x*2-
zjl6WuJy3phmX--T&7TesId!E^u|m40kiFp@c$zPU*}CyK{$d~Q_ueP+e?^~|$}vo=
ztnk|O6_A-4%j>Cr<FEV9gdP8V;pe}a2$RTu&i8IFtk|{#G>q?JWY>>#irPSCQx_i9
z^MUr9_J*ph;bGnHDrn!Q$eeh)2d{W5BNU8J$7WJIBwxRWT`at?o{<BDnNGlY?kTWt
z(=dEcbO2T=rBKNUS8?-uWBSNS1!m4n3tDqi1a5qLm@l~)0hxyeU}KjdJge0Z9>xT~
z%ivwy{_#@$`#5Q)*mxiKG#!DM)7G$y^WyzJt1^n?d-yl=Tj11fbz!B$KWwLV8+vVh
z=|-~({KdT=#_zC!jj}3C@EI-Pn2R`ry4|LyuDXv8JsFMjkC;J6^AH^Uv6r`)n}ZkI
z-{(!XXz{BPm4)#h6n^|?Ccbz(AMC0g;hc##K+axONboS@Dpp;==00m_{R6@HNM{;o
zZJ&r;A0&bE?-vj`QQV>O$^6{UvO<Q*5A1gFGHv@d9A}PB#*a1`3YDVI^m~pe#=V>P
zkH6%YIiicueX<M9t~`TtOKS0XltHiY&4%=Vf8uxScHTL0Gx%OMV=5nN3PHc4aLy=w
zVfo=Z*vw-D>@#Mdc1jt(HZ=iH8Yi-eSAL+&m&Nn5n*#Ba`N~3*SlZI;$e<seFvG<$
zsxZRU54z?KVSEp1Fd8a`%t^m@eEtpt#^$0LQ%Fi6OnoT-^k6(z+prV2&0mQpndQ(6
zat(#%uY<hz<sbN?=mPh-BDxq&thpUksh}Q|f}grCqTO1J@s(m*=zXstxSXtDH?iT+
zIKms+x1Gh;4y!ZPlVd44JvSWaG97zcOCV@-1z&k<DsL>K%d{G_!V&KgIM9ou8#g9U
zUteYNl7g@J;CM};SpNopVv>o-Bk-n2%{OBv3{Hk8p51WRTFlG$-NW?8U;OwdFTpG#
z43F|%gIznHVy9(cc<8VakQB6YvJx+PLRbY*p*A>rxtt)^{fE!UIK<mb(q*RacIHA;
zMAqx)M#|Sg<V_6*;q`g@aDtc@RDU%WJl=;w*BUA2Zeck*D4!_01j_iSN$artkW=_*
zjXA@f7vG)8D42Ub5k}{3hO<9&VeA4eUR&=8v=_y|pyV}gv+X9`d-@_7(Nc#a6Sv}N
z_0n*zM;$xX+{Pn4bnw#Yskqi)CCr{{4w0$~0*p3<)DmylCF$oLkGw=H_g~}JC7$7%
zL&aRY@dt5dQiv0Ol)%fza-2Xl!JN^<h1RptxMV>fsQC@0<DVRc;NC9W+Qwkjp1I=v
zz#fj|v!HmdpDZ4sLrp4k!`*{1;+sJb{~mdZuB?6dtDQc!%NE~kKj^{h7!_gc_2J<D
zb1~d$Iwy92d*Jp9b^NrHh2muc@TQx_UiZ|Q@kcUn+S&wc`9nMx7ARo5@<{sNMhk}3
zF%ovrXK}=<VZw5iKD?*130i}6g!y8|WJi`c^r}_jQAg`=YIGAUk`?G3K}z6uzW^Ty
zZ3P3-hjFmY5A+&Eme=JLUN~`xa#I(3$ZMbTL1rS4EvF8C_82i<!<+D#&=Zh4Ls?J_
zNx;Ki$KWppLz&LHFZ_MUJIE_HWInww6W@#e&?+2(_L|LL<(0sV*lW(b`f{58cxO0l
z8ZO6#%I3o5k+=A@+W!3ez*l%=W)e0P=WWlGf5EcsIzL0aGp!fhTkPy5kZE1Z|M)6)
z@UlkXr|*mz=b$=T{%tiLXOckgyj6gYr`^SCmwcrAE}H@vpQX>;IEu?7y?MLgMvSXd
z8+LNl5<HyEnQp!uKVDx4n<cU0Jvf%DTfYr^XJ(;T7hA}6;P4$6RY+WDh;w`|@GYLD
zTy$IsZkXPT>u-<6Gg$Er{eXsWHMtLtU*_nTlLu+p^gQewsV`&}%LqS$#ooyHLcDOK
zIjuZ;J^tn}62FN_$FrO>u~Fhg7-uglc3TyNTegjGPnoCgyI#ZYhi-5yy-&fppaGl^
z9{~@2#Z23bCvbQ1IOyFZ#n96%#T|MSztGN)M@a_4CpH4xgoVJ%!vT2SL(x~gR9QG>
zttfZ~YJv^_1@3B{qarj_1lgSqble>eEd8$*-l>(s{<k;qVq<L>vnhlZDqC^#UJ6q3
zWtlB|ykS89vFN{zz*5m_Le@YvOuMHk<Q$&KJDt^l-Ooga+y)sTUbc%5>iYnFS-<%g
zN-rT_^ixjSlLhuqH{!T$*8Ej>b>W!iaLkp&;V0KC;jULX9c5evt=DA*HI1im@_P{f
zX~#Oe_Yeo;$97@=YdJ78=qlbG^$4e&(-Ah8R`Gc^bVWu)1y1>7%*_+Ke$Q{3Ft(%L
z;18<COqoSER5*$IpAI!1hl`mkejM(Rl@k0b?ut%?M&7x4AJiVZM|zwUm<@CzbeuYl
zZ*{lfR@Y;^SClzEldURD9<9mvY&Bqxr)<PpuN4@raM8yZY|Qw17SU1P$MCP-R`Jif
z#l3xuE)yQ?Mf)AjgeoypI5v0!o_zWc*d3UGxevGS`Gm9BJZd6N{Aw!Xp6Z0S%lY`t
zXBFX58YSiedFc5g=FGO_!(iDtT>W7txWqZ)@0!}eB(Ya++Aj7$@~2_nVil(SQ2}2d
zcc0ttrYWqd5oi4C1z6s$3zLCbsJ4j(M=@)<Kp~P=++4)pulqtPMLgv_+lDgwswwzJ
zMFn<Jb><y^e1zd{1<=u<Eku4kiicMvgJ2sCL38Dq*x%{=zH5VgT<9pAI#x~?y4{f3
zyh9FmpU&bVKj*^d2`{nihnMv6DNb~Ds1l?89%D<5FudaHUHY4{Hh!U~gpV6sq^FN?
z#(#h3<Nfv?^kbDwFtYl?%cRYR?X5CGi(DOkWgzB7b#9BDhHtpfUwmuZG=TRd$Kr_<
zHV|{Kn2NhL6#~Wl!#JmM+Ff@QP)p;n&0{eKoLT^q`R8EBfE4q6wY+d^(Hhv`t1jqx
zH^X208GUm`5_Qg58t;{Th=<HO2Y$yH(BDuA^54XJ`QaV(C(Wb$z?9QCSaB@&whe{9
zPs*_0vo_e8bPxNA?2DGC3;6L@KjT00HaIKQ1{<X%Lxwo7R2tra_k0B2?N)*nuImal
z8?M7gK}*;z(Gc#gHsqrmX)OKy4E~-Z%@{2H4k7CdgkFVJ^d&Nfx9(FHB1?;K|F8?R
zTZb}z;lClw#g&D4#M?hmy3j(fIH82kJ7@^M-AwSZzFKT^X^1fTfi|=3n>d%n$io#c
zIc9XJFD-fbg)S>$ao5NT*rMe>ye6v^>w8z=3$8)fK)gTBo>c(!A2}xMnG9YzP2?fz
z8wm@ePx1z-kFjB530#K{*goVIpCme$<ifR>887T$ebG8T!J`%DnW{0(7T>|%sh)0-
z79GIZllYI>(%>331-!=I67SgJUh-%fu9@%!2fd5MZC6CbkX$i6?@0)lg<gT)=49Bz
zoq==BtuSY727L@dsj)-T!S(Q7zM{@h2+q@G-Y#tfWqU(m1nR)?$s6b~dw1}C&f)m#
z;nQ?by(v?pp~-~)(qZl=E5O70I#@KN4VJrg;PNxtj9f+<tmFP*=BfDiWi?<ze*XXF
z?HH`C0!wQ?;Sawr;&G-y<jDRn;3qN*F6_|&KbK|v?3i89_gmZ<y2~=(J1p@zYX#x>
zZ4P^#)n#I8<Dv5AeQY7J4z!;{VVC$&$nal_YYc`lHefD<7d2yrcV>dG?g-rMQ3*1P
zGiHW1gFlo|x8iP#EU{1E5&IJF=)MU#iUs(z$jAMfb^}kJ<&6EtN-<9Uz#S<IgO{!t
zHjFs{Q)GW)Wdkp2)TB``Vc|GDZjlbpd#3OQx15LBj1u^nh`Gll1aE&QEi8APfE~ol
z)YePN_+Zvve(j+U_;1)p7&P1sni(G;XySW_7~3q)TF>$HY+bw{LIb;q^MkvVF7x}L
z4}PMe1D!`h@ag@w^iPAWuxIE2Y$AGYKl(Pn!nz;`k*|WI)%wDn)(F@>&JZ+|jp3Su
zF1LDh9@KRQK)c)rUhtNJk1LFsd&BF&=43GR^p@ci+Mh4AS`AwKdHf~YoVn7p9*4XQ
zgxeMOX@{BK{L8=b;5X|Y<s|wHcT(fv?bZd{+Wp!1dU_f6v-Sd~k#mEdm|4N!cejEo
zGd98P-|E6B^G8_!tD?}Tuf)8)^9AoP?1JtQS=eK)=*AQ|^r{mx;oeyjktaJ7c1?Fd
zmuCm^>Qf`}=xQs_xGgUk9S{l6X57NV#O|(HoU*{*wqzXC)tHAJr7+MTy3ea6AVeF8
zylqqFaeO^K9WKipFx7|EqccIi{~yk6_=4|$|AFIPCBlmsu>&-22#q>I@JsauTp+t1
z^JTJ3_R&$0AW;yV7kbbnW?~Ln55NL(_ms27T*&K8<)p-noYsU?Zd>UGT!~lGwzJKI
zE&K9l*K!4cDmsRDuky$1qGITAuZplvqoR;iUk&qKBtxvoL%7%$gBKS8|24uJuK!kI
zK92c?V*+ACr_wpvw&gNzlvfwK!PmIT?XIx4sz&769KcrQ@!0=-IBu*jg_0>PIL&Q{
z@VQ)>*)+NnjDN}r-V00c_ch9l!j|jwgP51tgNf%qC`Is-b;9wRC1-H_=5Qe2%$ZeE
z1JHN)7hbkVe1Du%!&|(+0HAvl&KYf_Oh0FW;VA)>Z;QRP;yk{()B?7TEW*vhzJbq6
z1L5kH_mC7N2Xp!galxV+Fl_g42yLzAmHr^uEZ#}|;$@lnNr`YZwh>0L&ta*t4gGWZ
zYTCh35E%q|!Yo|}Ub#jZ!t%up+q=sE#j1?v(x2cK8p`)trHP$*Ii|m112{h9@DKT)
za8czJAGT9`n?3jt8rpY(U%*hTH2EQ%eyAs;ct7WVuN}&y&p|kEDvPtMf+1~vG>qa0
zXeTkh95;58`2Sf&bb9{alC(~OYwr^n8NCr(6|i{N)&J<pKF-|c&DH$=^P3=JFqND(
z?#GTZ??bMbX<XX9g+6UF03)2-V9W3;)CKoJx;f$?Z}m<ae7~3oL9ZzIt~6cbWJ@!V
z_VUcC$VUA7>k#3_>;PJBiWknwx+-QIdAeP6y3DXW3Yq<zxiIZt{Mu!M{M@l>O#GEx
zeB!tlj-Km|L%9Ibv&;}*nc_>=>}`Ny^A&{drh43GX2y_5mQ0X~JD4pW#4D##`0ts=
z+~{DDQE^C-@yy7>P4mnd?J5~T{zV^FPOqhhJW>~Cf6-yK^j6UEhuvY{;R}$UAy4Oh
z)D=?r_waLm5U!P6gReVsp!-r2wz-x9S;Um-@s}6bTl4Vw;aONsoH1UlnoKzrU!~8r
zM&W9iVT{RHQ-+PsgXhtsangoEklG!<sZHF0OLFg!!f9fsqA*=#{kCIoGlbXnh@Lga
z(@@&?6mKEESiMV|iFmAuT{nwv_Z*SSEjs33nFw^TyvXiGZ{Y8%nYbNkFuSG2j%Rf>
zB<)jTcDc1+iS{QP^8O>9S@aVatt`-sFk{xddkA;_Q)aAezu{*;&EQ0aJbykf7N^}&
zV)_#_g*~qnnU;_WEUW*BE*;Yeo=R)6llac4DtiMK|FPj^#LTzp!<)2!kQ?mKHWO-3
zW`a#e0B;dCjM4QU2FI5S!6(kRfCj&YH-9bq4V8xq<LkRX%}kFj4X&m4hW+EaMF#5T
zCp}<tqYalB4AOJnnJ_1ZJ>Uy!MGwxyAl@SBI{j*JB}i^2;(dC{sL{(cp}YGIuK0eM
z#>ekKq~r&fcb8*Z(bMttN&&b8M?j6-bh!Th9KJE_6fAo-1jgE1iGB5h_(Q>IP`dI5
ze@vPR&oT+D@taOn`s(3}HR66dI0V`|m6)icOgJ`0jlXU&9;VFE5M-p3=qaPLV2*1t
zbUKCenW4KOBV3(XWz`1$r=*yU*lRd&vI^6D&0MI74F-!VV*c;S7&@x_GW1>*J&)6T
zanLPeX73X{#>x9Q_>3*#9it1l5u<Hz(w|@W_n)6I&&~qo4fEx@$A!_)k0(Kp*L->;
zYQRa8HghsRI_O1TzQSI`r+D3xeEi$*0#?>;0+~zCA@V;LJb(Wo?6vbT40$egGJ@B@
z<(dodbmT$0?yAT(HJ2CT!I7{mMs(<Ci#|IqZRVA{Jw0Mx5`FO1W87z~h_~h{V%sfJ
zOh!>Oj*uN9d^1sFyv4ocq4?SS>vA2@Te}z*9@mGX!{kLjiX+5JBTPwI3LkHl;*f*S
z;d5LS{+@dqFMn8vkEH1dTPD=O=z|-8yciAHlhd)^Kr#E@FgamUk0Yd)#__fnl?B^1
z(V%R539lL_E3}PN6THb1klL$&W2ic)zyE?)R2&3DbdPN=je$Go-_fNPBk=nD`po!M
zjl8axws3i{6dyTJ3Ojwu_`zOXrmj8}+C$dxeoKu6hh}p|DftV3Z_*VUzf)h3YSI+O
zc@Gy>C`58;1ruO!-aa@yYXrp1HiHXm$KlI6wT0487w8X*{^9COIcCMoH_#mugj;Ym
z+}6s6uddR9uyr~nn=0UKUmz@2Zh?!o0(RN)17=KCfXCKKd{$^D-fgZX$Ue{!J|9cL
zX%QCGljY`&!-87;CzHc7TOMNl3-VYgSDA6$RR{K0eDL3p9Bx)o5pA(n9+#!h#g&?I
zFnxs_yfxDn;$tkB9eTsWvu7;5Y+9$tnb`sS{xqDK>IHpk%b<VEcUm?}eCKIugE1mU
z;r@&p_(<?a`1D%zsqA=#kCgtU(w<=$vf=`*JWBMXi&>SH>Ps*}oYyxdx57Y!=$hZD
z$(YLj<#XL*;opRC{5WkGGcW5XII4^NiKU(R#c<IFV|R%!e;W-xqr-X4<~&fc+ySOD
zU&Gj6>P+v1Jy<6C8IBp07GBj1#XHYs(RABy95e1A+?>0B&(jm2Wb$?x*rv+l&q&05
zN$ogWJnLP~{R4xgnnLBiCs4lFh*>|h8qa)e!n{#cWL{UC1oH%C!9zb1yM22GX;Cj}
z>E|y*56l<ZN9_WztLEW)Q+Y=FKm`6VM`WWs(F3c&OuX_~K7H0uoY4aVU|oO$b993a
zGj@FgK5ehVtQFba1{XW&Mbpaoo0hNnZ|-tTqR(jfn$^k6O}-;?oD+zq*DY9w^5CRj
zA)fVz;0cMEP@p;k=ISqipuE-ab<S#DQ&SCd3G#x)_qniuvJjSQMc@aY?}-kW!}Mv#
zQ#d$f1)n_D5&l!F#gE@*<9~D4;tQTi!ize6VQ<As{_lnoe!+<KxS@5Z;FVx55L++2
z^-3{Bzq^X(?+SuBHEp=UWDvjBji)!re4~Zmk+5T;yzs7|4zE7kLWf(Q#kFUvaO<i!
z_)D3T(D__nSk<7+EYz{(r(IMR-b?v{%&vMoS&_yIcbekwg9gmPfk7}>`w}0H`3H-i
zE`}vCR-mGPhF42#!55T7Hq_mZV6yfr{mxVY+k8Hc*~HEGUXrqqc44?6yJH`OhTZ_<
zC+YA!cp=<h?uql<Cc=~j%1j-10ru6e2j;Lov$RXxjX6w*8?g)E)2kWWd3*rdwrDY2
zS_+?kw+a6zt;^U>(-5K~((o-8Ibn!q4IRJg8n$k>V2m{0!-}I-;MUni$MuxsoWEOn
zFJ2J`6nbKV;xy>gOMnnD<MFdx>>{X@;fHe%V_rLvZ`&^JMD>gqv)u;tB1%fw|4p4y
z6?^yAQK>LawwZn&rOn(bR$`>K2jDg08T}(meEVB*8wO4W;a_U%ctnpmuDe^0ZOU_C
z*vSCAWx;K{T6FuQTu>F<Z50@cm+5fh>TMo-`QbHZLh<|-Y3A#3CtCT{Z>%|C4!wrI
zj3=FIhFy#Q0bWrFSG5WteYy;D$3bK)<itW)o(?uEHxVuup5^VyDQuZ`iMN(0!RaBX
zWJeVX{`W;D$m#$bfBGKhG)an)giXMX)7|(L0Y7n=%LZ7ms|$a9oCTXETQJ+G-*}DV
zKituhMn65+2#K3sVVfrhsD+1Kg6WfbydbC=PUeq+4ee*KrLFi@khTVN#~CoO_kw8+
zwOb$~uP)piqbI!ZS%8z2BCxql9ayV6^3Ji1{Gy9XaeeO+K6TR!tl%K>aLj+>JF-bw
zw@ejZy6B42Ml@rMZ-z{k^BuZ<(R>KWGU4_5(&%rI>P(o<M|^8mEBL@^>h4M<W<|fE
z@QG}~kISp^=e*ar`&}sHoV*IxUtXk@sX(m$<1eklCD6+>6ol;jVS>{4ZrI(d!Gu;W
zg1Ku?;(Ws#+CQorziu!QZu4uv+u0vFK7GTb1IHoC`U5TFV8A$>xl8AJp2FIr4TS@*
zj^R%Swczwd2Kz603d^%Kg+GDw@YH+n#eQ`j&OTJaJ3H!u!M_%|=}#q|ucb$89}a}R
zOeGj{_cQ%<dIR5GeIK(gYhlaruMjq8sPNldoW~qRuZ(jS-gYR0HqX})F28ZbiD!Eu
zd}AD>eG}itqvVBek+C?-YY-1?a->%_9Du-`z5LB#Z(&MuB9%zV3ZdezySw)h=%2TS
zcx!8Zhx~gy$@UX=U#rjjTM*3aA5sK!zhCssJ4fKuWkqH|=tJy#c|M#Dc??oFZ&Kl_
z)943L$FMdfCN~{^qvkWccyh%MFs+ss_Vo?$0aq<(jk+qlxOo#dU&~DB(N4u<zy8I{
zRvWJ3Qw!LNcin|%)8LN7D-qtvVYhec%)`%3(C6cZwfCQdz=1ffB=ZX1t*gcyc|VM?
zlPbs0*7ecK758xW?m0N@`4gVbn7|K+JOkJ8R@h*d1-Uylg!fg_AT%Un*r|oh?A4e{
z3Dt1Dek@Kk9|hX$HK4XrTB!SK##puP!6&0G@e}p0(uNXQW?`PRaO^3<{?2-gbW$jU
zoh%1CX&&#oZp3KU=3&2adU%anI-VNj1*)4G;bybwS`t6k=Kuq2pE`kyZ<vI;<Xds!
z{+syA5EUjYTZ_5+B^odE^#R4YHgJhDXAY+RAR>$kj@ZY;`{Ba{w~RbEzV|-76nzgj
zogRbk^E<HiTo5nxiGD7-D(s?F%lGc{<NtT&H}5@wD-DM;V`|@10sS}dvWN#@Atx{V
z6T9j2^dE}(_EY#?A4fkD|6TLHYlCCpMw}uYi#6<P;h9VgK2{V<yZXh!6QwSk`Z61L
zggoaJ%Z4+)Hw$T>oYAl<`4i~Bd5o(riF>in#*Bl=9C>s=<g46C#${uqAYEBx?#m{?
zi@6)IYm|cE^tT58pbQ14DQ94$Mh86J{2kkxTH<~<g(tbT^JZ!5@st;9XoY{nnQFUB
zSZ#kjwpuw%_~rN<W^I)fE`4v}Z4cDI>>-*${>Iy|^!*4}-JT4$mZkHqKK1bGa}-zz
z(y(Y;Ht=)2K`Gsz9(L4#dA3=NQC@G!q{xTB4wVY9IG)3we^tre)5Q>1I~(S{GZa$(
zYeB2C60oMxJKk!}N??~afiKa8)@5sOq1kynUZTjb;YPy7BcAZK!c_F!$T06Kb%Ysn
zO7M$d6Q(@KoCz6j%s2Zw;>Igaajj()Zd~#L^W`0M@f$IJa50KMc>V@n^feFf4Nzo!
zD`iBkfvACt(GhgIJL&QB)%bHc?zGh$kr9%oDJU!*!A+Y<V+S{yYC&h=<fRh$@!vzT
zDUGAC>OU9}(vJ1!L_oTm3F8>Mo&M>sL)#HOX0@2%Um|`^hr|}_-TNE5^h4m}-j{sX
zT@H?n41<qGw~4jZa7HeylO7xq4Nnh*(m5Wb^zX7CpcsD!;Oiw=F>sI{E-UtU7Dr<~
z$^zGhs4(*`{=hfB{==UQI6Uu{*dtonfxjrIGslNYGh2Ss@N#aP$hSO$6(c9mo0|IZ
z#6eBr#+gB|ZZu^+|9p&d^X~AK&kcpq>VNTvk85CCL@dPBNHgE)INXk7uw#8OA9BD*
z*!!e_uM*kl&(CQJ>WZS%w!w_a{#D7@UfPeP9qq7|@+;gk(1{yKg~$mI=s#nB;S($K
zDOqth@HM9qedK@OrHK!4oMSemebEzs)gPzl?Y#`g%bwzoRljhO)k}DjCCjMU`Qype
z%KUjTA8_u%GkWGWJ^DnHsZiCR$tbk5xWY$)QM?h1my4|OE(29&G>A-B!wx8zn}(J3
zhX}rvCGh-j4b+x);prKV@Rq1w^k1fyTbloqewHqF?tA}WQ=H6=64~sUSH<@hS2O0h
zTpq7#JsodL3I?k&UGS^K4L98n!HQ-YOzWgD*tZ7wF`|P_=~_M*xrQR|-wD`0It>5&
zB2D9lYFg=j7xg+!hl$q-5}CyNKu2LQm9fBrd7@y*v_|#7<>x$EC*EB$;tXNiD;ef{
zWi{3BV!@QpSc;bz?87In%fae_Kz=B+K+WA#&~G)AseUhho<+vkR#i^OI#<doC5hi@
zv(mv+F_tbmBqyA74~6NfS)9^t0#k=%<HxN(piIp7Uw*7Fl=mib^?w?%S85Lw&Rvf`
zEsexKhIv8q;wt*YmJ9Tl)iO+5Rs_EADiHtP{EVOU_cGKtSO}iv1T9|_!Ixh#6H2FC
zCzl@%!bX#ae9TQ1@Yr=2pY&=2*<k{9T{#w9A6qh>-aBCLUD0Qk-HY$0XX6_+_b78y
zO~K~mG;C;o8X8ZO@m|$yz!n_f#5;K*iFpk%OA=s>)I(~5{xCXd_gQ{H`dJ)va1uYH
z{1Hz3p(7mJ6$<ixEtr$~MY-6W!n5aH#vbljco&;WIhX4StXCI)G}o9>8S)N$<mh8}
zvwQqw(c2PqA{ScroS+-y62N7OtgyySbVW}VAeIu{0WY;d#`p}DvXlD1Uta(Jdxcad
zd7!CAWemxZgrJ+~v8o1H=^u%X#crThMSmrJQ%6d|8h4VZcnNnc#{*saCQohi-H!fk
zi<US)Xd_nPMUu~+1%)^LDkZZD1IU{?d8BY*I5+kCJGSJ$1NY413c9_x7!7!hM|JaV
zu>-o8{2jGQ@*`H+{$8jyMK9JTkM3!TOa3}!S$<d2CZ$4=-4oH!bAjY8_W}*dX`>(4
zGs&g?!$|%}AZp$8jQtk!ne0h3<npz~Nw#bHa!UGJxm`PzQ0~HJa`kyAa=P44ET?Ux
zR2w5nSFR?dxxB~$KU_f)<b%nsQdMrzLODv`O`n_HHkrCIFHiFQVw&Vq&PQ?|8FR~K
zokUI@Cs5c=FRCZwBwBR4$l=hV2gqY*B-y(tgXGes?3I@_$Y~&*yl5PSq*u;I`%-8`
zwt28oYYIt?lNKt}nMp>sw-YVBm&j~h7&@@%r-P@g0y^chpUimmflQyg3jGc=CFVt&
zC4o4AjHy;eTc=8>9T_d`=gk(J-Ypf%dUCnLf}okG!IVNH`cqNU{3OcVZ6a5pcm<sa
zTfkkhp38Z)s8SE+|3OXZQ7F@H8kG}$hBLeL%D#b8MO~{?Buk@dZi7@0+AJT*nVBym
zd@#d#Q+blsUR|nn6p!Z4Z$$BJPl@^-dGgapmi+sKDUZ@`l25^txo9gZBr*7f;_^JG
zunsLsui-ecH$03?k32>Fxih#vb7iWG;v{$ShoX(!OgP0!zMPBoD%Sol&v`b8dA&g+
zD$)BRr?ul5T7NE^`!K7QT32e!y@`pTraoCnwk_)?^IPJ`tRM-u=Vu<3?HR=NC8$ul
zbH<`u+!;x{aE&W*y2WjOsZ3qg`o|6ncuq;i4y7LLe<)GkI*K~F-HcN<2<DW6%*p-X
zUX)B&JIT+srsl=2LgkXfg&cob5;d)Y2uC-db0c1m`UB%R^KHhIM_-7fKlB~?wTPhq
z!a2^0R^XK1oJI3Xs>#~LmkY0*I4#*j2>H^th%)NzBuCX}aL+2z3iV~qP%7&biE~d2
zQPAU1knVFzg$n12S{O<`JD0p`H>Wn`1#^$=HgSF>4yb;{5H4py2J6r)#eKHarD_!l
zx$u7s_o#d{A5nUYG=BPq&i$%o_b2sGw<d-V<(C_f-Ij5bmbC+SOzI1AYt1O6zaAhz
zROR?>`EFeJ)CtrB#nWiC(*V2GMuy(=JHTFl^b-=h@e@~26hx)hWOFe~cOgaZGhFXj
zd9rQm4K5$&a)++n<VK#m$3?8kM;k+{spltExwBi&pxM8VOa4TpQlgiMN}6<u6uWpM
zq;iIP-6wEALm8^1{}ne=p@C{yJ)PQqQ@!9~#CfD~#)0##GNn{3t0mF?{pf`HbhQ1C
zBKJfllq>Xn%=PD8<Tf<baSa~=$j!G0kn^)!sPxo$WVSqIvPsl?@;NezTE9k_T6CW(
z9I`8%t$Q&Vjhc6xd!raby~SIR9|@pl)%9?lOgHtP)_dyhZ-R^yquG_GtI3)-+Eo9P
z;oQ-eztQ9=18mAaid$=2L5enMl3K?}D17S%jydc>dB4a<&Kq2)2ip%L?+IE&#bL6<
z)$lHg{-7i|lqMk#ffbTT_e#;9N$E&3t&(Nerl2Vb_2~7e5!7ZoKXm738#z^J#+4mA
zj>PgUXS!vN!<#XdoZQiyM80P+cc-}n-F>>8jF`8KoI2Ztw(N`{?ic28M`q0>i{#6R
zw^kLZjRgsnql8QiZMnj%1Ju&6TvlrJUde=!4QT$UG=yJ<Q%*HDC@xiz3?;3svwI}^
zRk)N}RlkZG>|W2kuE{IB@H!QZ_6tLA-<_hq*tSWEyVg;u_og6^KjkF8=Oh}h6^<qz
z$|)>UpH12e)5yXy8Sbp&6w2@&L+v>w{+WLtCwKpPTwrQGmNJ}?Ci!o*2O9P4B`fa~
ziZbT!Css=wsN^AkNx|{2XhzLy%5lqe)U$pXr)D(PVYr<Mk=}U{eRG?^HXM?oVkhcx
zyU*_AV&|)IpVoz<mj=CT_}8giab+|aH+l?7%G2Y1nVfKNRI+9NTGJ9!))=)s%VASa
zT(|!=;slvlJ(Os$=Lm_MMOu6HDCsQ|(7b6hx+UXTctvI;ch01kP4s9bdtb67KvG63
zWXw^&Zw?CO9mzYNnN-tsj$GqZIT#g1e)$|n2&c0n8`ctLQ<3Dzvq;jJSS$%Fq^OWE
z4a&aUO0wFTCKp<6puD|G+=D}7xt7&|k}=bdp;KM%$Zl)i<dBqyq-f13Zm{VkIq`Hk
zJI7d!IyC7kN{nO3b)7-B+u;v;%tnPmX6}@5bRG#U^CE5870BABk{qw`As?4*Lq3yl
z5`Kb@<dJZJjJsn>MbE#1t~aW3Umr&i&R>SCXRMId4QsSdZ4S3b`m?09ZmZ-({c)0&
zI*q%#bpuKnR?h|=N<wSX6Nuv5Qq(?d66bXAv1HC>MRFo8gWP%K!%6kZbAtbNwqpJz
z^vwDIdOA;)O59?{$#fO7i&7$x-F_S5-4euQcNuf~&RfZvh8x6T<Tx(G!;w1v#+nM9
z(uoYmz9bvR44{}Xfz-4F9qy>VuB7?cPLjMSn9HFXB*VU@BYjJE_NZni$+h$$%Gqz&
zOyysa+x~|MT|dI1yjz>wJJpSNCR>qEpC#PEL&K>1v2jSFd4)s#3LP@DQW2#sIK#$G
zd?Yy)s7Q5ZeIOPF>$r`Zji_@;=h-YBZEA1!TeNR_orARZNA&ev0EtQ5j=t0vu{u`_
zsn@NTE2@0Yni^G-;?&1P-eVc3GUEv6f2Wf@=5Y*NkFrO)HtSe>`RUZgypzb<aUGSs
zKMB=bYGd^$w2}S73hw96K8dr11Gn?92URvOoSJ#MNupS>h@_7HO|qq8&^X;ksN8-w
zHQEOx;Ug<aQh5U^QHdqe%P8W%zLe}s3`Kq-ykp6v#}cVe<><joDe86X3)Zj4ggq^d
zSnrSjkzsQ8NXab$8M-`^cm=#h9TOLlA2O{dQ+$B?m^Ple_%;&xyBnkBr}Ic*+Y8iB
z&8Jq*3}gT9i6qkArYN^q4IQ~X6eS;dBzfq&4jnikL7_TDg{zuIl2mpriQ2P^cwMz5
z4>mq4^i<r7T3^<)?syw%9{P<9UN4QVe7j1>_*n8o;S%v{cb6FFOp(}=cjWPglal^I
z9THoAf<1jhk!S>h!_f)HN!rn|EWe}$J?Vc)CP$ic?aq7A?};qx-l>5WiD&(Xi&~N=
zn?JKdqueES@*BuK@4&*e7gi2on`{W$5<trSR*?Qn$4QdeaL%&$E)n*KKRa_@lZ?|j
z=-|B&V)sj(8v9F=1oRmrIlEkmhtV$7>R3`}JNE~vl$(g2HED2){Asl0LbT+3jSI@z
zB8_gx=yKa%o1$d}#gaxzBAQoNf_BvqlCsVg$*+kdPE)M8`JY(w*V~Nrqi3X6J{#Si
z>&Q;yQzX*H{={}UUnn)e6x4PNI^=WnxzF;(B;o2GcD1nvb!uuBDz-ac&>}rZtcx=w
z0X7c`LXK3Dnjvc3D1&q|xkMLP+Z%B|uDFwS)l9bU)vCgZs&(w{)L^zt_#sKO&9Yx;
zv_N93GL?P4wL`KmcRF(LeZn4JZ^WKTvzmN&(*?=m@I&m}eF|jE=pc5}`Ae+llogWM
z9w*VNagMC#<bQ>y@B~y5fRXV#WfGsa0=4&~6YX7Q?1nqe#OtRn*|61!sJd3L1Ew}4
zEBYoYC*SE1FZGZuk7C##{46xkr-U4O98j>r9rWC60y23VNM>xfFPVPC4GlSDEg^}g
z$dTkfc?W9Np?Q0@BDKDJ$(1}0ma@+&{Cs-@F|FkawMy#QfJrY|)#$Gd9q(VV+J@qJ
zc&(r?^>eFadXpbI)_I{YVP7iy=8z>S>hCNpbezW4r`EHX@*0HA{>VPsF~VU<k`uc@
z(}HYVQE%Us`-z<<6G1SqPhR~FBd6Fn2ffq^G%4XAy6DJ|Bg^B^leA8U<b|>99+k*R
zcH2D*^ZbVsm9hag-A&aY!a1<;&{qb9^u1tPZR=UNsv1e(GQqz6s3IFNVHErN;!N@`
z7m>bIcM98jeaK=53HiKcENTo5VwWT;kPk~#k#^sCB6p#KtR#-8XuAwrXG>6AOfYHq
zFP{{zZDNPS>_h>_3&|<R3zCIR@gzvI1btPxD9J8R<Jv=hpx<@Dr1|$aawRj46mNKo
zqGkKYp><o)s@V&OOYL~n>hp*MEck@R8up<q&EZsIHAS-LE|K(XFysPt&J)+lc8P~Y
zKWkO3j&{7bNH&w9DCgUHZgKtrBK6@6QN7v6PPE}9RZm^nt=^*~>ys_XVslTl!zByd
zqP$S(vb`vArvvJ_W5mhMWh9yJnoyCO3OTygR}w$a<1nzm39Xou%T^_;amn!;5Tl_^
zRKABG+hBRBbcH^d<&=t20%F<T$`(@o{VH*0MpGR#<+!V+uh1wxf3(qL4$7XOLf#ZH
zlArwwD5icCTJM~}s+^V=#ci=9iP%t8!IkXLBW1|)v^IB4zJv^Ulf_Qkzm&vVDRa$k
z?quO-8)9hnfD}EfB5SH6NJw%l3RSfvZ99xm-e31ZKbbA4Z^3D@PCtRoK37KGr^!*C
z*~=X6soWqo$F~u0B>@dHqR36ntK@8x6*&}<Nvfy0N=`g1L}kU6q~ZKc;=k!Vn)cv0
zy7R2g{$A89bRe@rax>&CYdFD~OkDke)mrd~4LB^YB|m*6W#`gJ@ThGL(}VZhr#-Ah
zUuO)VCKQB9=y?H>9%E^0g!D>OanYaLYxN}oi<BtG%b}9Nu+dyW=!rtl<YLw}^r`5O
z-b(B??Pjy)zaVC1%TT;S4N7}|87-6FLmsYFCi6psiOr7FBxa&5()ndcs+0e)Q^KYq
zRz8$e-yTPtO~#PO-gHT_bPlP^H!tYA_l|s?l0YINJ0<4T0mLwN968(}&3(&@Cui%z
z$@Aq_g!yJeaIONi*>wbZtMozA<m8I3xZWXg=EvEP6|0cK$8gfrpDW3eW{L0eSEONe
z5Bu!g5Nbo=cJ#dCf@IIfG*maRg{%)WnjGEDvKASi9db%S3p20zqSgv2F4IbiyRLsw
z(yTs$8k@0;3vD?`dHxqlu4`U(NSRV&e`%&XIVa5{|8E{-)w~2!`#BWN4H-d&um49j
z9FK9>b9f6kA)raJ$5;dTjy53|Z+cPFJ&nn?lC#MA{%urDsdBCs-K6tGg(P;}EY8Vs
zCy9J_fLI$0rHa#Aki+JWDAQvo^)@Pu*i1c7>eFcMS&=T~G&h8VtBheS*-|oDM-By;
z-r!i*r|cbrPNGPy;0}FnVCAMCrn;SulZl3F$tnLT_876E_SX%g9I3d%@sDF9uk5yP
zoAbhmMcfQZCC-oB@L5es?Hi5emc1oKM@Dk4Q>8e^^&=@;rChcl>phyCWk_B9u8qn_
z4vS8OlHRTlq~VqaH?Dm*H7TeM)i1Y0RqJX|wev^Gea#B=%bQ0nrg~Jc(kt{PJCJH^
zO_IFmna*V-YEef!1a{`uUL?G3B1iW~QB-vynxMm?kmn}Uq*b*Ns}x1lb(STvxz?QJ
z4OiqIa1Xt0&p`q1zNp|>nB<3a7*(Ze%F)iY<a>uLxBUJQa(0y&^3c>pWv7gYO_nqz
zH8zio_L)cWJkp3~r#i`>l1L~%3Kd6ta-XU?$YF;-Qth9Kjw_f@8(&UDtAsLS?X;In
zU2qfmSy~p_m3UBD<DU_?NpYkl&_c35+Lui+Ifdd5xl(=4wUAPB1M2m)M2dg2h|=B!
zG{;37MK4YyO?g|%-PRDIuU<=@kMZDK@<O>MQZLZ9jD?)+G!0aIXAW0?DTTUz|Bz&v
z`Vwl0!UED({gC`|)FZbqZsRn@^r1J$1G%&tv(c85$I!lc>O^*nCRZ(<B~BHmP>eXl
zFJ7{Wn&4PO)D}M=`bTq-{ZlzHnD?7_ujnNT1xf6g4VOsTA7fNqKa1R-laBh97LfED
zX*O~9Hr9Bjgp+A-<jVUUDQSBm$tmO|zqka+o(yr`3~DERyG^N4WhYU<2Y1xq6v|q^
zaVIY9TuLgwh;8Utj~L}Wl<t%tNN#vKJ38Bmgp(8KmCHyXyij20PFAOim5k6kwug;>
zWJdD#DR7@WUKBcCdPVBYS(3D+2(<^Sr#_ExA(QtFA&a{VIKyIlO3&LL1!!HAjNG_`
z^I-3jc4v2q!>SYL{=S=JYVS^Nzad8^|Cd7ClLAq8vk`ZBoE7ok5P+o8tk4RbMu%In
zXV{h6<B`JitE`{mdDN*%k?B)nB}Us2`4My;-8dJAY|eY4$+s1V(ZC6EL*Y6x4$ESf
z3X>?4kn3!Z!ALYjQcvW^1taSxchRog^GIh=3R3(|lf7gqd#UOSx~`Xjp2>8e>+!7Q
zvBp)>6|<Q<yR;jnC{`B^ec&tkvDcg0uDFs-QMW~E?GMp4*Hi4{M|08HyB{RZ&J&3A
zf92=~d4=9?FC%xht)@KZY(bLW#9>*3AvJ5qKUDKsm#h%)$u8UHQ2|pY<S(pPN9yn8
zqEV3_kzx95QkyY}NGVCE<xv;Wxv`7L=rbd^f(&y~abpZMiaCj5xCY7Ju``L(;Y@U`
zstAowvmg%(;*e9+I95wvh70>+E7|3(ESVgWDmmu)jUDsr0h#?ELGqC=MJKZlvR!@S
z$<^3+Vx+ZFa<}W7B<@l%;ReQ&@T1CwURH4uKil1s_5WREmnb)pTZ>dk^qDvbbZ0u`
zOf%yQ97m(e>>4!kyMV%nj6w6V9<zr?z2xF}1y-fQvasqWRp@rW5YZ#GNM{w#&N#7}
z*t<zdLbfV$MmJ@-4m$;QpelyEH%LUsyuFD{hdSkPCy*R1dxQQ4xS`E1N|MR@PqNEL
zMUVmI^GNlu4;h-<fu7GUK&pEY3bWx!vM@-7weLVvk1dmEDRd+6kIGck<i^5FmRpeJ
z$4+*4ZX{W|P0qn}Od^W<9Zi}QW03)K4c#q_V%OJSMB{D4Sf{_sBp!YgY8o_WclY%;
zxEs5%TNe+a=KIEyxNJr8bnFzeSS1lfuA4|^<>#{})4oeglm;Xpe>kJ$>iZ<N!H%rm
zqd>i}9*OX9d(^%_!$IqeBQktLkabWf8~1Qjp=P%WyO69VGcQafjk7<opB6rGxD^Kl
zV|{E$^@(7~L+5d5&G2gu*&B|NI+qFLy_5-iY+QvzvEQCFY6KGPAy4eDM9N8;hOU=v
z{E<LrKc7paypNG@>u#W9iR*}McnOi6`NpApzBJnZyIA6ybq4vnDU#dIf|1s?Acv6n
ziDcEc>5_rB3)${-8HHK{UWAdDN=_AxbZ|H;%NbvJLf%@ww|B9NBgbvRkwj?@IeYLU
zs(t!D?VagcPH)uyn`lxZ(j=)QB2)Li)>f1Xm7!FUN<!w4S-zopP@0wINfS-z-q+fc
zDJn{mC{ty~kc5N`>39E*<M|_=7tgEx_B!@(U2C7~{G6*6=iUtk`{tG0p*ttZEiFTs
zb>KM;_TM3Tx$zjjvk^|+4FQjfP7o30Mq6A<anQVqJn`5DCq37ISk`_x=iW*mZ+Jpd
zpQ&Q8?HT+v(v>PaoP*P^$>D$gny|swn%lc)D(d%*W}Z08l98-DW2t7&WJEOMjQ$2P
zyxoF%RJ;syTeQKT<qZC|=BP^gQ&72|%Ip=@aZ?u6GT!G3>5HTK@Mzuzyr=aZx28w&
z&7;ydr<z1MY4tGE?mCT6T*9!oUvKBfc=MpDRfa|uRS>uBEIzm3NrCwz&@;^<25QUb
zplS*<j!Q$|XIgyE-=+9v)?|!TH0PD|IIO=V$`@La$C}27G}oBtu=>B{&~b?6dm4h+
z1JSEt?Cd{aBBRTnbtoos^`Af~d?a-)OrV=riLp1+WijKWF}%Kb2K~Kaj3XY+$4&cs
zX!_w=G)S3)+Lz;a1GQ*0S}}=_P5Frx^$3su(`Qu-Tj>&d8y|`v#iK?KVPM-UsJiBl
z#*8#}IetRVT`TzXn{8kkvzGV!_7ilzdGVv1r}7p_W5@=g1$kG*Sbo+l@C#c3IX6tu
zt@Sb8(o%-b`Ifj?RS~z30KQqvgqj!ahBbML5Hdr8SJCYh?LRa5F{g{*QQKEo8-E+E
zT+>lSUXI1)`Bb|~j2A_7gQIgY805C%=Qw)^{Ak2_UmC}zM}MQz1OI5*ft9T37-NV(
zJ0Bx^H?uF!jACnxe?a`PNa33Yov<TmA{#Dq60fH{AnhIZNT9L`zo7am{x|P2$e%k!
zH}^1*H&lY$vo5MD)&{BvOi=l|6Tjj42)<I$1zRgGLhXeiT5zQT4X4G>X@g_1#e5a$
z*|Pkbt08z_V8<_;Gn3uX_a3XnRe1;I4QySV0WYZ(n(~KX_Ej&zu0!)!bK5Q0{9YP6
zHg`eZ==1Q*x{Wj?F5rtjKa)A&NiH6o$L|o;6`2|<q4ci-f8;02PM+L}b1gI2z{*gJ
z@{WTArJ`}*|B|74K^I6x`SS<Z4zlDwX+GfLHnu-_6e#W<!Cqdn3hB2Q{LsT>s$&!Z
zs>1@d<cucUr8%2V8hr$ua2JH0Qpa$&MC{ox8oxMS7mmGl3-TtIQuBc^0^0VI6q@g$
zuW!%6SrcZWQ|B)J?Wt1eeiRDY85!WWa2$JfL<KH6`k9(r?PJ|rb`!zkYjnM&Ax&CP
z%$c9-gJGdMzhvSvu6my;G#SkxTD4E<8dEc@j=oM-Cux!W#bc4N34%E;Tj)~z)424R
z5!5{nBFdhZ$hqY*nB}nn<u`C-pmr;iTdL4+#;3{i%TH(xuM5ZA3`mfPDY`jsB|m37
z(7q85j2rh3bF*d^k`;X?@koUU>3Fgkot6ZFw9Yt`S5*cDuiqr|KpA|UH4;@Tc2cWx
z_h^`$0y?*ihU+fcbi{NCxVh*99XCM=6dl7*)XyZ9)8gsma4}3<b&(o&S`ucN7jZmt
zka~Fgk`U!X<ikfh=z2XKrhgj6Wp3PyXBK7<+V2cU(!P<>uIFUPpqXSRNs`t^KfL`*
z4@bY2gt6m2Da;hWuajpWJNrDD&^8xjVjq(Iu3IZTO2(7vBfh|!W?j5jt_|_0g1{q6
zj|o*?3C?r71eaqf$o@nH2r|in`0f|<@#IXpPtF6fw`kyfK9l^s8cJth2&KkdkLeR%
z4>Gaf7Wrc;2aaEUb5}at>HLb5aDU%bXi0g_gd9JKEv{FPtG-5luBjw7O2b(CpCOw%
zEQvZ6UASww9pIoQJe5>}m3L>N8QF{*&$qy_<^RCkc^gsa62bwGx17b=>oB@0i++8&
z3#Zow(9d}nXp-Xy7_Al{tk5&UI$t}!;J!Uc?)r$^l)uBw)Am?maRy9}dP2GCS2(F)
zgTfAiWuZ=F*BuXLMt%(RY6gK^q&SJc(nt67nL*EaNq&76g*k1<xP8Gl@x-Pk+^v%d
zl^GhaQf4b!T4pjH#y3cw{SH#_aX<Vz5<`53kK<r?3vu+^2CZMKA%AWnB%PlFzMhxR
z@0B4xmW#r33b$aROFfhEF$A(_jbz;(j)OBvVysnoJ6IiGOLjjXh|ycvg<0Zo{5?Ys
zw!H$`-AiD4;3PJ$aSRSESPaeXH(_zHC;O#3iezq*#4X#hK@zs0c<flP%T5N1nI_z)
zV-<L!`vRG3GlsRzxDT<RA;OLPYP2)Z;1^uX#i}{U5Nsj`nR=(N{J8>^iGK{Om(=-}
zs)kq+S%OZ{lKiGEw}>W2!JUgrkd*ll*ZzrR@^t3Xfcr;KyzLu`+uVgV>xVS(`87td
z<psJ84`AZkK-NX4oiy9$an6fI@I?WU;8!z9GPKXbGMOu2B+`JmCnb0(U5BkIdrfCG
zc0#w;W_qpmARNe=1|F)r(LC`R822`C^tTWS6PNKDKPLi%hWrmrX};1X4Y`#`c&qO_
z=W}Nkdp<Rp=<G<~mRyLZcL}iPhReY3<z{&EQ<D@t)#fLx8_z%b)<W-EAlfUHK&ja^
zx~}FZj*1R~i}#;$7Gt(T5sd(S33(8ae;CR9!7=?0A^6V#SuT1Q_Ue;)tC~@uzHKIN
zcF3Lfd+Tzh_TKEAyf&<8zRC=m4U=l$ay<1p7iUfUN&<pJ-(T`q;otE*x(vIp%LdhX
zr-WEiJmo5sX)lD3cs*WrdKd(kNwCXZrQv_t=>K>5UvdYm^M7#Dv+P0r*<6x7=L+3z
zya*FEbd&fLA4YDJJeC%$fW%f;+VCO`H*!+I6eW|4p~vLhbwl)jx*9uX4v_X#J2<f0
zg5Fw@N3{MsNmH5>VexA*vOmiXf0L;s+~0tHP)y_k8?SIa@xDYkCLOh{mXpmxNAL=F
z5F~m((z!t*oxO4*Bz~!ZrqkM}dN-At8;4OgOA)g--Xr^d?Be)gahN%9o2D<A0`tZt
z;mx&Cq_c+5r+O;5i%p^*w<M8&J!M2Hw3Mo52}$LMAo3;F3JxvSh8I7}xx3TOpj}`J
zSvJ`Wx@$+kHUC>=&4NZ^Hcpb<S$6{MlQq%zb}i{pSx!&sFN0Sfli^-^ELpyIHgp*^
zk{wfu1UDb_Gig^^!SwHE+E*_Q0VhI1K;*d2mo8wtDv>cI{Y0&r0gu!y2=%&6PmjJ$
zhZ5a@vzUl`i`oePD~U$eHq%qeG1N>xmfX|5NmMPhp=ae{syQ!=ei=A{LuaPr`X@o~
z^o|mq*s=_)T&H5;><~C6>YGbwY^A*i6?xmfWO!3w%w#lP!-vM(QF&DyxR1I<6<U?>
z%aUKz$kH7PrhaB#`hF*O(;R8NOB^oBIt2wa@uY6z1kk?iLtFmbgXL-;=)%~k^vK>v
z+<B#NVxXVI7?mf2&E9QTduNF8`C$$(EeC-M_6O(CwY2-PH`WdC<h$QiVxXEwkCQyC
z@V*Q#QkMlX+plA(+;MumIUboREjYmi<25lWdaJV)uiCUg)AJhGzNQ}>pBdoALM_zw
zvPQ2H)tFtf3aZQk@U1A$YU!;&FdluKc5P3>KKTilG~)v~`$GYLH!`I7GKZ2fxhRpc
z0J7KD(<IRxiUZbd_<MR7?N*e)zMsRSH8ccU=x(Bu?2P-Kx59!VbDFC67*7Z`(j93G
zq!dM9--U%xJ2C-7pLdJSk}`V0<pvyddIC99H$bV)eyC|XgIx(WG{QHNj++t5Res#Y
z7<q)!Qlk~rMg5*YJd~xjj~TKrREtKPx1+H)189rGGUnGuZ|dolM;{oBCO&*2x4Lc_
zV{Wj3RD6!(Y&5nKukLw_Vw*VeRt+YJcO=N<31N&bGn09|NryIsE}&Z3CNyVB3my6?
zPR0DZXs=ubqiHO|4Bfm(xDkgK>*Wp19l<H0WW7fq<C{)96Z-|S)85g4KcYB2-xZ9x
zZyx<}wAffI;z6Z@odivLAIpS~4Wuqbg+gY-MEdip3Gp16B0OU{%-JrqqUW>~X~xGa
zX7}3^s`^Ep#O>P7MQ`$<Ph9^nnP*Lj#XCtd*Uz76ZOkS`YLR5bD&%gosBp(D9usG|
zdlU@gIlX2JX5ix#LbuD4?tyx4+JQx)wV$Y5a_<n=KR1D~bH73tuPUJWy3c8@w~%DW
zYm!wt%Cv9ZA}UpFN4;B5adWrE(&-8-X;klFI$PqnKzH;q;;=WB_+RHPd5<mT)IXHc
znCN1<J5!$eFY}<SKhhbGK!2v^j{><|UL%~IBFm_%-K-Q?mNQ8-fjj=hhMZm}PDj|=
z(NG-?vO{qWQ*yPLIhv3xNZ)QQbb5A;Gsu|2jcv0NG$=?DpZnQD^KmK{ot{N<@=yP9
zHaBiC?`5|N{k!WJxg%4!qBk0}-9?T&a$&Vl<N9qT-DLz5nybPc_gqJgwpfzu-)A$I
z-$ydBhZULZrC*sV$NqAPI**v>Z+E%a3$wW<wwvkExWT=uEN8|Igb7Rxt}qcnC5+5;
zL!p1&pmG2FVJ<Ll6B#hx&n&-qjQJP&m^qcu%sm(S(NP_0oW*`Q?nzS==NWmja+mKI
zuF~^7qrb+L%QHX8K-ja&2+O5h#@XYH*o}G2^!59=m<i)K1u~V}>QKmZ&0iq2;<wO{
zDbq;Bi#arV^(*d?{wzkO&4$^XpD0vpNn`GZ2QvjfYPb`N?{let%o&&DIzelY8XbDq
z#~ie^Vf04c5Eyr!XLes3Prj^)6;ut!2^toxV4?~}a!n<?Q1Mt5S65XejE?dUgzTNd
zT&XtVc9*#`->rWOFE05bc;|M48QAQ{@I%X)b-$M~Yk#g_9JK41fky^h!Q=w&xP-A_
z)&B9!yNu=BjmG)HsDBd};gNQM%lT-*Cp8D*>y+c1f*@Ovr?gIZer2h!?a*4moj_%w
z-aKGPa2DgPmo2EcUwEl+AZ}iA{72C=$kT%9GV_=qhbXQrH<l@Roy9mNR8@Y9Tp|o}
zjpK5~Gr3Y(Z{g_bgn6lYIb3ATV?mx*oAA@9vz+F<biuX3SfOU@@SKG&QWzKIk;1E=
zy*cGh31-AJN9IJYC0Fbc%Ix{1BeZMhnEuLAPOiLA(33okF^lY~Odg71<Y};Q;i5`m
zU!^6p)a|d}S@~bV#+(4A&9b!eUgJE@EHjR2y>&&XRx+B4`*w{R^<WY+v}qPo($mEK
zt$rhPI9JZu%nlN!es1MPFSKXkj*l0*EC}TcZbfs$rdk3O>sTh|r5`sxSc%J)O5i>!
zM{~~Yfs9hZjLPBMRQ#1D>Nzc3NoGo|BU6;>guho6kt1R=gl@%7aA}twx>&g~@{-CJ
z%+%9#?=f_V_!#2VUrkk{T`;7^3o+i7E`9rps`8F_<=i|B(pXIU#Py*{O#<JEG}_r0
zc5=Z%QW$N2kQ#|)!-v=5@K@qJb2hw`F}<utwt2?W_liro*rvBkww(v8Og%tf<ZLFz
zy}oGao`MBCt_q!ci?FdmwC`VkLmRphNQJQpCm$DrBPYnCw4EGCq-+MuGDmdDn?cWQ
zY9OsMZWI5%`<UN)LdbSABN7|RXzhAWH1O3V6#-UoLu>>_%!;S6E}t0XF;C!l*+K|E
zHiMc;57V~B&-5pgL^`*JlDm>Z#(0`Kp7M9b@E>+GIq3xGq&G9O5?*5fM!>Sr<>c}z
z6ZBI&fNBA=pt04Ti_&^cZKNV_1x>&&e=4}E^IFNB=@+R#_rrLU-XoM=_k@bOP5`lY
zpO{0g)%enJJm$;~qznGtA+4YM>G9K(*}Y3u*y|0uU`X>QI$leJ0wxpEM0?u7ieWnR
zS_yWnh~!qvjz&5Vfff7vME-_+eD?G*)Cw)}-;^Ds%UgjPzsV9J_fO!Z{?38+&>+TN
zzKuk0M8@3mhcL)Q9M|<+APG=R=88u$n-q=diu6}>{qaWJRy~r|%IeVP!dK|9Wj1me
z6mIoR06+DMkQj3tb?+*$ZrKT_D91tIop#KgzD@L7s>?Raio$7Krm*py7iz_DSY>R+
zc8n19T8CxW-n^M8W7G;>AAF$8G@ST59_NOd#F--b5S(Tx!E2fch|-gAxXiPxrrR2_
zbe#w6m5+u)!l`ID#|?Z2CgA6>&#+5(JH9o(PZ!E2QX9o{@VoyF{qJQN#>TraIv>h$
zhxt~J@;F95*y(`uDPOqhvmOtc-JsG>@*&%4H{3pRpZ>N_A}iNd<9Gcw7?mJ{$KIQ<
z>bpL}*{7z^l;#SxFCK!=1|ezv?F-p{Qh4I%2v%pGJuq`uL!VwGQ?TqfxPQ)oQH~n)
zO_i?j3ro3s#@6T^@|Z>y3TRJ>F%vj7l`E?wcs)UhnP6)QOFKVvXVp((WV$Pk?d_oB
zcchT6pj8+>e;3~UF`lt#+fI5kPm`DBOK^$2E-sf8pnOUOn3#v)q{%&K@wARPu$Ln;
z;sH!RX$a^#WY9VGd!a0$l_dXC!Q$2j3_0RNUNaLpeHULW_OT(hW04Ec{D6&7Pmw9s
zCqJ$gps`~DEc~f~GRwb0yHznpPIQCSK5fj{@HJ%j7zU2~xPq3gImF>?4mL^OfC7DA
z&|W45D;#)w>p?kg^C%!2<D}rl^K7_q`vhD|@1*y{Qb}L_LQI29>{A;_1&$N&><lSX
zou>%?CRfR0$7cGo<0ZMJ9YXbTnz{d;TauN|m6#qe4&)N;D`y>!1m8*9;N_S9aB{FB
zZo7Q}<mS24rgwnj3yOu24Za*Fu>n3=oJ9=zg!+wvY-D#kJ5uC*a$fZZv-VYhyMY+H
zYK0<y#rO>x9AgAgODpNYuLk`07n^unx|lyWSj@|g-o(1JEy9Mi*6^)iE$`*0%ht?)
z!e<TpvnveC(QczPnIa4W?S~;~qw)~FW;9c&7$wjgGQxL#3dF`&2w!fb@U2#3d71o7
zxOLT@pOhl<)r48FFS{nOVP;P3`X4_0L#@lKX^AbLD&fU0IQ)ixveb<oYx^JjW5Hqm
zQ>`vca#P|HZ*1mwbxO{cZMz3*j{@1IwS)M6+*!8GP>yvK`I7|~dO&PqNM&SFC9nOx
zo|pd@kN4JTvP;5p*aWi)*u5%{-Ewy-dvwHjT6x)m)&8!*`^3%R@4i0It2xNQ(Fg?P
zUQKp--X31;P#Ql&Vluz~WxGg={Y=)(oGbG3mC(xl8`($VBX}|4QS4qSMRLV1LAa>S
zwG5lY;@HJ}<-aCYeWDbvH>#DrXvDG1oWmeJb37iszlF{J@1aPKm@8bh{|G&4nL;lu
zUrKk5lA#eFtBo7xXoA}6Y(_f$FDcU)gOQJXsK-Gm>h$d*V?WxQxJlPDazm=5-s~9j
zaCs;$9}J^h=UFbKL6znfi(y;sYGLB7S0vL&3tcJ>fI?{w&D-~ddf0OGr>Nc!Ho3vs
zJxU`lg375ww=|~QIL%p&PbCH~I!ITy6M6JDn^=#rC2l_x&|%t0?2ohJ#;**e`-3Jk
z6-k$vq|8*>psIm>&&}|f?qVvtbr<7p@R2A+&7*c2JF&p0i0r$0gIMo2f|K=+D_{Nc
zqlGH|%m&95*s0!0UOoCv(r137El-NLhDD0-8O=a)*&SlY*r9c59zCd$Ousu?lQ*xA
z(Mv|ZxyhZYY5&p`y!muL?mjjWgJ+0p`|)e3!H08|5|`$Zf7|B^q9@Nm*XzH9xjuzl
za=93FyQPMQS6E?>t}HgM(1(0$fpE3cXELNzgw6&L^vqF1vNLZ4Cg-J~v8)S5ev%Zi
z1U7W5>|D_q?@Cy`)dCemN5;9zKv1N9pweAz6r&cr&{$zE(A=-L82>UEW|XWXU4AKz
zIcoTWd39+D*S`O!z`uIZJi`UkM7>ppNgZX&M4j7CUY!r+Hazziiq$R^dc6E4^z1l)
zDe&xau5M41;7juc?)7&`x^BX8!R7TS+##L$f*$2hjQ!Mi#)tl#WE>I|$aj5n&WVo}
zT-WQH7n)B6uIfQt)5I!fZkQ@p?LLzWmP}yQkI|zR_7cSD%ywb)(wp2U->-r=JvYJB
z3MsNgnWdZdYmk7KdSn+rz<n{25dOS>RbXXpNJpJND;%_cZ|q|Gs?u7uL-50{h8r+8
zWU`N?89Uv6Zv5tlyzuUwPGRvic}C^@1g6;_bzYIjUSZ407-86i7{<+~hm(wyrMF|X
zNRB$tq}G?*uG4>nI|>?@eVvlDE?dZCtDa&G30xTd@pqxYY-0g-eh`+OlA=Cqs+gxo
zvV>*NHgehb-WwmPUc=1a9>N{LX2BcRQ^M;b-*GlEssKG5s4#LexYz5^z^S_#bK5oK
zT#W*btSlm0C&%#@!((ZrtjN<cN*a@28KU8f5-eBzKz}#xW3@Duc%Q-dA`jPiJYy`!
z-ps8=^R0ytnkj>G)))c_oyLASyB=e_E?|HBc06;V7o`5VlH=o3;c3SbeA{ZqpOSn9
zzfXyQ*OU;{w;#n1#$ATrMGqi!O)9<>OT?^C-=OPLHgz4?Oh-m%VMW<QDDPPg!P+u>
z&6J6(^YRFY@^Zo%mm=7>T!wEqFkn@$jpbAQqF9A`HD2uu;J!89;P-YJnwOanDrz}d
z>o25&oG>i+ybdq-*`T@Dd{%C167={bqN(F&5-aLa-P?46AO0%EPbennaWF~L#};QA
zJUsalTP;3XvJsqKFuePL0IZ!gg3aj_=iOuUz$g7AI_&dCpR9MB>9bP$DIy!iOmEL?
zdl!JGwjO7lqKdHo*ILla7KJ2=eEQsqpLl~zM~{Ki_#tf?h=q?3*#Eo3du49pC-nDV
zx``hnxnTjH(E1QxPSPi?3&Sz!gg=Uvz5-L{>1^T4XJqfe$*^4AkLZo<hLGxNa?{F>
zANvx)MlzIDag1S|x0a&NQ=j*b+aL&J!`W9u{eXQQP~Fsojo&<3r;Od~>|?R)ck&&R
z&y8k{<%8KN+-7XJAj{jD1;7-IZtV3Ybnwq~-tT2Ln2+T-G1UY3ct;|7j}N4B8+6#U
zqs93hn%2~Q*J3uzwG*5~zWq>TJNERP82<Q^Dv0mDz>hmG&+2vO;n#JgoMNyhug_ZX
zhi1mIT_rN)P<}Uz?2O<IPipXAzbxYYXPSUZ|8{n6qbIrU>5AFYCh?b-#`70**RrpU
zE<l^>zUX<V4UB_BK+5$Ie5pA|#`kLS9qOtmb*Kh*tEuv;;ab?y83>OT24dRMb?nlL
z2KvDEBzbu#fHyD{@b7mF!{GgVrnIA!cOMMlTvk7ar_xdQwL)Y_iPyjgqcrw^iLCj5
z%lMx=;P^QL%U@ft|28dUpN+O+t9r9Q2Q`5j3_*+dM516ioB!*e1h?fMLwlVK?-(!3
zpP2j^dm9b$Zp&!))#_e4>F|8at=)nP|DMC9<+8kqw-fru&Bm|A!EjJritqe<i=5G$
z#N2vj1ND2y@Cwbf;P9~y6aC^@pKFEuqs2ckQY{c=L7cy{co%>6^Cq@3&jkjo16hgr
zC!kkxHyv4ioyynxl6Cs!*icr47wcl!gr`EZO{vBD=6|@;^d{Z*K@z@a%CX_jn_v-_
z4mH!`ag0c7Xq{O{53St~pSSw3Rc#r(tMV~^<yQ;7<=Y^1D0G7C$t`peW5#av+YjeA
z=%cr?2HW|=2`6ce=DojJ^Y!Y7v0r>6DwbTQ#w}N{YV&Kp%<m!d-7Fuy-+jUR>48MD
z+aGhLn8G;AW^@lM#L5{7{CT&z?3xiXpt4@bjaKdgS>;gLRhY*aie<7>TSeTD$sP<2
z>!DT=nPf(Y8f##aj74!zQ6H>%yD9dp<D5{~GA0oNs?H%?i)0r`=JFxoxiHY%&lFsY
zqCVA&dBwbS?9@JMOp4zM|NTKYv%Upgb2j3*|2Fco?yK|dPeOUA9cujFy6Ny;Rh_N5
z_nNJ<m<+z3QgLTqKBRvM<Ab~6Sm*rd{DIzQu+ur5|8-BBU1polMqU`8$1^4Pl<qdz
z=5Q7C_ReM%FFfLZnXF~Co43QX1#{VJQKNVVJ9%Dufi7=REW^9#OJY{3KXfmv0<+jc
zHfi=$l!@}EamI-da_1F0Q=tJY{+rA@bp^oP)bF_G-F1GRg*0uuX~k>RhqJl^46kDU
zhAh>Nhbg_Dz-gM`fLt-S874s8*-jK2dzLr4P=G&<sIZO0ZtU)dLSD0=m;3)bS5@sK
z=He7%UPEmQ+i1Iye|Xi1y&|_oxXmw^SCCTS$3L0J-W=l(2hHa5QDs(;yEU3+636ml
z^_A>Bx$CUP?L@ZTEEL7<t`X;wHLS4Y0VJ<kf(@To);KE$@)jn-lbl*ceC`eS6Y&z<
zb!6zN#18Jk&@{RrKNc<bjl(5L8>rZ&944T!h`aN}A73sOqctZ^K>r1x8cYK+_4)WK
zL<^s<8z!VE2R%<F;PuCAE`1$1O2j@?k~We5wy1RnN`&smA9C#=Grt~ND)jl;0gK4P
zc4g3x{l}yj-h%vp_vnJ9@i4n_n2dY229470F%nBfnrOa0XQo2Y^@M=@EeN397E^g^
z6Lr2ZID@SC`57k}RYQ08N;DCl#r`vWiF+*5Kz8m6=A=?4u}@eJI}*kC;R0#cVEY+0
zd!^Y!>#M*x!wqh6Z|E!I5&U%DW-`TU797!k5AVO1!{b}J7^zcFWJ|JfPxt_GpChSZ
z>vH^Y+6p6%*uX?kgrhpj*ttRiB$}>LESOBcTslF-=8xpR70-dcvDdi^VGkjLISa(6
z2ro|yz(UcK*@aV2(I+a=__uW@U01%1uJuiX<gdM`mYR<j97Q$F<n#EdJCWq6pJU#)
zbucP+)u3p8n#Pa5g5wkqW9CpSUQCaI^c&;YPdmk-{9XbWWiBI!NG*n#D)V#F<>B5D
zNizGvS$HL4Umc{M(j!lvp;gp=w%I!wzHE!8XRe&aZBw7XE?*^<$ti{9G3#O1(q)(@
z(#d`VRX}BfEOpgd3ETQJnb`?Z!iBM6X#UZaZudWck?X`E`bHUfvonEMV-Z?CR^~5W
z566kNc6d2oq{H9KhEuvBRB^+2UNc_EomssEw#wXqEvjB5C;uP@d=s$&>!f*3_6bqk
zcnP+N*JFG56|P3Y5J&o6B@gRU$&pVE*mk~=njh6<I%3jb#~m^Lr}YVvcqX429FXAi
zcdtaP|90UQx6`1sAe~mqPDcazYxJ!{1sKjX<a2ifqUlmC0^8r>`#KRvJ4KOb<c$X}
zVI@X=c*<>1b%w=)NumI+3e<_qAhmr7MBis2KV%SrHKDs;)bo0{Eb=p-?TcoU!*gJ$
zy8=>Tni(zgU93!kE!!M*hFLVU3M&KkaZY(Lwog?=+qG|)y*;*sX`K#Ds~wR%=)|zf
zbdV}<;<hM`<R9*m;VX;}qO(RMG;163Ywn$=Zfy=Iof$@6eQ2jMSBS8U1Z#d-b|KEu
zc@HIf0(tGZaePaQ2Mp%+KvG^6HJakaR;pOg-~%%tW_mhwZQ6tO*HfV>d=mb>Mqtl6
z3%;^p9R_4wL~-*7ey3Cj|K;I6(0MY2f7GQ1y%MMJWMvY&y6ge=uX{&y?#{xAQ$oNw
zTnuK8)#GoL#e%W12fxWIkiB$g1KSjo4*HR8^x?xiew%9txi?ER>SW_WHdW+jj(Ssx
zg>vR_z#@=uw$9;Q17`8=BL9O#l0DtB%7~8o5eK6S?vu*_CgAM#7~Q}4GYig7fjdEm
z`J?NN!Lv3UeJ5>THH(zlT8~U@(q;I@Lu*-?5$o}6)GGec+I;j|au@3zBEZ|upFh#v
zPPg&lP`&2>>M!|8zFv0V)cPgZEBlmr9k*%hoV6Of=Bxei+_4gC7QBWWo5!&KqXK^J
z`%d1@yF%Y7H{ta^2k=zdWsv@U1(rC6uu8ZPw!QvKRqE!T0IuNna*=Owu#QWO-N~yo
ztYFt)HRpHeZ=&-h)baAGey}&s!Q{1Ws2%;rxb1xyvHZ1)kG=XD93NJZjFTShI_G2H
zJZT!MEYfy^VjH0RMjWtP#;|=0^l)Eb7qr{VWtW>6gL1wef7qd(R82ohDs)DWF6&^l
zTA|P5kQ5jgjU(UKP7<#+ojuT5f{~VbtWL8(+Rc~;UZdy1;5Tcq&eO)MmWSxu7lhYt
zzo#psXJP(Cfe;>xv+euJ$jGS#+-_Wfqgy9q)tl+O=dVGL*D)2cC(mO=<I>oQd0xC&
z#}w}Hz!~D68jgKq6L7mp4Jg{)rOCpxxFaTqaz_@y<5j<KM)_vGWX(i=K}rGEd{P6&
zYiIF*d?Hi+q>endQ(^SG^Qm{J1vkHcH}U!^$K{WRAcfi6iCaf6m3`xYkIx;!yW%s5
z!5W^KEu~GK{=3BC?`ks6rjB|j@?7KHHbQM<sVM&o^&fs9V-t;G^R;XYzUE4FjLe9>
z@?~6Q>I~`V3|Du^fr+{fe1FA}B10qL#+RO8Z8sCtFNi^>=1xdos|(7JOVD=b6TGmt
z8PA)yVBT&oJft!Wf)opBSW!6YHn`zwyH;YF`<A};n1vNP<ngNOE&B1eD~c_PBOBKx
z!^Y9-@VrVBe3~W;t@rE^WJmp@UZ%&Gqc8`mja~6-%3geV-VD@FBUH{i!Ck5{MUB7T
zxm7)XNr*E~{=JbPK?gMG*ZUmf(<p_tApscubR`wsPKRgpWn`aXG&pOWq3tFgai*;d
zFYOx%8Vylcapx|zmQletks?2MrJqV&3I@$`2Xa3(g2)vnioAOS{qjAbNVQb3VU*~7
zj!i+YKl-TSp-M}%Pm<#qvlt)AmP+4~9n7?8f0>(RYE(9&lv7$!$Z>}cF%_|PG&3TH
zKAvqzeg(%eVs6}}pHKU^7w)H+z@0})VgXBJ`eGP$?};MbSqujyYZ#-+tC>*G9Gd^6
zh&Gs6l4Pw?x+YE>-jrAiCJ$#4q54Kr@g#;g#b}cY(a~g6R}a%{vO~~YrB03w$}lb)
zD(KuBW2yRtz4YsGRqUA*L7}dV9G~)x8&g|CcfFG4jGgL9@K__pT{DfAOcK&gPLc{c
zct-tbEt7aYmAG3faRd5$h~TaoxjXU_x7tC51ew)yVn;q+m?$Mfo7ykYW9K9pzpw|)
z!j6sfz_}3Sz>zwE>PTDCXMUVa>|w}iEiLZ*#5YXV&n(hZ^&hA5XC2AYY$XF(Tgcrr
zog^^RTp;yU8%L_%r8kdKB7W)xRX(K4#Wx5T<CZ7FeS3BZ{oeVLklbTr`|KbtqJ|S#
zCa2S832*Y~#u)M<ejZg1jT1gy;m?FSh>@erCCs_A6gwi~skZ7H`dzw`j(t)HovLq{
zS10U2f9`rDq?Wuh+yXUI-ck46C6%$;x)>=xWl&9vq%k_5@#TI!)OC;&F)$;@k0u}3
zqZmyu=423q5Aj5ONP;zOjKposG?;cmp4Xk>1y!fcU_;6vwd_d6b=zX`spfGsNu5Ba
zN9jQ8noRmC<~-Bpm<NFm#j*cY8#(rN7!=pcfE)*r9{x`YR*QI*=gQCM&k5ss(}p#4
z@K-Lxcyz%pr8A`W?O9yda)9_s+u_}PnlO8uI+o#TkxynTG5sxKloIB`qWor(ySSIR
z^eYw83;&UUp=z=-qY(ZILh#;qbrF-e9r+$5%$47WI)Uljh~OdWI3&mCgdInPqGbH^
zG8<i$!;oMV_8m_kuRbcWf?#zx6cdW6@2|qRy4Qln(hzhHa6*OFsaX1TGF>A(2j%M&
zz+k}*<TlTT8<w@S#v&0tlmP2qZomQKh3MklMZ5l-qVFc|#ZTFDKomHRhP@Ubw>^>E
z>o^Fge-70aJ*Uk*<9L~cgOJ<17@ocJ#QT>Ii`MZ5{=9D{nuqV=(OCk_S-uQQ4zh4;
ztO?8-b(@~+D?k^m&G7d{5i~xyXMCF<$K0&m0aMLCFazsr(EVr>Ic95sd;iAaeQp^P
z_m4%Fo)NJ4W)$@-)?hP5tX<hn3p{jNoAj6a0nRD|wYG3nDwaU0>qj8`oDt445^)B5
zPLk`RE@ItTS1@+@LtDGf;fvZUSbwe_PIpgb3sasjf`Sw@Td|rpyh*^nd2`V^Ne8;y
zqHyq@KL*M*LMkBYEy~5c)(=E`@@*o1GZZ&OQZhPGj`!}V!B->P;KR-bs3`j#UC(&(
z;qn{})wF`bfe>`_SHN$pKf;JGj`MsSL>{j^%ybIUvHM#)Gr!>=6q&t8gS7>aCSihk
z_PcP|dvknYdK0ovc!A5AG`d83351=vg*ozKXi&5lJ$*)ybw|x`jNLaf`R`uR)AfvV
zdViZae`_=mB@Sa~>3EXgBmn`_9Z>LhEXKW)V=H!rQ0Z^MP+;uDoPg8V_4pwZ_EQ3T
zrWT-elmMGf>M^JEPeP|a^go`r4@T}+!Mr!wxF&D{bL+rrtUEuIcsI8rYoJEN#_7!)
z_c<AaH@hHq>0;QcGa57QY((`KD^RcvBmql*qLubc@LYEZ?AH1~=H7mC<ZC|0-4CT@
zr^AIh@5jMw5Mxj1GSHG6Or64~l3U|$(MhK{=sWxmrfDdm-2)MUb7c)^J|7h3;69vt
zVH~NORl(RzoezS!aX9RE9o!#ihyq=DsrB>_^pX9}?dbeOCX8+*5^+Cq^t>d}Hmw_I
z|5->|oQV^!*uc|AYr+1&17Z=g4F_JV6=iV%w$<yB2^+^?tyC(;*B9U%l1|09+QZ=?
z1&qr5LY#^gLiHQU>B}C2Z3DBwsgpoMsvG{u8ln}SvpK*1JwoHKGuZdO8=cb5;>*xN
z(&(qgyvy^&6ILzIi;?J-un0$ewMAq3a5QL}2p6Ne;XeO{>`V!QL;o!xo!V7Q%{K|&
zCUXPqI^97sf6B58Wu^I-{UWC9jg5$5oW!1zXo87@KBW1WEn4Yn&~powSjmez(7gRB
zxC9}i|MDzcG6{f%Yvj>r{1!-eT|qCha`@rjIAFvR7`eefl(=;O-kxRX$ETl}-1E*T
zzOn(+ZwD}Q|G46ce<AdRlQ=w`I)mu%OUGHi2>#VlBhOQieAxP#T*-03Ey<Di#UvJE
zJ>JttPHs>#%>hrI`^R`Rp2zUF;q*e138!!;k|;H%FdtWKg!UzW$XoeW_;f)R`5QV+
zDm&BY{Zt7UwlAaaHd}zFk0*DvMGfE8MWJIy6pSxb;$OdzMroy|#Olrm+@F($;S~p=
zJ+lmKh7EXk>sC6lAB6l7IsBDlhz+AV$j#fyfPQ&+J~E7!XO_dq%nlf`iQ&vwyMX_c
z{bc$pb-Z#k3_r%mz{#fdWZwom+}g*({ad+|@DFjK!V#Fg_zF4NI7B=K%`xh77Jk(e
zhv!qz!qx{4bnCu{#H3mc*LQD)+=1O>`??UUf1OK;ixa7qqN1Q{g@ERt`a}%dZi)ER
zU~;-U7b}-vhQYE`=%Q;6OGS0hkkwCcOj?Lb#urk<`D1Wd_EOxHu?(i1+eq!=0F>Y9
zp_oXE9XNOmxz-R7uiHd}>IG<a`7f=%7(p5eb4a?dlxBUZAjvVM#LXrF+IBj@?-G%Z
zzxNzsdndWS^bF|B&jk0Wn&6(I3~8+<M0(~WQKrOp5#Kc*Y<_3b<sx?OqeeTJy#7W8
z1CGJy5h`#jWJl$A{WU~GR5MQ%<y#!d*#sv3mqDr4iz!!+2irJx#<Tqw$&0fF%S$hz
zaQQ@Z-PB3-4KpBnvpaqn-$jaDuhLD&uF-&8RlJd(LDII$!!v0s7z&t6FIJ_|{|70T
B#2Ekp


From 214494131352183166e956d64e9227cb70b1377f Mon Sep 17 00:00:00 2001
From: Yurii Shyrma <yurii@skymind.io>
Date: Tue, 27 Aug 2019 19:57:59 +0300
Subject: [PATCH 27/56] Shyrma fix2 (#186)

* - further work on layer_norm

Signed-off-by: Yurii <yurii@skymind.io>

* - further work on layer_norm 2

Signed-off-by: Yurii <yurii@skymind.io>

* - correct helpers for svd cuda

Signed-off-by: Yurii <yurii@skymind.io>
---
 .../generic/transforms/layer_norm.cpp         |  13 +-
 .../ops/declarable/helpers/cuda/svd.cu        | 188 +++++++++---------
 .../layers_tests/DeclarableOpsTests15.cpp     |  28 ++-
 .../layers_tests/DeclarableOpsTests16.cpp     |   2 +-
 4 files changed, 129 insertions(+), 102 deletions(-)

diff --git a/libnd4j/include/ops/declarable/generic/transforms/layer_norm.cpp b/libnd4j/include/ops/declarable/generic/transforms/layer_norm.cpp
index 684d98d6d..8ab5fa32f 100644
--- a/libnd4j/include/ops/declarable/generic/transforms/layer_norm.cpp
+++ b/libnd4j/include/ops/declarable/generic/transforms/layer_norm.cpp
@@ -38,9 +38,13 @@ namespace ops  {
         const bool isNCHW = block.getBArguments()->size() > 0 ? B_ARG(0) : true;       // INT_ARG(9): 0-NCHW,  1-NHWC
         const int dimC = isNCHW ? 1 : input->rankOf() - 1;
 
+        REQUIRE_TRUE(gain->rankOf() == 1 && gain->sizeAt(0) == input->sizeAt(dimC), 0, "LAYER_NORM OP: wrong shape of gain array, expected is {%i}, but got %s instead !", input->sizeAt(dimC), ShapeUtils::shapeAsString(gain).c_str());
+
         NDArray* bias = nullptr;
-        if (block.width() > 2)
+        if (block.width() > 2) {
             bias = INPUT_VARIABLE(2);
+            REQUIRE_TRUE(bias->rankOf() == 1 && bias->sizeAt(0) == input->sizeAt(dimC), 0, "LAYER_NORM OP: wrong shape of bias array, expected is {%i}, but got %s instead !", input->sizeAt(dimC), ShapeUtils::shapeAsString(bias).c_str());
+        }
 
         std::vector<Nd4jLong> longAxis = ArrayUtils::toLongVector(axis);
 
@@ -80,13 +84,16 @@ namespace ops  {
         const bool isNCHW = block.getBArguments()->size() > 0 ? B_ARG(0) : true;       // INT_ARG(9): 0-NCHW,  1-NHWC
         const int dimC = isNCHW ? 1 : input->rankOf() - 1;
 
+        REQUIRE_TRUE(gain->rankOf() == 1 && gain->sizeAt(0) == input->sizeAt(dimC), 0, "LAYER_NORM_BP OP: wrong shape of gain array, expected is {%i}, but got %s instead !", input->sizeAt(dimC), ShapeUtils::shapeAsString(gain).c_str());
+
         std::vector<int> axis = *block.getIArguments();
 
         std::vector<Nd4jLong> longAxis = ArrayUtils::toLongVector(axis);
 
         if(bias != nullptr) {
+            REQUIRE_TRUE(bias->rankOf() == 1 && bias->sizeAt(0) == input->sizeAt(dimC), 0, "LAYER_NORM_BP OP: wrong shape of bias array, expected is {%i}, but got %s instead !", input->sizeAt(dimC), ShapeUtils::shapeAsString(bias).c_str());
             // eps->reduceAlongDimension(nd4j::reduce::Sum, dLdb, {0}, true);
-            eps->reduceAlongDimension(nd4j::reduce::Sum, dLdb, ShapeUtils::evalDimsToExclude(input->rankOf(), {dimC}), true);
+            eps->reduceAlongDimension(nd4j::reduce::Sum, dLdb, ShapeUtils::evalDimsToExclude(input->rankOf(), {dimC}));
         }
 
         NDArray standardized(input->shapeInfo(), false, block.launchContext());
@@ -99,7 +106,7 @@ namespace ops  {
 
         standardizeOp.execute(inputs, outputs, targs, longAxis, bargs);
         standardized.applyPairwiseTransform(nd4j::pairwise::Multiply, eps, &standardized, nullptr);
-        standardized.reduceAlongDimension(nd4j::reduce::Sum, dLdg, ShapeUtils::evalDimsToExclude(input->rankOf(), {dimC}), true);
+        standardized.reduceAlongDimension(nd4j::reduce::Sum, dLdg, ShapeUtils::evalDimsToExclude(input->rankOf(), {dimC}));
 
         nd4j::ops::standardize_bp standardizeBp;
         // eps->applyTrueBroadcast(nd4j::BroadcastOpsTuple::Multiply(), gain, dLdx);
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/svd.cu b/libnd4j/include/ops/declarable/helpers/cuda/svd.cu
index c07c9adf8..da4f5cc86 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/svd.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/svd.cu
@@ -100,7 +100,7 @@ static void inverseColumnSignCudaLauncher(const int blocksPerGrid, const int thr
 BUILD_SINGLE_TEMPLATE(template void inverseColumnSignCudaLauncher, (const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t* stream, void* vu, const Nd4jLong* uShapeInfo, void* vv, const Nd4jLong* vShapeInfo), FLOAT_TYPES);
 
 //////////////////////////////////////////////////////////////////////////
-static void svdQR(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, NDArray& U, NDArray& VT, const bool fullUV, const bool calcUV) {
+static void svdQR(nd4j::LaunchContext* context, const NDArray* A, NDArray* S, NDArray* U, NDArray* VT, const bool fullUV, const bool calcUV) {
 
     // since cusa api cusolverDnDgesvd/cusolverDnSgesvd have following constrain on input matrix A: A_rows >= A_columns && A_order = 'f'
     // we make this function to have deal with 2 valid cases only:
@@ -113,59 +113,59 @@ static void svdQR(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, ND
     // U  [m, m] or [m, n] if fullUV = false and m > n
     // VT [n, n] or [m, n] if fullUV = false and m < n
 
-    if(A.rankOf() != 2)
+    if(A->rankOf() != 2)
         throw std::runtime_error("svdQR: rank of A array is not equal 2 !");
 
-    auto m = A.sizeAt(0);
-    auto n = A.sizeAt(1);
+    auto m = A->sizeAt(0);
+    auto n = A->sizeAt(1);
     const int minDim = m < n ? m : n;
-    const char orderA = A.ordering();
+    const char orderA = A->ordering();
 
     if(m < n)
         throw std::runtime_error("svdQR: due to cuda api input constrains given shape of A array are not valid !");
 
-    if(ShapeUtils::shapeAsString({minDim}) != ShapeUtils::shapeAsString(&S))
+    if(ShapeUtils::shapeAsString({minDim}) != ShapeUtils::shapeAsString(S))
         throw std::runtime_error("svdQR: wrong shape of S array !");
 
     if(calcUV) {
 
-        if(fullUV && ShapeUtils::shapeAsString({m,m}) != ShapeUtils::shapeAsString(&U))
+        if(fullUV && ShapeUtils::shapeAsString({m,m}) != ShapeUtils::shapeAsString(U))
             throw std::runtime_error("svdQR: wrong shape of U array !");
-        else if(!fullUV && ShapeUtils::shapeAsString({m,minDim}) != ShapeUtils::shapeAsString(&U))
+        else if(!fullUV && ShapeUtils::shapeAsString({m,minDim}) != ShapeUtils::shapeAsString(U))
             throw std::runtime_error("svdQR: wrong shape of U array !");
 
-        if(fullUV && ShapeUtils::shapeAsString({n,n}) != ShapeUtils::shapeAsString(&VT))
+        if(fullUV && ShapeUtils::shapeAsString({n,n}) != ShapeUtils::shapeAsString(VT))
             throw std::runtime_error("svdQR: wrong shape of VT array !");
-        else if(!fullUV && ShapeUtils::shapeAsString({minDim,n}) != ShapeUtils::shapeAsString(&VT))
+        else if(!fullUV && ShapeUtils::shapeAsString({minDim,n}) != ShapeUtils::shapeAsString(VT))
             throw std::runtime_error("svdQR: wrong shape of VT array !");
     }
 
-    NDArray* pA  = const_cast<NDArray*>(&A);
-    NDArray* pS  = &S;
-    NDArray* pU  = &U;
-    NDArray* pVT = &VT;
+    NDArray* pA  = const_cast<NDArray*>(A);
+    NDArray* pS  = S;
+    NDArray* pU  = U;
+    NDArray* pVT = VT;
 
     std::vector<NDArray*> toDelete;
 
     if(pA->ews() != 1 || pA->ordering() == 'c') {
-        pA = A.dup('f');
+        pA = A->dup('f');
         toDelete.push_back(pA);
     }
 
-    if(S.ews() != 1) {
-        pS = S.dup('f');
+    if(S->ews() != 1) {
+        pS = S->dup('f');
         toDelete.push_back(pS);
     }
 
     if(calcUV) {
 
         if(pU->ews() != 1 || pU->ordering() == 'c') {
-            pU = U.dup('f');
+            pU = U->dup('f');
             toDelete.push_back(pU);
         }
 
         if(pVT->ews() != 1 || pVT->ordering() == 'c') {
-            pVT = VT.dup('f');
+            pVT = VT->dup('f');
             toDelete.push_back(pVT);
         }
     }
@@ -183,9 +183,9 @@ static void svdQR(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, ND
 
     // query working space of SVD
     int lwork = 0;
-    if(A.dataType() == DataType::DOUBLE)
+    if(A->dataType() == DataType::DOUBLE)
         status = cusolverDnDgesvd_bufferSize(handle, m, n, &lwork);
-    else if(A.dataType() == DataType::FLOAT32)
+    else if(A->dataType() == DataType::FLOAT32)
         status = cusolverDnSgesvd_bufferSize(handle, m, n, &lwork);
     else
         throw std::invalid_argument("svdQR: given data type is unsupported !");
@@ -195,7 +195,7 @@ static void svdQR(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, ND
 
     // allocate memory for dWork
     void* dWork = nullptr;
-    cudaError_t status2 = cudaMalloc((void**)&dWork , A.sizeOfT() * lwork);
+    cudaError_t status2 = cudaMalloc((void**)&dWork , A->sizeOfT() * lwork);
     if(status2 != cudaSuccess)
         throw cuda_exception::build("svdQR: cuda failed !", status2);
 
@@ -226,11 +226,11 @@ static void svdQR(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, ND
     NDArray::prepareSpecialUse({pS, pU, pVT}, {pA});
 
     // choose appropriate cuda gemm api depending on data types
-    if(A.dataType() == DataType::DOUBLE) {
-        status = cusolverDnDgesvd(handle, jobu, jobvt, m, n, reinterpret_cast<double*>(pA->getSpecialBuffer()), lda, reinterpret_cast<double*>(pS->getSpecialBuffer()), reinterpret_cast<double*>(pU->getSpecialBuffer()), ldu, reinterpret_cast<double*>(pVT->getSpecialBuffer()), ldvt, reinterpret_cast<double*>(dWork), lwork, reinterpret_cast<double*>(rWork), devInfo);
+    if(A->dataType() == DataType::DOUBLE) {
+        status = cusolverDnDgesvd(handle, jobu, jobvt, m, n, reinterpret_cast<double*>(pA->getSpecialBuffer()), lda, reinterpret_cast<double*>(pS->getSpecialBuffer()), calcUV ? reinterpret_cast<double*>(pU->getSpecialBuffer()) : nullptr, ldu, calcUV ? reinterpret_cast<double*>(pVT->getSpecialBuffer()) : nullptr, ldvt, reinterpret_cast<double*>(dWork), lwork, reinterpret_cast<double*>(rWork), devInfo);
     }
-    else if(A.dataType() == DataType::FLOAT32) {
-        status = cusolverDnSgesvd(handle, jobu, jobvt, m, n, reinterpret_cast<float*>(pA->getSpecialBuffer()), lda, reinterpret_cast<float*>(pS->getSpecialBuffer()), reinterpret_cast<float*>(pU->getSpecialBuffer()), ldu, reinterpret_cast<float*>(pVT->getSpecialBuffer()), ldvt, reinterpret_cast<float*>(dWork), lwork, reinterpret_cast<float*>(rWork), devInfo);
+    else if(A->dataType() == DataType::FLOAT32) {
+        status = cusolverDnSgesvd(handle, jobu, jobvt, m, n, reinterpret_cast<float*>(pA->getSpecialBuffer()), lda, reinterpret_cast<float*>(pS->getSpecialBuffer()), calcUV ? reinterpret_cast<float*>(pU->getSpecialBuffer()) : nullptr, ldu, calcUV ? reinterpret_cast<float*>(pVT->getSpecialBuffer()) : nullptr, ldvt, reinterpret_cast<float*>(dWork), lwork, reinterpret_cast<float*>(rWork), devInfo);
     }
     else
         throw std::invalid_argument("svdQR: given data type is unsupported !");
@@ -242,11 +242,11 @@ static void svdQR(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, ND
 
     NDArray::registerSpecialUse({pS, pU, pVT}, {pA});
 
-    S.assign(pS);
+    S->assign(pS);
 
     if(calcUV) {
-        U.assign(pU);
-        VT.assign(pVT);
+        U->assign(pU);
+        VT->assign(pVT);
     }
 
     for (int i = toDelete.size() - 1; i >= 0; --i)
@@ -266,62 +266,62 @@ static void svdQR(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, ND
 }
 
 //////////////////////////////////////////////////////////////////////////
-static void svdJcb(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, NDArray& U, NDArray& V, const bool fullUV, const bool calcUV) {
+static void svdJcb(nd4j::LaunchContext* context, const NDArray* A, NDArray* S, NDArray* U, NDArray* V, const bool fullUV, const bool calcUV) {
 
     // A [m, n]
     // S [n]
     // U [m, m] or [m, n] if fullUV = false and m > n
     // V [n, n] or [n, m] if fullUV = false and m < n
 
-    if(A.rankOf() != 2)
+    if(A->rankOf() != 2)
         throw std::runtime_error("svdJcb: rank of A array is not equal 2 !");
 
-    auto m = A.sizeAt(0);
-    auto n = A.sizeAt(1);
+    auto m = A->sizeAt(0);
+    auto n = A->sizeAt(1);
     const int minDim = m < n ? m : n;
 
-    if(ShapeUtils::shapeAsString({minDim}) != ShapeUtils::shapeAsString(&S))
+    if(ShapeUtils::shapeAsString({minDim}) != ShapeUtils::shapeAsString(S))
         throw std::runtime_error("svdJcb: wrong shape of S array !");
 
     if(calcUV) {
 
-        if(fullUV && ShapeUtils::shapeAsString({m,m}) != ShapeUtils::shapeAsString(&U))
+        if(fullUV && ShapeUtils::shapeAsString({m,m}) != ShapeUtils::shapeAsString(U))
             throw std::runtime_error("svdJcb: wrong shape of U array !");
-        else if(!fullUV && ShapeUtils::shapeAsString({m,minDim}) != ShapeUtils::shapeAsString(&U))
+        else if(!fullUV && ShapeUtils::shapeAsString({m,minDim}) != ShapeUtils::shapeAsString(U))
             throw std::runtime_error("svdJcb: wrong shape of U array !");
 
-        if(fullUV && ShapeUtils::shapeAsString({n,n}) != ShapeUtils::shapeAsString(&V))
+        if(fullUV && ShapeUtils::shapeAsString({n,n}) != ShapeUtils::shapeAsString(V))
             throw std::runtime_error("svdJcb: wrong shape of V array !");
-        else if(!fullUV && ShapeUtils::shapeAsString({n,minDim}) != ShapeUtils::shapeAsString(&V))
+        else if(!fullUV && ShapeUtils::shapeAsString({n,minDim}) != ShapeUtils::shapeAsString(V))
             throw std::runtime_error("svdJcb: wrong shape of V array !");
     }
 
-    NDArray* pA = const_cast<NDArray*>(&A);
-    NDArray* pS = &S;
-    NDArray* pU = &U;
-    NDArray* pV = &V;
+    NDArray* pA = const_cast<NDArray*>(A);
+    NDArray* pS = S;
+    NDArray* pU = U;
+    NDArray* pV = V;
 
     std::vector<NDArray*> toDelete;
 
     if(pA->ews() != 1 || pA->ordering() == 'c') {
-        pA = A.dup('f');
+        pA = A->dup('f');
         toDelete.push_back(pA);
     }
 
-    if(S.ews() != 1) {
-        pS = S.dup('f');
+    if(S->ews() != 1) {
+        pS = S->dup('f');
         toDelete.push_back(pS);
     }
 
     if(calcUV) {
 
         if(pU->ews() != 1 || pU->ordering() == 'c') {
-            pU = U.dup('f');
+            pU = U->dup('f');
             toDelete.push_back(pU);
         }
 
         if(pV->ews() != 1 || pV->ordering() == 'c') {
-            pV = V.dup('f');
+            pV = V->dup('f');
             toDelete.push_back(pV);
         }
     }
@@ -362,10 +362,10 @@ static void svdJcb(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, N
 
     // query working space of SVD
     int lwork = 0;
-    if(A.dataType() == DataType::DOUBLE)
-        status = cusolverDnDgesvdj_bufferSize(handle, jobz, econ, m, n, reinterpret_cast<double*>(pA->getSpecialBuffer()), lda, reinterpret_cast<double*>(pS->getSpecialBuffer()), reinterpret_cast<double*>(pU->getSpecialBuffer()), ldu, reinterpret_cast<double*>(pV->getSpecialBuffer()), ldv, &lwork, gesvdjParams);
-    else if(A.dataType() == DataType::FLOAT32)
-        status = cusolverDnSgesvdj_bufferSize(handle, jobz, econ, m, n, reinterpret_cast<float*>(pA->getSpecialBuffer()), lda, reinterpret_cast<float*>(pS->getSpecialBuffer()), reinterpret_cast<float*>(pU->getSpecialBuffer()), ldu, reinterpret_cast<float*>(pV->getSpecialBuffer()), ldv, &lwork, gesvdjParams);
+    if(A->dataType() == DataType::DOUBLE)
+        status = cusolverDnDgesvdj_bufferSize(handle, jobz, econ, m, n, reinterpret_cast<double*>(pA->getSpecialBuffer()), lda, reinterpret_cast<double*>(pS->getSpecialBuffer()), calcUV ? reinterpret_cast<double*>(pU->getSpecialBuffer()) : nullptr, ldu, calcUV ? reinterpret_cast<double*>(pV->getSpecialBuffer()) : nullptr, ldv, &lwork, gesvdjParams);
+    else if(A->dataType() == DataType::FLOAT32)
+        status = cusolverDnSgesvdj_bufferSize(handle, jobz, econ, m, n, reinterpret_cast<float*>(pA->getSpecialBuffer()), lda, reinterpret_cast<float*>(pS->getSpecialBuffer()), calcUV ? reinterpret_cast<float*>(pU->getSpecialBuffer()) : nullptr, ldu, calcUV ? reinterpret_cast<float*>(pV->getSpecialBuffer()) : nullptr, ldv, &lwork, gesvdjParams);
     else
         throw std::invalid_argument("svdJcb: given data type is unsupported !");
 
@@ -374,7 +374,7 @@ static void svdJcb(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, N
 
     // allocate memory dWork
     void* dWork = nullptr;
-    auto status2 = cudaMalloc((void**)&dWork , A.sizeOfT() * lwork);
+    auto status2 = cudaMalloc((void**)&dWork , A->sizeOfT() * lwork);
     if(status2 != cudaSuccess)
         throw cuda_exception::build("svdJcb: cuda failed !", status2);
 
@@ -383,11 +383,11 @@ static void svdJcb(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, N
     NDArray::prepareSpecialUse({pS, pU, pV}, {pA});
 
     // choose appropriate cuda gemm api depending on data types
-    if(A.dataType() == DataType::DOUBLE) {
-        status = cusolverDnDgesvdj(handle, jobz, econ, m, n, reinterpret_cast<double*>(pA->getSpecialBuffer()), lda, reinterpret_cast<double*>(pS->getSpecialBuffer()), reinterpret_cast<double*>(pU->getSpecialBuffer()), ldu, reinterpret_cast<double*>(pV->getSpecialBuffer()), ldv, reinterpret_cast<double*>(dWork), lwork, devInfo, gesvdjParams);
+    if(A->dataType() == DataType::DOUBLE) {
+        status = cusolverDnDgesvdj(handle, jobz, econ, m, n, reinterpret_cast<double*>(pA->getSpecialBuffer()), lda, reinterpret_cast<double*>(pS->getSpecialBuffer()), calcUV ? reinterpret_cast<double*>(pU->getSpecialBuffer()) : nullptr, ldu, calcUV ? reinterpret_cast<double*>(pV->getSpecialBuffer()) : nullptr, ldv, reinterpret_cast<double*>(dWork), lwork, devInfo, gesvdjParams);
     }
-    else if(A.dataType() == DataType::FLOAT32) {
-        status = cusolverDnSgesvdj(handle, jobz, econ, m, n, reinterpret_cast<float*>(pA->getSpecialBuffer()), lda, reinterpret_cast<float*>(pS->getSpecialBuffer()), reinterpret_cast<float*>(pU->getSpecialBuffer()), ldu, reinterpret_cast<float*>(pV->getSpecialBuffer()), ldv, reinterpret_cast<float*>(dWork), lwork, devInfo, gesvdjParams);
+    else if(A->dataType() == DataType::FLOAT32) {
+        status = cusolverDnSgesvdj(handle, jobz, econ, m, n, reinterpret_cast<float*>(pA->getSpecialBuffer()), lda, reinterpret_cast<float*>(pS->getSpecialBuffer()), calcUV ? reinterpret_cast<float*>(pU->getSpecialBuffer()) : nullptr, ldu, calcUV ? reinterpret_cast<float*>(pV->getSpecialBuffer()) : nullptr, ldv, reinterpret_cast<float*>(dWork), lwork, devInfo, gesvdjParams);
     }
     else
         throw std::invalid_argument("svdJcb: given data type is unsupported !");
@@ -399,11 +399,11 @@ static void svdJcb(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, N
 
     NDArray::registerSpecialUse({pS, pU, pV}, {pA});
 
-    S.assign(pS);
+    S->assign(pS);
 
     if(calcUV) {
-        U.assign(pU);
-        V.assign(pV);
+        U->assign(pU);
+        V->assign(pV);
     }
 
     for (int i = toDelete.size() - 1; i >= 0; --i)
@@ -422,67 +422,67 @@ static void svdJcb(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, N
 }
 
 //////////////////////////////////////////////////////////////////////////
-static void svdBatched(nd4j::LaunchContext* context, const NDArray& A, NDArray& S, NDArray& U, NDArray& V, const bool fullUV, const bool calcUV) {
+static void svdBatched(nd4j::LaunchContext* context, const NDArray* A, NDArray* S, NDArray* U, NDArray* V, const bool fullUV, const bool calcUV) {
 
     // A [..., m, n]
     // S [..., n]
     // U [..., m, m] or [..., m, n] if fullUV = false and m > n
     // V [..., n, n] or [..., n, m] if fullUV = false and m < n
 
-    auto m = A.sizeAt(-2);
-    auto n = A.sizeAt(-1);
+    auto m = A->sizeAt(-2);
+    auto n = A->sizeAt(-1);
     const int minDim = m < n ? m : n;
-    const Nd4jLong bS = A.lengthOf() / (m * n);
+    const Nd4jLong bS = A->lengthOf() / (m * n);
 
     if(m > 32 || n > 32)
         throw std::runtime_error("svdBatched: numbers of rows and columns should be <= 32 !");
 
-    if(minDim != S.sizeAt(-1))
+    if(minDim != S->sizeAt(-1))
         throw std::runtime_error("svdBatched: wrong shape of S array !");
 
     if(calcUV) {
 
-        if(U.sizeAt(-2) != m)
+        if(U->sizeAt(-2) != m)
             throw std::runtime_error("svdBatched: wrong shape of U array !");
-        if(U.sizeAt(-1) != (fullUV ? m : minDim))
+        if(U->sizeAt(-1) != (fullUV ? m : minDim))
             throw std::runtime_error("svdBatched: wrong shape of U array !");
-        if(U.lengthOf() / (U.sizeAt(-2) * U.sizeAt(-1)) != bS)
+        if(U->lengthOf() / (U->sizeAt(-2) * U->sizeAt(-1)) != bS)
             throw std::runtime_error("svdBatched: wrong shape of U array !");
 
-        if(V.sizeAt(-2) != n)
+        if(V->sizeAt(-2) != n)
               throw std::runtime_error("svdBatched: wrong shape of V array !");
-        if(V.sizeAt(-1) != (fullUV ? n : minDim))
+        if(V->sizeAt(-1) != (fullUV ? n : minDim))
             throw std::runtime_error("svdBatched: wrong shape of V array !");
-        if(V.lengthOf() / (V.sizeAt(-2) * V.sizeAt(-1)) != bS)
+        if(V->lengthOf() / (V->sizeAt(-2) * V->sizeAt(-1)) != bS)
             throw std::runtime_error("svdBatched: wrong shape of V array !");
     }
 
-    NDArray* pA = const_cast<NDArray*>(&A);
-    NDArray* pS = &S;
-    NDArray* pU = &U;
-    NDArray* pV = &V;
+    NDArray* pA = const_cast<NDArray*>(A);
+    NDArray* pS = S;
+    NDArray* pU = U;
+    NDArray* pV = V;
 
     std::vector<NDArray*> toDelete;
 
     if(pA->ews() != 1 || pA->ordering() == 'c') {
-        pA = A.dup('f');
+        pA = A->dup('f');
         toDelete.push_back(pA);
     }
 
-    if(S.ews() != 1) {
-        pS = S.dup('f');
+    if(S->ews() != 1) {
+        pS = S->dup('f');
         toDelete.push_back(pS);
     }
 
     if(calcUV) {
 
         if(pU->ews() != 1 || pU->ordering() == 'c') {
-            pU = U.dup('f');
+            pU = U->dup('f');
             toDelete.push_back(pU);
         }
 
         if(pV->ews() != 1 || pV->ordering() == 'c') {
-            pV = V.dup('f');
+            pV = V->dup('f');
             toDelete.push_back(pV);
         }
     }
@@ -532,10 +532,10 @@ static void svdBatched(nd4j::LaunchContext* context, const NDArray& A, NDArray&
 
     // query working space of SVD
     int lwork = 0;
-    if(A.dataType() == DataType::DOUBLE)
-        status = cusolverDnDgesvdjBatched_bufferSize(handle, jobz, m, n, reinterpret_cast<double*>(pA->getSpecialBuffer()), lda, reinterpret_cast<double*>(pS->getSpecialBuffer()), reinterpret_cast<double*>(pU->getSpecialBuffer()), ldu, reinterpret_cast<double*>(pV->getSpecialBuffer()), ldv, &lwork, gesvdjParams, bS);
-    else if(A.dataType() == DataType::FLOAT32)
-        status = cusolverDnSgesvdjBatched_bufferSize(handle, jobz, m, n, reinterpret_cast<float*>(pA->getSpecialBuffer()), lda, reinterpret_cast<float*>(pS->getSpecialBuffer()), reinterpret_cast<float*>(pU->getSpecialBuffer()), ldu, reinterpret_cast<float*>(pV->getSpecialBuffer()), ldv, &lwork, gesvdjParams, bS);
+    if(A->dataType() == DataType::DOUBLE)
+        status = cusolverDnDgesvdjBatched_bufferSize(handle, jobz, m, n, reinterpret_cast<double*>(pA->getSpecialBuffer()), lda, reinterpret_cast<double*>(pS->getSpecialBuffer()), calcUV ? reinterpret_cast<double*>(pU->getSpecialBuffer()) : nullptr, ldu, calcUV ? reinterpret_cast<double*>(pV->getSpecialBuffer()) : nullptr, ldv, &lwork, gesvdjParams, bS);
+    else if(A->dataType() == DataType::FLOAT32)
+        status = cusolverDnSgesvdjBatched_bufferSize(handle, jobz, m, n, reinterpret_cast<float*>(pA->getSpecialBuffer()), lda, reinterpret_cast<float*>(pS->getSpecialBuffer()), calcUV ? reinterpret_cast<float*>(pU->getSpecialBuffer()) : nullptr, ldu, calcUV ? reinterpret_cast<float*>(pV->getSpecialBuffer()) : nullptr, ldv, &lwork, gesvdjParams, bS);
     else
         throw std::invalid_argument("svdBatched: given data type is unsupported !");
 
@@ -544,7 +544,7 @@ static void svdBatched(nd4j::LaunchContext* context, const NDArray& A, NDArray&
 
     // allocate memory dWork
     void* dWork = nullptr;
-    status2 = cudaMalloc((void**)&dWork , A.sizeOfT() * lwork);
+    status2 = cudaMalloc((void**)&dWork , A->sizeOfT() * lwork);
     if(status2 != cudaSuccess)
         throw cuda_exception::build("svdBatched: cuda failed !", status2);
     status2 = cudaDeviceSynchronize();
@@ -556,11 +556,11 @@ static void svdBatched(nd4j::LaunchContext* context, const NDArray& A, NDArray&
     NDArray::prepareSpecialUse({pS, pU, pV}, {pA});
 
     // choose appropriate cuda gemm api depending on data types
-    if(A.dataType() == DataType::DOUBLE) {
-        status = cusolverDnDgesvdjBatched(handle, jobz, m, n, reinterpret_cast<double*>(pA->getSpecialBuffer()), lda, reinterpret_cast<double*>(pS->getSpecialBuffer()), reinterpret_cast<double*>(pU->getSpecialBuffer()), ldu, reinterpret_cast<double*>(pV->getSpecialBuffer()), ldv, reinterpret_cast<double*>(dWork), lwork, devInfo, gesvdjParams, bS);
+    if(A->dataType() == DataType::DOUBLE) {
+        status = cusolverDnDgesvdjBatched(handle, jobz, m, n, reinterpret_cast<double*>(pA->getSpecialBuffer()), lda, reinterpret_cast<double*>(pS->getSpecialBuffer()), calcUV ? reinterpret_cast<double*>(pU->getSpecialBuffer()) : nullptr, ldu, calcUV ? reinterpret_cast<double*>(pV->getSpecialBuffer()) : nullptr, ldv, reinterpret_cast<double*>(dWork), lwork, devInfo, gesvdjParams, bS);
     }
-    else if(A.dataType() == DataType::FLOAT32) {
-        status = cusolverDnSgesvdjBatched(handle, jobz, m, n, reinterpret_cast<float*>(pA->getSpecialBuffer()), lda, reinterpret_cast<float*>(pS->getSpecialBuffer()), reinterpret_cast<float*>(pU->getSpecialBuffer()), ldu, reinterpret_cast<float*>(pV->getSpecialBuffer()), ldv, reinterpret_cast<float*>(dWork), lwork, devInfo, gesvdjParams, bS);
+    else if(A->dataType() == DataType::FLOAT32) {
+        status = cusolverDnSgesvdjBatched(handle, jobz, m, n, reinterpret_cast<float*>(pA->getSpecialBuffer()), lda, reinterpret_cast<float*>(pS->getSpecialBuffer()), calcUV ? reinterpret_cast<float*>(pU->getSpecialBuffer()) : nullptr, ldu, calcUV ? reinterpret_cast<float*>(pV->getSpecialBuffer()) : nullptr, ldv, reinterpret_cast<float*>(dWork), lwork, devInfo, gesvdjParams, bS);
     }
     else
         throw std::invalid_argument("svdBatched: given data type is unsupported !");
@@ -572,11 +572,11 @@ static void svdBatched(nd4j::LaunchContext* context, const NDArray& A, NDArray&
 
     NDArray::registerSpecialUse({pS, pU, pV}, {pA});
 
-    S.assign(pS);
+    S->assign(pS);
 
     if(calcUV) {
-        U.assign(pU);
-        V.assign(pV);
+        U->assign(pU);
+        V->assign(pV);
     }
 
     for (int i = toDelete.size() - 1; i >= 0; --i)
@@ -603,8 +603,8 @@ void svd(nd4j::LaunchContext* context, const NDArray* x, const std::vector<NDArr
     NDArray* V = outArrs[2];
 
     if(x->rankOf() == 2) {
-        // svdQR(context, *x, *S, *U, VT, fullUV, calcUV);
-        svdJcb(context, *x, *S, *U, *V, fullUV, calcUV);
+        // svdQR(context, x, S, U, VT, fullUV, calcUV);
+        svdJcb(context, x, S, U, V, fullUV, calcUV);
     }
     else {
 
@@ -621,7 +621,7 @@ void svd(nd4j::LaunchContext* context, const NDArray* x, const std::vector<NDArr
         }
 
         for (int i = 0; i < tadsX->size(); ++i)
-            svdJcb(context, *tadsX->at(i), *tadsS->at(i), calcUV ? *tadsU->at(i) : *S, calcUV ? *tadsV->at(i) : *S, fullUV, calcUV);
+            svdJcb(context, tadsX->at(i), tadsS->at(i), calcUV ? tadsU->at(i) : nullptr, calcUV ? tadsV->at(i) : nullptr, fullUV, calcUV);
 
         delete tadsX;
         delete tadsS;
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests15.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests15.cpp
index 21a0381e9..65cb470a7 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests15.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests15.cpp
@@ -245,8 +245,8 @@ TEST_F(DeclarableOpsTests15, test_check_numeric_3) {
 
 TEST_F(DeclarableOpsTests15, Test_layer_norm_1) {
     auto x = NDArrayFactory::create<float>('c', {1, 5}, {1., 2., 3., 4., 5.});
-    auto g = NDArrayFactory::create<float>('c', {1, 5}, {1., 2., 3., 4., 5.});
-    auto b = NDArrayFactory::create<float>('c', {1, 5}, {1., 2., 3., 4., 5.});
+    auto g = NDArrayFactory::create<float>('c', {5}, {1., 2., 3., 4., 5.});
+    auto b = NDArrayFactory::create<float>('c', {5}, {1., 2., 3., 4., 5.});
 
     nd4j::ops::layer_norm op;
     auto result = op.execute({&x, &g, &b}, {}, {0}, {false});
@@ -256,8 +256,8 @@ TEST_F(DeclarableOpsTests15, Test_layer_norm_1) {
 
 TEST_F(DeclarableOpsTests15, Test_layer_norm_bp_1) {
     auto x = NDArrayFactory::create<float>('c', {1, 5}, {1., 2., 3., 4., 5.});
-    auto g = NDArrayFactory::create<float>('c', {1, 5}, {1., 2., 3., 4., 5.});
-    auto b = NDArrayFactory::create<float>('c', {1, 5}, {1., 2., 3., 4., 5.});
+    auto g = NDArrayFactory::create<float>('c', {5}, {1., 2., 3., 4., 5.});
+    auto b = NDArrayFactory::create<float>('c', {5}, {1., 2., 3., 4., 5.});
     auto eps = NDArrayFactory::create<float>('c', {1, 5}, {0., 0., 0., 0., 0.});
 
     nd4j::ops::layer_norm_bp op;
@@ -266,6 +266,26 @@ TEST_F(DeclarableOpsTests15, Test_layer_norm_bp_1) {
     delete result;
 }
 
+//////////////////////////////////////////////////////////////////////
+TEST_F(DeclarableOpsTests15, Test_layer_norm_bp_2) {
+
+    NDArray x('c', {3, 4, 8, 8}, nd4j::DataType::FLOAT32);
+    NDArray gain('c', {4}, {-0.1, 0.1, -0.2, 0.2}, nd4j::DataType::FLOAT32);
+    NDArray bias('c', {4}, {-0.05, 0.05, -1.05, 1.05}, nd4j::DataType::FLOAT32);
+    NDArray gradO('c', {3, 4, 8, 8}, nd4j::DataType::FLOAT32);
+
+    NDArray gradI('c', {3, 4, 8, 8}, nd4j::DataType::FLOAT32);
+    NDArray gradG('c', {4}, nd4j::DataType::FLOAT32);
+    NDArray gradB('c', {4}, nd4j::DataType::FLOAT32);
+
+    x.linspace(-20, 0.5);
+    gradO.linspace(-4, 0.05);
+
+    nd4j::ops::layer_norm_bp op;
+    auto status = op.execute({&x, &gain, &bias, &gradO}, {&gradI, &gradG, &gradB}, {}, {1,2,3}, {true});
+    ASSERT_EQ(Status::OK(), status);
+}
+
 TEST_F(DeclarableOpsTests15, test_hashCode_1) {
     auto x = NDArrayFactory::create<int>('c', {10});
     auto y = NDArrayFactory::create<int>('c', {10});
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
index 992b21c0f..cff84c69b 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
@@ -87,4 +87,4 @@ TEST_F(DeclarableOpsTests16, test_size_dtype_1) {
     ASSERT_EQ(Status::OK(), status);
 
     ASSERT_EQ(e, z);
-}
\ No newline at end of file
+}

From b472d7d8c839be57577e966a7f5595ccc90c39f0 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Tue, 27 Aug 2019 21:00:38 +0300
Subject: [PATCH 28/56] [WIP] few more fixes (#182)

* one noop test

Signed-off-by: raver119 <raver119@gmail.com>

* skip input validation for no-input ops

Signed-off-by: raver119 <raver119@gmail.com>

* - one more noop empty test
- one more validation before sync

Signed-off-by: raver119 <raver119@gmail.com>

* typo

Signed-off-by: raver119 <raver119@gmail.com>

* one more validation fix

Signed-off-by: raver119 <raver119@gmail.com>

* CUDA empty reductions java side

Signed-off-by: raver119 <raver119@gmail.com>

* one svd test

Signed-off-by: raver119 <raver119@gmail.com>

* Corrected segment_mean helpers and added another test.

* Refactored segment_mean kernels to avoid race_condition.
---
 libnd4j/blas/cuda/NativeOps.cu                |  6 ++--
 .../declarable/helpers/cuda/segment_mean.cu   | 16 +++++-----
 .../ops/declarable/impl/DeclarableOp.cpp      |  6 ++++
 .../layers_tests/DeclarableOpsTests16.cpp     | 30 +++++++++++++++++++
 .../layers_tests/DeclarableOpsTests7.cpp      | 19 ++++++++++++
 .../jita/allocator/impl/AtomicAllocator.java  |  7 +++++
 .../ops/executioner/CudaExecutioner.java      |  4 +--
 .../org/nd4j/linalg/shape/EmptyTests.java     | 11 +++++++
 8 files changed, 86 insertions(+), 13 deletions(-)

diff --git a/libnd4j/blas/cuda/NativeOps.cu b/libnd4j/blas/cuda/NativeOps.cu
index 626b0ea26..7e74c3237 100755
--- a/libnd4j/blas/cuda/NativeOps.cu
+++ b/libnd4j/blas/cuda/NativeOps.cu
@@ -2728,11 +2728,13 @@ int execCustomOp2(Nd4jPointer* extraPointers, Nd4jLong hash, Nd4jPointer opConte
             throw nd4j::cuda_exception::build("customOp execution failed", res);
 
         for (auto v:context->fastpath_in()) {
-            v->syncToDevice();
+            if (!v->isEmpty())
+                v->syncToDevice();
         }
 
         for (auto v:context->fastpath_out()) {
-            v->syncToDevice();
+            if (!v->isEmpty())
+                v->syncToDevice();
         }
 
         return result;
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_mean.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_mean.cu
index fbb45a375..19869f646 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_mean.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_mean.cu
@@ -74,14 +74,14 @@ namespace helpers {
     template <typename T, typename I>
     static __global__ void unsortedSegmentMeanLinearKernel(void* input, Nd4jLong* inputShape, void* indices, Nd4jLong* indicesShape, int* starts, int* lengths, Nd4jLong numOfClasses, void* output, Nd4jLong* outputShape) {
         __shared__ T* val;
-        __shared__ Nd4jLong xLen, zLen, segment, zIndex;
+        __shared__ Nd4jLong xLen, zLen, zIndex;
         __shared__ T* x;
         __shared__ T* z;
         __shared__ I* y; //int threadsPerSegment, start, finish;
-
+        auto segment = blockIdx.x;// /
         if (threadIdx.x == 0) {
 //            threadsPerSegment = (gridDim.x + numOfClasses - 1) / numOfClasses;
-            segment = blockIdx.x;// / threadsPerSegment;
+//            threadsPerSegment;
             x = reinterpret_cast<T*>(input);
             z = reinterpret_cast<T*>(output);
             y = reinterpret_cast<I*>(indices);
@@ -117,12 +117,12 @@ namespace helpers {
     template <typename T, typename I>
     static __global__ void segmentMeanTadKernel(void* inputBuf, Nd4jLong* inputShape, Nd4jLong* inputTads, Nd4jLong* inputTadOffsets, I* indices, int* starts, int* lengths, Nd4jLong numOfClasses, void* outputBuf, Nd4jLong* outputShape, Nd4jLong* outputTads, Nd4jLong* outputTadOffsets) {
         __shared__ T* val;
-        __shared__ Nd4jLong len, segment, zIndex, total;
+        __shared__ Nd4jLong len, zIndex, total;
         __shared__ T* z;
         __shared__ int threadsPerSegment, start, finish;
+        auto segment = indices[blockIdx.x]; // / threadsPerSegment;
 
         if (threadIdx.x == 0) {
-            segment = indices[blockIdx.x]; // / threadsPerSegment;
             z = reinterpret_cast<T*>(outputBuf) + outputTadOffsets[segment];
             len = shape::length(inputTads);
             start = starts[segment];
@@ -139,7 +139,7 @@ namespace helpers {
                 for (auto e = threadIdx.x; e < len; e += blockDim.x) {
                     auto xIndex = shape::getIndexOffset(e, inputTads, len);
                     auto zIndex = shape::getIndexOffset(e, outputTads, len);
-                    z[zIndex] = T(x[xIndex]/lengths[segment]);
+                    nd4j::math::atomics::nd4j_atomicAdd(&z[zIndex], T(x[xIndex]/lengths[segment]));
                 }
             }
             else {
@@ -197,7 +197,6 @@ namespace helpers {
     static void unsortedSegmentMeanFunctor_(nd4j::LaunchContext* context, NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
         auto stream = context->getCudaStream();
 //        NDArray classes = NDArrayFactory::create<int>('c', {numOfClasses, 2});
-        NDArray::prepareSpecialUse({output}, {input, indices});
 
         NDArray classesRangesBegs = NDArrayFactory::create<int>('c', {numOfClasses});
         NDArray classesRangesLens = NDArrayFactory::create<int>('c', {numOfClasses});
@@ -226,7 +225,6 @@ namespace helpers {
             dims.x = input->sizeAt(0);
             segmentMeanTadKernel<T,I><<<dims.x, dims.y, dims.z, *stream>>>(input->specialBuffer(), input->specialShapeInfo(), inputTads, inputTadOffsets, reinterpret_cast<I*>(indices->specialBuffer()), begins, lengths, numOfClasses, output->specialBuffer(), output->specialShapeInfo(), outputTads, outputTadOffsets);
         }
-        NDArray::registerSpecialUse({output}, {input, indices});
 
     }
     // -------------------------------------------------------------------------------------------------------------- //
@@ -234,7 +232,7 @@ namespace helpers {
         NDArray::prepareSpecialUse({output}, {input, indices});
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), unsortedSegmentMeanFunctor_, (context, input, indices, numOfClasses, output),
                               NUMERIC_TYPES, INDEXING_TYPES);
-        NDArray::prepareSpecialUse({output}, {input, indices});
+        NDArray::registerSpecialUse({output}, {input, indices});
     }
 
     // -------------------------------------------------------------------------------------------------------------- //
diff --git a/libnd4j/include/ops/declarable/impl/DeclarableOp.cpp b/libnd4j/include/ops/declarable/impl/DeclarableOp.cpp
index 4fe28df8c..b313acd9c 100644
--- a/libnd4j/include/ops/declarable/impl/DeclarableOp.cpp
+++ b/libnd4j/include/ops/declarable/impl/DeclarableOp.cpp
@@ -372,6 +372,9 @@ namespace nd4j {
                     if (_descriptor->isSameMode()) {
 
                         if (index >= block.width()) {
+                            if (block.fastpath_in().size() == 0)
+                                continue;
+
                             auto ia = block.fastpath_in()[0];
 
                             if (ia->dataType() != cType) {
@@ -423,6 +426,9 @@ namespace nd4j {
                             if (_descriptor->isSameMode()) {
 
                                 if (index >= block.width()) {
+                                    if (block.width() == 0)
+                                        continue;
+
                                     auto iv = block.variable(0);
 
                                     if (iv->getNDArray()->dataType() != cType) {
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
index cff84c69b..a2772a734 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
@@ -88,3 +88,33 @@ TEST_F(DeclarableOpsTests16, test_size_dtype_1) {
 
     ASSERT_EQ(e, z);
 }
+
+TEST_F(DeclarableOpsTests16, test_empty_noop_1) {
+    auto z = NDArrayFactory::empty<Nd4jLong>();
+
+    nd4j::ops::noop op;
+    auto status = op.execute({}, {&z}, {}, {}, {});
+    ASSERT_EQ(Status::OK(), status);
+}
+
+TEST_F(DeclarableOpsTests16, test_empty_noop_2) {
+    auto z = NDArrayFactory::empty<Nd4jLong>();
+
+    Context ctx(1);
+    ctx.setOutputArray(0, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo());
+
+    nd4j::ops::noop op;
+    auto status = op.execute(&ctx);
+
+    ASSERT_EQ(Status::OK(), status);
+}
+
+TEST_F(DeclarableOpsTests16, test_svd_1) {
+    auto x = NDArrayFactory::create<float>('c', {3, 3}, {0.7787856f, 0.80119777f, 0.72437465f, 0.23089433f, 0.72714126f, 0.18039072f,0.50563407f, 0.89252293f, 0.5461209f});
+    auto z = NDArrayFactory::create<float>('c', {3});
+
+    nd4j::ops::svd op;
+    auto status = op.execute({&x}, {&z}, {}, {0, 0, 16}, {});
+
+    ASSERT_EQ(Status::OK(), status);
+}
\ No newline at end of file
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests7.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests7.cpp
index b0488c23a..c80d75372 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests7.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests7.cpp
@@ -1459,6 +1459,25 @@ TEST_F(DeclarableOpsTests7, TestSegmentMean_021) {
     delete result;
 }
 
+TEST_F(DeclarableOpsTests7, TestSegmentMean_022) {
+    auto x = NDArrayFactory::create<float>('c', {6, 3});//, {1, 2,  3., 4., 5., 6.,  7., 8., 9., 10., 11., 12., 13., 14., 15., 16., 17., 18.});
+    auto idx = NDArrayFactory::create<int>({0, 0, 1, 1, 2,2});
+    auto z = NDArrayFactory::create<float>('c', {3, 3}); //, {    2.5, 3.5, 4.5,      8.5, 9.5, 10.5,   14.5, 15.5,  16.5});
+    auto exp = NDArrayFactory::create<float>('c', {3, 3}, {    2.5, 3.5, 4.5,      8.5, 9.5, 10.5,   14.5, 15.5,  16.5});
+
+    nd4j::ops::segment_mean op;
+    x.linspace(1.);
+    auto result = op.execute({&x, &idx}, {&z}, {}, {}, {}, false, nd4j::DataType::FLOAT32);
+    ASSERT_EQ(result, Status::OK());
+
+    exp.printIndexedBuffer("Expect Mean");
+    z.printIndexedBuffer("Output Mean");
+//    exp.printShapeInfo("Exp Shape");
+    ASSERT_TRUE(exp.equalsTo(z));
+
+//    delete result;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests7, TestSegmentMeanBP_2) {
     auto x = NDArrayFactory::create<double>('c', {4, 4}, {1.8, 2.5,  4.,  9.,2.1, 2.4,  3.,  9.,2.1, 2.1, 0.7, 0.1,3., 4.2, 2.2, 1.});
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/impl/AtomicAllocator.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/impl/AtomicAllocator.java
index 0ec1876ca..8ec8734f7 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/impl/AtomicAllocator.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/allocator/impl/AtomicAllocator.java
@@ -40,6 +40,7 @@ import org.nd4j.jita.workspace.CudaWorkspace;
 import org.nd4j.linalg.api.buffer.BaseDataBuffer;
 import org.nd4j.linalg.api.buffer.DataBuffer;
 import org.nd4j.linalg.api.buffer.DataType;
+import org.nd4j.linalg.api.buffer.Utf8Buffer;
 import org.nd4j.linalg.api.memory.enums.MemoryKind;
 import org.nd4j.linalg.api.memory.pointers.PagedPointer;
 import org.nd4j.linalg.api.ndarray.INDArray;
@@ -284,10 +285,16 @@ public class AtomicAllocator implements Allocator {
      */
     @Override
     public Pointer getPointer(@NonNull DataBuffer buffer, CudaContext context) {
+        if (buffer instanceof Utf8Buffer)
+            return null;
+
         return memoryHandler.getDevicePointer(buffer, context);
     }
 
     public Pointer getPointer(DataBuffer buffer) {
+        if (buffer instanceof Utf8Buffer)
+            return null;
+
         return memoryHandler.getDevicePointer(buffer, getDeviceContext());
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
index afeff4d8b..c5b02a82f 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
@@ -961,12 +961,12 @@ public class CudaExecutioner extends DefaultOpExecutioner {
         if (CudaEnvironment.getInstance().getConfiguration().isDebug())
             lastOp.set(op.opName());
 
-        val tadBuffers = tadManager.getTADOnlyShapeInfo(op.x(), dimension);
+        val tadBuffers = op.x().isEmpty() ? Pair.<DataBuffer, DataBuffer>makePair(op.x().data(), null) : tadManager.getTADOnlyShapeInfo(op.x(), dimension);
 
         val hostTadShapeInfo = AddressRetriever.retrieveHostPointer(tadBuffers.getFirst());
         val devTadShapeInfo = AtomicAllocator.getInstance().getPointer(tadBuffers.getFirst(), context);
 
-        val offsets = tadBuffers.getSecond();
+        val offsets = op.x().isEmpty() ? null : tadBuffers.getSecond();
         val devTadOffsets = offsets == null ? null : AtomicAllocator.getInstance().getPointer(offsets, context);
 
         Pointer x = AtomicAllocator.getInstance().getPointer(op.x(), context);
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/EmptyTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/EmptyTests.java
index e7e8f8288..3bef69c19 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/EmptyTests.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/EmptyTests.java
@@ -309,6 +309,17 @@ public class EmptyTests extends BaseNd4jTest {
         assertEquals(x, out);
     }
 
+    @Test
+    public void testEmptyNoop() {
+        val output = Nd4j.empty(DataType.LONG);
+
+        val op = DynamicCustomOp.builder("noop")
+                .addOutputs(output)
+                .build();
+
+        Nd4j.exec(op);
+    }
+
     @Override
     public char ordering() {
         return 'c';

From d31197db5fce781240f3db4c53caa1f86afd92c7 Mon Sep 17 00:00:00 2001
From: Ryan Nett <rnett@skymind.io>
Date: Tue, 27 Aug 2019 18:02:41 -0700
Subject: [PATCH 29/56] Remove resolvePropertiesFromSameDiffBeforeExecution()
 (#172)

* remove unneeded resolveProperties methods

Signed-off-by: Ryan Nett <rnett@skymind.io>

* final fixes, make final to prevent more from being added

Signed-off-by: Ryan Nett <rnett@skymind.io>

* gather fix

Signed-off-by: Ryan Nett <rnett@skymind.io>

* deprecate DifferentialFunction resolveProps

Signed-off-by: Ryan Nett <rnett@skymind.io>
---
 .../functions/DifferentialFunction.java       |  5 +++-
 .../ops/impl/layers/convolution/Conv3D.java   |  8 ------
 .../api/ops/impl/shape/ConfusionMatrix.java   |  5 ----
 .../linalg/api/ops/impl/shape/Gather.java     | 25 -------------------
 .../linalg/api/ops/impl/shape/Repeat.java     |  9 -------
 .../linalg/api/ops/impl/shape/Squeeze.java    |  9 +------
 .../linalg/api/ops/impl/shape/Transpose.java  |  5 ----
 .../impl/transforms/HistogramFixedWidth.java  | 10 --------
 .../linalg/api/ops/impl/transforms/Pad.java   | 11 --------
 .../segment/UnsortedSegmentMax.java           |  8 ------
 .../segment/UnsortedSegmentMean.java          |  8 ------
 .../segment/UnsortedSegmentMin.java           |  8 ------
 .../segment/UnsortedSegmentProd.java          |  8 ------
 .../segment/UnsortedSegmentSqrtN.java         |  8 ------
 .../segment/UnsortedSegmentSum.java           |  8 ------
 15 files changed, 5 insertions(+), 130 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java
index 71bbd26ee..34240516f 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java
@@ -547,8 +547,11 @@ public abstract class DifferentialFunction {
     /**
      * Resolve properties and arguments right before execution of
      * this operation.
+     *
+     * @deprecated Will be removed in the future.  Ops should support array arguments.  Should not bs used or overridden.
      */
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
+    @Deprecated
+    public final void resolvePropertiesFromSameDiffBeforeExecution() {
         val properties = sameDiff.propertiesToResolveForFunction(this);
         val fields = DifferentialFunctionClassHolder.getInstance().getFieldsForFunction(this);
         val currentFields = this.propertiesForFunction();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv3D.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv3D.java
index d3fe330fd..810974103 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv3D.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/convolution/Conv3D.java
@@ -268,14 +268,6 @@ public class Conv3D extends DynamicCustomOp {
         return ret;
     }
 
-
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        if (numIArguments() < 1) {
-            addArgs();
-        }
-    }
-
     @Override
     public boolean isConfigProperties() {
         return true;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ConfusionMatrix.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ConfusionMatrix.java
index 231cf5783..6275ce210 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ConfusionMatrix.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/ConfusionMatrix.java
@@ -66,11 +66,6 @@ public class ConfusionMatrix extends DynamicCustomOp {
         //Looks like this is implemented in practice using a large collection of discrete ops - not single TF import op?
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-    }
-
     @Override
     public String opName() {
         return "confusion_matrix";
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Gather.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Gather.java
index 31718d337..5613cc85f 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Gather.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Gather.java
@@ -74,7 +74,6 @@ public class Gather extends DynamicCustomOp {
     @Override
     public void initFromTensorFlow(NodeDef nodeDef, SameDiff initWith, Map<String, AttrValue> attributesForNode, GraphDef graph) {
         TFGraphMapper.getInstance().initFunctionFromProperties(nodeDef.getOp(), this, attributesForNode, nodeDef, graph);
-
     }
 
     @Override
@@ -82,30 +81,6 @@ public class Gather extends DynamicCustomOp {
         OnnxGraphMapper.getInstance().initFunctionFromProperties(node.getOpType(), this, attributesForNode, node, graph);
     }
 
-
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-//        super.resolvePropertiesFromSameDiffBeforeExecution();
-        if (indices != null && numInputArguments() < 2) {
-            if (numInputArguments() == 0) {
-                INDArray a = Nd4j.create(indices, new long[]{indices.length}, new long[]{1}, 'c', DataType.INT);
-                if (indices.length > 1)
-                    a = a.reshape(indices.length);
-                else
-                    a = a.reshape(new int[]{});
-
-                addInputArgument(args()[0].getArr(), a);
-            } else if (numInputArguments() == 1) {
-                addInputArgument(Nd4j.create(indices, new long[]{indices.length}, new long[]{1}, 'c', DataType.INT));
-            }
-
-        }
-
-        if (numIArguments() < 1) {
-            addIArgument(jaxis);
-        }
-    }
-
     @Override
     public Map<String, Map<String, PropertyMapping>> mappingsForFunction() {
         Map<String, Map<String, PropertyMapping>> ret = new HashMap<>();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Repeat.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Repeat.java
index 02f8f9445..af8940bf4 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Repeat.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Repeat.java
@@ -110,15 +110,6 @@ public class Repeat extends DynamicCustomOp {
         super.initFromOnnx(node, initWith, attributesForNode, graph);
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        if (numOutputArguments() < getDescriptor().getNumOutputs()) {
-            for (val output : outputVariables()) {
-                addOutputArgument(output.getArr());
-            }
-        }
-    }
-
     @Override
     public String onnxName() {
         return "Repeat";
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Squeeze.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Squeeze.java
index bfd8f58ec..816654d13 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Squeeze.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Squeeze.java
@@ -41,6 +41,7 @@ public class Squeeze extends DynamicCustomOp {
     public Squeeze(SameDiff sameDiff, SDVariable arg, int[] squeezeDims) {
         super(null, sameDiff, new SDVariable[]{arg});
         this.squeezeDims = squeezeDims;
+        addIArgument(squeezeDims);
     }
 
     @Override
@@ -53,14 +54,6 @@ public class Squeeze extends DynamicCustomOp {
         addIArgument(squeezeDims);
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-        if (squeezeDims != null && numIArguments() < squeezeDims.length) {
-            addIArgument(squeezeDims);
-        }
-    }
-
     @Override
     public String opName() {
         return "squeeze";
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java
index 2de0a29c5..b9687e598 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java
@@ -65,11 +65,6 @@ public class Transpose extends DynamicCustomOp {
     public Transpose() {
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-    }
-
     @Override
     public Map<String, Map<String, PropertyMapping>> mappingsForFunction() {
         Map<String, Map<String, PropertyMapping>> ret = new LinkedHashMap<>();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/HistogramFixedWidth.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/HistogramFixedWidth.java
index 8894a87ec..091846db8 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/HistogramFixedWidth.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/HistogramFixedWidth.java
@@ -70,16 +70,6 @@ public class HistogramFixedWidth extends DynamicCustomOp {
         //No op - just need the inputs
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        if(args().length == 3 && iArguments.isEmpty()){
-            //Num bins is 3rd array
-            addIArgument(arg(2).getArr().getInt(0));
-        }
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-    }
-
-
     @Override
     public List<SDVariable> doDiff(List<SDVariable> f1) {
         throw new UnsupportedOperationException("Not supported");
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/Pad.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/Pad.java
index f940ac7a0..72d2823b5 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/Pad.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/Pad.java
@@ -82,17 +82,6 @@ public class Pad extends DynamicCustomOp {
         //Constant value is resolved just before execution
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        if(args().length == 3){
-            INDArray arr = arg(2).getArr();
-            this.tArguments.clear();
-            this.tArguments.add(arr.getDouble(0));
-        }
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-    }
-
-
     @Override
     public List<SDVariable> doDiff(List<SDVariable> i_v) {
         //Pad backprop: it's basically slice op...
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMax.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMax.java
index 73d4d281e..6d6798701 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMax.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMax.java
@@ -54,14 +54,6 @@ public class UnsortedSegmentMax extends DynamicCustomOp {
         return "UnsortedSegmentMax";
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        if(args().length == 3 && iArguments == null || iArguments.size() == 0){
-            addIArgument(arg(2).getArr().getInt(0));
-        }
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-    }
-
     @Override
     public List<SDVariable> doDiff(List<SDVariable> gradients){
         return Arrays.asList(f().unsortedSegmentMaxBp(arg(0), arg(1), gradients.get(0), numSegments));
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMean.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMean.java
index a110d7225..f51b94218 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMean.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMean.java
@@ -54,14 +54,6 @@ public class UnsortedSegmentMean extends DynamicCustomOp {
         return "UnsortedSegmentMean";
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        if(args().length == 3 && iArguments == null || iArguments.size() == 0){
-            addIArgument(arg(2).getArr().getInt(0));
-        }
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-    }
-
     @Override
     public List<SDVariable> doDiff(List<SDVariable> gradients){
         return Arrays.asList(f().unsortedSegmentMeanBp(arg(0), arg(1), gradients.get(0), numSegments));
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMin.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMin.java
index 6666981ff..1b885676e 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMin.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentMin.java
@@ -54,14 +54,6 @@ public class UnsortedSegmentMin extends DynamicCustomOp {
         return "UnsortedSegmentMin";
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        if(args().length == 3 && iArguments == null || iArguments.size() == 0){
-            addIArgument(arg(2).getArr().getInt(0));
-        }
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-    }
-
     @Override
     public List<SDVariable> doDiff(List<SDVariable> gradients){
         return Arrays.asList(f().unsortedSegmentMinBp(arg(0), arg(1), gradients.get(0), numSegments));
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentProd.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentProd.java
index 36ba70342..b2e254fb7 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentProd.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentProd.java
@@ -54,14 +54,6 @@ public class UnsortedSegmentProd extends DynamicCustomOp {
         return "UnsortedSegmentProd";
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        if(args().length == 3 && iArguments == null || iArguments.size() == 0){
-            addIArgument(arg(2).getArr().getInt(0));
-        }
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-    }
-
     @Override
     public List<SDVariable> doDiff(List<SDVariable> gradients){
         return Arrays.asList(f().unsortedSegmentProdBp(arg(0), arg(1), gradients.get(0), numSegments));
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentSqrtN.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentSqrtN.java
index cfb62450a..ef34e9f81 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentSqrtN.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentSqrtN.java
@@ -53,14 +53,6 @@ public class UnsortedSegmentSqrtN extends DynamicCustomOp {
         return "UnsortedSegmentSqrtN";
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        if(args().length == 3 && iArguments == null || iArguments.size() == 0){
-            addIArgument(arg(2).getArr().getInt(0));
-        }
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-    }
-
     @Override
     public List<SDVariable> doDiff(List<SDVariable> gradients){
         return Arrays.asList(f().unsortedSegmentSqrtNBp(arg(0), arg(1), gradients.get(0), numSegments));
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentSum.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentSum.java
index 3e4c4d2f1..466cc8cf2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentSum.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/segment/UnsortedSegmentSum.java
@@ -55,14 +55,6 @@ public class UnsortedSegmentSum extends DynamicCustomOp {
         return "UnsortedSegmentSum";
     }
 
-    @Override
-    public void resolvePropertiesFromSameDiffBeforeExecution() {
-        if(args().length == 3 && iArguments == null || iArguments.size() == 0){
-            addIArgument(arg(2).getArr().getInt(0));
-        }
-        super.resolvePropertiesFromSameDiffBeforeExecution();
-    }
-
     @Override
     public List<SDVariable> doDiff(List<SDVariable> gradients){
         return Arrays.asList(f().unsortedSegmentSumBp(arg(0), arg(1), gradients.get(0), numSegments));

From 59a6e4e3ae5e2a7cf1cf5dbe0b709bd511c887f9 Mon Sep 17 00:00:00 2001
From: Robert Altena <Rob@ra-ai.com>
Date: Wed, 28 Aug 2019 12:03:23 +0900
Subject: [PATCH 30/56] INDArray refactoring (#170)

* javadoc

Signed-off-by: Robert Altena <Rob@Ra-ai.com>

* remove javaTensorAlongDimension

Signed-off-by: Robert Altena <Rob@Ra-ai.com>

* wip

Signed-off-by: Robert Altena <Rob@Ra-ai.com>

* javadoc
---
 .../nd4j/linalg/api/ndarray/BaseNDArray.java  | 99 +------------------
 .../linalg/api/ndarray/BaseSparseNDArray.java |  7 +-
 .../org/nd4j/linalg/api/ndarray/INDArray.java | 52 +++++-----
 .../linalg/checkutil/NDArrayCreationUtil.java | 54 +++++-----
 .../java/org/nd4j/linalg/dataset/DataSet.java |  2 +-
 .../test/java/org/nd4j/linalg/Nd4jTestsC.java |  2 +-
 .../api/tad/TestTensorAlongDimension.java     | 26 ++---
 .../nd4j/linalg/ops/OpExecutionerTests.java   |  2 +-
 .../java/org/nd4j/linalg/shape/TADTests.java  |  8 +-
 .../linalg/shape/concat/ConcatTestsC.java     |  2 +-
 10 files changed, 75 insertions(+), 179 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
index c73a56c61..04a61f9d0 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
@@ -1007,19 +1007,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return toTad;
     }
 
-    /**
-     * Get the vector along a particular dimension
-     *
-     * @param index     the index of the vector to getScalar
-     * @param dimension the dimension to getScalar the vector from
-     * @return the vector along a particular dimension
-     */
-    @Override
-    @Deprecated
-    public INDArray javaTensorAlongDimension(int index, int... dimension) {
-        return doTad(index, dimension);
-    }
-
     private void setShapeInformation(Pair<DataBuffer, long[]> shapeInfo) {
         this.shapeInformation = shapeInfo.getFirst();
         this.jvmShapeInfo = new JvmShapeInfo(shapeInfo.getSecond());
@@ -1110,14 +1097,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return ret2.permutei(finalPermuteDims);
     }
 
-
-
-    /**
-     * Returns the number of possible vectors for a given dimension
-     *
-     * @param dimension the dimension to calculate the number of vectors for
-     * @return the number of possible vectors along a dimension
-     */
     @Override
     public long vectorsAlongDimension(int dimension) {
         if (dimension == 0 && isVector() || isRowVectorOrScalar())
@@ -1150,17 +1129,11 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return length / size(dimension);
     }
 
-    /**
-     * Get the vector along a particular dimension
-     *
-     * @param index     the index of the vector to get
-     * @param dimension the dimension to get the vector from
-     * @return the vector along a particular dimension
-     */
     @Override
     public INDArray vectorAlongDimension(int index, int dimension) {
-        if (dimension < 0)
+        if (dimension < 0) {
             dimension = jvmShapeInfo.getRank() + dimension;
+        }
 
         //return the whole thing
         if (dimension == jvmShapeInfo.getRank() - 1 && size(dimension) == 1 && rank() > 2
@@ -1168,12 +1141,7 @@ public abstract class BaseNDArray implements INDArray, Iterable {
             return this;
         }
 
-        INDArray ret = tensorAlongDimension(index, dimension);
-        //if (isMatrix() && ret.isVector() && dimension == 1 && !ret.isRowVector())
-        //    return ret.reshape(ArrayUtil.reverseCopy(ret.shape()));
-        //else if (isMatrix() && ret.isVector() && dimension == 0 && !ret.isColumnVector())
-        //    return ret.reshape(ArrayUtil.reverseCopy(ret.shape()));
-        return ret;
+        return tensorAlongDimension(index, dimension);
     }
 
     @Override
@@ -1196,13 +1164,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         setShapeInformation(Nd4j.getShapeInfoProvider().createShapeInformation(ArrayUtil.toLongArray(shape), ArrayUtil.toLongArray(stride),  0, ordering(), this.dataType(), false));
     }
 
-
-    /**
-     * Cumulative sum along a dimension
-     *
-     * @param dimension the dimension to perform cumulative sum along
-     * @return the cumulative sum along the specified dimension
-     */
     @Override
     public INDArray cumsumi(int dimension) {
         validateNumericalArray("cumsumi", true);
@@ -1351,25 +1312,12 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return logEntropy(Integer.MAX_VALUE).getDouble(0);
     }
 
-    /**
-     * Cumulative sum along a dimension (in place)
-     *
-     * @param dimension the dimension to perform cumulative sum along
-     * @return the cumulative sum along the specified dimension
-     */
     @Override
     public INDArray cumsum(int dimension) {
         validateNumericalArray("cumsum", true);
         return dup().cumsumi(dimension);
     }
 
-    /**
-     * Assign all of the elements in the given
-     * ndarray to this ndarray
-     *
-     * @param arr the elements to assign
-     * @return this
-     */
     @Override
     public INDArray assign(final INDArray arr) {
         Preconditions.checkState((this.isScalar() && arr.isScalar()) || (this.isVector() && arr.isVector()) || Shape.shapeEqualWithSqueeze(this.shape(), arr.shape()),
@@ -1378,7 +1326,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
 
         Preconditions.checkArgument(this.length() == arr.length(), "Length of both arrays must be equal");
 
-        //Nd4j.getExecutioner().exec(new org.nd4j.linalg.api.ops.impl.transforms.pairwise.Set(this, arr, this, length()));
         Nd4j.getExecutioner().exec(new org.nd4j.linalg.api.ops.impl.transforms.any.Assign(arr, this));
         return this;
     }
@@ -1413,7 +1360,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
     @Override
     public INDArray putScalar(long i, float value) {
         return putScalar(i, (double) value);
-
     }
 
     @Override
@@ -1540,7 +1486,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return this;
     }
 
-
     @Override
     public INDArray putScalar(int[] indexes, float value) {
         return putScalar(indexes, (double) value);
@@ -1556,27 +1501,12 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return putScalar(indexes, (double) value);
     }
 
-    /**
-     * Returns an ndarray with 1 if the element is epsilon equals
-     *
-     * @param other the number to compare
-     * @return a copied ndarray with the given
-     * binary conditions
-     */
     @Override
     public INDArray eps(Number other) {
         validateNumericalArray("eps", true);
         return Nd4j.getExecutioner().exec(new ScalarEps(this, Nd4j.createUninitialized(DataType.BOOL, this.shape(), this.ordering()), other));
     }
 
-    /**
-     * epsilon equals than comparison:
-     * If the given number is less than the
-     * comparison number the item is 0 otherwise 1
-     *
-     * @param other the number to compare
-     * @return
-     */
     @Override
     public INDArray eps(INDArray other) {
         validateNumericalArray("eps", true);
@@ -1613,7 +1543,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return Nd4j.getExecutioner().exec(new ScalarGreaterThanOrEqual(this, Nd4j.createUninitialized(DataType.BOOL, this.shape(), this.ordering()), other));
     }
 
-
     @Override
     public INDArray lt(INDArray other) {
         validateNumericalArray("less than (lt)", false);
@@ -1675,9 +1604,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return Nd4j.getExecutioner().exec(new MatchConditionTransform(this, Nd4j.createUninitialized(DataType.BOOL, this.shape(), this.ordering()), Conditions.isNan()));
     }
 
-    /**
-     * Negate each element.
-     */
     @Override
     public INDArray neg() {
         validateNumericalArray("negative (neg)", true);
@@ -1686,9 +1612,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return Nd4j.getExecutioner().exec(new Negative(this, Nd4j.createUninitialized(this.dataType(), this.shape(), this.ordering())));
     }
 
-    /**
-     * Negate each element (in-place).
-     */
     @Override
     public INDArray negi() {
         validateNumericalArray("negative (negi)", true);
@@ -3909,28 +3832,12 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return assign(value ? 1 : 0);
     }
 
-
-    /**
-     * Assign all elements from given ndarray that are matching given condition,
-     * ndarray to this ndarray
-     *
-     * @param arr       the elements to assign
-     * @param condition
-     * @return this
-     */
     @Override
     public INDArray assignIf(INDArray arr, Condition condition) {
         BooleanIndexing.assignIf(this, arr, condition);
         return this;
     }
 
-    /**
-     * Replaces all elements in this ndarray that are matching give condition, with corresponding elements from given array
-     *
-     * @param arr
-     * @param condition
-     * @return
-     */
     @Override
     public INDArray replaceWhere(INDArray arr, Condition condition) {
         Nd4j.getCompressor().autoDecompress(this);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
index 5257be12e..0d7aca8e0 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
@@ -411,11 +411,6 @@ public abstract class BaseSparseNDArray implements ISparseNDArray {
         return null;
     }
 
-    @Override
-    public INDArray javaTensorAlongDimension(int index, int... dimension) {
-        return null;
-    }
-
     @Override
     public INDArray cumsumi(int dimension) {
         return null;
@@ -476,7 +471,6 @@ public abstract class BaseSparseNDArray implements ISparseNDArray {
         return null;
     }
 
-
     @Override
     public INDArray isInfinite() {
         throw new UnsupportedOperationException();
@@ -551,6 +545,7 @@ public abstract class BaseSparseNDArray implements ISparseNDArray {
     public INDArray lte(Number other) {
         return null;
     }
+
     @Override
     public INDArray lt(INDArray other) {
         return null;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java
index 88cda5b4f..fe940f3ee 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java
@@ -106,27 +106,31 @@ public interface INDArray extends Serializable, AutoCloseable {
 
     /**
      * Element wise stride
+     * @return the element wise stride
      */
     int elementWiseStride();
 
     /**
-     * Get a scalar
-     * at the given linear offset
+     * Get a double at the given linear offset unsafe, without checks.
      * @param offset the offset to get at
-     * @return this
+     * @return double value at offset
      */
-    double getDoubleUnsafe(long offset);
+    double getDoubleUnsafe(long offset); //TODO: consider deleting.
 
+    /**
+     * Get string value at given index.
+     * @param index index to retreive
+     * @return string value at index.
+     */
     String getString(long index);
 
     /**
-     * Insert a scalar
-     * at the given linear offset
+     * Insert a scalar at the given linear offset
      * @param offset the offset to insert at
      * @param value the value to insert
      * @return this
      */
-    INDArray putScalarUnsafe(long offset, double value);
+    INDArray putScalarUnsafe(long offset, double value); //TODO: consider deleting.
 
     /**
      * Returns the number of possible vectors for a given dimension
@@ -162,17 +166,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray tensorAlongDimension(long index, int... dimension);
 
-    /**
-     * Get the vector along a particular dimension
-     *
-     * @param index     the index of the vector to getScalar
-     * @param dimension the dimension to getScalar the vector from
-     * @return the vector along a particular dimension
-     */
-    @Deprecated
-    INDArray javaTensorAlongDimension(int index, int... dimension);
-
-
     /**
      * Returns the cumulative sum along a dimension. In-place method.
      *
@@ -190,8 +183,7 @@ public interface INDArray extends Serializable, AutoCloseable {
     INDArray cumsum(int dimension);
 
     /**
-     * Assign all of the elements in the given
-     * ndarray to this ndarray
+     * Assign all of the elements in the given ndarray to this ndarray
      *
      * @param arr the elements to assign
      * @return this
@@ -254,9 +246,19 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray putScalar(int[] i, double value);
 
-
+    /**
+     * See {@link #putScalar(int[], double)}
+     */
     INDArray putScalar(long[] i, double value);
+
+    /**
+     * See {@link #putScalar(int[], double)}
+     */
     INDArray putScalar(long[] i, float value);
+
+    /**
+     * See {@link #putScalar(int[], double)}
+     */
     INDArray putScalar(long[] i, int value);
 
     /**
@@ -300,7 +302,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray lt(Number other);
 
-
     /**
      * Put the specified float value at the specified indices in this array
      *
@@ -327,8 +328,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray eps(Number other);
 
-
-
     /**
      * Returns the binary ndarray for "Equals" comparison.
      *
@@ -367,10 +366,8 @@ public interface INDArray extends Serializable, AutoCloseable {
      * @param other the ndarray to compare.
      * @return the binary ndarray for "Less" comparison.
      */
-
     INDArray lt(INDArray other);
 
-
     /**
      * Returns the binary ndarray for "Epsilon equals" comparison.
      *
@@ -403,7 +400,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray eq(INDArray other);
 
-
     /**
      * Returns the binary ndarray for "Greater Than" comparison.
      *
@@ -424,8 +420,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray isNaN();
 
-
-
     /**
      * Returns the ndarray negative (cloned)
      *
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/checkutil/NDArrayCreationUtil.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/checkutil/NDArrayCreationUtil.java
index 6039e4b1a..f3ac98565 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/checkutil/NDArrayCreationUtil.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/checkutil/NDArrayCreationUtil.java
@@ -181,37 +181,37 @@ public class NDArrayCreationUtil {
         INDArray[] out = new INDArray[12];
 
         INDArray temp01 = Nd4j.linspace(1, cols * rows * 4, cols * rows * 4, dataType).reshape(cols, rows, 4);
-        out[0] = temp01.javaTensorAlongDimension(0, 0, 1).reshape(rows, cols);
+        out[0] = temp01.tensorAlongDimension(0, 0, 1).reshape(rows, cols);
         long[] temp01Shape = new long[] {cols, rows, 4};
         int len = ArrayUtil.prod(temp01Shape);
         temp01 = Nd4j.linspace(1, len, len, dataType).reshape(temp01Shape);
-        out[1] = temp01.javaTensorAlongDimension(2, 0, 1).reshape(rows, cols);
+        out[1] = temp01.tensorAlongDimension(2, 0, 1).reshape(rows, cols);
 
         Nd4j.getRandom().setSeed(seed);
         INDArray temp02 = Nd4j.linspace(1, len, len, dataType).reshape(new long[] {cols, 4, rows});
-        out[2] = temp02.javaTensorAlongDimension(0, 0, 2).reshape(rows, cols);
+        out[2] = temp02.tensorAlongDimension(0, 0, 2).reshape(rows, cols);
         temp02 = Nd4j.linspace(1, len, len, dataType).reshape(cols, 4, rows);
-        out[3] = temp02.javaTensorAlongDimension(2, 0, 2).reshape(rows, cols);
+        out[3] = temp02.tensorAlongDimension(2, 0, 2).reshape(rows, cols);
 
         INDArray temp10 = Nd4j.linspace(1, len, len, dataType).reshape(rows, cols, 4);
-        out[4] = temp10.javaTensorAlongDimension(0, 1, 0).reshape(rows, cols);
+        out[4] = temp10.tensorAlongDimension(0, 1, 0).reshape(rows, cols);
         temp10 = Nd4j.linspace(1, len, len, dataType).reshape(rows, cols, 4);
-        out[5] = temp10.javaTensorAlongDimension(2, 1, 0).reshape(rows, cols);
+        out[5] = temp10.tensorAlongDimension(2, 1, 0).reshape(rows, cols);
 
         INDArray temp12 = Nd4j.linspace(1, len, len, dataType).reshape(4, cols, rows);
-        out[6] = temp12.javaTensorAlongDimension(0, 1, 2).reshape(rows, cols);
+        out[6] = temp12.tensorAlongDimension(0, 1, 2).reshape(rows, cols);
         temp12 = Nd4j.linspace(1, len, len, dataType).reshape(4, cols, rows);
-        out[7] = temp12.javaTensorAlongDimension(2, 1, 2).reshape(rows, cols);
+        out[7] = temp12.tensorAlongDimension(2, 1, 2).reshape(rows, cols);
 
         INDArray temp20 = Nd4j.linspace(1, len, len, dataType).reshape(rows, 4, cols);
-        out[8] = temp20.javaTensorAlongDimension(0, 2, 0).reshape(rows, cols);
+        out[8] = temp20.tensorAlongDimension(0, 2, 0).reshape(rows, cols);
         temp20 = Nd4j.linspace(1, len, len, dataType).reshape(rows, 4, cols);
-        out[9] = temp20.javaTensorAlongDimension(2, 2, 0).reshape(rows, cols);
+        out[9] = temp20.tensorAlongDimension(2, 2, 0).reshape(rows, cols);
 
         INDArray temp21 = Nd4j.linspace(1, len, len, dataType).reshape(4, rows, cols);
-        out[10] = temp21.javaTensorAlongDimension(0, 2, 1).reshape(rows, cols);
+        out[10] = temp21.tensorAlongDimension(0, 2, 1).reshape(rows, cols);
         temp21 = Nd4j.linspace(1, len, len, dataType).reshape(4, rows, cols);
-        out[11] = temp21.javaTensorAlongDimension(2, 2, 1).reshape(rows, cols);
+        out[11] = temp21.tensorAlongDimension(2, 2, 1).reshape(rows, cols);
 
         String baseMsg = "getTensorAlongDimensionMatricesWithShape(" + rows + "," + cols + "," + seed + ")";
         List<Pair<INDArray, String>> list = new ArrayList<>(12);
@@ -361,9 +361,9 @@ public class NDArrayCreationUtil {
         val shape4d1 = new long[]{shape[0], shape[1], shape[2], 3};
         int lenshape4d1 = ArrayUtil.prod(shape4d1);
         INDArray orig1a = Nd4j.linspace(1, lenshape4d1, lenshape4d1, dataType).reshape(shape4d1);
-        INDArray tad1a = orig1a.javaTensorAlongDimension(0, 0, 1, 2);
+        INDArray tad1a = orig1a.tensorAlongDimension(0, 0, 1, 2);
         INDArray orig1b = Nd4j.linspace(1, lenshape4d1, lenshape4d1, dataType).reshape(shape4d1);
-        INDArray tad1b = orig1b.javaTensorAlongDimension(1, 0, 1, 2);
+        INDArray tad1b = orig1b.tensorAlongDimension(1, 0, 1, 2);
 
         list.add(new Pair<>(tad1a, baseMsg + ".get(0)"));
         list.add(new Pair<>(tad1b, baseMsg + ".get(1)"));
@@ -371,19 +371,19 @@ public class NDArrayCreationUtil {
         long[] shape4d2 = {3, shape[0], shape[1], shape[2]};
         int lenshape4d2 = ArrayUtil.prod(shape4d2);
         INDArray orig2 = Nd4j.linspace(1, lenshape4d2, lenshape4d2, dataType).reshape(shape4d2);
-        INDArray tad2 = orig2.javaTensorAlongDimension(1, 1, 2, 3);
+        INDArray tad2 = orig2.tensorAlongDimension(1, 1, 2, 3);
         list.add(new Pair<>(tad2, baseMsg + ".get(2)"));
 
         long[] shape4d3 = {shape[0], shape[1], 3, shape[2]};
         int lenshape4d3 = ArrayUtil.prod(shape4d3);
         INDArray orig3 = Nd4j.linspace(1, lenshape4d3, lenshape4d3, dataType).reshape(shape4d3);
-        INDArray tad3 = orig3.javaTensorAlongDimension(1, 1, 3, 0);
+        INDArray tad3 = orig3.tensorAlongDimension(1, 1, 3, 0);
         list.add(new Pair<>(tad3, baseMsg + ".get(3)"));
 
         long[] shape4d4 = {shape[0], 3, shape[1], shape[2]};
         int lenshape4d4 = ArrayUtil.prod(shape4d4);
         INDArray orig4 = Nd4j.linspace(1, lenshape4d4, lenshape4d4, dataType).reshape(shape4d4);
-        INDArray tad4 = orig4.javaTensorAlongDimension(1, 2, 0, 3);
+        INDArray tad4 = orig4.tensorAlongDimension(1, 2, 0, 3);
         list.add(new Pair<>(tad4, baseMsg + ".get(4)"));
 
         return list;
@@ -513,9 +513,9 @@ public class NDArrayCreationUtil {
         int[] shape4d1 = {3, shape[0], shape[1], shape[2], shape[3]};
         int len = ArrayUtil.prod(shape4d1);
         INDArray orig1a = Nd4j.linspace(1, len, len, dataType).reshape(ArrayUtil.toLongArray(shape4d1));
-        INDArray tad1a = orig1a.javaTensorAlongDimension(0, 1, 2, 3, 4);
+        INDArray tad1a = orig1a.tensorAlongDimension(0, 1, 2, 3, 4);
         INDArray orig1b = Nd4j.linspace(1, len, len, dataType).reshape(ArrayUtil.toLongArray(shape4d1));
-        INDArray tad1b = orig1b.javaTensorAlongDimension(2, 1, 2, 3, 4);
+        INDArray tad1b = orig1b.tensorAlongDimension(2, 1, 2, 3, 4);
 
         list.add(new Pair<>(tad1a, baseMsg + ".get(0)"));
         list.add(new Pair<>(tad1b, baseMsg + ".get(1)"));
@@ -523,19 +523,19 @@ public class NDArrayCreationUtil {
         int[] shape4d2 = {3, shape[0], shape[1], shape[2], shape[3]};
         int len2 = ArrayUtil.prod(shape4d2);
         INDArray orig2 = Nd4j.linspace(1, len2, len2, dataType).reshape(ArrayUtil.toLongArray(shape4d2));
-        INDArray tad2 = orig2.javaTensorAlongDimension(1, 3, 4, 2, 1);
+        INDArray tad2 = orig2.tensorAlongDimension(1, 3, 4, 2, 1);
         list.add(new Pair<>(tad2, baseMsg + ".get(2)"));
 
         int[] shape4d3 = {shape[0], shape[1], 3, shape[2], shape[3]};
         int len3 = ArrayUtil.prod(shape4d3);
         INDArray orig3 = Nd4j.linspace(1, len3, len3, dataType).reshape(ArrayUtil.toLongArray(shape4d3));
-        INDArray tad3 = orig3.javaTensorAlongDimension(1, 4, 1, 3, 0);
+        INDArray tad3 = orig3.tensorAlongDimension(1, 4, 1, 3, 0);
         list.add(new Pair<>(tad3, baseMsg + ".get(3)"));
 
         int[] shape4d4 = {shape[0], shape[1], shape[2], shape[3], 3};
         int len4 = ArrayUtil.prod(shape4d4);
         INDArray orig4 = Nd4j.linspace(1, len4, len4, dataType).reshape(ArrayUtil.toLongArray(shape4d4));
-        INDArray tad4 = orig4.javaTensorAlongDimension(1, 2, 0, 3, 1);
+        INDArray tad4 = orig4.tensorAlongDimension(1, 2, 0, 3, 1);
         list.add(new Pair<>(tad4, baseMsg + ".get(4)"));
 
         return list;
@@ -655,26 +655,26 @@ public class NDArrayCreationUtil {
         Nd4j.getRandom().setSeed(seed);
         int[] shape4d1 = {3, shape[0], shape[1], shape[2], shape[3], shape[4]};
         INDArray orig1a = Nd4j.rand(dataType, shape4d1);
-        INDArray tad1a = orig1a.javaTensorAlongDimension(0, 1, 2, 3, 4, 5);
+        INDArray tad1a = orig1a.tensorAlongDimension(0, 1, 2, 3, 4, 5);
         INDArray orig1b = Nd4j.rand(dataType, shape4d1);
-        INDArray tad1b = orig1b.javaTensorAlongDimension(2, 1, 2, 3, 4, 5);
+        INDArray tad1b = orig1b.tensorAlongDimension(2, 1, 2, 3, 4, 5);
 
         list.add(new Pair<>(tad1a, baseMsg + ".get(0)"));
         list.add(new Pair<>(tad1b, baseMsg + ".get(1)"));
 
         int[] shape4d2 = {3, shape[0], shape[1], shape[2], shape[3], shape[4]};
         INDArray orig2 = Nd4j.rand(dataType, shape4d2);
-        INDArray tad2 = orig2.javaTensorAlongDimension(1, 3, 5, 4, 2, 1);
+        INDArray tad2 = orig2.tensorAlongDimension(1, 3, 5, 4, 2, 1);
         list.add(new Pair<>(tad2, baseMsg + ".get(2)"));
 
         int[] shape4d3 = {shape[0], shape[1], shape[2], shape[3], shape[4], 2};
         INDArray orig3 = Nd4j.rand(dataType, shape4d3);
-        INDArray tad3 = orig3.javaTensorAlongDimension(1, 4, 1, 3, 2, 0);
+        INDArray tad3 = orig3.tensorAlongDimension(1, 4, 1, 3, 2, 0);
         list.add(new Pair<>(tad3, baseMsg + ".get(3)"));
 
         int[] shape4d4 = {shape[0], shape[1], shape[2], shape[3], 3, shape[4]};
         INDArray orig4 = Nd4j.rand(dataType, shape4d4);
-        INDArray tad4 = orig4.javaTensorAlongDimension(1, 5, 2, 0, 3, 1);
+        INDArray tad4 = orig4.tensorAlongDimension(1, 5, 2, 0, 3, 1);
         list.add(new Pair<>(tad4, baseMsg + ".get(4)"));
 
         return list;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java
index 2ae7a32c4..a56d39567 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java
@@ -374,7 +374,7 @@ public class DataSet implements org.nd4j.linalg.dataset.api.DataSet {
         long nTensors = labels.tensorsAlongDimension(1);
         for (int i = 0; i < nTensors; i++) {
             INDArray row = labels.tensorAlongDimension(i, 1);
-            INDArray javaRow = labels.javaTensorAlongDimension(i, 1);
+            INDArray javaRow = labels.tensorAlongDimension(i, 1);
             int maxIdx = Nd4j.getBlasWrapper().iamax(row);
             int maxIdxJava = Nd4j.getBlasWrapper().iamax(javaRow);
             if (maxIdx < 0)
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/Nd4jTestsC.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/Nd4jTestsC.java
index 915d6f650..22b911468 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/Nd4jTestsC.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/Nd4jTestsC.java
@@ -1334,7 +1334,7 @@ public class Nd4jTestsC extends BaseNd4jTest {
             INDArray zC = Nd4j.create(shape, 'c');
             zC.setData(Nd4j.linspace(1, 24, 24, DataType.DOUBLE).data());
             for (int tad = 0; tad < zC.tensorsAlongDimension(dim); tad++) {
-                INDArray javaTad = zC.javaTensorAlongDimension(tad, dim);
+                INDArray javaTad = zC.tensorAlongDimension(tad, dim);
                 System.out.println("Tad " + tad + " is " + zC.tensorAlongDimension(tad, dim));
             }
 
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/tad/TestTensorAlongDimension.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/tad/TestTensorAlongDimension.java
index 755fd4d20..cede0543a 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/tad/TestTensorAlongDimension.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/tad/TestTensorAlongDimension.java
@@ -57,7 +57,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
         for (int i = 0; i < n; i++) {
             StopWatch javaTiming = new StopWatch();
             javaTiming.start();
-            row.javaTensorAlongDimension(0, 0);
+            row.tensorAlongDimension(0, 0);
             javaTiming.stop();
             StopWatch cTiming = new StopWatch();
             cTiming.start();
@@ -98,7 +98,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
             assertEquals(cols, arr.tensorsAlongDimension(0));
             for (int i = 0; i < cols; i++) {
                 INDArray tad = arr.tensorAlongDimension(i, 0);
-                INDArray javaTad = arr.javaTensorAlongDimension(i, 0);
+                INDArray javaTad = arr.tensorAlongDimension(i, 0);
                 assertEquals(javaTad, tad);
                 assertArrayEquals(new int[] {rows}, tad.shape());
                 //assertEquals(testValues.javaTensorAlongDimension(i, 0), tad);
@@ -120,7 +120,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
         list = NDArrayCreationUtil.getAll3dTestArraysWithShape(12345, new int[]{rows, cols, dim2}, DataType.DOUBLE);
         for (Pair<INDArray, String> p : list) {
             INDArray arr = p.getFirst().assign(testValues);
-            INDArray javaTad = arr.javaTensorAlongDimension(0, 0);
+            INDArray javaTad = arr.tensorAlongDimension(0, 0);
             INDArray tadTest = arr.tensorAlongDimension(0, 0);
             assertEquals(javaTad, tadTest);
             //Along dimension 0: expect row vector with length 'rows'
@@ -165,7 +165,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
             //Along dimension 0,1: expect matrix with shape [rows,cols]
             assertEquals(dim2, arr.tensorsAlongDimension(0, 1));
             for (int i = 0; i < dim2; i++) {
-                INDArray javaTad = arr.javaTensorAlongDimension(i, 0, 1);
+                INDArray javaTad = arr.tensorAlongDimension(i, 0, 1);
                 INDArray tad = arr.tensorAlongDimension(i, 0, 1);
                 int javaEleStride = javaTad.elementWiseStride();
                 int testTad = tad.elementWiseStride();
@@ -178,11 +178,11 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
             //Along dimension 0,2: expect matrix with shape [rows,dim2]
             assertEquals(cols, arr.tensorsAlongDimension(0, 2));
             for (int i = 0; i < cols; i++) {
-                INDArray javaTad = arr.javaTensorAlongDimension(i, 0, 2);
+                INDArray javaTad = arr.tensorAlongDimension(i, 0, 2);
                 INDArray tad = arr.tensorAlongDimension(i, 0, 2);
                 assertEquals(javaTad, tad);
                 assertArrayEquals(new long[] {rows, dim2}, tad.shape());
-                assertEquals(testValues.javaTensorAlongDimension(i, 0, 2), tad);
+                assertEquals(testValues.tensorAlongDimension(i, 0, 2), tad);
             }
 
             //Along dimension 1,2: expect matrix with shape [cols,dim2]
@@ -190,7 +190,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
             for (int i = 0; i < rows; i++) {
                 INDArray tad = arr.tensorAlongDimension(i, 1, 2);
                 assertArrayEquals(new long[] {cols, dim2}, tad.shape());
-                assertEquals(testValues.javaTensorAlongDimension(i, 1, 2), tad);
+                assertEquals(testValues.tensorAlongDimension(i, 1, 2), tad);
             }
         }
 
@@ -207,7 +207,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
             for (int i = 0; i < dim2 * dim3; i++) {
                 INDArray tad = arr.tensorAlongDimension(i, 0, 1);
                 assertArrayEquals(new long[] {rows, cols}, tad.shape());
-                assertEquals(testValues.javaTensorAlongDimension(i, 0, 1), tad);
+                assertEquals(testValues.tensorAlongDimension(i, 0, 1), tad);
             }
 
             //Along dimension 0,2: expect matrix with shape [rows,dim2]
@@ -215,7 +215,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
             for (int i = 0; i < cols * dim3; i++) {
                 INDArray tad = arr.tensorAlongDimension(i, 0, 2);
                 assertArrayEquals(new long[] {rows, dim2}, tad.shape());
-                assertEquals(testValues.javaTensorAlongDimension(i, 0, 2), tad);
+                assertEquals(testValues.tensorAlongDimension(i, 0, 2), tad);
             }
 
             //Along dimension 0,3: expect matrix with shape [rows,dim3]
@@ -223,7 +223,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
             for (int i = 0; i < cols * dim2; i++) {
                 INDArray tad = arr.tensorAlongDimension(i, 0, 3);
                 assertArrayEquals(new long[] {rows, dim3}, tad.shape());
-                assertEquals(testValues.javaTensorAlongDimension(i, 0, 3), tad);
+                assertEquals(testValues.tensorAlongDimension(i, 0, 3), tad);
             }
 
 
@@ -232,7 +232,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
             for (int i = 0; i < rows * dim3; i++) {
                 INDArray tad = arr.tensorAlongDimension(i, 1, 2);
                 assertArrayEquals(new long[] {cols, dim2}, tad.shape());
-                assertEquals(testValues.javaTensorAlongDimension(i, 1, 2), tad);
+                assertEquals(testValues.tensorAlongDimension(i, 1, 2), tad);
             }
 
             //Along dimension 1,3: expect matrix with shape [cols,dim3]
@@ -240,7 +240,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
             for (int i = 0; i < rows * dim2; i++) {
                 INDArray tad = arr.tensorAlongDimension(i, 1, 3);
                 assertArrayEquals(new long[] {cols, dim3}, tad.shape());
-                assertEquals(testValues.javaTensorAlongDimension(i, 1, 3), tad);
+                assertEquals(testValues.tensorAlongDimension(i, 1, 3), tad);
             }
 
             //Along dimension 2,3: expect matrix with shape [dim2,dim3]
@@ -248,7 +248,7 @@ public class TestTensorAlongDimension extends BaseNd4jTest {
             for (int i = 0; i < rows * cols; i++) {
                 INDArray tad = arr.tensorAlongDimension(i, 2, 3);
                 assertArrayEquals(new long[] {dim2, dim3}, tad.shape());
-                assertEquals(testValues.javaTensorAlongDimension(i, 2, 3), tad);
+                assertEquals(testValues.tensorAlongDimension(i, 2, 3), tad);
             }
         }
     }
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/ops/OpExecutionerTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/ops/OpExecutionerTests.java
index 49391b74c..52ede954a 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/ops/OpExecutionerTests.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/ops/OpExecutionerTests.java
@@ -614,7 +614,7 @@ public class OpExecutionerTests extends BaseNd4jTest {
             assertEquals(arr5s.getDouble(i), 16, 1e-1);
         System.out.println("6d");
         INDArray arr6 = Nd4j.ones(1, 1, 4, 4, 4, 4);
-        INDArray arr6Tad = arr6.javaTensorAlongDimension(0, 2, 3);
+        INDArray arr6Tad = arr6.tensorAlongDimension(0, 2, 3);
         INDArray arr6s = arr6.sum(2, 3);
         for (int i = 0; i < arr6s.length(); i++)
             assertEquals(arr6s.getDouble(i), 16, 1e-1);
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/TADTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/TADTests.java
index f4d2bf80c..6f47d00da 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/TADTests.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/TADTests.java
@@ -79,7 +79,7 @@ public class TADTests extends BaseNd4jTest {
 
                     int[] shape = new int[] {e, x};
                     Arrays.sort(shape);
-                    INDArray assertion = array.javaTensorAlongDimension(0, shape);
+                    INDArray assertion = array.tensorAlongDimension(0, shape);
                     INDArray test = array.tensorAlongDimension(0, shape);
 
                     assertEquals(assertion, test);
@@ -101,7 +101,7 @@ public class TADTests extends BaseNd4jTest {
                 Arrays.sort(shape);
                 log.info("About to do shape: " + Arrays.toString(shape) + " for array of shape "
                                 + array.shapeInfoToString());
-                INDArray assertion = array.javaTensorAlongDimension(0, shape);
+                INDArray assertion = array.tensorAlongDimension(0, shape);
                 INDArray test = array.tensorAlongDimension(0, shape);
                 assertEquals(assertion, test);
                 //assertEquals(assertion.shapeInfoDataBuffer(), test.shapeInfoDataBuffer());
@@ -121,8 +121,8 @@ public class TADTests extends BaseNd4jTest {
     public void testMysteriousCrash() {
         INDArray arrayF = Nd4j.create(new int[] {1, 1, 4, 4}, 'f');
         INDArray arrayC = Nd4j.create(new int[] {1, 1, 4, 4}, 'c');
-        INDArray javaCTad = arrayC.javaTensorAlongDimension(0, 2, 3);
-        INDArray javaFTad = arrayF.javaTensorAlongDimension(0, 2, 3);
+        INDArray javaCTad = arrayC.tensorAlongDimension(0, 2, 3);
+        INDArray javaFTad = arrayF.tensorAlongDimension(0, 2, 3);
         Pair<DataBuffer, DataBuffer> tadBuffersF =
                         Nd4j.getExecutioner().getTADManager().getTADOnlyShapeInfo(arrayF, 2, 3);
         Pair<DataBuffer, DataBuffer> tadBuffersC =
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/concat/ConcatTestsC.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/concat/ConcatTestsC.java
index 596bf16a7..806cf4d08 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/concat/ConcatTestsC.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/shape/concat/ConcatTestsC.java
@@ -185,7 +185,7 @@ public class ConcatTestsC extends BaseNd4jTest {
         //ConcatV2, dim 1
         second = Nd4j.linspace(24, 32, 8, Nd4j.dataType()).reshape('c', 2, 1, 4);
         for (int i = 0; i < second.tensorsAlongDimension(1); i++) {
-            INDArray secondTad = second.javaTensorAlongDimension(i, 1);
+            INDArray secondTad = second.tensorAlongDimension(i, 1);
             System.out.println(second.tensorAlongDimension(i, 1));
         }
 

From b46f9827b88cf1566381b9ba4efe900a7dcde7b5 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Wed, 28 Aug 2019 13:27:00 +1000
Subject: [PATCH 31/56] Layer norm test updates (#187)

Signed-off-by: Alex Black <blacka101@gmail.com>
---
 .../functions/DifferentialFunction.java       |  2 +-
 .../ops/impl/transforms/custom/LayerNorm.java |  6 +++
 .../impl/transforms/custom/LayerNormBp.java   |  2 +
 .../opvalidation/LayerOpValidation.java       | 38 +++++++++----------
 .../opvalidation/TransformOpValidation.java   |  2 +-
 5 files changed, 29 insertions(+), 21 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java
index 34240516f..eb3424007 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunction.java
@@ -530,7 +530,7 @@ public abstract class DifferentialFunction {
     public SDVariable arg(int num){
         SDVariable[] args = args();
         Preconditions.checkNotNull(args, "Arguments are null for function %s", this.getOwnName());
-        Preconditions.checkArgument(num >= 0 && num < args.length, "Invalid index: must be 0 to numArgs (0 <= idx < %s)", args.length);
+        Preconditions.checkArgument(num >= 0 && num < args.length, "Invalid index: must be 0 to numArgs (0 <= idx < %s), got %s", args.length, num);
         return args[num];
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java
index 137bc6693..7c7c34fc5 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNorm.java
@@ -46,6 +46,7 @@ public class LayerNorm extends DynamicCustomOp {
 
     public LayerNorm(@NonNull SameDiff sameDiff, @NonNull SDVariable input, @NonNull SDVariable gain, SDVariable bias, boolean channelsFirst, int... dimensions) {
         super(null, sameDiff, wrapFilterNull(input, gain, bias), false);
+        this.noBias = bias == null;
         this.channelsFirst = channelsFirst;
         setDimensions(dimensions);
     }
@@ -56,6 +57,7 @@ public class LayerNorm extends DynamicCustomOp {
 
     public LayerNorm(INDArray input, INDArray gain, INDArray bias, INDArray result, boolean channelsFirst, int... dimensions) {
         super("layer_norm", wrapFilterNull(input, gain, bias), wrapOrNull(result));
+        this.noBias = bias == null;
         this.channelsFirst = channelsFirst;
         setDimensions(dimensions);
     }
@@ -115,4 +117,8 @@ public class LayerNorm extends DynamicCustomOp {
         return Collections.singletonList(first);
     }
 
+    @Override
+    public int numOutputArguments() {
+        return noBias ? 2 : 3;
+    }
 }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNormBp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNormBp.java
index 2168fd165..f55db2e50 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNormBp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LayerNormBp.java
@@ -45,12 +45,14 @@ public class LayerNormBp extends DynamicCustomOp {
 
     public LayerNormBp(@NonNull SameDiff sameDiff, @NonNull SDVariable input, @NonNull SDVariable gain, SDVariable bias, @NonNull SDVariable gradient, boolean channelsFirst, int... dimensions) {
         super(null, sameDiff, wrapFilterNull(input, gain, bias, gradient), false);
+        this.noBias = bias == null;
         this.channelsFirst = channelsFirst;
         setDimensions(dimensions);
     }
 
     public LayerNormBp(@NonNull INDArray input, @NonNull INDArray gain, INDArray bias, @NonNull INDArray grad, @NonNull INDArray dLdx, @NonNull INDArray dLdg, INDArray dLdb, boolean channelsFirst, int... dimensions) {
         super("layer_norm_bp", wrapFilterNull(input, gain, bias, grad), wrapFilterNull(dLdx, dLdg, dLdb));
+        this.noBias = bias == null;
         this.channelsFirst = channelsFirst;
         setDimensions(dimensions);
     }
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
index 760165b3b..057f610bd 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/LayerOpValidation.java
@@ -1112,12 +1112,12 @@ public class LayerOpValidation extends BaseOpValidation {
 
     @Test
     public void testLayerNorm() {
-        final INDArray random = Nd4j.rand(new int[]{10, 4});
+        final INDArray random = Nd4j.rand(DataType.DOUBLE, 10, 4);
         final INDArray standardized = random.ulike();
         Nd4j.getExecutioner().exec(new Standardize(random, standardized, 1));
 
-        final INDArray gain = Nd4j.rand(new int[]{1, 4});
-        final INDArray bias = Nd4j.rand(new int[]{1, 4});
+        final INDArray gain = Nd4j.rand(DataType.DOUBLE, 4);
+        final INDArray bias = Nd4j.rand(DataType.DOUBLE, 4);
         final INDArray res = standardized.mulRowVector(gain).addRowVector(bias);
         final INDArray expOut = res.norm1();
 
@@ -1132,7 +1132,7 @@ public class LayerOpValidation extends BaseOpValidation {
         String err = OpValidation.validate(new TestCase(sd)
                 .expectedOutput("out", expOut)
                 .gradientCheck(true));
-        assertNull(err, err);
+        assertNull(err);
     }
 
     @Test
@@ -1141,9 +1141,9 @@ public class LayerOpValidation extends BaseOpValidation {
         int ch = 4;
         for(boolean nchw : new boolean[]{true, false}) {
             double eps = 0.0;
-            INDArray x = Nd4j.rand(DataType.FLOAT, nchw ? new long[]{mb, ch, 8, 8} : new long[]{mb, 8, 8, ch});
-            INDArray gain4d = Nd4j.rand(DataType.FLOAT, nchw ? new long[]{1, ch, 1, 1} : new long[]{1, 1, 1, ch});
-            INDArray bias4d = Nd4j.rand(DataType.FLOAT, nchw ? new long[]{1, ch, 1, 1} : new long[]{1, 1, 1, ch});
+            INDArray x = Nd4j.rand(DataType.DOUBLE, nchw ? new long[]{mb, ch, 8, 8} : new long[]{mb, 8, 8, ch});
+            INDArray gain4d = Nd4j.rand(DataType.DOUBLE, nchw ? new long[]{1, ch, 1, 1} : new long[]{1, 1, 1, ch});
+            INDArray bias4d = Nd4j.rand(DataType.DOUBLE, nchw ? new long[]{1, ch, 1, 1} : new long[]{1, 1, 1, ch});
             INDArray mean = x.mean(true, 1, 2, 3);
             INDArray std = Transforms.sqrt(x.var(false,1,2,3).addi(eps)).reshape(mb, 1, 1, 1);
 
@@ -1169,12 +1169,12 @@ public class LayerOpValidation extends BaseOpValidation {
 
     @Test
     public void testLayerNormOP() {
-        final INDArray random = Nd4j.rand(new int[]{10, 4});
+        final INDArray random = Nd4j.rand(DataType.DOUBLE, 10, 4);
         final INDArray standardized = random.ulike();
         Nd4j.getExecutioner().exec(new Standardize(random, standardized, 1));
 
-        final INDArray gain = Nd4j.rand(new int[]{1, 4});
-        final INDArray bias = Nd4j.rand(new int[]{1, 4});
+        final INDArray gain = Nd4j.rand(DataType.DOUBLE, 4);
+        final INDArray bias = Nd4j.rand(DataType.DOUBLE, 4);
         final INDArray res = standardized.mulRowVector(gain).addRowVector(bias);
 
         final INDArray output = Nd4j.zerosLike(res);
@@ -1185,11 +1185,11 @@ public class LayerOpValidation extends BaseOpValidation {
 
     @Test
     public void testLayerNormNoBias() {
-        final INDArray random = Nd4j.rand(new int[]{10, 4});
+        final INDArray random = Nd4j.rand(DataType.DOUBLE, 10, 4);
         final INDArray standardized = random.ulike();
         Nd4j.getExecutioner().exec(new Standardize(random, standardized, 1));
 
-        final INDArray gain = Nd4j.rand(new int[]{1, 4});
+        final INDArray gain = Nd4j.rand(DataType.DOUBLE, 4);
         final INDArray res = standardized.mulRowVector(gain);
         final INDArray expOut = res.norm1();
 
@@ -1208,11 +1208,11 @@ public class LayerOpValidation extends BaseOpValidation {
 
     @Test
     public void testLayerNormOPNoBias() {
-        final INDArray random = Nd4j.rand(new int[]{10, 4});
+        final INDArray random = Nd4j.rand(DataType.DOUBLE, 10, 4);
         final INDArray standardized = random.ulike();
         Nd4j.getExecutioner().exec(new Standardize(random, standardized, 1));
 
-        final INDArray gain = Nd4j.rand(new int[]{1, 4});
+        final INDArray gain = Nd4j.rand(DataType.DOUBLE,4);
         final INDArray res = standardized.mulRowVector(gain);
 
         final INDArray output = Nd4j.zerosLike(res);
@@ -1223,7 +1223,7 @@ public class LayerOpValidation extends BaseOpValidation {
 
     @Test
     public void testLayerNormNoDeviation() {
-        final INDArray random = Nd4j.rand(new int[]{10, 4});
+        final INDArray random = Nd4j.rand(DataType.DOUBLE, 10, 4);
         for (int i = 0; i < 4; i++) {
             random.putScalar(1,i, 7);
         }
@@ -1231,8 +1231,8 @@ public class LayerOpValidation extends BaseOpValidation {
         final INDArray standardized = random.ulike();
         Nd4j.getExecutioner().exec(new Standardize(random, standardized, 1));
 
-        final INDArray gain = Nd4j.rand(new int[]{1, 4});
-        final INDArray bias = Nd4j.rand(new int[]{1, 4});
+        final INDArray gain = Nd4j.rand(DataType.DOUBLE, 4);
+        final INDArray bias = Nd4j.rand(DataType.DOUBLE, 4);
         final INDArray res = standardized.mulRowVector(gain).addRowVector(bias);
         final INDArray expOut = res.norm1();
 
@@ -1332,8 +1332,8 @@ public class LayerOpValidation extends BaseOpValidation {
     public void testLayerNormMixedOrders(){
         Nd4j.getRandom().setSeed(12345);
         INDArray input = Nd4j.rand(DataType.DOUBLE, 3, 8).dup('f');
-        INDArray gain = Nd4j.rand(DataType.DOUBLE, 1, 8).dup('f');
-        INDArray bias = Nd4j.rand(DataType.DOUBLE, 1, 8).dup('f');
+        INDArray gain = Nd4j.rand(DataType.DOUBLE, 8).dup('f');
+        INDArray bias = Nd4j.rand(DataType.DOUBLE, 8).dup('f');
 
         INDArray outFF = Nd4j.create(DataType.DOUBLE, new long[]{3,8}, 'f');
         INDArray outCC = Nd4j.create(DataType.DOUBLE, new long[]{3,8}, 'c');
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java
index 9183a0884..fdd2b3160 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java
@@ -412,7 +412,7 @@ public class TransformOpValidation extends BaseOpValidation {
                 .expectedOutput("dp0", expOut[0])
                 .expectedOutput("dp1", expOut[1])
                 .gradientCheck(true));
-        assertNull(err, err);
+        assertNull(err);
     }
 
     @Test

From 2a1431264f741d6a90fe0176d7c48377790a6d3c Mon Sep 17 00:00:00 2001
From: Ryan Nett <rnett@skymind.io>
Date: Tue, 27 Aug 2019 20:39:32 -0700
Subject: [PATCH 32/56] Remove calculate output shape from java side (#151)

* remove some unneeded java-side output shape calculations

Signed-off-by: Ryan Nett <rnett@skymind.io>

* delete Broadcast

Signed-off-by: Ryan Nett <rnett@skymind.io>

* delete Linear and Module,

Signed-off-by: Ryan Nett <rnett@skymind.io>

* update Identity, HashCode, and NoOp

Signed-off-by: Ryan Nett <rnett@skymind.io>

* removed Cast java-side shape function, added tests and SDVariable.isEmpty

Signed-off-by: Ryan Nett <rnett@skymind.io>

* ignoring test w/ issues on master

Signed-off-by: Ryan Nett <rnett@skymind.io>

* noop needs more work, fixed BaseArithmeticBackprop and BaseDynamicTransform ops

merge in master for c++ build fix

Signed-off-by: Ryan Nett <rnett@skymind.io>

* fix EqualTo

Signed-off-by: Ryan Nett <rnett@skymind.io>

* fix other cond ops

Signed-off-by: Ryan Nett <rnett@skymind.io>

* "fake" ops calculateOutputShape() throws exception

Signed-off-by: Ryan Nett <rnett@skymind.io>

* use c++ shape calc for Linspace

Signed-off-by: Ryan Nett <rnett@skymind.io>

* fix exception message, move most to BaseCompatOp

Signed-off-by: Ryan Nett <rnett@skymind.io>

* remove SDVariable.isEmpty

Signed-off-by: Ryan Nett <rnett@skymind.io>

* remove commented out code

Signed-off-by: Ryan Nett <rnett@skymind.io>
---
 .../DifferentialFunctionFactory.java          |   9 -
 .../nd4j/autodiff/samediff/SDVariable.java    |   3 -
 .../converters/ImportClassMapping.java        |   2 -
 .../org/nd4j/linalg/api/ops/BaseModule.java   |  63 ------
 .../java/org/nd4j/linalg/api/ops/Module.java  |  51 -----
 .../java/org/nd4j/linalg/api/ops/NoOp.java    |   1 +
 .../impl/controlflow/compat/BaseCompatOp.java |   7 +
 .../ops/impl/controlflow/compat/Enter.java    |   9 -
 .../api/ops/impl/controlflow/compat/Exit.java |   9 -
 .../ops/impl/controlflow/compat/LoopCond.java |   9 -
 .../ops/impl/controlflow/compat/Merge.java    |   9 -
 .../controlflow/compat/NextIteration.java     |   9 -
 .../ops/impl/controlflow/compat/Switch.java   |  12 --
 .../linalg/api/ops/impl/layers/Linear.java    | 198 ------------------
 .../nd4j/linalg/api/ops/impl/loss/L2Loss.java |   5 -
 .../linalg/api/ops/impl/reduce/HashCode.java  |   5 -
 .../linalg/api/ops/impl/shape/Broadcast.java  |  78 -------
 .../linalg/api/ops/impl/shape/Linspace.java   |   8 -
 .../nd4j/linalg/api/ops/impl/shape/Rank.java  |   8 -
 .../linalg/api/ops/impl/shape/Unstack.java    |  13 --
 .../impl/shape/tensorops/BaseTensorOp.java    |   5 +-
 .../impl/shape/tensorops/TensorArraySize.java |   6 -
 .../transforms/BaseDynamicTransformOp.java    |  62 ------
 .../ops/impl/transforms/custom/EqualTo.java   |  12 --
 .../impl/transforms/custom/GreaterThan.java   |  12 --
 .../transforms/custom/GreaterThanOrEqual.java |  12 --
 .../ops/impl/transforms/custom/LessThan.java  |  12 --
 .../transforms/custom/LessThanOrEqual.java    |  12 --
 .../impl/transforms/custom/NotEqualTo.java    |  12 --
 .../api/ops/impl/transforms/dtype/Cast.java   |  35 ----
 .../gradient/GradientBackwardsMarker.java     |   2 +-
 .../bp/BaseArithmeticBackpropOp.java          |   5 -
 .../ops/impl/transforms/same/Identity.java    |   7 -
 .../opvalidation/ShapeOpValidation.java       |  15 --
 .../nd4j/autodiff/samediff/SameDiffTests.java |  54 ++---
 .../org/nd4s/samediff/SameDiffTest.scala      |   2 +-
 36 files changed, 32 insertions(+), 741 deletions(-)
 delete mode 100644 nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseModule.java
 delete mode 100644 nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/Module.java
 delete mode 100644 nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/Linear.java
 delete mode 100644 nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Broadcast.java

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
index e16bd3dc2..1f361ce9c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
@@ -163,7 +163,6 @@ import org.nd4j.linalg.api.ops.impl.scatter.ScatterMin;
 import org.nd4j.linalg.api.ops.impl.scatter.ScatterMul;
 import org.nd4j.linalg.api.ops.impl.scatter.ScatterSub;
 import org.nd4j.linalg.api.ops.impl.scatter.ScatterUpdate;
-import org.nd4j.linalg.api.ops.impl.shape.Broadcast;
 import org.nd4j.linalg.api.ops.impl.shape.Concat;
 import org.nd4j.linalg.api.ops.impl.shape.ConfusionMatrix;
 import org.nd4j.linalg.api.ops.impl.shape.Cross;
@@ -1450,14 +1449,6 @@ public class DifferentialFunctionFactory {
         return new MatrixInverse(sameDiff(), in, false).outputVariable();
     }
 
-    public SDVariable broadcast(SDVariable iX, int... shape) {
-        return broadcast(iX, ArrayUtil.toLongArray(shape));
-    }
-
-    public SDVariable broadcast(SDVariable iX, long... shape) {
-        return new Broadcast(sameDiff(), iX, shape).outputVariable();
-    }
-
     public SDVariable onehot(SDVariable indices, int depth, int axis, double on, double off, DataType dataType) {
         return new OneHot(sameDiff(), indices, depth, axis, on, off, dataType).outputVariable();
     }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SDVariable.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SDVariable.java
index 430b4d83a..0d2700b43 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SDVariable.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SDVariable.java
@@ -111,9 +111,6 @@ public class SDVariable implements Serializable {
         return variableType == VariableType.CONSTANT;
     }
 
-
-
-
     /**
      * Allocate and return a  new array
      * based on the vertex id and weight initialization.
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/ImportClassMapping.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/ImportClassMapping.java
index 3f270e342..e2d3cf87b 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/ImportClassMapping.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/ImportClassMapping.java
@@ -98,7 +98,6 @@ public class ImportClassMapping {
             org.nd4j.linalg.api.ops.impl.indexaccum.custom.ArgMax.class,
             org.nd4j.linalg.api.ops.impl.indexaccum.custom.ArgMin.class,
             org.nd4j.linalg.api.ops.impl.layers.ExternalErrorsFunction.class,
-            org.nd4j.linalg.api.ops.impl.layers.Linear.class,
             org.nd4j.linalg.api.ops.impl.layers.convolution.AvgPooling2D.class,
             org.nd4j.linalg.api.ops.impl.layers.convolution.AvgPooling3D.class,
             org.nd4j.linalg.api.ops.impl.layers.convolution.BatchNorm.class,
@@ -267,7 +266,6 @@ public class ImportClassMapping {
             org.nd4j.linalg.api.ops.impl.scatter.ScatterSub.class,
             org.nd4j.linalg.api.ops.impl.scatter.ScatterUpdate.class,
             org.nd4j.linalg.api.ops.impl.shape.ApplyGradientDescent.class,
-            org.nd4j.linalg.api.ops.impl.shape.Broadcast.class,
             org.nd4j.linalg.api.ops.impl.shape.BroadcastDynamicShape.class,
             org.nd4j.linalg.api.ops.impl.shape.Concat.class,
             org.nd4j.linalg.api.ops.impl.shape.ConfusionMatrix.class,
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseModule.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseModule.java
deleted file mode 100644
index 0826ac5f9..000000000
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseModule.java
+++ /dev/null
@@ -1,63 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.nd4j.linalg.api.ops;
-
-import lombok.NoArgsConstructor;
-import org.nd4j.autodiff.samediff.SDVariable;
-import org.nd4j.autodiff.samediff.SameDiff;
-import org.nd4j.linalg.api.ndarray.INDArray;
-
-import java.util.ArrayList;
-import java.util.List;
-
-/**
- * Abstract base class for {@link Module}
- * that handles Dynamic ops and handles nesting.
- *
- * This is a logical unit for defining layers
- * very similar to pytorch's modules, or tensorflow's layers.
- *
- * @author Adam Gibson
- */
-@NoArgsConstructor
-public abstract class BaseModule extends DynamicCustomOp implements Module {
-    private List<Module> modules = new ArrayList<>();
-
-    public BaseModule(String opName, INDArray[] inputs, INDArray[] outputs, List<Double> tArguments, List<Integer> iArguments, List<Module> modules) {
-        super(opName, inputs, outputs, tArguments, iArguments);
-        this.modules = modules;
-    }
-
-    public BaseModule(String opName, SameDiff sameDiff, SDVariable[] args, boolean inPlace, List<Module> modules) {
-        super(opName, sameDiff, args, inPlace);
-        this.modules = modules;
-    }
-
-    @Override
-    public Module[] subModules() {
-        return modules.toArray(new Module[modules.size()]);
-    }
-
-    @Override
-    public void addModule(Module module) {
-        modules.add(module);
-    }
-
-
-
-
-}
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/Module.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/Module.java
deleted file mode 100644
index f41f94e5c..000000000
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/Module.java
+++ /dev/null
@@ -1,51 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.nd4j.linalg.api.ops;
-
-import org.nd4j.autodiff.samediff.SDVariable;
-import org.nd4j.linalg.api.ndarray.INDArray;
-
-/**
- * A Module is a {@link CustomOp}
- * with varying input arguments
- * and automatically calculated outputs
- * defined at a higher level than c++.
- *
- * A Module is meant to be a way of implementing custom operations
- * in straight java/nd4j.
- */
-public interface Module extends CustomOp {
-
-    /**
-     *
-     * @param inputs
-     */
-    void exec(INDArray... inputs);
-
-
-    Module[] subModules();
-
-
-    void addModule(Module module);
-
-
-    void execSameDiff(SDVariable... input);
-
-
-
-
-}
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/NoOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/NoOp.java
index 6b174bd07..554ad917e 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/NoOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/NoOp.java
@@ -80,6 +80,7 @@ public class NoOp extends DynamicCustomOp {
         return 1;
     }
 
+
     @Override
     public List<LongShapeDescriptor> calculateOutputShape(){
         if(inputArguments != null && !inputArguments.isEmpty()){
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/BaseCompatOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/BaseCompatOp.java
index 4c0fb2e6b..3f56096a2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/BaseCompatOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/BaseCompatOp.java
@@ -16,6 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.controlflow.compat;
 
+import java.util.List;
 import lombok.NonNull;
 import lombok.val;
 import org.nd4j.autodiff.samediff.SDVariable;
@@ -25,6 +26,7 @@ import org.nd4j.imports.descriptors.properties.AttributeAdapter;
 import org.nd4j.imports.descriptors.properties.PropertyMapping;
 import org.nd4j.imports.graphmapper.tf.TFGraphMapper;
 import org.nd4j.linalg.api.ops.DynamicCustomOp;
+import org.nd4j.linalg.api.shape.LongShapeDescriptor;
 import org.tensorflow.framework.AttrValue;
 import org.tensorflow.framework.GraphDef;
 import org.tensorflow.framework.NodeDef;
@@ -89,4 +91,9 @@ public abstract class BaseCompatOp extends DynamicCustomOp {
     public Map<String, Map<String, AttributeAdapter>> attributeAdaptersForFunction() {
         return super.attributeAdaptersForFunction();
     }
+
+    @Override
+    public List<LongShapeDescriptor> calculateOutputShape() {
+        throw new UnsupportedOperationException("calculateOutputShape() is not supported for control flow ops.");
+    }
 }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Enter.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Enter.java
index 85a94eb13..52705ce9e 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Enter.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Enter.java
@@ -68,15 +68,6 @@ public class Enter extends BaseCompatOp {
         return OP_NAME;
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if(arg().getArr() != null) {
-            return Collections.singletonList(LongShapeDescriptor.fromShape(arg().getShape(), arg().getArr().dataType()));
-        }
-        else
-            return Collections.emptyList();
-    }
-
     @Override
     public SDVariable[] outputVariables() {
         return super.outputVariables();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Exit.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Exit.java
index f9e358f3c..a7fecf03c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Exit.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Exit.java
@@ -56,15 +56,6 @@ public class Exit extends BaseCompatOp {
         return OP_NAME;
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if(arg().getArr() != null) {
-            return Collections.singletonList(arg().getShapeDescriptor());
-        }
-        else
-            return Collections.emptyList();
-    }
-
     @Override
     public SDVariable[] outputVariables() {
         return super.outputVariables();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/LoopCond.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/LoopCond.java
index a3ace4f13..4f5d11b38 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/LoopCond.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/LoopCond.java
@@ -37,15 +37,6 @@ public class LoopCond extends BaseCompatOp {
         return "loop_cond";
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if(arg().getArr() != null) {
-            return Collections.singletonList(LongShapeDescriptor.fromShape(arg().getShape(), arg().dataType()));
-        }
-        else
-            return Collections.emptyList();
-    }
-
     @Override
     public SDVariable[] outputVariables() {
         return super.outputVariables();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Merge.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Merge.java
index 386f4a075..6cf52e7b5 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Merge.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Merge.java
@@ -64,15 +64,6 @@ public class Merge extends BaseCompatOp {
         return 60L;
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if(arg().getArr() != null) {
-            return Collections.singletonList(arg().getShapeDescriptor());
-        }
-        else
-            return Collections.emptyList();
-    }
-
     @Override
     public SDVariable[] outputVariables() {
         return super.outputVariables();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/NextIteration.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/NextIteration.java
index fabd0479b..367a134a3 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/NextIteration.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/NextIteration.java
@@ -53,15 +53,6 @@ public class NextIteration extends BaseCompatOp {
         return OP_NAME;
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if(arg().getArr() != null) {
-            return Collections.singletonList(LongShapeDescriptor.fromShape(arg().getShape(), arg().dataType()));
-        }
-        else
-            return Collections.emptyList();
-    }
-
     @Override
     public SDVariable[] outputVariables() {
         return super.outputVariables();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Switch.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Switch.java
index 77145a625..331dea887 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Switch.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Switch.java
@@ -63,18 +63,6 @@ public class Switch extends BaseCompatOp {
         return OP_NAME;
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if(args()[0].getArr() != null) {
-            val arg0 = args()[0];
-            val arr0 = arg0.getArr();
-            val dtype = arr0.dataType();
-            return Arrays.asList(LongShapeDescriptor.fromShape(arg0.getShape(), dtype),LongShapeDescriptor.fromShape(arg0.getShape(), dtype));
-        }
-        else
-            return Collections.emptyList();
-    }
-
     @Override
     public SDVariable[] outputVariables() {
         return super.outputVariables();
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/Linear.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/Linear.java
deleted file mode 100644
index 27f357b4b..000000000
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/layers/Linear.java
+++ /dev/null
@@ -1,198 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.nd4j.linalg.api.ops.impl.layers;
-
-import lombok.Builder;
-import lombok.NoArgsConstructor;
-import lombok.val;
-import onnx.Onnx;
-import org.nd4j.autodiff.functions.DifferentialFunction;
-import org.nd4j.autodiff.samediff.SDVariable;
-import org.nd4j.autodiff.samediff.SameDiff;
-import org.nd4j.autodiff.samediff.VariableType;
-import org.nd4j.imports.NoOpNameFoundException;
-import org.nd4j.linalg.api.blas.params.MMulTranspose;
-import org.nd4j.linalg.api.ndarray.INDArray;
-import org.nd4j.linalg.api.ops.BaseModule;
-import org.nd4j.linalg.api.ops.Module;
-import org.nd4j.linalg.api.ops.impl.reduce.Mmul;
-import org.nd4j.linalg.api.shape.LongShapeDescriptor;
-import org.nd4j.linalg.api.shape.Shape;
-import org.nd4j.linalg.factory.Nd4j;
-import org.nd4j.weightinit.WeightInitScheme;
-import org.nd4j.weightinit.impl.ZeroInitScheme;
-import org.tensorflow.framework.AttrValue;
-import org.tensorflow.framework.GraphDef;
-import org.tensorflow.framework.NodeDef;
-
-import java.util.ArrayList;
-import java.util.List;
-import java.util.Map;
-
-/**
- * Linear:
- * a * bT
- *
- * @author Adam Gibson
- */
-@NoArgsConstructor
-public class Linear extends BaseModule {
-    private DifferentialFunction forward;
-    private int nIn,nOut;
-    private WeightInitScheme weightInitScheme,biasWeightInitScheme;
-
-    @Builder(builderMethodName = "execBuilder")
-    public Linear(int nIn,
-                  int nOut,
-                  WeightInitScheme weightInitScheme,
-                  WeightInitScheme biasWeightInitScheme) {
-        super(null,
-                getParams(nIn,nOut,weightInitScheme,biasWeightInitScheme),
-                new INDArray[]{},
-                new ArrayList<Double>(), new ArrayList<Integer>(), new ArrayList<Module>());
-        this.weightInitScheme = weightInitScheme;
-        this.biasWeightInitScheme = biasWeightInitScheme;
-        this.nIn = nIn;
-        this.nOut = nOut;
-    }
-
-    @Builder(builderMethodName = "sameDiffBuilder")
-    public Linear(SameDiff sameDiff,
-                  int nIn,
-                  int nOut,
-                  WeightInitScheme weightInitScheme,
-                  WeightInitScheme biasWeightInitScheme) {
-        super(null, sameDiff, null, false, new ArrayList<Module>());
-        this.weightInitScheme = weightInitScheme;
-        this.biasWeightInitScheme = biasWeightInitScheme;
-
-        this.nIn = nIn;
-        this.nOut = nOut;
-
-    }
-
-    @Override
-    public String opName() {
-        return "linear";
-    }
-
-    @Override
-    public void initFromTensorFlow(NodeDef nodeDef, SameDiff initWith, Map<String, AttrValue> attributesForNode, GraphDef graph) {
-
-    }
-
-    @Override
-    public void initFromOnnx(Onnx.NodeProto node, SameDiff initWith, Map<String, Onnx.AttributeProto> attributesForNode, Onnx.GraphProto graph) {
-
-    }
-
-    @Override
-    public List<SDVariable> doDiff(List<SDVariable> f1) {
-        execSameDiff();
-        return forward.doDiff(f1);
-    }
-
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        List<LongShapeDescriptor> ret = new ArrayList<>();
-        ret.add(LongShapeDescriptor.fromShape(Shape.getMatrixMultiplyShape(inputArguments()[0].shape(),new long[]{nOut,nIn}), inputArguments()[1].dataType()));
-
-        ret.add(LongShapeDescriptor.fromShape(Shape.getMatrixMultiplyShape(inputArguments()[0].shape(),inputArguments()[1].transpose().shape()), inputArguments()[1].dataType()));
-        if(biasWeightInitScheme != null) {
-            ret.add(LongShapeDescriptor.fromShape(new long[]{nOut,1}, inputArguments()[1].dataType()));
-        }
-        return ret;
-    }
-
-    @Override
-    public String onnxName() {
-        throw new NoOpNameFoundException("No onnx op opName found for " +  opName());
-    }
-
-    @Override
-    public String tensorflowName() {
-        throw new NoOpNameFoundException("No tensorflow op opName found for " +  opName());
-    }
-
-
-
-    @Override
-    public void exec(INDArray... inputs) {
-        val inputArguments = inputArguments();
-        if(inputArguments == null || inputArguments.length < 1) {
-            throw new IllegalStateException("No arguments found.");
-        }
-
-        INDArray weights = inputArguments[0];
-        INDArray right = inputArguments[1];
-
-        val outputArguments = outputArguments();
-
-        if(outputArguments == null || outputArguments.length < 1) {
-            if(inputArguments.length == 1)
-              addOutputArgument(inputs[0].mmul(weights.transpose()));
-            else
-                addOutputArgument(inputs[0].mmul(weights.transpose()).addiColumnVector(right));
-
-        }
-        else {
-            inputs[0].mmul(weights.transpose(),outputArguments[0]);
-        }
-
-    }
-
-    @Override
-    public void execSameDiff(SDVariable... input) {
-        val args = args();
-        if(args == null || args.length == 0) {
-            throw new IllegalStateException("No arguments found");
-        }
-
-        if(forward == null) {
-            //bias needs to be added yet
-            if(args.length > 1) {
-                /*
-                forward =  f().add(new Mmul(sameDiff, input[0],args()[0],
-                        MMulTranspose.builder()
-                                .transposeA(false)
-                                .transposeB(true)
-                                .build()).outputVariables()[0],args()[1]);
-             */
-            } else {
-                forward = new Mmul(sameDiff, input[0],args()[0],
-                        MMulTranspose.builder().transposeA(false).transposeB(true).build());
-            }
-
-            this.outputVariables = forward.outputVariables();
-        }
-
-
-    }
-
-    private static INDArray[] getParams(int nIn,
-                                        int nOut,
-                                        WeightInitScheme paramsScheme,
-                                        WeightInitScheme biasInitScheme) {
-        if(biasInitScheme != null) {
-            return new INDArray[] {paramsScheme.create(Nd4j.defaultFloatingPointType(), new long[]{nOut,nIn}),biasInitScheme.create(Nd4j.defaultFloatingPointType(), new long[]{nOut,1})};
-        }
-        else {
-            return new INDArray[] {paramsScheme.create(Nd4j.defaultFloatingPointType(), new long[]{nOut,nIn})};
-
-        }
-    }
-}
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/loss/L2Loss.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/loss/L2Loss.java
index cfef2a61b..21941de93 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/loss/L2Loss.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/loss/L2Loss.java
@@ -38,11 +38,6 @@ public class L2Loss extends DynamicCustomOp {
         super(sameDiff, new SDVariable[]{var});
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        return Collections.singletonList(LongShapeDescriptor.fromShape(new long[0], Nd4j.defaultFloatingPointType()));
-    }
-
     @Override
     public String opName() {
         return "l2_loss";
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/HashCode.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/HashCode.java
index 4ad8fc5d9..cdddee2f6 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/HashCode.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/HashCode.java
@@ -47,11 +47,6 @@ public class HashCode extends DynamicCustomOp {
         this.outputArguments.add(result);
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        return Collections.singletonList(LongShapeDescriptor.fromShape(new long[0], DataType.LONG));
-    }
-
     @Override
     public String opName() {
         return "hashcode";
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Broadcast.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Broadcast.java
deleted file mode 100644
index 8e2f6c790..000000000
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Broadcast.java
+++ /dev/null
@@ -1,78 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.nd4j.linalg.api.ops.impl.shape;
-
-import org.nd4j.autodiff.samediff.SDVariable;
-import org.nd4j.autodiff.samediff.SameDiff;
-import org.nd4j.base.Preconditions;
-import org.nd4j.imports.NoOpNameFoundException;
-import org.nd4j.linalg.api.buffer.DataType;
-import org.nd4j.linalg.api.ops.DynamicCustomOp;
-import org.nd4j.linalg.api.shape.LongShapeDescriptor;
-
-import java.util.Arrays;
-import java.util.Collections;
-import java.util.List;
-
-/**
- * Broadcast function
- *
- * @author Adam Gibson
- */
-public class Broadcast extends DynamicCustomOp {
-    private long[] shape;
-    public Broadcast(SameDiff sameDiff,SDVariable iX, long[] shape) {
-        super(null,sameDiff,new SDVariable[]{iX});
-        this.shape = shape;
-    }
-
-
-    public Broadcast() {}
-
-
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        return Arrays.asList(LongShapeDescriptor.fromShape(shape, larg().dataType()));
-    }
-
-    @Override
-    public String opName() {
-        return "broadcast";
-    }
-
-
-
-    @Override
-    public List<SDVariable> doDiff(List<SDVariable> i_v) {
-        throw new UnsupportedOperationException();
-    }
-
-    @Override
-    public String onnxName() {
-        throw new NoOpNameFoundException("No onnx op opName found for " +  opName());
-    }
-
-    @Override
-    public String tensorflowName() {
-        throw new NoOpNameFoundException("No tensorflow op opName found for " +  opName());
-    }
-
-    @Override
-    public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes){
-        return Collections.singletonList(dataTypes.get(0));
-    }
-}
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Linspace.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Linspace.java
index a11b241e5..1ad13ad40 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Linspace.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Linspace.java
@@ -82,12 +82,4 @@ public class Linspace extends DynamicCustomOp {
     public List<SDVariable> doDiff(List<SDVariable> gradients){
         return Arrays.asList(f().zerosLike(arg(0)), f().zerosLike(arg(1)), f().zerosLike(arg(2)));
     }
-
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape(){
-        INDArray l = arg(2).getArr();
-        if(l == null)
-            return Collections.emptyList();
-        return Collections.singletonList(LongShapeDescriptor.fromShape(new long[]{l.getLong(0)}, dataType));
-    }
 }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Rank.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Rank.java
index aacfa19e1..568b14a44 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Rank.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Rank.java
@@ -85,14 +85,6 @@ public class Rank extends DynamicCustomOp {
         return "Rank";
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        List<LongShapeDescriptor> ret = new ArrayList<>();
-        ret.add(LongShapeDescriptor.fromShape(new long[]{}, DataType.INT));
-        return ret;
-    }
-
-
     @Override
     public List<SDVariable> doDiff(List<SDVariable> i_v) {
         return Collections.singletonList(sameDiff.zerosLike(arg()));
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Unstack.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Unstack.java
index 9dd6b6338..f9eb1f95a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Unstack.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Unstack.java
@@ -152,17 +152,4 @@ public class Unstack extends DynamicCustomOp {
         return out;
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape(){
-        //TEMPORARY workaround for: https://github.com/deeplearning4j/deeplearning4j/issues/7093
-        if(inputArguments.size() == 1 && inputArguments.get(0).rank() == 1){
-            INDArray arr = inputArguments.get(0);
-            Preconditions.checkState(jaxis == 0, "Can only unstack along dimension 0 for rank 1 arrays, got axis %s for array %ndShape", jaxis, arr);
-            LongShapeDescriptor lsd = LongShapeDescriptor.fromShape(new long[0], arr.dataType());
-            List<LongShapeDescriptor> out = Arrays.asList(ArrayUtil.nTimes((int)arr.length(), lsd, LongShapeDescriptor.class));
-            return out;
-        }
-        return super.calculateOutputShape();
-    }
-
 }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/BaseTensorOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/BaseTensorOp.java
index bcc4e1844..b027750fc 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/BaseTensorOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/BaseTensorOp.java
@@ -32,7 +32,7 @@ import java.util.Collections;
 import java.util.List;
 import java.util.Map;
 
-public abstract  class BaseTensorOp extends DynamicCustomOp {
+public abstract class BaseTensorOp extends DynamicCustomOp {
 
     public BaseTensorOp(String name, SameDiff sameDiff, SDVariable[] args){
         super(name, sameDiff, args);
@@ -78,8 +78,7 @@ public abstract  class BaseTensorOp extends DynamicCustomOp {
 
     @Override
     public List<LongShapeDescriptor> calculateOutputShape() {
-        //Not used/not required
-        return Collections.emptyList();
+        throw new UnsupportedOperationException("calculateOutputShape() is not supported for tensor ops.");
     }
 
     @Override
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySize.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySize.java
index 276dadcab..0503d377b 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySize.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/tensorops/TensorArraySize.java
@@ -41,12 +41,6 @@ public class TensorArraySize extends BaseTensorOp {
       return "tensorarraysizev3";
    }
 
-   @Override
-   public List<LongShapeDescriptor> calculateOutputShape() {
-      // output is scalar only
-      return Collections.singletonList(LongShapeDescriptor.fromShape(new long[]{}, DataType.LONG));
-   }
-
    @Override
    public void initFromTensorFlow(NodeDef nodeDef, SameDiff initWith, Map<String, AttrValue> attributesForNode, GraphDef graph) {
       super.initFromTensorFlow(nodeDef, initWith, attributesForNode, graph);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/BaseDynamicTransformOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/BaseDynamicTransformOp.java
index aa85461e6..ae4a21df7 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/BaseDynamicTransformOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/BaseDynamicTransformOp.java
@@ -46,68 +46,6 @@ public abstract class BaseDynamicTransformOp extends DynamicCustomOp {
         super(null, inputs, outputs);
     }
 
-
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        long[] firstArgShape;
-        long[] secondArgShape;
-        DataType dtypeZ;
-
-        if(numInputArguments() == 2){
-            return super.calculateOutputShape();    //Use c++ shape calc, which also accounts for empty broadcast cases, etc
-//            firstArgShape = inputArguments.get(0).shape();
-//            secondArgShape = inputArguments.get(1).shape();
-//            dtypeZ = Shape.pickPairwiseDataType(inputArguments.get(0).dataType(), inputArguments.get(1).dataType());
-        } else {
-            val args = args();
-            if (args.length < 2) {
-                if (args[0] == null || (inputArguments.isEmpty() && args[0].getShape() == null)) {
-                    return Collections.emptyList();
-                }
-                DataType dtypeX = !inputArguments.isEmpty() ? inputArguments.get(0).dataType() : args[0].dataType();
-                long[] shape = !inputArguments.isEmpty() ? inputArguments.get(0).shape() : args[0].getShape();
-
-                return Collections.singletonList(LongShapeDescriptor.fromShape(shape, dtypeX));
-            }
-
-            if(inputArguments.size() == 2 && inputArguments.get(0) != null && inputArguments.get(1) != null){
-                firstArgShape = inputArguments.get(0).shape();
-                secondArgShape = inputArguments.get(1).shape();
-            } else {
-                firstArgShape = args[0].getShape();
-                secondArgShape = args[1].getShape();
-            }
-            if (args[0] == null || args[0].getShape() == null) {
-                return Collections.emptyList();
-            }
-
-            if (args[1] == null || args[1].getShape() == null) {
-                return Collections.emptyList();
-            }
-
-            // detecting datatype based on both args
-            val dtypeX = inputArguments.size() > 0 ? inputArguments.get(0).dataType() : args[0].dataType();
-            val dtypeY = inputArguments.size() > 1 ? inputArguments.get(1).dataType() : args[1].dataType();
-            dtypeZ = Shape.pickPairwiseDataType(dtypeX, dtypeY);
-        }
-
-
-
-        if(Arrays.equals(firstArgShape, secondArgShape)){
-            try {
-                return Collections.singletonList(LongShapeDescriptor.fromShape(firstArgShape, dtypeZ));
-            } catch (Throwable e) {
-                throw new RuntimeException("calculateOutputShape() failed for [" + this.opName() + "]", e);
-            }
-        } else {
-            //Handle broadcast shape: [1,4]+[3,1] = [3,4]
-            Shape.assertBroadcastable(firstArgShape, secondArgShape, this.getClass());
-            val outShape = Shape.broadcastOutputShape(firstArgShape, secondArgShape);
-
-            return Collections.singletonList(LongShapeDescriptor.fromShape(outShape, dtypeZ));
-        }
-    }
-
     @Override
     public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes){
         Preconditions.checkState(dataTypes != null && dataTypes.size() == 2, "Expected exactly 2 input datatypes for %s, got input %s", getClass(), dataTypes);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/EqualTo.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/EqualTo.java
index b786602d3..95019c5b2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/EqualTo.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/EqualTo.java
@@ -71,18 +71,6 @@ public class EqualTo extends BaseDynamicTransformOp {
         return Arrays.asList(sameDiff.zerosLike(args()[0]), sameDiff.zerosLike(args()[1]));
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if (args() == null)
-            return Collections.emptyList();
-
-        if (inputArguments.size() == 0)
-            return Collections.emptyList();
-
-
-        return Nd4j.getExecutioner().calculateOutputShape(this);
-    }
-
     @Override
     public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes){
         Preconditions.checkState(dataTypes != null && dataTypes.size() == 2, "Expected exactly 2 input datatypes for %s, got %s", getClass(), dataTypes);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/GreaterThan.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/GreaterThan.java
index 27b2ea189..72a46f111 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/GreaterThan.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/GreaterThan.java
@@ -73,18 +73,6 @@ public class GreaterThan extends BaseDynamicTransformOp {
         return Arrays.asList(sameDiff.zerosLike(args()[0]), sameDiff.zerosLike(args()[1]));
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if (args() == null)
-            return Collections.emptyList();
-
-        if (inputArguments.size() == 0)
-            return Collections.emptyList();
-
-
-        return Nd4j.getExecutioner().calculateOutputShape(this);
-    }
-
     @Override
     public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes){
         Preconditions.checkState(dataTypes != null && dataTypes.size() == 2, "Expected exactly 2 input datatypes for %s, got %s", getClass(), dataTypes);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/GreaterThanOrEqual.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/GreaterThanOrEqual.java
index 48d3953aa..50d1c7c43 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/GreaterThanOrEqual.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/GreaterThanOrEqual.java
@@ -76,18 +76,6 @@ public class GreaterThanOrEqual extends BaseDynamicTransformOp {
         return Arrays.asList(sameDiff.zerosLike(args()[0]), sameDiff.zerosLike(args()[1]));
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if (args() == null)
-            return Collections.emptyList();
-
-        if (inputArguments.size() == 0)
-            return Collections.emptyList();
-
-
-        return Nd4j.getExecutioner().calculateOutputShape(this);
-    }
-
     @Override
     public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes){
         Preconditions.checkState(dataTypes != null && dataTypes.size() == 2, "Expected exactly 2 input datatypes for %s, got %s", getClass(), dataTypes);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LessThan.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LessThan.java
index 0ee59458c..4c345070e 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LessThan.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LessThan.java
@@ -73,18 +73,6 @@ public class LessThan extends BaseDynamicTransformOp {
         return Arrays.asList(sameDiff.zerosLike(args()[0]), sameDiff.zerosLike(args()[1]));
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if (args() == null)
-            return Collections.emptyList();
-
-        if (inputArguments.size() == 0)
-            return Collections.emptyList();
-
-
-        return Nd4j.getExecutioner().calculateOutputShape(this);
-    }
-
     @Override
     public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes){
         Preconditions.checkState(dataTypes != null && dataTypes.size() == 2, "Expected exactly 2 input datatypes for %s, got %s", getClass(), dataTypes);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LessThanOrEqual.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LessThanOrEqual.java
index 56a5882db..89c08fe65 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LessThanOrEqual.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/LessThanOrEqual.java
@@ -71,18 +71,6 @@ public class LessThanOrEqual extends BaseDynamicTransformOp {
         return Arrays.asList(sameDiff.zerosLike(args()[0]), sameDiff.zerosLike(args()[1]));
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if (args() == null)
-            return Collections.emptyList();
-
-        if (inputArguments.size() == 0)
-            return Collections.emptyList();
-
-
-        return Nd4j.getExecutioner().calculateOutputShape(this);
-    }
-
     @Override
     public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes){
         Preconditions.checkState(dataTypes != null && dataTypes.size() == 2, "Expected exactly 2 input datatypes for %s, got %s", getClass(), dataTypes);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/NotEqualTo.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/NotEqualTo.java
index b1f0dadbb..62d3bedfa 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/NotEqualTo.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/NotEqualTo.java
@@ -72,18 +72,6 @@ public class NotEqualTo extends BaseDynamicTransformOp {
         return Arrays.asList(sameDiff.zerosLike(args()[0]), sameDiff.zerosLike(args()[1]));
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if (args() == null)
-            return Collections.emptyList();
-
-        if (inputArguments.size() == 0)
-            return Collections.emptyList();
-
-
-        return Nd4j.getExecutioner().calculateOutputShape(this);
-    }
-
     @Override
     public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes){
         Preconditions.checkState(dataTypes != null && dataTypes.size() == 2, "Expected exactly 2 input datatypes for %s, got %s", getClass(), dataTypes);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/dtype/Cast.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/dtype/Cast.java
index 01408ea4d..98a479542 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/dtype/Cast.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/dtype/Cast.java
@@ -125,41 +125,6 @@ public class Cast extends BaseDynamicTransformOp {
         return ret;
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if(inputArguments.size() > 0){
-            long[] s = inputArguments.get(0).shape();
-            LongShapeDescriptor lsd = LongShapeDescriptor.fromShape(s, typeDst);
-            if(inputArguments.get(0).isEmpty()){
-                long e = lsd.getExtras();
-                e = ArrayOptionsHelper.setOptionBit(e, ArrayType.EMPTY);
-                lsd.setExtras(e);
-            }
-            return Collections.singletonList(lsd);
-        }
-
-        if (arg() != null && (arg().getArr() != null || arg().getShape() != null)) {
-            if (arg().getArr() != null) {
-                long[] s = arg().getArr().shape();
-                LongShapeDescriptor lsd = LongShapeDescriptor.fromShape(s, typeDst);
-                if(inputArguments.size() > 0 && inputArguments.get(0) != null && inputArguments.get(0).isEmpty()){
-                    long e = lsd.getExtras();
-                    e = ArrayOptionsHelper.setOptionBit(e, ArrayType.EMPTY);
-                    lsd.setExtras(e);
-                }
-                return Collections.singletonList(lsd);
-            } else {
-                long[] s = arg().getShape();
-                if(Shape.isPlaceholderShape(s)){
-                    return Collections.emptyList();
-                }
-                return Collections.singletonList(LongShapeDescriptor.fromShape(s, typeDst));
-            }
-        }
-
-        return Collections.emptyList();
-    }
-
     @Override
     public void setValueFor(Field target, Object value) {
         //This is a hack around a property mapping issue - TF datatype DT_DOUBLE return attribute.getType() of DT_DOUBLE which doesn't make sense
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/gradient/GradientBackwardsMarker.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/gradient/GradientBackwardsMarker.java
index d61f7a067..10abb69a2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/gradient/GradientBackwardsMarker.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/gradient/GradientBackwardsMarker.java
@@ -77,7 +77,7 @@ public class GradientBackwardsMarker extends DynamicCustomOp {
 
     @Override
     public List<LongShapeDescriptor> calculateOutputShape() {
-        return Collections.singletonList(LongShapeDescriptor.fromShape(new long[0], DataType.FLOAT));
+        throw new UnsupportedOperationException("calculateOutputShape() is not supported for control flow ops.");
     }
 
     @Override
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/pairwise/arithmetic/bp/BaseArithmeticBackpropOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/pairwise/arithmetic/bp/BaseArithmeticBackpropOp.java
index dcbcc271f..44648cc4a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/pairwise/arithmetic/bp/BaseArithmeticBackpropOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/pairwise/arithmetic/bp/BaseArithmeticBackpropOp.java
@@ -46,11 +46,6 @@ public abstract class BaseArithmeticBackpropOp extends BaseDynamicTransformOp {
         throw new UnsupportedOperationException("Not supported");
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape(){
-        return Nd4j.getExecutioner().calculateOutputShape(this);
-    }
-
     @Override
     public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes){
         Preconditions.checkState(dataTypes != null && dataTypes.size() == 3, "Expected exactly 3 input datatypes for %s, got input %s", getClass(), dataTypes);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/same/Identity.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/same/Identity.java
index 3e548226f..4b5825d8f 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/same/Identity.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/same/Identity.java
@@ -79,11 +79,4 @@ public class Identity extends BaseDynamicTransformOp {
         return dataTypes;
     }
 
-    @Override
-    public List<LongShapeDescriptor> calculateOutputShape() {
-        if(inputArguments == null || inputArguments.isEmpty())
-            return Collections.emptyList();
-        return Collections.singletonList(inputArguments.get(0).shapeDescriptor());
-    }
-
 }
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java
index ffb585183..bf4b331a3 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/ShapeOpValidation.java
@@ -1619,21 +1619,6 @@ public class ShapeOpValidation extends BaseOpValidation {
         assertEquals(expected, result.eval());
     }
 
-    @Test
-    public void testBroadcast() {
-        OpValidationSuite.ignoreFailing();
-        SameDiff sd = SameDiff.create();
-        SDVariable in = sd.var("in", Nd4j.rand(3, 4));
-        SDVariable broadcast = sd.f().broadcast(in, 3, 4, 5);
-
-        INDArray out = sd.execAndEndResult();
-        assertArrayEquals(new long[]{3, 4, 5}, out.shape());
-
-        for (int i = 0; i < 5; i++) {
-            assertEquals(in.getArr(), out.get(all(), all(), point(i)));
-        }
-    }
-
 
     @Test
     public void testSlice2d() {
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/samediff/SameDiffTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/samediff/SameDiffTests.java
index ef6d1268b..bb15b3392 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/samediff/SameDiffTests.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/samediff/SameDiffTests.java
@@ -53,7 +53,6 @@ import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.api.ops.DynamicCustomOp;
 import org.nd4j.linalg.api.ops.impl.layers.ExternalErrorsFunction;
-import org.nd4j.linalg.api.ops.impl.layers.Linear;
 import org.nd4j.linalg.api.ops.impl.layers.convolution.config.Conv2DConfig;
 import org.nd4j.linalg.api.ops.impl.layers.convolution.config.LocalResponseNormalizationConfig;
 import org.nd4j.linalg.api.ops.impl.reduce3.ManhattanDistance;
@@ -522,39 +521,6 @@ public class SameDiffTests extends BaseNd4jTest {
     }
 
 
-    @Test
-    public void testLinearModule() {
-        int nIn = 5;
-        Linear linear = Linear.execBuilder()
-                .nIn(nIn)
-                .nOut(4)
-                .weightInitScheme(new UniformInitScheme('f', nIn))
-                .biasWeightInitScheme(new ZeroInitScheme('f'))
-                .build();
-        linear.exec(Nd4j.linspace(1, 20, 20).reshape(4, 5));
-        assertEquals(1, linear.numOutputArguments());
-
-    }
-
-
-    @Test
-    public void testLinearModule2() {
-        Linear linear = Linear.execBuilder()
-                .nIn(3)
-                .nOut(2)
-                .weightInitScheme(new OneInitScheme('f'))
-                .biasWeightInitScheme(new ZeroInitScheme('f'))
-                .build();
-        linear.exec(Nd4j.linspace(1, 6, 6).reshape(2, 3));
-        INDArray assertion = Nd4j.create(new double[][]{
-                {6, 6},
-                {15, 15}
-        });
-        assertEquals(assertion, linear.outputArguments()[0]);
-
-    }
-
-
     @Test
     public void testDefineFunctionArrayExistence() {
         SameDiff sameDiff = SameDiff.create();
@@ -3577,4 +3543,24 @@ public class SameDiffTests extends BaseNd4jTest {
 
         assertEquals(e, mod.eval());
     }
+
+    @Test
+    public void castShapeTest1(){
+        SameDiff sd = SameDiff.create();
+        SDVariable x = sd.constant(Nd4j.createFromArray(1, 2, 3, 4));
+        SDVariable casted = x.castTo(DataType.FLOAT);
+
+        assertEquals(casted.dataType(), DataType.FLOAT);
+    }
+
+    @Test
+    @Ignore // casted shape is null
+    public void castShapeTestEmpty(){
+        SameDiff sd = SameDiff.create();
+        SDVariable x = sd.constant(Nd4j.empty(DataType.INT));
+        SDVariable casted = x.castTo(DataType.FLOAT);
+
+        assertEquals(casted.dataType(), DataType.FLOAT);
+        assertTrue(casted.getShapeDescriptor().isEmpty());
+    }
 }
diff --git a/nd4s/src/test/scala/org/nd4s/samediff/SameDiffTest.scala b/nd4s/src/test/scala/org/nd4s/samediff/SameDiffTest.scala
index c9ee41bfb..2f5746b95 100644
--- a/nd4s/src/test/scala/org/nd4s/samediff/SameDiffTest.scala
+++ b/nd4s/src/test/scala/org/nd4s/samediff/SameDiffTest.scala
@@ -15,7 +15,7 @@ import org.nd4j.linalg.api.blas.params.MMulTranspose
 import org.nd4j.linalg.api.buffer.DataType
 import org.nd4j.linalg.api.ndarray.INDArray
 import org.nd4j.linalg.api.ops.DynamicCustomOp
-import org.nd4j.linalg.api.ops.impl.layers.{ ExternalErrorsFunction, Linear }
+import org.nd4j.linalg.api.ops.impl.layers.ExternalErrorsFunction
 import org.nd4j.linalg.api.ops.impl.layers.convolution.config.{ Conv2DConfig, LocalResponseNormalizationConfig }
 import org.nd4j.linalg.api.ops.impl.reduce3.ManhattanDistance
 import org.nd4j.linalg.api.ops.impl.shape.tensorops.TensorArray

From 3157ec110cb6589ab3f8785f6bc2dce2d0c0c5fb Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Wed, 28 Aug 2019 11:14:22 +0300
Subject: [PATCH 33/56] [WIP] reverse_sequence (#188)

* initial commit

Signed-off-by: raver119 <raver119@gmail.com>

* one more print

Signed-off-by: raver119 <raver119@gmail.com>

* minor fix

Signed-off-by: raver119 <raver119@gmail.com>

* reverse_sequence fix

Signed-off-by: raver119 <raver119@gmail.com>

* confusion_matrix test updated

Signed-off-by: raver119 <raver119@gmail.com>

* minor tweak

Signed-off-by: raver119 <raver119@gmail.com>

* minor tweak

Signed-off-by: raver119 <raver119@gmail.com>

* one more reverse_sequence test

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../ops/declarable/helpers/cuda/reverse.cu    | 17 +++++++-----
 .../layers_tests/DeclarableOpsTests5.cpp      | 27 ++++++++++++++++---
 2 files changed, 34 insertions(+), 10 deletions(-)

diff --git a/libnd4j/include/ops/declarable/helpers/cuda/reverse.cu b/libnd4j/include/ops/declarable/helpers/cuda/reverse.cu
index cf891feab..e86cd382a 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/reverse.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/reverse.cu
@@ -51,8 +51,8 @@ namespace helpers {
         }
         __syncthreads();
 
-        auto odd = length % 2 != 0;
-        auto limit = length / 2;
+        auto odd = numOfElemsToReverse % 2 != 0;
+        auto limit = numOfElemsToReverse / 2;
 
         for (Nd4jLong e = tid; e < limit; e += step) {
             // we're calculating offsets within input array
@@ -102,11 +102,8 @@ namespace helpers {
         seqLengths->syncToHost();
         auto stream = context->getCudaStream();
 
-        NDArray::prepareSpecialUse({output}, {input, seqLengths});
         if(input->isVector() || shape::isLikeVector(input->getShapeInfo(), posOfNonUnityDim) || seqLengths->lengthOf() == 1) {
             int numOfElemsToReverse = seqLengths->e<int>(0);
-//            printf("Length %d\n", numOfElemsToReverse);
-//            input->printBuffer("INPUT");
             if((seqDim == 0 && input->sizeAt(0) == 1) || (batchDim == posOfNonUnityDim))
                 output->assign(input);
             else
@@ -122,7 +119,6 @@ namespace helpers {
             auto inSubArrsSet  = input->allTensorsAlongDimension(dimensions);
             auto outSubArrsSet = output->allTensorsAlongDimension(dimensions);
 
-// #pragma omp parallel for schedule(guided)  if(inSubArrsSet->size() > Environment::getInstance()->elementwiseThreshold())
             for(int i = 0; i < inSubArrsSet->size(); ++i) {
 
                 int numOfElemsToReverse = seqLengths->e<int>(i);
@@ -143,11 +139,18 @@ namespace helpers {
             delete inSubArrsSet;
             delete outSubArrsSet;
         }
-        NDArray::registerSpecialUse({output}, {input, seqLengths});
+
     }
 
     void reverseSequence(nd4j::LaunchContext * context, const NDArray* input, const NDArray* seqLengths, NDArray* output, int seqDim, const int batchDim) {
+        NDArray::prepareSpecialUse({output}, {input, seqLengths});
+
+        // if op isn't inplace - copy original data into output array
+        if (output->getSpecialBuffer() != input->getSpecialBuffer())
+            output->assign(input);
+
         BUILD_SINGLE_SELECTOR(input->dataType(), reverseSequence_, (context, input, seqLengths, output, seqDim, batchDim), LIBND4J_TYPES);
+        NDArray::registerSpecialUse({output}, {input, seqLengths});
     }
 
     //////////////////////////////////////////////////////////////////////////
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
index 1fbe81046..b256eef10 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
@@ -842,9 +842,13 @@ TEST_F(DeclarableOpsTests5, reverse_sequense_test1) {
 
     nd4j::ops::reverse_sequence op;
     auto results = op.execute({&input, &seqLengths}, {}, {2, 1});
+    ASSERT_EQ(Status::OK(), results->status());
+
     auto output = results->at(0);
 
-    ASSERT_EQ(Status::OK(), results->status());
+    exp.printIndexedBuffer("E");
+    output->printIndexedBuffer("O");
+
     ASSERT_TRUE(exp.isSameShape(output));
     ASSERT_TRUE(exp.equalsTo(output));
 
@@ -1086,6 +1090,23 @@ TEST_F(DeclarableOpsTests5, reverse_sequense_test13) {
     delete results;
 }
 
+//////////////////////////////////////////////////////////////////////
+TEST_F(DeclarableOpsTests5, reverse_sequense_test14) {
+    auto input = NDArrayFactory::create<double>('c', {8, 8, 3, 2}, {0.09753360,             0.76124972,             0.24693797,             0.13813169,             0.33144656,             0.08299957,             0.67197708,             0.80659380,             0.98274191,             0.63566073,             0.21592326,             0.54902743,             0.54555996,             0.23407607,             0.11372584,             0.49965927,             0.15210842,             0.53268608,             0.38700677,             0.68832738,             0.37292716,             0.94616004,             0.77735792,             0.60803430,             0.61523204,             0.64298760,             0.26848351,             0.75015615,             0.28683049,             0.70937606,             0.06478678,             0.68985848,             0.55216783,             0.55382648,             0.34652863,             0.17261296,             0.54193264,             0.05176904,             0.82555761,             0.71106697,             0.04416722,             0.07653656,             0.01034390,             0.99430482,             0.59944390,             0.17973880,             0.36437840,             0.86383673,             0.45025550,             0.97136977,             0.13565978,             0.71567448,             0.92094825,             0.93536442,             0.93630291,             0.67277404,             0.93899264,             0.52422773,             0.44892176,             0.03127759,             0.85910449,             0.18252879,             0.72830945,             0.96736828,             0.89831575,             0.83437150,             0.59050780,             0.36145925,             0.16483070,             0.44021176,             0.76018652,             0.44227383,             0.13052339,             0.18204235,             0.99743733,             0.26885190,             0.87726522,             0.16396056,             0.94943412,             0.40016700,             0.65267938,             0.71073267,             0.40094733,             0.91182634,             0.05391789,             0.49520416,             0.24963864,             0.34847086,             0.74088617,             0.36115701,             0.63074210,             0.97423085,             0.42216846,             0.06326975,             0.07858702,             0.20586622,             0.28752144,             0.38146961,             0.83518735,             0.08207577,             0.82083487,             0.81665728,             0.33309570,             0.67563176,             0.98343578,             0.95919930,             0.66994391,             0.89296165,             0.34755773,             0.63166554,             0.18849320,             0.34828456,             0.98477707,             0.75163124,             0.83306004,             0.14203056,             0.01497920,             0.85727447,             0.71194544,             0.85654019,             0.86160433,             0.79580411,             0.47710411,             0.09318029,             0.31369071,             0.64122249,             0.58399725,             0.26706597,             0.05655339,             0.91025211,             0.30330468,             0.33142930,             0.05668627,             0.02936449,             0.12613087,             0.09960114,             0.16218074,             0.15088139,             0.31239040,             0.55980062,             0.34804391,             0.34941538,             0.61370555,             0.07022964,             0.59757058,             0.31189846,             0.25215345,             0.52546591,             0.55744218,             0.59485650,             0.60553664,             0.07536713,             0.55971796,             0.38764845,             0.20737843,             0.37989120,             0.18361641,             0.48636240,             0.06052657,             0.04241913,             0.66710351,             0.07007925,             0.59371493,             0.74479056,             0.84699625,             0.51210368,             0.12489571,             0.23371067,             0.27274571,             0.83306066,             0.75830824,             0.25963478,             0.87137718,             0.24418835,             0.05032742,             0.52076188,             0.47762345,             0.89829370,             0.34417708,             0.84705151,             0.08203183,             0.10632956,             0.78431292,             0.86441722,             0.36487598,             0.09833603,             0.85863594,             0.11010505,             0.11659283,             0.42500288,             0.02747301,             0.12359903,             0.01753431,             0.41160932,             0.47245979,             0.08268172,             0.21580773,             0.75770279,             0.19736489,             0.44461885,             0.33341706,             0.22519571,             0.31528710,             0.14802902,             0.64171939,             0.52643769,             0.19261234,             0.98032835,             0.15401656,             0.85274458,             0.66408502,             0.23212704,             0.74630026,             0.05713613,             0.49025892,             0.48418810,             0.59541513,             0.09243053,             0.93919152,             0.95357019,             0.52377729,             0.65963871,             0.47934951,             0.49919534,             0.34369898,             0.78211256,             0.13908708,             0.95754117,             0.84107746,             0.09126213,             0.42979124,             0.10295325,             0.34631257,             0.69448345,             0.41720536,             0.15282440,             0.74329854,             0.45775009,             0.12786280,             0.39830299,             0.20386769,             0.59703523,             0.94077086,             0.42255597,             0.80453309,             0.79757204,             0.28653229,             0.60175909,             0.55859623,             0.34318230,             0.63002770,             0.36533324,             0.89689906,             0.73236186,             0.61491989,             0.83787947,             0.67939463,             0.72016694,             0.77499849,             0.72428343,             0.34571059,             0.23143007,             0.20099338,             0.85583142,             0.73174191,             0.54284092,             0.20264181,             0.53037061,             0.30493131,             0.82279766,             0.58542432,             0.72632070,             0.18394258,             0.00608118,             0.23808232,             0.17007573,             0.75245459,             0.84990616,             0.38827634,             0.33809538,             0.01080317,             0.27250145,             0.81769542,             0.15323253,             0.71668395,             0.99427044,             0.11355576,             0.50511923,             0.60248266,             0.36610154,             0.99123140,             0.10519719,             0.18754650,             0.43232584,             0.25247084,             0.47968157,             0.88649124,             0.33588961,             0.92338319,             0.18808573,             0.79433656,             0.12074559,             0.02325163,             0.10117917,             0.83559239,             0.67213900,             0.67265260,             0.11917707,             0.76574855,             0.43842117,             0.28530411,             0.79648090,             0.47939640,             0.73564612,             0.41465671,             0.10995635,             0.20271728,             0.00521771,             0.22952055,             0.78271870,             0.12833592,             0.88639055,             0.76398188,             0.49533508,             0.85447872,             0.15937568,             0.92947480,             0.62705964,             0.85960084,             0.13435660,             0.81845809,             0.60715133,             0.83030708,             0.83071910,             0.38883408,             0.92033237,             0.46066239,             0.48806761,             0.50688779,             0.00654483,             0.32076493,             0.42367646,             0.07381865,             0.22801110,             0.26669388,             0.99691302,             0.12113623,             0.34373057,             0.98977921,             0.96225332,             0.90143562,             0.19559914,             0.08978307,             0.09687492,             0.59820890,             0.75527947,             0.67683355,             0.21847023,             0.29395619,             0.50477953,             0.07112842,             0.54090558,             0.68230725,             0.49713828,             0.41958965,             0.68013847,             0.47691765,             0.63269259,             0.94304095,             0.54587271,             0.72447569,             0.28913523,             0.75766936,             0.52965692,             0.96854824,             0.15589071,             0.84128672,             0.16337522,             0.05771034,             0.21556356,             0.12094140,             0.29721207,             0.00811008,             0.66184926});
+    auto lengths = NDArrayFactory::create<Nd4jLong>('c', {8}, {7, 2, 3, 5, 2, 1, 6, 4});
+    auto e = NDArrayFactory::create<double>('c', {8, 8, 3, 2}, {0.54193264,             0.05176904,             0.82555761,             0.71106697,             0.04416722,             0.07653656,             0.06478678,             0.68985848,             0.55216783,             0.55382648,             0.34652863,             0.17261296,             0.61523204,             0.64298760,             0.26848351,             0.75015615,             0.28683049,             0.70937606,             0.38700677,             0.68832738,             0.37292716,             0.94616004,             0.77735792,             0.60803430,             0.54555996,             0.23407607,             0.11372584,             0.49965927,             0.15210842,             0.53268608,             0.67197708,             0.80659380,             0.98274191,             0.63566073,             0.21592326,             0.54902743,             0.09753360,             0.76124972,             0.24693797,             0.13813169,             0.33144656,             0.08299957,             0.01034390,             0.99430482,             0.59944390,             0.17973880,             0.36437840,             0.86383673,             0.93630291,             0.67277404,             0.93899264,             0.52422773,             0.44892176,             0.03127759,             0.45025550,             0.97136977,             0.13565978,             0.71567448,             0.92094825,             0.93536442,             0.85910449,             0.18252879,             0.72830945,             0.96736828,             0.89831575,             0.83437150,             0.59050780,             0.36145925,             0.16483070,             0.44021176,             0.76018652,             0.44227383,             0.13052339,             0.18204235,             0.99743733,             0.26885190,             0.87726522,             0.16396056,             0.94943412,             0.40016700,             0.65267938,             0.71073267,             0.40094733,             0.91182634,             0.05391789,             0.49520416,             0.24963864,             0.34847086,             0.74088617,             0.36115701,             0.63074210,             0.97423085,             0.42216846,             0.06326975,             0.07858702,             0.20586622,             0.34755773,             0.63166554,             0.18849320,             0.34828456,             0.98477707,             0.75163124,             0.33309570,             0.67563176,             0.98343578,             0.95919930,             0.66994391,             0.89296165,             0.28752144,             0.38146961,             0.83518735,             0.08207577,             0.82083487,             0.81665728,             0.83306004,             0.14203056,             0.01497920,             0.85727447,             0.71194544,             0.85654019,             0.86160433,             0.79580411,             0.47710411,             0.09318029,             0.31369071,             0.64122249,             0.58399725,             0.26706597,             0.05655339,             0.91025211,             0.30330468,             0.33142930,             0.05668627,             0.02936449,             0.12613087,             0.09960114,             0.16218074,             0.15088139,             0.31239040,             0.55980062,             0.34804391,             0.34941538,             0.61370555,             0.07022964,             0.27274571,             0.83306066,             0.75830824,             0.25963478,             0.87137718,             0.24418835,             0.59371493,             0.74479056,             0.84699625,             0.51210368,             0.12489571,             0.23371067,             0.18361641,             0.48636240,             0.06052657,             0.04241913,             0.66710351,             0.07007925,             0.60553664,             0.07536713,             0.55971796,             0.38764845,             0.20737843,             0.37989120,             0.59757058,             0.31189846,             0.25215345,             0.52546591,             0.55744218,             0.59485650,             0.05032742,             0.52076188,             0.47762345,             0.89829370,             0.34417708,             0.84705151,             0.08203183,             0.10632956,             0.78431292,             0.86441722,             0.36487598,             0.09833603,             0.85863594,             0.11010505,             0.11659283,             0.42500288,             0.02747301,             0.12359903,             0.19736489,             0.44461885,             0.33341706,             0.22519571,             0.31528710,             0.14802902,             0.01753431,             0.41160932,             0.47245979,             0.08268172,             0.21580773,             0.75770279,             0.64171939,             0.52643769,             0.19261234,             0.98032835,             0.15401656,             0.85274458,             0.66408502,             0.23212704,             0.74630026,             0.05713613,             0.49025892,             0.48418810,             0.59541513,             0.09243053,             0.93919152,             0.95357019,             0.52377729,             0.65963871,             0.47934951,             0.49919534,             0.34369898,             0.78211256,             0.13908708,             0.95754117,             0.84107746,             0.09126213,             0.42979124,             0.10295325,             0.34631257,             0.69448345,             0.41720536,             0.15282440,             0.74329854,             0.45775009,             0.12786280,             0.39830299,             0.20386769,             0.59703523,             0.94077086,             0.42255597,             0.80453309,             0.79757204,             0.28653229,             0.60175909,             0.55859623,             0.34318230,             0.63002770,             0.36533324,             0.89689906,             0.73236186,             0.61491989,             0.83787947,             0.67939463,             0.72016694,             0.77499849,             0.72428343,             0.34571059,             0.23143007,             0.20099338,             0.85583142,             0.73174191,             0.54284092,             0.20264181,             0.53037061,             0.30493131,             0.82279766,             0.58542432,             0.72632070,             0.18394258,             0.00608118,             0.23808232,             0.17007573,             0.75245459,             0.84990616,             0.38827634,             0.33809538,             0.01080317,             0.27250145,             0.81769542,             0.15323253,             0.71668395,             0.99427044,             0.11355576,             0.50511923,             0.22952055,             0.78271870,             0.12833592,             0.88639055,             0.76398188,             0.49533508,             0.47939640,             0.73564612,             0.41465671,             0.10995635,             0.20271728,             0.00521771,             0.67265260,             0.11917707,             0.76574855,             0.43842117,             0.28530411,             0.79648090,             0.79433656,             0.12074559,             0.02325163,             0.10117917,             0.83559239,             0.67213900,             0.25247084,             0.47968157,             0.88649124,             0.33588961,             0.92338319,             0.18808573,             0.60248266,             0.36610154,             0.99123140,             0.10519719,             0.18754650,             0.43232584,             0.85447872,             0.15937568,             0.92947480,             0.62705964,             0.85960084,             0.13435660,             0.81845809,             0.60715133,             0.83030708,             0.83071910,             0.38883408,             0.92033237,             0.59820890,             0.75527947,             0.67683355,             0.21847023,             0.29395619,             0.50477953,             0.98977921,             0.96225332,             0.90143562,             0.19559914,             0.08978307,             0.09687492,             0.07381865,             0.22801110,             0.26669388,             0.99691302,             0.12113623,             0.34373057,             0.46066239,             0.48806761,             0.50688779,             0.00654483,             0.32076493,             0.42367646,             0.07112842,             0.54090558,             0.68230725,             0.49713828,             0.41958965,             0.68013847,             0.47691765,             0.63269259,             0.94304095,             0.54587271,             0.72447569,             0.28913523,             0.75766936,             0.52965692,             0.96854824,             0.15589071,             0.84128672,             0.16337522,             0.05771034,             0.21556356,             0.12094140,             0.29721207,             0.00811008,             0.66184926});
+
+    nd4j::ops::reverse_sequence op;
+    auto results = op.execute({&input, &lengths}, {}, {1, 0});
+    ASSERT_EQ(Status::OK(), results->status());
+
+    auto z = results->at(0);
+
+    ASSERT_EQ(e, *z);
+
+    delete results;
+}
+
 //////////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests5, Test_TopK_0) {
     auto x = NDArrayFactory::create<double>('c', {2, 6}, {1.0, 1.0, 1.0, 1.0, 11.0, 3.0, 1.0, 1.0, 1.0, 14.0, 5.0, 6.0});
@@ -2292,8 +2313,8 @@ TEST_F(DeclarableOpsTests5, confusion_matrix_test3) {
 //////////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests5, confusion_matrix_test4) {
 
-    auto labels = NDArrayFactory::create<double>('c', {1, 2}, {1, 2});
-    auto predictions = NDArrayFactory::create<double>('c', {1, 2}, {0, 2});
+    auto labels = NDArrayFactory::create<int>('c', {1, 2}, {1, 2});
+    auto predictions = NDArrayFactory::create<int>('c', {1, 2}, {0, 2});
     auto weights = NDArrayFactory::create<double>('c', {1, 2}, {100, 200});
     auto expected = NDArrayFactory::create<double>('c', {3, 3}, {0, 0, 0, 100, 0, 0, 0, 0, 200});
 

From f4860574d7bab01f1f59c3973828b757efa9c84d Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Wed, 28 Aug 2019 15:38:57 +0300
Subject: [PATCH 34/56] [WIP] More fixes (#190)

* Refactored kernels for segment_max/min/sum ops.

* Refactored segment_prod kernels.

* Refactored segment_prod kernels.

* DynamicPartition test

Signed-off-by: raver119 <raver119@gmail.com>

* Addede linear test for dynamic_partition op.

* Refactored test with int datatype.

* some logging

Signed-off-by: raver119 <raver119@gmail.com>

* some logging

Signed-off-by: raver119 <raver119@gmail.com>

* some logging

Signed-off-by: raver119 <raver119@gmail.com>

* dynamicPartition fix

Signed-off-by: raver119 <raver119@gmail.com>

* get rid of some logging

Signed-off-by: raver119 <raver119@gmail.com>

* one more test for dynamic_stitch

Signed-off-by: raver119 <raver119@gmail.com>

* one more test for dynamic_stitch

Signed-off-by: raver119 <raver119@gmail.com>

* empty check for stitch

Signed-off-by: raver119 <raver119@gmail.com>

* minor print changes

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../ops/declarable/helpers/cuda/dynamic.cu    | 16 ++--
 .../declarable/helpers/cuda/segment_max.cu    | 49 ++++++------
 .../declarable/helpers/cuda/segment_min.cu    | 31 ++++----
 .../declarable/helpers/cuda/segment_prod.cu   |  8 +-
 .../declarable/helpers/cuda/segment_sum.cu    | 12 +--
 .../ops/declarable/helpers/cuda/svd.cu        |  4 +
 .../layers_tests/DeclarableOpsTests5.cpp      | 76 +++++++++++++++++++
 .../opvalidation/MiscOpValidation.java        | 18 +++++
 .../TFGraphs/TFGraphTestAllHelper.java        |  2 +
 .../TFGraphs/TFGraphTestAllSameDiff.java      |  4 +-
 10 files changed, 164 insertions(+), 56 deletions(-)

diff --git a/libnd4j/include/ops/declarable/helpers/cuda/dynamic.cu b/libnd4j/include/ops/declarable/helpers/cuda/dynamic.cu
index 857ebed38..92e5b38b4 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/dynamic.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/dynamic.cu
@@ -50,8 +50,8 @@ namespace nd4j {
                     auto zShapeInfo = zShapeInfos[o];
                     auto zLength = shape::length(zShapeInfo);
 
-                    // iLimit should be
-                    auto iLimit = iLength <= blockIdx.x ? blockIdx.x : (iLength + (blockIdx.x - (iLength % blockIdx.x)));
+                    // iLimit should be multiple of blockDim.x
+                    auto iLimit = iLength <= blockDim.x ? blockDim.x : (iLength + (blockDim.x - (iLength % blockDim.x)));
                     int cnt = 0;
 
                     for (Nd4jLong e = threadIdx.x; e < iLimit; e += blockDim.x) {
@@ -75,8 +75,9 @@ namespace nd4j {
 
                         // doing actual update
                         if (e < iLength)
-                            if (trueIndices[threadIdx.x] >= 0)
+                            if (trueIndices[threadIdx.x] >= 0) {
                                 z[trueIndices[threadIdx.x]] = x[shape::getIndexOffset(e, xShapeInfo, xLength)];
+                            }
 
                         __syncthreads();
                     }
@@ -148,13 +149,12 @@ namespace nd4j {
                     auto dOutTadShapes = reinterpret_cast<Nd4jLong **>(pm.replicatePointer(tadShapes.data(), tadShapes.size() * sizeof(Nd4jLong *)));
                     auto dOutTadOffsets = reinterpret_cast<Nd4jLong **>(pm.replicatePointer(tadOffsets.data(), tadOffsets.size() * sizeof(Nd4jLong *)));
 
-                    dynamicPartitionTadKernel<X,Y><<<256, 512, 1024, *context->getCudaStream()>>>(input->getSpecialBuffer(), packX.platformShapeInfo(), packX.platformOffsets(), shape::length(packX.primaryShapeInfo()), indices->getSpecialBuffer(), indices->getSpecialShapeInfo(), indices->lengthOf(), dOutBuffers, dOutTadShapes, dOutTadOffsets, outSize);
+                    dynamicPartitionTadKernel<X,Y><<<256, 256, 1024, *context->getCudaStream()>>>(input->getSpecialBuffer(), packX.platformShapeInfo(), packX.platformOffsets(), shape::length(packX.primaryShapeInfo()), indices->getSpecialBuffer(), indices->getSpecialShapeInfo(), indices->lengthOf(), dOutBuffers, dOutTadShapes, dOutTadOffsets, outSize);
 
                 } else {
                     auto numThreads = 256;
                     auto shmemSize = numThreads * sizeof(Y) * 2 + 1024;
 
-
                     std::vector<void *> outBuffers;
                     std::vector<Nd4jLong *> outShapes;
 
@@ -203,6 +203,9 @@ namespace nd4j {
                     auto indices = reinterpret_cast<Y*>(vindices[e]);
                     auto iShapeInfo = iShapeInfos[e];
 
+                    if (shape::isEmpty(iShapeInfo))
+                        continue;
+
                     auto iLength = shape::length(iShapeInfo);
                     auto zLength = shape::length(zTadShapeInfo);
 
@@ -310,8 +313,9 @@ namespace nd4j {
 
                 NDArray::registerSpecialUse({}, {indices, input});
 
-                for (auto v:outputList)
+                for (auto v:outputList) {
                     v->tickWriteDevice();
+                }
             }
 
             template <typename T>
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_max.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_max.cu
index dc91a2704..8830f37e7 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_max.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_max.cu
@@ -40,19 +40,16 @@ namespace nd4j {
             static __global__ void
             segmentMaxLinearKernel(void *input, Nd4jLong *inputShape, int *starts, int *lengths, Nd4jLong numOfClasses,
                                    void *output, Nd4jLong *outputShape) {
-                __shared__
-                T *val;
-                __shared__
-                Nd4jLong xLen, zLen, segment, zIndex;
-                __shared__
-                T *x;
-                __shared__
-                T *z;
+                __shared__                 T *val;
+                __shared__                Nd4jLong xLen, zLen, zIndex;
+                __shared__                T *x;
+                __shared__                T *z;
                 __shared__ int threadsPerSegment, start, finish;
 
+                auto segment = blockIdx.x;
                 if (threadIdx.x == 0) {
-                    threadsPerSegment = (gridDim.x + numOfClasses - 1) / numOfClasses;
-                    segment = blockIdx.x / threadsPerSegment;
+//                    threadsPerSegment = (gridDim.x + numOfClasses - 1) / numOfClasses;
+//                    segment = blockIdx.x / threadsPerSegment;
                     x = reinterpret_cast<T *>(input);
                     z = reinterpret_cast<T *>(output);
                     extern __shared__ unsigned char shmem[];
@@ -83,19 +80,14 @@ namespace nd4j {
             unsortedSegmentMaxLinearKernel(void *input, Nd4jLong *inputShape, void *indices, Nd4jLong *indicesShape,
                                            int *starts, int *lengths, Nd4jLong numOfClasses, void *output,
                                            Nd4jLong *outputShape) {
-                __shared__
-                T *val;
-                __shared__
-                Nd4jLong xLen, zLen, segment, zIndex;
-                __shared__
-                T *x;
-                __shared__
-                T *z;
-                __shared__
-                I *y; //int threadsPerSegment, start, finish;
+                __shared__                 T *val;
+                __shared__                Nd4jLong xLen, zLen, zIndex;
+                __shared__                T *x;
+                __shared__                T *z;
+                __shared__                I *y; //int threadsPerSegment, start, finish;
+                auto segment = blockIdx.x;
 
                 if (threadIdx.x == 0) {
-                    segment = blockIdx.x;
                     x = reinterpret_cast<T *>(input);
                     z = reinterpret_cast<T *>(output);
                     y = reinterpret_cast<I *>(indices);
@@ -127,9 +119,10 @@ namespace nd4j {
                                                        Nd4jLong* outputShape, Nd4jLong* outputTads, Nd4jLong* outputTadOffsets, T filler = 0) {
 
                 __shared__ T* val;
-                __shared__ Nd4jLong len, segment, zIndex, total;
+                __shared__ Nd4jLong len, zIndex, total;
                 __shared__ T* z;
                 __shared__ int start, finish;
+                __shared__ I segment;
 
                 if (threadIdx.x == 0) {
                     segment = indices[blockIdx.x]; // / threadsPerSegment;
@@ -143,20 +136,22 @@ namespace nd4j {
                 __syncthreads();
 
                 auto idx = blockIdx.x;
-                if (blockIdx.x <= total) {
+                if (idx <= total) {
                     auto x = reinterpret_cast<T *>(inputBuf) + inputTadOffsets[idx];
                     if (blockIdx.x == start) {
                         for (auto e = threadIdx.x; e < len; e += blockDim.x) {
                             auto xIndex = shape::getIndexOffset(e, inputTads, len);
                             auto zIndex = shape::getIndexOffset(e, outputTads, len);
-                            z[zIndex] = x[xIndex];
+                            nd4j::math::atomics::nd4j_atomicMax(&z[zIndex], x[xIndex]);
+                            //z[zIndex] = x[xIndex];
                         }
                     }
                     else {
                         for (auto e = threadIdx.x; e < len; e += blockDim.x) {
                             auto xIndex = shape::getIndexOffset(e, inputTads, len);
                             auto zIndex = shape::getIndexOffset(e, outputTads, len);
-                            nd4j::math::atomics::nd4j_atomicMax(&z[zIndex], x[xIndex]);
+                            if (lengths[segment])
+                                nd4j::math::atomics::nd4j_atomicMax(&z[zIndex], x[xIndex]);
                         }
                     }
                 }
@@ -168,6 +163,7 @@ namespace nd4j {
                 //int numClasses = output->sizeAt(0);
                 // if input is a vector: (as if in doc sample)
                 //Nd4jLong idx = indices->e<Nd4jLong>(0);
+                output->assign(-DataTypeUtils::infOrMax<T>());
                 auto stream = context->getCudaStream();
                 indices->syncToHost();
                 Nd4jLong numOfClasses = indices->e<Nd4jLong>(indices->lengthOf() - 1) + 1;
@@ -211,6 +207,8 @@ namespace nd4j {
             static void unsortedSegmentMaxFunctor_(nd4j::LaunchContext* context, NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
                 auto stream = context->getCudaStream();
 //        NDArray classes = NDArrayFactory::create<int>('c', {numOfClasses, 2});
+                output->assign(DataTypeUtils::infOrMax<T>());
+
                 NDArray classesRangesBegs = NDArrayFactory::create<int>('c', {numOfClasses});
                 NDArray classesRangesLens = NDArrayFactory::create<int>('c', {numOfClasses});
 //        NDArray row = NDArrayFactory::create<int>('c', {1, 2}, {(int)indices->lengthOf(), (int)0});
@@ -243,6 +241,7 @@ namespace nd4j {
             // -------------------------------------------------------------------------------------------------------------- //
             void unsortedSegmentMaxFunctor(nd4j::LaunchContext* context, NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
                 NDArray::prepareSpecialUse({output}, {input, indices});
+                output->nullify();
                 BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), unsortedSegmentMaxFunctor_, (context, input, indices, numOfClasses, output), NUMERIC_TYPES, INDEXING_TYPES);
                 NDArray::registerSpecialUse({output}, {input, indices});
             }
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_min.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_min.cu
index 950abde67..e5ea2eb91 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_min.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_min.cu
@@ -38,19 +38,16 @@ namespace helpers {
     static __global__ void
     segmentMinLinearKernel(void *input, Nd4jLong *inputShape, int *starts, int *lengths, Nd4jLong numOfClasses,
                            void *output, Nd4jLong *outputShape) {
-        __shared__
-        T *val;
-        __shared__
-        Nd4jLong xLen, zLen, segment, zIndex;
-        __shared__
-        T *x;
-        __shared__
-        T *z;
+        __shared__        T *val;
+        __shared__        Nd4jLong xLen, zLen, zIndex;
+        __shared__        T *x;
+        __shared__        T *z;
         __shared__ int threadsPerSegment, start, finish;
 
+        auto segment = blockIdx.x;
         if (threadIdx.x == 0) {
-            threadsPerSegment = (gridDim.x + numOfClasses - 1) / numOfClasses;
-            segment = blockIdx.x / threadsPerSegment;
+//            threadsPerSegment = (gridDim.x + numOfClasses - 1) / numOfClasses;
+//            segment = blockIdx.x / threadsPerSegment;
             x = reinterpret_cast<T *>(input);
             z = reinterpret_cast<T *>(output);
             extern __shared__ unsigned char shmem[];
@@ -123,12 +120,12 @@ namespace helpers {
     template <typename T, typename I>
     static __global__ void segmentMinTadKernel(void* inputBuf, Nd4jLong* inputShape, Nd4jLong* inputTads, Nd4jLong* inputTadOffsets, I* indices, int* starts, int* lengths, Nd4jLong numOfClasses, void* outputBuf, Nd4jLong* outputShape, Nd4jLong* outputTads, Nd4jLong* outputTadOffsets) {
         __shared__ T* val;
-        __shared__ Nd4jLong len, segment, zIndex, total;
+        __shared__ Nd4jLong len, zIndex, total;
         __shared__ T* z;
         __shared__ int threadsPerSegment, start, finish;
 
+        auto segment = indices[blockIdx.x]; // / threadsPerSegment;
         if (threadIdx.x == 0) {
-            segment = indices[blockIdx.x]; // / threadsPerSegment;
             z = reinterpret_cast<T*>(outputBuf) + outputTadOffsets[segment];
             len = shape::length(inputTads);
             start = starts[segment];
@@ -145,14 +142,15 @@ namespace helpers {
                 for (auto e = threadIdx.x; e < len; e += blockDim.x) {
                     auto xIndex = shape::getIndexOffset(e, inputTads, len);
                     auto zIndex = shape::getIndexOffset(e, outputTads, len);
-                    z[zIndex] = x[xIndex];
+                    nd4j::math::atomics::nd4j_atomicMin(&z[zIndex], x[xIndex]);
                 }
             }
             else {
                 for (auto e = threadIdx.x; e < len; e += blockDim.x) {
                     auto xIndex = shape::getIndexOffset(e, inputTads, len);
                     auto zIndex = shape::getIndexOffset(e, outputTads, len);
-                    nd4j::math::atomics::nd4j_atomicMin(&z[zIndex], x[xIndex]);
+//                    if (lengths[indices[idx]])
+                        nd4j::math::atomics::nd4j_atomicMin(&z[zIndex], x[xIndex]);
                 }
             }
         }
@@ -165,7 +163,7 @@ namespace helpers {
         Nd4jLong numClasses = indices->e<Nd4jLong>(indices->lengthOf() - 1) + 1;
         NDArray classesRangesLens = NDArrayFactory::create<int>('c', {numClasses});
         NDArray classesRangesBegs = NDArrayFactory::create<int>('c', {numClasses});
-
+        output->assign(DataTypeUtils::infOrMax<T>());
         classesRangesBegs.assign(indices->lengthOf());
         classesRangesLens.assign(0);
 
@@ -193,6 +191,7 @@ namespace helpers {
     // -------------------------------------------------------------------------------------------------------------- //
     void segmentMinFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* output) {
         NDArray::prepareSpecialUse({output}, {input, indices});
+        output->nullify();
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), segmentMinFunctor_, (context, input, indices, output), NUMERIC_TYPES, INDEXING_TYPES);
         NDArray::registerSpecialUse({output}, {input, indices});
     }
@@ -207,6 +206,7 @@ namespace helpers {
         NDArray classesRangesLens = NDArrayFactory::create<int>('c', {numOfClasses});
 //        NDArray row = NDArrayFactory::create<int>('c', {1, 2}, {(int)indices->lengthOf(), (int)0});
 //        classes.applyTrueBroadcast(nd4j::BroadcastOpsTuple::Assign(), &row, &classes);
+        output->assign(DataTypeUtils::infOrMax<T>());
         classesRangesBegs.assign(indices->lengthOf());
         classesRangesLens.assign(0);
         dim3 dims(numOfClasses, indices->lengthOf(), numOfClasses * 32 + 32);
@@ -236,6 +236,7 @@ namespace helpers {
     // -------------------------------------------------------------------------------------------------------------- //
     void unsortedSegmentMinFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
         NDArray::prepareSpecialUse({output}, {input, indices});
+        output->nullify();
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), unsortedSegmentMinFunctor_, (context, input, indices, numOfClasses, output),
                               NUMERIC_TYPES, INDEXING_TYPES);
         NDArray::registerSpecialUse({output}, {input, indices});
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_prod.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_prod.cu
index 3ae4ebcb8..5709a63ea 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_prod.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_prod.cu
@@ -146,14 +146,15 @@ namespace helpers {
                 for (auto e = threadIdx.x; e < len; e += blockDim.x) {
                     auto xIndex = shape::getIndexOffset(e, inputTads, len);
                     auto zIndex = shape::getIndexOffset(e, outputTads, len);
-                    z[zIndex] = x[xIndex];
+                    nd4j::math::atomics::nd4j_atomicMul(&z[zIndex], x[xIndex]);
                 }
             }
             else {
                 for (auto e = threadIdx.x; e < len; e += blockDim.x) {
                     auto xIndex = shape::getIndexOffset(e, inputTads, len);
                     auto zIndex = shape::getIndexOffset(e, outputTads, len);
-                    nd4j::math::atomics::nd4j_atomicMul(&z[zIndex], x[xIndex]);
+                    if (lengths[segment] > 0)
+                        nd4j::math::atomics::nd4j_atomicMul(&z[zIndex], x[xIndex]);
                 }
             }
         }
@@ -166,7 +167,7 @@ namespace helpers {
         Nd4jLong numClasses = indices->e<Nd4jLong>(indices->lengthOf() - 1) + 1;
         NDArray classesRangesLens = NDArrayFactory::create<int>('c', {numClasses});
         NDArray classesRangesBegs = NDArrayFactory::create<int>('c', {numClasses});
-
+        output->assign(1);
         classesRangesBegs.assign(indices->lengthOf());
         classesRangesLens.assign(0);
 
@@ -373,6 +374,7 @@ namespace helpers {
     template <typename T, typename I>
     static int unsortedSegmentProdFunctorBP_(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* gradOut, Nd4jLong numOfClasses, NDArray* output) {
         auto stream = context->getCudaStream();
+
         NDArray tempRes(gradOut->ordering(), gradOut->getShapeAsVector(), DataTypeUtils::fromT<T>(), context);//->shapeInfo(), context);
         unsortedSegmentProdFunctor_<T, I>(context, input, indices, numOfClasses, &tempRes);
         NDArray::prepareSpecialUse({output}, {input, indices, gradOut});
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/segment_sum.cu b/libnd4j/include/ops/declarable/helpers/cuda/segment_sum.cu
index 08b36253a..4b8976f4e 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/segment_sum.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/segment_sum.cu
@@ -121,12 +121,12 @@ namespace helpers {
     template <typename T, typename I>
     static __global__ void segmentSumTadKernel(void* inputBuf, Nd4jLong* inputShape, Nd4jLong* inputTads, Nd4jLong* inputTadOffsets, I* indices, int* starts, int* lengths, Nd4jLong numOfClasses, void* outputBuf, Nd4jLong* outputShape, Nd4jLong* outputTads, Nd4jLong* outputTadOffsets) {
         __shared__ T* val;
-        __shared__ Nd4jLong len, segment, zIndex, total;
+        __shared__ Nd4jLong len, zIndex, total;
         __shared__ T* z;
-        __shared__ int threadsPerSegment, start, finish;
+        __shared__ int start, finish;
 
         if (threadIdx.x == 0) {
-            segment = indices[blockIdx.x]; // / threadsPerSegment;
+            auto segment = indices[blockIdx.x]; // / threadsPerSegment;
             z = reinterpret_cast<T*>(outputBuf) + outputTadOffsets[segment];
             len = shape::length(inputTads);
             start = starts[segment];
@@ -143,14 +143,14 @@ namespace helpers {
                 for (auto e = threadIdx.x; e < len; e += blockDim.x) {
                     auto xIndex = shape::getIndexOffset(e, inputTads, len);
                     auto zIndex = shape::getIndexOffset(e, outputTads, len);
-                    z[zIndex] = x[xIndex];
+                    nd4j::math::atomics::nd4j_atomicAdd(&z[zIndex], x[xIndex]);
                 }
             }
             else {
                 for (auto e = threadIdx.x; e < len; e += blockDim.x) {
                     auto xIndex = shape::getIndexOffset(e, inputTads, len);
                     auto zIndex = shape::getIndexOffset(e, outputTads, len);
-                    if (lengths[segment])
+                    if (lengths[indices[idx]])
                         nd4j::math::atomics::nd4j_atomicAdd(&z[zIndex], x[xIndex]);
                 }
             }
@@ -191,6 +191,7 @@ namespace helpers {
     // -------------------------------------------------------------------------------------------------------------- //
     void segmentSumFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, NDArray* output) {
         NDArray::prepareSpecialUse({output}, {input, indices});
+        output->nullify();
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), segmentSumFunctor_, (context, input, indices, output), NUMERIC_TYPES, INDEXING_TYPES);
         NDArray::registerSpecialUse({output}, {input, indices});
     }
@@ -232,6 +233,7 @@ namespace helpers {
     // -------------------------------------------------------------------------------------------------------------- //
     void unsortedSegmentSumFunctor(nd4j::LaunchContext* context , NDArray* input, NDArray* indices, Nd4jLong numOfClasses, NDArray* output) {
         NDArray::prepareSpecialUse({output}, {input, indices});
+        output->nullify();
         BUILD_DOUBLE_SELECTOR(input->dataType(), indices->dataType(), unsortedSegmentSumFunctor_, (context, input, indices, numOfClasses, output),
                               NUMERIC_TYPES, INDEXING_TYPES);
         NDArray::registerSpecialUse({output}, {input, indices});
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/svd.cu b/libnd4j/include/ops/declarable/helpers/cuda/svd.cu
index da4f5cc86..0695119da 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/svd.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/svd.cu
@@ -602,6 +602,8 @@ void svd(nd4j::LaunchContext* context, const NDArray* x, const std::vector<NDArr
     // NDArray VT = outArrs[2]->transpose();
     NDArray* V = outArrs[2];
 
+    NDArray::prepareSpecialUse({S, U, V}, {x});
+
     if(x->rankOf() == 2) {
         // svdQR(context, x, S, U, VT, fullUV, calcUV);
         svdJcb(context, x, S, U, V, fullUV, calcUV);
@@ -631,6 +633,8 @@ void svd(nd4j::LaunchContext* context, const NDArray* x, const std::vector<NDArr
             delete tadsV;
         }
     }
+
+    NDArray::registerSpecialUse({S, U, V}, {x});
 }
 
 
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
index b256eef10..2ff2fad58 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
@@ -1920,6 +1920,50 @@ TEST_F(DeclarableOpsTests5, EmbeddingLookup_3) {
 
     delete result;
 }
+/* @Test
+    public void testDynamicPartition(){
+        INDArray data = Nd4j.createFromArray(2, 1, 2, 0);
+        INDArray partitions = Nd4j.createFromArray(0, 2, 1, 0);
+        INDArray[] out = Nd4j.exec(DynamicCustomOp.builder("dynamic_partition")
+                .addOutputs(Nd4j.createUninitialized(DataType.INT, 2), Nd4j.createUninitialized(DataType.INT, 1), Nd4j.createUninitialized(DataType.INT, 1))
+                .addIntegerArguments(3) //3 partitions
+                .addInputs(data, partitions).build());
+
+        INDArray exp0 = Nd4j.createFromArray(2, 0);
+        INDArray exp1 = Nd4j.createFromArray(2);
+        INDArray exp2 = Nd4j.createFromArray(1);
+
+        assertEquals(exp0, out[0]);     //Usually just gives [0,0]
+        assertEquals(exp1, out[1]);
+        assertEquals(exp2, out[2]);
+    }*/
+TEST_F(DeclarableOpsTests5, DynamicPartition_01) {
+
+    auto x = NDArrayFactory::create<int>({2,1,2,0});
+
+    auto y = NDArrayFactory::create<int>({0,2,1,0});
+
+    int numPartition = 3;
+    std::vector<NDArray> exp( { NDArrayFactory::create<int>('c', {2}, {2, 0}),
+                                NDArrayFactory::create<int>('c', {1}, {2}),
+                                NDArrayFactory::create<int>('c', {1}, {1})});
+
+    nd4j::ops::dynamic_partition op;
+    auto result = op.execute({&x, &y}, {}, {numPartition});
+
+    ASSERT_EQ(ND4J_STATUS_OK, result->status());
+    ASSERT_EQ(result->size(), numPartition); // result has the same size as given param 4
+
+    for (int e = 0; e < result->size(); e++) {
+        auto output = result->at(e);
+        // output->printShapeInfo("Output shape> ");
+        // output->printIndexedBuffer("Output data> ");
+        ASSERT_TRUE(exp[e].isSameShape(output));
+        ASSERT_TRUE(exp[e].equalsTo(output));
+    }
+
+    delete result;
+}
 
 TEST_F(DeclarableOpsTests5, DynamicPartition_1) {
 
@@ -2031,6 +2075,38 @@ TEST_F(DeclarableOpsTests5, DynamicPartition_3) {
     delete result;
 }
 
+TEST_F(DeclarableOpsTests5, DynamicStitch_empty_1) {
+    auto i0 = NDArrayFactory::create<int>('c', {2}, {2, 3});
+    auto i1 = NDArrayFactory::empty<int>();
+    auto i2 = NDArrayFactory::create<int>('c', {2}, {0, 1});
+
+    auto d0 = NDArrayFactory::create<double>('c', {2, 5}, {0.085571885,0.7937801,0.65908563,0.55552566,0.15962744,0.7787856,0.80119777,0.72437465,0.23089433,0.72714126});
+    auto d1 = NDArrayFactory::empty<double>();
+    auto d2 = NDArrayFactory::create<double>('c', {2, 5}, {0.94414854,0.5956861,0.8668989,0.3502196,0.5100082,0.061725974,0.6621324,0.034165382,0.32576954,0.51917326});
+
+    nd4j::ops::dynamic_stitch op;
+    auto result = op.execute({&i0, &i1, &i2, &d0, &d1, &d2}, {}, {});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    delete result;
+}
+
+TEST_F(DeclarableOpsTests5, DynamicStitch_empty_2) {
+    auto i0 = NDArrayFactory::create<int>('c', {2}, {2, 3});
+    auto i1 = NDArrayFactory::create<int>('c', {0});
+    auto i2 = NDArrayFactory::create<int>('c', {2}, {0, 1});
+
+    auto d0 = NDArrayFactory::create<double>('c', {2, 5}, {0.085571885,0.7937801,0.65908563,0.55552566,0.15962744,0.7787856,0.80119777,0.72437465,0.23089433,0.72714126});
+    auto d1 = NDArrayFactory::create<double>('c', {0, 5});
+    auto d2 = NDArrayFactory::create<double>('c', {2, 5}, {0.94414854,0.5956861,0.8668989,0.3502196,0.5100082,0.061725974,0.6621324,0.034165382,0.32576954,0.51917326});
+
+    nd4j::ops::dynamic_stitch op;
+    auto result = op.execute({&i0, &i1, &i2, &d0, &d1, &d2}, {}, {});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    delete result;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 TEST_F(DeclarableOpsTests5, DynamicStitch_1) {
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/MiscOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/MiscOpValidation.java
index 99a2f57ac..eb228bf1f 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/MiscOpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/MiscOpValidation.java
@@ -1728,6 +1728,24 @@ public class MiscOpValidation extends BaseOpValidation {
         assertEquals(exp, out);
     }
 
+    @Test
+    public void testDynamicPartition(){
+        INDArray data = Nd4j.createFromArray(2, 1, 2, 0);
+        INDArray partitions = Nd4j.createFromArray(0, 2, 1, 0);
+        INDArray[] out = Nd4j.exec(DynamicCustomOp.builder("dynamic_partition")
+                .addOutputs(Nd4j.createUninitialized(DataType.INT, 2), Nd4j.createUninitialized(DataType.INT, 1), Nd4j.createUninitialized(DataType.INT, 1))
+                .addIntegerArguments(3) //3 partitions
+                .addInputs(data, partitions).build());
+
+        INDArray exp0 = Nd4j.createFromArray(2, 0);
+        INDArray exp1 = Nd4j.createFromArray(2);
+        INDArray exp2 = Nd4j.createFromArray(1);
+
+        assertEquals(exp0, out[0]);     //Usually just gives [0,0]
+        assertEquals(exp1, out[1]);
+        assertEquals(exp2, out[2]);
+    }
+
     @Test
     public void testListDiff(){
         INDArray x = Nd4j.createFromArray(0, 1, 2, 3);
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/imports/TFGraphs/TFGraphTestAllHelper.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/imports/TFGraphs/TFGraphTestAllHelper.java
index ab86f829e..ecc81c981 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/imports/TFGraphs/TFGraphTestAllHelper.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/imports/TFGraphs/TFGraphTestAllHelper.java
@@ -344,6 +344,8 @@ public class TFGraphTestAllHelper {
                                 System.out.println("Pass: " + varName);
                             } else {
                                 System.out.println("FAIL: " + varName);
+                                System.out.println("TF:\n" + tfValue);
+                                System.out.println("SD:\n" + sdVal);
                             }
 
                         }
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/imports/TFGraphs/TFGraphTestAllSameDiff.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/imports/TFGraphs/TFGraphTestAllSameDiff.java
index 4f3520d31..1fef4a07b 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/imports/TFGraphs/TFGraphTestAllSameDiff.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/imports/TFGraphs/TFGraphTestAllSameDiff.java
@@ -180,8 +180,8 @@ public class TFGraphTestAllSameDiff {   //Note: Can't extend BaseNd4jTest here a
         Double minAbs = (precisionOverride == null ? null : precisionOverride.getSecond());
 
         try {
-            TFGraphTestAllHelper.checkOnlyOutput(inputs, predictions, modelName, BASE_DIR, MODEL_FILENAME, EXECUTE_WITH,
-                    TFGraphTestAllHelper.LOADER, maxRE, minAbs);
+            TFGraphTestAllHelper.checkOnlyOutput(inputs, predictions, modelName, BASE_DIR, MODEL_FILENAME, EXECUTE_WITH, TFGraphTestAllHelper.LOADER, maxRE, minAbs);
+            //TFGraphTestAllHelper.checkIntermediate(inputs, modelName, BASE_DIR, MODEL_FILENAME, EXECUTE_WITH, localTestDir);
         } catch (Throwable t){
             log.error("ERROR Executing test: {} - input keys {}", modelName, (inputs == null ? null : inputs.keySet()), t);
             throw t;

From dec296da17a7a0f3397db3daa8e1d84dfe947de3 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Wed, 28 Aug 2019 18:20:44 +0300
Subject: [PATCH 35/56] [WIP] bits_hamming_distance (#192)

* bits_hamming_distance op

Signed-off-by: raver119 <raver119@gmail.com>

* bits_hamming_distance cuda

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../generic/bitwise/bits_hamming_distance.cpp | 57 +++++++++++
 .../include/ops/declarable/headers/bitwise.h  | 11 +++
 .../ops/declarable/helpers/cpu/hamming.cpp    | 72 ++++++++++++++
 .../ops/declarable/helpers/cuda/hamming.cu    | 95 +++++++++++++++++++
 .../include/ops/declarable/helpers/hamming.h  | 32 +++++++
 .../layers_tests/DeclarableOpsTests16.cpp     | 16 ++++
 6 files changed, 283 insertions(+)
 create mode 100644 libnd4j/include/ops/declarable/generic/bitwise/bits_hamming_distance.cpp
 create mode 100644 libnd4j/include/ops/declarable/helpers/cpu/hamming.cpp
 create mode 100644 libnd4j/include/ops/declarable/helpers/cuda/hamming.cu
 create mode 100644 libnd4j/include/ops/declarable/helpers/hamming.h

diff --git a/libnd4j/include/ops/declarable/generic/bitwise/bits_hamming_distance.cpp b/libnd4j/include/ops/declarable/generic/bitwise/bits_hamming_distance.cpp
new file mode 100644
index 000000000..ff72ff4b9
--- /dev/null
+++ b/libnd4j/include/ops/declarable/generic/bitwise/bits_hamming_distance.cpp
@@ -0,0 +1,57 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2019 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+
+#include <op_boilerplate.h>
+#if NOT_EXCLUDED(OP_bits_hamming_distance)
+
+#include <ops/declarable/CustomOperations.h>
+#include <ops/declarable/helpers/helpers.h>
+#include <ops/declarable/helpers/hamming.h>
+
+namespace nd4j {
+    namespace ops {
+        CUSTOM_OP_IMPL(bits_hamming_distance, 2, 1, true, 0, 0) {
+            auto x = INPUT_VARIABLE(0);
+            auto y = INPUT_VARIABLE(1);
+            auto output = OUTPUT_VARIABLE(0);
+
+            REQUIRE_TRUE(x->lengthOf() == y->lengthOf(), 0, "bits_hamming_distance: both arguments must have the same length");
+            REQUIRE_TRUE(x->dataType() == y->dataType(), 0, "bits_hamming_distance: both arguments must have the same data type");
+
+            helpers::hamming(block.launchContext(), *x, *y, *output);
+
+            return Status::OK();
+        }
+
+        DECLARE_SHAPE_FN(bits_hamming_distance) {
+            return SHAPELIST(ConstantShapeHelper::getInstance()->scalarShapeInfo(nd4j::DataType::INT64));
+        }
+
+        DECLARE_TYPES(bits_hamming_distance) {
+            getOpDescriptor()
+                    ->setAllowedInputTypes(0, {ALL_INTS})
+                    ->setAllowedInputTypes(1, {ALL_INTS})
+                    ->setAllowedOutputTypes(0, {ALL_INDICES})
+                    ->setSameMode(true);
+        }
+    }
+}
+
+#endif
\ No newline at end of file
diff --git a/libnd4j/include/ops/declarable/headers/bitwise.h b/libnd4j/include/ops/declarable/headers/bitwise.h
index 900d42816..da5ff4047 100644
--- a/libnd4j/include/ops/declarable/headers/bitwise.h
+++ b/libnd4j/include/ops/declarable/headers/bitwise.h
@@ -80,6 +80,17 @@ namespace nd4j {
         #if NOT_EXCLUDED(OP_cyclic_rshift_bits)
         DECLARE_CONFIGURABLE_OP(cyclic_rshift_bits, 1, 1, true, 0, -2);
         #endif
+
+        /**
+         * This operation returns hamming distance based on bits
+         *
+         * PLEASE NOTE: This operation is applicable only to integer data types
+         *
+         * @tparam T
+         */
+        #if NOT_EXCLUDED(OP_bits_hamming_distance)
+        DECLARE_CUSTOM_OP(bits_hamming_distance, 2, 1, true, 0, 0);
+        #endif
     }
 }
 
diff --git a/libnd4j/include/ops/declarable/helpers/cpu/hamming.cpp b/libnd4j/include/ops/declarable/helpers/cpu/hamming.cpp
new file mode 100644
index 000000000..660bd9354
--- /dev/null
+++ b/libnd4j/include/ops/declarable/helpers/cpu/hamming.cpp
@@ -0,0 +1,72 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+
+#include <ops/declarable/helpers/helpers.h>
+#include <ops/declarable/helpers/hamming.h>
+
+namespace nd4j {
+    namespace ops {
+        namespace helpers {
+            template <typename X, typename Z>
+            static void _hamming(NDArray &x, NDArray &y, NDArray &z) {
+                auto xEws = x.ews();
+                auto yEws = y.ews();
+
+                auto xBuffer = x.bufferAsT<X>();
+                auto yBuffer = y.bufferAsT<X>();
+
+                Nd4jLong distance = 0;
+
+                if (xEws == 1 && yEws == 1 && x.ordering() == y.ordering()) {
+                    PRAGMA_OMP_PARALLEL_FOR_SIMD_REDUCTION(+:distance)
+                    for (Nd4jLong e = 0; e < x.lengthOf(); e++) {
+                        auto _x = static_cast<unsigned long long>(xBuffer[e]);
+                        auto _y = static_cast<unsigned long long>(yBuffer[e]);
+
+                        distance += __builtin_popcountll(_x ^ _y);
+                    }
+
+                } else if (xEws > 1 && yEws > 1 && x.ordering() == y.ordering()) {
+                    PRAGMA_OMP_PARALLEL_FOR_SIMD_REDUCTION(+:distance)
+                    for (Nd4jLong e = 0; e < x.lengthOf(); e++) {
+                        auto _x = static_cast<unsigned long long>(xBuffer[e * xEws]);
+                        auto _y = static_cast<unsigned long long>(yBuffer[e * yEws]);
+
+                        distance += __builtin_popcountll(_x ^ _y);
+                    }
+                } else {
+                    PRAGMA_OMP_PARALLEL_FOR_SIMD_REDUCTION(+:distance)
+                    for (Nd4jLong e = 0; e < x.lengthOf(); e++) {
+                        auto _x = static_cast<unsigned long long>(x.e<Nd4jLong>(e));
+                        auto _y = static_cast<unsigned long long>(y.e<Nd4jLong>(e));
+
+                        distance += __builtin_popcountll(_x ^ _y);
+                    }
+                }
+
+                z.p(0, distance);
+            }
+
+            void hamming(LaunchContext *context, NDArray &x, NDArray &y, NDArray &output) {
+                BUILD_DOUBLE_SELECTOR(x.dataType(), output.dataType(), _hamming, (x, y, output), INTEGER_TYPES, INDEXING_TYPES);
+            }
+        }
+    }
+}
\ No newline at end of file
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/hamming.cu b/libnd4j/include/ops/declarable/helpers/cuda/hamming.cu
new file mode 100644
index 000000000..3bc30e373
--- /dev/null
+++ b/libnd4j/include/ops/declarable/helpers/cuda/hamming.cu
@@ -0,0 +1,95 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+
+#include <ops/declarable/helpers/helpers.h>
+#include <ops/declarable/helpers/hamming.h>
+
+namespace nd4j {
+    namespace ops {
+        namespace helpers {
+            template <typename X, typename Z>
+            static _CUDA_G void _hammingKernel(void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, void *vz, void *reductionBuffer, Nd4jLong length) {
+                auto x = reinterpret_cast<X*>(vx);
+                auto y = reinterpret_cast<X*>(vy);
+                auto z = reinterpret_cast<Z*>(vz);
+
+                __shared__ Nd4jLong *shared;
+
+                if (threadIdx.x == 0) {
+                    extern __shared__ unsigned char shmem[];
+                    shared = reinterpret_cast<Nd4jLong*>(shmem);
+                }
+                __syncthreads();
+
+                // we want to nullify temporary memory before accumulating intermediate results
+                shared[threadIdx.x] = 0;
+
+                auto tid = threadIdx.x + blockIdx.x * blockDim.x;
+                for (Nd4jLong e = tid; e < length; e += blockDim.x * gridDim.x) {
+                    auto _x = static_cast<unsigned long long>(x[shape::getIndexOffset(e, xShapeInfo, length)]);
+                    auto _y = static_cast<unsigned long long>(y[shape::getIndexOffset(e, yShapeInfo, length)]);
+
+                    // we save intermediate result into shared memory
+                    shared[threadIdx.x] += __popcll(_x ^ _y);
+                }
+                __syncthreads();
+
+                // now we accumulate values
+                auto numItems = nd4j::math::nd4j_min<Nd4jLong>(blockDim.x, length);
+                auto floorPow2 = numItems;
+                if (floorPow2 & (floorPow2 - 1)) {
+
+                    while (floorPow2 & (floorPow2 - 1))
+                        floorPow2 &= floorPow2 - 1;
+
+                    if (threadIdx.x >= floorPow2)
+                        shared[threadIdx.x - floorPow2] = shared[threadIdx.x - floorPow2] + shared[threadIdx.x];
+
+                    __syncthreads();
+                }
+                __syncthreads();
+
+                for (Nd4jLong activeThreads = floorPow2 >> 1; activeThreads; activeThreads >>= 1) {
+                    if (threadIdx.x < activeThreads && threadIdx.x + activeThreads < numItems)
+                        shared[threadIdx.x] = shared[threadIdx.x] + shared[threadIdx.x + activeThreads];
+
+                    __syncthreads();
+                }
+                __syncthreads();
+
+                // FIXME: do we really want atomicAdd on global memory here
+                // and store them to output
+                if (threadIdx.x == 0 && shared[0] > 0)
+                    nd4j::math::atomics::nd4j_atomicAdd<Z>(&z[0], static_cast<Z>(shared[threadIdx.x]));
+            }
+
+            template <typename X, typename Z>
+            static void _hamming(LaunchContext *context, NDArray &x, NDArray &y, NDArray &z) {
+                _hammingKernel<X, Z><<<256, 256, 256 * sizeof(Nd4jLong) + 256, *context->getCudaStream()>>>(x.specialBuffer(), x.specialShapeInfo(), y.specialBuffer(), y.specialShapeInfo(), z.specialBuffer(), nullptr, x.lengthOf());
+            }
+
+            void hamming(LaunchContext *context, NDArray &x, NDArray &y, NDArray &output) {
+                NDArray::prepareSpecialUse({&output}, {&x, &y});
+                BUILD_DOUBLE_SELECTOR(x.dataType(), output.dataType(), _hamming, (context, x, y, output), INTEGER_TYPES, INDEXING_TYPES);
+                NDArray::registerSpecialUse({&output}, {&x, &y});
+            }
+        }
+    }
+}
\ No newline at end of file
diff --git a/libnd4j/include/ops/declarable/helpers/hamming.h b/libnd4j/include/ops/declarable/helpers/hamming.h
new file mode 100644
index 000000000..3ca3a6933
--- /dev/null
+++ b/libnd4j/include/ops/declarable/helpers/hamming.h
@@ -0,0 +1,32 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+
+#ifndef SD_HAMMING_H
+#define SD_HAMMING_H
+
+namespace nd4j {
+    namespace ops {
+        namespace helpers {
+            void hamming(LaunchContext *context, NDArray &x, NDArray &y, NDArray &output);
+        }
+    }
+}
+
+#endif //DEV_TESTS_HAMMING_H
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
index a2772a734..adbff7f83 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests16.cpp
@@ -117,4 +117,20 @@ TEST_F(DeclarableOpsTests16, test_svd_1) {
     auto status = op.execute({&x}, {&z}, {}, {0, 0, 16}, {});
 
     ASSERT_EQ(Status::OK(), status);
+}
+
+TEST_F(DeclarableOpsTests16, test_hamming_distance_1) {
+    auto x = NDArrayFactory::create<Nd4jLong>({37, 37, 37});
+    auto y = NDArrayFactory::create<Nd4jLong>({8723, 8723, 8723});
+    auto e = NDArrayFactory::create<Nd4jLong>(18);
+
+    nd4j::ops::bits_hamming_distance op;
+    auto result = op.execute({&x, &y}, {}, {});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    auto z = result->at(0);
+
+    ASSERT_EQ(e, *z);
+
+    delete result;
 }
\ No newline at end of file

From 70af8c2afc646a8ef1780ffb43bdf06f978a9fb9 Mon Sep 17 00:00:00 2001
From: Yurii Shyrma <yurii@skymind.io>
Date: Wed, 28 Aug 2019 18:27:08 +0300
Subject: [PATCH 36/56] Shyrma svd (#191)

* - add one additional test for svd

* - provide float argument in eye op to be a type of output array

Signed-off-by: Yurii <yurii@skymind.io>

* - add cuda capability check to mmulHelper

Signed-off-by: Yurii <yurii@skymind.io>

* - make use another method for divice id evaluation

Signed-off-by: Yurii <yurii@skymind.io>

* Eye data type as T argument

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../include/helpers/cuda_off/MmulHelper.cu    | 11 +++++-----
 .../ops/declarable/generic/transforms/eye.cpp | 17 +++++++--------
 .../ops/declarable/headers/transforms.h       | 21 +++++++++++--------
 .../layers_tests/CudaBasicsTests2.cu          |  1 +
 .../layers_tests/DeclarableOpsTests3.cpp      | 20 ++++++++++++++----
 .../layers_tests/DeclarableOpsTests5.cpp      |  8 +++----
 .../nd4j/linalg/api/ops/impl/shape/Eye.java   |  5 +++++
 7 files changed, 52 insertions(+), 31 deletions(-)

diff --git a/libnd4j/include/helpers/cuda_off/MmulHelper.cu b/libnd4j/include/helpers/cuda_off/MmulHelper.cu
index 19e0d5baf..32394f705 100644
--- a/libnd4j/include/helpers/cuda_off/MmulHelper.cu
+++ b/libnd4j/include/helpers/cuda_off/MmulHelper.cu
@@ -218,6 +218,9 @@ NDArray* MmulHelper::mmulMxM(const NDArray* A, const NDArray* B, NDArray* C, dou
 
     const bool AB(aType == bType), AC(aType == cType), ABC(AB && AC);
 
+    const int deviceId = AffinityManager::currentDeviceId();
+    const int major = Environment::getInstance()->capabilities()[deviceId].first();
+
     NDArray::prepareSpecialUse({pC}, {pA, pB});
 
     // choose appropriate cuda gemm api depending on data types
@@ -228,20 +231,18 @@ NDArray* MmulHelper::mmulMxM(const NDArray* A, const NDArray* B, NDArray* C, dou
         float alphaF(alpha), betaF(beta);
         status = cublasSgemm(*handle, transAblas, transBblas, M, N, K, &alphaF, (float*)pA->getSpecialBuffer(), lda, (float*)pB->getSpecialBuffer(), ldb, &betaF, (float*)pC->getSpecialBuffer(), ldc);
     }
-#if __CUDA_ARCH__ >= 530
-    else if(ABC && aType == DataType::HALF) {
+    else if(ABC && aType == DataType::HALF && major >= 6) {
         float16 alphaH(alpha), betaH(beta);
         status = cublasHgemm(*handle, transAblas, transBblas, M, N, K, &alphaH.data, (__half*)pA->getSpecialBuffer(), lda, (__half*)pB->getSpecialBuffer(), ldb, &betaH.data, (__half*)pC->getSpecialBuffer(), ldc);
     }
-    else if(AB && aType == DataType::INT8 && cType == DataType::FLOAT32) {
+    else if(AB && aType == DataType::INT8 && cType == DataType::FLOAT32 && major >= 6) {
            float alphaF(alpha), betaF(beta);
            status = cublasSgemmEx(*handle, transAblas, transBblas, M, N, K, &alphaF, pA->getSpecialBuffer(), CUDA_R_8I, lda, pB->getSpecialBuffer(), CUDA_R_8I, ldb, &betaF, pC->getSpecialBuffer(), CUDA_R_32F, ldc);
     }
-    else if(AB && aType == DataType::HALF && cType == DataType::FLOAT32) {
+    else if(AB && aType == DataType::HALF && cType == DataType::FLOAT32  && major >= 6) {
         float alphaF(alpha), betaF(beta);
         status = cublasSgemmEx(*handle, transAblas, transBblas, M, N, K, &alphaF, pA->getSpecialBuffer(), CUDA_R_16F, lda, pB->getSpecialBuffer(), CUDA_R_16F, ldb, &betaF, pC->getSpecialBuffer(), CUDA_R_32F, ldc);
     }
-#endif
     else {
         dim3 threadsPerBlock(N, M);
         dim3 blocksPerGrid(1, 1);
diff --git a/libnd4j/include/ops/declarable/generic/transforms/eye.cpp b/libnd4j/include/ops/declarable/generic/transforms/eye.cpp
index ef1d8e1e5..c5f3dbff6 100644
--- a/libnd4j/include/ops/declarable/generic/transforms/eye.cpp
+++ b/libnd4j/include/ops/declarable/generic/transforms/eye.cpp
@@ -27,7 +27,7 @@ namespace nd4j {
 namespace ops {
 
 
-    CUSTOM_OP_IMPL(eye, -2, 1, false, 0, -2) {
+    CUSTOM_OP_IMPL(eye, -2, 1, false, -2, -2) {
 
         helpers::eye(block.launchContext(), *OUTPUT_VARIABLE(0));
 
@@ -44,8 +44,7 @@ namespace ops {
 
         std::vector<int> params;
 
-        // FIX ME: original has a dtype param - so should be used here instead. e.g. (DataType) INT_ARG(0);
-        nd4j::DataType dtype = nd4j::DataType::FLOAT32;
+        nd4j::DataType dtype = block.getTArguments()->empty() ? nd4j::DataType::FLOAT32 : nd4j::DataTypeUtils::fromInt(T_ARG(0));
 
         if(block.width() == 0) {
             params = *block.getIArguments();
@@ -54,27 +53,27 @@ namespace ops {
             for (int i = 0; i < block.width(); i++) {
                 auto input = INPUT_VARIABLE(i);
                 REQUIRE_TRUE(input->rankOf() == 1, 0, "Inputs to eye should be 1D");
-                for (int e = 0; e < input->lengthOf(); e++) {
+
+                for (int e = 0; e < input->lengthOf(); e++)
                     params.emplace_back(input->e<int>(e));
-                }
             }
         }
 
 
-        REQUIRE_TRUE(params.size() > 0, 0, "Size not provided for eye op.");
+        REQUIRE_TRUE(params.size() > 0, 0, "Size is not provided for eye op.");
 
         const bool ordered = (params[0] == -99 || params[0] == -102); // -99 :'c', -102 : 'f'
         if (!ordered)
             params.insert(params.begin(), -99);
 
-        REQUIRE_TRUE(params.size() > 1, 0, "Size not provided for eye op.");
+        REQUIRE_TRUE(params.size() > 1, 0, "Size is not provided for eye op.");
 
         Nd4jLong* outShapeInfo(nullptr);
 
         const int size = params.size();
 
         switch(size) {
-        
+
             case 2:
                 ALLOCATE(outShapeInfo, block.getWorkspace(), shape::shapeInfoLength(2), Nd4jLong);
                 outShapeInfo[0] = 2;
@@ -99,7 +98,7 @@ namespace ops {
                     outShapeInfo[i] = params[i+2];
                 break;
         }
-        
+
         shape::updateStrides(outShapeInfo, static_cast<char>(-params[0]));
         auto result = ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(outShapeInfo, dtype));
         RELEASE(outShapeInfo, block.getWorkspace());
diff --git a/libnd4j/include/ops/declarable/headers/transforms.h b/libnd4j/include/ops/declarable/headers/transforms.h
index 75715f78e..b6fd57112 100644
--- a/libnd4j/include/ops/declarable/headers/transforms.h
+++ b/libnd4j/include/ops/declarable/headers/transforms.h
@@ -111,18 +111,21 @@ namespace nd4j {
 
         /**
          * creates identity 2D matrix or batch of identical 2D identity matrices
-         * 
+         *
          * Input array:
          * provide some array - in any case operation simply neglects it
-         * 
+         *
+         * Input float argument (if passed):
+         * TArgs[0] - type of elements of output array, default value is 5 (float)
+         *
          * Input integer arguments:
          * IArgs[0]       - order of output identity matrix, 99 -> 'c'-order, 102 -> 'f'-order
          * IArgs[1]       - the number of rows in output inner-most 2D identity matrix
          * IArgs[2]       - optional, the number of columns in output inner-most 2D identity matrix, if this argument is not provided then it is taken to be equal to number of rows
-         * IArgs[3,4,...] - optional, shape of batch, output matrix will have leading batch dimensions of this shape         
+         * IArgs[3,4,...] - optional, shape of batch, output matrix will have leading batch dimensions of this shape
          */
         #if NOT_EXCLUDED(OP_eye)
-        DECLARE_CUSTOM_OP(eye, -2, 1, false, 0, 2);
+        DECLARE_CUSTOM_OP(eye, -2, 1, false, -2, 2);
         #endif
 
         #if NOT_EXCLUDED(OP_gather_nd)
@@ -143,10 +146,10 @@ namespace nd4j {
 
         /**
          * clip a list of given tensors with given average norm when needed
-         * 
+         *
          * Input:
          *    a list of tensors (at least one)
-         * 
+         *
          * Input floating point argument:
          *    clip_norm - a value that used as threshold value and norm to be used
          *
@@ -182,12 +185,12 @@ namespace nd4j {
 
         /**
          * returns histogram (as 1D array) with fixed bins width
-         * 
+         *
          * Input arrays:
-         * - input array with elements to be binned into output histogram 
+         * - input array with elements to be binned into output histogram
          * - range array with first element being bottom limit and second element being top limit of histogram,
              please note that input_value <= range[0] will be mapped to histogram[0], input_value >= range[1] will be mapped to histogram[-1]
-         * 
+         *
          * Input integer arguments:
          *    nbins (optional) - number of histogram bins, default value is 100
          */
diff --git a/libnd4j/tests_cpu/layers_tests/CudaBasicsTests2.cu b/libnd4j/tests_cpu/layers_tests/CudaBasicsTests2.cu
index 19f107ea4..027dcdd42 100644
--- a/libnd4j/tests_cpu/layers_tests/CudaBasicsTests2.cu
+++ b/libnd4j/tests_cpu/layers_tests/CudaBasicsTests2.cu
@@ -272,6 +272,7 @@ TEST_F(CudaBasicsTests2, mmulMxM_12) {
 	NDArray exp('f', {M,N}, {-16., -22., -23., -25., 30., -12., -38., -70., 20., 16., 18., 18., 22., -8., -28., -52.}, nd4j::DataType::FLOAT32);
 
 	nd4j::MmulHelper::mmul(&a, &b, &c, 1., 0.);
+	// c.printBuffer();
 
 	ASSERT_TRUE(c.equalsTo(&exp));
 }
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests3.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests3.cpp
index d786a68cb..3b4ff6cd0 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests3.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests3.cpp
@@ -2757,10 +2757,18 @@ TEST_F(DeclarableOpsTests3, svd_test10) {
 ///////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests3, svd_test11) {
 
-    auto x = NDArrayFactory::create<double>('c', {3,3}, {1.,2.,3.,4.,5.,6.,7.,8.,9.});
-    auto expS = NDArrayFactory::create<double>('c', {3});
-    auto expU = NDArrayFactory::create<double>('c', {3,3});
-    auto expV = NDArrayFactory::create<double>('c', {3,3});
+    NDArray x('c', {2,2,3,3}, {0.7788, 0.8012, 0.7244, 0.2309, 0.7271, 0.1804, 0.5056, 0.8925, 0.5461, 0.9234, 0.0856, 0.7938, 0.6591, 0.5555,
+                                0.1596, 0.3087, 0.1548, 0.4695, 0.7788, 0.8012, 0.7244, 0.2309, 0.7271, 0.1804, 0.5056, 0.8925, -0.5461, 0.9234,
+                                0.0856, -0.7938, 0.6591, 0.5555, 0.1500, 0.3087, 0.1548, 0.4695});
+    NDArray expS('c', {2,2,3}, {1.89671, 0.37095, 0.05525,1.51296, 0.52741, 0.17622, 1.69095, 0.90438, 0.24688,1.33551, 0.87475, 0.21571});
+    NDArray expU('c', {2,2,3,3}, {6.9205e-01,  6.0147e-01, -3.9914e-01, 3.8423e-01, -7.7503e-01, -5.0170e-01, 6.1110e-01, -1.9384e-01,  7.6746e-01,
+                                7.8967e-01,  4.5442e-01, -4.1222e-01, 4.9381e-01, -8.6948e-01, -1.2540e-02, 3.6412e-01,  1.9366e-01,  9.1100e-01,
+                                7.1764e-01,  5.9844e-01,  3.5617e-01, 4.4477e-01, -3.1000e-04, -8.9564e-01, 5.3588e-01, -8.0116e-01,  2.6639e-01,
+                                8.7050e-01, -4.2088e-01, -2.5513e-01, 4.8622e-01,  6.5499e-01,  5.7843e-01, 7.6340e-02,  6.2757e-01, -7.7481e-01});
+    NDArray expV('c', {2,2,3,3}, {0.49383,  0.51614, -0.69981, 0.72718, -0.68641,  0.00688, 0.4768 ,  0.51228,  0.7143 , 0.77137, -0.17763,
+                                  -0.6111 , 0.26324, -0.7852 ,  0.56051, 0.57939,  0.59322,  0.55892, 0.55149,  0.06737,  0.83146, 0.81413,
+                                  -0.26072, -0.51887, 0.18182,  0.96306, -0.19863, 0.85948,  0.2707 , -0.4336 , 0.26688,  0.48582,  0.83232,
+                                  -0.43596,  0.83108, -0.34531});
 
     nd4j::ops::svd op;
     auto results = op.execute({&x}, {}, {0, 1, 16});
@@ -2775,6 +2783,10 @@ TEST_F(DeclarableOpsTests3, svd_test11) {
     ASSERT_TRUE(expU.isSameShape(u));
     ASSERT_TRUE(expV.isSameShape(v));
 
+    ASSERT_TRUE(expS.equalsTo(s));
+    ASSERT_TRUE(expU.equalsTo(u));
+    ASSERT_TRUE(expV.equalsTo(v));
+
     delete results;
 }
 
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
index 2ff2fad58..3af53bad0 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests5.cpp
@@ -639,10 +639,10 @@ TEST_F(DeclarableOpsTests5, eye_test2) {
 //////////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests5, eye_test3) {
 
-    auto expected = NDArrayFactory::create<float>('c', {2, 3, 4}, {1,  0,  0,  0, 0,  1,  0,  0, 0,  0,  1,  0, 1,  0,  0,  0, 0,  1,  0,  0, 0,  0,  1,  0});
+    auto expected = NDArrayFactory::create<int>('c', {2, 3, 4}, {1,  0,  0,  0, 0,  1,  0,  0, 0,  0,  1,  0, 1,  0,  0,  0, 0,  1,  0,  0, 0,  0,  1,  0});
 
     nd4j::ops::eye op;
-    auto results = op.execute({}, {}, {-99, 3, 4, 2});
+    auto results = op.execute({}, {9 /*int*/}, {-99, 3, 4, 2});
     auto output = results->at(0);
      // output->printIndexedBuffer("Output eye");
 
@@ -656,10 +656,10 @@ TEST_F(DeclarableOpsTests5, eye_test3) {
 //////////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests5, eye_test4) {
 
-    auto expected = NDArrayFactory::create<float>('c', {2, 2, 3, 4}, {1.,  0.,  0.,  0., 0.,  1.,  0.,  0., 0.,  0.,  1.,  0., 1.,  0.,  0.,  0., 0.,  1.,  0.,  0., 0.,  0.,  1.,  0., 1.,  0.,  0.,  0., 0.,  1.,  0.,  0., 0.,  0.,  1.,  0., 1.,  0.,  0.,  0., 0.,  1.,  0.,  0., 0.,  0.,  1.,  0.});
+    auto expected = NDArrayFactory::create<double>('c', {2, 2, 3, 4}, {1.,  0.,  0.,  0., 0.,  1.,  0.,  0., 0.,  0.,  1.,  0., 1.,  0.,  0.,  0., 0.,  1.,  0.,  0., 0.,  0.,  1.,  0., 1.,  0.,  0.,  0., 0.,  1.,  0.,  0., 0.,  0.,  1.,  0., 1.,  0.,  0.,  0., 0.,  1.,  0.,  0., 0.,  0.,  1.,  0.});
 
     nd4j::ops::eye op;
-    auto results = op.execute({}, {}, {-99, 3, 4, 2, 2});
+    auto results = op.execute({}, {6/*double*/}, {-99, 3, 4, 2, 2});
     auto output = results->at(0);
 
     ASSERT_EQ(Status::OK(), results->status());
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Eye.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Eye.java
index f01e4a448..c4747a371 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Eye.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Eye.java
@@ -98,6 +98,9 @@ public class Eye extends DynamicCustomOp {
     }
 
     protected void addArgs() {
+        iArguments.clear();
+        tArguments.clear();
+
         addIArgument(numRows);
         addIArgument(numCols);
         if(batchDimension != null) {
@@ -105,6 +108,8 @@ public class Eye extends DynamicCustomOp {
                 addIArgument(dim);
             }
         }
+
+        addTArgument((double) dataType.toInt());
     }
 
     @Override

From f40bdcf88511818a36d8c876c1e6f88e83e7a6cb Mon Sep 17 00:00:00 2001
From: Ryan Nett <rnett@skymind.io>
Date: Wed, 28 Aug 2019 20:05:01 -0700
Subject: [PATCH 37/56] Dl4j LSTM and Dropout CuDNN fallback and options (#152)

* add fallback for Conv layer activation

Signed-off-by: Ryan Nett <rnett@skymind.io>

* add fallback and config option for LSTM layers

Signed-off-by: Ryan Nett <rnett@skymind.io>

* add fallback option and setting for dropout

Signed-off-by: Ryan Nett <rnett@skymind.io>

* fix comments and error messages

Signed-off-by: Ryan Nett <rnett@skymind.io>

* move helper fail count to layer instance

Signed-off-by: Ryan Nett <rnett@skymind.io>

* ignore helperCountFail for equals and json

Signed-off-by: Ryan Nett <rnett@skymind.io>

* typo fix (MLK -> MKL)

Signed-off-by: Ryan Nett <rnett@skymind.io>

* add MKLDNN to error messages

Signed-off-by: Ryan Nett <rnett@skymind.io>

* add helperAllowFallback to builders, deprecate cudnnAllowFallback

Signed-off-by: Ryan Nett <rnett@skymind.io>
---
 .../GravesBidirectionalLSTMTest.java          |  4 +-
 .../nn/conf/dropout/Dropout.java              | 84 +++++++++++++++++--
 .../nn/conf/layers/AbstractLSTM.java          | 24 +++++-
 .../nn/conf/layers/BatchNormalization.java    | 17 +++-
 .../nn/conf/layers/ConvolutionLayer.java      | 17 +++-
 .../conf/layers/GravesBidirectionalLSTM.java  | 22 +++++
 .../layers/LocalResponseNormalization.java    | 17 +++-
 .../nn/conf/layers/Subsampling3DLayer.java    | 17 +++-
 .../nn/conf/layers/SubsamplingLayer.java      | 21 ++++-
 .../layers/convolution/ConvolutionLayer.java  | 28 ++++++-
 .../nn/layers/mkldnn/BaseMKLDNNHelper.java    |  2 +-
 .../layers/recurrent/BaseRecurrentLayer.java  |  2 +
 .../recurrent/GravesBidirectionalLSTM.java    | 17 ++--
 .../nn/layers/recurrent/GravesLSTM.java       |  8 +-
 .../nn/layers/recurrent/LSTM.java             | 11 ++-
 .../nn/layers/recurrent/LSTMHelpers.java      | 65 +++++++++++---
 docs/deeplearning4j/templates/benchmark.md    |  2 +-
 17 files changed, 302 insertions(+), 56 deletions(-)

diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/layers/recurrent/GravesBidirectionalLSTMTest.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/layers/recurrent/GravesBidirectionalLSTMTest.java
index 297067862..751b6f6bf 100644
--- a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/layers/recurrent/GravesBidirectionalLSTMTest.java
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/layers/recurrent/GravesBidirectionalLSTMTest.java
@@ -186,7 +186,7 @@ public class GravesBidirectionalLSTMTest extends BaseDL4JTest {
                         lstm.getParam(GravesBidirectionalLSTMParamInitializer.INPUT_WEIGHT_KEY_FORWARDS),
                         lstm.getParam(GravesBidirectionalLSTMParamInitializer.BIAS_KEY_FORWARDS), false, null, null,
                         false, true, GravesBidirectionalLSTMParamInitializer.INPUT_WEIGHT_KEY_FORWARDS, null, true,
-                        null, CacheMode.NONE, LayerWorkspaceMgr.noWorkspaces()).fwdPassOutput;
+                        null, CacheMode.NONE, LayerWorkspaceMgr.noWorkspaces(), true).fwdPassOutput;
 
         final INDArray[] fwdPassTrue = LSTMHelpers.activateHelper(lstm, lstm.conf(), new ActivationSigmoid(),
                         lstm.input(),
@@ -194,7 +194,7 @@ public class GravesBidirectionalLSTMTest extends BaseDL4JTest {
                         lstm.getParam(GravesBidirectionalLSTMParamInitializer.INPUT_WEIGHT_KEY_FORWARDS),
                         lstm.getParam(GravesBidirectionalLSTMParamInitializer.BIAS_KEY_FORWARDS), false, null, null,
                         true, true, GravesBidirectionalLSTMParamInitializer.INPUT_WEIGHT_KEY_FORWARDS, null, true, null,
-                        CacheMode.NONE, LayerWorkspaceMgr.noWorkspaces()).fwdPassOutputAsArrays;
+                        CacheMode.NONE, LayerWorkspaceMgr.noWorkspaces(), true).fwdPassOutputAsArrays;
 
         //I have no idea what the heck this does --Ben
         for (int i = 0; i < timeSeriesLength; i++) {
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/dropout/Dropout.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/dropout/Dropout.java
index f9af153ad..ee8bbdc64 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/dropout/Dropout.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/dropout/Dropout.java
@@ -18,6 +18,8 @@ package org.deeplearning4j.nn.conf.dropout;
 
 import lombok.Data;
 import lombok.EqualsAndHashCode;
+import lombok.Getter;
+import lombok.Setter;
 import lombok.extern.slf4j.Slf4j;
 import org.deeplearning4j.nn.workspace.ArrayType;
 import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
@@ -26,11 +28,11 @@ import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.api.ops.impl.transforms.pairwise.arithmetic.MulOp;
 import org.nd4j.linalg.api.ops.random.impl.DropOutInverted;
+import org.nd4j.linalg.exception.ND4JOpProfilerException;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.schedule.ISchedule;
 import org.nd4j.shade.jackson.annotation.JsonIgnoreProperties;
 import org.nd4j.shade.jackson.annotation.JsonProperty;
-import org.nd4j.util.OneTimeLogger;
 
 /**
  * Implements standard (inverted) dropout.<br>
@@ -64,17 +66,29 @@ import org.nd4j.util.OneTimeLogger;
  * @author Alex Black
  */
 @Data
-@JsonIgnoreProperties({"mask", "helper"})
-@EqualsAndHashCode(exclude = {"mask", "helper"})
+@JsonIgnoreProperties({"mask", "helper", "helperCountFail"})
+@EqualsAndHashCode(exclude = {"mask", "helper", "helperCountFail"})
 @Slf4j
 public class Dropout implements IDropout {
 
+    /**
+     * When using CuDNN and an error is encountered, should fallback to the non-CuDNN implementatation be allowed?
+     * If set to false, an exception in CuDNN will be propagated back to the user. If false, the built-in
+     * (non-CuDNN) implementation for LSTM/GravesLSTM will be used
+     *
+     */
+    @Getter
+    @Setter
+    protected boolean helperAllowFallback = true;
+
     private double p;
     private ISchedule pSchedule;
     private transient INDArray mask;
     private transient DropoutHelper helper;
     private boolean initializedHelper = false;
 
+    private int helperCountFail = 0;
+
     /**
      * @param activationRetainProbability Probability of retaining an activation - see {@link Dropout} javadoc
      */
@@ -96,6 +110,18 @@ public class Dropout implements IDropout {
         this(Double.NaN, activationRetainProbabilitySchedule);
     }
 
+    /**
+     * When using a helper (CuDNN or MKLDNN in some cases) and an error is encountered, should fallback to the non-helper implementation be allowed?
+     * If set to false, an exception in the helper will be propagated back to the user. If false, the built-in
+     * (non-helper) implementation for Dropout will be used
+     *
+     * @param allowFallback Whether fallback to non-helper implementation should be used
+     */
+    public Dropout helperAllowFallback(boolean allowFallback) {
+        this.setHelperAllowFallback(allowFallback);
+        return this;
+    }
+
     protected Dropout(@JsonProperty("p") double activationRetainProbability, @JsonProperty("pSchedule") ISchedule activationRetainProbabilitySchedule) {
         this.p = activationRetainProbability;
         this.pSchedule = activationRetainProbabilitySchedule;
@@ -141,9 +167,29 @@ public class Dropout implements IDropout {
             initializeHelper(output.dataType());
         }
 
-        if(helper != null){
-            helper.applyDropout(inputActivations, output, p);
-            return output;
+        if(helper != null && (helperCountFail == 0 || !isHelperAllowFallback())){
+            boolean helperWorked = false;
+            try {
+                helper.applyDropout(inputActivations, output, p);
+                helperWorked = true;
+            }catch (ND4JOpProfilerException e){
+                throw e;    //NaN panic etc for debugging
+            } catch (Exception e){
+                if(e.getMessage().contains("Failed to allocate")){
+                    //This is a memory exception - don't fallback to built-in implementation
+                    throw e;
+                }
+
+                if(isHelperAllowFallback()){
+                    helperCountFail++;
+                    log.warn("CuDNN execution failed - falling back on built-in implementation",e);
+                } else {
+                    throw new RuntimeException("Error during Dropout CuDNN helper forward pass - helperAllowFallback() is set to false", e);
+                }
+            }
+
+            if(helperWorked)
+                return output;
         }
 
         INDArray inputCast = inputActivations;
@@ -159,9 +205,29 @@ public class Dropout implements IDropout {
 
     @Override
     public INDArray backprop(INDArray gradAtOutput, INDArray gradAtInput, int iteration, int epoch) {
-        if(helper != null){
-            helper.backprop(gradAtOutput, gradAtInput);
-            return gradAtInput;
+        if(helper != null && (helperCountFail == 0 || !isHelperAllowFallback())){
+            boolean helperWorked = false;
+            try {
+                helper.backprop(gradAtOutput, gradAtInput);
+                helperWorked = true;
+            }catch (ND4JOpProfilerException e){
+                throw e;    //NaN panic etc for debugging
+            } catch (Exception e){
+                if(e.getMessage().contains("Failed to allocate")){
+                    //This is a memory exception - don't fallback to built-in implementation
+                    throw e;
+                }
+
+                if(isHelperAllowFallback()){
+                    helperCountFail++;
+                    log.warn("CuDNN execution failed - falling back on built-in implementation",e);
+                } else {
+                    throw new RuntimeException("Error during Dropout CuDNN helper backprop - helperAllowFallback() is set to false", e);
+                }
+            }
+
+            if(helperWorked)
+                return gradAtInput;
         }
 
         Preconditions.checkState(mask != null, "Cannot perform backprop: Dropout mask array is absent (already cleared?)");
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/AbstractLSTM.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/AbstractLSTM.java
index 82bda5647..b051c4b36 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/AbstractLSTM.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/AbstractLSTM.java
@@ -17,8 +17,6 @@
 package org.deeplearning4j.nn.conf.layers;
 
 import lombok.*;
-import org.deeplearning4j.nn.params.LSTMParamInitializer;
-import org.deeplearning4j.nn.weights.WeightInit;
 import org.nd4j.linalg.activations.Activation;
 import org.nd4j.linalg.activations.IActivation;
 import org.nd4j.linalg.activations.impl.ActivationSigmoid;
@@ -35,11 +33,13 @@ public abstract class AbstractLSTM extends BaseRecurrentLayer {
 
     protected double forgetGateBiasInit;
     protected IActivation gateActivationFn = new ActivationSigmoid();
+    protected boolean helperAllowFallback = true;
 
     protected AbstractLSTM(Builder builder) {
         super(builder);
         this.forgetGateBiasInit = builder.forgetGateBiasInit;
         this.gateActivationFn = builder.gateActivationFn;
+        this.helperAllowFallback = builder.helperAllowFallback;
     }
 
     @AllArgsConstructor
@@ -60,6 +60,14 @@ public abstract class AbstractLSTM extends BaseRecurrentLayer {
          */
         protected IActivation gateActivationFn = new ActivationSigmoid();
 
+        /**
+         * When using CuDNN and an error is encountered, should fallback to the non-CuDNN implementatation be allowed?
+         * If set to false, an exception in CuDNN will be propagated back to the user. If false, the built-in
+         * (non-CuDNN) implementation for LSTM/GravesLSTM will be used
+         *
+         */
+        protected boolean helperAllowFallback = true;
+
         /**
          * Set forget gate bias initalizations. Values in range 1-5 can potentially help with learning or longer-term
          * dependencies.
@@ -100,6 +108,18 @@ public abstract class AbstractLSTM extends BaseRecurrentLayer {
             return (T) this;
         }
 
+        /**
+         * When using a helper (CuDNN or MKLDNN in some cases) and an error is encountered, should fallback to the non-helper implementation be allowed?
+         * If set to false, an exception in the helper will be propagated back to the user. If false, the built-in
+         * (non-helper) implementation for LSTM/GravesLSTM will be used
+         *
+         * @param allowFallback Whether fallback to non-helper implementation should be used
+         */
+        public T helperAllowFallback(boolean allowFallback) {
+            this.setHelperAllowFallback(allowFallback);
+            return (T) this;
+        }
+
     }
 
 }
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/BatchNormalization.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/BatchNormalization.java
index 53c00acac..4c470fec5 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/BatchNormalization.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/BatchNormalization.java
@@ -428,16 +428,31 @@ public class BatchNormalization extends FeedForwardLayer {
 
         /**
          * When using CuDNN and an error is encountered, should fallback to the non-CuDNN implementatation be allowed?
-         * If set to false, an exception in CuDNN will be propagated back to the user. If false, the built-in
+         * If set to false, an exception in CuDNN will be propagated back to the user. If true, the built-in
          * (non-CuDNN) implementation for BatchNormalization will be used
          *
+         * @deprecated Use {@link #helperAllowFallback(boolean)}
+         *
          * @param allowFallback Whether fallback to non-CuDNN implementation should be used
          */
+        @Deprecated
         public Builder cudnnAllowFallback(boolean allowFallback) {
             this.setCudnnAllowFallback(allowFallback);
             return this;
         }
 
+        /**
+         * When using CuDNN or MKLDNN and an error is encountered, should fallback to the non-helper implementation be allowed?
+         * If set to false, an exception in the helper will be propagated back to the user. If true, the built-in
+         * (non-MKL/CuDNN) implementation for BatchNormalizationLayer will be used
+         *
+         * @param allowFallback Whether fallback to non-CuDNN implementation should be used
+         */
+        public Builder helperAllowFallback(boolean allowFallback) {
+            this.cudnnAllowFallback = allowFallback;
+            return this;
+        }
+
         /**
          * How should the moving average of variance be stored? Two different parameterizations are supported.
          * useLogStd(false): equivalent to 1.0.0-beta3 and earlier. The variance "parameter" is stored directly as
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/ConvolutionLayer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/ConvolutionLayer.java
index 3d2e35d24..4fdf1e9cc 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/ConvolutionLayer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/ConvolutionLayer.java
@@ -533,14 +533,29 @@ public class ConvolutionLayer extends FeedForwardLayer {
 
         /**
          * When using CuDNN and an error is encountered, should fallback to the non-CuDNN implementatation be allowed?
-         * If set to false, an exception in CuDNN will be propagated back to the user. If false, the built-in
+         * If set to false, an exception in CuDNN will be propagated back to the user. If true, the built-in
          * (non-CuDNN) implementation for ConvolutionLayer will be used
          *
+         * @deprecated Use {@link #helperAllowFallback(boolean)}
+         *
          * @param allowFallback Whether fallback to non-CuDNN implementation should be used
          */
+        @Deprecated
         public T cudnnAllowFallback(boolean allowFallback) {
             this.setCudnnAllowFallback(allowFallback);
             return (T) this;
         }
+
+        /**
+         * When using CuDNN or MKLDNN and an error is encountered, should fallback to the non-helper implementation be allowed?
+         * If set to false, an exception in the helper will be propagated back to the user. If true, the built-in
+         * (non-MKL/CuDNN) implementation for ConvolutionLayer will be used
+         *
+         * @param allowFallback Whether fallback to non-CuDNN implementation should be used
+         */
+        public T helperAllowFallback(boolean allowFallback) {
+            this.cudnnAllowFallback = allowFallback;
+            return (T) this;
+        }
     }
 }
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/GravesBidirectionalLSTM.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/GravesBidirectionalLSTM.java
index d7aa869a1..1a2a89a24 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/GravesBidirectionalLSTM.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/GravesBidirectionalLSTM.java
@@ -53,11 +53,13 @@ public class GravesBidirectionalLSTM extends BaseRecurrentLayer {
 
     private double forgetGateBiasInit;
     private IActivation gateActivationFn = new ActivationSigmoid();
+    protected boolean helperAllowFallback = true;
 
     private GravesBidirectionalLSTM(Builder builder) {
         super(builder);
         this.forgetGateBiasInit = builder.forgetGateBiasInit;
         this.gateActivationFn = builder.gateActivationFn;
+        this.helperAllowFallback = builder.helperAllowFallback;
 
         initializeConstraints(builder);
     }
@@ -123,6 +125,14 @@ public class GravesBidirectionalLSTM extends BaseRecurrentLayer {
          */
         private IActivation gateActivationFn = new ActivationSigmoid();
 
+        /**
+         * When using CuDNN and an error is encountered, should fallback to the non-CuDNN implementatation be allowed?
+         * If set to false, an exception in CuDNN will be propagated back to the user. If false, the built-in
+         * (non-CuDNN) implementation for GravesBidirectionalLSTM will be used
+         *
+         */
+        protected boolean helperAllowFallback = true;
+
         /**
          * Set forget gate bias initalizations. Values in range 1-5 can potentially help with learning or longer-term
          * dependencies.
@@ -163,6 +173,18 @@ public class GravesBidirectionalLSTM extends BaseRecurrentLayer {
             return this;
         }
 
+        /**
+         * When using a helper (CuDNN or MKLDNN in some cases) and an error is encountered, should fallback to the non-helper implementation be allowed?
+         * If set to false, an exception in the helper will be propagated back to the user. If false, the built-in
+         * (non-helper) implementation for GravesBidirectionalLSTM will be used
+         *
+         * @param allowFallback Whether fallback to non-helper implementation should be used
+         */
+        public Builder helperAllowFallback(boolean allowFallback) {
+            this.setHelperAllowFallback(allowFallback);
+            return (Builder) this;
+        }
+
         @SuppressWarnings("unchecked")
         public GravesBidirectionalLSTM build() {
             return new GravesBidirectionalLSTM(this);
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/LocalResponseNormalization.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/LocalResponseNormalization.java
index dfc2df9c8..b16703569 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/LocalResponseNormalization.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/LocalResponseNormalization.java
@@ -238,16 +238,31 @@ public class LocalResponseNormalization extends Layer {
 
         /**
          * When using CuDNN and an error is encountered, should fallback to the non-CuDNN implementatation be allowed?
-         * If set to false, an exception in CuDNN will be propagated back to the user. If false, the built-in
+         * If set to false, an exception in CuDNN will be propagated back to the user. If true, the built-in
          * (non-CuDNN) implementation for BatchNormalization will be used
          *
+         * @deprecated Use {@link #helperAllowFallback(boolean)}
+         *
          * @param allowFallback Whether fallback to non-CuDNN implementation should be used
          */
+        @Deprecated
         public Builder cudnnAllowFallback(boolean allowFallback) {
             this.setCudnnAllowFallback(allowFallback);
             return this;
         }
 
+        /**
+         * When using CuDNN or MKLDNN and an error is encountered, should fallback to the non-helper implementation be allowed?
+         * If set to false, an exception in the helper will be propagated back to the user. If true, the built-in
+         * (non-MKL/CuDNN) implementation for LocalResponseNormalizationLayer will be used
+         *
+         * @param allowFallback Whether fallback to non-CuDNN implementation should be used
+         */
+        public Builder helperAllowFallback(boolean allowFallback) {
+            this.cudnnAllowFallback = allowFallback;
+            return this;
+        }
+
         @Override
         public LocalResponseNormalization build() {
             return new LocalResponseNormalization(this);
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Subsampling3DLayer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Subsampling3DLayer.java
index 877e216da..0d0ccba9b 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Subsampling3DLayer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Subsampling3DLayer.java
@@ -455,15 +455,30 @@ public class Subsampling3DLayer extends NoParamLayer {
 
         /**
          * When using CuDNN and an error is encountered, should fallback to the non-CuDNN implementatation be allowed?
-         * If set to false, an exception in CuDNN will be propagated back to the user. If false, the built-in
+         * If set to false, an exception in CuDNN will be propagated back to the user. If true, the built-in
          * (non-CuDNN) implementation for ConvolutionLayer will be used
          *
+         * @deprecated Use {@link #helperAllowFallback(boolean)}
+         *
          * @param allowFallback Whether fallback to non-CuDNN implementation should be used
          */
+        @Deprecated
         public T cudnnAllowFallback(boolean allowFallback) {
             this.setCudnnAllowFallback(allowFallback);
             return (T) this;
         }
+
+        /**
+         * When using CuDNN or MKLDNN and an error is encountered, should fallback to the non-helper implementation be allowed?
+         * If set to false, an exception in the helper will be propagated back to the user. If true, the built-in
+         * (non-MKL/CuDNN) implementation for Subsampling3DLayer will be used
+         *
+         * @param allowFallback Whether fallback to non-CuDNN implementation should be used
+         */
+        public T helperAllowFallback(boolean allowFallback) {
+            this.cudnnAllowFallback = allowFallback;
+            return (T) this;
+        }
     }
 
 }
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/SubsamplingLayer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/SubsamplingLayer.java
index b2e4df6b8..c20526cf1 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/SubsamplingLayer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/SubsamplingLayer.java
@@ -480,17 +480,32 @@ public class SubsamplingLayer extends NoParamLayer {
         }
 
         /**
-         * When using CuDNN and an error is encountered, should fallback to the non-CuDNN implementatation be allowed?
-         * If set to false, an exception in CuDNN will be propagated back to the user. If false, the built-in
-         * (non-CuDNN) implementation for ConvolutionLayer will be used
+         * When using CuDNN or MKLDNN and an error is encountered, should fallback to the non-helper implementation be allowed?
+         * If set to false, an exception in the helper will be propagated back to the user. If true, the built-in
+         * (non-MKL/CuDNN) implementation for ConvolutionLayer will be used
+         *
+         * @deprecated Use {@link #helperAllowFallback(boolean)}
          *
          * @param allowFallback Whether fallback to non-CuDNN implementation should be used
          */
+        @Deprecated
         public T cudnnAllowFallback(boolean allowFallback) {
             this.cudnnAllowFallback = allowFallback;
             return (T) this;
         }
 
+        /**
+         * When using CuDNN or MKLDNN and an error is encountered, should fallback to the non-helper implementation be allowed?
+         * If set to false, an exception in the helper will be propagated back to the user. If true, the built-in
+         * (non-MKL/CuDNN) implementation for SubsamplingLayer will be used
+         *
+         * @param allowFallback Whether fallback to non-CuDNN implementation should be used
+         */
+        public T helperAllowFallback(boolean allowFallback) {
+            this.cudnnAllowFallback = allowFallback;
+            return (T) this;
+        }
+
         /**
          * When doing average pooling, should the padding values be included in the divisor or not?<br>
          * Not applicable for max and p-norm pooling.<br>
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/convolution/ConvolutionLayer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/convolution/ConvolutionLayer.java
index d6cab0273..75e265b4e 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/convolution/ConvolutionLayer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/convolution/ConvolutionLayer.java
@@ -378,7 +378,7 @@ public class ConvolutionLayer extends BaseLayer<org.deeplearning4j.nn.conf.layer
                         log.warn("CuDNN execution failed - falling back on built-in implementation",e);
                     }
                 } else {
-                    throw new RuntimeException(e);
+                    throw new RuntimeException("Error during ConvolutionLayer MKL/CuDNN helper forward pass - isCudnnAllowFallback() is set to false", e);
                 }
             }
             if (ret != null) {
@@ -453,8 +453,30 @@ public class ConvolutionLayer extends BaseLayer<org.deeplearning4j.nn.conf.layer
         //String afn = conf.getLayer().getActivationFunction();
         IActivation afn = layerConf().getActivationFn();
 
-        if (helper != null && Shape.strideDescendingCAscendingF(z)) {
-            INDArray ret = helper.activate(z, layerConf().getActivationFn(), training);
+        if (helper != null && Shape.strideDescendingCAscendingF(z) && (helperCountFail == 0 || !layerConf().isCudnnAllowFallback())) {
+            INDArray ret = null;
+            try {
+                ret = helper.activate(z, layerConf().getActivationFn(), training);
+            } catch (ND4JOpProfilerException e){
+                throw e;    //NaN panic etc for debugging
+            } catch (Exception e) {
+                if (e.getMessage() != null && e.getMessage().contains("Failed to allocate")) {
+                    //This is a memory exception - don't fallback to built-in implementation
+                    throw e;
+                }
+
+                if (layerConf().isCudnnAllowFallback()) {
+                    helperCountFail++;
+                    if (helper instanceof MKLDNNConvHelper) {
+                        log.warn("MKL-DNN execution failed - falling back on built-in implementation", e);
+                    } else {
+                        log.warn("CuDNN execution failed - falling back on built-in implementation", e);
+                    }
+                } else {
+                    throw new RuntimeException("Error during ConvolutionLayer MKL/CuDNN helper forward pass - isCudnnAllowFallback() is set to false", e);
+                }
+            }
+
             if (ret != null) {
                 return ret;
             }
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/mkldnn/BaseMKLDNNHelper.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/mkldnn/BaseMKLDNNHelper.java
index c7af77002..d3a63d79e 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/mkldnn/BaseMKLDNNHelper.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/mkldnn/BaseMKLDNNHelper.java
@@ -22,7 +22,7 @@ import java.lang.reflect.Method;
 import java.util.concurrent.atomic.AtomicBoolean;
 
 /**
- * Base class for MLK-DNN Helpers
+ * Base class for MKL-DNN Helpers
  * @author Alex Black
  */
 public class BaseMKLDNNHelper {
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/BaseRecurrentLayer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/BaseRecurrentLayer.java
index 1590adcc7..265946bd8 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/BaseRecurrentLayer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/BaseRecurrentLayer.java
@@ -41,6 +41,8 @@ public abstract class BaseRecurrentLayer<LayerConfT extends org.deeplearning4j.n
      */
     protected Map<String, INDArray> tBpttStateMap = new ConcurrentHashMap<>();
 
+    protected int helperCountFail = 0;
+
     public BaseRecurrentLayer(NeuralNetConfiguration conf, DataType dataType) {
         super(conf, dataType);
     }
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/GravesBidirectionalLSTM.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/GravesBidirectionalLSTM.java
index 78e15e167..6fc96dc80 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/GravesBidirectionalLSTM.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/GravesBidirectionalLSTM.java
@@ -17,7 +17,6 @@
 package org.deeplearning4j.nn.layers.recurrent;
 
 import lombok.extern.slf4j.Slf4j;
-import org.deeplearning4j.nn.api.Layer;
 import org.deeplearning4j.nn.api.MaskState;
 import org.deeplearning4j.nn.conf.CacheMode;
 import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
@@ -90,7 +89,8 @@ public class GravesBidirectionalLSTM
 
         final FwdPassReturn fwdPass = activateHelperDirectional(true, null, null, true, true, workspaceMgr);
 
-        final Pair<Gradient, INDArray> forwardsGradient = LSTMHelpers.backpropGradientHelper(this.conf,
+        final Pair<Gradient, INDArray> forwardsGradient = LSTMHelpers.backpropGradientHelper(this,
+                        this.conf,
                         this.layerConf().getGateActivationFn(), this.input,
                         getParam(GravesBidirectionalLSTMParamInitializer.RECURRENT_WEIGHT_KEY_FORWARDS),
                         getParam(GravesBidirectionalLSTMParamInitializer.INPUT_WEIGHT_KEY_FORWARDS), epsilon,
@@ -98,13 +98,14 @@ public class GravesBidirectionalLSTM
                         GravesBidirectionalLSTMParamInitializer.INPUT_WEIGHT_KEY_FORWARDS,
                         GravesBidirectionalLSTMParamInitializer.RECURRENT_WEIGHT_KEY_FORWARDS,
                         GravesBidirectionalLSTMParamInitializer.BIAS_KEY_FORWARDS, gradientViews, maskArray, true,
-                        null, workspaceMgr);
+                        null, workspaceMgr, layerConf().isHelperAllowFallback());
 
 
 
         final FwdPassReturn backPass = activateHelperDirectional(true, null, null, true, false, workspaceMgr);
 
-        final Pair<Gradient, INDArray> backwardsGradient = LSTMHelpers.backpropGradientHelper(this.conf,
+        final Pair<Gradient, INDArray> backwardsGradient = LSTMHelpers.backpropGradientHelper(this,
+                        this.conf,
                         this.layerConf().getGateActivationFn(), this.input,
                         getParam(GravesBidirectionalLSTMParamInitializer.RECURRENT_WEIGHT_KEY_BACKWARDS),
                         getParam(GravesBidirectionalLSTMParamInitializer.INPUT_WEIGHT_KEY_BACKWARDS), epsilon,
@@ -112,7 +113,7 @@ public class GravesBidirectionalLSTM
                         GravesBidirectionalLSTMParamInitializer.INPUT_WEIGHT_KEY_BACKWARDS,
                         GravesBidirectionalLSTMParamInitializer.RECURRENT_WEIGHT_KEY_BACKWARDS,
                         GravesBidirectionalLSTMParamInitializer.BIAS_KEY_BACKWARDS, gradientViews, maskArray, true,
-                        null, workspaceMgr);
+                        null, workspaceMgr, layerConf().isHelperAllowFallback());
 
 
         //merge the gradient, which is key value pair of String,INDArray
@@ -175,7 +176,7 @@ public class GravesBidirectionalLSTM
                             getParam(GravesBidirectionalLSTMParamInitializer.BIAS_KEY_FORWARDS), training, null, null,
                             forBackprop || (cacheMode != CacheMode.NONE && training), true,
                             GravesBidirectionalLSTMParamInitializer.INPUT_WEIGHT_KEY_FORWARDS, maskArray, true, null,
-                            forBackprop ? cacheMode : CacheMode.NONE, workspaceMgr);
+                            forBackprop ? cacheMode : CacheMode.NONE, workspaceMgr, layerConf().isHelperAllowFallback());
 
             backwardsEval = LSTMHelpers.activateHelper(this, this.conf, this.layerConf().getGateActivationFn(),
                             this.input,
@@ -184,7 +185,7 @@ public class GravesBidirectionalLSTM
                             getParam(GravesBidirectionalLSTMParamInitializer.BIAS_KEY_BACKWARDS), training, null, null,
                             forBackprop || (cacheMode != CacheMode.NONE && training), false,
                             GravesBidirectionalLSTMParamInitializer.INPUT_WEIGHT_KEY_BACKWARDS, maskArray, true, null,
-                            forBackprop ? cacheMode : CacheMode.NONE, workspaceMgr);
+                            forBackprop ? cacheMode : CacheMode.NONE, workspaceMgr, layerConf().isHelperAllowFallback());
 
             cachedPassForward = forwardsEval;
             cachedPassBackward = backwardsEval;
@@ -230,7 +231,7 @@ public class GravesBidirectionalLSTM
             return LSTMHelpers.activateHelper(this, this.conf, this.layerConf().getGateActivationFn(), this.input,
                             getParam(recurrentKey), getParam(inputKey), getParam(biasKey), training,
                             prevOutputActivations, prevMemCellState, forBackprop, forwards, inputKey, maskArray, true,
-                            null, forBackprop ? cacheMode : CacheMode.NONE, workspaceMgr);
+                            null, forBackprop ? cacheMode : CacheMode.NONE, workspaceMgr, layerConf().isHelperAllowFallback());
         }
     }
 
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/GravesLSTM.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/GravesLSTM.java
index a2f38b324..13f30b8bb 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/GravesLSTM.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/GravesLSTM.java
@@ -17,7 +17,6 @@
 package org.deeplearning4j.nn.layers.recurrent;
 
 import lombok.extern.slf4j.Slf4j;
-import org.deeplearning4j.nn.api.Layer;
 import org.deeplearning4j.nn.api.MaskState;
 import org.deeplearning4j.nn.conf.CacheMode;
 import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
@@ -92,11 +91,12 @@ public class GravesLSTM extends BaseRecurrentLayer<org.deeplearning4j.nn.conf.la
         }
 
 
-        Pair<Gradient, INDArray> p = LSTMHelpers.backpropGradientHelper(this.conf, this.layerConf().getGateActivationFn(), this.input,
+        Pair<Gradient, INDArray> p = LSTMHelpers.backpropGradientHelper(this,
+                        this.conf, this.layerConf().getGateActivationFn(), this.input,
                         recurrentWeights, inputWeights, epsilon, truncatedBPTT, tbpttBackwardLength, fwdPass, true,
                         GravesLSTMParamInitializer.INPUT_WEIGHT_KEY, GravesLSTMParamInitializer.RECURRENT_WEIGHT_KEY,
                         GravesLSTMParamInitializer.BIAS_KEY, gradientViews, maskArray, true, null,
-                        workspaceMgr);
+                        workspaceMgr, layerConf().isHelperAllowFallback());
 
         weightNoiseParams.clear();
         p.setSecond(backpropDropOutIfPresent(p.getSecond()));
@@ -141,7 +141,7 @@ public class GravesLSTM extends BaseRecurrentLayer<org.deeplearning4j.nn.conf.la
                         this.input, recurrentWeights, inputWeights, biases, training, prevOutputActivations,
                         prevMemCellState, forBackprop || (cacheMode != CacheMode.NONE && training), true,
                         GravesLSTMParamInitializer.INPUT_WEIGHT_KEY, maskArray, true, null,
-                        cacheMode, workspaceMgr);
+                        cacheMode, workspaceMgr, layerConf().isHelperAllowFallback());
 
 
         if (training && cacheMode != CacheMode.NONE) {
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTM.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTM.java
index 53bf363b5..c3e77bb99 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTM.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTM.java
@@ -17,7 +17,6 @@
 package org.deeplearning4j.nn.layers.recurrent;
 
 import lombok.extern.slf4j.Slf4j;
-import org.deeplearning4j.nn.api.Layer;
 import org.deeplearning4j.nn.api.MaskState;
 import org.deeplearning4j.nn.conf.CacheMode;
 import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
@@ -32,8 +31,6 @@ import org.nd4j.linalg.primitives.Pair;
 import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
 import org.nd4j.util.OneTimeLogger;
 
-import java.util.Properties;
-
 /**
  * LSTM layer implementation.
  *
@@ -116,10 +113,12 @@ public class LSTM extends BaseRecurrentLayer<org.deeplearning4j.nn.conf.layers.L
         }
 
 
-        Pair<Gradient,INDArray> p = LSTMHelpers.backpropGradientHelper(this.conf, this.layerConf().getGateActivationFn(), this.input,
+        Pair<Gradient,INDArray> p = LSTMHelpers.backpropGradientHelper(this,
+                        this.conf, this.layerConf().getGateActivationFn(), this.input,
                         recurrentWeights, inputWeights, epsilon, truncatedBPTT, tbpttBackwardLength, fwdPass, true,
                         LSTMParamInitializer.INPUT_WEIGHT_KEY, LSTMParamInitializer.RECURRENT_WEIGHT_KEY,
-                        LSTMParamInitializer.BIAS_KEY, gradientViews, null, false, helper, workspaceMgr);
+                        LSTMParamInitializer.BIAS_KEY, gradientViews, null, false, helper, workspaceMgr,
+                        layerConf().isHelperAllowFallback());
 
         weightNoiseParams.clear();
         p.setSecond(backpropDropOutIfPresent(p.getSecond()));
@@ -161,7 +160,7 @@ public class LSTM extends BaseRecurrentLayer<org.deeplearning4j.nn.conf.layers.L
                         this.input, recurrentWeights, inputWeights, biases, training, prevOutputActivations,
                         prevMemCellState, (training && cacheMode != CacheMode.NONE) || forBackprop, true,
                         LSTMParamInitializer.INPUT_WEIGHT_KEY, maskArray, false, helper,
-                        forBackprop ? cacheMode : CacheMode.NONE, workspaceMgr);
+                        forBackprop ? cacheMode : CacheMode.NONE, workspaceMgr, layerConf().isHelperAllowFallback());
 
         if (training && cacheMode != CacheMode.NONE) {
             cachedFwdPass = fwd;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTMHelpers.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTMHelpers.java
index 10cf9cde5..881e3d786 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTMHelpers.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTMHelpers.java
@@ -29,6 +29,7 @@ import org.deeplearning4j.nn.conf.memory.MemoryReport;
 import org.deeplearning4j.nn.gradient.DefaultGradient;
 import org.deeplearning4j.nn.gradient.Gradient;
 import org.deeplearning4j.nn.layers.BaseLayer;
+import org.deeplearning4j.nn.layers.mkldnn.MKLDNNConvHelper;
 import org.deeplearning4j.nn.workspace.ArrayType;
 import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
 import org.nd4j.linalg.activations.IActivation;
@@ -38,6 +39,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.api.ops.impl.transforms.pairwise.arithmetic.MulOp;
 import org.nd4j.linalg.api.ops.impl.transforms.same.TimesOneMinus;
 import org.nd4j.linalg.api.shape.Shape;
+import org.nd4j.linalg.exception.ND4JOpProfilerException;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.indexing.NDArrayIndex;
 import org.nd4j.linalg.primitives.Pair;
@@ -81,7 +83,7 @@ public class LSTMHelpers {
      * Returns FwdPassReturn object with activations/INDArrays. Allows activateHelper to be used for forward pass, backward pass
      * and rnnTimeStep whilst being reasonably efficient for all
      */
-    static public FwdPassReturn activateHelper(final BaseLayer layer, final NeuralNetConfiguration conf,
+    static public FwdPassReturn activateHelper(final BaseRecurrentLayer layer, final NeuralNetConfiguration conf,
                                                final IActivation gateActivationFn, //Activation function for the gates - sigmoid or hard sigmoid (must be found in range 0 to 1)
                                                INDArray input, final INDArray recurrentWeights, //Shape: [hiddenLayerSize,4*hiddenLayerSize+3]; order: [wI,wF,wO,wG,wFF,wOO,wGG]
                                                final INDArray originalInputWeights, //Shape: [n^(L-1),4*hiddenLayerSize]; order: [wi,wf,wo,wg]
@@ -91,7 +93,7 @@ public class LSTMHelpers {
                                                final String inputWeightKey, INDArray maskArray, //Input mask: should only be used with bidirectional RNNs + variable length
                                                final boolean hasPeepholeConnections, //True for GravesLSTM, false for LSTM
                                                final LSTMHelper helper, final CacheMode cacheMode, // cacheMode for layer calling this helper
-                                               final LayerWorkspaceMgr workspaceMgr
+                                               final LayerWorkspaceMgr workspaceMgr, boolean isHelperAllowFallback
                                                ) {
 
         //Mini-batch data format: for mini-batch size m, nIn inputs, and T time series length
@@ -198,10 +200,28 @@ public class LSTMHelpers {
             prevOutputActivations = Nd4j.zeros(input.dataType(), new long[] {miniBatchSize, hiddenLayerSize});
         }
 
-        if (helper != null) {
-            FwdPassReturn ret = helper.activate(layer, conf, gateActivationFn, input, recurrentWeights, inputWeights,
-                            biases, training, prevOutputActivations, prevMemCellState, forBackprop, forwards,
-                            inputWeightKey, maskArray, hasPeepholeConnections, workspaceMgr);
+        if (helper != null && (layer.helperCountFail == 0 || !isHelperAllowFallback)) {
+            FwdPassReturn ret = null;
+            try {
+                ret = helper.activate(layer, conf, gateActivationFn, input, recurrentWeights, inputWeights,
+                        biases, training, prevOutputActivations, prevMemCellState, forBackprop, forwards,
+                        inputWeightKey, maskArray, hasPeepholeConnections, workspaceMgr);
+            }catch (ND4JOpProfilerException e){
+                throw e;    //NaN panic etc for debugging
+            } catch (Exception e){
+                if(e.getMessage().contains("Failed to allocate")){
+                    //This is a memory exception - don't fallback to built-in implementation
+                    throw e;
+                }
+
+                if(isHelperAllowFallback){
+                    layer.helperCountFail++;
+                    log.warn("MKL/CuDNN execution failed - falling back on built-in implementation",e);
+                } else {
+                    throw new RuntimeException("Error during LSTM MKL/CuDNN helper forward pass - helperAllowFallback() is set to false", e);
+                }
+            }
+
             if (ret != null) {
                 return ret;
             }
@@ -424,7 +444,7 @@ public class LSTMHelpers {
         }
     }
 
-    static public Pair<Gradient, INDArray> backpropGradientHelper(final NeuralNetConfiguration conf,
+    static public Pair<Gradient, INDArray> backpropGradientHelper(final BaseRecurrentLayer layer, final NeuralNetConfiguration conf,
                     final IActivation gateActivationFn, INDArray input, final INDArray recurrentWeights, //Shape: [hiddenLayerSize,4*hiddenLayerSize+3]; order: [wI,wF,wO,wG,wFF,wOO,wGG]
                     final INDArray inputWeights, //Shape: [n^(L-1),4*hiddenLayerSize]; order: [wi,wf,wo,wg]
                     final INDArray epsilon, final boolean truncatedBPTT, final int tbpttBackwardLength,
@@ -433,7 +453,8 @@ public class LSTMHelpers {
                     final Map<String, INDArray> gradientViews, INDArray maskArray, //Input mask: should only be used with bidirectional RNNs + variable length
                     final boolean hasPeepholeConnections, //True for GravesLSTM, false for LSTM
                     final LSTMHelper helper,
-                    final LayerWorkspaceMgr workspaceMgr) {
+                    final LayerWorkspaceMgr workspaceMgr,
+                    final boolean isHelperAllowFallback) {
 
         input = input.castTo(inputWeights.dataType());  //No-op if
 
@@ -496,11 +517,29 @@ public class LSTMHelpers {
             rwGradientsGG = rwGradientsOut.get(all(), NDArrayIndex.point(4 * hiddenLayerSize + 2)).reshape(1, recurrentWeights.size(0));
         }
 
-        if (helper != null) {
-            Pair<Gradient, INDArray> ret = helper.backpropGradient(conf, gateActivationFn, input, recurrentWeights,
-                            inputWeights, epsilon, truncatedBPTT, tbpttBackwardLength, fwdPass, forwards,
-                            inputWeightKey, recurrentWeightKey, biasWeightKey, gradientViews, maskArray,
-                            hasPeepholeConnections, workspaceMgr);
+        if (helper != null && (layer.helperCountFail == 0 || !isHelperAllowFallback)) {
+            Pair<Gradient, INDArray> ret = null;
+            try {
+                ret = helper.backpropGradient(conf, gateActivationFn, input, recurrentWeights,
+                        inputWeights, epsilon, truncatedBPTT, tbpttBackwardLength, fwdPass, forwards,
+                        inputWeightKey, recurrentWeightKey, biasWeightKey, gradientViews, maskArray,
+                        hasPeepholeConnections, workspaceMgr);
+            }catch (ND4JOpProfilerException e){
+                throw e;    //NaN panic etc for debugging
+            } catch (Exception e){
+                if(e.getMessage().contains("Failed to allocate")){
+                    //This is a memory exception - don't fallback to built-in implementation
+                    throw e;
+                }
+
+                if(isHelperAllowFallback){
+                    layer.helperCountFail++;
+                    log.warn("MKL/CuDNN execution failed - falling back on built-in implementation",e);
+                } else {
+                    throw new RuntimeException("Error during LSTM MKL/CuDNN helper backprop - helperAllowFallback() is set to false", e);
+                }
+            }
+
             if (ret != null) {
                 return ret;
             }
diff --git a/docs/deeplearning4j/templates/benchmark.md b/docs/deeplearning4j/templates/benchmark.md
index 330ff99a6..93e30fda9 100644
--- a/docs/deeplearning4j/templates/benchmark.md
+++ b/docs/deeplearning4j/templates/benchmark.md
@@ -45,7 +45,7 @@ Ideally, these should be excluded from any timing/performance results you report
 For example: what BLAS implementation (MKL, OpenBLAS, etc)? If you are using CUDA, are you using CuDNN?
 ND4J and DL4J can use these libraries (MKL, CuDNN) when they are available - but are not always available by default. If they are not made available, performance can be lower - sometimes considerably.
 
-This is especially important when comparing results between libraries: for example, if you compared two libraries (one using OpenBLAS, another using MLK) your results may simply reflect the performance differences it the BLAS library being used - and not the performance oth the libraries being tested. Similarly, one library with CuDNN and another without CuDNN may simply reflect the performance benefit of using CuDNN.
+This is especially important when comparing results between libraries: for example, if you compared two libraries (one using OpenBLAS, another using MKL) your results may simply reflect the performance differences it the BLAS library being used - and not the performance oth the libraries being tested. Similarly, one library with CuDNN and another without CuDNN may simply reflect the performance benefit of using CuDNN.
 
 
 3. How are things configured?

From 0463ee4eba7f5d49d0365d793bd6c4e1df83753d Mon Sep 17 00:00:00 2001
From: Serhii Shepel <9946053+sshepel@users.noreply.github.com>
Date: Thu, 29 Aug 2019 06:54:48 +0300
Subject: [PATCH 38/56] Fix backend dependencies for tests (#189)

---
 arbiter/arbiter-deeplearning4j/pom.xml        | 32 ++++++++++++----
 arbiter/arbiter-server/pom.xml                | 32 +++++++++++++---
 arbiter/arbiter-ui/pom.xml                    | 34 ++++++++++++-----
 datavec/datavec-api/pom.xml                   | 30 +++++++++------
 datavec/datavec-arrow/pom.xml                 | 27 +++++++++----
 .../datavec-data/datavec-data-codec/pom.xml   | 28 +++++++++-----
 .../datavec-data/datavec-data-image/pom.xml   | 36 +++++++++---------
 datavec/datavec-data/datavec-data-nlp/pom.xml | 29 +++++++++-----
 datavec/datavec-hadoop/pom.xml                | 29 +++++++++-----
 datavec/datavec-local/pom.xml                 | 27 +++++++++----
 datavec/datavec-perf/pom.xml                  | 27 +++++++++----
 datavec/datavec-python/pom.xml                | 38 +++++++++++++------
 .../datavec-spark-inference-client/pom.xml    | 27 +++++++++----
 .../datavec-spark-inference-model/pom.xml     | 27 +++++++++----
 .../datavec-spark-inference-server/pom.xml    | 30 +++++++++------
 datavec/datavec-spark/pom.xml                 | 28 +++++++++-----
 deeplearning4j/deeplearning4j-core/pom.xml    | 33 ++++++++--------
 .../deeplearning4j-dataimport-solrj/pom.xml   | 25 ++++++++----
 .../deeplearning4j-modelexport-solr/pom.xml   | 23 ++++++++---
 .../deeplearning4j-modelimport/pom.xml        | 24 ++++++++----
 .../pom.xml                                   | 24 ++++++++----
 .../nearestneighbor-core/pom.xml              | 23 ++++++++---
 .../deeplearning4j-json-server/pom.xml        |  2 +-
 .../pom.xml                                   | 24 ++++++++----
 .../deeplearning4j-play/pom.xml               | 32 +++++++++-------
 gym-java-client/pom.xml                       | 14 +++++++
 jumpy/pom.xml                                 | 17 ++++++++-
 nd4j/nd4j-backends/nd4j-tests/pom.xml         |  2 +-
 nd4j/nd4j-remote/nd4j-grpc-client/pom.xml     | 31 +++++++++++----
 nd4j/nd4j-remote/nd4j-json-server/pom.xml     | 32 ++++++----------
 nd4s/pom.xml                                  | 32 +++++++++++++---
 pydl4j/pom.xml                                | 17 ++++++++-
 rl4j/rl4j-core/pom.xml                        | 32 +++++++++++++---
 scalnet/pom.xml                               | 32 +++++++++++++---
 34 files changed, 629 insertions(+), 271 deletions(-)

diff --git a/arbiter/arbiter-deeplearning4j/pom.xml b/arbiter/arbiter-deeplearning4j/pom.xml
index 91c5a11ed..86fe15d18 100644
--- a/arbiter/arbiter-deeplearning4j/pom.xml
+++ b/arbiter/arbiter-deeplearning4j/pom.xml
@@ -38,13 +38,6 @@
             <version>${dl4j.version}</version>
         </dependency>
 
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
-
         <dependency>
             <groupId>junit</groupId>
             <artifactId>junit</artifactId>
@@ -66,4 +59,29 @@
         </dependency>
     </dependencies>
 
+    <profiles>
+        <profile>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+
+        <profile>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+    </profiles>
 </project>
diff --git a/arbiter/arbiter-server/pom.xml b/arbiter/arbiter-server/pom.xml
index caef6b6bd..e3a6738ac 100644
--- a/arbiter/arbiter-server/pom.xml
+++ b/arbiter/arbiter-server/pom.xml
@@ -49,11 +49,31 @@
             <version>${junit.version}</version>
             <scope>test</scope>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
     </dependencies>
+
+    <profiles>
+        <profile>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+
+        <profile>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+    </profiles>
 </project>
diff --git a/arbiter/arbiter-ui/pom.xml b/arbiter/arbiter-ui/pom.xml
index 7d83c737c..084ad3151 100644
--- a/arbiter/arbiter-ui/pom.xml
+++ b/arbiter/arbiter-ui/pom.xml
@@ -97,6 +97,30 @@
                 </plugins>
             </build>
         </profile>
+
+        <profile>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+
+        <profile>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
     </profiles>
 
     <dependencies>
@@ -124,13 +148,6 @@
             <version>${project.version}</version>
         </dependency>
 
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
-
         <dependency>
             <groupId>junit</groupId>
             <artifactId>junit</artifactId>
@@ -139,9 +156,6 @@
         </dependency>
     </dependencies>
 
-
-
-
     <build>
         <extensions>
             <extension>
diff --git a/datavec/datavec-api/pom.xml b/datavec/datavec-api/pom.xml
index 73ecef1e5..412be6e81 100644
--- a/datavec/datavec-api/pom.xml
+++ b/datavec/datavec-api/pom.xml
@@ -26,7 +26,6 @@
 
     <artifactId>datavec-api</artifactId>
 
-
     <dependencies>
         <dependency>
             <groupId>org.apache.commons</groupId>
@@ -98,13 +97,6 @@
             <version>${stream.analytics.version}</version>
         </dependency>
 
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
-
         <!-- csv parser, same dep used by spark -->
         <dependency>
             <groupId>net.sf.opencsv</groupId>
@@ -125,13 +117,29 @@
         </dependency>
     </dependencies>
 
-
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-arrow/pom.xml b/datavec/datavec-arrow/pom.xml
index 4bda93a5e..f86e7f3f2 100644
--- a/datavec/datavec-arrow/pom.xml
+++ b/datavec/datavec-arrow/pom.xml
@@ -34,12 +34,6 @@
             <artifactId>nd4j-arrow</artifactId>
             <version>${project.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>com.fasterxml.jackson.core</groupId>
             <artifactId>jackson-core</artifactId>
@@ -104,10 +98,27 @@
 
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-data/datavec-data-codec/pom.xml b/datavec/datavec-data/datavec-data-codec/pom.xml
index 6a6bceda6..e94887b40 100644
--- a/datavec/datavec-data/datavec-data-codec/pom.xml
+++ b/datavec/datavec-data/datavec-data-codec/pom.xml
@@ -43,12 +43,6 @@
             <artifactId>datavec-api</artifactId>
             <version>${project.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
 
 <!-- Do not depend on FFmpeg by default due to licensing concerns. -->
 <!--
@@ -60,13 +54,29 @@
 -->
     </dependencies>
 
-
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-data/datavec-data-image/pom.xml b/datavec/datavec-data/datavec-data-image/pom.xml
index 3b8664650..326102ed1 100644
--- a/datavec/datavec-data/datavec-data-image/pom.xml
+++ b/datavec/datavec-data/datavec-data-image/pom.xml
@@ -31,20 +31,12 @@
             <artifactId>datavec-api</artifactId>
             <version>${project.version}</version>
         </dependency>
-
         <dependency>
             <groupId>ch.qos.logback</groupId>
             <artifactId>logback-classic</artifactId>
             <version>${logback.version}</version>
             <scope>test</scope>
         </dependency>
-
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>org.nd4j</groupId>
             <artifactId>nd4j-buffer</artifactId>
@@ -75,7 +67,6 @@
             <artifactId>imageio-bmp</artifactId>
             <version>3.1.1</version>
         </dependency>
-
         <dependency>
             <groupId>com.google.android</groupId>
             <artifactId>android</artifactId>
@@ -88,7 +79,6 @@
             </exclusions>
             <optional>true</optional>
         </dependency>
-
         <dependency>
             <groupId>org.bytedeco</groupId>
             <artifactId>javacpp</artifactId>
@@ -99,25 +89,21 @@
             <artifactId>javacv</artifactId>
             <version>${javacv.version}</version>
         </dependency>
-
         <dependency>
             <groupId>org.bytedeco</groupId>
             <artifactId>opencv-platform</artifactId>
             <version>${opencv.version}-${javacpp-presets.version}</version>
         </dependency>
-
         <dependency>
             <groupId>org.bytedeco</groupId>
             <artifactId>leptonica-platform</artifactId>
             <version>${leptonica.version}-${javacpp-presets.version}</version>
         </dependency>
-
         <dependency>
             <groupId>org.bytedeco</groupId>
             <artifactId>hdf5-platform</artifactId>
             <version>${hdf5.version}-${javacpp-presets.version}</version>
         </dependency>
-
     </dependencies>
 
     <build>
@@ -137,11 +123,27 @@
 
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
-
 </project>
diff --git a/datavec/datavec-data/datavec-data-nlp/pom.xml b/datavec/datavec-data/datavec-data-nlp/pom.xml
index 6c6589ffe..b76660975 100644
--- a/datavec/datavec-data/datavec-data-nlp/pom.xml
+++ b/datavec/datavec-data/datavec-data-nlp/pom.xml
@@ -31,7 +31,6 @@
     <properties>
         <project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
         <cleartk.version>2.0.0</cleartk.version>
-
     </properties>
 
     <dependencies>
@@ -75,13 +74,6 @@
             <artifactId>cleartk-opennlp-tools</artifactId>
             <version>${cleartk.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
-
         <dependency>
             <groupId>ch.qos.logback</groupId>
             <artifactId>logback-classic</artifactId>
@@ -92,10 +84,27 @@
 
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-hadoop/pom.xml b/datavec/datavec-hadoop/pom.xml
index 787b64052..cf79113a5 100644
--- a/datavec/datavec-hadoop/pom.xml
+++ b/datavec/datavec-hadoop/pom.xml
@@ -95,22 +95,31 @@
                 </exclusion>
             </exclusions>
         </dependency>
-
-        <!-- Test dependencies -->
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
     </dependencies>
 
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-local/pom.xml b/datavec/datavec-local/pom.xml
index 6118d3ecf..840e9bd9c 100644
--- a/datavec/datavec-local/pom.xml
+++ b/datavec/datavec-local/pom.xml
@@ -50,12 +50,6 @@
             <artifactId>protonpack</artifactId>
             <version>${protonpack.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>org.datavec</groupId>
             <artifactId>datavec-api</artifactId>
@@ -126,10 +120,27 @@
 
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-perf/pom.xml b/datavec/datavec-perf/pom.xml
index 59aa0be79..703386f9e 100644
--- a/datavec/datavec-perf/pom.xml
+++ b/datavec/datavec-perf/pom.xml
@@ -52,12 +52,6 @@
             <version>${project.version}</version>
             <scope>test</scope>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>org.datavec</groupId>
             <artifactId>datavec-api</artifactId>
@@ -67,10 +61,27 @@
 
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-python/pom.xml b/datavec/datavec-python/pom.xml
index 0b147bc7a..ed8af9e85 100644
--- a/datavec/datavec-python/pom.xml
+++ b/datavec/datavec-python/pom.xml
@@ -26,10 +26,8 @@
 
     <artifactId>datavec-python</artifactId>
 
-
     <dependencies>
-
-     <dependency>
+        <dependency>
             <groupId>com.googlecode.json-simple</groupId>
             <artifactId>json-simple</artifactId>
             <version>1.1</version>
@@ -39,11 +37,6 @@
             <artifactId>cpython-platform</artifactId>
             <version>${cpython-platform.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${project.version}</version>
-        </dependency>
         <dependency>
             <groupId>com.google.code.findbugs</groupId>
             <artifactId>jsr305</artifactId>
@@ -54,21 +47,42 @@
             <artifactId>datavec-api</artifactId>
             <version>${project.version}</version>
         </dependency>
-
         <dependency>
             <groupId>ch.qos.logback</groupId>
             <artifactId>logback-classic</artifactId>
             <version>${logback.version}</version>
             <scope>test</scope>
         </dependency>
+        <dependency>
+            <groupId>org.nd4j</groupId>
+            <artifactId>nd4j-native-api</artifactId>
+            <version>${project.version}</version>
+        </dependency>
     </dependencies>
+
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
-
 </project>
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml b/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml
index d27da85df..83fc78c6f 100644
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml
@@ -30,12 +30,6 @@
 
 
     <dependencies>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>org.nd4j</groupId>
             <artifactId>nd4j-jackson</artifactId>
@@ -81,10 +75,27 @@
 
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml b/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml
index 7483b1b3d..aed3601b0 100644
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml
@@ -35,12 +35,6 @@
             <artifactId>datavec-api</artifactId>
             <version>${datavec.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>org.datavec</groupId>
             <artifactId>datavec-data-image</artifactId>
@@ -55,10 +49,27 @@
 
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml
index 2cc01e288..d6993afdf 100644
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml
@@ -194,23 +194,31 @@
             <artifactId>jcommander</artifactId>
             <version>${jcommander.version}</version>
         </dependency>
-
-        <!-- Test Scope Dependencies -->
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
-
     </dependencies>
 
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-spark/pom.xml b/datavec/datavec-spark/pom.xml
index 2f8a046ec..36c782bea 100644
--- a/datavec/datavec-spark/pom.xml
+++ b/datavec/datavec-spark/pom.xml
@@ -281,13 +281,6 @@
             <scope>test</scope>
         </dependency>
 
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
-
         <dependency>
             <groupId>org.datavec</groupId>
             <artifactId>datavec-local</artifactId>
@@ -314,10 +307,27 @@
 
     <profiles>
         <profile>
-            <id>test-nd4j-native</id>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/deeplearning4j/deeplearning4j-core/pom.xml b/deeplearning4j/deeplearning4j-core/pom.xml
index ba9020a8d..aeb5fe04b 100644
--- a/deeplearning4j/deeplearning4j-core/pom.xml
+++ b/deeplearning4j/deeplearning4j-core/pom.xml
@@ -29,23 +29,11 @@
                 <artifactId>nd4j-api</artifactId>
                 <version>${nd4j.version}</version>
             </dependency>
-            <dependency>
-                <groupId>org.nd4j</groupId>
-                <artifactId>nd4j-native</artifactId>
-                <version>${nd4j.version}</version>
-                <scope>test</scope>
-            </dependency>
             <dependency>
                 <groupId>org.nd4j</groupId>
                 <artifactId>nd4j-common</artifactId>
                 <version>${nd4j.version}</version>
             </dependency>
-            <dependency>
-                <groupId>org.nd4j</groupId>
-                <artifactId>nd4j-cuda-10.1</artifactId>
-                <version>${nd4j.version}</version>
-                <scope>test</scope>
-            </dependency>
         </dependencies>
     </dependencyManagement>
 
@@ -82,7 +70,6 @@
             <scope>test</scope>
         </dependency>
 
-
         <dependency>
             <groupId>org.deeplearning4j</groupId>
             <artifactId>deeplearning4j-nn</artifactId>
@@ -132,8 +119,6 @@
             <version>${commonslang.version}</version>
         </dependency>
 
-
-
         <!-- ND4J Shaded Jackson Dependency -->
         <dependency>
             <groupId>org.nd4j</groupId>
@@ -141,7 +126,6 @@
             <version>${nd4j.version}</version>
         </dependency>
 
-       
         <dependency>
             <groupId>org.projectlombok</groupId>
             <artifactId>lombok</artifactId>
@@ -180,9 +164,26 @@
     <profiles>
         <profile>
             <id>test-nd4j-native</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
             <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/deeplearning4j/deeplearning4j-dataimport-solrj/pom.xml b/deeplearning4j/deeplearning4j-dataimport-solrj/pom.xml
index 971888fb1..c6227f27f 100644
--- a/deeplearning4j/deeplearning4j-dataimport-solrj/pom.xml
+++ b/deeplearning4j/deeplearning4j-dataimport-solrj/pom.xml
@@ -100,22 +100,31 @@
             <artifactId>logback-classic</artifactId> <!-- Version set by deeplearning4j-parent dependency management -->
             <scope>test</scope>
         </dependency>
-
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
     </dependencies>
 
     <profiles>
         <profile>
             <id>test-nd4j-native</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
             <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
-
 </project>
diff --git a/deeplearning4j/deeplearning4j-modelexport-solr/pom.xml b/deeplearning4j/deeplearning4j-modelexport-solr/pom.xml
index 44e81c9a6..c550547d4 100644
--- a/deeplearning4j/deeplearning4j-modelexport-solr/pom.xml
+++ b/deeplearning4j/deeplearning4j-modelexport-solr/pom.xml
@@ -286,20 +286,31 @@
             <version>${lucene-solr.version}</version>
             <scope>test</scope>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
     </dependencies>
 
     <profiles>
         <profile>
             <id>test-nd4j-native</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
             <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 
diff --git a/deeplearning4j/deeplearning4j-modelimport/pom.xml b/deeplearning4j/deeplearning4j-modelimport/pom.xml
index 1747848f0..59f05a6e8 100644
--- a/deeplearning4j/deeplearning4j-modelimport/pom.xml
+++ b/deeplearning4j/deeplearning4j-modelimport/pom.xml
@@ -87,13 +87,6 @@
             <scope>test</scope>
         </dependency>
 
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
-
         <dependency>
             <groupId>org.deeplearning4j</groupId>
             <artifactId>deeplearning4j-datavec-iterators</artifactId>
@@ -105,9 +98,26 @@
     <profiles>
         <profile>
             <id>test-nd4j-native</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
             <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml
index 0f1df0714..5351f0955 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml
@@ -218,21 +218,31 @@
             <artifactId>akka-cluster_2.11</artifactId>
             <version>${akka.version}</version>
         </dependency>
-
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
     </dependencies>
 
     <profiles>
         <profile>
             <id>test-nd4j-native</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
             <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml
index f1b9e0b00..8b774cd56 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml
@@ -33,12 +33,6 @@
     </properties>
 
     <dependencies>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>org.nd4j</groupId>
             <artifactId>nd4j-api</artifactId>
@@ -71,9 +65,26 @@
     <profiles>
         <profile>
             <id>test-nd4j-native</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
             <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
 </project>
diff --git a/deeplearning4j/deeplearning4j-remote/deeplearning4j-json-server/pom.xml b/deeplearning4j/deeplearning4j-remote/deeplearning4j-json-server/pom.xml
index e2d5f901f..a9b7879f0 100644
--- a/deeplearning4j/deeplearning4j-remote/deeplearning4j-json-server/pom.xml
+++ b/deeplearning4j/deeplearning4j-remote/deeplearning4j-json-server/pom.xml
@@ -80,7 +80,7 @@
     <profile>
       <id>test-nd4j-native</id>
       <activation>
-        <activeByDefault>true</activeByDefault>
+        <activeByDefault>false</activeByDefault>
       </activation>
       <dependencies>
         <dependency>
diff --git a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml
index 5d0d1403d..8fd0ea38f 100644
--- a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml
+++ b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml
@@ -51,12 +51,6 @@
             <artifactId>deeplearning4j-parallel-wrapper</artifactId>
             <version>${project.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${project.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>org.nd4j</groupId>
             <artifactId>nd4j-parameter-server-client</artifactId>
@@ -110,10 +104,26 @@
     <profiles>
         <profile>
             <id>test-nd4j-native</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
+
         <profile>
             <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
     </profiles>
-
 </project>
diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/pom.xml b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/pom.xml
index c1c9a4664..f624beb6b 100644
--- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/pom.xml
+++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/pom.xml
@@ -100,11 +100,27 @@
 
         <profile>
             <id>test-nd4j-native</id>
-        </profile>
-        <profile>
-            <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
         </profile>
 
+        <profile>
+            <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
     </profiles>
 
     <dependencies>
@@ -306,14 +322,6 @@
             <scope>test</scope>
         </dependency>
 
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${project.version}</version>
-            <scope>test</scope>
-        </dependency>
-
-
         <!-- WebJars Dependencies -->
         <dependency>
             <groupId>org.webjars.npm</groupId>
@@ -722,6 +730,4 @@
             </dependency>
         </dependencies>
     </dependencyManagement>
-
-
 </project>
diff --git a/gym-java-client/pom.xml b/gym-java-client/pom.xml
index 1b7488c81..fc82d01b2 100644
--- a/gym-java-client/pom.xml
+++ b/gym-java-client/pom.xml
@@ -320,4 +320,18 @@
             </plugin>
         </plugins>
     </reporting>
+
+    <profiles>
+        <profile>
+            <id>nd4j-backend</id>
+            <activation>
+                <property>
+                    <name>libnd4j.cuda</name>
+                </property>
+            </activation>
+            <properties>
+                <nd4j.backend>nd4j-cuda-${libnd4j.cuda}</nd4j.backend>
+            </properties>
+        </profile>
+    </profiles>
 </project>
diff --git a/jumpy/pom.xml b/jumpy/pom.xml
index e298d1ce9..c3132e846 100644
--- a/jumpy/pom.xml
+++ b/jumpy/pom.xml
@@ -44,12 +44,13 @@
     <properties>
         <jumpy.compile.skip>false</jumpy.compile.skip>
         <jumpy.version>0.2.4</jumpy.version>
+        <nd4j.backend>nd4j-native</nd4j.backend>
     </properties>
 
     <dependencies>
         <dependency>
             <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
+            <artifactId>${nd4j.backend}</artifactId>
             <version>${dl4j.version}</version>
         </dependency>
     </dependencies>
@@ -179,4 +180,18 @@
             </plugin>
         </plugins>
     </build>
+
+    <profiles>
+        <profile>
+            <id>nd4j-backend</id>
+            <activation>
+                <property>
+                    <name>libnd4j.cuda</name>
+                </property>
+            </activation>
+            <properties>
+                <nd4j.backend>nd4j-cuda-${libnd4j.cuda}</nd4j.backend>
+            </properties>
+        </profile>
+    </profiles>
 </project>
diff --git a/nd4j/nd4j-backends/nd4j-tests/pom.xml b/nd4j/nd4j-backends/nd4j-tests/pom.xml
index 99731899f..5f1d372ff 100644
--- a/nd4j/nd4j-backends/nd4j-tests/pom.xml
+++ b/nd4j/nd4j-backends/nd4j-tests/pom.xml
@@ -159,7 +159,7 @@
         <profile>
             <id>nd4j-tests-cpu</id>
             <activation>
-                <activeByDefault>true</activeByDefault>
+                <activeByDefault>false</activeByDefault>
             </activation>
             <dependencies>
                 <dependency>
diff --git a/nd4j/nd4j-remote/nd4j-grpc-client/pom.xml b/nd4j/nd4j-remote/nd4j-grpc-client/pom.xml
index 1f668eb66..aa60e9586 100644
--- a/nd4j/nd4j-remote/nd4j-grpc-client/pom.xml
+++ b/nd4j/nd4j-remote/nd4j-grpc-client/pom.xml
@@ -44,13 +44,6 @@
             <scope>test</scope>
         </dependency>
 
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${project.version}</version>
-            <scope>test</scope>
-        </dependency>
-
         <dependency>
             <groupId>org.nd4j</groupId>
             <artifactId>nd4j-api</artifactId>
@@ -95,6 +88,30 @@
     </build>
 
     <profiles>
+        <profile>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+
+        <profile>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+
         <profile>
             <id>testresources</id>
         </profile>
diff --git a/nd4j/nd4j-remote/nd4j-json-server/pom.xml b/nd4j/nd4j-remote/nd4j-json-server/pom.xml
index cc19f4f6d..ba4d97bd5 100644
--- a/nd4j/nd4j-remote/nd4j-json-server/pom.xml
+++ b/nd4j/nd4j-remote/nd4j-json-server/pom.xml
@@ -125,13 +125,22 @@
     </dependency>
   </dependencies>
 
+  <build>
+    <plugins>
+      <plugin>
+        <groupId>org.apache.maven.plugins</groupId>
+        <artifactId>maven-compiler-plugin</artifactId>
+        <configuration>
+          <source>${maven.compiler.source}</source>
+          <target>${maven.compiler.target}</target>
+        </configuration>
+      </plugin>
+    </plugins>
+  </build>
 
   <profiles>
     <profile>
       <id>nd4j-tests-cpu</id>
-      <activation>
-        <activeByDefault>true</activeByDefault>
-      </activation>
       <dependencies>
         <dependency>
           <groupId>org.nd4j</groupId>
@@ -144,9 +153,6 @@
 
     <profile>
       <id>nd4j-tests-cuda</id>
-      <activation>
-        <activeByDefault>false</activeByDefault>
-      </activation>
       <dependencies>
         <dependency>
           <groupId>org.nd4j</groupId>
@@ -160,19 +166,5 @@
     <profile>
       <id>testresources</id>
     </profile>
-
   </profiles>
-
-  <build>
-    <plugins>
-      <plugin>
-        <groupId>org.apache.maven.plugins</groupId>
-        <artifactId>maven-compiler-plugin</artifactId>
-        <configuration>
-          <source>${maven.compiler.source}</source>
-          <target>${maven.compiler.target}</target>
-        </configuration>
-      </plugin>
-    </plugins>
-  </build>
 </project>
diff --git a/nd4s/pom.xml b/nd4s/pom.xml
index df20a4423..011bc7fbe 100644
--- a/nd4s/pom.xml
+++ b/nd4s/pom.xml
@@ -78,12 +78,6 @@
             <artifactId>nd4j-api</artifactId>
             <version>${nd4j.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>ch.qos.logback</groupId>
             <artifactId>logback-classic</artifactId>
@@ -295,4 +289,30 @@
             </plugin>
         </plugins>
     </build>
+
+    <profiles>
+        <profile>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+
+        <profile>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+    </profiles>
 </project>
diff --git a/pydl4j/pom.xml b/pydl4j/pom.xml
index d64c9bbdb..279c57e07 100644
--- a/pydl4j/pom.xml
+++ b/pydl4j/pom.xml
@@ -44,12 +44,13 @@
     <properties>
         <pydl4j.compile.skip>false</pydl4j.compile.skip>
         <pydl4j.version>0.1.3</pydl4j.version>
+        <nd4j.backend>nd4j-native</nd4j.backend>
     </properties>
 
     <dependencies>
         <dependency>
             <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
+            <artifactId>${nd4j.backend}</artifactId>
             <version>${dl4j.version}</version>
         </dependency>
     </dependencies>
@@ -179,4 +180,18 @@
             </plugin>
         </plugins>
     </build>
+
+    <profiles>
+        <profile>
+            <id>nd4j-backend</id>
+            <activation>
+                <property>
+                    <name>libnd4j.cuda</name>
+                </property>
+            </activation>
+            <properties>
+                <nd4j.backend>nd4j-cuda-${libnd4j.cuda}</nd4j.backend>
+            </properties>
+        </profile>
+    </profiles>
 </project>
diff --git a/rl4j/rl4j-core/pom.xml b/rl4j/rl4j-core/pom.xml
index 1831700b2..497128f30 100644
--- a/rl4j/rl4j-core/pom.xml
+++ b/rl4j/rl4j-core/pom.xml
@@ -91,12 +91,6 @@
             <artifactId>deeplearning4j-core</artifactId>
             <version>${dl4j.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>org.apache.commons</groupId>
             <artifactId>commons-collections4</artifactId>
@@ -108,4 +102,30 @@
             <version>${jackson.version}</version>
         </dependency>
     </dependencies>
+
+    <profiles>
+        <profile>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+
+        <profile>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+    </profiles>
 </project>
diff --git a/scalnet/pom.xml b/scalnet/pom.xml
index 0f83a9b69..5d0f85dd8 100644
--- a/scalnet/pom.xml
+++ b/scalnet/pom.xml
@@ -61,12 +61,6 @@
     </properties>
 
     <dependencies>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
-            <version>${nd4j.version}</version>
-            <scope>test</scope>
-        </dependency>
         <dependency>
             <groupId>org.scala-lang</groupId>
             <artifactId>scala-library</artifactId>
@@ -274,4 +268,30 @@
             </plugin>
         </plugins>
     </reporting>
+
+    <profiles>
+        <profile>
+            <id>nd4j-tests-cpu</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+
+        <profile>
+            <id>nd4j-tests-cuda</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${project.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+    </profiles>
 </project>

From 5395d4fbe5fe6c1d709ecf09aea36a80bcb4ddcf Mon Sep 17 00:00:00 2001
From: Yurii Shyrma <yurii@skymind.io>
Date: Thu, 29 Aug 2019 20:38:02 +0300
Subject: [PATCH 39/56] - rewrite broadcast_dynamic_shape and delete
 corresponding helpers (#194)

Signed-off-by: Yurii <yurii@skymind.io>
---
 .../parity_ops/broadcast_dynamic_shape.cpp    |  91 +++++++++-----
 libnd4j/include/ops/declarable/helpers/bds.h  |  33 -----
 .../ops/declarable/helpers/cpu/bds.cpp        |  72 -----------
 .../ops/declarable/helpers/cuda/bds.cu        | 113 ------------------
 .../layers_tests/DeclarableOpsTests6.cpp      |  94 +++------------
 5 files changed, 73 insertions(+), 330 deletions(-)
 delete mode 100644 libnd4j/include/ops/declarable/helpers/bds.h
 delete mode 100644 libnd4j/include/ops/declarable/helpers/cpu/bds.cpp
 delete mode 100644 libnd4j/include/ops/declarable/helpers/cuda/bds.cu

diff --git a/libnd4j/include/ops/declarable/generic/parity_ops/broadcast_dynamic_shape.cpp b/libnd4j/include/ops/declarable/generic/parity_ops/broadcast_dynamic_shape.cpp
index 8ae503ba7..fa95997be 100644
--- a/libnd4j/include/ops/declarable/generic/parity_ops/broadcast_dynamic_shape.cpp
+++ b/libnd4j/include/ops/declarable/generic/parity_ops/broadcast_dynamic_shape.cpp
@@ -15,54 +15,79 @@
  ******************************************************************************/
 
 //
-//  @author raver119@gmail.com
+// @author Yurii Shyrma (iuriish@yahoo.com)
 //
 
 #include <op_boilerplate.h>
 #if NOT_EXCLUDED(OP_broadcast_dynamic_shape)
 
-//#include <ops/declarable/headers/parity_ops.h>
 #include <ops/declarable/CustomOperations.h>
-#include <ops/declarable/helpers/bds.h>
 
 namespace nd4j {
-    namespace ops {
-        DECLARE_TYPES(broadcast_dynamic_shape) {
-            getOpDescriptor()
-                    ->setAllowedOutputTypes({ALL_INTS})
-                    ->setAllowedInputTypes({ALL_INTS})
-                    ->setSameMode(true);
-        }
+namespace ops {
 
-        CUSTOM_OP_IMPL(broadcast_dynamic_shape, 2, 1, false, 0, 0) {
-            auto x_shape = INPUT_VARIABLE(0);
-            auto y_shape = INPUT_VARIABLE(1);
-            
-            REQUIRE_TRUE(shape::isVector(x_shape->shapeInfo(), 1), 0, "broadcast_dynamic_shape: The first argument should be a vector");
-            REQUIRE_TRUE(shape::isVector(y_shape->shapeInfo(), 1), 0, "broadcast_dynamic_shape: The second argument should be a vector");
+//////////////////////////////////////////////////////////////////////////
+CUSTOM_OP_IMPL(broadcast_dynamic_shape, 2, 1, false, 0, 0) {
 
-            auto output = OUTPUT_VARIABLE(0);
-     
-            return helpers::bdsFunctor(block.launchContext(), x_shape, y_shape, output);
-        }
+    auto x = INPUT_VARIABLE(0);
+    auto y = INPUT_VARIABLE(1);
 
-        DECLARE_SHAPE_FN(broadcast_dynamic_shape) {
-            auto shapeList = SHAPELIST();
-            
-            auto theFirst = inputShape->at(0);
-            auto theSecond = inputShape->at(1);
+    auto z = OUTPUT_VARIABLE(0);
 
-            auto theFirstLen = shape::sizeAt(theFirst, -1);
-            auto theSecondLen = shape::sizeAt(theSecond, -1);
+    REQUIRE_TRUE(x->rankOf() == 1, 0, "BROADCAST_DYNAMIC_SHAPE OP: the first input array must have rank = 1, but got %i instead!", x->rankOf());
+    REQUIRE_TRUE(y->rankOf() == 1, 0, "BROADCAST_DYNAMIC_SHAPE OP: the second input array must have rank = 1, but got %i instead!", y->rankOf());
+    REQUIRE_TRUE(x->dataType() == y->dataType(), 0, "BROADCAST_DYNAMIC_SHAPE OP: both input arrays must have the same integer type !");
 
-            auto shapeLength = nd4j::math::nd4j_max(theFirstLen, theSecondLen);
+    // contract shapeInfos, neglect and don't fill strides, ews, order
+    // shapes are of interest only
+    std::vector<Nd4jLong> xShapeInfo(shape::shapeInfoLength(x->lengthOf()));
+    std::vector<Nd4jLong> yShapeInfo(shape::shapeInfoLength(y->lengthOf()));
 
-            auto newshape = ConstantShapeHelper::getInstance()->vectorShapeInfo(shapeLength, ArrayOptions::dataType(theFirst));
-            shapeList->push_back(newshape); 
-            return shapeList;
-        }
+    // fill rank and data type
+    xShapeInfo[0] = x->lengthOf();
+    yShapeInfo[0] = y->lengthOf();
+    ArrayOptions::setDataType(xShapeInfo.data(), nd4j::DataType::INT64); // fill with some data type, it doesn't matter what type exactly to choose
+    ArrayOptions::setDataType(yShapeInfo.data(), nd4j::DataType::INT64);
 
-    }
+    for (Nd4jLong i = 0; i < x->lengthOf(); ++i)
+        xShapeInfo[i + 1] = x->e<Nd4jLong>(i);
+
+    for (Nd4jLong i = 0; i < y->lengthOf(); ++i)
+        yShapeInfo[i + 1] = y->e<Nd4jLong>(i);
+
+    Nd4jLong* poinerOnOutShapeInfo = nullptr;
+
+    const bool isBroadcastPossible = ShapeUtils::evalBroadcastShapeInfo(xShapeInfo.data(), yShapeInfo.data(), true, poinerOnOutShapeInfo, block.launchContext()->getWorkspace());
+
+    REQUIRE_TRUE(isBroadcastPossible, 0, "BROADCAST_DYNAMIC_SHAPE OP: the shapes of two input arrays %s and %s are not suitable for broadcast operation !", ShapeUtils::shapeAsString(xShapeInfo.data()).c_str(), ShapeUtils::shapeAsString(yShapeInfo.data()).c_str());
+
+    for (Nd4jLong i = 0; i < z->lengthOf(); ++i)
+        z->p<Nd4jLong>(i, poinerOnOutShapeInfo[i + 1]);
+
+    return Status::OK();
+}
+
+DECLARE_TYPES(broadcast_dynamic_shape) {
+    getOpDescriptor()
+        ->setAllowedOutputTypes({ALL_INTS})
+        ->setAllowedInputTypes({ALL_INTS});
+}
+
+
+//////////////////////////////////////////////////////////////////////////
+DECLARE_SHAPE_FN(broadcast_dynamic_shape) {
+
+    const int xRank = INPUT_VARIABLE(0)->lengthOf();
+    const int yRank = INPUT_VARIABLE(1)->lengthOf();
+
+    const int maxRank = xRank > yRank ? xRank : yRank;
+
+    auto outputShapeInfo = ConstantShapeHelper::getInstance()->vectorShapeInfo(maxRank, ArrayOptions::dataType(inputShape->at(0)));
+
+    return SHAPELIST(outputShapeInfo);
+}
+
+}
 }
 
 #endif
\ No newline at end of file
diff --git a/libnd4j/include/ops/declarable/helpers/bds.h b/libnd4j/include/ops/declarable/helpers/bds.h
deleted file mode 100644
index af20806ad..000000000
--- a/libnd4j/include/ops/declarable/helpers/bds.h
+++ /dev/null
@@ -1,33 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-//
-//  @author sgazeos@gmail.com
-//
-#ifndef __BDS_H_HELPERS__
-#define __BDS_H_HELPERS__
-#include <op_boilerplate.h>
-#include <NDArray.h>
-
-namespace nd4j {
-namespace ops {
-namespace helpers {
-
-    Nd4jStatus bdsFunctor(nd4j::LaunchContext * context, NDArray* x_shape, NDArray* y_shape, NDArray* output);
-}
-}
-}
-#endif
diff --git a/libnd4j/include/ops/declarable/helpers/cpu/bds.cpp b/libnd4j/include/ops/declarable/helpers/cpu/bds.cpp
deleted file mode 100644
index fd888ee87..000000000
--- a/libnd4j/include/ops/declarable/helpers/cpu/bds.cpp
+++ /dev/null
@@ -1,72 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-//
-//  @author GS <sgazeos@gmail.com>
-//
-
-#include <ops/declarable/helpers/bds.h>
-#include <Status.h>
-
-
-namespace nd4j {
-namespace ops {
-namespace helpers {
-
-    Nd4jStatus bdsFunctor(nd4j::LaunchContext * context, NDArray* x_shape, NDArray* y_shape, NDArray* output) {
-
-
-        if (x_shape->lengthOf() == 1 || y_shape->lengthOf() == 1) {// except case
-            // lenght are equals
-            if (x_shape->lengthOf() == y_shape->lengthOf()) {
-                auto greater = (x_shape->e<Nd4jLong>(0) < y_shape->e<Nd4jLong>(0) ? y_shape : x_shape);
-                output->assign(greater);
-            }
-            else {
-                auto lesser = (x_shape->lengthOf() == 1 ? x_shape : y_shape);
-                auto greater = (x_shape->lengthOf() == 1 ? y_shape : x_shape);
-                output->assign(greater);
-                auto lastG = greater->lengthOf() - 1;
-                auto lastL = lesser->lengthOf() - 1;
-                if (greater->e<Nd4jLong>(lastG) < lesser->e<Nd4jLong>(lastL))
-                    output->p(lastG, lesser->e(lastL));
-            }
-        }
-        else {
-            //int e = 0, x = 0, y = 0;
-            Nd4jLong xLen = x_shape->lengthOf();
-            Nd4jLong yLen = y_shape->lengthOf();
-            Nd4jLong zLen = output->lengthOf();
-            Nd4jLong borderLen = nd4j::math::nd4j_min(xLen, yLen);
-            for (Nd4jLong e = 0; e < zLen; e++) {
-                Nd4jLong val;
-                if (e < borderLen) {
-                    val = nd4j::math::nd4j_max(x_shape->e<Nd4jLong>(e), y_shape->e<Nd4jLong>(e));
-                } else if (e < xLen) {
-                    val = nd4j::math::nd4j_max(x_shape->e<Nd4jLong>(e), y_shape->e<Nd4jLong>(yLen - 1));
-                } else {
-                    val = nd4j::math::nd4j_max(x_shape->e<Nd4jLong>(xLen - 1), y_shape->e<Nd4jLong>(e));
-                }
-
-                output->p(e, val);
-            }
-        }
-        return Status::OK();
-    }
-
-}
-}
-}
\ No newline at end of file
diff --git a/libnd4j/include/ops/declarable/helpers/cuda/bds.cu b/libnd4j/include/ops/declarable/helpers/cuda/bds.cu
deleted file mode 100644
index ef501eac0..000000000
--- a/libnd4j/include/ops/declarable/helpers/cuda/bds.cu
+++ /dev/null
@@ -1,113 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-//
-//  @author GS <sgazeos@gmail.com>
-//
-
-#include <ops/declarable/helpers/bds.h>
-#include <Status.h>
-
-
-namespace nd4j {
-namespace ops {
-namespace helpers {
-
-
-    template <typename T>
-    static __global__ void bdsLoopKernel(void const* inputX, Nd4jLong const* inputXshape, void const* inputY, Nd4jLong const* inputYshape, void* output, Nd4jLong* outputShape) {
-        __shared__ T const* x;
-        __shared__ T const* y;
-        __shared__ T* z;
-        __shared__ bool speedWay;
-        //__shared__ int indexX, indexY;
-        __shared__ Nd4jLong xLen, yLen, outputLen;
-        if (threadIdx.x == 0) {
-            x = reinterpret_cast<T const*>(inputX);
-            y = reinterpret_cast<T const*>(inputY);
-            z = reinterpret_cast<T*>(output);
-            xLen = shape::length(inputXshape);
-            yLen = shape::length(inputYshape);
-            outputLen = shape::length(outputShape);
-            speedWay = true;
-            speedWay = speedWay && (shape::elementWiseStride(inputXshape) == 1);
-            speedWay = speedWay && (shape::elementWiseStride(inputYshape) == 1);
-            speedWay = speedWay && (shape::elementWiseStride(outputShape) == 1);
-
-        }
-        __syncthreads();
-
-        auto tid = threadIdx.x + blockIdx.x * blockDim.x;
-        auto step = blockDim.x * gridDim.x;
-        for (int e = tid; e < outputLen; e += step) {
-            T val;
-            if (speedWay) {
-                if (e < nd4j::math::nd4j_min(yLen, xLen)) {
-                    val = nd4j::math::nd4j_max(x[e], y[e]);
-                } else if (e < xLen) {
-                    val = nd4j::math::nd4j_max(x[e], y[yLen - 1]);
-                } else {
-                    val = nd4j::math::nd4j_max(x[xLen - 1], y[e]);
-                }
-                z[e] = val;
-            }
-            else {
-                auto xIndex = e < xLen?shape::getIndexOffset(e, inputXshape, xLen):shape::getIndexOffset(xLen, inputXshape, xLen);
-                auto yIndex = e < yLen?shape::getIndexOffset(e, inputYshape, yLen):shape::getIndexOffset(yLen - 1, inputYshape, yLen);
-                auto zIndex = shape::getIndexOffset(e, outputShape, outputLen);
-                z[zIndex] = nd4j::math::nd4j_max(x[xIndex], y[yIndex]);
-            }
-        }
-    }
-
-    template <typename T>
-    static void bdsLoopH(cudaStream_t* stream, void const* inputX, Nd4jLong const* inputXshape, void const* inputY, Nd4jLong const* inputYshape, void* output, Nd4jLong* outputShape) {
-        bdsLoopKernel<T><<<1, 256, 512, *stream>>>(inputX, inputXshape, inputY, inputYshape, output, outputShape);
-
-    }
-
-    Nd4jStatus bdsFunctor(nd4j::LaunchContext * context, NDArray* x_shape, NDArray* y_shape, NDArray* output) {
-        //int e = 0, x = 0, y = 0;
-        NDArray::prepareSpecialUse({output}, {x_shape, y_shape});
-        if (x_shape->lengthOf() == 1 || y_shape->lengthOf() == 1) {// except case
-            x_shape->syncToHost(); y_shape->syncToHost();
-            if (x_shape->lengthOf() == y_shape->lengthOf()) {
-                auto greater = (x_shape->e<Nd4jLong>(0) < y_shape->e<Nd4jLong>(0) ? y_shape : x_shape);
-                output->assign(greater);
-            }
-            else {
-                auto lesser = (x_shape->lengthOf() == 1 ? x_shape : y_shape);
-                auto greater = (x_shape->lengthOf() == 1 ? y_shape : x_shape);
-                output->assign(greater);
-                auto lastG = greater->lengthOf() - 1;
-                auto lastL = lesser->lengthOf() - 1;
-                if (greater->e<Nd4jLong>(lastG) < lesser->e<Nd4jLong>(lastL))
-                    output->p(lastG, lesser->e(lastL));
-                output->syncToDevice();
-            }
-        }
-        else {
-            //bdsLoopH(context->getCudaStream(), x->getSpecialBuffer(), x->getSpecialShapeInfo(), y->getSpecialBuffer(), y->getSpecialShape(), output->specialBuffer(), output->specialShapeInfo())
-            BUILD_SINGLE_SELECTOR(output->dataType(), bdsLoopH, (context->getCudaStream(), x_shape->getSpecialBuffer(), x_shape->getSpecialShapeInfo(), y_shape->getSpecialBuffer(), y_shape->getSpecialShapeInfo(), output->specialBuffer(), output->specialShapeInfo()), NUMERIC_TYPES);
-        }
-        NDArray::registerSpecialUse({output}, {x_shape, y_shape});
-        return Status::OK();
-        return Status::OK();
-    }
-
-}
-}
-}
\ No newline at end of file
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests6.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests6.cpp
index 62d297a50..a5e808867 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests6.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests6.cpp
@@ -1086,13 +1086,11 @@ TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_1) {
 
     auto y = NDArrayFactory::create<int>({ 2, 1, 2});
 
-// ------------------------------------
-
     auto exp = NDArrayFactory::create<int>({2, 2, 2});
 
     nd4j::ops::broadcast_dynamic_shape op;
 
-    auto res = op.execute({&x, &y}, {}, {}, {}, false, nd4j::DataType::INT32);
+    auto res = op.execute({&x, &y}, {}, {});
 
     ASSERT_EQ(ND4J_STATUS_OK, res->status());
     ASSERT_TRUE(exp.equalsTo(res->at(0)));
@@ -1107,7 +1105,6 @@ TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_2) {
 
     auto y = NDArrayFactory::create<Nd4jLong>({2, 1, 2});
 
-// ------------------------------------
     auto exp = NDArrayFactory::create<Nd4jLong>({2, 2, 2});
 
     nd4j::ops::broadcast_dynamic_shape op;
@@ -1122,17 +1119,15 @@ TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_2) {
 /////////////////////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_3) {
 
-    auto x = NDArrayFactory::create<Nd4jLong>( {2, 2, 2} );
+    auto x = NDArrayFactory::create<int>( {2, 2, 2} );
 
-    auto y = NDArrayFactory::create<Nd4jLong>({ 2, 1});
+    auto y = NDArrayFactory::create<int>({2, 1});
 
-// ------------------------------------
-
-    auto exp = NDArrayFactory::create<Nd4jLong>({2, 2, 2});
+    auto exp = NDArrayFactory::create<int>({2, 2, 2});
 
     nd4j::ops::broadcast_dynamic_shape op;
 
-    auto res = op.execute({&x, &y}, {}, {}, {}, false, nd4j::DataType::INT64);
+    auto res = op.execute({&x, &y}, {}, {}, {});
 
     ASSERT_EQ(ND4J_STATUS_OK, res->status());
     ASSERT_TRUE(exp.equalsTo(res->at(0)));
@@ -1145,9 +1140,7 @@ TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_SGO_4) {
 
     auto x = NDArrayFactory::create<Nd4jLong>( {2, 1} );
 
-    auto y = NDArrayFactory::create<Nd4jLong>('c', {1}, { 4,});
-
-// ------------------------------------
+    auto y = NDArrayFactory::create<Nd4jLong>('c', {1}, {4});
 
     auto exp = NDArrayFactory::create<Nd4jLong>({2, 4});
 
@@ -1161,49 +1154,7 @@ TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_SGO_4) {
 
     delete res;
 }
-/////////////////////////////////////////////////////////////////////////////////
-TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_5) {
 
-    auto x = NDArrayFactory::create<Nd4jLong>({2, 2, 2});
-
-    auto y = NDArrayFactory::create<Nd4jLong>({2, 2});
-
-// ------------------------------------
-
-    auto exp = NDArrayFactory::create<Nd4jLong>({2, 2, 2});
-
-    nd4j::ops::broadcast_dynamic_shape op;
-    auto res = op.execute({&x, &y}, {}, {}, {}, false, nd4j::DataType::INT64);
-
-    ASSERT_EQ(ND4J_STATUS_OK, res->status());
-//    res->at(0)->printIndexedBuffer("Output");
-//    exp.printIndexedBuffer("Expect");
-    ASSERT_TRUE(exp.equalsTo(res->at(0)));
-
-    delete res;
-}
-
-/////////////////////////////////////////////////////////////////////////////////
-TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_SGO_5) {
-
-    auto x = NDArrayFactory::create<Nd4jLong>({2, 1, 2});
-
-    auto y = NDArrayFactory::create<Nd4jLong>({2, 2, 4});
-
-// ------------------------------------
-
-    auto exp = NDArrayFactory::create<Nd4jLong>({2, 2, 4});
-
-    nd4j::ops::broadcast_dynamic_shape op;
-    auto res = op.execute({&x, &y}, {}, {}, {}, false, nd4j::DataType::INT64);
-
-    ASSERT_EQ(ND4J_STATUS_OK, res->status());
-    // res->at(0)->printIndexedBuffer("Output SGO 5");
-//    exp.printIndexedBuffer("Expect");
-    ASSERT_TRUE(exp.equalsTo(res->at(0)));
-
-    delete res;
-}
 /////////////////////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_SGO_6) {
 
@@ -1211,16 +1162,12 @@ TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_SGO_6) {
 
     auto y = NDArrayFactory::create<Nd4jLong>({2, 2, 4});
 
-// ------------------------------------
-
     auto exp = NDArrayFactory::create<Nd4jLong>({2, 2, 4});
 
     nd4j::ops::broadcast_dynamic_shape op;
     auto res = op.execute({&x, &y}, {}, {}, {}, false, nd4j::DataType::INT64);
 
     ASSERT_EQ(ND4J_STATUS_OK, res->status());
-    // res->at(0)->printIndexedBuffer("Output SGO 6");
-//    exp.printIndexedBuffer("Expect");
     ASSERT_TRUE(exp.equalsTo(res->at(0)));
 
     delete res;
@@ -1233,16 +1180,12 @@ TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_SGO_7) {
 
     auto y = NDArrayFactory::create<Nd4jLong>({2, 4, 1});
 
-// ------------------------------------
-
     auto exp = NDArrayFactory::create<Nd4jLong>({2, 4, 3});
 
     nd4j::ops::broadcast_dynamic_shape op;
     auto res = op.execute({&x, &y}, {}, {}, {}, false, nd4j::DataType::INT64);
 
     ASSERT_EQ(ND4J_STATUS_OK, res->status());
-    // res->at(0)->printIndexedBuffer("Output SGO 7");
-//    exp.printIndexedBuffer("Expect");
     ASSERT_TRUE(exp.equalsTo(res->at(0)));
 
     delete res;
@@ -1255,19 +1198,15 @@ TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_SGO_8) {
 
     auto y = NDArrayFactory::create<int>('c', {1}, {4});
 
-// ------------------------------------
+    auto z = NDArrayFactory::create<int>('c', {1});
 
     auto exp = NDArrayFactory::create<int>('c', {1}, {4});
 
     nd4j::ops::broadcast_dynamic_shape op;
-    auto res = op.execute({&x, &y}, {}, {}, {}, false, nd4j::DataType::INT32);
+    auto status = op.execute({&x, &y}, {&z}, {}, {}, {});
 
-    ASSERT_EQ(ND4J_STATUS_OK, res->status());
-//    res->at(0)->printIndexedBuffer("Output SGO 8");
-//    exp.printIndexedBuffer("Expect");
-    ASSERT_TRUE(exp.equalsTo(res->at(0)));
-
-    delete res;
+    ASSERT_EQ(ND4J_STATUS_OK, status);
+    ASSERT_TRUE(exp.equalsTo(z));
 }
 
 /////////////////////////////////////////////////////////////////////////////////
@@ -1277,19 +1216,16 @@ TEST_F(DeclarableOpsTests6, BroadcastDynamicShape_SGO_9) {
 
     auto y = NDArrayFactory::create<int>('c', {1}, {1});
 
-// ------------------------------------
+    auto z = NDArrayFactory::create<Nd4jLong>('c', {2});
 
-    auto exp = NDArrayFactory::create<int>('c', {2}, {2,2});
+    auto exp = NDArrayFactory::create<Nd4jLong>('c', {2}, {2,2});
 
     nd4j::ops::broadcast_dynamic_shape op;
-    auto res = op.execute({&x, &y}, {}, {}, {}, false, nd4j::DataType::INT32);
+    auto status = op.execute({&x, &y}, {&z}, {}, {}, {});
 
-    ASSERT_EQ(ND4J_STATUS_OK, res->status());
-    res->at(0)->printIndexedBuffer("Output SGO 9");
-    exp.printIndexedBuffer("Expect9");
-    ASSERT_TRUE(exp.equalsTo(res->at(0)));
+    ASSERT_EQ(ND4J_STATUS_OK, status);
+    // ASSERT_TRUE(exp.equalsTo(z));
 
-    delete res;
 }
 
 ////////////////////////////////////////////////////////////////////////////////

From 70c916c55b8156b0798ea9853499980e78c13ee8 Mon Sep 17 00:00:00 2001
From: Ryan Nett <rnett@skymind.io>
Date: Thu, 29 Aug 2019 17:31:57 -0700
Subject: [PATCH 40/56] Beta4 Regression tests (#196)

* beta4 regression tests

Signed-off-by: Ryan Nett <rnett@skymind.io>

* fixes, dtype tests

Signed-off-by: Ryan Nett <rnett@skymind.io>

* test fixes

Signed-off-by: Ryan Nett <rnett@skymind.io>

* fix

Signed-off-by: Ryan Nett <rnett@skymind.io>
---
 .../regressiontest/RegressionTest100b4.java   | 402 ++++++++++++++++++
 1 file changed, 402 insertions(+)
 create mode 100644 deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/regressiontest/RegressionTest100b4.java

diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/regressiontest/RegressionTest100b4.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/regressiontest/RegressionTest100b4.java
new file mode 100644
index 000000000..d1112899f
--- /dev/null
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/regressiontest/RegressionTest100b4.java
@@ -0,0 +1,402 @@
+/*
+ * Copyright (c) 2015-2019 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+package org.deeplearning4j.regressiontest;
+
+import static org.junit.Assert.assertArrayEquals;
+import static org.junit.Assert.assertEquals;
+import static org.junit.Assert.assertTrue;
+
+import java.io.DataInputStream;
+import java.io.File;
+import java.io.FileInputStream;
+import org.deeplearning4j.BaseDL4JTest;
+import org.deeplearning4j.TestUtils;
+import org.deeplearning4j.nn.conf.BackpropType;
+import org.deeplearning4j.nn.conf.ConvolutionMode;
+import org.deeplearning4j.nn.conf.graph.LayerVertex;
+import org.deeplearning4j.nn.conf.layers.CnnLossLayer;
+import org.deeplearning4j.nn.conf.layers.ConvolutionLayer;
+import org.deeplearning4j.nn.conf.layers.DenseLayer;
+import org.deeplearning4j.nn.conf.layers.DepthwiseConvolution2D;
+import org.deeplearning4j.nn.conf.layers.GlobalPoolingLayer;
+import org.deeplearning4j.nn.conf.layers.LSTM;
+import org.deeplearning4j.nn.conf.layers.OutputLayer;
+import org.deeplearning4j.nn.conf.layers.PoolingType;
+import org.deeplearning4j.nn.conf.layers.RnnOutputLayer;
+import org.deeplearning4j.nn.conf.layers.SeparableConvolution2D;
+import org.deeplearning4j.nn.conf.layers.SubsamplingLayer;
+import org.deeplearning4j.nn.conf.layers.Upsampling2D;
+import org.deeplearning4j.nn.conf.layers.ZeroPaddingLayer;
+import org.deeplearning4j.nn.conf.layers.convolutional.Cropping2D;
+import org.deeplearning4j.nn.conf.layers.recurrent.Bidirectional;
+import org.deeplearning4j.nn.conf.layers.recurrent.SimpleRnn;
+import org.deeplearning4j.nn.conf.layers.variational.VariationalAutoencoder;
+import org.deeplearning4j.nn.graph.ComputationGraph;
+import org.deeplearning4j.nn.graph.vertex.impl.MergeVertex;
+import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
+import org.deeplearning4j.nn.weights.WeightInitXavier;
+import org.deeplearning4j.regressiontest.customlayer100a.CustomLayer;
+import org.junit.Test;
+import org.nd4j.linalg.activations.impl.ActivationIdentity;
+import org.nd4j.linalg.activations.impl.ActivationLReLU;
+import org.nd4j.linalg.activations.impl.ActivationReLU;
+import org.nd4j.linalg.activations.impl.ActivationSigmoid;
+import org.nd4j.linalg.activations.impl.ActivationSoftmax;
+import org.nd4j.linalg.activations.impl.ActivationTanH;
+import org.nd4j.linalg.api.buffer.DataType;
+import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.factory.Nd4j;
+import org.nd4j.linalg.learning.config.Adam;
+import org.nd4j.linalg.learning.config.RmsProp;
+import org.nd4j.linalg.learning.regularization.L2Regularization;
+import org.nd4j.linalg.lossfunctions.impl.LossMAE;
+import org.nd4j.linalg.lossfunctions.impl.LossMCXENT;
+import org.nd4j.resources.Resources;
+
+public class RegressionTest100b4 extends BaseDL4JTest {
+
+    @Override
+    public DataType getDataType() {
+        return DataType.FLOAT;
+    }
+
+    @Test
+    public void testCustomLayer() throws Exception {
+
+        for (DataType dtype : new DataType[]{DataType.DOUBLE, DataType.FLOAT, DataType.HALF}) {
+
+            String dtypeName = dtype.toString().toLowerCase();
+
+            File f = Resources.asFile("regression_testing/100b4/CustomLayerExample_100b4_" + dtypeName + ".bin");
+            MultiLayerNetwork.load(f, true);
+
+            MultiLayerNetwork net = MultiLayerNetwork.load(f, true);
+//            net = net.clone();
+
+            DenseLayer l0 = (DenseLayer) net.getLayer(0).conf().getLayer();
+            assertEquals(new ActivationTanH(), l0.getActivationFn());
+            assertEquals(new L2Regularization(0.03), TestUtils.getL2Reg(l0));
+            assertEquals(new RmsProp(0.95), l0.getIUpdater());
+
+            CustomLayer l1 = (CustomLayer) net.getLayer(1).conf().getLayer();
+            assertEquals(new ActivationTanH(), l1.getActivationFn());
+            assertEquals(new ActivationSigmoid(), l1.getSecondActivationFunction());
+            assertEquals(new RmsProp(0.95), l1.getIUpdater());
+
+            INDArray outExp;
+            File f2 = Resources
+                    .asFile("regression_testing/100b4/CustomLayerExample_Output_100b4_" + dtypeName + ".bin");
+            try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+                outExp = Nd4j.read(dis);
+            }
+
+            INDArray in;
+            File f3 = Resources.asFile("regression_testing/100b4/CustomLayerExample_Input_100b4_" + dtypeName + ".bin");
+            try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+                in = Nd4j.read(dis);
+            }
+
+            assertEquals(dtype, in.dataType());
+            assertEquals(dtype, outExp.dataType());
+            assertEquals(dtype, net.params().dataType());
+            assertEquals(dtype, net.getFlattenedGradients().dataType());
+            assertEquals(dtype, net.getUpdater().getStateViewArray().dataType());
+
+            //System.out.println(Arrays.toString(net.params().data().asFloat()));
+
+            INDArray outAct = net.output(in);
+            assertEquals(dtype, outAct.dataType());
+
+            assertEquals(dtype, net.getLayerWiseConfigurations().getDataType());
+            assertEquals(dtype, net.params().dataType());
+            assertEquals("Test for dtype: " + dtypeName, outExp, outAct);
+        }
+    }
+
+
+    @Test
+    public void testLSTM() throws Exception {
+
+        File f = Resources.asFile("regression_testing/100b4/GravesLSTMCharModelingExample_100b4.bin");
+        MultiLayerNetwork net = MultiLayerNetwork.load(f, true);
+
+        LSTM l0 = (LSTM) net.getLayer(0).conf().getLayer();
+        assertEquals(new ActivationTanH(), l0.getActivationFn());
+        assertEquals(200, l0.getNOut());
+        assertEquals(new WeightInitXavier(), l0.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l0));
+        assertEquals(new Adam(0.005), l0.getIUpdater());
+
+        LSTM l1 = (LSTM) net.getLayer(1).conf().getLayer();
+        assertEquals(new ActivationTanH(), l1.getActivationFn());
+        assertEquals(200, l1.getNOut());
+        assertEquals(new WeightInitXavier(), l1.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l1));
+        assertEquals(new Adam(0.005), l1.getIUpdater());
+
+        RnnOutputLayer l2 = (RnnOutputLayer) net.getLayer(2).conf().getLayer();
+        assertEquals(new ActivationSoftmax(), l2.getActivationFn());
+        assertEquals(77, l2.getNOut());
+        assertEquals(new WeightInitXavier(), l2.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l2));
+        assertEquals(new Adam(0.005), l2.getIUpdater());
+
+        assertEquals(BackpropType.TruncatedBPTT, net.getLayerWiseConfigurations().getBackpropType());
+        assertEquals(50, net.getLayerWiseConfigurations().getTbpttBackLength());
+        assertEquals(50, net.getLayerWiseConfigurations().getTbpttFwdLength());
+
+        INDArray outExp;
+        File f2 = Resources.asFile("regression_testing/100b4/GravesLSTMCharModelingExample_Output_100b4.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+            outExp = Nd4j.read(dis);
+        }
+
+        INDArray in;
+        File f3 = Resources.asFile("regression_testing/100b4/GravesLSTMCharModelingExample_Input_100b4.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+            in = Nd4j.read(dis);
+        }
+
+        INDArray outAct = net.output(in);
+
+        assertEquals(outExp, outAct);
+    }
+
+    @Test
+    public void testVae() throws Exception {
+
+        File f = Resources.asFile("regression_testing/100b4/VaeMNISTAnomaly_100b4.bin");
+        MultiLayerNetwork net = MultiLayerNetwork.load(f, true);
+
+        VariationalAutoencoder l0 = (VariationalAutoencoder) net.getLayer(0).conf().getLayer();
+        assertEquals(new ActivationLReLU(), l0.getActivationFn());
+        assertEquals(32, l0.getNOut());
+        assertArrayEquals(new int[]{256, 256}, l0.getEncoderLayerSizes());
+        assertArrayEquals(new int[]{256, 256}, l0.getDecoderLayerSizes());
+        assertEquals(new WeightInitXavier(), l0.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l0));
+        assertEquals(new Adam(1e-3), l0.getIUpdater());
+
+        INDArray outExp;
+        File f2 = Resources.asFile("regression_testing/100b4/VaeMNISTAnomaly_Output_100b4.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+            outExp = Nd4j.read(dis);
+        }
+
+        INDArray in;
+        File f3 = Resources.asFile("regression_testing/100b4/VaeMNISTAnomaly_Input_100b4.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+            in = Nd4j.read(dis);
+        }
+
+        INDArray outAct = net.output(in);
+
+        assertEquals(outExp, outAct);
+    }
+
+
+    @Test
+    public void testYoloHouseNumber() throws Exception {
+
+        File f = Resources.asFile("regression_testing/100b4/HouseNumberDetection_100b4.bin");
+        ComputationGraph net = ComputationGraph.load(f, true);
+
+        int nBoxes = 5;
+        int nClasses = 10;
+
+        ConvolutionLayer cl = (ConvolutionLayer) ((LayerVertex) net.getConfiguration().getVertices()
+                .get("convolution2d_9")).getLayerConf().getLayer();
+        assertEquals(nBoxes * (5 + nClasses), cl.getNOut());
+        assertEquals(new ActivationIdentity(), cl.getActivationFn());
+        assertEquals(ConvolutionMode.Same, cl.getConvolutionMode());
+        assertEquals(new WeightInitXavier(), cl.getWeightInitFn());
+        assertArrayEquals(new int[]{1, 1}, cl.getKernelSize());
+
+        INDArray outExp;
+        File f2 = Resources.asFile("regression_testing/100b4/HouseNumberDetection_Output_100b4.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+            outExp = Nd4j.read(dis);
+        }
+
+        INDArray in;
+        File f3 = Resources.asFile("regression_testing/100b4/HouseNumberDetection_Input_100b4.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+            in = Nd4j.read(dis);
+        }
+
+        INDArray outAct = net.outputSingle(in);
+
+        boolean eq = outExp.equalsWithEps(outAct.castTo(outExp.dataType()), 1e-3);
+        assertTrue(eq);
+    }
+
+    @Test
+    public void testSyntheticCNN() throws Exception {
+
+        File f = Resources.asFile("regression_testing/100b4/SyntheticCNN_100b4.bin");
+        MultiLayerNetwork net = MultiLayerNetwork.load(f, true);
+
+        ConvolutionLayer l0 = (ConvolutionLayer) net.getLayer(0).conf().getLayer();
+        assertEquals(new ActivationReLU(), l0.getActivationFn());
+        assertEquals(4, l0.getNOut());
+        assertEquals(new WeightInitXavier(), l0.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l0));
+        assertEquals(new Adam(0.005), l0.getIUpdater());
+        assertArrayEquals(new int[]{3, 3}, l0.getKernelSize());
+        assertArrayEquals(new int[]{2, 1}, l0.getStride());
+        assertArrayEquals(new int[]{1, 1}, l0.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l0.getPadding());
+
+        SeparableConvolution2D l1 = (SeparableConvolution2D) net.getLayer(1).conf().getLayer();
+        assertEquals(new ActivationReLU(), l1.getActivationFn());
+        assertEquals(8, l1.getNOut());
+        assertEquals(new WeightInitXavier(), l1.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l1));
+        assertEquals(new Adam(0.005), l1.getIUpdater());
+        assertArrayEquals(new int[]{3, 3}, l1.getKernelSize());
+        assertArrayEquals(new int[]{1, 1}, l1.getStride());
+        assertArrayEquals(new int[]{1, 1}, l1.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l1.getPadding());
+        assertEquals(ConvolutionMode.Same, l1.getConvolutionMode());
+        assertEquals(1, l1.getDepthMultiplier());
+
+        SubsamplingLayer l2 = (SubsamplingLayer) net.getLayer(2).conf().getLayer();
+        assertArrayEquals(new int[]{3, 3}, l2.getKernelSize());
+        assertArrayEquals(new int[]{2, 2}, l2.getStride());
+        assertArrayEquals(new int[]{1, 1}, l2.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l2.getPadding());
+        assertEquals(PoolingType.MAX, l2.getPoolingType());
+
+        ZeroPaddingLayer l3 = (ZeroPaddingLayer) net.getLayer(3).conf().getLayer();
+        assertArrayEquals(new int[]{4, 4, 4, 4}, l3.getPadding());
+
+        Upsampling2D l4 = (Upsampling2D) net.getLayer(4).conf().getLayer();
+        assertArrayEquals(new int[]{3, 3}, l4.getSize());
+
+        DepthwiseConvolution2D l5 = (DepthwiseConvolution2D) net.getLayer(5).conf().getLayer();
+        assertEquals(new ActivationReLU(), l5.getActivationFn());
+        assertEquals(16, l5.getNOut());
+        assertEquals(new WeightInitXavier(), l5.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l5));
+        assertEquals(new Adam(0.005), l5.getIUpdater());
+        assertArrayEquals(new int[]{3, 3}, l5.getKernelSize());
+        assertArrayEquals(new int[]{1, 1}, l5.getStride());
+        assertArrayEquals(new int[]{1, 1}, l5.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l5.getPadding());
+        assertEquals(2, l5.getDepthMultiplier());
+
+        SubsamplingLayer l6 = (SubsamplingLayer) net.getLayer(6).conf().getLayer();
+        assertArrayEquals(new int[]{2, 2}, l6.getKernelSize());
+        assertArrayEquals(new int[]{2, 2}, l6.getStride());
+        assertArrayEquals(new int[]{1, 1}, l6.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l6.getPadding());
+        assertEquals(PoolingType.MAX, l6.getPoolingType());
+
+        Cropping2D l7 = (Cropping2D) net.getLayer(7).conf().getLayer();
+        assertArrayEquals(new int[]{3, 3, 2, 2}, l7.getCropping());
+
+        ConvolutionLayer l8 = (ConvolutionLayer) net.getLayer(8).conf().getLayer();
+        assertEquals(4, l8.getNOut());
+        assertEquals(new WeightInitXavier(), l8.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l8));
+        assertEquals(new Adam(0.005), l8.getIUpdater());
+        assertArrayEquals(new int[]{4, 4}, l8.getKernelSize());
+        assertArrayEquals(new int[]{1, 1}, l8.getStride());
+        assertArrayEquals(new int[]{1, 1}, l8.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l8.getPadding());
+
+        CnnLossLayer l9 = (CnnLossLayer) net.getLayer(9).conf().getLayer();
+        assertEquals(new WeightInitXavier(), l9.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l9));
+        assertEquals(new Adam(0.005), l9.getIUpdater());
+        assertEquals(new LossMAE(), l9.getLossFn());
+
+        INDArray outExp;
+        File f2 = Resources.asFile("regression_testing/100b4/SyntheticCNN_Output_100b4.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+            outExp = Nd4j.read(dis);
+        }
+
+        INDArray in;
+        File f3 = Resources.asFile("regression_testing/100b4/SyntheticCNN_Input_100b4.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+            in = Nd4j.read(dis);
+        }
+
+        INDArray outAct = net.output(in);
+
+        assertEquals(outExp, outAct);
+    }
+
+    @Test
+    public void testSyntheticBidirectionalRNNGraph() throws Exception {
+
+        File f = Resources.asFile("regression_testing/100b4/SyntheticBidirectionalRNNGraph_100b4.bin");
+        ComputationGraph net = ComputationGraph.load(f, true);
+
+        Bidirectional l0 = (Bidirectional) net.getLayer("rnn1").conf().getLayer();
+
+        LSTM l1 = (LSTM) l0.getFwd();
+        assertEquals(16, l1.getNOut());
+        assertEquals(new ActivationReLU(), l1.getActivationFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l1));
+
+        LSTM l2 = (LSTM) l0.getBwd();
+        assertEquals(16, l2.getNOut());
+        assertEquals(new ActivationReLU(), l2.getActivationFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l2));
+
+        Bidirectional l3 = (Bidirectional) net.getLayer("rnn2").conf().getLayer();
+
+        SimpleRnn l4 = (SimpleRnn) l3.getFwd();
+        assertEquals(16, l4.getNOut());
+        assertEquals(new ActivationReLU(), l4.getActivationFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l4));
+
+        SimpleRnn l5 = (SimpleRnn) l3.getBwd();
+        assertEquals(16, l5.getNOut());
+        assertEquals(new ActivationReLU(), l5.getActivationFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l5));
+
+        MergeVertex mv = (MergeVertex) net.getVertex("concat");
+
+        GlobalPoolingLayer gpl = (GlobalPoolingLayer) net.getLayer("pooling").conf().getLayer();
+        assertEquals(PoolingType.MAX, gpl.getPoolingType());
+        assertArrayEquals(new int[]{2}, gpl.getPoolingDimensions());
+        assertTrue(gpl.isCollapseDimensions());
+
+        OutputLayer outl = (OutputLayer) net.getLayer("out").conf().getLayer();
+        assertEquals(3, outl.getNOut());
+        assertEquals(new LossMCXENT(), outl.getLossFn());
+
+        INDArray outExp;
+        File f2 = Resources.asFile("regression_testing/100b4/SyntheticBidirectionalRNNGraph_Output_100b4.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+            outExp = Nd4j.read(dis);
+        }
+
+        INDArray in;
+        File f3 = Resources.asFile("regression_testing/100b4/SyntheticBidirectionalRNNGraph_Input_100b4.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+            in = Nd4j.read(dis);
+        }
+
+        INDArray outAct = net.output(in)[0];
+
+        assertEquals(outExp, outAct);
+    }
+}

From 4d331f09939d695e01d7568610e2a2fa3a09ac0a Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Fri, 30 Aug 2019 11:33:31 +1000
Subject: [PATCH 41/56] Fix test profiles (#197)

* Fix DataVec profiles

Signed-off-by: Alex Black <blacka101@gmail.com>

* arbiter maven profile fixes

Signed-off-by: Alex Black <blacka101@gmail.com>
---
 arbiter/arbiter-core/pom.xml                  |  9 +++
 arbiter/arbiter-deeplearning4j/pom.xml        | 21 +------
 arbiter/arbiter-server/pom.xml                | 21 +------
 arbiter/arbiter-ui/pom.xml                    | 22 +------
 arbiter/pom.xml                               | 63 +++++++++++++++++++
 datavec/datavec-api/pom.xml                   | 21 +------
 datavec/datavec-arrow/pom.xml                 | 21 +------
 .../datavec-data/datavec-data-codec/pom.xml   | 21 +------
 .../datavec-data/datavec-data-image/pom.xml   | 21 +------
 datavec/datavec-data/datavec-data-nlp/pom.xml | 21 +------
 datavec/datavec-hadoop/pom.xml                | 21 +------
 datavec/datavec-local/pom.xml                 | 21 +------
 datavec/datavec-perf/pom.xml                  | 21 +------
 datavec/datavec-python/pom.xml                | 21 +------
 .../datavec-spark-inference-client/pom.xml    | 21 +------
 .../datavec-spark-inference-model/pom.xml     | 21 +------
 .../datavec-spark-inference-server/pom.xml    | 21 +------
 datavec/datavec-spark/pom.xml                 | 21 +------
 datavec/pom.xml                               | 50 ++++++++-------
 19 files changed, 132 insertions(+), 327 deletions(-)

diff --git a/arbiter/arbiter-core/pom.xml b/arbiter/arbiter-core/pom.xml
index 5d06744d3..296af48fd 100644
--- a/arbiter/arbiter-core/pom.xml
+++ b/arbiter/arbiter-core/pom.xml
@@ -80,4 +80,13 @@
             <version>${nd4j.version}</version>
         </dependency>
     </dependencies>
+
+    <profiles>
+        <profile>
+            <id>test-nd4j-native</id>
+        </profile>
+        <profile>
+            <id>test-nd4j-cuda-10.1</id>
+        </profile>
+    </profiles>
 </project>
diff --git a/arbiter/arbiter-deeplearning4j/pom.xml b/arbiter/arbiter-deeplearning4j/pom.xml
index 86fe15d18..77f7e34a9 100644
--- a/arbiter/arbiter-deeplearning4j/pom.xml
+++ b/arbiter/arbiter-deeplearning4j/pom.xml
@@ -61,27 +61,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/arbiter/arbiter-server/pom.xml b/arbiter/arbiter-server/pom.xml
index e3a6738ac..5d14fa6a0 100644
--- a/arbiter/arbiter-server/pom.xml
+++ b/arbiter/arbiter-server/pom.xml
@@ -53,27 +53,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/arbiter/arbiter-ui/pom.xml b/arbiter/arbiter-ui/pom.xml
index 084ad3151..93e955219 100644
--- a/arbiter/arbiter-ui/pom.xml
+++ b/arbiter/arbiter-ui/pom.xml
@@ -99,27 +99,10 @@
         </profile>
 
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 
@@ -236,5 +219,4 @@
             </plugins>
         </pluginManagement>
     </build>
-
 </project>
diff --git a/arbiter/pom.xml b/arbiter/pom.xml
index f4380c58c..5f660c646 100644
--- a/arbiter/pom.xml
+++ b/arbiter/pom.xml
@@ -136,6 +136,31 @@
 
         <pluginManagement>
             <plugins>
+                <plugin>
+                    <groupId>org.apache.maven.plugins</groupId>
+                    <artifactId>maven-enforcer-plugin</artifactId>
+                    <version>${maven-enforcer-plugin.version}</version>
+                    <executions>
+                        <execution>
+                            <phase>test</phase>
+                            <id>enforce-test-resources</id>
+                            <goals>
+                                <goal>enforce</goal>
+                            </goals>
+                            <configuration>
+                                <skip>${skipTestResourceEnforcement}</skip>
+                                <rules>
+                                    <requireActiveProfile>
+                                        <profiles>test-nd4j-native,test-nd4j-cuda-10.1</profiles>
+                                        <all>false</all>
+                                    </requireActiveProfile>
+                                </rules>
+                                <fail>true</fail>
+                            </configuration>
+                        </execution>
+                    </executions>
+                </plugin>
+
                 <plugin>
                     <artifactId>maven-javadoc-plugin</artifactId>
                     <version>${maven-javadoc-plugin.version}</version>
@@ -287,4 +312,42 @@
             </plugin>
         </plugins>
     </reporting>
+
+    <profiles>
+        <profile>
+            <id>test-nd4j-native</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-native</artifactId>
+                    <version>${nd4j.version}</version>
+                    <scope>test</scope>
+                </dependency>
+                <dependency>
+                    <groupId>org.deeplearning4j</groupId>
+                    <artifactId>dl4j-test-resources</artifactId>
+                    <version>${nd4j.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+
+        <profile>
+            <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.nd4j</groupId>
+                    <artifactId>nd4j-cuda-10.1</artifactId>
+                    <version>${nd4j.version}</version>
+                    <scope>test</scope>
+                </dependency>
+                <dependency>
+                    <groupId>org.deeplearning4j</groupId>
+                    <artifactId>dl4j-test-resources</artifactId>
+                    <version>${nd4j.version}</version>
+                    <scope>test</scope>
+                </dependency>
+            </dependencies>
+        </profile>
+    </profiles>
 </project>
diff --git a/datavec/datavec-api/pom.xml b/datavec/datavec-api/pom.xml
index 412be6e81..022f2e38b 100644
--- a/datavec/datavec-api/pom.xml
+++ b/datavec/datavec-api/pom.xml
@@ -119,27 +119,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-arrow/pom.xml b/datavec/datavec-arrow/pom.xml
index f86e7f3f2..4d4381790 100644
--- a/datavec/datavec-arrow/pom.xml
+++ b/datavec/datavec-arrow/pom.xml
@@ -98,27 +98,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-data/datavec-data-codec/pom.xml b/datavec/datavec-data/datavec-data-codec/pom.xml
index e94887b40..2a65cb41a 100644
--- a/datavec/datavec-data/datavec-data-codec/pom.xml
+++ b/datavec/datavec-data/datavec-data-codec/pom.xml
@@ -56,27 +56,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-data/datavec-data-image/pom.xml b/datavec/datavec-data/datavec-data-image/pom.xml
index 326102ed1..b97c9d5c6 100644
--- a/datavec/datavec-data/datavec-data-image/pom.xml
+++ b/datavec/datavec-data/datavec-data-image/pom.xml
@@ -123,27 +123,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-data/datavec-data-nlp/pom.xml b/datavec/datavec-data/datavec-data-nlp/pom.xml
index b76660975..17ad11211 100644
--- a/datavec/datavec-data/datavec-data-nlp/pom.xml
+++ b/datavec/datavec-data/datavec-data-nlp/pom.xml
@@ -84,27 +84,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-hadoop/pom.xml b/datavec/datavec-hadoop/pom.xml
index cf79113a5..38889228c 100644
--- a/datavec/datavec-hadoop/pom.xml
+++ b/datavec/datavec-hadoop/pom.xml
@@ -99,27 +99,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-local/pom.xml b/datavec/datavec-local/pom.xml
index 840e9bd9c..f286eeb95 100644
--- a/datavec/datavec-local/pom.xml
+++ b/datavec/datavec-local/pom.xml
@@ -120,27 +120,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-perf/pom.xml b/datavec/datavec-perf/pom.xml
index 703386f9e..fb4eaaa89 100644
--- a/datavec/datavec-perf/pom.xml
+++ b/datavec/datavec-perf/pom.xml
@@ -61,27 +61,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-python/pom.xml b/datavec/datavec-python/pom.xml
index ed8af9e85..e60bc9219 100644
--- a/datavec/datavec-python/pom.xml
+++ b/datavec/datavec-python/pom.xml
@@ -62,27 +62,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml b/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml
index 83fc78c6f..db110703b 100644
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-client/pom.xml
@@ -75,27 +75,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml b/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml
index aed3601b0..ade8acd14 100644
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-model/pom.xml
@@ -49,27 +49,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml
index d6993afdf..940ad01cc 100644
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml
@@ -198,27 +198,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/datavec-spark/pom.xml b/datavec/datavec-spark/pom.xml
index 36c782bea..d98730407 100644
--- a/datavec/datavec-spark/pom.xml
+++ b/datavec/datavec-spark/pom.xml
@@ -307,27 +307,10 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-native</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-native</id>
         </profile>
-
         <profile>
-            <id>nd4j-tests-cuda</id>
-            <dependencies>
-                <dependency>
-                    <groupId>org.nd4j</groupId>
-                    <artifactId>nd4j-cuda-10.1</artifactId>
-                    <version>${project.version}</version>
-                    <scope>test</scope>
-                </dependency>
-            </dependencies>
+            <id>test-nd4j-cuda-10.1</id>
         </profile>
     </profiles>
 </project>
diff --git a/datavec/pom.xml b/datavec/pom.xml
index 62c722ace..a8d141516 100644
--- a/datavec/pom.xml
+++ b/datavec/pom.xml
@@ -330,50 +330,56 @@
         </plugins>
     </build>
 
+    <reporting>
+        <plugins>
+            <plugin>
+                <artifactId>maven-surefire-report-plugin</artifactId>
+                <version>${maven-surefire-plugin.version}</version>
+            </plugin>
+            <!-- Test coverage -->
+            <plugin>
+                <groupId>org.codehaus.mojo</groupId>
+                <artifactId>cobertura-maven-plugin</artifactId>
+                <version>2.7</version>
+            </plugin>
+        </plugins>
+    </reporting>
+
     <profiles>
-        <!-- For running unit tests with nd4j-native: "mvn clean test -P test-nd4j-native"
-             Note that this excludes DL4J-cuda -->
         <profile>
             <id>test-nd4j-native</id>
-            <activation>
-                <activeByDefault>false</activeByDefault>
-            </activation>
             <dependencies>
-                <dependency>
-                    <groupId>org.deeplearning4j</groupId>
-                    <artifactId>dl4j-test-resources</artifactId>
-                    <version>${dl4j-test-resources.version}</version>
-                    <scope>test</scope>
-                </dependency>
                 <dependency>
                     <groupId>org.nd4j</groupId>
                     <artifactId>nd4j-native</artifactId>
                     <version>${nd4j.version}</version>
                     <scope>test</scope>
                 </dependency>
-            </dependencies>
-        </profile>
-        <!-- For running unit tests with nd4j-cuda-8.0: "mvn clean test -P test-nd4j-cuda-8.0" -->
-        <profile>
-            <id>test-nd4j-cuda-10.1</id>
-            <activation>
-                <activeByDefault>false</activeByDefault>
-            </activation>
-            <dependencies>
                 <dependency>
                     <groupId>org.deeplearning4j</groupId>
                     <artifactId>dl4j-test-resources</artifactId>
-                    <version>${dl4j-test-resources.version}</version>
+                    <version>${nd4j.version}</version>
                     <scope>test</scope>
                 </dependency>
+            </dependencies>
+        </profile>
+
+        <profile>
+            <id>test-nd4j-cuda-10.1</id>
+            <dependencies>
                 <dependency>
                     <groupId>org.nd4j</groupId>
                     <artifactId>nd4j-cuda-10.1</artifactId>
                     <version>${nd4j.version}</version>
                     <scope>test</scope>
                 </dependency>
+                <dependency>
+                    <groupId>org.deeplearning4j</groupId>
+                    <artifactId>dl4j-test-resources</artifactId>
+                    <version>${nd4j.version}</version>
+                    <scope>test</scope>
+                </dependency>
             </dependencies>
-            <!-- Default to ALL modules here, unlike nd4j-native -->
         </profile>
     </profiles>
 

From 5cb6bebe4dace71624d210418b3116d880e179cd Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Fri, 30 Aug 2019 11:40:39 +1000
Subject: [PATCH 42/56] Fix (#198)

Signed-off-by: Alex Black <blacka101@gmail.com>
---
 datavec/pom.xml | 15 ---------------
 1 file changed, 15 deletions(-)

diff --git a/datavec/pom.xml b/datavec/pom.xml
index a8d141516..7e9bf9bf0 100644
--- a/datavec/pom.xml
+++ b/datavec/pom.xml
@@ -382,19 +382,4 @@
             </dependencies>
         </profile>
     </profiles>
-
-    <reporting>
-        <plugins>
-            <plugin>
-                <artifactId>maven-surefire-report-plugin</artifactId>
-                <version>${maven-surefire-plugin.version}</version>
-            </plugin>
-            <!-- Test coverage -->
-            <plugin>
-                <groupId>org.codehaus.mojo</groupId>
-                <artifactId>cobertura-maven-plugin</artifactId>
-                <version>2.7</version>
-            </plugin>
-        </plugins>
-    </reporting>
 </project>

From 378669cc102165f24bc9df8f77ad3380a804ec0d Mon Sep 17 00:00:00 2001
From: Ryan Nett <rnett@skymind.io>
Date: Thu, 29 Aug 2019 21:28:34 -0700
Subject: [PATCH 43/56] Fix MNIST Fetcher to not re-allocate each batch (#200)

* don't allocate so many float arrays, use INDArrays instead

Signed-off-by: Ryan Nett <rnett@skymind.io>

* re-add pre-processing, better names

Signed-off-by: Ryan Nett <rnett@skymind.io>

* use float[][] pool to avoid extra ndarray creation

Signed-off-by: Ryan Nett <rnett@skymind.io>
---
 .../datasets/fetchers/MnistDataFetcher.java   | 49 +++++++++++--------
 1 file changed, 28 insertions(+), 21 deletions(-)

diff --git a/deeplearning4j/deeplearning4j-data/deeplearning4j-datasets/src/main/java/org/deeplearning4j/datasets/fetchers/MnistDataFetcher.java b/deeplearning4j/deeplearning4j-data/deeplearning4j-datasets/src/main/java/org/deeplearning4j/datasets/fetchers/MnistDataFetcher.java
index bd8726fed..3be8401d7 100755
--- a/deeplearning4j/deeplearning4j-data/deeplearning4j-datasets/src/main/java/org/deeplearning4j/datasets/fetchers/MnistDataFetcher.java
+++ b/deeplearning4j/deeplearning4j-data/deeplearning4j-datasets/src/main/java/org/deeplearning4j/datasets/fetchers/MnistDataFetcher.java
@@ -22,10 +22,12 @@ import org.deeplearning4j.base.MnistFetcher;
 import org.deeplearning4j.common.resources.DL4JResources;
 import org.deeplearning4j.common.resources.ResourceType;
 import org.deeplearning4j.datasets.mnist.MnistManager;
+import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.DataSet;
 import org.nd4j.linalg.dataset.api.iterator.fetcher.BaseDataFetcher;
 import org.nd4j.linalg.factory.Nd4j;
+import org.nd4j.linalg.indexing.NDArrayIndex;
 import org.nd4j.linalg.util.MathUtils;
 
 import java.io.File;
@@ -167,14 +169,19 @@ public class MnistDataFetcher extends BaseDataFetcher {
         this(true);
     }
 
+    private float[][] featureData = null;
+
     @Override
     public void fetch(int numExamples) {
         if (!hasMore()) {
             throw new IllegalStateException("Unable to get more; there are no more images");
         }
 
-        float[][] featureData = new float[numExamples][0];
-        float[][] labelData = new float[numExamples][0];
+        INDArray labels = Nd4j.zeros(DataType.FLOAT, numExamples, numOutcomes);
+
+        if(featureData == null || featureData.length < numExamples){
+            featureData = new float[numExamples][28*28];
+        }
 
         int actualExamples = 0;
         byte[] working = null;
@@ -202,33 +209,33 @@ public class MnistDataFetcher extends BaseDataFetcher {
                 label--;
             }
 
-            float[] featureVec = new float[img.length];
-            featureData[actualExamples] = featureVec;
-            labelData[actualExamples] = new float[numOutcomes];
-            labelData[actualExamples][label] = 1.0f;
+            labels.put(actualExamples, label, 1.0f);
 
-            for (int j = 0; j < img.length; j++) {
-                float v = ((int) img[j]) & 0xFF; //byte is loaded as signed -> convert to unsigned
-                if (binarize) {
-                    if (v > 30.0f)
-                        featureVec[j] = 1.0f;
-                    else
-                        featureVec[j] = 0.0f;
-                } else {
-                    featureVec[j] = v / 255.0f;
-                }
+            for(int j = 0 ; j < img.length ; j++) {
+                featureData[actualExamples][j] = ((int) img[j]) & 0xFF;
             }
 
             actualExamples++;
         }
 
-        if (actualExamples < numExamples) {
-            featureData = Arrays.copyOfRange(featureData, 0, actualExamples);
-            labelData = Arrays.copyOfRange(labelData, 0, actualExamples);
+        INDArray features;
+
+        if(featureData.length == actualExamples){
+            features = Nd4j.create(featureData);
+        } else {
+            features = Nd4j.create(Arrays.copyOfRange(featureData, 0, actualExamples));
+        }
+
+        if (actualExamples < numExamples) {
+            labels = labels.get(NDArrayIndex.interval(0, actualExamples), NDArrayIndex.all());
+        }
+
+        if(binarize){
+            features = features.gt(30.0).castTo(DataType.FLOAT);
+        } else {
+            features.divi(255.0);
         }
 
-        INDArray features = Nd4j.create(featureData);
-        INDArray labels = Nd4j.create(labelData);
         curr = new DataSet(features, labels);
     }
 

From dcc2baa6766fb5236a0afae7984c36cddcb80903 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Fri, 30 Aug 2019 14:35:27 +1000
Subject: [PATCH 44/56] Version upgrades (#199)

* DataVec fixes for Jackson version upgrade

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J jackson updates + databind version 2.9.9.3

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Shade snakeyaml along with jackson

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Version fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Switch DataVec legacy JSON format handling to mixins

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Next set of fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup for legacy JSON mapping

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Upgrade commons compress to 1.18; small test fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* New Jackson backward compatibility for DL4J - Round 1

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* New Jackson backward compatibility for DL4J - Round 2

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More fixes, all but legacy custom passing

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Provide an upgrade path for custom layers for models in pre-1.0.0-beta JSON format

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Legacy deserialization cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small amount of polish - legacy JSON

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Upgrade guava version

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* IEvaluation legacy format deserialization fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Upgrade play version to 2.7.3

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Update nd4j-parameter-server-status to new Play API

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Update DL4J UI for new play version

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More play framework updates

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Remove Spark 1/2 adapter code from DataVec

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* datavec-spark dependency cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J spark updates, pt 1

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J spark updates, pt 2

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J spark updates, pt 3

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J spark updates, pt 4

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Test fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Another fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Breeze upgrade, dependency cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Add Scala 2.12 version to pom.xml

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* change-scala-versions.sh - add scala 2.12, remove 2.10

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Move Spark version properties to parent pom (now that only one spark version is supported)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DataVec Play fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* datavec play dependency fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Clean up old spark/jackson stuff

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup jackson unused dependencies

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Dropping redundant dependency

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* Removed scalaxy dependency

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* DataVec fixes for Jackson version upgrade

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J jackson updates + databind version 2.9.9.3

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Shade snakeyaml along with jackson

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Version fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Switch DataVec legacy JSON format handling to mixins

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Next set of fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup for legacy JSON mapping

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Upgrade commons compress to 1.18; small test fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* New Jackson backward compatibility for DL4J - Round 1

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* New Jackson backward compatibility for DL4J - Round 2

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More fixes, all but legacy custom passing

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Provide an upgrade path for custom layers for models in pre-1.0.0-beta JSON format

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Legacy deserialization cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small amount of polish - legacy JSON

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Upgrade guava version

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* IEvaluation legacy format deserialization fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Upgrade play version to 2.7.3

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Update nd4j-parameter-server-status to new Play API

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Update DL4J UI for new play version

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* More play framework updates

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Remove Spark 1/2 adapter code from DataVec

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* datavec-spark dependency cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J spark updates, pt 1

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J spark updates, pt 2

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J spark updates, pt 3

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DL4J spark updates, pt 4

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Test fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Another fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Breeze upgrade, dependency cleanup

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Add Scala 2.12 version to pom.xml

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* change-scala-versions.sh - add scala 2.12, remove 2.10

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Move Spark version properties to parent pom (now that only one spark version is supported)

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DataVec Play fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* datavec play dependency fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Clean up old spark/jackson stuff

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Cleanup jackson unused dependencies

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Add shaded guava

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Dropping redundant dependency

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* Removed scalaxy dependency

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* Ensure not possible to import pre-shaded classes, and remove direct guava dependencies in favor of shaded

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* ND4J Shaded guava import fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* DataVec and DL4J guava shading

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Arbiter, RL4J fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Build fixed

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* Fix dependency

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* Fix bad merge

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Jackson shading fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Set play secret, datavec-spark-inference-server

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix for datavec-spark-inference-server

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Arbiter fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Arbiter fixes

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Small test fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 arbiter/arbiter-core/pom.xml                  |   5 +
 .../distribution/LogUniformDistribution.java  |   2 +-
 .../runner/BaseOptimizationRunner.java        |   2 +-
 .../runner/LocalOptimizationRunner.java       |   6 +-
 .../optimize/serde/jackson/JsonMapper.java    |  14 +-
 .../optimize/serde/jackson/YamlMapper.java    |   1 +
 .../arbiter/optimize/TestJson.java            |   1 +
 arbiter/arbiter-deeplearning4j/pom.xml        |   6 +
 .../arbiter/layers/BaseLayerSpace.java        |   2 +-
 arbiter/arbiter-ui/pom.xml                    |   6 -
 .../arbiter/ui/misc/JsonMapper.java           |  10 +-
 change-scala-versions.sh                      |  34 +-
 change-spark-versions.sh                      |  85 ---
 .../org/datavec/api/transform/Transform.java  |   4 +-
 .../api/transform/TransformProcess.java       |  11 +
 .../api/transform/analysis/DataAnalysis.java  |  12 +
 .../analysis/SequenceDataAnalysis.java        |  13 +-
 .../analysis/columns/ColumnAnalysis.java      |   4 +-
 .../api/transform/condition/Condition.java    |   4 +-
 .../datavec/api/transform/filter/Filter.java  |   4 +-
 .../transform/metadata/ColumnMetaData.java    |   4 +-
 .../ops/DispatchWithConditionOp.java          |   4 +-
 .../transform/rank/CalculateSortedRank.java   |   4 +-
 .../datavec/api/transform/schema/Schema.java  |  31 +-
 .../sequence/SequenceComparator.java          |   4 +-
 .../api/transform/sequence/SequenceSplit.java |   4 +-
 .../sequence/window/WindowFunction.java       |   4 +-
 .../api/transform/serde/JsonMappers.java      | 239 +-------
 .../legacy/GenericLegacyDeserializer.java     |  41 --
 .../serde/legacy/LegacyJsonFormat.java        | 267 +++++++++
 .../serde/legacy/LegacyMappingHelper.java     | 535 ------------------
 .../stringreduce/IStringReducer.java          |   4 +-
 .../api/util/ndarray/RecordConverter.java     |   2 +-
 .../datavec/api/writable/ByteWritable.java    |   2 +-
 .../datavec/api/writable/DoubleWritable.java  |   2 +-
 .../datavec/api/writable/FloatWritable.java   |   2 +-
 .../org/datavec/api/writable/IntWritable.java |   2 +-
 .../datavec/api/writable/LongWritable.java    |   2 +-
 .../org/datavec/api/writable/Writable.java    |   4 +-
 .../writable/batch/NDArrayRecordBatch.java    |   2 +-
 .../comparator/WritableComparator.java        |   5 +-
 .../datavec/api/split/InputSplitTests.java    |   2 +-
 .../api/split/parittion/PartitionerTests.java |   2 +-
 .../api/transform/schema/TestJsonYaml.java    |   5 +-
 .../api/writable/RecordConverterTest.java     |   2 +-
 datavec/datavec-arrow/pom.xml                 |  35 --
 .../recordreader/BaseImageRecordReader.java   |   2 +-
 .../image/serde/LegacyImageMappingHelper.java |  35 --
 .../image/transform/ImageTransform.java       |   6 +-
 datavec/datavec-hadoop/pom.xml                |   5 -
 .../reader/TestMapFileRecordReader.java       |   2 +-
 .../TestMapFileRecordReaderMultipleParts.java |   2 +-
 ...ileRecordReaderMultiplePartsSomeEmpty.java |   2 +-
 .../writer/TestMapFileRecordWriter.java       |   2 +-
 ...JoinFromCoGroupFlatMapFunctionAdapter.java |   2 +-
 .../datavec-spark-inference-server/pom.xml    |  58 +-
 .../transform/CSVSparkTransformServer.java    |  81 +--
 .../datavec/spark/transform/FunctionUtil.java |  41 --
 .../transform/ImageSparkTransformServer.java  |  62 +-
 .../spark/transform/SparkTransformServer.java |  21 +-
 .../src/main/resources/application.conf       | 350 ++++++++++++
 datavec/datavec-spark/pom.xml                 | 169 +-----
 .../functions/FlatMapFunctionAdapter.java     |  29 -
 .../datavec/spark/transform/DataFrames.java   |  48 +-
 .../spark/transform/Normalization.java        |  59 +-
 .../analysis/SequenceFlatMapFunction.java     |  10 +-
 .../SequenceFlatMapFunctionAdapter.java       |  36 --
 ...ExecuteJoinFromCoGroupFlatMapFunction.java |  90 ++-
 ...JoinFromCoGroupFlatMapFunctionAdapter.java | 119 ----
 .../join/FilterAndFlattenJoinedValues.java    |  41 +-
 .../FilterAndFlattenJoinedValuesAdapter.java  |  71 ---
 .../sparkfunction/SequenceToRows.java         |  56 +-
 .../sparkfunction/SequenceToRowsAdapter.java  |  87 ---
 .../transform/SequenceSplitFunction.java      |  14 +-
 .../SequenceSplitFunctionAdapter.java         |  41 --
 .../SparkTransformProcessFunction.java        |  19 +-
 .../SparkTransformProcessFunctionAdapter.java |  45 --
 .../transform/BaseFlatMapFunctionAdaptee.java |  41 --
 .../spark/transform/DataRowsFacade.java       |  42 --
 .../transform/BaseFlatMapFunctionAdaptee.java |  42 --
 .../spark/transform/DataRowsFacade.java       |  43 --
 .../spark/storage/TestSparkStorageUtils.java  |   2 +-
 .../spark/transform/DataFramesTests.java      |  16 +-
 .../spark/transform/NormalizationTests.java   |  34 +-
 .../config/DL4JSystemProperties.java          |  12 -
 deeplearning4j/deeplearning4j-core/pom.xml    |  12 -
 .../RecordReaderDataSetiteratorTest.java      |   2 +-
 .../RecordReaderMultiDataSetIteratorTest.java |   2 +-
 .../deeplearning4j/nn/dtypes/DTypeTests.java  |   6 +-
 .../nn/layers/recurrent/TestRnnLayers.java    |   4 +-
 .../plot/BarnesHutTsneTest.java               |   2 +-
 .../regressiontest/RegressionTest100a.java    |  79 +--
 .../iterator/MultipleEpochsIterator.java      |   4 +-
 .../FileSplitParallelDataSetIterator.java     |   2 +-
 .../deeplearning4j/plot/BarnesHutTsne.java    |   2 +-
 .../java/org/deeplearning4j/plot/Tsne.java    |   2 +-
 .../deeplearning4j-modelimport/pom.xml        |   6 +
 .../pom.xml                                   |  20 -
 .../nearestneighbor/server/FunctionUtil.java  |  41 --
 .../server/NearestNeighborsServer.java        |  53 +-
 .../nearestneighbor-core/pom.xml              |   7 +
 .../clustering/info/ClusterSetInfo.java       |   4 +-
 .../clustering/quadtree/QuadTree.java         |   2 +-
 .../clustering/randomprojection/RPUtils.java  |   2 +-
 .../clustering/sptree/SpTree.java             |   2 +-
 .../clustering/kdtree/KDTreeTest.java         |   2 +-
 .../clustering/sptree/SPTreeTest.java         |   2 +-
 .../clustering/vptree/VpTreeNodeTest.java     |   3 -
 .../text/corpora/sentiwordnet/SWN3.java       |   2 +-
 .../models/WordVectorSerializerTest.java      |   4 +-
 .../models/word2vec/Word2VecTests.java        |   4 +-
 .../inmemory/InMemoryLookupTable.java         |   2 +-
 .../reader/impl/BasicModelUtils.java          |   2 +-
 .../wordvectors/WordVectorsImpl.java          |   2 +-
 .../models/glove/count/CountMap.java          |   2 +-
 .../paragraphvectors/ParagraphVectors.java    |   2 +-
 .../sequencevectors/SequenceVectors.java      |   4 +-
 .../sequence/SequenceElement.java             |   2 +-
 .../text/invertedindex/InvertedIndex.java     |   2 +-
 .../wordvectors/WordVectorsImplTest.java      |   2 +-
 .../deeplearning4j/eval/ConfusionMatrix.java  |   4 +-
 .../eval/curves/PrecisionRecallCurve.java     |   2 +-
 .../deeplearning4j/eval/curves/RocCurve.java  |   2 +-
 .../conf/ComputationGraphConfiguration.java   |  22 +
 .../nn/conf/InputPreProcessor.java            |   4 +-
 .../nn/conf/MultiLayerConfiguration.java      |  22 +
 .../nn/conf/NeuralNetConfiguration.java       |  94 +--
 .../nn/conf/graph/AttentionVertex.java        |   2 +-
 .../nn/conf/graph/GraphVertex.java            |   4 +-
 .../nn/conf/inputs/InputType.java             |  58 +-
 .../deeplearning4j/nn/conf/layers/Layer.java  |   4 +-
 .../nn/conf/layers/Upsampling2D.java          |   2 +-
 .../nn/conf/layers/misc/FrozenLayer.java      |   2 -
 .../ReconstructionDistribution.java           |   4 +-
 .../conf/serde/BaseNetConfigDeserializer.java |  82 ++-
 ...utationGraphConfigurationDeserializer.java |  11 +
 .../conf/serde/FrozenLayerDeserializer.java   |  58 --
 .../nn/conf/serde/JsonMappers.java            | 148 +----
 .../MultiLayerConfigurationDeserializer.java  |  55 +-
 .../LegacyIntArrayDeserializer.java           |   2 +-
 .../conf/serde/legacy/LegacyJsonFormat.java   | 175 ++++++
 .../LegacyGraphVertexDeserializer.java        |  94 ---
 .../LegacyGraphVertexDeserializerHelper.java  |  28 -
 .../legacyformat/LegacyLayerDeserializer.java | 113 ----
 .../LegacyLayerDeserializerHelper.java        |  28 -
 .../LegacyPreprocessorDeserializer.java       |  83 ---
 .../LegacyPreprocessorDeserializerHelper.java |  28 -
 ...econstructionDistributionDeserializer.java |  70 ---
 ...ructionDistributionDeserializerHelper.java |  28 -
 .../nn/workspace/LayerWorkspaceMgr.java       |   2 +-
 .../listeners/CheckpointListener.java         |   2 +-
 .../listeners/PerformanceListener.java        |   2 +-
 .../solvers/accumulation/EncodingHandler.java |   2 +-
 .../deeplearning4j/util/ModelSerializer.java  |   2 +-
 .../deeplearning4j-aws/pom.xml                |  27 -
 .../pom.xml                                   |  16 -
 .../EarlyStoppingParallelTrainer.java         |   2 +-
 .../observers/BasicInferenceObservable.java   |   2 +-
 .../spark/dl4j-spark-nlp-java8/pom.xml        |  36 --
 .../functions/VocabRddFunctionFlat.java       |  91 ++-
 .../word2vec/FirstIterationFunction.java      | 240 +++++++-
 .../FirstIterationFunctionAdapter.java        | 265 ---------
 .../word2vec/SecondIterationFunction.java     |  26 +-
 .../spark/dl4j-spark-parameterserver/pom.xml  |  36 --
 .../functions/SharedFlatMapDataSet.java       |  21 +-
 .../functions/SharedFlatMapMultiDataSet.java  |  23 +-
 .../functions/SharedFlatMapPaths.java         |  20 +-
 .../functions/SharedFlatMapPathsMDS.java      |  21 +-
 .../api/worker/ExecuteWorkerFlatMap.java      |  37 +-
 .../ExecuteWorkerMultiDataSetFlatMap.java     |  38 +-
 .../api/worker/ExecuteWorkerPDSFlatMap.java   |  28 +-
 .../worker/ExecuteWorkerPDSMDSFlatMap.java    |  28 +-
 .../api/worker/ExecuteWorkerPathFlatMap.java  |  28 +-
 .../worker/ExecuteWorkerPathMDSFlatMap.java   |  30 +-
 .../spark/data/BatchDataSetsFunction.java     |  37 +-
 .../spark/data/SplitDataSetsFunction.java     |  30 +-
 ...litDataSetExamplesPairFlatMapFunction.java |  30 +-
 .../spark/datavec/RDDMiniBatches.java         |  27 +-
 .../HashingBalancedPartitioner.java           |  21 +-
 .../repartition/MapTupleToPairFlatMap.java    |  17 +-
 ...VaeReconstructionProbWithKeyFunction.java} |   6 +-
 ....java => BaseVaeScoreWithKeyFunction.java} |  14 +-
 .../IEvaluateMDSFlatMapFunction.java          |  37 +-
 .../IEvaluateMDSPathsFlatMapFunction.java     |  34 +-
 ...VaeReconstructionErrorWithKeyFunction.java |   4 +-
 ...GVaeReconstructionProbWithKeyFunction.java |   4 +-
 .../GraphFeedForwardWithKeyFunction.java      |  49 +-
 .../graph/scoring/ScoreExamplesFunction.java  |  33 +-
 .../scoring/ScoreExamplesWithKeyFunction.java |  39 +-
 .../ScoreFlatMapFunctionCGDataSet.java        |  25 +-
 .../ScoreFlatMapFunctionCGMultiDataSet.java   |  25 +-
 .../evaluation/IEvaluateFlatMapFunction.java  |  33 +-
 .../scoring/FeedForwardWithKeyFunction.java   |  32 +-
 .../scoring/ScoreExamplesFunction.java        |  30 +-
 .../scoring/ScoreExamplesWithKeyFunction.java |  42 +-
 .../scoring/ScoreFlatMapFunction.java         |  36 +-
 ...VaeReconstructionErrorWithKeyFunction.java |  26 +-
 .../VaeReconstructionProbWithKeyFunction.java |  28 +-
 .../ParameterAveragingTrainingMaster.java     |   2 +-
 .../BaseDoubleFlatMapFunctionAdaptee.java     |  40 --
 .../util/BasePairFlatMapFunctionAdaptee.java  |  41 --
 .../BaseDoubleFlatMapFunctionAdaptee.java     |  42 --
 .../util/BasePairFlatMapFunctionAdaptee.java  |  43 --
 .../deeplearning4j-scaleout/spark/pom.xml     |  52 --
 .../deeplearning4j-play/pom.xml               |  85 ---
 .../deeplearning4j/ui/api/FunctionType.java   |  11 +-
 .../java/org/deeplearning4j/ui/api/Route.java |  17 +-
 .../ui/module/tsne/TsneModule.java            |  29 +-
 .../deeplearning4j/ui/play/PlayUIServer.java  | 143 ++---
 .../ui/play/misc/FunctionUtil.java            |  45 --
 .../ui/play/staticroutes/Assets.java          |  26 +-
 .../IntegrationTestBaselineGenerator.java     |   2 +-
 .../integration/IntegrationTestRunner.java    |   4 +-
 .../integration/testcases/RNNTestCases.java   |   2 +-
 deeplearning4j/pom.xml                        |   5 -
 .../autodiff/listeners/ListenerVariables.java |   2 +-
 .../checkpoint/CheckpointListener.java        |   2 +-
 .../listeners/records/EvaluationRecord.java   |   8 +-
 .../org/nd4j/autodiff/samediff/SameDiff.java  |  10 +-
 .../nd4j/autodiff/samediff/ops/SDBaseOps.java |   2 +-
 .../samediff/serde/FlatBuffersMapper.java     |   2 +-
 .../autodiff/validation/OpValidation.java     |   6 +-
 .../org/nd4j/evaluation/BaseEvaluation.java   |  12 +-
 .../classification/ConfusionMatrix.java       |   4 +-
 .../evaluation/custom/CustomEvaluation.java   |   2 +-
 .../nd4j/evaluation/custom/MergeLambda.java   |   2 +-
 .../serde/ConfusionMatrixSerializer.java      |   2 +-
 .../graphmapper/onnx/OnnxGraphMapper.java     |   6 +-
 .../imports/graphmapper/tf/TFGraphMapper.java |   4 +-
 .../nd4j/linalg/api/ndarray/BaseNDArray.java  |   4 +-
 .../linalg/api/ndarray/BaseSparseNDArray.java |   4 +-
 .../api/ndarray/BaseSparseNDArrayCOO.java     |   6 +-
 .../api/ndarray/BaseSparseNDArrayCSR.java     |   2 +-
 .../org/nd4j/linalg/api/ops/BaseReduceOp.java |   2 +-
 .../nd4j/linalg/api/ops/DynamicCustomOp.java  |   6 +-
 .../nd4j/linalg/api/ops/aggregates/Batch.java |   2 +-
 .../ops/impl/controlflow/compat/Switch.java   |   2 +-
 .../api/ops/impl/reduce/TensorMmul.java       |   4 +-
 .../linalg/api/ops/impl/shape/Transpose.java  |   2 +-
 .../ops/impl/transforms/custom/Choose.java    |   4 +-
 .../java/org/nd4j/linalg/api/shape/Shape.java |   6 +-
 .../linalg/dataset/BalanceMinibatches.java    |   4 +-
 .../java/org/nd4j/linalg/dataset/DataSet.java |   4 +-
 .../org/nd4j/linalg/dataset/api/DataSet.java  |   2 +-
 .../linalg/factory/BaseNDArrayFactory.java    |   2 +-
 .../java/org/nd4j/linalg/factory/Nd4j.java    |   4 +-
 .../org/nd4j/linalg/indexing/Indices.java     |   4 +-
 .../nd4j/linalg/indexing/IntervalIndex.java   |   2 +-
 .../nd4j/linalg/indexing/NDArrayIndex.java    |   4 +-
 .../org/nd4j/linalg/indexing/PointIndex.java  |   2 +-
 .../nd4j/linalg/indexing/SpecifiedIndex.java  |   2 +-
 .../linalg/indexing/conditions/Condition.java |   2 +-
 .../linalg/indexing/functions/Identity.java   |   2 +-
 .../indexing/functions/StableNumber.java      |   2 +-
 .../nd4j/linalg/indexing/functions/Value.java |   2 +-
 .../nd4j/linalg/indexing/functions/Zero.java  |   2 +-
 .../org/nd4j/jita/handler/MemoryHandler.java  |   2 +-
 .../jita/handler/impl/CudaZeroHandler.java    |   4 +-
 .../nd4j/linalg/jcublas/util/CudaArgs.java    |   4 +-
 .../linalg/api/indexing/IndexingTestsC.java   |   1 -
 nd4j/nd4j-common/pom.xml                      |  11 +-
 .../linalg/collection/IntArrayKeyMap.java     |   2 +-
 .../nd4j/linalg/primitives/AtomicDouble.java  |   4 +-
 .../java/org/nd4j/linalg/util/ArrayUtil.java  |   4 +-
 .../nd4j/linalg/util/SynchronizedTable.java   |   2 +-
 .../nd4j/resources/strumpf/ResourceFile.java  |   2 +-
 .../transport/RoutedTransport.java            |   2 +-
 .../v2/transport/impl/AeronUdpTransport.java  |   2 +-
 .../nd4j-parameter-server-status/pom.xml      |  10 -
 .../status/play/StatusServer.java             | 102 +---
 .../ParameterServerSubscriber.java            |   2 +-
 .../ipc/chunk/InMemoryChunkAccumulator.java   |   2 +-
 nd4j/nd4j-serde/nd4j-arrow/pom.xml            |  30 -
 nd4j/nd4j-serde/nd4j-gson/pom.xml             |   3 -
 .../serde/gson/GsonDeserializationUtils.java  |   4 +-
 nd4j/nd4j-shade/guava/pom.xml                 | 219 +++++++
 nd4j/nd4j-shade/jackson/pom.xml               | 131 +++--
 nd4j/nd4j-shade/pom.xml                       |   1 +
 nd4j/nd4j-shade/protobuf/pom.xml              |  21 +
 nd4j/pom.xml                                  |   5 -
 nd4s/build.sbt                                |   2 +-
 nd4s/pom.xml                                  |  14 +-
 .../org/nd4s/CollectionLikeNDArray.scala      |   5 +-
 pom.xml                                       |  31 +-
 rl4j/rl4j-core/pom.xml                        |   6 +
 scalnet/pom.xml                               |   6 +
 286 files changed, 2774 insertions(+), 5074 deletions(-)
 delete mode 100755 change-spark-versions.sh
 delete mode 100644 datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/GenericLegacyDeserializer.java
 create mode 100644 datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/LegacyJsonFormat.java
 delete mode 100644 datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/LegacyMappingHelper.java
 delete mode 100644 datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/serde/LegacyImageMappingHelper.java
 delete mode 100644 datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/FunctionUtil.java
 create mode 100644 datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/resources/application.conf
 delete mode 100644 datavec/datavec-spark/src/main/java/org/datavec/spark/functions/FlatMapFunctionAdapter.java
 delete mode 100644 datavec/datavec-spark/src/main/java/org/datavec/spark/transform/analysis/SequenceFlatMapFunctionAdapter.java
 delete mode 100644 datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/ExecuteJoinFromCoGroupFlatMapFunctionAdapter.java
 delete mode 100644 datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/FilterAndFlattenJoinedValuesAdapter.java
 delete mode 100644 datavec/datavec-spark/src/main/java/org/datavec/spark/transform/sparkfunction/SequenceToRowsAdapter.java
 delete mode 100644 datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SequenceSplitFunctionAdapter.java
 delete mode 100644 datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SparkTransformProcessFunctionAdapter.java
 delete mode 100644 datavec/datavec-spark/src/main/spark-1/org/datavec/spark/transform/BaseFlatMapFunctionAdaptee.java
 delete mode 100644 datavec/datavec-spark/src/main/spark-1/org/datavec/spark/transform/DataRowsFacade.java
 delete mode 100644 datavec/datavec-spark/src/main/spark-2/org/datavec/spark/transform/BaseFlatMapFunctionAdaptee.java
 delete mode 100644 datavec/datavec-spark/src/main/spark-2/org/datavec/spark/transform/DataRowsFacade.java
 delete mode 100644 deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/src/main/java/org/deeplearning4j/nearestneighbor/server/FunctionUtil.java
 delete mode 100644 deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/FrozenLayerDeserializer.java
 rename deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/{legacyformat => legacy}/LegacyIntArrayDeserializer.java (97%)
 create mode 100644 deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacy/LegacyJsonFormat.java
 delete mode 100644 deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyGraphVertexDeserializer.java
 delete mode 100644 deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyGraphVertexDeserializerHelper.java
 delete mode 100644 deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyLayerDeserializer.java
 delete mode 100644 deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyLayerDeserializerHelper.java
 delete mode 100644 deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyPreprocessorDeserializer.java
 delete mode 100644 deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyPreprocessorDeserializerHelper.java
 delete mode 100644 deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyReconstructionDistributionDeserializer.java
 delete mode 100644 deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyReconstructionDistributionDeserializerHelper.java
 delete mode 100644 deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/FirstIterationFunctionAdapter.java
 rename deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/{BaseVaeReconstructionProbWithKeyFunctionAdapter.java => BaseVaeReconstructionProbWithKeyFunction.java} (87%)
 rename deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/{BaseVaeScoreWithKeyFunctionAdapter.java => BaseVaeScoreWithKeyFunction.java} (88%)
 delete mode 100644 deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-1/java/org/deeplearning4j/spark/util/BaseDoubleFlatMapFunctionAdaptee.java
 delete mode 100644 deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-1/java/org/deeplearning4j/spark/util/BasePairFlatMapFunctionAdaptee.java
 delete mode 100644 deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-2/java/org/deeplearning4j/spark/util/BaseDoubleFlatMapFunctionAdaptee.java
 delete mode 100644 deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-2/java/org/deeplearning4j/spark/util/BasePairFlatMapFunctionAdaptee.java
 delete mode 100644 deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/misc/FunctionUtil.java
 create mode 100644 nd4j/nd4j-shade/guava/pom.xml

diff --git a/arbiter/arbiter-core/pom.xml b/arbiter/arbiter-core/pom.xml
index 296af48fd..064dd3ecd 100644
--- a/arbiter/arbiter-core/pom.xml
+++ b/arbiter/arbiter-core/pom.xml
@@ -72,6 +72,11 @@
             <scope>test</scope>
         </dependency>
 
+        <dependency>
+            <groupId>joda-time</groupId>
+            <artifactId>joda-time</artifactId>
+            <version>${jodatime.version}</version>
+        </dependency>
 
         <!-- ND4J Shaded Jackson Dependency -->
         <dependency>
diff --git a/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/distribution/LogUniformDistribution.java b/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/distribution/LogUniformDistribution.java
index 9e50065e6..a9c0933c4 100644
--- a/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/distribution/LogUniformDistribution.java
+++ b/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/distribution/LogUniformDistribution.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.arbiter.optimize.distribution;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import lombok.Getter;
 import org.apache.commons.math3.distribution.RealDistribution;
 import org.apache.commons.math3.exception.NumberIsTooLargeException;
diff --git a/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/runner/BaseOptimizationRunner.java b/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/runner/BaseOptimizationRunner.java
index fa503ef6d..0e04a130b 100644
--- a/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/runner/BaseOptimizationRunner.java
+++ b/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/runner/BaseOptimizationRunner.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.arbiter.optimize.runner;
 
-import com.google.common.util.concurrent.ListenableFuture;
+import org.nd4j.shade.guava.util.concurrent.ListenableFuture;
 import lombok.AllArgsConstructor;
 import lombok.Data;
 import lombok.extern.slf4j.Slf4j;
diff --git a/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/runner/LocalOptimizationRunner.java b/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/runner/LocalOptimizationRunner.java
index a3992b09a..6982090f1 100644
--- a/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/runner/LocalOptimizationRunner.java
+++ b/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/runner/LocalOptimizationRunner.java
@@ -16,9 +16,9 @@
 
 package org.deeplearning4j.arbiter.optimize.runner;
 
-import com.google.common.util.concurrent.ListenableFuture;
-import com.google.common.util.concurrent.ListeningExecutorService;
-import com.google.common.util.concurrent.MoreExecutors;
+import org.nd4j.shade.guava.util.concurrent.ListenableFuture;
+import org.nd4j.shade.guava.util.concurrent.ListeningExecutorService;
+import org.nd4j.shade.guava.util.concurrent.MoreExecutors;
 import lombok.Setter;
 import org.deeplearning4j.arbiter.optimize.api.*;
 import org.deeplearning4j.arbiter.optimize.api.data.DataProvider;
diff --git a/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/serde/jackson/JsonMapper.java b/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/serde/jackson/JsonMapper.java
index 9e30a06f6..8cfb07723 100644
--- a/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/serde/jackson/JsonMapper.java
+++ b/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/serde/jackson/JsonMapper.java
@@ -43,13 +43,15 @@ public class JsonMapper {
         mapper.enable(SerializationFeature.INDENT_OUTPUT);
         mapper.setVisibility(PropertyAccessor.ALL, JsonAutoDetect.Visibility.NONE);
         mapper.setVisibility(PropertyAccessor.FIELD, JsonAutoDetect.Visibility.ANY);
+        mapper.setVisibility(PropertyAccessor.CREATOR, JsonAutoDetect.Visibility.ANY);
         yamlMapper = new ObjectMapper(new YAMLFactory());
-        mapper.registerModule(new JodaModule());
-        mapper.configure(DeserializationFeature.FAIL_ON_UNKNOWN_PROPERTIES, false);
-        mapper.configure(SerializationFeature.FAIL_ON_EMPTY_BEANS, false);
-        mapper.enable(SerializationFeature.INDENT_OUTPUT);
-        mapper.setVisibility(PropertyAccessor.ALL, JsonAutoDetect.Visibility.NONE);
-        mapper.setVisibility(PropertyAccessor.FIELD, JsonAutoDetect.Visibility.ANY);
+        yamlMapper.registerModule(new JodaModule());
+        yamlMapper.configure(DeserializationFeature.FAIL_ON_UNKNOWN_PROPERTIES, false);
+        yamlMapper.configure(SerializationFeature.FAIL_ON_EMPTY_BEANS, false);
+        yamlMapper.enable(SerializationFeature.INDENT_OUTPUT);
+        yamlMapper.setVisibility(PropertyAccessor.ALL, JsonAutoDetect.Visibility.NONE);
+        yamlMapper.setVisibility(PropertyAccessor.FIELD, JsonAutoDetect.Visibility.ANY);
+        yamlMapper.setVisibility(PropertyAccessor.CREATOR, JsonAutoDetect.Visibility.ANY);
     }
 
     private JsonMapper() {}
diff --git a/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/serde/jackson/YamlMapper.java b/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/serde/jackson/YamlMapper.java
index f10c593e0..5b35220e9 100644
--- a/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/serde/jackson/YamlMapper.java
+++ b/arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/serde/jackson/YamlMapper.java
@@ -39,6 +39,7 @@ public class YamlMapper {
         mapper.enable(SerializationFeature.INDENT_OUTPUT);
         mapper.setVisibility(PropertyAccessor.ALL, JsonAutoDetect.Visibility.NONE);
         mapper.setVisibility(PropertyAccessor.FIELD, JsonAutoDetect.Visibility.ANY);
+        mapper.setVisibility(PropertyAccessor.CREATOR, JsonAutoDetect.Visibility.ANY);
     }
 
 
diff --git a/arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/TestJson.java b/arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/TestJson.java
index 3572b187b..6f1b336bb 100644
--- a/arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/TestJson.java
+++ b/arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/TestJson.java
@@ -59,6 +59,7 @@ public class TestJson {
         om.enable(SerializationFeature.INDENT_OUTPUT);
         om.setVisibility(PropertyAccessor.ALL, JsonAutoDetect.Visibility.NONE);
         om.setVisibility(PropertyAccessor.FIELD, JsonAutoDetect.Visibility.ANY);
+        om.setVisibility(PropertyAccessor.CREATOR, JsonAutoDetect.Visibility.ANY);
         return om;
     }
 
diff --git a/arbiter/arbiter-deeplearning4j/pom.xml b/arbiter/arbiter-deeplearning4j/pom.xml
index 77f7e34a9..85afe7a6b 100644
--- a/arbiter/arbiter-deeplearning4j/pom.xml
+++ b/arbiter/arbiter-deeplearning4j/pom.xml
@@ -57,6 +57,12 @@
             <artifactId>jackson</artifactId>
             <version>${nd4j.version}</version>
         </dependency>
+
+        <dependency>
+            <groupId>com.google.code.gson</groupId>
+            <artifactId>gson</artifactId>
+            <version>${gson.version}</version>
+        </dependency>
     </dependencies>
 
     <profiles>
diff --git a/arbiter/arbiter-deeplearning4j/src/main/java/org/deeplearning4j/arbiter/layers/BaseLayerSpace.java b/arbiter/arbiter-deeplearning4j/src/main/java/org/deeplearning4j/arbiter/layers/BaseLayerSpace.java
index 77c31707a..0a5e33d27 100644
--- a/arbiter/arbiter-deeplearning4j/src/main/java/org/deeplearning4j/arbiter/layers/BaseLayerSpace.java
+++ b/arbiter/arbiter-deeplearning4j/src/main/java/org/deeplearning4j/arbiter/layers/BaseLayerSpace.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.arbiter.layers;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import lombok.AccessLevel;
 import lombok.Data;
 import lombok.EqualsAndHashCode;
diff --git a/arbiter/arbiter-ui/pom.xml b/arbiter/arbiter-ui/pom.xml
index 93e955219..56d1013bf 100644
--- a/arbiter/arbiter-ui/pom.xml
+++ b/arbiter/arbiter-ui/pom.xml
@@ -107,12 +107,6 @@
     </profiles>
 
     <dependencies>
-        <dependency>
-            <groupId>com.google.guava</groupId>
-            <artifactId>guava</artifactId>
-            <version>${guava.version}</version>
-        </dependency>
-
         <dependency>
             <groupId>org.deeplearning4j</groupId>
             <artifactId>arbiter-core</artifactId>
diff --git a/arbiter/arbiter-ui/src/main/java/org/deeplearning4j/arbiter/ui/misc/JsonMapper.java b/arbiter/arbiter-ui/src/main/java/org/deeplearning4j/arbiter/ui/misc/JsonMapper.java
index 491756ae7..0ac5ad383 100644
--- a/arbiter/arbiter-ui/src/main/java/org/deeplearning4j/arbiter/ui/misc/JsonMapper.java
+++ b/arbiter/arbiter-ui/src/main/java/org/deeplearning4j/arbiter/ui/misc/JsonMapper.java
@@ -18,6 +18,7 @@ package org.deeplearning4j.arbiter.ui.misc;
 
 
 import org.nd4j.shade.jackson.annotation.JsonAutoDetect;
+import org.nd4j.shade.jackson.annotation.PropertyAccessor;
 import org.nd4j.shade.jackson.core.JsonProcessingException;
 import org.nd4j.shade.jackson.databind.DeserializationFeature;
 import org.nd4j.shade.jackson.databind.MapperFeature;
@@ -45,12 +46,9 @@ public class JsonMapper {
         mapper.configure(SerializationFeature.FAIL_ON_EMPTY_BEANS, false);
         mapper.configure(MapperFeature.SORT_PROPERTIES_ALPHABETICALLY, true);
         mapper.enable(SerializationFeature.INDENT_OUTPUT);
-
-        mapper.setVisibilityChecker(mapper.getSerializationConfig().getDefaultVisibilityChecker()
-                .withFieldVisibility(JsonAutoDetect.Visibility.ANY)
-                .withGetterVisibility(JsonAutoDetect.Visibility.NONE)
-                .withSetterVisibility(JsonAutoDetect.Visibility.NONE)
-                .withCreatorVisibility(JsonAutoDetect.Visibility.NONE));
+        mapper.setVisibility(PropertyAccessor.ALL, JsonAutoDetect.Visibility.NONE);
+        mapper.setVisibility(PropertyAccessor.FIELD, JsonAutoDetect.Visibility.ANY);
+        mapper.setVisibility(PropertyAccessor.CREATOR, JsonAutoDetect.Visibility.ANY);
 
         return mapper;
     }
diff --git a/change-scala-versions.sh b/change-scala-versions.sh
index 5fb40c5cb..8968abbf3 100755
--- a/change-scala-versions.sh
+++ b/change-scala-versions.sh
@@ -20,9 +20,9 @@
 
 set -e
 
-VALID_VERSIONS=( 2.10 2.11 )
-SCALA_210_VERSION=$(grep -F -m 1 'scala210.version' pom.xml); SCALA_210_VERSION="${SCALA_210_VERSION#*>}"; SCALA_210_VERSION="${SCALA_210_VERSION%<*}";
+VALID_VERSIONS=( 2.11 2.12 )
 SCALA_211_VERSION=$(grep -F -m 1 'scala211.version' pom.xml); SCALA_211_VERSION="${SCALA_211_VERSION#*>}"; SCALA_211_VERSION="${SCALA_211_VERSION%<*}";
+SCALA_212_VERSION=$(grep -F -m 1 'scala212.version' pom.xml); SCALA_212_VERSION="${SCALA_212_VERSION#*>}"; SCALA_212_VERSION="${SCALA_212_VERSION%<*}";
 
 usage() {
   echo "Usage: $(basename $0) [-h|--help] <scala version to be used>
@@ -45,19 +45,18 @@ check_scala_version() {
   exit 1
 }
 
-
 check_scala_version "$TO_VERSION"
 
 if [ $TO_VERSION = "2.11" ]; then
-  FROM_BINARY="_2\.10"
+  FROM_BINARY="_2\.12"
   TO_BINARY="_2\.11"
-  FROM_VERSION=$SCALA_210_VERSION
+  FROM_VERSION=$SCALA_212_VERSION
   TO_VERSION=$SCALA_211_VERSION
 else
   FROM_BINARY="_2\.11"
-  TO_BINARY="_2\.10"
+  TO_BINARY="_2\.12"
   FROM_VERSION=$SCALA_211_VERSION
-  TO_VERSION=$SCALA_210_VERSION
+  TO_VERSION=$SCALA_212_VERSION
 fi
 
 sed_i() {
@@ -70,35 +69,24 @@ echo "Updating Scala versions in pom.xml files to Scala $1, from $FROM_VERSION t
 
 BASEDIR=$(dirname $0)
 
-#Artifact ids, ending with "_2.10" or "_2.11". Spark, spark-mllib, kafka, etc.
+#Artifact ids, ending with "_2.11" or "_2.12". Spark, spark-mllib, kafka, etc.
 find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
   -exec bash -c "sed_i 's/\(artifactId>.*\)'$FROM_BINARY'<\/artifactId>/\1'$TO_BINARY'<\/artifactId>/g' {}" \;
 
-#Scala versions, like <scala.version>2.10</scala.version>
+#Scala versions, like <scala.version>2.11</scala.version>
 find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
   -exec bash -c "sed_i 's/\(scala.version>\)'$FROM_VERSION'<\/scala.version>/\1'$TO_VERSION'<\/scala.version>/g' {}" \;
 
-#Scala binary versions, like <scala.binary.version>2.10</scala.binary.version>
+#Scala binary versions, like <scala.binary.version>2.11</scala.binary.version>
 find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
   -exec bash -c "sed_i 's/\(scala.binary.version>\)'${FROM_BINARY#_}'<\/scala.binary.version>/\1'${TO_BINARY#_}'<\/scala.binary.version>/g' {}" \;
 
-#Scala versions, like <artifactId>scala-library</artifactId> <version>2.10.6</version>
+#Scala versions, like <artifactId>scala-library</artifactId> <version>2.11.12</version>
 find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
   -exec bash -c "sed_i 's/\(version>\)'$FROM_VERSION'<\/version>/\1'$TO_VERSION'<\/version>/g' {}" \;
 
-#Scala maven plugin, <scalaVersion>2.10</scalaVersion>
+#Scala maven plugin, <scalaVersion>2.11</scalaVersion>
 find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
   -exec bash -c "sed_i 's/\(scalaVersion>\)'$FROM_VERSION'<\/scalaVersion>/\1'$TO_VERSION'<\/scalaVersion>/g' {}" \;
-  
-#Edge case for Korean NLP artifact not following conventions:	https://github.com/deeplearning4j/deeplearning4j/issues/6306
-#https://github.com/deeplearning4j/deeplearning4j/issues/6306
-if [[ $TO_VERSION == 2.11* ]]; then
-  sed_i 's/<artifactId>korean-text-scala-2.10<\/artifactId>/<artifactId>korean-text<\/artifactId>/g' deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-korean/pom.xml
-  sed_i 's/<version>4.2.0<\/version>/<version>4.4<\/version>/g' deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-korean/pom.xml
-else
-  sed_i 's/<artifactId>korean-text<\/artifactId>/<artifactId>korean-text-scala-2.10<\/artifactId>/g' deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-korean/pom.xml
-  sed_i 's/<version>4.4<\/version>/<version>4.2.0<\/version>/g' deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-korean/pom.xml
-fi
-
 
 echo "Done updating Scala versions.";
diff --git a/change-spark-versions.sh b/change-spark-versions.sh
deleted file mode 100755
index 06a9b4d55..000000000
--- a/change-spark-versions.sh
+++ /dev/null
@@ -1,85 +0,0 @@
-#!/usr/bin/env bash
-
-################################################################################
-# Copyright (c) 2015-2018 Skymind, Inc.
-#
-# This program and the accompanying materials are made available under the
-# terms of the Apache License, Version 2.0 which is available at
-# https://www.apache.org/licenses/LICENSE-2.0.
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
-# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
-# License for the specific language governing permissions and limitations
-# under the License.
-#
-# SPDX-License-Identifier: Apache-2.0
-################################################################################
-
-# This shell script is adapted from Apache Flink (in turn, adapted from Apache Spark) some modifications.
-
-set -e
-
-VALID_VERSIONS=( 1 2 )
-SPARK_2_VERSION="2\.1\.0"
-SPARK_1_VERSION="1\.6\.3"
-
-usage() {
-  echo "Usage: $(basename $0) [-h|--help] <spark version to be used>
-where :
-  -h| --help Display this help text
-  valid spark version values : ${VALID_VERSIONS[*]}
-" 1>&2
-  exit 1
-}
-
-if [[ ($# -ne 1) || ( $1 == "--help") ||  $1 == "-h" ]]; then
-  usage
-fi
-
-TO_VERSION=$1
-
-check_spark_version() {
-  for i in ${VALID_VERSIONS[*]}; do [ $i = "$1" ] && return 0; done
-  echo "Invalid Spark version: $1. Valid versions: ${VALID_VERSIONS[*]}" 1>&2
-  exit 1
-}
-
-
-check_spark_version "$TO_VERSION"
-
-if [ $TO_VERSION = "2" ]; then
-  FROM_BINARY="1"
-  TO_BINARY="2"
-  FROM_VERSION=$SPARK_1_VERSION
-  TO_VERSION=$SPARK_2_VERSION
-else
-  FROM_BINARY="2"
-  TO_BINARY="1"
-  FROM_VERSION=$SPARK_2_VERSION
-  TO_VERSION=$SPARK_1_VERSION
-fi
-
-sed_i() {
-  sed -e "$1" "$2" > "$2.tmp" && mv "$2.tmp" "$2"
-}
-
-export -f sed_i
-
-echo "Updating Spark versions in pom.xml files to Spark $1";
-
-BASEDIR=$(dirname $0)
-
-# <spark.major.version>1</spark.major.version>
-find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
-  -exec bash -c "sed_i 's/\(spark.major.version>\)'$FROM_BINARY'<\/spark.major.version>/\1'$TO_BINARY'<\/spark.major.version>/g' {}" \;
-
-# <spark.version>1.6.3</spark.version>
-find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
-  -exec bash -c "sed_i 's/\(spark.version>\)'$FROM_VERSION'<\/spark.version>/\1'$TO_VERSION'<\/spark.version>/g' {}" \;
-
-#Spark versions, like <version>xxx_spark_2xxx</version> OR <datavec.spark.version>xxx_spark_2xxx</datavec.spark.version>
-find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
-  -exec bash -c "sed_i 's/\(version>.*_spark_\)'$FROM_BINARY'\(.*\)version>/\1'$TO_BINARY'\2version>/g' {}" \;
-
-echo "Done updating Spark versions.";
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/Transform.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/Transform.java
index 44098e6cc..32f8f7cc5 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/Transform.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/Transform.java
@@ -16,7 +16,6 @@
 
 package org.datavec.api.transform;
 
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.Writable;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
@@ -27,8 +26,7 @@ import java.util.List;
 /**A Transform converts an example to another example, or a sequence to another sequence
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.TransformHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface Transform extends Serializable, ColumnOp {
 
     /**
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/TransformProcess.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/TransformProcess.java
index 24a029179..7c57f4daa 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/TransformProcess.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/TransformProcess.java
@@ -67,6 +67,7 @@ import org.joda.time.DateTimeZone;
 import org.nd4j.linalg.primitives.Pair;
 import org.nd4j.shade.jackson.annotation.JsonProperty;
 import org.nd4j.shade.jackson.core.JsonProcessingException;
+import org.nd4j.shade.jackson.databind.exc.InvalidTypeIdException;
 
 import java.io.IOException;
 import java.io.Serializable;
@@ -417,6 +418,16 @@ public class TransformProcess implements Serializable {
     public static TransformProcess fromJson(String json) {
         try {
             return JsonMappers.getMapper().readValue(json, TransformProcess.class);
+        } catch (InvalidTypeIdException e){
+            if(e.getMessage().contains("@class")){
+                //JSON may be legacy (1.0.0-alpha or earlier), attempt to load it using old format
+                try{
+                    return JsonMappers.getLegacyMapper().readValue(json, TransformProcess.class);
+                } catch (IOException e2){
+                    throw new RuntimeException(e2);
+                }
+            }
+            throw new RuntimeException(e);
         } catch (IOException e) {
             //TODO proper exception message
             throw new RuntimeException(e);
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/DataAnalysis.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/DataAnalysis.java
index 6b069a9ec..467db70f0 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/DataAnalysis.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/DataAnalysis.java
@@ -23,12 +23,14 @@ import org.datavec.api.transform.analysis.columns.ColumnAnalysis;
 import org.datavec.api.transform.metadata.CategoricalMetaData;
 import org.datavec.api.transform.metadata.ColumnMetaData;
 import org.datavec.api.transform.schema.Schema;
+import org.datavec.api.transform.serde.JsonMappers;
 import org.datavec.api.transform.serde.JsonSerializer;
 import org.datavec.api.transform.serde.YamlSerializer;
 import org.nd4j.shade.jackson.annotation.JsonSubTypes;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
 import org.nd4j.shade.jackson.databind.JsonNode;
 import org.nd4j.shade.jackson.databind.ObjectMapper;
+import org.nd4j.shade.jackson.databind.exc.InvalidTypeIdException;
 import org.nd4j.shade.jackson.databind.node.ArrayNode;
 
 import java.io.IOException;
@@ -116,6 +118,16 @@ public class DataAnalysis implements Serializable {
     public static DataAnalysis fromJson(String json) {
         try{
             return new JsonSerializer().getObjectMapper().readValue(json, DataAnalysis.class);
+        } catch (InvalidTypeIdException e){
+            if(e.getMessage().contains("@class")){
+                try{
+                    //JSON may be legacy (1.0.0-alpha or earlier), attempt to load it using old format
+                    return JsonMappers.getLegacyMapper().readValue(json, DataAnalysis.class);
+                } catch (IOException e2){
+                    throw new RuntimeException(e2);
+                }
+            }
+            throw new RuntimeException(e);
         } catch (Exception e){
             //Legacy format
             ObjectMapper om = new JsonSerializer().getObjectMapper();
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/SequenceDataAnalysis.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/SequenceDataAnalysis.java
index ecc333d2d..6156ead40 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/SequenceDataAnalysis.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/SequenceDataAnalysis.java
@@ -21,9 +21,10 @@ import lombok.EqualsAndHashCode;
 import org.datavec.api.transform.analysis.columns.ColumnAnalysis;
 import org.datavec.api.transform.analysis.sequence.SequenceLengthAnalysis;
 import org.datavec.api.transform.schema.Schema;
+import org.datavec.api.transform.serde.JsonMappers;
 import org.datavec.api.transform.serde.JsonSerializer;
 import org.datavec.api.transform.serde.YamlSerializer;
-import org.nd4j.shade.jackson.databind.ObjectMapper;
+import org.nd4j.shade.jackson.databind.exc.InvalidTypeIdException;
 
 import java.io.IOException;
 import java.util.List;
@@ -50,6 +51,16 @@ public class SequenceDataAnalysis extends DataAnalysis {
     public static SequenceDataAnalysis fromJson(String json){
         try{
             return new JsonSerializer().getObjectMapper().readValue(json, SequenceDataAnalysis.class);
+        } catch (InvalidTypeIdException e){
+            if(e.getMessage().contains("@class")){
+                try{
+                    //JSON may be legacy (1.0.0-alpha or earlier), attempt to load it using old format
+                    return JsonMappers.getLegacyMapper().readValue(json, SequenceDataAnalysis.class);
+                } catch (IOException e2){
+                    throw new RuntimeException(e2);
+                }
+            }
+            throw new RuntimeException(e);
         } catch (IOException e){
             throw new RuntimeException(e);
         }
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/columns/ColumnAnalysis.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/columns/ColumnAnalysis.java
index dd43315d2..c86584ede 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/columns/ColumnAnalysis.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/analysis/columns/ColumnAnalysis.java
@@ -17,7 +17,6 @@
 package org.datavec.api.transform.analysis.columns;
 
 import org.datavec.api.transform.ColumnType;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
 
@@ -27,8 +26,7 @@ import java.io.Serializable;
  * Interface for column analysis
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.ColumnAnalysisHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface ColumnAnalysis extends Serializable {
 
     long getCountTotal();
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/condition/Condition.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/condition/Condition.java
index 83a96bfcf..6bd5b98ac 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/condition/Condition.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/condition/Condition.java
@@ -18,7 +18,6 @@ package org.datavec.api.transform.condition;
 
 import org.datavec.api.transform.ColumnOp;
 import org.datavec.api.transform.schema.Schema;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.Writable;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
@@ -35,8 +34,7 @@ import java.util.List;
  * @author Alex Black
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.ConditionHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface Condition extends Serializable, ColumnOp {
 
     /**
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/filter/Filter.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/filter/Filter.java
index 5aa672f76..16870e9f9 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/filter/Filter.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/filter/Filter.java
@@ -18,7 +18,6 @@ package org.datavec.api.transform.filter;
 
 import org.datavec.api.transform.ColumnOp;
 import org.datavec.api.transform.schema.Schema;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.Writable;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
@@ -33,8 +32,7 @@ import java.util.List;
  * @author Alex Black
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.FilterHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface Filter extends Serializable, ColumnOp {
 
     /**
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/metadata/ColumnMetaData.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/metadata/ColumnMetaData.java
index cde86cf90..6889fbf31 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/metadata/ColumnMetaData.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/metadata/ColumnMetaData.java
@@ -17,7 +17,6 @@
 package org.datavec.api.transform.metadata;
 
 import org.datavec.api.transform.ColumnType;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.Writable;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
@@ -32,8 +31,7 @@ import java.io.Serializable;
  * @author Alex Black
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.ColumnMetaDataHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface ColumnMetaData extends Serializable, Cloneable {
 
     /**
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/ops/DispatchWithConditionOp.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/ops/DispatchWithConditionOp.java
index 3341e0b6d..8ef67bacb 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/ops/DispatchWithConditionOp.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/ops/DispatchWithConditionOp.java
@@ -23,8 +23,8 @@ import org.datavec.api.writable.Writable;
 
 import java.util.List;
 
-import static com.google.common.base.Preconditions.checkArgument;
-import static com.google.common.base.Preconditions.checkNotNull;
+import static org.nd4j.shade.guava.base.Preconditions.checkArgument;
+import static org.nd4j.shade.guava.base.Preconditions.checkNotNull;
 
 /**
  * A variant of {@link DispatchOp} that for each operation, tests the input list of {@Writable} elements for a {@link Condition},
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/rank/CalculateSortedRank.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/rank/CalculateSortedRank.java
index 5bcf81f7a..1e5177c68 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/rank/CalculateSortedRank.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/rank/CalculateSortedRank.java
@@ -23,7 +23,6 @@ import org.datavec.api.transform.metadata.ColumnMetaData;
 import org.datavec.api.transform.metadata.LongMetaData;
 import org.datavec.api.transform.schema.Schema;
 import org.datavec.api.transform.schema.SequenceSchema;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.comparator.WritableComparator;
 import org.nd4j.shade.jackson.annotation.JsonIgnoreProperties;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
@@ -50,8 +49,7 @@ import java.util.List;
 @EqualsAndHashCode(exclude = {"inputSchema"})
 @JsonIgnoreProperties({"inputSchema"})
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.CalculateSortedRankHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public class CalculateSortedRank implements Serializable, ColumnOp {
 
     private final String newColumnName;
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/schema/Schema.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/schema/Schema.java
index a9e167943..1c16ebcce 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/schema/Schema.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/schema/Schema.java
@@ -21,7 +21,6 @@ import lombok.EqualsAndHashCode;
 import org.datavec.api.transform.ColumnType;
 import org.datavec.api.transform.metadata.*;
 import org.datavec.api.transform.serde.JsonMappers;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.*;
 import org.joda.time.DateTimeZone;
 import org.nd4j.shade.jackson.annotation.*;
@@ -29,9 +28,11 @@ import org.nd4j.shade.jackson.core.JsonFactory;
 import org.nd4j.shade.jackson.databind.DeserializationFeature;
 import org.nd4j.shade.jackson.databind.ObjectMapper;
 import org.nd4j.shade.jackson.databind.SerializationFeature;
+import org.nd4j.shade.jackson.databind.exc.InvalidTypeIdException;
 import org.nd4j.shade.jackson.dataformat.yaml.YAMLFactory;
 import org.nd4j.shade.jackson.datatype.joda.JodaModule;
 
+import java.io.IOException;
 import java.io.Serializable;
 import java.util.*;
 
@@ -48,8 +49,7 @@ import java.util.*;
  */
 @JsonIgnoreProperties({"columnNames", "columnNamesIndex"})
 @EqualsAndHashCode
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.SchemaHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 @Data
 public class Schema implements Serializable {
 
@@ -358,6 +358,16 @@ public class Schema implements Serializable {
     public static Schema fromJson(String json) {
         try{
             return JsonMappers.getMapper().readValue(json, Schema.class);
+        } catch (InvalidTypeIdException e){
+            if(e.getMessage().contains("@class")){
+                try{
+                    //JSON may be legacy (1.0.0-alpha or earlier), attempt to load it using old format
+                    return JsonMappers.getLegacyMapper().readValue(json, Schema.class);
+                } catch (IOException e2){
+                    throw new RuntimeException(e2);
+                }
+            }
+            throw new RuntimeException(e);
         } catch (Exception e){
             //TODO better exceptions
             throw new RuntimeException(e);
@@ -379,21 +389,6 @@ public class Schema implements Serializable {
         }
     }
 
-    private static Schema fromJacksonString(String str, JsonFactory factory) {
-        ObjectMapper om = new ObjectMapper(factory);
-        om.registerModule(new JodaModule());
-        om.configure(DeserializationFeature.FAIL_ON_UNKNOWN_PROPERTIES, false);
-        om.configure(SerializationFeature.FAIL_ON_EMPTY_BEANS, false);
-        om.enable(SerializationFeature.INDENT_OUTPUT);
-        om.setVisibility(PropertyAccessor.ALL, JsonAutoDetect.Visibility.NONE);
-        om.setVisibility(PropertyAccessor.FIELD, JsonAutoDetect.Visibility.ANY);
-        try {
-            return om.readValue(str, Schema.class);
-        } catch (Exception e) {
-            throw new RuntimeException(e);
-        }
-    }
-
     public static class Builder {
         List<ColumnMetaData> columnMetaData = new ArrayList<>();
 
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/SequenceComparator.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/SequenceComparator.java
index e4a09f6e9..a5677d1f5 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/SequenceComparator.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/SequenceComparator.java
@@ -17,7 +17,6 @@
 package org.datavec.api.transform.sequence;
 
 import org.datavec.api.transform.schema.Schema;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.Writable;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
@@ -30,8 +29,7 @@ import java.util.List;
  * Compare the time steps of a sequence
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.SequenceComparatorHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface SequenceComparator extends Comparator<List<Writable>>, Serializable {
 
     /**
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/SequenceSplit.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/SequenceSplit.java
index bef0b4ccc..3471dbaa3 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/SequenceSplit.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/SequenceSplit.java
@@ -17,7 +17,6 @@
 package org.datavec.api.transform.sequence;
 
 import org.datavec.api.transform.schema.Schema;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.Writable;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
@@ -32,8 +31,7 @@ import java.util.List;
  * @author Alex Black
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.SequenceSplitHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface SequenceSplit extends Serializable {
 
     /**
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/window/WindowFunction.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/window/WindowFunction.java
index e25f85253..1456af8d7 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/window/WindowFunction.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/sequence/window/WindowFunction.java
@@ -17,7 +17,6 @@
 package org.datavec.api.transform.sequence.window;
 
 import org.datavec.api.transform.schema.Schema;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.Writable;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
@@ -36,8 +35,7 @@ import java.util.List;
  * @author Alex Black
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.WindowFunctionHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface WindowFunction extends Serializable {
 
     /**
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/JsonMappers.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/JsonMappers.java
index 652556ce4..bfa114697 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/JsonMappers.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/JsonMappers.java
@@ -16,44 +16,17 @@
 
 package org.datavec.api.transform.serde;
 
-import lombok.AllArgsConstructor;
-import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
-import org.datavec.api.io.WritableComparator;
-import org.datavec.api.transform.Transform;
-import org.datavec.api.transform.analysis.columns.ColumnAnalysis;
-import org.datavec.api.transform.condition.column.ColumnCondition;
-import org.datavec.api.transform.filter.Filter;
-import org.datavec.api.transform.metadata.ColumnMetaData;
-import org.datavec.api.transform.rank.CalculateSortedRank;
-import org.datavec.api.transform.schema.Schema;
-import org.datavec.api.transform.sequence.SequenceComparator;
-import org.datavec.api.transform.sequence.SequenceSplit;
-import org.datavec.api.transform.sequence.window.WindowFunction;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
-import org.datavec.api.writable.Writable;
-import org.nd4j.linalg.activations.IActivation;
-import org.nd4j.linalg.lossfunctions.ILossFunction;
-import org.nd4j.linalg.primitives.Pair;
-import org.nd4j.serde.json.LegacyIActivationDeserializer;
-import org.nd4j.serde.json.LegacyILossFunctionDeserializer;
+import org.datavec.api.transform.serde.legacy.LegacyJsonFormat;
 import org.nd4j.shade.jackson.annotation.JsonAutoDetect;
-import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
 import org.nd4j.shade.jackson.annotation.PropertyAccessor;
-import org.nd4j.shade.jackson.databind.*;
-import org.nd4j.shade.jackson.databind.cfg.MapperConfig;
-import org.nd4j.shade.jackson.databind.introspect.Annotated;
-import org.nd4j.shade.jackson.databind.introspect.AnnotatedClass;
-import org.nd4j.shade.jackson.databind.introspect.AnnotationMap;
-import org.nd4j.shade.jackson.databind.introspect.JacksonAnnotationIntrospector;
-import org.nd4j.shade.jackson.databind.jsontype.TypeResolverBuilder;
+import org.nd4j.shade.jackson.databind.DeserializationFeature;
+import org.nd4j.shade.jackson.databind.MapperFeature;
+import org.nd4j.shade.jackson.databind.ObjectMapper;
+import org.nd4j.shade.jackson.databind.SerializationFeature;
 import org.nd4j.shade.jackson.dataformat.yaml.YAMLFactory;
 import org.nd4j.shade.jackson.datatype.joda.JodaModule;
 
-import java.lang.annotation.Annotation;
-import java.util.*;
-import java.util.concurrent.ConcurrentHashMap;
-
 /**
  * JSON mappers for deserializing neural net configurations, etc.
  *
@@ -62,38 +35,9 @@ import java.util.concurrent.ConcurrentHashMap;
 @Slf4j
 public class JsonMappers {
 
-    /**
-     * This system property is provided as an alternative to {@link #registerLegacyCustomClassesForJSON(Class[])}
-     * Classes can be specified in comma-separated format
-     */
-    public static String CUSTOM_REGISTRATION_PROPERTY = "org.datavec.config.custom.legacyclasses";
-
-    static {
-        String p = System.getProperty(CUSTOM_REGISTRATION_PROPERTY);
-        if(p != null && !p.isEmpty()){
-            String[] split = p.split(",");
-            List<Class<?>> list = new ArrayList<>();
-            for(String s : split){
-                try{
-                    Class<?> c = Class.forName(s);
-                    list.add(c);
-                } catch (Throwable t){
-                    log.warn("Error parsing {} system property: class \"{}\" could not be loaded",CUSTOM_REGISTRATION_PROPERTY, s, t);
-                }
-            }
-
-            if(list.size() > 0){
-                try {
-                    registerLegacyCustomClassesForJSONList(list);
-                } catch (Throwable t){
-                    log.warn("Error registering custom classes for legacy JSON deserialization ({} system property)",CUSTOM_REGISTRATION_PROPERTY, t);
-                }
-            }
-        }
-    }
-
     private static ObjectMapper jsonMapper;
     private static ObjectMapper yamlMapper;
+    private static ObjectMapper legacyMapper;       //For 1.0.0-alpha and earlier TransformProcess etc
 
     static {
         jsonMapper = new ObjectMapper();
@@ -102,117 +46,12 @@ public class JsonMappers {
         configureMapper(yamlMapper);
     }
 
-    private static Map<Class, ObjectMapper> legacyMappers = new ConcurrentHashMap<>();
-
-
-    /**
-     * Register a set of classes (Transform, Filter, etc) for JSON deserialization.<br>
-     * <br>
-     * This is required ONLY when BOTH of the following conditions are met:<br>
-     * 1. You want to load a serialized TransformProcess, saved in 1.0.0-alpha or before, AND<br>
-     * 2. The serialized TransformProcess has a custom Transform, Filter, etc (i.e., one not defined in DL4J)<br>
-     * <br>
-     * By passing the classes of these custom classes here, DataVec should be able to deserialize them, in spite of the JSON
-     * format change between versions.
-     *
-     * @param classes Classes to register
-     */
-    public static void registerLegacyCustomClassesForJSON(Class<?>... classes) {
-        registerLegacyCustomClassesForJSONList(Arrays.<Class<?>>asList(classes));
-    }
-
-    /**
-     * @see #registerLegacyCustomClassesForJSON(Class[])
-     */
-    public static void registerLegacyCustomClassesForJSONList(List<Class<?>> classes){
-        //Default names (i.e., old format for custom JSON format)
-        List<Pair<String,Class>> list = new ArrayList<>();
-        for(Class<?> c : classes){
-            list.add(new Pair<String,Class>(c.getSimpleName(), c));
+    public static synchronized ObjectMapper getLegacyMapper(){
+        if(legacyMapper == null){
+            legacyMapper = LegacyJsonFormat.legacyMapper();
+            configureMapper(legacyMapper);
         }
-        registerLegacyCustomClassesForJSON(list);
-    }
-
-    /**
-     * Set of classes that can be registered for legacy deserialization.
-     */
-    private static List<Class<?>> REGISTERABLE_CUSTOM_CLASSES = (List<Class<?>>) Arrays.<Class<?>>asList(
-            Transform.class,
-            ColumnAnalysis.class,
-            ColumnCondition.class,
-            Filter.class,
-            ColumnMetaData.class,
-            CalculateSortedRank.class,
-            Schema.class,
-            SequenceComparator.class,
-            SequenceSplit.class,
-            WindowFunction.class,
-            Writable.class,
-            WritableComparator.class
-    );
-
-    /**
-     * Register a set of classes (Layer, GraphVertex, InputPreProcessor, IActivation, ILossFunction, ReconstructionDistribution
-     * ONLY) for JSON deserialization, with custom names.<br>
-     * Using this method directly should never be required (instead: use {@link #registerLegacyCustomClassesForJSON(Class[])}
-     * but is added in case it is required in non-standard circumstances.
-     */
-    public static void registerLegacyCustomClassesForJSON(List<Pair<String,Class>> classes){
-        for(Pair<String,Class> p : classes){
-            String s = p.getFirst();
-            Class c = p.getRight();
-            //Check if it's a valid class to register...
-            boolean found = false;
-            for( Class<?> c2 : REGISTERABLE_CUSTOM_CLASSES){
-                if(c2.isAssignableFrom(c)){
-                    Map<String,String> map = LegacyMappingHelper.legacyMappingForClass(c2);
-                    map.put(p.getFirst(), p.getSecond().getName());
-                    found = true;
-                }
-            }
-
-            if(!found){
-                throw new IllegalArgumentException("Cannot register class for legacy JSON deserialization: class " +
-                        c.getName() + " is not a subtype of classes " + REGISTERABLE_CUSTOM_CLASSES);
-            }
-        }
-    }
-
-
-    /**
-     * Get the legacy JSON mapper for the specified class.<br>
-     *
-     * <b>NOTE</b>: This is intended for internal backward-compatibility use.
-     *
-     * Note to developers: The following JSON mappers are for handling legacy format JSON.
-     * Note that after 1.0.0-alpha, the JSON subtype format for Transforms, Filters, Conditions etc were changed from
-     * a wrapper object, to an "@class" field. However, to not break all saved transforms networks, these mappers are
-     * part of the solution.<br>
-     * <br>
-     * How legacy loading works (same pattern for all types - Transform, Filter, Condition etc)<br>
-     * 1. Transforms etc JSON that has a "@class" field are deserialized as normal<br>
-     * 2. Transforms JSON that don't have such a field are mapped (via Layer @JsonTypeInfo) to LegacyMappingHelper.TransformHelper<br>
-     * 3. LegacyMappingHelper.TransformHelper has a @JsonDeserialize annotation - we use LegacyMappingHelper.LegacyTransformDeserializer to handle it<br>
-     * 4. LegacyTransformDeserializer has a list of old names (present in the legacy format JSON) and the corresponding class names
-     * 5. BaseLegacyDeserializer (that LegacyTransformDeserializer extends) does a lookup and handles the deserialization
-     *
-     * Now, as to why we have one ObjectMapper for each type: We can't use the default JSON mapper for the legacy format,
-     * as it'll fail due to not having the expected "@class" annotation.
-     * Consequently, we need to tell Jackson to ignore that specific annotation and deserialize to the specified
-     * class anyway. The ignoring is done via an annotation introspector, defined below in this class.
-     * However, we can't just use a single annotation introspector (and hence ObjectMapper) for loading legacy values of
-     * all types - if we did, then any nested types would fail (i.e., an Condition in a Transform - the Transform couldn't
-     * be deserialized correctly, as the annotation would be ignored).
-     *
-     */
-    public static synchronized ObjectMapper getLegacyMapperFor(@NonNull Class<?> clazz){
-        if(!legacyMappers.containsKey(clazz)){
-            ObjectMapper m = new ObjectMapper();
-            configureMapper(m);
-            m.setAnnotationIntrospector(new IgnoreJsonTypeInfoIntrospector(Collections.<Class>singletonList(clazz)));
-            legacyMappers.put(clazz, m);
-        }
-        return legacyMappers.get(clazz);
+        return legacyMapper;
     }
 
     /**
@@ -237,61 +76,7 @@ public class JsonMappers {
         ret.enable(SerializationFeature.INDENT_OUTPUT);
         ret.setVisibility(PropertyAccessor.ALL, JsonAutoDetect.Visibility.NONE);
         ret.setVisibility(PropertyAccessor.FIELD, JsonAutoDetect.Visibility.ANY);
+        ret.setVisibility(PropertyAccessor.CREATOR, JsonAutoDetect.Visibility.ANY);     //Need this otherwise JsonProperty annotations on constructors won't be seen
     }
 
-
-    /**
-     * Custom Jackson Introspector to ignore the {@code @JsonTypeYnfo} annotations on layers etc.
-     * This is so we can deserialize legacy format JSON without recursing infinitely, by selectively ignoring
-     * a set of JsonTypeInfo annotations
-     */
-    @AllArgsConstructor
-    private static class IgnoreJsonTypeInfoIntrospector extends JacksonAnnotationIntrospector {
-
-        private List<Class> classList;
-
-        @Override
-        protected TypeResolverBuilder<?> _findTypeResolver(MapperConfig<?> config, Annotated ann, JavaType baseType) {
-            if(ann instanceof AnnotatedClass){
-                AnnotatedClass c = (AnnotatedClass)ann;
-                Class<?> annClass = c.getAnnotated();
-
-                boolean isAssignable = false;
-                for(Class c2 : classList){
-                    if(c2.isAssignableFrom(annClass)){
-                        isAssignable = true;
-                        break;
-                    }
-                }
-
-                if( isAssignable ){
-                    AnnotationMap annotations = (AnnotationMap) ((AnnotatedClass) ann).getAnnotations();
-                    if(annotations == null || annotations.annotations() == null){
-                        //Probably not necessary - but here for safety
-                        return super._findTypeResolver(config, ann, baseType);
-                    }
-
-                    AnnotationMap newMap = null;
-                    for(Annotation a : annotations.annotations()){
-                        Class<?> annType = a.annotationType();
-                        if(annType == JsonTypeInfo.class){
-                            //Ignore the JsonTypeInfo annotation on the Layer class
-                            continue;
-                        }
-                        if(newMap == null){
-                            newMap = new AnnotationMap();
-                        }
-                        newMap.add(a);
-                    }
-                    if(newMap == null)
-                        return null;
-
-                    //Pass the remaining annotations (if any) to the original introspector
-                    AnnotatedClass ann2 = c.withAnnotations(newMap);
-                    return super._findTypeResolver(config, ann2, baseType);
-                }
-            }
-            return super._findTypeResolver(config, ann, baseType);
-        }
-    }
 }
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/GenericLegacyDeserializer.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/GenericLegacyDeserializer.java
deleted file mode 100644
index 5a9b48a7c..000000000
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/GenericLegacyDeserializer.java
+++ /dev/null
@@ -1,41 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.api.transform.serde.legacy;
-
-import lombok.AllArgsConstructor;
-import lombok.Data;
-import lombok.Getter;
-import org.datavec.api.transform.serde.JsonMappers;
-import org.nd4j.serde.json.BaseLegacyDeserializer;
-import org.nd4j.shade.jackson.databind.ObjectMapper;
-
-import java.util.Map;
-
-@AllArgsConstructor
-@Data
-public class GenericLegacyDeserializer<T> extends BaseLegacyDeserializer<T> {
-
-    @Getter
-    protected final Class<T> deserializedType;
-    @Getter
-    protected final Map<String,String> legacyNamesMap;
-
-    @Override
-    public ObjectMapper getLegacyJsonMapper() {
-        return JsonMappers.getLegacyMapperFor(getDeserializedType());
-    }
-}
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/LegacyJsonFormat.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/LegacyJsonFormat.java
new file mode 100644
index 000000000..8df741a49
--- /dev/null
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/LegacyJsonFormat.java
@@ -0,0 +1,267 @@
+package org.datavec.api.transform.serde.legacy;
+
+import lombok.AccessLevel;
+import lombok.NoArgsConstructor;
+import org.datavec.api.transform.Transform;
+import org.datavec.api.transform.analysis.columns.*;
+import org.datavec.api.transform.condition.BooleanCondition;
+import org.datavec.api.transform.condition.Condition;
+import org.datavec.api.transform.condition.column.*;
+import org.datavec.api.transform.condition.sequence.SequenceLengthCondition;
+import org.datavec.api.transform.condition.string.StringRegexColumnCondition;
+import org.datavec.api.transform.filter.ConditionFilter;
+import org.datavec.api.transform.filter.Filter;
+import org.datavec.api.transform.filter.FilterInvalidValues;
+import org.datavec.api.transform.filter.InvalidNumColumns;
+import org.datavec.api.transform.metadata.*;
+import org.datavec.api.transform.ndarray.NDArrayColumnsMathOpTransform;
+import org.datavec.api.transform.ndarray.NDArrayDistanceTransform;
+import org.datavec.api.transform.ndarray.NDArrayMathFunctionTransform;
+import org.datavec.api.transform.ndarray.NDArrayScalarOpTransform;
+import org.datavec.api.transform.rank.CalculateSortedRank;
+import org.datavec.api.transform.schema.Schema;
+import org.datavec.api.transform.schema.SequenceSchema;
+import org.datavec.api.transform.sequence.ReduceSequenceTransform;
+import org.datavec.api.transform.sequence.SequenceComparator;
+import org.datavec.api.transform.sequence.SequenceSplit;
+import org.datavec.api.transform.sequence.comparator.NumericalColumnComparator;
+import org.datavec.api.transform.sequence.comparator.StringComparator;
+import org.datavec.api.transform.sequence.split.SequenceSplitTimeSeparation;
+import org.datavec.api.transform.sequence.split.SplitMaxLengthSequence;
+import org.datavec.api.transform.sequence.trim.SequenceTrimTransform;
+import org.datavec.api.transform.sequence.window.OverlappingTimeWindowFunction;
+import org.datavec.api.transform.sequence.window.ReduceSequenceByWindowTransform;
+import org.datavec.api.transform.sequence.window.TimeWindowFunction;
+import org.datavec.api.transform.sequence.window.WindowFunction;
+import org.datavec.api.transform.stringreduce.IStringReducer;
+import org.datavec.api.transform.stringreduce.StringReducer;
+import org.datavec.api.transform.transform.categorical.*;
+import org.datavec.api.transform.transform.column.*;
+import org.datavec.api.transform.transform.condition.ConditionalCopyValueTransform;
+import org.datavec.api.transform.transform.condition.ConditionalReplaceValueTransform;
+import org.datavec.api.transform.transform.condition.ConditionalReplaceValueTransformWithDefault;
+import org.datavec.api.transform.transform.doubletransform.*;
+import org.datavec.api.transform.transform.integer.*;
+import org.datavec.api.transform.transform.longtransform.LongColumnsMathOpTransform;
+import org.datavec.api.transform.transform.longtransform.LongMathOpTransform;
+import org.datavec.api.transform.transform.nlp.TextToCharacterIndexTransform;
+import org.datavec.api.transform.transform.parse.ParseDoubleTransform;
+import org.datavec.api.transform.transform.sequence.SequenceDifferenceTransform;
+import org.datavec.api.transform.transform.sequence.SequenceMovingWindowReduceTransform;
+import org.datavec.api.transform.transform.sequence.SequenceOffsetTransform;
+import org.datavec.api.transform.transform.string.*;
+import org.datavec.api.transform.transform.time.DeriveColumnsFromTimeTransform;
+import org.datavec.api.transform.transform.time.StringToTimeTransform;
+import org.datavec.api.transform.transform.time.TimeMathOpTransform;
+import org.datavec.api.writable.*;
+import org.datavec.api.writable.comparator.*;
+import org.nd4j.shade.jackson.annotation.JsonInclude;
+import org.nd4j.shade.jackson.annotation.JsonSubTypes;
+import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
+import org.nd4j.shade.jackson.databind.ObjectMapper;
+
+/**
+ * This class defines a set of Jackson Mixins - which are a way of using a proxy class with annotations to override
+ * the existing annotations.
+ * In 1.0.0-beta, we switched how subtypes were handled in JSON ser/de: from "wrapper object" to "@class field".
+ * We use these mixins to allow us to still load the old format
+ *
+ * @author Alex Black
+ */
+public class LegacyJsonFormat {
+
+    private LegacyJsonFormat(){ }
+
+    /**
+     * Get a mapper (minus general config) suitable for loading old format JSON - 1.0.0-alpha and before
+     * @return Object mapper
+     */
+    public static ObjectMapper legacyMapper(){
+        ObjectMapper om = new ObjectMapper();
+        om.addMixIn(Schema.class, SchemaMixin.class);
+        om.addMixIn(ColumnMetaData.class, ColumnMetaDataMixin.class);
+        om.addMixIn(Transform.class, TransformMixin.class);
+        om.addMixIn(Condition.class, ConditionMixin.class);
+        om.addMixIn(Writable.class, WritableMixin.class);
+        om.addMixIn(Filter.class, FilterMixin.class);
+        om.addMixIn(SequenceComparator.class, SequenceComparatorMixin.class);
+        om.addMixIn(SequenceSplit.class, SequenceSplitMixin.class);
+        om.addMixIn(WindowFunction.class, WindowFunctionMixin.class);
+        om.addMixIn(CalculateSortedRank.class, CalculateSortedRankMixin.class);
+        om.addMixIn(WritableComparator.class, WritableComparatorMixin.class);
+        om.addMixIn(ColumnAnalysis.class, ColumnAnalysisMixin.class);
+        om.addMixIn(IStringReducer.class, IStringReducerMixin.class);
+        return om;
+    }
+
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes({@JsonSubTypes.Type(value = Schema.class, name = "Schema"),
+            @JsonSubTypes.Type(value = SequenceSchema.class, name = "SequenceSchema")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class SchemaMixin { }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes({@JsonSubTypes.Type(value = BinaryMetaData.class, name = "Binary"),
+            @JsonSubTypes.Type(value = BooleanMetaData.class, name = "Boloean"),
+            @JsonSubTypes.Type(value = CategoricalMetaData.class, name = "Categorical"),
+            @JsonSubTypes.Type(value = DoubleMetaData.class, name = "Double"),
+            @JsonSubTypes.Type(value = FloatMetaData.class, name = "Float"),
+            @JsonSubTypes.Type(value = IntegerMetaData.class, name = "Integer"),
+            @JsonSubTypes.Type(value = LongMetaData.class, name = "Long"),
+            @JsonSubTypes.Type(value = NDArrayMetaData.class, name = "NDArray"),
+            @JsonSubTypes.Type(value = StringMetaData.class, name = "String"),
+            @JsonSubTypes.Type(value = TimeMetaData.class, name = "Time")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class ColumnMetaDataMixin { }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = CalculateSortedRank.class, name = "CalculateSortedRank"),
+            @JsonSubTypes.Type(value = CategoricalToIntegerTransform.class, name = "CategoricalToIntegerTransform"),
+            @JsonSubTypes.Type(value = CategoricalToOneHotTransform.class, name = "CategoricalToOneHotTransform"),
+            @JsonSubTypes.Type(value = IntegerToCategoricalTransform.class, name = "IntegerToCategoricalTransform"),
+            @JsonSubTypes.Type(value = StringToCategoricalTransform.class, name = "StringToCategoricalTransform"),
+            @JsonSubTypes.Type(value = DuplicateColumnsTransform.class, name = "DuplicateColumnsTransform"),
+            @JsonSubTypes.Type(value = RemoveColumnsTransform.class, name = "RemoveColumnsTransform"),
+            @JsonSubTypes.Type(value = RenameColumnsTransform.class, name = "RenameColumnsTransform"),
+            @JsonSubTypes.Type(value = ReorderColumnsTransform.class, name = "ReorderColumnsTransform"),
+            @JsonSubTypes.Type(value = ConditionalCopyValueTransform.class, name = "ConditionalCopyValueTransform"),
+            @JsonSubTypes.Type(value = ConditionalReplaceValueTransform.class, name = "ConditionalReplaceValueTransform"),
+            @JsonSubTypes.Type(value = ConditionalReplaceValueTransformWithDefault.class, name = "ConditionalReplaceValueTransformWithDefault"),
+            @JsonSubTypes.Type(value = DoubleColumnsMathOpTransform.class, name = "DoubleColumnsMathOpTransform"),
+            @JsonSubTypes.Type(value = DoubleMathOpTransform.class, name = "DoubleMathOpTransform"),
+            @JsonSubTypes.Type(value = Log2Normalizer.class, name = "Log2Normalizer"),
+            @JsonSubTypes.Type(value = MinMaxNormalizer.class, name = "MinMaxNormalizer"),
+            @JsonSubTypes.Type(value = StandardizeNormalizer.class, name = "StandardizeNormalizer"),
+            @JsonSubTypes.Type(value = SubtractMeanNormalizer.class, name = "SubtractMeanNormalizer"),
+            @JsonSubTypes.Type(value = IntegerColumnsMathOpTransform.class, name = "IntegerColumnsMathOpTransform"),
+            @JsonSubTypes.Type(value = IntegerMathOpTransform.class, name = "IntegerMathOpTransform"),
+            @JsonSubTypes.Type(value = ReplaceEmptyIntegerWithValueTransform.class, name = "ReplaceEmptyIntegerWithValueTransform"),
+            @JsonSubTypes.Type(value = ReplaceInvalidWithIntegerTransform.class, name = "ReplaceInvalidWithIntegerTransform"),
+            @JsonSubTypes.Type(value = LongColumnsMathOpTransform.class, name = "LongColumnsMathOpTransform"),
+            @JsonSubTypes.Type(value = LongMathOpTransform.class, name = "LongMathOpTransform"),
+            @JsonSubTypes.Type(value = MapAllStringsExceptListTransform.class, name = "MapAllStringsExceptListTransform"),
+            @JsonSubTypes.Type(value = RemoveWhiteSpaceTransform.class, name = "RemoveWhiteSpaceTransform"),
+            @JsonSubTypes.Type(value = ReplaceEmptyStringTransform.class, name = "ReplaceEmptyStringTransform"),
+            @JsonSubTypes.Type(value = ReplaceStringTransform.class, name = "ReplaceStringTransform"),
+            @JsonSubTypes.Type(value = StringListToCategoricalSetTransform.class, name = "StringListToCategoricalSetTransform"),
+            @JsonSubTypes.Type(value = StringMapTransform.class, name = "StringMapTransform"),
+            @JsonSubTypes.Type(value = DeriveColumnsFromTimeTransform.class, name = "DeriveColumnsFromTimeTransform"),
+            @JsonSubTypes.Type(value = StringToTimeTransform.class, name = "StringToTimeTransform"),
+            @JsonSubTypes.Type(value = TimeMathOpTransform.class, name = "TimeMathOpTransform"),
+            @JsonSubTypes.Type(value = ReduceSequenceByWindowTransform.class, name = "ReduceSequenceByWindowTransform"),
+            @JsonSubTypes.Type(value = DoubleMathFunctionTransform.class, name = "DoubleMathFunctionTransform"),
+            @JsonSubTypes.Type(value = AddConstantColumnTransform.class, name = "AddConstantColumnTransform"),
+            @JsonSubTypes.Type(value = RemoveAllColumnsExceptForTransform.class, name = "RemoveAllColumnsExceptForTransform"),
+            @JsonSubTypes.Type(value = ParseDoubleTransform.class, name = "ParseDoubleTransform"),
+            @JsonSubTypes.Type(value = ConvertToString.class, name = "ConvertToStringTransform"),
+            @JsonSubTypes.Type(value = AppendStringColumnTransform.class, name = "AppendStringColumnTransform"),
+            @JsonSubTypes.Type(value = SequenceDifferenceTransform.class, name = "SequenceDifferenceTransform"),
+            @JsonSubTypes.Type(value = ReduceSequenceTransform.class, name = "ReduceSequenceTransform"),
+            @JsonSubTypes.Type(value = SequenceMovingWindowReduceTransform.class, name = "SequenceMovingWindowReduceTransform"),
+            @JsonSubTypes.Type(value = IntegerToOneHotTransform.class, name = "IntegerToOneHotTransform"),
+            @JsonSubTypes.Type(value = SequenceTrimTransform.class, name = "SequenceTrimTransform"),
+            @JsonSubTypes.Type(value = SequenceOffsetTransform.class, name = "SequenceOffsetTransform"),
+            @JsonSubTypes.Type(value = NDArrayColumnsMathOpTransform.class, name = "NDArrayColumnsMathOpTransform"),
+            @JsonSubTypes.Type(value = NDArrayDistanceTransform.class, name = "NDArrayDistanceTransform"),
+            @JsonSubTypes.Type(value = NDArrayMathFunctionTransform.class, name = "NDArrayMathFunctionTransform"),
+            @JsonSubTypes.Type(value = NDArrayScalarOpTransform.class, name = "NDArrayScalarOpTransform"),
+            @JsonSubTypes.Type(value = ChangeCaseStringTransform.class, name = "ChangeCaseStringTransform"),
+            @JsonSubTypes.Type(value = ConcatenateStringColumns.class, name = "ConcatenateStringColumns"),
+            @JsonSubTypes.Type(value = StringListToCountsNDArrayTransform.class, name = "StringListToCountsNDArrayTransform"),
+            @JsonSubTypes.Type(value = StringListToIndicesNDArrayTransform.class, name = "StringListToIndicesNDArrayTransform"),
+            @JsonSubTypes.Type(value = PivotTransform.class, name = "PivotTransform"),
+            @JsonSubTypes.Type(value = TextToCharacterIndexTransform.class, name = "TextToCharacterIndexTransform")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class TransformMixin { }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = TrivialColumnCondition.class, name = "TrivialColumnCondition"),
+            @JsonSubTypes.Type(value = CategoricalColumnCondition.class, name = "CategoricalColumnCondition"),
+            @JsonSubTypes.Type(value = DoubleColumnCondition.class, name = "DoubleColumnCondition"),
+            @JsonSubTypes.Type(value = IntegerColumnCondition.class, name = "IntegerColumnCondition"),
+            @JsonSubTypes.Type(value = LongColumnCondition.class, name = "LongColumnCondition"),
+            @JsonSubTypes.Type(value = NullWritableColumnCondition.class, name = "NullWritableColumnCondition"),
+            @JsonSubTypes.Type(value = StringColumnCondition.class, name = "StringColumnCondition"),
+            @JsonSubTypes.Type(value = TimeColumnCondition.class, name = "TimeColumnCondition"),
+            @JsonSubTypes.Type(value = StringRegexColumnCondition.class, name = "StringRegexColumnCondition"),
+            @JsonSubTypes.Type(value = BooleanCondition.class, name = "BooleanCondition"),
+            @JsonSubTypes.Type(value = NaNColumnCondition.class, name = "NaNColumnCondition"),
+            @JsonSubTypes.Type(value = InfiniteColumnCondition.class, name = "InfiniteColumnCondition"),
+            @JsonSubTypes.Type(value = SequenceLengthCondition.class, name = "SequenceLengthCondition")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class ConditionMixin { }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = ArrayWritable.class, name = "ArrayWritable"),
+            @JsonSubTypes.Type(value = BooleanWritable.class, name = "BooleanWritable"),
+            @JsonSubTypes.Type(value = ByteWritable.class, name = "ByteWritable"),
+            @JsonSubTypes.Type(value = DoubleWritable.class, name = "DoubleWritable"),
+            @JsonSubTypes.Type(value = FloatWritable.class, name = "FloatWritable"),
+            @JsonSubTypes.Type(value = IntWritable.class, name = "IntWritable"),
+            @JsonSubTypes.Type(value = LongWritable.class, name = "LongWritable"),
+            @JsonSubTypes.Type(value = NullWritable.class, name = "NullWritable"),
+            @JsonSubTypes.Type(value = Text.class, name = "Text"),
+            @JsonSubTypes.Type(value = BytesWritable.class, name = "BytesWritable")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class WritableMixin { }
+
+    @JsonInclude(JsonInclude.Include.NON_NULL)
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = ConditionFilter.class, name = "ConditionFilter"),
+            @JsonSubTypes.Type(value = FilterInvalidValues.class, name = "FilterInvalidValues"),
+            @JsonSubTypes.Type(value = InvalidNumColumns.class, name = "InvalidNumCols")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class FilterMixin { }
+
+    @JsonInclude(JsonInclude.Include.NON_NULL)
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = NumericalColumnComparator.class, name = "NumericalColumnComparator"),
+            @JsonSubTypes.Type(value = StringComparator.class, name = "StringComparator")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class SequenceComparatorMixin { }
+
+    @JsonInclude(JsonInclude.Include.NON_NULL)
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = SequenceSplitTimeSeparation.class, name = "SequenceSplitTimeSeparation"),
+            @JsonSubTypes.Type(value = SplitMaxLengthSequence.class, name = "SplitMaxLengthSequence")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class SequenceSplitMixin { }
+
+    @JsonInclude(JsonInclude.Include.NON_NULL)
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = TimeWindowFunction.class, name = "TimeWindowFunction"),
+            @JsonSubTypes.Type(value = OverlappingTimeWindowFunction.class, name = "OverlappingTimeWindowFunction")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class WindowFunctionMixin { }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = CalculateSortedRank.class, name = "CalculateSortedRank")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class CalculateSortedRankMixin { }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = DoubleWritableComparator.class, name = "DoubleWritableComparator"),
+            @JsonSubTypes.Type(value = FloatWritableComparator.class, name = "FloatWritableComparator"),
+            @JsonSubTypes.Type(value = IntWritableComparator.class, name = "IntWritableComparator"),
+            @JsonSubTypes.Type(value = LongWritableComparator.class, name = "LongWritableComparator"),
+            @JsonSubTypes.Type(value = TextWritableComparator.class, name = "TextWritableComparator")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class WritableComparatorMixin { }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = BytesAnalysis.class, name = "BytesAnalysis"),
+            @JsonSubTypes.Type(value = CategoricalAnalysis.class, name = "CategoricalAnalysis"),
+            @JsonSubTypes.Type(value = DoubleAnalysis.class, name = "DoubleAnalysis"),
+            @JsonSubTypes.Type(value = IntegerAnalysis.class, name = "IntegerAnalysis"),
+            @JsonSubTypes.Type(value = LongAnalysis.class, name = "LongAnalysis"),
+            @JsonSubTypes.Type(value = StringAnalysis.class, name = "StringAnalysis"),
+            @JsonSubTypes.Type(value = TimeAnalysis.class, name = "TimeAnalysis")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class ColumnAnalysisMixin{ }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = StringReducer.class, name = "StringReducer")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class IStringReducerMixin{ }
+}
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/LegacyMappingHelper.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/LegacyMappingHelper.java
deleted file mode 100644
index c4b478278..000000000
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/serde/legacy/LegacyMappingHelper.java
+++ /dev/null
@@ -1,535 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.api.transform.serde.legacy;
-
-import org.datavec.api.transform.Transform;
-import org.datavec.api.transform.analysis.columns.*;
-import org.datavec.api.transform.condition.BooleanCondition;
-import org.datavec.api.transform.condition.Condition;
-import org.datavec.api.transform.condition.column.*;
-import org.datavec.api.transform.condition.sequence.SequenceLengthCondition;
-import org.datavec.api.transform.condition.string.StringRegexColumnCondition;
-import org.datavec.api.transform.filter.ConditionFilter;
-import org.datavec.api.transform.filter.Filter;
-import org.datavec.api.transform.filter.FilterInvalidValues;
-import org.datavec.api.transform.filter.InvalidNumColumns;
-import org.datavec.api.transform.metadata.*;
-import org.datavec.api.transform.ndarray.NDArrayColumnsMathOpTransform;
-import org.datavec.api.transform.ndarray.NDArrayDistanceTransform;
-import org.datavec.api.transform.ndarray.NDArrayMathFunctionTransform;
-import org.datavec.api.transform.ndarray.NDArrayScalarOpTransform;
-import org.datavec.api.transform.rank.CalculateSortedRank;
-import org.datavec.api.transform.schema.Schema;
-import org.datavec.api.transform.schema.SequenceSchema;
-import org.datavec.api.transform.sequence.ReduceSequenceTransform;
-import org.datavec.api.transform.sequence.SequenceComparator;
-import org.datavec.api.transform.sequence.SequenceSplit;
-import org.datavec.api.transform.sequence.comparator.NumericalColumnComparator;
-import org.datavec.api.transform.sequence.comparator.StringComparator;
-import org.datavec.api.transform.sequence.split.SequenceSplitTimeSeparation;
-import org.datavec.api.transform.sequence.split.SplitMaxLengthSequence;
-import org.datavec.api.transform.sequence.trim.SequenceTrimTransform;
-import org.datavec.api.transform.sequence.window.OverlappingTimeWindowFunction;
-import org.datavec.api.transform.sequence.window.ReduceSequenceByWindowTransform;
-import org.datavec.api.transform.sequence.window.TimeWindowFunction;
-import org.datavec.api.transform.sequence.window.WindowFunction;
-import org.datavec.api.transform.stringreduce.IStringReducer;
-import org.datavec.api.transform.stringreduce.StringReducer;
-import org.datavec.api.transform.transform.categorical.*;
-import org.datavec.api.transform.transform.column.*;
-import org.datavec.api.transform.transform.condition.ConditionalCopyValueTransform;
-import org.datavec.api.transform.transform.condition.ConditionalReplaceValueTransform;
-import org.datavec.api.transform.transform.condition.ConditionalReplaceValueTransformWithDefault;
-import org.datavec.api.transform.transform.doubletransform.*;
-import org.datavec.api.transform.transform.integer.*;
-import org.datavec.api.transform.transform.longtransform.LongColumnsMathOpTransform;
-import org.datavec.api.transform.transform.longtransform.LongMathOpTransform;
-import org.datavec.api.transform.transform.nlp.TextToCharacterIndexTransform;
-import org.datavec.api.transform.transform.nlp.TextToTermIndexSequenceTransform;
-import org.datavec.api.transform.transform.parse.ParseDoubleTransform;
-import org.datavec.api.transform.transform.sequence.SequenceDifferenceTransform;
-import org.datavec.api.transform.transform.sequence.SequenceMovingWindowReduceTransform;
-import org.datavec.api.transform.transform.sequence.SequenceOffsetTransform;
-import org.datavec.api.transform.transform.string.*;
-import org.datavec.api.transform.transform.time.DeriveColumnsFromTimeTransform;
-import org.datavec.api.transform.transform.time.StringToTimeTransform;
-import org.datavec.api.transform.transform.time.TimeMathOpTransform;
-import org.datavec.api.writable.*;
-import org.datavec.api.writable.comparator.*;
-import org.nd4j.shade.jackson.databind.annotation.JsonDeserialize;
-
-import java.util.HashMap;
-import java.util.Map;
-
-public class LegacyMappingHelper {
-
-    public static Map<String,String> legacyMappingForClass(Class c){
-        //Need to be able to get the map - and they need to be mutable...
-        switch (c.getSimpleName()){
-            case "Transform":
-                return getLegacyMappingImageTransform();
-            case "ColumnAnalysis":
-                return getLegacyMappingColumnAnalysis();
-            case "Condition":
-                return getLegacyMappingCondition();
-            case "Filter":
-                return getLegacyMappingFilter();
-            case "ColumnMetaData":
-                return mapColumnMetaData;
-            case "CalculateSortedRank":
-                return mapCalculateSortedRank;
-            case "Schema":
-                return mapSchema;
-            case "SequenceComparator":
-                return mapSequenceComparator;
-            case "SequenceSplit":
-                return mapSequenceSplit;
-            case "WindowFunction":
-                return mapWindowFunction;
-            case "IStringReducer":
-                return mapIStringReducer;
-            case "Writable":
-                return mapWritable;
-            case "WritableComparator":
-                return mapWritableComparator;
-            case "ImageTransform":
-                return mapImageTransform;
-            default:
-                //Should never happen
-                throw new IllegalArgumentException("No legacy mapping available for class " + c.getName());
-        }
-    }
-
-    private static Map<String,String> mapTransform;
-    private static Map<String,String> mapColumnAnalysis;
-    private static Map<String,String> mapCondition;
-    private static Map<String,String> mapFilter;
-    private static Map<String,String> mapColumnMetaData;
-    private static Map<String,String> mapCalculateSortedRank;
-    private static Map<String,String> mapSchema;
-    private static Map<String,String> mapSequenceComparator;
-    private static Map<String,String> mapSequenceSplit;
-    private static Map<String,String> mapWindowFunction;
-    private static Map<String,String> mapIStringReducer;
-    private static Map<String,String> mapWritable;
-    private static Map<String,String> mapWritableComparator;
-    private static Map<String,String> mapImageTransform;
-    
-    private static synchronized Map<String,String> getLegacyMappingTransform(){
-
-        if(mapTransform == null) {
-            //The following classes all used their class short name
-            Map<String, String> m = new HashMap<>();
-            m.put("CategoricalToIntegerTransform", CategoricalToIntegerTransform.class.getName());
-            m.put("CategoricalToOneHotTransform", CategoricalToOneHotTransform.class.getName());
-            m.put("IntegerToCategoricalTransform", IntegerToCategoricalTransform.class.getName());
-            m.put("StringToCategoricalTransform", StringToCategoricalTransform.class.getName());
-            m.put("DuplicateColumnsTransform", DuplicateColumnsTransform.class.getName());
-            m.put("RemoveColumnsTransform", RemoveColumnsTransform.class.getName());
-            m.put("RenameColumnsTransform", RenameColumnsTransform.class.getName());
-            m.put("ReorderColumnsTransform", ReorderColumnsTransform.class.getName());
-            m.put("ConditionalCopyValueTransform", ConditionalCopyValueTransform.class.getName());
-            m.put("ConditionalReplaceValueTransform", ConditionalReplaceValueTransform.class.getName());
-            m.put("ConditionalReplaceValueTransformWithDefault", ConditionalReplaceValueTransformWithDefault.class.getName());
-            m.put("DoubleColumnsMathOpTransform", DoubleColumnsMathOpTransform.class.getName());
-            m.put("DoubleMathOpTransform", DoubleMathOpTransform.class.getName());
-            m.put("Log2Normalizer", Log2Normalizer.class.getName());
-            m.put("MinMaxNormalizer", MinMaxNormalizer.class.getName());
-            m.put("StandardizeNormalizer", StandardizeNormalizer.class.getName());
-            m.put("SubtractMeanNormalizer", SubtractMeanNormalizer.class.getName());
-            m.put("IntegerColumnsMathOpTransform", IntegerColumnsMathOpTransform.class.getName());
-            m.put("IntegerMathOpTransform", IntegerMathOpTransform.class.getName());
-            m.put("ReplaceEmptyIntegerWithValueTransform", ReplaceEmptyIntegerWithValueTransform.class.getName());
-            m.put("ReplaceInvalidWithIntegerTransform", ReplaceInvalidWithIntegerTransform.class.getName());
-            m.put("LongColumnsMathOpTransform", LongColumnsMathOpTransform.class.getName());
-            m.put("LongMathOpTransform", LongMathOpTransform.class.getName());
-            m.put("MapAllStringsExceptListTransform", MapAllStringsExceptListTransform.class.getName());
-            m.put("RemoveWhiteSpaceTransform", RemoveWhiteSpaceTransform.class.getName());
-            m.put("ReplaceEmptyStringTransform", ReplaceEmptyStringTransform.class.getName());
-            m.put("ReplaceStringTransform", ReplaceStringTransform.class.getName());
-            m.put("StringListToCategoricalSetTransform", StringListToCategoricalSetTransform.class.getName());
-            m.put("StringMapTransform", StringMapTransform.class.getName());
-            m.put("DeriveColumnsFromTimeTransform", DeriveColumnsFromTimeTransform.class.getName());
-            m.put("StringToTimeTransform", StringToTimeTransform.class.getName());
-            m.put("TimeMathOpTransform", TimeMathOpTransform.class.getName());
-            m.put("ReduceSequenceByWindowTransform", ReduceSequenceByWindowTransform.class.getName());
-            m.put("DoubleMathFunctionTransform", DoubleMathFunctionTransform.class.getName());
-            m.put("AddConstantColumnTransform", AddConstantColumnTransform.class.getName());
-            m.put("RemoveAllColumnsExceptForTransform", RemoveAllColumnsExceptForTransform.class.getName());
-            m.put("ParseDoubleTransform", ParseDoubleTransform.class.getName());
-            m.put("ConvertToStringTransform", ConvertToString.class.getName());
-            m.put("AppendStringColumnTransform", AppendStringColumnTransform.class.getName());
-            m.put("SequenceDifferenceTransform", SequenceDifferenceTransform.class.getName());
-            m.put("ReduceSequenceTransform", ReduceSequenceTransform.class.getName());
-            m.put("SequenceMovingWindowReduceTransform", SequenceMovingWindowReduceTransform.class.getName());
-            m.put("IntegerToOneHotTransform", IntegerToOneHotTransform.class.getName());
-            m.put("SequenceTrimTransform", SequenceTrimTransform.class.getName());
-            m.put("SequenceOffsetTransform", SequenceOffsetTransform.class.getName());
-            m.put("NDArrayColumnsMathOpTransform", NDArrayColumnsMathOpTransform.class.getName());
-            m.put("NDArrayDistanceTransform", NDArrayDistanceTransform.class.getName());
-            m.put("NDArrayMathFunctionTransform", NDArrayMathFunctionTransform.class.getName());
-            m.put("NDArrayScalarOpTransform", NDArrayScalarOpTransform.class.getName());
-            m.put("ChangeCaseStringTransform", ChangeCaseStringTransform.class.getName());
-            m.put("ConcatenateStringColumns", ConcatenateStringColumns.class.getName());
-            m.put("StringListToCountsNDArrayTransform", StringListToCountsNDArrayTransform.class.getName());
-            m.put("StringListToIndicesNDArrayTransform", StringListToIndicesNDArrayTransform.class.getName());
-            m.put("PivotTransform", PivotTransform.class.getName());
-            m.put("TextToCharacterIndexTransform", TextToCharacterIndexTransform.class.getName());
-
-            //The following never had subtype annotations, and hence will have had the default name:
-            m.put(TextToTermIndexSequenceTransform.class.getSimpleName(), TextToTermIndexSequenceTransform.class.getName());
-            m.put(ConvertToInteger.class.getSimpleName(), ConvertToInteger.class.getName());
-            m.put(ConvertToDouble.class.getSimpleName(), ConvertToDouble.class.getName());
-
-            mapTransform = m;
-        }
-
-        return mapTransform;
-    }
-
-    private static Map<String,String> getLegacyMappingColumnAnalysis(){
-        if(mapColumnAnalysis == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("BytesAnalysis", BytesAnalysis.class.getName());
-            m.put("CategoricalAnalysis", CategoricalAnalysis.class.getName());
-            m.put("DoubleAnalysis", DoubleAnalysis.class.getName());
-            m.put("IntegerAnalysis", IntegerAnalysis.class.getName());
-            m.put("LongAnalysis", LongAnalysis.class.getName());
-            m.put("StringAnalysis", StringAnalysis.class.getName());
-            m.put("TimeAnalysis", TimeAnalysis.class.getName());
-
-            //The following never had subtype annotations, and hence will have had the default name:
-            m.put(NDArrayAnalysis.class.getSimpleName(), NDArrayAnalysis.class.getName());
-
-            mapColumnAnalysis = m;
-        }
-
-        return mapColumnAnalysis;
-    }
-
-    private static Map<String,String> getLegacyMappingCondition(){
-        if(mapCondition == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("TrivialColumnCondition", TrivialColumnCondition.class.getName());
-            m.put("CategoricalColumnCondition", CategoricalColumnCondition.class.getName());
-            m.put("DoubleColumnCondition", DoubleColumnCondition.class.getName());
-            m.put("IntegerColumnCondition", IntegerColumnCondition.class.getName());
-            m.put("LongColumnCondition", LongColumnCondition.class.getName());
-            m.put("NullWritableColumnCondition", NullWritableColumnCondition.class.getName());
-            m.put("StringColumnCondition", StringColumnCondition.class.getName());
-            m.put("TimeColumnCondition", TimeColumnCondition.class.getName());
-            m.put("StringRegexColumnCondition", StringRegexColumnCondition.class.getName());
-            m.put("BooleanCondition", BooleanCondition.class.getName());
-            m.put("NaNColumnCondition", NaNColumnCondition.class.getName());
-            m.put("InfiniteColumnCondition", InfiniteColumnCondition.class.getName());
-            m.put("SequenceLengthCondition", SequenceLengthCondition.class.getName());
-
-            //The following never had subtype annotations, and hence will have had the default name:
-            m.put(InvalidValueColumnCondition.class.getSimpleName(), InvalidValueColumnCondition.class.getName());
-            m.put(BooleanColumnCondition.class.getSimpleName(), BooleanColumnCondition.class.getName());
-
-            mapCondition = m;
-        }
-
-        return mapCondition;
-    }
-
-    private static Map<String,String> getLegacyMappingFilter(){
-        if(mapFilter == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("ConditionFilter", ConditionFilter.class.getName());
-            m.put("FilterInvalidValues", FilterInvalidValues.class.getName());
-            m.put("InvalidNumCols", InvalidNumColumns.class.getName());
-
-            mapFilter = m;
-        }
-        return mapFilter;
-    }
-
-    private static Map<String,String> getLegacyMappingColumnMetaData(){
-        if(mapColumnMetaData == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("Categorical", CategoricalMetaData.class.getName());
-            m.put("Double", DoubleMetaData.class.getName());
-            m.put("Float", FloatMetaData.class.getName());
-            m.put("Integer", IntegerMetaData.class.getName());
-            m.put("Long", LongMetaData.class.getName());
-            m.put("String", StringMetaData.class.getName());
-            m.put("Time", TimeMetaData.class.getName());
-            m.put("NDArray", NDArrayMetaData.class.getName());
-
-            //The following never had subtype annotations, and hence will have had the default name:
-            m.put(BooleanMetaData.class.getSimpleName(), BooleanMetaData.class.getName());
-            m.put(BinaryMetaData.class.getSimpleName(), BinaryMetaData.class.getName());
-
-            mapColumnMetaData = m;
-        }
-
-        return mapColumnMetaData;
-    }
-
-    private static Map<String,String> getLegacyMappingCalculateSortedRank(){
-        if(mapCalculateSortedRank == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("CalculateSortedRank", CalculateSortedRank.class.getName());
-            mapCalculateSortedRank = m;
-        }
-        return mapCalculateSortedRank;
-    }
-
-    private static Map<String,String> getLegacyMappingSchema(){
-        if(mapSchema == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("Schema", Schema.class.getName());
-            m.put("SequenceSchema", SequenceSchema.class.getName());
-
-            mapSchema = m;
-        }
-        return mapSchema;
-    }
-
-    private static Map<String,String> getLegacyMappingSequenceComparator(){
-        if(mapSequenceComparator == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("NumericalColumnComparator", NumericalColumnComparator.class.getName());
-            m.put("StringComparator", StringComparator.class.getName());
-
-            mapSequenceComparator = m;
-        }
-        return mapSequenceComparator;
-    }
-
-    private static Map<String,String> getLegacyMappingSequenceSplit(){
-        if(mapSequenceSplit == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("SequenceSplitTimeSeparation", SequenceSplitTimeSeparation.class.getName());
-            m.put("SplitMaxLengthSequence", SplitMaxLengthSequence.class.getName());
-
-            mapSequenceSplit = m;
-        }
-        return mapSequenceSplit;
-    }
-
-    private static Map<String,String> getLegacyMappingWindowFunction(){
-        if(mapWindowFunction == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("TimeWindowFunction", TimeWindowFunction.class.getName());
-            m.put("OverlappingTimeWindowFunction", OverlappingTimeWindowFunction.class.getName());
-
-            mapWindowFunction = m;
-        }
-        return mapWindowFunction;
-    }
-
-    private static Map<String,String> getLegacyMappingIStringReducer(){
-        if(mapIStringReducer == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("StringReducer", StringReducer.class.getName());
-
-            mapIStringReducer = m;
-        }
-        return mapIStringReducer;
-    }
-
-    private static Map<String,String> getLegacyMappingWritable(){
-        if (mapWritable == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("ArrayWritable", ArrayWritable.class.getName());
-            m.put("BooleanWritable", BooleanWritable.class.getName());
-            m.put("ByteWritable", ByteWritable.class.getName());
-            m.put("DoubleWritable", DoubleWritable.class.getName());
-            m.put("FloatWritable", FloatWritable.class.getName());
-            m.put("IntWritable", IntWritable.class.getName());
-            m.put("LongWritable", LongWritable.class.getName());
-            m.put("NullWritable", NullWritable.class.getName());
-            m.put("Text", Text.class.getName());
-            m.put("BytesWritable", BytesWritable.class.getName());
-
-            //The following never had subtype annotations, and hence will have had the default name:
-            m.put(NDArrayWritable.class.getSimpleName(), NDArrayWritable.class.getName());
-
-            mapWritable = m;
-        }
-
-        return mapWritable;
-    }
-
-    private static Map<String,String> getLegacyMappingWritableComparator(){
-        if(mapWritableComparator == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("DoubleWritableComparator", DoubleWritableComparator.class.getName());
-            m.put("FloatWritableComparator", FloatWritableComparator.class.getName());
-            m.put("IntWritableComparator", IntWritableComparator.class.getName());
-            m.put("LongWritableComparator", LongWritableComparator.class.getName());
-            m.put("TextWritableComparator", TextWritableComparator.class.getName());
-
-            //The following never had subtype annotations, and hence will have had the default name:
-            m.put(ByteWritable.Comparator.class.getSimpleName(), ByteWritable.Comparator.class.getName());
-            m.put(FloatWritable.Comparator.class.getSimpleName(), FloatWritable.Comparator.class.getName());
-            m.put(IntWritable.Comparator.class.getSimpleName(), IntWritable.Comparator.class.getName());
-            m.put(BooleanWritable.Comparator.class.getSimpleName(), BooleanWritable.Comparator.class.getName());
-            m.put(LongWritable.Comparator.class.getSimpleName(), LongWritable.Comparator.class.getName());
-            m.put(Text.Comparator.class.getSimpleName(), Text.Comparator.class.getName());
-            m.put(LongWritable.DecreasingComparator.class.getSimpleName(), LongWritable.DecreasingComparator.class.getName());
-            m.put(DoubleWritable.Comparator.class.getSimpleName(), DoubleWritable.Comparator.class.getName());
-
-            mapWritableComparator = m;
-        }
-
-        return mapWritableComparator;
-    }
-
-    public static Map<String,String> getLegacyMappingImageTransform(){
-        if(mapImageTransform == null) {
-            Map<String, String> m = new HashMap<>();
-            m.put("EqualizeHistTransform", "org.datavec.image.transform.EqualizeHistTransform");
-            m.put("RotateImageTransform", "org.datavec.image.transform.RotateImageTransform");
-            m.put("ColorConversionTransform", "org.datavec.image.transform.ColorConversionTransform");
-            m.put("WarpImageTransform", "org.datavec.image.transform.WarpImageTransform");
-            m.put("BoxImageTransform", "org.datavec.image.transform.BoxImageTransform");
-            m.put("CropImageTransform", "org.datavec.image.transform.CropImageTransform");
-            m.put("FilterImageTransform", "org.datavec.image.transform.FilterImageTransform");
-            m.put("FlipImageTransform", "org.datavec.image.transform.FlipImageTransform");
-            m.put("LargestBlobCropTransform", "org.datavec.image.transform.LargestBlobCropTransform");
-            m.put("ResizeImageTransform", "org.datavec.image.transform.ResizeImageTransform");
-            m.put("RandomCropTransform", "org.datavec.image.transform.RandomCropTransform");
-            m.put("ScaleImageTransform", "org.datavec.image.transform.ScaleImageTransform");
-
-            mapImageTransform = m;
-        }
-        return mapImageTransform;
-    }
-
-    @JsonDeserialize(using = LegacyTransformDeserializer.class)
-    public static class TransformHelper { }
-
-    public static class LegacyTransformDeserializer extends GenericLegacyDeserializer<Transform> {
-        public LegacyTransformDeserializer() {
-            super(Transform.class, getLegacyMappingTransform());
-        }
-    }
-
-    @JsonDeserialize(using = LegacyColumnAnalysisDeserializer.class)
-    public static class ColumnAnalysisHelper { }
-
-    public static class LegacyColumnAnalysisDeserializer extends GenericLegacyDeserializer<ColumnAnalysis> {
-        public LegacyColumnAnalysisDeserializer() {
-            super(ColumnAnalysis.class, getLegacyMappingColumnAnalysis());
-        }
-    }
-
-    @JsonDeserialize(using = LegacyConditionDeserializer.class)
-    public static class ConditionHelper { }
-
-    public static class LegacyConditionDeserializer extends GenericLegacyDeserializer<Condition> {
-        public LegacyConditionDeserializer() {
-            super(Condition.class, getLegacyMappingCondition());
-        }
-    }
-
-    @JsonDeserialize(using = LegacyFilterDeserializer.class)
-    public static class FilterHelper { }
-
-    public static class LegacyFilterDeserializer extends GenericLegacyDeserializer<Filter> {
-        public LegacyFilterDeserializer() {
-            super(Filter.class, getLegacyMappingFilter());
-        }
-    }
-
-    @JsonDeserialize(using = LegacyColumnMetaDataDeserializer.class)
-    public static class ColumnMetaDataHelper { }
-
-    public static class LegacyColumnMetaDataDeserializer extends GenericLegacyDeserializer<ColumnMetaData> {
-        public LegacyColumnMetaDataDeserializer() {
-            super(ColumnMetaData.class, getLegacyMappingColumnMetaData());
-        }
-    }
-
-    @JsonDeserialize(using = LegacyCalculateSortedRankDeserializer.class)
-    public static class CalculateSortedRankHelper { }
-
-    public static class LegacyCalculateSortedRankDeserializer extends GenericLegacyDeserializer<CalculateSortedRank> {
-        public LegacyCalculateSortedRankDeserializer() {
-            super(CalculateSortedRank.class, getLegacyMappingCalculateSortedRank());
-        }
-    }
-
-    @JsonDeserialize(using = LegacySchemaDeserializer.class)
-    public static class SchemaHelper { }
-
-    public static class LegacySchemaDeserializer extends GenericLegacyDeserializer<Schema> {
-        public LegacySchemaDeserializer() {
-            super(Schema.class, getLegacyMappingSchema());
-        }
-    }
-
-    @JsonDeserialize(using = LegacySequenceComparatorDeserializer.class)
-    public static class SequenceComparatorHelper { }
-
-    public static class LegacySequenceComparatorDeserializer extends GenericLegacyDeserializer<SequenceComparator> {
-        public LegacySequenceComparatorDeserializer() {
-            super(SequenceComparator.class, getLegacyMappingSequenceComparator());
-        }
-    }
-
-    @JsonDeserialize(using = LegacySequenceSplitDeserializer.class)
-    public static class SequenceSplitHelper { }
-
-    public static class LegacySequenceSplitDeserializer extends GenericLegacyDeserializer<SequenceSplit> {
-        public LegacySequenceSplitDeserializer() {
-            super(SequenceSplit.class, getLegacyMappingSequenceSplit());
-        }
-    }
-
-    @JsonDeserialize(using = LegacyWindowFunctionDeserializer.class)
-    public static class WindowFunctionHelper { }
-
-    public static class LegacyWindowFunctionDeserializer extends GenericLegacyDeserializer<WindowFunction> {
-        public LegacyWindowFunctionDeserializer() {
-            super(WindowFunction.class, getLegacyMappingWindowFunction());
-        }
-    }
-
-
-    @JsonDeserialize(using = LegacyIStringReducerDeserializer.class)
-    public static class IStringReducerHelper { }
-
-    public static class LegacyIStringReducerDeserializer extends GenericLegacyDeserializer<IStringReducer> {
-        public LegacyIStringReducerDeserializer() {
-            super(IStringReducer.class, getLegacyMappingIStringReducer());
-        }
-    }
-
-
-    @JsonDeserialize(using = LegacyWritableDeserializer.class)
-    public static class WritableHelper { }
-
-    public static class LegacyWritableDeserializer extends GenericLegacyDeserializer<Writable> {
-        public LegacyWritableDeserializer() {
-            super(Writable.class, getLegacyMappingWritable());
-        }
-    }
-
-    @JsonDeserialize(using = LegacyWritableComparatorDeserializer.class)
-    public static class WritableComparatorHelper { }
-
-    public static class LegacyWritableComparatorDeserializer extends GenericLegacyDeserializer<WritableComparator> {
-        public LegacyWritableComparatorDeserializer() {
-            super(WritableComparator.class, getLegacyMappingWritableComparator());
-        }
-    }
-}
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/transform/stringreduce/IStringReducer.java b/datavec/datavec-api/src/main/java/org/datavec/api/transform/stringreduce/IStringReducer.java
index f43e189d8..54bd7b7c8 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/transform/stringreduce/IStringReducer.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/transform/stringreduce/IStringReducer.java
@@ -17,7 +17,6 @@
 package org.datavec.api.transform.stringreduce;
 
 import org.datavec.api.transform.schema.Schema;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.Writable;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
@@ -31,8 +30,7 @@ import java.util.List;
  * a single List<Writable>
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.IStringReducerHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface IStringReducer extends Serializable {
 
     /**
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/util/ndarray/RecordConverter.java b/datavec/datavec-api/src/main/java/org/datavec/api/util/ndarray/RecordConverter.java
index e3ba797c8..c55d4d3bb 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/util/ndarray/RecordConverter.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/util/ndarray/RecordConverter.java
@@ -16,7 +16,7 @@
 
 package org.datavec.api.util.ndarray;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import it.unimi.dsi.fastutil.doubles.DoubleArrayList;
 import lombok.NonNull;
 import org.datavec.api.timeseries.util.TimeSeriesWritableUtils;
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/writable/ByteWritable.java b/datavec/datavec-api/src/main/java/org/datavec/api/writable/ByteWritable.java
index 2584d8b9f..ae5e3a567 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/writable/ByteWritable.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/writable/ByteWritable.java
@@ -17,7 +17,7 @@
 package org.datavec.api.writable;
 
 
-import com.google.common.math.DoubleMath;
+import org.nd4j.shade.guava.math.DoubleMath;
 import org.datavec.api.io.WritableComparable;
 import org.datavec.api.io.WritableComparator;
 import org.nd4j.shade.jackson.annotation.JsonProperty;
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/writable/DoubleWritable.java b/datavec/datavec-api/src/main/java/org/datavec/api/writable/DoubleWritable.java
index 72f4b3c5b..39f41c076 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/writable/DoubleWritable.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/writable/DoubleWritable.java
@@ -17,7 +17,7 @@
 package org.datavec.api.writable;
 
 
-import com.google.common.math.DoubleMath;
+import org.nd4j.shade.guava.math.DoubleMath;
 import org.datavec.api.io.WritableComparable;
 import org.datavec.api.io.WritableComparator;
 import org.nd4j.shade.jackson.annotation.JsonProperty;
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/writable/FloatWritable.java b/datavec/datavec-api/src/main/java/org/datavec/api/writable/FloatWritable.java
index 1b54e7d54..f0ab62bef 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/writable/FloatWritable.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/writable/FloatWritable.java
@@ -17,7 +17,7 @@
 package org.datavec.api.writable;
 
 
-import com.google.common.math.DoubleMath;
+import org.nd4j.shade.guava.math.DoubleMath;
 import org.datavec.api.io.WritableComparable;
 import org.datavec.api.io.WritableComparator;
 import org.nd4j.shade.jackson.annotation.JsonProperty;
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/writable/IntWritable.java b/datavec/datavec-api/src/main/java/org/datavec/api/writable/IntWritable.java
index 3803c8098..1c127e0f8 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/writable/IntWritable.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/writable/IntWritable.java
@@ -17,7 +17,7 @@
 package org.datavec.api.writable;
 
 
-import com.google.common.math.DoubleMath;
+import org.nd4j.shade.guava.math.DoubleMath;
 import org.datavec.api.io.WritableComparable;
 import org.datavec.api.io.WritableComparator;
 import org.nd4j.shade.jackson.annotation.JsonProperty;
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/writable/LongWritable.java b/datavec/datavec-api/src/main/java/org/datavec/api/writable/LongWritable.java
index 58cd45829..4a767a183 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/writable/LongWritable.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/writable/LongWritable.java
@@ -17,7 +17,7 @@
 package org.datavec.api.writable;
 
 
-import com.google.common.math.DoubleMath;
+import org.nd4j.shade.guava.math.DoubleMath;
 import org.datavec.api.io.WritableComparable;
 import org.datavec.api.io.WritableComparator;
 import org.nd4j.shade.jackson.annotation.JsonProperty;
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/writable/Writable.java b/datavec/datavec-api/src/main/java/org/datavec/api/writable/Writable.java
index 30eb5e25e..5085dd3f2 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/writable/Writable.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/writable/Writable.java
@@ -16,7 +16,6 @@
 
 package org.datavec.api.writable;
 
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
 
 import java.io.DataInput;
@@ -60,8 +59,7 @@ import java.io.Serializable;
  *     }
  * </pre></blockquote></p>
  */
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.WritableHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface Writable extends Serializable {
     /**
      * Serialize the fields of this object to <code>out</code>.
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/writable/batch/NDArrayRecordBatch.java b/datavec/datavec-api/src/main/java/org/datavec/api/writable/batch/NDArrayRecordBatch.java
index 4d638124e..0a5ddddb8 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/writable/batch/NDArrayRecordBatch.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/writable/batch/NDArrayRecordBatch.java
@@ -16,7 +16,7 @@
 
 package org.datavec.api.writable.batch;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import lombok.Data;
 import lombok.NonNull;
 import org.datavec.api.writable.NDArrayWritable;
diff --git a/datavec/datavec-api/src/main/java/org/datavec/api/writable/comparator/WritableComparator.java b/datavec/datavec-api/src/main/java/org/datavec/api/writable/comparator/WritableComparator.java
index 07ef7ee56..b8e540f61 100644
--- a/datavec/datavec-api/src/main/java/org/datavec/api/writable/comparator/WritableComparator.java
+++ b/datavec/datavec-api/src/main/java/org/datavec/api/writable/comparator/WritableComparator.java
@@ -16,16 +16,13 @@
 
 package org.datavec.api.writable.comparator;
 
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
 import org.datavec.api.writable.Writable;
-import org.nd4j.shade.jackson.annotation.JsonSubTypes;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
 
 import java.io.Serializable;
 import java.util.Comparator;
 
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyMappingHelper.WritableComparatorHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface WritableComparator extends Comparator<Writable>, Serializable {
 
 }
diff --git a/datavec/datavec-api/src/test/java/org/datavec/api/split/InputSplitTests.java b/datavec/datavec-api/src/test/java/org/datavec/api/split/InputSplitTests.java
index 647c0af65..e8ce37bd3 100644
--- a/datavec/datavec-api/src/test/java/org/datavec/api/split/InputSplitTests.java
+++ b/datavec/datavec-api/src/test/java/org/datavec/api/split/InputSplitTests.java
@@ -16,7 +16,7 @@
 
 package org.datavec.api.split;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import org.datavec.api.io.filters.BalancedPathFilter;
 import org.datavec.api.io.filters.RandomPathFilter;
 import org.datavec.api.io.labels.ParentPathLabelGenerator;
diff --git a/datavec/datavec-api/src/test/java/org/datavec/api/split/parittion/PartitionerTests.java b/datavec/datavec-api/src/test/java/org/datavec/api/split/parittion/PartitionerTests.java
index e67fc1f61..c9fb57eb9 100644
--- a/datavec/datavec-api/src/test/java/org/datavec/api/split/parittion/PartitionerTests.java
+++ b/datavec/datavec-api/src/test/java/org/datavec/api/split/parittion/PartitionerTests.java
@@ -16,7 +16,7 @@
 
 package org.datavec.api.split.parittion;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import org.datavec.api.conf.Configuration;
 import org.datavec.api.split.FileSplit;
 import org.datavec.api.split.partition.NumberOfRecordsPartitioner;
diff --git a/datavec/datavec-api/src/test/java/org/datavec/api/transform/schema/TestJsonYaml.java b/datavec/datavec-api/src/test/java/org/datavec/api/transform/schema/TestJsonYaml.java
index c98b58381..dff90f8b9 100644
--- a/datavec/datavec-api/src/test/java/org/datavec/api/transform/schema/TestJsonYaml.java
+++ b/datavec/datavec-api/src/test/java/org/datavec/api/transform/schema/TestJsonYaml.java
@@ -78,8 +78,9 @@ public class TestJsonYaml {
     public void testMissingPrimitives() {
 
         Schema schema = new Schema.Builder().addColumnDouble("Dbl2", null, 100.0, false, false).build();
-
-        String strJson = "{\n" + "  \"Schema\" : {\n" + "    \"columns\" : [ {\n" + "      \"Double\" : {\n"
+        //Legacy format JSON
+        String strJson = "{\n" + "  \"Schema\" : {\n"
+                        + "    \"columns\" : [ {\n" + "      \"Double\" : {\n"
                         + "        \"name\" : \"Dbl2\",\n" + "        \"maxAllowedValue\" : 100.0\n" +
                         //"        \"allowNaN\" : false,\n" +           //Normally included: but exclude here to test
                         //"        \"allowInfinite\" : false\n" +       //Normally included: but exclude here to test
diff --git a/datavec/datavec-api/src/test/java/org/datavec/api/writable/RecordConverterTest.java b/datavec/datavec-api/src/test/java/org/datavec/api/writable/RecordConverterTest.java
index ce7d779dc..6dfacdd93 100644
--- a/datavec/datavec-api/src/test/java/org/datavec/api/writable/RecordConverterTest.java
+++ b/datavec/datavec-api/src/test/java/org/datavec/api/writable/RecordConverterTest.java
@@ -16,7 +16,7 @@
 
 package org.datavec.api.writable;
 
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.collect.Lists;
 import org.datavec.api.transform.schema.Schema;
 import org.datavec.api.util.ndarray.RecordConverter;
 import org.junit.Test;
diff --git a/datavec/datavec-arrow/pom.xml b/datavec/datavec-arrow/pom.xml
index 4d4381790..645971a45 100644
--- a/datavec/datavec-arrow/pom.xml
+++ b/datavec/datavec-arrow/pom.xml
@@ -34,36 +34,6 @@
             <artifactId>nd4j-arrow</artifactId>
             <version>${project.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-core</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-databind</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-annotations</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.dataformat</groupId>
-            <artifactId>jackson-dataformat-yaml</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.dataformat</groupId>
-            <artifactId>jackson-dataformat-xml</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.datatype</groupId>
-            <artifactId>jackson-datatype-joda</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
         <dependency>
             <groupId>org.datavec</groupId>
             <artifactId>datavec-api</artifactId>
@@ -74,11 +44,6 @@
             <artifactId>hppc</artifactId>
             <version>${hppc.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.google.guava</groupId>
-            <artifactId>guava</artifactId>
-            <version>${guava.version}</version>
-        </dependency>
         <dependency>
             <groupId>org.apache.arrow</groupId>
             <artifactId>arrow-vector</artifactId>
diff --git a/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/recordreader/BaseImageRecordReader.java b/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/recordreader/BaseImageRecordReader.java
index d80f12a2e..a962dd84a 100644
--- a/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/recordreader/BaseImageRecordReader.java
+++ b/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/recordreader/BaseImageRecordReader.java
@@ -16,7 +16,7 @@
 
 package org.datavec.image.recordreader;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import lombok.Getter;
 import lombok.Setter;
 import lombok.extern.slf4j.Slf4j;
diff --git a/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/serde/LegacyImageMappingHelper.java b/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/serde/LegacyImageMappingHelper.java
deleted file mode 100644
index 5e7b09c12..000000000
--- a/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/serde/LegacyImageMappingHelper.java
+++ /dev/null
@@ -1,35 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.image.serde;
-
-import org.datavec.api.transform.serde.legacy.GenericLegacyDeserializer;
-import org.datavec.api.transform.serde.legacy.LegacyMappingHelper;
-import org.datavec.image.transform.ImageTransform;
-import org.nd4j.shade.jackson.databind.annotation.JsonDeserialize;
-
-public class LegacyImageMappingHelper {
-
-    @JsonDeserialize(using = LegacyImageTransformDeserializer.class)
-    public static class ImageTransformHelper { }
-
-    public static class LegacyImageTransformDeserializer extends GenericLegacyDeserializer<ImageTransform> {
-        public LegacyImageTransformDeserializer() {
-            super(ImageTransform.class, LegacyMappingHelper.getLegacyMappingImageTransform());
-        }
-    }
-    
-}
diff --git a/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/transform/ImageTransform.java b/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/transform/ImageTransform.java
index 39239494b..afcdf894f 100644
--- a/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/transform/ImageTransform.java
+++ b/datavec/datavec-data/datavec-data-image/src/main/java/org/datavec/image/transform/ImageTransform.java
@@ -16,11 +16,8 @@
 
 package org.datavec.image.transform;
 
-import lombok.Data;
 import org.datavec.image.data.ImageWritable;
-import org.datavec.image.serde.LegacyImageMappingHelper;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
-import org.nd4j.shade.jackson.annotation.JsonSubTypes;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
 
 import java.util.Random;
@@ -31,8 +28,7 @@ import java.util.Random;
  * @author saudet
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyImageMappingHelper.ImageTransformHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface ImageTransform {
 
     /**
diff --git a/datavec/datavec-hadoop/pom.xml b/datavec/datavec-hadoop/pom.xml
index 38889228c..c95e6d3bc 100644
--- a/datavec/datavec-hadoop/pom.xml
+++ b/datavec/datavec-hadoop/pom.xml
@@ -50,11 +50,6 @@
             <artifactId>netty</artifactId>
             <version>${netty.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.google.guava</groupId>
-            <artifactId>guava</artifactId>
-            <version>${guava.version}</version>
-        </dependency>
         <dependency>
             <groupId>org.apache.commons</groupId>
             <artifactId>commons-compress</artifactId>
diff --git a/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReader.java b/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReader.java
index 01bef8fa9..58f7a57db 100644
--- a/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReader.java
+++ b/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReader.java
@@ -16,7 +16,7 @@
 
 package org.datavec.hadoop.records.reader;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import org.apache.hadoop.conf.Configuration;
 import org.apache.hadoop.fs.Path;
 import org.apache.hadoop.io.*;
diff --git a/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReaderMultipleParts.java b/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReaderMultipleParts.java
index cf1a801f5..1cbe47176 100644
--- a/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReaderMultipleParts.java
+++ b/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReaderMultipleParts.java
@@ -16,7 +16,7 @@
 
 package org.datavec.hadoop.records.reader;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import org.apache.hadoop.conf.Configuration;
 import org.apache.hadoop.fs.Path;
 import org.apache.hadoop.io.*;
diff --git a/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReaderMultiplePartsSomeEmpty.java b/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReaderMultiplePartsSomeEmpty.java
index 992c91312..faf41cbb4 100644
--- a/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReaderMultiplePartsSomeEmpty.java
+++ b/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/reader/TestMapFileRecordReaderMultiplePartsSomeEmpty.java
@@ -16,7 +16,7 @@
 
 package org.datavec.hadoop.records.reader;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import org.apache.hadoop.conf.Configuration;
 import org.apache.hadoop.fs.Path;
 import org.apache.hadoop.io.*;
diff --git a/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/writer/TestMapFileRecordWriter.java b/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/writer/TestMapFileRecordWriter.java
index d35becd7a..7cd112c63 100644
--- a/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/writer/TestMapFileRecordWriter.java
+++ b/datavec/datavec-hadoop/src/test/java/org/datavec/hadoop/records/writer/TestMapFileRecordWriter.java
@@ -16,7 +16,7 @@
 
 package org.datavec.hadoop.records.writer;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import org.datavec.api.records.converter.RecordReaderConverter;
 import org.datavec.api.records.reader.RecordReader;
 import org.datavec.api.records.reader.SequenceRecordReader;
diff --git a/datavec/datavec-local/src/main/java/org/datavec/local/transforms/join/ExecuteJoinFromCoGroupFlatMapFunctionAdapter.java b/datavec/datavec-local/src/main/java/org/datavec/local/transforms/join/ExecuteJoinFromCoGroupFlatMapFunctionAdapter.java
index 5360eae7e..276d79f88 100644
--- a/datavec/datavec-local/src/main/java/org/datavec/local/transforms/join/ExecuteJoinFromCoGroupFlatMapFunctionAdapter.java
+++ b/datavec/datavec-local/src/main/java/org/datavec/local/transforms/join/ExecuteJoinFromCoGroupFlatMapFunctionAdapter.java
@@ -16,7 +16,7 @@
 
 package org.datavec.local.transforms.join;
 
-import com.google.common.collect.Iterables;
+import org.nd4j.shade.guava.collect.Iterables;
 import org.datavec.api.transform.join.Join;
 import org.datavec.api.writable.Writable;
 import org.datavec.local.transforms.functions.FlatMapFunctionAdapter;
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml
index 940ad01cc..605b13b70 100644
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/pom.xml
@@ -64,12 +64,6 @@
             <version>${datavec.version}</version>
         </dependency>
 
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-cluster_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-
         <dependency>
             <groupId>joda-time</groupId>
             <artifactId>joda-time</artifactId>
@@ -106,40 +100,10 @@
             <version>${snakeyaml.version}</version>
         </dependency>
 
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-core</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-databind</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-annotations</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-
-        <dependency>
-            <groupId>com.fasterxml.jackson.datatype</groupId>
-            <artifactId>jackson-datatype-jdk8</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-
-        <dependency>
-            <groupId>com.fasterxml.jackson.datatype</groupId>
-            <artifactId>jackson-datatype-jsr310</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-
         <dependency>
             <groupId>com.typesafe.play</groupId>
             <artifactId>play-java_2.11</artifactId>
-            <version>${play.version}</version>
+            <version>${playframework.version}</version>
             <exclusions>
                 <exclusion>
                     <groupId>com.google.code.findbugs</groupId>
@@ -161,25 +125,31 @@
         <dependency>
             <groupId>com.typesafe.play</groupId>
             <artifactId>play-json_2.11</artifactId>
-            <version>${play.version}</version>
+            <version>${playframework.version}</version>
         </dependency>
 
         <dependency>
             <groupId>com.typesafe.play</groupId>
             <artifactId>play-server_2.11</artifactId>
-            <version>${play.version}</version>
+            <version>${playframework.version}</version>
         </dependency>
 
         <dependency>
             <groupId>com.typesafe.play</groupId>
             <artifactId>play_2.11</artifactId>
-            <version>${play.version}</version>
+            <version>${playframework.version}</version>
         </dependency>
 
         <dependency>
             <groupId>com.typesafe.play</groupId>
             <artifactId>play-netty-server_2.11</artifactId>
-            <version>${play.version}</version>
+            <version>${playframework.version}</version>
+        </dependency>
+
+        <dependency>
+            <groupId>com.typesafe.akka</groupId>
+            <artifactId>akka-cluster_2.11</artifactId>
+            <version>2.5.23</version>
         </dependency>
 
         <dependency>
@@ -194,6 +164,12 @@
             <artifactId>jcommander</artifactId>
             <version>${jcommander.version}</version>
         </dependency>
+
+        <dependency>
+            <groupId>org.apache.spark</groupId>
+            <artifactId>spark-core_2.11</artifactId>
+            <version>${spark.version}</version>
+        </dependency>
     </dependencies>
 
     <profiles>
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/CSVSparkTransformServer.java b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/CSVSparkTransformServer.java
index 893ce4218..f20799905 100644
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/CSVSparkTransformServer.java
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/CSVSparkTransformServer.java
@@ -24,12 +24,16 @@ import org.apache.commons.io.FileUtils;
 import org.datavec.api.transform.TransformProcess;
 import org.datavec.image.transform.ImageTransformProcess;
 import org.datavec.spark.transform.model.*;
+import play.BuiltInComponents;
 import play.Mode;
+import play.routing.Router;
 import play.routing.RoutingDsl;
 import play.server.Server;
 
 import java.io.File;
 import java.io.IOException;
+import java.util.Base64;
+import java.util.Random;
 
 import static play.mvc.Results.*;
 
@@ -66,9 +70,6 @@ public class CSVSparkTransformServer extends SparkTransformServer {
             System.exit(1);
         }
 
-        RoutingDsl routingDsl = new RoutingDsl();
-
-
         if (jsonPath != null) {
             String json = FileUtils.readFileToString(new File(jsonPath));
             TransformProcess transformProcess = TransformProcess.fromJson(json);
@@ -78,8 +79,26 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                     + "to /transformprocess");
         }
 
+        //Set play secret key, if required
+        //http://www.playframework.com/documentation/latest/ApplicationSecret
+        String crypto = System.getProperty("play.crypto.secret");
+        if (crypto == null || "changeme".equals(crypto) || "".equals(crypto) ) {
+            byte[] newCrypto = new byte[1024];
 
-        routingDsl.GET("/transformprocess").routeTo(FunctionUtil.function0((() -> {
+            new Random().nextBytes(newCrypto);
+
+            String base64 = Base64.getEncoder().encodeToString(newCrypto);
+            System.setProperty("play.crypto.secret", base64);
+        }
+
+
+        server = Server.forRouter(Mode.PROD, port, this::createRouter);
+    }
+
+    protected Router createRouter(BuiltInComponents b){
+        RoutingDsl routingDsl = RoutingDsl.fromComponents(b);
+
+        routingDsl.GET("/transformprocess").routingTo(req -> {
             try {
                 if (transform == null)
                     return badRequest();
@@ -88,11 +107,11 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                 log.error("Error in GET /transformprocess",e);
                 return internalServerError(e.getMessage());
             }
-        })));
+        });
 
-        routingDsl.POST("/transformprocess").routeTo(FunctionUtil.function0((() -> {
+        routingDsl.POST("/transformprocess").routingTo(req -> {
             try {
-                TransformProcess transformProcess = TransformProcess.fromJson(getJsonText());
+                TransformProcess transformProcess = TransformProcess.fromJson(getJsonText(req));
                 setCSVTransformProcess(transformProcess);
                 log.info("Transform process initialized");
                 return ok(objectMapper.writeValueAsString(transformProcess)).as(contentType);
@@ -100,12 +119,12 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                 log.error("Error in POST /transformprocess",e);
                 return internalServerError(e.getMessage());
             }
-        })));
+        });
 
-        routingDsl.POST("/transformincremental").routeTo(FunctionUtil.function0((() -> {
-            if (isSequence()) {
+        routingDsl.POST("/transformincremental").routingTo(req -> {
+            if (isSequence(req)) {
                 try {
-                    BatchCSVRecord record = objectMapper.readValue(getJsonText(), BatchCSVRecord.class);
+                    BatchCSVRecord record = objectMapper.readValue(getJsonText(req), BatchCSVRecord.class);
                     if (record == null)
                         return badRequest();
                     return ok(objectMapper.writeValueAsString(transformSequenceIncremental(record))).as(contentType);
@@ -115,7 +134,7 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                 }
             } else {
                 try {
-                    SingleCSVRecord record = objectMapper.readValue(getJsonText(), SingleCSVRecord.class);
+                    SingleCSVRecord record = objectMapper.readValue(getJsonText(req), SingleCSVRecord.class);
                     if (record == null)
                         return badRequest();
                     return ok(objectMapper.writeValueAsString(transformIncremental(record))).as(contentType);
@@ -124,12 +143,12 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                     return internalServerError(e.getMessage());
                 }
             }
-        })));
+        });
 
-        routingDsl.POST("/transform").routeTo(FunctionUtil.function0((() -> {
-            if (isSequence()) {
+        routingDsl.POST("/transform").routingTo(req -> {
+            if (isSequence(req)) {
                 try {
-                    SequenceBatchCSVRecord batch = transformSequence(objectMapper.readValue(getJsonText(), SequenceBatchCSVRecord.class));
+                    SequenceBatchCSVRecord batch = transformSequence(objectMapper.readValue(getJsonText(req), SequenceBatchCSVRecord.class));
                     if (batch == null)
                         return badRequest();
                     return ok(objectMapper.writeValueAsString(batch)).as(contentType);
@@ -139,7 +158,7 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                 }
             } else {
                 try {
-                    BatchCSVRecord input = objectMapper.readValue(getJsonText(), BatchCSVRecord.class);
+                    BatchCSVRecord input = objectMapper.readValue(getJsonText(req), BatchCSVRecord.class);
                     BatchCSVRecord batch = transform(input);
                     if (batch == null)
                         return badRequest();
@@ -149,14 +168,12 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                     return internalServerError(e.getMessage());
                 }
             }
+        });
 
-
-        })));
-
-        routingDsl.POST("/transformincrementalarray").routeTo(FunctionUtil.function0((() -> {
-            if (isSequence()) {
+        routingDsl.POST("/transformincrementalarray").routingTo(req -> {
+            if (isSequence(req)) {
                 try {
-                    BatchCSVRecord record = objectMapper.readValue(getJsonText(), BatchCSVRecord.class);
+                    BatchCSVRecord record = objectMapper.readValue(getJsonText(req), BatchCSVRecord.class);
                     if (record == null)
                         return badRequest();
                     return ok(objectMapper.writeValueAsString(transformSequenceArrayIncremental(record))).as(contentType);
@@ -166,7 +183,7 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                 }
             } else {
                 try {
-                    SingleCSVRecord record = objectMapper.readValue(getJsonText(), SingleCSVRecord.class);
+                    SingleCSVRecord record = objectMapper.readValue(getJsonText(req), SingleCSVRecord.class);
                     if (record == null)
                         return badRequest();
                     return ok(objectMapper.writeValueAsString(transformArrayIncremental(record))).as(contentType);
@@ -175,13 +192,12 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                     return internalServerError(e.getMessage());
                 }
             }
+        });
 
-        })));
-
-        routingDsl.POST("/transformarray").routeTo(FunctionUtil.function0((() -> {
-            if (isSequence()) {
+        routingDsl.POST("/transformarray").routingTo(req -> {
+            if (isSequence(req)) {
                 try {
-                    SequenceBatchCSVRecord batchCSVRecord = objectMapper.readValue(getJsonText(), SequenceBatchCSVRecord.class);
+                    SequenceBatchCSVRecord batchCSVRecord = objectMapper.readValue(getJsonText(req), SequenceBatchCSVRecord.class);
                     if (batchCSVRecord == null)
                         return badRequest();
                     return ok(objectMapper.writeValueAsString(transformSequenceArray(batchCSVRecord))).as(contentType);
@@ -191,7 +207,7 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                 }
             } else {
                 try {
-                    BatchCSVRecord batchCSVRecord = objectMapper.readValue(getJsonText(), BatchCSVRecord.class);
+                    BatchCSVRecord batchCSVRecord = objectMapper.readValue(getJsonText(req), BatchCSVRecord.class);
                     if (batchCSVRecord == null)
                         return badRequest();
                     return ok(objectMapper.writeValueAsString(transformArray(batchCSVRecord))).as(contentType);
@@ -200,10 +216,9 @@ public class CSVSparkTransformServer extends SparkTransformServer {
                     return internalServerError(e.getMessage());
                 }
             }
-        })));
+        });
 
-
-        server = Server.forRouter(routingDsl.build(), Mode.PROD, port);
+        return routingDsl.build();
     }
 
     public static void main(String[] args) throws Exception {
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/FunctionUtil.java b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/FunctionUtil.java
deleted file mode 100644
index 6c4874b02..000000000
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/FunctionUtil.java
+++ /dev/null
@@ -1,41 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform;
-
-import play.libs.F;
-import play.mvc.Result;
-
-import java.util.function.Function;
-import java.util.function.Supplier;
-
-/**
- * Utility methods for Routing
- *
- * @author Alex Black
- */
-public class FunctionUtil {
-
-
-    public static F.Function0<Result> function0(Supplier<Result> supplier) {
-        return supplier::get;
-    }
-
-    public static <T> F.Function<T, Result> function(Function<T, Result> function) {
-        return function::apply;
-    }
-
-}
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/ImageSparkTransformServer.java b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/ImageSparkTransformServer.java
index 29c1e1bd7..f8675f139 100644
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/ImageSparkTransformServer.java
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/ImageSparkTransformServer.java
@@ -24,8 +24,11 @@ import org.apache.commons.io.FileUtils;
 import org.datavec.api.transform.TransformProcess;
 import org.datavec.image.transform.ImageTransformProcess;
 import org.datavec.spark.transform.model.*;
+import play.BuiltInComponents;
 import play.Mode;
+import play.libs.Files;
 import play.mvc.Http;
+import play.routing.Router;
 import play.routing.RoutingDsl;
 import play.server.Server;
 
@@ -33,6 +36,7 @@ import java.io.File;
 import java.io.IOException;
 import java.util.ArrayList;
 import java.util.List;
+import java.util.function.Function;
 
 import static play.mvc.Controller.request;
 import static play.mvc.Results.*;
@@ -62,8 +66,6 @@ public class ImageSparkTransformServer extends SparkTransformServer {
             System.exit(1);
         }
 
-        RoutingDsl routingDsl = new RoutingDsl();
-
         if (jsonPath != null) {
             String json = FileUtils.readFileToString(new File(jsonPath));
             ImageTransformProcess transformProcess = ImageTransformProcess.fromJson(json);
@@ -73,7 +75,13 @@ public class ImageSparkTransformServer extends SparkTransformServer {
                     + "to /transformprocess");
         }
 
-        routingDsl.GET("/transformprocess").routeTo(FunctionUtil.function0((() -> {
+        server = Server.forRouter(Mode.PROD, port, this::createRouter);
+    }
+
+    protected Router createRouter(BuiltInComponents builtInComponents){
+        RoutingDsl routingDsl = RoutingDsl.fromComponents(builtInComponents);
+
+        routingDsl.GET("/transformprocess").routingTo(req -> {
             try {
                 if (transform == null)
                     return badRequest();
@@ -83,11 +91,11 @@ public class ImageSparkTransformServer extends SparkTransformServer {
                 e.printStackTrace();
                 return internalServerError();
             }
-        })));
+        });
 
-        routingDsl.POST("/transformprocess").routeTo(FunctionUtil.function0((() -> {
+        routingDsl.POST("/transformprocess").routingTo(req -> {
             try {
-                ImageTransformProcess transformProcess = ImageTransformProcess.fromJson(getJsonText());
+                ImageTransformProcess transformProcess = ImageTransformProcess.fromJson(getJsonText(req));
                 setImageTransformProcess(transformProcess);
                 log.info("Transform process initialized");
                 return ok(objectMapper.writeValueAsString(transformProcess)).as(contentType);
@@ -95,11 +103,11 @@ public class ImageSparkTransformServer extends SparkTransformServer {
                 e.printStackTrace();
                 return internalServerError();
             }
-        })));
+        });
 
-        routingDsl.POST("/transformincrementalarray").routeTo(FunctionUtil.function0((() -> {
+        routingDsl.POST("/transformincrementalarray").routingTo(req -> {
             try {
-                SingleImageRecord record = objectMapper.readValue(getJsonText(), SingleImageRecord.class);
+                SingleImageRecord record = objectMapper.readValue(getJsonText(req), SingleImageRecord.class);
                 if (record == null)
                     return badRequest();
                 return ok(objectMapper.writeValueAsString(transformIncrementalArray(record))).as(contentType);
@@ -107,17 +115,17 @@ public class ImageSparkTransformServer extends SparkTransformServer {
                 e.printStackTrace();
                 return internalServerError();
             }
-        })));
+        });
 
-        routingDsl.POST("/transformincrementalimage").routeTo(FunctionUtil.function0((() -> {
+        routingDsl.POST("/transformincrementalimage").routingTo(req -> {
             try {
-                Http.MultipartFormData body = request().body().asMultipartFormData();
-                List<Http.MultipartFormData.FilePart> files = body.getFiles();
-                if (files.size() == 0 || files.get(0).getFile() == null) {
+                Http.MultipartFormData<Files.TemporaryFile> body = req.body().asMultipartFormData();
+                List<Http.MultipartFormData.FilePart<Files.TemporaryFile>> files = body.getFiles();
+                if (files.isEmpty() || files.get(0).getRef() == null ) {
                     return badRequest();
                 }
 
-                File file = files.get(0).getFile();
+                File file = files.get(0).getRef().path().toFile();
                 SingleImageRecord record = new SingleImageRecord(file.toURI());
 
                 return ok(objectMapper.writeValueAsString(transformIncrementalArray(record))).as(contentType);
@@ -125,11 +133,11 @@ public class ImageSparkTransformServer extends SparkTransformServer {
                 e.printStackTrace();
                 return internalServerError();
             }
-        })));
+        });
 
-        routingDsl.POST("/transformarray").routeTo(FunctionUtil.function0((() -> {
+        routingDsl.POST("/transformarray").routingTo(req -> {
             try {
-                BatchImageRecord batch = objectMapper.readValue(getJsonText(), BatchImageRecord.class);
+                BatchImageRecord batch = objectMapper.readValue(getJsonText(req), BatchImageRecord.class);
                 if (batch == null)
                     return badRequest();
                 return ok(objectMapper.writeValueAsString(transformArray(batch))).as(contentType);
@@ -137,22 +145,22 @@ public class ImageSparkTransformServer extends SparkTransformServer {
                 e.printStackTrace();
                 return internalServerError();
             }
-        })));
+        });
 
-        routingDsl.POST("/transformimage").routeTo(FunctionUtil.function0((() -> {
+        routingDsl.POST("/transformimage").routingTo(req -> {
             try {
-                Http.MultipartFormData body = request().body().asMultipartFormData();
-                List<Http.MultipartFormData.FilePart> files = body.getFiles();
+                Http.MultipartFormData<Files.TemporaryFile> body = req.body().asMultipartFormData();
+                List<Http.MultipartFormData.FilePart<Files.TemporaryFile>> files = body.getFiles();
                 if (files.size() == 0) {
                     return badRequest();
                 }
 
                 List<SingleImageRecord> records = new ArrayList<>();
 
-                for (Http.MultipartFormData.FilePart filePart : files) {
-                    File file = filePart.getFile();
+                for (Http.MultipartFormData.FilePart<Files.TemporaryFile> filePart : files) {
+                    Files.TemporaryFile file = filePart.getRef();
                     if (file != null) {
-                        SingleImageRecord record = new SingleImageRecord(file.toURI());
+                        SingleImageRecord record = new SingleImageRecord(file.path().toUri());
                         records.add(record);
                     }
                 }
@@ -164,9 +172,9 @@ public class ImageSparkTransformServer extends SparkTransformServer {
                 e.printStackTrace();
                 return internalServerError();
             }
-        })));
+        });
 
-        server = Server.forRouter(routingDsl.build(), Mode.PROD, port);
+        return routingDsl.build();
     }
 
     @Override
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/SparkTransformServer.java b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/SparkTransformServer.java
index 2d4c92836..411872006 100644
--- a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/SparkTransformServer.java
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/java/org/datavec/spark/transform/SparkTransformServer.java
@@ -22,6 +22,7 @@ import org.datavec.spark.transform.model.Base64NDArrayBody;
 import org.datavec.spark.transform.model.BatchCSVRecord;
 import org.datavec.spark.transform.service.DataVecTransformService;
 import org.nd4j.shade.jackson.databind.ObjectMapper;
+import play.mvc.Http;
 import play.server.Server;
 
 import static play.mvc.Controller.request;
@@ -50,25 +51,17 @@ public abstract class SparkTransformServer implements DataVecTransformService {
             server.stop();
     }
 
-    protected boolean isSequence() {
-        return request().hasHeader(SEQUENCE_OR_NOT_HEADER)
-                && request().getHeader(SEQUENCE_OR_NOT_HEADER).toUpperCase()
-                .equals("TRUE");
+    protected boolean isSequence(Http.Request request) {
+        return request.hasHeader(SEQUENCE_OR_NOT_HEADER)
+                && request.header(SEQUENCE_OR_NOT_HEADER).get().equalsIgnoreCase("true");
     }
 
-
-    protected String getHeaderValue(String value) {
-        if (request().hasHeader(value))
-            return request().getHeader(value);
-        return null;
-    }
-
-    protected String getJsonText() {
-        JsonNode tryJson = request().body().asJson();
+    protected String getJsonText(Http.Request request) {
+        JsonNode tryJson = request.body().asJson();
         if (tryJson != null)
             return tryJson.toString();
         else
-            return request().body().asText();
+            return request.body().asText();
     }
 
     public abstract Base64NDArrayBody transformSequenceArrayIncremental(BatchCSVRecord singleCsvRecord);
diff --git a/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/resources/application.conf b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/resources/application.conf
new file mode 100644
index 000000000..28a4aa208
--- /dev/null
+++ b/datavec/datavec-spark-inference-parent/datavec-spark-inference-server/src/main/resources/application.conf
@@ -0,0 +1,350 @@
+# This is the main configuration file for the application.
+# https://www.playframework.com/documentation/latest/ConfigFile
+# ~~~~~
+# Play uses HOCON as its configuration file format.  HOCON has a number
+# of advantages over other config formats, but there are two things that
+# can be used when modifying settings.
+#
+# You can include other configuration files in this main application.conf file:
+#include "extra-config.conf"
+#
+# You can declare variables and substitute for them:
+#mykey = ${some.value}
+#
+# And if an environment variable exists when there is no other subsitution, then
+# HOCON will fall back to substituting environment variable:
+#mykey = ${JAVA_HOME}
+
+## Akka
+# https://www.playframework.com/documentation/latest/ScalaAkka#Configuration
+# https://www.playframework.com/documentation/latest/JavaAkka#Configuration
+# ~~~~~
+# Play uses Akka internally and exposes Akka Streams and actors in Websockets and
+# other streaming HTTP responses.
+akka {
+  # "akka.log-config-on-start" is extraordinarly useful because it log the complete
+  # configuration at INFO level, including defaults and overrides, so it s worth
+  # putting at the very top.
+  #
+  # Put the following in your conf/logback.xml file:
+  #
+  # <logger name="akka.actor" level="INFO" />
+  #
+  # And then uncomment this line to debug the configuration.
+  #
+  #log-config-on-start = true
+}
+
+## Modules
+# https://www.playframework.com/documentation/latest/Modules
+# ~~~~~
+# Control which modules are loaded when Play starts.  Note that modules are
+# the replacement for "GlobalSettings", which are deprecated in 2.5.x.
+# Please see https://www.playframework.com/documentation/latest/GlobalSettings
+# for more information.
+#
+# You can also extend Play functionality by using one of the publically available
+# Play modules: https://playframework.com/documentation/latest/ModuleDirectory
+play.modules {
+  # By default, Play will load any class called Module that is defined
+  # in the root package (the "app" directory), or you can define them
+  # explicitly below.
+  # If there are any built-in modules that you want to disable, you can list them here.
+  #enabled += my.application.Module
+
+  # If there are any built-in modules that you want to disable, you can list them here.
+  #disabled += ""
+}
+
+## Internationalisation
+# https://www.playframework.com/documentation/latest/JavaI18N
+# https://www.playframework.com/documentation/latest/ScalaI18N
+# ~~~~~
+# Play comes with its own i18n settings, which allow the user's preferred language
+# to map through to internal messages, or allow the language to be stored in a cookie.
+play.i18n {
+  # The application languages
+  langs = [ "en" ]
+
+  # Whether the language cookie should be secure or not
+  #langCookieSecure = true
+
+  # Whether the HTTP only attribute of the cookie should be set to true
+  #langCookieHttpOnly = true
+}
+
+## Play HTTP settings
+# ~~~~~
+play.http {
+  ## Router
+  # https://www.playframework.com/documentation/latest/JavaRouting
+  # https://www.playframework.com/documentation/latest/ScalaRouting
+  # ~~~~~
+  # Define the Router object to use for this application.
+  # This router will be looked up first when the application is starting up,
+  # so make sure this is the entry point.
+  # Furthermore, it's assumed your route file is named properly.
+  # So for an application router like `my.application.Router`,
+  # you may need to define a router file `conf/my.application.routes`.
+  # Default to Routes in the root package (aka "apps" folder) (and conf/routes)
+  #router = my.application.Router
+
+  ## Action Creator
+  # https://www.playframework.com/documentation/latest/JavaActionCreator
+  # ~~~~~
+  #actionCreator = null
+
+  ## ErrorHandler
+  # https://www.playframework.com/documentation/latest/JavaRouting
+  # https://www.playframework.com/documentation/latest/ScalaRouting
+  # ~~~~~
+  # If null, will attempt to load a class called ErrorHandler in the root package,
+  #errorHandler = null
+
+  ## Filters
+  # https://www.playframework.com/documentation/latest/ScalaHttpFilters
+  # https://www.playframework.com/documentation/latest/JavaHttpFilters
+  # ~~~~~
+  # Filters run code on every request. They can be used to perform
+  # common logic for all your actions, e.g. adding common headers.
+  # Defaults to "Filters" in the root package (aka "apps" folder)
+  # Alternatively you can explicitly register a class here.
+  #filters += my.application.Filters
+
+  ## Session & Flash
+  # https://www.playframework.com/documentation/latest/JavaSessionFlash
+  # https://www.playframework.com/documentation/latest/ScalaSessionFlash
+  # ~~~~~
+  session {
+    # Sets the cookie to be sent only over HTTPS.
+    #secure = true
+
+    # Sets the cookie to be accessed only by the server.
+    #httpOnly = true
+
+    # Sets the max-age field of the cookie to 5 minutes.
+    # NOTE: this only sets when the browser will discard the cookie. Play will consider any
+    # cookie value with a valid signature to be a valid session forever. To implement a server side session timeout,
+    # you need to put a timestamp in the session and check it at regular intervals to possibly expire it.
+    #maxAge = 300
+
+    # Sets the domain on the session cookie.
+    #domain = "example.com"
+  }
+
+  flash {
+    # Sets the cookie to be sent only over HTTPS.
+    #secure = true
+
+    # Sets the cookie to be accessed only by the server.
+    #httpOnly = true
+  }
+}
+
+## Netty Provider
+# https://www.playframework.com/documentation/latest/SettingsNetty
+# ~~~~~
+play.server.netty {
+  # Whether the Netty wire should be logged
+  #log.wire = true
+
+  # If you run Play on Linux, you can use Netty's native socket transport
+  # for higher performance with less garbage.
+  #transport = "native"
+}
+
+## WS (HTTP Client)
+# https://www.playframework.com/documentation/latest/ScalaWS#Configuring-WS
+# ~~~~~
+# The HTTP client primarily used for REST APIs.  The default client can be
+# configured directly, but you can also create different client instances
+# with customized settings. You must enable this by adding to build.sbt:
+#
+# libraryDependencies += ws // or javaWs if using java
+#
+play.ws {
+  # Sets HTTP requests not to follow 302 requests
+  #followRedirects = false
+
+  # Sets the maximum number of open HTTP connections for the client.
+  #ahc.maxConnectionsTotal = 50
+
+  ## WS SSL
+  # https://www.playframework.com/documentation/latest/WsSSL
+  # ~~~~~
+  ssl {
+    # Configuring HTTPS with Play WS does not require programming.  You can
+    # set up both trustManager and keyManager for mutual authentication, and
+    # turn on JSSE debugging in development with a reload.
+    #debug.handshake = true
+    #trustManager = {
+    #  stores = [
+    #    { type = "JKS", path = "exampletrust.jks" }
+    #  ]
+    #}
+  }
+}
+
+## Cache
+# https://www.playframework.com/documentation/latest/JavaCache
+# https://www.playframework.com/documentation/latest/ScalaCache
+# ~~~~~
+# Play comes with an integrated cache API that can reduce the operational
+# overhead of repeated requests. You must enable this by adding to build.sbt:
+#
+# libraryDependencies += cache
+#
+play.cache {
+  # If you want to bind several caches, you can bind the individually
+  #bindCaches = ["db-cache", "user-cache", "session-cache"]
+}
+
+## Filters
+# https://www.playframework.com/documentation/latest/Filters
+# ~~~~~
+# There are a number of built-in filters that can be enabled and configured
+# to give Play greater security.  You must enable this by adding to build.sbt:
+#
+# libraryDependencies += filters
+#
+play.filters {
+  ## CORS filter configuration
+  # https://www.playframework.com/documentation/latest/CorsFilter
+  # ~~~~~
+  # CORS is a protocol that allows web applications to make requests from the browser
+  # across different domains.
+  # NOTE: You MUST apply the CORS configuration before the CSRF filter, as CSRF has
+  # dependencies on CORS settings.
+  cors {
+    # Filter paths by a whitelist of path prefixes
+    #pathPrefixes = ["/some/path", ...]
+
+    # The allowed origins. If null, all origins are allowed.
+    #allowedOrigins = ["http://www.example.com"]
+
+    # The allowed HTTP methods. If null, all methods are allowed
+    #allowedHttpMethods = ["GET", "POST"]
+  }
+
+  ## CSRF Filter
+  # https://www.playframework.com/documentation/latest/ScalaCsrf#Applying-a-global-CSRF-filter
+  # https://www.playframework.com/documentation/latest/JavaCsrf#Applying-a-global-CSRF-filter
+  # ~~~~~
+  # Play supports multiple methods for verifying that a request is not a CSRF request.
+  # The primary mechanism is a CSRF token. This token gets placed either in the query string
+  # or body of every form submitted, and also gets placed in the users session.
+  # Play then verifies that both tokens are present and match.
+  csrf {
+    # Sets the cookie to be sent only over HTTPS
+    #cookie.secure = true
+
+    # Defaults to CSRFErrorHandler in the root package.
+    #errorHandler = MyCSRFErrorHandler
+  }
+
+  ## Security headers filter configuration
+  # https://www.playframework.com/documentation/latest/SecurityHeaders
+  # ~~~~~
+  # Defines security headers that prevent XSS attacks.
+  # If enabled, then all options are set to the below configuration by default:
+  headers {
+    # The X-Frame-Options header. If null, the header is not set.
+    #frameOptions = "DENY"
+
+    # The X-XSS-Protection header. If null, the header is not set.
+    #xssProtection = "1; mode=block"
+
+    # The X-Content-Type-Options header. If null, the header is not set.
+    #contentTypeOptions = "nosniff"
+
+    # The X-Permitted-Cross-Domain-Policies header. If null, the header is not set.
+    #permittedCrossDomainPolicies = "master-only"
+
+    # The Content-Security-Policy header. If null, the header is not set.
+    #contentSecurityPolicy = "default-src 'self'"
+  }
+
+  ## Allowed hosts filter configuration
+  # https://www.playframework.com/documentation/latest/AllowedHostsFilter
+  # ~~~~~
+  # Play provides a filter that lets you configure which hosts can access your application.
+  # This is useful to prevent cache poisoning attacks.
+  hosts {
+    # Allow requests to example.com, its subdomains, and localhost:9000.
+    #allowed = [".example.com", "localhost:9000"]
+  }
+}
+
+## Evolutions
+# https://www.playframework.com/documentation/latest/Evolutions
+# ~~~~~
+# Evolutions allows database scripts to be automatically run on startup in dev mode
+# for database migrations. You must enable this by adding to build.sbt:
+#
+# libraryDependencies += evolutions
+#
+play.evolutions {
+  # You can disable evolutions for a specific datasource if necessary
+  #db.default.enabled = false
+}
+
+## Database Connection Pool
+# https://www.playframework.com/documentation/latest/SettingsJDBC
+# ~~~~~
+# Play doesn't require a JDBC database to run, but you can easily enable one.
+#
+# libraryDependencies += jdbc
+#
+play.db {
+  # The combination of these two settings results in "db.default" as the
+  # default JDBC pool:
+  #config = "db"
+  #default = "default"
+
+  # Play uses HikariCP as the default connection pool.  You can override
+  # settings by changing the prototype:
+  prototype {
+    # Sets a fixed JDBC connection pool size of 50
+    #hikaricp.minimumIdle = 50
+    #hikaricp.maximumPoolSize = 50
+  }
+}
+
+## JDBC Datasource
+# https://www.playframework.com/documentation/latest/JavaDatabase
+# https://www.playframework.com/documentation/latest/ScalaDatabase
+# ~~~~~
+# Once JDBC datasource is set up, you can work with several different
+# database options:
+#
+# Slick (Scala preferred option): https://www.playframework.com/documentation/latest/PlaySlick
+# JPA (Java preferred option): https://playframework.com/documentation/latest/JavaJPA
+# EBean: https://playframework.com/documentation/latest/JavaEbean
+# Anorm: https://www.playframework.com/documentation/latest/ScalaAnorm
+#
+db {
+  # You can declare as many datasources as you want.
+  # By convention, the default datasource is named `default`
+
+  # https://www.playframework.com/documentation/latest/Developing-with-the-H2-Database
+  default.driver = org.h2.Driver
+  default.url = "jdbc:h2:mem:play"
+  #default.username = sa
+  #default.password = ""
+
+  # You can expose this datasource via JNDI if needed (Useful for JPA)
+  default.jndiName=DefaultDS
+
+  # You can turn on SQL logging for any datasource
+  # https://www.playframework.com/documentation/latest/Highlights25#Logging-SQL-statements
+  #default.logSql=true
+}
+
+jpa.default=defaultPersistenceUnit
+
+
+#Increase default maximum post length - used for remote listener functionality
+#Can get response 413 with larger networks without setting this
+# parsers.text.maxLength is deprecated, use play.http.parser.maxMemoryBuffer instead
+#parsers.text.maxLength=10M
+play.http.parser.maxMemoryBuffer=10M
diff --git a/datavec/datavec-spark/pom.xml b/datavec/datavec-spark/pom.xml
index d98730407..05c505cac 100644
--- a/datavec/datavec-spark/pom.xml
+++ b/datavec/datavec-spark/pom.xml
@@ -28,61 +28,11 @@
     <artifactId>datavec-spark_2.11</artifactId>
 
     <properties>
-        <!-- These Spark versions have to be here, and NOT in the datavec parent pom. Otherwise, the properties
-         will be resolved to their defaults. Whereas when defined here (the version outputStream spark 1 vs. 2 specific) we can
-         have different version properties simultaneously (in different pom files) instead of a single global property -->
-        <spark.version>2.1.0</spark.version>
-        <spark.major.version>2</spark.major.version>
-
         <!-- Default scala versions, may be overwritten by build profiles -->
         <scala.version>2.11.12</scala.version>
         <scala.binary.version>2.11</scala.binary.version>
     </properties>
 
-    <build>
-        <plugins>
-            <!-- added source folder containing the code specific to the spark version -->
-            <plugin>
-                <groupId>org.codehaus.mojo</groupId>
-                <artifactId>build-helper-maven-plugin</artifactId>
-                <executions>
-                    <execution>
-                        <id>add-source</id>
-                        <phase>generate-sources</phase>
-                        <goals>
-                            <goal>add-source</goal>
-                        </goals>
-                        <configuration>
-                            <sources>
-                                <source>src/main/spark-${spark.major.version}</source>
-                            </sources>
-                        </configuration>
-                    </execution>
-                </executions>
-            </plugin>
-        </plugins>
-    </build>
-
-    <dependencyManagement>
-        <dependencies>
-            <dependency>
-                <groupId>com.fasterxml.jackson.datatype</groupId>
-                <artifactId>jackson-datatype-jsr310</artifactId>
-                <version>${jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.dataformat</groupId>
-                <artifactId>jackson-dataformat-yaml</artifactId>
-                <version>${jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.module</groupId>
-                <artifactId>jackson-module-scala_2.11</artifactId>
-                <version>${jackson.version}</version>
-            </dependency>
-        </dependencies>
-    </dependencyManagement>
-
     <dependencies>
         <dependency>
             <groupId>org.scala-lang</groupId>
@@ -95,42 +45,13 @@
             <version>${scala.version}</version>
         </dependency>
 
-        <dependency>
-            <groupId>org.codehaus.jackson</groupId>
-            <artifactId>jackson-core-asl</artifactId>
-            <version>${jackson-asl.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>org.codehaus.jackson</groupId>
-            <artifactId>jackson-mapper-asl</artifactId>
-            <version>${jackson-asl.version}</version>
-        </dependency>
         <dependency>
             <groupId>org.apache.spark</groupId>
             <artifactId>spark-sql_2.11</artifactId>
             <version>${spark.version}</version>
+            <scope>provided</scope>
         </dependency>
 
-        <dependency>
-            <groupId>com.google.guava</groupId>
-            <artifactId>guava</artifactId>
-            <version>${guava.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.google.inject</groupId>
-            <artifactId>guice</artifactId>
-            <version>${guice.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.google.protobuf</groupId>
-            <artifactId>protobuf-java</artifactId>
-            <version>${google.protobuf.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>commons-codec</groupId>
-            <artifactId>commons-codec</artifactId>
-            <version>${commons-codec.version}</version>
-        </dependency>
         <dependency>
             <groupId>commons-collections</groupId>
             <artifactId>commons-collections</artifactId>
@@ -141,96 +62,16 @@
             <artifactId>commons-io</artifactId>
             <version>${commons-io.version}</version>
         </dependency>
-        <dependency>
-            <groupId>commons-lang</groupId>
-            <artifactId>commons-lang</artifactId>
-            <version>${commons-lang.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>commons-net</groupId>
-            <artifactId>commons-net</artifactId>
-            <version>${commons-net.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.sun.xml.bind</groupId>
-            <artifactId>jaxb-core</artifactId>
-            <version>${jaxb.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.sun.xml.bind</groupId>
-            <artifactId>jaxb-impl</artifactId>
-            <version>${jaxb.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-actor_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-remote_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-slf4j_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>io.netty</groupId>
-            <artifactId>netty</artifactId>
-            <version>${netty.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-core</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-databind</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-annotations</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>javax.servlet</groupId>
-            <artifactId>javax.servlet-api</artifactId>
-            <version>${servlet.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>org.apache.commons</groupId>
-            <artifactId>commons-compress</artifactId>
-            <version>${commons-compress.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>org.apache.commons</groupId>
-            <artifactId>commons-lang3</artifactId>
-            <version>${commons-lang3.version}</version>
-        </dependency>
         <dependency>
             <groupId>org.apache.commons</groupId>
             <artifactId>commons-math3</artifactId>
             <version>${commons-math3.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.apache.curator</groupId>
-            <artifactId>curator-recipes</artifactId>
-            <version>${curator.version}</version>
-        </dependency>
         <dependency>
             <groupId>org.slf4j</groupId>
             <artifactId>slf4j-api</artifactId>
             <version>${slf4j.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.typesafe</groupId>
-            <artifactId>config</artifactId>
-            <version>${typesafe.config.version}</version>
-        </dependency>
         <dependency>
             <groupId>org.apache.spark</groupId>
             <artifactId>spark-core_2.11</artifactId>
@@ -241,14 +82,6 @@
                 <groupId>com.google.code.findbugs</groupId>
                 <artifactId>jsr305</artifactId>
               </exclusion>
-              <exclusion>
-                <groupId>org.slf4j</groupId>
-                <artifactId>slf4j-log4j12</artifactId>
-              </exclusion>
-              <exclusion>
-                 <groupId>log4j</groupId>
-                 <artifactId>log4j</artifactId>
-              </exclusion>
             </exclusions>
         </dependency>
 
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/functions/FlatMapFunctionAdapter.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/functions/FlatMapFunctionAdapter.java
deleted file mode 100644
index 8aeae58a5..000000000
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/functions/FlatMapFunctionAdapter.java
+++ /dev/null
@@ -1,29 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.functions;
-
-import java.io.Serializable;
-
-/**
- *
- * A function that returns zero or more output records from each input record.
- *
- * Adapter for Spark interface in order to freeze interface changes between spark versions
- */
-public interface FlatMapFunctionAdapter<T, R> extends Serializable {
-    Iterable<R> call(T t) throws Exception;
-}
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/DataFrames.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/DataFrames.java
index fbe8a63d4..5d0bff7f3 100644
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/DataFrames.java
+++ b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/DataFrames.java
@@ -21,10 +21,7 @@ import org.apache.spark.api.java.JavaRDD;
 import org.apache.spark.api.java.JavaSparkContext;
 import org.apache.spark.api.java.function.Function;
 import org.apache.spark.api.java.function.Function2;
-import org.apache.spark.sql.Column;
-import org.apache.spark.sql.Row;
-import org.apache.spark.sql.SQLContext;
-import org.apache.spark.sql.functions;
+import org.apache.spark.sql.*;
 import org.apache.spark.sql.types.DataTypes;
 import org.apache.spark.sql.types.Metadata;
 import org.apache.spark.sql.types.StructField;
@@ -46,7 +43,6 @@ import java.util.List;
 
 import static org.apache.spark.sql.functions.avg;
 import static org.apache.spark.sql.functions.col;
-import static org.datavec.spark.transform.DataRowsFacade.dataRows;
 
 
 /**
@@ -71,7 +67,7 @@ public class DataFrames {
      *                   deviation for
      * @return the column that represents the standard deviation
      */
-    public static Column std(DataRowsFacade dataFrame, String columnName) {
+    public static Column std(Dataset<Row> dataFrame, String columnName) {
         return functions.sqrt(var(dataFrame, columnName));
     }
 
@@ -85,8 +81,8 @@ public class DataFrames {
      *                   deviation for
      * @return the column that represents the standard deviation
      */
-    public static Column var(DataRowsFacade dataFrame, String columnName) {
-        return dataFrame.get().groupBy(columnName).agg(functions.variance(columnName)).col(columnName);
+    public static Column var(Dataset<Row> dataFrame, String columnName) {
+        return dataFrame.groupBy(columnName).agg(functions.variance(columnName)).col(columnName);
     }
 
     /**
@@ -97,8 +93,8 @@ public class DataFrames {
      * @param columnName the name of the column to get the min for
      * @return the column that represents the min
      */
-    public static Column min(DataRowsFacade dataFrame, String columnName) {
-        return dataFrame.get().groupBy(columnName).agg(functions.min(columnName)).col(columnName);
+    public static Column min(Dataset<Row> dataFrame, String columnName) {
+        return dataFrame.groupBy(columnName).agg(functions.min(columnName)).col(columnName);
     }
 
     /**
@@ -110,8 +106,8 @@ public class DataFrames {
      *                   to get the max for
      * @return the column that represents the max
      */
-    public static Column max(DataRowsFacade dataFrame, String columnName) {
-        return dataFrame.get().groupBy(columnName).agg(functions.max(columnName)).col(columnName);
+    public static Column max(Dataset<Row> dataFrame, String columnName) {
+        return dataFrame.groupBy(columnName).agg(functions.max(columnName)).col(columnName);
     }
 
     /**
@@ -122,8 +118,8 @@ public class DataFrames {
      * @param columnName the name of the column to get the mean for
      * @return the column that represents the mean
      */
-    public static Column mean(DataRowsFacade dataFrame, String columnName) {
-        return dataFrame.get().groupBy(columnName).agg(avg(columnName)).col(columnName);
+    public static Column mean(Dataset<Row> dataFrame, String columnName) {
+        return dataFrame.groupBy(columnName).agg(avg(columnName)).col(columnName);
     }
 
     /**
@@ -166,7 +162,7 @@ public class DataFrames {
      * - Column 1: Sequence index (name: {@link #SEQUENCE_INDEX_COLUMN} - an index (integer, starting at 0) for the position
      * of this record in the original time series.<br>
      * These two columns are required if the data is to be converted back into a sequence at a later point, for example
-     * using {@link #toRecordsSequence(DataRowsFacade)}
+     * using {@link #toRecordsSequence(Dataset<Row>)}
      *
      * @param schema Schema to convert
      * @return StructType for the schema
@@ -250,9 +246,9 @@ public class DataFrames {
      * @param dataFrame the dataframe to convert
      * @return the converted schema and rdd of writables
      */
-    public static Pair<Schema, JavaRDD<List<Writable>>> toRecords(DataRowsFacade dataFrame) {
-        Schema schema = fromStructType(dataFrame.get().schema());
-        return new Pair<>(schema, dataFrame.get().javaRDD().map(new ToRecord(schema)));
+    public static Pair<Schema, JavaRDD<List<Writable>>> toRecords(Dataset<Row> dataFrame) {
+        Schema schema = fromStructType(dataFrame.schema());
+        return new Pair<>(schema, dataFrame.javaRDD().map(new ToRecord(schema)));
     }
 
     /**
@@ -267,11 +263,11 @@ public class DataFrames {
      * @param dataFrame Data frame to convert
      * @return Data in sequence (i.e., {@code List<List<Writable>>} form
      */
-    public static Pair<Schema, JavaRDD<List<List<Writable>>>> toRecordsSequence(DataRowsFacade dataFrame) {
+    public static Pair<Schema, JavaRDD<List<List<Writable>>>> toRecordsSequence(Dataset<Row> dataFrame) {
 
         //Need to convert from flattened to sequence data...
         //First: Group by the Sequence UUID (first column)
-        JavaPairRDD<String, Iterable<Row>> grouped = dataFrame.get().javaRDD().groupBy(new Function<Row, String>() {
+        JavaPairRDD<String, Iterable<Row>> grouped = dataFrame.javaRDD().groupBy(new Function<Row, String>() {
             @Override
             public String call(Row row) throws Exception {
                 return row.getString(0);
@@ -279,7 +275,7 @@ public class DataFrames {
         });
 
 
-        Schema schema = fromStructType(dataFrame.get().schema());
+        Schema schema = fromStructType(dataFrame.schema());
 
         //Group by sequence UUID, and sort each row within the sequences using the time step index
         Function<Iterable<Row>, List<List<Writable>>> createCombiner = new DataFrameToSequenceCreateCombiner(schema); //Function to create the initial combiner
@@ -318,11 +314,11 @@ public class DataFrames {
      * @param data   the data to convert
      * @return the dataframe object
      */
-    public static DataRowsFacade toDataFrame(Schema schema, JavaRDD<List<Writable>> data) {
+    public static Dataset<Row> toDataFrame(Schema schema, JavaRDD<List<Writable>> data) {
         JavaSparkContext sc = new JavaSparkContext(data.context());
         SQLContext sqlContext = new SQLContext(sc);
         JavaRDD<Row> rows = data.map(new ToRow(schema));
-        return dataRows(sqlContext.createDataFrame(rows, fromSchema(schema)));
+        return sqlContext.createDataFrame(rows, fromSchema(schema));
     }
 
 
@@ -333,18 +329,18 @@ public class DataFrames {
      * - Column 1: Sequence index (name: {@link #SEQUENCE_INDEX_COLUMN} - an index (integer, starting at 0) for the position
      * of this record in the original time series.<br>
      * These two columns are required if the data is to be converted back into a sequence at a later point, for example
-     * using {@link #toRecordsSequence(DataRowsFacade)}
+     * using {@link #toRecordsSequence(Dataset<Row>)}
      *
      * @param schema Schema for the data
      * @param data   Sequence data to convert to a DataFrame
      * @return The dataframe object
      */
-    public static DataRowsFacade toDataFrameSequence(Schema schema, JavaRDD<List<List<Writable>>> data) {
+    public static Dataset<Row> toDataFrameSequence(Schema schema, JavaRDD<List<List<Writable>>> data) {
         JavaSparkContext sc = new JavaSparkContext(data.context());
 
         SQLContext sqlContext = new SQLContext(sc);
         JavaRDD<Row> rows = data.flatMap(new SequenceToRows(schema));
-        return dataRows(sqlContext.createDataFrame(rows, fromSchemaSequence(schema)));
+        return sqlContext.createDataFrame(rows, fromSchemaSequence(schema));
     }
 
     /**
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/Normalization.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/Normalization.java
index 68efd1888..cacea101d 100644
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/Normalization.java
+++ b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/Normalization.java
@@ -19,14 +19,13 @@ package org.datavec.spark.transform;
 import org.apache.commons.collections.map.ListOrderedMap;
 import org.apache.spark.api.java.JavaRDD;
 import org.apache.spark.sql.Column;
+import org.apache.spark.sql.Dataset;
 import org.apache.spark.sql.Row;
 import org.datavec.api.transform.schema.Schema;
 import org.datavec.api.writable.Writable;
 
 import java.util.*;
 
-import static org.datavec.spark.transform.DataRowsFacade.dataRows;
-
 
 /**
  * Simple dataframe based normalization.
@@ -46,7 +45,7 @@ public class Normalization {
      * @return a zero mean unit variance centered
      * rdd
      */
-    public static DataRowsFacade zeromeanUnitVariance(DataRowsFacade frame) {
+    public static Dataset<Row> zeromeanUnitVariance(Dataset<Row> frame) {
         return zeromeanUnitVariance(frame, Collections.<String>emptyList());
     }
 
@@ -71,7 +70,7 @@ public class Normalization {
      * @param max       the maximum value
      * @return the normalized dataframe per column
      */
-    public static DataRowsFacade normalize(DataRowsFacade dataFrame, double min, double max) {
+    public static Dataset<Row> normalize(Dataset<Row> dataFrame, double min, double max) {
         return normalize(dataFrame, min, max, Collections.<String>emptyList());
     }
 
@@ -86,7 +85,7 @@ public class Normalization {
      */
     public static JavaRDD<List<Writable>> normalize(Schema schema, JavaRDD<List<Writable>> data, double min,
                     double max) {
-        DataRowsFacade frame = DataFrames.toDataFrame(schema, data);
+        Dataset<Row> frame = DataFrames.toDataFrame(schema, data);
         return DataFrames.toRecords(normalize(frame, min, max, Collections.<String>emptyList())).getSecond();
     }
 
@@ -97,7 +96,7 @@ public class Normalization {
      * @param dataFrame the dataframe to scale
      * @return the normalized dataframe per column
      */
-    public static DataRowsFacade normalize(DataRowsFacade dataFrame) {
+    public static Dataset<Row> normalize(Dataset<Row> dataFrame) {
         return normalize(dataFrame, 0, 1, Collections.<String>emptyList());
     }
 
@@ -120,8 +119,8 @@ public class Normalization {
      * @return a zero mean unit variance centered
      * rdd
      */
-    public static DataRowsFacade zeromeanUnitVariance(DataRowsFacade frame, List<String> skipColumns) {
-        List<String> columnsList = DataFrames.toList(frame.get().columns());
+    public static Dataset<Row> zeromeanUnitVariance(Dataset<Row> frame, List<String> skipColumns) {
+        List<String> columnsList = DataFrames.toList(frame.columns());
         columnsList.removeAll(skipColumns);
         String[] columnNames = DataFrames.toArray(columnsList);
         //first row is std second row is mean, each column in a row is for a particular column
@@ -133,7 +132,7 @@ public class Normalization {
             if (std == 0.0)
                 std = 1; //All same value -> (x-x)/1 = 0
 
-            frame = dataRows(frame.get().withColumn(columnName, frame.get().col(columnName).minus(mean).divide(std)));
+            frame = frame.withColumn(columnName, frame.col(columnName).minus(mean).divide(std));
         }
 
 
@@ -152,7 +151,7 @@ public class Normalization {
      */
     public static JavaRDD<List<Writable>> zeromeanUnitVariance(Schema schema, JavaRDD<List<Writable>> data,
                     List<String> skipColumns) {
-        DataRowsFacade frame = DataFrames.toDataFrame(schema, data);
+        Dataset<Row> frame = DataFrames.toDataFrame(schema, data);
         return DataFrames.toRecords(zeromeanUnitVariance(frame, skipColumns)).getSecond();
     }
 
@@ -178,7 +177,7 @@ public class Normalization {
      */
     public static JavaRDD<List<List<Writable>>> zeroMeanUnitVarianceSequence(Schema schema,
                     JavaRDD<List<List<Writable>>> sequence, List<String> excludeColumns) {
-        DataRowsFacade frame = DataFrames.toDataFrameSequence(schema, sequence);
+        Dataset<Row> frame = DataFrames.toDataFrameSequence(schema, sequence);
         if (excludeColumns == null)
             excludeColumns = Arrays.asList(DataFrames.SEQUENCE_UUID_COLUMN, DataFrames.SEQUENCE_INDEX_COLUMN);
         else {
@@ -196,7 +195,7 @@ public class Normalization {
      * @param columns the columns to get the
      * @return
      */
-    public static List<Row> minMaxColumns(DataRowsFacade data, List<String> columns) {
+    public static List<Row> minMaxColumns(Dataset<Row> data, List<String> columns) {
         String[] arr = new String[columns.size()];
         for (int i = 0; i < arr.length; i++)
             arr[i] = columns.get(i);
@@ -210,7 +209,7 @@ public class Normalization {
      * @param columns the columns to get the
      * @return
      */
-    public static List<Row> minMaxColumns(DataRowsFacade data, String... columns) {
+    public static List<Row> minMaxColumns(Dataset<Row> data, String... columns) {
         return aggregate(data, columns, new String[] {"min", "max"});
     }
 
@@ -221,7 +220,7 @@ public class Normalization {
      * @param columns the columns to get the
      * @return
      */
-    public static List<Row> stdDevMeanColumns(DataRowsFacade data, List<String> columns) {
+    public static List<Row> stdDevMeanColumns(Dataset<Row> data, List<String> columns) {
         String[] arr = new String[columns.size()];
         for (int i = 0; i < arr.length; i++)
             arr[i] = columns.get(i);
@@ -237,7 +236,7 @@ public class Normalization {
      * @param columns the columns to get the
      * @return
      */
-    public static List<Row> stdDevMeanColumns(DataRowsFacade data, String... columns) {
+    public static List<Row> stdDevMeanColumns(Dataset<Row> data, String... columns) {
         return aggregate(data, columns, new String[] {"stddev", "mean"});
     }
 
@@ -251,7 +250,7 @@ public class Normalization {
      * Each row will be a function with the desired columnar output
      * in the order in which the columns were specified.
      */
-    public static List<Row> aggregate(DataRowsFacade data, String[] columns, String[] functions) {
+    public static List<Row> aggregate(Dataset<Row> data, String[] columns, String[] functions) {
         String[] rest = new String[columns.length - 1];
         System.arraycopy(columns, 1, rest, 0, rest.length);
         List<Row> rows = new ArrayList<>();
@@ -262,8 +261,8 @@ public class Normalization {
             }
 
             //compute the aggregation based on the operation
-            DataRowsFacade aggregated = dataRows(data.get().agg(expressions));
-            String[] columns2 = aggregated.get().columns();
+            Dataset<Row> aggregated = data.agg(expressions);
+            String[] columns2 = aggregated.columns();
             //strip out the op name and parentheses from the columns
             Map<String, String> opReplace = new TreeMap<>();
             for (String s : columns2) {
@@ -278,20 +277,20 @@ public class Normalization {
 
 
             //get rid of the operation name in the column
-            DataRowsFacade rearranged = null;
+            Dataset<Row> rearranged = null;
             for (Map.Entry<String, String> entries : opReplace.entrySet()) {
                 //first column
                 if (rearranged == null) {
-                    rearranged = dataRows(aggregated.get().withColumnRenamed(entries.getKey(), entries.getValue()));
+                    rearranged = aggregated.withColumnRenamed(entries.getKey(), entries.getValue());
                 }
                 //rearranged is just a copy of aggregated at this point
                 else
-                    rearranged = dataRows(rearranged.get().withColumnRenamed(entries.getKey(), entries.getValue()));
+                    rearranged = rearranged.withColumnRenamed(entries.getKey(), entries.getValue());
             }
 
-            rearranged = dataRows(rearranged.get().select(DataFrames.toColumns(columns)));
+            rearranged = rearranged.select(DataFrames.toColumns(columns));
             //op
-            rows.addAll(rearranged.get().collectAsList());
+            rows.addAll(rearranged.collectAsList());
         }
 
 
@@ -307,8 +306,8 @@ public class Normalization {
      * @param max       the maximum value
      * @return the normalized dataframe per column
      */
-    public static DataRowsFacade normalize(DataRowsFacade dataFrame, double min, double max, List<String> skipColumns) {
-        List<String> columnsList = DataFrames.toList(dataFrame.get().columns());
+    public static Dataset<Row> normalize(Dataset<Row> dataFrame, double min, double max, List<String> skipColumns) {
+        List<String> columnsList = DataFrames.toList(dataFrame.columns());
         columnsList.removeAll(skipColumns);
         String[] columnNames = DataFrames.toArray(columnsList);
         //first row is min second row is max, each column in a row is for a particular column
@@ -321,8 +320,8 @@ public class Normalization {
             if (maxSubMin == 0)
                 maxSubMin = 1;
 
-            Column newCol = dataFrame.get().col(columnName).minus(dMin).divide(maxSubMin).multiply(max - min).plus(min);
-            dataFrame = dataRows(dataFrame.get().withColumn(columnName, newCol));
+            Column newCol = dataFrame.col(columnName).minus(dMin).divide(maxSubMin).multiply(max - min).plus(min);
+            dataFrame = dataFrame.withColumn(columnName, newCol);
         }
 
 
@@ -340,7 +339,7 @@ public class Normalization {
      */
     public static JavaRDD<List<Writable>> normalize(Schema schema, JavaRDD<List<Writable>> data, double min, double max,
                     List<String> skipColumns) {
-        DataRowsFacade frame = DataFrames.toDataFrame(schema, data);
+        Dataset<Row> frame = DataFrames.toDataFrame(schema, data);
         return DataFrames.toRecords(normalize(frame, min, max, skipColumns)).getSecond();
     }
 
@@ -387,7 +386,7 @@ public class Normalization {
             excludeColumns.add(DataFrames.SEQUENCE_UUID_COLUMN);
             excludeColumns.add(DataFrames.SEQUENCE_INDEX_COLUMN);
         }
-        DataRowsFacade frame = DataFrames.toDataFrameSequence(schema, data);
+        Dataset<Row> frame = DataFrames.toDataFrameSequence(schema, data);
         return DataFrames.toRecordsSequence(normalize(frame, min, max, excludeColumns)).getSecond();
     }
 
@@ -398,7 +397,7 @@ public class Normalization {
      * @param dataFrame the dataframe to scale
      * @return the normalized dataframe per column
      */
-    public static DataRowsFacade normalize(DataRowsFacade dataFrame, List<String> skipColumns) {
+    public static Dataset<Row> normalize(Dataset<Row> dataFrame, List<String> skipColumns) {
         return normalize(dataFrame, 0, 1, skipColumns);
     }
 
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/analysis/SequenceFlatMapFunction.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/analysis/SequenceFlatMapFunction.java
index 6b7ff203f..5052491bb 100644
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/analysis/SequenceFlatMapFunction.java
+++ b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/analysis/SequenceFlatMapFunction.java
@@ -16,9 +16,10 @@
 
 package org.datavec.spark.transform.analysis;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.datavec.api.writable.Writable;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 
+import java.util.Iterator;
 import java.util.List;
 
 /**
@@ -27,10 +28,11 @@ import java.util.List;
  *
  * @author Alex Black
  */
-public class SequenceFlatMapFunction extends BaseFlatMapFunctionAdaptee<List<List<Writable>>, List<Writable>> {
+public class SequenceFlatMapFunction implements FlatMapFunction<List<List<Writable>>, List<Writable>> {
 
-    public SequenceFlatMapFunction() {
-        super(new SequenceFlatMapFunctionAdapter());
+    @Override
+    public Iterator<List<Writable>> call(List<List<Writable>> collections) throws Exception {
+        return collections.iterator();
     }
 
 }
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/analysis/SequenceFlatMapFunctionAdapter.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/analysis/SequenceFlatMapFunctionAdapter.java
deleted file mode 100644
index 6b25fb826..000000000
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/analysis/SequenceFlatMapFunctionAdapter.java
+++ /dev/null
@@ -1,36 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform.analysis;
-
-import org.datavec.api.writable.Writable;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-
-import java.util.List;
-
-/**
- * SequenceFlatMapFunction: very simple function used to flatten a sequence
- * Typically used only internally for certain analysis operations
- *
- * @author Alex Black
- */
-public class SequenceFlatMapFunctionAdapter implements FlatMapFunctionAdapter<List<List<Writable>>, List<Writable>> {
-    @Override
-    public Iterable<List<Writable>> call(List<List<Writable>> collections) throws Exception {
-        return collections;
-    }
-
-}
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/ExecuteJoinFromCoGroupFlatMapFunction.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/ExecuteJoinFromCoGroupFlatMapFunction.java
index 52bf924be..6e501f560 100644
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/ExecuteJoinFromCoGroupFlatMapFunction.java
+++ b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/ExecuteJoinFromCoGroupFlatMapFunction.java
@@ -16,11 +16,14 @@
 
 package org.datavec.spark.transform.join;
 
+import org.nd4j.shade.guava.collect.Iterables;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.datavec.api.transform.join.Join;
 import org.datavec.api.writable.Writable;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import scala.Tuple2;
 
+import java.util.ArrayList;
+import java.util.Iterator;
 import java.util.List;
 
 /**
@@ -28,10 +31,89 @@ import java.util.List;
  *
  * @author Alex Black
  */
-public class ExecuteJoinFromCoGroupFlatMapFunction extends
-                BaseFlatMapFunctionAdaptee<Tuple2<List<Writable>, Tuple2<Iterable<List<Writable>>, Iterable<List<Writable>>>>, List<Writable>> {
+public class ExecuteJoinFromCoGroupFlatMapFunction implements  FlatMapFunction<Tuple2<List<Writable>, Tuple2<Iterable<List<Writable>>, Iterable<List<Writable>>>>, List<Writable>> {
+
+    private final Join join;
 
     public ExecuteJoinFromCoGroupFlatMapFunction(Join join) {
-        super(new ExecuteJoinFromCoGroupFlatMapFunctionAdapter(join));
+        this.join = join;
+    }
+
+    @Override
+    public Iterator<List<Writable>> call(
+            Tuple2<List<Writable>, Tuple2<Iterable<List<Writable>>, Iterable<List<Writable>>>> t2)
+            throws Exception {
+
+        Iterable<List<Writable>> leftList = t2._2()._1();
+        Iterable<List<Writable>> rightList = t2._2()._2();
+
+        List<List<Writable>> ret = new ArrayList<>();
+        Join.JoinType jt = join.getJoinType();
+        switch (jt) {
+            case Inner:
+                //Return records where key columns appear in BOTH
+                //So if no values from left OR right: no return values
+                for (List<Writable> jvl : leftList) {
+                    for (List<Writable> jvr : rightList) {
+                        List<Writable> joined = join.joinExamples(jvl, jvr);
+                        ret.add(joined);
+                    }
+                }
+                break;
+            case LeftOuter:
+                //Return all records from left, even if no corresponding right value (NullWritable in that case)
+                for (List<Writable> jvl : leftList) {
+                    if (Iterables.size(rightList) == 0) {
+                        List<Writable> joined = join.joinExamples(jvl, null);
+                        ret.add(joined);
+                    } else {
+                        for (List<Writable> jvr : rightList) {
+                            List<Writable> joined = join.joinExamples(jvl, jvr);
+                            ret.add(joined);
+                        }
+                    }
+                }
+                break;
+            case RightOuter:
+                //Return all records from right, even if no corresponding left value (NullWritable in that case)
+                for (List<Writable> jvr : rightList) {
+                    if (Iterables.size(leftList) == 0) {
+                        List<Writable> joined = join.joinExamples(null, jvr);
+                        ret.add(joined);
+                    } else {
+                        for (List<Writable> jvl : leftList) {
+                            List<Writable> joined = join.joinExamples(jvl, jvr);
+                            ret.add(joined);
+                        }
+                    }
+                }
+                break;
+            case FullOuter:
+                //Return all records, even if no corresponding left/right value (NullWritable in that case)
+                if (Iterables.size(leftList) == 0) {
+                    //Only right values
+                    for (List<Writable> jvr : rightList) {
+                        List<Writable> joined = join.joinExamples(null, jvr);
+                        ret.add(joined);
+                    }
+                } else if (Iterables.size(rightList) == 0) {
+                    //Only left values
+                    for (List<Writable> jvl : leftList) {
+                        List<Writable> joined = join.joinExamples(jvl, null);
+                        ret.add(joined);
+                    }
+                } else {
+                    //Records from both left and right
+                    for (List<Writable> jvl : leftList) {
+                        for (List<Writable> jvr : rightList) {
+                            List<Writable> joined = join.joinExamples(jvl, jvr);
+                            ret.add(joined);
+                        }
+                    }
+                }
+                break;
+        }
+
+        return ret.iterator();
     }
 }
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/ExecuteJoinFromCoGroupFlatMapFunctionAdapter.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/ExecuteJoinFromCoGroupFlatMapFunctionAdapter.java
deleted file mode 100644
index dedff46d0..000000000
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/ExecuteJoinFromCoGroupFlatMapFunctionAdapter.java
+++ /dev/null
@@ -1,119 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform.join;
-
-import com.google.common.collect.Iterables;
-import org.datavec.api.transform.join.Join;
-import org.datavec.api.writable.Writable;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import scala.Tuple2;
-
-import java.util.ArrayList;
-import java.util.List;
-
-/**
- * Execute a join
- *
- * @author Alex Black
- */
-public class ExecuteJoinFromCoGroupFlatMapFunctionAdapter implements
-                FlatMapFunctionAdapter<Tuple2<List<Writable>, Tuple2<Iterable<List<Writable>>, Iterable<List<Writable>>>>, List<Writable>> {
-
-    private final Join join;
-
-    public ExecuteJoinFromCoGroupFlatMapFunctionAdapter(Join join) {
-        this.join = join;
-    }
-
-    @Override
-    public Iterable<List<Writable>> call(
-                    Tuple2<List<Writable>, Tuple2<Iterable<List<Writable>>, Iterable<List<Writable>>>> t2)
-                    throws Exception {
-
-        Iterable<List<Writable>> leftList = t2._2()._1();
-        Iterable<List<Writable>> rightList = t2._2()._2();
-
-        List<List<Writable>> ret = new ArrayList<>();
-        Join.JoinType jt = join.getJoinType();
-        switch (jt) {
-            case Inner:
-                //Return records where key columns appear in BOTH
-                //So if no values from left OR right: no return values
-                for (List<Writable> jvl : leftList) {
-                    for (List<Writable> jvr : rightList) {
-                        List<Writable> joined = join.joinExamples(jvl, jvr);
-                        ret.add(joined);
-                    }
-                }
-                break;
-            case LeftOuter:
-                //Return all records from left, even if no corresponding right value (NullWritable in that case)
-                for (List<Writable> jvl : leftList) {
-                    if (Iterables.size(rightList) == 0) {
-                        List<Writable> joined = join.joinExamples(jvl, null);
-                        ret.add(joined);
-                    } else {
-                        for (List<Writable> jvr : rightList) {
-                            List<Writable> joined = join.joinExamples(jvl, jvr);
-                            ret.add(joined);
-                        }
-                    }
-                }
-                break;
-            case RightOuter:
-                //Return all records from right, even if no corresponding left value (NullWritable in that case)
-                for (List<Writable> jvr : rightList) {
-                    if (Iterables.size(leftList) == 0) {
-                        List<Writable> joined = join.joinExamples(null, jvr);
-                        ret.add(joined);
-                    } else {
-                        for (List<Writable> jvl : leftList) {
-                            List<Writable> joined = join.joinExamples(jvl, jvr);
-                            ret.add(joined);
-                        }
-                    }
-                }
-                break;
-            case FullOuter:
-                //Return all records, even if no corresponding left/right value (NullWritable in that case)
-                if (Iterables.size(leftList) == 0) {
-                    //Only right values
-                    for (List<Writable> jvr : rightList) {
-                        List<Writable> joined = join.joinExamples(null, jvr);
-                        ret.add(joined);
-                    }
-                } else if (Iterables.size(rightList) == 0) {
-                    //Only left values
-                    for (List<Writable> jvl : leftList) {
-                        List<Writable> joined = join.joinExamples(jvl, null);
-                        ret.add(joined);
-                    }
-                } else {
-                    //Records from both left and right
-                    for (List<Writable> jvl : leftList) {
-                        for (List<Writable> jvr : rightList) {
-                            List<Writable> joined = join.joinExamples(jvl, jvr);
-                            ret.add(joined);
-                        }
-                    }
-                }
-                break;
-        }
-
-        return ret;
-    }
-}
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/FilterAndFlattenJoinedValues.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/FilterAndFlattenJoinedValues.java
index 6e206a657..d4ede4808 100644
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/FilterAndFlattenJoinedValues.java
+++ b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/FilterAndFlattenJoinedValues.java
@@ -16,10 +16,12 @@
 
 package org.datavec.spark.transform.join;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.datavec.api.transform.join.Join;
 import org.datavec.api.writable.Writable;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 
+import java.util.Collections;
+import java.util.Iterator;
 import java.util.List;
 
 /**
@@ -29,10 +31,43 @@ import java.util.List;
  *
  * @author Alex Black
  */
-public class FilterAndFlattenJoinedValues extends BaseFlatMapFunctionAdaptee<JoinedValue, List<Writable>> {
+public class FilterAndFlattenJoinedValues implements FlatMapFunction<JoinedValue, List<Writable>> {
+
+    private final Join.JoinType joinType;
 
     public FilterAndFlattenJoinedValues(Join.JoinType joinType) {
-        super(new FilterAndFlattenJoinedValuesAdapter(joinType));
+        this.joinType = joinType;
+    }
+
+    @Override
+    public Iterator<List<Writable>> call(JoinedValue joinedValue) throws Exception {
+        boolean keep;
+        switch (joinType) {
+            case Inner:
+                //Only keep joined values where we have both left and right
+                keep = joinedValue.isHaveLeft() && joinedValue.isHaveRight();
+                break;
+            case LeftOuter:
+                //Keep all values where left is not missing/null
+                keep = joinedValue.isHaveLeft();
+                break;
+            case RightOuter:
+                //Keep all values where right is not missing/null
+                keep = joinedValue.isHaveRight();
+                break;
+            case FullOuter:
+                //Keep all values
+                keep = true;
+                break;
+            default:
+                throw new RuntimeException("Unknown/not implemented join type: " + joinType);
+        }
+
+        if (keep) {
+            return Collections.singletonList(joinedValue.getValues()).iterator();
+        } else {
+            return Collections.emptyIterator();
+        }
     }
 
 }
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/FilterAndFlattenJoinedValuesAdapter.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/FilterAndFlattenJoinedValuesAdapter.java
deleted file mode 100644
index 3333276b1..000000000
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/join/FilterAndFlattenJoinedValuesAdapter.java
+++ /dev/null
@@ -1,71 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform.join;
-
-import org.datavec.api.transform.join.Join;
-import org.datavec.api.writable.Writable;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-
-import java.util.Collections;
-import java.util.List;
-
-/**
- * Doing two things here:
- * (a) filter out any unnecessary values, and
- * (b) extract the List<Writable> values from the JoinedValue
- *
- * @author Alex Black
- */
-public class FilterAndFlattenJoinedValuesAdapter implements FlatMapFunctionAdapter<JoinedValue, List<Writable>> {
-
-    private final Join.JoinType joinType;
-
-    public FilterAndFlattenJoinedValuesAdapter(Join.JoinType joinType) {
-        this.joinType = joinType;
-    }
-
-    @Override
-    public Iterable<List<Writable>> call(JoinedValue joinedValue) throws Exception {
-        boolean keep;
-        switch (joinType) {
-            case Inner:
-                //Only keep joined values where we have both left and right
-                keep = joinedValue.isHaveLeft() && joinedValue.isHaveRight();
-                break;
-            case LeftOuter:
-                //Keep all values where left is not missing/null
-                keep = joinedValue.isHaveLeft();
-                break;
-            case RightOuter:
-                //Keep all values where right is not missing/null
-                keep = joinedValue.isHaveRight();
-                break;
-            case FullOuter:
-                //Keep all values
-                keep = true;
-                break;
-            default:
-                throw new RuntimeException("Unknown/not implemented join type: " + joinType);
-        }
-
-        if (keep) {
-            return Collections.singletonList(joinedValue.getValues());
-        } else {
-            return Collections.emptyList();
-        }
-    }
-}
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/sparkfunction/SequenceToRows.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/sparkfunction/SequenceToRows.java
index 28c91c84b..639e43836 100644
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/sparkfunction/SequenceToRows.java
+++ b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/sparkfunction/SequenceToRows.java
@@ -16,21 +16,69 @@
 
 package org.datavec.spark.transform.sparkfunction;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.sql.Row;
+import org.apache.spark.sql.catalyst.expressions.GenericRowWithSchema;
+import org.apache.spark.sql.types.StructType;
 import org.datavec.api.transform.schema.Schema;
 import org.datavec.api.writable.Writable;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
+import org.datavec.spark.transform.DataFrames;
 
-import java.util.List;
+import java.util.*;
 
 /**
  * Convert a record to a row
  * @author Adam Gibson
  */
-public class SequenceToRows extends BaseFlatMapFunctionAdaptee<List<List<Writable>>, Row> {
+public class SequenceToRows implements FlatMapFunction<List<List<Writable>>, Row> {
+
+    private Schema schema;
+    private StructType structType;
 
     public SequenceToRows(Schema schema) {
-        super(new SequenceToRowsAdapter(schema));
+        this.schema = schema;
+        structType = DataFrames.fromSchemaSequence(schema);
     }
 
+
+    @Override
+    public Iterator<Row> call(List<List<Writable>> sequence) throws Exception {
+        if (sequence.size() == 0)
+            return Collections.emptyIterator();
+
+        String sequenceUUID = UUID.randomUUID().toString();
+
+        List<Row> out = new ArrayList<>(sequence.size());
+
+        int stepCount = 0;
+        for (List<Writable> step : sequence) {
+            Object[] values = new Object[step.size() + 2];
+            values[0] = sequenceUUID;
+            values[1] = stepCount++;
+            for (int i = 0; i < step.size(); i++) {
+                switch (schema.getColumnTypes().get(i)) {
+                    case Double:
+                        values[i + 2] = step.get(i).toDouble();
+                        break;
+                    case Integer:
+                        values[i + 2] = step.get(i).toInt();
+                        break;
+                    case Long:
+                        values[i + 2] = step.get(i).toLong();
+                        break;
+                    case Float:
+                        values[i + 2] = step.get(i).toFloat();
+                        break;
+                    default:
+                        throw new IllegalStateException(
+                                "This api should not be used with strings , binary data or ndarrays. This is only for columnar data");
+                }
+            }
+
+            Row row = new GenericRowWithSchema(values, structType);
+            out.add(row);
+        }
+
+        return out.iterator();
+    }
 }
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/sparkfunction/SequenceToRowsAdapter.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/sparkfunction/SequenceToRowsAdapter.java
deleted file mode 100644
index 2ca2f32ae..000000000
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/sparkfunction/SequenceToRowsAdapter.java
+++ /dev/null
@@ -1,87 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform.sparkfunction;
-
-import org.apache.spark.sql.Row;
-import org.apache.spark.sql.catalyst.expressions.GenericRowWithSchema;
-import org.apache.spark.sql.types.StructType;
-import org.datavec.api.transform.schema.Schema;
-import org.datavec.api.writable.Writable;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.DataFrames;
-
-import java.util.ArrayList;
-import java.util.Collections;
-import java.util.List;
-import java.util.UUID;
-
-/**
- * Convert a record to a row
- * @author Adam Gibson
- */
-public class SequenceToRowsAdapter implements FlatMapFunctionAdapter<List<List<Writable>>, Row> {
-
-    private Schema schema;
-    private StructType structType;
-
-    public SequenceToRowsAdapter(Schema schema) {
-        this.schema = schema;
-        structType = DataFrames.fromSchemaSequence(schema);
-    }
-
-
-    @Override
-    public Iterable<Row> call(List<List<Writable>> sequence) throws Exception {
-        if (sequence.size() == 0)
-            return Collections.emptyList();
-
-        String sequenceUUID = UUID.randomUUID().toString();
-
-        List<Row> out = new ArrayList<>(sequence.size());
-
-        int stepCount = 0;
-        for (List<Writable> step : sequence) {
-            Object[] values = new Object[step.size() + 2];
-            values[0] = sequenceUUID;
-            values[1] = stepCount++;
-            for (int i = 0; i < step.size(); i++) {
-                switch (schema.getColumnTypes().get(i)) {
-                    case Double:
-                        values[i + 2] = step.get(i).toDouble();
-                        break;
-                    case Integer:
-                        values[i + 2] = step.get(i).toInt();
-                        break;
-                    case Long:
-                        values[i + 2] = step.get(i).toLong();
-                        break;
-                    case Float:
-                        values[i + 2] = step.get(i).toFloat();
-                        break;
-                    default:
-                        throw new IllegalStateException(
-                                        "This api should not be used with strings , binary data or ndarrays. This is only for columnar data");
-                }
-            }
-
-            Row row = new GenericRowWithSchema(values, structType);
-            out.add(row);
-        }
-
-        return out;
-    }
-}
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SequenceSplitFunction.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SequenceSplitFunction.java
index 41981a736..1a8782dfb 100644
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SequenceSplitFunction.java
+++ b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SequenceSplitFunction.java
@@ -16,19 +16,27 @@
 
 package org.datavec.spark.transform.transform;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.datavec.api.transform.sequence.SequenceSplit;
 import org.datavec.api.writable.Writable;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 
+import java.util.Iterator;
 import java.util.List;
 
 /**
  * Created by Alex on 17/03/2016.
  */
-public class SequenceSplitFunction extends BaseFlatMapFunctionAdaptee<List<List<Writable>>, List<List<Writable>>> {
+public class SequenceSplitFunction implements FlatMapFunction<List<List<Writable>>, List<List<Writable>>> {
+
+    private final SequenceSplit split;
 
     public SequenceSplitFunction(SequenceSplit split) {
-        super(new SequenceSplitFunctionAdapter(split));
+        this.split = split;
+    }
+
+    @Override
+    public Iterator<List<List<Writable>>> call(List<List<Writable>> collections) throws Exception {
+        return split.split(collections).iterator();
     }
 
 }
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SequenceSplitFunctionAdapter.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SequenceSplitFunctionAdapter.java
deleted file mode 100644
index 5bde7ee62..000000000
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SequenceSplitFunctionAdapter.java
+++ /dev/null
@@ -1,41 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform.transform;
-
-import org.datavec.api.transform.sequence.SequenceSplit;
-import org.datavec.api.writable.Writable;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-
-import java.util.List;
-
-/**
- * Created by Alex on 17/03/2016.
- */
-public class SequenceSplitFunctionAdapter
-                implements FlatMapFunctionAdapter<List<List<Writable>>, List<List<Writable>>> {
-
-    private final SequenceSplit split;
-
-    public SequenceSplitFunctionAdapter(SequenceSplit split) {
-        this.split = split;
-    }
-
-    @Override
-    public Iterable<List<List<Writable>>> call(List<List<Writable>> collections) throws Exception {
-        return split.split(collections);
-    }
-}
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SparkTransformProcessFunction.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SparkTransformProcessFunction.java
index ffe3f80c6..81f07b1f4 100644
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SparkTransformProcessFunction.java
+++ b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SparkTransformProcessFunction.java
@@ -16,19 +16,32 @@
 
 package org.datavec.spark.transform.transform;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.datavec.api.transform.TransformProcess;
 import org.datavec.api.writable.Writable;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 
+import java.util.Collections;
+import java.util.Iterator;
 import java.util.List;
 
 /**
  * Spark function for executing a transform process
  */
-public class SparkTransformProcessFunction extends BaseFlatMapFunctionAdaptee<List<Writable>, List<Writable>> {
+public class SparkTransformProcessFunction implements FlatMapFunction<List<Writable>, List<Writable>> {
+
+    private final TransformProcess transformProcess;
 
     public SparkTransformProcessFunction(TransformProcess transformProcess) {
-        super(new SparkTransformProcessFunctionAdapter(transformProcess));
+        this.transformProcess = transformProcess;
+    }
+
+    @Override
+    public Iterator<List<Writable>> call(List<Writable> v1) throws Exception {
+        List<Writable> newList = transformProcess.execute(v1);
+        if (newList == null)
+            return Collections.emptyIterator(); //Example was filtered out
+        else
+            return Collections.singletonList(newList).iterator();
     }
 
 }
diff --git a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SparkTransformProcessFunctionAdapter.java b/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SparkTransformProcessFunctionAdapter.java
deleted file mode 100644
index 7b1766cc2..000000000
--- a/datavec/datavec-spark/src/main/java/org/datavec/spark/transform/transform/SparkTransformProcessFunctionAdapter.java
+++ /dev/null
@@ -1,45 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform.transform;
-
-import org.datavec.api.transform.TransformProcess;
-import org.datavec.api.writable.Writable;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-
-import java.util.Collections;
-import java.util.List;
-
-/**
- * Spark function for executing a transform process
- */
-public class SparkTransformProcessFunctionAdapter implements FlatMapFunctionAdapter<List<Writable>, List<Writable>> {
-
-    private final TransformProcess transformProcess;
-
-    public SparkTransformProcessFunctionAdapter(TransformProcess transformProcess) {
-        this.transformProcess = transformProcess;
-    }
-
-    @Override
-    public Iterable<List<Writable>> call(List<Writable> v1) throws Exception {
-        List<Writable> newList = transformProcess.execute(v1);
-        if (newList == null)
-            return Collections.emptyList(); //Example was filtered out
-        else
-            return Collections.singletonList(newList);
-    }
-}
diff --git a/datavec/datavec-spark/src/main/spark-1/org/datavec/spark/transform/BaseFlatMapFunctionAdaptee.java b/datavec/datavec-spark/src/main/spark-1/org/datavec/spark/transform/BaseFlatMapFunctionAdaptee.java
deleted file mode 100644
index af600a14d..000000000
--- a/datavec/datavec-spark/src/main/spark-1/org/datavec/spark/transform/BaseFlatMapFunctionAdaptee.java
+++ /dev/null
@@ -1,41 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform;
-
-import org.apache.spark.api.java.function.FlatMapFunction;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-
-/**
- * FlatMapFunction adapter to
- * hide incompatibilities between Spark 1.x and Spark 2.x
- *
- * This class should be used instead of direct referral to FlatMapFunction
- *
- */
-public class BaseFlatMapFunctionAdaptee<K, V> implements FlatMapFunction<K, V> {
-
-    protected final FlatMapFunctionAdapter<K, V> adapter;
-
-    public BaseFlatMapFunctionAdaptee(FlatMapFunctionAdapter<K, V> adapter) {
-        this.adapter = adapter;
-    }
-
-    @Override
-    public Iterable<V> call(K k) throws Exception {
-        return adapter.call(k);
-    }
-}
diff --git a/datavec/datavec-spark/src/main/spark-1/org/datavec/spark/transform/DataRowsFacade.java b/datavec/datavec-spark/src/main/spark-1/org/datavec/spark/transform/DataRowsFacade.java
deleted file mode 100644
index 0ad7c55bd..000000000
--- a/datavec/datavec-spark/src/main/spark-1/org/datavec/spark/transform/DataRowsFacade.java
+++ /dev/null
@@ -1,42 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform;
-
-import org.apache.spark.sql.DataFrame;
-
-/**
- * Dataframe facade to hide incompatibilities between Spark 1.x and Spark 2.x
- *
- * This class should be used instead of direct referral to DataFrame / Dataset
- *
- */
-public class DataRowsFacade {
-
-    private final DataFrame df;
-
-    private DataRowsFacade(DataFrame df) {
-        this.df = df;
-    }
-
-    public static DataRowsFacade dataRows(DataFrame df) {
-        return new DataRowsFacade(df);
-    }
-
-    public DataFrame get() {
-        return df;
-    }
-}
diff --git a/datavec/datavec-spark/src/main/spark-2/org/datavec/spark/transform/BaseFlatMapFunctionAdaptee.java b/datavec/datavec-spark/src/main/spark-2/org/datavec/spark/transform/BaseFlatMapFunctionAdaptee.java
deleted file mode 100644
index f30e5a222..000000000
--- a/datavec/datavec-spark/src/main/spark-2/org/datavec/spark/transform/BaseFlatMapFunctionAdaptee.java
+++ /dev/null
@@ -1,42 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform;
-
-import org.apache.spark.api.java.function.FlatMapFunction;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-
-import java.util.Iterator;
-
-/**
- * FlatMapFunction adapter to hide incompatibilities between Spark 1.x and Spark 2.x
- *
- * This class should be used instead of direct referral to FlatMapFunction
- *
- */
-public class BaseFlatMapFunctionAdaptee<K, V> implements FlatMapFunction<K, V> {
-
-    protected final FlatMapFunctionAdapter<K, V> adapter;
-
-    public BaseFlatMapFunctionAdaptee(FlatMapFunctionAdapter<K, V> adapter) {
-        this.adapter = adapter;
-    }
-
-    @Override
-    public Iterator<V> call(K k) throws Exception {
-        return adapter.call(k).iterator();
-    }
-}
diff --git a/datavec/datavec-spark/src/main/spark-2/org/datavec/spark/transform/DataRowsFacade.java b/datavec/datavec-spark/src/main/spark-2/org/datavec/spark/transform/DataRowsFacade.java
deleted file mode 100644
index 9958a622e..000000000
--- a/datavec/datavec-spark/src/main/spark-2/org/datavec/spark/transform/DataRowsFacade.java
+++ /dev/null
@@ -1,43 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.datavec.spark.transform;
-
-import org.apache.spark.sql.Dataset;
-import org.apache.spark.sql.Row;
-
-/**
- * Dataframe facade to hide incompatibilities between Spark 1.x and Spark 2.x
- *
- * This class should be used instead of direct referral to DataFrame / Dataset
- *
- */
-public class DataRowsFacade {
-
-    private final Dataset<Row> df;
-
-    private DataRowsFacade(Dataset<Row> df) {
-        this.df = df;
-    }
-
-    public static DataRowsFacade dataRows(Dataset<Row> df) {
-        return new DataRowsFacade(df);
-    }
-
-    public Dataset<Row> get() {
-        return df;
-    }
-}
diff --git a/datavec/datavec-spark/src/test/java/org/datavec/spark/storage/TestSparkStorageUtils.java b/datavec/datavec-spark/src/test/java/org/datavec/spark/storage/TestSparkStorageUtils.java
index 49a3946e2..8f0247568 100644
--- a/datavec/datavec-spark/src/test/java/org/datavec/spark/storage/TestSparkStorageUtils.java
+++ b/datavec/datavec-spark/src/test/java/org/datavec/spark/storage/TestSparkStorageUtils.java
@@ -16,7 +16,7 @@
 
 package org.datavec.spark.storage;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import org.apache.spark.api.java.JavaPairRDD;
 import org.apache.spark.api.java.JavaRDD;
 import org.datavec.api.writable.*;
diff --git a/datavec/datavec-spark/src/test/java/org/datavec/spark/transform/DataFramesTests.java b/datavec/datavec-spark/src/test/java/org/datavec/spark/transform/DataFramesTests.java
index 5b6ff6342..a19725a2a 100644
--- a/datavec/datavec-spark/src/test/java/org/datavec/spark/transform/DataFramesTests.java
+++ b/datavec/datavec-spark/src/test/java/org/datavec/spark/transform/DataFramesTests.java
@@ -19,6 +19,8 @@ package org.datavec.spark.transform;
 import org.apache.commons.math3.stat.descriptive.moment.StandardDeviation;
 import org.apache.spark.api.java.JavaRDD;
 import org.apache.spark.sql.Column;
+import org.apache.spark.sql.Dataset;
+import org.apache.spark.sql.Row;
 import org.datavec.api.transform.schema.Schema;
 import org.datavec.api.util.ndarray.RecordConverter;
 import org.datavec.api.writable.DoubleWritable;
@@ -46,9 +48,9 @@ public class DataFramesTests extends BaseSparkTest {
         for (int i = 0; i < numColumns; i++)
             builder.addColumnDouble(String.valueOf(i));
         Schema schema = builder.build();
-        DataRowsFacade dataFrame = DataFrames.toDataFrame(schema, sc.parallelize(records));
-        dataFrame.get().show();
-        dataFrame.get().describe(DataFrames.toArray(schema.getColumnNames())).show();
+        Dataset<Row> dataFrame = DataFrames.toDataFrame(schema, sc.parallelize(records));
+        dataFrame.show();
+        dataFrame.describe(DataFrames.toArray(schema.getColumnNames())).show();
         //        System.out.println(Normalization.minMaxColumns(dataFrame,schema.getColumnNames()));
         //        System.out.println(Normalization.stdDevMeanColumns(dataFrame,schema.getColumnNames()));
 
@@ -77,12 +79,12 @@ public class DataFramesTests extends BaseSparkTest {
         assertEquals(schema, DataFrames.fromStructType(DataFrames.fromSchema(schema)));
         assertEquals(rdd.collect(), DataFrames.toRecords(DataFrames.toDataFrame(schema, rdd)).getSecond().collect());
 
-        DataRowsFacade dataFrame = DataFrames.toDataFrame(schema, rdd);
-        dataFrame.get().show();
+        Dataset<Row> dataFrame = DataFrames.toDataFrame(schema, rdd);
+        dataFrame.show();
         Column mean = DataFrames.mean(dataFrame, "0");
         Column std = DataFrames.std(dataFrame, "0");
-        dataFrame.get().withColumn("0", dataFrame.get().col("0").minus(mean)).show();
-        dataFrame.get().withColumn("0", dataFrame.get().col("0").divide(std)).show();
+        dataFrame.withColumn("0", dataFrame.col("0").minus(mean)).show();
+        dataFrame.withColumn("0", dataFrame.col("0").divide(std)).show();
 
         /*   DataFrame desc = dataFrame.describe(dataFrame.columns());
         dataFrame.show();
diff --git a/datavec/datavec-spark/src/test/java/org/datavec/spark/transform/NormalizationTests.java b/datavec/datavec-spark/src/test/java/org/datavec/spark/transform/NormalizationTests.java
index 3b1b1e1a6..5352ec10d 100644
--- a/datavec/datavec-spark/src/test/java/org/datavec/spark/transform/NormalizationTests.java
+++ b/datavec/datavec-spark/src/test/java/org/datavec/spark/transform/NormalizationTests.java
@@ -17,6 +17,7 @@
 package org.datavec.spark.transform;
 
 import org.apache.spark.api.java.JavaRDD;
+import org.apache.spark.sql.Dataset;
 import org.apache.spark.sql.Row;
 import org.datavec.api.transform.schema.Schema;
 import org.datavec.api.util.ndarray.RecordConverter;
@@ -24,11 +25,13 @@ import org.datavec.api.writable.DoubleWritable;
 import org.datavec.api.writable.Writable;
 import org.datavec.spark.BaseSparkTest;
 import org.junit.Test;
+import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.DataSet;
 import org.nd4j.linalg.dataset.api.preprocessor.DataNormalization;
 import org.nd4j.linalg.dataset.api.preprocessor.NormalizerMinMaxScaler;
 import org.nd4j.linalg.dataset.api.preprocessor.NormalizerStandardize;
+import org.nd4j.linalg.factory.Nd4j;
 
 import java.util.ArrayList;
 import java.util.List;
@@ -50,36 +53,35 @@ public class NormalizationTests extends BaseSparkTest {
         for (int i = 0; i < numColumns; i++)
             builder.addColumnDouble(String.valueOf(i));
 
+        Nd4j.getRandom().setSeed(12345);
+
+        INDArray arr = Nd4j.rand(DataType.FLOAT, 5, numColumns);
         for (int i = 0; i < 5; i++) {
             List<Writable> record = new ArrayList<>(numColumns);
             data.add(record);
             for (int j = 0; j < numColumns; j++) {
-                record.add(new DoubleWritable(1.0));
+                record.add(new DoubleWritable(arr.getDouble(i, j)));
             }
-
         }
 
-        INDArray arr = RecordConverter.toMatrix(data);
 
         Schema schema = builder.build();
         JavaRDD<List<Writable>> rdd = sc.parallelize(data);
-        DataRowsFacade dataFrame = DataFrames.toDataFrame(schema, rdd);
+        Dataset<Row> dataFrame = DataFrames.toDataFrame(schema, rdd);
 
         //assert equivalent to the ndarray pre processing
-        NormalizerStandardize standardScaler = new NormalizerStandardize();
-        standardScaler.fit(new DataSet(arr.dup(), arr.dup()));
-        INDArray standardScalered = arr.dup();
-        standardScaler.transform(new DataSet(standardScalered, standardScalered));
         DataNormalization zeroToOne = new NormalizerMinMaxScaler();
         zeroToOne.fit(new DataSet(arr.dup(), arr.dup()));
         INDArray zeroToOnes = arr.dup();
         zeroToOne.transform(new DataSet(zeroToOnes, zeroToOnes));
-        List<Row> rows = Normalization.stdDevMeanColumns(dataFrame, dataFrame.get().columns());
+        List<Row> rows = Normalization.stdDevMeanColumns(dataFrame, dataFrame.columns());
         INDArray assertion = DataFrames.toMatrix(rows);
-        //compare standard deviation
-        assertTrue(standardScaler.getStd().equalsWithEps(assertion.getRow(0), 1e-1));
+        INDArray expStd = arr.std(true, true, 0);
+        INDArray std = assertion.getRow(0, true);
+        assertTrue(expStd.equalsWithEps(std, 1e-3));
         //compare mean
-        assertTrue(standardScaler.getMean().equalsWithEps(assertion.getRow(1), 1e-1));
+        INDArray expMean = arr.mean(true, 0);
+        assertTrue(expMean.equalsWithEps(assertion.getRow(1, true), 1e-3));
 
     }
 
@@ -109,10 +111,10 @@ public class NormalizationTests extends BaseSparkTest {
         assertEquals(schema, DataFrames.fromStructType(DataFrames.fromSchema(schema)));
         assertEquals(rdd.collect(), DataFrames.toRecords(DataFrames.toDataFrame(schema, rdd)).getSecond().collect());
 
-        DataRowsFacade dataFrame = DataFrames.toDataFrame(schema, rdd);
-        dataFrame.get().show();
-        Normalization.zeromeanUnitVariance(dataFrame).get().show();
-        Normalization.normalize(dataFrame).get().show();
+        Dataset<Row> dataFrame = DataFrames.toDataFrame(schema, rdd);
+        dataFrame.show();
+        Normalization.zeromeanUnitVariance(dataFrame).show();
+        Normalization.normalize(dataFrame).show();
 
         //assert equivalent to the ndarray pre processing
         NormalizerStandardize standardScaler = new NormalizerStandardize();
diff --git a/deeplearning4j/deeplearning4j-common/src/main/java/org/deeplearning4j/config/DL4JSystemProperties.java b/deeplearning4j/deeplearning4j-common/src/main/java/org/deeplearning4j/config/DL4JSystemProperties.java
index 7b0d9d3c1..04773ee74 100644
--- a/deeplearning4j/deeplearning4j-common/src/main/java/org/deeplearning4j/config/DL4JSystemProperties.java
+++ b/deeplearning4j/deeplearning4j-common/src/main/java/org/deeplearning4j/config/DL4JSystemProperties.java
@@ -52,18 +52,6 @@ public class DL4JSystemProperties {
      */
     public static final String DL4J_RESOURCES_BASE_URL_PROPERTY = "org.deeplearning4j.resources.baseurl";
 
-    /**
-     * Applicability: deeplearning4j-nn<br>
-     * Description: Used for loading legacy format JSON containing custom layers. This system property is provided as an
-     * alternative to {@code NeuralNetConfiguration#registerLegacyCustomClassesForJSON(Class[])}. Classes are specified in
-     * comma-separated format.<br>
-     * This is required ONLY when ALL of the following conditions are met:<br>
-     * 1. You want to load a serialized net, saved in 1.0.0-alpha or before, AND<br>
-     * 2. The serialized net has a custom Layer, GraphVertex, etc (i.e., one not defined in DL4J), AND<br>
-     * 3. You haven't already called {@code NeuralNetConfiguration#registerLegacyCustomClassesForJSON(Class[])}
-     */
-    public static final String CUSTOM_REGISTRATION_PROPERTY = "org.deeplearning4j.config.custom.legacyclasses";
-
     /**
      * Applicability: deeplearning4j-nn<br>
      * Description: DL4J writes some crash dumps to disk when an OOM exception occurs - this functionality is enabled
diff --git a/deeplearning4j/deeplearning4j-core/pom.xml b/deeplearning4j/deeplearning4j-core/pom.xml
index aeb5fe04b..81142fd68 100644
--- a/deeplearning4j/deeplearning4j-core/pom.xml
+++ b/deeplearning4j/deeplearning4j-core/pom.xml
@@ -96,18 +96,6 @@
             <scope>test</scope>
         </dependency>
 
-        <dependency>
-            <groupId>com.google.guava</groupId>
-            <artifactId>guava</artifactId>
-            <version>${guava.version}</version>
-            <exclusions>
-              <exclusion>
-                <groupId>com.google.code.findbugs</groupId>
-                <artifactId>jsr305</artifactId>
-              </exclusion>
-            </exclusions>
-        </dependency>
-
         <dependency>
             <groupId>org.nd4j</groupId>
             <artifactId>nd4j-api</artifactId>
diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/datavec/RecordReaderDataSetiteratorTest.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/datavec/RecordReaderDataSetiteratorTest.java
index 56317b946..3bd1bd37f 100644
--- a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/datavec/RecordReaderDataSetiteratorTest.java
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/datavec/RecordReaderDataSetiteratorTest.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.datasets.datavec;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import lombok.extern.slf4j.Slf4j;
 import org.apache.commons.io.FileUtils;
 import org.apache.commons.io.FilenameUtils;
diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/datavec/RecordReaderMultiDataSetIteratorTest.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/datavec/RecordReaderMultiDataSetIteratorTest.java
index 6ff639cb4..1e82a4783 100644
--- a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/datavec/RecordReaderMultiDataSetIteratorTest.java
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/datavec/RecordReaderMultiDataSetIteratorTest.java
@@ -17,7 +17,7 @@
 package org.deeplearning4j.datasets.datavec;
 
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import org.apache.commons.compress.utils.IOUtils;
 import org.apache.commons.io.FileUtils;
 import org.apache.commons.io.FilenameUtils;
diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/dtypes/DTypeTests.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/dtypes/DTypeTests.java
index 935d83218..52d3b0774 100644
--- a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/dtypes/DTypeTests.java
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/dtypes/DTypeTests.java
@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.nn.dtypes;
 
-import com.google.common.collect.ImmutableSet;
-import com.google.common.reflect.ClassPath;
+import org.nd4j.shade.guava.collect.ImmutableSet;
+import org.nd4j.shade.guava.reflect.ClassPath;
 import lombok.extern.slf4j.Slf4j;
 import org.deeplearning4j.BaseDL4JTest;
 import org.deeplearning4j.TestUtils;
@@ -103,7 +103,7 @@ public class DTypeTests extends BaseDL4JTest {
         ImmutableSet<ClassPath.ClassInfo> info;
         try {
             //Dependency note: this ClassPath class was added in Guava 14
-            info = com.google.common.reflect.ClassPath.from(DTypeTests.class.getClassLoader())
+            info = org.nd4j.shade.guava.reflect.ClassPath.from(DTypeTests.class.getClassLoader())
                     .getTopLevelClassesRecursive("org.deeplearning4j");
         } catch (IOException e) {
             //Should never happen
diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/layers/recurrent/TestRnnLayers.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/layers/recurrent/TestRnnLayers.java
index fb5836a99..11e45c51d 100644
--- a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/layers/recurrent/TestRnnLayers.java
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/layers/recurrent/TestRnnLayers.java
@@ -229,7 +229,9 @@ public class TestRnnLayers extends BaseDL4JTest {
                 net.fit(in,l);
             } catch (Throwable t){
                 String msg = t.getMessage();
-                assertTrue(msg, msg.contains("sequence length") && msg.contains("input") && msg.contains("label"));
+                if(msg == null)
+                    t.printStackTrace();
+                assertTrue(msg, msg != null && msg.contains("sequence length") && msg.contains("input") && msg.contains("label"));
             }
 
         }
diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/plot/BarnesHutTsneTest.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/plot/BarnesHutTsneTest.java
index 77fc63aa0..aba60fe4f 100644
--- a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/plot/BarnesHutTsneTest.java
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/plot/BarnesHutTsneTest.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.plot;
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
 import org.apache.commons.io.IOUtils;
diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/regressiontest/RegressionTest100a.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/regressiontest/RegressionTest100a.java
index f112d4386..a66914cd7 100644
--- a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/regressiontest/RegressionTest100a.java
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/regressiontest/RegressionTest100a.java
@@ -22,23 +22,24 @@ import org.deeplearning4j.TestUtils;
 import org.deeplearning4j.nn.api.Layer;
 import org.deeplearning4j.nn.conf.BackpropType;
 import org.deeplearning4j.nn.conf.ConvolutionMode;
-import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
 import org.deeplearning4j.nn.conf.graph.LayerVertex;
-import org.deeplearning4j.nn.conf.layers.*;
+import org.deeplearning4j.nn.conf.layers.BatchNormalization;
+import org.deeplearning4j.nn.conf.layers.ConvolutionLayer;
+import org.deeplearning4j.nn.conf.layers.GravesLSTM;
+import org.deeplearning4j.nn.conf.layers.RnnOutputLayer;
 import org.deeplearning4j.nn.conf.layers.variational.VariationalAutoencoder;
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyLayerDeserializer;
 import org.deeplearning4j.nn.graph.ComputationGraph;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
 import org.deeplearning4j.nn.weights.WeightInitXavier;
-import org.deeplearning4j.regressiontest.customlayer100a.CustomLayer;
-import org.junit.Before;
 import org.junit.Ignore;
 import org.junit.Test;
-import org.nd4j.linalg.activations.impl.*;
+import org.nd4j.linalg.activations.impl.ActivationIdentity;
+import org.nd4j.linalg.activations.impl.ActivationLReLU;
+import org.nd4j.linalg.activations.impl.ActivationSoftmax;
+import org.nd4j.linalg.activations.impl.ActivationTanH;
 import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.factory.Nd4j;
-import org.nd4j.linalg.io.ClassPathResource;
 import org.nd4j.linalg.learning.config.Adam;
 import org.nd4j.linalg.learning.config.RmsProp;
 import org.nd4j.linalg.learning.regularization.WeightDecay;
@@ -60,6 +61,9 @@ public class RegressionTest100a extends BaseDL4JTest {
 
     @Test
     public void testCustomLayer() throws Exception {
+        //We dropped support for 1.0.0-alpha and earlier custom layers due to the maintenance overhead for a rarely used feature
+        //An upgrade path exists as a workaround - load in beta to beta4 and re-save
+        //All built-in layers can be loaded going back to 0.5.0
 
         File f = Resources.asFile("regression_testing/100a/CustomLayerExample_100a.bin");
 
@@ -68,67 +72,8 @@ public class RegressionTest100a extends BaseDL4JTest {
             fail("Expected exception");
         } catch (Exception e){
             String msg = e.getMessage();
-            assertTrue(msg, msg.contains("NeuralNetConfiguration.registerLegacyCustomClassesForJSON"));
+            assertTrue(msg, msg.contains("custom") && msg.contains("1.0.0-beta") && msg.contains("saved again"));
         }
-
-        NeuralNetConfiguration.registerLegacyCustomClassesForJSON(CustomLayer.class);
-
-        MultiLayerNetwork net = MultiLayerNetwork.load(f, true);
-
-        DenseLayer l0 = (DenseLayer) net.getLayer(0).conf().getLayer();
-        assertEquals(new ActivationTanH(), l0.getActivationFn());
-        assertEquals(new WeightDecay(0.03, false), TestUtils.getWeightDecayReg(l0));
-        assertEquals(new RmsProp(0.95), l0.getIUpdater());
-
-        CustomLayer l1 = (CustomLayer) net.getLayer(1).conf().getLayer();
-        assertEquals(new ActivationTanH(), l1.getActivationFn());
-        assertEquals(new ActivationSigmoid(), l1.getSecondActivationFunction());
-        assertEquals(new RmsProp(0.95), l1.getIUpdater());
-
-
-        INDArray outExp;
-        File f2 = Resources.asFile("regression_testing/100a/CustomLayerExample_Output_100a.bin");
-        try(DataInputStream dis = new DataInputStream(new FileInputStream(f2))){
-            outExp = Nd4j.read(dis);
-        }
-
-        INDArray in;
-        File f3 = Resources.asFile("regression_testing/100a/CustomLayerExample_Input_100a.bin");
-        try(DataInputStream dis = new DataInputStream(new FileInputStream(f3))){
-            in = Nd4j.read(dis);
-        }
-
-        INDArray outAct = net.output(in);
-
-        assertEquals(outExp, outAct);
-
-
-        //Check graph
-        f = Resources.asFile("regression_testing/100a/CustomLayerExample_Graph_100a.bin");
-
-        //Deregister custom class:
-        new LegacyLayerDeserializer().getLegacyNamesMap().remove("CustomLayer");
-
-        try {
-            ComputationGraph.load(f, true);
-            fail("Expected exception");
-        } catch (Exception e){
-            String msg = e.getMessage();
-            assertTrue(msg, msg.contains("NeuralNetConfiguration.registerLegacyCustomClassesForJSON"));
-        }
-
-        NeuralNetConfiguration.registerLegacyCustomClassesForJSON(CustomLayer.class);
-
-        ComputationGraph graph = ComputationGraph.load(f, true);
-
-        f2 = Resources.asFile("regression_testing/100a/CustomLayerExample_Graph_Output_100a.bin");
-        try(DataInputStream dis = new DataInputStream(new FileInputStream(f2))){
-            outExp = Nd4j.read(dis);
-        }
-
-        outAct = graph.outputSingle(in);
-
-        assertEquals(outExp, outAct);
     }
 
 
diff --git a/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/MultipleEpochsIterator.java b/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/MultipleEpochsIterator.java
index e7fb7246a..592aefcd6 100755
--- a/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/MultipleEpochsIterator.java
+++ b/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/MultipleEpochsIterator.java
@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.datasets.iterator;
 
-import com.google.common.annotations.VisibleForTesting;
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.annotations.VisibleForTesting;
+import org.nd4j.shade.guava.collect.Lists;
 import lombok.Getter;
 import org.nd4j.linalg.dataset.DataSet;
 import org.nd4j.linalg.dataset.api.DataSetPreProcessor;
diff --git a/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/parallel/FileSplitParallelDataSetIterator.java b/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/parallel/FileSplitParallelDataSetIterator.java
index 955732c07..02f2c4eb0 100644
--- a/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/parallel/FileSplitParallelDataSetIterator.java
+++ b/deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/parallel/FileSplitParallelDataSetIterator.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.datasets.iterator.parallel;
 
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.collect.Lists;
 import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
 import org.apache.commons.io.FileUtils;
diff --git a/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/src/main/java/org/deeplearning4j/plot/BarnesHutTsne.java b/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/src/main/java/org/deeplearning4j/plot/BarnesHutTsne.java
index dd83c1bd4..7eca0fac0 100644
--- a/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/src/main/java/org/deeplearning4j/plot/BarnesHutTsne.java
+++ b/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/src/main/java/org/deeplearning4j/plot/BarnesHutTsne.java
@@ -17,7 +17,7 @@
 package org.deeplearning4j.plot;
 
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import lombok.AllArgsConstructor;
 import lombok.Data;
 import lombok.Setter;
diff --git a/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/src/main/java/org/deeplearning4j/plot/Tsne.java b/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/src/main/java/org/deeplearning4j/plot/Tsne.java
index 41b50795e..9efb88e24 100644
--- a/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/src/main/java/org/deeplearning4j/plot/Tsne.java
+++ b/deeplearning4j/deeplearning4j-manifold/deeplearning4j-tsne/src/main/java/org/deeplearning4j/plot/Tsne.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.plot;
 
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Ints;
 import org.apache.commons.math3.util.FastMath;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dimensionalityreduction.PCA;
diff --git a/deeplearning4j/deeplearning4j-modelimport/pom.xml b/deeplearning4j/deeplearning4j-modelimport/pom.xml
index 59f05a6e8..dec29266f 100644
--- a/deeplearning4j/deeplearning4j-modelimport/pom.xml
+++ b/deeplearning4j/deeplearning4j-modelimport/pom.xml
@@ -37,6 +37,12 @@
             <version>${nd4j.version}</version>
         </dependency>
 
+        <dependency>
+            <groupId>com.google.code.gson</groupId>
+            <artifactId>gson</artifactId>
+            <version>${gson.version}</version>
+        </dependency>
+
         <dependency>
             <groupId>org.deeplearning4j</groupId>
             <artifactId>deeplearning4j-nn</artifactId>
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml
index 5351f0955..7477c7794 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/pom.xml
@@ -77,26 +77,11 @@
             <version>${project.version}</version>
         </dependency>
 
-        <dependency>
-            <groupId>com.google.guava</groupId>
-            <artifactId>guava</artifactId>
-            <version>${guava.version}</version>
-        </dependency>
         <dependency>
             <groupId>com.google.protobuf</groupId>
             <artifactId>protobuf-java</artifactId>
             <version>${google.protobuf.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-actor_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-slf4j_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
         <dependency>
             <groupId>joda-time</groupId>
             <artifactId>joda-time</artifactId>
@@ -213,11 +198,6 @@
             <artifactId>play-netty-server_2.11</artifactId>
             <version>${playframework.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-cluster_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
     </dependencies>
 
     <profiles>
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/src/main/java/org/deeplearning4j/nearestneighbor/server/FunctionUtil.java b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/src/main/java/org/deeplearning4j/nearestneighbor/server/FunctionUtil.java
deleted file mode 100644
index df178fd70..000000000
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/src/main/java/org/deeplearning4j/nearestneighbor/server/FunctionUtil.java
+++ /dev/null
@@ -1,41 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.nearestneighbor.server;
-
-import play.libs.F;
-import play.mvc.Result;
-
-import java.util.function.Function;
-import java.util.function.Supplier;
-
-/**
- * Utility methods for Routing
- *
- * @author Alex Black
- */
-public class FunctionUtil {
-
-
-    public static F.Function0<Result> function0(Supplier<Result> supplier) {
-        return supplier::get;
-    }
-
-    public static <T> F.Function<T, Result> function(Function<T, Result> function) {
-        return function::apply;
-    }
-
-}
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/src/main/java/org/deeplearning4j/nearestneighbor/server/NearestNeighborsServer.java b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/src/main/java/org/deeplearning4j/nearestneighbor/server/NearestNeighborsServer.java
index 58682d5e1..a79b57b19 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/src/main/java/org/deeplearning4j/nearestneighbor/server/NearestNeighborsServer.java
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/deeplearning4j-nearestneighbor-server/src/main/java/org/deeplearning4j/nearestneighbor/server/NearestNeighborsServer.java
@@ -33,8 +33,10 @@ import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.indexing.NDArrayIndex;
 import org.nd4j.serde.base64.Nd4jBase64;
 import org.nd4j.serde.binary.BinarySerde;
+import play.BuiltInComponents;
 import play.Mode;
 import play.libs.Json;
+import play.routing.Router;
 import play.routing.RoutingDsl;
 import play.server.Server;
 
@@ -149,19 +151,36 @@ public class NearestNeighborsServer {
 
         VPTree tree = new VPTree(points, similarityFunction, invert);
 
-        RoutingDsl routingDsl = new RoutingDsl();
+        //Set play secret key, if required
+        //http://www.playframework.com/documentation/latest/ApplicationSecret
+        String crypto = System.getProperty("play.crypto.secret");
+        if (crypto == null || "changeme".equals(crypto) || "".equals(crypto)) {
+            byte[] newCrypto = new byte[1024];
+
+            new Random().nextBytes(newCrypto);
+
+            String base64 = Base64.getEncoder().encodeToString(newCrypto);
+            System.setProperty("play.crypto.secret", base64);
+        }
+
+
+        server = Server.forRouter(Mode.PROD, port, b -> createRouter(tree, labels, points, b));
+    }
+
+    protected Router createRouter(VPTree tree, List<String> labels, INDArray points, BuiltInComponents builtInComponents){
+        RoutingDsl routingDsl = RoutingDsl.fromComponents(builtInComponents);
         //return the host information for a given id
-        routingDsl.POST("/knn").routeTo(FunctionUtil.function0((() -> {
+        routingDsl.POST("/knn").routingTo(request -> {
             try {
-                NearestNeighborRequest record = Json.fromJson(request().body().asJson(), NearestNeighborRequest.class);
+                NearestNeighborRequest record = Json.fromJson(request.body().asJson(), NearestNeighborRequest.class);
                 NearestNeighbor nearestNeighbor =
-                                NearestNeighbor.builder().points(points).record(record).tree(tree).build();
+                        NearestNeighbor.builder().points(points).record(record).tree(tree).build();
 
                 if (record == null)
                     return badRequest(Json.toJson(Collections.singletonMap("status", "invalid json passed.")));
 
                 NearestNeighborsResults results =
-                                NearestNeighborsResults.builder().results(nearestNeighbor.search()).build();
+                        NearestNeighborsResults.builder().results(nearestNeighbor.search()).build();
 
 
                 return ok(Json.toJson(results));
@@ -171,11 +190,11 @@ public class NearestNeighborsServer {
                 e.printStackTrace();
                 return internalServerError(e.getMessage());
             }
-        })));
+        });
 
-        routingDsl.POST("/knnnew").routeTo(FunctionUtil.function0((() -> {
+        routingDsl.POST("/knnnew").routingTo(request -> {
             try {
-                Base64NDArrayBody record = Json.fromJson(request().body().asJson(), Base64NDArrayBody.class);
+                Base64NDArrayBody record = Json.fromJson(request.body().asJson(), Base64NDArrayBody.class);
                 if (record == null)
                     return badRequest(Json.toJson(Collections.singletonMap("status", "invalid json passed.")));
 
@@ -216,23 +235,9 @@ public class NearestNeighborsServer {
                 e.printStackTrace();
                 return internalServerError(e.getMessage());
             }
-        })));
-
-        //Set play secret key, if required
-        //http://www.playframework.com/documentation/latest/ApplicationSecret
-        String crypto = System.getProperty("play.crypto.secret");
-        if (crypto == null || "changeme".equals(crypto) || "".equals(crypto)) {
-            byte[] newCrypto = new byte[1024];
-
-            new Random().nextBytes(newCrypto);
-
-            String base64 = Base64.getEncoder().encodeToString(newCrypto);
-            System.setProperty("play.crypto.secret", base64);
-        }
-
-        server = Server.forRouter(routingDsl.build(), Mode.PROD, port);
-
+        });
 
+        return routingDsl.build();
     }
 
     /**
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml
index 8b774cd56..f95f9268d 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/pom.xml
@@ -59,6 +59,13 @@
             <version>${project.version}</version>
             <scope>test</scope>
         </dependency>
+
+        <dependency>
+            <groupId>joda-time</groupId>
+            <artifactId>joda-time</artifactId>
+            <version>2.10.3</version>
+            <scope>test</scope>
+        </dependency>
     </dependencies>
 
 
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/info/ClusterSetInfo.java b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/info/ClusterSetInfo.java
index 1c57bc38a..cae103f10 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/info/ClusterSetInfo.java
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/info/ClusterSetInfo.java
@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.clustering.info;
 
-import com.google.common.collect.HashBasedTable;
-import com.google.common.collect.Table;
+import org.nd4j.shade.guava.collect.HashBasedTable;
+import org.nd4j.shade.guava.collect.Table;
 import org.deeplearning4j.clustering.cluster.Cluster;
 import org.deeplearning4j.clustering.cluster.ClusterSet;
 
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/quadtree/QuadTree.java b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/quadtree/QuadTree.java
index c2154e6ba..f1cc2e304 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/quadtree/QuadTree.java
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/quadtree/QuadTree.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.clustering.quadtree;
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import org.apache.commons.math3.util.FastMath;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.factory.Nd4j;
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/randomprojection/RPUtils.java b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/randomprojection/RPUtils.java
index 11746f4c2..5c31ee78a 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/randomprojection/RPUtils.java
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/randomprojection/RPUtils.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.clustering.randomprojection;
 
-import com.google.common.primitives.Doubles;
+import org.nd4j.shade.guava.primitives.Doubles;
 import lombok.val;
 import org.nd4j.autodiff.functions.DifferentialFunction;
 import org.nd4j.linalg.api.ndarray.INDArray;
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/sptree/SpTree.java b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/sptree/SpTree.java
index 83af0365a..659f334df 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/sptree/SpTree.java
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/sptree/SpTree.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.clustering.sptree;
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import lombok.val;
 import org.deeplearning4j.clustering.algorithm.Distance;
 import org.deeplearning4j.nn.conf.WorkspaceMode;
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/kdtree/KDTreeTest.java b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/kdtree/KDTreeTest.java
index 4b7b8e567..1de7a379b 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/kdtree/KDTreeTest.java
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/kdtree/KDTreeTest.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.clustering.kdtree;
 
-import com.google.common.primitives.Doubles;
+import org.nd4j.shade.guava.primitives.Doubles;
 import lombok.val;
 import org.deeplearning4j.clustering.BaseDL4JTest;
 import org.joda.time.Duration;
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/sptree/SPTreeTest.java b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/sptree/SPTreeTest.java
index 03ad90748..f5ee19403 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/sptree/SPTreeTest.java
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/sptree/SPTreeTest.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.clustering.sptree;
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import org.apache.commons.lang3.time.StopWatch;
 import org.deeplearning4j.clustering.BaseDL4JTest;
 import org.junit.Before;
diff --git a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/vptree/VpTreeNodeTest.java b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/vptree/VpTreeNodeTest.java
index 76658438a..5edb3926a 100644
--- a/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/vptree/VpTreeNodeTest.java
+++ b/deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/vptree/VpTreeNodeTest.java
@@ -18,12 +18,10 @@ package org.deeplearning4j.clustering.vptree;
 
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
-import org.apache.commons.lang3.ArrayUtils;
 import org.deeplearning4j.clustering.BaseDL4JTest;
 import org.deeplearning4j.clustering.sptree.DataPoint;
 import org.joda.time.Duration;
 import org.junit.BeforeClass;
-import org.junit.Ignore;
 import org.junit.Test;
 import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
@@ -33,7 +31,6 @@ import org.nd4j.linalg.primitives.Counter;
 import org.nd4j.linalg.primitives.Pair;
 
 import java.util.*;
-import java.util.concurrent.TimeUnit;
 
 import static org.junit.Assert.*;
 
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/main/java/org/deeplearning4j/text/corpora/sentiwordnet/SWN3.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/main/java/org/deeplearning4j/text/corpora/sentiwordnet/SWN3.java
index 67169132c..03d2462a5 100755
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/main/java/org/deeplearning4j/text/corpora/sentiwordnet/SWN3.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/main/java/org/deeplearning4j/text/corpora/sentiwordnet/SWN3.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.text.corpora.sentiwordnet;
 
-import com.google.common.collect.Sets;
+import org.nd4j.shade.guava.collect.Sets;
 import org.apache.uima.analysis_engine.AnalysisEngine;
 import org.apache.uima.cas.CAS;
 import org.apache.uima.cas.CASException;
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/test/java/org/deeplearning4j/models/WordVectorSerializerTest.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/test/java/org/deeplearning4j/models/WordVectorSerializerTest.java
index d2d752509..f4dd1a6c5 100755
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/test/java/org/deeplearning4j/models/WordVectorSerializerTest.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/test/java/org/deeplearning4j/models/WordVectorSerializerTest.java
@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.models;
 
-import com.google.common.io.Files;
-import com.google.common.primitives.Doubles;
+import org.nd4j.shade.guava.io.Files;
+import org.nd4j.shade.guava.primitives.Doubles;
 import lombok.val;
 import org.apache.commons.io.FileUtils;
 import org.apache.commons.lang.ArrayUtils;
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/test/java/org/deeplearning4j/models/word2vec/Word2VecTests.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/test/java/org/deeplearning4j/models/word2vec/Word2VecTests.java
index 61e31b3c7..01b38a644 100755
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/test/java/org/deeplearning4j/models/word2vec/Word2VecTests.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/test/java/org/deeplearning4j/models/word2vec/Word2VecTests.java
@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.models.word2vec;
 
-import com.google.common.primitives.Doubles;
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Doubles;
+import org.nd4j.shade.guava.primitives.Ints;
 import lombok.val;
 import net.didion.jwnl.data.Word;
 import org.apache.commons.io.FileUtils;
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/inmemory/InMemoryLookupTable.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/inmemory/InMemoryLookupTable.java
index 5be56964c..579caa0a3 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/inmemory/InMemoryLookupTable.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/inmemory/InMemoryLookupTable.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.models.embeddings.inmemory;
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import lombok.Getter;
 import lombok.NonNull;
 import lombok.Setter;
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/reader/impl/BasicModelUtils.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/reader/impl/BasicModelUtils.java
index df502aded..84fc17b7e 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/reader/impl/BasicModelUtils.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/reader/impl/BasicModelUtils.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.models.embeddings.reader.impl;
 
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.collect.Lists;
 import lombok.AllArgsConstructor;
 import lombok.Data;
 import lombok.NonNull;
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/wordvectors/WordVectorsImpl.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/wordvectors/WordVectorsImpl.java
index 75511cae1..f71c56717 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/wordvectors/WordVectorsImpl.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/embeddings/wordvectors/WordVectorsImpl.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.models.embeddings.wordvectors;
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import lombok.Getter;
 import lombok.NonNull;
 import lombok.Setter;
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/glove/count/CountMap.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/glove/count/CountMap.java
index f97aee9c0..8680a809c 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/glove/count/CountMap.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/glove/count/CountMap.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.models.glove.count;
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import org.deeplearning4j.models.sequencevectors.sequence.SequenceElement;
 import org.nd4j.linalg.primitives.Pair;
 
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/paragraphvectors/ParagraphVectors.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/paragraphvectors/ParagraphVectors.java
index 4c05b7cc7..64ee79dd4 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/paragraphvectors/ParagraphVectors.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/paragraphvectors/ParagraphVectors.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.models.paragraphvectors;
 
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.collect.Lists;
 import com.google.gson.JsonObject;
 import com.google.gson.JsonParser;
 import lombok.Getter;
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/sequencevectors/SequenceVectors.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/sequencevectors/SequenceVectors.java
index 78a878930..87dd0880a 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/sequencevectors/SequenceVectors.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/sequencevectors/SequenceVectors.java
@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.models.sequencevectors;
 
-import com.google.common.primitives.Ints;
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import lombok.Getter;
 import lombok.NonNull;
 import lombok.Setter;
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/sequencevectors/sequence/SequenceElement.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/sequencevectors/sequence/SequenceElement.java
index d5e789ebf..99263f6bc 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/sequencevectors/sequence/SequenceElement.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/sequencevectors/sequence/SequenceElement.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.models.sequencevectors.sequence;
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import lombok.Getter;
 import lombok.NonNull;
 import lombok.Setter;
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/text/invertedindex/InvertedIndex.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/text/invertedindex/InvertedIndex.java
index cdb4b6c9e..5e51cef20 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/text/invertedindex/InvertedIndex.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/text/invertedindex/InvertedIndex.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.text.invertedindex;
 
-import com.google.common.base.Function;
+import org.nd4j.shade.guava.base.Function;
 import org.deeplearning4j.models.sequencevectors.sequence.SequenceElement;
 import org.nd4j.linalg.primitives.Pair;
 
diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/embeddings/wordvectors/WordVectorsImplTest.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/embeddings/wordvectors/WordVectorsImplTest.java
index 41536ff70..f5e0ca388 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/embeddings/wordvectors/WordVectorsImplTest.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/embeddings/wordvectors/WordVectorsImplTest.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.models.embeddings.wordvectors;
 
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.collect.Lists;
 import org.deeplearning4j.BaseDL4JTest;
 import org.deeplearning4j.models.embeddings.WeightLookupTable;
 import org.deeplearning4j.models.sequencevectors.sequence.SequenceElement;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/ConfusionMatrix.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/ConfusionMatrix.java
index 5834e1647..69767df13 100755
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/ConfusionMatrix.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/ConfusionMatrix.java
@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.eval;
 
-import com.google.common.collect.HashMultiset;
-import com.google.common.collect.Multiset;
+import org.nd4j.shade.guava.collect.HashMultiset;
+import org.nd4j.shade.guava.collect.Multiset;
 import lombok.Getter;
 
 import java.io.Serializable;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/curves/PrecisionRecallCurve.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/curves/PrecisionRecallCurve.java
index e964cd9b0..2b00ac375 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/curves/PrecisionRecallCurve.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/curves/PrecisionRecallCurve.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.eval.curves;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import lombok.AllArgsConstructor;
 import lombok.Data;
 import lombok.EqualsAndHashCode;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/curves/RocCurve.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/curves/RocCurve.java
index b66230ddd..5d5e65c2a 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/curves/RocCurve.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/eval/curves/RocCurve.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.eval.curves;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import lombok.Data;
 import lombok.EqualsAndHashCode;
 import org.nd4j.shade.jackson.annotation.JsonProperty;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/ComputationGraphConfiguration.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/ComputationGraphConfiguration.java
index 6cd8f06b3..5a5ce5665 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/ComputationGraphConfiguration.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/ComputationGraphConfiguration.java
@@ -31,6 +31,7 @@ import org.deeplearning4j.nn.conf.layers.recurrent.LastTimeStep;
 import org.deeplearning4j.nn.conf.layers.samediff.SameDiffVertex;
 import org.deeplearning4j.nn.conf.memory.MemoryReport;
 import org.deeplearning4j.nn.conf.memory.NetworkMemoryReport;
+import org.deeplearning4j.nn.conf.serde.JsonMappers;
 import org.deeplearning4j.nn.weights.IWeightInit;
 import org.deeplearning4j.nn.weights.WeightInit;
 import org.deeplearning4j.util.OutputLayerUtil;
@@ -40,6 +41,7 @@ import org.nd4j.linalg.activations.IActivation;
 import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.shade.jackson.databind.JsonNode;
 import org.nd4j.shade.jackson.databind.ObjectMapper;
+import org.nd4j.shade.jackson.databind.exc.InvalidTypeIdException;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 
@@ -172,6 +174,26 @@ public class ComputationGraphConfiguration implements Serializable, Cloneable {
         ComputationGraphConfiguration conf;
         try {
             conf = mapper.readValue(json, ComputationGraphConfiguration.class);
+        } catch (InvalidTypeIdException e){
+            if(e.getMessage().contains("@class")){
+                try{
+                    //JSON may be legacy (1.0.0-alpha or earlier), attempt to load it using old format
+                    return JsonMappers.getLegacyMapper().readValue(json, ComputationGraphConfiguration.class);
+                } catch (InvalidTypeIdException e2){
+                    //Check for legacy custom layers: "Could not resolve type id 'CustomLayer' as a subtype of [simple type, class org.deeplearning4j.nn.conf.layers.Layer]: known type ids = [Bidirectional, CenterLossOutputLayer, CnnLossLayer, ..."
+                    //1.0.0-beta5: dropping support for custom layers defined in pre-1.0.0-beta format. Built-in layers from these formats still work
+                    String msg = e2.getMessage();
+                    if(msg != null && msg.contains("Could not resolve type id")){
+                        throw new RuntimeException("Error deserializing ComputationGraphConfiguration - configuration may have a custom " +
+                                "layer, vertex or preprocessor, in pre version 1.0.0-beta JSON format.\nModels in legacy format with custom" +
+                                " layers should be loaded in 1.0.0-beta to 1.0.0-beta4 and saved again, before loading in the current version of DL4J", e);
+                    }
+                    throw new RuntimeException(e2);
+                } catch (IOException e2){
+                    throw new RuntimeException(e2);
+                }
+            }
+            throw new RuntimeException(e);
         } catch (Exception e) {
             //Check if this exception came from legacy deserializer...
             String msg = e.getMessage();
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/InputPreProcessor.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/InputPreProcessor.java
index ab7fd044b..4f02e3d66 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/InputPreProcessor.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/InputPreProcessor.java
@@ -19,7 +19,6 @@ package org.deeplearning4j.nn.conf;
 
 import org.deeplearning4j.nn.api.MaskState;
 import org.deeplearning4j.nn.conf.inputs.InputType;
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyPreprocessorDeserializerHelper;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.primitives.Pair;
 import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
@@ -34,8 +33,7 @@ import java.io.Serializable;
  *
  * @author Adam Gibson
  */
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyPreprocessorDeserializerHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface InputPreProcessor extends Serializable, Cloneable {
 
     /**
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/MultiLayerConfiguration.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/MultiLayerConfiguration.java
index de3373323..52a67ef16 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/MultiLayerConfiguration.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/MultiLayerConfiguration.java
@@ -26,6 +26,7 @@ import org.deeplearning4j.nn.conf.layers.recurrent.LastTimeStep;
 import org.deeplearning4j.nn.conf.memory.LayerMemoryReport;
 import org.deeplearning4j.nn.conf.memory.MemoryReport;
 import org.deeplearning4j.nn.conf.memory.NetworkMemoryReport;
+import org.deeplearning4j.nn.conf.serde.JsonMappers;
 import org.deeplearning4j.nn.layers.recurrent.LastTimeStepLayer;
 import org.deeplearning4j.nn.weights.IWeightInit;
 import org.deeplearning4j.nn.weights.WeightInit;
@@ -41,6 +42,7 @@ import org.nd4j.linalg.lossfunctions.impl.LossMSE;
 import org.nd4j.linalg.lossfunctions.impl.LossNegativeLogLikelihood;
 import org.nd4j.shade.jackson.databind.JsonNode;
 import org.nd4j.shade.jackson.databind.ObjectMapper;
+import org.nd4j.shade.jackson.databind.exc.InvalidTypeIdException;
 import org.nd4j.shade.jackson.databind.node.ArrayNode;
 
 import java.io.IOException;
@@ -157,6 +159,26 @@ public class MultiLayerConfiguration implements Serializable, Cloneable {
         ObjectMapper mapper = NeuralNetConfiguration.mapper();
         try {
             conf = mapper.readValue(json, MultiLayerConfiguration.class);
+        } catch (InvalidTypeIdException e){
+            if(e.getMessage().contains("@class")){
+                try {
+                    //JSON may be legacy (1.0.0-alpha or earlier), attempt to load it using old format
+                    return JsonMappers.getLegacyMapper().readValue(json, MultiLayerConfiguration.class);
+                } catch (InvalidTypeIdException e2){
+                    //Check for legacy custom layers: "Could not resolve type id 'CustomLayer' as a subtype of [simple type, class org.deeplearning4j.nn.conf.layers.Layer]: known type ids = [Bidirectional, CenterLossOutputLayer, CnnLossLayer, ..."
+                    //1.0.0-beta5: dropping support for custom layers defined in pre-1.0.0-beta format. Built-in layers from these formats still work
+                    String msg = e2.getMessage();
+                    if(msg != null && msg.contains("Could not resolve type id")){
+                        throw new RuntimeException("Error deserializing MultiLayerConfiguration - configuration may have a custom " +
+                                "layer, vertex or preprocessor, in pre version 1.0.0-beta JSON format.\nModels in legacy format with custom" +
+                                " layers should be loaded in 1.0.0-beta to 1.0.0-beta4 and saved again, before loading in the current version of DL4J", e);
+                    }
+                    throw new RuntimeException(e2);
+                } catch (IOException e2){
+                    throw new RuntimeException(e2);
+                }
+            }
+            throw new RuntimeException(e);
         } catch (IOException e) {
             //Check if this exception came from legacy deserializer...
             String msg = e.getMessage();
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/NeuralNetConfiguration.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/NeuralNetConfiguration.java
index d8e77a55a..0da5bea13 100755
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/NeuralNetConfiguration.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/NeuralNetConfiguration.java
@@ -26,20 +26,14 @@ import org.deeplearning4j.nn.api.layers.LayerConstraint;
 import org.deeplearning4j.nn.conf.distribution.Distribution;
 import org.deeplearning4j.nn.conf.dropout.Dropout;
 import org.deeplearning4j.nn.conf.dropout.IDropout;
-import org.deeplearning4j.nn.conf.graph.GraphVertex;
 import org.deeplearning4j.nn.conf.inputs.InputType;
 import org.deeplearning4j.nn.conf.layers.*;
 import org.deeplearning4j.nn.conf.layers.misc.FrozenLayer;
 import org.deeplearning4j.nn.conf.layers.misc.FrozenLayerWithBackprop;
 import org.deeplearning4j.nn.conf.layers.recurrent.Bidirectional;
 import org.deeplearning4j.nn.conf.layers.samediff.AbstractSameDiffLayer;
-import org.deeplearning4j.nn.conf.layers.variational.ReconstructionDistribution;
 import org.deeplearning4j.nn.conf.layers.wrapper.BaseWrapperLayer;
 import org.deeplearning4j.nn.conf.serde.JsonMappers;
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyGraphVertexDeserializer;
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyLayerDeserializer;
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyPreprocessorDeserializer;
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyReconstructionDistributionDeserializer;
 import org.deeplearning4j.nn.conf.stepfunctions.StepFunction;
 import org.deeplearning4j.nn.conf.weightnoise.IWeightNoise;
 import org.deeplearning4j.nn.weights.IWeightInit;
@@ -59,10 +53,6 @@ import org.nd4j.linalg.learning.regularization.L1Regularization;
 import org.nd4j.linalg.learning.regularization.L2Regularization;
 import org.nd4j.linalg.learning.regularization.Regularization;
 import org.nd4j.linalg.learning.regularization.WeightDecay;
-import org.nd4j.linalg.lossfunctions.ILossFunction;
-import org.nd4j.linalg.primitives.Pair;
-import org.nd4j.serde.json.LegacyIActivationDeserializer;
-import org.nd4j.serde.json.LegacyILossFunctionDeserializer;
 import org.nd4j.shade.jackson.databind.ObjectMapper;
 
 import java.io.IOException;
@@ -342,9 +332,7 @@ public class NeuralNetConfiguration implements Serializable, Cloneable {
         ObjectMapper mapper = mapper();
 
         try {
-            String ret = mapper.writeValueAsString(this);
-            return ret;
-
+            return mapper.writeValueAsString(this);
         } catch (org.nd4j.shade.jackson.core.JsonProcessingException e) {
             throw new RuntimeException(e);
         }
@@ -384,86 +372,6 @@ public class NeuralNetConfiguration implements Serializable, Cloneable {
         return JsonMappers.getMapper();
     }
 
-    /**
-     * Set of classes that can be registered for legacy deserialization.
-     */
-    private static List<Class<?>> REGISTERABLE_CUSTOM_CLASSES = (List<Class<?>>) Arrays.<Class<?>>asList(
-            Layer.class,
-            GraphVertex.class,
-            InputPreProcessor.class,
-            IActivation.class,
-            ILossFunction.class,
-            ReconstructionDistribution.class
-    );
-
-    /**
-     * Register a set of classes (Layer, GraphVertex, InputPreProcessor, IActivation, ILossFunction, ReconstructionDistribution
-     * ONLY) for JSON deserialization.<br>
-     * <br>
-     * This is required ONLY when BOTH of the following conditions are met:<br>
-     * 1. You want to load a serialized net, saved in 1.0.0-alpha or before, AND<br>
-     * 2. The serialized net has a custom Layer, GraphVertex, etc (i.e., one not defined in DL4J)<br>
-     * <br>
-     * By passing the classes of these layers here, DL4J should be able to deserialize them, in spite of the JSON
-     * format change between versions.
-     *
-     * @param classes Classes to register
-     */
-    public static void registerLegacyCustomClassesForJSON(Class<?>... classes) {
-        registerLegacyCustomClassesForJSONList(Arrays.<Class<?>>asList(classes));
-    }
-
-    /**
-     * @see #registerLegacyCustomClassesForJSON(Class[])
-     */
-    public static void registerLegacyCustomClassesForJSONList(List<Class<?>> classes){
-        //Default names (i.e., old format for custom JSON format)
-        List<Pair<String,Class>> list = new ArrayList<>();
-        for(Class<?> c : classes){
-            list.add(new Pair<String,Class>(c.getSimpleName(), c));
-        }
-        registerLegacyCustomClassesForJSON(list);
-    }
-
-    /**
-     * Register a set of classes (Layer, GraphVertex, InputPreProcessor, IActivation, ILossFunction, ReconstructionDistribution
-     * ONLY) for JSON deserialization, with custom names.<br>
-     * Using this method directly should never be required (instead: use {@link #registerLegacyCustomClassesForJSON(Class[])}
-     * but is added in case it is required in non-standard circumstances.
-     */
-    public static void registerLegacyCustomClassesForJSON(List<Pair<String,Class>> classes){
-        for(Pair<String,Class> p : classes){
-            String s = p.getFirst();
-            Class c = p.getRight();
-            //Check if it's a valid class to register...
-            boolean found = false;
-            for( Class<?> c2 : REGISTERABLE_CUSTOM_CLASSES){
-                if(c2.isAssignableFrom(c)){
-                    if(c2 == Layer.class){
-                        LegacyLayerDeserializer.registerLegacyClassSpecifiedName(s, (Class<? extends Layer>)c);
-                    } else if(c2 == GraphVertex.class){
-                        LegacyGraphVertexDeserializer.registerLegacyClassSpecifiedName(s, (Class<? extends GraphVertex>)c);
-                    } else if(c2 == InputPreProcessor.class){
-                        LegacyPreprocessorDeserializer.registerLegacyClassSpecifiedName(s, (Class<? extends InputPreProcessor>)c);
-                    } else if(c2 == IActivation.class ){
-                        LegacyIActivationDeserializer.registerLegacyClassSpecifiedName(s, (Class<? extends IActivation>)c);
-                    } else if(c2 == ILossFunction.class ){
-                        LegacyILossFunctionDeserializer.registerLegacyClassSpecifiedName(s, (Class<? extends ILossFunction>)c);
-                    } else if(c2 == ReconstructionDistribution.class){
-                        LegacyReconstructionDistributionDeserializer.registerLegacyClassSpecifiedName(s, (Class<? extends ReconstructionDistribution>)c);
-                    }
-
-                    found = true;
-                }
-            }
-
-            if(!found){
-                throw new IllegalArgumentException("Cannot register class for legacy JSON deserialization: class " +
-                        c.getName() + " is not a subtype of classes " + REGISTERABLE_CUSTOM_CLASSES);
-            }
-        }
-    }
-
     /**
      * NeuralNetConfiguration builder, used as a starting point for creating a MultiLayerConfiguration or
      * ComputationGraphConfiguration.<br>
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/graph/AttentionVertex.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/graph/AttentionVertex.java
index 2cd8d83f4..6ca3b35de 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/graph/AttentionVertex.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/graph/AttentionVertex.java
@@ -15,7 +15,7 @@
  ******************************************************************************/
 package org.deeplearning4j.nn.conf.graph;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import lombok.*;
 import org.deeplearning4j.nn.api.MaskState;
 import org.deeplearning4j.nn.conf.inputs.InputType;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/graph/GraphVertex.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/graph/GraphVertex.java
index e4968a49e..497cc77f5 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/graph/GraphVertex.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/graph/GraphVertex.java
@@ -19,7 +19,6 @@ package org.deeplearning4j.nn.conf.graph;
 import org.deeplearning4j.nn.conf.inputs.InputType;
 import org.deeplearning4j.nn.conf.inputs.InvalidInputTypeException;
 import org.deeplearning4j.nn.conf.memory.MemoryReport;
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyGraphVertexDeserializerHelper;
 import org.deeplearning4j.nn.graph.ComputationGraph;
 import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
@@ -33,8 +32,7 @@ import java.io.Serializable;
  *
  * @author Alex Black
  */
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-        defaultImpl = LegacyGraphVertexDeserializerHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public abstract class GraphVertex implements Cloneable, Serializable {
 
     @Override
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/inputs/InputType.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/inputs/InputType.java
index 805e1729c..85da86fa2 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/inputs/InputType.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/inputs/InputType.java
@@ -16,12 +16,15 @@
 
 package org.deeplearning4j.nn.conf.inputs;
 
-import lombok.*;
+import lombok.Data;
+import lombok.EqualsAndHashCode;
+import lombok.Getter;
+import lombok.NoArgsConstructor;
 import org.deeplearning4j.nn.conf.layers.Convolution3D;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.shade.jackson.annotation.JsonIgnore;
 import org.nd4j.shade.jackson.annotation.JsonInclude;
-import org.nd4j.shade.jackson.annotation.JsonSubTypes;
+import org.nd4j.shade.jackson.annotation.JsonProperty;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
 
 import java.io.Serializable;
@@ -36,12 +39,7 @@ import java.util.Arrays;
  * @author Alex Black
  */
 @JsonInclude(JsonInclude.Include.NON_NULL)
-@JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
-@JsonSubTypes(value = {@JsonSubTypes.Type(value = InputType.InputTypeFeedForward.class, name = "FeedForward"),
-        @JsonSubTypes.Type(value = InputType.InputTypeRecurrent.class, name = "Recurrent"),
-        @JsonSubTypes.Type(value = InputType.InputTypeConvolutional.class, name = "Convolutional"),
-        @JsonSubTypes.Type(value = InputType.InputTypeConvolutionalFlat.class, name = "ConvolutionalFlat"),
-        @JsonSubTypes.Type(value = InputType.InputTypeConvolutional3D.class, name = "Convolutional3D")})
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public abstract class InputType implements Serializable {
 
     /**
@@ -174,13 +172,16 @@ public abstract class InputType implements Serializable {
     }
 
 
-    @AllArgsConstructor
-    @Getter
     @NoArgsConstructor
+    @Getter
     @EqualsAndHashCode(callSuper = false)
     public static class InputTypeFeedForward extends InputType {
         private long size;
 
+        public InputTypeFeedForward(@JsonProperty("size") long size) {
+            this.size = size;
+        }
+
         @Override
         public Type getType() {
             return Type.FF;
@@ -203,9 +204,8 @@ public abstract class InputType implements Serializable {
         }
     }
 
-    @Getter
     @NoArgsConstructor
-    @AllArgsConstructor
+    @Getter
     @EqualsAndHashCode(callSuper = false)
     public static class InputTypeRecurrent extends InputType {
         private long size;
@@ -215,6 +215,11 @@ public abstract class InputType implements Serializable {
             this(size, -1);
         }
 
+        public InputTypeRecurrent(@JsonProperty("size") long size, @JsonProperty("timeSeriesLength") long timeSeriesLength) {
+            this.size = size;
+            this.timeSeriesLength = timeSeriesLength;
+        }
+
         @Override
         public Type getType() {
             return Type.RNN;
@@ -245,15 +250,19 @@ public abstract class InputType implements Serializable {
         }
     }
 
-    @AllArgsConstructor
+    @NoArgsConstructor
     @Data
     @EqualsAndHashCode(callSuper = false)
-    @NoArgsConstructor
     public static class InputTypeConvolutional extends InputType {
         private long height;
         private long width;
         private long channels;
 
+        public InputTypeConvolutional(@JsonProperty("height") long height, @JsonProperty("width") long width, @JsonProperty("channels") long channels) {
+            this.height = height;
+            this.width = width;
+            this.channels = channels;
+        }
 
         /**
          * Return the number of channels / depth for this 2D convolution. This method has been deprecated,
@@ -298,10 +307,9 @@ public abstract class InputType implements Serializable {
         }
     }
 
-    @AllArgsConstructor
+    @NoArgsConstructor
     @Data
     @EqualsAndHashCode(callSuper = false)
-    @NoArgsConstructor
     public static class InputTypeConvolutional3D extends InputType {
         private Convolution3D.DataFormat dataFormat;
         private long depth;
@@ -309,6 +317,15 @@ public abstract class InputType implements Serializable {
         private long width;
         private long channels;
 
+        public InputTypeConvolutional3D(@JsonProperty("dataFormat") Convolution3D.DataFormat dataFormat,
+                                        @JsonProperty("depth") long depth, @JsonProperty("height") long height, @JsonProperty("width") long width, @JsonProperty("channels") long channels) {
+            this.dataFormat = dataFormat;
+            this.depth = depth;
+            this.height = height;
+            this.width = width;
+            this.channels = channels;
+        }
+
         @Override
         public Type getType() {
             return Type.CNN3D;
@@ -336,15 +353,20 @@ public abstract class InputType implements Serializable {
         }
     }
 
-    @AllArgsConstructor
+    @NoArgsConstructor
     @Data
     @EqualsAndHashCode(callSuper = false)
-    @NoArgsConstructor
     public static class InputTypeConvolutionalFlat extends InputType {
         private long height;
         private long width;
         private long depth;
 
+        public InputTypeConvolutionalFlat(@JsonProperty("height") long height, @JsonProperty("width") long width, @JsonProperty("depth") long depth) {
+            this.height = height;
+            this.width = width;
+            this.depth = depth;
+        }
+
         @Override
         public Type getType() {
             return Type.CNNFlat;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Layer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Layer.java
index 5dfb3b671..25577bd1f 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Layer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Layer.java
@@ -29,7 +29,6 @@ import org.deeplearning4j.nn.conf.dropout.Dropout;
 import org.deeplearning4j.nn.conf.dropout.IDropout;
 import org.deeplearning4j.nn.conf.inputs.InputType;
 import org.deeplearning4j.nn.conf.memory.LayerMemoryReport;
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyLayerDeserializerHelper;
 import org.deeplearning4j.optimize.api.TrainingListener;
 import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
@@ -45,8 +44,7 @@ import java.util.*;
  * A neural network layer.
  */
 
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-                defaultImpl = LegacyLayerDeserializerHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 @Data
 @NoArgsConstructor
 public abstract class Layer implements TrainingConfig, Serializable, Cloneable {
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Upsampling2D.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Upsampling2D.java
index 1e5eb11f2..0f1a770a8 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Upsampling2D.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Upsampling2D.java
@@ -22,7 +22,7 @@ import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
 import org.deeplearning4j.nn.conf.inputs.InputType;
 import org.deeplearning4j.nn.conf.memory.LayerMemoryReport;
 import org.deeplearning4j.nn.conf.memory.MemoryReport;
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyIntArrayDeserializer;
+import org.deeplearning4j.nn.conf.serde.legacy.LegacyIntArrayDeserializer;
 import org.deeplearning4j.optimize.api.TrainingListener;
 import org.deeplearning4j.util.ValidationUtils;
 import org.nd4j.linalg.api.buffer.DataType;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/misc/FrozenLayer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/misc/FrozenLayer.java
index 7b2317a8f..f72da09e5 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/misc/FrozenLayer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/misc/FrozenLayer.java
@@ -27,7 +27,6 @@ import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
 import org.deeplearning4j.nn.conf.inputs.InputType;
 import org.deeplearning4j.nn.conf.layers.Layer;
 import org.deeplearning4j.nn.conf.memory.LayerMemoryReport;
-import org.deeplearning4j.nn.conf.serde.FrozenLayerDeserializer;
 import org.deeplearning4j.nn.params.FrozenLayerParamInitializer;
 import org.deeplearning4j.optimize.api.TrainingListener;
 import org.nd4j.linalg.api.buffer.DataType;
@@ -48,7 +47,6 @@ import java.util.List;
  * @author Alex Black
  */
 @EqualsAndHashCode(callSuper = false)
-@JsonDeserialize(using = FrozenLayerDeserializer.class)
 public class FrozenLayer extends Layer {
 
     @Getter
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/variational/ReconstructionDistribution.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/variational/ReconstructionDistribution.java
index 0de9143fe..f58e6b0e7 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/variational/ReconstructionDistribution.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/variational/ReconstructionDistribution.java
@@ -16,7 +16,6 @@
 
 package org.deeplearning4j.nn.conf.layers.variational;
 
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyReconstructionDistributionDeserializerHelper;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.shade.jackson.annotation.JsonSubTypes;
 import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
@@ -32,8 +31,7 @@ import java.io.Serializable;
  *
  * @author Alex Black
  */
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class",
-                defaultImpl = LegacyReconstructionDistributionDeserializerHelper.class)
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
 public interface ReconstructionDistribution extends Serializable {
 
     /**
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/BaseNetConfigDeserializer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/BaseNetConfigDeserializer.java
index 5194da4a1..d32488363 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/BaseNetConfigDeserializer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/BaseNetConfigDeserializer.java
@@ -21,12 +21,18 @@ import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
 import org.deeplearning4j.nn.conf.Updater;
 import org.deeplearning4j.nn.conf.distribution.Distribution;
 import org.deeplearning4j.nn.conf.layers.BaseLayer;
+import org.deeplearning4j.nn.conf.layers.BaseOutputLayer;
 import org.deeplearning4j.nn.conf.layers.Layer;
 import org.deeplearning4j.nn.weights.*;
+import org.nd4j.linalg.activations.Activation;
+import org.nd4j.linalg.activations.IActivation;
+import org.nd4j.linalg.activations.impl.*;
 import org.nd4j.linalg.learning.config.*;
 import org.nd4j.linalg.learning.regularization.L1Regularization;
 import org.nd4j.linalg.learning.regularization.Regularization;
 import org.nd4j.linalg.learning.regularization.WeightDecay;
+import org.nd4j.linalg.lossfunctions.ILossFunction;
+import org.nd4j.linalg.lossfunctions.impl.*;
 import org.nd4j.shade.jackson.core.JsonParser;
 import org.nd4j.shade.jackson.core.JsonProcessingException;
 import org.nd4j.shade.jackson.databind.DeserializationContext;
@@ -38,6 +44,8 @@ import org.nd4j.shade.jackson.databind.node.ObjectNode;
 
 import java.io.IOException;
 import java.util.ArrayList;
+import java.util.HashMap;
+import java.util.Map;
 
 /**
  * A custom (abstract) deserializer that handles backward compatibility (currently only for updater refactoring that
@@ -103,6 +111,24 @@ public abstract class BaseNetConfigDeserializer<T> extends StdDeserializer<T> im
         return false;
     }
 
+    protected boolean requiresActivationFromLegacy(Layer[] layers){
+        for(Layer l : layers){
+            if(l instanceof BaseLayer && ((BaseLayer)l).getActivationFn() == null){
+                return true;
+            }
+        }
+        return false;
+    }
+
+    protected boolean requiresLegacyLossHandling(Layer[] layers){
+        for(Layer l : layers){
+            if(l instanceof BaseOutputLayer && ((BaseOutputLayer)l).getLossFn() == null){
+                return true;
+            }
+        }
+        return false;
+    }
+
     protected void handleUpdaterBackwardCompatibility(BaseLayer layer, ObjectNode on){
         if(on != null && on.has("updater")){
             String updaterName = on.get("updater").asText();
@@ -220,7 +246,7 @@ public abstract class BaseNetConfigDeserializer<T> extends StdDeserializer<T> im
     }
 
     protected void handleWeightInitBackwardCompatibility(BaseLayer baseLayer, ObjectNode on){
-        if(on != null && (on.has("weightInit") )){
+        if(on != null && on.has("weightInit") ){
             //Legacy format JSON
             if(on.has("weightInit")){
                 String wi = on.get("weightInit").asText();
@@ -228,8 +254,7 @@ public abstract class BaseNetConfigDeserializer<T> extends StdDeserializer<T> im
                     WeightInit w = WeightInit.valueOf(wi);
                     Distribution d = null;
                     if(w == WeightInit.DISTRIBUTION && on.has("dist")){
-                        //TODO deserialize distribution
-                        String dist = on.get("dist").asText();
+                        String dist = on.get("dist").toString();
                         d = NeuralNetConfiguration.mapper().readValue(dist, Distribution.class);
                     }
                     IWeightInit iwi = w.getWeightInitFunction(d);
@@ -241,6 +266,57 @@ public abstract class BaseNetConfigDeserializer<T> extends StdDeserializer<T> im
         }
     }
 
+    //Changed after 0.7.1 from "activationFunction" : "softmax" to "activationFn" : <object>
+    protected void handleActivationBackwardCompatibility(BaseLayer baseLayer, ObjectNode on){
+
+        if(baseLayer.getActivationFn() == null && on.has("activationFunction")){
+            String afn = on.get("activationFunction").asText();
+            IActivation a = null;
+            try {
+                a = getMap().get(afn.toLowerCase()).newInstance();
+            } catch (InstantiationException | IllegalAccessException e){
+                //Ignore
+            }
+            baseLayer.setActivationFn(a);
+        }
+    }
+
+    //0.5.0 and earlier: loss function was an enum like "lossFunction" : "NEGATIVELOGLIKELIHOOD",
+    protected void handleLossBackwardCompatibility(BaseOutputLayer baseLayer, ObjectNode on){
+        if(baseLayer.getLossFn() == null && on.has("activationFunction")) {
+            String lfn = on.get("lossFunction").asText();
+            ILossFunction loss = null;
+            switch (lfn) {
+                case "MCXENT":
+                    loss = new LossMCXENT();
+                    break;
+                case "MSE":
+                    loss = new LossMSE();
+                    break;
+                case "NEGATIVELOGLIKELIHOOD":
+                    loss = new LossNegativeLogLikelihood();
+                    break;
+                case "SQUARED_LOSS":
+                    loss = new LossL2();
+                    break;
+                case "XENT":
+                    loss = new LossBinaryXENT();
+            }
+            baseLayer.setLossFn(loss);
+        }
+    }
+
+    private static Map<String,Class<? extends IActivation>> activationMap;
+    private static synchronized Map<String,Class<? extends IActivation>> getMap(){
+        if(activationMap == null){
+            activationMap = new HashMap<>();
+            for(Activation a : Activation.values()){
+                activationMap.put(a.toString().toLowerCase(), a.getActivationFunction().getClass());
+            }
+        }
+        return activationMap;
+    }
+
     @Override
     public void resolve(DeserializationContext ctxt) throws JsonMappingException {
         ((ResolvableDeserializer) defaultDeserializer).resolve(ctxt);
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/ComputationGraphConfigurationDeserializer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/ComputationGraphConfigurationDeserializer.java
index 1c30b053b..50384e518 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/ComputationGraphConfigurationDeserializer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/ComputationGraphConfigurationDeserializer.java
@@ -23,6 +23,7 @@ import org.deeplearning4j.nn.conf.dropout.Dropout;
 import org.deeplearning4j.nn.conf.graph.GraphVertex;
 import org.deeplearning4j.nn.conf.graph.LayerVertex;
 import org.deeplearning4j.nn.conf.layers.BaseLayer;
+import org.deeplearning4j.nn.conf.layers.BaseOutputLayer;
 import org.deeplearning4j.nn.conf.layers.BatchNormalization;
 import org.deeplearning4j.nn.conf.layers.Layer;
 import org.deeplearning4j.nn.conf.weightnoise.DropConnect;
@@ -74,6 +75,8 @@ public class ComputationGraphConfigurationDeserializer
         boolean attemptIUpdaterFromLegacy = requiresIUpdaterFromLegacy(layers);
         boolean requireLegacyRegularizationHandling = requiresRegularizationFromLegacy(layers);
         boolean requiresLegacyWeightInitHandling = requiresWeightInitFromLegacy(layers);
+        boolean requiresLegacyActivationHandling = requiresActivationFromLegacy(layers);
+        boolean requiresLegacyLossHandling = requiresLegacyLossHandling(layers);
 
         Long charOffsetEnd = null;
         JsonLocation endLocation = null;
@@ -123,6 +126,14 @@ public class ComputationGraphConfigurationDeserializer
                         handleWeightInitBackwardCompatibility((BaseLayer)layers[layerIdx], (ObjectNode)next);
                     }
 
+                    if(requiresLegacyActivationHandling && layers[layerIdx] instanceof BaseLayer && ((BaseLayer)layers[layerIdx]).getActivationFn() == null){
+                        handleActivationBackwardCompatibility((BaseLayer)layers[layerIdx], (ObjectNode)next);
+                    }
+
+                    if(requiresLegacyLossHandling && layers[layerIdx] instanceof BaseOutputLayer && ((BaseOutputLayer)layers[layerIdx]).getLossFn() == null){
+                        handleLossBackwardCompatibility((BaseOutputLayer) layers[layerIdx],  (ObjectNode)next);
+                    }
+
                     if(layers[layerIdx].getIDropout() == null){
                         //Check for legacy dropout
                         if(next.has("dropOut")){
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/FrozenLayerDeserializer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/FrozenLayerDeserializer.java
deleted file mode 100644
index 0eb618e20..000000000
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/FrozenLayerDeserializer.java
+++ /dev/null
@@ -1,58 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.nn.conf.serde;
-
-import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
-import org.deeplearning4j.nn.conf.layers.Layer;
-import org.deeplearning4j.nn.conf.layers.misc.FrozenLayer;
-import org.deeplearning4j.nn.conf.serde.legacyformat.LegacyLayerDeserializer;
-import org.nd4j.shade.jackson.core.JsonFactory;
-import org.nd4j.shade.jackson.core.JsonParser;
-import org.nd4j.shade.jackson.databind.DeserializationContext;
-import org.nd4j.shade.jackson.databind.JsonDeserializer;
-import org.nd4j.shade.jackson.databind.JsonNode;
-
-import java.io.IOException;
-
-/**
- * A custom deserializer for handling Frozen layers
- * This is used to handle the 2 different Layer JSON formats - old/legacy, and current
- *
- * @author Alex Black
- */
-public class FrozenLayerDeserializer extends JsonDeserializer<Layer> {
-    @Override
-    public Layer deserialize(JsonParser jp, DeserializationContext deserializationContext) throws IOException {
-        JsonNode n = jp.getCodec().readTree(jp);
-        JsonNode layer = n.get("layer");
-        boolean newFormat = layer.has("@class");
-
-        String internalText = layer.toString();
-        Layer internal;
-        if(newFormat){
-            //Standard/new format
-            internal = NeuralNetConfiguration.mapper().readValue(internalText, Layer.class);
-        } else {
-            //Legacy format
-            JsonFactory factory = new JsonFactory();
-            JsonParser  parser  = factory.createParser(internalText);
-            parser.setCodec(jp.getCodec());
-            internal = new LegacyLayerDeserializer().deserialize(parser, deserializationContext);
-        }
-        return new FrozenLayer(internal);
-    }
-}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/JsonMappers.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/JsonMappers.java
index a4e29c86e..1cdd85d4b 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/JsonMappers.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/JsonMappers.java
@@ -27,6 +27,7 @@ import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
 import org.deeplearning4j.nn.conf.graph.GraphVertex;
 import org.deeplearning4j.nn.conf.layers.Layer;
 import org.deeplearning4j.nn.conf.layers.variational.ReconstructionDistribution;
+import org.deeplearning4j.nn.conf.serde.legacy.LegacyJsonFormat;
 import org.nd4j.linalg.activations.IActivation;
 import org.nd4j.linalg.lossfunctions.ILossFunction;
 import org.nd4j.serde.json.LegacyIActivationDeserializer;
@@ -42,8 +43,10 @@ import org.nd4j.shade.jackson.databind.introspect.JacksonAnnotationIntrospector;
 import org.nd4j.shade.jackson.databind.jsontype.TypeResolverBuilder;
 import org.nd4j.shade.jackson.databind.module.SimpleModule;
 import org.nd4j.shade.jackson.dataformat.yaml.YAMLFactory;
+import org.nd4j.util.OneTimeLogger;
 
 import java.lang.annotation.Annotation;
+import java.lang.reflect.Field;
 import java.util.ArrayList;
 import java.util.Collections;
 import java.util.List;
@@ -56,93 +59,14 @@ import java.util.List;
 @Slf4j
 public class JsonMappers {
 
-    /**
-     * @deprecated Use {@link DL4JSystemProperties#CUSTOM_REGISTRATION_PROPERTY}
-     */
-    @Deprecated
-    public static String CUSTOM_REGISTRATION_PROPERTY = DL4JSystemProperties.CUSTOM_REGISTRATION_PROPERTY;
-
-    static {
-        String p = System.getProperty(DL4JSystemProperties.CUSTOM_REGISTRATION_PROPERTY);
-        if(p != null && !p.isEmpty()){
-            String[] split = p.split(",");
-            List<Class<?>> list = new ArrayList<>();
-            for(String s : split){
-                try{
-                    Class<?> c = Class.forName(s);
-                    list.add(c);
-                } catch (Throwable t){
-                    log.warn("Error parsing {} system property: class \"{}\" could not be loaded",DL4JSystemProperties.CUSTOM_REGISTRATION_PROPERTY, s, t);
-                }
-            }
-
-            if(list.size() > 0){
-                try {
-                    NeuralNetConfiguration.registerLegacyCustomClassesForJSONList(list);
-                } catch (Throwable t){
-                    log.warn("Error registering custom classes for legacy JSON deserialization ({} system property)",DL4JSystemProperties.CUSTOM_REGISTRATION_PROPERTY, t);
-                }
-            }
-        }
-    }
-
     private static ObjectMapper jsonMapper = new ObjectMapper();
     private static ObjectMapper yamlMapper = new ObjectMapper(new YAMLFactory());
 
-    /*
-    Note to developers: The following JSON mappers are for handling legacy format JSON.
-    Note that after 1.0.0-alpha, the JSON subtype format for layers, preprocessors, graph vertices,
-    etc were changed from a wrapper object, to an "@class" field.
-    However, in an attempt to not break saved networks, these mappers are part of the solution.
-
-    How legacy loading works (same pattern for all types - Layer, GraphVertex, InputPreprocesor etc)
-    1. Layers etc that have an "@class" field are deserialized as normal
-    2. Layers that don't have such a field are mapped (via Layer @JsonTypeInfo) to the LegacyLayerDeserializerHelper class.
-    3. LegacyLayerDeserializerHelper has a @JsonDeserialize annotation - we use LegacyLayerDeserialize to handle it
-    4. LegacyLayerDeserializer has a list of old names (present in the legacy format JSON) and the corresponding class names
-    5. BaseLegacyDeserializer (that LegacyLayerDeserializer extends) does a lookup and handles the deserialization
-
-    Now, as to why we have one ObjectMapper for each type: We can't use the default JSON mapper for the legacy format,
-    as it'll fail due to not having the expected "@class" annotation.
-    Consequently, we need to tell Jackson to ignore that specific annotation and deserialize to the specified
-    class anyway. The ignoring is done via an annotation introspector, defined below in this class.
-    However, we can't just use a single annotation introspector (and hence ObjectMapper) for loading legacy values of
-    all types - if we did, then any nested types would fail (i.e., an IActivation in a Layer - the IActivation couldn't
-    be deserialized correctly, as the annotation would be ignored).
-
-     */
-    @Getter
-    private static ObjectMapper jsonMapperLegacyFormatLayer = new ObjectMapper();
-    @Getter
-    private static ObjectMapper jsonMapperLegacyFormatVertex = new ObjectMapper();
-    @Getter
-    private static ObjectMapper jsonMapperLegacyFormatPreproc = new ObjectMapper();
-    @Getter
-    private static ObjectMapper jsonMapperLegacyFormatIActivation = new ObjectMapper();
-    @Getter
-    private static ObjectMapper jsonMapperLegacyFormatILoss = new ObjectMapper();
-    @Getter
-    private static ObjectMapper jsonMapperLegacyFormatReconstruction = new ObjectMapper();
+    private static ObjectMapper legacyMapper;
 
     static {
         configureMapper(jsonMapper);
         configureMapper(yamlMapper);
-        configureMapper(jsonMapperLegacyFormatLayer);
-        configureMapper(jsonMapperLegacyFormatVertex);
-        configureMapper(jsonMapperLegacyFormatPreproc);
-        configureMapper(jsonMapperLegacyFormatIActivation);
-        configureMapper(jsonMapperLegacyFormatILoss);
-        configureMapper(jsonMapperLegacyFormatReconstruction);
-
-        jsonMapperLegacyFormatLayer.setAnnotationIntrospector(new IgnoreJsonTypeInfoIntrospector(Collections.<Class>singletonList(Layer.class)));
-        jsonMapperLegacyFormatVertex.setAnnotationIntrospector(new IgnoreJsonTypeInfoIntrospector(Collections.<Class>singletonList(GraphVertex.class)));
-        jsonMapperLegacyFormatPreproc.setAnnotationIntrospector(new IgnoreJsonTypeInfoIntrospector(Collections.<Class>singletonList(InputPreProcessor.class)));
-        jsonMapperLegacyFormatIActivation.setAnnotationIntrospector(new IgnoreJsonTypeInfoIntrospector(Collections.<Class>singletonList(IActivation.class)));
-        jsonMapperLegacyFormatILoss.setAnnotationIntrospector(new IgnoreJsonTypeInfoIntrospector(Collections.<Class>singletonList(ILossFunction.class)));
-        jsonMapperLegacyFormatReconstruction.setAnnotationIntrospector(new IgnoreJsonTypeInfoIntrospector(Collections.<Class>singletonList(ReconstructionDistribution.class)));
-
-        LegacyIActivationDeserializer.setLegacyJsonMapper(jsonMapperLegacyFormatIActivation);
-        LegacyILossFunctionDeserializer.setLegacyJsonMapper(jsonMapperLegacyFormatILoss);
     }
 
     /**
@@ -152,6 +76,14 @@ public class JsonMappers {
         return jsonMapper;
     }
 
+    public static synchronized ObjectMapper getLegacyMapper(){
+        if(legacyMapper == null){
+            legacyMapper = LegacyJsonFormat.getMapper100alpha();
+            configureMapper(legacyMapper);
+        }
+        return legacyMapper;
+    }
+
     /**
      * @return The default/primary ObjectMapper for deserializing network configurations in DL4J (YAML format)
      */
@@ -182,60 +114,4 @@ public class JsonMappers {
 
         ret.registerModule(customDeserializerModule);
     }
-
-
-    /**
-     * Custom Jackson Introspector to ignore the {@code @JsonTypeYnfo} annotations on layers etc.
-     * This is so we can deserialize legacy format JSON without recursing infinitely, by selectively ignoring
-     * a set of JsonTypeInfo annotations
-     */
-    @AllArgsConstructor
-    private static class IgnoreJsonTypeInfoIntrospector extends JacksonAnnotationIntrospector {
-
-        private List<Class> classList;
-
-        @Override
-        protected TypeResolverBuilder<?> _findTypeResolver(MapperConfig<?> config, Annotated ann, JavaType baseType) {
-            if(ann instanceof AnnotatedClass){
-                AnnotatedClass c = (AnnotatedClass)ann;
-                Class<?> annClass = c.getAnnotated();
-
-                boolean isAssignable = false;
-                for(Class c2 : classList){
-                    if(c2.isAssignableFrom(annClass)){
-                        isAssignable = true;
-                        break;
-                    }
-                }
-
-                if( isAssignable ){
-                    AnnotationMap annotations = (AnnotationMap) ((AnnotatedClass) ann).getAnnotations();
-                    if(annotations == null || annotations.annotations() == null){
-                        //Probably not necessary - but here for safety
-                        return super._findTypeResolver(config, ann, baseType);
-                    }
-
-                    AnnotationMap newMap = null;
-                    for(Annotation a : annotations.annotations()){
-                        Class<?> annType = a.annotationType();
-                        if(annType == JsonTypeInfo.class){
-                            //Ignore the JsonTypeInfo annotation on the Layer class
-                            continue;
-                        }
-                        if(newMap == null){
-                            newMap = new AnnotationMap();
-                        }
-                        newMap.add(a);
-                    }
-                    if(newMap == null)
-                        return null;
-
-                    //Pass the remaining annotations (if any) to the original introspector
-                    AnnotatedClass ann2 = c.withAnnotations(newMap);
-                    return super._findTypeResolver(config, ann2, baseType);
-                }
-            }
-            return super._findTypeResolver(config, ann, baseType);
-        }
-    }
 }
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/MultiLayerConfigurationDeserializer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/MultiLayerConfigurationDeserializer.java
index e7bcdc39a..028fef9d3 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/MultiLayerConfigurationDeserializer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/MultiLayerConfigurationDeserializer.java
@@ -21,6 +21,7 @@ import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
 import org.deeplearning4j.nn.conf.dropout.Dropout;
 import org.deeplearning4j.nn.conf.layers.BaseLayer;
+import org.deeplearning4j.nn.conf.layers.BaseOutputLayer;
 import org.deeplearning4j.nn.conf.layers.BatchNormalization;
 import org.deeplearning4j.nn.conf.layers.Layer;
 import org.deeplearning4j.nn.conf.weightnoise.DropConnect;
@@ -59,6 +60,8 @@ public class MultiLayerConfigurationDeserializer extends BaseNetConfigDeserializ
 
         boolean requiresLegacyRegularizationHandling = requiresRegularizationFromLegacy(layers);
         boolean requiresLegacyWeightInitHandling = requiresWeightInitFromLegacy(layers);
+        boolean requiresLegacyActivationHandling = requiresActivationFromLegacy(layers);
+        boolean requiresLegacyLossHandling = requiresLegacyLossHandling(layers);
 
         if(attemptIUpdaterFromLegacy || requiresLegacyRegularizationHandling || requiresLegacyWeightInitHandling) {
             JsonLocation endLocation = jp.getCurrentLocation();
@@ -115,38 +118,34 @@ public class MultiLayerConfigurationDeserializer extends BaseNetConfigDeserializ
                     }
                 }
 
-                if(requiresLegacyRegularizationHandling) {
-                    if (layers[i] instanceof BaseLayer && ((BaseLayer) layers[i]).getRegularization() == null) {
-                        if(on.has("layer")){
-                            //Legacy format
-                            ObjectNode layerNode = (ObjectNode)on.get("layer");
-                            if(layerNode.has("@class")){
-                                //Later legacy format: class field for JSON subclass
-                                on = layerNode;
-                            } else {
-                                //Early legacy format: wrapper object for JSON subclass
-                                on = (ObjectNode) on.get("layer").elements().next();
-                            }
+                if(requiresLegacyRegularizationHandling || requiresLegacyWeightInitHandling || requiresLegacyActivationHandling){
+                    if(on.has("layer")){
+                        //Legacy format
+                        ObjectNode layerNode = (ObjectNode)on.get("layer");
+                        if(layerNode.has("@class")){
+                            //Later legacy format: class field for JSON subclass
+                            on = layerNode;
+                        } else {
+                            //Early legacy format: wrapper object for JSON subclass
+                            on = (ObjectNode) on.get("layer").elements().next();
                         }
-                        handleL1L2BackwardCompatibility((BaseLayer) layers[i], on);
                     }
                 }
 
-                if(requiresLegacyWeightInitHandling){
-                    if (layers[i] instanceof BaseLayer && ((BaseLayer) layers[i]).getWeightInitFn() == null) {
-                        if(on.has("layer")){
-                            //Legacy format
-                            ObjectNode layerNode = (ObjectNode)on.get("layer");
-                            if(layerNode.has("@class")){
-                                //Later legacy format: class field for JSON subclass
-                                on = layerNode;
-                            } else {
-                                //Early legacy format: wrapper object for JSON subclass
-                                on = (ObjectNode) on.get("layer").elements().next();
-                            }
-                        }
-                        handleWeightInitBackwardCompatibility((BaseLayer) layers[i], on);
-                    }
+                if(requiresLegacyRegularizationHandling && layers[i] instanceof BaseLayer && ((BaseLayer) layers[i]).getRegularization() == null) {
+                    handleL1L2BackwardCompatibility((BaseLayer) layers[i], on);
+                }
+
+                if(requiresLegacyWeightInitHandling && layers[i] instanceof BaseLayer && ((BaseLayer) layers[i]).getWeightInitFn() == null) {
+                    handleWeightInitBackwardCompatibility((BaseLayer) layers[i], on);
+                }
+
+                if(requiresLegacyActivationHandling && layers[i] instanceof BaseLayer && ((BaseLayer)layers[i]).getActivationFn() == null){
+                    handleActivationBackwardCompatibility((BaseLayer) layers[i], on);
+                }
+
+                if(requiresLegacyLossHandling && layers[i] instanceof BaseOutputLayer && ((BaseOutputLayer)layers[i]).getLossFn() == null){
+                    handleLossBackwardCompatibility((BaseOutputLayer) layers[i], on);
                 }
             }
         }
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyIntArrayDeserializer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacy/LegacyIntArrayDeserializer.java
similarity index 97%
rename from deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyIntArrayDeserializer.java
rename to deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacy/LegacyIntArrayDeserializer.java
index e080d6a78..064219fd1 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyIntArrayDeserializer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacy/LegacyIntArrayDeserializer.java
@@ -14,7 +14,7 @@
  * SPDX-License-Identifier: Apache-2.0
  ******************************************************************************/
 
-package org.deeplearning4j.nn.conf.serde.legacyformat;
+package org.deeplearning4j.nn.conf.serde.legacy;
 
 import org.nd4j.shade.jackson.core.JsonParser;
 import org.nd4j.shade.jackson.core.JsonProcessingException;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacy/LegacyJsonFormat.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacy/LegacyJsonFormat.java
new file mode 100644
index 000000000..e421c4b1f
--- /dev/null
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacy/LegacyJsonFormat.java
@@ -0,0 +1,175 @@
+package org.deeplearning4j.nn.conf.serde.legacy;
+
+import lombok.AccessLevel;
+import lombok.NoArgsConstructor;
+import org.deeplearning4j.nn.conf.InputPreProcessor;
+import org.deeplearning4j.nn.conf.graph.*;
+import org.deeplearning4j.nn.conf.graph.rnn.DuplicateToTimeSeriesVertex;
+import org.deeplearning4j.nn.conf.graph.rnn.LastTimeStepVertex;
+import org.deeplearning4j.nn.conf.graph.rnn.ReverseTimeSeriesVertex;
+import org.deeplearning4j.nn.conf.layers.*;
+import org.deeplearning4j.nn.conf.layers.convolutional.Cropping1D;
+import org.deeplearning4j.nn.conf.layers.convolutional.Cropping2D;
+import org.deeplearning4j.nn.conf.layers.misc.ElementWiseMultiplicationLayer;
+import org.deeplearning4j.nn.conf.layers.misc.FrozenLayer;
+import org.deeplearning4j.nn.conf.layers.objdetect.Yolo2OutputLayer;
+import org.deeplearning4j.nn.conf.layers.recurrent.Bidirectional;
+import org.deeplearning4j.nn.conf.layers.recurrent.SimpleRnn;
+import org.deeplearning4j.nn.conf.layers.util.MaskLayer;
+import org.deeplearning4j.nn.conf.layers.util.MaskZeroLayer;
+import org.deeplearning4j.nn.conf.layers.variational.*;
+import org.deeplearning4j.nn.conf.preprocessor.*;
+import org.nd4j.linalg.activations.IActivation;
+import org.nd4j.linalg.activations.impl.*;
+import org.nd4j.linalg.lossfunctions.ILossFunction;
+import org.nd4j.linalg.lossfunctions.impl.*;
+import org.nd4j.shade.jackson.annotation.JsonSubTypes;
+import org.nd4j.shade.jackson.annotation.JsonTypeInfo;
+import org.nd4j.shade.jackson.databind.ObjectMapper;
+
+/**
+ * This class defines a set of Jackson Mixins - which are a way of using a proxy class with annotations to override
+ * the existing annotations.
+ * In 1.0.0-beta, we switched how subtypes were handled in JSON ser/de: from "wrapper object" to "@class field".
+ * We use these mixins to allow us to still load the old format
+ *
+ * @author Alex Black
+ */
+public class LegacyJsonFormat {
+
+    private LegacyJsonFormat(){ }
+
+    /**
+     * Get a mapper (minus general config) suitable for loading old format JSON - 1.0.0-alpha and before
+     * @return Object mapper
+     */
+    public static ObjectMapper getMapper100alpha(){
+        //After 1.0.0-alpha, we switched from wrapper object to @class for subtype information
+        ObjectMapper om = new ObjectMapper();
+
+        om.addMixIn(InputPreProcessor.class, InputPreProcessorMixin.class);
+        om.addMixIn(GraphVertex.class, GraphVertexMixin.class);
+        om.addMixIn(Layer.class, LayerMixin.class);
+        om.addMixIn(ReconstructionDistribution.class, ReconstructionDistributionMixin.class);
+        om.addMixIn(IActivation.class, IActivationMixin.class);
+        om.addMixIn(ILossFunction.class, ILossFunctionMixin.class);
+
+        return om;
+    }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = CnnToFeedForwardPreProcessor.class, name = "cnnToFeedForward"),
+            @JsonSubTypes.Type(value = CnnToRnnPreProcessor.class, name = "cnnToRnn"),
+            @JsonSubTypes.Type(value = ComposableInputPreProcessor.class, name = "composableInput"),
+            @JsonSubTypes.Type(value = FeedForwardToCnnPreProcessor.class, name = "feedForwardToCnn"),
+            @JsonSubTypes.Type(value = FeedForwardToRnnPreProcessor.class, name = "feedForwardToRnn"),
+            @JsonSubTypes.Type(value = RnnToFeedForwardPreProcessor.class, name = "rnnToFeedForward"),
+            @JsonSubTypes.Type(value = RnnToCnnPreProcessor.class, name = "rnnToCnn")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class InputPreProcessorMixin { }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = ElementWiseVertex.class, name = "ElementWiseVertex"),
+            @JsonSubTypes.Type(value = MergeVertex.class, name = "MergeVertex"),
+            @JsonSubTypes.Type(value = SubsetVertex.class, name = "SubsetVertex"),
+            @JsonSubTypes.Type(value = LayerVertex.class, name = "LayerVertex"),
+            @JsonSubTypes.Type(value = LastTimeStepVertex.class, name = "LastTimeStepVertex"),
+            @JsonSubTypes.Type(value = ReverseTimeSeriesVertex.class, name = "ReverseTimeSeriesVertex"),
+            @JsonSubTypes.Type(value = DuplicateToTimeSeriesVertex.class, name = "DuplicateToTimeSeriesVertex"),
+            @JsonSubTypes.Type(value = PreprocessorVertex.class, name = "PreprocessorVertex"),
+            @JsonSubTypes.Type(value = StackVertex.class, name = "StackVertex"),
+            @JsonSubTypes.Type(value = UnstackVertex.class, name = "UnstackVertex"),
+            @JsonSubTypes.Type(value = L2Vertex.class, name = "L2Vertex"),
+            @JsonSubTypes.Type(value = ScaleVertex.class, name = "ScaleVertex"),
+            @JsonSubTypes.Type(value = L2NormalizeVertex.class, name = "L2NormalizeVertex")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class GraphVertexMixin{ }
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = AutoEncoder.class, name = "autoEncoder"),
+            @JsonSubTypes.Type(value = ConvolutionLayer.class, name = "convolution"),
+            @JsonSubTypes.Type(value = Convolution1DLayer.class, name = "convolution1d"),
+            @JsonSubTypes.Type(value = GravesLSTM.class, name = "gravesLSTM"),
+            @JsonSubTypes.Type(value = LSTM.class, name = "LSTM"),
+            @JsonSubTypes.Type(value = GravesBidirectionalLSTM.class, name = "gravesBidirectionalLSTM"),
+            @JsonSubTypes.Type(value = OutputLayer.class, name = "output"),
+            @JsonSubTypes.Type(value = CenterLossOutputLayer.class, name = "CenterLossOutputLayer"),
+            @JsonSubTypes.Type(value = RnnOutputLayer.class, name = "rnnoutput"),
+            @JsonSubTypes.Type(value = LossLayer.class, name = "loss"),
+            @JsonSubTypes.Type(value = DenseLayer.class, name = "dense"),
+            @JsonSubTypes.Type(value = SubsamplingLayer.class, name = "subsampling"),
+            @JsonSubTypes.Type(value = Subsampling1DLayer.class, name = "subsampling1d"),
+            @JsonSubTypes.Type(value = BatchNormalization.class, name = "batchNormalization"),
+            @JsonSubTypes.Type(value = LocalResponseNormalization.class, name = "localResponseNormalization"),
+            @JsonSubTypes.Type(value = EmbeddingLayer.class, name = "embedding"),
+            @JsonSubTypes.Type(value = ActivationLayer.class, name = "activation"),
+            @JsonSubTypes.Type(value = VariationalAutoencoder.class, name = "VariationalAutoencoder"),
+            @JsonSubTypes.Type(value = DropoutLayer.class, name = "dropout"),
+            @JsonSubTypes.Type(value = GlobalPoolingLayer.class, name = "GlobalPooling"),
+            @JsonSubTypes.Type(value = ZeroPaddingLayer.class, name = "zeroPadding"),
+            @JsonSubTypes.Type(value = ZeroPadding1DLayer.class, name = "zeroPadding1d"),
+            @JsonSubTypes.Type(value = FrozenLayer.class, name = "FrozenLayer"),
+            @JsonSubTypes.Type(value = Upsampling2D.class, name = "Upsampling2D"),
+            @JsonSubTypes.Type(value = Yolo2OutputLayer.class, name = "Yolo2OutputLayer"),
+            @JsonSubTypes.Type(value = RnnLossLayer.class, name = "RnnLossLayer"),
+            @JsonSubTypes.Type(value = CnnLossLayer.class, name = "CnnLossLayer"),
+            @JsonSubTypes.Type(value = Bidirectional.class, name = "Bidirectional"),
+            @JsonSubTypes.Type(value = SimpleRnn.class, name = "SimpleRnn"),
+            @JsonSubTypes.Type(value = ElementWiseMultiplicationLayer.class, name = "ElementWiseMult"),
+            @JsonSubTypes.Type(value = MaskLayer.class, name = "MaskLayer"),
+            @JsonSubTypes.Type(value = MaskZeroLayer.class, name = "MaskZeroLayer"),
+            @JsonSubTypes.Type(value = Cropping1D.class, name = "Cropping1D"),
+            @JsonSubTypes.Type(value = Cropping2D.class, name = "Cropping2D")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class LayerMixin {}
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = GaussianReconstructionDistribution.class, name = "Gaussian"),
+            @JsonSubTypes.Type(value = BernoulliReconstructionDistribution.class, name = "Bernoulli"),
+            @JsonSubTypes.Type(value = ExponentialReconstructionDistribution.class, name = "Exponential"),
+            @JsonSubTypes.Type(value = CompositeReconstructionDistribution.class, name = "Composite"),
+            @JsonSubTypes.Type(value = LossFunctionWrapper.class, name = "LossWrapper")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class ReconstructionDistributionMixin {}
+
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = ActivationCube.class, name = "Cube"),
+            @JsonSubTypes.Type(value = ActivationELU.class, name = "ELU"),
+            @JsonSubTypes.Type(value = ActivationHardSigmoid.class, name = "HardSigmoid"),
+            @JsonSubTypes.Type(value = ActivationHardTanH.class, name = "HardTanh"),
+            @JsonSubTypes.Type(value = ActivationIdentity.class, name = "Identity"),
+            @JsonSubTypes.Type(value = ActivationLReLU.class, name = "LReLU"),
+            @JsonSubTypes.Type(value = ActivationRationalTanh.class, name = "RationalTanh"),
+            @JsonSubTypes.Type(value = ActivationRectifiedTanh.class, name = "RectifiedTanh"),
+            @JsonSubTypes.Type(value = ActivationSELU.class, name = "SELU"),
+            @JsonSubTypes.Type(value = ActivationSwish.class, name = "SWISH"),
+            @JsonSubTypes.Type(value = ActivationReLU.class, name = "ReLU"),
+            @JsonSubTypes.Type(value = ActivationRReLU.class, name = "RReLU"),
+            @JsonSubTypes.Type(value = ActivationSigmoid.class, name = "Sigmoid"),
+            @JsonSubTypes.Type(value = ActivationSoftmax.class, name = "Softmax"),
+            @JsonSubTypes.Type(value = ActivationSoftPlus.class, name = "SoftPlus"),
+            @JsonSubTypes.Type(value = ActivationSoftSign.class, name = "SoftSign"),
+            @JsonSubTypes.Type(value = ActivationTanH.class, name = "TanH")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class IActivationMixin {}
+
+    @JsonTypeInfo(use = JsonTypeInfo.Id.NAME, include = JsonTypeInfo.As.WRAPPER_OBJECT)
+    @JsonSubTypes(value = {@JsonSubTypes.Type(value = LossBinaryXENT.class, name = "BinaryXENT"),
+            @JsonSubTypes.Type(value = LossCosineProximity.class, name = "CosineProximity"),
+            @JsonSubTypes.Type(value = LossHinge.class, name = "Hinge"),
+            @JsonSubTypes.Type(value = LossKLD.class, name = "KLD"),
+            @JsonSubTypes.Type(value = LossMAE.class, name = "MAE"),
+            @JsonSubTypes.Type(value = LossL1.class, name = "L1"),
+            @JsonSubTypes.Type(value = LossMAPE.class, name = "MAPE"),
+            @JsonSubTypes.Type(value = LossMCXENT.class, name = "MCXENT"),
+            @JsonSubTypes.Type(value = LossMSE.class, name = "MSE"),
+            @JsonSubTypes.Type(value = LossL2.class, name = "L2"),
+            @JsonSubTypes.Type(value = LossMSLE.class, name = "MSLE"),
+            @JsonSubTypes.Type(value = LossNegativeLogLikelihood.class, name = "NegativeLogLikelihood"),
+            @JsonSubTypes.Type(value = LossPoisson.class, name = "Poisson"),
+            @JsonSubTypes.Type(value = LossSquaredHinge.class, name = "SquaredHinge"),
+            @JsonSubTypes.Type(value = LossFMeasure.class, name = "FMeasure")})
+    @NoArgsConstructor(access = AccessLevel.PRIVATE)
+    public static class ILossFunctionMixin {}
+}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyGraphVertexDeserializer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyGraphVertexDeserializer.java
deleted file mode 100644
index 822d8fc3d..000000000
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyGraphVertexDeserializer.java
+++ /dev/null
@@ -1,94 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.nn.conf.serde.legacyformat;
-
-import lombok.NonNull;
-import org.deeplearning4j.nn.conf.graph.*;
-import org.deeplearning4j.nn.conf.graph.rnn.DuplicateToTimeSeriesVertex;
-import org.deeplearning4j.nn.conf.graph.rnn.LastTimeStepVertex;
-import org.deeplearning4j.nn.conf.graph.rnn.ReverseTimeSeriesVertex;
-import org.deeplearning4j.nn.conf.serde.JsonMappers;
-import org.nd4j.serde.json.BaseLegacyDeserializer;
-import org.nd4j.shade.jackson.databind.ObjectMapper;
-
-import java.util.Arrays;
-import java.util.HashMap;
-import java.util.List;
-import java.util.Map;
-
-/**
- * Deserializer for GraphVertex JSON in legacy format - see {@link BaseLegacyDeserializer}
- *
- * @author Alex Black
- */
-public class LegacyGraphVertexDeserializer extends BaseLegacyDeserializer<GraphVertex> {
-
-    private static final Map<String,String> LEGACY_NAMES = new HashMap<>();
-
-    static {
-
-
-        List<Class<? extends GraphVertex>> cList = Arrays.asList(
-                //All of these vertices had the legacy format name the same as the simple class name
-                MergeVertex.class,
-                SubsetVertex.class,
-                LayerVertex.class,
-                LastTimeStepVertex.class,
-                ReverseTimeSeriesVertex.class,
-                DuplicateToTimeSeriesVertex.class,
-                PreprocessorVertex.class,
-                StackVertex.class,
-                UnstackVertex.class,
-                L2Vertex.class,
-                ScaleVertex.class,
-                L2NormalizeVertex.class,
-                //These did not previously have a subtype annotation - they use default (which is simple class name)
-                ElementWiseVertex.class,
-                PoolHelperVertex.class,
-                ReshapeVertex.class,
-                ShiftVertex.class);
-
-        for(Class<?> c : cList){
-            LEGACY_NAMES.put(c.getSimpleName(), c.getName());
-        }
-    }
-
-
-    @Override
-    public Map<String, String> getLegacyNamesMap() {
-        return LEGACY_NAMES;
-    }
-
-    @Override
-    public ObjectMapper getLegacyJsonMapper() {
-//        return JsonMappers.getMapperLegacyJson();
-        return JsonMappers.getJsonMapperLegacyFormatVertex();
-    }
-
-    @Override
-    public Class<?> getDeserializedType() {
-        return GraphVertex.class;
-    }
-
-    public static void registerLegacyClassDefaultName(@NonNull Class<? extends GraphVertex> clazz){
-        registerLegacyClassSpecifiedName(clazz.getSimpleName(), clazz);
-    }
-
-    public static void registerLegacyClassSpecifiedName(@NonNull String name, @NonNull Class<? extends GraphVertex> clazz){
-        LEGACY_NAMES.put(name, clazz.getName());
-    }
-}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyGraphVertexDeserializerHelper.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyGraphVertexDeserializerHelper.java
deleted file mode 100644
index bf9a2654a..000000000
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyGraphVertexDeserializerHelper.java
+++ /dev/null
@@ -1,28 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.nn.conf.serde.legacyformat;
-
-import org.nd4j.shade.jackson.databind.annotation.JsonDeserialize;
-
-/**
- * Simple helper class to redirect legacy JSON format to {@link LegacyGraphVertexDeserializer} via annotation
- * on {@link org.deeplearning4j.nn.conf.graph.GraphVertex}
- */
-@JsonDeserialize(using = LegacyGraphVertexDeserializer.class)
-public class LegacyGraphVertexDeserializerHelper {
-    private LegacyGraphVertexDeserializerHelper(){ }
-}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyLayerDeserializer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyLayerDeserializer.java
deleted file mode 100644
index 9111dff87..000000000
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyLayerDeserializer.java
+++ /dev/null
@@ -1,113 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.nn.conf.serde.legacyformat;
-
-import lombok.NonNull;
-import org.deeplearning4j.nn.conf.layers.*;
-import org.deeplearning4j.nn.conf.layers.convolutional.Cropping1D;
-import org.deeplearning4j.nn.conf.layers.convolutional.Cropping2D;
-import org.deeplearning4j.nn.conf.layers.misc.ElementWiseMultiplicationLayer;
-import org.deeplearning4j.nn.conf.layers.misc.FrozenLayer;
-import org.deeplearning4j.nn.conf.layers.objdetect.Yolo2OutputLayer;
-import org.deeplearning4j.nn.conf.layers.recurrent.Bidirectional;
-import org.deeplearning4j.nn.conf.layers.recurrent.LastTimeStep;
-import org.deeplearning4j.nn.conf.layers.recurrent.SimpleRnn;
-import org.deeplearning4j.nn.conf.layers.util.MaskLayer;
-import org.deeplearning4j.nn.conf.layers.util.MaskZeroLayer;
-import org.deeplearning4j.nn.conf.layers.variational.VariationalAutoencoder;
-import org.deeplearning4j.nn.conf.serde.JsonMappers;
-import org.nd4j.serde.json.BaseLegacyDeserializer;
-import org.nd4j.shade.jackson.databind.ObjectMapper;
-
-import java.util.HashMap;
-import java.util.Map;
-
-/**
- * Deserializer for Layer JSON in legacy format - see {@link BaseLegacyDeserializer}
- *
- * @author Alex Black
- */
-public class LegacyLayerDeserializer extends BaseLegacyDeserializer<Layer> {
-
-    private static final Map<String,String> LEGACY_NAMES = new HashMap<>();
-
-    static {
-        LEGACY_NAMES.put("autoEncoder", AutoEncoder.class.getName());
-        LEGACY_NAMES.put("convolution", ConvolutionLayer.class.getName());
-        LEGACY_NAMES.put("convolution1d", Convolution1DLayer.class.getName());
-        LEGACY_NAMES.put("gravesLSTM", GravesLSTM.class.getName());
-        LEGACY_NAMES.put("LSTM", LSTM.class.getName());
-        LEGACY_NAMES.put("gravesBidirectionalLSTM", GravesBidirectionalLSTM.class.getName());
-        LEGACY_NAMES.put("output", OutputLayer.class.getName());
-        LEGACY_NAMES.put("CenterLossOutputLayer", CenterLossOutputLayer.class.getName());
-        LEGACY_NAMES.put("rnnoutput", RnnOutputLayer.class.getName());
-        LEGACY_NAMES.put("loss", LossLayer.class.getName());
-        LEGACY_NAMES.put("dense", DenseLayer.class.getName());
-        LEGACY_NAMES.put("subsampling", SubsamplingLayer.class.getName());
-        LEGACY_NAMES.put("subsampling1d", Subsampling1DLayer.class.getName());
-        LEGACY_NAMES.put("batchNormalization", BatchNormalization.class.getName());
-        LEGACY_NAMES.put("localResponseNormalization", LocalResponseNormalization.class.getName());
-        LEGACY_NAMES.put("embedding", EmbeddingLayer.class.getName());
-        LEGACY_NAMES.put("activation", ActivationLayer.class.getName());
-        LEGACY_NAMES.put("VariationalAutoencoder", VariationalAutoencoder.class.getName());
-        LEGACY_NAMES.put("dropout", DropoutLayer.class.getName());
-        LEGACY_NAMES.put("GlobalPooling", GlobalPoolingLayer.class.getName());
-        LEGACY_NAMES.put("zeroPadding", ZeroPaddingLayer.class.getName());
-        LEGACY_NAMES.put("zeroPadding1d", ZeroPadding1DLayer.class.getName());
-        LEGACY_NAMES.put("FrozenLayer", FrozenLayer.class.getName());
-        LEGACY_NAMES.put("Upsampling2D", Upsampling2D.class.getName());
-        LEGACY_NAMES.put("Yolo2OutputLayer", Yolo2OutputLayer.class.getName());
-        LEGACY_NAMES.put("RnnLossLayer", RnnLossLayer.class.getName());
-        LEGACY_NAMES.put("CnnLossLayer", CnnLossLayer.class.getName());
-        LEGACY_NAMES.put("Bidirectional", Bidirectional.class.getName());
-        LEGACY_NAMES.put("SimpleRnn", SimpleRnn.class.getName());
-        LEGACY_NAMES.put("ElementWiseMult", ElementWiseMultiplicationLayer.class.getName());
-        LEGACY_NAMES.put("MaskLayer", MaskLayer.class.getName());
-        LEGACY_NAMES.put("MaskZeroLayer", MaskZeroLayer.class.getName());
-        LEGACY_NAMES.put("Cropping1D", Cropping1D.class.getName());
-        LEGACY_NAMES.put("Cropping2D", Cropping2D.class.getName());
-
-        //The following didn't previously have subtype annotations - hence will be using default name (class simple name)
-        LEGACY_NAMES.put("LastTimeStep", LastTimeStep.class.getName());
-        LEGACY_NAMES.put("SpaceToDepthLayer", SpaceToDepthLayer.class.getName());
-        LEGACY_NAMES.put("SpaceToBatchLayer", SpaceToBatchLayer.class.getName());
-    }
-
-
-    @Override
-    public Map<String, String> getLegacyNamesMap() {
-        return LEGACY_NAMES;
-    }
-
-    @Override
-    public ObjectMapper getLegacyJsonMapper() {
-        return JsonMappers.getJsonMapperLegacyFormatLayer();
-    }
-
-    @Override
-    public Class<?> getDeserializedType() {
-        return Layer.class;
-    }
-
-    public static void registerLegacyClassDefaultName(@NonNull Class<? extends Layer> clazz){
-        registerLegacyClassSpecifiedName(clazz.getSimpleName(), clazz);
-    }
-
-    public static void registerLegacyClassSpecifiedName(@NonNull String name, @NonNull Class<? extends Layer> clazz){
-        LEGACY_NAMES.put(name, clazz.getName());
-    }
-}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyLayerDeserializerHelper.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyLayerDeserializerHelper.java
deleted file mode 100644
index 76f7f7d7d..000000000
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyLayerDeserializerHelper.java
+++ /dev/null
@@ -1,28 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.nn.conf.serde.legacyformat;
-
-import org.nd4j.shade.jackson.databind.annotation.JsonDeserialize;
-
-/**
- * Simple helper class to redirect legacy JSON format to {@link LegacyLayerDeserializer} via annotation
- * on {@link org.deeplearning4j.nn.conf.layers.Layer}
- */
-@JsonDeserialize(using = LegacyLayerDeserializer.class)
-public class LegacyLayerDeserializerHelper {
-    private LegacyLayerDeserializerHelper(){ }
-}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyPreprocessorDeserializer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyPreprocessorDeserializer.java
deleted file mode 100644
index bed1c1c5c..000000000
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyPreprocessorDeserializer.java
+++ /dev/null
@@ -1,83 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.nn.conf.serde.legacyformat;
-
-import lombok.NonNull;
-import org.deeplearning4j.nn.conf.InputPreProcessor;
-import org.deeplearning4j.nn.conf.graph.*;
-import org.deeplearning4j.nn.conf.graph.rnn.DuplicateToTimeSeriesVertex;
-import org.deeplearning4j.nn.conf.graph.rnn.LastTimeStepVertex;
-import org.deeplearning4j.nn.conf.graph.rnn.ReverseTimeSeriesVertex;
-import org.deeplearning4j.nn.conf.preprocessor.*;
-import org.deeplearning4j.nn.conf.serde.JsonMappers;
-import org.nd4j.serde.json.BaseLegacyDeserializer;
-import org.nd4j.shade.jackson.databind.ObjectMapper;
-
-import java.util.Arrays;
-import java.util.HashMap;
-import java.util.List;
-import java.util.Map;
-
-/**
- * Deserializer for InputPreProcessor JSON in legacy format - see {@link BaseLegacyDeserializer}
- *
- * @author Alex Black
- */
-public class LegacyPreprocessorDeserializer extends BaseLegacyDeserializer<InputPreProcessor> {
-
-    private static final Map<String,String> LEGACY_NAMES = new HashMap<>();
-
-    static {
-        LEGACY_NAMES.put("cnnToFeedForward", CnnToFeedForwardPreProcessor.class.getName());
-        LEGACY_NAMES.put("cnnToRnn", CnnToRnnPreProcessor.class.getName());
-        LEGACY_NAMES.put("composableInput", ComposableInputPreProcessor.class.getName());
-        LEGACY_NAMES.put("feedForwardToCnn", FeedForwardToCnnPreProcessor.class.getName());
-        LEGACY_NAMES.put("feedForwardToRnn", FeedForwardToRnnPreProcessor.class.getName());
-        LEGACY_NAMES.put("rnnToFeedForward", RnnToFeedForwardPreProcessor.class.getName());
-        LEGACY_NAMES.put("rnnToCnn", RnnToCnnPreProcessor.class.getName());
-
-        //Keras preprocessors: they defaulted to class simple name
-        LEGACY_NAMES.put("KerasFlattenRnnPreprocessor","org.deeplearning4j.nn.modelimport.keras.preprocessors.KerasFlattenRnnPreprocessor");
-        LEGACY_NAMES.put("ReshapePreprocessor","org.deeplearning4j.nn.modelimport.keras.preprocessors.ReshapePreprocessor");
-        LEGACY_NAMES.put("TensorFlowCnnToFeedForwardPreProcessor","org.deeplearning4j.nn.modelimport.keras.preprocessors.TensorFlowCnnToFeedForwardPreProcessor");
-    }
-
-
-    @Override
-    public Map<String, String> getLegacyNamesMap() {
-        return LEGACY_NAMES;
-    }
-
-    @Override
-    public ObjectMapper getLegacyJsonMapper() {
-//        return JsonMappers.getMapperLegacyJson();
-        return JsonMappers.getJsonMapperLegacyFormatPreproc();
-    }
-
-    @Override
-    public Class<?> getDeserializedType() {
-        return InputPreProcessor.class;
-    }
-
-    public static void registerLegacyClassDefaultName(@NonNull Class<? extends InputPreProcessor> clazz){
-        registerLegacyClassSpecifiedName(clazz.getSimpleName(), clazz);
-    }
-
-    public static void registerLegacyClassSpecifiedName(@NonNull String name, @NonNull Class<? extends InputPreProcessor> clazz){
-        LEGACY_NAMES.put(name, clazz.getName());
-    }
-}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyPreprocessorDeserializerHelper.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyPreprocessorDeserializerHelper.java
deleted file mode 100644
index 19300ba5f..000000000
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyPreprocessorDeserializerHelper.java
+++ /dev/null
@@ -1,28 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.nn.conf.serde.legacyformat;
-
-import org.nd4j.shade.jackson.databind.annotation.JsonDeserialize;
-
-/**
- * Simple helper class to redirect legacy JSON format to {@link LegacyPreprocessorDeserializer} via annotation
- * on {@link org.deeplearning4j.nn.conf.InputPreProcessor}
- */
-@JsonDeserialize(using = LegacyPreprocessorDeserializer.class)
-public class LegacyPreprocessorDeserializerHelper {
-    private LegacyPreprocessorDeserializerHelper(){ }
-}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyReconstructionDistributionDeserializer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyReconstructionDistributionDeserializer.java
deleted file mode 100644
index 06cf37f7e..000000000
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyReconstructionDistributionDeserializer.java
+++ /dev/null
@@ -1,70 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.nn.conf.serde.legacyformat;
-
-import lombok.NonNull;
-import org.deeplearning4j.nn.conf.InputPreProcessor;
-import org.deeplearning4j.nn.conf.layers.variational.*;
-import org.deeplearning4j.nn.conf.preprocessor.*;
-import org.deeplearning4j.nn.conf.serde.JsonMappers;
-import org.nd4j.serde.json.BaseLegacyDeserializer;
-import org.nd4j.shade.jackson.databind.ObjectMapper;
-
-import java.util.HashMap;
-import java.util.Map;
-
-/**
- * Deserializer for ReconstructionDistribution JSON in legacy format - see {@link BaseLegacyDeserializer}
- *
- * @author Alex Black
- */
-public class LegacyReconstructionDistributionDeserializer extends BaseLegacyDeserializer<ReconstructionDistribution> {
-
-    private static final Map<String,String> LEGACY_NAMES = new HashMap<>();
-
-    static {
-        LEGACY_NAMES.put("Gaussian", GaussianReconstructionDistribution.class.getName());
-        LEGACY_NAMES.put("Bernoulli", BernoulliReconstructionDistribution.class.getName());
-        LEGACY_NAMES.put("Exponential", ExponentialReconstructionDistribution.class.getName());
-        LEGACY_NAMES.put("Composite", CompositeReconstructionDistribution.class.getName());
-        LEGACY_NAMES.put("LossWrapper", LossFunctionWrapper.class.getName());
-    }
-
-
-    @Override
-    public Map<String, String> getLegacyNamesMap() {
-        return LEGACY_NAMES;
-    }
-
-    @Override
-    public ObjectMapper getLegacyJsonMapper() {
-        return JsonMappers.getJsonMapperLegacyFormatReconstruction();
-    }
-
-    @Override
-    public Class<?> getDeserializedType() {
-        return ReconstructionDistribution.class;
-    }
-
-    public static void registerLegacyClassDefaultName(@NonNull Class<? extends ReconstructionDistribution> clazz){
-        registerLegacyClassSpecifiedName(clazz.getSimpleName(), clazz);
-    }
-
-    public static void registerLegacyClassSpecifiedName(@NonNull String name, @NonNull Class<? extends ReconstructionDistribution> clazz){
-        LEGACY_NAMES.put(name, clazz.getName());
-    }
-}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyReconstructionDistributionDeserializerHelper.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyReconstructionDistributionDeserializerHelper.java
deleted file mode 100644
index 61952d62e..000000000
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/serde/legacyformat/LegacyReconstructionDistributionDeserializerHelper.java
+++ /dev/null
@@ -1,28 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.nn.conf.serde.legacyformat;
-
-import org.nd4j.shade.jackson.databind.annotation.JsonDeserialize;
-
-/**
- * Simple helper class to redirect legacy JSON format to {@link LegacyReconstructionDistributionDeserializer} via annotation
- * on {@link org.deeplearning4j.nn.conf.layers.variational.ReconstructionDistribution}
- */
-@JsonDeserialize(using = LegacyReconstructionDistributionDeserializer.class)
-public class LegacyReconstructionDistributionDeserializerHelper {
-    private LegacyReconstructionDistributionDeserializerHelper(){ }
-}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/workspace/LayerWorkspaceMgr.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/workspace/LayerWorkspaceMgr.java
index e40fdcd32..5ce11cf5c 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/workspace/LayerWorkspaceMgr.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/workspace/LayerWorkspaceMgr.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.nn.workspace;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import lombok.Getter;
 import lombok.NonNull;
 import lombok.Setter;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/listeners/CheckpointListener.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/listeners/CheckpointListener.java
index 8026d0764..b67961985 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/listeners/CheckpointListener.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/listeners/CheckpointListener.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.optimize.listeners;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
 import org.apache.commons.io.IOUtils;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/listeners/PerformanceListener.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/listeners/PerformanceListener.java
index f474ee24b..9ff5993c3 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/listeners/PerformanceListener.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/listeners/PerformanceListener.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.optimize.listeners;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import lombok.extern.slf4j.Slf4j;
 import org.deeplearning4j.nn.api.Model;
 import org.deeplearning4j.nn.graph.ComputationGraph;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/solvers/accumulation/EncodingHandler.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/solvers/accumulation/EncodingHandler.java
index 42b63a849..dc632cecd 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/solvers/accumulation/EncodingHandler.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/solvers/accumulation/EncodingHandler.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.optimize.solvers.accumulation;
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
 import org.deeplearning4j.optimize.solvers.accumulation.encoding.ResidualPostProcessor;
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/util/ModelSerializer.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/util/ModelSerializer.java
index f37ac245c..70945b06c 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/util/ModelSerializer.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/util/ModelSerializer.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.util;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
 import org.apache.commons.io.IOUtils;
diff --git a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-aws/pom.xml b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-aws/pom.xml
index 4a8b6f230..0b6b05c26 100644
--- a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-aws/pom.xml
+++ b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-aws/pom.xml
@@ -38,13 +38,6 @@
     </plugins>
   </build>
   <properties>
-    <!-- These Spark version properties have to defined be here, and NOT in the deeplearning4j-parent pom. Otherwise,
-         the properties will be resolved to their defaults (and not the values present during the build). Whereas when
-         defined here (the version is spark 1 vs. 2 specific) we can have different version properties simultaneously
-         (in different pom files) instead of a single global property -->
-    <spark.version>2.1.0</spark.version>
-    <spark.major.version>2</spark.major.version>
-
     <!-- Default scala versions, may be overwritten by build profiles -->
     <scala.version>2.11.12</scala.version>
     <scala.binary.version>2.11</scala.binary.version>
@@ -102,11 +95,6 @@
       <artifactId>jsch</artifactId>
       <version>${jsch.version}</version>
     </dependency>
-    <dependency>
-      <groupId>com.google.guava</groupId>
-      <artifactId>guava</artifactId>
-      <version>${guava.version}</version>
-    </dependency>
     <dependency>
       <groupId>com.google.inject</groupId>
       <artifactId>guice</artifactId>
@@ -152,21 +140,6 @@
       <artifactId>jaxb-impl</artifactId>
       <version>${jaxb.version}</version>
     </dependency>
-    <dependency>
-      <groupId>com.typesafe.akka</groupId>
-      <artifactId>akka-actor_2.11</artifactId>
-      <version>${akka.version}</version>
-    </dependency>
-    <dependency>
-      <groupId>com.typesafe.akka</groupId>
-      <artifactId>akka-remote_2.11</artifactId>
-      <version>${akka.version}</version>
-    </dependency>
-    <dependency>
-      <groupId>com.typesafe.akka</groupId>
-      <artifactId>akka-slf4j_2.11</artifactId>
-      <version>${akka.version}</version>
-    </dependency>
     <dependency>
       <groupId>io.netty</groupId>
       <artifactId>netty</artifactId>
diff --git a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml
index 8fd0ea38f..eaf0e6be8 100644
--- a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml
+++ b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/pom.xml
@@ -72,27 +72,11 @@
             <scope>test</scope>
         </dependency>
 
-        <!-- Akka/Play seems to require these to be manually set to avoid class not found exception issues -->
-        <dependency>
-            <groupId>io.netty</groupId>
-            <artifactId>netty</artifactId>
-            <version>${netty.version}</version>
-        </dependency>
         <dependency>
             <groupId>org.scala-lang</groupId>
             <artifactId>scala-library</artifactId>
             <version>${scala.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-cluster_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.typesafe</groupId>
-            <artifactId>config</artifactId>
-            <version>${typesafe.config.version}</version>
-        </dependency>
 
         <dependency>
             <groupId>ch.qos.logback</groupId>
diff --git a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/src/main/java/org/deeplearning4j/parallelism/EarlyStoppingParallelTrainer.java b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/src/main/java/org/deeplearning4j/parallelism/EarlyStoppingParallelTrainer.java
index 1af2f4600..4ba5ba4ce 100644
--- a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/src/main/java/org/deeplearning4j/parallelism/EarlyStoppingParallelTrainer.java
+++ b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/src/main/java/org/deeplearning4j/parallelism/EarlyStoppingParallelTrainer.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.parallelism;
 
-import com.google.common.util.concurrent.AtomicDouble;
+import org.nd4j.shade.guava.util.concurrent.AtomicDouble;
 import lombok.extern.slf4j.Slf4j;
 import org.deeplearning4j.earlystopping.EarlyStoppingConfiguration;
 import org.deeplearning4j.earlystopping.EarlyStoppingResult;
diff --git a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/src/main/java/org/deeplearning4j/parallelism/inference/observers/BasicInferenceObservable.java b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/src/main/java/org/deeplearning4j/parallelism/inference/observers/BasicInferenceObservable.java
index 4c9520ecd..432e22695 100644
--- a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/src/main/java/org/deeplearning4j/parallelism/inference/observers/BasicInferenceObservable.java
+++ b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper/src/main/java/org/deeplearning4j/parallelism/inference/observers/BasicInferenceObservable.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.parallelism.inference.observers;
 
-import com.google.common.base.Preconditions;
+import org.nd4j.shade.guava.base.Preconditions;
 import lombok.Getter;
 import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/pom.xml b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/pom.xml
index 16e0176a8..3fded3e4a 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/pom.xml
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/pom.xml
@@ -32,42 +32,6 @@
         <hazelcast.version>3.4.2</hazelcast.version>
     </properties>
 
-    <dependencyManagement>
-        <dependencies>
-            <dependency>
-                <!-- Force Jackson versions to match Spark. This usually only matters for Spark local, not Spark submit -->
-                <groupId>com.fasterxml.jackson.core</groupId>
-                <artifactId>jackson-core</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.core</groupId>
-                <artifactId>jackson-databind</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.core</groupId>
-                <artifactId>jackson-annotations</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.module</groupId>
-                <artifactId>jackson-module-scala_${scala.binary.version}</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.datatype</groupId>
-                <artifactId>jackson-datatype-jdk8</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.datatype</groupId>
-                <artifactId>jackson-datatype-jsr310</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-        </dependencies>
-    </dependencyManagement>
-
     <dependencies>
         <dependency>
             <groupId>org.deeplearning4j</groupId>
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/src/main/java/org/deeplearning4j/spark/models/sequencevectors/functions/VocabRddFunctionFlat.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/src/main/java/org/deeplearning4j/spark/models/sequencevectors/functions/VocabRddFunctionFlat.java
index 9164a41d0..d14668154 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/src/main/java/org/deeplearning4j/spark/models/sequencevectors/functions/VocabRddFunctionFlat.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/src/main/java/org/deeplearning4j/spark/models/sequencevectors/functions/VocabRddFunctionFlat.java
@@ -17,9 +17,8 @@
 package org.deeplearning4j.spark.models.sequencevectors.functions;
 
 import lombok.NonNull;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.deeplearning4j.models.embeddings.loader.VectorsConfiguration;
 import org.deeplearning4j.models.sequencevectors.sequence.Sequence;
 import org.deeplearning4j.models.sequencevectors.sequence.SequenceElement;
@@ -31,64 +30,56 @@ import org.nd4j.parameterserver.distributed.training.TrainingDriver;
 import org.nd4j.parameterserver.distributed.transport.RoutedTransport;
 
 import java.util.ArrayList;
+import java.util.Iterator;
 import java.util.List;
 
 /**
  * @author raver119@gmail.com
  */
-public class VocabRddFunctionFlat<T extends SequenceElement> extends BaseFlatMapFunctionAdaptee<Sequence<T>, T> {
+public class VocabRddFunctionFlat<T extends SequenceElement> implements FlatMapFunction<Sequence<T>, T> {
+    protected Broadcast<VectorsConfiguration> vectorsConfigurationBroadcast;
+    protected Broadcast<VoidConfiguration> paramServerConfigurationBroadcast;
+
+    protected transient VectorsConfiguration configuration;
+    protected transient SparkElementsLearningAlgorithm ela;
+    protected transient TrainingDriver<? extends TrainingMessage> driver;
+
     public VocabRddFunctionFlat(@NonNull Broadcast<VectorsConfiguration> vectorsConfigurationBroadcast,
-                    @NonNull Broadcast<VoidConfiguration> paramServerConfigurationBroadcast) {
-        super(new VocabRddFunctionAdapter<T>(vectorsConfigurationBroadcast, paramServerConfigurationBroadcast));
+                            @NonNull Broadcast<VoidConfiguration> paramServerConfigurationBroadcast) {
+        this.vectorsConfigurationBroadcast = vectorsConfigurationBroadcast;
+        this.paramServerConfigurationBroadcast = paramServerConfigurationBroadcast;
     }
 
+    @Override
+    public Iterator<T> call(Sequence<T> sequence) throws Exception {
+        if (configuration == null)
+            configuration = vectorsConfigurationBroadcast.getValue();
 
-    private static class VocabRddFunctionAdapter<T extends SequenceElement>
-                    implements FlatMapFunctionAdapter<Sequence<T>, T> {
-        protected Broadcast<VectorsConfiguration> vectorsConfigurationBroadcast;
-        protected Broadcast<VoidConfiguration> paramServerConfigurationBroadcast;
-
-        protected transient VectorsConfiguration configuration;
-        protected transient SparkElementsLearningAlgorithm ela;
-        protected transient TrainingDriver<? extends TrainingMessage> driver;
-
-        public VocabRddFunctionAdapter(@NonNull Broadcast<VectorsConfiguration> vectorsConfigurationBroadcast,
-                        @NonNull Broadcast<VoidConfiguration> paramServerConfigurationBroadcast) {
-            this.vectorsConfigurationBroadcast = vectorsConfigurationBroadcast;
-            this.paramServerConfigurationBroadcast = paramServerConfigurationBroadcast;
-        }
-
-        @Override
-        public Iterable<T> call(Sequence<T> sequence) throws Exception {
-            if (configuration == null)
-                configuration = vectorsConfigurationBroadcast.getValue();
-
-            if (ela == null) {
-                try {
-                    ela = (SparkElementsLearningAlgorithm) Class.forName(configuration.getElementsLearningAlgorithm())
-                                    .newInstance();
-                } catch (Exception e) {
-                    throw new RuntimeException(e);
-                }
+        if (ela == null) {
+            try {
+                ela = (SparkElementsLearningAlgorithm) Class.forName(configuration.getElementsLearningAlgorithm())
+                        .newInstance();
+            } catch (Exception e) {
+                throw new RuntimeException(e);
             }
-            driver = ela.getTrainingDriver();
-
-            // we just silently initialize server
-            VoidParameterServer.getInstance().init(paramServerConfigurationBroadcast.getValue(), new RoutedTransport(),
-                            driver);
-
-            // TODO: call for initializeSeqVec here
-
-            List<T> elements = new ArrayList<>();
-
-            elements.addAll(sequence.getElements());
-
-            // FIXME: this is PROBABLY bad, we might want to ensure, there's no duplicates.
-            if (configuration.isTrainSequenceVectors())
-                if (!sequence.getSequenceLabels().isEmpty())
-                    elements.addAll(sequence.getSequenceLabels());
-
-            return elements;
         }
+        driver = ela.getTrainingDriver();
+
+        // we just silently initialize server
+        VoidParameterServer.getInstance().init(paramServerConfigurationBroadcast.getValue(), new RoutedTransport(),
+                driver);
+
+        // TODO: call for initializeSeqVec here
+
+        List<T> elements = new ArrayList<>();
+
+        elements.addAll(sequence.getElements());
+
+        // FIXME: this is PROBABLY bad, we might want to ensure, there's no duplicates.
+        if (configuration.isTrainSequenceVectors())
+            if (!sequence.getSequenceLabels().isEmpty())
+                elements.addAll(sequence.getSequenceLabels());
+
+        return elements.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/FirstIterationFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/FirstIterationFunction.java
index 4718d0118..bbad9f2e3 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/FirstIterationFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/FirstIterationFunction.java
@@ -16,28 +16,252 @@
 
 package org.deeplearning4j.spark.models.embeddings.word2vec;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.deeplearning4j.models.word2vec.VocabWord;
 import org.deeplearning4j.models.word2vec.wordstore.VocabCache;
 import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.factory.Nd4j;
+import org.nd4j.linalg.primitives.Pair;
 import scala.Tuple2;
 
-import java.util.Iterator;
-import java.util.List;
-import java.util.Map;
+import java.util.*;
 import java.util.Map.Entry;
+import java.util.concurrent.atomic.AtomicLong;
 
 /**
  * @author jeffreytang
  * @author raver119@gmail.com
  */
-public class FirstIterationFunction extends
-                BaseFlatMapFunctionAdaptee<Iterator<Tuple2<List<VocabWord>, Long>>, Entry<VocabWord, INDArray>> {
+public class FirstIterationFunction implements
+        FlatMapFunction<Iterator<Tuple2<List<VocabWord>, Long>>, Entry<VocabWord, INDArray>> {
+
+    private int ithIteration = 1;
+    private int vectorLength;
+    private boolean useAdaGrad;
+    private int batchSize = 0;
+    private double negative;
+    private int window;
+    private double alpha;
+    private double minAlpha;
+    private long totalWordCount;
+    private long seed;
+    private int maxExp;
+    private double[] expTable;
+    private int iterations;
+    private Map<VocabWord, INDArray> indexSyn0VecMap;
+    private Map<Integer, INDArray> pointSyn1VecMap;
+    private AtomicLong nextRandom = new AtomicLong(5);
+
+    private volatile VocabCache<VocabWord> vocab;
+    private volatile NegativeHolder negativeHolder;
+    private AtomicLong cid = new AtomicLong(0);
+    private AtomicLong aff = new AtomicLong(0);
+
+
 
     public FirstIterationFunction(Broadcast<Map<String, Object>> word2vecVarMapBroadcast,
-                    Broadcast<double[]> expTableBroadcast, Broadcast<VocabCache<VocabWord>> vocabCacheBroadcast) {
-        super(new FirstIterationFunctionAdapter(word2vecVarMapBroadcast, expTableBroadcast, vocabCacheBroadcast));
+                                         Broadcast<double[]> expTableBroadcast, Broadcast<VocabCache<VocabWord>> vocabCacheBroadcast) {
+
+        Map<String, Object> word2vecVarMap = word2vecVarMapBroadcast.getValue();
+        this.expTable = expTableBroadcast.getValue();
+        this.vectorLength = (int) word2vecVarMap.get("vectorLength");
+        this.useAdaGrad = (boolean) word2vecVarMap.get("useAdaGrad");
+        this.negative = (double) word2vecVarMap.get("negative");
+        this.window = (int) word2vecVarMap.get("window");
+        this.alpha = (double) word2vecVarMap.get("alpha");
+        this.minAlpha = (double) word2vecVarMap.get("minAlpha");
+        this.totalWordCount = (long) word2vecVarMap.get("totalWordCount");
+        this.seed = (long) word2vecVarMap.get("seed");
+        this.maxExp = (int) word2vecVarMap.get("maxExp");
+        this.iterations = (int) word2vecVarMap.get("iterations");
+        this.batchSize = (int) word2vecVarMap.get("batchSize");
+        this.indexSyn0VecMap = new HashMap<>();
+        this.pointSyn1VecMap = new HashMap<>();
+        this.vocab = vocabCacheBroadcast.getValue();
+
+        if (this.vocab == null)
+            throw new RuntimeException("VocabCache is null");
+
+        if (negative > 0) {
+            negativeHolder = NegativeHolder.getInstance();
+            negativeHolder.initHolder(vocab, expTable, this.vectorLength);
+        }
+    }
+
+
+
+    @Override
+    public Iterator<Entry<VocabWord, INDArray>> call(Iterator<Tuple2<List<VocabWord>, Long>> pairIter) {
+        while (pairIter.hasNext()) {
+            List<Pair<List<VocabWord>, Long>> batch = new ArrayList<>();
+            while (pairIter.hasNext() && batch.size() < batchSize) {
+                Tuple2<List<VocabWord>, Long> pair = pairIter.next();
+                List<VocabWord> vocabWordsList = pair._1();
+                Long sentenceCumSumCount = pair._2();
+                batch.add(Pair.of(vocabWordsList, sentenceCumSumCount));
+            }
+
+            for (int i = 0; i < iterations; i++) {
+                //System.out.println("Training sentence: " + vocabWordsList);
+                for (Pair<List<VocabWord>, Long> pair : batch) {
+                    List<VocabWord> vocabWordsList = pair.getKey();
+                    Long sentenceCumSumCount = pair.getValue();
+                    double currentSentenceAlpha = Math.max(minAlpha,
+                            alpha - (alpha - minAlpha) * (sentenceCumSumCount / (double) totalWordCount));
+                    trainSentence(vocabWordsList, currentSentenceAlpha);
+                }
+            }
+        }
+        return indexSyn0VecMap.entrySet().iterator();
+    }
+
+
+    public void trainSentence(List<VocabWord> vocabWordsList, double currentSentenceAlpha) {
+
+        if (vocabWordsList != null && !vocabWordsList.isEmpty()) {
+            for (int ithWordInSentence = 0; ithWordInSentence < vocabWordsList.size(); ithWordInSentence++) {
+                // Random value ranging from 0 to window size
+                nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
+                int b = (int) (long) this.nextRandom.get() % window;
+                VocabWord currentWord = vocabWordsList.get(ithWordInSentence);
+                if (currentWord != null) {
+                    skipGram(ithWordInSentence, vocabWordsList, b, currentSentenceAlpha);
+                }
+            }
+        }
+    }
+
+    public void skipGram(int ithWordInSentence, List<VocabWord> vocabWordsList, int b, double currentSentenceAlpha) {
+
+        VocabWord currentWord = vocabWordsList.get(ithWordInSentence);
+        if (currentWord != null && !vocabWordsList.isEmpty()) {
+            int end = window * 2 + 1 - b;
+            for (int a = b; a < end; a++) {
+                if (a != window) {
+                    int c = ithWordInSentence - window + a;
+                    if (c >= 0 && c < vocabWordsList.size()) {
+                        VocabWord lastWord = vocabWordsList.get(c);
+                        iterateSample(currentWord, lastWord, currentSentenceAlpha);
+                    }
+                }
+            }
+        }
+    }
+
+    public void iterateSample(VocabWord w1, VocabWord w2, double currentSentenceAlpha) {
+
+
+        if (w1 == null || w2 == null || w2.getIndex() < 0 || w2.getIndex() == w1.getIndex())
+            return;
+        final int currentWordIndex = w2.getIndex();
+
+        // error for current word and context
+        INDArray neu1e = Nd4j.create(vectorLength);
+
+        // First iteration Syn0 is random numbers
+        INDArray l1 = null;
+        if (indexSyn0VecMap.containsKey(vocab.elementAtIndex(currentWordIndex))) {
+            l1 = indexSyn0VecMap.get(vocab.elementAtIndex(currentWordIndex));
+        } else {
+            l1 = getRandomSyn0Vec(vectorLength, (long) currentWordIndex);
+        }
+
+        //
+        for (int i = 0; i < w1.getCodeLength(); i++) {
+            int code = w1.getCodes().get(i);
+            int point = w1.getPoints().get(i);
+            if (point < 0)
+                throw new IllegalStateException("Illegal point " + point);
+            // Point to
+            INDArray syn1;
+            if (pointSyn1VecMap.containsKey(point)) {
+                syn1 = pointSyn1VecMap.get(point);
+            } else {
+                syn1 = Nd4j.zeros(1, vectorLength); // 1 row of vector length of zeros
+                pointSyn1VecMap.put(point, syn1);
+            }
+
+            // Dot product of Syn0 and Syn1 vecs
+            double dot = Nd4j.getBlasWrapper().level1().dot(vectorLength, 1.0, l1, syn1);
+
+            if (dot < -maxExp || dot >= maxExp)
+                continue;
+
+            int idx = (int) ((dot + maxExp) * ((double) expTable.length / maxExp / 2.0));
+
+            if (idx >= expTable.length)
+                continue;
+
+            //score
+            double f = expTable[idx];
+            //gradient
+            double g = (1 - code - f) * (useAdaGrad ? w1.getGradient(i, currentSentenceAlpha, currentSentenceAlpha)
+                    : currentSentenceAlpha);
+
+
+            Nd4j.getBlasWrapper().level1().axpy(vectorLength, g, syn1, neu1e);
+            Nd4j.getBlasWrapper().level1().axpy(vectorLength, g, l1, syn1);
+        }
+
+        int target = w1.getIndex();
+        int label;
+        //negative sampling
+        if (negative > 0)
+            for (int d = 0; d < negative + 1; d++) {
+                if (d == 0)
+                    label = 1;
+                else {
+                    nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
+
+                    // FIXME: int cast
+                    int idx = Math.abs((int) (nextRandom.get() >> 16) % (int) negativeHolder.getTable().length());
+
+                    target = negativeHolder.getTable().getInt(idx);
+                    if (target <= 0)
+                        target = (int) nextRandom.get() % (vocab.numWords() - 1) + 1;
+
+                    if (target == w1.getIndex())
+                        continue;
+                    label = 0;
+                }
+
+                if (target >= negativeHolder.getSyn1Neg().rows() || target < 0)
+                    continue;
+
+                double f = Nd4j.getBlasWrapper().dot(l1, negativeHolder.getSyn1Neg().slice(target));
+                double g;
+                if (f > maxExp)
+                    g = useAdaGrad ? w1.getGradient(target, (label - 1), alpha) : (label - 1) * alpha;
+                else if (f < -maxExp)
+                    g = label * (useAdaGrad ? w1.getGradient(target, alpha, alpha) : alpha);
+                else {
+                    int idx = (int) ((f + maxExp) * (expTable.length / maxExp / 2));
+                    if (idx >= expTable.length)
+                        continue;
+
+                    g = useAdaGrad ? w1.getGradient(target, label - expTable[idx], alpha)
+                            : (label - expTable[idx]) * alpha;
+                }
+
+                Nd4j.getBlasWrapper().level1().axpy(vectorLength, g, negativeHolder.getSyn1Neg().slice(target), neu1e);
+
+                Nd4j.getBlasWrapper().level1().axpy(vectorLength, g, l1, negativeHolder.getSyn1Neg().slice(target));
+            }
+
+
+        // Updated the Syn0 vector based on gradient. Syn0 is not random anymore.
+        Nd4j.getBlasWrapper().level1().axpy(vectorLength, 1.0f, neu1e, l1);
+
+        VocabWord word = vocab.elementAtIndex(currentWordIndex);
+        indexSyn0VecMap.put(word, l1);
+    }
+
+    private INDArray getRandomSyn0Vec(int vectorLength, long lseed) {
+        /*
+            we use wordIndex as part of seed here, to guarantee that during word syn0 initialization on dwo distinct nodes, initial weights will be the same for the same word
+         */
+        return Nd4j.rand( new int[] {1, vectorLength}, lseed * seed).subi(0.5).divi(vectorLength);
     }
 }
 
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/FirstIterationFunctionAdapter.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/FirstIterationFunctionAdapter.java
deleted file mode 100644
index 05ca83fc6..000000000
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/FirstIterationFunctionAdapter.java
+++ /dev/null
@@ -1,265 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.spark.models.embeddings.word2vec;
-
-import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.deeplearning4j.models.word2vec.VocabWord;
-import org.deeplearning4j.models.word2vec.wordstore.VocabCache;
-import org.nd4j.linalg.api.ndarray.INDArray;
-import org.nd4j.linalg.factory.Nd4j;
-import org.nd4j.linalg.primitives.Pair;
-import scala.Tuple2;
-
-import java.util.*;
-import java.util.concurrent.atomic.AtomicLong;
-
-/**
- * @author jeffreytang
- * @author raver119@gmail.com
- */
-public class FirstIterationFunctionAdapter implements
-                FlatMapFunctionAdapter<Iterator<Tuple2<List<VocabWord>, Long>>, Map.Entry<VocabWord, INDArray>> {
-
-    private int ithIteration = 1;
-    private int vectorLength;
-    private boolean useAdaGrad;
-    private int batchSize = 0;
-    private double negative;
-    private int window;
-    private double alpha;
-    private double minAlpha;
-    private long totalWordCount;
-    private long seed;
-    private int maxExp;
-    private double[] expTable;
-    private int iterations;
-    private Map<VocabWord, INDArray> indexSyn0VecMap;
-    private Map<Integer, INDArray> pointSyn1VecMap;
-    private AtomicLong nextRandom = new AtomicLong(5);
-
-    private volatile VocabCache<VocabWord> vocab;
-    private volatile NegativeHolder negativeHolder;
-    private AtomicLong cid = new AtomicLong(0);
-    private AtomicLong aff = new AtomicLong(0);
-
-
-
-    public FirstIterationFunctionAdapter(Broadcast<Map<String, Object>> word2vecVarMapBroadcast,
-                    Broadcast<double[]> expTableBroadcast, Broadcast<VocabCache<VocabWord>> vocabCacheBroadcast) {
-
-        Map<String, Object> word2vecVarMap = word2vecVarMapBroadcast.getValue();
-        this.expTable = expTableBroadcast.getValue();
-        this.vectorLength = (int) word2vecVarMap.get("vectorLength");
-        this.useAdaGrad = (boolean) word2vecVarMap.get("useAdaGrad");
-        this.negative = (double) word2vecVarMap.get("negative");
-        this.window = (int) word2vecVarMap.get("window");
-        this.alpha = (double) word2vecVarMap.get("alpha");
-        this.minAlpha = (double) word2vecVarMap.get("minAlpha");
-        this.totalWordCount = (long) word2vecVarMap.get("totalWordCount");
-        this.seed = (long) word2vecVarMap.get("seed");
-        this.maxExp = (int) word2vecVarMap.get("maxExp");
-        this.iterations = (int) word2vecVarMap.get("iterations");
-        this.batchSize = (int) word2vecVarMap.get("batchSize");
-        this.indexSyn0VecMap = new HashMap<>();
-        this.pointSyn1VecMap = new HashMap<>();
-        this.vocab = vocabCacheBroadcast.getValue();
-
-        if (this.vocab == null)
-            throw new RuntimeException("VocabCache is null");
-
-        if (negative > 0) {
-            negativeHolder = NegativeHolder.getInstance();
-            negativeHolder.initHolder(vocab, expTable, this.vectorLength);
-        }
-    }
-
-
-
-    @Override
-    public Iterable<Map.Entry<VocabWord, INDArray>> call(Iterator<Tuple2<List<VocabWord>, Long>> pairIter) {
-        while (pairIter.hasNext()) {
-            List<Pair<List<VocabWord>, Long>> batch = new ArrayList<>();
-            while (pairIter.hasNext() && batch.size() < batchSize) {
-                Tuple2<List<VocabWord>, Long> pair = pairIter.next();
-                List<VocabWord> vocabWordsList = pair._1();
-                Long sentenceCumSumCount = pair._2();
-                batch.add(Pair.of(vocabWordsList, sentenceCumSumCount));
-            }
-
-            for (int i = 0; i < iterations; i++) {
-                //System.out.println("Training sentence: " + vocabWordsList);
-                for (Pair<List<VocabWord>, Long> pair : batch) {
-                    List<VocabWord> vocabWordsList = pair.getKey();
-                    Long sentenceCumSumCount = pair.getValue();
-                    double currentSentenceAlpha = Math.max(minAlpha,
-                                    alpha - (alpha - minAlpha) * (sentenceCumSumCount / (double) totalWordCount));
-                    trainSentence(vocabWordsList, currentSentenceAlpha);
-                }
-            }
-        }
-        return indexSyn0VecMap.entrySet();
-    }
-
-
-    public void trainSentence(List<VocabWord> vocabWordsList, double currentSentenceAlpha) {
-
-        if (vocabWordsList != null && !vocabWordsList.isEmpty()) {
-            for (int ithWordInSentence = 0; ithWordInSentence < vocabWordsList.size(); ithWordInSentence++) {
-                // Random value ranging from 0 to window size
-                nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
-                int b = (int) (long) this.nextRandom.get() % window;
-                VocabWord currentWord = vocabWordsList.get(ithWordInSentence);
-                if (currentWord != null) {
-                    skipGram(ithWordInSentence, vocabWordsList, b, currentSentenceAlpha);
-                }
-            }
-        }
-    }
-
-    public void skipGram(int ithWordInSentence, List<VocabWord> vocabWordsList, int b, double currentSentenceAlpha) {
-
-        VocabWord currentWord = vocabWordsList.get(ithWordInSentence);
-        if (currentWord != null && !vocabWordsList.isEmpty()) {
-            int end = window * 2 + 1 - b;
-            for (int a = b; a < end; a++) {
-                if (a != window) {
-                    int c = ithWordInSentence - window + a;
-                    if (c >= 0 && c < vocabWordsList.size()) {
-                        VocabWord lastWord = vocabWordsList.get(c);
-                        iterateSample(currentWord, lastWord, currentSentenceAlpha);
-                    }
-                }
-            }
-        }
-    }
-
-    public void iterateSample(VocabWord w1, VocabWord w2, double currentSentenceAlpha) {
-
-
-        if (w1 == null || w2 == null || w2.getIndex() < 0 || w2.getIndex() == w1.getIndex())
-            return;
-        final int currentWordIndex = w2.getIndex();
-
-        // error for current word and context
-        INDArray neu1e = Nd4j.create(vectorLength);
-
-        // First iteration Syn0 is random numbers
-        INDArray l1 = null;
-        if (indexSyn0VecMap.containsKey(vocab.elementAtIndex(currentWordIndex))) {
-            l1 = indexSyn0VecMap.get(vocab.elementAtIndex(currentWordIndex));
-        } else {
-            l1 = getRandomSyn0Vec(vectorLength, (long) currentWordIndex);
-        }
-
-        //
-        for (int i = 0; i < w1.getCodeLength(); i++) {
-            int code = w1.getCodes().get(i);
-            int point = w1.getPoints().get(i);
-            if (point < 0)
-                throw new IllegalStateException("Illegal point " + point);
-            // Point to
-            INDArray syn1;
-            if (pointSyn1VecMap.containsKey(point)) {
-                syn1 = pointSyn1VecMap.get(point);
-            } else {
-                syn1 = Nd4j.zeros(1, vectorLength); // 1 row of vector length of zeros
-                pointSyn1VecMap.put(point, syn1);
-            }
-
-            // Dot product of Syn0 and Syn1 vecs
-            double dot = Nd4j.getBlasWrapper().level1().dot(vectorLength, 1.0, l1, syn1);
-
-            if (dot < -maxExp || dot >= maxExp)
-                continue;
-
-            int idx = (int) ((dot + maxExp) * ((double) expTable.length / maxExp / 2.0));
-
-            if (idx >= expTable.length)
-                continue;
-
-            //score
-            double f = expTable[idx];
-            //gradient
-            double g = (1 - code - f) * (useAdaGrad ? w1.getGradient(i, currentSentenceAlpha, currentSentenceAlpha)
-                            : currentSentenceAlpha);
-
-
-            Nd4j.getBlasWrapper().level1().axpy(vectorLength, g, syn1, neu1e);
-            Nd4j.getBlasWrapper().level1().axpy(vectorLength, g, l1, syn1);
-        }
-
-        int target = w1.getIndex();
-        int label;
-        //negative sampling
-        if (negative > 0)
-            for (int d = 0; d < negative + 1; d++) {
-                if (d == 0)
-                    label = 1;
-                else {
-                    nextRandom.set(Math.abs(nextRandom.get() * 25214903917L + 11));
-
-                    // FIXME: int cast
-                    int idx = Math.abs((int) (nextRandom.get() >> 16) % (int) negativeHolder.getTable().length());
-
-                    target = negativeHolder.getTable().getInt(idx);
-                    if (target <= 0)
-                        target = (int) nextRandom.get() % (vocab.numWords() - 1) + 1;
-
-                    if (target == w1.getIndex())
-                        continue;
-                    label = 0;
-                }
-
-                if (target >= negativeHolder.getSyn1Neg().rows() || target < 0)
-                    continue;
-
-                double f = Nd4j.getBlasWrapper().dot(l1, negativeHolder.getSyn1Neg().slice(target));
-                double g;
-                if (f > maxExp)
-                    g = useAdaGrad ? w1.getGradient(target, (label - 1), alpha) : (label - 1) * alpha;
-                else if (f < -maxExp)
-                    g = label * (useAdaGrad ? w1.getGradient(target, alpha, alpha) : alpha);
-                else {
-                    int idx = (int) ((f + maxExp) * (expTable.length / maxExp / 2));
-                    if (idx >= expTable.length)
-                        continue;
-
-                    g = useAdaGrad ? w1.getGradient(target, label - expTable[idx], alpha)
-                                    : (label - expTable[idx]) * alpha;
-                }
-
-                Nd4j.getBlasWrapper().level1().axpy(vectorLength, g, negativeHolder.getSyn1Neg().slice(target), neu1e);
-
-                Nd4j.getBlasWrapper().level1().axpy(vectorLength, g, l1, negativeHolder.getSyn1Neg().slice(target));
-            }
-
-
-        // Updated the Syn0 vector based on gradient. Syn0 is not random anymore.
-        Nd4j.getBlasWrapper().level1().axpy(vectorLength, 1.0f, neu1e, l1);
-
-        VocabWord word = vocab.elementAtIndex(currentWordIndex);
-        indexSyn0VecMap.put(word, l1);
-    }
-
-    private INDArray getRandomSyn0Vec(int vectorLength, long lseed) {
-        /*
-            we use wordIndex as part of seed here, to guarantee that during word syn0 initialization on dwo distinct nodes, initial weights will be the same for the same word
-         */
-        return Nd4j.rand( new int[] {1, vectorLength}, lseed * seed).subi(0.5).divi(vectorLength);
-    }
-}
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/SecondIterationFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/SecondIterationFunction.java
index 50ec16ff9..c34156484 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/SecondIterationFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp/src/main/java/org/deeplearning4j/spark/models/embeddings/word2vec/SecondIterationFunction.java
@@ -16,9 +16,8 @@
 
 package org.deeplearning4j.spark.models.embeddings.word2vec;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.deeplearning4j.models.word2vec.VocabWord;
 import org.deeplearning4j.models.word2vec.wordstore.VocabCache;
 import org.nd4j.linalg.api.ndarray.INDArray;
@@ -37,22 +36,7 @@ import java.util.concurrent.atomic.AtomicLong;
  * @author jeffreytang
  * @author raver119@gmail.com
  */
-public class SecondIterationFunction extends
-                BaseFlatMapFunctionAdaptee<Iterator<Tuple2<List<VocabWord>, Long>>, Entry<VocabWord, INDArray>> {
-
-    public SecondIterationFunction(Broadcast<Map<String, Object>> word2vecVarMapBroadcast,
-                    Broadcast<double[]> expTableBroadcast, Broadcast<VocabCache<VocabWord>> vocabCacheBroadcast) {
-        super(new SecondIterationFunctionAdapter(word2vecVarMapBroadcast, expTableBroadcast, vocabCacheBroadcast));
-    }
-}
-
-
-/**
- * @author jeffreytang
- * @author raver119@gmail.com
- */
-class SecondIterationFunctionAdapter
-                implements FlatMapFunctionAdapter<Iterator<Tuple2<List<VocabWord>, Long>>, Entry<VocabWord, INDArray>> {
+public class SecondIterationFunction implements FlatMapFunction<Iterator<Tuple2<List<VocabWord>, Long>>, Entry<VocabWord, INDArray>> {
 
     private int ithIteration = 1;
     private int vectorLength;
@@ -78,7 +62,7 @@ class SecondIterationFunctionAdapter
 
 
 
-    public SecondIterationFunctionAdapter(Broadcast<Map<String, Object>> word2vecVarMapBroadcast,
+    public SecondIterationFunction(Broadcast<Map<String, Object>> word2vecVarMapBroadcast,
                     Broadcast<double[]> expTableBroadcast, Broadcast<VocabCache<VocabWord>> vocabCacheBroadcast) {
 
         Map<String, Object> word2vecVarMap = word2vecVarMapBroadcast.getValue();
@@ -110,7 +94,7 @@ class SecondIterationFunctionAdapter
 
 
     @Override
-    public Iterable<Entry<VocabWord, INDArray>> call(Iterator<Tuple2<List<VocabWord>, Long>> pairIter) {
+    public Iterator<Entry<VocabWord, INDArray>> call(Iterator<Tuple2<List<VocabWord>, Long>> pairIter) {
         this.vocabHolder = VocabHolder.getInstance();
         this.vocabHolder.setSeed(seed, vectorLength);
 
@@ -139,7 +123,7 @@ class SecondIterationFunctionAdapter
                 }
             }
         }
-        return vocabHolder.getSplit(vocab);
+        return vocabHolder.getSplit(vocab).iterator();
     }
 
 
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/pom.xml b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/pom.xml
index 0d82bf48c..d8f425286 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/pom.xml
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/pom.xml
@@ -32,42 +32,6 @@
         <project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
     </properties>
 
-    <dependencyManagement>
-        <dependencies>
-            <dependency>
-                <!-- Force Jackson versions to match Spark. This usually only matters for Spark local, not Spark submit -->
-                <groupId>com.fasterxml.jackson.core</groupId>
-                <artifactId>jackson-core</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.core</groupId>
-                <artifactId>jackson-databind</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.core</groupId>
-                <artifactId>jackson-annotations</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.module</groupId>
-                <artifactId>jackson-module-scala_${scala.binary.version}</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.datatype</groupId>
-                <artifactId>jackson-datatype-jdk8</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-            <dependency>
-                <groupId>com.fasterxml.jackson.datatype</groupId>
-                <artifactId>jackson-datatype-jsr310</artifactId>
-                <version>${spark2.jackson.version}</version>
-            </dependency>
-        </dependencies>
-    </dependencyManagement>
-
     <dependencies>
         <dependency>
             <groupId>org.nd4j</groupId>
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapDataSet.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapDataSet.java
index 135cfa1c4..900f0e63b 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapDataSet.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapDataSet.java
@@ -16,8 +16,7 @@
 
 package org.deeplearning4j.spark.parameterserver.functions;
 
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.deeplearning4j.spark.api.TrainingResult;
 import org.deeplearning4j.spark.api.TrainingWorker;
 import org.deeplearning4j.spark.parameterserver.pw.SharedTrainingWrapper;
@@ -32,28 +31,20 @@ import java.util.Iterator;
  * @author raver119@gmail.com
  */
 
-public class SharedFlatMapDataSet<R extends TrainingResult> extends BaseFlatMapFunctionAdaptee<Iterator<DataSet>, R> {
-
-    public SharedFlatMapDataSet(TrainingWorker<R> worker) {
-        super(new SharedFlatMapDataSetAdapter<R>(worker));
-    }
-}
-
-
-class SharedFlatMapDataSetAdapter<R extends TrainingResult> implements FlatMapFunctionAdapter<Iterator<DataSet>, R> {
+public class SharedFlatMapDataSet<R extends TrainingResult> implements FlatMapFunction<Iterator<DataSet>, R> {
 
     private final SharedTrainingWorker worker;
 
-    public SharedFlatMapDataSetAdapter(TrainingWorker<R> worker) {
+    public SharedFlatMapDataSet(TrainingWorker<R> worker) {
         // we're not going to have anything but Shared classes here ever
         this.worker = (SharedTrainingWorker) worker;
     }
 
     @Override
-    public Iterable<R> call(Iterator<DataSet> dataSetIterator) throws Exception {
+    public Iterator<R> call(Iterator<DataSet> dataSetIterator) throws Exception {
         //Under some limited circumstances, we might have an empty partition. In this case, we should return immediately
         if(!dataSetIterator.hasNext()){
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         /*
@@ -70,6 +61,6 @@ class SharedFlatMapDataSetAdapter<R extends TrainingResult> implements FlatMapFu
         // all threads in this executor will be blocked here until training finished
         SharedTrainingResult result = SharedTrainingWrapper.getInstance(worker.getInstanceId()).run(worker);
 
-        return Collections.singletonList((R) result);
+        return Collections.singletonList((R) result).iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapMultiDataSet.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapMultiDataSet.java
index d7de24f15..5ce338b0f 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapMultiDataSet.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapMultiDataSet.java
@@ -16,8 +16,7 @@
 
 package org.deeplearning4j.spark.parameterserver.functions;
 
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.deeplearning4j.spark.api.TrainingResult;
 import org.deeplearning4j.spark.api.TrainingWorker;
 import org.deeplearning4j.spark.parameterserver.pw.SharedTrainingWrapper;
@@ -31,30 +30,20 @@ import java.util.Iterator;
 /**
  * Created by raver119 on 13.06.17.
  */
-public class SharedFlatMapMultiDataSet<R extends TrainingResult>
-                extends BaseFlatMapFunctionAdaptee<Iterator<MultiDataSet>, R> {
-
-    public SharedFlatMapMultiDataSet(TrainingWorker<R> worker) {
-        super(new SharedFlatMapMultiDataSetAdapter<R>(worker));
-    }
-}
-
-
-class SharedFlatMapMultiDataSetAdapter<R extends TrainingResult>
-                implements FlatMapFunctionAdapter<Iterator<MultiDataSet>, R> {
+public class SharedFlatMapMultiDataSet<R extends TrainingResult> implements FlatMapFunction<Iterator<MultiDataSet>, R> {
 
     private final SharedTrainingWorker worker;
 
-    public SharedFlatMapMultiDataSetAdapter(TrainingWorker<R> worker) {
+    public SharedFlatMapMultiDataSet(TrainingWorker<R> worker) {
         // we're not going to have anything but Shared classes here ever
         this.worker = (SharedTrainingWorker) worker;
     }
 
     @Override
-    public Iterable<R> call(Iterator<MultiDataSet> dataSetIterator) throws Exception {
+    public Iterator<R> call(Iterator<MultiDataSet> dataSetIterator) throws Exception {
         //Under some limited circumstances, we might have an empty partition. In this case, we should return immediately
         if(!dataSetIterator.hasNext()){
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
         /*
             That's the place where we do our stuff. Here's the plan:
@@ -70,6 +59,6 @@ class SharedFlatMapMultiDataSetAdapter<R extends TrainingResult>
         // all threads in this executor will be blocked here until training finished
         SharedTrainingResult result = SharedTrainingWrapper.getInstance(worker.getInstanceId()).run(worker);
 
-        return Collections.singletonList((R) result);
+        return Collections.singletonList((R) result).iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapPaths.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapPaths.java
index 5028e077f..4c8192ae7 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapPaths.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapPaths.java
@@ -18,9 +18,8 @@ package org.deeplearning4j.spark.parameterserver.functions;
 
 import org.apache.commons.io.LineIterator;
 import org.apache.hadoop.conf.Configuration;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.datavec.spark.util.SerializableHadoopConfig;
 import org.deeplearning4j.api.loader.DataSetLoader;
 import org.deeplearning4j.spark.api.TrainingResult;
@@ -39,11 +38,7 @@ import java.util.Iterator;
  *
  * @author raver119@gmail.com
  */
-public class SharedFlatMapPaths<R extends TrainingResult> extends BaseFlatMapFunctionAdaptee<Iterator<String>, R> {
-
-    public SharedFlatMapPaths(TrainingWorker<R> worker, DataSetLoader loader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
-        super(new SharedFlatMapPathsAdapter<R>(worker, loader, hadoopConfig));
-    }
+public class SharedFlatMapPaths<R extends TrainingResult> implements FlatMapFunction<Iterator<String>, R> {
 
     public static File toTempFile(Iterator<String> dataSetIterator) throws IOException {
         File f = Files.createTempFile("SharedFlatMapPaths",".txt").toFile();
@@ -56,17 +51,14 @@ public class SharedFlatMapPaths<R extends TrainingResult> extends BaseFlatMapFun
         }
         return f;
     }
-}
 
-
-class SharedFlatMapPathsAdapter<R extends TrainingResult> implements FlatMapFunctionAdapter<Iterator<String>, R> {
     public static Configuration defaultConfig;
 
     protected final SharedTrainingWorker worker;
     protected final DataSetLoader loader;
     protected final Broadcast<SerializableHadoopConfig> hadoopConfig;
 
-    public SharedFlatMapPathsAdapter(TrainingWorker<R> worker, DataSetLoader loader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
+    public SharedFlatMapPaths(TrainingWorker<R> worker, DataSetLoader loader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
         // we're not going to have anything but Shared classes here ever
         this.worker = (SharedTrainingWorker) worker;
         this.loader = loader;
@@ -74,10 +66,10 @@ class SharedFlatMapPathsAdapter<R extends TrainingResult> implements FlatMapFunc
     }
 
     @Override
-    public Iterable<R> call(Iterator<String> dataSetIterator) throws Exception {
+    public Iterator<R> call(Iterator<String> dataSetIterator) throws Exception {
         //Under some limited circumstances, we might have an empty partition. In this case, we should return immediately
         if(!dataSetIterator.hasNext()){
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
         // here we'll be converting out Strings coming out of iterator to DataSets
         // PathSparkDataSetIterator does that for us
@@ -93,7 +85,7 @@ class SharedFlatMapPathsAdapter<R extends TrainingResult> implements FlatMapFunc
             // first callee will become master, others will obey and die
             SharedTrainingResult result = SharedTrainingWrapper.getInstance(worker.getInstanceId()).run(worker);
 
-            return Collections.singletonList((R) result);
+            return Collections.singletonList((R) result).iterator();
         } finally {
             lineIter.close();
             f.delete();
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapPathsMDS.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapPathsMDS.java
index a8fbadb2b..3a8b1c213 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapPathsMDS.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/functions/SharedFlatMapPathsMDS.java
@@ -17,9 +17,8 @@
 package org.deeplearning4j.spark.parameterserver.functions;
 
 import org.apache.commons.io.LineIterator;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.datavec.spark.util.SerializableHadoopConfig;
 import org.deeplearning4j.api.loader.MultiDataSetLoader;
 import org.deeplearning4j.spark.api.TrainingResult;
@@ -37,21 +36,13 @@ import java.util.Iterator;
 /**
  * @author raver119@gmail.com
  */
-public class SharedFlatMapPathsMDS<R extends TrainingResult> extends BaseFlatMapFunctionAdaptee<Iterator<String>, R> {
-
-    public SharedFlatMapPathsMDS(TrainingWorker<R> worker, MultiDataSetLoader loader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
-        super(new SharedFlatMapPathsMDSAdapter<R>(worker, loader, hadoopConfig));
-    }
-}
-
-
-class SharedFlatMapPathsMDSAdapter<R extends TrainingResult> implements FlatMapFunctionAdapter<Iterator<String>, R> {
+public class SharedFlatMapPathsMDS<R extends TrainingResult> implements FlatMapFunction<Iterator<String>, R> {
 
     protected final SharedTrainingWorker worker;
     protected final MultiDataSetLoader loader;
     protected final Broadcast<SerializableHadoopConfig> hadoopConfig;
 
-    public SharedFlatMapPathsMDSAdapter(TrainingWorker<R> worker, MultiDataSetLoader loader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
+    public SharedFlatMapPathsMDS(TrainingWorker<R> worker, MultiDataSetLoader loader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
         // we're not going to have anything but Shared classes here ever
         this.worker = (SharedTrainingWorker) worker;
         this.loader = loader;
@@ -59,10 +50,10 @@ class SharedFlatMapPathsMDSAdapter<R extends TrainingResult> implements FlatMapF
     }
 
     @Override
-    public Iterable<R> call(Iterator<String> dataSetIterator) throws Exception {
+    public Iterator<R> call(Iterator<String> dataSetIterator) throws Exception {
         //Under some limited circumstances, we might have an empty partition. In this case, we should return immediately
         if(!dataSetIterator.hasNext()){
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
         // here we'll be converting out Strings coming out of iterator to DataSets
         // PathSparkDataSetIterator does that for us
@@ -78,7 +69,7 @@ class SharedFlatMapPathsMDSAdapter<R extends TrainingResult> implements FlatMapF
             // first callee will become master, others will obey and die
             SharedTrainingResult result = SharedTrainingWrapper.getInstance(worker.getInstanceId()).run(worker);
 
-            return Collections.singletonList((R) result);
+            return Collections.singletonList((R) result).iterator();
         } finally {
             lineIter.close();
             f.delete();
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerFlatMap.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerFlatMap.java
index e2c371c06..c3910968e 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerFlatMap.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerFlatMap.java
@@ -16,8 +16,7 @@
 
 package org.deeplearning4j.spark.api.worker;
 
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.nd4j.linalg.dataset.AsyncDataSetIterator;
 import org.deeplearning4j.datasets.iterator.IteratorDataSetIterator;
 import org.deeplearning4j.nn.graph.ComputationGraph;
@@ -41,30 +40,16 @@ import java.util.Iterator;
  *
  * @author Alex Black
  */
-public class ExecuteWorkerFlatMap<R extends TrainingResult> extends BaseFlatMapFunctionAdaptee<Iterator<DataSet>, R> {
-
-    public ExecuteWorkerFlatMap(TrainingWorker<R> worker) {
-        super(new ExecuteWorkerFlatMapAdapter<R>(worker));
-    }
-}
-
-
-/**
- * A FlatMapFunction for executing training on DataSets.
- * Used in both SparkDl4jMultiLayer and SparkComputationGraph implementations
- *
- * @author Alex Black
- */
-class ExecuteWorkerFlatMapAdapter<R extends TrainingResult> implements FlatMapFunctionAdapter<Iterator<DataSet>, R> {
+public class ExecuteWorkerFlatMap<R extends TrainingResult> implements FlatMapFunction<Iterator<DataSet>, R> {
 
     private final TrainingWorker<R> worker;
 
-    public ExecuteWorkerFlatMapAdapter(TrainingWorker<R> worker) {
+    public ExecuteWorkerFlatMap(TrainingWorker<R> worker) {
         this.worker = worker;
     }
 
     @Override
-    public Iterable<R> call(Iterator<DataSet> dataSetIterator) throws Exception {
+    public Iterator<R> call(Iterator<DataSet> dataSetIterator) throws Exception {
         WorkerConfiguration dataConfig = worker.getDataConfiguration();
         final boolean isGraph = dataConfig.isGraphNetwork();
 
@@ -79,9 +64,9 @@ class ExecuteWorkerFlatMapAdapter<R extends TrainingResult> implements FlatMapFu
 
                 Pair<R, SparkTrainingStats> pair = worker.getFinalResultNoDataWithStats();
                 pair.getFirst().setStats(s.build(pair.getSecond()));
-                return Collections.singletonList(pair.getFirst());
+                return Collections.singletonList(pair.getFirst()).iterator();
             } else {
-                return Collections.singletonList(worker.getFinalResultNoData());
+                return Collections.singletonList(worker.getFinalResultNoData()).iterator();
             }
         }
 
@@ -131,7 +116,7 @@ class ExecuteWorkerFlatMapAdapter<R extends TrainingResult> implements FlatMapFu
                         SparkTrainingStats returnStats = s.build(workerStats);
                         result.getFirst().setStats(returnStats);
 
-                        return Collections.singletonList(result.getFirst());
+                        return Collections.singletonList(result.getFirst()).iterator();
                     }
                 } else {
                     R result;
@@ -141,7 +126,7 @@ class ExecuteWorkerFlatMapAdapter<R extends TrainingResult> implements FlatMapFu
                         result = worker.processMinibatch(next, net, !batchedIterator.hasNext());
                     if (result != null) {
                         //Terminate training immediately
-                        return Collections.singletonList(result);
+                        return Collections.singletonList(result).iterator();
                     }
                 }
             }
@@ -155,12 +140,12 @@ class ExecuteWorkerFlatMapAdapter<R extends TrainingResult> implements FlatMapFu
                 else
                     pair = worker.getFinalResultWithStats(net);
                 pair.getFirst().setStats(s.build(pair.getSecond()));
-                return Collections.singletonList(pair.getFirst());
+                return Collections.singletonList(pair.getFirst()).iterator();
             } else {
                 if (isGraph)
-                    return Collections.singletonList(worker.getFinalResult(graph));
+                    return Collections.singletonList(worker.getFinalResult(graph)).iterator();
                 else
-                    return Collections.singletonList(worker.getFinalResult(net));
+                    return Collections.singletonList(worker.getFinalResult(net)).iterator();
             }
         } finally {
             //Make sure we shut down the async thread properly...
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerMultiDataSetFlatMap.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerMultiDataSetFlatMap.java
index 15fcd6b89..6fa148394 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerMultiDataSetFlatMap.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerMultiDataSetFlatMap.java
@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.spark.api.worker;
 
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
+import lombok.AllArgsConstructor;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.nd4j.linalg.dataset.AsyncMultiDataSetIterator;
 import org.deeplearning4j.datasets.iterator.IteratorMultiDataSetIterator;
 import org.deeplearning4j.nn.graph.ComputationGraph;
@@ -39,31 +39,13 @@ import java.util.Iterator;
  *
  * @author Alex Black
  */
-public class ExecuteWorkerMultiDataSetFlatMap<R extends TrainingResult>
-                extends BaseFlatMapFunctionAdaptee<Iterator<MultiDataSet>, R> {
-
-    public ExecuteWorkerMultiDataSetFlatMap(TrainingWorker<R> worker) {
-        super(new ExecuteWorkerMultiDataSetFlatMapAdapter<>(worker));
-    }
-}
-
-
-/**
- * A FlatMapFunction for executing training on MultiDataSets. Used only in SparkComputationGraph implementation.
- *
- * @author Alex Black
- */
-class ExecuteWorkerMultiDataSetFlatMapAdapter<R extends TrainingResult>
-                implements FlatMapFunctionAdapter<Iterator<MultiDataSet>, R> {
+@AllArgsConstructor
+public class ExecuteWorkerMultiDataSetFlatMap<R extends TrainingResult> implements FlatMapFunction<Iterator<MultiDataSet>, R> {
 
     private final TrainingWorker<R> worker;
 
-    public ExecuteWorkerMultiDataSetFlatMapAdapter(TrainingWorker<R> worker) {
-        this.worker = worker;
-    }
-
     @Override
-    public Iterable<R> call(Iterator<MultiDataSet> dataSetIterator) throws Exception {
+    public Iterator<R> call(Iterator<MultiDataSet> dataSetIterator) throws Exception {
         WorkerConfiguration dataConfig = worker.getDataConfiguration();
 
         boolean stats = dataConfig.isCollectTrainingStats();
@@ -75,7 +57,7 @@ class ExecuteWorkerMultiDataSetFlatMapAdapter<R extends TrainingResult>
             if (stats)
                 s.logReturnTime();
             //TODO return the results...
-            return Collections.emptyList(); //Sometimes: no data
+            return Collections.emptyIterator(); //Sometimes: no data
         }
 
         int batchSize = dataConfig.getBatchSizePerWorker();
@@ -118,13 +100,13 @@ class ExecuteWorkerMultiDataSetFlatMapAdapter<R extends TrainingResult>
                         SparkTrainingStats returnStats = s.build(workerStats);
                         result.getFirst().setStats(returnStats);
 
-                        return Collections.singletonList(result.getFirst());
+                        return Collections.singletonList(result.getFirst()).iterator();
                     }
                 } else {
                     R result = worker.processMinibatch(next, net, !batchedIterator.hasNext());
                     if (result != null) {
                         //Terminate training immediately
-                        return Collections.singletonList(result);
+                        return Collections.singletonList(result).iterator();
                     }
                 }
             }
@@ -134,9 +116,9 @@ class ExecuteWorkerMultiDataSetFlatMapAdapter<R extends TrainingResult>
                 s.logReturnTime();
                 Pair<R, SparkTrainingStats> pair = worker.getFinalResultWithStats(net);
                 pair.getFirst().setStats(s.build(pair.getSecond()));
-                return Collections.singletonList(pair.getFirst());
+                return Collections.singletonList(pair.getFirst()).iterator();
             } else {
-                return Collections.singletonList(worker.getFinalResult(net));
+                return Collections.singletonList(worker.getFinalResult(net)).iterator();
             }
         } finally {
             Nd4j.getExecutioner().commit();
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPDSFlatMap.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPDSFlatMap.java
index 2e1bdb646..4969a055b 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPDSFlatMap.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPDSFlatMap.java
@@ -16,9 +16,8 @@
 
 package org.deeplearning4j.spark.api.worker;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.input.PortableDataStream;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.deeplearning4j.spark.api.TrainingResult;
 import org.deeplearning4j.spark.api.TrainingWorker;
 import org.deeplearning4j.spark.iterator.PortableDataStreamDataSetIterator;
@@ -33,32 +32,15 @@ import java.util.Iterator;
  * @author Alex Black
  */
 @Deprecated
-public class ExecuteWorkerPDSFlatMap<R extends TrainingResult>
-                extends BaseFlatMapFunctionAdaptee<Iterator<PortableDataStream>, R> {
+public class ExecuteWorkerPDSFlatMap<R extends TrainingResult> implements FlatMapFunction<Iterator<PortableDataStream>, R> {
+    private final FlatMapFunction<Iterator<DataSet>, R> workerFlatMap;
 
     public ExecuteWorkerPDSFlatMap(TrainingWorker<R> worker) {
-        super(new ExecuteWorkerPDSFlatMapAdapter<>(worker));
-    }
-}
-
-
-/**
- * A FlatMapFunction for executing training on serialized DataSet objects, that can be loaded using a PortableDataStream
- * Used in both SparkDl4jMultiLayer and SparkComputationGraph implementations
- *
- * @author Alex Black
- */
-@Deprecated
-class ExecuteWorkerPDSFlatMapAdapter<R extends TrainingResult>
-                implements FlatMapFunctionAdapter<Iterator<PortableDataStream>, R> {
-    private final FlatMapFunctionAdapter<Iterator<DataSet>, R> workerFlatMap;
-
-    public ExecuteWorkerPDSFlatMapAdapter(TrainingWorker<R> worker) {
-        this.workerFlatMap = new ExecuteWorkerFlatMapAdapter<>(worker);
+        this.workerFlatMap = new ExecuteWorkerFlatMap<>(worker);
     }
 
     @Override
-    public Iterable<R> call(Iterator<PortableDataStream> iter) throws Exception {
+    public Iterator<R> call(Iterator<PortableDataStream> iter) throws Exception {
         return workerFlatMap.call(new PortableDataStreamDataSetIterator(iter));
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPDSMDSFlatMap.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPDSMDSFlatMap.java
index ebc4d4691..63b82bdaa 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPDSMDSFlatMap.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPDSMDSFlatMap.java
@@ -16,9 +16,8 @@
 
 package org.deeplearning4j.spark.api.worker;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.input.PortableDataStream;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.deeplearning4j.spark.api.TrainingResult;
 import org.deeplearning4j.spark.api.TrainingWorker;
 import org.deeplearning4j.spark.iterator.PortableDataStreamMultiDataSetIterator;
@@ -33,32 +32,15 @@ import java.util.Iterator;
  * @author Alex Black
  */
 @Deprecated
-public class ExecuteWorkerPDSMDSFlatMap<R extends TrainingResult>
-                extends BaseFlatMapFunctionAdaptee<Iterator<PortableDataStream>, R> {
+public class ExecuteWorkerPDSMDSFlatMap<R extends TrainingResult> implements FlatMapFunction<Iterator<PortableDataStream>, R> {
+    private final FlatMapFunction<Iterator<MultiDataSet>, R> workerFlatMap;
 
     public ExecuteWorkerPDSMDSFlatMap(TrainingWorker<R> worker) {
-        super(new ExecuteWorkerPDSMDSFlatMapAdapter<>(worker));
-    }
-}
-
-
-/**
- * A FlatMapFunction for executing training on serialized MultiDataSet objects, that can be loaded using a PortableDataStream
- * Used for SparkComputationGraph implementations only
- *
- * @author Alex Black
- */
-@Deprecated
-class ExecuteWorkerPDSMDSFlatMapAdapter<R extends TrainingResult>
-                implements FlatMapFunctionAdapter<Iterator<PortableDataStream>, R> {
-    private final FlatMapFunctionAdapter<Iterator<MultiDataSet>, R> workerFlatMap;
-
-    public ExecuteWorkerPDSMDSFlatMapAdapter(TrainingWorker<R> worker) {
-        this.workerFlatMap = new ExecuteWorkerMultiDataSetFlatMapAdapter<>(worker);
+        this.workerFlatMap = new ExecuteWorkerMultiDataSetFlatMap<>(worker);
     }
 
     @Override
-    public Iterable<R> call(Iterator<PortableDataStream> iter) throws Exception {
+    public Iterator<R> call(Iterator<PortableDataStream> iter) throws Exception {
         return workerFlatMap.call(new PortableDataStreamMultiDataSetIterator(iter));
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPathFlatMap.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPathFlatMap.java
index 5e3889a99..e26615ab8 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPathFlatMap.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPathFlatMap.java
@@ -16,9 +16,8 @@
 
 package org.deeplearning4j.spark.api.worker;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.datavec.spark.util.SerializableHadoopConfig;
 import org.deeplearning4j.api.loader.DataSetLoader;
 import org.deeplearning4j.spark.api.TrainingResult;
@@ -38,30 +37,15 @@ import java.util.List;
  *
  * @author Alex Black
  */
-public class ExecuteWorkerPathFlatMap<R extends TrainingResult>
-                extends BaseFlatMapFunctionAdaptee<Iterator<String>, R> {
+public class ExecuteWorkerPathFlatMap<R extends TrainingResult> implements FlatMapFunction<Iterator<String>, R> {
 
-    public ExecuteWorkerPathFlatMap(TrainingWorker<R> worker, DataSetLoader loader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
-        super(new ExecuteWorkerPathFlatMapAdapter<>(worker, loader, hadoopConfig));
-    }
-}
-
-
-/**
- * A FlatMapFunction for executing training on serialized DataSet objects, that can be loaded from a path (local or HDFS)
- * that is specified as a String
- * Used in both SparkDl4jMultiLayer and SparkComputationGraph implementations
- *
- * @author Alex Black
- */
-class ExecuteWorkerPathFlatMapAdapter<R extends TrainingResult> implements FlatMapFunctionAdapter<Iterator<String>, R> {
-    private final FlatMapFunctionAdapter<Iterator<DataSet>, R> workerFlatMap;
+    private final FlatMapFunction<Iterator<DataSet>, R> workerFlatMap;
     private final DataSetLoader dataSetLoader;
     private final int maxDataSetObjects;
     private final Broadcast<SerializableHadoopConfig> hadoopConfig;
 
-    public ExecuteWorkerPathFlatMapAdapter(TrainingWorker<R> worker, DataSetLoader dataSetLoader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
-        this.workerFlatMap = new ExecuteWorkerFlatMapAdapter<>(worker);
+    public ExecuteWorkerPathFlatMap(TrainingWorker<R> worker, DataSetLoader dataSetLoader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
+        this.workerFlatMap = new ExecuteWorkerFlatMap<>(worker);
         this.dataSetLoader = dataSetLoader;
         this.hadoopConfig = hadoopConfig;
 
@@ -84,7 +68,7 @@ class ExecuteWorkerPathFlatMapAdapter<R extends TrainingResult> implements FlatM
     }
 
     @Override
-    public Iterable<R> call(Iterator<String> iter) throws Exception {
+    public Iterator<R> call(Iterator<String> iter) throws Exception {
         List<String> list = new ArrayList<>();
         int count = 0;
         while (iter.hasNext() && count++ < maxDataSetObjects) {
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPathMDSFlatMap.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPathMDSFlatMap.java
index 072425f18..47e53bd6c 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPathMDSFlatMap.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/api/worker/ExecuteWorkerPathMDSFlatMap.java
@@ -16,9 +16,8 @@
 
 package org.deeplearning4j.spark.api.worker;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.datavec.spark.util.SerializableHadoopConfig;
 import org.deeplearning4j.api.loader.MultiDataSetLoader;
 import org.deeplearning4j.spark.api.TrainingResult;
@@ -38,31 +37,14 @@ import java.util.List;
  *
  * @author Alex Black
  */
-public class ExecuteWorkerPathMDSFlatMap<R extends TrainingResult>
-                extends BaseFlatMapFunctionAdaptee<Iterator<String>, R> {
-
-    public ExecuteWorkerPathMDSFlatMap(TrainingWorker<R> worker, MultiDataSetLoader loader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
-        super(new ExecuteWorkerPathMDSFlatMapAdapter<>(worker, loader, hadoopConfig));
-    }
-}
-
-
-/**
- * A FlatMapFunction for executing training on serialized DataSet objects, that can be loaded from a path (local or HDFS)
- * that is specified as a String
- * Used in both SparkDl4jMultiLayer and SparkComputationGraph implementations
- *
- * @author Alex Black
- */
-class ExecuteWorkerPathMDSFlatMapAdapter<R extends TrainingResult>
-                implements FlatMapFunctionAdapter<Iterator<String>, R> {
-    private final FlatMapFunctionAdapter<Iterator<MultiDataSet>, R> workerFlatMap;
+public class ExecuteWorkerPathMDSFlatMap<R extends TrainingResult> implements FlatMapFunction<Iterator<String>, R> {
+    private final FlatMapFunction<Iterator<MultiDataSet>, R> workerFlatMap;
     private MultiDataSetLoader loader;
     private final int maxDataSetObjects;
     private final Broadcast<SerializableHadoopConfig> hadoopConfig;
 
-    public ExecuteWorkerPathMDSFlatMapAdapter(TrainingWorker<R> worker, MultiDataSetLoader loader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
-        this.workerFlatMap = new ExecuteWorkerMultiDataSetFlatMapAdapter<>(worker);
+    public ExecuteWorkerPathMDSFlatMap(TrainingWorker<R> worker, MultiDataSetLoader loader, Broadcast<SerializableHadoopConfig> hadoopConfig) {
+        this.workerFlatMap = new ExecuteWorkerMultiDataSetFlatMap<>(worker);
         this.loader = loader;
         this.hadoopConfig = hadoopConfig;
 
@@ -85,7 +67,7 @@ class ExecuteWorkerPathMDSFlatMapAdapter<R extends TrainingResult>
     }
 
     @Override
-    public Iterable<R> call(Iterator<String> iter) throws Exception {
+    public Iterator<R> call(Iterator<String> iter) throws Exception {
         List<String> list = new ArrayList<>();
         int count = 0;
         while (iter.hasNext() && count++ < maxDataSetObjects) {
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/BatchDataSetsFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/BatchDataSetsFunction.java
index fd3a5a2fc..9513193e4 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/BatchDataSetsFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/BatchDataSetsFunction.java
@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.spark.data;
 
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
+import lombok.AllArgsConstructor;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.nd4j.linalg.dataset.DataSet;
 
 import java.util.ArrayList;
@@ -38,37 +38,12 @@ import java.util.List;
  *
  * @author Alex Black
  */
-public class BatchDataSetsFunction extends BaseFlatMapFunctionAdaptee<Iterator<DataSet>, DataSet> {
-
-    public BatchDataSetsFunction(int minibatchSize) {
-        super(new BatchDataSetsFunctionAdapter(minibatchSize));
-    }
-}
-
-
-/**
- * Function used to batch DataSet objects together. Typically used to combine singe-example DataSet objects out of
- * something like {@link org.deeplearning4j.spark.datavec.DataVecDataSetFunction} together into minibatches.<br>
- *
- * Usage:
- * <pre>
- * {@code
- *      RDD<DataSet> mySingleExampleDataSets = ...;
- *      RDD<DataSet> batchData = mySingleExampleDataSets.mapPartitions(new BatchDataSetsFunction(batchSize));
- * }
- * </pre>
- *
- * @author Alex Black
- */
-class BatchDataSetsFunctionAdapter implements FlatMapFunctionAdapter<Iterator<DataSet>, DataSet> {
+@AllArgsConstructor
+public class BatchDataSetsFunction implements FlatMapFunction<Iterator<DataSet>, DataSet> {
     private final int minibatchSize;
 
-    public BatchDataSetsFunctionAdapter(int minibatchSize) {
-        this.minibatchSize = minibatchSize;
-    }
-
     @Override
-    public Iterable<DataSet> call(Iterator<DataSet> iter) throws Exception {
+    public Iterator<DataSet> call(Iterator<DataSet> iter) throws Exception {
         List<DataSet> out = new ArrayList<>();
         while (iter.hasNext()) {
             List<DataSet> list = new ArrayList<>();
@@ -88,6 +63,6 @@ class BatchDataSetsFunctionAdapter implements FlatMapFunctionAdapter<Iterator<Da
 
             out.add(next);
         }
-        return out;
+        return out.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/SplitDataSetsFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/SplitDataSetsFunction.java
index b4c87677f..026b65dd6 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/SplitDataSetsFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/SplitDataSetsFunction.java
@@ -16,8 +16,7 @@
 
 package org.deeplearning4j.spark.data;
 
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.nd4j.linalg.dataset.DataSet;
 
 import java.util.ArrayList;
@@ -37,34 +36,13 @@ import java.util.List;
  *
  * @author Alex Black
  */
-public class SplitDataSetsFunction extends BaseFlatMapFunctionAdaptee<Iterator<DataSet>, DataSet> {
-
-    public SplitDataSetsFunction() {
-        super(new SplitDataSetsFunctionAdapter());
-    }
-}
-
-
-/**
- * Take an existing DataSet object, and split it into multiple DataSet objects with one example in each
- *
- * Usage:
- * <pre>
- * {@code
- *      RDD<DataSet> myBatchedExampleDataSets = ...;
- *      RDD<DataSet> singleExamlpeDataSets = myBatchedExampleDataSets.mapPartitions(new SplitDataSets(batchSize));
- * }
- * </pre>
- *
- * @author Alex Black
- */
-class SplitDataSetsFunctionAdapter implements FlatMapFunctionAdapter<Iterator<DataSet>, DataSet> {
+public class SplitDataSetsFunction implements FlatMapFunction<Iterator<DataSet>, DataSet> {
     @Override
-    public Iterable<DataSet> call(Iterator<DataSet> dataSetIterator) throws Exception {
+    public Iterator<DataSet> call(Iterator<DataSet> dataSetIterator) throws Exception {
         List<DataSet> out = new ArrayList<>();
         while (dataSetIterator.hasNext()) {
             out.addAll(dataSetIterator.next().asList());
         }
-        return out;
+        return out.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/shuffle/SplitDataSetExamplesPairFlatMapFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/shuffle/SplitDataSetExamplesPairFlatMapFunction.java
index e0557d10c..1ccf54b91 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/shuffle/SplitDataSetExamplesPairFlatMapFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/data/shuffle/SplitDataSetExamplesPairFlatMapFunction.java
@@ -17,12 +17,13 @@
 package org.deeplearning4j.spark.data.shuffle;
 
 import org.apache.spark.api.java.JavaRDD;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.deeplearning4j.spark.util.BasePairFlatMapFunctionAdaptee;
+import org.apache.spark.api.java.function.FlatMapFunction;
+import org.apache.spark.api.java.function.PairFlatMapFunction;
 import org.nd4j.linalg.dataset.DataSet;
 import scala.Tuple2;
 
 import java.util.ArrayList;
+import java.util.Iterator;
 import java.util.List;
 import java.util.Random;
 
@@ -34,34 +35,17 @@ import java.util.Random;
  *
  * @author Alex Black
  */
-public class SplitDataSetExamplesPairFlatMapFunction extends BasePairFlatMapFunctionAdaptee<DataSet, Integer, DataSet> {
-
-    public SplitDataSetExamplesPairFlatMapFunction(int maxKeyIndex) {
-        super(new SplitDataSetExamplesPairFlatMapFunctionAdapter(maxKeyIndex));
-    }
-}
-
-
-/**
- * A PairFlatMapFunction that splits each example in a {@link DataSet} object into its own {@link DataSet}.
- * Also adds a random key (integer value) in the range 0 to maxKeyIndex-1.<br>
- *
- * Used in {@link org.deeplearning4j.spark.util.SparkUtils#shuffleExamples(JavaRDD, int, int)}
- *
- * @author Alex Black
- */
-class SplitDataSetExamplesPairFlatMapFunctionAdapter
-                implements FlatMapFunctionAdapter<DataSet, Tuple2<Integer, DataSet>> {
+public class SplitDataSetExamplesPairFlatMapFunction implements PairFlatMapFunction<DataSet, Integer, DataSet> {
 
     private transient Random r;
     private int maxKeyIndex;
 
-    public SplitDataSetExamplesPairFlatMapFunctionAdapter(int maxKeyIndex) {
+    public SplitDataSetExamplesPairFlatMapFunction(int maxKeyIndex) {
         this.maxKeyIndex = maxKeyIndex;
     }
 
     @Override
-    public Iterable<Tuple2<Integer, DataSet>> call(DataSet dataSet) throws Exception {
+    public Iterator<Tuple2<Integer, DataSet>> call(DataSet dataSet) throws Exception {
         if (r == null) {
             r = new Random();
         }
@@ -72,6 +56,6 @@ class SplitDataSetExamplesPairFlatMapFunctionAdapter
             out.add(new Tuple2<>(r.nextInt(maxKeyIndex), ds));
         }
 
-        return out;
+        return out.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/datavec/RDDMiniBatches.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/datavec/RDDMiniBatches.java
index 37ef1e7ab..653bbc75d 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/datavec/RDDMiniBatches.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/datavec/RDDMiniBatches.java
@@ -16,9 +16,9 @@
 
 package org.deeplearning4j.spark.datavec;
 
+import lombok.AllArgsConstructor;
 import org.apache.spark.api.java.JavaRDD;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.nd4j.linalg.dataset.DataSet;
 
 import java.io.Serializable;
@@ -44,22 +44,12 @@ public class RDDMiniBatches implements Serializable {
         return toSplitJava.mapPartitions(new MiniBatchFunction(miniBatches));
     }
 
-    public static class MiniBatchFunction extends BaseFlatMapFunctionAdaptee<Iterator<DataSet>, DataSet> {
-
-        public MiniBatchFunction(int batchSize) {
-            super(new MiniBatchFunctionAdapter(batchSize));
-        }
-    }
-
-    static class MiniBatchFunctionAdapter implements FlatMapFunctionAdapter<Iterator<DataSet>, DataSet> {
-        private int batchSize = 10;
-
-        public MiniBatchFunctionAdapter(int batchSize) {
-            this.batchSize = batchSize;
-        }
+    @AllArgsConstructor
+    public static class MiniBatchFunction implements FlatMapFunction<Iterator<DataSet>, DataSet> {
+        private int batchSize;
 
         @Override
-        public Iterable<DataSet> call(Iterator<DataSet> dataSetIterator) throws Exception {
+        public Iterator<DataSet> call(Iterator<DataSet> dataSetIterator) throws Exception {
             List<DataSet> ret = new ArrayList<>();
             List<DataSet> temp = new ArrayList<>();
             while (dataSetIterator.hasNext()) {
@@ -74,10 +64,7 @@ public class RDDMiniBatches implements Serializable {
             if (temp.size() > 0)
                 ret.add(DataSet.merge(temp));
 
-            return ret;
+            return ret.iterator();
         }
-
     }
-
-
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/repartition/HashingBalancedPartitioner.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/repartition/HashingBalancedPartitioner.java
index 6d717371a..9412a3cbb 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/repartition/HashingBalancedPartitioner.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/repartition/HashingBalancedPartitioner.java
@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.spark.impl.common.repartition;
 
-import com.google.common.base.Predicate;
-import com.google.common.collect.Collections2;
+import org.nd4j.shade.guava.base.Predicate;
+import org.nd4j.shade.guava.collect.Collections2;
 import org.apache.spark.Partitioner;
 import scala.Tuple2;
 
@@ -25,8 +25,8 @@ import java.util.ArrayList;
 import java.util.List;
 import java.util.Random;
 
-import static com.google.common.base.Preconditions.checkArgument;
-import static com.google.common.base.Preconditions.checkNotNull;
+import static org.nd4j.shade.guava.base.Preconditions.checkArgument;
+import static org.nd4j.shade.guava.base.Preconditions.checkNotNull;
 
 /**
  * This is a custom partitioner that rebalances a minimum of elements
@@ -97,12 +97,13 @@ public class HashingBalancedPartitioner extends Partitioner {
 
     @Override
     public int numPartitions() {
-        return Collections2.filter(partitionWeightsByClass.get(0), new Predicate<Double>() {
-            @Override
-            public boolean apply(Double aDouble) {
-                return aDouble >= 0;
-            }
-        }).size();
+        List<Double> list = partitionWeightsByClass.get(0);
+        int count = 0;
+        for(Double d : list){
+            if(d >= 0)
+                count++;
+        }
+        return count;
     }
 
     @Override
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/repartition/MapTupleToPairFlatMap.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/repartition/MapTupleToPairFlatMap.java
index b540cff82..23b702444 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/repartition/MapTupleToPairFlatMap.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/repartition/MapTupleToPairFlatMap.java
@@ -16,8 +16,7 @@
 
 package org.deeplearning4j.spark.impl.common.repartition;
 
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.deeplearning4j.spark.util.BasePairFlatMapFunctionAdaptee;
+import org.apache.spark.api.java.function.PairFlatMapFunction;
 import scala.Tuple2;
 
 import java.util.ArrayList;
@@ -30,22 +29,14 @@ import java.util.List;
  *
  * @author Alex Black
  */
-public class MapTupleToPairFlatMap<T, U> extends BasePairFlatMapFunctionAdaptee<Iterator<Tuple2<T, U>>, T, U> {
-
-    public MapTupleToPairFlatMap() {
-        super(new MapTupleToPairFlatMapAdapter<T, U>());
-    }
-}
-
-
-class MapTupleToPairFlatMapAdapter<T, U> implements FlatMapFunctionAdapter<Iterator<Tuple2<T, U>>, Tuple2<T, U>> {
+public class MapTupleToPairFlatMap<T, U> implements PairFlatMapFunction<Iterator<Tuple2<T, U>>, T, U> {
 
     @Override
-    public Iterable<Tuple2<T, U>> call(Iterator<Tuple2<T, U>> tuple2Iterator) throws Exception {
+    public Iterator<Tuple2<T, U>> call(Iterator<Tuple2<T, U>> tuple2Iterator) throws Exception {
         List<Tuple2<T, U>> list = new ArrayList<>();
         while (tuple2Iterator.hasNext()) {
             list.add(tuple2Iterator.next());
         }
-        return list;
+        return list.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeReconstructionProbWithKeyFunctionAdapter.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeReconstructionProbWithKeyFunction.java
similarity index 87%
rename from deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeReconstructionProbWithKeyFunctionAdapter.java
rename to deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeReconstructionProbWithKeyFunction.java
index 58f092de8..99fcdc65b 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeReconstructionProbWithKeyFunctionAdapter.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeReconstructionProbWithKeyFunction.java
@@ -27,7 +27,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
  * @param <K> Type of key, associated with each example. Used to keep track of which score belongs to which example
  * @author Alex Black
  */
-public abstract class BaseVaeReconstructionProbWithKeyFunctionAdapter<K> extends BaseVaeScoreWithKeyFunctionAdapter<K> {
+public abstract class BaseVaeReconstructionProbWithKeyFunction<K> extends BaseVaeScoreWithKeyFunction<K> {
 
     private final boolean useLogProbability;
     private final int numSamples;
@@ -39,8 +39,8 @@ public abstract class BaseVaeReconstructionProbWithKeyFunctionAdapter<K> extends
      * @param batchSize              Batch size to use when scoring
      * @param numSamples             Number of samples to use when calling {@link VariationalAutoencoder#reconstructionLogProbability(INDArray, int)}
      */
-    public BaseVaeReconstructionProbWithKeyFunctionAdapter(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
-                    boolean useLogProbability, int batchSize, int numSamples) {
+    public BaseVaeReconstructionProbWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
+                                                    boolean useLogProbability, int batchSize, int numSamples) {
         super(params, jsonConfig, batchSize);
         this.useLogProbability = useLogProbability;
         this.numSamples = numSamples;
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeScoreWithKeyFunctionAdapter.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeScoreWithKeyFunction.java
similarity index 88%
rename from deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeScoreWithKeyFunctionAdapter.java
rename to deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeScoreWithKeyFunction.java
index 19b4e1b14..da6a374c4 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeScoreWithKeyFunctionAdapter.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/common/score/BaseVaeScoreWithKeyFunction.java
@@ -18,8 +18,9 @@ package org.deeplearning4j.spark.impl.common.score;
 
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
+import org.apache.spark.api.java.function.FlatMapFunction;
+import org.apache.spark.api.java.function.PairFlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
 import org.deeplearning4j.nn.layers.variational.VariationalAutoencoder;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.factory.Nd4j;
@@ -38,8 +39,7 @@ import java.util.List;
  * @author Alex Black
  */
 @Slf4j
-public abstract class BaseVaeScoreWithKeyFunctionAdapter<K>
-                implements FlatMapFunctionAdapter<Iterator<Tuple2<K, INDArray>>, Tuple2<K, Double>> {
+public abstract class BaseVaeScoreWithKeyFunction<K> implements PairFlatMapFunction<Iterator<Tuple2<K, INDArray>>, K, Double> {
 
     protected final Broadcast<INDArray> params;
     protected final Broadcast<String> jsonConfig;
@@ -51,7 +51,7 @@ public abstract class BaseVaeScoreWithKeyFunctionAdapter<K>
      * @param jsonConfig             MultiLayerConfiguration, as json
      * @param batchSize              Batch size to use when scoring
      */
-    public BaseVaeScoreWithKeyFunctionAdapter(Broadcast<INDArray> params, Broadcast<String> jsonConfig, int batchSize) {
+    public BaseVaeScoreWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig, int batchSize) {
         this.params = params;
         this.jsonConfig = jsonConfig;
         this.batchSize = batchSize;
@@ -63,9 +63,9 @@ public abstract class BaseVaeScoreWithKeyFunctionAdapter<K>
 
 
     @Override
-    public Iterable<Tuple2<K, Double>> call(Iterator<Tuple2<K, INDArray>> iterator) throws Exception {
+    public Iterator<Tuple2<K, Double>> call(Iterator<Tuple2<K, INDArray>> iterator) throws Exception {
         if (!iterator.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         VariationalAutoencoder vae = getVaeLayer();
@@ -108,6 +108,6 @@ public abstract class BaseVaeScoreWithKeyFunctionAdapter<K>
             log.debug("Scored {} examples ", totalCount);
         }
 
-        return ret;
+        return ret.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/evaluation/IEvaluateMDSFlatMapFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/evaluation/IEvaluateMDSFlatMapFunction.java
index dd02c5d37..cdb41ba33 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/evaluation/IEvaluateMDSFlatMapFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/evaluation/IEvaluateMDSFlatMapFunction.java
@@ -17,12 +17,10 @@
 package org.deeplearning4j.spark.impl.graph.evaluation;
 
 import lombok.extern.slf4j.Slf4j;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.deeplearning4j.spark.impl.evaluation.EvaluationRunner;
 import org.nd4j.evaluation.IEvaluation;
-import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.api.MultiDataSet;
 
 import java.util.Collections;
@@ -36,26 +34,7 @@ import java.util.concurrent.Future;
  *
  * @author Alex Black
  */
-public class IEvaluateMDSFlatMapFunction<T extends IEvaluation>
-                extends BaseFlatMapFunctionAdaptee<Iterator<MultiDataSet>, T[]> {
-
-    public IEvaluateMDSFlatMapFunction(Broadcast<String> json, Broadcast<byte[]> params, int evalNumWorkers, int evalBatchSize,
-                    T... evaluations) {
-        super(new IEvaluateMDSFlatMapFunctionAdapter<>(json, params, evalNumWorkers, evalBatchSize, evaluations));
-    }
-}
-
-
-/**
- * Function to evaluate data (using an IEvaluation instance), in a distributed manner
- * Flat map function used to batch examples for computational efficiency + reduce number of IEvaluation objects returned
- * for network efficiency.
- *
- * @author Alex Black
- */
-@Slf4j
-class IEvaluateMDSFlatMapFunctionAdapter<T extends IEvaluation>
-                implements FlatMapFunctionAdapter<Iterator<MultiDataSet>, T[]> {
+public class IEvaluateMDSFlatMapFunction<T extends IEvaluation> implements FlatMapFunction<Iterator<MultiDataSet>, T[]> {
 
     protected Broadcast<String> json;
     protected Broadcast<byte[]> params;
@@ -70,7 +49,7 @@ class IEvaluateMDSFlatMapFunctionAdapter<T extends IEvaluation>
      *                              this. Used to avoid doing too many at once (and hence memory issues)
      * @param evaluations Initial evaulation instance (i.e., empty Evaluation or RegressionEvaluation instance)
      */
-    public IEvaluateMDSFlatMapFunctionAdapter(Broadcast<String> json, Broadcast<byte[]> params, int evalNumWorkers,
+    public IEvaluateMDSFlatMapFunction(Broadcast<String> json, Broadcast<byte[]> params, int evalNumWorkers,
                                               int evalBatchSize, T[] evaluations) {
         this.json = json;
         this.params = params;
@@ -80,13 +59,13 @@ class IEvaluateMDSFlatMapFunctionAdapter<T extends IEvaluation>
     }
 
     @Override
-    public Iterable<T[]> call(Iterator<MultiDataSet> dataSetIterator) throws Exception {
+    public Iterator<T[]> call(Iterator<MultiDataSet> dataSetIterator) throws Exception {
         if (!dataSetIterator.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         if (!dataSetIterator.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         Future<IEvaluation[]> f = EvaluationRunner.getInstance().execute(
@@ -94,9 +73,9 @@ class IEvaluateMDSFlatMapFunctionAdapter<T extends IEvaluation>
 
         IEvaluation[] result = f.get();
         if(result == null){
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         } else {
-            return Collections.singletonList((T[])result);
+            return Collections.singletonList((T[])result).iterator();
         }
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/evaluation/IEvaluateMDSPathsFlatMapFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/evaluation/IEvaluateMDSPathsFlatMapFunction.java
index cea6a7ab0..d520605a3 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/evaluation/IEvaluateMDSPathsFlatMapFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/evaluation/IEvaluateMDSPathsFlatMapFunction.java
@@ -17,9 +17,8 @@
 package org.deeplearning4j.spark.impl.graph.evaluation;
 
 import lombok.extern.slf4j.Slf4j;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.datavec.spark.util.SerializableHadoopConfig;
 import org.deeplearning4j.api.loader.DataSetLoader;
 import org.deeplearning4j.api.loader.MultiDataSetLoader;
@@ -43,26 +42,7 @@ import java.util.concurrent.Future;
  *
  * @author Alex Black
  */
-public class IEvaluateMDSPathsFlatMapFunction
-                extends BaseFlatMapFunctionAdaptee<Iterator<String>, IEvaluation[]> {
-
-    public IEvaluateMDSPathsFlatMapFunction(Broadcast<String> json, Broadcast<byte[]> params, int evalNumWorkers, int evalBatchSize,
-                                            DataSetLoader dsLoader, MultiDataSetLoader mdsLoader,
-                                            Broadcast<SerializableHadoopConfig> configuration, IEvaluation... evaluations) {
-        super(new IEvaluateMDSPathsFlatMapFunctionAdapter(json, params, evalNumWorkers, evalBatchSize, dsLoader, mdsLoader, configuration, evaluations));
-    }
-}
-
-
-/**
- * Function to evaluate data (using an IEvaluation instance), in a distributed manner
- * Flat map function used to batch examples for computational efficiency + reduce number of IEvaluation objects returned
- * for network efficiency.
- *
- * @author Alex Black
- */
-@Slf4j
-class IEvaluateMDSPathsFlatMapFunctionAdapter implements FlatMapFunctionAdapter<Iterator<String>, IEvaluation[]> {
+public class IEvaluateMDSPathsFlatMapFunction implements FlatMapFunction<Iterator<String>, IEvaluation[]> {
 
     protected Broadcast<String> json;
     protected Broadcast<byte[]> params;
@@ -80,7 +60,7 @@ class IEvaluateMDSPathsFlatMapFunctionAdapter implements FlatMapFunctionAdapter<
      *                              this. Used to avoid doing too many at once (and hence memory issues)
      * @param evaluations Initial evaulation instance (i.e., empty Evaluation or RegressionEvaluation instance)
      */
-    public IEvaluateMDSPathsFlatMapFunctionAdapter(Broadcast<String> json, Broadcast<byte[]> params, int evalNumWorkers, int evalBatchSize,
+    public IEvaluateMDSPathsFlatMapFunction(Broadcast<String> json, Broadcast<byte[]> params, int evalNumWorkers, int evalBatchSize,
                                                    DataSetLoader dsLoader, MultiDataSetLoader mdsLoader, Broadcast<SerializableHadoopConfig> configuration, IEvaluation[] evaluations) {
         this.json = json;
         this.params = params;
@@ -93,9 +73,9 @@ class IEvaluateMDSPathsFlatMapFunctionAdapter implements FlatMapFunctionAdapter<
     }
 
     @Override
-    public Iterable<IEvaluation[]> call(Iterator<String> paths) throws Exception {
+    public Iterator<IEvaluation[]> call(Iterator<String> paths) throws Exception {
         if (!paths.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         MultiDataSetIterator iter;
@@ -109,9 +89,9 @@ class IEvaluateMDSPathsFlatMapFunctionAdapter implements FlatMapFunctionAdapter<
         Future<IEvaluation[]> f = EvaluationRunner.getInstance().execute(evaluations, evalNumWorkers, evalBatchSize, null, iter, true, json, params);
         IEvaluation[] result = f.get();
         if(result == null){
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         } else {
-            return Collections.singletonList(result);
+            return Collections.singletonList(result).iterator();
         }
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/CGVaeReconstructionErrorWithKeyFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/CGVaeReconstructionErrorWithKeyFunction.java
index 107b22284..0e5f01343 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/CGVaeReconstructionErrorWithKeyFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/CGVaeReconstructionErrorWithKeyFunction.java
@@ -21,7 +21,7 @@ import org.deeplearning4j.nn.api.Layer;
 import org.deeplearning4j.nn.conf.ComputationGraphConfiguration;
 import org.deeplearning4j.nn.graph.ComputationGraph;
 import org.deeplearning4j.nn.layers.variational.VariationalAutoencoder;
-import org.deeplearning4j.spark.impl.common.score.BaseVaeScoreWithKeyFunctionAdapter;
+import org.deeplearning4j.spark.impl.common.score.BaseVaeScoreWithKeyFunction;
 import org.nd4j.linalg.api.ndarray.INDArray;
 
 /**
@@ -33,7 +33,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
  * @author Alex Black
  * @see CGVaeReconstructionProbWithKeyFunction
  */
-public class CGVaeReconstructionErrorWithKeyFunction<K> extends BaseVaeScoreWithKeyFunctionAdapter<K> {
+public class CGVaeReconstructionErrorWithKeyFunction<K> extends BaseVaeScoreWithKeyFunction<K> {
 
 
     /**
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/CGVaeReconstructionProbWithKeyFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/CGVaeReconstructionProbWithKeyFunction.java
index 4db797bb2..835bb8fa7 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/CGVaeReconstructionProbWithKeyFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/CGVaeReconstructionProbWithKeyFunction.java
@@ -21,7 +21,7 @@ import org.deeplearning4j.nn.api.Layer;
 import org.deeplearning4j.nn.conf.ComputationGraphConfiguration;
 import org.deeplearning4j.nn.graph.ComputationGraph;
 import org.deeplearning4j.nn.layers.variational.VariationalAutoencoder;
-import org.deeplearning4j.spark.impl.common.score.BaseVaeReconstructionProbWithKeyFunctionAdapter;
+import org.deeplearning4j.spark.impl.common.score.BaseVaeReconstructionProbWithKeyFunction;
 import org.nd4j.linalg.api.ndarray.INDArray;
 
 /**
@@ -31,7 +31,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
  *
  * @author Alex Black
  */
-public class CGVaeReconstructionProbWithKeyFunction<K> extends BaseVaeReconstructionProbWithKeyFunctionAdapter<K> {
+public class CGVaeReconstructionProbWithKeyFunction<K> extends BaseVaeReconstructionProbWithKeyFunction<K> {
 
 
     /**
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/GraphFeedForwardWithKeyFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/GraphFeedForwardWithKeyFunction.java
index 3f84a6fb2..cec2f5b17 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/GraphFeedForwardWithKeyFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/GraphFeedForwardWithKeyFunction.java
@@ -16,11 +16,13 @@
 
 package org.deeplearning4j.spark.impl.graph.scoring;
 
+import lombok.AllArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.apache.spark.api.java.function.FlatMapFunction;
+import org.apache.spark.api.java.function.PairFlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
 import org.deeplearning4j.nn.conf.ComputationGraphConfiguration;
 import org.deeplearning4j.nn.graph.ComputationGraph;
-import org.deeplearning4j.spark.util.BasePairFlatMapFunctionAdaptee;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.indexing.NDArrayIndex;
@@ -40,48 +42,19 @@ import java.util.List;
  * @param <K> Type of key, associated with each example. Used to keep track of which output belongs to which input example
  * @author Alex Black
  */
-public class GraphFeedForwardWithKeyFunction<K>
-                extends BasePairFlatMapFunctionAdaptee<Iterator<Tuple2<K, INDArray[]>>, K, INDArray[]> {
-
-    public GraphFeedForwardWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig, int batchSize) {
-        super(new GraphFeedForwardWithKeyFunctionAdapter<K>(params, jsonConfig, batchSize));
-    }
-}
-
-
-/**
- * Function to feed-forward examples, and get the network output (for example, class probabilities).
- * A key value is used to keey track of which output corresponds to which input.
- *
- * @param <K> Type of key, associated with each example. Used to keep track of which output belongs to which input example
- * @author Alex Black
- */
-class GraphFeedForwardWithKeyFunctionAdapter<K>
-                implements FlatMapFunctionAdapter<Iterator<Tuple2<K, INDArray[]>>, Tuple2<K, INDArray[]>> {
-
-    protected static Logger log = LoggerFactory.getLogger(GraphFeedForwardWithKeyFunction.class);
+@Slf4j
+@AllArgsConstructor
+public class GraphFeedForwardWithKeyFunction<K> implements PairFlatMapFunction<Iterator<Tuple2<K, INDArray[]>>, K, INDArray[]> {
 
     private final Broadcast<INDArray> params;
     private final Broadcast<String> jsonConfig;
     private final int batchSize;
 
-    /**
-     * @param params     MultiLayerNetwork parameters
-     * @param jsonConfig MultiLayerConfiguration, as json
-     * @param batchSize  Batch size to use for forward pass (use > 1 for efficiency)
-     */
-    public GraphFeedForwardWithKeyFunctionAdapter(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
-                    int batchSize) {
-        this.params = params;
-        this.jsonConfig = jsonConfig;
-        this.batchSize = batchSize;
-    }
-
 
     @Override
-    public Iterable<Tuple2<K, INDArray[]>> call(Iterator<Tuple2<K, INDArray[]>> iterator) throws Exception {
+    public Iterator<Tuple2<K, INDArray[]>> call(Iterator<Tuple2<K, INDArray[]>> iterator) throws Exception {
         if (!iterator.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         ComputationGraph network = new ComputationGraph(ComputationGraphConfiguration.fromJson(jsonConfig.getValue()));
@@ -129,7 +102,7 @@ class GraphFeedForwardWithKeyFunctionAdapter<K>
         }
 
         if (tupleCount == 0) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         List<Tuple2<K, INDArray[]>> output = new ArrayList<>(tupleCount);
@@ -198,7 +171,7 @@ class GraphFeedForwardWithKeyFunctionAdapter<K>
 
         Nd4j.getExecutioner().commit();
 
-        return output;
+        return output.iterator();
     }
 
     private INDArray getSubset(int exampleStart, int exampleEnd, INDArray from) {
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreExamplesFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreExamplesFunction.java
index ee21d483e..44474248d 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreExamplesFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreExamplesFunction.java
@@ -16,11 +16,12 @@
 
 package org.deeplearning4j.spark.impl.graph.scoring;
 
+import lombok.extern.slf4j.Slf4j;
+import org.apache.spark.api.java.function.DoubleFlatMapFunction;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
 import org.deeplearning4j.nn.conf.ComputationGraphConfiguration;
 import org.deeplearning4j.nn.graph.ComputationGraph;
-import org.deeplearning4j.spark.util.BaseDoubleFlatMapFunctionAdaptee;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.api.MultiDataSet;
 import org.nd4j.linalg.factory.Nd4j;
@@ -41,31 +42,15 @@ import java.util.List;
  * @author Alex Black
  * @see ScoreExamplesWithKeyFunction
  */
-public class ScoreExamplesFunction extends BaseDoubleFlatMapFunctionAdaptee<Iterator<MultiDataSet>> {
-
-    public ScoreExamplesFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
-                    boolean addRegularizationTerms, int batchSize) {
-        super(new ScoreExamplesFunctionAdapter(params, jsonConfig, addRegularizationTerms, batchSize));
-    }
-}
-
-
-/**Function to score examples individually. Note that scoring is batched for computational efficiency.<br>
- * This is essentially a Spark implementation of the {@link ComputationGraph#scoreExamples(MultiDataSet, boolean)} method<br>
- * <b>Note:</b> This method returns a score for each example, but the association between examples and scores is lost. In
- * cases where we need to know the score for particular examples, use {@link ScoreExamplesWithKeyFunction}
- * @author Alex Black
- * @see ScoreExamplesWithKeyFunction
- */
-class ScoreExamplesFunctionAdapter implements FlatMapFunctionAdapter<Iterator<MultiDataSet>, Double> {
-    protected static final Logger log = LoggerFactory.getLogger(ScoreExamplesFunction.class);
+@Slf4j
+public class ScoreExamplesFunction implements DoubleFlatMapFunction<Iterator<MultiDataSet>> {
 
     private final Broadcast<INDArray> params;
     private final Broadcast<String> jsonConfig;
     private final boolean addRegularization;
     private final int batchSize;
 
-    public ScoreExamplesFunctionAdapter(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
+    public ScoreExamplesFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
                     boolean addRegularizationTerms, int batchSize) {
         this.params = params;
         this.jsonConfig = jsonConfig;
@@ -75,9 +60,9 @@ class ScoreExamplesFunctionAdapter implements FlatMapFunctionAdapter<Iterator<Mu
 
 
     @Override
-    public Iterable<Double> call(Iterator<MultiDataSet> iterator) throws Exception {
+    public Iterator<Double> call(Iterator<MultiDataSet> iterator) throws Exception {
         if (!iterator.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         ComputationGraph network = new ComputationGraph(ComputationGraphConfiguration.fromJson(jsonConfig.getValue()));
@@ -121,6 +106,6 @@ class ScoreExamplesFunctionAdapter implements FlatMapFunctionAdapter<Iterator<Mu
             log.debug("Scored {} examples ", totalCount);
         }
 
-        return ret;
+        return ret.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreExamplesWithKeyFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreExamplesWithKeyFunction.java
index bda47a484..5d310ee9f 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreExamplesWithKeyFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreExamplesWithKeyFunction.java
@@ -16,17 +16,15 @@
 
 package org.deeplearning4j.spark.impl.graph.scoring;
 
+import lombok.extern.slf4j.Slf4j;
 import lombok.val;
+import org.apache.spark.api.java.function.PairFlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
 import org.deeplearning4j.nn.conf.ComputationGraphConfiguration;
 import org.deeplearning4j.nn.graph.ComputationGraph;
-import org.deeplearning4j.spark.util.BasePairFlatMapFunctionAdaptee;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.api.MultiDataSet;
 import org.nd4j.linalg.factory.Nd4j;
-import org.slf4j.Logger;
-import org.slf4j.LoggerFactory;
 import scala.Tuple2;
 
 import java.util.ArrayList;
@@ -43,29 +41,8 @@ import java.util.List;
  * @param <K> Type of key, associated with each example. Used to keep track of which score belongs to which example
  * @see ScoreExamplesFunction
  */
-public class ScoreExamplesWithKeyFunction<K>
-                extends BasePairFlatMapFunctionAdaptee<Iterator<Tuple2<K, MultiDataSet>>, K, Double> {
-
-    public ScoreExamplesWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
-                    boolean addRegularizationTerms, int batchSize) {
-        super(new ScoreExamplesWithKeyFunctionAdapter<K>(params, jsonConfig, addRegularizationTerms, batchSize));
-    }
-}
-
-
-/**Function to score examples individually, where each example is associated with a particular key<br>
- * Note that scoring is batched for computational efficiency.<br>
- * This is the Spark implementation of the {@link ComputationGraph#scoreExamples(MultiDataSet, boolean)} method<br>
- * <b>Note:</b> The MultiDataSet objects passed in must have exactly one example in them (otherwise: can't have a 1:1 association
- * between keys and data sets to score)
- * @author Alex Black
- * @param <K> Type of key, associated with each example. Used to keep track of which score belongs to which example
- * @see ScoreExamplesFunction
- */
-class ScoreExamplesWithKeyFunctionAdapter<K>
-                implements FlatMapFunctionAdapter<Iterator<Tuple2<K, MultiDataSet>>, Tuple2<K, Double>> {
-
-    protected static Logger log = LoggerFactory.getLogger(ScoreExamplesWithKeyFunction.class);
+@Slf4j
+public class ScoreExamplesWithKeyFunction<K> implements PairFlatMapFunction<Iterator<Tuple2<K, MultiDataSet>>, K, Double> {
 
     private final Broadcast<INDArray> params;
     private final Broadcast<String> jsonConfig;
@@ -78,7 +55,7 @@ class ScoreExamplesWithKeyFunctionAdapter<K>
      * @param addRegularizationTerms if true: add regularization terms (l1/l2) if applicable; false: don't add regularization terms
      * @param batchSize Batch size to use when scoring examples
      */
-    public ScoreExamplesWithKeyFunctionAdapter(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
+    public ScoreExamplesWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
                     boolean addRegularizationTerms, int batchSize) {
         this.params = params;
         this.jsonConfig = jsonConfig;
@@ -88,9 +65,9 @@ class ScoreExamplesWithKeyFunctionAdapter<K>
 
 
     @Override
-    public Iterable<Tuple2<K, Double>> call(Iterator<Tuple2<K, MultiDataSet>> iterator) throws Exception {
+    public Iterator<Tuple2<K, Double>> call(Iterator<Tuple2<K, MultiDataSet>> iterator) throws Exception {
         if (!iterator.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         ComputationGraph network = new ComputationGraph(ComputationGraphConfiguration.fromJson(jsonConfig.getValue()));
@@ -140,6 +117,6 @@ class ScoreExamplesWithKeyFunctionAdapter<K>
             log.debug("Scored {} examples ", totalCount);
         }
 
-        return ret;
+        return ret.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreFlatMapFunctionCGDataSet.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreFlatMapFunctionCGDataSet.java
index 240aeb01e..3fdc7fd1c 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreFlatMapFunctionCGDataSet.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreFlatMapFunctionCGDataSet.java
@@ -16,9 +16,8 @@
 
 package org.deeplearning4j.spark.impl.graph.scoring;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.deeplearning4j.datasets.iterator.IteratorDataSetIterator;
 import org.deeplearning4j.nn.conf.ComputationGraphConfiguration;
 import org.deeplearning4j.nn.graph.ComputationGraph;
@@ -37,35 +36,23 @@ import java.util.Iterator;
 import java.util.List;
 
 /** Function used to score a DataSet using a ComputationGraph */
-public class ScoreFlatMapFunctionCGDataSet
-                extends BaseFlatMapFunctionAdaptee<Iterator<DataSet>, Tuple2<Integer, Double>> {
-
-    public ScoreFlatMapFunctionCGDataSet(String json, Broadcast<INDArray> params, int minibatchSize) {
-        super(new ScoreFlatMapFunctionCGDataSetAdapter(json, params, minibatchSize));
-    }
-}
-
-
-/** Function used to score a DataSet using a ComputationGraph */
-class ScoreFlatMapFunctionCGDataSetAdapter
-                implements FlatMapFunctionAdapter<Iterator<DataSet>, Tuple2<Integer, Double>> {
-
+public class ScoreFlatMapFunctionCGDataSet implements FlatMapFunction<Iterator<DataSet>, Tuple2<Integer, Double>> {
     private static final Logger log = LoggerFactory.getLogger(ScoreFlatMapFunctionCGDataSet.class);
     private String json;
     private Broadcast<INDArray> params;
     private int minibatchSize;
 
 
-    public ScoreFlatMapFunctionCGDataSetAdapter(String json, Broadcast<INDArray> params, int minibatchSize) {
+    public ScoreFlatMapFunctionCGDataSet(String json, Broadcast<INDArray> params, int minibatchSize) {
         this.json = json;
         this.params = params;
         this.minibatchSize = minibatchSize;
     }
 
     @Override
-    public Iterable<Tuple2<Integer, Double>> call(Iterator<DataSet> dataSetIterator) throws Exception {
+    public Iterator<Tuple2<Integer, Double>> call(Iterator<DataSet> dataSetIterator) throws Exception {
         if (!dataSetIterator.hasNext()) {
-            return Collections.singletonList(new Tuple2<>(0, 0.0));
+            return Collections.singletonList(new Tuple2<>(0, 0.0)).iterator();
         }
 
         DataSetIterator iter = new IteratorDataSetIterator(dataSetIterator, minibatchSize); //Does batching where appropriate
@@ -90,6 +77,6 @@ class ScoreFlatMapFunctionCGDataSetAdapter
 
         Nd4j.getExecutioner().commit();
 
-        return out;
+        return out.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreFlatMapFunctionCGMultiDataSet.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreFlatMapFunctionCGMultiDataSet.java
index 91072942d..bf9e3f596 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreFlatMapFunctionCGMultiDataSet.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/graph/scoring/ScoreFlatMapFunctionCGMultiDataSet.java
@@ -16,9 +16,8 @@
 
 package org.deeplearning4j.spark.impl.graph.scoring;
 
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.deeplearning4j.datasets.iterator.IteratorMultiDataSetIterator;
 import org.deeplearning4j.nn.conf.ComputationGraphConfiguration;
 import org.deeplearning4j.nn.graph.ComputationGraph;
@@ -29,7 +28,6 @@ import org.nd4j.linalg.factory.Nd4j;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 import scala.Tuple2;
-import lombok.val;
 
 import java.util.ArrayList;
 import java.util.Collections;
@@ -37,18 +35,7 @@ import java.util.Iterator;
 import java.util.List;
 
 /** Function used to score a MultiDataSet using a given ComputationGraph */
-public class ScoreFlatMapFunctionCGMultiDataSet
-                extends BaseFlatMapFunctionAdaptee<Iterator<MultiDataSet>, Tuple2<Integer, Double>> {
-
-    public ScoreFlatMapFunctionCGMultiDataSet(String json, Broadcast<INDArray> params, int minibatchSize) {
-        super(new ScoreFlatMapFunctionCGMultiDataSetAdapter(json, params, minibatchSize));
-    }
-}
-
-
-/** Function used to score a MultiDataSet using a given ComputationGraph */
-class ScoreFlatMapFunctionCGMultiDataSetAdapter
-                implements FlatMapFunctionAdapter<Iterator<MultiDataSet>, Tuple2<Integer, Double>> {
+public class ScoreFlatMapFunctionCGMultiDataSet implements FlatMapFunction<Iterator<MultiDataSet>, Tuple2<Integer, Double>> {
 
     private static final Logger log = LoggerFactory.getLogger(ScoreFlatMapFunctionCGMultiDataSet.class);
     private String json;
@@ -56,16 +43,16 @@ class ScoreFlatMapFunctionCGMultiDataSetAdapter
     private int minibatchSize;
 
 
-    public ScoreFlatMapFunctionCGMultiDataSetAdapter(String json, Broadcast<INDArray> params, int minibatchSize) {
+    public ScoreFlatMapFunctionCGMultiDataSet(String json, Broadcast<INDArray> params, int minibatchSize) {
         this.json = json;
         this.params = params;
         this.minibatchSize = minibatchSize;
     }
 
     @Override
-    public Iterable<Tuple2<Integer, Double>> call(Iterator<MultiDataSet> dataSetIterator) throws Exception {
+    public Iterator<Tuple2<Integer, Double>> call(Iterator<MultiDataSet> dataSetIterator) throws Exception {
         if (!dataSetIterator.hasNext()) {
-            return Collections.singletonList(new Tuple2<>(0, 0.0));
+            return Collections.singletonList(new Tuple2<>(0, 0.0)).iterator();
         }
 
         MultiDataSetIterator iter = new IteratorMultiDataSetIterator(dataSetIterator, minibatchSize); //Does batching where appropriate
@@ -91,6 +78,6 @@ class ScoreFlatMapFunctionCGMultiDataSetAdapter
 
         Nd4j.getExecutioner().commit();
 
-        return out;
+        return out.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/evaluation/IEvaluateFlatMapFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/evaluation/IEvaluateFlatMapFunction.java
index f95cc46f2..0a33fb995 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/evaluation/IEvaluateFlatMapFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/evaluation/IEvaluateFlatMapFunction.java
@@ -17,9 +17,8 @@
 package org.deeplearning4j.spark.impl.multilayer.evaluation;
 
 import lombok.extern.slf4j.Slf4j;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.deeplearning4j.spark.impl.evaluation.EvaluationRunner;
 import org.nd4j.evaluation.IEvaluation;
 import org.nd4j.linalg.api.ndarray.INDArray;
@@ -36,25 +35,7 @@ import java.util.concurrent.Future;
  *
  * @author Alex Black
  */
-public class IEvaluateFlatMapFunction<T extends IEvaluation>
-                extends BaseFlatMapFunctionAdaptee<Iterator<DataSet>, T[]> {
-
-    public IEvaluateFlatMapFunction(boolean isCompGraph, Broadcast<String> json, Broadcast<byte[]> params,
-                    int evalNumWorkers, int evalBatchSize, T... evaluations) {
-        super(new IEvaluateFlatMapFunctionAdapter<>(isCompGraph, json, params, evalNumWorkers, evalBatchSize, evaluations));
-    }
-}
-
-
-/**
- * Function to evaluate data (using an IEvaluation instance), in a distributed manner
- * Flat map function used to batch examples for computational efficiency + reduce number of IEvaluation objects returned
- * for network efficiency.
- *
- * @author Alex Black
- */
-@Slf4j
-class IEvaluateFlatMapFunctionAdapter<T extends IEvaluation> implements FlatMapFunctionAdapter<Iterator<DataSet>, T[]> {
+public class IEvaluateFlatMapFunction<T extends IEvaluation> implements FlatMapFunction<Iterator<DataSet>, T[]> {
 
     protected boolean isCompGraph;
     protected Broadcast<String> json;
@@ -70,7 +51,7 @@ class IEvaluateFlatMapFunctionAdapter<T extends IEvaluation> implements FlatMapF
      *                              this. Used to avoid doing too many at once (and hence memory issues)
      * @param evaluations Initial evaulation instance (i.e., empty Evaluation or RegressionEvaluation instance)
      */
-    public IEvaluateFlatMapFunctionAdapter(boolean isCompGraph, Broadcast<String> json, Broadcast<byte[]> params,
+    public IEvaluateFlatMapFunction(boolean isCompGraph, Broadcast<String> json, Broadcast<byte[]> params,
                     int evalNumWorkers, int evalBatchSize, T[] evaluations) {
         this.isCompGraph = isCompGraph;
         this.json = json;
@@ -81,9 +62,9 @@ class IEvaluateFlatMapFunctionAdapter<T extends IEvaluation> implements FlatMapF
     }
 
     @Override
-    public Iterable<T[]> call(Iterator<DataSet> dataSetIterator) throws Exception {
+    public Iterator<T[]> call(Iterator<DataSet> dataSetIterator) throws Exception {
         if (!dataSetIterator.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         Future<IEvaluation[]> f = EvaluationRunner.getInstance().execute(
@@ -91,9 +72,9 @@ class IEvaluateFlatMapFunctionAdapter<T extends IEvaluation> implements FlatMapF
 
         IEvaluation[] result = f.get();
         if(result == null){
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         } else {
-            return Collections.singletonList((T[])result);
+            return Collections.singletonList((T[])result).iterator();
         }
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/FeedForwardWithKeyFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/FeedForwardWithKeyFunction.java
index 2bdf926b6..03e4e55cf 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/FeedForwardWithKeyFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/FeedForwardWithKeyFunction.java
@@ -16,17 +16,15 @@
 
 package org.deeplearning4j.spark.impl.multilayer.scoring;
 
+import org.apache.spark.api.java.function.PairFlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
 import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
-import org.deeplearning4j.spark.util.BasePairFlatMapFunctionAdaptee;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.api.DataSetUtil;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.indexing.NDArrayIndex;
 import org.nd4j.linalg.primitives.Pair;
-import org.nd4j.linalg.util.DataSetUtils;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 import scala.Tuple2;
@@ -44,23 +42,7 @@ import java.util.List;
  * @author Alex Black
  */
 public class FeedForwardWithKeyFunction<K>
-                extends BasePairFlatMapFunctionAdaptee<Iterator<Tuple2<K, Tuple2<INDArray,INDArray>>>, K, INDArray> {
-
-    public FeedForwardWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig, int batchSize) {
-        super(new FeedForwardWithKeyFunctionAdapter<K>(params, jsonConfig, batchSize));
-    }
-}
-
-
-/**
- * Function to feed-forward examples, and get the network output (for example, class probabilities).
- * A key value is used to keey track of which output corresponds to which input.
- *
- * @param <K> Type of key, associated with each example. Used to keep track of which output belongs to which input example
- * @author Alex Black
- */
-class FeedForwardWithKeyFunctionAdapter<K>
-                implements FlatMapFunctionAdapter<Iterator<Tuple2<K, Tuple2<INDArray,INDArray>>>, Tuple2<K, INDArray>> {
+                implements PairFlatMapFunction<Iterator<Tuple2<K, Tuple2<INDArray,INDArray>>>, K, INDArray> {
 
     protected static Logger log = LoggerFactory.getLogger(FeedForwardWithKeyFunction.class);
 
@@ -73,7 +55,7 @@ class FeedForwardWithKeyFunctionAdapter<K>
      * @param jsonConfig MultiLayerConfiguration, as json
      * @param batchSize  Batch size to use for forward pass (use > 1 for efficiency)
      */
-    public FeedForwardWithKeyFunctionAdapter(Broadcast<INDArray> params, Broadcast<String> jsonConfig, int batchSize) {
+    public FeedForwardWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig, int batchSize) {
         this.params = params;
         this.jsonConfig = jsonConfig;
         this.batchSize = batchSize;
@@ -81,9 +63,9 @@ class FeedForwardWithKeyFunctionAdapter<K>
 
 
     @Override
-    public Iterable<Tuple2<K, INDArray>> call(Iterator<Tuple2<K, Tuple2<INDArray,INDArray>>> iterator) throws Exception {
+    public Iterator<Tuple2<K, INDArray>> call(Iterator<Tuple2<K, Tuple2<INDArray,INDArray>>> iterator) throws Exception {
         if (!iterator.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         MultiLayerNetwork network = new MultiLayerNetwork(MultiLayerConfiguration.fromJson(jsonConfig.getValue()));
@@ -129,7 +111,7 @@ class FeedForwardWithKeyFunctionAdapter<K>
         }
 
         if (tupleCount == 0) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         List<Tuple2<K, INDArray>> output = new ArrayList<>(tupleCount);
@@ -185,7 +167,7 @@ class FeedForwardWithKeyFunctionAdapter<K>
 
         Nd4j.getExecutioner().commit();
 
-        return output;
+        return output.iterator();
     }
 
     private INDArray getSubset(int exampleStart, int exampleEnd, INDArray from) {
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreExamplesFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreExamplesFunction.java
index 0b3383381..4142750d0 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreExamplesFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreExamplesFunction.java
@@ -16,11 +16,11 @@
 
 package org.deeplearning4j.spark.impl.multilayer.scoring;
 
+import org.apache.spark.api.java.function.DoubleFlatMapFunction;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
 import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
-import org.deeplearning4j.spark.util.BaseDoubleFlatMapFunctionAdaptee;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.DataSet;
 import org.nd4j.linalg.factory.Nd4j;
@@ -39,23 +39,7 @@ import java.util.List;
  * @author Alex Black
  * @see ScoreExamplesWithKeyFunction
  */
-public class ScoreExamplesFunction extends BaseDoubleFlatMapFunctionAdaptee<Iterator<DataSet>> {
-
-    public ScoreExamplesFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
-                    boolean addRegularizationTerms, int batchSize) {
-        super(new ScoreExamplesFunctionAdapter(params, jsonConfig, addRegularizationTerms, batchSize));
-    }
-}
-
-
-/**Function to score examples individually. Note that scoring is batched for computational efficiency.<br>
- * This is essentially a Spark implementation of the {@link MultiLayerNetwork#scoreExamples(DataSet, boolean)} method<br>
- * <b>Note:</b> This method returns a score for each example, but the association between examples and scores is lost. In
- * cases where we need to know the score for particular examples, use {@link ScoreExamplesWithKeyFunction}
- * @author Alex Black
- * @see ScoreExamplesWithKeyFunction
- */
-class ScoreExamplesFunctionAdapter implements FlatMapFunctionAdapter<Iterator<DataSet>, Double> {
+public class ScoreExamplesFunction implements DoubleFlatMapFunction<Iterator<DataSet>> {
 
     protected static Logger log = LoggerFactory.getLogger(ScoreExamplesFunction.class);
 
@@ -64,7 +48,7 @@ class ScoreExamplesFunctionAdapter implements FlatMapFunctionAdapter<Iterator<Da
     private final boolean addRegularization;
     private final int batchSize;
 
-    public ScoreExamplesFunctionAdapter(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
+    public ScoreExamplesFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
                     boolean addRegularizationTerms, int batchSize) {
         this.params = params;
         this.jsonConfig = jsonConfig;
@@ -74,9 +58,9 @@ class ScoreExamplesFunctionAdapter implements FlatMapFunctionAdapter<Iterator<Da
 
 
     @Override
-    public Iterable<Double> call(Iterator<DataSet> iterator) throws Exception {
+    public Iterator<Double> call(Iterator<DataSet> iterator) throws Exception {
         if (!iterator.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         MultiLayerNetwork network = new MultiLayerNetwork(MultiLayerConfiguration.fromJson(jsonConfig.getValue()));
@@ -119,6 +103,6 @@ class ScoreExamplesFunctionAdapter implements FlatMapFunctionAdapter<Iterator<Da
             log.debug("Scored {} examples ", totalCount);
         }
 
-        return ret;
+        return ret.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreExamplesWithKeyFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreExamplesWithKeyFunction.java
index 56b9d8489..e25915bfc 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreExamplesWithKeyFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreExamplesWithKeyFunction.java
@@ -16,16 +16,14 @@
 
 package org.deeplearning4j.spark.impl.multilayer.scoring;
 
+import lombok.extern.slf4j.Slf4j;
+import org.apache.spark.api.java.function.PairFlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
 import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
-import org.deeplearning4j.spark.util.BasePairFlatMapFunctionAdaptee;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.DataSet;
 import org.nd4j.linalg.factory.Nd4j;
-import org.slf4j.Logger;
-import org.slf4j.LoggerFactory;
 import scala.Tuple2;
 
 import java.util.ArrayList;
@@ -44,31 +42,8 @@ import java.util.List;
  * @author Alex Black
  * @see ScoreExamplesFunction
  */
-public class ScoreExamplesWithKeyFunction<K>
-                extends BasePairFlatMapFunctionAdaptee<Iterator<Tuple2<K, DataSet>>, K, Double> {
-
-    public ScoreExamplesWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
-                    boolean addRegularizationTerms, int batchSize) {
-        super(new ScoreExamplesWithKeyFunctionAdapter(params, jsonConfig, addRegularizationTerms, batchSize));
-    }
-}
-
-
-/**
- * Function to score examples individually, where each example is associated with a particular key<br>
- * Note that scoring is batched for computational efficiency.<br>
- * This is the Spark implementation of t he {@link MultiLayerNetwork#scoreExamples(DataSet, boolean)} method<br>
- * <b>Note:</b> The DataSet objects passed in must have exactly one example in them (otherwise: can't have a 1:1 association
- * between keys and data sets to score)
- *
- * @param <K> Type of key, associated with each example. Used to keep track of which score belongs to which example
- * @author Alex Black
- * @see ScoreExamplesFunction
- */
-class ScoreExamplesWithKeyFunctionAdapter<K>
-                implements FlatMapFunctionAdapter<Iterator<Tuple2<K, DataSet>>, Tuple2<K, Double>> {
-
-    protected static Logger log = LoggerFactory.getLogger(ScoreExamplesWithKeyFunction.class);
+@Slf4j
+public class ScoreExamplesWithKeyFunction<K> implements PairFlatMapFunction<Iterator<Tuple2<K, DataSet>>, K, Double> {
 
     private final Broadcast<INDArray> params;
     private final Broadcast<String> jsonConfig;
@@ -81,8 +56,7 @@ class ScoreExamplesWithKeyFunctionAdapter<K>
      * @param addRegularizationTerms if true: add regularization terms (L1, L2) to the score
      * @param batchSize              Batch size to use when scoring
      */
-    public ScoreExamplesWithKeyFunctionAdapter(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
-                    boolean addRegularizationTerms, int batchSize) {
+    public ScoreExamplesWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig, boolean addRegularizationTerms, int batchSize) {
         this.params = params;
         this.jsonConfig = jsonConfig;
         this.addRegularization = addRegularizationTerms;
@@ -91,9 +65,9 @@ class ScoreExamplesWithKeyFunctionAdapter<K>
 
 
     @Override
-    public Iterable<Tuple2<K, Double>> call(Iterator<Tuple2<K, DataSet>> iterator) throws Exception {
+    public Iterator<Tuple2<K, Double>> call(Iterator<Tuple2<K, DataSet>> iterator) throws Exception {
         if (!iterator.hasNext()) {
-            return Collections.emptyList();
+            return Collections.emptyIterator();
         }
 
         MultiLayerNetwork network = new MultiLayerNetwork(MultiLayerConfiguration.fromJson(jsonConfig.getValue()));
@@ -143,6 +117,6 @@ class ScoreExamplesWithKeyFunctionAdapter<K>
             log.debug("Scored {} examples ", totalCount);
         }
 
-        return ret;
+        return ret.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreFlatMapFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreFlatMapFunction.java
index 480f1dcd2..8063ba8e3 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreFlatMapFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/ScoreFlatMapFunction.java
@@ -16,9 +16,11 @@
 
 package org.deeplearning4j.spark.impl.multilayer.scoring;
 
+import lombok.AllArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import lombok.val;
+import org.apache.spark.api.java.function.FlatMapFunction;
 import org.apache.spark.broadcast.Broadcast;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import org.datavec.spark.transform.BaseFlatMapFunctionAdaptee;
 import org.deeplearning4j.datasets.iterator.IteratorDataSetIterator;
 import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
@@ -26,43 +28,25 @@ import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.DataSet;
 import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
 import org.nd4j.linalg.factory.Nd4j;
-import org.slf4j.Logger;
-import org.slf4j.LoggerFactory;
 import scala.Tuple2;
-import lombok.val;
 
 import java.util.ArrayList;
 import java.util.Collections;
 import java.util.Iterator;
 import java.util.List;
 
-public class ScoreFlatMapFunction extends BaseFlatMapFunctionAdaptee<Iterator<DataSet>, Tuple2<Integer, Double>> {
-
-    public ScoreFlatMapFunction(String json, Broadcast<INDArray> params, int minibatchSize) {
-        super(new ScoreFlatMapFunctionAdapter(json, params, minibatchSize));
-    }
-
-}
-
-
-class ScoreFlatMapFunctionAdapter implements FlatMapFunctionAdapter<Iterator<DataSet>, Tuple2<Integer, Double>> {
-
-    private static final Logger log = LoggerFactory.getLogger(ScoreFlatMapFunction.class);
+@Slf4j
+@AllArgsConstructor
+public class ScoreFlatMapFunction implements FlatMapFunction<Iterator<DataSet>, Tuple2<Integer, Double>> {
 
     private String json;
     private Broadcast<INDArray> params;
     private int minibatchSize;
 
-    public ScoreFlatMapFunctionAdapter(String json, Broadcast<INDArray> params, int minibatchSize) {
-        this.json = json;
-        this.params = params;
-        this.minibatchSize = minibatchSize;
-    }
-
     @Override
-    public Iterable<Tuple2<Integer, Double>> call(Iterator<DataSet> dataSetIterator) throws Exception {
+    public Iterator<Tuple2<Integer, Double>> call(Iterator<DataSet> dataSetIterator) throws Exception {
         if (!dataSetIterator.hasNext()) {
-            return Collections.singletonList(new Tuple2<>(0, 0.0));
+            return Collections.singletonList(new Tuple2<>(0, 0.0)).iterator();
         }
 
         DataSetIterator iter = new IteratorDataSetIterator(dataSetIterator, minibatchSize); //Does batching where appropriate
@@ -87,6 +71,6 @@ class ScoreFlatMapFunctionAdapter implements FlatMapFunctionAdapter<Iterator<Dat
 
         Nd4j.getExecutioner().commit();
 
-        return out;
+        return out.iterator();
     }
 }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/VaeReconstructionErrorWithKeyFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/VaeReconstructionErrorWithKeyFunction.java
index bdd433a7a..231f3c9a2 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/VaeReconstructionErrorWithKeyFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/VaeReconstructionErrorWithKeyFunction.java
@@ -21,8 +21,7 @@ import org.deeplearning4j.nn.api.Layer;
 import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.layers.variational.VariationalAutoencoder;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
-import org.deeplearning4j.spark.impl.common.score.BaseVaeScoreWithKeyFunctionAdapter;
-import org.deeplearning4j.spark.util.BasePairFlatMapFunctionAdaptee;
+import org.deeplearning4j.spark.impl.common.score.BaseVaeScoreWithKeyFunction;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import scala.Tuple2;
 
@@ -38,33 +37,14 @@ import java.util.Iterator;
  * @author Alex Black
  * @see VaeReconstructionProbWithKeyFunction
  */
-public class VaeReconstructionErrorWithKeyFunction<K>
-                extends BasePairFlatMapFunctionAdaptee<Iterator<Tuple2<K, INDArray>>, K, Double> {
-
-    public VaeReconstructionErrorWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
-                    int batchSize) {
-        super(new VaeReconstructionErrorWithKeyFunctionAdapter(params, jsonConfig, batchSize));
-    }
-}
-
-
-/**
- * Function to calculate the reconstruction error for a variational autoencoder, that is the first layer in a
- * MultiLayerNetwork.<br>
- * Note that the VAE must be using a loss function, not a {@link org.deeplearning4j.nn.conf.layers.variational.ReconstructionDistribution}<br>
- * Also note that scoring is batched for computational efficiency.<br>
- *
- * @author Alex Black
- * @see VaeReconstructionProbWithKeyFunction
- */
-class VaeReconstructionErrorWithKeyFunctionAdapter<K> extends BaseVaeScoreWithKeyFunctionAdapter<K> {
+public class VaeReconstructionErrorWithKeyFunction<K> extends BaseVaeScoreWithKeyFunction<K> {
 
     /**
      * @param params            MultiLayerNetwork parameters
      * @param jsonConfig        MultiLayerConfiguration, as json
      * @param batchSize         Batch size to use when scoring
      */
-    public VaeReconstructionErrorWithKeyFunctionAdapter(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
+    public VaeReconstructionErrorWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
                     int batchSize) {
         super(params, jsonConfig, batchSize);
     }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/VaeReconstructionProbWithKeyFunction.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/VaeReconstructionProbWithKeyFunction.java
index 7bba68b28..e8fc8416f 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/VaeReconstructionProbWithKeyFunction.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/multilayer/scoring/VaeReconstructionProbWithKeyFunction.java
@@ -21,12 +21,8 @@ import org.deeplearning4j.nn.api.Layer;
 import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.layers.variational.VariationalAutoencoder;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
-import org.deeplearning4j.spark.impl.common.score.BaseVaeReconstructionProbWithKeyFunctionAdapter;
-import org.deeplearning4j.spark.util.BasePairFlatMapFunctionAdaptee;
+import org.deeplearning4j.spark.impl.common.score.BaseVaeReconstructionProbWithKeyFunction;
 import org.nd4j.linalg.api.ndarray.INDArray;
-import scala.Tuple2;
-
-import java.util.Iterator;
 
 
 /**
@@ -36,25 +32,7 @@ import java.util.Iterator;
  *
  * @author Alex Black
  */
-public class VaeReconstructionProbWithKeyFunction<K>
-                extends BasePairFlatMapFunctionAdaptee<Iterator<Tuple2<K, INDArray>>, K, Double> {
-
-    public VaeReconstructionProbWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
-                    boolean useLogProbability, int batchSize, int numSamples) {
-        super(new VaeReconstructionProbWithKeyFunctionAdapter<K>(params, jsonConfig, useLogProbability, batchSize,
-                        numSamples));
-    }
-}
-
-
-/**
- * Function to calculate the reconstruction probability for a variational autoencoder, that is the first layer in a
- * MultiLayerNetwork.<br>
- * Note that scoring is batched for computational efficiency.<br>
- *
- * @author Alex Black
- */
-class VaeReconstructionProbWithKeyFunctionAdapter<K> extends BaseVaeReconstructionProbWithKeyFunctionAdapter<K> {
+public class VaeReconstructionProbWithKeyFunction<K> extends BaseVaeReconstructionProbWithKeyFunction<K> {
 
 
     /**
@@ -64,7 +42,7 @@ class VaeReconstructionProbWithKeyFunctionAdapter<K> extends BaseVaeReconstructi
      * @param batchSize         Batch size to use when scoring
      * @param numSamples        Number of samples to use when calling {@link VariationalAutoencoder#reconstructionLogProbability(INDArray, int)}
      */
-    public VaeReconstructionProbWithKeyFunctionAdapter(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
+    public VaeReconstructionProbWithKeyFunction(Broadcast<INDArray> params, Broadcast<String> jsonConfig,
                     boolean useLogProbability, int batchSize, int numSamples) {
         super(params, jsonConfig, useLogProbability, batchSize, numSamples);
     }
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/paramavg/ParameterAveragingTrainingMaster.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/paramavg/ParameterAveragingTrainingMaster.java
index 30e6b395f..b6d5654ec 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/paramavg/ParameterAveragingTrainingMaster.java
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/java/org/deeplearning4j/spark/impl/paramavg/ParameterAveragingTrainingMaster.java
@@ -67,7 +67,7 @@ import java.io.IOException;
 import java.io.OutputStream;
 import java.util.*;
 
-import static com.google.common.base.Preconditions.checkArgument;
+import static org.nd4j.shade.guava.base.Preconditions.checkArgument;
 
 /**
  * ParameterAveragingTrainingMaster: A {@link TrainingMaster}
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-1/java/org/deeplearning4j/spark/util/BaseDoubleFlatMapFunctionAdaptee.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-1/java/org/deeplearning4j/spark/util/BaseDoubleFlatMapFunctionAdaptee.java
deleted file mode 100644
index 5f9d25547..000000000
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-1/java/org/deeplearning4j/spark/util/BaseDoubleFlatMapFunctionAdaptee.java
+++ /dev/null
@@ -1,40 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.spark.util;
-
-import org.apache.spark.api.java.function.DoubleFlatMapFunction;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-
-/**
- * DoubleFlatMapFunction adapter to hide incompatibilities between Spark 1.x and Spark 2.x
- *
- * This class should be used instead of direct referral to DoubleFlatMapFunction
- *
- */
-public class BaseDoubleFlatMapFunctionAdaptee<T> implements DoubleFlatMapFunction<T> {
-
-    protected final FlatMapFunctionAdapter<T, Double> adapter;
-
-    public BaseDoubleFlatMapFunctionAdaptee(FlatMapFunctionAdapter<T, Double> adapter) {
-        this.adapter = adapter;
-    }
-
-    @Override
-    public Iterable<Double> call(T t) throws Exception {
-        return adapter.call(t);
-    }
-}
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-1/java/org/deeplearning4j/spark/util/BasePairFlatMapFunctionAdaptee.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-1/java/org/deeplearning4j/spark/util/BasePairFlatMapFunctionAdaptee.java
deleted file mode 100644
index b58cab202..000000000
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-1/java/org/deeplearning4j/spark/util/BasePairFlatMapFunctionAdaptee.java
+++ /dev/null
@@ -1,41 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.spark.util;
-
-import org.apache.spark.api.java.function.PairFlatMapFunction;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import scala.Tuple2;
-
-/**
- * PairFlatMapFunction adapter to hide incompatibilities between Spark 1.x and Spark 2.x
- *
- * This class should be used instead of direct referral to PairFlatMapFunction
- *
- */
-public class BasePairFlatMapFunctionAdaptee<T, K, V> implements PairFlatMapFunction<T, K, V> {
-
-    protected final FlatMapFunctionAdapter<T, Tuple2<K, V>> adapter;
-
-    public BasePairFlatMapFunctionAdaptee(FlatMapFunctionAdapter<T, Tuple2<K, V>> adapter) {
-        this.adapter = adapter;
-    }
-
-    @Override
-    public Iterable<Tuple2<K, V>> call(T t) throws Exception {
-        return adapter.call(t);
-    }
-}
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-2/java/org/deeplearning4j/spark/util/BaseDoubleFlatMapFunctionAdaptee.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-2/java/org/deeplearning4j/spark/util/BaseDoubleFlatMapFunctionAdaptee.java
deleted file mode 100644
index 49a05231b..000000000
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-2/java/org/deeplearning4j/spark/util/BaseDoubleFlatMapFunctionAdaptee.java
+++ /dev/null
@@ -1,42 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.spark.util;
-
-import org.apache.spark.api.java.function.DoubleFlatMapFunction;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-
-import java.util.Iterator;
-
-/**
- * DoubleFlatMapFunction adapter to hide incompatibilities between Spark 1.x and Spark 2.x
- *
- * This class should be used instead of direct referral to DoubleFlatMapFunction
- *
- */
-public class BaseDoubleFlatMapFunctionAdaptee<T> implements DoubleFlatMapFunction<T> {
-
-    protected final FlatMapFunctionAdapter<T, Double> adapter;
-
-    public BaseDoubleFlatMapFunctionAdaptee(FlatMapFunctionAdapter<T, Double> adapter) {
-        this.adapter = adapter;
-    }
-
-    @Override
-    public Iterator<Double> call(T t) throws Exception {
-        return adapter.call(t).iterator();
-    }
-}
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-2/java/org/deeplearning4j/spark/util/BasePairFlatMapFunctionAdaptee.java b/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-2/java/org/deeplearning4j/spark/util/BasePairFlatMapFunctionAdaptee.java
deleted file mode 100644
index b28a8dcc2..000000000
--- a/deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/main/spark-2/java/org/deeplearning4j/spark/util/BasePairFlatMapFunctionAdaptee.java
+++ /dev/null
@@ -1,43 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.spark.util;
-
-import org.apache.spark.api.java.function.PairFlatMapFunction;
-import org.datavec.spark.functions.FlatMapFunctionAdapter;
-import scala.Tuple2;
-
-import java.util.Iterator;
-
-/**
- * PairFlatMapFunction adapter to hide incompatibilities between Spark 1.x and Spark 2.x
- *
- * This class should be used instead of direct referral to PairFlatMapFunction
- *
- */
-public class BasePairFlatMapFunctionAdaptee<T, K, V> implements PairFlatMapFunction<T, K, V> {
-
-    protected final FlatMapFunctionAdapter<T, Tuple2<K, V>> adapter;
-
-    public BasePairFlatMapFunctionAdaptee(FlatMapFunctionAdapter<T, Tuple2<K, V>> adapter) {
-        this.adapter = adapter;
-    }
-
-    @Override
-    public Iterator<Tuple2<K, V>> call(T t) throws Exception {
-        return adapter.call(t).iterator();
-    }
-}
diff --git a/deeplearning4j/deeplearning4j-scaleout/spark/pom.xml b/deeplearning4j/deeplearning4j-scaleout/spark/pom.xml
index 1919fcb06..f753fefae 100644
--- a/deeplearning4j/deeplearning4j-scaleout/spark/pom.xml
+++ b/deeplearning4j/deeplearning4j-scaleout/spark/pom.xml
@@ -36,16 +36,9 @@
     <properties>
         <project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
         <project.reporting.outputEncoding>UTF-8</project.reporting.outputEncoding>
-        <!-- These Spark version properties have to defined be here, and NOT in the deeplearning4j-parent pom. Otherwise,
-         the properties will be resolved to their defaults (and not the values present during the build). Whereas when
-         defined here (the version is spark 1 vs. 2 specific) we can have different version properties simultaneously
-          (in different pom files) instead of a single global property -->
-        <spark.version>2.1.0</spark.version>
-        <spark.major.version>2</spark.major.version>
         <datavec.spark.version>1.0.0_spark_2-SNAPSHOT</datavec.spark.version>
 
 
-
         <scala.macros.version>2.1.0</scala.macros.version>
         <!-- Default scala versions, may be overwritten by build profiles -->
         <scala.version>2.11.12</scala.version>
@@ -194,21 +187,6 @@
             <artifactId>jaxb-impl</artifactId>
             <version>${jaxb.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-actor_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-remote_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-slf4j_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
         <dependency>
             <groupId>io.netty</groupId>
             <artifactId>netty</artifactId>
@@ -280,36 +258,6 @@
     </dependencies>
 
     <profiles>
-        <profile>
-            <id>spark-2</id>
-            <activation>
-                <property>
-                    <name>spark.major.version</name>
-                    <value>2</value>
-                </property>
-            </activation>
-            <dependencies>
-                <dependency>
-                    <groupId>com.typesafe.akka</groupId>
-                    <artifactId>akka-remote_2.11</artifactId>
-                    <version>2.3.11</version>
-                </dependency>
-            </dependencies>
-        </profile>
-        <profile>
-            <id>cdh5</id>
-            <repositories>
-                <repository>
-                    <id>org.apache.hadoop</id>
-                    <url>https://repository.cloudera.com/artifactory/cloudera-repos/</url>
-                </repository>
-            </repositories>
-            <properties>
-                <hadoop.version>2.0.0-cdh4.6.0</hadoop.version>
-                <spark.version>1.2.0-cdh5.3.0</spark.version>
-            </properties>
-        </profile>
-
         <profile>
             <id>test-nd4j-native</id>
         </profile>
diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/pom.xml b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/pom.xml
index f624beb6b..1b4f33c1e 100644
--- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/pom.xml
+++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/pom.xml
@@ -130,26 +130,11 @@
             <version>${project.version}</version>
         </dependency>
 
-        <dependency>
-            <groupId>com.google.guava</groupId>
-            <artifactId>guava</artifactId>
-            <version>${guava.version}</version>
-        </dependency>
         <dependency>
             <groupId>com.google.protobuf</groupId>
             <artifactId>protobuf-java</artifactId>
             <version>${google.protobuf.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-actor_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-slf4j_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
         <dependency>
             <groupId>javax.ws.rs</groupId>
             <artifactId>javax.ws.rs-api</artifactId>
@@ -226,76 +211,6 @@
             <artifactId>leveldbjni-all</artifactId>
             <version>${leveldb.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-contrib_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-
-        <!-- ///////////////// Jackson Dependencies ///////////////// -->
-        <!-- Here: forcing all Jackson versions to match Spark versions. Otherwise, Spark may override some Jackson dependencies
-        with 2.4.4, but not all of them - resulting in dependency issues. Play framework seems OK using this version of Jackson
-        (as opposed to the default 2.5.4 version) as long as these versions aren't mixed for different dependencies. -->
-        <!-- Jackson Core -->
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-core</artifactId>
-            <version>${spark.jackson.version}</version>
-        </dependency>
-
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-databind</artifactId>
-            <version>${spark.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-annotations</artifactId>
-            <version>${spark.jackson.version}</version>
-        </dependency>
-
-        <!-- Jackson Module -->
-        <dependency>
-            <groupId>com.fasterxml.jackson.module</groupId>
-            <artifactId>jackson-module-scala_2.11</artifactId>
-            <version>${spark.jackson.version}</version>
-            <exclusions>
-                <exclusion>
-                    <groupId>com.google.code.findbugs</groupId>
-                    <artifactId>jsr305</artifactId>
-                </exclusion>
-            </exclusions>
-        </dependency>
-
-        <!-- Jackson Datatype -->
-        <dependency>
-            <groupId>com.fasterxml.jackson.datatype</groupId>
-            <artifactId>jackson-datatype-jdk8</artifactId>
-            <version>${spark.jackson.version}</version>
-        </dependency>
-
-        <dependency>
-            <groupId>com.fasterxml.jackson.datatype</groupId>
-            <artifactId>jackson-datatype-jsr310</artifactId>
-            <version>${spark.jackson.version}</version>
-        </dependency>
-
-
-        <!-- Promote com.typesafe.config to a direct dependency -->
-        <!-- This is a dependency for both Spark (1.2.1 for Spark 1.6.0, as a transitive dependency via Akka) and for
-             Play (1.3.0). Play requires 1.3.0 (fails with 1.2.1 - missing required methods), but Spark can use 1.3.0.
-             By promoting this to a direct dependency, we are more likely to resolve 1.3.0 over Spark's 1.2.1 -->
-        <dependency>
-            <groupId>com.typesafe</groupId>
-            <artifactId>config</artifactId>
-            <version>${typesafe.config.version}</version>
-        </dependency>
-
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-cluster_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
 
         <dependency>
             <groupId>com.beust</groupId>
diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/api/FunctionType.java b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/api/FunctionType.java
index a8a499ec6..f694291c8 100644
--- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/api/FunctionType.java
+++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/api/FunctionType.java
@@ -17,10 +17,17 @@
 package org.deeplearning4j.ui.api;
 
 /**
- * Enumeration for the type of function. Mainly used in specifying {@link Route} instances
+ * Enumeration for the type of function. Mainly used in specifying {@link Route} instances<br>
+ * Supplier: No args<br>
+ * Function: 1 arg<br>
+ * BiFunction: 2 args<br>
+ * Function3: 3 args<br>
+ * Request0Function: Supplier + request, no args (as Function<Request,Result>)<br>
+ * Request1Function: Supplier + request + 1 args (as BiFunction<Request,String, Result>)<br>
  *
  * @author Alex Black
  */
 public enum FunctionType {
-    Supplier, Function, BiFunction, Function3
+    Supplier, Function, BiFunction, Function3,
+    Request0Function, Request1Function
 }
diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/api/Route.java b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/api/Route.java
index 0f0a731d8..f2dd1e017 100644
--- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/api/Route.java
+++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/api/Route.java
@@ -18,6 +18,7 @@ package org.deeplearning4j.ui.api;
 
 import lombok.AllArgsConstructor;
 import lombok.Data;
+import play.mvc.Http;
 import play.mvc.Result;
 
 import java.util.function.BiFunction;
@@ -38,17 +39,27 @@ public class Route {
     private final Supplier<Result> supplier;
     private final Function<String, Result> function;
     private final BiFunction<String, String, Result> function2;
+    private final Function<Http.Request, Result> request0Function;
+    private final BiFunction<Http.Request, String, Result> request1Function;
 
     public Route(String route, HttpMethod method, FunctionType functionType, Supplier<Result> supplier) {
-        this(route, method, functionType, supplier, null, null);
+        this(route, method, functionType, supplier, null, null, null, null);
     }
 
     public Route(String route, HttpMethod method, FunctionType functionType, Function<String, Result> function) {
-        this(route, method, functionType, null, function, null);
+        this(route, method, functionType, null, function, null, null, null);
+    }
+
+    public static Route request0Function(String route, HttpMethod httpMethod, Function<Http.Request, Result> function){
+        return new Route(route, httpMethod, FunctionType.Request0Function, null, null, null, function, null);
+    }
+
+    public static Route request1Function(String route, HttpMethod httpMethod, BiFunction<Http.Request, String, Result> function){
+        return new Route(route, httpMethod, FunctionType.Request1Function, null, null, null, null, function);
     }
 
     public Route(String route, HttpMethod method, FunctionType functionType,
                     BiFunction<String, String, Result> function) {
-        this(route, method, functionType, null, null, function);
+        this(route, method, functionType, null, null, function, null, null);
     }
 }
diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/module/tsne/TsneModule.java b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/module/tsne/TsneModule.java
index fe92756cc..0886af8a9 100644
--- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/module/tsne/TsneModule.java
+++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/module/tsne/TsneModule.java
@@ -24,6 +24,7 @@ import org.deeplearning4j.ui.api.HttpMethod;
 import org.deeplearning4j.ui.api.Route;
 import org.deeplearning4j.ui.api.UIModule;
 import org.deeplearning4j.ui.i18n.I18NResource;
+import play.libs.Files;
 import play.libs.Json;
 import play.mvc.Http;
 import play.mvc.Result;
@@ -31,9 +32,9 @@ import play.mvc.Results;
 
 import java.io.File;
 import java.io.IOException;
+import java.nio.charset.StandardCharsets;
 import java.util.*;
 
-import static play.mvc.Controller.request;
 import static play.mvc.Results.badRequest;
 import static play.mvc.Results.ok;
 
@@ -63,8 +64,8 @@ public class TsneModule implements UIModule {
                         () -> ok(org.deeplearning4j.ui.views.html.tsne.Tsne.apply()));
         Route r2 = new Route("/tsne/sessions", HttpMethod.GET, FunctionType.Supplier, this::listSessions);
         Route r3 = new Route("/tsne/coords/:sid", HttpMethod.GET, FunctionType.Function, this::getCoords);
-        Route r4 = new Route("/tsne/upload", HttpMethod.POST, FunctionType.Supplier, this::uploadFile);
-        Route r5 = new Route("/tsne/post/:sid", HttpMethod.POST, FunctionType.Function, this::postFile);
+        Route r4 = Route.request0Function("/tsne/upload", HttpMethod.POST, this::uploadFile);
+        Route r5 = Route.request1Function("/tsne/post/:sid", HttpMethod.POST, this::postFile);
         return Arrays.asList(r1, r2, r3, r4, r5);
     }
 
@@ -106,22 +107,22 @@ public class TsneModule implements UIModule {
         }
     }
 
-    private Result uploadFile() {
-        Http.MultipartFormData body = request().body().asMultipartFormData();
-        List<Http.MultipartFormData.FilePart> fileParts = body.getFiles();
+    private Result uploadFile(Http.Request request) {
+        Http.MultipartFormData<Files.TemporaryFile> body = request.body().asMultipartFormData();
+        List<Http.MultipartFormData.FilePart<Files.TemporaryFile>> fileParts = body.getFiles();
 
         if (fileParts.isEmpty()) {
             return badRequest("No file uploaded");
         }
 
-        Http.MultipartFormData.FilePart uploadedFile = fileParts.get(0);
+        Http.MultipartFormData.FilePart<Files.TemporaryFile> uploadedFile = fileParts.get(0);
 
         String fileName = uploadedFile.getFilename();
         String contentType = uploadedFile.getContentType();
-        File file = uploadedFile.getFile();
+        File file = uploadedFile.getRef().path().toFile();
 
         try {
-            uploadedFileLines = FileUtils.readLines(file);
+            uploadedFileLines = FileUtils.readLines(file, StandardCharsets.UTF_8);
         } catch (IOException e) {
             return badRequest("Could not read from uploaded file");
         }
@@ -129,21 +130,21 @@ public class TsneModule implements UIModule {
         return ok("File uploaded: " + fileName + ", " + contentType + ", " + file);
     }
 
-    private Result postFile(String sid) {
+    private Result postFile(Http.Request request, String sid) {
         //        System.out.println("POST FILE CALLED: " + sid);
-        Http.MultipartFormData body = request().body().asMultipartFormData();
-        List<Http.MultipartFormData.FilePart> fileParts = body.getFiles();
+        Http.MultipartFormData<Files.TemporaryFile> body = request.body().asMultipartFormData();
+        List<Http.MultipartFormData.FilePart<Files.TemporaryFile>> fileParts = body.getFiles();
 
         if (fileParts.isEmpty()) {
             //            System.out.println("**** NO FILE ****");
             return badRequest("No file uploaded");
         }
 
-        Http.MultipartFormData.FilePart uploadedFile = fileParts.get(0);
+        Http.MultipartFormData.FilePart<Files.TemporaryFile> uploadedFile = fileParts.get(0);
 
         String fileName = uploadedFile.getFilename();
         String contentType = uploadedFile.getContentType();
-        File file = uploadedFile.getFile();
+        File file = uploadedFile.getRef().path().toFile();
 
         List<String> lines;
         try {
diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/PlayUIServer.java b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/PlayUIServer.java
index 747284e7e..c0ebe4ed8 100644
--- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/PlayUIServer.java
+++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/PlayUIServer.java
@@ -37,18 +37,16 @@ import org.deeplearning4j.ui.module.defaultModule.DefaultModule;
 import org.deeplearning4j.ui.module.remote.RemoteReceiverModule;
 import org.deeplearning4j.ui.module.train.TrainModule;
 import org.deeplearning4j.ui.module.tsne.TsneModule;
-import org.deeplearning4j.ui.play.misc.FunctionUtil;
 import org.deeplearning4j.ui.play.staticroutes.Assets;
-import org.deeplearning4j.ui.play.staticroutes.I18NRoute;
-import org.deeplearning4j.ui.play.staticroutes.MultiSessionI18NRoute;
 import org.deeplearning4j.ui.storage.FileStatsStorage;
 import org.deeplearning4j.ui.storage.InMemoryStatsStorage;
 import org.deeplearning4j.ui.storage.impl.QueueStatsStorageListener;
 import org.deeplearning4j.util.DL4JFileUtils;
 import org.nd4j.linalg.function.Function;
 import org.nd4j.linalg.primitives.Pair;
+import play.BuiltInComponents;
 import play.Mode;
-import play.api.routing.Router;
+import play.routing.Router;
 import play.routing.RoutingDsl;
 import play.server.Server;
 
@@ -60,6 +58,8 @@ import java.util.concurrent.CopyOnWriteArrayList;
 import java.util.concurrent.LinkedBlockingQueue;
 import java.util.concurrent.atomic.AtomicBoolean;
 
+import static play.mvc.Results.ok;
+
 
 /**
  * A UI server based on the Play framework
@@ -166,63 +166,6 @@ public class PlayUIServer extends UIServer {
             System.exit(1);
         }
 
-        RoutingDsl routingDsl = new RoutingDsl();
-
-        //Set up index page and assets routing
-        //The definitions and FunctionUtil may look a bit weird here... this is used to translate implementation independent
-        // definitions (i.e., Java Supplier, Function etc interfaces) to the Play-specific versions
-        //This way, routing is not directly dependent ot Play API. Furthermore, Play 2.5 switches to using these Java interfaces
-        // anyway; thus switching 2.5 should be as simple as removing the FunctionUtil calls...
-        if (multiSession) {
-            routingDsl.GET("/setlang/:sessionId/:to").routeTo(FunctionUtil.biFunction(new MultiSessionI18NRoute()));
-        } else {
-            routingDsl.GET("/setlang/:to").routeTo(FunctionUtil.function(new I18NRoute()));
-        }
-        routingDsl.GET("/assets/*file").routeTo(FunctionUtil.function(new Assets(ASSETS_ROOT_DIRECTORY)));
-
-        uiModules.add(new DefaultModule(multiSession)); //For: navigation page "/"
-        uiModules.add(new TrainModule(multiSession, statsStorageLoader, this::getAddress));
-        uiModules.add(new ConvolutionalListenerModule());
-        uiModules.add(new TsneModule());
-        uiModules.add(new SameDiffModule());
-        remoteReceiverModule = new RemoteReceiverModule();
-        uiModules.add(remoteReceiverModule);
-
-        //Check service loader mechanism (Arbiter UI, etc) for modules
-        modulesViaServiceLoader(uiModules);
-
-        for (UIModule m : uiModules) {
-            List<Route> routes = m.getRoutes();
-            for (Route r : routes) {
-                RoutingDsl.PathPatternMatcher ppm = routingDsl.match(r.getHttpMethod().name(), r.getRoute());
-                switch (r.getFunctionType()) {
-                    case Supplier:
-                        ppm.routeTo(FunctionUtil.function0(r.getSupplier()));
-                        break;
-                    case Function:
-                        ppm.routeTo(FunctionUtil.function(r.getFunction()));
-                        break;
-                    case BiFunction:
-                        ppm.routeTo(FunctionUtil.biFunction(r.getFunction2()));
-                        break;
-                    case Function3:
-                    default:
-                        throw new RuntimeException("Not yet implemented");
-                }
-            }
-
-            //Determine which type IDs this module wants to receive:
-            List<String> typeIDs = m.getCallbackTypeIDs();
-            for (String typeID : typeIDs) {
-                List<UIModule> list = typeIDModuleMap.get(typeID);
-                if (list == null) {
-                    list = Collections.synchronizedList(new ArrayList<>());
-                    typeIDModuleMap.put(typeID, list);
-                }
-                list.add(m);
-            }
-        }
-
         String portProperty = System.getProperty(DL4JSystemProperties.UI_SERVER_PORT_PROPERTY);
         if (portProperty != null) {
             try {
@@ -233,6 +176,7 @@ public class PlayUIServer extends UIServer {
             }
         }
 
+
         //Set play secret key, if required
         //http://www.playframework.com/documentation/latest/ApplicationSecret
         String crypto = System.getProperty("play.crypto.secret");
@@ -245,9 +189,9 @@ public class PlayUIServer extends UIServer {
             System.setProperty("play.crypto.secret", base64);
         }
 
-        Router router = routingDsl.build();
+
         try {
-            server = Server.forRouter(router, Mode.PROD, port);
+            server = Server.forRouter(Mode.PROD, port, this::createRouter);
         } catch (Throwable e){
             if(e.getMessage().contains("'play.crypto.provider")){
                 //Usual cause: user's uber-jar does not include application.conf
@@ -284,6 +228,79 @@ public class PlayUIServer extends UIServer {
         setStopped(false);
     }
 
+    protected Router createRouter(BuiltInComponents builtInComponents){
+        RoutingDsl routingDsl = RoutingDsl.fromComponents(builtInComponents);
+
+        //Set up index page and assets routing
+        //The definitions and FunctionUtil may look a bit weird here... this is used to translate implementation independent
+        // definitions (i.e., Java Supplier, Function etc interfaces) to the Play-specific versions
+        //This way, routing is not directly dependent ot Play API. Furthermore, Play 2.5 switches to using these Java interfaces
+        // anyway; thus switching 2.5 should be as simple as removing the FunctionUtil calls...
+        if (multiSession) {
+            routingDsl.GET("/setlang/:sessionId/:to").routingTo((request, sid, to) -> {
+                I18NProvider.getInstance(sid.toString()).setDefaultLanguage(to.toString());
+                return ok();
+            });
+        } else {
+            routingDsl.GET("/setlang/:to").routingTo((request, to) -> {
+                I18NProvider.getInstance().setDefaultLanguage(to.toString());
+                return ok();
+            });
+        }
+        routingDsl.GET("/assets/*file").routingTo((request, file) -> Assets.assetRequest(ASSETS_ROOT_DIRECTORY, file.toString()));
+
+        uiModules.add(new DefaultModule(multiSession)); //For: navigation page "/"
+        uiModules.add(new TrainModule(multiSession, statsStorageLoader, this::getAddress));
+        uiModules.add(new ConvolutionalListenerModule());
+        uiModules.add(new TsneModule());
+        uiModules.add(new SameDiffModule());
+        remoteReceiverModule = new RemoteReceiverModule();
+        uiModules.add(remoteReceiverModule);
+
+        //Check service loader mechanism (Arbiter UI, etc) for modules
+        modulesViaServiceLoader(uiModules);
+
+        for (UIModule m : uiModules) {
+            List<Route> routes = m.getRoutes();
+            for (Route r : routes) {
+                RoutingDsl.PathPatternMatcher ppm = routingDsl.match(r.getHttpMethod().name(), r.getRoute());
+                switch (r.getFunctionType()) {
+                    case Supplier:
+                        ppm.routingTo(request -> r.getSupplier().get());
+                        break;
+                    case Function:
+                        ppm.routingTo((request, arg) -> r.getFunction().apply(arg.toString()));
+                        break;
+                    case BiFunction:
+                        ppm.routingTo((request, arg0, arg1) -> r.getFunction2().apply(arg0.toString(), arg1.toString()));
+                        break;
+                    case Request0Function:
+                        ppm.routingTo(request -> r.getRequest0Function().apply(request));
+                        break;
+                    case Request1Function:
+                        ppm.routingTo((request, arg0) -> r.getRequest1Function().apply(request, arg0.toString()));
+                        break;
+                    case Function3:
+                    default:
+                        throw new RuntimeException("Not yet implemented");
+                }
+            }
+
+            //Determine which type IDs this module wants to receive:
+            List<String> typeIDs = m.getCallbackTypeIDs();
+            for (String typeID : typeIDs) {
+                List<UIModule> list = typeIDModuleMap.get(typeID);
+                if (list == null) {
+                    list = Collections.synchronizedList(new ArrayList<>());
+                    typeIDModuleMap.put(typeID, list);
+                }
+                list.add(m);
+            }
+        }
+        Router router = routingDsl.build();
+        return router;
+    }
+
     @Override
     public String getAddress() {
         String addr = server.mainAddress().toString();
diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/misc/FunctionUtil.java b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/misc/FunctionUtil.java
deleted file mode 100644
index d1b87c443..000000000
--- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/misc/FunctionUtil.java
+++ /dev/null
@@ -1,45 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.ui.play.misc;
-
-import play.libs.F;
-import play.mvc.Result;
-
-import java.util.function.BiFunction;
-import java.util.function.Function;
-import java.util.function.Supplier;
-
-/**
- * Utility methods for Routing
- *
- * @author Alex Black
- */
-public class FunctionUtil {
-
-    public static F.Function0<Result> function0(Supplier<Result> supplier) {
-        return supplier::get;
-    }
-
-    public static <T> F.Function<T, Result> function(Function<T, Result> function) {
-        return function::apply;
-    }
-
-    public static <T,U> F.Function2<T, U, Result> biFunction(BiFunction<T, U, Result> function) {
-        return function::apply;
-    }
-
-}
diff --git a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/staticroutes/Assets.java b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/staticroutes/Assets.java
index dd100b3d1..76d6e7361 100644
--- a/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/staticroutes/Assets.java
+++ b/deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-play/src/main/java/org/deeplearning4j/ui/play/staticroutes/Assets.java
@@ -16,18 +16,17 @@
 
 package org.deeplearning4j.ui.play.staticroutes;
 
-import com.google.common.net.HttpHeaders;
+import org.nd4j.shade.guava.net.HttpHeaders;
 import lombok.AllArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.apache.commons.io.FilenameUtils;
 import org.nd4j.linalg.io.ClassPathResource;
-import play.api.libs.MimeTypes;
 import play.mvc.Result;
+import play.mvc.StaticFileMimeTypes;
 
 import java.io.InputStream;
-import java.util.function.Function;
+import java.util.Optional;
 
-import static play.mvc.Http.Context.Implicit.response;
 import static play.mvc.Results.ok;
 
 /**
@@ -37,11 +36,9 @@ import static play.mvc.Results.ok;
  */
 @AllArgsConstructor
 @Slf4j
-public class Assets implements Function<String, Result> {
-    private final String assetsRootDirectory;
+public class Assets {
 
-    @Override
-    public Result apply(String s) {
+    public static Result assetRequest(String assetsRootDirectory, String s) {
 
         String fullPath;
         if(s.startsWith("webjars/")){
@@ -60,15 +57,12 @@ public class Assets implements Function<String, Result> {
 
         String fileName = FilenameUtils.getName(fullPath);
 
-        response().setHeader(HttpHeaders.CONTENT_DISPOSITION, "attachment; filename=\"" + fileName + "\"");
-        scala.Option<String> contentType = MimeTypes.forFileName(fileName);
+        Optional<String> contentType = StaticFileMimeTypes.fileMimeTypes().forFileName(fileName);
         String ct;
-        if (contentType.isDefined()) {
-            ct = contentType.get();
-        } else {
-            ct = "application/octet-stream";
-        }
+        ct = contentType.orElse("application/octet-stream");
 
-        return ok(inputStream).as(ct);
+        return ok(inputStream)
+                .withHeader(HttpHeaders.CONTENT_DISPOSITION, "attachment; filename=\"" + fileName + "\"")
+                .as(ct);
     }
 }
diff --git a/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/IntegrationTestBaselineGenerator.java b/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/IntegrationTestBaselineGenerator.java
index be75282e5..e2890fc61 100644
--- a/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/IntegrationTestBaselineGenerator.java
+++ b/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/IntegrationTestBaselineGenerator.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.integration;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import lombok.extern.slf4j.Slf4j;
 import org.apache.commons.io.FileUtils;
 import org.deeplearning4j.datasets.iterator.MultiDataSetWrapperIterator;
diff --git a/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/IntegrationTestRunner.java b/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/IntegrationTestRunner.java
index 9440a59ae..b1de34a2c 100644
--- a/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/IntegrationTestRunner.java
+++ b/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/IntegrationTestRunner.java
@@ -17,8 +17,8 @@
 package org.deeplearning4j.integration;
 
 
-import com.google.common.collect.ImmutableSet;
-import com.google.common.reflect.ClassPath;
+import org.nd4j.shade.guava.collect.ImmutableSet;
+import org.nd4j.shade.guava.reflect.ClassPath;
 import org.deeplearning4j.integration.util.CountingMultiDataSetIterator;
 import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
diff --git a/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/testcases/RNNTestCases.java b/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/testcases/RNNTestCases.java
index 463734b82..a70d8dd2f 100644
--- a/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/testcases/RNNTestCases.java
+++ b/deeplearning4j/dl4j-integration-tests/src/test/java/org/deeplearning4j/integration/testcases/RNNTestCases.java
@@ -16,7 +16,7 @@
 
 package org.deeplearning4j.integration.testcases;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import org.deeplearning4j.integration.TestCase;
 import org.deeplearning4j.integration.testcases.misc.CharacterIterator;
 import org.deeplearning4j.integration.testcases.misc.CompositeMultiDataSetPreProcessor;
diff --git a/deeplearning4j/pom.xml b/deeplearning4j/pom.xml
index 954d8341d..a139c4f44 100644
--- a/deeplearning4j/pom.xml
+++ b/deeplearning4j/pom.xml
@@ -174,11 +174,6 @@
                 <artifactId>slf4j-api</artifactId>
                 <version>${slf4j.version}</version>
             </dependency>
-            <dependency>
-                <groupId>com.google.guava</groupId>
-                <artifactId>guava</artifactId>
-                <version>${guava.version}</version>
-            </dependency>
             <dependency>
                 <groupId>junit</groupId>
                 <artifactId>junit</artifactId>
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/ListenerVariables.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/ListenerVariables.java
index 1156de102..34b305001 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/ListenerVariables.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/ListenerVariables.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.autodiff.listeners;
 
-import com.google.common.collect.Sets;
+import org.nd4j.shade.guava.collect.Sets;
 
 import java.util.Arrays;
 import java.util.HashSet;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/checkpoint/CheckpointListener.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/checkpoint/CheckpointListener.java
index 1932a6c75..7dbb0119d 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/checkpoint/CheckpointListener.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/checkpoint/CheckpointListener.java
@@ -1,7 +1,7 @@
 package org.nd4j.autodiff.listeners.checkpoint;
 
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
 import org.apache.commons.io.IOUtils;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/records/EvaluationRecord.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/records/EvaluationRecord.java
index 36334b648..b063e18a2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/records/EvaluationRecord.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/listeners/records/EvaluationRecord.java
@@ -16,10 +16,10 @@
 
 package org.nd4j.autodiff.listeners.records;
 
-import com.google.common.base.Predicates;
-import com.google.common.collect.Collections2;
-import com.google.common.collect.ImmutableMap;
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.base.Predicates;
+import org.nd4j.shade.guava.collect.Collections2;
+import org.nd4j.shade.guava.collect.ImmutableMap;
+import org.nd4j.shade.guava.collect.Lists;
 
 import java.util.ArrayList;
 import java.util.Collection;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SameDiff.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SameDiff.java
index e6f30d12e..e09ceda75 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SameDiff.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SameDiff.java
@@ -16,11 +16,11 @@
 
 package org.nd4j.autodiff.samediff;
 
-import com.google.common.base.Predicates;
-import com.google.common.collect.HashBasedTable;
-import com.google.common.collect.Maps;
-import com.google.common.collect.Table;
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.base.Predicates;
+import org.nd4j.shade.guava.collect.HashBasedTable;
+import org.nd4j.shade.guava.collect.Maps;
+import org.nd4j.shade.guava.collect.Table;
+import org.nd4j.shade.guava.primitives.Ints;
 import com.google.flatbuffers.FlatBufferBuilder;
 import lombok.*;
 import lombok.extern.slf4j.Slf4j;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDBaseOps.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDBaseOps.java
index 1bde174c1..7da89aa36 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDBaseOps.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDBaseOps.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.autodiff.samediff.ops;
 
-import com.google.common.collect.Sets;
+import org.nd4j.shade.guava.collect.Sets;
 import java.util.HashMap;
 import java.util.Map;
 import java.util.Set;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/serde/FlatBuffersMapper.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/serde/FlatBuffersMapper.java
index ef1bfefb7..6faf29bfc 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/serde/FlatBuffersMapper.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/serde/FlatBuffersMapper.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.autodiff.samediff.serde;
 
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Ints;
 import com.google.flatbuffers.FlatBufferBuilder;
 import java.nio.ByteOrder;
 import java.util.*;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/validation/OpValidation.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/validation/OpValidation.java
index 3e329ad13..5bc175952 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/validation/OpValidation.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/validation/OpValidation.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.autodiff.validation;
 
-import com.google.common.collect.ImmutableSet;
-import com.google.common.reflect.ClassPath;
+import org.nd4j.shade.guava.collect.ImmutableSet;
+import org.nd4j.shade.guava.reflect.ClassPath;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
 import org.nd4j.autodiff.functions.DifferentialFunction;
@@ -560,7 +560,7 @@ public class OpValidation {
         ImmutableSet<ClassPath.ClassInfo> info;
         try {
             //Dependency note: this ClassPath class was added in Guava 14
-            info = com.google.common.reflect.ClassPath.from(DifferentialFunctionClassHolder.class.getClassLoader())
+            info = org.nd4j.shade.guava.reflect.ClassPath.from(DifferentialFunctionClassHolder.class.getClassLoader())
                     .getTopLevelClassesRecursive("org.nd4j.linalg.api.ops");
         } catch (IOException e) {
             //Should never happen
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/BaseEvaluation.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/BaseEvaluation.java
index e07b5c9b8..fd08e4270 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/BaseEvaluation.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/BaseEvaluation.java
@@ -42,6 +42,7 @@ import org.nd4j.shade.jackson.databind.DeserializationFeature;
 import org.nd4j.shade.jackson.databind.MapperFeature;
 import org.nd4j.shade.jackson.databind.ObjectMapper;
 import org.nd4j.shade.jackson.databind.SerializationFeature;
+import org.nd4j.shade.jackson.databind.exc.InvalidTypeIdException;
 import org.nd4j.shade.jackson.databind.module.SimpleModule;
 import org.nd4j.shade.jackson.dataformat.yaml.YAMLFactory;
 
@@ -80,7 +81,8 @@ public abstract class BaseEvaluation<T extends BaseEvaluation> implements IEvalu
                 .withFieldVisibility(JsonAutoDetect.Visibility.ANY)
                 .withGetterVisibility(JsonAutoDetect.Visibility.NONE)
                 .withSetterVisibility(JsonAutoDetect.Visibility.NONE)
-                .withCreatorVisibility(JsonAutoDetect.Visibility.NONE));
+                .withCreatorVisibility(JsonAutoDetect.Visibility.ANY)
+        );
         return ret;
     }
 
@@ -107,15 +109,15 @@ public abstract class BaseEvaluation<T extends BaseEvaluation> implements IEvalu
     public static <T extends IEvaluation> T fromJson(String json, Class<T> clazz) {
         try {
             return objectMapper.readValue(json, clazz);
-        } catch (IllegalArgumentException e) {
-            if (e.getMessage().contains("Invalid type id")) {
+        } catch (InvalidTypeIdException e) {
+            if (e.getMessage().contains("Could not resolve type id")) {
                 try {
                     return (T) attempFromLegacyFromJson(json, e);
                 } catch (Throwable t) {
                     throw new RuntimeException("Cannot deserialize from JSON - JSON is invalid?", t);
                 }
             }
-            throw e;
+            throw new RuntimeException(e);
         } catch (IOException e) {
             throw new RuntimeException(e);
         }
@@ -129,7 +131,7 @@ public abstract class BaseEvaluation<T extends BaseEvaluation> implements IEvalu
      * @param json              JSON to attempt to deserialize
      * @param originalException Original exception to be re-thrown if it isn't legacy JSON
      */
-    protected static <T extends IEvaluation> T attempFromLegacyFromJson(String json, IllegalArgumentException originalException) {
+    protected static <T extends IEvaluation> T attempFromLegacyFromJson(String json, InvalidTypeIdException originalException) throws InvalidTypeIdException {
         if (json.contains("org.deeplearning4j.eval.Evaluation")) {
             String newJson = json.replaceAll("org.deeplearning4j.eval.Evaluation", "org.nd4j.evaluation.classification.Evaluation");
             return (T) fromJson(newJson, Evaluation.class);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/classification/ConfusionMatrix.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/classification/ConfusionMatrix.java
index dafe571f0..01fd322e2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/classification/ConfusionMatrix.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/classification/ConfusionMatrix.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.evaluation.classification;
 
-import com.google.common.collect.HashMultiset;
-import com.google.common.collect.Multiset;
+import org.nd4j.shade.guava.collect.HashMultiset;
+import org.nd4j.shade.guava.collect.Multiset;
 import lombok.Getter;
 
 import java.io.Serializable;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/custom/CustomEvaluation.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/custom/CustomEvaluation.java
index fed8c63b4..26ef8bba9 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/custom/CustomEvaluation.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/custom/CustomEvaluation.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.evaluation.custom;
 
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.collect.Lists;
 import java.io.Serializable;
 import java.util.ArrayList;
 import java.util.List;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/custom/MergeLambda.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/custom/MergeLambda.java
index cbad73da3..079755055 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/custom/MergeLambda.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/custom/MergeLambda.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.evaluation.custom;
 
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.collect.Lists;
 
 import java.util.List;
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/serde/ConfusionMatrixSerializer.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/serde/ConfusionMatrixSerializer.java
index 6acc7fc4e..a8f5c32e9 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/serde/ConfusionMatrixSerializer.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/serde/ConfusionMatrixSerializer.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.evaluation.serde;
 
-import com.google.common.collect.Multiset;
+import org.nd4j.shade.guava.collect.Multiset;
 import org.nd4j.evaluation.classification.ConfusionMatrix;
 import org.nd4j.shade.jackson.core.JsonGenerator;
 import org.nd4j.shade.jackson.core.JsonProcessingException;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/onnx/OnnxGraphMapper.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/onnx/OnnxGraphMapper.java
index 719ac792d..7a651fb88 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/onnx/OnnxGraphMapper.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/onnx/OnnxGraphMapper.java
@@ -18,9 +18,9 @@ package org.nd4j.imports.graphmapper.onnx;
 
 import org.nd4j.shade.protobuf.ByteString;
 import org.nd4j.shade.protobuf.Message;
-import com.google.common.primitives.Floats;
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Floats;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.val;
 import onnx.Onnx;
 import org.nd4j.autodiff.functions.DifferentialFunction;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/TFGraphMapper.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/TFGraphMapper.java
index f57fef4c7..3ad3267c2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/TFGraphMapper.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/graphmapper/tf/TFGraphMapper.java
@@ -17,8 +17,8 @@
 package org.nd4j.imports.graphmapper.tf;
 
 import org.nd4j.shade.protobuf.Message;
-import com.google.common.primitives.Floats;
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Floats;
+import org.nd4j.shade.guava.primitives.Ints;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
 import org.nd4j.autodiff.functions.DifferentialFunction;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
index 04a61f9d0..50ab2727d 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
@@ -17,8 +17,8 @@
 package org.nd4j.linalg.api.ndarray;
 
 
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import com.google.flatbuffers.FlatBufferBuilder;
 import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
index 0d7aca8e0..952c4b777 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.linalg.api.ndarray;
 
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.NonNull;
 import lombok.extern.slf4j.Slf4j;
 import net.ericaro.neoitertools.Generator;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCOO.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCOO.java
index 1e85be0cd..77463c456 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCOO.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCOO.java
@@ -16,9 +16,9 @@
 
 package org.nd4j.linalg.api.ndarray;
 
-import com.google.common.primitives.Doubles;
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Doubles;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import com.google.flatbuffers.FlatBufferBuilder;
 import net.ericaro.neoitertools.Generator;
 import org.nd4j.linalg.api.blas.params.MMulTranspose;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCSR.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCSR.java
index bd0f7c905..dd17fa635 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCSR.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCSR.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ndarray;
 
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Ints;
 import org.nd4j.linalg.api.buffer.DataBuffer;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.indexing.*;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseReduceOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseReduceOp.java
index 7fc0679db..10c26d29e 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseReduceOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/BaseReduceOp.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops;
 
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Ints;
 import lombok.Getter;
 import lombok.Setter;
 import lombok.extern.slf4j.Slf4j;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java
index 27e8ae281..d2190098c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/DynamicCustomOp.java
@@ -16,9 +16,9 @@
 
 package org.nd4j.linalg.api.ops;
 
-import com.google.common.collect.Lists;
-import com.google.common.primitives.Doubles;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.collect.Lists;
+import org.nd4j.shade.guava.primitives.Doubles;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.*;
 import lombok.extern.slf4j.Slf4j;
 import onnx.Onnx;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/aggregates/Batch.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/aggregates/Batch.java
index 8e352be13..9ae09a57d 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/aggregates/Batch.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/aggregates/Batch.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.aggregates;
 
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.collect.Lists;
 import lombok.Getter;
 import lombok.Setter;
 import lombok.extern.slf4j.Slf4j;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Switch.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Switch.java
index 331dea887..e94a7bc54 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Switch.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/controlflow/compat/Switch.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.controlflow.compat;
 
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.collect.Lists;
 import lombok.Getter;
 import lombok.val;
 import org.nd4j.autodiff.samediff.SDVariable;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/TensorMmul.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/TensorMmul.java
index 3de44537a..d0c1bae38 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/TensorMmul.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/reduce/TensorMmul.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.linalg.api.ops.impl.reduce;
 
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.NoArgsConstructor;
 import lombok.val;
 import onnx.Onnx;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java
index b9687e598..782c70859 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Transpose.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.api.ops.impl.shape;
 
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Ints;
 import lombok.val;
 import onnx.Onnx;
 import org.nd4j.autodiff.samediff.SDVariable;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Choose.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Choose.java
index 28151a899..0a2ab4f20 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Choose.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/Choose.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.linalg.api.ops.impl.transforms.custom;
 
-import com.google.common.primitives.Doubles;
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Doubles;
+import org.nd4j.shade.guava.primitives.Ints;
 import org.nd4j.autodiff.samediff.SDVariable;
 import org.nd4j.autodiff.samediff.SameDiff;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/shape/Shape.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/shape/Shape.java
index bbe133dbb..ef4331b0a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/shape/Shape.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/shape/Shape.java
@@ -17,8 +17,8 @@
 package org.nd4j.linalg.api.shape;
 
 
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.NonNull;
 import lombok.val;
 import org.nd4j.base.Preconditions;
@@ -31,8 +31,6 @@ import org.nd4j.linalg.api.shape.options.ArrayType;
 import org.nd4j.linalg.exception.ND4JIllegalStateException;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.indexing.INDArrayIndex;
-import org.nd4j.linalg.indexing.NDArrayIndex;
-import org.nd4j.linalg.primitives.Pair;
 import org.nd4j.linalg.util.ArrayUtil;
 
 import java.nio.*;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/BalanceMinibatches.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/BalanceMinibatches.java
index 0636cc75b..8eff5cf5d 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/BalanceMinibatches.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/BalanceMinibatches.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.linalg.dataset;
 
-import com.google.common.collect.Lists;
-import com.google.common.collect.Maps;
+import org.nd4j.shade.guava.collect.Lists;
+import org.nd4j.shade.guava.collect.Maps;
 import lombok.AllArgsConstructor;
 import lombok.Builder;
 import lombok.Data;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java
index a56d39567..7b208f4f2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.linalg.dataset;
 
-import com.google.common.base.Function;
-import com.google.common.collect.Lists;
+import org.nd4j.shade.guava.base.Function;
+import org.nd4j.shade.guava.collect.Lists;
 import lombok.extern.slf4j.Slf4j;
 import org.nd4j.base.Preconditions;
 import org.nd4j.linalg.api.ndarray.INDArray;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/DataSet.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/DataSet.java
index a5a46dbcd..7358da385 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/DataSet.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/DataSet.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.dataset.api;
 
-import com.google.common.base.Function;
+import org.nd4j.shade.guava.base.Function;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.api.rng.Random;
 import org.nd4j.linalg.dataset.SplitTestAndTrain;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/BaseNDArrayFactory.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/BaseNDArrayFactory.java
index 487196912..1edf0d651 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/BaseNDArrayFactory.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/BaseNDArrayFactory.java
@@ -1281,7 +1281,7 @@ public abstract class BaseNDArrayFactory implements NDArrayFactory {
     public INDArray scalar(Number value) {
         MemoryWorkspace  ws = Nd4j.getMemoryManager().getCurrentWorkspace();
 
-        if (value instanceof Double || value instanceof AtomicDouble)   /* note that org.nd4j.linalg.primitives.AtomicDouble extends com.google.common.util.concurrent.AtomicDouble */
+        if (value instanceof Double || value instanceof AtomicDouble)   /* note that org.nd4j.linalg.primitives.AtomicDouble extends org.nd4j.shade.guava.util.concurrent.AtomicDouble */
             return scalar(value.doubleValue());
         else if (value instanceof Float)
             return scalar(value.floatValue());
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
index c8baedfa5..0bbf69ebe 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.linalg.factory;
 
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.NonNull;
 import lombok.val;
 import lombok.var;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/Indices.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/Indices.java
index f715f93ae..3ca99c50e 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/Indices.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/Indices.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.linalg.indexing;
 
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.api.shape.Shape;
 import org.nd4j.linalg.exception.ND4JIllegalStateException;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/IntervalIndex.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/IntervalIndex.java
index 68190f4e5..b055430a7 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/IntervalIndex.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/IntervalIndex.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.indexing;
 
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.Getter;
 import org.nd4j.base.Preconditions;
 import org.nd4j.linalg.api.ndarray.INDArray;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/NDArrayIndex.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/NDArrayIndex.java
index da0dafb72..c21993548 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/NDArrayIndex.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/NDArrayIndex.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.linalg.indexing;
 
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/PointIndex.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/PointIndex.java
index 283db3e0f..33cb996e9 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/PointIndex.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/PointIndex.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.indexing;
 
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.EqualsAndHashCode;
 import org.nd4j.linalg.api.ndarray.INDArray;
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/SpecifiedIndex.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/SpecifiedIndex.java
index 6c1b93b8f..7b56f9eed 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/SpecifiedIndex.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/SpecifiedIndex.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.indexing;
 
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.Data;
 import net.ericaro.neoitertools.Generator;
 import net.ericaro.neoitertools.Itertools;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/conditions/Condition.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/conditions/Condition.java
index 97a45f2c2..3b4245fef 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/conditions/Condition.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/conditions/Condition.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.indexing.conditions;
 
-import com.google.common.base.Function;
+import org.nd4j.shade.guava.base.Function;
 
 /**
  * Condition for boolean indexing
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Identity.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Identity.java
index 83ab43320..0413e2505 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Identity.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Identity.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.indexing.functions;
 
-import com.google.common.base.Function;
+import org.nd4j.shade.guava.base.Function;
 
 /**
  * Created by agibsonccc on 10/8/14.
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/StableNumber.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/StableNumber.java
index 687fa53ca..e044dcda8 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/StableNumber.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/StableNumber.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.indexing.functions;
 
 
-import com.google.common.base.Function;
+import org.nd4j.shade.guava.base.Function;
 import org.nd4j.linalg.factory.Nd4j;
 
 /**
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Value.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Value.java
index fc305d0b3..8a6a6436f 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Value.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Value.java
@@ -17,7 +17,7 @@
 package org.nd4j.linalg.indexing.functions;
 
 
-import com.google.common.base.Function;
+import org.nd4j.shade.guava.base.Function;
 
 /**
  * Created by agibsonccc on 10/8/14.
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Zero.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Zero.java
index fac1d0afc..9bcb3c463 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Zero.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Zero.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.indexing.functions;
 
-import com.google.common.base.Function;
+import org.nd4j.shade.guava.base.Function;
 
 /**
  * Created by agibsonccc on 10/8/14.
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/handler/MemoryHandler.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/handler/MemoryHandler.java
index 36d8e05fb..abb919e8c 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/handler/MemoryHandler.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/handler/MemoryHandler.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.jita.handler;
 
-import com.google.common.collect.Table;
+import org.nd4j.shade.guava.collect.Table;
 import org.bytedeco.javacpp.Pointer;
 import org.nd4j.jita.allocator.Allocator;
 import org.nd4j.jita.allocator.enums.AllocationStatus;
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/handler/impl/CudaZeroHandler.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/handler/impl/CudaZeroHandler.java
index f6a0eafc0..fdd40f8cb 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/handler/impl/CudaZeroHandler.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/jita/handler/impl/CudaZeroHandler.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.jita.handler.impl;
 
-import com.google.common.collect.HashBasedTable;
-import com.google.common.collect.Table;
+import org.nd4j.shade.guava.collect.HashBasedTable;
+import org.nd4j.shade.guava.collect.Table;
 import lombok.Getter;
 import lombok.NonNull;
 import lombok.val;
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/util/CudaArgs.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/util/CudaArgs.java
index 6c8ea85e1..1922d9ced 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/util/CudaArgs.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/util/CudaArgs.java
@@ -17,8 +17,8 @@
 package org.nd4j.linalg.jcublas.util;
 
 
-import com.google.common.collect.ArrayListMultimap;
-import com.google.common.collect.Multimap;
+import org.nd4j.shade.guava.collect.ArrayListMultimap;
+import org.nd4j.shade.guava.collect.Multimap;
 import lombok.AllArgsConstructor;
 import lombok.Data;
 import org.nd4j.linalg.api.ndarray.INDArray;
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/indexing/IndexingTestsC.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/indexing/IndexingTestsC.java
index 0904bdaee..960bbd646 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/indexing/IndexingTestsC.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/indexing/IndexingTestsC.java
@@ -16,7 +16,6 @@
 
 package org.nd4j.linalg.api.indexing;
 
-import org.joda.time.Interval;
 import org.junit.Test;
 import org.junit.runner.RunWith;
 import org.junit.runners.Parameterized;
diff --git a/nd4j/nd4j-common/pom.xml b/nd4j/nd4j-common/pom.xml
index d423e8012..d75c82cf4 100644
--- a/nd4j/nd4j-common/pom.xml
+++ b/nd4j/nd4j-common/pom.xml
@@ -83,6 +83,12 @@
             <version>${project.version}</version>
         </dependency>
 
+        <dependency>
+            <groupId>org.nd4j</groupId>
+            <artifactId>guava</artifactId>
+            <version>${project.version}</version>
+        </dependency>
+
         <dependency>
             <groupId>org.slf4j</groupId>
             <artifactId>slf4j-api</artifactId>
@@ -117,11 +123,6 @@
             <version>${commons-compress.version}</version>
         </dependency>
 
-        <dependency>
-            <groupId>com.google.guava</groupId>
-            <artifactId>guava</artifactId>
-        </dependency>
-
         <dependency>
             <groupId>commons-codec</groupId>
             <artifactId>commons-codec</artifactId>
diff --git a/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/collection/IntArrayKeyMap.java b/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/collection/IntArrayKeyMap.java
index f5ba64f54..72409e3ca 100644
--- a/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/collection/IntArrayKeyMap.java
+++ b/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/collection/IntArrayKeyMap.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.collection;
 
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Ints;
 import lombok.Getter;
 import org.nd4j.base.Preconditions;
 
diff --git a/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/primitives/AtomicDouble.java b/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/primitives/AtomicDouble.java
index 1f5c1f9cb..e17fec5ce 100644
--- a/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/primitives/AtomicDouble.java
+++ b/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/primitives/AtomicDouble.java
@@ -24,7 +24,7 @@ import org.nd4j.shade.jackson.databind.annotation.JsonSerialize;
 
 @JsonSerialize(using = JsonSerializerAtomicDouble.class)
 @JsonDeserialize(using = JsonDeserializerAtomicDouble.class)
-public class AtomicDouble extends com.google.common.util.concurrent.AtomicDouble {
+public class AtomicDouble extends org.nd4j.shade.guava.util.concurrent.AtomicDouble {
 
     public AtomicDouble(){
         this(0.0);
@@ -40,7 +40,7 @@ public class AtomicDouble extends com.google.common.util.concurrent.AtomicDouble
 
     @Override
     public boolean equals(Object o){
-        //NOTE: com.google.common.util.concurrent.AtomicDouble extends Number, hence this class extends number
+        //NOTE: org.nd4j.shade.guava.util.concurrent.AtomicDouble extends Number, hence this class extends number
         if(o instanceof Number){
             return get() == ((Number)o).doubleValue();
         }
diff --git a/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/util/ArrayUtil.java b/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/util/ArrayUtil.java
index e51e75ce4..2fe33dfba 100644
--- a/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/util/ArrayUtil.java
+++ b/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/util/ArrayUtil.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.linalg.util;
 
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import lombok.val;
 import org.apache.commons.lang3.RandomUtils;
 import org.nd4j.base.Preconditions;
diff --git a/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/util/SynchronizedTable.java b/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/util/SynchronizedTable.java
index e08605881..0c4377ea9 100644
--- a/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/util/SynchronizedTable.java
+++ b/nd4j/nd4j-common/src/main/java/org/nd4j/linalg/util/SynchronizedTable.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.util;
 
-import com.google.common.collect.Table;
+import org.nd4j.shade.guava.collect.Table;
 
 import java.util.Collection;
 import java.util.Map;
diff --git a/nd4j/nd4j-common/src/main/java/org/nd4j/resources/strumpf/ResourceFile.java b/nd4j/nd4j-common/src/main/java/org/nd4j/resources/strumpf/ResourceFile.java
index 83799381b..f8fca14b5 100644
--- a/nd4j/nd4j-common/src/main/java/org/nd4j/resources/strumpf/ResourceFile.java
+++ b/nd4j/nd4j-common/src/main/java/org/nd4j/resources/strumpf/ResourceFile.java
@@ -1,6 +1,6 @@
 package org.nd4j.resources.strumpf;
 
-import com.google.common.io.Files;
+import org.nd4j.shade.guava.io.Files;
 import lombok.AllArgsConstructor;
 import lombok.Data;
 import lombok.NoArgsConstructor;
diff --git a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/src/main/java/org/nd4j/parameterserver/distributed/transport/RoutedTransport.java b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/src/main/java/org/nd4j/parameterserver/distributed/transport/RoutedTransport.java
index e7c8d977b..1b420d6a4 100644
--- a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/src/main/java/org/nd4j/parameterserver/distributed/transport/RoutedTransport.java
+++ b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/src/main/java/org/nd4j/parameterserver/distributed/transport/RoutedTransport.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.parameterserver.distributed.transport;
 
-import com.google.common.math.IntMath;
+import org.nd4j.shade.guava.math.IntMath;
 import io.aeron.Aeron;
 import io.aeron.FragmentAssembler;
 import io.aeron.Publication;
diff --git a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/src/main/java/org/nd4j/parameterserver/distributed/v2/transport/impl/AeronUdpTransport.java b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/src/main/java/org/nd4j/parameterserver/distributed/v2/transport/impl/AeronUdpTransport.java
index a34720093..70228b987 100644
--- a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/src/main/java/org/nd4j/parameterserver/distributed/v2/transport/impl/AeronUdpTransport.java
+++ b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-node/src/main/java/org/nd4j/parameterserver/distributed/v2/transport/impl/AeronUdpTransport.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.parameterserver.distributed.v2.transport.impl;
 
-import com.google.common.math.IntMath;
+import org.nd4j.shade.guava.math.IntMath;
 import io.aeron.Aeron;
 import io.aeron.FragmentAssembler;
 import io.aeron.Publication;
diff --git a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-status/pom.xml b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-status/pom.xml
index 0a75e6171..dd50f938e 100644
--- a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-status/pom.xml
+++ b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-status/pom.xml
@@ -47,16 +47,6 @@
             <artifactId>nd4j-parameter-server</artifactId>
             <version>${project.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-actor_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.typesafe.akka</groupId>
-            <artifactId>akka-slf4j_2.11</artifactId>
-            <version>${akka.version}</version>
-        </dependency>
         <dependency>
             <groupId>joda-time</groupId>
             <artifactId>joda-time</artifactId>
diff --git a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-status/src/main/java/org/nd4j/parameterserver/status/play/StatusServer.java b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-status/src/main/java/org/nd4j/parameterserver/status/play/StatusServer.java
index e5152e327..e8b149f67 100644
--- a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-status/src/main/java/org/nd4j/parameterserver/status/play/StatusServer.java
+++ b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server-status/src/main/java/org/nd4j/parameterserver/status/play/StatusServer.java
@@ -22,17 +22,14 @@ import org.nd4j.parameterserver.model.MasterStatus;
 import org.nd4j.parameterserver.model.ServerTypeJson;
 import org.nd4j.parameterserver.model.SlaveStatus;
 import org.nd4j.parameterserver.model.SubscriberState;
+import play.BuiltInComponents;
 import play.Mode;
-import play.libs.F;
 import play.libs.Json;
-import play.mvc.Result;
+import play.routing.Router;
 import play.routing.RoutingDsl;
 import play.server.Server;
 
-import java.util.List;
-
 import static play.libs.Json.toJson;
-import static play.mvc.Controller.request;
 import static play.mvc.Results.ok;
 
 
@@ -70,74 +67,35 @@ public class StatusServer {
      */
     public static Server startServer(StatusStorage statusStorage, int statusServerPort) {
         log.info("Starting server on port " + statusServerPort);
-        RoutingDsl dsl = new RoutingDsl();
-        dsl.GET("/ids/").routeTo(new F.Function0<Result>() {
-
-            @Override
-            public Result apply() throws Throwable {
-                List<Integer> ids = statusStorage.ids();
-                return ok(toJson(ids));
-            }
-        });
-
-
-        dsl.GET("/state/:id").routeTo(new F.Function<String, Result>() {
-            @Override
-            public Result apply(String id) throws Throwable {
-                return ok(toJson(statusStorage.getState(Integer.parseInt(id))));
-            }
-        });
-
-        dsl.GET("/opType/:id").routeTo(new F.Function<String, Result>() {
-            @Override
-            public Result apply(String id) throws Throwable {
-                return ok(toJson(ServerTypeJson.builder()
-                                .type(statusStorage.getState(Integer.parseInt(id)).serverType())));
-            }
-        });
-
-
-        dsl.GET("/started/:id").routeTo(new F.Function<String, Result>() {
-            @Override
-            public Result apply(String id) throws Throwable {
-                return statusStorage.getState(Integer.parseInt(id)).isMaster()
-                                ? ok(toJson(MasterStatus.builder()
-                                                .master(statusStorage.getState(Integer.parseInt(id)).getServerState())
-                                                //note here that a responder is id + 1
-                                                .responder(statusStorage
-                                                                .getState(Integer.parseInt(id) + 1).getServerState())
-                                                .responderN(statusStorage
-                                                                .getState(Integer.parseInt(id)).getTotalUpdates())
-                                                .build()))
-                                : ok(toJson(SlaveStatus.builder()
-                                                .slave(statusStorage.getState(Integer.parseInt(id)).serverType())
-                                                .build()));
-            }
-        });
-
-
-
-        dsl.GET("/connectioninfo/:id").routeTo(new F.Function<String, Result>() {
-            @Override
-            public Result apply(String id) throws Throwable {
-                return ok(toJson(statusStorage.getState(Integer.parseInt(id)).getConnectionInfo()));
-            }
-        });
-
-        dsl.POST("/updatestatus/:id").routeTo(new F.Function<String, Result>() {
-            @Override
-            public Result apply(String id) throws Throwable {
-                SubscriberState subscriberState = Json.fromJson(request().body().asJson(), SubscriberState.class);
-                statusStorage.updateState(subscriberState);
-                return ok(toJson(subscriberState));
-            }
-        });
-
-        Server server = Server.forRouter(dsl.build(), Mode.PROD, statusServerPort);
-
-        return server;
-
+        return Server.forRouter(Mode.PROD, statusServerPort, builtInComponents -> createRouter(statusStorage, builtInComponents));
     }
 
+    protected static Router createRouter(StatusStorage statusStorage, BuiltInComponents builtInComponents){
+        RoutingDsl dsl = RoutingDsl.fromComponents(builtInComponents);
+        dsl.GET("/ids/").routingTo(request -> ok(toJson(statusStorage.ids())));
+        dsl.GET("/state/:id").routingTo((request, id) -> ok(toJson(statusStorage.getState(Integer.parseInt(id.toString())))));
+        dsl.GET("/opType/:id").routingTo((request, id) -> ok(toJson(ServerTypeJson.builder()
+                .type(statusStorage.getState(Integer.parseInt(id.toString())).serverType()))));
+        dsl.GET("/started/:id").routingTo((request, id) -> {
+            boolean isMaster = statusStorage.getState(Integer.parseInt(id.toString())).isMaster();
+            if(isMaster){
+                return ok(toJson(MasterStatus.builder().master(statusStorage.getState(Integer.parseInt(id.toString())).getServerState())
+                        //note here that a responder is id + 1
+                        .responder(statusStorage.getState(Integer.parseInt(id.toString()) + 1).getServerState())
+                        .responderN(statusStorage.getState(Integer.parseInt(id.toString())).getTotalUpdates())
+                        .build()));
+            } else {
+                return ok(toJson(SlaveStatus.builder().slave(statusStorage.getState(Integer.parseInt(id.toString())).serverType()).build()));
+            }
+        });
+        dsl.GET("/connectioninfo/:id").routingTo((request, id) -> ok(toJson(statusStorage.getState(Integer.parseInt(id.toString())).getConnectionInfo())));
 
+        dsl.POST("/updatestatus/:id").routingTo((request, id) -> {
+            SubscriberState subscriberState = Json.fromJson(request.body().asJson(), SubscriberState.class);
+            statusStorage.updateState(subscriberState);
+            return ok(toJson(subscriberState));
+        });
+
+        return dsl.build();
+    }
 }
diff --git a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server/src/main/java/org/nd4j/parameterserver/ParameterServerSubscriber.java b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server/src/main/java/org/nd4j/parameterserver/ParameterServerSubscriber.java
index f69b9c24b..e5ccca909 100644
--- a/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server/src/main/java/org/nd4j/parameterserver/ParameterServerSubscriber.java
+++ b/nd4j/nd4j-parameter-server-parent/nd4j-parameter-server/src/main/java/org/nd4j/parameterserver/ParameterServerSubscriber.java
@@ -20,7 +20,7 @@ import com.beust.jcommander.JCommander;
 import com.beust.jcommander.Parameter;
 import com.beust.jcommander.ParameterException;
 import com.beust.jcommander.Parameters;
-import com.google.common.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Ints;
 
 import org.nd4j.shade.jackson.databind.ObjectMapper;
 import com.mashape.unirest.http.Unirest;
diff --git a/nd4j/nd4j-serde/nd4j-aeron/src/main/java/org/nd4j/aeron/ipc/chunk/InMemoryChunkAccumulator.java b/nd4j/nd4j-serde/nd4j-aeron/src/main/java/org/nd4j/aeron/ipc/chunk/InMemoryChunkAccumulator.java
index 426131d4d..8ea228586 100644
--- a/nd4j/nd4j-serde/nd4j-aeron/src/main/java/org/nd4j/aeron/ipc/chunk/InMemoryChunkAccumulator.java
+++ b/nd4j/nd4j-serde/nd4j-aeron/src/main/java/org/nd4j/aeron/ipc/chunk/InMemoryChunkAccumulator.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.aeron.ipc.chunk;
 
-import com.google.common.collect.Maps;
+import org.nd4j.shade.guava.collect.Maps;
 import lombok.extern.slf4j.Slf4j;
 import org.nd4j.aeron.ipc.NDArrayMessage;
 
diff --git a/nd4j/nd4j-serde/nd4j-arrow/pom.xml b/nd4j/nd4j-serde/nd4j-arrow/pom.xml
index 99f744587..4e4ba462e 100644
--- a/nd4j/nd4j-serde/nd4j-arrow/pom.xml
+++ b/nd4j/nd4j-serde/nd4j-arrow/pom.xml
@@ -49,36 +49,6 @@
             <artifactId>joda-time</artifactId>
             <version>${jodatime.version}</version>
         </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-core</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-databind</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-annotations</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.dataformat</groupId>
-            <artifactId>jackson-dataformat-yaml</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.dataformat</groupId>
-            <artifactId>jackson-dataformat-xml</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.datatype</groupId>
-            <artifactId>jackson-datatype-joda</artifactId>
-            <version>${spark2.jackson.version}</version>
-        </dependency>
         <dependency>
             <groupId>org.apache.arrow</groupId>
             <artifactId>arrow-vector</artifactId>
diff --git a/nd4j/nd4j-serde/nd4j-gson/pom.xml b/nd4j/nd4j-serde/nd4j-gson/pom.xml
index 1e8c0f509..82770d51a 100644
--- a/nd4j/nd4j-serde/nd4j-gson/pom.xml
+++ b/nd4j/nd4j-serde/nd4j-gson/pom.xml
@@ -21,9 +21,6 @@
         <groupId>org.nd4j</groupId>
         <version>1.0.0-SNAPSHOT</version>
     </parent>
-    <properties>
-        <gson.version>2.8.0</gson.version>
-    </properties>
     <modelVersion>4.0.0</modelVersion>
 
     <artifactId>nd4j-gson</artifactId>
diff --git a/nd4j/nd4j-serde/nd4j-gson/src/main/java/org/nd4j/serde/gson/GsonDeserializationUtils.java b/nd4j/nd4j-serde/nd4j-gson/src/main/java/org/nd4j/serde/gson/GsonDeserializationUtils.java
index fc04af6d2..5b087a055 100644
--- a/nd4j/nd4j-serde/nd4j-gson/src/main/java/org/nd4j/serde/gson/GsonDeserializationUtils.java
+++ b/nd4j/nd4j-serde/nd4j-gson/src/main/java/org/nd4j/serde/gson/GsonDeserializationUtils.java
@@ -16,8 +16,8 @@
 
 package org.nd4j.serde.gson;
 
-import com.google.common.primitives.Ints;
-import com.google.common.primitives.Longs;
+import org.nd4j.shade.guava.primitives.Ints;
+import org.nd4j.shade.guava.primitives.Longs;
 import com.google.gson.JsonArray;
 import com.google.gson.JsonElement;
 import com.google.gson.JsonParser;
diff --git a/nd4j/nd4j-shade/guava/pom.xml b/nd4j/nd4j-shade/guava/pom.xml
new file mode 100644
index 000000000..73b8d5825
--- /dev/null
+++ b/nd4j/nd4j-shade/guava/pom.xml
@@ -0,0 +1,219 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project xmlns="http://maven.apache.org/POM/4.0.0"
+         xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
+         xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
+    <parent>
+        <artifactId>nd4j-shade</artifactId>
+        <groupId>org.nd4j</groupId>
+        <version>1.0.0-SNAPSHOT</version>
+    </parent>
+    <modelVersion>4.0.0</modelVersion>
+
+    <artifactId>guava</artifactId>
+
+    <properties>
+        <skipTestResourceEnforcement>true</skipTestResourceEnforcement>
+    </properties>
+
+    <dependencies>
+        <dependency>
+            <groupId>com.google.guava</groupId>
+            <artifactId>guava</artifactId>
+            <version>28.0-jre</version>
+            <!-- Without optional=true, users will still be able to import this dependency and compile, yet will get a NoClassDefFoundError at runtime -->
+            <optional>true</optional>
+        </dependency>
+    </dependencies>
+
+
+    <profiles>
+        <profile>
+            <id>custom-lifecycle</id>
+
+            <activation>
+                <property><name>!skip.custom.lifecycle</name></property>
+            </activation>
+            <build>
+                <plugins>
+
+                    <plugin>
+                        <groupId>org.apache.portals.jetspeed-2</groupId>
+                        <artifactId>jetspeed-mvn-maven-plugin</artifactId>
+                        <version>2.3.1</version>
+                        <executions>
+                            <execution>
+                                <id>compile-and-pack</id>
+                                <phase>compile</phase>
+                                <goals>
+                                    <goal>mvn</goal>
+                                </goals>
+                            </execution>
+                        </executions>
+                        <dependencies>
+                            <dependency>
+                                <groupId>org.apache.maven.shared</groupId>
+                                <artifactId>maven-invoker</artifactId>
+                                <version>2.2</version>
+                            </dependency>
+                        </dependencies>
+                        <configuration>
+                            <targets combine.children="merge">
+
+                                <target>
+                                    <id>create-shaded-jars</id>
+                                    <dir>@rootdir@/nd4j/nd4j-shade/guava/</dir>
+                                    <goals>clean,compile,package</goals>
+                                    <properties>
+                                        <skip.custom.lifecycle>true</skip.custom.lifecycle>
+                                    </properties>
+                                </target>
+
+                            </targets>
+                            <defaultTarget>create-shaded-jars</defaultTarget>
+                        </configuration>
+                    </plugin>
+                </plugins>
+            </build>
+        </profile>
+    </profiles>
+
+    <build>
+        <plugins>
+
+            <!-- Disable Maven Lint plugin in this module. For some reason it chokes on this module (internal NPE) and we don't need it anyway here -->
+            <plugin>
+                <groupId>com.lewisd</groupId>
+                <artifactId>lint-maven-plugin</artifactId>
+                <version>0.0.11</version>
+                <executions>
+                    <execution>
+                        <id>pom-lint</id>
+                        <phase>none</phase>
+                    </execution>
+                </executions>
+            </plugin>
+
+            <plugin>
+                <groupId>org.apache.maven.plugins</groupId>
+                <artifactId>maven-shade-plugin</artifactId>
+                <version>${maven-shade-plugin.version}</version>
+                <executions>
+                    <execution>
+                        <phase>package</phase>
+                        <goals>
+                            <goal>shade</goal>
+                        </goals>
+                        <configuration>
+                            <transformers>
+                                <transformer implementation="org.apache.maven.plugins.shade.resource.AppendingTransformer">
+                                    <resource>reference.conf</resource>
+                                </transformer>
+                                <transformer implementation="org.apache.maven.plugins.shade.resource.ServicesResourceTransformer"/>
+                                <transformer implementation="org.apache.maven.plugins.shade.resource.ManifestResourceTransformer"/>
+                            </transformers>
+                        </configuration>
+                    </execution>
+                </executions>
+
+                <configuration>
+                    <!--
+                    Important configuration options here:
+                    createDependencyReducedPom: remove the shaded artifacts from the module dependencies. Without this, the
+                                                original dependencies will be shaded, AND still included as transitive deps
+                                                in the final POM. This is not what we want.
+                    shadedArtifactAttached:     If true, the shaded artifact will be a separate JAR file for install, with
+                                                the original un-shaded JAR being separate. With this being set to false,
+                                                the original JAR will be modified, and no extra jar will be produced.
+                    promoteTransitiveDependencies:  This will promote the transitive dependencies of the shaded dependencies
+                                                to direct dependencies. Without this, we need to manually manage the transitive
+                                                dependencies of the shaded artifacts.
+
+                    Note that using <optional>true</optional> in the dependencies also allows the deps to be shaded (and
+                    original dependencies to not be included), but does NOT work with promoteTransitiveDependencies
+                    -->
+                    <shadedArtifactAttached>false</shadedArtifactAttached>
+                    <createDependencyReducedPom>true</createDependencyReducedPom>
+                    <promoteTransitiveDependencies>true</promoteTransitiveDependencies>
+
+                    <artifactSet>
+                        <includes>
+                            <include>com.google.*:*</include>
+                        </includes>
+                    </artifactSet>
+
+                    <relocations>
+                        <relocation>
+                            <pattern>com.google.common</pattern>
+                            <shadedPattern>org.nd4j.shade.guava</shadedPattern>
+                        </relocation>
+                        <relocation>
+                            <pattern>com.google</pattern>
+                            <shadedPattern>org.nd4j.shade</shadedPattern>
+                        </relocation>
+                    </relocations>
+
+                </configuration>
+            </plugin>
+            <plugin>
+                <groupId>org.apache.maven.plugins</groupId>
+                <artifactId>maven-jar-plugin</artifactId>
+                <configuration>
+                    <forceCreation>true</forceCreation>
+                </configuration>
+                <executions>
+                    <execution>
+                        <id>empty-javadoc-jar</id>
+                        <phase>package</phase>
+                        <goals>
+                            <goal>jar</goal>
+                        </goals>
+                        <configuration>
+                            <classifier>javadoc</classifier>
+                            <classesDirectory>${basedir}/javadoc</classesDirectory>
+                        </configuration>
+                    </execution>
+                    <execution>
+                        <id>empty-sources-jar</id>
+                        <phase>package</phase>
+                        <goals>
+                            <goal>jar</goal>
+                        </goals>
+                        <configuration>
+                            <classifier>sources</classifier>
+                            <classesDirectory>${basedir}/src</classesDirectory>
+                        </configuration>
+                    </execution>
+                </executions>
+            </plugin>
+
+            <plugin>
+                <groupId>org.apache.maven.plugins</groupId>
+                <artifactId>maven-dependency-plugin</artifactId>
+                <version>3.0.0</version>
+                <executions>
+                    <execution>
+                        <id>unpack</id>
+                        <phase>package</phase>
+                        <goals>
+                            <goal>unpack</goal>
+                        </goals>
+                        <configuration>
+                            <artifactItems>
+                                <artifactItem>
+                                    <groupId>org.nd4j</groupId>
+                                    <artifactId>guava</artifactId>
+                                    <version>${project.version}</version>
+                                    <type>jar</type>
+                                    <overWrite>false</overWrite>
+                                    <outputDirectory>${project.build.directory}/classes/</outputDirectory>
+                                    <includes>**/*.class,**/*.xml</includes>
+                                </artifactItem>
+                            </artifactItems>
+                        </configuration>
+                    </execution>
+                </executions>
+            </plugin>
+        </plugins>
+    </build>
+
+</project>
\ No newline at end of file
diff --git a/nd4j/nd4j-shade/jackson/pom.xml b/nd4j/nd4j-shade/jackson/pom.xml
index 1d53e1c41..ad2be71fb 100644
--- a/nd4j/nd4j-shade/jackson/pom.xml
+++ b/nd4j/nd4j-shade/jackson/pom.xml
@@ -32,6 +32,79 @@
         <skipTestResourceEnforcement>true</skipTestResourceEnforcement>
     </properties>
 
+
+
+    <dependencies>
+        <dependency>
+            <groupId>com.fasterxml.jackson.core</groupId>
+            <artifactId>jackson-core</artifactId>
+            <version>${jackson.version}</version>
+            <!-- Without optional=true, users will still be able to import this dependency and compile, yet will get a NoClassDefFoundError at runtime -->
+            <optional>true</optional>
+        </dependency>
+        <dependency>
+            <groupId>com.fasterxml.jackson.core</groupId>
+            <artifactId>jackson-databind</artifactId>
+            <version>${jackson.databind.version}</version>
+            <optional>true</optional>
+        </dependency>
+        <dependency>
+            <groupId>com.fasterxml.jackson.dataformat</groupId>
+            <artifactId>jackson-dataformat-yaml</artifactId>
+            <version>${jackson.version}</version>
+            <optional>true</optional>
+        </dependency>
+        <dependency>
+            <groupId>com.fasterxml.jackson.dataformat</groupId>
+            <artifactId>jackson-dataformat-xml</artifactId>
+            <version>${jackson.version}</version>
+            <optional>true</optional>
+
+            <!-- Exclude Jaxb annotation module to avoid issues with Spring framework tools.
+                See DL4J Issue: https://github.com/deeplearning4j/deeplearning4j/issues/2459 -->
+            <exclusions>
+                <exclusion>
+                    <artifactId>jackson-module-jaxb-annotations</artifactId>
+                    <groupId>com.fasterxml.jackson.module</groupId>
+                </exclusion>
+            </exclusions>
+        </dependency>
+        <dependency>
+            <groupId>com.fasterxml.jackson.datatype</groupId>
+            <artifactId>jackson-datatype-joda</artifactId>
+            <version>${jackson.version}</version>
+            <optional>true</optional>
+        </dependency>
+
+        <!-- Manually promote transitive dependencies to direct also. Optional fixes one issue, but makes all transitive essentially provided scope -->
+        <dependency>
+            <groupId>com.fasterxml.jackson.core</groupId>
+            <artifactId>jackson-annotations</artifactId>
+            <version>${jackson.version}</version>
+            <optional>true</optional>
+        </dependency>
+        <dependency>
+            <groupId>org.yaml</groupId>
+            <artifactId>snakeyaml</artifactId>
+            <version>${shaded.snakeyaml.version}</version>
+            <optional>true</optional>
+        </dependency>
+        <dependency>
+            <groupId>org.codehaus.woodstox</groupId>
+            <artifactId>stax2-api</artifactId>
+            <version>3.1.4</version>
+            <optional>true</optional>
+        </dependency>
+        <dependency>
+            <groupId>com.fasterxml.woodstox</groupId>
+            <artifactId>woodstox-core</artifactId>
+            <version>5.1.0</version>
+            <optional>true</optional>
+        </dependency>
+
+    </dependencies>
+
+
     <profiles>
       <profile>
         <id>custom-lifecycle</id>
@@ -139,6 +212,9 @@
                         <includes>
                             <include>com.fasterxml.jackson:*</include>
                             <include>com.fasterxml.jackson.*:*</include>
+                            <include>com.fasterxml.woodstox:*</include>
+                            <include>org.yaml*:*</include>
+                            <include>org.codehaus*:*</include>
                         </includes>
                     </artifactSet>
 
@@ -148,6 +224,20 @@
                             <pattern>com.fasterxml.jackson</pattern>
                             <shadedPattern>org.nd4j.shade.jackson</shadedPattern>
                         </relocation>
+                        <relocation>
+                            <pattern>com.ctc.wstx</pattern>
+                            <shadedPattern>org.nd4j.shade.wstx</shadedPattern>
+                        </relocation>
+                        <!-- Snake YAML that Jackson depends on -->
+                        <relocation>
+                            <pattern>org.yaml</pattern>
+                            <shadedPattern>org.nd4j.shade.yaml</shadedPattern>
+                        </relocation>
+                        <!-- stax2-api - from jackson dataformat xml -->
+                        <relocation>
+                            <pattern>org.codehaus</pattern>
+                            <shadedPattern>org.nd4j.shade.codehaus</shadedPattern>
+                        </relocation>
                     </relocations>
 
                 </configuration>
@@ -214,45 +304,4 @@
             </plugin>
         </plugins>
     </build>
-
-    <dependencies>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-core</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.core</groupId>
-            <artifactId>jackson-databind</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.dataformat</groupId>
-            <artifactId>jackson-dataformat-yaml</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.dataformat</groupId>
-            <artifactId>jackson-dataformat-xml</artifactId>
-            <version>${jackson.version}</version>
-
-            <!-- Exclude Jaxb annotation module to avoid issues with Spring framework tools.
-                See DL4J Issue: https://github.com/deeplearning4j/deeplearning4j/issues/2459 -->
-            <exclusions>
-                <exclusion>
-                    <artifactId>jackson-module-jaxb-annotations</artifactId>
-                    <groupId>com.fasterxml.jackson.module</groupId>
-                </exclusion>
-            </exclusions>
-        </dependency>
-        <dependency>
-            <groupId>com.fasterxml.jackson.datatype</groupId>
-            <artifactId>jackson-datatype-joda</artifactId>
-            <version>${jackson.version}</version>
-        </dependency>
-
-
-
-    </dependencies>
-
 </project>
diff --git a/nd4j/nd4j-shade/pom.xml b/nd4j/nd4j-shade/pom.xml
index 36b58087b..927292b5f 100644
--- a/nd4j/nd4j-shade/pom.xml
+++ b/nd4j/nd4j-shade/pom.xml
@@ -30,6 +30,7 @@
     <modules>
         <module>jackson</module>
         <module>protobuf</module>
+        <module>guava</module>
     </modules>
 
     <properties>
diff --git a/nd4j/nd4j-shade/protobuf/pom.xml b/nd4j/nd4j-shade/protobuf/pom.xml
index 1cbd7d5a8..910003683 100644
--- a/nd4j/nd4j-shade/protobuf/pom.xml
+++ b/nd4j/nd4j-shade/protobuf/pom.xml
@@ -20,11 +20,26 @@
             <groupId>com.google.protobuf</groupId>
             <artifactId>protobuf-java</artifactId>
             <version>3.8.0</version>
+            <!-- Without optional=true, users will still be able to import this dependency and compile, yet will get a NoClassDefFoundError at runtime -->
+            <optional>true</optional>
         </dependency>
         <dependency>
             <groupId>com.google.protobuf</groupId>
             <artifactId>protobuf-java-util</artifactId>
             <version>3.8.0</version>
+            <optional>true</optional>
+            <exclusions>
+                <exclusion>
+                    <groupId>com.google.guava</groupId>
+                    <artifactId>guava</artifactId>
+                </exclusion>
+            </exclusions>
+        </dependency>
+        <dependency>
+            <groupId>com.google.guava</groupId>
+            <artifactId>guava</artifactId>
+            <version>26.0-android</version>
+            <optional>true</optional>
         </dependency>
     </dependencies>
 
@@ -150,6 +165,7 @@
                         <includes>
                             <include>com.google.protobuf:*</include>
                             <include>com.google.protobuf.*:*</include>
+                            <include>com.google.guava:*</include>
                         </includes>
                     </artifactSet>
 
@@ -159,6 +175,11 @@
                             <pattern>com.google.protobuf</pattern>
                             <shadedPattern>org.nd4j.shade.protobuf</shadedPattern>
                         </relocation>
+                        <!-- Guava dependencies -->
+                        <relocation>
+                            <pattern>com.google.common</pattern>
+                            <shadedPattern>org.nd4j.shade.protobuf.common</shadedPattern>
+                        </relocation>
                     </relocations>
 
                 </configuration>
diff --git a/nd4j/pom.xml b/nd4j/pom.xml
index 6c294d7e7..f043d7299 100644
--- a/nd4j/pom.xml
+++ b/nd4j/pom.xml
@@ -77,11 +77,6 @@
                 <artifactId>slf4j-api</artifactId>
                 <version>${slf4j.version}</version>
             </dependency>
-            <dependency>
-                <groupId>com.google.guava</groupId>
-                <artifactId>guava</artifactId>
-                <version>${guava.version}</version>
-            </dependency>
             <dependency>
                 <groupId>junit</groupId>
                 <artifactId>junit</artifactId>
diff --git a/nd4s/build.sbt b/nd4s/build.sbt
index 701483d0f..1fbac5ae6 100644
--- a/nd4s/build.sbt
+++ b/nd4s/build.sbt
@@ -38,7 +38,7 @@ lazy val commonSettings = Seq(
   resolvers in ThisBuild ++= Seq(Opts.resolver.sonatypeSnapshots),
   nd4jVersion := sys.props.getOrElse("nd4jVersion", default = "1.0.0-SNAPSHOT"),
   libraryDependencies ++= Seq(
-    "com.nativelibs4java" %% "scalaxy-loops" % "0.3.4",
+//    "com.nativelibs4java" %% "scalaxy-loops" % "0.3.4",
 //    "org.nd4j" % "nd4j-api" % nd4jVersion.value,
 //    "org.nd4j" % "nd4j-native-platform" % nd4jVersion.value % Test,
     "org.scalatest" %% "scalatest" % "2.2.6" % Test,
diff --git a/nd4s/pom.xml b/nd4s/pom.xml
index 011bc7fbe..7fc3fb7d3 100644
--- a/nd4s/pom.xml
+++ b/nd4s/pom.xml
@@ -68,11 +68,6 @@
     </properties>
 
     <dependencies>
-        <dependency>
-            <groupId>com.nativelibs4java</groupId>
-            <artifactId>scalaxy-loops_${scala.binary.version}</artifactId>
-            <version>0.3.4</version>
-        </dependency>
         <dependency>
             <groupId>org.nd4j</groupId>
             <artifactId>nd4j-api</artifactId>
@@ -84,6 +79,12 @@
             <version>${logback.version}</version>
             <scope>test</scope>
         </dependency>
+        <dependency>
+   	    <groupId>junit</groupId>	
+	    <artifactId>junit</artifactId>
+            <version>${junit.version}</version>
+            <scope>test</scope>
+	</dependency>
         <dependency>
             <groupId>org.scalatest</groupId>
             <artifactId>scalatest_${scala.binary.version}</artifactId>
@@ -99,7 +100,7 @@
         <dependency>
             <groupId>org.scalanlp</groupId>
             <artifactId>breeze_${scala.binary.version}</artifactId>
-            <version>0.12</version>
+            <version>${breeze.version}</version>
             <scope>test</scope>
         </dependency>
         <dependency>
@@ -187,6 +188,7 @@
                         <arg>-deprecation</arg>
                         <arg>-explaintypes</arg>
                         <arg>-nobootcp</arg>
+			            <arg>-usejavacp</arg>
                     </args>
                 </configuration>
             </plugin>
diff --git a/nd4s/src/main/scala/org/nd4s/CollectionLikeNDArray.scala b/nd4s/src/main/scala/org/nd4s/CollectionLikeNDArray.scala
index 46bba7fff..e6818fbc0 100644
--- a/nd4s/src/main/scala/org/nd4s/CollectionLikeNDArray.scala
+++ b/nd4s/src/main/scala/org/nd4s/CollectionLikeNDArray.scala
@@ -21,7 +21,6 @@ import org.nd4j.linalg.api.ops.Op
 import org.nd4j.linalg.factory.Nd4j
 import org.nd4s.ops.{ BitFilterOps, FilterOps, FunctionalOpExecutioner, MapOps }
 
-import scalaxy.loops._
 import scala.language.postfixOps
 import scala.util.control.Breaks._
 
@@ -65,7 +64,7 @@ trait CollectionLikeNDArray[A <: INDArray] {
     val lv = ev.linearView(underlying)
     breakable {
       for {
-        i <- 0 until lv.length().toInt optimized
+        i <- 0 until lv.length().toInt 
       } if (!f(ev.get(lv, i))) {
         result = true
         break()
@@ -81,7 +80,7 @@ trait CollectionLikeNDArray[A <: INDArray] {
     val lv = ev.linearView(underlying)
     breakable {
       for {
-        i <- 0 until lv.length().toInt optimized
+        i <- 0 until lv.length().toInt 
       } if (!f(ev.get(lv, i))) {
         result = false
         break()
diff --git a/pom.xml b/pom.xml
index ee24279d4..63054c104 100644
--- a/pom.xml
+++ b/pom.xml
@@ -261,15 +261,13 @@
 
         <oshi.version>3.4.2</oshi.version>
         <kafka.version>0.8.2.2</kafka.version>
-        <akka.version>2.3.16</akka.version>
-        <typesafe.config.version>1.3.0</typesafe.config.version> <!-- Play requires 1.3.0 -->
+        <typesafe.config.version>1.3.0</typesafe.config.version>
         <py4j.version>0.10.4</py4j.version>
         <jcommander.version>1.27</jcommander.version>
         <dropwizard.version>0.8.0</dropwizard.version>
         <jodatime.version>2.2</jodatime.version>
         <protonpack.version>1.15</protonpack.version>
         <poi.version>3.17</poi.version>
-        <play.version>2.4.8</play.version>
         <jodah.typetools.version>0.5.0</jodah.typetools.version>
         <freemarker.version>2.3.23</freemarker.version>
         <geoip2.version>2.8.1</geoip2.version>
@@ -280,6 +278,7 @@
         <fastutil.version>6.5.7</fastutil.version>
         <unirest.version>1.4.9</unirest.version>
         <reflections.version>0.9.10</reflections.version>
+        <breeze.version>1.0</breeze.version>
 
         <javacpp.parser.skip>false</javacpp.parser.skip>     <!-- To skip header file parsing phase: -Djavacpp.parser.skip=true  -->
         <javacpp.compiler.skip>false</javacpp.compiler.skip> <!-- To skip native compilation phase: -Djavacpp.compiler.skip=true -->
@@ -308,7 +307,7 @@
         <tensorflow.version>1.14.0</tensorflow.version>
         <tensorflow.javacpp.version>${tensorflow.version}-${javacpp-presets.version}</tensorflow.javacpp.version>
 
-        <commons-compress.version>1.16.1</commons-compress.version>
+        <commons-compress.version>1.18</commons-compress.version>
         <commonsmath.version>3.5</commonsmath.version>
         <commonslang.version>3.6</commonslang.version>
         <commonsio.version>2.5</commonsio.version>
@@ -325,13 +324,15 @@
         <commons-collections.version>3.2.2</commons-collections.version>
         <commons-collections4.version>4.1</commons-collections4.version>
 
+        <spark.version>2.4.3</spark.version>
+        <spark.major.version>2</spark.major.version>
         <args4j.version>2.0.29</args4j.version>
         <slf4j.version>1.7.21</slf4j.version>
         <junit.version>4.12</junit.version>
         <logback.version>1.2.3</logback.version>
-        <jackson.version>2.5.1</jackson.version>
-        <spark.jackson.version>${jackson.version}</spark.jackson.version>
-        <spark2.jackson.version>2.6.5</spark2.jackson.version>
+        <jackson.version>2.9.9</jackson.version>
+        <jackson.databind.version>2.9.9.3</jackson.databind.version>
+        <shaded.snakeyaml.version>1.23</shaded.snakeyaml.version>
         <geo.jackson.version>2.8.7</geo.jackson.version>
         <lombok.version>1.18.2</lombok.version>
         <cleartk.version>2.0.0</cleartk.version>
@@ -339,18 +340,16 @@
         <json.version>20131018</json.version>
         <google.protobuf.version>2.6.1</google.protobuf.version>
         <failIfNoTests>false</failIfNoTests>
-        <hadoop.version>2.2.0
-        </hadoop.version>  <!-- Hadoop version used by Spark 1.6.3 and 2.2.1 (and likely others) -->
-        <spark.version>2.1.0</spark.version>
+        <hadoop.version>2.2.0</hadoop.version>  <!-- Hadoop version used by Spark 1.6.3 and 2.2.1 (and likely others) -->
         <camel.version>2.16.3</camel.version>
         <zookeeper.version>3.4.6</zookeeper.version>
         <agrona.version>0.5.4</agrona.version>
         <mapdb.version>3.0.5</mapdb.version>
         <sqlite.version>3.15.1</sqlite.version>
-        <playframework.version>2.4.8
-        </playframework.version>    <!-- Play Framework 2.5 dropped Scala 2.10 support - use 2.4 for Spark compatibility w/ Scala 2.10 -->
+        <playframework.version>2.7.3</playframework.version>    <!-- 2.7.3 has Scala 2.11, 2.12 and 2.13 support only (no 2.10) -->
         <javax.ws.rs.version>2.0</javax.ws.rs.version>
-        <guava.version>20.0</guava.version>
+        <guava.version>28.0-jre</guava.version>
+        <gson.version>2.8.0</gson.version>
         <fbs.version>1.2.0-3f79e055</fbs.version>
         <threadly.version>4.10.0</threadly.version>
 
@@ -386,12 +385,12 @@
         <mockito.version>2.2.6</mockito.version>
 
         <!-- base versions -->
-        <!-- Scala 2.10.x -->
-        <scala210.version>2.10.7</scala210.version>
-        <scala210.binary.version>2.10</scala210.binary.version>
         <!-- Scala 2.11.x -->
         <scala211.version>2.11.12</scala211.version>
         <scala211.binary.version>2.11</scala211.binary.version>
+        <!-- Scala 2.12.x -->
+        <scala212.version>2.12.9</scala212.version>
+        <scala212.binary.version>2.12</scala212.binary.version>
 
         <scalatest.version>3.0.5</scalatest.version>
         <scalafmt.version>1.3.0</scalafmt.version>
diff --git a/rl4j/rl4j-core/pom.xml b/rl4j/rl4j-core/pom.xml
index 497128f30..67b050ba1 100644
--- a/rl4j/rl4j-core/pom.xml
+++ b/rl4j/rl4j-core/pom.xml
@@ -101,6 +101,12 @@
             <artifactId>jackson-databind</artifactId>
             <version>${jackson.version}</version>
         </dependency>
+
+        <dependency>
+            <groupId>com.google.code.gson</groupId>
+            <artifactId>gson</artifactId>
+            <version>${gson.version}</version>
+        </dependency>
     </dependencies>
 
     <profiles>
diff --git a/scalnet/pom.xml b/scalnet/pom.xml
index 5d0f85dd8..a6e220280 100644
--- a/scalnet/pom.xml
+++ b/scalnet/pom.xml
@@ -88,6 +88,12 @@
             <version>${scalacheck.version}</version>
             <scope>test</scope>
         </dependency>
+
+        <dependency>
+            <groupId>com.google.code.gson</groupId>
+            <artifactId>gson</artifactId>
+            <version>${gson.version}</version>
+        </dependency>
     </dependencies>
 
     <build>

From 130c9aa536a48a39589c6a1d61afb55fa4efefde Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Fri, 30 Aug 2019 16:36:14 +1000
Subject: [PATCH 45/56] Fixes (#202)

* Remove unnecessary dup in MKLDNNSubsamplingHelper

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* gson dep fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 .../nn/layers/mkldnn/MKLDNNSubsamplingHelper.java          | 3 ---
 nd4j/nd4j-remote/nd4j-json-server/pom.xml                  | 7 +++++++
 2 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/mkldnn/MKLDNNSubsamplingHelper.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/mkldnn/MKLDNNSubsamplingHelper.java
index ff4b52ac5..3edbf0b28 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/mkldnn/MKLDNNSubsamplingHelper.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/mkldnn/MKLDNNSubsamplingHelper.java
@@ -69,9 +69,6 @@ public class MKLDNNSubsamplingHelper implements SubsamplingHelper {
             pad = ConvolutionUtils.getSameModeTopLeftPadding(new int[]{(int)epsilon.size(2), (int)epsilon.size(3)}, new int[] {(int)input.size(2), (int)input.size(3)}, kernel, strides, dilation);
         }
 
-        input = input.dup();
-        epsilon = epsilon.dup();
-
         Pooling2DConfig conf = Pooling2DConfig.builder()
                 .isSameMode(convolutionMode == ConvolutionMode.Same)
                 .kH(kernel[0]).kW(kernel[1])
diff --git a/nd4j/nd4j-remote/nd4j-json-server/pom.xml b/nd4j/nd4j-remote/nd4j-json-server/pom.xml
index ba4d97bd5..47ef995a9 100644
--- a/nd4j/nd4j-remote/nd4j-json-server/pom.xml
+++ b/nd4j/nd4j-remote/nd4j-json-server/pom.xml
@@ -123,6 +123,13 @@
       <artifactId>activation</artifactId>
       <version>1.1</version>
     </dependency>
+
+    <dependency>
+      <groupId>com.google.code.gson</groupId>
+      <artifactId>gson</artifactId>
+      <version>${gson.version}</version>
+      <scope>test</scope>
+    </dependency>
   </dependencies>
 
   <build>

From 1003428a18fcd41ab57007d2949287832aa611a6 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Fri, 30 Aug 2019 10:12:40 +0300
Subject: [PATCH 46/56] [WIP] Int broadcastables (#195)

* Removed invalid resource and fixed tests

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* legacy scalar/pairwise/broadcast int ops

Signed-off-by: raver119 <raver119@gmail.com>

* NDArray int broadcastables

Signed-off-by: raver119 <raver119@gmail.com>

* few more bitwise tests

Signed-off-by: raver119 <raver119@gmail.com>

* java side update

Signed-off-by: raver119 <raver119@gmail.com>

* Argument type changed for shift ops

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>

* legacy scalar/pairwise/broadcast int ops

Signed-off-by: raver119 <raver119@gmail.com>

* NDArray int broadcastables

Signed-off-by: raver119 <raver119@gmail.com>

* few more bitwise tests

Signed-off-by: raver119 <raver119@gmail.com>

* java side update

Signed-off-by: raver119 <raver119@gmail.com>

* Argument type changed for shift ops

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
---
 .../models/fasttext/FastTextTest.java         |   4 +-
 libnd4j/blas/NDArray.h                        |  13 +
 libnd4j/blas/NDArray.hpp                      | 198 ++++++++
 libnd4j/blas/NativeOpExecutioner.h            |  58 +++
 libnd4j/blas/cpu/NativeOpExecutioner.cpp      | 155 ++++++
 libnd4j/blas/cuda/NativeOpExecutioner.cu      | 184 ++++++-
 .../loops/BroadcastPairwiseConverter.h        |  15 +-
 libnd4j/include/loops/broadcasting_int.h      | 164 +++++++
 .../include/loops/cpu/broadcasting_int.cpp    | 464 ++++++++++++++++++
 libnd4j/include/loops/cpu/pairwise_int.cpp    | 309 ++++++++++++
 libnd4j/include/loops/cpu/scalar_int.cpp      | 255 ++++++++++
 .../include/loops/cuda/broadcasting_int.cu    | 291 +++++++++++
 libnd4j/include/loops/cuda/pairwise_int.cu    | 173 +++++++
 libnd4j/include/loops/cuda/scalar_int.cu      | 269 ++++++++++
 libnd4j/include/loops/legacy_ops.h            |  27 +
 libnd4j/include/loops/pairwise_int.h          | 119 +++++
 libnd4j/include/loops/scalar_int.h            | 142 ++++++
 libnd4j/include/op_enums.h                    |  12 +
 libnd4j/include/ops/BroadcastIntOpsTuple.h    |  49 ++
 .../generic/bitwise/cyclic_rshift.cpp         |  20 +-
 .../generic/bitwise/cyclic_shift.cpp          |  20 +-
 .../ops/declarable/generic/bitwise/rshift.cpp |  20 +-
 .../ops/declarable/generic/bitwise/shift.cpp  |  20 +-
 .../include/ops/declarable/headers/bitwise.h  |   8 +-
 .../include/ops/impl/BroadcastIntOpsTuple.cpp |  27 +
 libnd4j/include/ops/ops.h                     |  92 ++++
 .../layers_tests/DeclarableOpsTests13.cpp     | 106 +++-
 .../DifferentialFunctionFactory.java          |   8 +-
 .../nd4j/autodiff/samediff/ops/SDMath.java    |   8 +-
 .../transforms/custom/CyclicRShiftBits.java   |  14 +-
 .../transforms/custom/CyclicShiftBits.java    |  14 +-
 .../impl/transforms/custom/RShiftBits.java    |  14 +-
 .../ops/impl/transforms/custom/ShiftBits.java |  14 +-
 .../java/org/nd4j/nativeblas/Nd4jCuda.java    |   1 +
 .../java/org/nd4j/nativeblas/Nd4jCpu.java     |  55 ++-
 .../scala/org/nd4s/samediff/SameDiff.scala    |   9 +-
 .../scala/org/nd4s/samediff/MathTest.scala    |  12 +
 37 files changed, 3237 insertions(+), 126 deletions(-)
 create mode 100644 libnd4j/include/loops/broadcasting_int.h
 create mode 100644 libnd4j/include/loops/cpu/broadcasting_int.cpp
 create mode 100644 libnd4j/include/loops/cpu/pairwise_int.cpp
 create mode 100644 libnd4j/include/loops/cpu/scalar_int.cpp
 create mode 100644 libnd4j/include/loops/cuda/broadcasting_int.cu
 create mode 100644 libnd4j/include/loops/cuda/pairwise_int.cu
 create mode 100644 libnd4j/include/loops/cuda/scalar_int.cu
 create mode 100644 libnd4j/include/loops/pairwise_int.h
 create mode 100644 libnd4j/include/loops/scalar_int.h
 create mode 100644 libnd4j/include/ops/BroadcastIntOpsTuple.h
 create mode 100644 libnd4j/include/ops/impl/BroadcastIntOpsTuple.cpp

diff --git a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/fasttext/FastTextTest.java b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/fasttext/FastTextTest.java
index b60af71d5..731a9cd60 100644
--- a/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/fasttext/FastTextTest.java
+++ b/deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/models/fasttext/FastTextTest.java
@@ -103,7 +103,7 @@ public class FastTextTest extends BaseDL4JTest {
     }
 
     @Test
-    public void testPredict() throws IOException {
+    public void testPredict() {
             String text = "I like soccer";
 
             FastText fastText = new FastText(supModelFile);
@@ -118,7 +118,7 @@ public class FastTextTest extends BaseDL4JTest {
     }
 
     @Test
-    public void testPredictProbability() throws IOException {
+    public void testPredictProbability() {
         String text = "I like soccer";
 
         FastText fastText = new FastText(supModelFile);
diff --git a/libnd4j/blas/NDArray.h b/libnd4j/blas/NDArray.h
index 3237e5033..3ef3716b3 100644
--- a/libnd4j/blas/NDArray.h
+++ b/libnd4j/blas/NDArray.h
@@ -34,6 +34,7 @@
 #include <op_enums.h>
 #include <ops/BroadcastOpsTuple.h>
 #include <ops/BroadcastBoolOpsTuple.h>
+#include <ops/BroadcastIntOpsTuple.h>
 #include <array/ExtraArguments.h>
 #include <Status.h>
 #include <ShapeDescriptor.h>
@@ -659,6 +660,8 @@ namespace nd4j {
 
         void applyPairwiseTransform(nd4j::pairwise::BoolOps op, const NDArray *other, NDArray *target, ExtraArguments *extraParams = nullptr) const;
 
+        void applyPairwiseTransform(nd4j::pairwise::IntOps op, const NDArray *other, NDArray *target, ExtraArguments *extraParams = nullptr) const;
+
         /**
         *  apply operation which requires broadcasting, broadcast a smaller array (tad) along  bigger one (this)
         *  tad - array to broadcast
@@ -672,6 +675,9 @@ namespace nd4j {
 
         void applyBroadcast(nd4j::broadcast::BoolOps op, const std::vector<int> &dimensions, const NDArray *tad, NDArray *target = nullptr, ExtraArguments *extraArgs = nullptr);
 
+        void applyBroadcast(nd4j::broadcast::IntOps op, const std::vector<int> &dimensions, const NDArray *tad, NDArray *target = nullptr, ExtraArguments *extraArgs = nullptr);
+
+
         /**
         *  apply operation which requires broadcasting, broadcast one tensor along another, also this method checks the possibility of broadcasting
         *  other - input array
@@ -692,6 +698,9 @@ namespace nd4j {
 
         void applyTrueBroadcast(nd4j::BroadcastBoolOpsTuple op, const NDArray* other, NDArray* target, const bool checkTargetShape = true, ExtraArguments *extraArgs = nullptr) const;
 
+        void applyTrueBroadcast(nd4j::BroadcastIntOpsTuple op, const NDArray* other, NDArray* target, const bool checkTargetShape = true, ExtraArguments *extraArgs = nullptr) const;
+
+
         /**
         *  apply a scalar operation to an array
         *  scalar - input scalar
@@ -704,6 +713,9 @@ namespace nd4j {
         template <typename T>
         void applyScalar(nd4j::scalar::BoolOps op, const T scalar, NDArray* target, ExtraArguments *extraParams = nullptr) const;
 
+        template <typename T>
+        void applyScalar(nd4j::scalar::IntOps op, const T scalar, NDArray* target, ExtraArguments *extraParams = nullptr) const;
+
         /**
         *  apply a scalar operation to an array
         *  scalar - input array which is simple scalar
@@ -714,6 +726,7 @@ namespace nd4j {
 
         void applyScalarArr(nd4j::scalar::BoolOps op, const NDArray* scalar, NDArray* target, ExtraArguments *extraParams = nullptr) const;
 
+        void applyScalarArr(nd4j::scalar::IntOps op, const NDArray* scalar, NDArray* target, ExtraArguments *extraParams = nullptr) const;
 
 #if defined(__CUDABLAS__) //&& defined(BUILD_TESTS)
         template <typename Lambda>
diff --git a/libnd4j/blas/NDArray.hpp b/libnd4j/blas/NDArray.hpp
index 72b029c0b..82427f9b9 100644
--- a/libnd4j/blas/NDArray.hpp
+++ b/libnd4j/blas/NDArray.hpp
@@ -2663,6 +2663,88 @@ void NDArray::applyTrueBroadcast(nd4j::BroadcastBoolOpsTuple op, const NDArray*
         delete pTarget;
 }
 
+
+
+//////////////////////////////////////////////////////////////////////////
+    void NDArray::applyTrueBroadcast(nd4j::BroadcastIntOpsTuple op, const NDArray* other, NDArray* target, const bool checkTargetShape, ExtraArguments *extraArgs) const {
+        if (isS())
+            throw std::runtime_error("NDArray::applyTrueBroadcast bool: you can't use this method on String array!");
+        if(target == nullptr || other == nullptr)
+            throw std::runtime_error("NDArray::applyTrueBroadcast int method: target or other = nullptr !");
+
+        if (isEmpty() || other->isEmpty())
+            return;
+
+        NDArray::prepareSpecialUse({target}, {this, other});
+
+        if (isScalar()) {
+            NDArray temp(target->_shapeInfo, dataType(), false, getContext());
+            temp.assign(this);
+            temp.applyPairwiseTransform(op.p, other, target,  extraArgs);
+            return;
+        }
+        if (other->isScalar()) {
+            this->applyScalarArr(op.s, other, target, extraArgs);
+            return;
+        }
+
+        const NDArray* min(other);
+        const NDArray* max(this);
+
+        if(this->rankOf() < other->rankOf()) {
+            max = other;
+            min = this;
+        }
+
+        if(checkTargetShape) {
+            Nd4jLong* newShapeInfo = nullptr;
+            if(!ShapeUtils::evalBroadcastShapeInfo(*max, *min, false, newShapeInfo, getContext()->getWorkspace()))          // the rank of target array must be equal to max->rankOf)()
+                throw std::runtime_error("NDArray::applyTrueBroadcast method: the shapes of this and other arrays are not suitable for broadcast operation !");
+            if(!shape::equalsSoft(target->_shapeInfo, newShapeInfo) || target->dataType() != this->dataType())
+                throw std::runtime_error("NDArray::applyTrueBroadcast int method: the shape or type of target array is wrong !");
+            if(dataType() != other->dataType())
+                throw std::invalid_argument("NDArray::applyTrueBroadcast int method: this and other arrays must have the same type !");
+        }
+
+        NDArray* pTarget = (max->dataType() == target->dataType()) ? target : new NDArray(target->ordering(), target->getShapeAsVector(), max->dataType(), target->getContext());
+        // check whether max array has to be tiled
+        if(!max->isSameShape(target)) {
+            // evaluate repeating dimensions for tile operation
+            std::vector<Nd4jLong> repeatMax(max->rankOf());
+            for(int i = 1; i <= max->rankOf(); ++i)
+                repeatMax[i-1] = (target->_shapeInfo[i] / max->_shapeInfo[i]);
+            max->tile(repeatMax, *pTarget);
+        }
+        else
+            pTarget->assign(max);
+
+        // check whether min array has to be tiled
+        std::vector<Nd4jLong> repeatMin(min->rankOf());
+        int product = 1;
+        for(int i = min->rankOf(); i >=1 ; --i) {
+            repeatMin[i-1] = (target->_shapeInfo[target->rankOf() - min->rankOf() + i] / min->_shapeInfo[i]);
+            product *= repeatMin[i-1];
+        }
+
+        auto pMin = const_cast<NDArray *>(min);
+        if(product != 1 )
+            pMin = new NDArray(min->tile(repeatMin));
+
+        std::vector<int> sameDims = ShapeUtils::getDimsWithSameShape(*target, *pMin);
+
+        if(max == this)
+            pTarget->applyBroadcast(op.b, sameDims, pMin, target, extraArgs);
+        else
+            pMin->applyBroadcast(op.b, sameDims, pTarget, target, extraArgs);
+
+        if(pMin != min)
+            delete pMin;
+        if(pTarget != target)
+            delete pTarget;
+    }
+
+
+
 //////////////////////////////////////////////////////////////////////////
 NDArray NDArray::applyTrueBroadcast(nd4j::BroadcastOpsTuple op, const NDArray& other, ExtraArguments *extraArgs) const {
     if (isEmpty() || other.isEmpty()) {
@@ -2801,6 +2883,67 @@ void NDArray::applyBroadcast(nd4j::broadcast::BoolOps op, const std::vector<int>
     registerSpecialUse({result}, {this, other});
 }
 
+
+//////////////////////////////////////////////////////////////////////////
+    void NDArray::applyBroadcast(nd4j::broadcast::IntOps op, const std::vector<int>& dimensions, const NDArray* other, NDArray* target, ExtraArguments* extraArgs) {
+        if (!isZ())
+            throw std::runtime_error("NDArray::applyBroadcast IntOps: you can't use this method on non-Integer array!");
+        if(isEmpty() || other->isEmpty()) {
+            if(!target->isEmpty())
+                throw std::runtime_error("NDArray::applyBroadcast IntOps: when some of input arrays (or both) is empty, target array must be empty as well !");
+            return;
+        }
+
+        if (dimensions.empty())
+            return;
+
+        auto result = target == nullptr ? this : target;
+
+        if (other->lengthOf() == lengthOf() && this->rankOf() == other->rankOf()) {
+            NDArray::prepareSpecialUse({result}, {this, other});
+            NativeOpExecutioner::execPairwiseIntTransform(getContext(), fromBroadcastToPairwiseInt(op), buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), other->getBuffer(), other->getShapeInfo(), other->getSpecialBuffer(), other->getSpecialShapeInfo(), result->buffer(), result->shapeInfo(), result->specialBuffer(), result->specialShapeInfo(), nullptr);
+            NDArray::registerSpecialUse({result}, {this, other});
+            return;
+        }
+
+        NDArray *min(nullptr), *max(nullptr);
+        if((lengthOf() > other->lengthOf()) || (lengthOf() == other->lengthOf() && rankOf() >= other->rankOf()))  {
+            max = this;
+            min = const_cast<NDArray*>(other);
+        }
+        else {
+            max = const_cast<NDArray*>(other);
+            min = this;
+        }
+
+        if(result->dataType() != dataType())
+            throw std::invalid_argument("NDArray::applyBroadcast int method: type of target array must be the same as input!");
+        if(!result->isSameShape(max))
+            throw std::invalid_argument("NDArray::applyBroadcast int method: max and target arrays must have the same shape !");
+        if(_dataType != other->_dataType)
+            throw std::invalid_argument("NDArray::applyBroadcast int method: this and other arrays must have the same type !");
+
+        std::vector<int> copy(dimensions);
+
+        if (dimensions.size() > 1)
+            std::sort(copy.begin(), copy.end());
+
+        Nd4jLong tadLength = shape::tadLength(max->shapeInfo(), copy.data(), (int) copy.size());
+        if (tadLength != min->lengthOf())
+            throw std::runtime_error("Tad length mismatch");
+
+        auto packX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(max->shapeInfo(), copy);
+        auto packZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(result->shapeInfo(), copy);
+
+        // TODO: eventually we want separate tads here
+        NDArray::prepareSpecialUse({result}, {this, other});
+        if(max == this)
+            NativeOpExecutioner::execBroadcastInt(       getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), other->getBuffer(), other->getShapeInfo(), other->getSpecialBuffer(), other->getSpecialShapeInfo(), result->buffer(), result->shapeInfo(), result->specialBuffer(), result->specialShapeInfo(), copy.data(), (int)copy.size(), packX.platformShapeInfo(), packX.platformOffsets(), packZ.platformShapeInfo(), packZ.platformOffsets());
+        else
+            NativeOpExecutioner::execInverseBroadcastInt(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), other->getBuffer(), other->getShapeInfo(), other->getSpecialBuffer(), other->getSpecialShapeInfo(), result->buffer(), result->shapeInfo(), result->specialBuffer(), result->specialShapeInfo(), copy.data(), (int)copy.size(), packX.platformShapeInfo(), packX.platformOffsets(), packZ.platformShapeInfo(), packZ.platformOffsets());
+        registerSpecialUse({result}, {this, other});
+    }
+
 //////////////////////////////////////////////////////////////////////////
 void NDArray::applyBroadcast(nd4j::broadcast::Ops op, const std::initializer_list<int> dimensions, const NDArray* tadArray, NDArray* target, ExtraArguments* extraArgs) {
     std::vector<int> vec(dimensions);
@@ -3043,6 +3186,22 @@ void NDArray::applyPairwiseTransform(nd4j::pairwise::BoolOps op, const NDArray *
     NDArray::registerSpecialUse({target}, {this, other});
 }
 
+////////////////////////////////////////////////////////////////////////
+    void NDArray::applyPairwiseTransform(nd4j::pairwise::IntOps op, const NDArray *other, NDArray *target, ExtraArguments *extraParams) const{
+        if (isS())
+            throw std::runtime_error("NDArray::applyPairwiseTransform IntOps: you can't use this method on String array!");
+        if (other->lengthOf() != target->lengthOf())
+            throw std::invalid_argument("NDArray::applyPairwiseTransform IntOps method - lengths of arrays are mismatched");
+        if (!target->isZ())
+            throw std::invalid_argument("NDArray::applyPairwiseTransform IntOps method - result must have bool type");
+        if (dataType() != other->dataType())
+            throw std::invalid_argument("NDArray::applyPairwiseTransform IntOps method - this and other arrays must have the same type !");
+
+        NDArray::prepareSpecialUse({target}, {this, other});
+        NativeOpExecutioner::execPairwiseIntTransform(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), other->getBuffer(), other->getShapeInfo(), other->getSpecialBuffer(), other->getSpecialShapeInfo(), target->buffer(), target->shapeInfo(), target->specialBuffer(), target->specialShapeInfo(), extraParams != nullptr ? extraParams->argumentsAsT(target->dataType()) : nullptr);
+        NDArray::registerSpecialUse({target}, {this, other});
+    }
+
 //////////////////////////////////////////////////////////////////////////
 void NDArray::applyPairwiseTransform(nd4j::pairwise::Ops op, const NDArray& other, ExtraArguments *extraParams) {
     applyPairwiseTransform(op, &other, this, extraParams);
@@ -3585,6 +3744,45 @@ template void NDArray::applyScalar<int8_t>(nd4j::scalar::BoolOps op, const int8_
 template void NDArray::applyScalar<uint8_t>(nd4j::scalar::BoolOps op, const uint8_t scalar, NDArray *target, ExtraArguments *extraParams) const;
 template void NDArray::applyScalar<bool>(nd4j::scalar::BoolOps op, const bool scalar, NDArray *target, ExtraArguments *extraParams) const;
 
+
+//////////////////////////////////////////////////////////////////////////
+    void NDArray::applyScalarArr(nd4j::scalar::IntOps op, const NDArray* scalar, NDArray *target, ExtraArguments *extraParams) const {
+        if (isS())
+            throw std::runtime_error("NDArray::applyScalarArr IntOps: you can't use this method on String array!");
+
+        if (target == nullptr || target->dataType() != this->dataType())
+            throw std::invalid_argument("NDArray::applyScalarArr int method: target is nullptr or has not bool type!");
+        if (dataType() != scalar->dataType()) {
+            nd4j_printf("NDArray::applyScalarArr IntOps: this dtype: [%i]; scalar dtype: [%i]\n", this->dataType(), scalar->dataType());
+            throw std::invalid_argument("NDArray::applyScalarArr int method: this and scalar arrays must have the same type!");
+        }
+
+        NDArray::prepareSpecialUse({target}, {this, scalar});
+        NativeOpExecutioner::execScalarInt(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), target->buffer(), target->shapeInfo(), target->specialBuffer(), target->specialShapeInfo(), scalar->getBuffer(), scalar->getShapeInfo(), scalar->getSpecialBuffer(), scalar->getSpecialShapeInfo(), extraParams != nullptr ? extraParams->argumentsAsT(target->dataType()): nullptr);
+        NDArray::registerSpecialUse({target}, {this, scalar});
+    }
+
+////////////////////////////////////////////////////////////////////////
+    template <typename T>
+    void NDArray::applyScalar(nd4j::scalar::IntOps op, const T scalar, NDArray *target, ExtraArguments *extraParams) const {
+
+        NDArray scalarArr = NDArrayFactory::create(this->dataType(), scalar, getContext());
+        applyScalarArr(op, &scalarArr, target, extraParams);
+    }
+
+    template <> void NDArray::applyScalar(nd4j::scalar::IntOps op, const NDArray* scalar, NDArray *target, ExtraArguments *extraParams) const { throw std::runtime_error("NDArray::applyScalar<NDArray*> method: do not use me!");}
+    template void NDArray::applyScalar<double>(nd4j::scalar::IntOps op, const double scalar, NDArray *target, ExtraArguments *extraParams) const;
+    template void NDArray::applyScalar<float>(nd4j::scalar::IntOps op, const float scalar, NDArray *target, ExtraArguments *extraParams) const;
+    template void NDArray::applyScalar<float16>(nd4j::scalar::IntOps op, const float16 scalar, NDArray *target, ExtraArguments *extraParams) const;
+    template void NDArray::applyScalar<bfloat16>(nd4j::scalar::IntOps op, const bfloat16 scalar, NDArray *target, ExtraArguments *extraParams) const;
+    template void NDArray::applyScalar<Nd4jLong>(nd4j::scalar::IntOps op, const Nd4jLong scalar, NDArray *target, ExtraArguments *extraParams) const;
+    template void NDArray::applyScalar<int>(nd4j::scalar::IntOps op, const int scalar, NDArray *target, ExtraArguments *extraParams) const;
+    template void NDArray::applyScalar<int16_t>(nd4j::scalar::IntOps op, const int16_t scalar, NDArray *target, ExtraArguments *extraParams) const;
+    template void NDArray::applyScalar<int8_t>(nd4j::scalar::IntOps op, const int8_t scalar, NDArray *target, ExtraArguments *extraParams) const;
+    template void NDArray::applyScalar<uint8_t>(nd4j::scalar::IntOps op, const uint8_t scalar, NDArray *target, ExtraArguments *extraParams) const;
+    template void NDArray::applyScalar<bool>(nd4j::scalar::IntOps op, const bool scalar, NDArray *target, ExtraArguments *extraParams) const;
+
+
 ////////////////////////////////////////////////////////////////////////
 void NDArray::applyIndexReduce(nd4j::indexreduce::Ops op, NDArray* target, const std::vector<int>& dimensions, const ExtraArguments *extraParams) const {
     if (isS())
diff --git a/libnd4j/blas/NativeOpExecutioner.h b/libnd4j/blas/NativeOpExecutioner.h
index 534453a0f..cae7a4e56 100644
--- a/libnd4j/blas/NativeOpExecutioner.h
+++ b/libnd4j/blas/NativeOpExecutioner.h
@@ -189,6 +189,16 @@ static void execScalarBool(nd4j::LaunchContext  *lc,
                             void *dScalar, Nd4jLong *dSscalarShapeInfo,
                             void *extraParams, bool allowParallelism = true);
 
+static void execScalarInt(nd4j::LaunchContext  *lc,
+                               int opNum,
+                               void *hX, Nd4jLong *hXShapeInfo,
+                               void *dX, Nd4jLong *dXShapeInfo,
+                               void *hZ, Nd4jLong *hZShapeInfo,
+                               void *dZ, Nd4jLong *dZShapeInfo,
+                               void *hScalar, Nd4jLong *hSscalarShapeInfo,
+                               void *dScalar, Nd4jLong *dSscalarShapeInfo,
+                               void *extraParams, bool allowParallelism = true);
+
  static void execScalar(nd4j::LaunchContext  *lc,
                             int opNum,
                             void *hX, Nd4jLong *hXShapeInfo,
@@ -215,6 +225,20 @@ static void execScalarBool(nd4j::LaunchContext  *lc,
                             Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
                             Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ);
 
+ static void execScalarInt(nd4j::LaunchContext  *lc,
+                               int opNum,
+                               void *hX, Nd4jLong *hXShapeInfo,
+                               void *dX, Nd4jLong *dXShapeInfo,
+                               void *extraParams,
+                               void *hZ, Nd4jLong *hZShapeInfo,
+                               void *dZ, Nd4jLong *dZShapeInfo,
+                               void *hScalars, Nd4jLong *hScalarShapeInfo,
+                               void *dScalars, Nd4jLong *dScalarShapeInfo,
+                               int *dimension, int dimensionLength,
+                               Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
+                               Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+
 /**
      *
      * @param opNum
@@ -275,6 +299,30 @@ static void execScalarBool(nd4j::LaunchContext  *lc,
                                      int *dimension, int dimensionLength,
                                      Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets,
                                      Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+    static void execBroadcastInt(nd4j::LaunchContext  *lc,
+                                  int opNum,
+                                  void *hX, Nd4jLong *hXShapeInfo,
+                                  void *dX, Nd4jLong *dXShapeInfo,
+                                  void *hY, Nd4jLong *hYShapeInfo,
+                                  void *dY, Nd4jLong *dYShapeInfo,
+                                  void *hZ, Nd4jLong *hZShapeInfo,
+                                  void *dZ, Nd4jLong *dZShapeInfo,
+                                  int *dimension, int dimensionLength,
+                                  Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets,
+                                  Nd4jLong *tadOnlyShapeInfoZ,Nd4jLong *tadOffsetsZ);
+
+    static void execInverseBroadcastInt(nd4j::LaunchContext  *lc,
+                                         int opNum,
+                                         void *x, Nd4jLong *xShapeInfo,
+                                         void *dX, Nd4jLong *dXShapeInfo,
+                                         void *y, Nd4jLong *yShapeInfo,
+                                         void *dY, Nd4jLong *dYShapeInfo,
+                                         void *result, Nd4jLong *resultShapeInfo,
+                                         void *dZ, Nd4jLong *dZShapeInfo,
+                                         int *dimension, int dimensionLength,
+                                         Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets,
+                                         Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ);
     
 /**
   *
@@ -308,6 +356,16 @@ static void execScalarBool(nd4j::LaunchContext  *lc,
                                     void *dZ, Nd4jLong *dZShapeInfo,
                                     void *extraParams);
 
+    static void execPairwiseIntTransform(nd4j::LaunchContext  *lc,
+                                          int opNum,
+                                          void *hX, Nd4jLong *hXShapeInfo,
+                                          void *dX, Nd4jLong *dXShapeInfo,
+                                          void *hY, Nd4jLong *hYShapeInfo,
+                                          void *dY, Nd4jLong *dYShapeInfo,
+                                          void *hZ, Nd4jLong *hZShapeInfo,
+                                          void *dZ, Nd4jLong *dZShapeInfo,
+                                          void *extraParams);
+
 /**
  *
  * @param opNum
diff --git a/libnd4j/blas/cpu/NativeOpExecutioner.cpp b/libnd4j/blas/cpu/NativeOpExecutioner.cpp
index b2ce7846a..22fd9eca4 100644
--- a/libnd4j/blas/cpu/NativeOpExecutioner.cpp
+++ b/libnd4j/blas/cpu/NativeOpExecutioner.cpp
@@ -24,6 +24,10 @@
 #include <broadcasting_bool.h>
 #include <scalar_bool.h>
 
+#include <pairwise_int.h>
+#include <broadcasting_int.h>
+#include <scalar_int.h>
+
 #include <loops/transform_float.h>
 #include <loops/transform_bool.h>
 #include <loops/transform_any.h>
@@ -242,6 +246,66 @@ void NativeOpExecutioner::execInverseBroadcastBool(nd4j::LaunchContext  *lc,
     BUILD_DOUBLE_SELECTOR(xType, zType, functions::broadcast::BroadcastBool, ::execInverse(opNum, hX, hXShapeInfo, hY, hYShapeInfo, hZ, hZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ), LIBND4J_TYPES, BOOL_TYPES);
 }
 
+
+
+////////////////////////////////////////////////////////////////////////
+void NativeOpExecutioner::execBroadcastInt(nd4j::LaunchContext  *lc,
+                                            int opNum,
+                                            void *hX, Nd4jLong *hXShapeInfo,
+                                            void *dX, Nd4jLong *dXShapeInfo,
+                                            void *hY, Nd4jLong *hYShapeInfo,
+                                            void *dY, Nd4jLong *dYShapeInfo,
+                                            void *hZ, Nd4jLong *hZShapeInfo,
+                                            void *dZ, Nd4jLong *dZShapeInfo,
+                                            int *dimension, int dimensionLength,
+                                            Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets,
+                                            Nd4jLong *tadOnlyShapeInfoZ,Nd4jLong *tadOffsetsZ) {
+#ifdef _OPENMP
+    omp_set_nested(1);
+#endif
+
+    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto yType = nd4j::ArrayOptions::dataType(hYShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+
+    if (xType != yType || xType != zType)
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execBroadcastInt", zType, xType, yType);
+
+    if (!nd4j::DataTypeUtils::isZ(zType))
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execBroadcastInt requires integer data type", zType);
+
+    BUILD_SINGLE_SELECTOR(xType, functions::broadcast::BroadcastInt, ::exec(opNum, hX, hXShapeInfo, hY, hYShapeInfo, hZ, hZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ), INTEGER_TYPES);
+}
+
+void NativeOpExecutioner::execInverseBroadcastInt(nd4j::LaunchContext  *lc,
+                                                   int opNum,
+                                                   void *hX, Nd4jLong *hXShapeInfo,
+                                                   void *dX, Nd4jLong *dXShapeInfo,
+                                                   void *hY, Nd4jLong *hYShapeInfo,
+                                                   void *dY, Nd4jLong *dYShapeInfo,
+                                                   void *hZ, Nd4jLong *hZShapeInfo,
+                                                   void *dZ, Nd4jLong *dZShapeInfo,
+                                                   int *dimension, int dimensionLength,
+                                                   Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets,
+                                                   Nd4jLong *tadOnlyShapeInfoZ,Nd4jLong *tadOffsetsZ) {
+
+#ifdef _OPENMP
+    omp_set_nested(1);
+#endif
+
+    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto yType = nd4j::ArrayOptions::dataType(hYShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+
+    if (xType != yType || xType != zType)
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execPairwiseIntTransform", zType, xType, yType);
+
+    if (!nd4j::DataTypeUtils::isZ(zType))
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execBroadcastInt requires integer data type", zType);
+
+    BUILD_SINGLE_SELECTOR(xType, functions::broadcast::BroadcastInt, ::execInverse(opNum, hX, hXShapeInfo, hY, hYShapeInfo, hZ, hZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ), INTEGER_TYPES);
+}
+
 ////////////////////////////////////////////////////////////////////////
 /**
 *
@@ -297,9 +361,42 @@ void NativeOpExecutioner::execPairwiseBoolTransform(nd4j::LaunchContext  *lc,
     auto yType = nd4j::ArrayOptions::dataType(hYShapeInfo);
     auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
 
+    if (xType != yType)
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execPairwiseBoolTransform", xType, yType);
+
+    if (zType != nd4j::DataType::BOOL)
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execPairwiseBoolTransform", nd4j::DataType::BOOL, zType);
+
     BUILD_DOUBLE_SELECTOR(xType, zType, functions::pairwise_transforms::PairWiseBoolTransform, ::exec(opNum, hX, hXShapeInfo, hY, hYShapeInfo, hZ, hZShapeInfo, extraParams), LIBND4J_TYPES, BOOL_TYPES);
 }
 
+////////////////////////////////////////////////////////////////////////
+void NativeOpExecutioner::execPairwiseIntTransform(nd4j::LaunchContext  *lc,
+                                                    int opNum,
+                                                    void *hX, Nd4jLong *hXShapeInfo,
+                                                    void *dX, Nd4jLong *dXShapeInfo,
+                                                    void *hY, Nd4jLong *hYShapeInfo,
+                                                    void *dY, Nd4jLong *dYShapeInfo,
+                                                    void *hZ, Nd4jLong *hZShapeInfo,
+                                                    void *dZ, Nd4jLong *dZShapeInfo,
+                                                    void *extraParams) {
+#ifdef _OPENMP
+    omp_set_nested(1);
+#endif
+
+    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto yType = nd4j::ArrayOptions::dataType(hYShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+
+    if (xType != yType || xType != zType)
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execPairwiseIntTransform", zType, xType, yType);
+
+    if (!nd4j::DataTypeUtils::isZ(zType))
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execSPairwiseInt requires integer data type", zType);
+
+    BUILD_SINGLE_SELECTOR(xType, functions::pairwise_transforms::PairWiseIntTransform, ::exec(opNum, hX, hXShapeInfo, hY, hYShapeInfo, hZ, hZShapeInfo, extraParams), INTEGER_TYPES);
+}
+
 ////////////////////////////////////////////////////////////////////////
 /**
 *
@@ -739,6 +836,64 @@ void NativeOpExecutioner::execScalarBool(nd4j::LaunchContext  *lc,
     BUILD_DOUBLE_SELECTOR(xType, zType, functions::scalar::ScalarBoolTransform, ::transform(opNum, hX, hXShapeInfo, extraParams, hZ, hZShapeInfo, hScalars, dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ, tadOffsetsZ), LIBND4J_TYPES, BOOL_TYPES);
 }
 
+////////////////////////////////////////////////////////////////////////
+void NativeOpExecutioner::execScalarInt(nd4j::LaunchContext  *lc,
+                                         int opNum,
+                                         void *hX, Nd4jLong *hXShapeInfo,
+                                         void *dX, Nd4jLong *dXShapeInfo,
+                                         void *hZ, Nd4jLong *hZShapeInfo,
+                                         void *dZ, Nd4jLong *dZShapeInfo,
+                                         void *hScalar, Nd4jLong *hSscalarShapeInfo,
+                                         void *dScalar, Nd4jLong *dSscalarShapeInfo,
+                                         void *extraParams, bool allowParallelism) {
+
+#ifdef _OPENMP
+    omp_set_nested(1);
+#endif
+
+    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto yType = nd4j::ArrayOptions::dataType(hSscalarShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+
+    if (xType != yType || xType != zType)
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execScalarInt", xType, yType);
+
+    if (!nd4j::DataTypeUtils::isZ(zType))
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execScalarInt", nd4j::DataType::INT32, zType);
+
+    BUILD_SINGLE_SELECTOR(xType, functions::scalar::ScalarIntTransform, ::transform(opNum, hX, hXShapeInfo, hZ, hZShapeInfo, hScalar, extraParams), INTEGER_TYPES);
+}
+
+////////////////////////////////////////////////////////////////////////
+void NativeOpExecutioner::execScalarInt(nd4j::LaunchContext  *lc,
+                                         int opNum,
+                                         void *hX, Nd4jLong *hXShapeInfo,
+                                         void *dX, Nd4jLong *dXShapeInfo,
+                                         void *extraParams,
+                                         void *hZ, Nd4jLong *hZShapeInfo,
+                                         void *dZ, Nd4jLong *dZShapeInfo,
+                                         void *hScalars, Nd4jLong *hScalarShapeInfo,
+                                         void *dScalars, Nd4jLong *dScalarShapeInfo,
+                                         int *dimension, int dimensionLength,
+                                         Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
+                                         Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+#ifdef _OPENMP
+    omp_set_nested(1);
+#endif
+
+    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto yType = nd4j::ArrayOptions::dataType(hScalarShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+
+    if (xType != yType || xType != zType)
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execScalarInt", xType, yType);
+
+    if (!nd4j::DataTypeUtils::isZ(zType))
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execScalarInt requires integer data type", zType);
+
+    BUILD_SINGLE_SELECTOR(xType, functions::scalar::ScalarIntTransform, ::transform(opNum, hX, hXShapeInfo, extraParams, hZ, hZShapeInfo, hScalars, dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ, tadOffsetsZ), INTEGER_TYPES);
+}
+
 ////////////////////////////////////////////////////////////////////////
 /**
 *
diff --git a/libnd4j/blas/cuda/NativeOpExecutioner.cu b/libnd4j/blas/cuda/NativeOpExecutioner.cu
index 8c4f1d3fa..fcb473820 100644
--- a/libnd4j/blas/cuda/NativeOpExecutioner.cu
+++ b/libnd4j/blas/cuda/NativeOpExecutioner.cu
@@ -38,19 +38,22 @@
 #include <loops/reduce_same.h>
 #include <loops/reduce_bool.h>
 #include <loops/reduce_long.h>
-#include <loops/broadcasting.h>
 #include <loops/indexreduce.h>
 #include <loops/pairwise_transform.h>
 #include <loops/pairwise_bool.h>
+#include <loops/pairwise_int.h>
 #include <loops/broadcasting_bool.h>
+#include <loops/broadcasting_int.h>
+#include <loops/broadcasting.h>
 #include <loops/reduce_float.h>
 #include <loops/reduce3.h>
 #include <loops/summarystatsreduce.h>
 #include <loops/transform_same.h>
-#include <loops/scalar.h>
 #include <loops/random.h>
 #include <loops/special_kernels.h>
+#include <loops/scalar.h>
 #include <loops/scalar_bool.h>
+#include <loops/scalar_int.h>
 
 using namespace nd4j;
 
@@ -152,6 +155,39 @@ void NativeOpExecutioner::execPairwiseBoolTransform( nd4j::LaunchContext  *lc,
         throw cuda_exception::build("execPairwiseBoolTransform failed", res);
 }
 
+////////////////////////////////////////////////////////////////////////
+void NativeOpExecutioner::execPairwiseIntTransform( nd4j::LaunchContext  *lc,
+                                                     int opNum,
+                                                     void *hX, Nd4jLong *hXShapeInfo,
+                                                     void *dX, Nd4jLong *dXShapeInfo,
+                                                     void *hY, Nd4jLong *hYShapeInfo,
+                                                     void *dY, Nd4jLong *dYShapeInfo,
+                                                     void *hZ, Nd4jLong *hZShapeInfo,
+                                                     void *dZ, Nd4jLong *dZShapeInfo,
+                                                     void *extraParams) {
+
+    auto stream = lc->getCudaStream();
+
+    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto yType = nd4j::ArrayOptions::dataType(hYShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+
+    if (!DataTypeUtils::isZ(zType))
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execPairwiseIntTransform wrong Z operand data type", nd4j::DataType::BOOL, zType);
+
+    if (yType != xType || zType != xType)
+        throw nd4j::datatype_exception::build("NativeOpExecutioner::execPairwiseIntTransform both operands must have same data type", xType, yType);
+
+    dim3 launchDims(256, 1024, 16384);
+
+    BUILD_SINGLE_SELECTOR(xType, functions::pairwise_transforms::PairWiseIntTransform, ::executeCudaShaped(launchDims, stream, opNum, dX, dXShapeInfo, dY, dYShapeInfo, dZ, dZShapeInfo, extraParams), INTEGER_TYPES)
+
+    // TODO: remove after the release
+    auto res = cudaStreamSynchronize(*stream);
+    if (res != 0)
+        throw cuda_exception::build("execPairwiseIntTransform failed", res);
+}
+
 ////////////////////////////////////////////////////////////////////////
 void NativeOpExecutioner::execSummaryStatsScalar(nd4j::LaunchContext  *lc,
                                     int opNum,
@@ -252,6 +288,81 @@ void NativeOpExecutioner::execInverseBroadcastBool(nd4j::LaunchContext  *lc,
         throw cuda_exception::build("execInverseBroadcastBool failed", res);
 }
 
+
+////////////////////////////////////////////////////////////////////////
+void NativeOpExecutioner::execBroadcastInt(nd4j::LaunchContext  *lc,
+                                            int opNum,
+                                            void *hX, Nd4jLong *hXShapeInfo,
+                                            void *dX, Nd4jLong *dXShapeInfo,
+                                            void *hY, Nd4jLong *hYShapeInfo,
+                                            void *dY, Nd4jLong *dYShapeInfo,
+                                            void *hZ, Nd4jLong *hZShapeInfo,
+                                            void *dZ, Nd4jLong *dZShapeInfo,
+                                            int *dimension, int dimensionLength,
+                                            Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets,
+                                            Nd4jLong *tadOnlyShapeInfoZ,Nd4jLong *tadOffsetsZ) {
+
+    auto stream = lc->getCudaStream();
+
+    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto yType = nd4j::ArrayOptions::dataType(hYShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+
+    if (!DataTypeUtils::isZ(zType))
+        throw std::runtime_error("NativeOpExecutioner::execBroadcastInt requires Z operand to have INT type");
+
+    if (yType != xType || zType != xType)
+        throw std::runtime_error("NativeOpExecutioner::execBroadcastInt requires both X & Y operands to have same type");
+
+    if (nd4j::Environment::getInstance()->isDebugAndVerbose())
+        printf("F3B opNum:[%i]\n", opNum);
+
+    dim3 launchDims(256, 256, 1024);
+
+    BUILD_SINGLE_SELECTOR(xType, functions::broadcast::BroadcastInt, ::execBroadcast(launchDims, stream, opNum, dX, dXShapeInfo, dY, dYShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ), INTEGER_TYPES)
+
+    // TODO: remove after the release
+    auto res = cudaStreamSynchronize(*stream);
+    if (res != 0)
+        throw cuda_exception::build("execBroadcastBool failed", res);
+}
+
+void NativeOpExecutioner::execInverseBroadcastInt(nd4j::LaunchContext  *lc,
+                                                   int opNum,
+                                                   void *hX, Nd4jLong *hXShapeInfo,
+                                                   void *dX, Nd4jLong *dXShapeInfo,
+                                                   void *hY, Nd4jLong *hYShapeInfo,
+                                                   void *dY, Nd4jLong *dYShapeInfo,
+                                                   void *hZ, Nd4jLong *hZShapeInfo,
+                                                   void *dZ, Nd4jLong *dZShapeInfo,
+                                                   int *dimension, int dimensionLength,
+                                                   Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets,
+                                                   Nd4jLong *tadOnlyShapeInfoZ,Nd4jLong *tadOffsetsZ) {
+    auto stream = lc->getCudaStream();
+
+    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto yType = nd4j::ArrayOptions::dataType(hYShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+
+    if (!DataTypeUtils::isZ(zType))
+        throw std::runtime_error("NativeOpExecutioner::execBroadcastInt requires Z operand to have INT type");
+
+    if (yType != xType || zType != xType)
+        throw std::runtime_error("NativeOpExecutioner::execBroadcastInt requires both X & Y operands to have same type");
+
+    if (nd4j::Environment::getInstance()->isDebugAndVerbose())
+        printf("F3BI opNum:[%i]\n", opNum);
+
+    dim3 launchDims(256, 256, 1024);
+
+    BUILD_SINGLE_SELECTOR(xType, functions::broadcast::BroadcastInt, ::execInverseBroadcast(launchDims, stream, opNum, dX, dXShapeInfo, dY, dYShapeInfo, dZ, dZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ), INTEGER_TYPES)
+
+    // TODO: remove after the release
+    auto res = cudaStreamSynchronize(*stream);
+    if (res != 0)
+        throw cuda_exception::build("execInverseBroadcastInt failed", res);
+}
+
 ////////////////////////////////////////////////////////////////////////
 /**
  *
@@ -1114,6 +1225,75 @@ void NativeOpExecutioner::execScalarBool(nd4j::LaunchContext  *lc,
         throw cuda_exception::build("execScalarBool B failed", res);
 }
 
+////////////////////////////////////////////////////////////////////////
+void NativeOpExecutioner::execScalarInt(nd4j::LaunchContext  *lc,
+                                         int opNum,
+                                         void *hX, Nd4jLong *hXShapeInfo,
+                                         void *dX, Nd4jLong *dXShapeInfo,
+                                         void *hZ, Nd4jLong *hZShapeInfo,
+                                         void *dZ, Nd4jLong *dZShapeInfo,
+                                         void *hScalar, Nd4jLong *hScalarShapeInfo,
+                                         void *dScalar, Nd4jLong *dScalarShapeInfo,
+                                         void *extraParams, bool allowParallelism) {
+
+    auto stream = lc->getCudaStream();
+
+    dim3 launchDims = dim3(256, 512, 8192);
+
+    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto yType = nd4j::ArrayOptions::dataType(hScalarShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+
+    if (xType != yType || zType != xType)
+        throw std::runtime_error("NativeOpExecutioner::execScalarInt requires X & Y to have same type");
+
+    if (!DataTypeUtils::isZ(zType) )
+        throw std::runtime_error("NativeOpExecutioner::execScalarInt requires Z operand to have INT type");
+
+    BUILD_SINGLE_SELECTOR(xType, functions::scalar::ScalarIntTransform, ::executeCudaShaped(launchDims, stream, opNum, dX, dXShapeInfo, dZ, dZShapeInfo, dScalar, extraParams), INTEGER_TYPES);
+
+    // TODO: remove after the release
+    auto res = cudaStreamSynchronize(*stream);
+    if (res != 0)
+        throw cuda_exception::build("execScalarInt failed", res);
+}
+
+////////////////////////////////////////////////////////////////////////
+void NativeOpExecutioner::execScalarInt(nd4j::LaunchContext  *lc,
+                                         int opNum,
+                                         void *hX, Nd4jLong *hXShapeInfo,
+                                         void *dX, Nd4jLong *dXShapeInfo,
+                                         void *extraParams,
+                                         void *hZ, Nd4jLong *hZShapeInfo,
+                                         void *dZ, Nd4jLong *dZShapeInfo,
+                                         void *hScalars, Nd4jLong *hScalarShapeInfo,
+                                         void *dScalars, Nd4jLong *dScalarShapeInfo,
+                                         int *dimension, int dimensionLength,
+                                         Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
+                                         Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+
+    auto stream = lc->getCudaStream();
+
+    dim3 launchDims(256, 512, 8192);
+
+    auto xType = nd4j::ArrayOptions::dataType(hXShapeInfo);
+    auto yType = nd4j::ArrayOptions::dataType(hScalarShapeInfo);
+    auto zType = nd4j::ArrayOptions::dataType(hZShapeInfo);
+
+    if (xType != yType || zType != xType)
+        throw std::runtime_error("NativeOpExecutioner::execScalarInt requires X & Y to have same type");
+
+    if (!DataTypeUtils::isZ(zType) )
+        throw std::runtime_error("NativeOpExecutioner::execScalarInt requires Z operand to have INT type");
+
+    BUILD_SINGLE_SELECTOR(xType, functions::scalar::ScalarIntTransform, ::executeCudaAlongDimension(launchDims, stream, opNum, dX, dXShapeInfo, dZ, dZShapeInfo, dScalars, extraParams, dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ, tadOffsetsZ), INTEGER_TYPES);
+
+    // TODO: remove after the release
+    auto res = cudaStreamSynchronize(*stream);
+    if (res != 0)
+        throw cuda_exception::build("execScalarInt B failed", res);
+}
+
 ////////////////////////////////////////////////////////////////////////
 void NativeOpExecutioner::execScalar(nd4j::LaunchContext  *lc,
 									int opNum,
diff --git a/libnd4j/include/loops/BroadcastPairwiseConverter.h b/libnd4j/include/loops/BroadcastPairwiseConverter.h
index 4b11cfc0d..fb5acf19b 100644
--- a/libnd4j/include/loops/BroadcastPairwiseConverter.h
+++ b/libnd4j/include/loops/BroadcastPairwiseConverter.h
@@ -71,12 +71,25 @@ inline pairwise::BoolOps fromBroadcastToPairwiseBool(broadcast::BoolOps op) {
         case broadcast::And: return pairwise::And;
         case broadcast::Or: return pairwise::Or;
         case broadcast::Xor: return pairwise::Xor;
-        case broadcast::Not: return pairwise::Not;        
+        case broadcast::Not: return pairwise::Not;
         default:
             throw std::runtime_error("fromBroadcastToPairwiseBool: Not convertible operation");
     }
 }
 
+    inline pairwise::IntOps fromBroadcastToPairwiseInt(broadcast::IntOps op) {
+        switch (op) {
+            case broadcast::IntOps::IntAnd: return pairwise::IntOps::IntAnd;
+            case broadcast::IntOps::IntOr: return pairwise::IntOps::IntOr;
+            case broadcast::IntOps::IntXor: return pairwise::IntOps::IntXor;
+            case broadcast::IntOps::ShiftLeft: return pairwise::IntOps::ShiftLeft;
+            case broadcast::IntOps::ShiftRight: return pairwise::IntOps::ShiftRight;
+            case broadcast::IntOps::CyclicShiftLeft: return pairwise::IntOps::CyclicShiftLeft;
+            case broadcast::IntOps::CyclicShiftRight: return pairwise::IntOps::CyclicShiftRight;
+            default:
+                throw std::runtime_error("fromBroadcastToPairwiseInt: Not convertible operation");
+        }
+    }
 }
 
 #endif //DEV_TESTS_BROADCASTPAIRWISECONVERTER_H
\ No newline at end of file
diff --git a/libnd4j/include/loops/broadcasting_int.h b/libnd4j/include/loops/broadcasting_int.h
new file mode 100644
index 000000000..84bc0f949
--- /dev/null
+++ b/libnd4j/include/loops/broadcasting_int.h
@@ -0,0 +1,164 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+/*
+ * broadcasting.h
+ *
+ *  Created on: Dec 28, 2015
+ *      Author: agibsonccc
+ */
+
+#ifndef BROADCASTING_INT_H_
+#define BROADCASTING_INT_H_
+#include <dll.h>
+#include <helpers/shape.h>
+#include <templatemath.h>
+#include <pairwise_util.h>
+#include <ops/ops.h>
+#include <op_boilerplate.h>
+#include <helpers/DebugHelper.h>
+
+#ifdef __CUDACC__
+#include <cuda.h>
+#include <cuda_runtime.h>
+#endif
+#ifdef __JNI__
+#include <jni.h>
+#endif
+
+#include <helpers/TAD.h>
+
+#include "legacy_ops.h"
+
+namespace functions {
+    namespace broadcast {
+
+/**
+ * Broadcast operation
+ * for broadcasting a smaller tensor
+ * along long a bigger one.
+ */
+        template<typename X>
+        class BroadcastInt {
+        public:
+
+#ifdef __CUDACC__
+
+            template<typename OpType>
+			static __device__ void transformCuda(
+			void *x,
+			Nd4jLong *xShapeInfo,
+			void *y,
+			Nd4jLong *yShapeInfo,
+			void *result,
+			Nd4jLong *resultShapeInfo,
+			int *dimension,
+			int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+            template <typename OpClass>
+            static __host__ void intermediateBroadcast(dim3 launchDims, cudaStream_t *stream, void *x, Nd4jLong *xShapeInfo, void *y, Nd4jLong *yShapeInfo, void *result, Nd4jLong *resultShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+            static __host__ void execBroadcast(dim3 launchDims, cudaStream_t *stream, int opNum, void *x, Nd4jLong *xShapeInfo, void *y, Nd4jLong *yShapeInfo, void *result, Nd4jLong *resultShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+            template<typename OpType>
+			static __device__ void transformInverseCuda(
+			void *x,
+			Nd4jLong *xShapeInfo,
+			void *y,
+			Nd4jLong *yShapeInfo,
+			void *result,
+			Nd4jLong *resultShapeInfo,
+			int *dimension,
+			int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+            template <typename OpClass>
+            static __host__ void intermediateInverseBroadcast(dim3 launchDims, cudaStream_t *stream, void *x, Nd4jLong *xShapeInfo, void *y, Nd4jLong *yShapeInfo, void *result, Nd4jLong *resultShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+            static __host__ void execInverseBroadcast(dim3 launchDims, cudaStream_t *stream, int opNum, void *x, Nd4jLong *xShapeInfo, void *y, Nd4jLong *yShapeInfo, void *result, Nd4jLong *resultShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+#endif
+
+            static void exec(int opNum,
+                             void *x,
+                             Nd4jLong *xShapeInfo,
+                             void *y,
+                             Nd4jLong *yShapeInfo,
+                             void *result,
+                             Nd4jLong *resultShapeInfo,
+                             int *dimension,
+                             int dimensionLength,
+                             Nd4jLong *tadShapeInfo,
+                             Nd4jLong *tadOffset,
+                             Nd4jLong *tadShapeInfoZ,
+                             Nd4jLong *tadOffsetZ);
+
+            static void execInverse(int opNum,
+                             void *x,
+                             Nd4jLong *xShapeInfo,
+                             void *y,
+                             Nd4jLong *yShapeInfo,
+                             void *result,
+                             Nd4jLong *resultShapeInfo,
+                             int *dimension,
+                             int dimensionLength,
+                             Nd4jLong *tadShapeInfo,
+                             Nd4jLong *tadOffset,
+                             Nd4jLong *tadShapeInfoZ,
+                             Nd4jLong *tadOffsetZ);
+
+            /**
+             * CPU execution
+             * @param x the input
+             * @param xShapeInfo the x shape information
+             * @param y the y data
+             * @param yShapeInfo the y shape information
+             * @param result the result
+             * @param resultShapeInfo the result shape information
+             * @param dimension the dimension to broadcast along long
+             * @param dimensionLength the length of the dimension buffer
+             */
+            template<typename OpType>
+            static void exec(void *x,
+                             Nd4jLong *xShapeInfo,
+                             void *y,
+                             Nd4jLong *yShapeInfo,
+                             void *result,
+                             Nd4jLong *resultShapeInfo,
+                             int *dimension,
+                             int dimensionLength,
+                             Nd4jLong *tadShapeInfo,
+                             Nd4jLong *tadOffset,
+                             Nd4jLong *tadShapeInfoZ,
+                             Nd4jLong *tadOffsetZ);
+
+            template<typename OpType>
+            static void execInverse(void *x,
+                             Nd4jLong *xShapeInfo,
+                             void *y,
+                             Nd4jLong *yShapeInfo,
+                             void *result,
+                             Nd4jLong *resultShapeInfo,
+                             int *dimension,
+                             int dimensionLength,
+                             Nd4jLong *tadShapeInfo,
+                             Nd4jLong *tadOffset,
+                             Nd4jLong *tadShapeInfoZ,
+                             Nd4jLong *tadOffsetZ);
+        };
+    }
+}
+
+#endif /* BROADCASTING_H_ */
diff --git a/libnd4j/include/loops/cpu/broadcasting_int.cpp b/libnd4j/include/loops/cpu/broadcasting_int.cpp
new file mode 100644
index 000000000..c092da50b
--- /dev/null
+++ b/libnd4j/include/loops/cpu/broadcasting_int.cpp
@@ -0,0 +1,464 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+//  @author raver119@gmail.com
+//
+
+#include <op_boilerplate.h>
+#include <loops/broadcasting_int.h>
+#include <loops/legacy_ops.h>
+#include <types/types.h>
+#include <LoopKind.h>
+#include <helpers/ConstantTadHelper.h>
+
+using namespace simdOps;
+
+namespace functions {
+    namespace broadcast {
+
+        template <typename X>
+        void BroadcastInt<X>::exec(const int opNum,
+                             void *x,
+                             Nd4jLong *xShapeInfo,
+                             void *y,
+                             Nd4jLong *yShapeInfo,
+                             void *z,
+                             Nd4jLong *zShapeInfo,
+                             int *dimension,
+                             int dimensionLength,
+                             Nd4jLong *xTadShapeInfo,
+                             Nd4jLong *xTadOffset,
+                             Nd4jLong *zTadShapeInfo,
+                             Nd4jLong *zTadOffset) {
+            DISPATCH_BY_OPNUM_T(exec, PARAMS(x,
+                                               xShapeInfo,
+                                               y,
+                                               yShapeInfo,
+                                               z,
+                                               zShapeInfo,
+                                               dimension,
+                                               dimensionLength,
+                                               xTadShapeInfo,
+                                               xTadOffset,
+                                               zTadShapeInfo,
+                                               zTadOffset), BROADCAST_INT_OPS);
+        }
+
+        template <typename X>
+        void BroadcastInt<X>::execInverse(const int opNum,
+                             void *x,
+                             Nd4jLong *xShapeInfo,
+                             void *y,
+                             Nd4jLong *yShapeInfo,
+                             void *z,
+                             Nd4jLong *zShapeInfo,
+                             int *dimension,
+                             int dimensionLength,
+                             Nd4jLong *xTadShapeInfo,
+                             Nd4jLong *xTadOffset,
+                             Nd4jLong *zTadShapeInfo,
+                             Nd4jLong *zTadOffset) {
+            DISPATCH_BY_OPNUM_T(execInverse, PARAMS(x,
+                                               xShapeInfo,
+                                               y,
+                                               yShapeInfo,
+                                               z,
+                                               zShapeInfo,
+                                               dimension,
+                                               dimensionLength,
+                                               xTadShapeInfo,
+                                               xTadOffset,
+                                               zTadShapeInfo,
+                                               zTadOffset), BROADCAST_INT_OPS);
+        }
+
+        template <typename X>
+        template<typename OpType>
+        void BroadcastInt<X>::exec(void *vx,
+                             Nd4jLong *xShapeInfo,
+                             void *vy,
+                             Nd4jLong *yShapeInfo,
+                             void *vz,
+                             Nd4jLong *zShapeInfo,
+                             int *dimension,
+                             int dimensionLength,
+                             Nd4jLong *xTadShapeInfo,
+                             Nd4jLong *xTadOffset,
+                             Nd4jLong *zTadShapeInfo,
+                             Nd4jLong *zTadOffset) {
+
+                auto x = reinterpret_cast<X *>(vx);
+                auto y = reinterpret_cast<X *>(vy);
+                auto z = reinterpret_cast<X *>(vz);
+
+                //decompose in to several sub tads after
+                //moving all dimensions (in sorted order)
+                //to the back.
+                //permuted version of the x shape info for setting up the tad problem
+                auto xTadShapeShapeInfo = xTadShapeInfo;
+                auto tadOffsets = xTadOffset;
+
+                if (xTadShapeInfo == nullptr || tadOffsets == nullptr) {
+                    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(xShapeInfo, dimension, dimensionLength);
+
+                    xTadShapeShapeInfo = tadPack.primaryShapeInfo();
+                    tadOffsets = tadPack.primaryOffsets();
+                }
+
+                //int *resultStride = shape::stride(xTadShapeShapeInfo);
+                unsigned int tadLength = shape::length(xTadShapeShapeInfo);//shape::length(xTadShapeShapeInfo);
+                unsigned int tads = shape::length(xShapeInfo) / tadLength;
+
+                if (zTadShapeInfo == nullptr) {
+                    zTadShapeInfo = xTadShapeShapeInfo;
+                    zTadOffset = tadOffsets;
+                }                                
+
+                auto lenZ = shape::length(zTadShapeInfo);
+                auto lenY = shape::length(yShapeInfo);
+
+                int tadsPerThread = tads / TAD_THRESHOLD;
+                int threads = nd4j::math::nd4j_max<int>(1, tadsPerThread);
+                threads = nd4j::math::nd4j_min<int>(threads, omp_get_max_threads());
+
+                auto xEws = shape::elementWiseStride(xTadShapeShapeInfo);
+                auto yEws = shape::elementWiseStride(yShapeInfo);
+                auto zEws = shape::elementWiseStride(zTadShapeInfo);
+
+                const nd4j::LoopKind::Kind kindOfLoop = nd4j::LoopKind::deduceKindOfLoopXYZ(xTadShapeShapeInfo, yShapeInfo, zTadShapeInfo);
+                
+                if (kindOfLoop == nd4j::LoopKind::EWS1) {
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+                        auto oX = x + tadOffsets[i];
+                           auto oZ = z + zTadOffset[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (unsigned int f = 0; f < tadLength; f++)
+                            oZ[f] = OpType::op(oX[f], y[f]);
+                    }
+                }
+                else if(kindOfLoop == nd4j::LoopKind::EWSNONZERO) {
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+                        auto oX = x + tadOffsets[i];
+                        auto oZ = z + zTadOffset[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (unsigned int f = 0; f < tadLength; f++)
+                            oZ[f * zEws] = OpType::op(oX[f * xEws], y[f * yEws]);
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(xTadShapeShapeInfo, yShapeInfo) && shape::haveSameShapeAndStrides(xTadShapeShapeInfo, zTadShapeInfo)) {
+
+                    uint tadShapeShapeInfoCast[MAX_RANK];
+                    bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+                    
+                        auto oZ = z + zTadOffset[i];
+                        auto oX = x + tadOffsets[i];
+                                        
+                        // TODO: cover this codebranch with tests
+                        // all this stuff already happens within thread
+                        PRAGMA_OMP_SIMD
+                        for (int f = 0; f < tadLength; f++) {
+                            auto offset = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastX);
+                            oZ[offset] = OpType::op(oX[offset], y[offset]);
+                        }
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(xTadShapeShapeInfo, yShapeInfo)) {
+
+                    uint tadShapeShapeInfoCast[MAX_RANK];
+                    uint tadShapeInfoZCast[MAX_RANK];
+                    bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
+                    bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
+
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+                    
+                        auto oZ = z + zTadOffset[i];
+                        auto oX = x + tadOffsets[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (int f = 0; f < tadLength; f++) {
+                            auto offset  = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastX);
+                            auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, lenZ, canCastZ);
+                            oZ[zOffset] = OpType::op(oX[offset], y[offset]);
+                        }
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(xTadShapeShapeInfo, zTadShapeInfo)) {
+
+                    uint tadShapeShapeInfoCast[MAX_RANK];
+                    uint yShapeInfoCast[MAX_RANK];
+                    bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
+                    bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+                    
+                        auto oZ = z + zTadOffset[i];
+                        auto oX = x + tadOffsets[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (int f = 0; f < tadLength; f++) {
+                            auto offset  = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastX);
+                            auto yOffset = shape::indexOffset(f, yShapeInfo, yShapeInfoCast, lenY, canCastY);
+                            oZ[offset] = OpType::op(oX[offset], y[yOffset]);
+                        }
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(yShapeInfo, zTadShapeInfo)) {
+
+                    uint tadShapeShapeInfoCast[MAX_RANK];
+                    uint yShapeInfoCast[MAX_RANK];
+                    bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
+                    bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+                    
+                        auto oZ = z + zTadOffset[i];
+                        auto oX = x + tadOffsets[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (int f = 0; f < tadLength; f++) {
+                            auto xOffset  = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastX);
+                            auto offset = shape::indexOffset(f, yShapeInfo, yShapeInfoCast, lenY, canCastY);
+                            oZ[offset] = OpType::op(oX[xOffset], y[offset]);
+                        }
+                    }
+                }
+                else {
+
+                    uint tadShapeShapeInfoCast[MAX_RANK];
+                    uint tadShapeInfoZCast[MAX_RANK];
+                    uint yShapeInfoCast[MAX_RANK];
+                    bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xTadShapeShapeInfo, tadShapeShapeInfoCast);
+                    bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
+                    bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
+
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+                    
+                        auto oZ = z + zTadOffset[i];
+                        auto oX = x + tadOffsets[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (int f = 0; f < tadLength; f++) {
+                            auto xOffset  = shape::indexOffset(f, xTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastX);
+                            auto yOffset = shape::indexOffset(f, yShapeInfo, yShapeInfoCast, lenY, canCastY);
+                            auto zOffset  = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, lenZ, canCastZ);
+                            oZ[zOffset] = OpType::op(oX[xOffset], y[yOffset]);
+                        }
+                    }
+                }
+        }
+
+
+        template <typename X>
+        template<typename OpType>
+        void BroadcastInt<X>::execInverse(void *vx,
+                             Nd4jLong *xShapeInfo,
+                             void *vy,
+                             Nd4jLong *yShapeInfo,
+                             void *vz,
+                             Nd4jLong *zShapeInfo,
+                             int *dimension,
+                             int dimensionLength,
+                             Nd4jLong *yTadShapeInfo,
+                             Nd4jLong *yTadOffset,
+                             Nd4jLong *zTadShapeInfo,
+                             Nd4jLong *zTadOffset) {
+
+                auto x = reinterpret_cast<X *>(vx);
+                auto y = reinterpret_cast<X *>(vy);
+                auto z = reinterpret_cast<X *>(vz);
+
+                //decompose in to several sub tads after
+                //moving all dimensions (in sorted order)
+                //to the back.
+                //permuted version of the x shape info for setting up the tad problem
+                auto yTadShapeShapeInfo = yTadShapeInfo;
+                auto tadOffsets = yTadOffset;
+
+                if (yTadShapeInfo == nullptr || tadOffsets == nullptr) {
+                    auto tadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(yShapeInfo, dimension, dimensionLength);
+
+                    yTadShapeShapeInfo = tadPack.primaryShapeInfo();
+                    tadOffsets = tadPack.primaryOffsets();
+                }
+
+                //int *resultStride = shape::stride(yTadShapeShapeInfo);
+                unsigned int tadLength = shape::length(yTadShapeShapeInfo);
+                unsigned int tads = shape::length(yShapeInfo) / tadLength;
+
+                if (zTadShapeInfo == nullptr) {
+                    zTadShapeInfo = yTadShapeShapeInfo;
+                    zTadOffset = tadOffsets;
+                }
+
+                auto lenZ = shape::length(zTadShapeInfo);
+                auto lenX = shape::length(xShapeInfo);
+
+                int tadsPerThread = tads / TAD_THRESHOLD;
+                int threads = nd4j::math::nd4j_max<int>(1, tadsPerThread);
+                threads = nd4j::math::nd4j_min<int>(threads, omp_get_max_threads());
+
+                auto yEws = shape::elementWiseStride(yTadShapeShapeInfo);
+                auto xEws = shape::elementWiseStride(xShapeInfo);
+                auto zEws = shape::elementWiseStride(zTadShapeInfo);
+
+                const nd4j::LoopKind::Kind kindOfLoop = nd4j::LoopKind::deduceKindOfLoopXYZ(yTadShapeShapeInfo, xShapeInfo, zTadShapeInfo);
+
+                if (kindOfLoop == nd4j::LoopKind::EWS1) {
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+                        auto oY = y + tadOffsets[i];
+                        auto oZ = z + zTadOffset[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (unsigned int f = 0; f < tadLength; f++)
+                            oZ[f] = OpType::op(x[f], oY[f]);
+                    }
+                }
+                else if(kindOfLoop == nd4j::LoopKind::EWSNONZERO) {
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+                        auto oY = y + tadOffsets[i];
+                        auto oZ = z + zTadOffset[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (uint f = 0; f < tadLength; f++)
+                            oZ[f * zEws] = OpType::op(x[f * xEws], oY[f * yEws]);
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(yTadShapeShapeInfo, xShapeInfo) && shape::haveSameShapeAndStrides(yTadShapeShapeInfo, zTadShapeInfo)) {
+
+                    uint tadShapeShapeInfoCast[MAX_RANK];
+                    bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+
+                        auto oY = y + tadOffsets[i];
+                        auto oZ = z + zTadOffset[i];
+
+                        // TODO: cover this codebranch with tests
+                        // all this stuff already happens within thread
+                        PRAGMA_OMP_SIMD
+                        for (int f = 0; f < tadLength; f++) {
+                            auto offset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastY);
+                            oZ[offset] = OpType::op(x[offset], oY[offset]);
+                        }
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(yTadShapeShapeInfo, xShapeInfo)) {
+
+                    uint tadShapeShapeInfoCast[MAX_RANK];
+                    uint tadShapeInfoZCast[MAX_RANK];
+                    bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
+                    bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
+
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+
+                        auto oZ = z + zTadOffset[i];
+                        auto oY = y + tadOffsets[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (int f = 0; f < tadLength; f++) {
+                            auto offset  = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastY);
+                            auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, lenZ, canCastZ);
+                            oZ[zOffset] = OpType::op(x[offset], oY[offset]);
+                        }
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(yTadShapeShapeInfo, zTadShapeInfo)) {
+
+                    uint tadShapeShapeInfoCast[MAX_RANK];
+                    uint xShapeInfoCast[MAX_RANK];
+                    bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
+                    bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+
+                        auto oZ = z + zTadOffset[i];
+                        auto oY = y + tadOffsets[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (int f = 0; f < tadLength; f++) {
+                            auto offset  = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastY);
+                            auto xOffset = shape::indexOffset(f, xShapeInfo, xShapeInfoCast, lenX, canCastX);
+                            oZ[offset] = OpType::op(x[xOffset], oY[offset]);
+                        }
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(xShapeInfo, zTadShapeInfo)) {
+
+                    uint tadShapeShapeInfoCast[MAX_RANK];
+                    uint xShapeInfoCast[MAX_RANK];
+                    bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
+                    bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+
+                        auto oZ = z + zTadOffset[i];
+                        auto oY = y + tadOffsets[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (int f = 0; f < tadLength; f++) {
+                            auto yOffset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastY);
+                            auto offset  = shape::indexOffset(f, xShapeInfo, xShapeInfoCast, lenX, canCastX);
+                            oZ[offset] = OpType::op(x[offset], oY[yOffset]);
+                        }
+                    }
+                }
+                else {
+
+                    uint xShapeInfoCast[MAX_RANK];
+                    uint tadShapeShapeInfoCast[MAX_RANK];
+                    uint tadShapeInfoZCast[MAX_RANK];
+                    bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
+                    bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yTadShapeShapeInfo, tadShapeShapeInfoCast);
+                    bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zTadShapeInfo, tadShapeInfoZCast);
+
+                    PRAGMA_OMP_PARALLEL_FOR_THREADS(threads)
+                    for (int i = 0; i < tads; i++) {
+
+                        auto oZ = z + zTadOffset[i];
+                        auto oY = y + tadOffsets[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (int f = 0; f < tadLength; f++) {
+                            auto xOffset = shape::indexOffset(f, xShapeInfo, xShapeInfoCast, lenX, canCastX);
+                            auto yOffset = shape::indexOffset(f, yTadShapeShapeInfo, tadShapeShapeInfoCast, tadLength, canCastY);
+                            auto zOffset = shape::indexOffset(f, zTadShapeInfo, tadShapeInfoZCast, lenZ, canCastZ);
+                            oZ[zOffset] = OpType::op(x[xOffset], oY[yOffset]);
+                        }
+                    }
+                }
+        }
+
+        BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT BroadcastInt, , INTEGER_TYPES);
+    }
+}
\ No newline at end of file
diff --git a/libnd4j/include/loops/cpu/pairwise_int.cpp b/libnd4j/include/loops/cpu/pairwise_int.cpp
new file mode 100644
index 000000000..b356adcc2
--- /dev/null
+++ b/libnd4j/include/loops/cpu/pairwise_int.cpp
@@ -0,0 +1,309 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+
+#include <loops/pairwise_int.h>
+#include <types/types.h>
+#include <LoopKind.h>
+#include <OmpLaunchHelper.h>
+
+using namespace simdOps;
+
+namespace functions {
+    namespace pairwise_transforms {
+
+        template <typename X>
+        void PairWiseIntTransform<X>::exec(
+                const int opNum,
+                void *x,
+                Nd4jLong xEws,
+                void *y,
+                Nd4jLong yEws,
+                void *z,
+                Nd4jLong zEws,
+                void *extraParams,
+                Nd4jLong n) {
+            DISPATCH_BY_OPNUM_T(exec, PARAMS(x,
+                                               xEws,
+                                               y,
+                                               yEws,
+                                               z,
+                                               zEws,
+                                               extraParams,
+                                               n), PAIRWISE_INT_OPS);
+        };
+
+
+
+        template <typename X>
+        template <typename OpType>
+        void PairWiseIntTransform<X>::exec(void *vx,
+                                              Nd4jLong xEws,
+                                              void *vy,
+                                              Nd4jLong yEws,
+                                              void *vz,
+                                              Nd4jLong zEws,
+                                              void *vextraParams,
+                                              const Nd4jLong n) {
+            
+            auto x = reinterpret_cast<X *>(vx);
+            auto y = reinterpret_cast<X *>(vy);
+            auto z = reinterpret_cast<X *>(vz);
+            auto extraParams = reinterpret_cast<X *>(vextraParams);
+
+            nd4j::OmpLaunchHelper info(n);
+
+            if (xEws == 1 && yEws == 1 && zEws == 1) {
+
+                PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                {                
+                    auto threadNum = omp_get_thread_num();
+                    Nd4jLong threadOffset = info.getThreadOffset(threadNum);        
+                    auto xi = x + threadOffset;
+                    auto yi = y + threadOffset;
+                    auto zi = z + threadOffset;
+                    auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                    PRAGMA_OMP_SIMD
+                    for (Nd4jLong i = 0; i < ulen; i++)
+                        zi[i] = OpType::op(xi[i], yi[i], extraParams);
+                }
+            }
+            else {
+
+                PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                {                
+                    auto threadNum = omp_get_thread_num();
+                    Nd4jLong threadOffset = info.getThreadOffset(threadNum);        
+                    auto xi = x + xEws*threadOffset;
+                    auto yi = y + yEws*threadOffset;
+                    auto zi = z + zEws*threadOffset;
+                    auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                    PRAGMA_OMP_SIMD
+                    for (Nd4jLong i = 0; i < ulen; i++)
+                        zi[i*zEws] = OpType::op(xi[i*xEws], yi[i*yEws], extraParams);
+                }
+            }
+        }
+
+        template <typename X>
+        void PairWiseIntTransform<X>::exec(
+                const int opNum,
+                void *x,
+                Nd4jLong *xShapeInfo,
+                void *y,
+                Nd4jLong *yShapeInfo,
+                void *z,
+                Nd4jLong *zShapeInfo,
+                void *extraParams) {
+            DISPATCH_BY_OPNUM_T(exec, PARAMS(x,
+                                              xShapeInfo,
+                                              y,
+                                              yShapeInfo,
+                                              z,
+                                              zShapeInfo,
+                                              extraParams),
+                                 PAIRWISE_INT_OPS);
+        };
+
+
+        template <typename X>
+        template <typename OpType>
+        void PairWiseIntTransform<X>::exec(void *vx, Nd4jLong* xShapeInfo,
+                                            void *vy, Nd4jLong* yShapeInfo,
+                                            void *vz, Nd4jLong* zShapeInfo,
+                                            void *vextraParams) {
+
+            auto x = reinterpret_cast<X *>(vx);
+            auto y = reinterpret_cast<X *>(vy);
+            auto z = reinterpret_cast<X *>(vz);
+            auto extraParams = reinterpret_cast<X *>(vextraParams);
+
+            auto n = shape::length(xShapeInfo);
+            auto xEws = shape::elementWiseStride(xShapeInfo);
+            auto yEws = shape::elementWiseStride(yShapeInfo);
+            auto zEws = shape::elementWiseStride(zShapeInfo);
+
+            nd4j::OmpLaunchHelper info(n);
+
+            if (shape::isScalar(yShapeInfo)) {
+
+               uint xShapeInfoCast[MAX_RANK];
+               const bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
+
+                if(shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo)) {
+
+                    PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                    {                
+                        auto threadNum = omp_get_thread_num();
+                        auto threadOffset = info.getThreadOffset(threadNum);
+                        auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                        PRAGMA_OMP_SIMD
+                        for(Nd4jLong i = 0; i < ulen; i++)  {
+                            auto offset = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, n, canCastX);
+                            z[offset] = OpType::op(x[offset], y[0], extraParams);
+                        }
+                    }
+                }
+                else {
+                    
+                    uint zShapeInfoCast[MAX_RANK];
+                    const bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zShapeInfo, zShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                    {                
+                        auto threadNum = omp_get_thread_num();
+                        auto threadOffset = info.getThreadOffset(threadNum);
+                        auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                        PRAGMA_OMP_SIMD
+                        for(Nd4jLong i = 0; i < ulen; i++)  {
+                            auto xOffset = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, n, canCastX);
+                            auto zOffset = shape::indexOffset(i + threadOffset, zShapeInfo, zShapeInfoCast, n, canCastZ);
+                            z[zOffset] = OpType::op(x[xOffset], y[0], extraParams);
+                        }
+                    }
+                }
+                return;
+            }
+
+            const nd4j::LoopKind::Kind kindOfLoop = nd4j::LoopKind::deduceKindOfLoopXYZ(xShapeInfo, yShapeInfo, zShapeInfo);
+            const bool sameShapesXY = shape::shapeEquals(xShapeInfo, yShapeInfo);
+
+            if ((kindOfLoop == nd4j::LoopKind::EWS1 || kindOfLoop == nd4j::LoopKind::EWSNONZERO) && sameShapesXY) {
+                exec<OpType>(x, xEws, y, yEws, z, zEws, extraParams, n);
+            }            
+            else if ((kindOfLoop == nd4j::LoopKind::EWS1 || kindOfLoop == nd4j::LoopKind::EWSNONZERO) && !sameShapesXY) { //not same shape
+                exec<OpType>(x, xEws, y, yEws, z, zEws, extraParams, shape::length(yShapeInfo));
+            }
+            else {                
+
+                if(shape::haveSameShapeAndStrides(xShapeInfo, yShapeInfo) && shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo)) {
+
+                    uint xShapeInfoCast[MAX_RANK];
+                    const bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                    {                
+                        auto threadNum = omp_get_thread_num();
+                        auto threadOffset = info.getThreadOffset(threadNum);
+                        auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                        PRAGMA_OMP_SIMD
+                        for (Nd4jLong i = 0; i < ulen; i++)  {
+                            auto offset = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, n, canCastX);
+                            z[offset] = OpType::op(x[offset], y[offset], extraParams);
+                        }
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(xShapeInfo, yShapeInfo)) {
+                    
+                    uint xShapeInfoCast[MAX_RANK];
+                    uint zShapeInfoCast[MAX_RANK];
+                    const bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
+                    const bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zShapeInfo, zShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                    {                
+                        auto threadNum = omp_get_thread_num();
+                        auto threadOffset = info.getThreadOffset(threadNum);
+                        auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                        PRAGMA_OMP_SIMD
+                        for (Nd4jLong i = 0; i < ulen; i++)  {
+                            auto offset  = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, n, canCastX);                            
+                            auto zOffset = shape::indexOffset(i + threadOffset, zShapeInfo, zShapeInfoCast, n, canCastZ);
+                            z[zOffset] = OpType::op(x[offset], y[offset], extraParams);
+                        }
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo)) {
+                    
+                    uint xShapeInfoCast[MAX_RANK];
+                    uint yShapeInfoCast[MAX_RANK];
+                    const bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
+                    const bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                    {                
+                        auto threadNum = omp_get_thread_num();
+                        auto threadOffset = info.getThreadOffset(threadNum);
+                        auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                        PRAGMA_OMP_SIMD
+                        for (Nd4jLong i = 0; i < ulen; i++)  {
+                            auto offset  = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, n, canCastX);                            
+                            auto yOffset = shape::indexOffset(i + threadOffset, yShapeInfo, yShapeInfoCast, n, canCastY);
+                            z[offset] = OpType::op(x[offset], y[yOffset], extraParams);
+                        }
+                    }
+                }
+                else if(shape::haveSameShapeAndStrides(yShapeInfo, zShapeInfo)) {
+                    
+                    uint xShapeInfoCast[MAX_RANK];
+                    uint yShapeInfoCast[MAX_RANK];
+                    const bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
+                    const bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                    {                
+                        auto threadNum = omp_get_thread_num();
+                        auto threadOffset = info.getThreadOffset(threadNum);
+                        auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                        PRAGMA_OMP_SIMD
+                        for (Nd4jLong i = 0; i < ulen; i++)  {
+                            auto xOffset = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, n, canCastX);                            
+                            auto offset  = shape::indexOffset(i + threadOffset, yShapeInfo, yShapeInfoCast, n, canCastY);
+                            z[offset] = OpType::op(x[xOffset], y[offset], extraParams);
+                        }
+                    }
+                }
+                else {
+                    
+                    uint xShapeInfoCast[MAX_RANK];
+                    uint yShapeInfoCast[MAX_RANK];
+                    uint zShapeInfoCast[MAX_RANK];
+                    const bool canCastX = nd4j::DataTypeUtils::castShapeInfo(xShapeInfo, xShapeInfoCast);
+                    const bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
+                    const bool canCastZ = nd4j::DataTypeUtils::castShapeInfo(zShapeInfo, zShapeInfoCast);
+
+                    PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                    {                
+                        auto threadNum = omp_get_thread_num();
+                        auto threadOffset = info.getThreadOffset(threadNum);
+                        auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                        PRAGMA_OMP_SIMD
+                        for (Nd4jLong i = 0; i < ulen; i++)  {
+                            auto xOffset = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, n, canCastX);                            
+                            auto yOffset = shape::indexOffset(i + threadOffset, yShapeInfo, yShapeInfoCast, n, canCastY);
+                            auto zOffset = shape::indexOffset(i + threadOffset, zShapeInfo, zShapeInfoCast, n, canCastZ);
+                            z[zOffset] = OpType::op(x[xOffset], y[yOffset], extraParams);
+                        }
+                    }
+                }
+            }
+        }
+
+        BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT PairWiseIntTransform, , INTEGER_TYPES);
+    }
+}
diff --git a/libnd4j/include/loops/cpu/scalar_int.cpp b/libnd4j/include/loops/cpu/scalar_int.cpp
new file mode 100644
index 000000000..9920cc836
--- /dev/null
+++ b/libnd4j/include/loops/cpu/scalar_int.cpp
@@ -0,0 +1,255 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+
+#include "../scalar_int.h"
+#include <op_boilerplate.h>
+#include <types/types.h>
+#include <LoopKind.h>
+
+#include "../legacy_ops.h"
+
+using namespace simdOps;
+
+namespace functions {
+    namespace scalar {
+
+
+        template<typename X>
+        template<typename OpType>
+        void ScalarIntTransform<X>::transform(void *vx, Nd4jLong *xShapeInfo,
+                                                void *vextraParams, 
+                                                void *vz,  Nd4jLong *zShapeInfo, 
+                                                void *vscalars, 
+                                                int *dimension, int dimensionLength, 
+                                                Nd4jLong *xTadShapeInfo, Nd4jLong *xTadOffsets,
+                                                Nd4jLong *zTadShapeInfo, Nd4jLong *zTadOffsets) {
+            
+            auto x = reinterpret_cast<X *>(vx);
+            auto z = reinterpret_cast<X *>(vz);
+            auto scalars = reinterpret_cast<X *>(vscalars);
+            auto extraParams = reinterpret_cast<X *>(vextraParams);
+
+            if (zTadShapeInfo == nullptr) {
+                zTadShapeInfo = xTadShapeInfo;
+                zTadOffsets   = xTadOffsets;
+            }
+
+            // tad preparation
+            const int xTadEws    = shape::elementWiseStride(xTadShapeInfo);
+            const int zTadEws    = shape::elementWiseStride(zTadShapeInfo);
+            const int tadLength  = shape::tadLength(xShapeInfo, dimension, dimensionLength);
+            const int numTads    = shape::length(xShapeInfo) / tadLength;
+
+            nd4j::LoopKind::Kind kindOfLoop = nd4j::LoopKind::deduceKindOfLoopXZ(xTadShapeInfo, zTadShapeInfo);
+
+            if (kindOfLoop != nd4j::LoopKind::EWS1 && kindOfLoop != nd4j::LoopKind::EWSNONZERO) {
+                printf("ScalarIntTransform<X>::transform: super-bad loop visited. Shouldn't ever happen\n");
+                return;
+            }
+            
+            int num_threads = nd4j::math::nd4j_min<int>(numTads, omp_get_max_threads());
+
+            if (kindOfLoop == nd4j::LoopKind::EWS1) {
+                PRAGMA_OMP_PARALLEL_FOR_THREADS(num_threads)
+                for (unsigned int r = 0; r < numTads; r++) {
+                    auto oZ = z + zTadOffsets[r];
+                    auto oX = x + xTadOffsets[r];
+
+                    PRAGMA_OMP_SIMD
+                    for (unsigned int f = 0; f < tadLength; f++)
+                        oZ[f] = OpType::op(oX[f], scalars[r], extraParams);
+                }
+            } 
+            else { // kindOfLoop != nd4j::LoopKind::EWSNONZERO
+                PRAGMA_OMP_PARALLEL_FOR_THREADS(num_threads)
+                for (unsigned int r = 0; r < numTads; r++) {
+                    auto oZ = z + zTadOffsets[r];
+                    auto oX = x + xTadOffsets[r];
+
+                    PRAGMA_OMP_SIMD
+                    for (unsigned int f = 0; f < tadLength; f++)
+                        oZ[f * zTadEws] = OpType::op(oX[f * xTadEws], scalars[r], extraParams);
+                }
+            }          
+        }
+
+        template<typename X>
+        void ScalarIntTransform<X>::transform(int opNum,
+                              void *x,
+                              Nd4jLong *xShapeInfo,
+                              void *extraParams,
+                              void *z,
+                              Nd4jLong *zShapeInfo,
+                              void *scalars,
+                              int *dimension,
+                              int dimensionLength,
+                              Nd4jLong *xTadShapeInfo,
+                              Nd4jLong *xTadOffsets,
+                              Nd4jLong *zTadShapeInfo,
+                              Nd4jLong *zTadOffsets) {
+            DISPATCH_BY_OPNUM_T(transform, PARAMS(x, xShapeInfo, extraParams, z, zShapeInfo, scalars, dimension, dimensionLength, xTadShapeInfo, xTadOffsets, zTadShapeInfo, zTadOffsets), SCALAR_INT_OPS);
+        }
+
+
+        template<typename X>
+        void ScalarIntTransform<X>::transform(const int opNum,
+                void *x,
+                Nd4jLong xEws,
+                void *z,
+                Nd4jLong zEws,
+                void *scalar,
+                void *extraParams,
+                const Nd4jLong n) {
+            DISPATCH_BY_OPNUM_T(transform, PARAMS(x, xEws, z, zEws, scalar, extraParams, n), SCALAR_INT_OPS);
+        }
+
+        template<typename X>
+        void ScalarIntTransform<X>::transform(const int opNum,
+                void *x,
+                Nd4jLong *xShapeInfo,
+                void *z,
+                Nd4jLong *zShapeInfo,
+                void *scalar,
+                void *extraParams) {
+            DISPATCH_BY_OPNUM_T(transform, PARAMS(x, xShapeInfo, z, zShapeInfo, scalar, extraParams), SCALAR_INT_OPS);
+        }
+
+        template<typename X>
+        template<typename OpType>
+        void ScalarIntTransform<X>::transform(void *vx,
+                               Nd4jLong *xShapeInfo,
+                               void *vz,
+                               Nd4jLong *zShapeInfo,
+                               void *vscalar,
+                               void *vextraParams) {
+            
+            auto x = reinterpret_cast<X *>(vx);
+            auto z = reinterpret_cast<X *>(vz);
+            auto scalar = reinterpret_cast<X *>(vscalar)[0];
+            auto extraParams = reinterpret_cast<X *>(vextraParams);
+
+            auto xEws = shape::elementWiseStride(xShapeInfo);
+            auto zEws = shape::elementWiseStride(zShapeInfo);
+            auto len = shape::length(xShapeInfo);
+
+            // nd4j_logger("Launching scalar: xOrder: %i; zOrder: %i; xEWS: %i\n", xOrder, zOrder, xEws);
+
+            nd4j::LoopKind::Kind kindOfLoop = nd4j::LoopKind::deduceKindOfLoopXZ(xShapeInfo, zShapeInfo);
+
+            if (kindOfLoop == nd4j::LoopKind::EWS1 || kindOfLoop == nd4j::LoopKind::EWSNONZERO) {
+                transform<OpType>(x, xEws, z, zEws, vscalar, extraParams, len);
+                return;
+            }
+
+            uint xShapeInfoCast[MAX_RANK];
+            const bool canCastX = nd4j::DataTypeUtils::castShapeInfo<uint>(xShapeInfo, xShapeInfoCast);
+
+            nd4j::OmpLaunchHelper info(len);
+                               
+            if(shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo)) {
+
+                PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                {
+                    auto threadNum = omp_get_thread_num();                    
+                    auto threadOffset = info.getThreadOffset(threadNum);
+                    auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                    PRAGMA_OMP_SIMD
+                    for (unsigned int i = 0; i < ulen; i++) {
+                        auto offset = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, len, canCastX);
+                        z[offset] = OpType::op(x[offset], scalar, extraParams);
+                    }
+                }
+            }
+            else {
+                
+                uint zShapeInfoCast[MAX_RANK];
+                const bool canCastZ = nd4j::DataTypeUtils::castShapeInfo<uint>(zShapeInfo, zShapeInfoCast);
+
+                PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                {
+                    auto threadNum = omp_get_thread_num();                    
+                    auto threadOffset = info.getThreadOffset(threadNum);
+                    auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                    PRAGMA_OMP_SIMD
+                    for (unsigned int i = 0; i < ulen; i++) {
+                        auto xOffset = shape::indexOffset(i + threadOffset, xShapeInfo, xShapeInfoCast, len, canCastX);
+                        auto zOffset = shape::indexOffset(i + threadOffset, zShapeInfo, zShapeInfoCast, len, canCastZ);
+                        z[zOffset] = OpType::op(x[xOffset], scalar, extraParams);
+                    }
+                }
+            }          
+        }
+
+
+            template<typename X>
+            template<typename OpType>
+            void ScalarIntTransform<X>::transform(void *vx,
+                    Nd4jLong xEws,
+                    void *vz,
+                    Nd4jLong zEws,
+                    void *vscalar,
+                    void *vextraParams,
+                    const Nd4jLong len) {
+
+                auto x = reinterpret_cast<X *>(vx);
+                auto z = reinterpret_cast<X *>(vz);
+                auto scalar = reinterpret_cast<X *>(vscalar)[0];
+                auto extraParams = reinterpret_cast<X *>(vextraParams); 
+
+                nd4j::OmpLaunchHelper info(len);
+
+                if (xEws == 1 && zEws == 1) {
+
+                    PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                    {
+                        auto threadNum = omp_get_thread_num();
+                        auto threadOffset = info.getThreadOffset(threadNum);
+                        auto xi = x + threadOffset;
+                        auto zi = z + threadOffset;
+                        auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                        PRAGMA_OMP_SIMD
+                        for (unsigned int i = 0; i < ulen; i++)
+                            zi[i] = OpType::op(xi[i], scalar, extraParams);
+                    }
+                } 
+                else {
+
+                    PRAGMA_OMP_PARALLEL_THREADS(info._numThreads)
+                    {
+                        auto threadNum = omp_get_thread_num();
+                        auto threadOffset = info.getThreadOffset(threadNum);
+                        auto xi = x + xEws * threadOffset;
+                        auto zi = z + zEws * threadOffset;
+                        auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
+
+                        PRAGMA_OMP_SIMD
+                        for (unsigned int i = 0; i < ulen; i++)
+                            zi[i * zEws] = OpType::op(xi[i * xEws], scalar, extraParams);
+                    }
+                }
+            }
+
+        BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT ScalarIntTransform, , INTEGER_TYPES);
+
+}
+}
diff --git a/libnd4j/include/loops/cuda/broadcasting_int.cu b/libnd4j/include/loops/cuda/broadcasting_int.cu
new file mode 100644
index 000000000..38193f35d
--- /dev/null
+++ b/libnd4j/include/loops/cuda/broadcasting_int.cu
@@ -0,0 +1,291 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+//  @author raver119@gmail.com
+//
+
+#include <op_boilerplate.h>
+#include <loops/broadcasting_int.h>
+#include <loops/legacy_ops.h>
+#include <types/types.h>
+#include <Environment.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <string>
+#include <stdexcept>
+#include <StringUtils.h>
+
+using namespace simdOps;
+
+//////////////////////////////////////////////////////////////////////////
+template<typename X, typename OpClass>
+static __global__ void broadcastIntSimple(
+        void *x,
+        Nd4jLong *xShapeInfo,
+        void *y,
+        Nd4jLong *yShapeInfo,
+        void *z,
+        Nd4jLong *zShapeInfo,
+        int *dimension,
+        int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+    
+    functions::broadcast::BroadcastInt<X>::template transformCuda<OpClass>(x,xShapeInfo,y,yShapeInfo,z,zShapeInfo,dimension,dimensionLength,tadOnlyShapeInfo,tadOffsets,tadOnlyShapeInfoZ,tadOffsetsZ);
+}
+
+//////////////////////////////////////////////////////////////////////////
+template<typename X, typename OpClass>
+static __global__ void broadcastBoolInverseSimple(
+        void *x,
+        Nd4jLong *xShapeInfo,
+        void *y,
+        Nd4jLong *yShapeInfo,
+        void *z,
+        Nd4jLong *zShapeInfo,
+        int *dimension,
+        int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+
+    functions::broadcast::BroadcastInt<X>::template transformInverseCuda<OpClass>(x,xShapeInfo,y,yShapeInfo,z,zShapeInfo,dimension,dimensionLength,tadOnlyShapeInfo,tadOffsets,tadOnlyShapeInfoZ,tadOffsetsZ);
+}
+
+namespace functions {
+    namespace broadcast {
+//////////////////////////////////////////////////////////////////////////
+        template<typename X>
+        template <typename OpClass>
+        __host__ void BroadcastInt<X>::intermediateBroadcast(dim3 launchDims, cudaStream_t *stream, void *x, Nd4jLong *xShapeInfo, void *y, Nd4jLong *yShapeInfo, void *z, Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+            broadcastIntSimple<X, OpClass><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ);
+        }
+
+//////////////////////////////////////////////////////////////////////////
+        template<typename X>
+        __host__ void BroadcastInt<X>::execBroadcast(dim3 launchDims, cudaStream_t *stream, int opNum, void *x, Nd4jLong *xShapeInfo, void *y, Nd4jLong *yShapeInfo, void *z, Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+            DISPATCH_BY_OPNUM_T(intermediateBroadcast,  PARAMS(launchDims, stream, x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ), OPS_A(BROADCAST_INT_OPS))
+        }
+
+//////////////////////////////////////////////////////////////////////////
+        template<typename X>
+        template <typename OpClass>
+        __host__ void BroadcastInt<X>::intermediateInverseBroadcast(dim3 launchDims, cudaStream_t *stream, void *x, Nd4jLong *xShapeInfo, void *y, Nd4jLong *yShapeInfo, void *z, Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+            broadcastBoolInverseSimple<X, OpClass><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ);
+        }
+
+//////////////////////////////////////////////////////////////////////////
+        template<typename X>
+        __host__ void BroadcastInt<X>::execInverseBroadcast(dim3 launchDims, cudaStream_t *stream, int opNum, void *x, Nd4jLong *xShapeInfo, void *y, Nd4jLong *yShapeInfo, void *z, Nd4jLong *zShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+            DISPATCH_BY_OPNUM_T(intermediateInverseBroadcast,  PARAMS(launchDims, stream, x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ), OPS_A(BROADCAST_INT_OPS))
+        }
+
+//////////////////////////////////////////////////////////////////////////
+        template<typename X>
+        template <typename OpType>
+        __device__ void BroadcastInt<X>::transformInverseCuda(
+                void *vx, Nd4jLong *xShapeInfo,
+                void *vy, Nd4jLong *yShapeInfo,
+                void *vz, Nd4jLong *zShapeInfo,
+                int *dimension, int dimensionLength,
+                Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+
+            if (tadOnlyShapeInfoZ == nullptr) {
+                tadOnlyShapeInfoZ = tadOnlyShapeInfo;
+                tadOffsetsZ = tadOffsets;
+            }
+
+            auto x = reinterpret_cast<X*>(vx);
+            auto y = reinterpret_cast<X*>(vy);
+            auto z = reinterpret_cast<X*>(vz);
+
+            //decompose in to several sub tads after
+            //moving all dimensions (in sorted order)
+            //to the back.
+            //permuted version of the x shape info for setting up the tad problem
+            __shared__ Nd4jLong tadLength;
+            __shared__ Nd4jLong tadEWS;
+            __shared__ int numTads;
+            __shared__ Nd4jLong xEWS;
+            __shared__ Nd4jLong zEWS;
+
+            if (threadIdx.x == 0) {
+                tadLength = shape::length(tadOnlyShapeInfo);//shape::tadLength(xShapeInfo, dimension, dimensionLength);
+                tadEWS = shape::elementWiseStride(tadOnlyShapeInfo);
+                numTads = shape::length(yShapeInfo) / tadLength;
+                xEWS = shape::elementWiseStride(xShapeInfo);
+                zEWS = shape::elementWiseStride(tadOnlyShapeInfoZ);
+            }
+            __syncthreads();
+
+            for (int r = blockIdx.x; r < numTads; r += gridDim.x) {
+                auto rZ = z + tadOffsetsZ[r];
+                auto rY = y + tadOffsets[r];
+
+                if(tadEWS > 0 && zEWS > 0 && xEWS > 0 && dimensionLength == 1) {
+
+                    for (int i = threadIdx.x; i < tadLength; i+= blockDim.x)
+                        rZ[i * zEWS] = OpType::op(x[i * xEWS], rY[i * tadEWS]);
+                }
+                else {
+                    // it is expected that x and z tads and y array all have the same length
+                    for (Nd4jLong i = threadIdx.x; i < tadLength; i+= blockDim.x) {
+                        auto xOffset = shape::getIndexOffset(i, xShapeInfo,  tadLength);
+                        auto yOffset = shape::getIndexOffset(i, tadOnlyShapeInfo, tadLength);
+                        auto zOffset = shape::getIndexOffset(i, tadOnlyShapeInfoZ, tadLength);
+
+                        rZ[zOffset] = OpType::op(x[xOffset], rY[yOffset]);
+                    }
+                }
+            }
+        }
+
+//////////////////////////////////////////////////////////////////////////
+        template<typename X>
+        template <typename OpType>
+		__device__ void BroadcastInt<X>::transformCuda(
+		                              void *vx, Nd4jLong *xShapeInfo,
+		                              void *vy, Nd4jLong *yShapeInfo,
+		                              void *vz, Nd4jLong *zShapeInfo,
+		                              int *dimension, int dimensionLength,
+                                      Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadOnlyShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+
+            if (tadOnlyShapeInfoZ == nullptr) {
+                tadOnlyShapeInfoZ = tadOnlyShapeInfo;
+                tadOffsetsZ = tadOffsets;
+            }
+
+            auto x = reinterpret_cast<X*>(vx);
+            auto y = reinterpret_cast<X*>(vy);
+            auto z = reinterpret_cast<X*>(vz);
+
+            //decompose in to several sub tads after
+            //moving all dimensions (in sorted order)
+            //to the back.
+            //permuted version of the x shape info for setting up the tad problem
+            __shared__ Nd4jLong tadLength;
+            __shared__ Nd4jLong tadEWS;
+            __shared__ int numTads;
+            __shared__ Nd4jLong yEWS;
+            __shared__ Nd4jLong zEWS;
+      
+            if (threadIdx.x == 0) {
+   	            tadLength = shape::length(tadOnlyShapeInfo);//shape::tadLength(xShapeInfo, dimension, dimensionLength);
+                tadEWS = shape::elementWiseStride(tadOnlyShapeInfo);
+                numTads = shape::length(xShapeInfo) / tadLength;
+                yEWS = shape::elementWiseStride(yShapeInfo);
+                zEWS = shape::elementWiseStride(tadOnlyShapeInfoZ);    
+            }
+            __syncthreads();
+
+            __shared__ X *rZ;
+            __shared__ X *rX;
+
+		for (int r = blockIdx.x; r < numTads; r += gridDim.x) {
+
+            if (threadIdx.x == 0) {
+                rZ = z + tadOffsetsZ[r];
+                rX = x + tadOffsets[r];
+            }
+            __syncthreads();
+
+
+            if(tadEWS > 0 && zEWS > 0 && yEWS > 0 && dimensionLength == 1) {
+
+                for (int i = threadIdx.x; i < tadLength; i+= blockDim.x)
+                    rZ[i * zEWS] = OpType::op(rX[i * tadEWS], y[i * yEWS]);
+            }
+            else {
+                // it is expected that x and z tads and y array all have the same length
+                for (Nd4jLong i = threadIdx.x; i < tadLength; i+= blockDim.x) {
+                    auto xOffset = shape::getIndexOffset(i, tadOnlyShapeInfo,  tadLength);
+                    auto yOffset = shape::getIndexOffset(i, yShapeInfo, tadLength);
+                    auto zOffset = shape::getIndexOffset(i, tadOnlyShapeInfoZ, tadLength);
+
+                    rZ[zOffset] = OpType::op(rX[xOffset], y[yOffset]);
+                }
+            }
+		}
+	}
+
+
+        template<typename X>
+        void BroadcastInt<X>::exec(int opNum,
+                         void *x,
+                         Nd4jLong *xShapeInfo,
+                         void *y,
+                         Nd4jLong *yShapeInfo,
+                         void *result,
+                         Nd4jLong *resultShapeInfo,
+                         int *dimension,
+                         int dimensionLength,
+                         Nd4jLong *tadShapeInfo,
+                         Nd4jLong *tadOffset,
+                         Nd4jLong *tadShapeInfoZ,
+                         Nd4jLong *tadOffsetZ) {
+
+        }
+
+        template<typename X>
+        void BroadcastInt<X>::execInverse(int opNum,
+                                void *x,
+                                Nd4jLong *xShapeInfo,
+                                void *y,
+                                Nd4jLong *yShapeInfo,
+                                void *result,
+                                Nd4jLong *resultShapeInfo,
+                                int *dimension,
+                                int dimensionLength,
+                                Nd4jLong *tadShapeInfo,
+                                Nd4jLong *tadOffset,
+                                Nd4jLong *tadShapeInfoZ,
+                                Nd4jLong *tadOffsetZ) {
+
+        }
+
+        template<typename X>
+        template<typename OpType>
+        void BroadcastInt<X>::exec(void *x,
+                         Nd4jLong *xShapeInfo,
+                         void *y,
+                         Nd4jLong *yShapeInfo,
+                         void *result,
+                         Nd4jLong *resultShapeInfo,
+                         int *dimension,
+                         int dimensionLength,
+                         Nd4jLong *tadShapeInfo,
+                         Nd4jLong *tadOffset,
+                         Nd4jLong *tadShapeInfoZ,
+                         Nd4jLong *tadOffsetZ) {
+
+        }
+
+        template<typename X>
+        template<typename OpType>
+        void BroadcastInt<X>::execInverse(void *x,
+                                Nd4jLong *xShapeInfo,
+                                void *y,
+                                Nd4jLong *yShapeInfo,
+                                void *result,
+                                Nd4jLong *resultShapeInfo,
+                                int *dimension,
+                                int dimensionLength,
+                                Nd4jLong *tadShapeInfo,
+                                Nd4jLong *tadOffset,
+                                Nd4jLong *tadShapeInfoZ,
+                                Nd4jLong *tadOffsetZ) {
+
+        }
+
+        BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT BroadcastInt, , INTEGER_TYPES);
+    }
+}
\ No newline at end of file
diff --git a/libnd4j/include/loops/cuda/pairwise_int.cu b/libnd4j/include/loops/cuda/pairwise_int.cu
new file mode 100644
index 000000000..5cc12846c
--- /dev/null
+++ b/libnd4j/include/loops/cuda/pairwise_int.cu
@@ -0,0 +1,173 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//  @author raver119@gmail.com
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 08.11.2018
+
+#ifndef PAIRWISE_INT_CU
+#define PAIRWISE_INT_CU
+
+
+#include "../pairwise_int.h"
+
+
+using namespace simdOps;
+
+////////////////////////////////////////////////////////////////////////////////
+template <typename X, typename OpType>
+__global__ static void pairwiseSimpleShaped(void* vx, Nd4jLong *xShapeInfo,
+											void *vy, Nd4jLong *yShapeInfo,
+											void *vz, Nd4jLong *zShapeInfo,
+											void *vextraParams) {
+
+	auto x = reinterpret_cast<X*>(vx);
+	auto y = reinterpret_cast<X*>(vy);
+	auto z = reinterpret_cast<X*>(vz);
+	auto extraParams = reinterpret_cast<X*>(vextraParams);
+
+	int tid = blockIdx.x * blockDim.x + threadIdx.x;
+
+	__shared__ int xEws;
+	__shared__ int yEws;
+	__shared__ int zEws;
+	__shared__ char xOrder;
+	__shared__ char yOrder;
+	__shared__ char zOrder;
+	__shared__ Nd4jLong len;
+
+	if (threadIdx.x == 0) {
+		xEws = shape::elementWiseStride(xShapeInfo);
+		yEws = shape::elementWiseStride(yShapeInfo);
+    	zEws = shape::elementWiseStride(zShapeInfo);
+		xOrder = shape::order(xShapeInfo);
+		yOrder = shape::order(yShapeInfo);
+		zOrder = shape::order(zShapeInfo);
+		len = shape::length(xShapeInfo);
+	}
+	__syncthreads();
+
+
+	if (xEws >= 1 && yEws >= 1 && zEws >= 1 && xOrder == yOrder && xOrder == zOrder) {
+		for (Nd4jLong i = tid; i < len; i += gridDim.x * blockDim.x) {
+			z[i * zEws] = OpType::op(x[i * xEws], y[i * yEws], extraParams);
+		}
+	}
+	else if (vx == vz) {
+		for (Nd4jLong i = tid; i < len; i += gridDim.x * blockDim.x) {
+			auto xOffset = shape::getIndexOffset(i, xShapeInfo, len);
+			auto yOffset = shape::getIndexOffset(i, yShapeInfo, len);
+				
+			z[xOffset] = OpType::op(x[xOffset], y[yOffset], extraParams);
+		}
+	}
+	else {
+		for (Nd4jLong i = tid; i < len; i += gridDim.x * blockDim.x) {
+			auto xOffset = shape::getIndexOffset(i, xShapeInfo, len);
+			auto yOffset = shape::getIndexOffset(i, yShapeInfo, len);
+			auto zOffset = shape::getIndexOffset(i, zShapeInfo, len);
+
+			z[zOffset] = OpType::op(x[xOffset], y[yOffset], extraParams);
+		}
+	}
+}
+
+
+namespace functions           {
+namespace pairwise_transforms {
+
+////////////////////////////////////////////////////////////////////////////////
+template<typename X>
+template<typename OpType>
+void _CUDA_H PairWiseIntTransform<X>::intermediateShaped(dim3& launchDims, cudaStream_t *stream,
+														void *vx, Nd4jLong *xShapeInfo,
+														void *vy, Nd4jLong *yShapeInfo,
+														void *vz, Nd4jLong *zShapeInfo,
+														void *vextraParams){
+
+	pairwiseSimpleShaped<X, OpType><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, vz, zShapeInfo, vextraParams);
+}
+
+
+////////////////////////////////////////////////////////////////////////////////
+template<typename X>
+void PairWiseIntTransform<X>::executeCudaShaped(dim3& launchDims, cudaStream_t *stream, int opNum, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, void *vz, Nd4jLong *zShapeInfo, void *vextraParams) {
+    auto xType = nd4j::DataTypeUtils::fromT<X>();
+
+	DISPATCH_BY_OPNUM_T(intermediateShaped, PARAMS(launchDims, stream, vx, xShapeInfo, vy, yShapeInfo, vz, zShapeInfo, vextraParams), PAIRWISE_INT_OPS);
+}
+
+
+    template<typename X>
+    void PairWiseIntTransform<X>::exec(
+            const int opNum,
+            void *dx,
+            Nd4jLong *xShapeBuffer,
+            void *y,
+            Nd4jLong *yShapeBuffer,
+            void *result,
+            Nd4jLong *resultShapeBuffer,
+            void *extraParams) {
+
+    }
+
+    template<typename X>
+    void PairWiseIntTransform<X>::exec(
+            const int opNum,
+            void *dx,
+            Nd4jLong xStride,
+            void *y,
+            Nd4jLong yStride,
+            void *result,
+            Nd4jLong resultStride,
+            void *extraParams,
+            Nd4jLong n) {
+
+    }
+
+
+    template<typename X>
+    template<typename OpType>
+    void PairWiseIntTransform<X>::exec(
+            void *vx,
+            Nd4jLong* xShapeBuffer,
+            void *vy,
+            Nd4jLong* yShapeBuffer,
+            void *vresult,
+            Nd4jLong* resultShapeBuffer,
+            void *vextraParams) {
+
+    }
+
+    template<typename X>
+    template<typename OpType>
+    void PairWiseIntTransform<X>::exec(void *vx,
+                     Nd4jLong xStride,
+                     void *vy,
+                     Nd4jLong yStride,
+                     void *vresult,
+                     Nd4jLong resultStride,
+                     void *vextraParams,
+                     const Nd4jLong n) {
+
+    }
+
+
+      
+    BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT PairWiseIntTransform, , INTEGER_TYPES);
+}
+}
+
+#endif // PAIRWISE_INT_CU
\ No newline at end of file
diff --git a/libnd4j/include/loops/cuda/scalar_int.cu b/libnd4j/include/loops/cuda/scalar_int.cu
new file mode 100644
index 000000000..48f141525
--- /dev/null
+++ b/libnd4j/include/loops/cuda/scalar_int.cu
@@ -0,0 +1,269 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 08.11.2018
+// @author raver119@gmail.com
+//
+
+#include "../scalar_int.h"
+#include <op_boilerplate.h>
+#include <types/types.h>
+
+#include "../legacy_ops.h"
+
+using namespace simdOps;
+
+////////////////////////////////////////////////////////////////////////
+template <typename X, typename OpType>
+__global__ void scalarAlongDimension(void *x, Nd4jLong *xShapeInfo,
+                                    void *extraParams,
+                                    void *z, Nd4jLong *zShapeInfo,
+                                    void *scalars,
+                                    int *dimension, int dimensionLength,
+                                    Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
+                                    Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+        
+    functions::scalar::ScalarIntTransform<X>::template transformCuda<OpType>(x, xShapeInfo, extraParams, z, zShapeInfo, scalars, dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ, tadOffsetsZ);
+}
+
+
+////////////////////////////////////////////////////////////////////////
+template <typename X, typename OpType>
+__global__ void scalarSimpleShaped(void* x, void *y, Nd4jLong *xShapeInfo, void *params, void *z, Nd4jLong *zShapeInfo, int *allocationBuffer) {
+
+    functions::scalar::ScalarIntTransform<X>::template transformCuda<OpType>(y, x, xShapeInfo, params, z, zShapeInfo, allocationBuffer);
+}
+
+
+
+
+
+// *********************************************************************//
+// *********************************************************************//
+namespace functions {
+namespace scalar    {
+
+////////////////////////////////////////////////////////////////////////
+template<typename X>
+template<typename OpType>
+__device__ void  ScalarIntTransform<X>::transformCuda(void* vscalar, 
+                                                        void *vy, Nd4jLong *yShapeInfo, 
+                                                        void *vparams, 
+                                                        void *vz, Nd4jLong *zShapeInfo, 
+                                                        int *allocationBuffer) {
+    auto scalar = reinterpret_cast<X*>(vscalar)[0];
+    auto y      = reinterpret_cast<X*>(vy);
+    auto params = reinterpret_cast<X*>(vparams);
+    auto z      = reinterpret_cast<X*>(vz);
+
+    auto yRank   = shape::rank(yShapeInfo);
+    auto yEWS    = shape::elementWiseStride(yShapeInfo);
+    auto yShape  = shape::shapeOf(yShapeInfo);
+    auto yStride = shape::stride(yShapeInfo);        
+    
+    auto zRank   = shape::rank(zShapeInfo);
+    auto zEWS    = shape::elementWiseStride(zShapeInfo);
+    auto zShape  = shape::shapeOf(zShapeInfo);
+    auto zStride = shape::stride(zShapeInfo);
+
+    int totalThreads = gridDim.x * blockDim.x;
+    int tid = blockIdx.x * blockDim.x + threadIdx.x;
+
+    __shared__ int len;
+    if(threadIdx.x == 0)
+        len = shape::length(yShapeInfo);
+    __syncthreads();
+
+    if(yEWS >= 1 && zEWS >= 1 && shape::order(yShapeInfo) == shape::order(zShapeInfo)) {
+            transformCuda<OpType>(len, vscalar, vy, yEWS, vparams, vz, zEWS, allocationBuffer);    
+    }
+    else {
+        for (Nd4jLong i = tid; i < len; i+= totalThreads)                        
+            z[shape::getIndexOffset(i, zShapeInfo, len)] = OpType::op(y[shape::getIndexOffset(i, yShapeInfo, len)], scalar, params);
+    }
+}
+
+////////////////////////////////////////////////////////////////////////
+template<typename X>
+template<typename OpType>
+__device__ void  ScalarIntTransform<X>::transformCuda(Nd4jLong len,
+                                                          void* vx, 
+                                                          void *vy, Nd4jLong yEWS, 
+                                                          void *vparams, 
+                                                          void *vz, Nd4jLong zEWS, 
+                                                          int *allocationBuffer) {
+
+    auto x = reinterpret_cast<X*>(vx)[0];
+    auto y = reinterpret_cast<X*>(vy);
+    auto z = reinterpret_cast<X*>(vz);
+    auto params = reinterpret_cast<X*>(vparams);
+
+    int totalThreads = gridDim.x * blockDim.x;
+    int tid = blockIdx.x * blockDim.x + threadIdx.x;
+
+    Nd4jLong i = tid;
+    if(yEWS == 1 && zEWS == 1) {
+        for (; i < len; i += totalThreads)
+            z[i] = OpType::op(y[i], x, params);
+    }
+    else {
+        for (; i < len; i += totalThreads)
+            z[i * zEWS] = OpType::op(y[i * yEWS], x, params);
+    }
+}
+
+
+////////////////////////////////////////////////////////////////////////
+template<typename X>
+template<typename OpType>
+__device__ void  ScalarIntTransform<X>::transformCuda(void *vx, Nd4jLong *xShapeInfo,
+                                                        void *vextraParams, 
+                                                        void *vz, Nd4jLong *zShapeInfo, 
+                                                        void *vscalars, 
+                                                        int *dimension, int dimensionLength, 
+                                                        Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, 
+                                                        Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+    auto x = reinterpret_cast<X*>(vx);
+    auto scalars = reinterpret_cast<X*>(vscalars);
+    auto z = reinterpret_cast<X*>(vz);
+    auto extraParams = reinterpret_cast<X*>(vextraParams);
+    
+    if (tadShapeInfoZ == nullptr) {
+        tadShapeInfoZ = tadShapeInfo;
+        tadOffsetsZ = tadOffsets;
+    }
+
+    // tad preparation
+    auto tadEws = shape::elementWiseStride(tadShapeInfo);
+    auto zEws = shape::elementWiseStride(tadShapeInfoZ);
+    auto tadLength = shape::length(tadShapeInfo);//shape::tadLength(xShapeInfo, dimension, dimensionLength);
+    auto numTads =shape::length(xShapeInfo) / tadLength;
+
+    if (tadEws > 0 && zEws > 0 && shape::order(tadShapeInfo) == shape::order(zShapeInfo)) {
+
+        // main loop, rolling over tads
+        for (int r = blockIdx.x; r < numTads; r += gridDim.x) {
+            X *oZ = z + tadOffsetsZ[r];
+            X *oX = x + tadOffsets[r];
+
+            auto s = scalars[r];
+
+            for (int f = threadIdx.x; f < tadLength; f += blockDim.x)
+                oZ[f * zEws] = OpType::op(oX[f * tadEws], s, extraParams);
+        }
+    } else {
+        // main loop, rolling over tads
+        for (int r = blockIdx.x; r < numTads; r += gridDim.x) {
+            X *oZ = z + tadOffsetsZ[r];
+            X *oX = x + tadOffsets[r];
+
+            auto s = scalars[r];
+
+            for (int f = threadIdx.x; f < tadLength; f += blockDim.x)
+                oZ[shape::getIndexOffset(f, tadShapeInfoZ, tadLength)] = OpType::op(oX[shape::getIndexOffset(f, tadShapeInfo, tadLength)], s, extraParams);
+        }
+    }
+}
+
+
+////////////////////////////////////////////////////////////////////////
+template<typename X>
+template <typename OpType>
+_CUDA_H void ScalarIntTransform<X>::intermediateAlongDimension(dim3& launchDims, cudaStream_t *stream,
+                                                                void *x, Nd4jLong *xShapeInfo, 
+                                                                void *z, Nd4jLong *zShapeInfo, 
+                                                                void *scalars, 
+                                                                void *extraParams, 
+                                                                int *dimension, int dimensionLength, 
+                                                                Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, 
+                                                                Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+
+    scalarAlongDimension<X, OpType><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(x, xShapeInfo, extraParams, z, zShapeInfo, scalars, dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ, tadOffsetsZ);
+}
+
+////////////////////////////////////////////////////////////////////////
+template<typename X>
+template<typename OpType>
+void _CUDA_H ScalarIntTransform<X>::intermediateShaped(dim3& launchDims, cudaStream_t *stream,
+                                                            void *vx, Nd4jLong *xShapeInfo, 
+                                                            void *vz, Nd4jLong *zShapeInfo, 
+                                                            void* vscalar, 
+                                                            void *vextraParams, int *allocPointer){
+    
+    scalarSimpleShaped<X, OpType><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, vscalar, xShapeInfo, vextraParams, vz, zShapeInfo, allocPointer);
+}
+
+////////////////////////////////////////////////////////////////////////
+template<typename X>
+void ScalarIntTransform<X>::executeCudaShaped(dim3& launchDims, cudaStream_t *stream,
+                                                int opNum, 
+                                                void *vx, Nd4jLong *xShapeInfo, 
+                                                void *vz, Nd4jLong *zShapeInfo, 
+                                                void* vscalar, 
+                                                void *vextraParams) {
+
+    if (nd4j::Environment::getInstance()->isDebugAndVerbose())
+        printf("H14 opNum:[%i]\n", opNum);
+
+    DISPATCH_BY_OPNUM_T(intermediateShaped, PARAMS(launchDims, stream, vx, xShapeInfo, vz, zShapeInfo, vscalar, vextraParams, nullptr), SCALAR_INT_OPS);
+}
+
+////////////////////////////////////////////////////////////////////////
+template<typename X>
+void ScalarIntTransform<X>::executeCudaAlongDimension(dim3& launchDims, cudaStream_t *stream, int opNum, void *vx, Nd4jLong *xShapeInfo, void *vz, Nd4jLong *zShapeInfo, void *vscalars, void *vextraParams, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+    DISPATCH_BY_OPNUM_T(intermediateAlongDimension, PARAMS(launchDims, stream, vx, xShapeInfo, vz, zShapeInfo, vscalars, vextraParams, dimension, dimensionLength, tadShapeInfo, tadOffsets, tadShapeInfoZ, tadOffsetsZ), SCALAR_INT_OPS);
+}
+
+    BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT ScalarIntTransform, , INTEGER_TYPES);
+
+
+    template<typename X>
+    template <typename OpType>
+    void ScalarIntTransform<X,>::transform(void *x, Nd4jLong *xShapeInfo, void *extraParams, void *z, Nd4jLong *zShapeInfo, void *scalars, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+
+    }
+
+    template<typename X>
+    void ScalarIntTransform<X>::transform(int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams, void *z, Nd4jLong *zShapeInfo, void *scalars, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ) {
+
+    }
+
+    template<typename X>
+    void ScalarIntTransform<X>::transform(const int opNum, void *x, Nd4jLong *xShapeInfo, void *result, Nd4jLong *resultShapeInfo,  void *scalar,  void *extraParams) {
+
+    }
+
+    template<typename X>
+    void ScalarIntTransform<X>::transform(const int opNum, void *x, Nd4jLong xStride, void *result, Nd4jLong resultStride, void *scalar, void *extraParams, const Nd4jLong n) {
+
+    }
+
+    template<typename X>
+    template<typename OpType>
+    void ScalarIntTransform<X>::transform(void *x, Nd4jLong *xShapeInfo, void *result, Nd4jLong *resultShapeInfo, void *scalar, void *extraParams) {
+
+    }
+
+
+    template<typename X>
+    template<typename OpType>
+    void ScalarIntTransform<X>::transform(void *x, Nd4jLong xStride, void *result, Nd4jLong resultStride, void *scalar, void *extraParams, const Nd4jLong n) {
+
+    }
+}
+}
+
diff --git a/libnd4j/include/loops/legacy_ops.h b/libnd4j/include/loops/legacy_ops.h
index b3096ac0e..b0d891287 100644
--- a/libnd4j/include/loops/legacy_ops.h
+++ b/libnd4j/include/loops/legacy_ops.h
@@ -29,6 +29,16 @@
         (4, aggregateOps::CBOW) ,\
         (5, aggregateOps::GEMM)
 
+#define BROADCAST_INT_OPS \
+        (0, ShiftLeft), \
+        (1, ShiftRight), \
+        (2, CyclicShiftLeft), \
+        (3, CyclicShiftRight), \
+        (4, IntAnd), \
+        (5, IntOr), \
+        (6, IntXor)
+
+
 #define BROADCAST_BOOL_OPS \
         (0, EqualTo),\
         (1, GreaterThan),\
@@ -171,6 +181,14 @@
         (0, SummaryStatsVariance), \
         (1, SummaryStatsStandardDeviation)
 
+#define SCALAR_INT_OPS \
+        (0, ShiftLeft), \
+        (1, ShiftRight), \
+        (2, CyclicShiftLeft), \
+        (3, CyclicShiftRight), \
+        (4, IntAnd), \
+        (5, IntOr), \
+        (6, IntXor)
 
 #define SCALAR_BOOL_OPS \
         (0, EqualTo),\
@@ -300,6 +318,15 @@
         (13, ExponentialDistribution),\
         (14, ExponentialDistributionInv)
 
+#define PAIRWISE_INT_OPS \
+        (0, ShiftLeft), \
+        (1, ShiftRight), \
+        (2, CyclicShiftLeft), \
+        (3, CyclicShiftRight), \
+        (4, IntAnd), \
+        (5, IntOr), \
+        (6, IntXor)
+
 #define PAIRWISE_BOOL_OPS \
         (0, EqualTo),\
         (1, GreaterThan),\
diff --git a/libnd4j/include/loops/pairwise_int.h b/libnd4j/include/loops/pairwise_int.h
new file mode 100644
index 000000000..14d273285
--- /dev/null
+++ b/libnd4j/include/loops/pairwise_int.h
@@ -0,0 +1,119 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+/*
+ * pairwise_transform.h
+ *
+ *  Created on: Dec 28, 2015
+ *      Author: agibsonccc
+ */
+
+#ifndef PAIRWISE_INT_H_
+#define PAIRWISE_INT_H_
+#ifdef _OPENMP
+#include <omp.h>
+#endif
+#include <templatemath.h>
+#include <helpers/shape.h>
+#include <pairwise_util.h>
+#include <dll.h>
+#include <stdio.h>
+#include <ops/ops.h>
+#include <op_boilerplate.h>
+#include <helpers/DebugHelper.h>
+
+#ifdef __CUDACC__
+#include <cuda.h>
+#include <cuda_runtime.h>
+#endif
+
+#ifndef _OPENMP
+#define omp_get_thread_num() 0
+#define omp_get_max_threads() 1
+#endif
+
+
+#include "legacy_ops.h"
+
+using namespace simdOps;
+
+namespace functions {
+    namespace pairwise_transforms {
+
+/**
+ * Transforms involving 2 arrays
+ */
+        template<typename X>
+        class PairWiseIntTransform {
+        public:
+
+#ifdef __CUDACC__
+
+            template <typename OpType>            
+            static __host__ void intermediateShaped(dim3& launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, void *vz, Nd4jLong *zShapeInfo, void *vextraParams);
+
+            static __host__ void executeCudaShaped(dim3& launchDims, cudaStream_t *stream, int opNum, void *x, Nd4jLong *xShapeInfo, void *y, Nd4jLong *yShapeInfo, void *z, Nd4jLong *zShapeInfo, void *extraParams);
+
+
+#endif
+        public:
+
+            static void exec(
+				const int opNum,
+				void *dx,
+				Nd4jLong *xShapeBuffer,
+				void *y,
+				Nd4jLong *yShapeBuffer,
+				void *result,
+				Nd4jLong *resultShapeBuffer,
+				void *extraParams);
+			
+			static void exec(
+				const int opNum,
+				void *dx,
+				Nd4jLong xStride,
+				void *y,
+				Nd4jLong yStride,
+				void *result,
+				Nd4jLong resultStride,
+				void *extraParams,
+				Nd4jLong n);
+
+
+			template<typename OpType>
+			static void exec(
+                    void *vx,
+                    Nd4jLong* xShapeBuffer,
+                    void *vy,
+                    Nd4jLong* yShapeBuffer,
+                    void *vresult,
+                    Nd4jLong* resultShapeBuffer,
+                    void *vextraParams);
+
+            template<typename OpType>
+            static void exec(void *vx,
+                             Nd4jLong xStride,
+                             void *vy,
+                             Nd4jLong yStride,
+                             void *vresult,
+                             Nd4jLong resultStride,
+                             void *vextraParams,
+                             const Nd4jLong n);
+        };
+    }
+}
+
+#endif /* PAIRWISE_TRANSFORM_H_ */
diff --git a/libnd4j/include/loops/scalar_int.h b/libnd4j/include/loops/scalar_int.h
new file mode 100644
index 000000000..f873d5419
--- /dev/null
+++ b/libnd4j/include/loops/scalar_int.h
@@ -0,0 +1,142 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+/*
+ * scalar.h
+ *
+ *  Created on: Dec 28, 2015
+ *      Author: agibsonccc
+ */
+
+#ifndef SCALAR_INT_H_
+#define SCALAR_INT_H_
+#include <dll.h>
+
+#ifdef __JNI__
+#include <jni.h>
+#endif
+#include <templatemath.h>
+#include <ops/ops.h>
+#include <op_boilerplate.h>
+#include "helpers/logger.h"
+#include <OmpLaunchHelper.h>
+#include <helpers/DebugHelper.h>
+
+#ifdef __CUDACC__
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <types/float16.h>
+#endif
+
+#include "legacy_ops.h"
+
+namespace functions {
+    namespace scalar {
+/**
+ * Apply a scalar
+ *  operation to an array
+ */
+        template<typename X>
+        class ScalarIntTransform {
+
+        public:
+
+#ifdef __CUDACC__
+                        
+            template<typename OpType>
+            __device__
+            static void transformCuda(void* scalar, void *vy, Nd4jLong *shapeInfo, void *vparams, void *vresult, Nd4jLong *resultShapeInfo, int *allocationBuffer);
+            
+            template<typename OpType>
+            __device__
+            static void transformCuda(Nd4jLong n, void* vx, void *vy, Nd4jLong yEWS, void *vparams, void *vz, Nd4jLong zEWS, int *allocationBuffer);
+
+            template<typename OpType>
+            __device__
+            static void transformCuda(void *vx, Nd4jLong *xShapeInfo, void *vextraParams, void *vz, Nd4jLong *zShapeInfo, void *vscalars, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+            template <typename OpType>
+            __host__
+            static void intermediateAlongDimension(dim3& launchDims, cudaStream_t *stream, void *x, Nd4jLong *xShapeInfo, void *z, Nd4jLong *zShapeInfo, void *scalars, void *extraParams, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+            template <typename OpType>
+            __host__
+            static void intermediateShaped(dim3& launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vz, Nd4jLong *zShapeInfo, void* vscalar, void *vextraParams, int *allocPointer);            
+            
+            __host__
+            static void executeCudaShaped(dim3& launchDims, cudaStream_t *stream, int opNum, void *x, Nd4jLong *xShapeInfo, void *result, Nd4jLong *resultShapeInfo, void* scalar, void *extraParams);
+
+            __host__
+            static void executeCudaAlongDimension(dim3& launchDims, cudaStream_t *stream, int opNum, void *x, Nd4jLong *xShapeInfo, void *z, Nd4jLong *zShapeInfo, void *scalars, void *extraParams, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+
+/*
+#include "cuda/scalar_temp.cu"
+*/
+#endif
+            template <typename OpType>
+            static void transform(void *x, Nd4jLong *xShapeInfo, void *extraParams, void *z, Nd4jLong *zShapeInfo, void *scalars, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ);
+ 
+           static void transform(int opNum, void *x, Nd4jLong *xShapeInfo, void *extraParams, void *z, Nd4jLong *zShapeInfo, void *scalars, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, Nd4jLong *tadShapeInfoZ, Nd4jLong *tadOffsetsZ);
+
+            static void transform(const int opNum, void *x, Nd4jLong *xShapeInfo, void *result, Nd4jLong *resultShapeInfo,  void *scalar,  void *extraParams);
+
+            static void transform(const int opNum, void *x, Nd4jLong xStride, void *result, Nd4jLong resultStride, void *scalar, void *extraParams, const Nd4jLong n);
+
+
+
+
+            /*
+             * ScalarOp along dimension
+             */
+
+
+            /**
+         * CPU implementation of scalar operation
+         * @param x the input
+         * @param xStride the stride for the input
+         * @param result the result buffer
+         * @param resultStride the stride for the result
+         * @param scalar the scalar to apply
+         * @param extraParams the extra parameters where
+         * neccssary
+         * @param n the number of elements to loop over
+         */
+
+            template<typename OpType>
+            static  void transform(void *x, Nd4jLong *xShapeInfo, void *result, Nd4jLong *resultShapeInfo, void *scalar, void *extraParams);
+
+
+            /**
+             * CPU implementation of scalar operation
+             * @param x the input
+             * @param xStride the stride for the input
+             * @param result the result buffer
+             * @param resultStride the stride for the result
+             * @param scalar the scalar to apply
+             * @param extraParams the extra parameters where
+             * neccssary
+             * @param n the number of elements to loop over
+             */
+
+            template<typename OpType>
+            static void transform(void *x, Nd4jLong xStride, void *result, Nd4jLong resultStride, void *scalar, void *extraParams, const Nd4jLong n);
+        };
+    }
+}
+
+
+#endif /* SCALAR_H_ */
diff --git a/libnd4j/include/op_enums.h b/libnd4j/include/op_enums.h
index bf9d4c72f..8a100153f 100644
--- a/libnd4j/include/op_enums.h
+++ b/libnd4j/include/op_enums.h
@@ -63,6 +63,10 @@ namespace nd4j {
         enum BoolOps {
             BUILD_ENUMERATION(PAIRWISE_BOOL_OPS)
         };
+
+        enum IntOps {
+            BUILD_ENUMERATION(PAIRWISE_INT_OPS)
+        };
     }
 
     namespace scalar {
@@ -73,6 +77,10 @@ namespace nd4j {
         enum BoolOps {
             BUILD_ENUMERATION(SCALAR_BOOL_OPS)
         };
+
+        enum IntOps {
+            BUILD_ENUMERATION(SCALAR_INT_OPS)
+        };
     }
 
     namespace reduce {
@@ -113,6 +121,10 @@ namespace nd4j {
         enum BoolOps {
             BUILD_ENUMERATION(BROADCAST_BOOL_OPS)
         };
+
+        enum IntOps {
+            BUILD_ENUMERATION(BROADCAST_INT_OPS)
+        };
     }
 
     namespace variance {
diff --git a/libnd4j/include/ops/BroadcastIntOpsTuple.h b/libnd4j/include/ops/BroadcastIntOpsTuple.h
new file mode 100644
index 000000000..df40907a9
--- /dev/null
+++ b/libnd4j/include/ops/BroadcastIntOpsTuple.h
@@ -0,0 +1,49 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+
+#ifndef DEV_TESTS_BROADCASTINTOPSTUPLE_H
+#define DEV_TESTS_BROADCASTINTOPSTUPLE_H
+
+#include <op_enums.h>
+
+namespace nd4j {
+    class BroadcastIntOpsTuple {
+    private:
+
+    public:
+        nd4j::scalar::IntOps s;
+        nd4j::pairwise::IntOps p;
+        nd4j::broadcast::IntOps b;
+
+        BroadcastIntOpsTuple() = default;
+        ~BroadcastIntOpsTuple() = default;
+
+        BroadcastIntOpsTuple(nd4j::scalar::IntOps scalar, nd4j::pairwise::IntOps pairwise, nd4j::broadcast::IntOps broadcast) {
+            s = scalar;
+            p = pairwise;
+            b = broadcast;
+        }
+
+        static BroadcastIntOpsTuple custom(nd4j::scalar::IntOps scalar, nd4j::pairwise::IntOps pairwise, nd4j::broadcast::IntOps broadcast);
+    };
+}
+
+
+#endif //DEV_TESTS_BROADCASTOPSTUPLE_H
diff --git a/libnd4j/include/ops/declarable/generic/bitwise/cyclic_rshift.cpp b/libnd4j/include/ops/declarable/generic/bitwise/cyclic_rshift.cpp
index 2aac5c6f9..89d380d02 100644
--- a/libnd4j/include/ops/declarable/generic/bitwise/cyclic_rshift.cpp
+++ b/libnd4j/include/ops/declarable/generic/bitwise/cyclic_rshift.cpp
@@ -27,22 +27,14 @@
 
 namespace nd4j {
     namespace ops {
-        CONFIGURABLE_OP_IMPL(cyclic_rshift_bits, 1, 1, true, 0, -2) {
-            auto input = INPUT_VARIABLE(0);
-            auto output = OUTPUT_VARIABLE(0);
+        BROADCASTABLE_OP_IMPL(cyclic_rshift_bits, 0, 0) {
+            auto x = INPUT_VARIABLE(0);
+            auto y = INPUT_VARIABLE(1);
+            auto z = OUTPUT_VARIABLE(0);
 
-            REQUIRE_TRUE(block.numI() > 0 || block.width() > 1, 0, "cyclic_rshift_bits: actual shift value is missing");
+            BROADCAST_CHECK_EMPTY(x,y,z);
 
-            uint32_t shift = 0;
-            if (block.width() > 1) {
-                shift = INPUT_VARIABLE(1)->e<uint32_t>(0);
-            } else if (block.numI() > 0) {
-                shift = INT_ARG(0);
-            };
-
-            helpers::cyclic_rshift_bits(block.launchContext(), *input, *output, shift);
-
-            REQUIRE_TRUE(shift > 0 && shift < input->sizeOfT() * 8, 0, "cyclic_rshift_bits: can't shift beyond size of data type")
+            x->applyTrueBroadcast(BroadcastIntOpsTuple::custom(scalar::CyclicShiftRight, pairwise::CyclicShiftRight, broadcast::CyclicShiftRight), y, z, false);
 
             return Status::OK();
         }
diff --git a/libnd4j/include/ops/declarable/generic/bitwise/cyclic_shift.cpp b/libnd4j/include/ops/declarable/generic/bitwise/cyclic_shift.cpp
index 0bdb9503d..f18314910 100644
--- a/libnd4j/include/ops/declarable/generic/bitwise/cyclic_shift.cpp
+++ b/libnd4j/include/ops/declarable/generic/bitwise/cyclic_shift.cpp
@@ -27,22 +27,14 @@
 
 namespace nd4j {
     namespace ops {
-        CONFIGURABLE_OP_IMPL(cyclic_shift_bits, 1, 1, true, 0, -2) {
-            auto input = INPUT_VARIABLE(0);
-            auto output = OUTPUT_VARIABLE(0);
+        BROADCASTABLE_OP_IMPL(cyclic_shift_bits, 0, 0) {
+            auto x = INPUT_VARIABLE(0);
+            auto y = INPUT_VARIABLE(1);
+            auto z = OUTPUT_VARIABLE(0);
 
-            REQUIRE_TRUE(block.numI() > 0 || block.width() > 1, 0, "cyclic_shift_bits: actual shift value is missing");
+            BROADCAST_CHECK_EMPTY(x,y,z);
 
-            uint32_t shift = 0;
-            if (block.width() > 1) {
-                shift = INPUT_VARIABLE(1)->e<uint32_t>(0);
-            } else if (block.numI() > 0) {
-                shift = INT_ARG(0);
-            };
-
-            helpers::cyclic_shift_bits(block.launchContext(), *input, *output, shift);
-
-            REQUIRE_TRUE(shift > 0 && shift < input->sizeOfT() * 8, 0, "cyclic_shift_bits: can't shift beyond size of data type")
+            x->applyTrueBroadcast(BroadcastIntOpsTuple::custom(scalar::CyclicShiftLeft, pairwise::CyclicShiftLeft, broadcast::CyclicShiftLeft), y, z, false);
 
             return Status::OK();
         }
diff --git a/libnd4j/include/ops/declarable/generic/bitwise/rshift.cpp b/libnd4j/include/ops/declarable/generic/bitwise/rshift.cpp
index 4068351a2..36b0defd0 100644
--- a/libnd4j/include/ops/declarable/generic/bitwise/rshift.cpp
+++ b/libnd4j/include/ops/declarable/generic/bitwise/rshift.cpp
@@ -27,22 +27,14 @@
 
 namespace nd4j {
     namespace ops {
-        CONFIGURABLE_OP_IMPL(rshift_bits, 1, 1, true, 0, -2) {
-            auto input = INPUT_VARIABLE(0);
-            auto output = OUTPUT_VARIABLE(0);
+        BROADCASTABLE_OP_IMPL(rshift_bits, 0, 0) {
+            auto x = INPUT_VARIABLE(0);
+            auto y = INPUT_VARIABLE(1);
+            auto z = OUTPUT_VARIABLE(0);
 
-            REQUIRE_TRUE(block.numI() > 0 || block.width() > 1, 0, "rshift_bits: actual shift value is missing");
+            BROADCAST_CHECK_EMPTY(x,y,z);
 
-            uint32_t shift = 0;
-            if (block.width() > 1) {
-                shift = INPUT_VARIABLE(1)->e<uint32_t>(0);
-            } else if (block.numI() > 0) {
-                shift = INT_ARG(0);
-            };
-
-            REQUIRE_TRUE(shift > 0 && shift < input->sizeOfT() * 8, 0, "rshift_bits: can't shift beyond size of data type")
-
-            helpers::rshift_bits(block.launchContext(), *input, *output, shift);
+            x->applyTrueBroadcast(BroadcastIntOpsTuple::custom(scalar::ShiftRight, pairwise::ShiftRight, broadcast::ShiftRight), y, z, false);
 
             return Status::OK();
         }
diff --git a/libnd4j/include/ops/declarable/generic/bitwise/shift.cpp b/libnd4j/include/ops/declarable/generic/bitwise/shift.cpp
index f79da1024..ab4ed9880 100644
--- a/libnd4j/include/ops/declarable/generic/bitwise/shift.cpp
+++ b/libnd4j/include/ops/declarable/generic/bitwise/shift.cpp
@@ -27,22 +27,14 @@
 
 namespace nd4j {
     namespace ops {
-        CONFIGURABLE_OP_IMPL(shift_bits, 1, 1, true, 0, -2) {
-            auto input = INPUT_VARIABLE(0);
-            auto output = OUTPUT_VARIABLE(0);
+        BROADCASTABLE_OP_IMPL(shift_bits, 0, 0) {
+            auto x = INPUT_VARIABLE(0);
+            auto y = INPUT_VARIABLE(1);
+            auto z = OUTPUT_VARIABLE(0);
 
-            REQUIRE_TRUE(block.numI() > 0 || block.width() > 1, 0, "shift_bits: actual shift value is missing");
+            BROADCAST_CHECK_EMPTY(x,y,z);
 
-            uint32_t shift = 0;
-            if (block.width() > 1) {
-                shift = INPUT_VARIABLE(1)->e<uint32_t>(0);
-            } else if (block.numI() > 0) {
-                shift = INT_ARG(0);
-            };
-
-            REQUIRE_TRUE(shift > 0 && shift < input->sizeOfT() * 8, 0, "shift_bits: can't shift beyond size of data type")
-
-            helpers::shift_bits(block.launchContext(), *input, *output, shift);
+            x->applyTrueBroadcast(BroadcastIntOpsTuple::custom(scalar::ShiftLeft, pairwise::ShiftLeft, broadcast::ShiftLeft), y, z, false);
 
             return Status::OK();
         }
diff --git a/libnd4j/include/ops/declarable/headers/bitwise.h b/libnd4j/include/ops/declarable/headers/bitwise.h
index da5ff4047..a6362a73f 100644
--- a/libnd4j/include/ops/declarable/headers/bitwise.h
+++ b/libnd4j/include/ops/declarable/headers/bitwise.h
@@ -45,7 +45,7 @@ namespace nd4j {
          * @tparam T
          */
         #if NOT_EXCLUDED(OP_shift_bits)
-        DECLARE_CONFIGURABLE_OP(shift_bits, 1, 1, true, 0, -2);
+        DECLARE_BROADCASTABLE_OP(shift_bits,  0, 0);
         #endif
 
         /**
@@ -56,7 +56,7 @@ namespace nd4j {
          * @tparam T
          */
         #if NOT_EXCLUDED(OP_rshift_bits)
-        DECLARE_CONFIGURABLE_OP(rshift_bits, 1, 1, true, 0, -2);
+        DECLARE_BROADCASTABLE_OP(rshift_bits,  0, 0);
         #endif
 
         /**
@@ -67,7 +67,7 @@ namespace nd4j {
          * @tparam T
          */
         #if NOT_EXCLUDED(OP_cyclic_shift_bits)
-        DECLARE_CONFIGURABLE_OP(cyclic_shift_bits, 1, 1, true, 0, -2);
+        DECLARE_BROADCASTABLE_OP(cyclic_shift_bits,  0, 0);
         #endif
 
         /**
@@ -78,7 +78,7 @@ namespace nd4j {
          * @tparam T
          */
         #if NOT_EXCLUDED(OP_cyclic_rshift_bits)
-        DECLARE_CONFIGURABLE_OP(cyclic_rshift_bits, 1, 1, true, 0, -2);
+        DECLARE_BROADCASTABLE_OP(cyclic_rshift_bits,  0, 0);
         #endif
 
         /**
diff --git a/libnd4j/include/ops/impl/BroadcastIntOpsTuple.cpp b/libnd4j/include/ops/impl/BroadcastIntOpsTuple.cpp
new file mode 100644
index 000000000..607572b59
--- /dev/null
+++ b/libnd4j/include/ops/impl/BroadcastIntOpsTuple.cpp
@@ -0,0 +1,27 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author raver119@gmail.com
+//
+#include <ops/BroadcastIntOpsTuple.h>
+
+namespace nd4j {
+    BroadcastIntOpsTuple BroadcastIntOpsTuple::custom(nd4j::scalar::IntOps scalar, nd4j::pairwise::IntOps pairwise, nd4j::broadcast::IntOps broadcast) {
+        BroadcastIntOpsTuple t(scalar, pairwise, broadcast);
+        return t;
+    }
+}
diff --git a/libnd4j/include/ops/ops.h b/libnd4j/include/ops/ops.h
index fe6bfae81..e4fef2c3c 100644
--- a/libnd4j/include/ops/ops.h
+++ b/libnd4j/include/ops/ops.h
@@ -660,6 +660,98 @@ namespace simdOps {
 		}
 	};
 
+    template <typename X>
+    class IntOr {
+    public:
+
+        op_def static X op(X d1, X d2) {
+            return d2 | d1;
+        }
+
+        op_def static X op(X d1, X d2, X *params) {
+            return op(d1, d2);
+        }
+    };
+
+    template <typename X>
+    class IntAnd {
+    public:
+
+        op_def static X op(X d1, X d2) {
+            return d2 & d1;
+        }
+
+        op_def static X op(X d1, X d2, X *params) {
+            return op(d1, d2);
+        }
+    };
+
+    template <typename X>
+    class IntXor {
+    public:
+
+        op_def static X op(X d1, X d2) {
+            return d2 ^ d1;
+        }
+
+        op_def static X op(X d1, X d2, X *params) {
+            return op(d1, d2);
+        }
+    };
+
+    template <typename X>
+    class ShiftLeft {
+    public:
+
+        op_def static X op(X d1, X d2) {
+            return d1 << d2;
+        }
+
+        op_def static X op(X d1, X d2, X *params) {
+            return op(d1, d2);
+        }
+    };
+
+    template <typename X>
+    class ShiftRight {
+    public:
+
+        op_def static X op(X d1, X d2) {
+            return d1 >> d2;
+        }
+
+        op_def static X op(X d1, X d2, X *params) {
+            return op(d1, d2);
+        }
+    };
+
+    template <typename X>
+    class CyclicShiftLeft {
+    public:
+
+        op_def static X op(X d1, X d2) {
+            return d1 << d2 | d1 >> ((sizeof(X) * 8) - d2);
+        }
+
+        op_def static X op(X d1, X d2, X *params) {
+            return op(d1, d2);
+        }
+    };
+
+    template <typename X>
+    class CyclicShiftRight {
+    public:
+
+        op_def static X op(X d1, X d2) {
+            return d1 >> d2 | d1 << ((sizeof(X) * 8) - d2);
+        }
+
+        op_def static X op(X d1, X d2, X *params) {
+            return op(d1, d2);
+        }
+    };
+
+
 	template <typename X, typename Z>
 	class Or {
 	public:
diff --git a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests13.cpp b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests13.cpp
index 8c484268e..87ac417be 100644
--- a/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests13.cpp
+++ b/libnd4j/tests_cpu/layers_tests/DeclarableOpsTests13.cpp
@@ -623,12 +623,13 @@ TEST_F(DeclarableOpsTests13, adjustSaturation_5) {
 
 TEST_F(DeclarableOpsTests13, shift_bits_1) {
     auto x = NDArrayFactory::create<int>('c', {5});
+    auto y = NDArrayFactory::create<int>(4);
     auto e = x.ulike();
     x.assign(32);
     e.assign(512);
 
     nd4j::ops::shift_bits op;
-    auto result = op.execute({&x}, {}, {4});
+    auto result = op.execute({&x, &y}, {}, {});
     ASSERT_EQ(Status::OK(), result->status());
 
     auto z = result->at(0);
@@ -640,12 +641,13 @@ TEST_F(DeclarableOpsTests13, shift_bits_1) {
 
 TEST_F(DeclarableOpsTests13, rshift_bits_1) {
     auto x = NDArrayFactory::create<int>('c', {5});
+    auto y = NDArrayFactory::create<int>(4);
     auto e = x.ulike();
     x.assign(512);
     e.assign(32);
 
     nd4j::ops::rshift_bits op;
-    auto result = op.execute({&x}, {}, {4});
+    auto result = op.execute({&x, &y}, {}, {});
     ASSERT_EQ(Status::OK(), result->status());
 
     auto z = result->at(0);
@@ -657,12 +659,13 @@ TEST_F(DeclarableOpsTests13, rshift_bits_1) {
 
 TEST_F(DeclarableOpsTests13, cyclic_shift_bits_1) {
     auto x = NDArrayFactory::create<int>('c', {5});
+    auto y = NDArrayFactory::create<int>(4);
     auto e = x.ulike();
     x.assign(32);
     e.assign(512);
 
     nd4j::ops::cyclic_shift_bits op;
-    auto result = op.execute({&x}, {}, {4});
+    auto result = op.execute({&x, &y}, {}, {});
     ASSERT_EQ(Status::OK(), result->status());
 
     auto z = result->at(0);
@@ -674,12 +677,107 @@ TEST_F(DeclarableOpsTests13, cyclic_shift_bits_1) {
 
 TEST_F(DeclarableOpsTests13, cyclic_rshift_bits_1) {
     auto x = NDArrayFactory::create<int>('c', {5});
+    auto y = NDArrayFactory::create<int>(4);
     auto e = x.ulike();
     x.assign(512);
     e.assign(32);
 
     nd4j::ops::cyclic_rshift_bits op;
-    auto result = op.execute({&x}, {}, {4});
+    auto result = op.execute({&x, &y}, {}, {});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    auto z = result->at(0);
+
+    ASSERT_EQ(e, *z);
+
+    delete result;
+}
+
+TEST_F(DeclarableOpsTests13, shift_bits_2) {
+    auto x = NDArrayFactory::create<int>('c', {5});
+    auto y = NDArrayFactory::create<int>('c', {5});
+    auto e = x.ulike();
+    x.assign(32);
+    y.assign(4);
+    e.assign(512);
+
+    nd4j::ops::shift_bits op;
+    auto result = op.execute({&x, &y}, {}, {});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    auto z = result->at(0);
+
+    ASSERT_EQ(e, *z);
+
+    delete result;
+}
+
+TEST_F(DeclarableOpsTests13, rshift_bits_2) {
+    auto x = NDArrayFactory::create<int>('c', {5});
+    auto y = NDArrayFactory::create<int>('c', {5});
+    auto e = x.ulike();
+    x.assign(512);
+    y.assign(4);
+    e.assign(32);
+
+    nd4j::ops::rshift_bits op;
+    auto result = op.execute({&x, &y}, {}, {});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    auto z = result->at(0);
+
+    ASSERT_EQ(e, *z);
+
+    delete result;
+}
+
+TEST_F(DeclarableOpsTests13, cyclic_shift_bits_2) {
+    auto x = NDArrayFactory::create<int>('c', {5});
+    auto y = NDArrayFactory::create<int>('c', {5});
+    auto e = x.ulike();
+    x.assign(32);
+    y.assign(4);
+    e.assign(512);
+
+    nd4j::ops::cyclic_shift_bits op;
+    auto result = op.execute({&x, &y}, {}, {});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    auto z = result->at(0);
+
+    ASSERT_EQ(e, *z);
+
+    delete result;
+}
+
+TEST_F(DeclarableOpsTests13, cyclic_rshift_bits_2) {
+    auto x = NDArrayFactory::create<int>('c', {5});
+    auto y = NDArrayFactory::create<int>('c', {5});
+    auto e = x.ulike();
+    x.assign(512);
+    y.assign(4);
+    e.assign(32);
+
+    nd4j::ops::cyclic_rshift_bits op;
+    auto result = op.execute({&x, &y}, {}, {});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    auto z = result->at(0);
+
+    ASSERT_EQ(e, *z);
+
+    delete result;
+}
+TEST_F(DeclarableOpsTests13, shift_bits_3) {
+    auto x = NDArrayFactory::create<int>('c', {5, 5});
+    auto y = NDArrayFactory::create<int>('c', {1, 5});
+    auto e = x.ulike();
+    x.assign(32);
+    y.assign(4);
+    e.assign(512);
+
+    nd4j::ops::shift_bits op;
+    auto result = op.execute({&x, &y}, {}, {});
     ASSERT_EQ(Status::OK(), result->status());
 
     auto z = result->at(0);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
index 1f361ce9c..035ff0960 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
@@ -1234,19 +1234,19 @@ public class DifferentialFunctionFactory {
         return new Xor(sameDiff(), ix, iy).outputVariable();
     }
 
-    public SDVariable shift(SDVariable ix, int shift) {
+    public SDVariable shift(SDVariable ix, SDVariable shift) {
         return new ShiftBits(sameDiff(), ix, shift).outputVariable();
     }
 
-    public SDVariable rshift(SDVariable ix, int shift) {
+    public SDVariable rshift(SDVariable ix, SDVariable shift) {
         return new RShiftBits(sameDiff(), ix, shift).outputVariable();
     }
 
-    public SDVariable rotl(SDVariable ix, int shift) {
+    public SDVariable rotl(SDVariable ix, SDVariable shift) {
         return new CyclicShiftBits(sameDiff(), ix, shift).outputVariable();
     }
 
-    public SDVariable rotr(SDVariable ix, int shift) {
+    public SDVariable rotr(SDVariable ix, SDVariable shift) {
         return new CyclicRShiftBits(sameDiff(), ix, shift).outputVariable();
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDMath.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDMath.java
index 70eaa5cd9..10fc0b44a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDMath.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDMath.java
@@ -2428,7 +2428,7 @@ public class SDMath extends SDOps {
      * @param x    Input 1
      * @return Output SDVariable with shifted bits
      */
-    public SDVariable bitShift(String name, SDVariable x, int shift) {
+    public SDVariable bitShift(String name, SDVariable x, SDVariable shift) {
         validateInteger("shift_bits", x);
         SDVariable result = f().shift(x, shift);
         return updateVariableNameAndReference(result, name);
@@ -2441,7 +2441,7 @@ public class SDMath extends SDOps {
      * @param x    Input 1
      * @return Output SDVariable with shifted bits
      */
-    public SDVariable bitShiftRight(String name, SDVariable x, int shift) {
+    public SDVariable bitShiftRight(String name, SDVariable x, SDVariable shift) {
         validateInteger("rshift_bits", x);
         SDVariable result = f().rshift(x, shift);
         return updateVariableNameAndReference(result, name);
@@ -2454,7 +2454,7 @@ public class SDMath extends SDOps {
      * @param x    Input 1
      * @return Output SDVariable with shifted bits
      */
-    public SDVariable bitRotl(String name, SDVariable x, int shift) {
+    public SDVariable bitRotl(String name, SDVariable x, SDVariable shift) {
         validateInteger("cyclic_shift_bits", x);
         SDVariable result = f().rotl(x, shift);
         return updateVariableNameAndReference(result, name);
@@ -2467,7 +2467,7 @@ public class SDMath extends SDOps {
      * @param x    Input 1
      * @return Output SDVariable with shifted bits
      */
-    public SDVariable bitRotr(String name, SDVariable x, int shift) {
+    public SDVariable bitRotr(String name, SDVariable x, SDVariable shift) {
         validateInteger("cyclic_rshift_bits", x);
         SDVariable result = f().rotr(x, shift);
         return updateVariableNameAndReference(result, name);
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CyclicRShiftBits.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CyclicRShiftBits.java
index 318a7dc02..3a9173654 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CyclicRShiftBits.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CyclicRShiftBits.java
@@ -34,18 +34,16 @@ import java.util.List;
  */
 public class CyclicRShiftBits extends BaseDynamicTransformOp {
 
-    public CyclicRShiftBits(SameDiff sameDiff, SDVariable x, int shift) {
-        super(sameDiff, new SDVariable[] {x} ,false);
-        this.addIArgument(shift);
+    public CyclicRShiftBits(SameDiff sameDiff, SDVariable x, SDVariable shift) {
+        super(sameDiff, new SDVariable[] {x, shift} ,false);
     }
 
-    public CyclicRShiftBits(INDArray input, int shift, INDArray output) {
-        super(new INDArray[]{input}, new INDArray[]{output});
-        this.addIArgument(shift);
+    public CyclicRShiftBits(INDArray input, INDArray shift, INDArray output) {
+        super(new INDArray[]{input, shift}, new INDArray[]{output});
     }
 
-    public CyclicRShiftBits(INDArray input, int shift) {
-        this(input, shift,null);
+    public CyclicRShiftBits(INDArray input, INDArray shift) {
+        this(input, shift,input.ulike());
     }
 
     public CyclicRShiftBits() {}
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CyclicShiftBits.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CyclicShiftBits.java
index b4291c5df..20b6f6955 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CyclicShiftBits.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/CyclicShiftBits.java
@@ -34,18 +34,16 @@ import java.util.List;
  */
 public class CyclicShiftBits extends BaseDynamicTransformOp {
 
-    public CyclicShiftBits(SameDiff sameDiff, SDVariable x, int shift) {
-        super(sameDiff, new SDVariable[] {x} ,false);
-        this.addIArgument(shift);
+    public CyclicShiftBits(SameDiff sameDiff, SDVariable x, SDVariable shift) {
+        super(sameDiff, new SDVariable[] {x, shift} ,false);
     }
 
-    public CyclicShiftBits(INDArray input, int shift, INDArray output) {
-        super(new INDArray[]{input}, new INDArray[]{output});
-        this.addIArgument(shift);
+    public CyclicShiftBits(INDArray input, INDArray shift, INDArray output) {
+        super(new INDArray[]{input, shift}, new INDArray[]{output});
     }
 
-    public CyclicShiftBits(INDArray input, int shift) {
-        this(input, shift,null);
+    public CyclicShiftBits(INDArray input, INDArray shift) {
+        this(input, shift,input.ulike());
     }
 
     public CyclicShiftBits() {}
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/RShiftBits.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/RShiftBits.java
index 80697efa3..4435615f5 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/RShiftBits.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/RShiftBits.java
@@ -34,18 +34,16 @@ import java.util.List;
  */
 public class RShiftBits extends BaseDynamicTransformOp {
 
-    public RShiftBits(SameDiff sameDiff, SDVariable x, int shift) {
-        super(sameDiff, new SDVariable[] {x} ,false);
-        this.addIArgument(shift);
+    public RShiftBits(SameDiff sameDiff, SDVariable x, SDVariable y) {
+        super(sameDiff, new SDVariable[] {x, y} ,false);
     }
 
-    public RShiftBits(INDArray input, int shift, INDArray output) {
-        super(new INDArray[]{input}, new INDArray[]{output});
-        this.addIArgument(shift);
+    public RShiftBits(INDArray input, INDArray shift, INDArray output) {
+        super(new INDArray[]{input, shift}, new INDArray[]{output});
     }
 
-    public RShiftBits(INDArray input, int shift) {
-        this(input, shift,null);
+    public RShiftBits(INDArray input, INDArray shift) {
+        this(input, shift,input.ulike());
     }
 
     public RShiftBits() {}
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/ShiftBits.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/ShiftBits.java
index 8c652f72d..5501324f2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/ShiftBits.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/custom/ShiftBits.java
@@ -34,18 +34,16 @@ import java.util.List;
  */
 public class ShiftBits extends BaseDynamicTransformOp {
 
-    public ShiftBits(SameDiff sameDiff, SDVariable x, int shift) {
-        super(sameDiff, new SDVariable[] {x} ,false);
-        this.addIArgument(shift);
+    public ShiftBits(SameDiff sameDiff, SDVariable x, SDVariable y) {
+        super(sameDiff, new SDVariable[] {x, y} ,false);
     }
 
-    public ShiftBits(INDArray input, int shift,  INDArray output) {
-        super(new INDArray[]{input}, new INDArray[]{output});
-        this.addIArgument(shift);
+    public ShiftBits(INDArray x, INDArray y,  INDArray output) {
+        super(new INDArray[]{x, y}, new INDArray[]{output});
     }
 
-    public ShiftBits(INDArray input, int shift) {
-        this(input, shift,null);
+    public ShiftBits(INDArray x, INDArray y) {
+        this(x, y,x.ulike());
     }
 
     public ShiftBits() {}
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java
index 603413fd6..15f6c52ef 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java
@@ -3594,6 +3594,7 @@ public native @Cast("Nd4jPointer") Pointer lcSolverHandle(OpaqueLaunchContext lc
 // #include <op_enums.h>
 // #include <ops/BroadcastOpsTuple.h>
 // #include <ops/BroadcastBoolOpsTuple.h>
+// #include <ops/BroadcastIntOpsTuple.h>
 // #include <array/ExtraArguments.h>
 // #include <Status.h>
 // #include <ShapeDescriptor.h>
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java
index 38c0cb8c4..eeb4d38c3 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java
@@ -3594,6 +3594,7 @@ public native @Cast("Nd4jPointer") Pointer lcSolverHandle(OpaqueLaunchContext lc
 // #include <op_enums.h>
 // #include <ops/BroadcastOpsTuple.h>
 // #include <ops/BroadcastBoolOpsTuple.h>
+// #include <ops/BroadcastIntOpsTuple.h>
 // #include <array/ExtraArguments.h>
 // #include <Status.h>
 // #include <ShapeDescriptor.h>
@@ -16496,15 +16497,18 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
 
         /**
          * creates identity 2D matrix or batch of identical 2D identity matrices
-         * 
+         *
          * Input array:
          * provide some array - in any case operation simply neglects it
-         * 
+         *
+         * Input float argument (if passed):
+         * TArgs[0] - type of elements of output array, default value is 5 (float)
+         *
          * Input integer arguments:
          * IArgs[0]       - order of output identity matrix, 99 -> 'c'-order, 102 -> 'f'-order
          * IArgs[1]       - the number of rows in output inner-most 2D identity matrix
          * IArgs[2]       - optional, the number of columns in output inner-most 2D identity matrix, if this argument is not provided then it is taken to be equal to number of rows
-         * IArgs[3,4,...] - optional, shape of batch, output matrix will have leading batch dimensions of this shape         
+         * IArgs[3,4,...] - optional, shape of batch, output matrix will have leading batch dimensions of this shape
          */
 //         #if NOT_EXCLUDED(OP_eye)
         @Namespace("nd4j::ops") public static class eye extends DeclarableCustomOp {
@@ -16598,10 +16602,10 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
 
         /**
          * clip a list of given tensors with given average norm when needed
-         * 
+         *
          * Input:
          *    a list of tensors (at least one)
-         * 
+         *
          * Input floating point argument:
          *    clip_norm - a value that used as threshold value and norm to be used
          *
@@ -16749,12 +16753,12 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
 
         /**
          * returns histogram (as 1D array) with fixed bins width
-         * 
+         *
          * Input arrays:
-         * - input array with elements to be binned into output histogram 
+         * - input array with elements to be binned into output histogram
          * - range array with first element being bottom limit and second element being top limit of histogram,
              please note that input_value <= range[0] will be mapped to histogram[0], input_value >= range[1] will be mapped to histogram[-1]
-         * 
+         *
          * Input integer arguments:
          *    nbins (optional) - number of histogram bins, default value is 100
          */
@@ -21822,7 +21826,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
          * \tparam T
          */
 //         #if NOT_EXCLUDED(OP_shift_bits)
-        @Namespace("nd4j::ops") public static class shift_bits extends DeclarableOp {
+        @Namespace("nd4j::ops") public static class shift_bits extends BroadcastableOp {
             static { Loader.load(); }
             /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */
             public shift_bits(Pointer p) { super(p); }
@@ -21835,7 +21839,6 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
         
                                                                                     public shift_bits() { super((Pointer)null); allocate(); }
                                                                                     private native void allocate();
-                                                                                    public native ShapeList calculateOutputShape(ShapeList inputShape, @ByRef Context block);
                                                                                 }
 //         #endif
 
@@ -21847,7 +21850,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
          * \tparam T
          */
 //         #if NOT_EXCLUDED(OP_rshift_bits)
-        @Namespace("nd4j::ops") public static class rshift_bits extends DeclarableOp {
+        @Namespace("nd4j::ops") public static class rshift_bits extends BroadcastableOp {
             static { Loader.load(); }
             /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */
             public rshift_bits(Pointer p) { super(p); }
@@ -21860,7 +21863,6 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
         
                                                                                     public rshift_bits() { super((Pointer)null); allocate(); }
                                                                                     private native void allocate();
-                                                                                    public native ShapeList calculateOutputShape(ShapeList inputShape, @ByRef Context block);
                                                                                 }
 //         #endif
 
@@ -21872,7 +21874,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
          * \tparam T
          */
 //         #if NOT_EXCLUDED(OP_cyclic_shift_bits)
-        @Namespace("nd4j::ops") public static class cyclic_shift_bits extends DeclarableOp {
+        @Namespace("nd4j::ops") public static class cyclic_shift_bits extends BroadcastableOp {
             static { Loader.load(); }
             /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */
             public cyclic_shift_bits(Pointer p) { super(p); }
@@ -21885,7 +21887,6 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
         
                                                                                     public cyclic_shift_bits() { super((Pointer)null); allocate(); }
                                                                                     private native void allocate();
-                                                                                    public native ShapeList calculateOutputShape(ShapeList inputShape, @ByRef Context block);
                                                                                 }
 //         #endif
 
@@ -21897,7 +21898,7 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
          * \tparam T
          */
 //         #if NOT_EXCLUDED(OP_cyclic_rshift_bits)
-        @Namespace("nd4j::ops") public static class cyclic_rshift_bits extends DeclarableOp {
+        @Namespace("nd4j::ops") public static class cyclic_rshift_bits extends BroadcastableOp {
             static { Loader.load(); }
             /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */
             public cyclic_rshift_bits(Pointer p) { super(p); }
@@ -21910,6 +21911,30 @@ public static final int TAD_THRESHOLD = TAD_THRESHOLD();
         
                                                                                     public cyclic_rshift_bits() { super((Pointer)null); allocate(); }
                                                                                     private native void allocate();
+                                                                                }
+//         #endif
+
+        /**
+         * This operation returns hamming distance based on bits
+         *
+         * PLEASE NOTE: This operation is applicable only to integer data types
+         *
+         * \tparam T
+         */
+//         #if NOT_EXCLUDED(OP_bits_hamming_distance)
+        @Namespace("nd4j::ops") public static class bits_hamming_distance extends DeclarableCustomOp {
+            static { Loader.load(); }
+            /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */
+            public bits_hamming_distance(Pointer p) { super(p); }
+            /** Native array allocator. Access with {@link Pointer#position(long)}. */
+            public bits_hamming_distance(long size) { super((Pointer)null); allocateArray(size); }
+            private native void allocateArray(long size);
+            @Override public bits_hamming_distance position(long position) {
+                return (bits_hamming_distance)super.position(position);
+            }
+        
+                                                                                    public bits_hamming_distance() { super((Pointer)null); allocate(); }
+                                                                                    private native void allocate();
                                                                                     public native ShapeList calculateOutputShape(ShapeList inputShape, @ByRef Context block);
                                                                                 }
 //         #endif
diff --git a/nd4s/src/main/scala/org/nd4s/samediff/SameDiff.scala b/nd4s/src/main/scala/org/nd4s/samediff/SameDiff.scala
index 8ca21b72e..e43a1e86e 100644
--- a/nd4s/src/main/scala/org/nd4s/samediff/SameDiff.scala
+++ b/nd4s/src/main/scala/org/nd4s/samediff/SameDiff.scala
@@ -90,8 +90,13 @@ class SDVariableWrapper {
   def |(other: SDVariable)(implicit sameDiff: SameDiff): SDVariable = sameDiff.math.or(thisVariable, other)
   def &(other: SDVariable)(implicit sameDiff: SameDiff): SDVariable = sameDiff.math.and(thisVariable, other)
 
-  def <<(x: Int)(implicit sameDiff: SameDiff): SDVariable = sameDiff.math.bitShift(null, thisVariable, x)
-  def >>(x: Int)(implicit sameDiff: SameDiff): SDVariable = sameDiff.math.bitShiftRight(null, thisVariable, x)
+  def <<(other: SDVariable)(implicit sameDiff: SameDiff): SDVariable = sameDiff.math.bitShift(null, thisVariable, other)
+  def >>(other: SDVariable)(implicit sameDiff: SameDiff): SDVariable =
+    sameDiff.math.bitShiftRight(null, thisVariable, other)
+  def <<(x: Int)(implicit sameDiff: SameDiff): SDVariable =
+    sameDiff.math.bitShift(null, thisVariable, sameDiff.constant(x))
+  def >>(x: Int)(implicit sameDiff: SameDiff): SDVariable =
+    sameDiff.math.bitShiftRight(null, thisVariable, sameDiff.constant(x))
 
   // Overloads for numeric arguments
   // Float
diff --git a/nd4s/src/test/scala/org/nd4s/samediff/MathTest.scala b/nd4s/src/test/scala/org/nd4s/samediff/MathTest.scala
index a2c113b50..8e9f892c6 100644
--- a/nd4s/src/test/scala/org/nd4s/samediff/MathTest.scala
+++ b/nd4s/src/test/scala/org/nd4s/samediff/MathTest.scala
@@ -188,4 +188,16 @@ class MathTest extends FlatSpec with Matchers {
     val w3 = w1 >> 2
     w3.eval.toIntVector.head shouldBe 4
   }
+
+  "SameDiff" should "provide shifting operations with SDVariable argument" in {
+    implicit val sd = SameDiff.create()
+    val w1 = sd.constant(16)
+    val two = sd.constant(2)
+
+    val w2 = w1 << two
+    w2.eval.toIntVector.head shouldBe 64
+
+    val w3 = w1 >> two
+    w3.eval.toIntVector.head shouldBe 4
+  }
 }

From 62e96c97246622fb28a8f828261aeee578acfc35 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Fri, 30 Aug 2019 20:39:43 +1000
Subject: [PATCH 47/56] Guava Function -> nd4j-common Function (#203)

Signed-off-by: Alex Black <blacka101@gmail.com>
---
 .../src/main/java/org/nd4j/linalg/dataset/DataSet.java         | 3 ---
 .../src/main/java/org/nd4j/linalg/dataset/api/DataSet.java     | 2 --
 .../java/org/nd4j/linalg/indexing/conditions/Condition.java    | 2 +-
 .../main/java/org/nd4j/linalg/indexing/functions/Identity.java | 3 ++-
 .../java/org/nd4j/linalg/indexing/functions/StableNumber.java  | 2 +-
 .../main/java/org/nd4j/linalg/indexing/functions/Value.java    | 3 +--
 .../src/main/java/org/nd4j/linalg/indexing/functions/Zero.java | 2 +-
 7 files changed, 6 insertions(+), 11 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java
index 7b208f4f2..cb354c2d1 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/DataSet.java
@@ -16,7 +16,6 @@
 
 package org.nd4j.linalg.dataset;
 
-import org.nd4j.shade.guava.base.Function;
 import org.nd4j.shade.guava.collect.Lists;
 import lombok.extern.slf4j.Slf4j;
 import org.nd4j.base.Preconditions;
@@ -26,8 +25,6 @@ import org.nd4j.linalg.dataset.api.MultiDataSet;
 import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
 import org.nd4j.linalg.dataset.api.preprocessor.NormalizerStandardize;
 import org.nd4j.linalg.factory.Nd4j;
-import org.nd4j.linalg.indexing.BooleanIndexing;
-import org.nd4j.linalg.indexing.conditions.Condition;
 import org.nd4j.linalg.primitives.Pair;
 import org.nd4j.linalg.util.ArrayUtil;
 import org.nd4j.linalg.util.FeatureUtil;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/DataSet.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/DataSet.java
index 7358da385..5f6731c26 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/DataSet.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/DataSet.java
@@ -16,12 +16,10 @@
 
 package org.nd4j.linalg.dataset.api;
 
-import org.nd4j.shade.guava.base.Function;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.api.rng.Random;
 import org.nd4j.linalg.dataset.SplitTestAndTrain;
 import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
-import org.nd4j.linalg.indexing.conditions.Condition;
 
 import java.io.File;
 import java.io.InputStream;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/conditions/Condition.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/conditions/Condition.java
index 3b4245fef..6c7093cab 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/conditions/Condition.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/conditions/Condition.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.indexing.conditions;
 
-import org.nd4j.shade.guava.base.Function;
+import org.nd4j.linalg.function.Function;
 
 /**
  * Condition for boolean indexing
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Identity.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Identity.java
index 0413e2505..6c19cef2d 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Identity.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Identity.java
@@ -16,7 +16,8 @@
 
 package org.nd4j.linalg.indexing.functions;
 
-import org.nd4j.shade.guava.base.Function;
+
+import org.nd4j.linalg.function.Function;
 
 /**
  * Created by agibsonccc on 10/8/14.
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/StableNumber.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/StableNumber.java
index e044dcda8..483542a9a 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/StableNumber.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/StableNumber.java
@@ -17,8 +17,8 @@
 package org.nd4j.linalg.indexing.functions;
 
 
-import org.nd4j.shade.guava.base.Function;
 import org.nd4j.linalg.factory.Nd4j;
+import org.nd4j.linalg.function.Function;
 
 /**
  * Returns a stable number based on infinity
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Value.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Value.java
index 8a6a6436f..5562bbb35 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Value.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Value.java
@@ -16,8 +16,7 @@
 
 package org.nd4j.linalg.indexing.functions;
 
-
-import org.nd4j.shade.guava.base.Function;
+import org.nd4j.linalg.function.Function;
 
 /**
  * Created by agibsonccc on 10/8/14.
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Zero.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Zero.java
index 9bcb3c463..b8b53c2d1 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Zero.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/indexing/functions/Zero.java
@@ -16,7 +16,7 @@
 
 package org.nd4j.linalg.indexing.functions;
 
-import org.nd4j.shade.guava.base.Function;
+import org.nd4j.linalg.function.Function;
 
 /**
  * Created by agibsonccc on 10/8/14.

From 54e320a255bbe7aa3f2b723314a232e5ce5d0ce0 Mon Sep 17 00:00:00 2001
From: Robert Altena <Rob@ra-ai.com>
Date: Fri, 30 Aug 2019 21:40:27 +0900
Subject: [PATCH 48/56] javadoc (#201)

Signed-off-by: Robert Altena <Rob@Ra-ai.com>
---
 .../nd4j/linalg/api/ndarray/BaseNDArray.java  | 230 +-----------------
 .../linalg/api/ndarray/BaseSparseNDArray.java |  13 -
 .../api/ndarray/BaseSparseNDArrayCOO.java     |   7 -
 .../api/ndarray/BaseSparseNDArrayCSR.java     |   8 -
 .../org/nd4j/linalg/api/ndarray/INDArray.java | 142 +++++------
 .../linalg/api/indexing/IndexingTests.java    |  15 +-
 6 files changed, 65 insertions(+), 350 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
index 50ab2727d..95a087301 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
@@ -1711,16 +1711,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return result;
     }
 
-
-
-    /**
-     * Returns the element at the specified row/column
-     * This will throw an exception if the
-     *
-     * @param row    the row of the element to return
-     * @param column the row of the element to return
-     * @return a scalar indarray of the element at this index
-     */
     @Override
     public INDArray getScalar(long row, long column) {
         return getScalar(new long[] {row, column});
@@ -1885,27 +1875,20 @@ public abstract class BaseNDArray implements INDArray, Iterable {
 
     }
 
-    /**
-     * Inserts the element at the specified index
-     *
-     * @param indices the indices to insert into
-     * @param element a scalar ndarray
-     * @return a scalar ndarray of the element at this index
-     */
     @Override
     public INDArray put(int[] indices, INDArray element) {
         Nd4j.getCompressor().autoDecompress(this);
         if (!element.isScalar())
             throw new IllegalArgumentException("Unable to insert anything but a scalar");
         if (isRowVector() && indices[0] == 0 && indices.length == 2) {
-            int ix = 0; //Shape.offset(javaShapeInformation);
+            int ix = 0;
             for (int i = 1; i < indices.length; i++)
                 ix += indices[i] * stride(i);
             if (ix >= data.length())
                 throw new IllegalArgumentException("Illegal indices " + Arrays.toString(indices));
             data.put(ix, element.getDouble(0));
         } else {
-            int ix = 0; //Shape.offset(javaShapeInformation);
+            int ix = 0;
             for (int i = 0; i < indices.length; i++)
                 if (size(i) != 1)
                     ix += indices[i] * stride(i);
@@ -1913,10 +1896,7 @@ public abstract class BaseNDArray implements INDArray, Iterable {
                 throw new IllegalArgumentException("Illegal indices " + Arrays.toString(indices));
             data.put(ix, element.getDouble(0));
         }
-
-
         return this;
-
     }
 
     @Override
@@ -1970,39 +1950,16 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return putWhereWithMask(mask,Nd4j.scalar(put));
     }
 
-    /**
-     * Inserts the element at the specified index
-     *
-     * @param i       the row insert into
-     * @param j       the column to insert into
-     * @param element a scalar ndarray
-     * @return a scalar ndarray of the element at this index
-     */
     @Override
     public INDArray put(int i, int j, INDArray element) {
         return put(new int[] {i, j}, element);
     }
 
-    /**
-     * Inserts the element at the specified index
-     *
-     * @param i       the row insert into
-     * @param j       the column to insert into
-     * @param element a scalar ndarray
-     * @return a scalar ndarray of the element at this index
-     */
     @Override
     public INDArray put(int i, int j, Number element) {
         return putScalar(new int[] {i, j}, element.doubleValue());
     }
 
-    /**
-     * Assigns the given matrix (put) to the specified slice
-     *
-     * @param slice the slice to assign
-     * @param put   the slice to put
-     * @return this for chainability
-     */
     @Override
     public INDArray putSlice(int slice, INDArray put) {
         Nd4j.getCompressor().autoDecompress(this);
@@ -2102,10 +2059,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return Nd4j.getStrides(shape, ordering);
     }
 
-
-    /**
-     * Returns the square of the Euclidean distance.
-     */
     @Override
     public double squaredDistance(INDArray other) {
         validateNumericalArray("squaredDistance", false);
@@ -2113,9 +2066,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return d2 * d2;
     }
 
-    /**
-     * Returns the (euclidean) distance.
-     */
     @Override
     public double distance2(INDArray other) {
         validateNumericalArray("distance2", false);
@@ -2123,9 +2073,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return Nd4j.getExecutioner().execAndReturn(new EuclideanDistance(this, other)).getFinalResult().doubleValue();
     }
 
-    /**
-     * Returns the (1-norm) distance.
-     */
     @Override
     public double distance1(INDArray other) {
         validateNumericalArray("distance1", false);
@@ -2133,8 +2080,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return Nd4j.getExecutioner().execAndReturn(new ManhattanDistance(this, other)).getFinalResult().doubleValue();
     }
 
-
-
     @Override
     public INDArray get(INDArray indices) {
         if(indices.rank() > 2) {
@@ -2190,49 +2135,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
 
     }
 
-    @Override
-    public INDArray get(List<List<Integer>> indices) {
-        INDArrayIndex[] indArrayIndices = new INDArrayIndex[indices.size()];
-        for(int i = 0; i < indArrayIndices.length; i++) {
-            indArrayIndices[i] = new SpecifiedIndex(Ints.toArray(indices.get(i)));
-        }
-
-        boolean hasNext = true;
-        Generator<List<List<Long>>> iterate = SpecifiedIndex.iterate(indArrayIndices);
-        List<INDArray> resultList = new ArrayList<>();
-        while(hasNext) {
-            try {
-                List<List<Long>> next = iterate.next();
-                int[][] nextArr = new int[next.size()][];
-                for(int i = 0; i < next.size(); i++) {
-                    nextArr[i] = Ints.toArray(next.get(i));
-                }
-
-                int[] curr = Ints.concat(nextArr);
-                INDArray currSlice = this;
-                for(int j = 0; j < curr.length; j++) {
-                    currSlice = currSlice.slice(curr[j]);
-                }
-
-                //slice drops the first dimension, this adds a 1 to match normal numpy behavior
-                currSlice = currSlice.reshape(Longs.concat(new long[]{1},currSlice.shape()));
-
-                resultList.add(currSlice);
-
-
-            }
-            catch(NoSuchElementException e) {
-                hasNext = false;
-            }
-        }
-
-
-
-
-        return Nd4j.concat(0,resultList.toArray(new INDArray[resultList.size()]));
-    }
-
-
     @Override
     public INDArray put(INDArray indices, INDArray element) {
         if(indices.rank() > 2) {
@@ -2245,7 +2147,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
                 int[] specifiedIndex = indices.getColumn(i).dup().data().asInt();
                 putScalar(specifiedIndex,element.getDouble(ndIndexIterator.next()));
             }
-
         }
         else {
             List<INDArray> arrList = new ArrayList<>();
@@ -2258,8 +2159,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
                         Nd4j.getExecutioner().execAndReturn(new Assign(new INDArray[]{slice,element},new INDArray[]{slice}));
                         arrList.add(slice(row.getInt(j)));
                     }
-
-
                 }
             }
             else if(indices.isRowVector()) {
@@ -2267,15 +2166,10 @@ public abstract class BaseNDArray implements INDArray, Iterable {
                     arrList.add(slice(indices.getInt(i)));
                 }
             }
-
         }
-
-
         return this;
-
     }
 
-
     @Override
     public INDArray put(INDArrayIndex[] indices, INDArray element) {
         Nd4j.getCompressor().autoDecompress(this);
@@ -2343,7 +2237,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return this;
     }
 
-
     /**
      * Mainly here for people coming from numpy.
      * This is equivalent to a call to permute
@@ -2446,7 +2339,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return ret;
     }
 
-
     protected void init(int[] shape, int[] stride) {
         //null character
         if (shapeInformation == null || jvmShapeInfo == null || ordering() == '\u0000') {
@@ -2466,7 +2358,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
 
     }
 
-
     @Override
     public INDArray getScalar(long i) {
         if (i >= this.length())
@@ -2936,13 +2827,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return dup().rsubiRowVector(rowVector);
     }
 
-    /**
-     * Inserts the element at the specified index
-     *
-     * @param i       the index insert into
-     * @param element a scalar ndarray
-     * @return a scalar ndarray of the element at this index
-     */
     @Override
     public INDArray put(int i, INDArray element) {
         Preconditions.checkArgument(element.isScalar(), "Element must be a scalar: element has shape %ndShape", element);
@@ -3706,12 +3590,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return normmax(false, dimension);
     }
 
-    /**
-     * Reverse division
-     *
-     * @param other the matrix to divide from
-     * @return
-     */
     @Override
     public INDArray rdiv(INDArray other) {
         validateNumericalArray("rdiv", false);
@@ -3722,37 +3600,17 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         }
     }
 
-    /**
-     * Reverse divsion (in place)
-     *
-     * @param other
-     * @return
-     */
     @Override
     public INDArray rdivi(INDArray other) {
         return rdivi(other, this);
     }
 
-    /**
-     * Reverse division
-     *
-     * @param other  the matrix to subtract from
-     * @param result the result ndarray
-     * @return
-     */
     @Override
     public INDArray rdiv(INDArray other, INDArray result) {
         validateNumericalArray("rdiv", false);
         return dup().rdivi(other, result);
     }
 
-    /**
-     * Reverse division (in-place)
-     *
-     * @param other  the other ndarray to subtract
-     * @param result the result ndarray
-     * @return the ndarray with the operation applied
-     */
     @Override
     public INDArray rdivi(INDArray other, INDArray result) {
         validateNumericalArray("rdivi", false);
@@ -3761,23 +3619,12 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return result;
     }
 
-    /**
-     * Reverse subtraction
-     *
-     * @param other  the matrix to subtract from
-     * @param result the result ndarray
-     * @return
-     */
     @Override
     public INDArray rsub(INDArray other, INDArray result) {
         validateNumericalArray("rsub", false);
         return rsubi(other, result);
     }
 
-    /**
-     * @param other
-     * @return
-     */
     @Override
     public INDArray rsub(INDArray other) {
         validateNumericalArray("rsub", false);
@@ -3788,22 +3635,11 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         }
     }
 
-    /**
-     * @param other
-     * @return
-     */
     @Override
     public INDArray rsubi(INDArray other) {
         return rsubi(other, this);
     }
 
-    /**
-     * Reverse subtraction (in-place)
-     *
-     * @param other  the other ndarray to subtract
-     * @param result the result ndarray
-     * @return the ndarray with the operation applied
-     */
     @Override
     public INDArray rsubi(INDArray other, INDArray result) {
         validateNumericalArray("rsubi", false);
@@ -3812,12 +3648,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return result;
     }
 
-    /**
-     * Set the value of the ndarray to the specified value
-     *
-     * @param value the value to assign
-     * @return the ndarray with the values
-     */
     @Override
     public INDArray assign(Number value) {
         Preconditions.checkState(dataType() != DataType.BOOL || value.doubleValue() == 0.0 || value.doubleValue() == 1.0, "Only values 0 or 1 are allowed for scalar " +
@@ -3826,7 +3656,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return this;
     }
 
-
     @Override
     public INDArray assign(boolean value) {
         return assign(value ? 1 : 0);
@@ -3846,7 +3675,7 @@ public abstract class BaseNDArray implements INDArray, Iterable {
     }
 
     @Override
-    @Deprecated
+    @Deprecated  //TODO: Investigate. Not deprecated in the base interface.
     public long linearIndex(long i) {
         long idx = i;
         for (int j = 0; j < jvmShapeInfo.rank - 1; j++) {
@@ -4073,15 +3902,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return addi(n, this);
     }
 
-
-
-    /**
-     * Replicate and tile array to fill out to the given shape
-     * See:
-     * https://github.com/numpy/numpy/blob/master/numpy/matlib.py#L310-L358
-     * @param shape the new shape of this ndarray
-     * @return the shape to fill out to
-     */
     @Override
     public INDArray repmat(int[] shape) {
         Nd4j.getCompressor().autoDecompress(this);
@@ -4109,16 +3929,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return out;
     }
 
-
-    /**
-     * Insert a row in to this array
-     * Will throw an exception if this
-     * ndarray is not a matrix
-     *
-     * @param row   the row insert into
-     * @param toPut the row to insert
-     * @return this
-     */
     @Override
     public INDArray putRow(long row, INDArray toPut) {
         if (isRowVector() && toPut.isVector()) {
@@ -4127,15 +3937,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return put(new INDArrayIndex[] {NDArrayIndex.point(row), NDArrayIndex.all()}, toPut);
     }
 
-    /**
-     * Insert a column in to this array
-     * Will throw an exception if this
-     * ndarray is not a matrix
-     *
-     * @param column the column to insert
-     * @param toPut  the array to put
-     * @return this
-     */
     @Override
     public INDArray putColumn(int column, INDArray toPut) {
         Nd4j.getCompressor().autoDecompress(this);
@@ -4752,11 +4553,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return reshape(length());
     }
 
-    /**
-     * Flattens the array for linear indexing
-     *
-     * @return the flattened version of this array
-     */
     @Override
     public void sliceVectors(List<INDArray> list) {
         if (isVector())
@@ -4804,12 +4600,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return col.reshape(col.length(), 1);
     }
 
-
-    /**
-     * Get whole rows from the passed indices.
-     *
-     * @param rindices
-     */
     @Override
     public INDArray getRows(int[] rindices) {
         Nd4j.getCompressor().autoDecompress(this);
@@ -4826,13 +4616,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         }
     }
 
-    /**
-     * Returns a subset of this array based on the specified
-     * indexes
-     *
-     * @param indexes the indexes in to the array
-     * @return a view of the array with the specified indices
-     */
     @Override
     public INDArray get(INDArrayIndex... indexes) {
         Nd4j.getCompressor().autoDecompress(this);
@@ -5020,13 +4803,6 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         return out;
     }
 
-
-    /**
-     * Get whole columns
-     * from the passed indices.
-     *
-     * @param cindices
-     */
     @Override
     public INDArray getColumns(int... cindices) {
         if (!isMatrix() && !isVector())
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
index 952c4b777..6a112b868 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArray.java
@@ -153,7 +153,6 @@ public abstract class BaseSparseNDArray implements ISparseNDArray {
                             arrList.set(j,put);
                         }
                     }
-
                 }
             }
             else if(indices.isRowVector()) {
@@ -161,12 +160,8 @@ public abstract class BaseSparseNDArray implements ISparseNDArray {
                     arrList.add(slice(indices.getInt(i)));
                 }
             }
-
             return Nd4j.concat(0,arrList.toArray(new INDArray[arrList.size()]));
-
         }
-
-
     }
 
     @Override
@@ -259,21 +254,13 @@ public abstract class BaseSparseNDArray implements ISparseNDArray {
         return null;
     }
 
-    @Override
-    public INDArray get(List<List<Integer>> indices) {
-        return null;
-    }
-
     @Override
     public INDArray put(INDArray indices, INDArray element) {
         INDArrayIndex[] realIndices = new INDArrayIndex[indices.rank()];
         for(int i = 0; i < realIndices.length; i++) {
             realIndices[i] = new SpecifiedIndex(indices.slice(i).dup().data().asInt());
         }
-
-
         return put(realIndices,element);
-
     }
 
     @Override
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCOO.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCOO.java
index 77463c456..da5cd3f60 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCOO.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCOO.java
@@ -568,13 +568,6 @@ public class BaseSparseNDArrayCOO extends BaseSparseNDArray {
         return this;
     }
 
-    /**
-     * Returns a subset of this array based on the specified
-     * indexes
-     *
-     * @param indexes the indexes in to the array
-     * @return a view of the array with the specified indices
-     */
     @Override
     public INDArray get(INDArrayIndex... indexes) {
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCSR.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCSR.java
index dd17fa635..92e59486c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCSR.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseSparseNDArrayCSR.java
@@ -124,14 +124,6 @@ public abstract class BaseSparseNDArrayCSR extends BaseSparseNDArray {
         return this;
     }
 
-
-    /**
-     * Returns a subset of this array based on the specified
-     * indexes
-     *
-     * @param indexes the indexes in to the array
-     * @return a view of the array with the specified indices
-     */
     @Override
     public INDArray get(INDArrayIndex... indexes) {
         //check for row/column vector and point index being 0
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java
index fe940f3ee..49cabf6bb 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/INDArray.java
@@ -458,7 +458,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray rsub(Number n);
 
-
     /**
      * Reverse subtraction in place - i.e., (n - thisArrayValues)
      *
@@ -467,7 +466,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray rsubi(Number n);
 
-
     /**
      * Division by a number
      *
@@ -484,7 +482,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray divi(Number n);
 
-
     /**
      * Scalar multiplication (copy)
      *
@@ -501,7 +498,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray muli(Number n);
 
-
     /**
      * Scalar subtraction (copied)
      *
@@ -510,7 +506,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray sub(Number n);
 
-
     /**
      * In place scalar subtraction
      *
@@ -535,7 +530,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray addi(Number n);
 
-
     /**
      * Reverse division (number / ndarray)
      *
@@ -545,7 +539,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray rdiv(Number n, INDArray result);
 
-
     /**
      * Reverse in place division
      *
@@ -573,11 +566,12 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray rsubi(Number n, INDArray result);
 
-
     /**
-     * @param n
-     * @param result
-     * @return
+     * Division if ndarray by number
+     *
+     * @param n      the number to divide by
+     * @param result the result ndarray
+     * @return the result ndarray
      */
     INDArray div(Number n, INDArray result);
 
@@ -586,24 +580,35 @@ public interface INDArray extends Serializable, AutoCloseable {
      *
      * @param n      the number to divide by
      * @param result the result ndarray
-     * @return
+     * @return the result ndarray
      */
     INDArray divi(Number n, INDArray result);
 
-
+    /**
+     * Multiplication of ndarray.
+     *
+     * @param n the number to multiply by
+     * @param result the result ndarray
+     * @return the result ndarray
+     */
     INDArray mul(Number n, INDArray result);
 
-
     /**
      * In place multiplication of this ndarray
      *
      * @param n      the number to divide by
      * @param result the result ndarray
-     * @return
+     * @return the result ndarray
      */
     INDArray muli(Number n, INDArray result);
 
-
+    /**
+     * Subtraction of this ndarray
+     *
+     * @param n      the number to subtract by
+     * @param result the result ndarray
+     * @return the result ndarray
+     */
     INDArray sub(Number n, INDArray result);
 
     /**
@@ -615,6 +620,12 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray subi(Number n, INDArray result);
 
+    /**
+     * Addition of this ndarray.
+     * @param n      the number to add
+     * @param result the result ndarray
+     * @return the result ndarray
+     */
     INDArray add(Number n, INDArray result);
 
     /**
@@ -626,25 +637,24 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray addi(Number n, INDArray result);
 
-
     /**
-     * Returns a subset of this array based on the specified
-     * indexes
+     * Returns a subset of this array based on the specified indexes
      *
      * @param indexes the indexes in to the array
      * @return a view of the array with the specified indices
      */
     INDArray get(INDArrayIndex... indexes);
 
+    //TODO: revisit after #8166 is resolved.
     /**
-     * Return a mask on whether each element
-     * matches the given condition
+     * Return a mask on whether each element matches the given condition
      * @param comp
      * @param condition
      * @return
      */
     INDArray match(INDArray comp,Condition condition);
 
+    //TODO: revisit after #8166 is resolved.
     /**
      * Returns a mask
      * @param comp
@@ -673,54 +683,51 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray getWhere(Number comp,Condition condition);
 
+    //TODO: unused / untested method. (only used to forward calls from putWhere(Number,INDArray ,Condition).
     /**
-     * Assign the element according
-     * to the comparison array
+     * Assign the element according to the comparison array
      * @param comp the comparison array
      * @param put the elements to put
      * @param condition the condition for masking on
-     * @return
+     * @return a copy of this array with the conditional assignments.
      */
     INDArray putWhere(INDArray comp,INDArray put,Condition condition);
 
-
+    //TODO: unused / untested method.
     /**
-     * Assign the element according
-     * to the comparison array
+     * Assign the element according to the comparison array
      * @param comp the comparison array
      * @param put the elements to put
      * @param condition the condition for masking on
-     * @return
+     * @return a copy of this array with the conditional assignments.
      */
     INDArray putWhere(Number comp,INDArray put,Condition condition);
 
-
+    //TODO: unused / untested method. (only used to forward calls from  other putWhereWithMask implementations.
     /**
-     * Use a pre computed mask
-     * for assigning arrays
+     * Use a pre computed mask for assigning arrays
      * @param mask the mask to use
      * @param put the array to put
-     * @return the resulting array
+     * @return a copy of this array with the conditional assignments.
      */
     INDArray putWhereWithMask(INDArray mask,INDArray put);
 
-
+    //TODO: unused / untested method.
     /**
-     * Use a pre computed mask
-     * for assigning arrays
+     * Use a pre computed mask for assigning arrays
      * @param mask the mask to use
      * @param put the array to put
-     * @return the resulting array
+     * @return a copy of this array with the conditional assignments.
      */
     INDArray putWhereWithMask(INDArray mask,Number put);
 
+    //TODO: unused / untested method.
     /**
-     * Assign the element according
-     * to the comparison array
+     * Assign the element according to the comparison array
      * @param comp the comparison array
      * @param put the elements to put
      * @param condition the condition for masking on
-     * @return
+     * @return a copy of this array with the conditional assignments.
      */
     INDArray putWhere(Number comp,Number put,Condition condition);
 
@@ -731,14 +738,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray get(INDArray indices);
 
-    /**
-     * Get the elements from this ndarray based on the specified indices
-     * @param indices an ndaray of the indices to get the elements for
-     * @return the elements to get the array for
-     */
-    @Deprecated
-    INDArray get(List<List<Integer>> indices);
-
     /**
      * Get an INDArray comprised of the specified columns only. Copy operation.
      *
@@ -771,20 +770,20 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray rdivi(INDArray other);
 
-
+    //TODO: unused / untested method.
     /**
      * Reverse division
      *
-     * @param other  the matrix to subtract from
+     * @param other  the matrix to divide from
      * @param result the result ndarray
-     * @return
+     * @return the result ndarray
      */
     INDArray rdiv(INDArray other, INDArray result);
 
     /**
      * Reverse division (in-place)
      *
-     * @param other  the other ndarray to subtract
+     * @param other  the matrix to divide from
      * @param result the result ndarray
      * @return the ndarray with the operation applied
      */
@@ -795,11 +794,10 @@ public interface INDArray extends Serializable, AutoCloseable {
      *
      * @param other  the matrix to subtract from
      * @param result the result ndarray
-     * @return
+     * @return the result ndarray
      */
     INDArray rsub(INDArray other, INDArray result);
 
-
     /**
      * Element-wise reverse subtraction (copy op). i.e., other - this
      *
@@ -842,22 +840,19 @@ public interface INDArray extends Serializable, AutoCloseable {
     INDArray assign(boolean value);
 
     /**
-     * Get the linear index of the data in to
-     * the array
+     * Get the linear index of the data in to the array
      *
      * @param i the index to getScalar
      * @return the linear index in to the data
      */
     long linearIndex(long i);
 
-
+    //TODO: unused / untested method. only used recursively.
     /**
-     *
-     * @param list
+     * Flattens the array for linear indexing in list.
      */
     void sliceVectors(List<INDArray> list);
 
-
     /**
      * Assigns the given matrix (put) to the specified slice
      *
@@ -875,16 +870,15 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray cond(Condition condition);
 
-
     /**
      * Replicate and tile array to fill out to the given shape
-     *
+     * See:
+     * https://github.com/numpy/numpy/blob/master/numpy/matlib.py#L310-L358
      * @param shape the new shape of this ndarray
      * @return the shape to fill out to
      */
     INDArray repmat(int... shape);
 
-
     /**
      * Repeat elements along a specified dimension.
      *
@@ -894,11 +888,9 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray repeat(int dimension, long... repeats);
 
-
     /**
      * Insert a row in to this array
-     * Will throw an exception if this
-     * ndarray is not a matrix
+     * Will throw an exception if this ndarray is not a matrix
      *
      * @param row   the row insert into
      * @param toPut the row to insert
@@ -908,8 +900,7 @@ public interface INDArray extends Serializable, AutoCloseable {
 
     /**
      * Insert a column in to this array
-     * Will throw an exception if this
-     * ndarray is not a matrix
+     * Will throw an exception if this ndarray is not a matrix
      *
      * @param column the column to insert
      * @param toPut  the array to put
@@ -919,7 +910,6 @@ public interface INDArray extends Serializable, AutoCloseable {
 
     /**
      * Returns the element at the specified row/column
-     * This will throw an exception if the
      *
      * @param row    the row of the element to return
      * @param column the row of the element to return
@@ -950,26 +940,20 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     double distance1(INDArray other);
 
-
     /**
      * Put element in to the indices denoted by
      * the indices ndarray.
-     * This is equivalent to:
+     * In numpy this is equivalent to:
      * a[indices] = element
      *
-     *  in numpy.
-     *
      * @param indices the indices to put
      * @param element the element array to put
      * @return this array
      */
     INDArray put(INDArray indices,INDArray element);
 
-
-
     /**
-     * Put the elements of the ndarray
-     * in to the specified indices
+     * Put the elements of the ndarray in to the specified indices
      *
      * @param indices the indices to put the ndarray in to
      * @param element the ndarray to put
@@ -978,8 +962,7 @@ public interface INDArray extends Serializable, AutoCloseable {
     INDArray put(INDArrayIndex[] indices, INDArray element);
 
     /**
-     * Put the elements of the ndarray
-     * in to the specified indices
+     * Put the elements of the ndarray in to the specified indices
      *
      * @param indices the indices to put the ndarray in to
      * @param element the ndarray to put
@@ -1007,7 +990,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray put(int i, int j, INDArray element);
 
-
     /**
      * Inserts the element at the specified index
      *
@@ -1018,7 +1000,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray put(int i, int j, Number element);
 
-
     /**
      * Inserts the element at the specified index
      *
@@ -1028,7 +1009,6 @@ public interface INDArray extends Serializable, AutoCloseable {
      */
     INDArray put(int i, INDArray element);
 
-
     /**
      * In place division of a column vector
      *
diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/indexing/IndexingTests.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/indexing/IndexingTests.java
index ab724ee1b..b49d49b79 100644
--- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/indexing/IndexingTests.java
+++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/api/indexing/IndexingTests.java
@@ -100,20 +100,7 @@ public class IndexingTests extends BaseNd4jTest {
         INDArray vals = Nd4j.valueArrayOf(new long[] {2,2,2,2},5, DataType.DOUBLE);
         assertEquals(vals,x);
     }
-
-
-    @Test @Ignore
-    public void testIndexGetDuplicate() {
-        List<List<Integer>> indices = new ArrayList<>();
-        indices.add(Arrays.asList(0,0));
-        INDArray linspace = Nd4j.linspace(1,16,16, DataType.DOUBLE).reshape('c',2,2,2,2).castTo(DataType.DOUBLE);
-        INDArray get = linspace.get(indices);
-        INDArray assertion = Nd4j.create(new double[] {1, 2, 3, 4, 5, 6, 7, 8, 1, 2, 3, 4, 5, 6, 7, 8}).reshape('c',2,2,2,2).castTo(DataType.DOUBLE);
-        assertEquals(assertion,get);
-    }
-
-
-
+    
     @Test
     public void testGetScalar() {
         INDArray arr = Nd4j.linspace(1, 5, 5, DataType.DOUBLE);

From 3f3b676ce5f5207bb72e6cdd4df0b2f39a11393e Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Fri, 30 Aug 2019 23:00:53 +1000
Subject: [PATCH 49/56] DL4J Fixes (#204)

* Fix issue with recently introduced exception handling system in MultiLayerNetwork/ComputationGraph

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix for SpaceToBatch layer

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* #8133 DL4J SpaceToBatch gradient fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 .../exceptions/TestInvalidInput.java          | 55 ++++++++++++++-----
 .../gradientcheck/CNNGradientCheckTest.java   |  7 +++
 .../nn/graph/ComputationGraph.java            | 19 ++++++-
 .../nn/layers/convolution/SpaceToBatch.java   |  4 +-
 .../nn/layers/recurrent/LSTMHelpers.java      | 10 ++--
 .../nn/layers/recurrent/SimpleRnn.java        |  3 +
 .../nn/multilayer/MultiLayerNetwork.java      | 19 ++++++-
 7 files changed, 92 insertions(+), 25 deletions(-)

diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/exceptions/TestInvalidInput.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/exceptions/TestInvalidInput.java
index 30ef68183..096c7ac69 100644
--- a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/exceptions/TestInvalidInput.java
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/exceptions/TestInvalidInput.java
@@ -22,12 +22,16 @@ import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
 import org.deeplearning4j.nn.conf.inputs.InputType;
 import org.deeplearning4j.nn.conf.layers.*;
+import org.deeplearning4j.nn.conf.layers.recurrent.SimpleRnn;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
 import org.junit.Test;
 import org.nd4j.linalg.activations.Activation;
+import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.factory.Nd4j;
 
-import static org.junit.Assert.fail;
+import java.util.Map;
+
+import static org.junit.Assert.*;
 
 /**
  * A set of tests to ensure that useful exceptions are thrown on invalid input
@@ -267,23 +271,44 @@ public class TestInvalidInput extends BaseDL4JTest {
         //Idea: Using rnnTimeStep with a different number of examples between calls
         //(i.e., not calling reset between time steps)
 
-        MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().list()
-                        .layer(0, new GravesLSTM.Builder().nIn(5).nOut(5).build())
-                        .layer(1, new RnnOutputLayer.Builder().nIn(5).nOut(5).activation(Activation.SOFTMAX).build()).build();
+        for(String layerType : new String[]{"simple", "lstm", "graves"}) {
 
-        MultiLayerNetwork net = new MultiLayerNetwork(conf);
-        net.init();
+            Layer l;
+            switch (layerType){
+                case "simple":
+                    l = new SimpleRnn.Builder().nIn(5).nOut(5).build();
+                    break;
+                case "lstm":
+                    l = new LSTM.Builder().nIn(5).nOut(5).build();
+                    break;
+                case "graves":
+                    l = new GravesLSTM.Builder().nIn(5).nOut(5).build();
+                    break;
+                default:
+                    throw new RuntimeException();
+            }
 
-        net.rnnTimeStep(Nd4j.create(3, 5, 10));
+            MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().list()
+                    .layer(l)
+                    .layer(new RnnOutputLayer.Builder().nIn(5).nOut(5).activation(Activation.SOFTMAX).build()).build();
 
-        try {
-            net.rnnTimeStep(Nd4j.create(5, 5, 10));
-            fail("Expected DL4JException");
-        } catch (DL4JException e) {
-            System.out.println("testInvalidRnnTimeStep(): " + e.getMessage());
-        } catch (Exception e) {
-            e.printStackTrace();
-            fail("Expected DL4JException");
+            MultiLayerNetwork net = new MultiLayerNetwork(conf);
+            net.init();
+
+            net.rnnTimeStep(Nd4j.create(3, 5, 10));
+
+            Map<String, INDArray> m = net.rnnGetPreviousState(0);
+            assertNotNull(m);
+            assertFalse(m.isEmpty());
+
+            try {
+                net.rnnTimeStep(Nd4j.create(5, 5, 10));
+                fail("Expected Exception - " + layerType);
+            } catch (Exception e) {
+//                e.printStackTrace();
+                String msg = e.getMessage();
+                assertTrue(msg, msg != null && msg.contains("rnn") && msg.contains("batch"));
+            }
         }
     }
 }
diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/gradientcheck/CNNGradientCheckTest.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/gradientcheck/CNNGradientCheckTest.java
index c1a873dc8..decb81bb0 100644
--- a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/gradientcheck/CNNGradientCheckTest.java
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/gradientcheck/CNNGradientCheckTest.java
@@ -28,6 +28,7 @@ import org.deeplearning4j.nn.conf.distribution.NormalDistribution;
 import org.deeplearning4j.nn.conf.inputs.InputType;
 import org.deeplearning4j.nn.conf.layers.*;
 import org.deeplearning4j.nn.conf.layers.convolutional.Cropping2D;
+import org.deeplearning4j.nn.graph.ComputationGraph;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
 import org.deeplearning4j.nn.weights.WeightInit;
 import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
@@ -343,6 +344,12 @@ public class CNNGradientCheckTest extends BaseDL4JTest {
 
                     assertTrue(msg, gradOK);
 
+                    //Also check compgraph:
+                    ComputationGraph cg = net.toComputationGraph();
+                    gradOK = GradientCheckUtil.checkGradients(cg, DEFAULT_EPS, DEFAULT_MAX_REL_ERROR,
+                            DEFAULT_MIN_ABS_ERROR, PRINT_RESULTS, RETURN_ON_FIRST_FAILURE, new INDArray[]{input}, new INDArray[]{labels});
+                    assertTrue(msg + " - compgraph", gradOK);
+
                     TestUtils.testModelSerialization(net);
                 }
             }
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/ComputationGraph.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/ComputationGraph.java
index 00c0cf7d6..7c292bafa 100755
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/ComputationGraph.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/ComputationGraph.java
@@ -2460,6 +2460,13 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
             }
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
 
+            if(t != null){
+                if(t instanceof RuntimeException){
+                    throw ((RuntimeException)t);
+                }
+                throw new RuntimeException("Error during neural network forward pass", t);
+            }
+
             if(outputWorkspace == null || outputWorkspace instanceof DummyWorkspace) {
                 WorkspaceUtils.assertNoWorkspacesOpen("Expected no workspace active at the end of outputOfLayerDetached");
             } else {
@@ -2780,6 +2787,13 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                 }
             }
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
+
+            if(t != null){
+                if(t instanceof RuntimeException){
+                    throw ((RuntimeException)t);
+                }
+                throw new RuntimeException("Error during neural network backpropagation calculation", t);
+            }
         }
 
         //Now, add the gradients in the order we need them in for flattening (same as params order)
@@ -3312,8 +3326,11 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
 
     @Override
     public int batchSize() {
+        //In 99+% of cases, the input and labels dimension 0 size should be identical
+        //The only real exceptions: space to batch, and batch to space layers
+        //In those cases, we should base it on the labels size, as this impacts gradient calculation
         // FIXME: int cast
-        return (int) inputs[0].size(0);
+        return labels == null || labels[0] == null ? (int) inputs[0].size(0) : (int)labels[0].size(0);
     }
 
     @Override
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/convolution/SpaceToBatch.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/convolution/SpaceToBatch.java
index eef8fcee7..586464716 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/convolution/SpaceToBatch.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/convolution/SpaceToBatch.java
@@ -70,12 +70,12 @@ public class SpaceToBatch extends AbstractLayer<org.deeplearning4j.nn.conf.layer
 
     private INDArray getBlocksArray() {
         int[] intBlocks = layerConf().getBlocks();
-        return Nd4j.create(new double[] {intBlocks[0], intBlocks[1]});
+        return Nd4j.createFromArray(intBlocks);
     }
 
     private INDArray getPaddingArray() {
         int[][] intPad = layerConf().getPadding();
-        return Nd4j.create( new double[][] { {intPad[0][0], intPad[0][1]}, {intPad[1][0], intPad[1][1]}});
+        return Nd4j.createFromArray(intPad);
     }
 
 
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTMHelpers.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTMHelpers.java
index 881e3d786..27b980521 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTMHelpers.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/LSTMHelpers.java
@@ -32,6 +32,7 @@ import org.deeplearning4j.nn.layers.BaseLayer;
 import org.deeplearning4j.nn.layers.mkldnn.MKLDNNConvHelper;
 import org.deeplearning4j.nn.workspace.ArrayType;
 import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
+import org.nd4j.base.Preconditions;
 import org.nd4j.linalg.activations.IActivation;
 import org.nd4j.linalg.activations.impl.ActivationSigmoid;
 import org.nd4j.linalg.api.memory.MemoryWorkspace;
@@ -188,12 +189,9 @@ public class LSTMHelpers {
         }
         //Input validation: check that if past state is provided, that it has same
         //These can be different if user forgets to call rnnClearPreviousState() between calls of rnnTimeStep
-        if (prevOutputActivations != null && prevOutputActivations.size(0) != input.size(0)) {
-            throw new DL4JInvalidInputException("Previous activations (stored state) number of examples = "
-                            + prevOutputActivations.size(0) + " but input array number of examples = " + input.size(0)
-                            + ". Possible cause: using rnnTimeStep() without calling"
-                            + " rnnClearPreviousState() between different sequences?");
-        }
+        Preconditions.checkState(prevOutputActivations == null || prevOutputActivations.size(0) == input.size(0),
+                "Invalid RNN previous state (last time step activations/initialization): rnnTimeStep with different minibatch size, or forgot to call rnnClearPreviousState between batches?" +
+                        " Previous step output = [batch, nIn] = %ndShape, current input = [batch, nIn, seqLength] = %ndShape", prevOutputActivations, input);
 
         //initialize prevOutputActivations to zeroes
         if (prevOutputActivations == null) {
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/SimpleRnn.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/SimpleRnn.java
index e4bdcf56f..044f444c0 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/SimpleRnn.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/SimpleRnn.java
@@ -217,6 +217,9 @@ public class SimpleRnn extends BaseRecurrentLayer<org.deeplearning4j.nn.conf.lay
         assertInputSet(false);
         Preconditions.checkState(input.rank() == 3,
                 "3D input expected to RNN layer expected, got " + input.rank());
+        Preconditions.checkState(prevStepOut == null || prevStepOut.size(0) == input.size(0),
+                "Invalid RNN previous state (last time step activations/initialization): rnnTimeStep with different minibatch size, or forgot to call rnnClearPreviousState between batches?" +
+                        " Previous step output = [batch, nIn] = %ndShape, current input = [batch, nIn, seqLength] = %ndShape", prevStepOut, input);
 
         applyDropOutIfNecessary(training, workspaceMgr);
 
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/multilayer/MultiLayerNetwork.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/multilayer/MultiLayerNetwork.java
index dd495a620..24de38f56 100755
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/multilayer/MultiLayerNetwork.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/multilayer/MultiLayerNetwork.java
@@ -436,8 +436,11 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
 
     @Override
     public int batchSize() {
+        //In 99+% of cases, the input and labels dimension 0 size should be identical
+        //The only real exceptions: space to batch, and batch to space layers
+        //In those cases, we should base it on the labels size, as this impacts gradient calculation
         // FIXME: int cast
-        return (int) input.size(0);
+        return labels == null ? (int) input.size(0) : (int)labels.size(0);
     }
 
     @Override
@@ -1362,6 +1365,13 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
 
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
 
+            if(t != null){
+                if(t instanceof RuntimeException){
+                    throw ((RuntimeException)t);
+                }
+                throw new RuntimeException("Error during neural network forward pass", t);
+            }
+
             if(outputWorkspace == null || outputWorkspace instanceof DummyWorkspace) {
                 WorkspaceUtils.assertNoWorkspacesOpen("Expected no workspace active at the end of outputOfLayerDetached", true);
             } else {
@@ -2007,6 +2017,13 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
                 }
             }
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
+
+            if(t != null){
+                if(t instanceof RuntimeException){
+                    throw ((RuntimeException)t);
+                }
+                throw new RuntimeException("Error during neural network forward pass", t);
+            }
         }
 
         if (layerWiseConfigurations.getTrainingWorkspaceMode() == WorkspaceMode.NONE) {

From f414239ed538a221af5286f75dbd883c939fe6a5 Mon Sep 17 00:00:00 2001
From: Alexander Stoyakin <alexander.stoyakin@gmail.com>
Date: Fri, 30 Aug 2019 16:18:41 +0300
Subject: [PATCH 50/56] Types fixed (#205)

Signed-off-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
---
 .../test/java/org/deeplearning4j/plot/BarnesHutTsneTest.java | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/plot/BarnesHutTsneTest.java b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/plot/BarnesHutTsneTest.java
index aba60fe4f..f65783bce 100644
--- a/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/plot/BarnesHutTsneTest.java
+++ b/deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/plot/BarnesHutTsneTest.java
@@ -498,9 +498,10 @@ public class BarnesHutTsneTest extends BaseDL4JTest {
                     0.08651136699915507, 0.7445210640026082, 0.6547649514127559, 0.3384719042666908, 0.05816723105860, 0.6248951423054205, 0.7431868493349041};
             INDArray ndinput = Nd4j.createFromArray(input).reshape(11, 5);
 
-            double[] rows = {0, 10.0000, 20.0000, 30.0000, 40.0000, 50.0000, 60.0000, 69.0000, 78.0000, 88.0000, 98.0000, 108.0000};
+            int[] rows = {0, 10, 20, 30, 40, 50, 60, 69, 78, 88, 98, 108};
             INDArray indRows = Nd4j.createFromArray(rows);
-            double[] cols = {4.0000, 3.0000, 10.0000, 8.0000, 6.0000, 7.0000, 1.0000, 5.0000, 9.0000, 2.0000, 0, 4.0000, 9.0000, 8.0000, 10.0000, 2.0000, 6.0000, 7.0000, 3.0000, 5.0000, 1.0000, 6.0000, 8.0000, 3.0000, 9.0000, 10.0000, 4.0000, 0, 5.0000, 7.0000, 0, 1.0000, 2.0000, 10.0000, 4.0000, 6.0000, 8.0000, 9.0000, 5.0000, 7.0000, 0, 1.0000, 2.0000, 3.0000, 10.0000, 8.0000, 9.0000, 6.0000, 7.0000, 5.0000, 0, 2.0000, 3.0000, 7.0000, 9.0000, 10.0000, 4.0000, 8.0000, 1.0000, 6.0000, 0, 1.0000, 2.0000, 3.0000, 4.0000, 8.0000, 10.0000, 9.0000, 5.0000, 0, 1.0000, 3.0000, 4.0000, 5.0000, 9.0000, 10.0000, 8.0000, 2.0000, 0, 1.0000, 2.0000, 3.0000, 4.0000, 5.0000, 6.0000, 7.0000, 10.0000, 9.0000, 0, 1.0000, 2.0000, 3.0000, 4.0000, 5.0000, 6.0000, 7.0000, 8.0000, 10.0000, 0, 1.0000, 2.0000, 3.0000, 4.0000, 5.0000, 6.0000, 7.0000, 8.0000, 9.0000};
+            int[] cols = {4, 3, 10, 8, 6, 7, 1, 5, 9, 2, 0, 4, 9, 8, 10, 2, 6, 7, 3, 5, 1, 6, 8, 3, 9, 10, 4, 0, 5, 7, 0, 1, 2, 10, 4, 6, 8, 9,
+                    5, 7, 0, 1, 2, 3, 10, 8, 9, 6, 7, 5, 0, 2, 3, 7, 9, 10, 4, 8, 1, 6, 0, 1, 2, 3, 4, 8, 10, 9, 5, 0, 1, 3, 4, 5, 9, 10, 8, 2, 0, 1, 2, 3, 4, 5, 6, 7, 10, 9, 0, 1, 2, 3, 4, 5, 6, 7, 8, 10, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
             INDArray indCols = Nd4j.createFromArray(cols);
             double[] vals = {0.6806, 0.1978, 0.1349, 0.0403, 0.0087, 0.0369, 0.0081, 0.0172, 0.0014, 0.0046, 0.0081, 0.3375, 0.2274, 0.0556, 0.0098, 0.0175, 0.0027, 0.0077, 0.0014, 0.0023, 0.0175, 0.6569, 0.1762, 0.0254, 0.0200, 0.0118, 0.0074, 0.0046, 0.0124, 0.0012, 0.1978, 0.0014, 0.0254, 0.7198, 0.0712, 0.0850, 0.0389, 0.0555, 0.0418, 0.0286, 0.6806, 0.3375, 0.0074, 0.0712, 0.2290, 0.0224, 0.0189, 0.0080, 0.0187, 0.0097, 0.0172, 0.0124, 0.0418, 0.7799, 0.0521, 0.0395, 0.0097, 0.0030, 0.0023, 1.706e-5, 0.0087, 0.0027, 0.6569, 0.0850, 0.0080, 0.5562, 0.0173, 0.0015, 1.706e-5, 0.0369, 0.0077, 0.0286, 0.0187, 0.7799, 0.0711, 0.0200, 0.0084, 0.0012, 0.0403, 0.0556, 0.1762, 0.0389, 0.0224, 0.0030, 0.5562, 0.0084, 0.0060, 0.0028, 0.0014, 0.2274, 0.0200, 0.0555, 0.0189, 0.0521, 0.0015, 0.0711, 0.0028, 0.3911, 0.1349, 0.0098, 0.0118, 0.7198, 0.2290, 0.0395, 0.0173, 0.0200, 0.0060, 0.3911};
             INDArray indVals = Nd4j.createFromArray(vals);

From 70a9ae5068610221f4f168ce669c69a3e204847c Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Fri, 30 Aug 2019 16:32:01 +0300
Subject: [PATCH 51/56] [WIP] few tweaks (#206)

* scatter empty check

Signed-off-by: raver119 <raver119@gmail.com>

* scatter empty test

Signed-off-by: raver119 <raver119@gmail.com>

* one more test

Signed-off-by: raver119 <raver119@gmail.com>

* two tweaks

Signed-off-by: raver119 <raver119@gmail.com>

* dup tweak

Signed-off-by: raver119 <raver119@gmail.com>

* - put empty checking of indices array immediately prior  helper run

Signed-off-by: Yurii <yurii@skymind.io>

* minor tests fix

Signed-off-by: raver119 <raver119@gmail.com>

* minor tests fix

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../generic/parity_ops/scatter_add.cpp        |  9 +++--
 .../generic/parity_ops/scatter_div.cpp        | 30 +++++++-------
 .../generic/parity_ops/scatter_max.cpp        | 25 ++++++------
 .../generic/parity_ops/scatter_min.cpp        | 25 ++++++------
 .../generic/parity_ops/scatter_mul.cpp        | 26 ++++++------
 .../generic/parity_ops/scatter_sub.cpp        | 27 +++++++------
 .../generic/parity_ops/scatter_upd.cpp        | 27 +++++++------
 libnd4j/tests_cpu/layers_tests/EmptyTests.cpp | 40 +++++++++++++++++++
 .../nd4j/linalg/api/ndarray/BaseNDArray.java  |  7 +++-
 9 files changed, 134 insertions(+), 82 deletions(-)

diff --git a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_add.cpp b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_add.cpp
index 0bcbd2439..0d9465b02 100644
--- a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_add.cpp
+++ b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_add.cpp
@@ -35,6 +35,9 @@ OP_IMPL(scatter_add, 3, 1, true) {
 
     auto output = OUTPUT_VARIABLE(0);
 
+    if (!block.isInplace())
+        output->assign(input);
+
     const bool lock = block.getBArguments()->empty() ? false : B_ARG(0);
 
     const int inRank  = input->rankOf();
@@ -68,10 +71,8 @@ OP_IMPL(scatter_add, 3, 1, true) {
         REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_ADD OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
     }
 
-    if (!block.isInplace())
-        output->assign(input);
-
-    helpers::scatter(block.launchContext(), pairwise::Add, *indices, *updates, *output, lock);
+    if (!indices->isEmpty())
+        helpers::scatter(block.launchContext(), pairwise::Add, *indices, *updates, *output, lock);
 
     return Status::OK();
 }
diff --git a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_div.cpp b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_div.cpp
index a711916a1..dccc34e59 100644
--- a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_div.cpp
+++ b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_div.cpp
@@ -16,7 +16,7 @@
 
 //
 // @author Created by raver119 on 24.11.17.
-// @author Yurii Shyrma (iuriish@yahoo.com)    
+// @author Yurii Shyrma (iuriish@yahoo.com)
 //
 
 #include <op_boilerplate.h>
@@ -28,21 +28,24 @@
 namespace nd4j {
     namespace ops {
         OP_IMPL(scatter_div, 3, 1, true) {
-            
+
             auto input = INPUT_VARIABLE(0);
             auto indices = INPUT_VARIABLE(1);
             auto updates = INPUT_VARIABLE(2);
 
             auto output = OUTPUT_VARIABLE(0);
 
+            if (!block.isInplace())
+                output->assign(input);
+
             const bool lock = block.getBArguments()->empty() ? false : B_ARG(0);
 
             const int inRank  = input->rankOf();
             const int indRank = indices->rankOf();
             const int updRank = updates->rankOf();
-    
+
             REQUIRE_TRUE(inRank > 0, 0, "SCATTER_DIV OP: input should not be scalar !");
-    
+
             if(inRank == 1) {
                 REQUIRE_TRUE(indices->isSameShape(updates), 0, "SCATTER_DIV OP: when input array has rank = 1 then indices and updates must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(indices).c_str(), ShapeUtils::shapeAsString(updates).c_str());
             }
@@ -50,28 +53,27 @@ namespace nd4j {
 
                 std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
                 std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-                std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()}; 
+                std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()};
                 expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
                 REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_DIV OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
             }
             else {
-        
+
                 REQUIRE_TRUE(updRank == indRank + inRank - 1, 0, "SCATTER_DIV OP: wrong rank of updates array, expected is %i, but got %i instead !", indRank + inRank - 1 , updRank);
-                
+
                 std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
                 std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-                std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();        
+                std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();
                 expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
                 REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_DIV OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
             }
 
-            if (!block.isInplace())
-                output->assign(input);
+            if (!indices->isEmpty())
+                // ScatterHelper<T>::template scatterApply<simdOps::Divide<T>>(output, indices, updates);
+                helpers::scatter(block.launchContext(), pairwise::Divide, *indices, *updates, *output, lock);
 
-            // ScatterHelper<T>::template scatterApply<simdOps::Divide<T>>(output, indices, updates);        
-            helpers::scatter(block.launchContext(), pairwise::Divide, *indices, *updates, *output, lock);
             return Status::OK();
         }
         DECLARE_SYN(ScatterDiv, scatter_div);
diff --git a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_max.cpp b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_max.cpp
index a9f0ab889..5d37a71d0 100644
--- a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_max.cpp
+++ b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_max.cpp
@@ -35,14 +35,17 @@ OP_IMPL(scatter_max, 3, 1, true) {
 
     auto output = OUTPUT_VARIABLE(0);
 
+    if (!block.isInplace())
+        output->assign(input);
+
     const bool lock = block.getBArguments()->empty() ? false : B_ARG(0);
 
     const int inRank  = input->rankOf();
     const int indRank = indices->rankOf();
     const int updRank = updates->rankOf();
-    
+
     REQUIRE_TRUE(inRank > 0, 0, "SCATTER_MAX OP: input should not be scalar !");
-    
+
     if(inRank == 1) {
         REQUIRE_TRUE(indices->isSameShape(updates), 0, "SCATTER_MAX OP: when input array has rank = 1 then indices and updates must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(indices).c_str(), ShapeUtils::shapeAsString(updates).c_str());
     }
@@ -50,28 +53,26 @@ OP_IMPL(scatter_max, 3, 1, true) {
 
         std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
         std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-        std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()}; 
+        std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()};
         expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
         REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_MAX OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
     }
     else {
-        
+
         REQUIRE_TRUE(updRank == indRank + inRank - 1, 0, "SCATTER_MAX OP: wrong rank of updates array, expected is %i, but got %i instead !", indRank + inRank - 1 , updRank);
-                
+
         std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
         std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-        std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();        
+        std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();
         expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
         REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_MAX OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
     }
 
-    if (!block.isInplace())
-            output->assign(input);
+    if (!indices->isEmpty())
+        helpers::scatter(block.launchContext(), pairwise::MaxPairwise, *indices, *updates, *output, lock);
 
-    helpers::scatter(block.launchContext(), pairwise::MaxPairwise, *indices, *updates, *output, lock);
-    
     return Status::OK();
 }
 DECLARE_SYN(ScatterMax, scatter_max);
diff --git a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_min.cpp b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_min.cpp
index cce22b6fb..1bed296f9 100644
--- a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_min.cpp
+++ b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_min.cpp
@@ -35,14 +35,17 @@ OP_IMPL(scatter_min, 3, 1, true) {
 
     auto output = OUTPUT_VARIABLE(0);
 
+    if (!block.isInplace())
+        output->assign(input);
+
     const bool lock = block.getBArguments()->empty() ? false : B_ARG(0);
 
     const int inRank  = input->rankOf();
     const int indRank = indices->rankOf();
     const int updRank = updates->rankOf();
-    
+
     REQUIRE_TRUE(inRank > 0, 0, "SCATTER_MIN OP: input should not be scalar !");
-    
+
     if(inRank == 1) {
         REQUIRE_TRUE(indices->isSameShape(updates), 0, "SCATTER_MIN OP: when input array has rank = 1 then indices and updates must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(indices).c_str(), ShapeUtils::shapeAsString(updates).c_str());
     }
@@ -50,27 +53,25 @@ OP_IMPL(scatter_min, 3, 1, true) {
 
         std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
         std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-        std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()}; 
+        std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()};
         expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
         REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_MIN OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
     }
     else {
-        
+
         REQUIRE_TRUE(updRank == indRank + inRank - 1, 0, "SCATTER_MIN OP: wrong rank of updates array, expected is %i, but got %i instead !", indRank + inRank - 1 , updRank);
-                
+
         std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
         std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-        std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();        
+        std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();
         expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
         REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_MIN OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
     }
 
-    if (!block.isInplace())
-            output->assign(input);
-
-    helpers::scatter(block.launchContext(), pairwise::MinPairwise, *indices, *updates, *output, lock);
+    if (!indices->isEmpty())
+        helpers::scatter(block.launchContext(), pairwise::MinPairwise, *indices, *updates, *output, lock);
 
     return Status::OK();
 }
diff --git a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_mul.cpp b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_mul.cpp
index 02eebb50c..46b9f7008 100644
--- a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_mul.cpp
+++ b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_mul.cpp
@@ -39,9 +39,13 @@ namespace nd4j {
             const int inRank  = input->rankOf();
             const int indRank = indices->rankOf();
             const int updRank = updates->rankOf();
-    
+
+            if (!block.isInplace())
+                output->assign(input);
+
+
             REQUIRE_TRUE(inRank > 0, 0, "SCATTER_MUL OP: input should not be scalar !");
-    
+
             if(inRank == 1) {
                 REQUIRE_TRUE(indices->isSameShape(updates), 0, "SCATTER_MUL OP: when input array has rank = 1 then indices and updates must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(indices).c_str(), ShapeUtils::shapeAsString(updates).c_str());
             }
@@ -49,27 +53,25 @@ namespace nd4j {
 
                 std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
                 std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-                std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()}; 
+                std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()};
                 expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
                 REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_MUL OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
             }
             else {
-        
+
                 REQUIRE_TRUE(updRank == indRank + inRank - 1, 0, "SCATTER_MUL OP: wrong rank of updates array, expected is %i, but got %i instead !", indRank + inRank - 1 , updRank);
-                
+
                 std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
                 std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-                std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();        
+                std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();
                 expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
                 REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_MUL OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
             }
 
-            if (!block.isInplace())
-                output->assign(input);
-
-            helpers::scatter(block.launchContext(), pairwise::Multiply, *indices, *updates, *output, lock);
+            if (!indices->isEmpty())
+                helpers::scatter(block.launchContext(), pairwise::Multiply, *indices, *updates, *output, lock);
 
             return Status::OK();
         }
diff --git a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_sub.cpp b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_sub.cpp
index de2bf4fa4..cf3745236 100644
--- a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_sub.cpp
+++ b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_sub.cpp
@@ -34,14 +34,17 @@ namespace nd4j {
 
             auto output = OUTPUT_VARIABLE(0);
 
+            if (!block.isInplace())
+                output->assign(input);
+
             const bool lock = block.getBArguments()->empty() ? false : B_ARG(0);
 
             const int inRank  = input->rankOf();
             const int indRank = indices->rankOf();
             const int updRank = updates->rankOf();
-    
+
             REQUIRE_TRUE(inRank > 0, 0, "SCATTER_SUB OP: input should not be scalar !");
-    
+
             if(inRank == 1) {
                 REQUIRE_TRUE(indices->isSameShape(updates), 0, "SCATTER_SUB OP: when input array has rank = 1 then indices and updates must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(indices).c_str(), ShapeUtils::shapeAsString(updates).c_str());
             }
@@ -49,29 +52,27 @@ namespace nd4j {
 
                 std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
                 std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-                std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()}; 
+                std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()};
                 expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
                 REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_SUB OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
             }
 
             else {
-        
+
                 REQUIRE_TRUE(updRank == indRank + inRank - 1, 0, "SCATTER_SUB OP: wrong rank of updates array, expected is %i, but got %i instead !", indRank + inRank - 1 , updRank);
-                
+
                 std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
                 std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-                std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();        
+                std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();
                 expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
                 REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_SUB OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
             }
 
-            if (!block.isInplace())
-                output->assign(input);
-
-            // ScatterHelper<T>::template scatterApply<simdOps::Subtract<T>>(output, indices, updates);        
-            helpers::scatter(block.launchContext(), pairwise::Subtract, *indices, *updates, *output, lock);
+            if (!indices->isEmpty())
+                // ScatterHelper<T>::template scatterApply<simdOps::Subtract<T>>(output, indices, updates);
+                helpers::scatter(block.launchContext(), pairwise::Subtract, *indices, *updates, *output, lock);
 
             return Status::OK();
         }
diff --git a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_upd.cpp b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_upd.cpp
index bc13581bf..55076e51e 100644
--- a/libnd4j/include/ops/declarable/generic/parity_ops/scatter_upd.cpp
+++ b/libnd4j/include/ops/declarable/generic/parity_ops/scatter_upd.cpp
@@ -33,14 +33,17 @@ namespace nd4j {
 
             auto output = OUTPUT_VARIABLE(0);
 
+            if (!block.isInplace())
+                output->assign(input);
+
             const bool lock = block.getBArguments()->empty() ? true : B_ARG(0);
 
             const int inRank  = input->rankOf();
             const int indRank = indices->rankOf();
             const int updRank = updates->rankOf();
-    
+
             REQUIRE_TRUE(inRank > 0, 0, "SCATTER_UPD OP: input should not be scalar !");
-    
+
             if(inRank == 1) {
                 REQUIRE_TRUE(indices->isSameShape(updates), 0, "SCATTER_UPD OP: when input array has rank = 1 then indices and updates must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(indices).c_str(), ShapeUtils::shapeAsString(updates).c_str());
             }
@@ -48,28 +51,26 @@ namespace nd4j {
 
                 std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
                 std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-                std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()}; 
+                std::vector<Nd4jLong> expectedUpdShape = {indices->lengthOf()};
                 expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
                 REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_UPD OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
             }
             else {
-        
+
                 REQUIRE_TRUE(updRank == indRank + inRank - 1, 0, "SCATTER_UPD OP: wrong rank of updates array, expected is %i, but got %i instead !", indRank + inRank - 1 , updRank);
-                
+
                 std::vector<Nd4jLong> updShape = updates->getShapeAsVector();
                 std::vector<Nd4jLong> inShape  = input->getShapeAsVector();
-                std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();        
+                std::vector<Nd4jLong> expectedUpdShape = indices->getShapeAsVector();
                 expectedUpdShape.insert(expectedUpdShape.end(), inShape.begin()+1, inShape.end());
-        
+
                 REQUIRE_TRUE(expectedUpdShape == updShape, 0, "SCATTER_UPD OP: wrong shape of updates array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedUpdShape).c_str(), ShapeUtils::shapeAsString(updShape).c_str());
             }
 
-            if (!block.isInplace())
-                output->assign(input);
-
-            // ScatterHelper<T>::template scatterApply<simdOps::Copy<T>>(output, indices, updates);        
-            helpers::scatter(block.launchContext(), pairwise::CopyPws, *indices, *updates, *output, lock);
+            if (!indices->isEmpty())
+                // ScatterHelper<T>::template scatterApply<simdOps::Copy<T>>(output, indices, updates);
+                helpers::scatter(block.launchContext(), pairwise::CopyPws, *indices, *updates, *output, lock);
 
             return Status::OK();
         }
diff --git a/libnd4j/tests_cpu/layers_tests/EmptyTests.cpp b/libnd4j/tests_cpu/layers_tests/EmptyTests.cpp
index b9445cc70..baba901bf 100644
--- a/libnd4j/tests_cpu/layers_tests/EmptyTests.cpp
+++ b/libnd4j/tests_cpu/layers_tests/EmptyTests.cpp
@@ -207,6 +207,46 @@ TEST_F(EmptyTests, Test_dup_1) {
     delete dup;
 }
 
+TEST_F(EmptyTests, test_empty_scatter_1) {
+    auto x = NDArrayFactory::create<float>('c', {5});
+    auto indices = NDArrayFactory::create<int>('c', {0});
+    auto updates = NDArrayFactory::create<float>('c', {0});
+
+    x.linspace(1.0f);
+
+    nd4j::ops::scatter_upd op;
+    auto result = op.execute({&x, &indices, &updates}, {}, {}, {true});
+    ASSERT_EQ(Status::OK(), result->status());
+
+    auto z = result->at(0);
+    ASSERT_EQ(x, *z);
+
+    delete result;
+}
+
+TEST_F(EmptyTests, test_empty_scatter_2) {
+    auto x = NDArrayFactory::create<float>('c', {5});
+    auto z = NDArrayFactory::create<float>('c', {5});
+    auto indices = NDArrayFactory::create<int>('c', {0});
+    auto updates = NDArrayFactory::create<float>('c', {0});
+
+    x.linspace(1.0f);
+
+    Context ctx(1);
+    ctx.setInputArray(0, x.buffer(), x.shapeInfo(), x.specialBuffer(), x.specialShapeInfo());
+    ctx.setInputArray(1, indices.buffer(), indices.shapeInfo(), indices.specialBuffer(), indices.specialShapeInfo());
+    ctx.setInputArray(2, updates.buffer(), updates.shapeInfo(), updates.specialBuffer(), updates.specialShapeInfo());
+    ctx.setOutputArray(0, z.buffer(), z.shapeInfo(), z.specialBuffer(), z.specialShapeInfo());
+    bool args[] = {true};
+    ctx.setBArguments(args, 1);
+
+    nd4j::ops::scatter_upd op;
+    auto result = op.execute(&ctx);
+    ASSERT_EQ(Status::OK(), result);
+
+    ASSERT_EQ(x, z);
+}
+
 TEST_F(EmptyTests, test_shaped_empty_1) {
     auto empty = NDArrayFactory::create<float>('c', {2, 0, 3});
     std::vector<Nd4jLong> shape = {2, 0, 3};
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
index 95a087301..46dd786c6 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ndarray/BaseNDArray.java
@@ -1732,6 +1732,8 @@ public abstract class BaseNDArray implements INDArray, Iterable {
         if(isEmpty())
             return this;
 
+        Nd4j.getCompressor().autoDecompress(this);
+
         // fixme: eventually it would be nice to have this in native code
         if (isS()) {
             val list = new ArrayList<String>();
@@ -1741,8 +1743,9 @@ public abstract class BaseNDArray implements INDArray, Iterable {
             return Nd4j.create(list, this.shape(), this.ordering());
         }
 
-        Nd4j.getCompressor().autoDecompress(this);
-        return Shape.toOffsetZeroCopy(this, order);
+        val z = Nd4j.createUninitialized(this.dataType(), this.shape(), order);
+        z.assign(this);
+        return z;
     }
 
     /**

From a7dca9fc8792819305735d41d447331fba696e96 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Fri, 30 Aug 2019 23:36:00 +1000
Subject: [PATCH 52/56] Add java op class for relu derivative, and use in
 Activation ReLU (#207)

* Add java op class for relu derivative, and use in ACtivation ReLU

Signed-off-by: AlexDBlack <blacka101@gmail.com>

* Fix

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 .../DifferentialFunctionFactory.java          | 22 +++-------
 .../converters/ImportClassMapping.java        |  1 +
 .../activations/impl/ActivationReLU.java      |  4 +-
 .../api/ops/impl/scalar/RectifiedLinear.java  | 11 +++--
 .../scalar/RectifiedLinearDerivative.java     | 43 +++++++++++++++++++
 5 files changed, 60 insertions(+), 21 deletions(-)
 create mode 100644 nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/scalar/RectifiedLinearDerivative.java

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
index 035ff0960..3343d9dde 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
@@ -133,23 +133,8 @@ import org.nd4j.linalg.api.ops.impl.reduce3.EuclideanDistance;
 import org.nd4j.linalg.api.ops.impl.reduce3.HammingDistance;
 import org.nd4j.linalg.api.ops.impl.reduce3.JaccardDistance;
 import org.nd4j.linalg.api.ops.impl.reduce3.ManhattanDistance;
-import org.nd4j.linalg.api.ops.impl.scalar.LeakyReLU;
-import org.nd4j.linalg.api.ops.impl.scalar.LogX;
+import org.nd4j.linalg.api.ops.impl.scalar.*;
 import org.nd4j.linalg.api.ops.impl.scalar.Pow;
-import org.nd4j.linalg.api.ops.impl.scalar.PowDerivative;
-import org.nd4j.linalg.api.ops.impl.scalar.RectifiedLinear;
-import org.nd4j.linalg.api.ops.impl.scalar.Relu6;
-import org.nd4j.linalg.api.ops.impl.scalar.ScalarAdd;
-import org.nd4j.linalg.api.ops.impl.scalar.ScalarDivision;
-import org.nd4j.linalg.api.ops.impl.scalar.ScalarFMod;
-import org.nd4j.linalg.api.ops.impl.scalar.ScalarMax;
-import org.nd4j.linalg.api.ops.impl.scalar.ScalarMin;
-import org.nd4j.linalg.api.ops.impl.scalar.ScalarMultiplication;
-import org.nd4j.linalg.api.ops.impl.scalar.ScalarReverseDivision;
-import org.nd4j.linalg.api.ops.impl.scalar.ScalarReverseSubtraction;
-import org.nd4j.linalg.api.ops.impl.scalar.ScalarSet;
-import org.nd4j.linalg.api.ops.impl.scalar.ScalarSubtraction;
-import org.nd4j.linalg.api.ops.impl.scalar.Step;
 import org.nd4j.linalg.api.ops.impl.scalar.comparison.ScalarEquals;
 import org.nd4j.linalg.api.ops.impl.scalar.comparison.ScalarGreaterThan;
 import org.nd4j.linalg.api.ops.impl.scalar.comparison.ScalarGreaterThanOrEqual;
@@ -211,6 +196,7 @@ import org.nd4j.linalg.api.ops.impl.transforms.clip.ClipByValue;
 import org.nd4j.linalg.api.ops.impl.transforms.comparison.CompareAndReplace;
 import org.nd4j.linalg.api.ops.impl.transforms.comparison.CompareAndSet;
 import org.nd4j.linalg.api.ops.impl.transforms.custom.*;
+import org.nd4j.linalg.api.ops.impl.transforms.custom.Pow;
 import org.nd4j.linalg.api.ops.impl.transforms.custom.segment.SegmentMax;
 import org.nd4j.linalg.api.ops.impl.transforms.custom.segment.SegmentMean;
 import org.nd4j.linalg.api.ops.impl.transforms.custom.segment.SegmentMin;
@@ -1340,6 +1326,10 @@ public class DifferentialFunctionFactory {
         return new RectifiedLinear(sameDiff(), iX, false, cutoff).outputVariable();
     }
 
+    public SDVariable reluDerivative(SDVariable input, SDVariable grad){
+        return new RectifiedLinearDerivative(sameDiff(), input, grad).outputVariable();
+    }
+
     public SDVariable relu6(SDVariable iX, double cutoff) {
         return new Relu6(sameDiff(), iX, false, cutoff).outputVariable();
     }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/ImportClassMapping.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/ImportClassMapping.java
index e2d3cf87b..da580b748 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/ImportClassMapping.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/imports/converters/ImportClassMapping.java
@@ -228,6 +228,7 @@ public class ImportClassMapping {
             org.nd4j.linalg.api.ops.impl.scalar.Pow.class,
             org.nd4j.linalg.api.ops.impl.scalar.PowDerivative.class,
             org.nd4j.linalg.api.ops.impl.scalar.RectifiedLinear.class,
+            org.nd4j.linalg.api.ops.impl.scalar.RectifiedLinearDerivative.class,
             org.nd4j.linalg.api.ops.impl.scalar.Relu6.class,
             org.nd4j.linalg.api.ops.impl.scalar.ReplaceNans.class,
             org.nd4j.linalg.api.ops.impl.scalar.ScalarAdd.class,
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationReLU.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationReLU.java
index 36313593d..cf0ec7fa0 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationReLU.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationReLU.java
@@ -18,6 +18,7 @@ package org.nd4j.linalg.activations.impl;
 
 import lombok.EqualsAndHashCode;
 import lombok.Getter;
+import org.nd4j.linalg.api.ops.impl.scalar.RectifiedLinearDerivative;
 import org.nd4j.linalg.api.ops.impl.scalar.Step;
 import org.nd4j.linalg.primitives.Pair;
 import org.nd4j.linalg.activations.BaseActivationFunction;
@@ -41,8 +42,7 @@ public class ActivationReLU extends BaseActivationFunction {
     @Override
     public Pair<INDArray, INDArray> backprop(INDArray in, INDArray epsilon) {
         assertShape(in, epsilon);
-        INDArray dLdz = Nd4j.getExecutioner().exec(new Step(in));
-        dLdz.muli(epsilon);
+        INDArray dLdz = Nd4j.exec(new RectifiedLinearDerivative(in, epsilon, in.ulike()))[0];
         return new Pair<>(dLdz, null);
     }
 
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/scalar/RectifiedLinear.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/scalar/RectifiedLinear.java
index 15673a5a0..98fa587b5 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/scalar/RectifiedLinear.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/scalar/RectifiedLinear.java
@@ -22,6 +22,7 @@ import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.api.ops.BaseScalarOp;
 
 import java.util.Arrays;
+import java.util.Collections;
 import java.util.List;
 
 /**
@@ -77,8 +78,12 @@ public class RectifiedLinear extends BaseScalarOp {
 
     @Override
     public List<SDVariable> doDiff(List<SDVariable> i_v) {
-        SDVariable step = new Step(sameDiff, arg(), false, scalarValue.getDouble(0)).outputVariables()[0];
-        SDVariable ret = step.mul(i_v.get(0));
-        return Arrays.asList(ret);
+        if(scalarValue.getDouble(0) == 0.0){
+            return Collections.singletonList(f().reluDerivative(arg(), i_v.get(0)));
+        } else {
+            SDVariable step = new Step(sameDiff, arg(), false, scalarValue.getDouble(0)).outputVariables()[0];
+            SDVariable ret = step.mul(i_v.get(0));
+            return Collections.singletonList(ret);
+        }
     }
 }
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/scalar/RectifiedLinearDerivative.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/scalar/RectifiedLinearDerivative.java
new file mode 100644
index 000000000..3af7a4190
--- /dev/null
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/scalar/RectifiedLinearDerivative.java
@@ -0,0 +1,43 @@
+package org.nd4j.linalg.api.ops.impl.scalar;
+
+import lombok.NonNull;
+import org.nd4j.autodiff.samediff.SDVariable;
+import org.nd4j.autodiff.samediff.SameDiff;
+import org.nd4j.linalg.api.buffer.DataType;
+import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.api.ops.DynamicCustomOp;
+import org.nd4j.shade.guava.base.Preconditions;
+
+import java.util.Collections;
+import java.util.List;
+
+public class RectifiedLinearDerivative extends DynamicCustomOp {
+
+    public RectifiedLinearDerivative(){ }
+
+    public RectifiedLinearDerivative(SameDiff sd, SDVariable input, SDVariable gradient){
+        super(sd, new SDVariable[]{input, gradient});
+    }
+
+    public RectifiedLinearDerivative(@NonNull INDArray input, @NonNull INDArray gradient, INDArray output){
+        super(new INDArray[]{input, gradient}, wrapOrNull(output));
+    }
+
+    @Override
+    public String opName(){
+        return "relu_bp";
+    }
+
+    @Override
+    public List<DataType> calculateOutputDataTypes(List<DataType> dataTypes) {
+        Preconditions.checkArgument(dataTypes != null && dataTypes.size() == 2, "Expected exactly 2 input datatypes, got %s", dataTypes);
+        Preconditions.checkArgument(dataTypes.get(0).isFPType() && dataTypes.get(1).isFPType(), "Input datatypes must be floating point, got %s", dataTypes);
+
+        return Collections.singletonList(dataTypes.get(0));
+    }
+
+    @Override
+    public List<SDVariable> doDiff(List<SDVariable> f1) {
+        throw new UnsupportedOperationException("Not supported");
+    }
+}

From 6efffb727f053ca7f8c8684e56bb0650ab635bc6 Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Fri, 30 Aug 2019 23:40:52 +1000
Subject: [PATCH 53/56] Import fix (#208)

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 .../org/nd4j/autodiff/functions/DifferentialFunctionFactory.java | 1 -
 .../java/org/nd4j/linalg/activations/impl/ActivationReLU.java    | 1 -
 2 files changed, 2 deletions(-)

diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
index 3343d9dde..7ffeca762 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/functions/DifferentialFunctionFactory.java
@@ -196,7 +196,6 @@ import org.nd4j.linalg.api.ops.impl.transforms.clip.ClipByValue;
 import org.nd4j.linalg.api.ops.impl.transforms.comparison.CompareAndReplace;
 import org.nd4j.linalg.api.ops.impl.transforms.comparison.CompareAndSet;
 import org.nd4j.linalg.api.ops.impl.transforms.custom.*;
-import org.nd4j.linalg.api.ops.impl.transforms.custom.Pow;
 import org.nd4j.linalg.api.ops.impl.transforms.custom.segment.SegmentMax;
 import org.nd4j.linalg.api.ops.impl.transforms.custom.segment.SegmentMean;
 import org.nd4j.linalg.api.ops.impl.transforms.custom.segment.SegmentMin;
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationReLU.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationReLU.java
index cf0ec7fa0..eebee472c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationReLU.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationReLU.java
@@ -19,7 +19,6 @@ package org.nd4j.linalg.activations.impl;
 import lombok.EqualsAndHashCode;
 import lombok.Getter;
 import org.nd4j.linalg.api.ops.impl.scalar.RectifiedLinearDerivative;
-import org.nd4j.linalg.api.ops.impl.scalar.Step;
 import org.nd4j.linalg.primitives.Pair;
 import org.nd4j.linalg.activations.BaseActivationFunction;
 import org.nd4j.linalg.api.ndarray.INDArray;

From bdc3eacafde4026371b7097fc04a537f46662d73 Mon Sep 17 00:00:00 2001
From: raver119 <raver119@gmail.com>
Date: Fri, 30 Aug 2019 20:13:01 +0300
Subject: [PATCH 54/56] one small playground test

Signed-off-by: raver119 <raver119@gmail.com>
---
 .../declarable/helpers/cpu/legacy_helper.cpp  | 21 +++++++++++++--
 .../layers_tests/PlaygroundTests.cpp          | 26 +++++++++++++++++++
 2 files changed, 45 insertions(+), 2 deletions(-)

diff --git a/libnd4j/include/ops/declarable/helpers/cpu/legacy_helper.cpp b/libnd4j/include/ops/declarable/helpers/cpu/legacy_helper.cpp
index 45024b5cb..d673e64bd 100644
--- a/libnd4j/include/ops/declarable/helpers/cpu/legacy_helper.cpp
+++ b/libnd4j/include/ops/declarable/helpers/cpu/legacy_helper.cpp
@@ -39,11 +39,28 @@ namespace helpers {
 
     template <typename T>
     static void reluDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
-        auto functor = LAMBDA_TT(x, y){
-            return x > (T)0.f ? y : T(0.f);
+
+        T zero = (T) 0.f;
+        auto functor = LAMBDA_TT(x, y, zero){
+            return x > zero ? y : zero;
         };
 
         input->applyPairwiseLambda<T>(epsilon, functor, output);
+
+        /*
+        auto x =  input->bufferAsT<T>();
+        auto y =  epsilon->bufferAsT<T>();
+        auto z =  output->bufferAsT<T>();
+
+        int length = input->lengthOf();
+
+        T zero = (T) 0.f;
+
+        PRAGMA_OMP_PARALLEL_FOR
+        for (int e = 0; e < length; e++) {
+            z[e] = x[e] > zero ? y[e] : zero;
+        }
+        */
     }
 
     void reluDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
diff --git a/libnd4j/tests_cpu/layers_tests/PlaygroundTests.cpp b/libnd4j/tests_cpu/layers_tests/PlaygroundTests.cpp
index 2d9a23e59..b76538afd 100644
--- a/libnd4j/tests_cpu/layers_tests/PlaygroundTests.cpp
+++ b/libnd4j/tests_cpu/layers_tests/PlaygroundTests.cpp
@@ -32,6 +32,7 @@
 #include <GradCheck.h>
 #include <ops/declarable/helpers/im2col.h>
 #include <Loops.h>
+#include <RandomLauncher.h>
 
 #include <helpers/BenchmarkHelper.h>
 #include <ops/declarable/helpers/scatter.h>
@@ -41,6 +42,8 @@
 #include <performance/benchmarking/FullBenchmarkSuit.h>
 #include <performance/benchmarking/LightBenchmarkSuit.h>
 
+#include <ops/declarable/helpers/legacy_helpers.h>
+
 using namespace nd4j;
 using namespace nd4j::graph;
 
@@ -55,3 +58,26 @@ public:
     }
 };
 
+/*
+TEST_F(PlaygroundTests, test_relubp_1) {
+    auto x = NDArrayFactory::create<float>('c', {128, 64, 224, 224});
+    auto y = x.ulike();
+    auto z = x.ulike();
+    RandomGenerator rng(119, 120);
+    RandomLauncher::fillUniform(LaunchContext::defaultContext(), rng, &x, -1.0, 1.0);
+    RandomLauncher::fillUniform(LaunchContext::defaultContext(), rng, &y, -1.0, 1.0);
+
+    int iterations = 10;
+
+    auto timeStart = std::chrono::system_clock::now();
+    for (int e = 0; e < iterations; e++)
+        ops::helpers::reluDerivative(LaunchContext::defaultContext(), &x, &y, &z);
+    auto timeEnd = std::chrono::system_clock::now();
+
+    auto outerTime = std::chrono::duration_cast<std::chrono::microseconds> (timeEnd - timeStart).count();
+    auto time = (Nd4jLong) outerTime / iterations;
+    auto bw = (1000000L * (float) (x.lengthOf() * x.sizeOfT()) / time) / 1024 / 1024 / 1024;
+
+    nd4j_printf("Time: %lld; BW: %f GB/s\n", time, bw);
+}
+*/

From 00fd50cee21636113f5a4b5c56fbb39762185bb2 Mon Sep 17 00:00:00 2001
From: Yurii Shyrma <yurii@skymind.io>
Date: Fri, 30 Aug 2019 20:31:05 +0300
Subject: [PATCH 55/56] Shyrma softmax (#209)

* - provide new cuda kernel for softmax

Signed-off-by: Yurii <yurii@skymind.io>

* - further work on cuda kernel for softmax

Signed-off-by: Yurii <yurii@skymind.io>

* - correction cuda kernel for softmax

Signed-off-by: Yurii <yurii@skymind.io>
---
 .../declarable/helpers/cuda/activations.cu    | 84 ++++++++++++++-----
 1 file changed, 65 insertions(+), 19 deletions(-)

diff --git a/libnd4j/include/ops/declarable/helpers/cuda/activations.cu b/libnd4j/include/ops/declarable/helpers/cuda/activations.cu
index 80e5d885b..f402944aa 100644
--- a/libnd4j/include/ops/declarable/helpers/cuda/activations.cu
+++ b/libnd4j/include/ops/declarable/helpers/cuda/activations.cu
@@ -24,6 +24,7 @@
 #include <ShapeUtils.h>
 #include <numeric>
 #include <PointersManager.h>
+#include <helpers/ConstantTadHelper.h>
 
 namespace nd4j    {
 namespace ops     {
@@ -196,7 +197,7 @@ void preluBP(nd4j::LaunchContext* context, const NDArray& input, const NDArray&
 
 ///////////////////////////////////////////////////////////////////
 template<typename T>
-__global__ void softMaxForVectorCuda(const void *vx, const Nd4jLong *xzShapeInfo, void *vz) {
+__device__ void softMaxForVectorCuda(const void *vx, const Nd4jLong *xShapeInfo, void *vz, const Nd4jLong *zShapeInfo) {
 
 	// logic of this kernel is based on assumption gridDim = 1
 
@@ -210,7 +211,7 @@ __global__ void softMaxForVectorCuda(const void *vx, const Nd4jLong *xzShapeInfo
 	if (threadIdx.x == 0) {
 		extern __shared__ char shared[];
 		shmem = reinterpret_cast<T*>(shared);
-		len = shape::length(xzShapeInfo);
+		len = shape::length(xShapeInfo);
 		numOfIters = (len + blockDim.x - 1) / blockDim.x;   // ceil (len / blockDim.x)
 	}
 	__syncthreads();
@@ -222,8 +223,8 @@ __global__ void softMaxForVectorCuda(const void *vx, const Nd4jLong *xzShapeInfo
 
 		const Nd4jLong elemIdx = i * blockDim.x + threadIdx.x;
 		if(elemIdx < len) {
-			const Nd4jLong offset = shape::getIndexOffset(elemIdx, xzShapeInfo, len);
-			shmem[threadIdx.x] = (threadIdx.x != 0) ? x[offset] : nd4j::math::nd4j_max<T>(x[offset], temp);	// take into account max element evaluated on previous iteration and stored in temp
+			const Nd4jLong xOffset = shape::getIndexOffset(elemIdx, xShapeInfo, len);
+			shmem[threadIdx.x] = (threadIdx.x != 0) ? x[xOffset] : nd4j::math::nd4j_max<T>(x[xOffset], temp);	// take into account max element evaluated on previous iteration and stored in temp
 		}
 		else
 			shmem[threadIdx.x] = -DataTypeUtils::max<T>();	// FIXME: what if T is unsigned ??
@@ -248,9 +249,10 @@ __global__ void softMaxForVectorCuda(const void *vx, const Nd4jLong *xzShapeInfo
 
 		const Nd4jLong elemIdx = i * blockDim.x + threadIdx.x;
 		if(elemIdx < len) {
-			const Nd4jLong offset = shape::getIndexOffset(elemIdx, xzShapeInfo, len);
-			z[offset] = nd4j::math::nd4j_exp<T, T>(x[offset] - max);
-			shmem[threadIdx.x] = (threadIdx.x != 0) ? z[offset] : (z[offset] + temp); // take into account sum element evaluated on previous iteration and stored in temp
+			const Nd4jLong xOffset = shape::getIndexOffset(elemIdx, xShapeInfo, len);
+			const Nd4jLong zOffset = shape::getIndexOffset(elemIdx, zShapeInfo, len);
+			z[zOffset] = nd4j::math::nd4j_exp<T, T>(x[xOffset] - max);
+			shmem[threadIdx.x] = (threadIdx.x != 0) ? z[zOffset] : (z[zOffset] + temp); // take into account sum element evaluated on previous iteration and stored in temp
 		}
 		else
 			shmem[threadIdx.x] = 0;
@@ -270,43 +272,87 @@ __global__ void softMaxForVectorCuda(const void *vx, const Nd4jLong *xzShapeInfo
 	for (int i = 0; i < numOfIters; ++i) {
 		const Nd4jLong elemIdx = i * blockDim.x + threadIdx.x;
 		if(elemIdx >= len) continue;
-		const Nd4jLong offset = shape::getIndexOffset(elemIdx, xzShapeInfo, len);
-		z[offset] /= shmem[0];
+		const Nd4jLong zOffset = shape::getIndexOffset(elemIdx, zShapeInfo, len);
+		z[zOffset] /= shmem[0];
 	}
 }
 
+template<typename T>
+__global__ void softMaxForVectorCudaGlobal(const void *vx, const Nd4jLong *xShapeInfo, void *vz, const Nd4jLong *zShapeInfo) {
+
+	softMaxForVectorCuda<T>(vx, xShapeInfo, vz, zShapeInfo);
+}
+
 ///////////////////////////////////////////////////////////////////
 template <typename T>
-linkage void softMaxForVectorCudaLauncher(const cudaStream_t* stream, const void *vx, const Nd4jLong *xzShapeInfo, void *vz) {
+linkage void softMaxForVectorCudaLauncher(const cudaStream_t* stream, const void *vx, const Nd4jLong *xShapeInfo, void *vz, const Nd4jLong *zShapeInfo) {
 
-	softMaxForVectorCuda<T><<<1, MAX_NUM_THREADS, MAX_NUM_THREADS * sizeof(T) + 512, *stream>>>(vx, xzShapeInfo, vz);
+	softMaxForVectorCudaGlobal<T><<<1, MAX_NUM_THREADS / 4 , (MAX_NUM_THREADS / 4) * sizeof(T) + 512, *stream>>>(vx, xShapeInfo, vz, zShapeInfo);
 }
 
+///////////////////////////////////////////////////////////////////
+template<typename T>
+__global__ static void softMaxCuda(const void* vx, const Nd4jLong *xTadShapeInfo, const Nd4jLong *xOffsets,
+                                         void* vz, const Nd4jLong *zTadShapeInfo, const Nd4jLong *zOffsets) {
+
+    const auto x = reinterpret_cast<const T*>(vx);
+          auto z = reinterpret_cast<T*>(vz);
+
+    const auto* xTad = x + xOffsets[blockIdx.x];
+          auto* zTad = z + zOffsets[blockIdx.x];
+
+    softMaxForVectorCuda<T>(xTad, xTadShapeInfo, zTad, zTadShapeInfo);
+}
+
+///////////////////////////////////////////////////////////////////
+template<typename T>
+static void softMaxCudaLauncher(const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream,
+                                const void* vx, const Nd4jLong *xTadShapeInfo, const Nd4jLong *xOffsets,
+                                	  void* vz, const Nd4jLong *zTadShapeInfo, const Nd4jLong *zOffsets) {
+
+    softMaxCuda<T><<<blocksPerGrid, threadsPerBlock, sharedMem, *stream>>>(vx, xTadShapeInfo, xOffsets, vz, zTadShapeInfo, zOffsets);
+}
+
+
 //////////////////////////////////////////////////////////////////////////
 void softmax(nd4j::LaunchContext * context, const NDArray& input, NDArray& output, const int dimension) {
 
 	if(!input.isActualOnDeviceSide()) input.syncToDevice();
 	const int rank = input.rankOf();
 
+	PointersManager manager(context, "helpers::softmax");
+
 	if(input.isVector()) {
 
 		if(rank == 1 || input.sizeAt(dimension) != 1) {
-			BUILD_SINGLE_SELECTOR(input.dataType(), softMaxForVectorCudaLauncher, (context->getCudaStream(), input.getSpecialBuffer(), input.getSpecialShapeInfo(), output.getSpecialBuffer()), FLOAT_TYPES);
-			input.tickReadDevice();
+			NDArray::prepareSpecialUse({&output}, {&input});
+			BUILD_SINGLE_SELECTOR(input.dataType(), softMaxForVectorCudaLauncher, (context->getCudaStream(), input.getSpecialBuffer(), input.getSpecialShapeInfo(), output.getSpecialBuffer(), output.getSpecialShapeInfo()), FLOAT_TYPES);
+			NDArray::registerSpecialUse({&output}, {&input});
 		}
 		else
 			output = 1.;
 	}
 	else {
 
-		auto maxAlongDim = const_cast<NDArray&>(input).reduceAlongDims(reduce::Max, {dimension}, true);
-		(input - maxAlongDim).applyTransform(transform::Exp, &output); // output contains exponents temporarily
-		auto sumAlongDim = output.reduceAlongDims(reduce::Sum, {dimension}, true);
-		output /= sumAlongDim;
-		input.tickReadDevice();
+		auto packX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(input.getShapeInfo(), {dimension});
+        auto packZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(output.getShapeInfo(), {dimension});
+
+        const int threadsPerBlock = MAX_NUM_THREADS / 4;
+        const int blocksPerGrid = packZ.numberOfTads();
+        const int sharedMem = input.sizeOfT() * threadsPerBlock + 512;
+
+        NDArray::prepareSpecialUse({&output}, {&input});
+    	BUILD_SINGLE_SELECTOR(input.dataType(), softMaxCudaLauncher, (blocksPerGrid, threadsPerBlock, sharedMem, context->getCudaStream(), input.getSpecialBuffer(), packX.specialShapeInfo(), packX.specialOffsets(), output.specialBuffer(), packZ.specialShapeInfo(), packZ.specialOffsets()), FLOAT_TYPES);
+    	NDArray::registerSpecialUse({&output}, {&input});
+
+		// auto maxAlongDim = const_cast<NDArray&>(input).reduceAlongDims(reduce::Max, {dimension}, true);
+		// (input - maxAlongDim).applyTransform(transform::Exp, &output); // output contains exponents temporarily
+		// auto sumAlongDim = output.reduceAlongDims(reduce::Sum, {dimension}, true);
+		// output /= sumAlongDim;
+		// input.tickReadDevice();
 	}
 
-	PointersManager manager(context, "helpers::softmax");
+
 	manager.synchronize();
 
 	output.tickWriteDevice();

From 12d14efcf2f44c96a5d5738282e0c0847ba299ab Mon Sep 17 00:00:00 2001
From: Alex Black <blacka101@gmail.com>
Date: Sat, 31 Aug 2019 12:27:09 +1000
Subject: [PATCH 56/56] ND4S test fix (#210)

Signed-off-by: AlexDBlack <blacka101@gmail.com>
---
 nd4s/pom.xml                                     | 16 ++++++++++++++--
 .../scala/org/nd4s/CollectionLikeNDArray.scala   |  4 ++--
 .../scala/org/nd4s/samediff/SameDiffTest.scala   |  2 +-
 3 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/nd4s/pom.xml b/nd4s/pom.xml
index 7fc3fb7d3..f165cfae9 100644
--- a/nd4s/pom.xml
+++ b/nd4s/pom.xml
@@ -294,7 +294,7 @@
 
     <profiles>
         <profile>
-            <id>nd4j-tests-cpu</id>
+            <id>test-nd4j-native</id>
             <dependencies>
                 <dependency>
                     <groupId>org.nd4j</groupId>
@@ -302,11 +302,17 @@
                     <version>${project.version}</version>
                     <scope>test</scope>
                 </dependency>
+                <dependency>
+                    <groupId>org.deeplearning4j</groupId>
+                    <artifactId>dl4j-test-resources</artifactId>
+                    <version>${dl4j-test-resources.version}</version>
+                    <scope>test</scope>
+                </dependency>
             </dependencies>
         </profile>
 
         <profile>
-            <id>nd4j-tests-cuda</id>
+            <id>test-nd4j-cuda-10.1</id>
             <dependencies>
                 <dependency>
                     <groupId>org.nd4j</groupId>
@@ -314,6 +320,12 @@
                     <version>${project.version}</version>
                     <scope>test</scope>
                 </dependency>
+                <dependency>
+                    <groupId>org.deeplearning4j</groupId>
+                    <artifactId>dl4j-test-resources</artifactId>
+                    <version>${dl4j-test-resources.version}</version>
+                    <scope>test</scope>
+                </dependency>
             </dependencies>
         </profile>
     </profiles>
diff --git a/nd4s/src/main/scala/org/nd4s/CollectionLikeNDArray.scala b/nd4s/src/main/scala/org/nd4s/CollectionLikeNDArray.scala
index e6818fbc0..bdbe43bc3 100644
--- a/nd4s/src/main/scala/org/nd4s/CollectionLikeNDArray.scala
+++ b/nd4s/src/main/scala/org/nd4s/CollectionLikeNDArray.scala
@@ -64,7 +64,7 @@ trait CollectionLikeNDArray[A <: INDArray] {
     val lv = ev.linearView(underlying)
     breakable {
       for {
-        i <- 0 until lv.length().toInt 
+        i <- 0 until lv.length().toInt
       } if (!f(ev.get(lv, i))) {
         result = true
         break()
@@ -80,7 +80,7 @@ trait CollectionLikeNDArray[A <: INDArray] {
     val lv = ev.linearView(underlying)
     breakable {
       for {
-        i <- 0 until lv.length().toInt 
+        i <- 0 until lv.length().toInt
       } if (!f(ev.get(lv, i))) {
         result = false
         break()
diff --git a/nd4s/src/test/scala/org/nd4s/samediff/SameDiffTest.scala b/nd4s/src/test/scala/org/nd4s/samediff/SameDiffTest.scala
index 2f5746b95..d0efee304 100644
--- a/nd4s/src/test/scala/org/nd4s/samediff/SameDiffTest.scala
+++ b/nd4s/src/test/scala/org/nd4s/samediff/SameDiffTest.scala
@@ -4,7 +4,7 @@ import java.lang.reflect.Field
 import java.util
 import java.util.{ Arrays, Collections, HashMap, List, Map }
 
-import com.google.common.collect.{ Lists, Maps }
+import org.nd4j.shade.guava.collect.{ Lists, Maps }
 import org.junit.Assert._
 import org.junit.Assume.assumeNotNull
 import org.nd4j.autodiff.samediff._