Merge master to upstream (#7945)

* Shugeo strided slice zeros (#14)

* Modified strided_slice op to properly work with empty-like shapes.

* Fixed test for reduce_mean with empty-like input.

* [WIP] Last merge (#15)

* correct logsoftmax looss (#2)

* Small SameDiff listener fix (#4)

* Various fixes (#6)

* #7839 Fix for asXMatrix and tests

* #7866 EmbeddingSequenceLayer dtype fix + test

* #7856 SameDiff save/load stream methods

* #7859 RegressionEvaluation rank 4 fix + tests + axis configuration

* EvaluationBinary 3d/4d

* More evaluation 3d/4d tests

* #7847 Evaluation empty checks

* Small test ifx

* #7848 Fix median edge case

* Improve DL4J samediff layer tests

* [WIP] FastText wrapper implemented (#8)

* FastText implemented

* Some fixes

* Fix shapes for wordsNearest

* Validation of input vectors

* Fixes

* Fixed test

* Thread tagged

* Some tweaks

* setContextClassLoader for DeallocatorServiceThread

* Numpy format tests (#1)

* Various fixes (#11)

* #7852 SameDiff gather fix

* #7892 SameDiff placeholder to constant conversion

* #7890 validate input rank for MLN/CG init methods

* Fix broken permute shape calculation

* Permute and gather fixes

* Tests

* #7850 LogSumExp fix + test

* Handful of test fixes

* Empty arrays with non-scalar shapes (#10)

* minor rearrangements for lambdas

* empty tensors with non-scalar shapes

* numpy empty tensors with non-scalar shapes

* few more empty tweaks

* Small fixes

* conv3d signature update

* micro fix in batchnorm mkldnn

* Import fixes

* Fix

* MKL-DNN update

* Small fill fix

* fill with empty input + test

* Fixes

* Small error improvement

* Fix

* one special test

* couple of fixes for lstm

* Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone

* Fixes

* FP16

* Unsigned

* BFloat16

* Fill op - empty tweaks

* - couple of fixes for empty arrays construction
- stack updated

* strided slice fix

* one transform test

* provide method for reducing shapeInfo in case of input array is empty

* Fixed reduceAlongDimensions to use empty input properly.

* couple of broadcast tests

* couple of tests broadcast tests + tweak to make them pass

* add check of non-empty to methods producing sub-arrays

* Fixed reshapeC with zeros in shape.

* complete empty check in reduce_... legacy ops

* Concat and cumsum/prod

* Tweak to empty shape inference on import

* add empty check to the rest of reduce legacy ops

* one more test

* correct typo in evalReduceShapeInfoEmpty

* Added tests for reduce_* ops to tests with zero shapes.

* few more tests for empty reductions

* Fixed strided_slice op with empty case and tests.

* one more empty reduction test

* Fixed strided_slice test.

* add empty check to NDArray::reshapei

* infOrMax

* empty min/max with infinity tests

* made unstack working correctly with empty arrays

* few IndexReduce tests + tweaks for empty shapes

* add test for empty concat

* few tests fixed

* Validation fix for reductions on empty shapes

* Reverse fix

* Reduction shape calc fixes

* SameDiff.generateOutputVariable: don't use shape function to determine number of outputs

* Range fix

* - NDArray constructor updated for scalars/empty arrays
- few tests fixed

* More fixes

* Empty creator fixes

* concat fix

* concat fix

* TF import tests: allow 'both all NaN' and 'both all inf' to pass

* Slice, zero fraction, and reshape fixes

* transpose, gather

* Zero fraction

* scalar cast fix

* Empty reduction axis support

* few more tests fixed

* Fixed input checks conforming with TF for concat op and tests.

* few tests fixed

* matmul scalar shape fix

* Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats.

* broadcast bool fix

* few more tests

* few more tests

* correct evalReduceShapeInfoEmpty

* argmax/argmin + tests

* one more empty edge case + one more test

* argmax/argmin/realdiv_bp tweaks

* empty reshape test + fix

* Helper fixes

* Small fixes

* Gather test fix

* Gather test fix

* Small fixes

* reduce scalar zero values

* scalar mean workaround

* Remove debug code

* along dim mean workaround

* one more test

* - equalsTo() tweak for empty arrays
- one more test

* broadcast tweaks

* [WIP] Fixing outstanding issues for NLP (#9)

* Avoid using not-inited objects

* Test fixed.

* Redundant method avoided for models like FastText

* KMeans++ implementation

* KMeans++ implementation

* Disable parallel execution

* KMeans++

* Tests

* Dev branch merge (#16)

* SameDiff: convertDataType and gradient check util improvements (#12)

* GradCheck util improvements

* StopGradient constructor + test

* SameDiff: Add datatype conversion

* Javadoc and add DataType.isNumerical()

* Small fix

* Fix SameDiff TF import test cases intermediate naming (workaround for bad default)

* TFGraphTestAllHelper: check intermediates in execution order

* Add missing debug listener

* [WIP] lstmBlock fix + other changes (#13)

- fixes lstmBlock issue
- changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer
- CheckNumerics op
- fixes for ReduceBool IsInfOrNan & IsFinite

* Small test fix

* CheckNumerics op wrapper

* Fix some issues on master (#17)

* Fix DataVec test issue

* Fix issue with dl4j SameDiff output layer

* Dtype fix for lambda layers

* #7912 BertIterator dtype fix (use float32 not global default)

* [WIP] Next set of CUDA stuff (#7)

New CUDA implementations and improvements

* bad file

* Dev branch master merge (#23)

* SameDiff: convertDataType and gradient check util improvements (#12)

* GradCheck util improvements

* StopGradient constructor + test

* SameDiff: Add datatype conversion

* Javadoc and add DataType.isNumerical()

* Small fix

* Fix SameDiff TF import test cases intermediate naming (workaround for bad default)

* TFGraphTestAllHelper: check intermediates in execution order

* Add missing debug listener

* [WIP] lstmBlock fix + other changes (#13)

- fixes lstmBlock issue
- changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer
- CheckNumerics op
- fixes for ReduceBool IsInfOrNan & IsFinite

* Small test fix

* CheckNumerics op wrapper

* Compatibility of deserialization (#18)

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

* SameDiff: add activation gradient checking support for debugging (#19)

* SameDiff gradient checker: first pass on activation gradient checks

* Fixes + tests for activation gradient checking

* Javadoc

* [WIP] Some nd4j data type corrections (#20)

* Adjust data type

* Set correct Data type.

* Size of proper data type.

* fix averaged cpu load (#22)

* SameDiff ops, TF import and fixes (#24)

* CheckNumerics tests + fixes + misc fixes

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

* Fake quant

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

* Fixes

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

* FakeQuantWithMinMaxArgs

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

* CheckNumerics fix

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

* Fix libnd4j ALL_INTS and ALL_FLOATS declaration (uint and bfloat types)

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

* Small fix

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

* Javadoc

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

* Exception tweak

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

* fix

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

* Fix for out of scope stack allocated var use

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

* Ignores

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

* Ignore for known failing test (already logged issue)

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

* Merge upstream to fork (#25)

* Add thousand-separator commas to TotalParams (#7915)

* Add thousand-separator commas to TotalParams

The number of parameters can be quite large, and it would help the reading of the summary printout to have the TotalParams column & values at the bottom have thousand-separator-commas in them.

* Add thousand-separator commas to MultiLayerNetwork

Corresponding change to MultiLayerNetwork

Signed-off-by: Jxtps Jxtps <jxtps435@gmail.com>

* Update contributing and issue/PR templates (#7934)

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

* Fix link to AdaDelta paper (#7942)

Fix link to AdaDelta paper hosted on matthewzeiler.com

Signed-off-by: Jxtps

* Fixes, and ignores for known/logged failing issues (#7943)

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

* SameDiff + DL4J/SameDiff: Multiple fixes (#28)

* #7919 HDF5 attribute buffer length fix

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

* #7909 Arbiter constructor exception ux improvements

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

* #7925 RNN output layer length checks

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

* #7939 Add listener for validating inputs are not incorrectly modified

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

* #7939 Integrate NonInplaceValidationListener into tests

* #7844 DL4J SameDiff fixes for variable minibatch size

* DL4J SameDiff fixes - ensure gradient for input placeholder is available

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

* Tweaks to ExternalErrorsFunction - use placeholders, make more robust

* Another fix

* More fixes

* More SameDiff/DL4J fixes

* Scope out scalar array creation in BaseScalarOp

* Remove debug code

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

* [WIP] Final dev branch merge (#29)

* SameDiff: convertDataType and gradient check util improvements (#12)

* GradCheck util improvements

* StopGradient constructor + test

* SameDiff: Add datatype conversion

* Javadoc and add DataType.isNumerical()

* Small fix

* Fix SameDiff TF import test cases intermediate naming (workaround for bad default)

* TFGraphTestAllHelper: check intermediates in execution order

* Add missing debug listener

* [WIP] lstmBlock fix + other changes (#13)

- fixes lstmBlock issue
- changes NDArray method reshape(), permute(), transpose() by making them return instance instead of pointer
- CheckNumerics op
- fixes for ReduceBool IsInfOrNan & IsFinite

* Small test fix

* CheckNumerics op wrapper

* Compatibility of deserialization (#18)

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

* SameDiff: add activation gradient checking support for debugging (#19)

* SameDiff gradient checker: first pass on activation gradient checks

* Fixes + tests for activation gradient checking

* Javadoc

* [WIP] Some nd4j data type corrections (#20)

* Adjust data type

* Set correct Data type.

* Size of proper data type.

* fix averaged cpu load (#22)

* [WIP] Multiple dataset iterators (#27)

* Splitting dataset into arbitrary number

* Fixes

* Multiple split of iterator

* Test

* Test

* Some fixes

* signature change

* one more tweak

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

* one more test for sequential use of DataSetIteratorSplitter

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

* Fixes

* Fixes

* one more test for Alexander

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

* Some fixes

* Some fixes

* one more test for Alexander

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

* minor test fix

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

* Some fixes

* Some fixes

* couple of assertions tweaked

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

* MDS splitter test :/

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

* Minor refactoring

* Multi dataset

* Some fixes

* More tests

* Small number of test fixes/improvements (failures on CI) (#31)

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

* [WIP] More CUDA stuff (#26)

* initial commit

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

* LRN BP CUDA

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

* less memory

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

* Fixed bug with crop_and_resize op helper.

* get rid of unnecessary index-calculation dunction

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

* Fixed sort with nth_element cuda-based helper.

* Refactored nth_element.

* Refactored nth_element op and tests.

* Modified usage of dim array with sortTad routine.

* Refactored main routine of helper for non_max_image_suppression op.

* non_max_image_suppression op helper with cuda kernel implementation. Initial revision.

* fix vol2col cuda kernel

* meh

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

* topK concept

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

* unsorted topK with scanWitdh of 1

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

* correct vol2col tests

* sorted/unsorted topK

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

* implementation and fixing col2im/col2vol

* Corrected usage flags with input/output with reverse op.

* dup is const now

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

* percentile op

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

* group tests for mapool2d

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

* special test for george

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

* less threads for sortTad

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

* provide conv2d for cuda

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

* remove auther in sort tad kernel code

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

* provide depthwise_conv2d for cuda

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

* - max_pooling_with_argmax
- null check for special use

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

* dts cuda

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

* provide sconv2d for cuda

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

* std cuda

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

* Refactored non_max_suppression op to conform TF implementation.

* Improved suppression helper.

* provide pooling3d for cuda

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

* minor lstm rearrangements

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

* more of minor lstm rearrangements

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

* (bi)dynamic_rnn

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

* templates init order

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

* Refactored non_max_suppression op.

* Added cuda kernel for non_max_suppression.

* CPU sort by key/value

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

* CPU sort TAD by key/value

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

* CPU sort TAD by key/value tests

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

* Eliminate compiler error with cuda implementation.

* - repaired gradCheck in cuda
- provide conv2d_bp for cuda

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

* missed signature

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

* provide depthwise_conv2d_bp for cuda

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

* Implementation of lup helper with cuda kernel. Initial commit.

* further work on backprops for convolutions

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

* CUDA linear sort by key/val

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

* CUDA tad sort by key/val

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

* start providing of backprop for pooling2d/3d

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

* Added atomicAdd for bool datatype.

* dynamic partition concept

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

* dynamic partition concept

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

* dynamic partition scalar CUDA

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

* important comment

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

* fix pooling2d/3d backprop helpers

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

* Added non-linear test with dynamic_partition.

* Improved test for dynamic_partition.

* dynamic_partition TAD concept

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

* - dynamic_partition TAD CUDA impl
- dynamic_partition TAD CPU fix

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

* - rewrite cpu code for usampling2d/3d
- write cuda code for usampling2d/3d

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

* dynamic_stitch CUDA vector case

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

* dynamic_stitch CUDA TAD case concept

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

* dynamic_stitch CUDA TAD case impl

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

* Added tests for dynamic_stitch 3D-4D cases.

* minor tests tweaks

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

* Fixed type check for dynamic stitch.

* min/max bp

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

* rewrite code for upsampling2d/3d cpu

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

* reduce min/max/norm_max bp

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

* lup implementation. Additional enhancements.

* provide code for upsamling2d/3d backprop

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

* weightedCrossEntropyWithLogits

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

* Fixed template math atomicMul for 64bit ints.

* Refactored dynamic_partition_bp op.

* inverseBroadcast fix

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

* DynamicPartitionBP test datatype fixed.

* - nd4j_atomicMul Windows fix
- cpu/NDArrayLambda.hpp excluded from CUDA

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

arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/api/TaskCreatorProvider.java

@@ -31,7 +31,7 @@ public class TaskCreatorProvider {
             }
             return c.newInstance();
         } catch (Exception e){
-            throw new RuntimeException("Could not create new instance of task creator class: " + c, e);
+            throw new RuntimeException("Could not create new instance of task creator class: " + c + " - missing no-arg constructor?", e);
         }
     }
 

arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/optimize/api/data/DataSetIteratorFactoryProvider.java

@@ -83,7 +83,7 @@ public class DataSetIteratorFactoryProvider implements DataProvider {
                             (Class<? extends DataSetIteratorFactory>) Class.forName(value);
             return clazz.newInstance();
         } catch (Exception e) {
-            throw new RuntimeException(e);
+            throw new RuntimeException("Could not create DataSetIteratorFactory instance - missing no-arg constructor?", e);
         }
     }
 }

arbiter/arbiter-deeplearning4j/src/main/java/org/deeplearning4j/arbiter/data/DataSetIteratorFactoryProvider.java

@@ -79,7 +79,7 @@ public class DataSetIteratorFactoryProvider implements DataProvider {
                             (Class<? extends DataSetIteratorFactory>) Class.forName(value);
             return clazz.newInstance();
         } catch (Exception e) {
-            throw new RuntimeException(e);
+            throw new RuntimeException("Could not create DataSetIteratorFactory instance - missing no-arg constructor?", e);
         }
     }
 }

arbiter/arbiter-deeplearning4j/src/main/java/org/deeplearning4j/arbiter/scoring/impl/BaseNetScoreFunction.java

@@ -54,7 +54,7 @@ public abstract class BaseNetScoreFunction implements ScoreFunction {
                 ds.configure(dataSourceProperties);
             }
         } catch (Exception e){
-            throw new RuntimeException(e);
+            throw new RuntimeException("Error creating DataSource instance - missing no-arg constructor?", e);
         }
         return score(model, ds.testData());
     }

arbiter/arbiter-deeplearning4j/src/main/java/org/deeplearning4j/arbiter/task/ComputationGraphTaskCreator.java

@@ -188,10 +188,15 @@ public class ComputationGraphTaskCreator implements TaskCreator {
             //For DataSetIterator: wraps in a MultiDataSetIterator, hence method can be used for both
             MultiDataSetIterator iterator;
             if(dataSource != null){
+                try {
                     DataSource dsInstance = dataSource.newInstance();
-                if(dataSourceProperties != null)
+                    if (dataSourceProperties != null)
                         dsInstance.configure(dataSourceProperties);
                     iterator = ScoreUtil.getMultiIterator(dsInstance.trainData());
+                } catch (Exception e){
+                    throw new RuntimeException("Error instantiating instance of DataSource for class " + dataSource.getName() +
+                            " - no zero-arg constructor?",e);
+                }
             } else {
                 iterator = ScoreUtil.getMultiIterator(dataProvider.trainData(candidate.getDataParameters()));
             }

arbiter/arbiter-deeplearning4j/src/main/java/org/deeplearning4j/arbiter/task/MultiLayerNetworkTaskCreator.java

@@ -190,7 +190,8 @@ public class MultiLayerNetworkTaskCreator implements TaskCreator {
                 try{
                     dsInstance = dataSource.newInstance();
                 } catch (Exception e){
-                    throw new RuntimeException("Error instantiating instance of DataSource for class " + dataSource.getName());
+                    throw new RuntimeException("Error instantiating instance of DataSource for class " + dataSource.getName() +
+                            " - no zero-arg constructor?",e);
                 }
                 if(dataSourceProperties != null)
                     dsInstance.configure(dataSourceProperties);

datavec/datavec-api/src/test/java/org/datavec/api/transform/transform/ndarray/TestNDArrayWritableTransforms.java

@@ -26,6 +26,7 @@ import org.datavec.api.writable.NDArrayWritable;
 import org.datavec.api.writable.Text;
 import org.datavec.api.writable.Writable;
 import org.junit.Test;
+import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.ops.transforms.Transforms;
@@ -78,14 +79,14 @@ public class TestNDArrayWritableTransforms {
         assertEquals(expColNames, tp.getFinalSchema().getColumnNames());
 
 
-        List<Writable> in = Arrays.<Writable>asList(new DoubleWritable(0), new NDArrayWritable(Nd4j.linspace(0, 9, 10)),
+        List<Writable> in = Arrays.<Writable>asList(new DoubleWritable(0), new NDArrayWritable(Nd4j.linspace(DataType.DOUBLE,0, 10, 1).reshape(1,10)),
-                        new NDArrayWritable(Nd4j.valueArrayOf(1, 10, 2.0)));
+                        new NDArrayWritable(Nd4j.valueArrayOf(1, 10, 2.0).castTo(DataType.DOUBLE)));
         List<Writable> out = tp.execute(in);
 
         List<Writable> exp =
-                        Arrays.<Writable>asList(new DoubleWritable(0), new NDArrayWritable(Nd4j.linspace(0, 9, 10)),
+                        Arrays.<Writable>asList(new DoubleWritable(0), new NDArrayWritable(Nd4j.linspace(DataType.DOUBLE,0, 10, 1).reshape(1,10)),
-                                        new NDArrayWritable(Nd4j.valueArrayOf(1, 10, 2.0)),
+                                        new NDArrayWritable(Nd4j.valueArrayOf(1, 10, 2.0).castTo(DataType.DOUBLE)),
-                                        new NDArrayWritable(Nd4j.linspace(0, 9, 10).addi(2.0)));
+                                        new NDArrayWritable(Nd4j.linspace(DataType.DOUBLE, 0, 10, 1).addi(2.0).reshape(1,10)));
 
         assertEquals(exp, out);
     }

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/iterator/DataSetSplitterTests.java

@@ -20,9 +20,15 @@ import lombok.val;
 import org.deeplearning4j.BaseDL4JTest;
 import org.deeplearning4j.datasets.iterator.tools.DataSetGenerator;
 import org.junit.Test;
+import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
 import org.nd4j.linalg.exception.ND4JIllegalStateException;
+import org.nd4j.linalg.factory.Nd4j;
 
-import static org.junit.Assert.assertEquals;
+import java.util.Collections;
+import java.util.List;
+import java.util.Random;
+
+import static org.junit.Assert.*;
 
 public class DataSetSplitterTests extends BaseDL4JTest {
     @Test
@@ -39,7 +45,7 @@ public class DataSetSplitterTests extends BaseDL4JTest {
         int gcntTest = 0;
         int global = 0;
         // emulating epochs here
-        for (int e = 0; e < numEpochs; e++){
+        for (int e = 0; e < numEpochs; e++) {
             int cnt = 0;
             while (train.hasNext()) {
                 val data = train.next().getFeatures();
@@ -79,7 +85,7 @@ public class DataSetSplitterTests extends BaseDL4JTest {
         int gcntTest = 0;
         int global = 0;
         // emulating epochs here
-        for (int e = 0; e < numEpochs; e++){
+        for (int e = 0; e < numEpochs; e++) {
             int cnt = 0;
             while (train.hasNext()) {
                 val data = train.next().getFeatures();
@@ -117,7 +123,7 @@ public class DataSetSplitterTests extends BaseDL4JTest {
         int gcntTest = 0;
         int global = 0;
         // emulating epochs here
-        for (int e = 0; e < numEpochs; e++){
+        for (int e = 0; e < numEpochs; e++) {
             int cnt = 0;
             while (train.hasNext()) {
                 val data = train.next().getFeatures();
@@ -144,4 +150,245 @@ public class DataSetSplitterTests extends BaseDL4JTest {
 
         assertEquals(1000 * numEpochs, global);
     }
+
+    @Test
+    public void testSplitter_4() {
+        val back = new DataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        val splitter = new DataSetIteratorSplitter(back, 1000, new double[]{0.5, 0.3, 0.2});
+        List<DataSetIterator> iteratorList = splitter.getIterators();
+        val numEpochs = 10;
+        int global = 0;
+        // emulating epochs here
+        for (int e = 0; e < numEpochs; e++) {
+            int iterNo = 0;
+            int perEpoch = 0;
+            for (val partIterator : iteratorList) {
+                int cnt = 0;
+                partIterator.reset();
+                while (partIterator.hasNext()) {
+                    val data = partIterator.next().getFeatures();
+                    assertEquals("Train failed on iteration " + cnt + "; epoch: " + e,
+                            (float) perEpoch, data.getFloat(0), 1e-5);
+                    //gcntTrain++;
+                    global++;
+                    cnt++;
+                    ++perEpoch;
+                }
+                ++iterNo;
+            }
+        }
+
+        assertEquals(1000* numEpochs, global);
+    }
+
+    @Test
+    public void testSplitter_5() {
+        val back = new DataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        val splitter = new DataSetIteratorSplitter(back, new int[]{900, 100});
+
+        List<DataSetIterator> iteratorList = splitter.getIterators();
+        val numEpochs = 10;
+
+        int global = 0;
+        // emulating epochs here
+        for (int e = 0; e < numEpochs; e++) {
+            int iterNo = 0;
+            int perEpoch = 0;
+            for (val partIterator : iteratorList) {
+                partIterator.reset();
+                while (partIterator.hasNext()) {
+                    int cnt = 0;
+                    val data = partIterator.next().getFeatures();
+
+                    assertEquals("Train failed on iteration " + cnt + "; epoch: " + e,
+                            (float) perEpoch, data.getFloat(0), 1e-5);
+                    //gcntTrain++;
+                    global++;
+                    cnt++;
+                    ++perEpoch;
+                }
+                ++iterNo;
+            }
+        }
+
+        assertEquals(1000 * numEpochs, global);
+    }
+
+    @Test
+    public void testSplitter_6() {
+        val back = new DataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        // we're going to mimic train+test+validation split
+        val splitter = new DataSetIteratorSplitter(back, new int[]{800, 100, 100});
+
+        assertEquals(3, splitter.getIterators().size());
+
+        val trainIter = splitter.getIterators().get(0);
+        val testIter = splitter.getIterators().get(1);
+        val validationIter = splitter.getIterators().get(2);
+
+        // we're going to have multiple epochs
+        int numEpochs = 10;
+        for (int e = 0; e < numEpochs; e++) {
+            int globalIter = 0;
+            trainIter.reset();
+            testIter.reset();
+            validationIter.reset();
+
+            boolean trained = false;
+            while (trainIter.hasNext()) {
+                trained = true;
+                val ds = trainIter.next();
+                assertNotNull(ds);
+
+                assertEquals("Failed at iteration [" + globalIter + "]", (double) globalIter, ds.getFeatures().getDouble(0), 1e-5f);
+                globalIter++;
+            }
+            assertTrue("Failed at epoch [" + e + "]", trained);
+            assertEquals(800, globalIter);
+
+
+            // test set is used every epoch
+            boolean tested = false;
+            //testIter.reset();
+            while (testIter.hasNext()) {
+                tested = true;
+                val ds = testIter.next();
+                assertNotNull(ds);
+
+                assertEquals("Failed at iteration [" + globalIter + "]", (double) globalIter, ds.getFeatures().getDouble(0), 1e-5f);
+                globalIter++;
+            }
+            assertTrue("Failed at epoch [" + e + "]", tested);
+            assertEquals(900, globalIter);
+
+            // validation set is used every 5 epochs
+            if (e % 5 == 0) {
+                boolean validated = false;
+                //validationIter.reset();
+                while (validationIter.hasNext()) {
+                    validated = true;
+                    val ds = validationIter.next();
+                    assertNotNull(ds);
+
+                    assertEquals("Failed at iteration [" + globalIter + "]", (double) globalIter, ds.getFeatures().getDouble(0), 1e-5f);
+                    globalIter++;
+                }
+                assertTrue("Failed at epoch [" + e + "]", validated);
+            }
+
+            // all 3 iterators have exactly 1000 elements combined
+            if (e % 5 == 0)
+                assertEquals(1000, globalIter);
+            else
+                assertEquals(900, globalIter);
+            trainIter.reset();
+        }
+    }
+
+    @Test
+    public void testUnorderedSplitter_1() {
+        val back = new DataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        val splitter = new DataSetIteratorSplitter(back, new int[]{500, 500});
+
+        List<DataSetIterator> iteratorList = splitter.getIterators();
+        val numEpochs = 10;
+
+        int global = 0;
+        // emulating epochs here
+        for (int e = 0; e < numEpochs; e++) {
+
+            // Get data from second part, then rewind for the first one.
+            int cnt = 0;
+            int partNumber = 1;
+            while (iteratorList.get(partNumber).hasNext()) {
+                int farCnt = (1000 / 2) * (partNumber) + cnt;
+                val data = iteratorList.get(partNumber).next().getFeatures();
+
+                assertEquals("Train failed on iteration " + cnt + "; epoch: " + e, (float) farCnt, data.getFloat(0), 1e-5);
+                cnt++;
+                global++;
+            }
+            iteratorList.get(partNumber).reset();
+            partNumber = 0;
+            cnt = 0;
+            while (iteratorList.get(0).hasNext()) {
+                val data = iteratorList.get(0).next().getFeatures();
+
+                assertEquals("Train failed on iteration " + cnt + "; epoch: " + e, (float) cnt++, data.getFloat(0), 1e-5);
+                global++;
+            }
+        }
+    }
+
+    @Test
+    public void testUnorderedSplitter_2() {
+        val back = new DataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        val splitter = new DataSetIteratorSplitter(back, new int[]{2});
+
+        List<DataSetIterator> iteratorList = splitter.getIterators();
+
+        for (int partNumber = 0 ; partNumber < iteratorList.size(); ++partNumber) {
+            int cnt = 0;
+            while (iteratorList.get(partNumber).hasNext()) {
+                val data = iteratorList.get(partNumber).next().getFeatures();
+
+                assertEquals("Train failed on iteration " + cnt, (float) (500*partNumber + cnt), data.getFloat(0), 1e-5);
+                cnt++;
+            }
+        }
+    }
+
+    @Test
+    public void testUnorderedSplitter_3() {
+        val back = new DataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        val splitter = new DataSetIteratorSplitter(back, new int[]{10});
+
+        List<DataSetIterator> iteratorList = splitter.getIterators();
+        Random random = new Random();
+        int[] indexes = new int[iteratorList.size()];
+        for (int i = 0; i < indexes.length; ++i) {
+            indexes[i] = random.nextInt(iteratorList.size());
+        }
+
+        for (int partNumber : indexes) {
+            int cnt = 0;
+            while (iteratorList.get(partNumber).hasNext()) {
+                val data = iteratorList.get(partNumber).next().getFeatures();
+
+                assertEquals("Train failed on iteration " + cnt, (float) (500*partNumber + cnt), data.getFloat(0), 1e-5);
+                cnt++;
+            }
+        }
+    }
+
+    @Test
+    public void testUnorderedSplitter_4() {
+        val back = new DataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        // we're going to mimic train+test+validation split
+        val splitter = new DataSetIteratorSplitter(back, new int[]{80, 10, 5});
+
+        assertEquals(3, splitter.getIterators().size());
+
+        val trainIter = splitter.getIterators().get(0);  // 0..79
+        val testIter = splitter.getIterators().get(1);   // 80 ..89
+        val validationIter = splitter.getIterators().get(2); // 90..94
+
+        // we're skipping train/test and go for validation first. we're that crazy, right.
+        int valCnt = 0;
+        while (validationIter.hasNext()) {
+            val ds = validationIter.next();
+            assertNotNull(ds);
+
+            assertEquals("Validation failed on iteration " + valCnt, (float) valCnt + 90, ds.getFeatures().getFloat(0), 1e-5);
+            valCnt++;
+        }
+        assertEquals(5, valCnt);
+    }
 }

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/iterator/MultiDataSetSplitterTests.java

@@ -18,11 +18,17 @@ package org.deeplearning4j.datasets.iterator;
 
 import lombok.val;
 import org.deeplearning4j.BaseDL4JTest;
+import org.deeplearning4j.datasets.iterator.tools.DataSetGenerator;
 import org.deeplearning4j.datasets.iterator.tools.MultiDataSetGenerator;
 import org.junit.Test;
+import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
+import org.nd4j.linalg.dataset.api.iterator.MultiDataSetIterator;
 import org.nd4j.linalg.exception.ND4JIllegalStateException;
 
-import static org.junit.Assert.assertEquals;
+import java.util.List;
+import java.util.Random;
+
+import static org.junit.Assert.*;
 
 /**
  *
@@ -150,4 +156,309 @@ public class MultiDataSetSplitterTests extends BaseDL4JTest {
 
         assertEquals(1000 * numEpochs, global);
     }
+
+    @Test
+    public void testMultiSplitter_1() {
+        val back = new MultiDataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        // we're going to mimic train+test+validation split
+        val splitter = new MultiDataSetIteratorSplitter(back, new int[]{800, 100, 100});
+
+        assertEquals(3, splitter.getIterators().size());
+
+        val trainIter = splitter.getIterators().get(0);
+        val testIter = splitter.getIterators().get(1);
+        val validationIter = splitter.getIterators().get(2);
+
+        // we're going to have multiple epochs
+        int numEpochs = 10;
+        for (int e = 0; e < numEpochs; e++) {
+            int globalIter = 0;
+            trainIter.reset();
+            testIter.reset();
+            validationIter.reset();
+
+            boolean trained = false;
+            while (trainIter.hasNext()) {
+                trained = true;
+                val ds = trainIter.next();
+                assertNotNull(ds);
+
+                for (int i = 0; i < ds.getFeatures().length; ++i) {
+                    assertEquals("Failed at iteration [" + globalIter + "]", (double) globalIter, ds.getFeatures()[i].getDouble(0), 1e-5f);
+                }
+                globalIter++;
+            }
+            assertTrue("Failed at epoch [" + e + "]", trained);
+            assertEquals(800, globalIter);
+
+
+            // test set is used every epoch
+            boolean tested = false;
+            //testIter.reset();
+            while (testIter.hasNext()) {
+                tested = true;
+                val ds = testIter.next();
+                assertNotNull(ds);
+
+                for (int i = 0; i < ds.getFeatures().length; ++i) {
+                    assertEquals("Failed at iteration [" + globalIter + "]", (double) globalIter, ds.getFeatures()[i].getDouble(0), 1e-5f);
+                }
+                globalIter++;
+            }
+            assertTrue("Failed at epoch [" + e + "]", tested);
+            assertEquals(900, globalIter);
+
+            // validation set is used every 5 epochs
+            if (e % 5 == 0) {
+                boolean validated = false;
+                //validationIter.reset();
+                while (validationIter.hasNext()) {
+                    validated = true;
+                    val ds = validationIter.next();
+                    assertNotNull(ds);
+
+                    for (int i = 0; i < ds.getFeatures().length; ++i) {
+                        assertEquals("Failed at iteration [" + globalIter + "]", (double) globalIter, ds.getFeatures()[i].getDouble(0), 1e-5f);
+                    }
+                    globalIter++;
+                }
+                assertTrue("Failed at epoch [" + e + "]", validated);
+            }
+
+            // all 3 iterators have exactly 1000 elements combined
+            if (e % 5 == 0)
+                assertEquals(1000, globalIter);
+            else
+                assertEquals(900, globalIter);
+            trainIter.reset();
+        }
+    }
+
+    @Test
+    public void testSplitter_5() {
+        val back = new MultiDataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        val splitter = new MultiDataSetIteratorSplitter(back, new int[]{900, 100});
+
+        List<MultiDataSetIterator> iteratorList = splitter.getIterators();
+        val numEpochs = 10;
+
+        int global = 0;
+        // emulating epochs here
+        for (int e = 0; e < numEpochs; e++) {
+            int iterNo = 0;
+            int perEpoch = 0;
+            for (val partIterator : iteratorList) {
+                partIterator.reset();
+                while (partIterator.hasNext()) {
+                    int cnt = 0;
+                    val data = partIterator.next().getFeatures();
+
+                    for (int i = 0; i < data.length; ++i) {
+                        assertEquals("Train failed on iteration " + cnt + "; epoch: " + e,
+                                (float) perEpoch, data[i].getFloat(0), 1e-5);
+                    }
+                    //gcntTrain++;
+                    global++;
+                    cnt++;
+                    ++perEpoch;
+                }
+                ++iterNo;
+            }
+        }
+
+        assertEquals(1000 * numEpochs, global);
+    }
+
+    @Test
+    public void testSplitter_6() {
+        val back = new MultiDataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        // we're going to mimic train+test+validation split
+        val splitter = new MultiDataSetIteratorSplitter(back, new int[]{800, 100, 100});
+
+        assertEquals(3, splitter.getIterators().size());
+
+        val trainIter = splitter.getIterators().get(0);
+        val testIter = splitter.getIterators().get(1);
+        val validationIter = splitter.getIterators().get(2);
+
+        // we're going to have multiple epochs
+        int numEpochs = 10;
+        for (int e = 0; e < numEpochs; e++) {
+            int globalIter = 0;
+            trainIter.reset();
+            testIter.reset();
+            validationIter.reset();
+
+            boolean trained = false;
+            while (trainIter.hasNext()) {
+                trained = true;
+                val ds = trainIter.next();
+                assertNotNull(ds);
+
+                for (int i = 0; i < ds.getFeatures().length; ++i) {
+                    assertEquals("Failed at iteration [" + globalIter + "]", (double) globalIter,
+                            ds.getFeatures()[i].getDouble(0), 1e-5f);
+                }
+                globalIter++;
+            }
+            assertTrue("Failed at epoch [" + e + "]", trained);
+            assertEquals(800, globalIter);
+
+
+            // test set is used every epoch
+            boolean tested = false;
+            //testIter.reset();
+            while (testIter.hasNext()) {
+                tested = true;
+                val ds = testIter.next();
+                assertNotNull(ds);
+                for (int i = 0; i < ds.getFeatures().length; ++i) {
+                    assertEquals("Failed at iteration [" + globalIter + "]", (double) globalIter, ds.getFeatures()[i].getDouble(0), 1e-5f);
+                }
+                globalIter++;
+            }
+            assertTrue("Failed at epoch [" + e + "]", tested);
+            assertEquals(900, globalIter);
+
+            // validation set is used every 5 epochs
+            if (e % 5 == 0) {
+                boolean validated = false;
+                //validationIter.reset();
+                while (validationIter.hasNext()) {
+                    validated = true;
+                    val ds = validationIter.next();
+                    assertNotNull(ds);
+
+                    for (int i = 0; i < ds.getFeatures().length; ++i) {
+                        assertEquals("Failed at iteration [" + globalIter + "]", (double) globalIter,
+                                ds.getFeatures()[i].getDouble(0), 1e-5f);
+                    }
+                    globalIter++;
+                }
+                assertTrue("Failed at epoch [" + e + "]", validated);
+            }
+
+            // all 3 iterators have exactly 1000 elements combined
+            if (e % 5 == 0)
+                assertEquals(1000, globalIter);
+            else
+                assertEquals(900, globalIter);
+            trainIter.reset();
+        }
+    }
+
+    @Test
+    public void testUnorderedSplitter_1() {
+        val back = new MultiDataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        val splitter = new MultiDataSetIteratorSplitter(back, new int[]{500, 500});
+
+        List<MultiDataSetIterator> iteratorList = splitter.getIterators();
+        val numEpochs = 10;
+
+        int global = 0;
+        // emulating epochs here
+        for (int e = 0; e < numEpochs; e++) {
+
+            // Get data from second part, then rewind for the first one.
+            int cnt = 0;
+            int partNumber = 1;
+            while (iteratorList.get(partNumber).hasNext()) {
+                int farCnt = (1000 / 2) * (partNumber) + cnt;
+                val data = iteratorList.get(partNumber).next().getFeatures();
+                for (int i = 0; i < data.length; ++i) {
+                    assertEquals("Train failed on iteration " + cnt + "; epoch: " + e, (float) farCnt, data[i].getFloat(0), 1e-5);
+                }
+                cnt++;
+                global++;
+            }
+            iteratorList.get(partNumber).reset();
+            partNumber = 0;
+            cnt = 0;
+            while (iteratorList.get(0).hasNext()) {
+                val data = iteratorList.get(0).next().getFeatures();
+                for (int i = 0; i < data.length; ++i) {
+                    assertEquals("Train failed on iteration " + cnt + "; epoch: " + e, (float) cnt++,
+                            data[i].getFloat(0), 1e-5);
+                }
+                global++;
+            }
+        }
+    }
+
+    @Test
+    public void testUnorderedSplitter_2() {
+        val back = new MultiDataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        val splitter = new MultiDataSetIteratorSplitter(back, new int[]{2});
+
+        List<MultiDataSetIterator> iteratorList = splitter.getIterators();
+
+        for (int partNumber = 0 ; partNumber < iteratorList.size(); ++partNumber) {
+            int cnt = 0;
+            while (iteratorList.get(partNumber).hasNext()) {
+                val data = iteratorList.get(partNumber).next().getFeatures();
+                for (int i = 0; i < data.length; ++i) {
+                    assertEquals("Train failed on iteration " + cnt, (float) (500 * partNumber + cnt), data[i].getFloat(0), 1e-5);
+                }
+                cnt++;
+            }
+        }
+    }
+
+    @Test
+    public void testUnorderedSplitter_3() {
+        val back = new MultiDataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        val splitter = new MultiDataSetIteratorSplitter(back, new int[]{10});
+
+        List<MultiDataSetIterator> iteratorList = splitter.getIterators();
+        Random random = new Random();
+        int[] indexes = new int[iteratorList.size()];
+        for (int i = 0; i < indexes.length; ++i) {
+            indexes[i] = random.nextInt(iteratorList.size());
+        }
+
+        for (int partNumber : indexes) {
+            int cnt = 0;
+            while (iteratorList.get(partNumber).hasNext()) {
+                val data = iteratorList.get(partNumber).next().getFeatures();
+                for (int i = 0; i < data.length; ++i) {
+                    assertEquals("Train failed on iteration " + cnt, (float) (500 * partNumber + cnt),
+                            data[i].getFloat(0), 1e-5);
+                }
+                cnt++;
+            }
+        }
+    }
+
+    @Test
+    public void testUnorderedSplitter_4() {
+        val back = new MultiDataSetGenerator(1000, new int[]{32, 100}, new int[]{32, 5});
+
+        // we're going to mimic train+test+validation split
+        val splitter = new MultiDataSetIteratorSplitter(back, new int[]{80, 10, 5});
+
+        assertEquals(3, splitter.getIterators().size());
+
+        val trainIter = splitter.getIterators().get(0);  // 0..79
+        val testIter = splitter.getIterators().get(1);   // 80 ..89
+        val validationIter = splitter.getIterators().get(2); // 90..94
+
+        // we're skipping train/test and go for validation first. we're that crazy, right.
+        int valCnt = 0;
+        while (validationIter.hasNext()) {
+            val ds = validationIter.next();
+            assertNotNull(ds);
+            for (int i = 0; i < ds.getFeatures().length; ++i) {
+                assertEquals("Validation failed on iteration " + valCnt, (float) valCnt + 90,
+                        ds.getFeatures()[i].getFloat(0), 1e-5);
+            }
+            valCnt++;
+        }
+        assertEquals(5, valCnt);
+    }
 }

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/layers/recurrent/TestRnnLayers.java

@@ -24,6 +24,7 @@ import org.deeplearning4j.nn.conf.dropout.TestDropout;
 import org.deeplearning4j.nn.conf.layers.GravesLSTM;
 import org.deeplearning4j.nn.conf.layers.LSTM;
 import org.deeplearning4j.nn.conf.layers.Layer;
+import org.deeplearning4j.nn.conf.layers.RnnLossLayer;
 import org.deeplearning4j.nn.conf.layers.RnnOutputLayer;
 import org.deeplearning4j.nn.conf.layers.recurrent.SimpleRnn;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
@@ -196,4 +197,43 @@ public class TestRnnLayers extends BaseDL4JTest {
         }
     }
 
+    @Test
+    public void testMismatchedInputLabelLength(){
+
+        for( int i=0; i<2; i++ ){
+
+            NeuralNetConfiguration.ListBuilder lb = new NeuralNetConfiguration.Builder()
+
+                    .list()
+                    .layer(new SimpleRnn.Builder().nIn(5).nOut(5).build());
+
+            switch (i){
+                case 0:
+                    lb.layer(new RnnOutputLayer.Builder().activation(Activation.SOFTMAX).lossFunction(LossFunctions.LossFunction.MCXENT).nIn(5).nOut(5).build());
+                    break;
+                case 1:
+                    lb.layer(new RnnLossLayer.Builder().activation(Activation.SOFTMAX).lossFunction(LossFunctions.LossFunction.MCXENT).build());
+                    break;
+                default:
+                    throw new RuntimeException();
+            }
+
+            MultiLayerConfiguration conf = lb.build();
+            MultiLayerNetwork net = new MultiLayerNetwork(conf);
+            net.init();
+
+            INDArray in = Nd4j.rand(DataType.FLOAT, 3, 5, 5);
+            INDArray l = TestUtils.randomOneHotTimeSeries(3, 5, 10);
+
+            try{
+                net.fit(in,l);
+            } catch (Throwable t){
+                String msg = t.getMessage();
+                assertTrue(msg, msg.contains("sequence length") && msg.contains("input") && msg.contains("label"));
+            }
+
+        }
+
+
+    }
 }

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/plot/BarnesHutTsneTest.java

@@ -249,7 +249,6 @@ public class BarnesHutTsneTest extends BaseDL4JTest {
     }
 
     @Test
-    @Ignore("AB 2019/05/31 - Failing on CI and locally - see issues 7820 and 7657")
     public void testCorrectness1() {
         DataTypeUtil.setDTypeForContext(DataType.DOUBLE);
         Nd4j.getRandom().setSeed(123);
@@ -270,30 +269,18 @@ public class BarnesHutTsneTest extends BaseDL4JTest {
                 .useAdaGrad(false).build();
 
         b.fit(data);
-        System.out.println(b.getData());
 
-        /*double[] expectedData = new double[]{15.5392794313924, 19.25226403656672, -5.194955746137196, -31.787679714614757, 48.8674725273665,
+        double[] expectedData = new double[]{  63.8206,   80.4013,  -19.4424, -140.4326,  198.7239,
-                24.92775755686273, -22.621939920239065, -29.790772278125395, 19.027362415188914, -16.013800175884274,
+                                              106.1148,  -96.6273, -124.3634,   78.4174,  -83.6621,
-                -27.454680593309185, 1.2929960811295493, -40.45000061571038, 61.23261682914338, 5.62278768938746,
+                                             -121.8706,    3.0888, -172.8560,  255.1262,   20.7021,
-                -28.16665244970911, -20.05502814088798, 12.803274346870865, -24.877262522905497, 45.115883138175874,
+                                             -120.7942,  -78.1829,   56.6021, -112.3294,  185.4084,
-                21.597495694710616, 18.63254779638783, -4.029728632528419, -0.4596087279592638, -42.35340705500429,
+                                               88.5330,   78.0497,  -18.8673,  -11.0155, -175.1564,
-                -69.24727547461491, 40.94332685199673, -24.60866142208024, 17.689874972878723, -3.6779759693605314,
+                                            -297.8463,  174.2511, -103.8793,   72.5455,  -15.8498,
-                -30.91803590368529, 10.645452930824145, 36.58583235020565, -64.74975614289316, -39.364099390585956,
+                                            -134.5235,   42.3300,  154.0391, -280.1010, -167.9765,
-                72.54886481127016, -35.30663155696714, 19.37116912936714, -7.790876543092118, 19.6586396288508,
+                                               306.9938, -150.9666,   83.4419,  -36.0877,   83.9992,
-                58.1332709511154, -18.49217368496203, -3.5050200971182424, 5.662891294031322, 39.69533295638775,
+                                               245.1813,  -81.5018,  -14.8430,   16.1557,  166.8651,
-                -15.114610550011662, -32.42366951357609, 17.039297537056537, 42.25610885633673, -2.7013781552769904,
+                                               -65.9247, -138.1783,   72.5444,  176.3088,  -25.6732,
-                -16.338582630617925, 41.734027526336874, 20.941332646863426, -3.2145240561108244, -45.36033539684912};*/
+                                               -69.6843,  167.3360,   87.6238,  -18.5874, -187.3806};
-        double[] expectedData = {40.93810899235225, 50.90183660191448, -14.298857560948981, -86.2012232604988, 129.51281793466023,
-                66.29136854264247, -61.650213611972326, -80.42836756633497, 50.28325210727952, -44.29008119040566,
-                -74.82748570869279, 2.0170536250746807, -109.21462846594635, 162.3973196127918, 14.000621153511705,
-                -76.30892822919527, -54.251704596942275, 33.99763310539589, -67.6307009607032, 119.50868525237786,
-                57.17786598853867, 49.1489174572297, -11.25663463504983, -2.38899196609398, -114.27194947404686,
-                -185.93832011474473, 108.9022579845252, -66.14099037301474, 47.13683038425694, -10.037893631405792,
-                -83.88458799629637, 26.985651418254996, 96.68139337135332, -174.2832443285551, -106.0999118697521,
-                193.02622700008175, -94.88003359113081, 51.39502524568139, -20.96021960048648, 52.32291574424741,
-                154.33973608321477, -50.90644802585217, -10.345744416395354, 13.721222143380892, 105.2111073677489,
-                -41.339268919407345, -87.73042354938127, 45.306865238870046, 112.53877133856602, -8.44454352074299,
-                -44.660828600669056, 110.72662022978719, 55.74660833987147, -9.613556053471232, -122.19953914048916};
 
         INDArray expectedArray = Nd4j.createFromArray(expectedData).reshape(11,5);
         for (int i = 0; i < expectedArray.rows(); ++i)

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/util/TimeSeriesUtilsTest.java

@@ -18,6 +18,7 @@ package org.deeplearning4j.util;
 
 import org.deeplearning4j.BaseDL4JTest;
 import org.junit.Test;
+import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.factory.Nd4j;
 
@@ -30,7 +31,7 @@ public class TimeSeriesUtilsTest extends BaseDL4JTest {
 
     @Test
     public void testMovingAverage() {
-        INDArray a = Nd4j.arange(0, 20);
+        INDArray a = Nd4j.arange(0, 20).castTo(DataType.DOUBLE);
         INDArray result = Nd4j.create(new double[] {1.5f, 2.5f, 3.5f, 4.5f, 5.5f, 6.5f, 7.5f, 8.5f, 9.5f, 10.5f, 11.5f,
                         12.5f, 13.5f, 14.5f, 15.5f, 16.5f, 17.5f});
 

deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/DataSetIteratorSplitter.java

@@ -24,6 +24,7 @@ import org.nd4j.linalg.dataset.api.DataSetPreProcessor;
 import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
 import org.nd4j.linalg.exception.ND4JIllegalStateException;
 
+import java.util.ArrayList;
 import java.util.List;
 import java.util.concurrent.atomic.AtomicBoolean;
 import java.util.concurrent.atomic.AtomicLong;
@@ -42,14 +43,20 @@ public class DataSetIteratorSplitter {
     protected DataSetIterator backedIterator;
     protected final long totalExamples;
     protected final double ratio;
+    protected final double[] ratios;
     protected final long numTrain;
     protected final long numTest;
+    protected final long numArbitrarySets;
+    protected final int[] splits;
+
 
     protected AtomicLong counter = new AtomicLong(0);
 
     protected AtomicBoolean resetPending = new AtomicBoolean(false);
     protected DataSet firstTrain = null;
 
+    protected int partNumber = 0;
+
     /**
      * The only constructor
      *
@@ -71,17 +78,94 @@ public class DataSetIteratorSplitter {
         this.backedIterator = baseIterator;
         this.totalExamples = totalBatches;
         this.ratio = ratio;
+        this.ratios = null;
         this.numTrain = (long) (totalExamples * ratio);
         this.numTest = totalExamples - numTrain;
+        this.numArbitrarySets = 2;
+        this.splits = null;
 
         log.warn("IteratorSplitter is used: please ensure you don't use randomization/shuffle in underlying iterator!");
     }
 
+    public DataSetIteratorSplitter(@NonNull DataSetIterator baseIterator, long totalBatches, double[] ratios) {
+        for (double ratio : ratios) {
+            if (!(ratio > 0.0 && ratio < 1.0))
+                throw new ND4JIllegalStateException("Ratio value should be in range of 0.0 > X < 1.0");
+        }
+
+        if (totalBatches < 0)
+            throw new ND4JIllegalStateException("totalExamples number should be positive value");
+
+        if (!baseIterator.resetSupported())
+            throw new ND4JIllegalStateException("Underlying iterator doesn't support reset, so it can't be used for runtime-split");
+
+
+        this.backedIterator = baseIterator;
+        this.totalExamples = totalBatches;
+        this.ratio = 0.0;
+        this.ratios = ratios;
+        this.numTrain = 0; //(long) (totalExamples * ratio);
+        this.numTest = 0; //totalExamples - numTrain;
+        this.numArbitrarySets = ratios.length;
+
+        this.splits = new int[this.ratios.length];
+        for (int i = 0; i < this.splits.length; ++i) {
+            this.splits[i] = (int)(totalExamples * ratios[i]);
+        }
+
+        log.warn("IteratorSplitter is used: please ensure you don't use randomization/shuffle in underlying iterator!");
+    }
+
+    public DataSetIteratorSplitter(@NonNull DataSetIterator baseIterator, int[] splits) {
+
+        /*if (!(simpleRatio > 0.0 && simpleRatio < 1.0))
+           throw new ND4JIllegalStateException("Ratio value should be in range of 0.0 > X < 1.0");*/
+
+        int totalBatches = 0;
+        for (val v:splits)
+            totalBatches += v;
+
+        if (totalBatches < 0)
+            throw new ND4JIllegalStateException("totalExamples number should be positive value");
+
+        if (!baseIterator.resetSupported())
+            throw new ND4JIllegalStateException("Underlying iterator doesn't support reset, so it can't be used for runtime-split");
+
+
+        this.backedIterator = baseIterator;
+        this.totalExamples = totalBatches;
+        this.ratio = 0.0;
+        this.ratios = null;
+
+        this.numTrain = 0; //(long) (totalExamples * ratio);
+        this.numTest = 0; //totalExamples - numTrain;
+        this.splits = splits;
+        this.numArbitrarySets = splits.length;
+
+        log.warn("IteratorSplitter is used: please ensure you don't use randomization/shuffle in underlying iterator!");
+    }
+
+    public List<DataSetIterator> getIterators() {
+        List<DataSetIterator> retVal = new ArrayList<>();
+        int partN = 0;
+        int bottom = 0;
+        for (final int split : splits) {
+                ScrollableDataSetIterator partIterator =
+                        new ScrollableDataSetIterator(partN++, backedIterator, counter, resetPending, firstTrain,
+                                new int[]{bottom,split});
+                bottom += split;
+                retVal.add(partIterator);
+        }
+        return retVal;
+    }
+
+
     /**
      * This method returns train iterator instance
      *
      * @return
      */
+    @Deprecated
     public DataSetIterator getTrainIterator() {
         return new DataSetIterator() {
             @Override
@@ -184,6 +268,7 @@ public class DataSetIteratorSplitter {
      *
      * @return
      */
+    @Deprecated
     public DataSetIterator getTestIterator() {
         return new DataSetIterator() {
             @Override

deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/MultiDataSetIteratorSplitter.java

@@ -21,9 +21,12 @@ import lombok.extern.slf4j.Slf4j;
 import lombok.val;
 import org.nd4j.linalg.dataset.api.MultiDataSet;
 import org.nd4j.linalg.dataset.api.MultiDataSetPreProcessor;
+import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
 import org.nd4j.linalg.dataset.api.iterator.MultiDataSetIterator;
 import org.nd4j.linalg.exception.ND4JIllegalStateException;
 
+import java.util.ArrayList;
+import java.util.List;
 import java.util.concurrent.atomic.AtomicBoolean;
 import java.util.concurrent.atomic.AtomicLong;
 
@@ -43,6 +46,9 @@ public class MultiDataSetIteratorSplitter {
     protected final double ratio;
     protected final long numTrain;
     protected final long numTest;
+    protected final double[] ratios;
+    protected final long numArbitrarySets;
+    protected final int[] splits;
 
     protected AtomicLong counter = new AtomicLong(0);
 
@@ -71,15 +77,87 @@ public class MultiDataSetIteratorSplitter {
         this.ratio = ratio;
         this.numTrain = (long) (totalExamples * ratio);
         this.numTest = totalExamples - numTrain;
+        this.ratios = null;
+        this.numArbitrarySets = 0;
+        this.splits = null;
 
         log.warn("IteratorSplitter is used: please ensure you don't use randomization/shuffle in underlying iterator!");
     }
 
+    public MultiDataSetIteratorSplitter(@NonNull MultiDataSetIterator baseIterator, long totalBatches, double[] ratios) {
+        for (double ratio : ratios) {
+            if (!(ratio > 0.0 && ratio < 1.0))
+                throw new ND4JIllegalStateException("Ratio value should be in range of 0.0 > X < 1.0");
+        }
+
+        if (totalBatches < 0)
+            throw new ND4JIllegalStateException("totalExamples number should be positive value");
+
+        if (!baseIterator.resetSupported())
+            throw new ND4JIllegalStateException("Underlying iterator doesn't support reset, so it can't be used for runtime-split");
+
+
+        this.backedIterator = baseIterator;
+        this.totalExamples = totalBatches;
+        this.ratio = 0.0;
+        this.numTrain = (long) (totalExamples * ratio);
+        this.numTest = totalExamples - numTrain;
+        this.ratios = null;
+        this.numArbitrarySets = ratios.length;
+
+        this.splits = new int[this.ratios.length];
+        for (int i = 0; i < this.splits.length; ++i) {
+            this.splits[i] = (int)(totalExamples * ratios[i]);
+        }
+
+        log.warn("IteratorSplitter is used: please ensure you don't use randomization/shuffle in underlying iterator!");
+    }
+
+    public MultiDataSetIteratorSplitter(@NonNull MultiDataSetIterator baseIterator, int[] splits) {
+
+        int totalBatches = 0;
+        for (val v:splits)
+            totalBatches += v;
+
+        if (totalBatches < 0)
+            throw new ND4JIllegalStateException("totalExamples number should be positive value");
+
+        if (!baseIterator.resetSupported())
+            throw new ND4JIllegalStateException("Underlying iterator doesn't support reset, so it can't be used for runtime-split");
+
+
+        this.backedIterator = baseIterator;
+        this.totalExamples = totalBatches;
+        this.ratio = 0.0;
+        this.numTrain = (long) (totalExamples * ratio);
+        this.numTest = totalExamples - numTrain;
+        this.ratios = null;
+        this.numArbitrarySets = splits.length;
+        this.splits = splits;
+
+        log.warn("IteratorSplitter is used: please ensure you don't use randomization/shuffle in underlying iterator!");
+    }
+
+    public List<MultiDataSetIterator> getIterators() {
+        List<MultiDataSetIterator> retVal = new ArrayList<>();
+        int partN = 0;
+        int bottom = 0;
+        for (final int split : splits) {
+            ScrollableMultiDataSetIterator partIterator =
+                    new ScrollableMultiDataSetIterator(partN++, backedIterator, counter, firstTrain,
+                            new int[]{bottom,split});
+            bottom += split;
+            retVal.add(partIterator);
+        }
+        return retVal;
+    }
+
     /**
      * This method returns train iterator instance
      *
      * @return
      */
+    @Deprecated
     public MultiDataSetIterator getTrainIterator() {
         return new MultiDataSetIterator() {
             @Override
@@ -162,6 +240,7 @@ public class MultiDataSetIteratorSplitter {
      *
      * @return
      */
+    @Deprecated
     public MultiDataSetIterator getTestIterator() {
         return new MultiDataSetIterator() {
             @Override

deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/ScrollableDataSetIterator.java

@@ -0,0 +1,158 @@
+package org.deeplearning4j.datasets.iterator;
+
+import lombok.val;
+import org.nd4j.linalg.dataset.DataSet;
+import org.nd4j.linalg.dataset.MultiDataSet;
+import org.nd4j.linalg.dataset.api.DataSetPreProcessor;
+import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
+import org.nd4j.linalg.dataset.api.iterator.MultiDataSetIterator;
+
+import java.util.List;
+import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.concurrent.atomic.AtomicLong;
+
+public class ScrollableDataSetIterator implements DataSetIterator {
+    private int thisPart = 0;
+    private int top = 0;
+    private int bottom = 0;
+    protected DataSetIterator backedIterator;
+    protected AtomicLong counter = new AtomicLong(0);
+
+    protected AtomicBoolean resetPending = new AtomicBoolean(false);
+    protected DataSet firstTrain = null;
+    protected MultiDataSet firstMultiTrain = null;
+    private double ratio;
+    private long totalExamples;
+    private long itemsPerPart;
+    private long current;
+
+
+    public ScrollableDataSetIterator(int num, DataSetIterator backedIterator, AtomicLong counter,
+                                     AtomicBoolean resetPending, DataSet firstTrain, double ratio,
+                                     int totalExamples) {
+        this.thisPart = num;
+        this.backedIterator = backedIterator;
+        this.counter = counter;
+        this.resetPending = resetPending;
+        this.firstTrain = firstTrain;
+        this.ratio = ratio;
+        this.totalExamples = totalExamples;
+        this.itemsPerPart = (long)(totalExamples * ratio);
+        this.current = 0;
+    }
+
+    public ScrollableDataSetIterator(int num, DataSetIterator backedIterator, AtomicLong counter,
+                                     AtomicBoolean resetPending, DataSet firstTrain,
+                                     int[] itemsPerPart) {
+        this.thisPart = num;
+        this.bottom = itemsPerPart[0];
+        this.top = bottom + itemsPerPart[1];
+        this.itemsPerPart = top;
+
+        this.backedIterator = backedIterator;
+        this.counter = counter;
+        //this.resetPending = resetPending;
+        this.firstTrain = firstTrain;
+        //this.totalExamples = totalExamples;
+        this.current = 0;
+    }
+
+    @Override
+    public DataSet next(int i) {
+        throw new UnsupportedOperationException();
+    }
+
+    @Override
+    public List<String> getLabels() {
+        return backedIterator.getLabels();
+    }
+
+    @Override
+    public int inputColumns() {
+        return backedIterator.inputColumns();
+    }
+
+    @Override
+    public void remove() {
+        throw new UnsupportedOperationException();
+    }
+
+    @Override
+    public int totalOutcomes() {
+        return backedIterator.totalOutcomes();
+    }
+
+    @Override
+    public boolean resetSupported() {
+        return backedIterator.resetSupported();
+    }
+
+    @Override
+    public boolean asyncSupported() {
+        return backedIterator.asyncSupported();
+    }
+
+    @Override
+    public void reset() {
+        resetPending.set(true);
+    }
+
+    @Override
+    public int batch() {
+        return backedIterator.batch();
+    }
+
+    @Override
+    public void setPreProcessor(DataSetPreProcessor dataSetPreProcessor) {
+        backedIterator.setPreProcessor(dataSetPreProcessor);
+    }
+
+    @Override
+    public DataSetPreProcessor getPreProcessor() {
+
+        return backedIterator.getPreProcessor();
+    }
+
+
+    @Override
+    public boolean hasNext() {
+        if (resetPending.get()) {
+            if (resetSupported()) {
+                backedIterator.reset();
+                counter.set(0);
+                current = 0;
+                resetPending.set(false);
+            } else
+                throw new UnsupportedOperationException("Reset isn't supported by underlying iterator");
+        }
+
+        boolean state = false;
+        if (current >= top)
+            return false;
+        state = backedIterator.hasNext();
+        if (!state)
+            return false;
+        if (state && counter.get() < itemsPerPart)
+            return true;
+        else
+            return false;
+
+    }
+
+    @Override
+    public DataSet next() {
+        counter.incrementAndGet();
+        if ((current == 0) && (bottom != 0)) {
+            backedIterator.reset();
+            long cnt = current;
+            for (; cnt < bottom; ++cnt) {
+                if (backedIterator.hasNext())
+                    backedIterator.next();
+            }
+            current = cnt+1;
+        }
+        else current++;
+        val p = backedIterator.next();
+        return p;
+    }
+}

deeplearning4j/deeplearning4j-data/deeplearning4j-utility-iterators/src/main/java/org/deeplearning4j/datasets/iterator/ScrollableMultiDataSetIterator.java

@@ -0,0 +1,121 @@
+package org.deeplearning4j.datasets.iterator;
+
+import lombok.val;
+import org.nd4j.linalg.dataset.DataSet;
+import org.nd4j.linalg.dataset.api.MultiDataSet;
+import org.nd4j.linalg.dataset.api.DataSetPreProcessor;
+import org.nd4j.linalg.dataset.api.MultiDataSetPreProcessor;
+import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
+import org.nd4j.linalg.dataset.api.iterator.MultiDataSetIterator;
+
+import javax.naming.OperationNotSupportedException;
+import java.util.List;
+import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.concurrent.atomic.AtomicLong;
+
+public class ScrollableMultiDataSetIterator implements MultiDataSetIterator {
+    private int thisPart = 0;
+    private int top = 0;
+    private int bottom = 0;
+    protected MultiDataSetIterator backedIterator;
+    protected AtomicLong counter = new AtomicLong(0);
+
+    protected AtomicBoolean resetPending = new AtomicBoolean(false);
+    protected DataSet firstTrain = null;
+    protected MultiDataSet firstMultiTrain = null;
+    private double ratio;
+    private long totalExamples;
+    private long itemsPerPart;
+    private long current;
+
+    public ScrollableMultiDataSetIterator(int num, MultiDataSetIterator backedIterator, AtomicLong counter,
+                                     MultiDataSet firstTrain,  int[] itemsPerPart) {
+        this.thisPart = num;
+        this.bottom = itemsPerPart[0];
+        this.top = bottom + itemsPerPart[1];
+        this.itemsPerPart = top;
+
+        this.counter = counter;
+        //this.resetPending = resetPending;
+        this.firstTrain = null;
+        this.firstMultiTrain = firstTrain;
+        //this.totalExamples = totalExamples;
+        this.current = 0;
+        this.backedIterator = backedIterator;
+        this.resetPending = resetPending;
+    }
+
+    @Override
+    public boolean resetSupported() {
+        return backedIterator.resetSupported();
+    }
+
+    @Override
+    public boolean asyncSupported() {
+        return backedIterator.asyncSupported();
+    }
+
+    @Override
+    public void reset() {
+        resetPending.set(true);
+    }
+
+    @Override
+    public void setPreProcessor(MultiDataSetPreProcessor dataSetPreProcessor) {
+        backedIterator.setPreProcessor(dataSetPreProcessor);
+    }
+
+    @Override
+    public MultiDataSetPreProcessor getPreProcessor() {
+
+        throw new UnsupportedOperationException();
+    }
+
+
+    @Override
+    public boolean hasNext() {
+        if (resetPending.get()) {
+            if (resetSupported()) {
+                backedIterator.reset();
+                counter.set(0);
+                current = 0;
+                resetPending.set(false);
+            } else
+                throw new UnsupportedOperationException("Reset isn't supported by underlying iterator");
+        }
+
+        boolean state = false;
+        if (current >= top)
+            return false;
+        state = backedIterator.hasNext();
+        if (!state)
+            return false;
+        if (state && counter.get() < itemsPerPart)
+            return true;
+        else
+            return false;
+
+    }
+
+    @Override
+    public MultiDataSet next() {
+        counter.incrementAndGet();
+        if ((current == 0) && (bottom != 0)) {
+            backedIterator.reset();
+            long cnt = current;
+            for (; cnt < bottom; ++cnt) {
+                if (backedIterator.hasNext())
+                    backedIterator.next();
+            }
+            current = cnt+1;
+        }
+        else current++;
+        val p = backedIterator.next();
+        return p;
+    }
+
+    @Override
+    public MultiDataSet next(int i) {
+        throw new UnsupportedOperationException();
+    }
+}

deeplearning4j/deeplearning4j-modelimport/src/main/java/org/deeplearning4j/nn/modelimport/keras/Hdf5Archive.java

@@ -47,6 +47,8 @@ import static org.bytedeco.hdf5.global.hdf5.*;
 @Slf4j
 public class Hdf5Archive implements Closeable {
 
+    public static final int MAX_BUFFER_SIZE_BYTES = (int)Math.pow(2, 28);       //256 MB
+
     /**
      * HDF5 library is not thread safe - possible to crash if multiple reads etc are performed concurrently
      * in multiple threads. This object is used for locking read etc activity using synchronized blocks
@@ -338,7 +340,7 @@ public class Hdf5Archive implements Closeable {
     private String readAttributeAsJson(Attribute attribute) throws UnsupportedKerasConfigurationException {
         synchronized (Hdf5Archive.LOCK_OBJECT) {
             VarLenType vl = attribute.getVarLenType();
-            int bufferSizeMult = 1;
+            int currBufferLength = 2048;
             String s;
             /* TODO: find a less hacky way to do this.
              * Reading variable length strings (from attributes) is a giant
@@ -349,8 +351,8 @@ public class Hdf5Archive implements Closeable {
              * buffer and repeat.
              */
             while (true) {
-                byte[] attrBuffer = new byte[bufferSizeMult * 2000];
+                byte[] attrBuffer = new byte[currBufferLength];
-                BytePointer attrPointer = new BytePointer(attrBuffer);
+                BytePointer attrPointer = new BytePointer(currBufferLength);
                 attribute.read(vl, attrPointer);
                 attrPointer.get(attrBuffer);
                 s = new String(attrBuffer);
@@ -362,9 +364,11 @@ public class Hdf5Archive implements Closeable {
                 } catch (IOException e) {
                     //OK - we don't know how long the buffer needs to be, so we'll try again with larger buffer
                 }
-                bufferSizeMult *= 2;
+
-                if (bufferSizeMult > 1024) {
+                if(currBufferLength == MAX_BUFFER_SIZE_BYTES){
-                    throw new UnsupportedKerasConfigurationException("Could not read abnormally long HDF5 attribute");
+                    throw new UnsupportedKerasConfigurationException("Could not read abnormally long HDF5 attribute: size exceeds " + currBufferLength + " bytes");
+                } else {
+                    currBufferLength = (int)Math.min(MAX_BUFFER_SIZE_BYTES, currBufferLength * 4L);
                 }
             }
             vl.deallocate();

deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/algorithm/BaseClusteringAlgorithm.java

@@ -21,6 +21,7 @@ import lombok.NoArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import lombok.val;
 import org.apache.commons.lang3.ArrayUtils;
+import org.apache.commons.math3.ml.clustering.KMeansPlusPlusClusterer;
 import org.deeplearning4j.clustering.cluster.Cluster;
 import org.deeplearning4j.clustering.cluster.ClusterSet;
 import org.deeplearning4j.clustering.cluster.ClusterUtils;
@@ -62,12 +63,13 @@ public class BaseClusteringAlgorithm implements ClusteringAlgorithm, Serializabl
     private ClusterSet clusterSet;
     private List<Point> initialPoints;
     private transient ExecutorService exec;
+    private boolean useKmeansPlusPlus;
 
 
-
+    protected BaseClusteringAlgorithm(ClusteringStrategy clusteringStrategy, boolean useKmeansPlusPlus) {
-    protected BaseClusteringAlgorithm(ClusteringStrategy clusteringStrategy) {
         this.clusteringStrategy = clusteringStrategy;
         this.exec = MultiThreadUtils.newExecutorService();
+        this.useKmeansPlusPlus = useKmeansPlusPlus;
     }
 
     /**
@@ -75,8 +77,8 @@ public class BaseClusteringAlgorithm implements ClusteringAlgorithm, Serializabl
      * @param clusteringStrategy
      * @return
      */
-    public static BaseClusteringAlgorithm setup(ClusteringStrategy clusteringStrategy) {
+    public static BaseClusteringAlgorithm setup(ClusteringStrategy clusteringStrategy, boolean useKmeansPlusPlus) {
-        return new BaseClusteringAlgorithm(clusteringStrategy);
+        return new BaseClusteringAlgorithm(clusteringStrategy, useKmeansPlusPlus);
     }
 
     /**
@@ -86,7 +88,7 @@ public class BaseClusteringAlgorithm implements ClusteringAlgorithm, Serializabl
      */
     public ClusterSet applyTo(List<Point> points) {
         resetState(points);
-        initClusters();
+        initClusters(useKmeansPlusPlus);
         iterations();
         return clusterSet;
     }
@@ -130,7 +132,7 @@ public class BaseClusteringAlgorithm implements ClusteringAlgorithm, Serializabl
      * Initialize the
      * cluster centers at random
      */
-    protected void initClusters() {
+    protected void initClusters(boolean kMeansPlusPlus) {
         log.info("Generating initial clusters");
         List<Point> points = new ArrayList<>(initialPoints);
 
@@ -152,7 +154,10 @@ public class BaseClusteringAlgorithm implements ClusteringAlgorithm, Serializabl
         //Thus, we are more likely to select (as a new cluster center) a point that is far from an existing cluster
         while (clusterSet.getClusterCount() < initialClusterCount && !points.isEmpty()) {
             dxs = ClusterUtils.computeSquareDistancesFromNearestCluster(clusterSet, points, dxs, exec);
-            double r = random.nextFloat() * dxs.maxNumber().doubleValue();
+            double summed = Nd4j.sum(dxs).getDouble(0);
+            double r = kMeansPlusPlus ? random.nextDouble() * summed:
+                                        random.nextFloat() * dxs.maxNumber().doubleValue();
+
             for (int i = 0; i < dxs.length(); i++) {
                 double distance = dxs.getDouble(i);
                 Preconditions.checkState(distance >= 0, "Encountered negative distance: distance function is not valid? Distance " +
@@ -170,6 +175,7 @@ public class BaseClusteringAlgorithm implements ClusteringAlgorithm, Serializabl
                         new IterationInfo(currentIteration, initialClusterSetInfo));
     }
 
+
     protected void applyClusteringStrategy() {
         if (!isStrategyApplicableNow())
             return;

deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/cluster/ClusterUtils.java

@@ -79,8 +79,8 @@ public class ClusterUtils {
         int nClusters = clusterSet.getClusterCount();
         for (int i = 0; i < nClusters; i++) {
             final Cluster cluster = clusterSet.getClusters().get(i);
-            tasks.add(new Runnable() {
+            //tasks.add(new Runnable() {
-                public void run() {
+            //    public void run() {
                     try {
                         final ClusterInfo clusterInfo = clusterSetInfo.getClusterInfo(cluster.getId());
                         refreshClusterCenter(cluster, clusterInfo);
@@ -88,10 +88,10 @@ public class ClusterUtils {
                     } catch (Throwable t) {
                         log.warn("Error refreshing cluster centers", t);
                     }
+            //    }
+            //});
         }
-            });
+        //MultiThreadUtils.parallelTasks(tasks, executorService);
-        }
-        MultiThreadUtils.parallelTasks(tasks, executorService);
     }
 
     public static void refreshClusterCenter(Cluster cluster, ClusterInfo clusterInfo) {
@@ -146,28 +146,29 @@ public class ClusterUtils {
         List<Runnable> tasks = new ArrayList<>();
         for (int i = 0; i < pointsCount; i++) {
             final int i2 = i;
-            tasks.add(new Runnable() {
+            //tasks.add(new Runnable() {
-                public void run() {
+            //    public void run() {
                     try {
                         Point point = points.get(i2);
                         double dist = clusterSet.isInverse() ? newCluster.getDistanceToCenter(point)
                                 : Math.pow(newCluster.getDistanceToCenter(point), 2);
-                        dxs.putScalar(i2, clusterSet.isInverse() ? dist : dist);
+                        dxs.putScalar(i2, /*clusterSet.isInverse() ? dist :*/ dist);
                     } catch (Throwable t) {
                         log.warn("Error computing squared distance from nearest cluster", t);
                     }
-                }
+            //    }
-            });
+            //});
 
         }
 
-        MultiThreadUtils.parallelTasks(tasks, executorService);
+        //MultiThreadUtils.parallelTasks(tasks, executorService);
-
         for (int i = 0; i < pointsCount; i++) {
             double previousMinDistance = previousDxs.getDouble(i);
             if (clusterSet.isInverse()) {
-                if (dxs.getDouble(i) < previousMinDistance)
+                if (dxs.getDouble(i) < previousMinDistance) {
+
                     dxs.putScalar(i, previousMinDistance);
+                }
             } else if (dxs.getDouble(i) > previousMinDistance)
                 dxs.putScalar(i, previousMinDistance);
         }
@@ -175,6 +176,23 @@ public class ClusterUtils {
         return dxs;
     }
 
+    public static INDArray computeWeightedProbaDistancesFromNearestCluster(final ClusterSet clusterSet,
+                                                                    final List<Point> points, INDArray previousDxs) {
+        final int pointsCount = points.size();
+        final INDArray dxs = Nd4j.create(pointsCount);
+        final Cluster newCluster = clusterSet.getClusters().get(clusterSet.getClusters().size() - 1);
+
+        Double sum = new Double(0);
+        for (int i = 0; i < pointsCount; i++) {
+
+                Point point = points.get(i);
+                double dist = Math.pow(newCluster.getDistanceToCenter(point), 2);
+                sum += dist;
+                dxs.putScalar(i, sum);
+        }
+
+        return dxs;
+    }
     /**
      *
      * @param clusterSet
@@ -194,27 +212,27 @@ public class ClusterUtils {
         List<Runnable> tasks = new ArrayList<>();
         for (int i = 0; i < clusterCount; i++) {
             final Cluster cluster = clusterSet.getClusters().get(i);
-            tasks.add(new Runnable() {
+            //tasks.add(new Runnable() {
-                public void run() {
+            //    public void run() {
                     try {
                         info.getClustersInfos().put(cluster.getId(),
                                 computeClusterInfos(cluster, clusterSet.getDistanceFunction()));
                     } catch (Throwable t) {
                         log.warn("Error computing cluster set info", t);
                     }
-                }
+                //}
-            });
+            //});
         }
 
 
-        MultiThreadUtils.parallelTasks(tasks, executorService);
+        //MultiThreadUtils.parallelTasks(tasks, executorService);
 
-        tasks = new ArrayList<>();
+        //tasks = new ArrayList<>();
         for (int i = 0; i < clusterCount; i++) {
             final int clusterIdx = i;
             final Cluster fromCluster = clusterSet.getClusters().get(i);
-            tasks.add(new Runnable() {
+            //tasks.add(new Runnable() {
-                public void run() {
+                //public void run() {
                     try {
                         for (int k = clusterIdx + 1, l = clusterSet.getClusterCount(); k < l; k++) {
                             Cluster toCluster = clusterSet.getClusters().get(k);
@@ -230,12 +248,12 @@ public class ClusterUtils {
                     } catch (Throwable t) {
                         log.warn("Error computing distances", t);
                     }
-                }
+            //    }
-            });
+            //});
 
         }
 
-        MultiThreadUtils.parallelTasks(tasks, executorService);
+        //MultiThreadUtils.parallelTasks(tasks, executorService);
 
         return info;
     }

deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/main/java/org/deeplearning4j/clustering/kmeans/KMeansClustering.java

@@ -37,8 +37,8 @@ public class KMeansClustering extends BaseClusteringAlgorithm {
      *
      * @param clusteringStrategy
      */
-    protected KMeansClustering(ClusteringStrategy clusteringStrategy) {
+    protected KMeansClustering(ClusteringStrategy clusteringStrategy, boolean useKMeansPlusPlus) {
-        super(clusteringStrategy);
+        super(clusteringStrategy, useKMeansPlusPlus);
     }
 
     /**
@@ -50,11 +50,11 @@ public class KMeansClustering extends BaseClusteringAlgorithm {
      * @return
      */
     public static KMeansClustering setup(int clusterCount, int maxIterationCount, Distance distanceFunction,
-                    boolean inverse) {
+                    boolean inverse, boolean useKMeansPlusPlus) {
         ClusteringStrategy clusteringStrategy =
                         FixedClusterCountStrategy.setup(clusterCount, distanceFunction, inverse);
         clusteringStrategy.endWhenIterationCountEquals(maxIterationCount);
-        return new KMeansClustering(clusteringStrategy);
+        return new KMeansClustering(clusteringStrategy, useKMeansPlusPlus);
     }
 
     /**
@@ -66,10 +66,10 @@ public class KMeansClustering extends BaseClusteringAlgorithm {
      * @return
      */
     public static KMeansClustering setup(int clusterCount, double minDistributionVariationRate, Distance distanceFunction,
-                    boolean inverse, boolean allowEmptyClusters) {
+                    boolean inverse, boolean allowEmptyClusters, boolean useKMeansPlusPlus) {
         ClusteringStrategy clusteringStrategy = FixedClusterCountStrategy.setup(clusterCount, distanceFunction, inverse)
                         .endWhenDistributionVariationRateLessThan(minDistributionVariationRate);
-        return new KMeansClustering(clusteringStrategy);
+        return new KMeansClustering(clusteringStrategy, useKMeansPlusPlus);
     }
 
 
@@ -81,8 +81,8 @@ public class KMeansClustering extends BaseClusteringAlgorithm {
      * @param distanceFunction the distance function to use for grouping
      * @return
      */
-    public static KMeansClustering setup(int clusterCount, int maxIterationCount, Distance distanceFunction) {
+    public static KMeansClustering setup(int clusterCount, int maxIterationCount, Distance distanceFunction, boolean useKMeansPlusPlus) {
-        return setup(clusterCount, maxIterationCount, distanceFunction, false);
+        return setup(clusterCount, maxIterationCount, distanceFunction, false, useKMeansPlusPlus);
     }
 
     /**
@@ -94,17 +94,17 @@ public class KMeansClustering extends BaseClusteringAlgorithm {
      * @return
      */
     public static KMeansClustering setup(int clusterCount, double minDistributionVariationRate, Distance distanceFunction,
-                    boolean allowEmptyClusters) {
+                    boolean allowEmptyClusters, boolean useKMeansPlusPlus) {
         ClusteringStrategy clusteringStrategy = FixedClusterCountStrategy.setup(clusterCount, distanceFunction, false);
         clusteringStrategy.endWhenDistributionVariationRateLessThan(minDistributionVariationRate);
-        return new KMeansClustering(clusteringStrategy);
+        return new KMeansClustering(clusteringStrategy, useKMeansPlusPlus);
     }
 
     public static KMeansClustering setup(int clusterCount, Distance distanceFunction,
-                                         boolean allowEmptyClusters) {
+                                         boolean allowEmptyClusters, boolean useKMeansPlusPlus) {
         ClusteringStrategy clusteringStrategy = FixedClusterCountStrategy.setup(clusterCount, distanceFunction, false);
         clusteringStrategy.endWhenDistributionVariationRateLessThan(VARIATION_TOLERANCE);
-        return new KMeansClustering(clusteringStrategy);
+        return new KMeansClustering(clusteringStrategy, useKMeansPlusPlus);
     }
 
 }

deeplearning4j/deeplearning4j-nearestneighbors-parent/nearestneighbor-core/src/test/java/org/deeplearning4j/clustering/kmeans/KMeansTest.java

@@ -16,6 +16,7 @@
 
 package org.deeplearning4j.clustering.kmeans;
 
+import lombok.val;
 import org.apache.commons.lang3.time.StopWatch;
 import org.deeplearning4j.clustering.BaseDL4JTest;
 import org.deeplearning4j.clustering.algorithm.Distance;
@@ -28,36 +29,40 @@ import org.nd4j.linalg.factory.Nd4j;
 
 import java.util.List;
 
-import static org.junit.Assert.assertEquals;
+import static org.junit.Assert.*;
-import static org.junit.Assert.fail;
 
 /**
  * Created by agibsonccc on 7/2/17.
  */
 public class KMeansTest extends BaseDL4JTest {
 
+    private boolean[] useKMeansPlusPlus = {true, false};
+
     @Test
     public void testKMeans() {
         Nd4j.getRandom().setSeed(7);
-        KMeansClustering kMeansClustering = KMeansClustering.setup(5, 5, Distance.EUCLIDEAN);
+        for (boolean mode : useKMeansPlusPlus) {
+            KMeansClustering kMeansClustering = KMeansClustering.setup(5, 5, Distance.EUCLIDEAN, mode);
             List<Point> points = Point.toPoints(Nd4j.randn(5, 5));
             ClusterSet clusterSet = kMeansClustering.applyTo(points);
             PointClassification pointClassification = clusterSet.classifyPoint(points.get(0));
             System.out.println(pointClassification);
         }
+    }
 
     @Test
     public void testKmeansCosine() {
 
         Nd4j.getRandom().setSeed(7);
         int numClusters = 5;
-        KMeansClustering kMeansClustering = KMeansClustering.setup(numClusters, 1000, Distance.COSINE_DISTANCE, true);
+        for (boolean mode : useKMeansPlusPlus) {
+            KMeansClustering kMeansClustering = KMeansClustering.setup(numClusters, 1000, Distance.COSINE_DISTANCE, mode);
             List<Point> points = Point.toPoints(Nd4j.rand(5, 300));
             ClusterSet clusterSet = kMeansClustering.applyTo(points);
             PointClassification pointClassification = clusterSet.classifyPoint(points.get(0));
 
 
-        KMeansClustering kMeansClusteringEuclidean = KMeansClustering.setup(numClusters, 1000, Distance.EUCLIDEAN);
+            KMeansClustering kMeansClusteringEuclidean = KMeansClustering.setup(numClusters, 1000, Distance.EUCLIDEAN, mode);
             ClusterSet clusterSetEuclidean = kMeansClusteringEuclidean.applyTo(points);
             PointClassification pointClassificationEuclidean = clusterSetEuclidean.classifyPoint(points.get(0));
             System.out.println("Cosine " + pointClassification);
@@ -66,6 +71,7 @@ public class KMeansTest extends BaseDL4JTest {
             assertEquals(pointClassification.getCluster().getPoints().get(0),
                     pointClassificationEuclidean.getCluster().getPoints().get(0));
         }
+    }
 
     @Ignore
     @Test
@@ -73,10 +79,11 @@ public class KMeansTest extends BaseDL4JTest {
         Nd4j.setDefaultDataTypes(DataType.DOUBLE, DataType.DOUBLE);
         Nd4j.getRandom().setSeed(7);
         int numClusters = 20;
+        for (boolean mode : useKMeansPlusPlus) {
             StopWatch watch = new StopWatch();
             watch.start();
-        KMeansClustering kMeansClustering = KMeansClustering.setup(numClusters, 1000, Distance.COSINE_DISTANCE, true);
+            KMeansClustering kMeansClustering = KMeansClustering.setup(numClusters, 1000, Distance.COSINE_DISTANCE, mode);
-        List<Point> points = Point.toPoints(Nd4j.linspace(0, 5000*300, 5000*300).reshape(5000,300 ));
+            List<Point> points = Point.toPoints(Nd4j.linspace(0, 5000 * 300, 5000 * 300).reshape(5000, 300));
 
             ClusterSet clusterSet = kMeansClustering.applyTo(points);
             watch.stop();
@@ -90,6 +97,7 @@ public class KMeansTest extends BaseDL4JTest {
             watch.stop();
             System.out.println("Elapsed for search: " + watch);
         }
+    }
 
     @Test
     @Ignore
@@ -97,11 +105,12 @@ public class KMeansTest extends BaseDL4JTest {
         Nd4j.setDefaultDataTypes(DataType.DOUBLE, DataType.DOUBLE);
         Nd4j.getRandom().setSeed(7);
         int numClusters = 20;
+        for (boolean mode : useKMeansPlusPlus) {
             StopWatch watch = new StopWatch();
             watch.start();
-        KMeansClustering kMeansClustering = KMeansClustering.setup(numClusters, Distance.COSINE_DISTANCE, false);
+            KMeansClustering kMeansClustering = KMeansClustering.setup(numClusters, Distance.COSINE_DISTANCE, false, mode);
 
-        List<Point> points = Point.toPoints(Nd4j.linspace(0, 10000*300, 10000*300).reshape(10000,300 ));
+            List<Point> points = Point.toPoints(Nd4j.linspace(0, 10000 * 300, 10000 * 300).reshape(10000, 300));
 
             ClusterSet clusterSet = kMeansClustering.applyTo(points);
             watch.stop();
@@ -117,9 +126,9 @@ public class KMeansTest extends BaseDL4JTest {
 
             watch.reset();
             watch.start();
-        kMeansClustering = KMeansClustering.setup(numClusters, 0.05, Distance.COSINE_DISTANCE, false);
+            kMeansClustering = KMeansClustering.setup(numClusters, 0.05, Distance.COSINE_DISTANCE, false, mode);
 
-        points = Point.toPoints(Nd4j.linspace(0, 10000*300, 10000*300).reshape(10000,300 ));
+            points = Point.toPoints(Nd4j.linspace(0, 10000 * 300, 10000 * 300).reshape(10000, 300));
 
             clusterSet = kMeansClustering.applyTo(points);
             watch.stop();
@@ -133,6 +142,7 @@ public class KMeansTest extends BaseDL4JTest {
             watch.stop();
             System.out.println("Elapsed for search: " + watch);
         }
+    }
 
     @Test
     public void testCorrectness() {
@@ -141,7 +151,8 @@ public class KMeansTest extends BaseDL4JTest {
             Nd4j.setDefaultDataTypes(DataType.DOUBLE, DataType.DOUBLE);
             Nd4j.getRandom().setSeed(7);
             int numClusters = 3;
-            KMeansClustering kMeansClustering = KMeansClustering.setup(numClusters, 1000, Distance.EUCLIDEAN, true);
+            for (boolean mode : useKMeansPlusPlus) {
+                KMeansClustering kMeansClustering = KMeansClustering.setup(numClusters, 1000, Distance.EUCLIDEAN, mode);
                 double[] data = new double[]{
                         15, 16,
                         16, 18.5,
@@ -181,6 +192,7 @@ public class KMeansTest extends BaseDL4JTest {
                     for (int i = 0; i < clusters.size(); ++i)
                         System.out.println("Choice: " + clusters.get(i).getCenter().getArray());
                 }
+            }
             /*assertEquals(Nd4j.createFromArray(new double[]{75.9348, 74.1990}),
                     pointClassification.getCluster().getCenter().getArray());*/
 
@@ -233,4 +245,39 @@ public class KMeansTest extends BaseDL4JTest {
             System.out.println();
         }
     }
+
+    @Test
+    public void testInitClusters() {
+        Nd4j.setDefaultDataTypes(DataType.DOUBLE, DataType.DOUBLE);
+        Nd4j.getRandom().setSeed(7);
+        {
+            KMeansClustering kMeansClustering = KMeansClustering.setup(5, 1, Distance.EUCLIDEAN, true);
+
+            double[][] dataArray = {{1000000.0, 2.8E7, 5.5E7, 8.2E7}, {2.8E7, 5.5E7, 8.2E7, 1.09E8}, {5.5E7, 8.2E7, 1.09E8, 1.36E8},
+                    {8.2E7, 1.09E8, 1.36E8, 1.63E8}, {1.09E8, 1.36E8, 1.63E8, 1.9E8}, {1.36E8, 1.63E8, 1.9E8, 2.17E8},
+                    {1.63E8, 1.9E8, 2.17E8, 2.44E8}, {1.9E8, 2.17E8, 2.44E8, 2.71E8}, {2.17E8, 2.44E8, 2.71E8, 2.98E8},
+                    {2.44E8, 2.71E8, 2.98E8, 3.25E8}, {2.71E8, 2.98E8, 3.25E8, 3.52E8}, {2.98E8, 3.25E8, 3.52E8, 3.79E8},
+                    {3.25E8, 3.52E8, 3.79E8, 4.06E8}, {3.52E8, 3.79E8, 4.06E8, 4.33E8}, {3.79E8, 4.06E8, 4.33E8, 4.6E8},
+                    {4.06E8, 4.33E8, 4.6E8, 4.87E8}, {4.33E8, 4.6E8, 4.87E8, 5.14E8}, {4.6E8, 4.87E8, 5.14E8, 5.41E8},
+                    {4.87E8, 5.14E8, 5.41E8, 5.68E8}, {5.14E8, 5.41E8, 5.68E8, 5.95E8}, {5.41E8, 5.68E8, 5.95E8, 6.22E8},
+                    {5.68E8, 5.95E8, 6.22E8, 6.49E8}, {5.95E8, 6.22E8, 6.49E8, 6.76E8}, {6.22E8, 6.49E8, 6.76E8, 7.03E8},
+                    {6.49E8, 6.76E8, 7.03E8, 7.3E8}, {6.76E8, 7.03E8, 7.3E8, 7.57E8}, {7.03E8, 7.3E8, 7.57E8, 7.84E8}};
+            INDArray data = Nd4j.createFromArray(dataArray);
+            List<Point> points = Point.toPoints(data);
+
+            ClusterSet clusterSet = kMeansClustering.applyTo(points);
+
+            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};
+
+            assertArrayEquals(centroid1, clusterSet.getClusters().get(0).getCenter().getArray().toDoubleVector(), 1e-4);
+            assertArrayEquals(centroid2, clusterSet.getClusters().get(1).getCenter().getArray().toDoubleVector(), 1e-4);
+            assertArrayEquals(centroid3, clusterSet.getClusters().get(2).getCenter().getArray().toDoubleVector(), 1e-4);
+            assertArrayEquals(centroid4, clusterSet.getClusters().get(3).getCenter().getArray().toDoubleVector(), 1e-4);
+            assertArrayEquals(centroid5, clusterSet.getClusters().get(4).getCenter().getArray().toDoubleVector(), 1e-4);
+        }
+    }
 }

deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-uima/src/test/java/org/deeplearning4j/models/WordVectorSerializerTest.java

@@ -23,6 +23,8 @@ import org.apache.commons.io.FileUtils;
 import org.apache.commons.lang.ArrayUtils;
 import org.apache.commons.lang3.RandomUtils;
 import org.deeplearning4j.BaseDL4JTest;
+import org.deeplearning4j.models.sequencevectors.SequenceVectors;
+import org.deeplearning4j.models.sequencevectors.serialization.VocabWordFactory;
 import org.junit.Rule;
 import org.junit.rules.TemporaryFolder;
 import org.nd4j.linalg.io.ClassPathResource;
@@ -857,4 +859,34 @@ public class WordVectorSerializerTest extends BaseDL4JTest {
         }
     }
 
+    @Test
+    public void testBackwardsCompatibleWord2Vec() {
+        File model_v3 = Resources.asFile("deeplearning4j-nlp/model_beta3.zip");
+        File model_v4 = Resources.asFile("deeplearning4j-nlp/model_beta4.zip");
+        Word2Vec word2Vec1 = WordVectorSerializer.readWord2VecModel(model_v3, true);
+        Word2Vec word2Vec2 = WordVectorSerializer.readWord2VecModel(model_v4, true);
+        try {
+            assertEquals(word2Vec1.toJson(), word2Vec2.toJson());
+        } catch (Exception e) {
+            fail(e.getMessage());
+        }
+    }
+
+    @Test
+    public void testBackwardsCompatibleSequenceVectors() {
+        File model_v3 = Resources.asFile("deeplearning4j-nlp/seqv_beta3.csv");
+        File model_v4 = Resources.asFile("deeplearning4j-nlp/seqv_beta4.csv");
+        try {
+            SequenceVectors vectors1 = WordVectorSerializer.readSequenceVectors(new VocabWordFactory(), model_v3);
+            SequenceVectors vectors2 = WordVectorSerializer.readSequenceVectors(new VocabWordFactory(), model_v4);
+
+            assertEquals(vectors1.vocab().numWords(), vectors2.vocab().numWords());
+            for (int i = 0; i < vectors1.vocab().numWords(); ++i) {
+                assertEquals(vectors1.vocab().words().toArray()[i], vectors2.vocab().words().toArray()[i]);
+            }
+        } catch (Exception e) {
+            fail(e.getMessage());
+        }
+    }
+
 }

deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/iterator/BertIterator.java

@@ -249,7 +249,7 @@ public class BertIterator implements MultiDataSetIterator {
             } else {
                 throw new RuntimeException();
             }
-            l[0] = Nd4j.create(Nd4j.defaultFloatingPointType(), mbPadded, numClasses);
+            l[0] = Nd4j.create(DataType.FLOAT, mbPadded, numClasses);
             for( int i=0; i<mb; i++ ){
                 l[0].putScalar(i, classLabels[i], 1.0);
             }
@@ -277,9 +277,9 @@ public class BertIterator implements MultiDataSetIterator {
             if(unsupervisedLabelFormat == UnsupervisedLabelFormat.RANK2_IDX){
                 labelArr = Nd4j.create(DataType.INT, mbPadded, outLength);
             } else if(unsupervisedLabelFormat == UnsupervisedLabelFormat.RANK3_NCL){
-                labelArr = Nd4j.create(Nd4j.defaultFloatingPointType(), mbPadded, vocabSize, outLength);
+                labelArr = Nd4j.create(DataType.FLOAT, mbPadded, vocabSize, outLength);
             } else if(unsupervisedLabelFormat == UnsupervisedLabelFormat.RANK3_LNC){
-                labelArr = Nd4j.create(Nd4j.defaultFloatingPointType(), outLength, mbPadded, vocabSize);
+                labelArr = Nd4j.create(DataType.FLOAT, outLength, mbPadded, vocabSize);
             } else {
                 throw new IllegalStateException("Unknown unsupervised label format: " + unsupervisedLabelFormat);
             }

deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/iterator/CnnSentenceDataSetIterator.java

@@ -201,7 +201,7 @@ public class CnnSentenceDataSetIterator implements DataSetIterator {
         List<String> tokens = new ArrayList<>();
         while (t.hasMoreTokens()) {
             String token = t.nextToken();
-            if (!wordVectors.hasWord(token)) {
+            if (!wordVectors.outOfVocabularySupported() && !wordVectors.hasWord(token)) {
                 switch (unknownWordHandling) {
                     case RemoveWord:
                         continue;

deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/sequencevectors/SequenceVectors.java

@@ -1312,10 +1312,12 @@ public class SequenceVectors<T extends SequenceElement> extends WordVectorsImpl<
                             int rest = batchSequences.size() % batchSize;
                             int chunks = ((batchSequences.size() >= batchSize) ? batchSequences.size() / batchSize : 0) + ((rest > 0)? 1 : 0);
                             for (int j = 0; j < chunks; ++j) {
+                                if (trainElementsVectors) {
                                     if (elementsLearningAlgorithm instanceof SkipGram)
-                                    ((SkipGram)elementsLearningAlgorithm).iterateSample(batchSequences.get(j));
+                                        ((SkipGram) elementsLearningAlgorithm).iterateSample(batchSequences.get(j));
                                     else if (elementsLearningAlgorithm instanceof CBOW)
-                                    ((CBOW)elementsLearningAlgorithm).iterateSample(batchSequences.get(j));
+                                        ((CBOW) elementsLearningAlgorithm).iterateSample(batchSequences.get(j));
+                                }
 
                                 if (trainSequenceVectors) {
                                     if (sequenceLearningAlgorithm instanceof DBOW)

deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/main/java/org/deeplearning4j/models/word2vec/VocabWord.java

@@ -32,7 +32,7 @@ import java.io.Serializable;
  *
  * @author Adam Gibson
  */
-@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class")
+@JsonTypeInfo(use = JsonTypeInfo.Id.CLASS, include = JsonTypeInfo.As.PROPERTY, property = "@class", defaultImpl =  VocabWord.class)
 @JsonAutoDetect(fieldVisibility = JsonAutoDetect.Visibility.ANY, getterVisibility = JsonAutoDetect.Visibility.NONE,
         setterVisibility = JsonAutoDetect.Visibility.NONE)
 public class VocabWord extends SequenceElement implements Serializable {

deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp/src/test/java/org/deeplearning4j/iterator/TestBertIterator.java

@@ -224,6 +224,7 @@ public class TestBertIterator extends BaseDL4JTest {
 
     @Test(timeout = 20000L)
     public void testMinibatchPadding() throws Exception {
+        Nd4j.setDefaultDataTypes(DataType.FLOAT, DataType.FLOAT);
         String toTokenize1 = "I saw a girl with a telescope.";
         String toTokenize2 = "Donaudampfschifffahrts Kapitänsmützeninnenfuttersaum";
         BertWordPieceTokenizerFactory t = new BertWordPieceTokenizerFactory(pathToVocab, false, false, c);

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/api/TrainingConfig.java

@@ -17,6 +17,7 @@
 package org.deeplearning4j.nn.api;
 
 import org.deeplearning4j.nn.conf.GradientNormalization;
+import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.learning.config.IUpdater;
 import org.nd4j.linalg.learning.regularization.Regularization;
 
@@ -73,4 +74,6 @@ public interface TrainingConfig {
      */
     double getGradientNormalizationThreshold();
 
+    void setDataType(DataType dataType);
+
 }

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/graph/GraphVertex.java

@@ -93,4 +93,9 @@ public abstract class GraphVertex implements Cloneable, Serializable {
      */
     public abstract MemoryReport getMemoryReport(InputType... inputTypes);
 
+
+    public void setDataType(DataType dataType) {
+        //No-op for most layers
+    }
+
 }

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/graph/LayerVertex.java

@@ -146,4 +146,9 @@ public class LayerVertex extends GraphVertex {
         //TODO preprocessor memory
         return layerConf.getLayer().getMemoryReport(it);
     }
+
+    @Override
+    public void setDataType(DataType dataType){
+        layerConf.getLayer().setDataType(dataType);
+    }
 }

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/Layer.java

@@ -223,6 +223,11 @@ public abstract class Layer implements TrainingConfig, Serializable, Cloneable {
                         "Not supported: all layers with parameters should override this method");
     }
 
+    @Override
+    public void setDataType(DataType dataType) {
+        //No-op for most layers
+    }
+
     /**
      * This is a report of the estimated memory consumption for the given layer
      *

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/samediff/SameDiffLambdaVertex.java

@@ -96,7 +96,7 @@ public abstract class SameDiffLambdaVertex extends SameDiffVertex {
 
             if (!map.containsKey(inputNum)) {
                 //Lazily define extra input variable as required
-                SDVariable var = sameDiff.var("var_" + inputNum, 1); //TODO is this shape safe?
+                SDVariable var = sameDiff.var("var_" + inputNum, dataType, -1); //TODO is this shape safe?
                 map.put(inputNum, var);
             }
 

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/layers/samediff/SameDiffVertex.java

@@ -62,6 +62,7 @@ public abstract class SameDiffVertex extends GraphVertex implements TrainingConf
     protected IUpdater biasUpdater;
     protected GradientNormalization gradientNormalization;
     protected double gradientNormalizationThreshold = Double.NaN;
+    protected DataType dataType;
 
     /**
      * Define the vertex
@@ -234,4 +235,9 @@ public abstract class SameDiffVertex extends GraphVertex implements TrainingConf
     public double getGradientNormalizationThreshold() {
         return gradientNormalizationThreshold;
     }
+
+    @Override
+    public void setDataType(DataType dataType) {
+        this.dataType = dataType;
+    }
 }

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/conf/misc/DummyConfig.java

@@ -19,6 +19,7 @@ package org.deeplearning4j.nn.conf.misc;
 import lombok.AllArgsConstructor;
 import org.deeplearning4j.nn.api.TrainingConfig;
 import org.deeplearning4j.nn.conf.GradientNormalization;
+import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.learning.config.IUpdater;
 import org.nd4j.linalg.learning.config.NoOp;
 import org.nd4j.linalg.learning.regularization.Regularization;
@@ -63,4 +64,9 @@ public class DummyConfig implements TrainingConfig {
     public double getGradientNormalizationThreshold() {
         return 1.0;
     }
+
+    @Override
+    public void setDataType(DataType dataType) {
+
+    }
 }

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/ComputationGraph.java

@@ -512,6 +512,7 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
         for(; i<topologicalOrder.length; i++ ){
             String name = indices.getIdxToName().get(i);
             org.deeplearning4j.nn.conf.graph.GraphVertex n = configVertexMap.get(name);
+            n.setDataType(netDtype);
             numParamsForVertex[i] = n.numParams(true);
             numParams += numParamsForVertex[i];
         }

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/RnnLossLayer.java

@@ -26,6 +26,7 @@ import org.deeplearning4j.nn.gradient.DefaultGradient;
 import org.deeplearning4j.nn.gradient.Gradient;
 import org.deeplearning4j.nn.layers.BaseLayer;
 import org.deeplearning4j.util.TimeSeriesUtils;
+import org.nd4j.base.Preconditions;
 import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.api.DataSet;
@@ -35,6 +36,7 @@ import org.nd4j.linalg.primitives.Pair;
 import org.deeplearning4j.nn.workspace.ArrayType;
 import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
 
+import java.util.Arrays;
 import java.util.List;
 
 /**
@@ -60,10 +62,16 @@ public class RnnLossLayer extends BaseLayer<org.deeplearning4j.nn.conf.layers.Rn
         assertInputSet(true);
         if (input.rank() != 3)
             throw new UnsupportedOperationException(
-                            "Input is not rank 3. Got input with rank " + input.rank() + " " + layerId());
+                            "Input is not rank 3. Expected rank 3 input of shape [minibatch, size, sequenceLength]. Got input with rank " +
+                                    input.rank() + " with shape " + Arrays.toString(input.shape()) + " for layer " + layerId());
         if (labels == null)
             throw new IllegalStateException("Labels are not set (null)");
 
+        Preconditions.checkState(labels.rank() == 3, "Expected rank 3 labels array, got label array with shape %ndShape", labels);
+        Preconditions.checkState(input.size(2) == labels.size(2), "Sequence lengths do not match for RnnOutputLayer input and labels:" +
+                "Arrays should be rank 3 with shape [minibatch, size, sequenceLength] - mismatch on dimension 2 (sequence length) - input=%ndShape vs. label=%ndShape", input, labels);
+
+
         INDArray input2d = TimeSeriesUtils.reshape3dTo2d(input, workspaceMgr, ArrayType.BP_WORKING_MEM);
         INDArray labels2d = TimeSeriesUtils.reshape3dTo2d(labels, workspaceMgr, ArrayType.BP_WORKING_MEM);
         INDArray maskReshaped;

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/recurrent/RnnOutputLayer.java

@@ -23,6 +23,7 @@ import org.deeplearning4j.nn.gradient.Gradient;
 import org.deeplearning4j.nn.layers.BaseOutputLayer;
 import org.deeplearning4j.nn.params.DefaultParamInitializer;
 import org.deeplearning4j.util.TimeSeriesUtils;
+import org.nd4j.base.Preconditions;
 import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.factory.Nd4j;
@@ -57,8 +58,13 @@ public class RnnOutputLayer extends BaseOutputLayer<org.deeplearning4j.nn.conf.l
                     "Input is not rank 3. RnnOutputLayer expects rank 3 input with shape [minibatch, layerInSize, sequenceLength]." +
                             " Got input with rank " + input.rank() + " and shape " + Arrays.toString(input.shape()) + " - " + layerId());
         }
+        Preconditions.checkState(labels.rank() == 3, "Expected rank 3 labels array, got label array with shape %ndShape", labels);
+        Preconditions.checkState(input.size(2) == labels.size(2), "Sequence lengths do not match for RnnOutputLayer input and labels:" +
+                "Arrays should be rank 3 with shape [minibatch, size, sequenceLength] - mismatch on dimension 2 (sequence length) - input=%ndShape vs. label=%ndShape", input, labels);
+
         INDArray inputTemp = input;
         this.input = TimeSeriesUtils.reshape3dTo2d(input, workspaceMgr, ArrayType.BP_WORKING_MEM);
+
         Pair<Gradient, INDArray> gradAndEpsilonNext = super.backpropGradient(epsilon, workspaceMgr);    //Also applies dropout
         this.input = inputTemp;
         INDArray epsilon2d = gradAndEpsilonNext.getSecond();

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/samediff/SameDiffGraphVertex.java

@@ -39,9 +39,7 @@ import org.nd4j.linalg.api.ops.impl.layers.ExternalErrorsFunction;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.primitives.Pair;
 
-import java.util.Arrays;
+import java.util.*;
-import java.util.LinkedHashMap;
-import java.util.Map;
 
 /**
  * Implementation of a SameDiff graph vertex.
@@ -96,12 +94,11 @@ public class SameDiffGraphVertex extends BaseGraphVertex {
 
     @Override
     public INDArray doForward(boolean training, LayerWorkspaceMgr workspaceMgr) {
+        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
             if(sameDiff == null){
                 doInit();
             }
 
-        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
-//            sameDiff.clearExecutionCache();
             config.validateInput(inputs);
             for(int i=0; i<inputs.length; i++ ){
                 String name = config.getVertexParams().getInputs().get(i);
@@ -121,6 +118,10 @@ public class SameDiffGraphVertex extends BaseGraphVertex {
             }
             Map<String,INDArray> out = sameDiff.exec(null, outputKey);
             INDArray result = out.get(outputKey);
+
+            //Clear placeholders and op inputs to ensure no out-of-scope arrays are still referenced anywhere
+            sameDiff.clearPlaceholders(true);
+            sameDiff.clearOpInputs();
             return workspaceMgr.dup(ArrayType.ACTIVATIONS, result);
         }
     }
@@ -131,27 +132,42 @@ public class SameDiffGraphVertex extends BaseGraphVertex {
 
         INDArray[] dLdIns;
         try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()){
-//            sameDiff.clearExecutionCache();
+            if(sameDiff == null){
+                doInit();
+            }
+
+            if(!sameDiff.hasGradientFunction()) {
+                //Create when scoped out, to ensure any arrays are not in WS
+                List<String> inputs = config.getVertexParams().getInputs();
+                String[] inArr = inputs.toArray(new String[inputs.size()]);
+                sameDiff.createGradFunction(inArr);
+            }
             config.validateInput(inputs);
-            //Set inputs
+            Map<String,INDArray> phMap = new HashMap<>();
-            for(int i=0; i<inputs.length; i++ ){
+            List<String> inputs = config.getVertexParams().getInputs();
-                String name = config.getVertexParams().getInputs().get(i);
+            int i=0;
+            for(String s : inputs){
+                phMap.put(s, this.inputs[i++]);
+            }
+            if(maskArrays != null){
+                for( int j=0; j<maskArrays.length; j++ ){
+                    String name = inputs.get(j);
                     final String maskName = name + "_mask";
-                sameDiff.associateArrayWithVariable(inputs[i].dup(), sameDiff.getVariable(name));
+                    if(maskArrays[j] != null) {
-                if(maskArrays != null && maskArrays[i] != null) {
+                        sameDiff.associateArrayWithVariable(maskArrays[j].dup(), maskName);
-                    sameDiff.associateArrayWithVariable(maskArrays[i].dup(), maskName);
-                }else{
-                    sameDiff.associateArrayWithVariable(createMask(dataType, inputs[i].shape()), maskName);
                     }
                 }
-            fn.updateVariable(outputVar.getVarName(), epsilon.dup());
+            }
+            String epsName = fn.getGradPlaceholderName();
+            phMap.put(epsName, epsilon);
+
 
             for(String s : paramTable.keySet() ){
                 //TODO this should only be necessary, in theory, once!
                 sameDiff.associateArrayWithVariable(paramTable.get(s), s);
             }
 
-            sameDiff.execBackwards(null);
+            sameDiff.execBackwards(phMap);
             for(String s : paramTable.keySet() ){
                 INDArray sdGrad = sameDiff.grad(s).getArr();
                 INDArray dl4jGrad = gradTable.get(s);
@@ -159,10 +175,10 @@ public class SameDiffGraphVertex extends BaseGraphVertex {
                 g.gradientForVariable().put(s, dl4jGrad);
             }
 
-            dLdIns = new INDArray[inputs.length];
+            dLdIns = new INDArray[inputs.size()];
-            for(int i=0; i<inputs.length; i++ ){
+            for(int j=0; j<inputs.size(); j++ ){
-                String name = config.getVertexParams().getInputs().get(i);
+                String name = inputs.get(j);
-                dLdIns[i] = sameDiff.grad(name).getArr();
+                dLdIns[j] = sameDiff.grad(name).getArr();
             }
         }
 
@@ -171,6 +187,9 @@ public class SameDiffGraphVertex extends BaseGraphVertex {
             dLdIns[i] = workspaceMgr.dup(ArrayType.ACTIVATION_GRAD, dLdIns[i]);
         }
 
+        //Clear placeholders and op inputs to ensure no out-of-scope arrays are still referenced anywhere
+        sameDiff.clearPlaceholders(true);
+        sameDiff.clearOpInputs();
         return new Pair<>(g, dLdIns);
     }
 

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/samediff/SameDiffLayer.java

@@ -35,6 +35,7 @@ import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.primitives.Pair;
 import org.deeplearning4j.nn.workspace.ArrayType;
 import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
+import org.nd4j.linalg.util.ArrayUtil;
 
 import java.util.*;
 
@@ -78,25 +79,32 @@ public class SameDiffLayer extends AbstractLayer<AbstractSameDiffLayer> {
     @Override
     public INDArray activate(boolean training, LayerWorkspaceMgr workspaceMgr) {
         assertInputSet(false);
+
+        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
             if(sameDiff == null){
                 doInit();
             }
 
-        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
             org.deeplearning4j.nn.conf.layers.samediff.SameDiffLayer bl = (org.deeplearning4j.nn.conf.layers.samediff.SameDiffLayer) layerConf();
             bl.validateInput(input);
-            sameDiff.associateArrayWithVariable(input.dup(), sameDiff.getVariable(INPUT_KEY));
+
+            Map<String,INDArray> phMap = new HashMap<>();
+            phMap.put(INPUT_KEY, input);
             if(maskArray != null){
-                sameDiff.associateArrayWithVariable(maskArray, sameDiff.getVariable(MASK_KEY));
+                phMap.put(MASK_KEY, maskArray);
-            }else{
-                sameDiff.associateArrayWithVariable(SameDiffGraphVertex.createMask(dataType, input.shape()), sameDiff.getVariable(MASK_KEY));
             }
+
             for(String s : paramTable.keySet() ) {
                 sameDiff.associateArrayWithVariable(paramTable.get(s), s);
             }
 
-            Map<String,INDArray> out = sameDiff.exec(null, outputKey);
+            Map<String,INDArray> out = sameDiff.exec(phMap, outputKey);
             INDArray result = out.get(outputKey);
+
+            //Clear placeholders and op inputs to ensure no out-of-scope arrays are still referenced anywhere
+            sameDiff.clearPlaceholders(true);
+            sameDiff.clearOpInputs();
+
             return workspaceMgr.dup(ArrayType.ACTIVATIONS, result);
         }
     }
@@ -110,24 +118,36 @@ public class SameDiffLayer extends AbstractLayer<AbstractSameDiffLayer> {
 
         INDArray dLdIn;
         try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()){
-//            sameDiff.clearExecutionCache();
+            if(sameDiff == null){
+                doInit();
+            }
+            if(!sameDiff.hasGradientFunction()) {
+                //Create when scoped out, to ensure any arrays are not in WS
+                sameDiff.createGradFunction(INPUT_KEY);
+            }
+
             org.deeplearning4j.nn.conf.layers.samediff.SameDiffLayer bl = (org.deeplearning4j.nn.conf.layers.samediff.SameDiffLayer) layerConf();
             bl.validateInput(input);
 
-            sameDiff.associateArrayWithVariable(input.dup(), sameDiff.getVariable(INPUT_KEY));
-            if(maskArray != null){
-                sameDiff.associateArrayWithVariable(maskArray, sameDiff.getVariable(MASK_KEY));
-            }else{
-                sameDiff.associateArrayWithVariable(SameDiffGraphVertex.createMask(dataType, input.shape()), sameDiff.getVariable(MASK_KEY));
-            }
-            fn.updateVariable(outputVar.getVarName(), epsilon.dup());
-
             for(String s : paramTable.keySet() ){
                 //TODO this should only be necessary, in theory, once!
                 sameDiff.associateArrayWithVariable(paramTable.get(s), s);
             }
 
-            sameDiff.execBackwards(Collections.<String, INDArray>emptyMap());
+            Map<String,INDArray> phMap = new HashMap<>();
+            phMap.put(INPUT_KEY, input);
+            phMap.put(fn.getGradPlaceholderName(), epsilon);
+            if(maskArray != null){
+                phMap.put(MASK_KEY, maskArray);
+            }
+
+            List<String> requiredGrads = new ArrayList<>(paramTable.size() + 1);
+            requiredGrads.add(sameDiff.grad(INPUT_KEY).getVarName());
+            for(String s : paramTable.keySet()){
+                requiredGrads.add(sameDiff.grad(s).getVarName());
+            }
+
+            sameDiff.execBackwards(phMap, requiredGrads);
             for(String s : paramTable.keySet() ){
                 INDArray sdGrad = sameDiff.grad(s).getArr();
                 INDArray dl4jGrad = gradTable.get(s);
@@ -138,6 +158,11 @@ public class SameDiffLayer extends AbstractLayer<AbstractSameDiffLayer> {
             dLdIn = sameDiff.grad(INPUT_KEY).getArr();
         }
 
+        //Clear placeholders and op inputs to ensure no out-of-scope arrays are still referenced anywhere
+        sameDiff.clearPlaceholders(true);
+        sameDiff.clearOpInputs();
+
+        System.out.println(dLdIn);
         return new Pair<>(g, workspaceMgr.dup(ArrayType.ACTIVATION_GRAD, dLdIn));   //TODO OPTIMIZE THIS
     }
 
@@ -225,8 +250,9 @@ public class SameDiffLayer extends AbstractLayer<AbstractSameDiffLayer> {
             sameDiff = SameDiff.create();
             Map<String, INDArray> p = paramTable();
 
-            val inputShape = input.shape().clone();
+            long[] inputShape = input.shape().clone();
-            SDVariable inputVar = sameDiff.var(INPUT_KEY, dataType, inputShape);
+            inputShape[0] = -1;
+            SDVariable inputVar = sameDiff.placeHolder(INPUT_KEY, dataType, inputShape);
             Map<String, long[]> paramShapes = layerConf().getLayerParams().getParamShapes();
             Map<String, SDVariable> params = new LinkedHashMap<>();
             for (String s : paramShapes.keySet()) {
@@ -235,7 +261,8 @@ public class SameDiffLayer extends AbstractLayer<AbstractSameDiffLayer> {
                 params.put(s, v);
             }
 
-            SDVariable mask = sameDiff.constant(MASK_KEY, SameDiffGraphVertex.createMask(dataType, inputShape));
+            long[] maskShape = ArrayUtil.nTimes((long)inputShape.length, -1);
+            SDVariable mask = sameDiff.placeHolder(MASK_KEY, dataType, maskShape);
 
             SDVariable layerOutput = bl.defineLayer(sameDiff, inputVar, params, mask);
             Preconditions.checkNotNull(layerOutput, "Invalid output: layer output is null");

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/samediff/SameDiffOutputLayer.java

@@ -87,35 +87,43 @@ public class SameDiffOutputLayer extends AbstractLayer<org.deeplearning4j.nn.con
     private INDArray activateHelper(boolean activations, LayerWorkspaceMgr workspaceMgr){
         assertInputSet(false);
 
-        //Check where the output occors. If it's a simple loss layer (no params) this could
+        //Check where the output occurs. If it's a simple loss layer (no params) this could
         // just be the input!
         if(activations && INPUT_KEY.equals(layerConf().activationsVertexName())){
             return workspaceMgr.leverageTo(ArrayType.ACTIVATIONS, input);
         }
 
+        //TODO optimize
+        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
             if(sameDiff == null){
                 doInit();
             }
 
-        //TODO optimize
-        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
-            sameDiff.associateArrayWithVariable(input.dup(), sameDiff.getVariable(INPUT_KEY));
-            if(layerConf().labelsRequired() && labels != null) {
-                sameDiff.associateArrayWithVariable(labels.dup(), sameDiff.getVariable(LABELS_KEY));
-            }
             for(String s : paramTable.keySet() ) {
                 sameDiff.associateArrayWithVariable(paramTable.get(s), s);
             }
 
-            INDArray score = sameDiff.execAndEndResult();
+            Map<String,INDArray> phMap = new HashMap<>();
+            phMap.put(INPUT_KEY, input);
+            if(!activations && layerConf().labelsRequired() && labels != null) {
+                phMap.put(LABELS_KEY, labels);
+            }
+
+            String s = activations ? layerConf().activationsVertexName() : outputVar.getVarName();
+
+            INDArray out = sameDiff.execSingle(phMap, s);
+
+            //Clear placeholders and op inputs to ensure no out-of-scope arrays are still referenced anywhere
+            sameDiff.clearPlaceholders(true);
+            sameDiff.clearOpInputs();
+
             if(activations) {
-                INDArray result = sameDiff.getArrForVarName(layerConf().activationsVertexName());
+                Preconditions.checkNotNull(out, "Activations (result) array for variable \"%s\" was " +
-                Preconditions.checkNotNull(result, "Activations (result) array for variable \"%s\" was " +
                         "null - error during execution or this variable (as defined by method activationsVertexName()) " +
                         "does not exist", layerConf().activationsVertexName());
-                return workspaceMgr.dup(ArrayType.ACTIVATIONS, result);
+                return workspaceMgr.dup(ArrayType.ACTIVATIONS, out);
             } else {
-                return score;
+                return out;
             }
         }
     }
@@ -127,23 +135,26 @@ public class SameDiffOutputLayer extends AbstractLayer<org.deeplearning4j.nn.con
         Preconditions.checkState(!layerConf().labelsRequired() || labels != null, "Cannot execute backprop: Labels are not set. " +
                 "If labels are not required for this SameDiff output layer, override SameDiffOutputLayer.labelsRequired()" +
                 " to return false instead");
+        Gradient g = new DefaultGradient();
 
+        INDArray dLdIn;
+        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()){
             if(sameDiff == null){
                 //Usually doInit will be called in forward pass; not necessarily the case in output layers
                 // (for efficiency, we skip output layer forward pass in MultiLayerNetwork/ComputationGraph)
                 doInit();
             }
+            if(!sameDiff.hasGradientFunction()) {
+                //Create when scoped out, to ensure any arrays are not in WS
+                sameDiff.createGradFunction(INPUT_KEY);
+            }
 
-        Gradient g = new DefaultGradient();
+            INDArray castInput = input.castTo(dataType);
-
-        INDArray dLdIn;
-        try(MemoryWorkspace ws = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()){
-            INDArray castInput = input.castTo(Nd4j.defaultFloatingPointType());
             if(castInput.isAttached())
                 castInput = castInput.dup();
             sameDiff.associateArrayWithVariable(castInput, sameDiff.getVariable(INPUT_KEY));
             if(layerConf().labelsRequired()) {
-                INDArray castLabels = labels.castTo(Nd4j.defaultFloatingPointType());
+                INDArray castLabels = labels.castTo(dataType);
                 if(castLabels.isAttached())
                     castLabels = castLabels.dup();
                 sameDiff.associateArrayWithVariable(castLabels, sameDiff.getVariable(LABELS_KEY));
@@ -154,7 +165,17 @@ public class SameDiffOutputLayer extends AbstractLayer<org.deeplearning4j.nn.con
                 sameDiff.associateArrayWithVariable(paramTable.get(s), s);
             }
 
-            sameDiff.execBackwards(Collections.<String, INDArray>emptyMap());
+            List<String> gradVarNames = new ArrayList<>();
+            for(String s : paramTable.keySet()){
+                gradVarNames.add(sameDiff.getVariable(s).getGradient().getVarName());
+            }
+            gradVarNames.add(sameDiff.grad(INPUT_KEY).getVarName());
+
+            Map<String,INDArray> phMap = new HashMap<>();
+            phMap.put(INPUT_KEY, input);
+            phMap.put(LABELS_KEY, labels);
+
+            sameDiff.execBackwards(phMap, gradVarNames);
             for(String s : paramTable.keySet() ){
                 INDArray sdGrad = sameDiff.grad(s).getArr();
                 INDArray dl4jGrad = gradTable.get(s);
@@ -165,6 +186,10 @@ public class SameDiffOutputLayer extends AbstractLayer<org.deeplearning4j.nn.con
             dLdIn = sameDiff.grad(INPUT_KEY).getArr();
         }
 
+        //Clear placeholders and op inputs to ensure no out-of-scope arrays are still referenced anywhere
+        sameDiff.clearPlaceholders(true);
+        sameDiff.clearOpInputs();
+
         return new Pair<>(g, workspaceMgr.dup(ArrayType.ACTIVATION_GRAD, dLdIn));   //TODO OPTIMIZE THIS
     }
 
@@ -252,18 +277,20 @@ public class SameDiffOutputLayer extends AbstractLayer<org.deeplearning4j.nn.con
             sameDiff = SameDiff.create();
             Map<String, INDArray> p = paramTable();
 
-            val inputShape = input.shape().clone();
+            long[] inputShape = input.shape().clone();
-            SDVariable inputVar = sameDiff.var(INPUT_KEY, dataType, inputShape);
+            inputShape[0] = -1;
+            SDVariable inputVar = sameDiff.placeHolder(INPUT_KEY, dataType, inputShape);
             SDVariable labelVar = null;
             if(layerConf().labelsRequired()){
-                long[] labelShape = labels == null ? new long[]{1} : labels.shape().clone();
+                long[] labelShape = labels == null ? new long[]{-1, -1} : labels.shape().clone();
-                labelVar = sameDiff.var(LABELS_KEY, dataType, labelShape);
+                labelShape[0] = -1;
+                labelVar = sameDiff.placeHolder(LABELS_KEY, dataType, labelShape);
             }
             Map<String, long[]> paramShapes = layerConf().getLayerParams().getParamShapes();
             Map<String, SDVariable> params = new LinkedHashMap<>();
             for (String s : paramShapes.keySet()) {
                 val ps = paramShapes.get(s);
-                SDVariable v = sameDiff.var(s, ps);
+                SDVariable v = sameDiff.var(s, dataType, ps);
                 params.put(s, v);
             }
             SDVariable layerOutput = bl.defineLayer(sameDiff, inputVar, labelVar, params);

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/multilayer/MultiLayerNetwork.java

@@ -660,6 +660,7 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
             val nParamsPerLayer = new long[nLayers];
             for (int i = 0; i < nLayers; i++) {
                 NeuralNetConfiguration conf = layerWiseConfigurations.getConf(i);
+                conf.getLayer().setDataType(netDtype);
                 nParamsPerLayer[i] = conf.getLayer().initializer().numParams(conf);
                 paramLength += nParamsPerLayer[i];
             }

deeplearning4j/dl4j-perf/src/main/java/org/deeplearning4j/perf/listener/HardwareMetric.java

@@ -152,7 +152,7 @@ public class HardwareMetric implements Serializable {
         return builder.logicalProcessorCount(processor.getLogicalProcessorCount())
                 .physicalProcessorCount(processor.getPhysicalProcessorCount())
                 .name(name)
-                .averagedCpuLoad((long) processor.getSystemCpuLoad() * 100)
+                .averagedCpuLoad((long)(processor.getSystemCpuLoad() * 100))
                 .ioWaitTime(iowait).gpuMetrics(gpuMetric)
                 .hostName(networkParams.getHostName()).diskInfo(diskInfoMap)
                 .currentMemoryUse(globalMemory.getTotal() - globalMemory.getAvailable())

libnd4j/blas/CMakeLists.txt

@@ -48,8 +48,6 @@ if(WIN32)
     SET(CMAKE_NINJA_FORCE_RESPONSE_FILE 1 CACHE INTERNAL "")
 endif()
 
-
-
 if ("${LIBND4J_ALL_OPS}")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DLIBND4J_ALL_OPS=true")
 else()
@@ -234,21 +232,21 @@ if(CUDA_BLAS)
             endif()
         endif()
 
-        if (NOT BUILD_TESTS)
         file(GLOB_RECURSE EXCEPTIONS_SOURCES false ../include/exceptions/*.cpp ../include/exceptions/*.h)
-            file(GLOB_RECURSE EXEC_SOURCES false ../include/execution/*.cpp ../include/execution/*.h)
+        file(GLOB_RECURSE EXEC_SOURCES false ../include/execution/impl/*.cpp ../include/execution/*.cu ../include/execution/*.h)
         file(GLOB_RECURSE TYPES_SOURCES false ../include/types/*.cpp ../include/types/*.h)
         file(GLOB_RECURSE ARRAY_SOURCES false ../include/array/impl/*.cpp ../include/array/cuda/*.cu ../include/array/*.h)
-            file(GLOB_RECURSE MEMORY_SOURCES false ../include/memory/*.cpp ../include/memory/*.h)
+        file(GLOB_RECURSE MEMORY_SOURCES false ../include/memory/impl/*.cpp ../include/memory/cuda/*.cu ../include/memory/*.h)
-            file(GLOB_RECURSE GRAPH_SOURCES false ../include/graph/*.cpp ../include/graph/*.h)
+        file(GLOB_RECURSE GRAPH_SOURCES false ../include/graph/*.cpp ../include/graph/*.cu ../include/graph/*.h)
         file(GLOB_RECURSE CUSTOMOPS_SOURCES false ../include/ops/declarable/generic/*.cpp)
-            file(GLOB_RECURSE CUSTOMOPS_HELPERS_SOURCES false ../include/ops/declarable/helpers/cpu/*.cpp)
+        file(GLOB_RECURSE CUSTOMOPS_HELPERS_SOURCES false ../include/ops/declarable/helpers/cuda/*.cu ../include/ops/declarable/helpers/impl/*.cpp)
         file(GLOB_RECURSE OPS_SOURCES false ../include/ops/impl/*.cpp ../include/ops/declarable/impl/*.cpp  ../include/ops/*.h)
+        file(GLOB_RECURSE HELPERS_SOURCES false ../include/helpers/impl/*.cpp ../include/helpers/*.cu ../include/helpers/*.cupp ../include/helpers/*.h)
         file(GLOB_RECURSE INDEXING_SOURCES false ../include/indexing/*.cpp ../include/indexing/*.h)
-            file(GLOB_RECURSE HELPERS_SOURCES false ../include/helpers/*.cpp ../include/helpers/cuda/*.cu ../include/helpers/*.h)
         file(GLOB_RECURSE LOOPS_SOURCES false ../include/loops/*.cpp ../include/loops/*.h)
         file(GLOB_RECURSE LOOPS_SOURCES_CUDA false ../include/loops/*.cu)
 
+        if (NOT BUILD_TESTS)
 			CUDA_ADD_LIBRARY(${LIBND4J_NAME} SHARED cuda/NativeOps.cu cuda/NativeOpExecutioner.cu ${LOOPS_SOURCES_CUDA}
                 ${CUSTOMOPS_HELPERS_SOURCES} ${HELPERS_SOURCES} ${EXEC_SOURCES}
                 ../include/cnpy/cnpy.cpp ../include/nd4jmemset.h ../include/nd4jmalloc.h
@@ -258,20 +256,6 @@ if(CUDA_BLAS)
 		else()
             set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DBUILD_TESTS=true")
 
-            file(GLOB_RECURSE EXCEPTIONS_SOURCES false ../include/exceptions/*.cpp ../include/exceptions/*.h)
-            file(GLOB_RECURSE EXEC_SOURCES false ../include/execution/impl/*.cpp ../include/execution/*.cu ../include/execution/*.h)
-            file(GLOB_RECURSE TYPES_SOURCES false ../include/types/*.cpp ../include/types/*.h)
-            file(GLOB_RECURSE ARRAY_SOURCES false ../include/array/impl/*.cpp ../include/array/cuda/*.cu ../include/array/*.h)
-            file(GLOB_RECURSE MEMORY_SOURCES false ../include/memory/impl/*.cpp ../include/memory/cuda/*.cu ../include/memory/*.h)
-            file(GLOB_RECURSE GRAPH_SOURCES false ../include/graph/*.cpp ../include/graph/*.cu ../include/graph/*.h)
-            file(GLOB_RECURSE CUSTOMOPS_SOURCES false ../include/ops/declarable/generic/*.cpp)
-            file(GLOB_RECURSE CUSTOMOPS_HELPERS_SOURCES false ../include/ops/declarable/helpers/cuda/*.cu)
-            file(GLOB_RECURSE OPS_SOURCES false ../include/ops/impl/*.cpp ../include/ops/declarable/impl/*.cpp  ../include/ops/*.h)
-            file(GLOB_RECURSE HELPERS_SOURCES false ../include/helpers/impl/*.cpp ../include/helpers/*.cu ../include/helpers/*.cupp ../include/helpers/*.h)
-            file(GLOB_RECURSE INDEXING_SOURCES false ../include/indexing/*.cpp ../include/indexing/*.h)
-            file(GLOB_RECURSE LOOPS_SOURCES false ../include/loops/*.cpp ../include/loops/*.h)
-            file(GLOB_RECURSE LOOPS_SOURCES_CUDA false ../include/loops/*.cu)
-
 			CUDA_ADD_LIBRARY(${LIBND4J_NAME} STATIC cuda/NativeOps.cu cuda/NativeOpExecutioner.cu ${LOOPS_SOURCES_CUDA}
                 ${CUSTOMOPS_HELPERS_SOURCES} ${HELPERS_SOURCES} ${EXEC_SOURCES}
                 ../include/cnpy/cnpy.cpp ../include/nd4jmemset.h ../include/nd4jmalloc.h
@@ -308,7 +292,7 @@ elseif(CPU_BLAS)
     file(GLOB_RECURSE MEMORY_SOURCES false ../include/memory/*.cpp ../include/memory/*.h)
     file(GLOB_RECURSE GRAPH_SOURCES false ../include/graph/*.cpp ../include/graph/*.h)
     file(GLOB_RECURSE CUSTOMOPS_SOURCES false ../include/ops/declarable/generic/*.cpp)
-    file(GLOB_RECURSE CUSTOMOPS_HELPERS_SOURCES false ../include/ops/declarable/helpers/cpu/*.cpp)
+    file(GLOB_RECURSE CUSTOMOPS_HELPERS_SOURCES false ../include/ops/declarable/helpers/cpu/*.cpp ../include/ops/declarable/helpers/impl/*.cpp)
     file(GLOB_RECURSE OPS_SOURCES false ../include/ops/impl/*.cpp ../include/ops/declarable/impl/*.cpp  ../include/ops/*.h)
     file(GLOB_RECURSE INDEXING_SOURCES false ../include/indexing/*.cpp ../include/indexing/*.h)
     file(GLOB_RECURSE HELPERS_SOURCES false ../include/helpers/*.cpp ../include/helpers/*.h)

libnd4j/blas/NDArray.h

@@ -372,8 +372,8 @@ namespace nd4j {
         /**
         *  if _bufferD==nullptr return _buffer, else return _bufferD
         */
-        FORCEINLINE void* specialBuffer();
+        void* specialBuffer();
-        FORCEINLINE void* getSpecialBuffer() const;
+        void* getSpecialBuffer() const;
 
         /**
         *   returns device buffer if compilation is for cuda case, otherwise returns host buffer
@@ -429,16 +429,16 @@ namespace nd4j {
         /**
         *  permutes the dimensions in array according to "dimensions" array, new array points on _buffer of this array
         */
-		NDArray* permute(const std::initializer_list<int>& dimensions) const;
+		NDArray permute(const std::initializer_list<int>& dimensions) const;
-        NDArray* permute(const std::vector<int>& dimensions) const;
+        NDArray permute(const std::vector<int>& dimensions) const;
-        NDArray* permute(const int* dimensions, const int rank) const;
+        NDArray permute(const int* dimensions, const int rank) const;
 
         void permute(const int* dimensions, const int rank, NDArray& target) const;
         void permute(const std::vector<int>& dimensions, NDArray& target) const;
 
-        NDArray* permute(const std::initializer_list<Nd4jLong>& dimensions) const;
+        NDArray permute(const std::initializer_list<Nd4jLong>& dimensions) const;
-        NDArray* permute(const std::vector<Nd4jLong>& dimensions) const;
+        NDArray permute(const std::vector<Nd4jLong>& dimensions) const;
-        NDArray* permute(const Nd4jLong* dimensions, const int rank) const;
+        NDArray permute(const Nd4jLong* dimensions, const int rank) const;
 
         void permute(const Nd4jLong* dimensions, const int rank, NDArray& target) const;
         void permute(const std::vector<Nd4jLong>& dimensions, NDArray& target) const;
@@ -508,7 +508,7 @@ namespace nd4j {
         /**
         *  returns new copy of this array, optionally in different order
         */
-        NDArray *dup(const char newOrder = 'a');
+        NDArray *dup(const char newOrder = 'a') const;
 
         /**
         *  returns sum of all elements of array
@@ -687,7 +687,7 @@ namespace nd4j {
         void applyScalarArr(nd4j::scalar::BoolOps op, const NDArray* scalar, NDArray* target, ExtraArguments *extraParams = nullptr) const;
 
 
-#if defined(__CUDABLAS__) && defined(BUILD_TESTS)
+#if defined(__CUDABLAS__) //&& defined(BUILD_TESTS)
         template <typename Lambda>
         FORCEINLINE void applyLambda(Lambda func, NDArray* target = nullptr);
 
@@ -790,8 +790,7 @@ namespace nd4j {
         /**
         *   apply transpose operation to the copy of this array, that is this array remains unaffected
         */
-        NDArray* transpose() const;
+        NDArray transpose() const;
-        NDArray  transp() const;
 
         /**
         *  perform transpose operation and store result in target, this array remains unaffected
@@ -915,7 +914,7 @@ namespace nd4j {
         *
         * if permute have been applied before or there are weird strides, then new buffer is allocated for new array
         */
-		NDArray* reshape(const char order, const std::vector<Nd4jLong>& shape) const;
+		NDArray reshape(const char order, const std::vector<Nd4jLong>& shape) const;
 
         /**
         *  calculate strides and set given order
@@ -2093,15 +2092,6 @@ Nd4jLong* NDArray::shapeInfo() {
     return _shapeInfo;
 }
 
-////////////////////////////////////////////////////////////////////////
-void* NDArray::specialBuffer() {
-
-    if (_buffer->special() == nullptr)
-        return getBuffer();
-    // FIXME: this should be fixed once CUDA backend added
-    return static_cast<int8_t*>(_buffer->special()) + (_offset * sizeOfT());
-}
-
 ////////////////////////////////////////////////////////////////////////
 Nd4jLong* NDArray::specialShapeInfo() {
     if (_shapeInfoD == nullptr)
@@ -2110,14 +2100,6 @@ Nd4jLong* NDArray::specialShapeInfo() {
     return _shapeInfoD;
 }
 
-////////////////////////////////////////////////////////////////////////
-void* NDArray::getSpecialBuffer() const {
-      if (_buffer->special() == nullptr)
-        return getBuffer();
-    // FIXME: this should be fixed once CUDA backend added
-    return static_cast<int8_t*>(_buffer->special()) + (_offset * sizeOfT());
-}
-
 ////////////////////////////////////////////////////////////////////////
 Nd4jLong NDArray::getBufferOffset() const {
     return _offset;
@@ -2137,7 +2119,7 @@ Nd4jLong* NDArray::getSpecialShapeInfo() const{
 }
 
 
-#if defined(__CUDACC__) && defined(BUILD_TESTS)
+#if defined(__CUDACC__) //&& defined(BUILD_TESTS)
 // for CUDA we need stil stuff inline
 #include "cuda/NDArrayLambda.hpp"
 #endif

libnd4j/blas/NDArray.hpp

@@ -39,9 +39,9 @@ NDArray* NDArray::asT() const{
     auto result = isScalar() ? new NDArray('c', {}, {0.}, DataTypeUtils::fromT<T>(), this->getContext()) : new NDArray(ordering(), getShapeAsVector(), DataTypeUtils::fromT<T>(), this->getContext());
     auto l = this->lengthOf();
 
-    prepareSpecialUse({result}, {this});
+    NDArray::prepareSpecialUse({result}, {this});
     NativeOpExecutioner::execTransformAny(getContext(), transform::AnyOps::Assign, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), result->getBuffer(), result->getShapeInfo(), result->getSpecialBuffer(), result->getSpecialShapeInfo(), nullptr, nullptr, nullptr);
-    registerSpecialUse({result}, {this});
+    NDArray::registerSpecialUse({result}, {this});
 
     return result;
 }
@@ -583,117 +583,130 @@ void NDArray::copyBuffersContinuouslyFrom(const NDArray& other, size_t sizeToCop
 void NDArray::assign(const double value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(dataType(), value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const float value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(dataType(), value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const float16 value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const bfloat16& value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const Nd4jLong value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const int value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(this->dataType(), value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), _shapeInfo, specialBuffer(), _shapeInfoD, buffer(), _shapeInfo, specialBuffer(), _shapeInfoD, temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp._shapeInfoD, nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const int16_t value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(this->dataType(), value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), _shapeInfo, specialBuffer(), _shapeInfoD, buffer(), _shapeInfo, specialBuffer(), _shapeInfoD, temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp._shapeInfoD, nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const uint8_t value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(this->dataType(), value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const uint16_t value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(this->dataType(), value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const uint32_t value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(this->dataType(), value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const uint64_t value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(this->dataType(), value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const int8_t value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(this->dataType(), value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::assign(const bool value) {
     // just fire scalar
     auto temp = NDArrayFactory::create(this->dataType(), value, this->getContext());
-    prepareSpecialUse({this}, {&temp});
+
+    NDArray::prepareSpecialUse({this}, {&temp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::CopyPws, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), temp.buffer(), temp.shapeInfo(), temp.specialBuffer(), temp.getSpecialShapeInfo(), nullptr);
-    registerSpecialUse({this}, {&temp});
+    NDArray::registerSpecialUse({this}, {&temp});
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -716,9 +729,9 @@ NDArray NDArray::varianceNumber(nd4j::variance::Ops op, bool biasCorrected) {
 
     NDArray res(DataTypeUtils::pickFloatingType(dataType()), getContext());
 
-    prepareSpecialUse({&res}, {this});
+    NDArray::prepareSpecialUse({&res}, {this});
     NativeOpExecutioner::execSummaryStatsScalar(getContext(), op, buffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), nullptr, res.buffer(), res.shapeInfo(), res.specialBuffer(), res.specialShapeInfo(), biasCorrected);
-    registerSpecialUse({&res}, {this});
+    NDArray::registerSpecialUse({&res}, {this});
 
     return res;
 }
@@ -918,9 +931,9 @@ NDArray NDArray::reduceNumber(nd4j::reduce::FloatOps op, void *extraParams) cons
     auto shape = ConstantShapeHelper::getInstance()->scalarShapeInfo(DataTypeUtils::pickFloatingType(dataType()));
     NDArray result(shape, true, this->getContext());
 
-    prepareSpecialUse({&result}, {this});
+    NDArray::prepareSpecialUse({&result}, {this});
     NativeOpExecutioner::execReduceFloatScalar(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), extraParams, result.buffer(), result.shapeInfo(), result.specialBuffer(), result.specialShapeInfo());
-    registerSpecialUse({&result}, {this});
+    NDArray::registerSpecialUse({&result}, {this});
 
     return result;
 }
@@ -932,9 +945,9 @@ NDArray NDArray::reduceNumber(nd4j::reduce::SameOps op, void *extraParams) const
 
     NDArray result(dataType(), getContext());
 
-    prepareSpecialUse({&result}, {this});
+    NDArray::prepareSpecialUse({&result}, {this});
     NativeOpExecutioner::execReduceSameScalar(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), extraParams, result.buffer(), result.shapeInfo(), result.specialBuffer(), result.specialShapeInfo());
-    registerSpecialUse({&result}, {this});
+    NDArray::registerSpecialUse({&result}, {this});
 
     return result;
 }
@@ -947,9 +960,9 @@ NDArray NDArray::reduceNumber(nd4j::reduce::BoolOps op, void *extraParams) const
     auto shape = ConstantShapeHelper::getInstance()->scalarShapeInfo(DataType::BOOL);
     NDArray result(shape, true, this->getContext());
 
-    prepareSpecialUse({&result}, {this});
+    NDArray::prepareSpecialUse({&result}, {this});
     NativeOpExecutioner::execReduceBoolScalar(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), extraParams, result.buffer(), result.shapeInfo(), result.specialBuffer(), result.specialShapeInfo());
-    registerSpecialUse({&result}, {this});
+    NDArray::registerSpecialUse({&result}, {this});
 
     return result;
 }
@@ -962,9 +975,9 @@ NDArray NDArray::reduceNumber(nd4j::reduce::LongOps op, void *extraParams) const
     auto shape = ConstantShapeHelper::getInstance()->scalarShapeInfo(DataType::INT64);
     NDArray result(shape, true, this->getContext());
 
-    prepareSpecialUse({&result}, {this});
+    NDArray::prepareSpecialUse({&result}, {this});
     NativeOpExecutioner::execReduceLongScalar(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), extraParams, result.buffer(), result.shapeInfo(), result.specialBuffer(), result.specialShapeInfo());
-    registerSpecialUse({&result}, {this});
+    NDArray::registerSpecialUse({&result}, {this});
 
     return result;
 }
@@ -976,9 +989,9 @@ void NDArray::reduceNumber(nd4j::reduce::FloatOps op, NDArray& target, void *ext
     if(!target.isScalar() || target.dataType() != DataTypeUtils::pickFloatingType(dataType()))
         throw std::invalid_argument("NDArray::reduceNumber FloatOps: target array should be scalar and have corresponding float type!");
 
-    prepareSpecialUse({&target}, {this});
+    NDArray::prepareSpecialUse({&target}, {this});
     NativeOpExecutioner::execReduceFloatScalar(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), extraParams, target.buffer(), target.shapeInfo(), target.specialBuffer(), target.specialShapeInfo());
-    registerSpecialUse({&target}, {this});
+    NDArray::registerSpecialUse({&target}, {this});
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -989,9 +1002,9 @@ void NDArray::reduceNumber(nd4j::reduce::SameOps op, NDArray& target, void *extr
     if(!target.isScalar() || target.dataType() != dataType())
         throw std::invalid_argument("NDArray::reduceNumber SameOps: target array should be scalar and have same type as this array!");
 
-    prepareSpecialUse({&target}, {this});
+    NDArray::prepareSpecialUse({&target}, {this});
     NativeOpExecutioner::execReduceSameScalar(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), extraParams, target.getBuffer(), target.getShapeInfo(), target.specialBuffer(), target.getSpecialShapeInfo());
-    registerSpecialUse({&target}, {this});
+    NDArray::registerSpecialUse({&target}, {this});
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -1002,9 +1015,9 @@ void NDArray::reduceNumber(nd4j::reduce::BoolOps op, NDArray& target, void *extr
     if(!target.isScalar() || target.dataType() != DataType::BOOL)
         throw std::invalid_argument("NDArray::reduceNumber BoolOps: target array should be scalar and have bool type!");
 
-    prepareSpecialUse({&target}, {this});
+    NDArray::prepareSpecialUse({&target}, {this});
     NativeOpExecutioner::execReduceBoolScalar(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), extraParams, target.getBuffer(), target.getShapeInfo(), target.specialBuffer(), target.getSpecialShapeInfo());
-    registerSpecialUse({&target}, {this});
+    NDArray::registerSpecialUse({&target}, {this});
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -1015,9 +1028,9 @@ void NDArray::reduceNumber(nd4j::reduce::LongOps op, NDArray& target, void *extr
     if(!target.isScalar() || target.dataType() != DataType::INT64)
         throw std::invalid_argument("NDArray::reduceNumber LongOps: target array should be scalar and have long type!");
 
-    prepareSpecialUse({&target}, {this});
+    NDArray::prepareSpecialUse({&target}, {this});
     NativeOpExecutioner::execReduceLongScalar(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), extraParams, target.getBuffer(), target.getShapeInfo(), target.specialBuffer(), target.getSpecialShapeInfo());
-    registerSpecialUse({&target}, {this});
+    NDArray::registerSpecialUse({&target}, {this});
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -1027,9 +1040,9 @@ NDArray NDArray::indexReduceNumber(nd4j::indexreduce::Ops op, ExtraArguments *ex
 
     auto res = NDArrayFactory::create<Nd4jLong>(0);
 
-    NDArray::prepareSpecialUse({&res}, {this});
+    NDArray::NDArray::prepareSpecialUse({&res}, {this});
     NativeOpExecutioner::execIndexReduceScalar(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), extraParams == nullptr ? nullptr : extraParams->argumentsAsT(this->dataType()), res.buffer(), res.shapeInfo(), res.specialBuffer(), res.specialShapeInfo());
-    NDArray::registerSpecialUse({&res}, {this});
+    NDArray::NDArray::registerSpecialUse({&res}, {this});
 
     return res;
 }
@@ -1240,17 +1253,10 @@ BUILD_SINGLE_TEMPLATE(template void* NDArray::templatedPointerShift, (const Nd4j
 
 //////////////////////////////////////////////////////////////////////////
 // method makes copy of this array and applies to the copy transpose operation, this array remains unaffected
-NDArray* NDArray::transpose() const {
+NDArray NDArray::transpose() const {
-    auto newArr = new NDArray(getBuffer(), getSpecialBuffer(), getShapeInfo(), getContext(), false, false);
+    NDArray newArr(getDataBuffer(), ShapeDescriptor(getShapeInfo()), getContext(), getBufferOffset());
-    newArr->transposei();
-
-    return newArr;
-}
-
-////////////////////////////////////////////////////////////////////////
-NDArray NDArray::transp() const {
-    NDArray newArr(getBuffer(), getShapeInfo(), getContext(), false);
     newArr.transposei();
+
     return newArr;
 }
 
@@ -1360,10 +1366,10 @@ Nd4jLong NDArray::argMax(std::initializer_list<int> dimensions) {
 
 //////////////////////////////////////////////////////////////////////////
 // create new array with corresponding order and shape, new array will point to the same _buffer as this array
-NDArray* NDArray::reshape(const char order, const std::vector<Nd4jLong>& shape) const {
+NDArray NDArray::reshape(const char order, const std::vector<Nd4jLong>& shape) const {
 
-    auto newArr = new NDArray(getDataBuffer(), ShapeDescriptor(getShapeInfo()), getContext());
+    NDArray newArr(getDataBuffer(), ShapeDescriptor(getShapeInfo()), getContext(), getBufferOffset());
-    newArr->reshapei(order, shape);
+    newArr.reshapei(order, shape);
 
     return newArr;
 }
@@ -1420,43 +1426,43 @@ bool NDArray::permutei(const std::vector<Nd4jLong>& dimensions) {
 }
 
 //////////////////////////////////////////////////////////////////////////
-NDArray* NDArray::permute(const int* dimensions, const int rank) const {
+NDArray NDArray::permute(const int* dimensions, const int rank) const {
 
     // evaluate shapeInfo for output (permuted) array ret
     auto shapeInfoPermuted = ShapeUtils::evalPermShapeInfo(dimensions, rank, *this, getContext()->getWorkspace());
-    auto ret = new NDArray(_buffer, ShapeDescriptor(shapeInfoPermuted), getContext(), getBufferOffset());
+    NDArray ret(getDataBuffer(), ShapeDescriptor(shapeInfoPermuted), getContext(), getBufferOffset());
-	ret->_isView = true;
+	ret._isView = true;
     return ret;
 }
 
 /////////////////////////////////////////////////////////////////////////
-NDArray* NDArray::permute(const Nd4jLong* dimensions, const int rank) const {
+NDArray NDArray::permute(const Nd4jLong* dimensions, const int rank) const {
     int tempDims[MAX_RANK];
     shape::convertT<Nd4jLong, int>(const_cast<Nd4jLong *>(dimensions), tempDims, rank);
     return permute(tempDims, rank);
 }
 
 //////////////////////////////////////////////////////////////////////////
-NDArray* NDArray::permute(const std::vector<int>& dimensions) const {
+NDArray NDArray::permute(const std::vector<int>& dimensions) const {
     auto data = dimensions.data();
     auto size = dimensions.size();
     return permute(data, size);
 }
 
 //////////////////////////////////////////////////////////////////////////
-NDArray* NDArray::permute(const std::vector<Nd4jLong>& dimensions) const {
+NDArray NDArray::permute(const std::vector<Nd4jLong>& dimensions) const {
     return permute(dimensions.data(), dimensions.size());
 }
 
 
 //////////////////////////////////////////////////////////////////////////
-NDArray* NDArray::permute(const std::initializer_list<int>& dimensions) const {
+NDArray NDArray::permute(const std::initializer_list<int>& dimensions) const {
     std::vector<int> vec(dimensions);
     return permute(vec);
 }
 
 //////////////////////////////////////////////////////////////////////////
-NDArray* NDArray::permute(const std::initializer_list<Nd4jLong>& dimensions) const {
+NDArray NDArray::permute(const std::initializer_list<Nd4jLong>& dimensions) const {
     std::vector<Nd4jLong> vec(dimensions);
     return permute(vec);
 }
@@ -1528,10 +1534,9 @@ bool NDArray::isUnitary() {
         throw std::runtime_error("isUnitary method: matrix must be square and have rank = 2 !");
 
     auto tr = this->transpose();
-    auto trMul = MmulHelper::mmul(this, tr, nullptr, 1.f, 0.f);
+    auto trMul = MmulHelper::mmul(this, &tr, nullptr, 1.f, 0.f);
 
     bool result = trMul->isIdentityMatrix();
-    delete tr;
     delete trMul;
 
     return result;
@@ -1777,11 +1782,11 @@ NDArray NDArray::operator*(const T& scalar) const {
 
     auto tmp = NDArrayFactory::create(dataType(), scalar, getContext());
     NDArray result(_shapeInfo, DataTypeUtils::pickPairwiseResultType(dataType(), DataTypeUtils::fromT<T>()), false, getContext());
+
     NDArray::prepareSpecialUse({&result}, {this, &tmp});
-
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::Multiply, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), result.buffer(), result.getShapeInfo(), result.specialBuffer(), result.getSpecialShapeInfo(), tmp.buffer(), tmp.shapeInfo(), tmp.specialBuffer(), tmp.specialShapeInfo(), nullptr);
-
     NDArray::registerSpecialUse({&result}, {this, &tmp});
+
     return result;
 }
 template NDArray NDArray::operator*(const double&   scalar) const;
@@ -1811,6 +1816,7 @@ NDArray NDArray::operator/(const T& scalar) const {
     NDArray::prepareSpecialUse({&result}, {this, &tmp});
     NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::Divide, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), result.buffer(), result.getShapeInfo(), result.specialBuffer(), result.getSpecialShapeInfo(), tmp.buffer(), tmp.shapeInfo(), tmp.specialBuffer(), tmp.specialShapeInfo(), nullptr);
     NDArray::registerSpecialUse({&result}, {this, &tmp});
+
     return result;
 }
 template NDArray NDArray::operator/(const double&   scalar) const;
@@ -2050,14 +2056,14 @@ void NDArray::operator+=(const NDArray& other) {
         throw nd4j::datatype_exception::build("NDArray operator+=: Cannot add different types", this->dataType(), other.dataType());
 
     if (!this->isScalar() && other.isScalar()) {
-        prepareSpecialUse({this}, {this, &other});
+        NDArray::prepareSpecialUse({this}, {this, &other});
         NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::Add, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({this}, {this, &other});
+        NDArray::registerSpecialUse({this}, {this, &other});
     }
     else if (other.lengthOf() == lengthOf() && this->rankOf() == other.rankOf()) {
-        prepareSpecialUse({this}, {this, &other});
+        NDArray::prepareSpecialUse({this}, {this, &other});
         NativeOpExecutioner::execPairwiseTransform(getContext(), nd4j::pairwise::Add, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({this}, {this, &other});
+        NDArray::registerSpecialUse({this}, {this, &other});
     }
     else{
         Nd4jLong *bShape = nullptr;
@@ -2084,14 +2090,14 @@ void NDArray::operator-=(const NDArray& other) {
         throw nd4j::datatype_exception::build("NDArray operator-=: Cannot subtract different types", this->dataType(), other.dataType());
 
     if (!this->isScalar() && other.isScalar()) {
-        prepareSpecialUse({this}, {this, &other});
+        NDArray::prepareSpecialUse({this}, {this, &other});
         NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::Subtract, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({this}, {this, &other});
+        NDArray::registerSpecialUse({this}, {this, &other});
     }
     else if (other.lengthOf() == lengthOf() && this->rankOf() == other.rankOf()) {
-        prepareSpecialUse({this}, {this, &other});
+        NDArray::prepareSpecialUse({this}, {this, &other});
         NativeOpExecutioner::execPairwiseTransform(getContext(), nd4j::pairwise::Subtract, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({this}, {this, &other});
+        NDArray::registerSpecialUse({this}, {this, &other});
     }
     else{
         Nd4jLong *bShape = nullptr;
@@ -2117,14 +2123,14 @@ void NDArray::operator*=(const NDArray& other) {
         throw nd4j::datatype_exception::build("NDArray operator*=: Cannot multiply different types", this->dataType(), other.dataType());
 
     if (!this->isScalar() && other.isScalar()) {
-        prepareSpecialUse({this}, {this, &other});
+        NDArray::prepareSpecialUse({this}, {this, &other});
         NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::Multiply, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({this}, {this, &other});
+        NDArray::registerSpecialUse({this}, {this, &other});
     }
     else if (other.lengthOf() == lengthOf() && this->rankOf() == other.rankOf()) {
-        prepareSpecialUse({this}, {this, &other});
+        NDArray::prepareSpecialUse({this}, {this, &other});
         NativeOpExecutioner::execPairwiseTransform(getContext(), nd4j::pairwise::Multiply, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({this}, {this, &other});
+        NDArray::registerSpecialUse({this}, {this, &other});
     }
     else{
         Nd4jLong *bShape = nullptr;
@@ -2154,14 +2160,14 @@ void NDArray::operator/=(const NDArray& other) {
     }
 
     if (!this->isScalar() && other.isScalar()) {
-        prepareSpecialUse({this}, {this, &other});
+        NDArray::prepareSpecialUse({this}, {this, &other});
         NativeOpExecutioner::execScalar(getContext(), nd4j::scalar::Divide, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({this}, {this, &other});
+        NDArray::registerSpecialUse({this}, {this, &other});
     }
     else if (other.lengthOf() == lengthOf() && this->rankOf() == other.rankOf()) {
-        prepareSpecialUse({this}, {this, &other});
+        NDArray::prepareSpecialUse({this}, {this, &other});
         NativeOpExecutioner::execPairwiseTransform(getContext(), nd4j::pairwise::Divide, buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), buffer(), getShapeInfo(), specialBuffer(), getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({this}, {this, &other});
+        NDArray::registerSpecialUse({this}, {this, &other});
     }
     else{
         Nd4jLong *bShape = nullptr;
@@ -2264,9 +2270,9 @@ NDArray NDArray::operator-(const NDArray& other) const {
     if (other.lengthOf() == lengthOf() && this->rankOf() == other.rankOf()) {
         NDArray result(getShapeInfo(), DataTypeUtils::pickPairwiseResultType(getShapeInfo(), other.getShapeInfo()), false, getContext());
 
-        prepareSpecialUse({&result}, {this, &other});
+        NDArray::prepareSpecialUse({&result}, {this, &other});
         NativeOpExecutioner::execPairwiseTransform(getContext(), nd4j::pairwise::Subtract, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), result.buffer(), result.getShapeInfo(), result.specialBuffer(), getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({&result}, {this, &other});
+        NDArray::registerSpecialUse({&result}, {this, &other});
 
         return result;
     }
@@ -2285,9 +2291,9 @@ NDArray NDArray::operator*(const NDArray& other) const {
     if (other.lengthOf() == lengthOf() && this->rankOf() == other.rankOf()) {
         NDArray result(getShapeInfo(), DataTypeUtils::pickPairwiseResultType(getShapeInfo(), other.getShapeInfo()), false, this->getContext());
 
-        prepareSpecialUse({&result}, {this, &other});
+        NDArray::prepareSpecialUse({&result}, {this, &other});
         NativeOpExecutioner::execPairwiseTransform(getContext(), nd4j::pairwise::Multiply, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), result.buffer(), result.getShapeInfo(), result.specialBuffer(), getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({&result}, {this, &other});
+        NDArray::registerSpecialUse({&result}, {this, &other});
 
         return result;
     }
@@ -2308,9 +2314,9 @@ NDArray NDArray::operator/(const NDArray& other) const {
     if (other.lengthOf() == lengthOf() && this->rankOf() == other.rankOf()) {
         NDArray result(getShapeInfo(), DataTypeUtils::pickPairwiseResultType(getShapeInfo(), other.getShapeInfo()), false, getContext());
 
-        prepareSpecialUse({&result}, {this, &other});
+        NDArray::prepareSpecialUse({&result}, {this, &other});
         NativeOpExecutioner::execPairwiseTransform(getContext(), nd4j::pairwise::Divide, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), result.buffer(), result.getShapeInfo(), result.specialBuffer(), getSpecialShapeInfo(), nullptr);
-        registerSpecialUse({&result}, {this, &other});
+        NDArray::registerSpecialUse({&result}, {this, &other});
 
         return result;
     }
@@ -2326,9 +2332,9 @@ NDArray NDArray::operator-() const {
 
     NDArray result(getShapeInfo(), false, getContext());
 
-    prepareSpecialUse({&result}, {this});
+    NDArray::prepareSpecialUse({&result}, {this});
     NativeOpExecutioner::execTransformSame(getContext(), nd4j::transform::Neg, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), result.buffer(), result.getShapeInfo(), result.specialBuffer(), result.getSpecialShapeInfo(), nullptr, nullptr, nullptr);
-    registerSpecialUse({&result}, {this});
+    NDArray::registerSpecialUse({&result}, {this});
 
     return result;
 }
@@ -2631,7 +2637,7 @@ void NDArray::applyBroadcast(nd4j::broadcast::Ops op, const std::vector<int>& di
     if (other->lengthOf() == lengthOf() && this->rankOf() == other->rankOf()) {
         NDArray::prepareSpecialUse({result}, {this, other});
         NativeOpExecutioner::execPairwiseTransform(getContext(), fromBroadcastToPairwise(op), buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), other->getBuffer(), other->getShapeInfo(), other->getSpecialBuffer(), other->getSpecialShapeInfo(), result->buffer(), result->shapeInfo(), result->specialBuffer(), result->specialShapeInfo(), nullptr);
-        registerSpecialUse({result}, {this, other});
+        NDArray::registerSpecialUse({result}, {this, other});
         return;
     }
 
@@ -2688,7 +2694,7 @@ void NDArray::applyBroadcast(nd4j::broadcast::BoolOps op, const std::vector<int>
     if (other->lengthOf() == lengthOf() && this->rankOf() == other->rankOf()) {
         NDArray::prepareSpecialUse({result}, {this, other});
         NativeOpExecutioner::execPairwiseBoolTransform(getContext(), fromBroadcastToPairwiseBool(op), buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), other->getBuffer(), other->getShapeInfo(), other->getSpecialBuffer(), other->getSpecialShapeInfo(), result->buffer(), result->shapeInfo(), result->specialBuffer(), result->specialShapeInfo(), nullptr);
-        registerSpecialUse({result}, {this, other});
+        NDArray::registerSpecialUse({result}, {this, other});
         return;
     }
 
@@ -2896,7 +2902,7 @@ bool NDArray::reshapei(const char order, const std::vector<Nd4jLong>& cshape) {
     Nd4jLong *shapeInfoNew;
     ALLOCATE(shapeInfoNew, getContext()->getWorkspace(), shape::shapeInfoLength(rank), Nd4jLong);
 
-    bool canReshape = shape::reshapeC(this->rankOf(), this->_shapeInfo, shape.size(), shape.data(), shapeInfoNew);
+    bool canReshape = shape::reshapeC(rankOf(), shapeInfo(), shape.size(), shape.data(), shapeInfoNew);
 
     // we can do this only if there was no permute applied, or there are no weird strides
     if (canReshape) {
@@ -2948,11 +2954,9 @@ void NDArray::applyPairwiseTransform(nd4j::pairwise::Ops op, const NDArray* othe
     if (target->dataType() != this->dataType() && target->dataType() != other->dataType())
         throw std::invalid_argument("NDArray::applyPairwiseTransform method - type of target array must be the same as type of this or other array !");
 
-    prepareSpecialUse({target}, {this, other});
+    NDArray::prepareSpecialUse({target}, {this, other});
-
     NativeOpExecutioner::execPairwiseTransform(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});
-    registerSpecialUse({target}, {this, other});
 
     if (extraParams != nullptr)
         synchronize("NDArray::applyPairwiseTransform");
@@ -2969,9 +2973,9 @@ void NDArray::applyPairwiseTransform(nd4j::pairwise::BoolOps op, const NDArray *
     if (dataType() != other->dataType())
         throw std::invalid_argument("NDArray::applyPairwiseTransform BoolOps method - this and other arrays must have the same type !");
 
-    prepareSpecialUse({target}, {this, other});
+    NDArray::prepareSpecialUse({target}, {this, other});
     NativeOpExecutioner::execPairwiseBoolTransform(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getShapeInfo(), other->getBuffer(), other->getShapeInfo(), other->getSpecialBuffer(), other->getSpecialShapeInfo(), target->buffer(), target->shapeInfo(), target->specialBuffer(), target->specialShapeInfo(), extraParams != nullptr ? extraParams->argumentsAsT(target->dataType()) : nullptr);
-    registerSpecialUse({target}, {this, other});
+    NDArray::registerSpecialUse({target}, {this, other});
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -3070,22 +3074,23 @@ void NDArray::assign(const NDArray& other) {
     if (other.isScalar()) {
 
         if(this->isScalar()) {
-            preparePrimaryUse({this}, {&other});
+            NDArray::preparePrimaryUse({this}, {&other});
             BUILD_DOUBLE_SELECTOR(dataType(), other.dataType(), templatedDoubleAssign, (buffer(), 0, other.getBuffer(), 0), LIBND4J_TYPES, LIBND4J_TYPES);
-            registerPrimaryUse({this}, {&other});
+            NDArray::registerPrimaryUse({this}, {&other});
+            this->syncToDevice();
         }
         else {
             if (dataType() != other.dataType()) {
                 auto tmp = other.cast(dataType());
-                prepareSpecialUse({this}, {tmp});
+                NDArray::prepareSpecialUse({this}, {tmp});
                 NativeOpExecutioner::execScalar(getContext(), scalar::CopyPws, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), tmp->getBuffer(), tmp->getShapeInfo(), tmp->getSpecialBuffer(), tmp->getSpecialShapeInfo(), nullptr);
-                registerSpecialUse({this}, {});
+                NDArray::registerSpecialUse({this}, {});
                 delete tmp;
             }
             else {
-                prepareSpecialUse({this}, {&other});
+                NDArray::prepareSpecialUse({this}, {&other});
                 NativeOpExecutioner::execScalar(getContext(), scalar::CopyPws, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), nullptr);
-                registerSpecialUse({this}, {&other});
+                NDArray::registerSpecialUse({this}, {&other});
             }
         }
     }
@@ -3101,16 +3106,16 @@ void NDArray::assign(const NDArray& other) {
         if (ordering() == other.ordering() && dataType() == other.dataType() && ews() == 1 && other.ews() == 1)
             copyBuffersContinuouslyFrom(other, other.lengthOf() * other.sizeOfT());
         else {
-            prepareSpecialUse({this}, {&other});
+            NDArray::prepareSpecialUse({this}, {&other});
             NativeOpExecutioner::execTransformAny(getContext(), transform::Assign, other.getBuffer(), other.getShapeInfo(), other.getSpecialBuffer(), other.getSpecialShapeInfo(), buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), nullptr, nullptr, nullptr);
-            registerSpecialUse({this}, {&other});
+            NDArray::registerSpecialUse({this}, {&other});
         }
     }
 }
 
 ////////////////////////////////////////////////////////////////////////
 // This method returns new copy of this NDArray, optionally in different order
-NDArray* NDArray::dup(const char newOrder) {
+NDArray* NDArray::dup(const char newOrder) const {
 
     if (isEmpty())
         return NDArrayFactory::empty_(dataType(), getContext());
@@ -3170,7 +3175,7 @@ std::string NDArray::e(const Nd4jLong i) const {
     if (!isS())
         throw std::runtime_error("Can't get std::string out of non-string array");
 
-    preparePrimaryUse({}, {this});
+    NDArray::preparePrimaryUse({}, {this});
 
     // getting "virtual" offset. it's not real though,since it doesn't take lengths into account
     auto offset = getOffset(i);
@@ -3208,8 +3213,8 @@ T NDArray::e(const Nd4jLong i) const {
 
     const auto rp = getOffset(i);
 
-    preparePrimaryUse({}, {this});
+    NDArray::preparePrimaryUse({}, {this});
-    registerPrimaryUse({}, {this});
+    NDArray::registerPrimaryUse({}, {this});
     BUILD_SINGLE_PARTIAL_SELECTOR(dataType(), return templatedGet<, T>(getBuffer(), rp), LIBND4J_TYPES);
 
 }
@@ -3226,8 +3231,8 @@ T NDArray::e(const Nd4jLong i, const Nd4jLong j) const {
     const Nd4jLong coords[2] = {i, j};
     const auto xOffset = shape::getOffset(0, shapeOf(), stridesOf(), coords, rankOf());
 
-    preparePrimaryUse({}, {this});
+    NDArray::preparePrimaryUse({}, {this});
-    registerPrimaryUse({}, {this});
+    NDArray::registerPrimaryUse({}, {this});
 
     BUILD_SINGLE_PARTIAL_SELECTOR(dataType(), return templatedGet<, T>(getBuffer(), xOffset), LIBND4J_TYPES);
 
@@ -3246,8 +3251,8 @@ T NDArray::e(const Nd4jLong i, const Nd4jLong j, const Nd4jLong k) const {
     const Nd4jLong coords[3] = {i, j, k};
     const auto xOffset = shape::getOffset(0, shapeOf(), stridesOf(), coords, rankOf());
 
-    preparePrimaryUse({}, {this});
+    NDArray::preparePrimaryUse({}, {this});
-    registerPrimaryUse({}, {this});
+    NDArray::registerPrimaryUse({}, {this});
 
     BUILD_SINGLE_PARTIAL_SELECTOR(dataType(), return templatedGet<, T>(getBuffer(), xOffset), LIBND4J_TYPES);
 
@@ -3266,8 +3271,8 @@ T NDArray::e(const Nd4jLong i, const Nd4jLong j, const Nd4jLong k, const Nd4jLon
     const Nd4jLong coords[4] = {i, j, k, l};
     const auto xOffset = shape::getOffset(0, shapeOf(), stridesOf(), coords, rankOf());
 
-    preparePrimaryUse({}, {this});
+    NDArray::preparePrimaryUse({}, {this});
-    registerPrimaryUse({}, {this});
+    NDArray::registerPrimaryUse({}, {this});
 
     BUILD_SINGLE_PARTIAL_SELECTOR(dataType(), return templatedGet<, T>(getBuffer(), xOffset), LIBND4J_TYPES);
 
@@ -3300,9 +3305,9 @@ void NDArray::applyTransform(nd4j::transform::FloatOps op, NDArray *target, Extr
     if (!target->isR())
         throw std::runtime_error("NDArray::applyTransform FloatOps: target array must have one of FLOAT types");
 
-    prepareSpecialUse({target}, {this});
+    NDArray::prepareSpecialUse({target}, {this});
     NativeOpExecutioner::execTransformFloat(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), target->buffer(), target->shapeInfo(), target->specialBuffer(), target->specialShapeInfo(), extraParams != nullptr ? extraParams->argumentsAsT(target->dataType()) : nullptr, nullptr, nullptr);
-    registerSpecialUse({target}, {this});
+    NDArray::registerSpecialUse({target}, {this});
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -3314,9 +3319,9 @@ void NDArray::applyTransform(nd4j::transform::AnyOps op, NDArray *target, ExtraA
     if (target == nullptr)
         target = this;
 
-    prepareSpecialUse({target}, {this});
+    NDArray::prepareSpecialUse({target}, {this});
     NativeOpExecutioner::execTransformAny(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), target->buffer(), target->shapeInfo(), target->specialBuffer(), target->specialShapeInfo(), extraParams != nullptr ? extraParams->argumentsAsT(target->dataType()) : nullptr, nullptr, nullptr);
-    registerSpecialUse({target}, {this});
+    NDArray::registerSpecialUse({target}, {this});
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -3331,9 +3336,9 @@ void NDArray::applyTransform(nd4j::transform::SameOps op, NDArray *target, Extra
     if (target->dataType() != dataType())
         throw std::runtime_error("NDArray::applyTransform SameOps: target array must have the same data type as original array");
 
-    prepareSpecialUse({target}, {this});
+    NDArray::prepareSpecialUse({target}, {this});
     NativeOpExecutioner::execTransformSame(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), target->buffer(), target->shapeInfo(), target->specialBuffer(), target->specialShapeInfo(), extraParams != nullptr ? extraParams->argumentsAsT(target->dataType()) : nullptr, nullptr, nullptr);
-    registerSpecialUse({target}, {this});
+    NDArray::registerSpecialUse({target}, {this});
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -3347,9 +3352,9 @@ void NDArray::applyTransform(nd4j::transform::StrictOps op, NDArray *target, Ext
     if (!this->isR() || !target->isR() || (this->dataType() != target->dataType()))
         throw std::runtime_error("NDArray::applyTransform StrictOps: both Source and Target array must have same FLOAT type !");
 
-    registerSpecialUse({target}, {this});
+    NDArray::prepareSpecialUse({target}, {this});
     NativeOpExecutioner::execTransformStrict(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), target->buffer(), target->shapeInfo(), target->specialBuffer(), target->specialShapeInfo(), extraParams != nullptr ? extraParams->argumentsAsT(target->dataType()) : nullptr, nullptr, nullptr);
-    prepareSpecialUse({target}, {this});
+    NDArray::registerSpecialUse({target}, {this});
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -3363,9 +3368,9 @@ void NDArray::applyTransform(nd4j::transform::BoolOps op, NDArray *target, Extra
     if (!target->isB())
         throw std::runtime_error("NDArray::applyTransform BoolOps: target array must have one of BOOL types");
 
-    prepareSpecialUse({target}, {this});
+    NDArray::prepareSpecialUse({target}, {this});
     NativeOpExecutioner::execTransformBool(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), target->buffer(), target->shapeInfo(), target->specialBuffer(), target->specialShapeInfo(), extraParams != nullptr ? extraParams->argumentsAsT(target->dataType()) : nullptr, nullptr, nullptr);
-    registerSpecialUse({target}, {this});
+    NDArray::registerSpecialUse({target}, {this});
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -3375,9 +3380,9 @@ NDArray NDArray::transform(nd4j::transform::FloatOps op, void *extraParams) cons
 
     NDArray result(ordering(), getShapeAsVector(), DataTypeUtils::pickFloatingType(dataType()), getContext());
 
-    registerSpecialUse({&result}, {this});
+    NDArray::prepareSpecialUse({&result}, {this});
     NativeOpExecutioner::execTransformFloat(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), result.buffer(), result.shapeInfo(), result.specialBuffer(), result.specialShapeInfo(), extraParams, nullptr, nullptr);
-    prepareSpecialUse({&result}, {this});
+    NDArray::registerSpecialUse({&result}, {this});
 
     return result;
 }
@@ -3389,9 +3394,9 @@ NDArray NDArray::transform(nd4j::transform::SameOps op, void *extraParams) const
 
     NDArray result(getShapeInfo(), false, getContext());
 
-    prepareSpecialUse({&result}, {this});
+    NDArray::prepareSpecialUse({&result}, {this});
     NativeOpExecutioner::execTransformSame(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), result.buffer(), result.shapeInfo(), result.specialBuffer(), result.specialShapeInfo(), extraParams, nullptr, nullptr);
-    registerSpecialUse({&result}, {this});
+    NDArray::registerSpecialUse({&result}, {this});
 
     return result;
 }
@@ -3403,9 +3408,9 @@ NDArray NDArray::transform(nd4j::transform::StrictOps op, void *extraParams) con
 
     NDArray result(getShapeInfo(), false, getContext());
 
-    prepareSpecialUse({&result}, {this});
+    NDArray::prepareSpecialUse({&result}, {this});
     NativeOpExecutioner::execTransformStrict(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), result.buffer(), result.shapeInfo(), result.specialBuffer(), result.specialShapeInfo(), extraParams, nullptr, nullptr);
-    registerSpecialUse({&result}, {this});
+    NDArray::registerSpecialUse({&result}, {this});
 
     return result;
 }
@@ -3417,9 +3422,9 @@ NDArray NDArray::transform(nd4j::transform::BoolOps op, void *extraParams) const
 
     NDArray result(ordering(), getShapeAsVector(), nd4j::DataType::BOOL, getContext());
 
-    prepareSpecialUse({&result}, {this});
+    NDArray::prepareSpecialUse({&result}, {this});
     NativeOpExecutioner::execTransformBool(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), result.buffer(), result.shapeInfo(), result.specialBuffer(), result.specialShapeInfo(), extraParams, nullptr, nullptr);
-    registerSpecialUse({&result}, {this});
+    NDArray::registerSpecialUse({&result}, {this});
 
     return result;
 }
@@ -3435,9 +3440,9 @@ void NDArray::applyScalarArr(nd4j::scalar::Ops op, const NDArray* scalar, NDArra
     if(target->dataType() != DataTypeUtils::pickPairwiseResultType(shapeInfo(), scalar->getShapeInfo()) && !(target->dataType() == dataType() || target->dataType() == scalar->dataType()))
         throw std::invalid_argument("NDArray::applyScalarArr method: wrong type of target array!");
 
-    prepareSpecialUse({target}, {this, scalar});
+    NDArray::prepareSpecialUse({target}, {this, scalar});
     NativeOpExecutioner::execScalar(getContext(), op, buffer(), shapeInfo(), specialBuffer(), specialShapeInfo(), target->buffer(), target->shapeInfo(), target->specialBuffer(), target->specialShapeInfo(), scalar->getBuffer(), scalar->getShapeInfo(), scalar->getSpecialBuffer(), scalar->getSpecialShapeInfo(), extraParams != nullptr ? extraParams->argumentsAsT(target->dataType()): nullptr);
-    registerSpecialUse({target}, {this, scalar});
+    NDArray::registerSpecialUse({target}, {this, scalar});
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -3471,10 +3476,9 @@ void NDArray::applyScalarArr(nd4j::scalar::BoolOps op, const NDArray* scalar, ND
         throw std::invalid_argument("NDArray::applyScalarArr bool method: this and scalar arrays must have the same type!");
     }
 
-    prepareSpecialUse({target}, {this, scalar});
+    NDArray::prepareSpecialUse({target}, {this, scalar});
     NativeOpExecutioner::execScalarBool(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});
-    registerSpecialUse({target}, {this, scalar});
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -3557,7 +3561,7 @@ NDArray* NDArray::applyReduce3(nd4j::reduce3::Ops op, const NDArray* other, cons
 
     NDArray::prepareSpecialUse({result}, {this, other});
     NativeOpExecutioner::execReduce3Scalar(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), params, other->getBuffer(), other->getShapeInfo(), other->getSpecialBuffer(), other->getSpecialShapeInfo(), result->buffer(), result->shapeInfo(), result->specialBuffer(), result->specialShapeInfo());
-    registerSpecialUse({result}, {this, other});
+    NDArray::registerSpecialUse({result}, {this, other});
 
     return result;
 }
@@ -3635,9 +3639,9 @@ NDArray* NDArray::applyAllReduce3(nd4j::reduce3::Ops op, const NDArray *other, c
 
     auto pDims = nd4j::Environment::getInstance()->isCPU() ? copy.data() : nullptr;
 
-    prepareSpecialUse({result}, {this, other});
+    NDArray::prepareSpecialUse({result}, {this, other});
     NativeOpExecutioner::execReduce3All(getContext(), op, getBuffer(), getShapeInfo(), getSpecialBuffer(), getSpecialShapeInfo(), params, other->getBuffer(), other->getShapeInfo(), other->getSpecialBuffer(), other->getSpecialShapeInfo(), result->buffer(), result->shapeInfo(), result->specialBuffer(), result->specialShapeInfo(), pDims, copy.size(), packX.platformShapeInfo(), packX.platformOffsets(), packY.platformShapeInfo(), packY.platformOffsets());
-    registerSpecialUse({result}, {this, other});
+    NDArray::registerSpecialUse({result}, {this, other});
 
     return result;
 }
@@ -3780,9 +3784,9 @@ void NDArray::p(const Nd4jLong i, const T value) {
     auto rp = getOffset(i);
     const void *pV = reinterpret_cast<const void*>(const_cast<T *>(&value));
 
-    preparePrimaryUse({this}, {}, true);
+    NDArray::preparePrimaryUse({this}, {}, true);
     BUILD_SINGLE_PARTIAL_SELECTOR(this->dataType(), templatedSet<, T>(this->getBuffer(), rp, pV), LIBND4J_TYPES);
-    registerPrimaryUse({this}, {});
+    NDArray::registerPrimaryUse({this}, {});
 }
 
 template void NDArray::p(const Nd4jLong i, const double value);
@@ -3811,9 +3815,9 @@ void NDArray::p(const Nd4jLong i, const Nd4jLong j, const T value) {
     Nd4jLong coords[2] = {i, j};
     auto xOffset = shape::getOffset(0, shapeOf(), stridesOf(), coords, rankOf());
 
-    preparePrimaryUse({this}, {}, true);
+    NDArray::preparePrimaryUse({this}, {}, true);
     BUILD_SINGLE_PARTIAL_SELECTOR(dataType(), templatedSet<, T>(this->getBuffer(), xOffset, p), LIBND4J_TYPES);
-    registerPrimaryUse({this}, {});
+    NDArray::registerPrimaryUse({this}, {});
 }
 template void NDArray::p(const Nd4jLong i, const Nd4jLong j, const double value);
 template void NDArray::p(const Nd4jLong i, const Nd4jLong j, const float value);
@@ -3837,13 +3841,13 @@ void NDArray::p(const Nd4jLong i, const Nd4jLong j, const Nd4jLong k, const T va
     if (rankOf() != 3 || i >= shapeOf()[0] || j >= shapeOf()[1] || k >= shapeOf()[2])
         throw std::invalid_argument("NDArray:pe(i,j,k, value): one of input indexes is out of array length or rank!=3 !");
 
-    preparePrimaryUse({this}, {}, true);
+    NDArray::preparePrimaryUse({this}, {}, true);
 
     void *p = reinterpret_cast<void *>(const_cast<T *>(&value));
     Nd4jLong coords[3] = {i, j, k};
     auto xOffset = shape::getOffset(0, shapeOf(), stridesOf(), coords, rankOf());
     BUILD_SINGLE_PARTIAL_SELECTOR(dataType(), templatedSet<, T>(this->getBuffer(), xOffset, p), LIBND4J_TYPES);
-    registerPrimaryUse({this}, {});
+    NDArray::registerPrimaryUse({this}, {});
 }
 template void NDArray::p(const Nd4jLong i, const Nd4jLong j, const Nd4jLong k, const double value);
 template void NDArray::p(const Nd4jLong i, const Nd4jLong j, const Nd4jLong k, const float value);
@@ -3870,9 +3874,9 @@ void NDArray::p(const Nd4jLong i, const Nd4jLong j, const Nd4jLong k, const Nd4j
     Nd4jLong coords[4] = {i, j, k, l};
     auto xOffset = shape::getOffset(0, shapeOf(), stridesOf(), coords, rankOf());
 
-    preparePrimaryUse({this}, {}, true);
+    NDArray::preparePrimaryUse({this}, {}, true);
     BUILD_SINGLE_PARTIAL_SELECTOR(dataType(), templatedSet<, T>(this->getBuffer(), xOffset, p), LIBND4J_TYPES);
-    registerPrimaryUse({this}, {});
+    NDArray::registerPrimaryUse({this}, {});
 }
 template void NDArray::p(const Nd4jLong i, const Nd4jLong j, const Nd4jLong k, const Nd4jLong l, const double value);
 template void NDArray::p(const Nd4jLong i, const Nd4jLong j, const Nd4jLong k, const Nd4jLong l, const float value);
@@ -3896,10 +3900,10 @@ void NDArray::p(const Nd4jLong i, const NDArray& scalar) {
     if (i >= _length)
         throw std::invalid_argument("NDArray::p(i, NDArray_scalar): input index is out of array length !");
 
-    preparePrimaryUse({this}, {&scalar}, true);
+    NDArray::preparePrimaryUse({this}, {&scalar}, true);
     auto rp = getOffset(i);
     BUILD_SINGLE_SELECTOR(scalar.dataType(), templatedSet, (getBuffer(), rp, scalar.dataType(), scalar.getBuffer()), LIBND4J_TYPES);
-    registerPrimaryUse({this}, {&scalar});
+    NDArray::registerPrimaryUse({this}, {&scalar});
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -4195,7 +4199,7 @@ ResultSet* NDArray::allTensorsAlongDimension(const std::vector<int> &dimensions)
 
 
     auto pack = ConstantTadHelper::getInstance()->tadForDimensions(_shapeInfo, const_cast<int*>(dimensions.data()), dimensions.size());
-    auto numTads = lengthOf() / shape::length(pack.primaryShapeInfo());
+    auto numTads = pack.numberOfTads();
 
     for (int idx = 0; idx < numTads; idx++ ) {
         auto array = new NDArray(_buffer, ShapeDescriptor(pack.primaryShapeInfo()), getContext(), pack.primaryOffsets()[idx] + getBufferOffset());

libnd4j/blas/NativeOps.h

@@ -1578,6 +1578,20 @@ public:
             void *dx, Nd4jLong *dxShapeInfo,
             bool descending);
 
+    void sortByKey(Nd4jPointer *extraPointers,
+                   void *x, Nd4jLong *xShapeInfo,
+                   void *dx, Nd4jLong *dxShapeInfo,
+                   void *y, Nd4jLong *yShapeInfo,
+                   void *dy, Nd4jLong *dyShapeInfo,
+                   bool descending);
+
+    void sortByValue(Nd4jPointer *extraPointers,
+                     void *x, Nd4jLong *xShapeInfo,
+                     void *dx, Nd4jLong *dxShapeInfo,
+                     void *y, Nd4jLong *yShapeInfo,
+                     void *dy, Nd4jLong *dyShapeInfo,
+                     bool descending);
+
     void sortTad(Nd4jPointer *extraPointers,
             void *x, Nd4jLong *xShapeInfo,
             void *dx, Nd4jLong *dxShapeInfo,
@@ -1587,6 +1601,24 @@ public:
             Nd4jLong *tadOffsets,
             bool descending);
 
+    void sortTadByKey(Nd4jPointer *extraPointers,
+                 void *x, Nd4jLong *xShapeInfo,
+                 void *dx, Nd4jLong *dxShapeInfo,
+                 void *y, Nd4jLong *yShapeInfo,
+                 void *dy, Nd4jLong *dyShapeInfo,
+                 int *dimension,
+                 int dimensionLength,
+                 bool descending);
+
+    void sortTadByValue(Nd4jPointer *extraPointers,
+                 void *x, Nd4jLong *xShapeInfo,
+                 void *dx, Nd4jLong *dxShapeInfo,
+                 void *y, Nd4jLong *yShapeInfo,
+                 void *dy, Nd4jLong *dyShapeInfo,
+                 int *dimension,
+                 int dimensionLength,
+                 bool descending);
+
 
     // special sort impl for sorting out COO indices and values
     void sortCooIndices(Nd4jPointer *extraPointers, Nd4jLong *indices, void *values, Nd4jLong length, int rank);

libnd4j/blas/cpu/NDArray.cpp

@@ -208,6 +208,23 @@ void* NDArray::specialBufferWithOffset(Nd4jLong offset) const {
     return nullptr;
 }
 
+////////////////////////////////////////////////////////////////////////
+void* NDArray::specialBuffer() {
+    if (_buffer->special() == nullptr)
+        return getBuffer();
+    // FIXME: this should be fixed once CUDA backend added
+    return static_cast<int8_t*>(_buffer->special()) + (_offset * sizeOfT());
+}
+
+////////////////////////////////////////////////////////////////////////
+void* NDArray::getSpecialBuffer() const {
+    if (_buffer->special() == nullptr)
+        return getBuffer();
+    // FIXME: this should be fixed once CUDA backend added
+    return static_cast<int8_t*>(_buffer->special()) + (_offset * sizeOfT());
+}
+
+
 //////////////////////////////////////////////////////////////////////////
 // change an array by repeating it the number of times given by reps.
 NDArray NDArray::tile(const std::vector<Nd4jLong>& reps) const {

libnd4j/blas/cpu/NDArrayFactory.cpp

@@ -27,6 +27,52 @@
 
 namespace nd4j {
 
+    ////////////////////////////////////////////////////////////////////////
+    template <>
+    NDArray NDArrayFactory::create<bool>(const char order, const std::vector<Nd4jLong> &shape, const std::vector<bool> &data, nd4j::LaunchContext * context) {
+
+        if ((int) shape.size() > MAX_RANK)
+            throw std::invalid_argument("NDArrayFactory::create: rank of NDArray can't exceed 32 !");
+
+        ShapeDescriptor descriptor(nd4j::DataType::BOOL, order, shape);
+
+        if (descriptor.arrLength() != data.size()) {
+            nd4j_printf("NDArrayFactory::create: data size [%i] doesn't match shape length [%lld]\n", data.size(), descriptor.arrLength());
+            throw std::runtime_error("NDArrayFactory::create: data size doesn't match shape");
+        }
+
+        bool* hostBuffer = nullptr;
+        ALLOCATE(hostBuffer, context->getWorkspace(), data.size(), bool);
+        std::copy(data.begin(), data.end(), hostBuffer);
+
+        std::shared_ptr<DataBuffer> buffer = std::make_shared<DataBuffer>(hostBuffer, data.size() * sizeof(bool), nd4j::DataType::BOOL, true, context->getWorkspace());
+
+        NDArray result(buffer, descriptor, context);
+
+        return result;
+    }
+
+    ////////////////////////////////////////////////////////////////////////
+    template <typename T>
+    NDArray NDArrayFactory::create(const char order, const std::vector<Nd4jLong> &shape, const std::vector<T> &data, nd4j::LaunchContext * context) {
+
+        if ((int) shape.size() > MAX_RANK)
+            throw std::invalid_argument("NDArrayFactory::create: rank of NDArray can't exceed 32 !");
+
+        ShapeDescriptor descriptor(DataTypeUtils::fromT<T>(), order, shape);
+
+        if (descriptor.arrLength() != data.size()) {
+            nd4j_printf("NDArrayFactory::create: data size [%i] doesn't match shape length [%lld]\n", data.size(), descriptor.arrLength());
+            throw std::runtime_error("NDArrayFactory::create: data size doesn't match shape");
+        }
+
+        std::shared_ptr<DataBuffer> buffer = std::make_shared<DataBuffer>(data.data(), DataTypeUtils::fromT<T>(), descriptor.arrLength() * sizeof(T), context->getWorkspace());
+
+        NDArray result(buffer, descriptor, context);
+
+        return result;
+
+    }
 
     NDArray NDArrayFactory::string(const char *str, nd4j::LaunchContext * context) {
         std::string s(str);
@@ -227,10 +273,13 @@ template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd
 template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<float16> &data, nd4j::LaunchContext * context);
 template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<bfloat16> &data, nd4j::LaunchContext * context);
 template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<int> &data, nd4j::LaunchContext * context);
+template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<unsigned int> &data, nd4j::LaunchContext * context);
+template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<unsigned long> &data, nd4j::LaunchContext * context);
 template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<Nd4jLong> &data, nd4j::LaunchContext * context);
 template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<int8_t> &data, nd4j::LaunchContext * context);
 template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<uint8_t> &data, nd4j::LaunchContext * context);
 template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<int16_t> &data, nd4j::LaunchContext * context);
+template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<uint16_t> &data, nd4j::LaunchContext * context);
 template NDArray* NDArrayFactory::create_(const char order, const std::vector<Nd4jLong> &shape, const std::vector<bool> &data, nd4j::LaunchContext * context);
 
 
@@ -391,6 +440,7 @@ template NDArray NDArrayFactory::create(const std::vector<bfloat16> &values, nd4
 template NDArray NDArrayFactory::create(const std::vector<Nd4jLong> &values, nd4j::LaunchContext * context);
 template NDArray NDArrayFactory::create(const std::vector<int> &values, nd4j::LaunchContext * context);
 template NDArray NDArrayFactory::create(const std::vector<int16_t> &values, nd4j::LaunchContext * context);
+template NDArray NDArrayFactory::create(const std::vector<uint16_t> &values, nd4j::LaunchContext * context);
 template NDArray NDArrayFactory::create(const std::vector<int8_t> &values, nd4j::LaunchContext * context);
 template NDArray NDArrayFactory::create(const std::vector<uint8_t> &values, nd4j::LaunchContext * context);
 template NDArray NDArrayFactory::create(const std::vector<bool> &values, nd4j::LaunchContext * context);
@@ -452,53 +502,6 @@ template NDArray NDArrayFactory::create(const std::vector<bool> &values, nd4j::L
         return new NDArray(order, shape, dataType, context);
     }
 
-////////////////////////////////////////////////////////////////////////
-    template <typename T>
-    NDArray NDArrayFactory::create(const char order, const std::vector<Nd4jLong> &shape, const std::vector<T> &data, nd4j::LaunchContext * context) {
-
-        if ((int) shape.size() > MAX_RANK)
-            throw std::invalid_argument("NDArrayFactory::create: rank of NDArray can't exceed 32 !");
-
-        ShapeDescriptor descriptor(DataTypeUtils::fromT<T>(), order, shape);
-
-        if (descriptor.arrLength() != data.size()) {
-            nd4j_printf("NDArrayFactory::create: data size [%i] doesn't match shape length [%lld]\n", data.size(), descriptor.arrLength());
-            throw std::runtime_error("NDArrayFactory::create: data size doesn't match shape");
-        }
-
-        std::shared_ptr<DataBuffer> buffer = std::make_shared<DataBuffer>(data.data(), DataTypeUtils::fromT<T>(), descriptor.arrLength() * sizeof(T), context->getWorkspace());
-
-        NDArray result(buffer, descriptor, context);
-
-        return result;
-
-    }
-    ////////////////////////////////////////////////////////////////////////
-    template <>
-    NDArray NDArrayFactory::create<bool>(const char order, const std::vector<Nd4jLong> &shape, const std::vector<bool> &data, nd4j::LaunchContext * context) {
-
-        if ((int) shape.size() > MAX_RANK)
-            throw std::invalid_argument("NDArrayFactory::create: rank of NDArray can't exceed 32 !");
-
-        ShapeDescriptor descriptor(nd4j::DataType::BOOL, order, shape);
-
-        if (descriptor.arrLength() != data.size()) {
-            nd4j_printf("NDArrayFactory::create: data size [%i] doesn't match shape length [%lld]\n", data.size(), descriptor.arrLength());
-            throw std::runtime_error("NDArrayFactory::create: data size doesn't match shape");
-        }
-
-        bool* hostBuffer = nullptr;
-        ALLOCATE(hostBuffer, context->getWorkspace(), data.size(), bool);
-        std::copy(data.begin(), data.end(), hostBuffer);
-
-        std::shared_ptr<DataBuffer> buffer = std::make_shared<DataBuffer>(hostBuffer, data.size() * sizeof(bool), nd4j::DataType::BOOL, true, context->getWorkspace());
-
-        NDArray result(buffer, descriptor, context);
-
-        return result;
-
-    }
-
 ////////////////////////////////////////////////////////////////////////
 template <typename T>
 NDArray NDArrayFactory::create(T* buffer, const char order, const std::initializer_list<Nd4jLong>& shape, nd4j::LaunchContext * context) {

libnd4j/blas/cpu/NativeOps.cpp

@@ -2736,6 +2736,60 @@ Nd4jPointer NativeOps::shapeBufferForNumpy(Nd4jPointer npyArray) {
     return reinterpret_cast<Nd4jPointer>(shapeBuffer);
 }
 
+void NativeOps::sortByKey(Nd4jPointer *extraPointers,
+                          void *x, Nd4jLong *xShapeInfo,
+                          void *dx, Nd4jLong *dxShapeInfo,
+                          void *y, Nd4jLong *yShapeInfo,
+                          void *dy, Nd4jLong *dyShapeInfo,
+                          bool descending) {
+    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);
+}
+
+void NativeOps::sortByValue(Nd4jPointer *extraPointers,
+                            void *x, Nd4jLong *xShapeInfo,
+                            void *dx, Nd4jLong *dxShapeInfo,
+                            void *y, Nd4jLong *yShapeInfo,
+                            void *dy, Nd4jLong *dyShapeInfo,
+                            bool descending) {
+
+    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);
+}
+
+void NativeOps::sortTadByKey(Nd4jPointer *extraPointers,
+                  void *x, Nd4jLong *xShapeInfo,
+                  void *dx, Nd4jLong *dxShapeInfo,
+                  void *y, Nd4jLong *yShapeInfo,
+                  void *dy, Nd4jLong *dyShapeInfo,
+                  int *dimension,
+                  int dimensionLength,
+                  bool descending) {
+    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);
+}
+
+void NativeOps::sortTadByValue(Nd4jPointer *extraPointers,
+                    void *x, Nd4jLong *xShapeInfo,
+                    void *dx, Nd4jLong *dxShapeInfo,
+                    void *y, Nd4jLong *yShapeInfo,
+                    void *dy, Nd4jLong *dyShapeInfo,
+                    int *dimension,
+                    int dimensionLength,
+                    bool descending) {
+    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_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);

libnd4j/blas/cuda/NDArray.cu

@@ -192,8 +192,8 @@ void NDArray::setIdentity() {
     if (isS())
         throw std::runtime_error("NDArray::setIdentity: you can't use this method on String array!");
 
-    if (rankOf() != 2)
+    // if (rankOf() != 2)
-        throw std::runtime_error("NDArray::setIdentity: method should work only for 2D tensors. But " + toStringValue(rankOf()) + " was given.");
+    //     throw std::runtime_error("NDArray::setIdentity: method should work only for 2D tensors. But " + toStringValue(rankOf()) + " was given.");
 
     const int threadsPerBlock = MAX_NUM_THREADS / 4;
     const int blocksPerGrid = (lengthOf() + threadsPerBlock - 1) / threadsPerBlock;
@@ -234,22 +234,27 @@ void NDArray::synchronize(const char* msg) const {
 void NDArray::prepareSpecialUse(const std::initializer_list<const NDArray*>& writeList, const std::initializer_list<const NDArray*>& readList, bool synchronizeWritables) {
 
     for (const auto& a : readList)
+        if(a != nullptr)
             a->syncToDevice();
 
     for (const auto& a : writeList) {
+        if (a != nullptr) {
             a->getDataBuffer()->allocateSpecial();
             if (synchronizeWritables)
                 a->syncToDevice();
         }
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
 void NDArray::registerSpecialUse(const std::initializer_list<const NDArray*>& writeList, const std::initializer_list<const NDArray*>& readList) {
 
     for (const auto& p : readList)
+        if(p != nullptr)
             p->tickReadDevice();
 
     for (const auto& p : writeList)
+        if (p != nullptr)
             p->tickWriteDevice();
 }
 
@@ -257,22 +262,27 @@ void NDArray::registerSpecialUse(const std::initializer_list<const NDArray*>& wr
 void NDArray::preparePrimaryUse(const std::initializer_list<const NDArray*>& writeList, const std::initializer_list<const NDArray*>& readList, bool synchronizeWritables) {
 
     for (const auto& a : readList)
+        if(a != nullptr)
             a->syncToHost();
 
     for (const auto& a : writeList) {
+        if (a != nullptr) {
             a->getDataBuffer()->allocatePrimary();
             if (synchronizeWritables)
                 a->syncToHost();
         }
+    }
 }
 
 ////////////////////////////////////////////////////////////////////////
 void NDArray::registerPrimaryUse(const std::initializer_list<const NDArray*>& writeList, const std::initializer_list<const NDArray*>& readList) {
 
     for (const auto& p : readList)
+        if(p != nullptr)
             p->tickReadHost();
 
     for (const auto& p : writeList)
+        if (p != nullptr)
             p->tickWriteHost();
 }
 
@@ -427,9 +437,26 @@ void NDArray::repeat(int dimension, NDArray& target) const {
     NDArray::registerSpecialUse({&target}, {this});
 }
 
+////////////////////////////////////////////////////////////////////////
+void* NDArray::specialBuffer() {
+
+    if (_buffer->special() == nullptr)
+        return getBuffer();
+    // FIXME: this should be fixed once CUDA backend added
+    return static_cast<int8_t*>(_buffer->special()) + (_offset * sizeOfT());
+}
+
+////////////////////////////////////////////////////////////////////////
+void* NDArray::getSpecialBuffer() const {
+    if (_buffer->special() == nullptr)
+        return getBuffer();
+    // FIXME: this should be fixed once CUDA backend added
+    return static_cast<int8_t*>(_buffer->special()) + (_offset * sizeOfT());
+}
+
 //////////////////////////////////////////////////////////////////////////
 template<typename T>
-void NDArray::printCurrentBuffer(const bool host, const char* msg, const int precision) const {\
+void NDArray::printCurrentBuffer(const bool host, const char* msg, const int precision) const {
 
     if(_length == 0)
             { printf("NDArray::printActualBuffer: array length is zero !\n"); return; }
@@ -477,7 +504,7 @@ template void NDArray::printCurrentBuffer<double>(const bool host, const char* m
 
 #if defined(__CUDACC__) && !defined(BUILD_TESTS)
 
-#include <cpu/NDArrayLambda.hpp>
+//#include <cpu/NDArrayLambda.hpp>
 
 #endif
 

libnd4j/blas/cuda/NativeOps.cu

@@ -2321,6 +2321,163 @@ void NativeOps::sort(Nd4jPointer *extraPointers,
 }
 
 
+void NativeOps::sortByKey(Nd4jPointer *extraPointers,
+               void *x, Nd4jLong *xShapeInfo,
+               void *dX, Nd4jLong *dXShapeInfo,
+               void *y, Nd4jLong *yShapeInfo,
+               void *dy, Nd4jLong *dyShapeInfo,
+               bool descending) {
+
+    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);
+
+
+    // 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);
+
+        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);
+        }
+    }
+}
+
+void NativeOps::sortByValue(Nd4jPointer *extraPointers,
+                 void *x, Nd4jLong *xShapeInfo,
+                 void *dX, Nd4jLong *dXShapeInfo,
+                 void *y, Nd4jLong *yShapeInfo,
+                 void *dy, Nd4jLong *dyShapeInfo,
+                 bool descending) {
+    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);
+
+
+    // 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);
+
+        for (int k = 2; k <= xLength; k = 2*k) {
+            for (int j = k >> 1; j > 0; j = j >> 1) {
+                BUILD_DOUBLE_SELECTOR(xType, yType, bitonicSortStepGenericValue, (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, bitonicArbitraryStepGenericValue, (launchDims, stream, dX, dXShapeInfo, dy, dyShapeInfo, n, xLength, rev, descending), LIBND4J_TYPES, LIBND4J_TYPES);
+                n>>=1;
+                rev = 1;
+            } while(n > 1);
+        }
+    }
+}
+
+
+
+void NativeOps::sortTadByKey(Nd4jPointer *extraPointers,
+                             void *x, Nd4jLong *xShapeInfo,
+                             void *dX, Nd4jLong *dXShapeInfo,
+                             void *y, Nd4jLong *yShapeInfo,
+                             void *dy, Nd4jLong *dyShapeInfo,
+                             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);
+
+    nd4j::DebugHelper::checkErrorCode(stream, "sortTadKey(...) failed");
+}
+
+void NativeOps::sortTadByValue(Nd4jPointer *extraPointers,
+                               void *x, Nd4jLong *xShapeInfo,
+                               void *dX, Nd4jLong *dXShapeInfo,
+                               void *y, Nd4jLong *yShapeInfo,
+                               void *dy, Nd4jLong *dyShapeInfo,
+                               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);
+
+    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");
+}
+
+
 void NativeOps::sortTad(Nd4jPointer *extraPointers,
 						void *x, Nd4jLong *xShapeInfo,
 						void *dX, Nd4jLong *dXShapeInfo,
@@ -2331,15 +2488,13 @@ void NativeOps::sortTad(Nd4jPointer *extraPointers,
 						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);
-    dim3 launchDims(tadPack.numberOfTads(), 1024, 33768);
-
 	auto xType = nd4j::ArrayOptions::dataType(xShapeInfo);
-    BUILD_SINGLE_SELECTOR(xType, oesTadGeneric, (launchDims, stream, dX, dXShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, descending), LIBND4J_TYPES);
+    BUILD_SINGLE_SELECTOR(xType, oesTadGeneric, (launchDims, stream, dX, dXShapeInfo, nullptr, dimensionLength, tadShapeInfo, tadOffsets, descending), LIBND4J_TYPES);
 
-    nd4j::DebugHelper::checkErrorCode(stream, "sortTadFloat(...) failed");
+    nd4j::DebugHelper::checkErrorCode(stream, "sortTad(...) failed");
 }
 
 void NativeOps::sortCooIndices(Nd4jPointer *extraPointers, Nd4jLong *indices, void *values, Nd4jLong length, int rank) {

libnd4j/include/array/ConstantDataBuffer.h

@@ -38,11 +38,11 @@ namespace nd4j {
         ConstantDataBuffer() = default;
         ~ConstantDataBuffer() = default;
 
-        Nd4jLong sizeOf();
+        Nd4jLong sizeOf() const;
-        Nd4jLong length();
+        Nd4jLong length() const;
 
-        Nd4jPointer primary();
+        Nd4jPointer primary() const;
-        Nd4jPointer special();
+        Nd4jPointer special() const;
 
         ConstantDataBuffer& operator=(const ConstantDataBuffer& other) = default;
         ConstantDataBuffer& operator=(ConstantDataBuffer&& other) noexcept = default;

libnd4j/include/array/DataBuffer.h

@@ -261,6 +261,8 @@ DataBuffer& DataBuffer::operator=(const DataBuffer& other) {
 
     allocateBuffers();
     copyBufferFrom(other);
+
+    return *this;
 }
 
 ////////////////////////////////////////////////////////////////////////
@@ -285,6 +287,8 @@ DataBuffer& DataBuffer::operator=(DataBuffer&& other) noexcept {
     other._primaryBuffer = other._specialBuffer = nullptr;
     other.setAllocFlags(false, false);
     other._lenInBytes = 0;
+
+    return *this;
 }
 
 ////////////////////////////////////////////////////////////////////////

libnd4j/include/array/DataTypeUtils.h

@@ -335,6 +335,8 @@ FORCEINLINE std::string DataTypeUtils::asString(DataType dataType) {
             return std::string("INT8");
         case INT16:
             return std::string("INT16");
+        case UINT16:
+            return std::string("UINT16");
         case INT32:
             return std::string("INT32");
         case INT64:
@@ -375,7 +377,7 @@ FORCEINLINE bool DataTypeUtils::castShapeInfo(const Nd4jLong *originalShapeInfo,
 ///////////////////////////////////////////////////////////////////
 // returns the difference between 1.0 and the next representable value of the given floating-point type
 template <typename T>
-FORCEINLINE T DataTypeUtils::eps() {
+FORCEINLINE _CUDA_HD T DataTypeUtils::eps() {
         if (std::is_same<T, double>::value)
             return std::numeric_limits<double>::epsilon();
         else if (std::is_same<T, float>::value)

libnd4j/include/array/ExtraArguments.h

@@ -26,6 +26,7 @@
 #include <vector>
 #include <array/DataType.h>
 #include <pointercast.h>
+#include <stdlib.h>
 
 namespace nd4j {
     class ND4J_EXPORT ExtraArguments {

libnd4j/include/array/TadPack.h

@@ -35,21 +35,21 @@ namespace nd4j {
         TadPack() = default;
         ~TadPack() = default;
 
-        Nd4jLong* primaryShapeInfo();
+        Nd4jLong* primaryShapeInfo() const;
-        Nd4jLong* primaryOffsets();
+        Nd4jLong* primaryOffsets() const;
 
-        Nd4jLong* specialShapeInfo();
+        Nd4jLong* specialShapeInfo() const;
-        Nd4jLong* specialOffsets();
+        Nd4jLong* specialOffsets() const;
 
-        Nd4jLong numberOfTads();
+        Nd4jLong numberOfTads() const;
-        int shapeInfoLength();
+        int shapeInfoLength() const;
 
         /**
          * These methods return either primary or special pointers depending on platform binaries were compiled for
          * @return
          */
-        Nd4jLong *platformShapeInfo();
+        Nd4jLong *platformShapeInfo() const;
-        Nd4jLong *platformOffsets();
+        Nd4jLong *platformOffsets() const;
     };
 }
 

libnd4j/include/array/impl/ConstantDataBuffer.cpp

@@ -28,19 +28,19 @@ namespace nd4j {
         _sizeOf = sizeOf;
     }
 
-    Nd4jPointer ConstantDataBuffer::primary() {
+    Nd4jPointer ConstantDataBuffer::primary() const {
         return _primaryBuffer;
     }
 
-    Nd4jPointer ConstantDataBuffer::special() {
+    Nd4jPointer ConstantDataBuffer::special() const {
         return _specialBuffer;
     }
 
-    Nd4jLong ConstantDataBuffer::sizeOf() {
+    Nd4jLong ConstantDataBuffer::sizeOf() const {
         return _sizeOf;
     }
 
-    Nd4jLong ConstantDataBuffer::length() {
+    Nd4jLong ConstantDataBuffer::length() const {
         return _length;
     }
 

libnd4j/include/array/impl/NDArrayList.cpp

@@ -54,7 +54,7 @@ namespace nd4j {
     NDArray* NDArrayList::readRaw(int idx) {
         if (_chunks.count(idx) < 1) {
             nd4j_printf("Non-existent chunk requested: [%i]\n", idx);
-            throw std::runtime_error("Bad index");
+            throw std::invalid_argument("Bad index");
         }
 
         return _chunks[idx];
@@ -120,7 +120,7 @@ namespace nd4j {
         // storing reference
         _chunks[idx] = array;
 
-        return ND4J_STATUS_OK;
+        return Status::OK();
     }
 
     std::vector<Nd4jLong>& NDArrayList::shape() {
@@ -152,8 +152,10 @@ namespace nd4j {
         std::vector<bool> bargs;
         int numElements = _elements.load();
 
-        for (int e = 0; e < numElements; e++)
+        for (int e = 0; e < numElements; e++) {
+            _chunks[e]->syncToDevice();
             inputs.emplace_back(_chunks[e]);
+        }
 
         iargs.push_back(_axis);
 

libnd4j/include/array/impl/TadPack.cpp

@@ -29,34 +29,34 @@ namespace nd4j {
         _numTads = numTads;
     }
 
-    Nd4jLong* TadPack::primaryShapeInfo() {
+    Nd4jLong* TadPack::primaryShapeInfo() const {
         return reinterpret_cast<Nd4jLong *>(_tadShape.primary());
     }
-    Nd4jLong* TadPack::primaryOffsets() {
+    Nd4jLong* TadPack::primaryOffsets() const {
         return reinterpret_cast<Nd4jLong *>(_tadOffsets.primary());
     }
 
-    Nd4jLong* TadPack::specialShapeInfo() {
+    Nd4jLong* TadPack::specialShapeInfo() const {
         return reinterpret_cast<Nd4jLong *>(_tadShape.special());
     }
 
-    Nd4jLong* TadPack::specialOffsets() {
+    Nd4jLong* TadPack::specialOffsets() const {
         return reinterpret_cast<Nd4jLong *>(_tadOffsets.special());
     }
 
-    Nd4jLong TadPack::numberOfTads() {
+    Nd4jLong TadPack::numberOfTads() const {
         return _numTads;
     }
 
-    Nd4jLong* TadPack::platformShapeInfo() {
+    Nd4jLong* TadPack::platformShapeInfo() const {
         return nd4j::Environment::getInstance()->isCPU() ? primaryShapeInfo() : specialShapeInfo();
     }
 
-    Nd4jLong* TadPack::platformOffsets() {
+    Nd4jLong* TadPack::platformOffsets() const {
         return nd4j::Environment::getInstance()->isCPU() ? primaryOffsets() : specialOffsets();
     }
 
-    int TadPack::shapeInfoLength() {
+    int TadPack::shapeInfoLength() const {
         return (int) shape::shapeInfoLength(primaryShapeInfo());
     }
 }

libnd4j/include/helpers/AttentionHelper.h

@@ -27,7 +27,7 @@ namespace nd4j {
     class AttentionHelper {
 
     public:
-        static nd4j::NDArray* multiHeadProject(const nd4j::NDArray* input, const nd4j::NDArray* projectionMatrix, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
+        static nd4j::NDArray multiHeadProject(const nd4j::NDArray* input, const nd4j::NDArray* projectionMatrix, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
         static void multiHeadProjectBp(const nd4j::NDArray* input, const nd4j::NDArray* projectionMatrix, const nd4j::NDArray* eps, nd4j::NDArray* dLdInput, nd4j::NDArray* dLdProjectionMatrix, nd4j::LaunchContext * context = nd4j::LaunchContext ::defaultContext());
     };
 }

libnd4j/include/helpers/benchmark/MatrixBenchmark.h

@@ -69,10 +69,10 @@ namespace nd4j {
         }
 
         void executeOnce() override {
-            auto xT = (_tA ? _x->transpose() : _x);
+            auto xT = (_tA ? _x->transpose() : *_x);
-            auto yT = (_tB ? _y->transpose() : _y);
+            auto yT = (_tB ? _y->transpose() : *_y);
 
-            MmulHelper::mmul(xT, yT, _z, _alpha, _beta);
+            MmulHelper::mmul(&xT, &yT, _z, _alpha, _beta);
         }
 
         std::string axis() override {

libnd4j/include/helpers/cpu/householder.cpp

@@ -133,9 +133,9 @@ void Householder<T>::mulLeft(NDArray& matrix, const NDArray& tail, const T coeff
 	// if(matrix.rankOf() != 2)
 	// 	throw "ops::helpers::Householder::mulLeft method: input array must be 2D matrix !";
 
-	if(matrix.sizeAt(0) == 1)   
+	if(matrix.sizeAt(0) == 1) {
-    	matrix *= (T)1.f - coeff;
+        matrix *= (T) 1.f - coeff;
-  	
+    }
     else if(coeff != (T)0.f) {
 
   		auto bottomPart = new NDArray(matrix({1,matrix.sizeAt(0), 0,0}, true));
@@ -145,13 +145,11 @@ void Householder<T>::mulLeft(NDArray& matrix, const NDArray& tail, const T coeff
 
 			auto column = tail;
 			auto row = tail.transpose();
-    		auto resultingRow = mmul(*row, bottomPartCopy);
+    		auto resultingRow = mmul(row, bottomPartCopy);
     		auto fistRow = matrix({0,1, 0,0}, true);
     		resultingRow += fistRow;
     		fistRow -= resultingRow * coeff;
     		*bottomPart -= mmul(column, resultingRow) * coeff;
-
-			delete row;
 		}
 		else {
 
@@ -161,9 +159,7 @@ void Householder<T>::mulLeft(NDArray& matrix, const NDArray& tail, const T coeff
     		auto fistRow = matrix({0,1, 0,0}, true);
     		resultingRow += fistRow;
     		fistRow -= resultingRow * coeff;
-    		*bottomPart -= mmul(*column, resultingRow) * coeff;    	
+    		*bottomPart -= mmul(column, resultingRow) * coeff;
-
-			delete column;
 		}
 		delete bottomPart;
 	}
@@ -193,21 +189,16 @@ void Householder<T>::mulRight(NDArray& matrix, const NDArray& tail, const T coef
     		auto resultingCol = mmul(rightPartCopy, column);
     		resultingCol += *fistCol;
     		*fistCol -= resultingCol * coeff;
-    		*rightPart -= mmul(resultingCol, *row) * coeff;    		
+    		*rightPart -= mmul(resultingCol, row) * coeff;
-
-			delete row;			
 		}
 		else {
 
 			auto row = tail;
 			auto column = tail.transpose();
-    		auto resultingCol = mmul(rightPartCopy, *column);
+    		auto resultingCol = mmul(rightPartCopy, column);
     		resultingCol += *fistCol;
     		*fistCol -= resultingCol * coeff;
     		*rightPart -= mmul(resultingCol, row) * coeff;
-
-			delete column;
-			
 		}
   		delete rightPart;
   		delete fistCol;

libnd4j/include/helpers/cpu/jacobiSVD.cpp

@@ -157,8 +157,7 @@ bool JacobiSVD<T>::isBlock2x2NotDiag(NDArray& block, int p, int q, T& maxElem) {
 
         if(_calcU) {
             auto temp2 = rotation.transpose();
-            mulRotationOnRight(p, q, _u, *temp2);
+            mulRotationOnRight(p, q, _u, temp2);
-            delete temp2;
         }
     }
 
@@ -251,9 +250,7 @@ void JacobiSVD<T>::svd2x2(const NDArray& block, int p, int q, NDArray& left, NDA
     m.p<T>(1, 1, _z);
 
     auto temp = right.transpose();
-    left.assign(mmul(rotation, *temp));
+    left.assign(mmul(rotation, temp));
-    delete temp;
-
 }
 
 
@@ -289,7 +286,7 @@ void JacobiSVD<T>::evalData(const NDArray& matrix) {
     else if(_rows < _cols) {
 
         auto matrixT = matrix.transpose();
-        HHcolPivQR qr(*matrixT / scale);
+        HHcolPivQR qr(matrixT / scale);
         _m.assign(qr._qr({0,_rows, 0,_rows}));
         _m.fillAsTriangular<T>(0., 0, 0, 'l');
         _m.transposei();
@@ -305,8 +302,6 @@ void JacobiSVD<T>::evalData(const NDArray& matrix) {
 
         if(_calcU)
             _u.assign(qr._permut);
-
-        delete matrixT;
     }
     else {
 
@@ -352,8 +347,7 @@ void JacobiSVD<T>::evalData(const NDArray& matrix) {
 
                         if(_calcU) {
                             auto temp = rotLeft.transpose();
-                            mulRotationOnRight(p, q, _u, *temp);
+                            mulRotationOnRight(p, q, _u, temp);
-                            delete temp;
                         }
 
                         mulRotationOnRight(p, q, _m, rotRight);

libnd4j/include/helpers/cpu/svd.cpp

@@ -920,7 +920,7 @@ void SVD<T>::evalData(const NDArray& matrix) {
     auto temp1 = biDiag._HHbidiag.transpose();
     auto temp2 = _m({0,_diagSize, 0,0}, true);
     temp2.assign(temp1);
-    delete temp1;   
+
 
     auto temp3 = _m({_m.sizeAt(0)-1,_m.sizeAt(0), 0,0}, true);
     temp3.assign(0.);

libnd4j/include/helpers/cuda_off/MmulHelper.cu

@@ -184,9 +184,9 @@ NDArray* MmulHelper::mmulMxM(const NDArray* A, const NDArray* B, NDArray* C, dou
 
     if(pC->ordering() != 'f') {
         auto temp = pA;
-        pA = pB  ->permute({1,0});
+        pA = new NDArray(pB  ->permute({1,0}));
-        pB = temp->permute({1,0});
+        pB = new NDArray(temp->permute({1,0}));
-        pC = pC  ->permute({1,0});
+        pC = new NDArray(pC  ->permute({1,0}));
         toDelete.push_back(pA);
         toDelete.push_back(pB);
         toDelete.push_back(pC);
@@ -251,7 +251,8 @@ NDArray* MmulHelper::mmulMxM(const NDArray* A, const NDArray* B, NDArray* C, dou
             blocksPerGrid.y = math::nd4j_ceil<double, int>(static_cast<double>(M) / threadsPerBlock.y);    // rows
         }
 
-        BUILD_TRIPLE_SELECTOR(aType, bType, cType, usualGemm, (blocksPerGrid, threadsPerBlock, stream, transA, transB, M, N, K, alpha, pA->getSpecialBuffer(), lda, pB->getSpecialBuffer(), ldb, beta, pC->getSpecialBuffer(), ldc), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
+        //BUILD_TRIPLE_SELECTOR(aType, bType, cType, usualGemm, (blocksPerGrid, threadsPerBlock, stream, transA, transB, M, N, K, alpha, pA->getSpecialBuffer(), lda, pB->getSpecialBuffer(), ldb, beta, pC->getSpecialBuffer(), ldc), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
+        BUILD_SINGLE_SELECTOR_THRICE(aType, usualGemm, (blocksPerGrid, threadsPerBlock, stream, transA, transB, M, N, K, alpha, pA->getSpecialBuffer(), lda, pB->getSpecialBuffer(), ldb, beta, pC->getSpecialBuffer(), ldc), LIBND4J_TYPES)
     }
 
     if (status != CUBLAS_STATUS_SUCCESS) throw cuda_exception::build("MmulHelper::mmulMxM cuda failed !", status);
@@ -339,7 +340,8 @@ NDArray* MmulHelper::mmulMxV(const NDArray* A, const NDArray* X, nd4j::NDArray*
             threadsPerBlock.x = 512;
             blocksPerGrid.x = math::nd4j_ceil<double, int>(static_cast<double>(M) / threadsPerBlock.x);    // rows
         }
-        BUILD_TRIPLE_SELECTOR(aType, xType, yType, usualGemv, (blocksPerGrid, threadsPerBlock, stream, transA, M, N, alpha, pA->getSpecialBuffer(), lda, X->getSpecialBuffer(), incx, beta, Y->getSpecialBuffer(), incy), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
+        //BUILD_TRIPLE_SELECTOR(aType, xType, yType, usualGemv, (blocksPerGrid, threadsPerBlock, stream, transA, M, N, alpha, pA->getSpecialBuffer(), lda, X->getSpecialBuffer(), incx, beta, Y->getSpecialBuffer(), incy), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
+        BUILD_SINGLE_SELECTOR_THRICE(xType, usualGemv, (blocksPerGrid, threadsPerBlock, stream, transA, M, N, alpha, pA->getSpecialBuffer(), lda, X->getSpecialBuffer(), incx, beta, Y->getSpecialBuffer(), incy), LIBND4J_TYPES)
     }
 
     if (status != CUBLAS_STATUS_SUCCESS) throw cuda_exception::build("MmulHelper::mmulMxV cuda failed !", status);
@@ -396,7 +398,8 @@ NDArray* MmulHelper::dot(const NDArray* X, const NDArray* Y, nd4j::NDArray* Z, c
 
     NDArray::prepareSpecialUse({Z}, {X, Y});
 
-    BUILD_TRIPLE_SELECTOR(xType, yType, zType, usualDot, (blocksPerGrid, threadsPerBlock, stream, length, alpha, X->getSpecialBuffer(), incx, Y->getSpecialBuffer(), incy, beta, Z->getSpecialBuffer()), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
+    //BUILD_TRIPLE_SELECTOR(xType, yType, zType, usualDot, (blocksPerGrid, threadsPerBlock, stream, length, alpha, X->getSpecialBuffer(), incx, Y->getSpecialBuffer(), incy, beta, Z->getSpecialBuffer()), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
+    BUILD_SINGLE_SELECTOR_THRICE(xType, usualDot, (blocksPerGrid, threadsPerBlock, stream, length, alpha, X->getSpecialBuffer(), incx, Y->getSpecialBuffer(), incy, beta, Z->getSpecialBuffer()), LIBND4J_TYPES)
 
     auto cudaResult = cudaStreamSynchronize(*stream);
     if (cudaResult != 0) throw cuda_exception::build("MmulHelper::dot cuda failed !", cudaResult);
@@ -406,8 +409,8 @@ NDArray* MmulHelper::dot(const NDArray* X, const NDArray* Y, nd4j::NDArray* Z, c
     return Z;
 }
 
-BUILD_TRIPLE_TEMPLATE(template void usualGemm, (const dim3 &blocksPerGrid, const dim3 &threadsPerBlock, cudaStream_t *stream, const bool transA, const bool transB, const int M, const int N, const int K, const double alpha, const void* vA, const int lda, const void* vB, const int ldb, const double beta, void* vC, const int ldc), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
+//BUILD_TRIPLE_TEMPLATE(template void usualGemm, (const dim3 &blocksPerGrid, const dim3 &threadsPerBlock, cudaStream_t *stream, const bool transA, const bool transB, const int M, const int N, const int K, const double alpha, const void* vA, const int lda, const void* vB, const int ldb, const double beta, void* vC, const int ldc), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
-BUILD_TRIPLE_TEMPLATE(template void usualGemv, (const dim3 &blocksPerGrid, const dim3 &threadsPerBlock, cudaStream_t *stream, const bool transA, const int M, const int N, const double alpha, const void* vA, const int lda, const void* vB, const int incx, const double beta, void* vC, const int incy), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
+//BUILD_TRIPLE_TEMPLATE(template void usualGemv, (const dim3 &blocksPerGrid, const dim3 &threadsPerBlock, cudaStream_t *stream, const bool transA, const int M, const int N, const double alpha, const void* vA, const int lda, const void* vB, const int incx, const double beta, void* vC, const int incy), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
-BUILD_TRIPLE_TEMPLATE(template void usualDot,  (const dim3 &blocksPerGrid, const dim3 &threadsPerBlock, cudaStream_t *stream, const Nd4jLong length, const double alpha, const void* vX, const Nd4jLong incx, const void* vY, const Nd4jLong incy, const double beta, void* vZ), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
+//BUILD_TRIPLE_TEMPLATE(template void usualDot,  (const dim3 &blocksPerGrid, const dim3 &threadsPerBlock, cudaStream_t *stream, const Nd4jLong length, const double alpha, const void* vX, const Nd4jLong incx, const void* vY, const Nd4jLong incy, const double beta, void* vZ), LIBND4J_TYPES, FLOAT_TYPES, FLOAT_TYPES);
 
 }

libnd4j/include/helpers/impl/AttentionHelper.cpp

@@ -28,33 +28,27 @@
 
 namespace nd4j {
 
-    nd4j::NDArray *
+    nd4j::NDArray AttentionHelper::multiHeadProject(const nd4j::NDArray *input, const nd4j::NDArray *projectionMatrix, nd4j::LaunchContext * context) {
-    AttentionHelper::multiHeadProject(const nd4j::NDArray *input, const nd4j::NDArray *projectionMatrix, nd4j::LaunchContext * context) {
         auto miniBatchSize = input->sizeAt(0);
         auto seqLength = input->sizeAt(2);
         auto numHeads = projectionMatrix->sizeAt(0);
         auto projectedSize = projectionMatrix->sizeAt(1);
 
         auto inputPerm = input->permute({1, 0, 2});
-        auto inputPrep = inputPerm->reshape('c', {input->sizeAt(1), (miniBatchSize * seqLength)});
+        auto inputPrep = inputPerm.reshape('c', {input->sizeAt(1), (miniBatchSize * seqLength)});
         auto projectionPrep = projectionMatrix->reshape('c', {numHeads * projectionMatrix->sizeAt(1), projectionMatrix->sizeAt(2)});
 
-        NDArray* projected = new NDArray('c', {numHeads * projectionMatrix->sizeAt(1), (miniBatchSize * seqLength)}, input->dataType(), context);
+        NDArray projected('c', {numHeads * projectionMatrix->sizeAt(1), (miniBatchSize * seqLength)}, input->dataType(), context);
         nd4j::ops::matmul mmul;
-        mmul.execute({projectionPrep, inputPrep}, {projected},  {}, {}, {});
+        mmul.execute({&projectionPrep, &inputPrep}, {&projected},  {}, {}, {});
 
-        projected->reshapei({numHeads, projectedSize, miniBatchSize, seqLength});
+        projected.reshapei({numHeads, projectedSize, miniBatchSize, seqLength});
-        projected->permutei({2, 0, 1, 3});
+        projected.permutei({2, 0, 1, 3});
-
-        delete inputPerm;
-        delete inputPrep;
-        delete projectionPrep;
 
         return projected;
     }
 
-    void
+    void AttentionHelper::multiHeadProjectBp(const nd4j::NDArray *input, const nd4j::NDArray *projectionMatrix,
-    AttentionHelper::multiHeadProjectBp(const nd4j::NDArray *input, const nd4j::NDArray *projectionMatrix,
                                         const nd4j::NDArray *eps, nd4j::NDArray *dLdInput,
                                         nd4j::NDArray *dLdProjectionMatrix, nd4j::LaunchContext * context) {
         auto miniBatchSize = input->sizeAt(0);
@@ -63,16 +57,16 @@ namespace nd4j {
         auto projectedSize = projectionMatrix->sizeAt(1);
 
         auto epsPerm = eps->permute({1, 2, 0, 3});
-        auto epsReshaped = epsPerm->reshape('c', {numHeads * projectedSize, miniBatchSize * seqLength});
+        auto epsReshaped = epsPerm.reshape('c', {numHeads * projectedSize, miniBatchSize * seqLength});
 
         auto inputPerm = input->permute({1, 0, 2});
-        auto inputPrep = inputPerm->reshape('c', {input->sizeAt(1), (miniBatchSize * seqLength)});
+        auto inputPrep = inputPerm.reshape('c', {input->sizeAt(1), (miniBatchSize * seqLength)});
         auto projectionPrep = projectionMatrix->reshape('c', {numHeads * projectionMatrix->sizeAt(1), projectionMatrix->sizeAt(2)});
 
         nd4j::ops::matmul_bp mmulBp;
-        NDArray dLdProjectionPrep(projectionPrep->shapeInfo(), false, context);
+        NDArray dLdProjectionPrep(projectionPrep.shapeInfo(), false, context);
-        NDArray dLdInputPrep(inputPrep->shapeInfo(), false, context);
+        NDArray dLdInputPrep(inputPrep.shapeInfo(), false, context);
-        mmulBp.execute({projectionPrep, inputPrep, epsReshaped}, {&dLdProjectionPrep, &dLdInputPrep}, {}, {}, {});
+        mmulBp.execute({&projectionPrep, &inputPrep, &epsReshaped}, {&dLdProjectionPrep, &dLdInputPrep}, {}, {}, {});
 
         dLdProjectionPrep.reshapei({numHeads, projectionMatrix->sizeAt(1), projectionMatrix->sizeAt(2)});
         dLdProjectionMatrix->assign(dLdProjectionPrep);
@@ -80,12 +74,6 @@ namespace nd4j {
         dLdInputPrep.reshapei({input->sizeAt(1), miniBatchSize, seqLength});
         dLdInputPrep.permutei({1, 0, 2});
         dLdInput->assign(dLdInputPrep);
-
-        delete inputPerm;
-        delete inputPrep;
-        delete epsPerm;
-        delete epsReshaped;
-        delete projectionPrep;
     }
 }
 

libnd4j/include/helpers/impl/GradCheck.cpp

@@ -53,7 +53,7 @@ void GradCheck::fillGradArrays(const LossFunc loss, const std::vector<NDArray*>&
 bool GradCheck::checkGrad(ops::DeclarableOp& opFF, ops::DeclarableOp& opBP, const OpArgsHolder& argsHolderFF, const OpArgsHolder& argsHolderBP,
 	                      const std::vector<bool>& whatArrsToCheck, const std::vector<double>& idxRange, const LossFunc loss ) {
 
-	const int numInArrsFF     = argsHolderFF.getNumInArrs();						// also numInArrsFF = number of output arrays in opBP
+	const int numInArrsFF     = argsHolderFF.getNumInArrs();						// at the same time numInArrsFF = number of output arrays in opBP
 	const int numInGradArrsBP = argsHolderBP.getNumInArrs() - numInArrsFF;  		// because argsHolderBP.getNumInArrs() = numInArrsFF + numInGradArrsBP
 	const std::vector<NDArray*>& inArrsFF = argsHolderFF.getInArrs();
 	const std::vector<NDArray*>& inArrsBP = argsHolderBP.getInArrs();
@@ -65,6 +65,7 @@ bool GradCheck::checkGrad(ops::DeclarableOp& opFF, ops::DeclarableOp& opBP, cons
 	ResultSet* outArrsBP = opBP.execute(argsHolderBP);		// number of output arrays in back prop = numInArrsFF;
 
 	NDArray tmpScalar(nd4j::DataType::DOUBLE, inArrsFF[0]->getContext()); // scalar = 0
+
 	for(int i = 0; i < numInArrsFF; ++i) {							// loop through input array
 
 		if(!whatArrsToCheck.empty() && static_cast<bool>(whatArrsToCheck[i]) == false)
@@ -75,39 +76,39 @@ bool GradCheck::checkGrad(ops::DeclarableOp& opFF, ops::DeclarableOp& opBP, cons
 
 		for(Nd4jLong j = idxStart; j < idxEnd; ++j) {			// loop through all elements for current array
 
-			double& elem = inArrsFF[i]->t<double>(j);
+			const double orig = inArrsFF[i]->e<double>(j);
-			const double orig = elem;
 
 			// add epsilon, feed forward
-			elem = orig + EPSILON;
+			inArrsFF[i]->p<double>(j, orig + EPSILON);
 			ResultSet* outArrsFF = opFF.execute(argsHolderFF);
 			int numOutArrs = outArrsFF->size();
 			double scorePlus = 0.;
-			for(int k = 0; k < numOutArrs; ++k) {                // loop through output array
+
+			for(int k = 0; k < numOutArrs; ++k) {                // loop through output arrays
 				if(loss == SUM)
-					NativeOpExecutioner::execReduceSameScalar(LaunchContext::defaultContext(), reduce::Sum, outArrsFF->at(k)->getBuffer(), outArrsFF->at(k)->getShapeInfo(), outArrsFF->at(k)->getSpecialBuffer(), outArrsFF->at(k)->getSpecialShapeInfo(), nullptr, tmpScalar.buffer(), tmpScalar.shapeInfo(), tmpScalar.specialBuffer(), tmpScalar.specialShapeInfo());
+					outArrsFF->at(k)->reduceNumber(reduce::Sum, tmpScalar);
 				else
-					NativeOpExecutioner::execReduceFloatScalar(LaunchContext::defaultContext(), reduce::Mean, outArrsFF->at(k)->getBuffer(), outArrsFF->at(k)->getShapeInfo(), outArrsFF->at(k)->getSpecialBuffer(), outArrsFF->at(k)->getSpecialShapeInfo(), nullptr, tmpScalar.buffer(), tmpScalar.shapeInfo(), tmpScalar.specialBuffer(), tmpScalar.specialShapeInfo());
+					outArrsFF->at(k)->reduceNumber(reduce::Mean, tmpScalar);
 				scorePlus += tmpScalar.e<double>(0);
 			}
 			delete outArrsFF;
 
 			// subtract epsilon, feed forward
-			elem = orig - EPSILON;
+			inArrsFF[i]->p<double>(j, orig - EPSILON);
 			outArrsFF = opFF.execute(argsHolderFF);
 			double scoreMinus = 0.;
 
-			for(int k = 0; k < numOutArrs; ++k) {            // loop through output array
+			for(int k = 0; k < numOutArrs; ++k) {            // loop through output arrays
 				if(loss == SUM)
-					NativeOpExecutioner::execReduceSameScalar(LaunchContext::defaultContext(), reduce::Sum, outArrsFF->at(k)->getBuffer(), outArrsFF->at(k)->getShapeInfo(), outArrsFF->at(k)->getSpecialBuffer(), outArrsFF->at(k)->getSpecialShapeInfo(), nullptr, tmpScalar.buffer(), tmpScalar.shapeInfo(), tmpScalar.specialBuffer(), tmpScalar.specialShapeInfo());
+					outArrsFF->at(k)->reduceNumber(reduce::Sum, tmpScalar);
 				else
-					NativeOpExecutioner::execReduceFloatScalar(LaunchContext::defaultContext(), reduce::Mean, outArrsFF->at(k)->getBuffer(), outArrsFF->at(k)->getShapeInfo(), outArrsFF->at(k)->getSpecialBuffer(), outArrsFF->at(k)->getSpecialShapeInfo(), nullptr, tmpScalar.buffer(), tmpScalar.shapeInfo(), tmpScalar.specialBuffer(), tmpScalar.specialShapeInfo());
+					outArrsFF->at(k)->reduceNumber(reduce::Mean, tmpScalar);
 				scoreMinus += tmpScalar.e<double>(0);
 			}
 			delete outArrsFF;
 
 			// restore initial element value
-			elem = orig;
+			inArrsFF[i]->p<double>(j, orig);
 
 			// calculate numerical gradient
 			const double numericalGrad = (scorePlus - scoreMinus) / (2 * EPSILON);

libnd4j/include/helpers/impl/MmulHelper.cpp

@@ -43,22 +43,19 @@ nd4j::NDArray* nd4j::MmulHelper::tensorDot(const nd4j::NDArray* a, const nd4j::N
 
     auto outShape = ShapeUtils::evalShapeForTensorDot(a, b, axes_0, axes_1, permutAt, permutBt, shapeAt, shapeBt);
 
-    NDArray* aPR = a->permute(permutAt);
+    NDArray aPR = a->permute(permutAt);
-    NDArray* bPR = b->permute(permutBt);
+    NDArray bPR = b->permute(permutBt);
 
     // check whether reshape is necessary
-    if(!aPR->isSameShape(shapeAt))
+    if(!aPR.isSameShape(shapeAt))
-        aPR->reshapei( shapeAt);
+        aPR.reshapei( shapeAt);
-    if(!bPR->isSameShape(shapeBt))
+    if(!bPR.isSameShape(shapeBt))
-        bPR->reshapei( shapeBt);
+        bPR.reshapei( shapeBt);
 
-    NDArray* c = mmul(aPR, bPR, nullptr, 1.0, 0.0);
+    NDArray* c = mmul(&aPR, &bPR, nullptr, 1.0, 0.0);
 
     c->reshapei(outShape);
 
-    delete aPR;
-    delete bPR;
-
     return c;
 }
 
@@ -74,21 +71,21 @@ void nd4j::MmulHelper::tensorDot(const nd4j::NDArray* a, const nd4j::NDArray* b,
 
     // check whether permutation is required
     if(!permutForC.empty())
-        cP = c->permute(permutForC);
+        cP = new NDArray(c->permute(permutForC));
 
     auto aPR = a->permute(permutAt);
     auto bPR = b->permute(permutBt);
 
     // check whether reshape is necessary
-    if(!aPR->isSameShape(shapeAt))
+    if(!aPR.isSameShape(shapeAt))
-            aPR->reshapei(shapeAt);
+            aPR.reshapei(shapeAt);
-    if(!bPR->isSameShape(shapeBt))
+    if(!bPR.isSameShape(shapeBt))
-            bPR->reshapei(shapeBt);
+            bPR.reshapei(shapeBt);
 
-    if(!cP->isSameShape({aPR->sizeAt(0), bPR->sizeAt(1)}))
+    if(!cP->isSameShape({aPR.sizeAt(0), bPR.sizeAt(1)}))
-        cPR = cP->reshape(cP->ordering(), {aPR->sizeAt(0), bPR->sizeAt(1)});
+        cPR = new NDArray(cP->reshape(cP->ordering(), {aPR.sizeAt(0), bPR.sizeAt(1)}));
 
-    mmul(aPR, bPR, cPR, 1.0, 0.0);
+    mmul(&aPR, &bPR, cPR, 1.0, 0.0);
 
     if(cPR->getBuffer() != cP->getBuffer() || cPR->getSpecialBuffer() != cP->getSpecialBuffer() )   // this means both permute and reshape have been performed on c, cP always points on c->getBuffer()
         cP->assign(cPR);
@@ -97,40 +94,42 @@ void nd4j::MmulHelper::tensorDot(const nd4j::NDArray* a, const nd4j::NDArray* b,
         delete cPR;
     if(cP != c)
         delete cP;
-    delete aPR;
-    delete bPR;
 }
 
 
 #ifndef __JAVACPP_HACK__
 //////////////////////////////////////////////////////////////////////////
 void nd4j::MmulHelper::tensorDot(const NDArray* a, const NDArray* b, NDArray* c, const std::vector<std::vector<Nd4jLong>>& modifA, const std::vector<std::vector<Nd4jLong>>& modifB, const std::vector<std::vector<Nd4jLong>>& modifC) {
+
     NDArray *aPR(const_cast<NDArray*>(a)), *bPR(const_cast<NDArray*>(b));
     std::string whatToDoWithA, whatToDoWithB, whatToDoWithC;         // "" - nothing; "p" - permutation; "r" - reshaping; "pr" - permutation+reshaping; "rp" - reshaping/permutation, and so on; if another string is produced - throw exception
+
     for(const auto& arr : modifA)
         whatToDoWithA = (std::find(arr.begin(), arr.end(), 0) != arr.end()) ? whatToDoWithA + "p" : whatToDoWithA + "r";        // when 0 is present in arr then it is permutation array, otherwise - it is reshaping array
     for(const auto& arr : modifB)
         whatToDoWithB = (std::find(arr.begin(), arr.end(), 0) != arr.end()) ? whatToDoWithB + "p" : whatToDoWithB + "r";
     for(const auto& arr : modifC)
         whatToDoWithC = (std::find(arr.begin(), arr.end(), 0) != arr.end()) ? whatToDoWithC + "p" : whatToDoWithC + "r";
+
     // first step for a array
     if(!whatToDoWithA.empty())
-        aPR = (whatToDoWithA[0] == 'p') ? a->permute(modifA[0]) : a->reshape(a->ordering(), modifA[0]);
+        aPR = (whatToDoWithA[0] == 'p') ? new NDArray(a->permute(modifA[0])) : new NDArray(a->reshape(a->ordering(), modifA[0]));
     // first step for b array
     if(!whatToDoWithB.empty())
-        bPR = (whatToDoWithB[0] == 'p') ? b->permute(modifB[0]) : b->reshape(b->ordering(), modifB[0]);
+        bPR = (whatToDoWithB[0] == 'p') ? new NDArray(b->permute(modifB[0])) : new NDArray(b->reshape(b->ordering(), modifB[0]));
     // rest steps for a array
     for(int i = 1; i < whatToDoWithA.size(); ++i)
         if(whatToDoWithA[i] == 'p') aPR->permutei(modifA[i]); else aPR->reshapei(modifA[i]);
     // rest steps for b array
     for(int i = 1; i < whatToDoWithB.size(); ++i)
         if(whatToDoWithB[i] == 'p') bPR->permutei(modifB[i]); else bPR->reshapei(modifB[i]);
+
     // now work with c array
     std::vector<NDArray*> cArrs = {c};
     if(!whatToDoWithC.empty()) {
         cArrs = std::vector<NDArray*>(whatToDoWithC.size()+1, c);
         for(int i = 0; i < cArrs.size()-1; ++i)
-            cArrs[i+1] = (whatToDoWithC[i] == 'p') ? cArrs[i]->permute(modifC[i]) : cArrs[i]->reshape(c->ordering(), modifC[i]);  // since we ignore first element in cArrs (that is cArrs[0]) then it is always equal to c
+            cArrs[i+1] = (whatToDoWithC[i] == 'p') ? new NDArray(cArrs[i]->permute(modifC[i])) : new NDArray(cArrs[i]->reshape(c->ordering(), modifC[i]));  // since we ignore first element in cArrs (that is cArrs[0]) then it is always equal to c
     }
 
     mmul(aPR, bPR, cArrs[cArrs.size()-1], 1.0, 0.0);
@@ -152,18 +151,21 @@ void nd4j::MmulHelper::tensorDot(const NDArray* a, const NDArray* b, NDArray* c,
 
 //////////////////////////////////////////////////////////////////////////
 NDArray* nd4j::MmulHelper::tensorDot(const nd4j::NDArray* a, const nd4j::NDArray* b, const std::vector<std::vector<Nd4jLong>>& modifA, const std::vector<std::vector<Nd4jLong>>& modifB) {
+
     NDArray *aPR(const_cast<NDArray*>(a)), *bPR(const_cast<NDArray*>(b));
     std::string whatToDoWithA, whatToDoWithB;         // "" - nothing; "p" - permutation only; "r" - reshaping only; "pr" - permutation+reshaping; "rp" - reshaping/permutation; another string - throw exception
+
     for(const auto& arr : modifA)
         whatToDoWithA = (std::find(arr.begin(), arr.end(), 0) != arr.end()) ? whatToDoWithA + "p" : whatToDoWithA + "r";        // when 0 is present in arr then it is permutation array, otherwise - it is reshaping array
     for(const auto& arr : modifB)
         whatToDoWithB = (std::find(arr.begin(), arr.end(), 0) != arr.end()) ? whatToDoWithB + "p" : whatToDoWithB + "r";
+
     // first step for a array
     if(!whatToDoWithA.empty())
-        aPR = (whatToDoWithA[0] == 'p') ? a->permute(modifA[0]) : a->reshape(a->ordering(), modifA[0]);
+        aPR = (whatToDoWithA[0] == 'p') ? new NDArray(a->permute(modifA[0])) : new NDArray(a->reshape(a->ordering(), modifA[0]));
     // first step for b array
     if(!whatToDoWithB.empty())
-        bPR = (whatToDoWithB[0] == 'p') ? b->permute(modifB[0]) : b->reshape(b->ordering(), modifB[0]);
+        bPR = (whatToDoWithB[0] == 'p') ? new NDArray(b->permute(modifB[0])) : new NDArray(b->reshape(b->ordering(), modifB[0]));
     // rest steps for a array
     for(int i = 1; i < whatToDoWithA.size(); ++i)
         if(whatToDoWithA[i] == 'p') aPR->permutei(modifA[i]); else aPR->reshapei(modifA[i]);
@@ -293,17 +295,17 @@ nd4j::NDArray* MmulHelper::mmul(const nd4j::NDArray* A, const nd4j::NDArray* B,
             permut[rank-1] = rank - 2;
 
             if(transX)
-                xT = x->permute(permut);
+                xT = new NDArray(x->permute(permut));
 
             if(transY)
-                yT = y->permute(permut);
+                yT = new NDArray(y->permute(permut));
         }
 
         if(xRank <= 2 && yRank <= 2) {  // dot (1Dx1D), vector-matrix (1Dx2D), matrix-vector (2Dx1D), matrix-matrix (2Dx2D) product cases
 
             if(xRank == 1 && yRank == 2) {   // reduce vector-matrix to matrix-matrix case
-                xT = x->reshape(x->ordering(), {1, x->lengthOf()}); // please note x is not transposed in this case (since xRank=1)
+                xT = new NDArray(x->reshape(x->ordering(), {1, x->lengthOf()})); // please note x is not transposed in this case (since xRank=1)
-                zT = z->reshape(z->ordering(), {1, z->lengthOf()});
+                zT = new NDArray(z->reshape(z->ordering(), {1, z->lengthOf()}));
             }
 
             mmul(xT, yT, zT, 1., 0.);

libnd4j/include/helpers/impl/ShapeUtils.cpp

@@ -473,19 +473,9 @@ bool ShapeUtils::evalBroadcastShapeInfo(Nd4jLong *max, Nd4jLong *min, const bool
     // FIXME: get rid of memcpy here
     memcpy(tmpShapeInfo, maxShapeInfo, shape::shapeInfoByteLength(maxRank));
     for (int i = 0; i < minRank; ++i)
-        if(maxShapeInfo[maxRank-i] < minShapeInfo[minRank-i])
+        if((maxShapeInfo[maxRank-i] != 0 && maxShapeInfo[maxRank-i] < minShapeInfo[minRank-i]) || minShapeInfo[minRank-i] == 0)
             tmpShapeInfo[maxRank - i] = minShapeInfo[minRank-i];
 
-    // nullify zero axis
-    for (int e = 0; e < maxRank; e++)
-        if (maxShapeInfo[e+1] == 0)
-            tmpShapeInfo[e+1] = 0;
-
-    int delta = maxRank - minRank;
-    for (int e = minRank - 1; e >= 0; e--)
-        if (minShapeInfo[e + 1] == 0)
-            tmpShapeInfo[e + 1 + delta] = 0;
-
     ShapeUtils::updateStridesAndType(tmpShapeInfo, DataTypeUtils::pickPairwiseResultType(maxShapeInfo, minShapeInfo), shape::order(maxShapeInfo));
 
     if (shape::isEmpty(max) || shape::isEmpty(min)) {

libnd4j/include/helpers/impl/logger.cpp

@@ -40,7 +40,7 @@ namespace nd4j {
 #ifdef __CUDACC__
     __host__
 #endif
-     void Logger::printv(const char *format, std::vector<int>& vec) {
+     void Logger::printv(const char *format, const std::vector<int>& vec) {
         printf("%s: {", format);
         for(int e = 0; e < vec.size(); e++) {
             auto v = vec[e];
@@ -55,7 +55,7 @@ namespace nd4j {
     #ifdef __CUDACC__
     __host__
 #endif
-     void Logger::printv(const char *format, std::vector<Nd4jLong>& vec) {
+     void Logger::printv(const char *format, const std::vector<Nd4jLong>& vec) {
         printf("%s: {", format);
         for(int e = 0; e < vec.size(); e++) {
             auto v = vec[e];

libnd4j/include/helpers/logger.h

@@ -55,8 +55,8 @@ namespace nd4j {
 
         static void _CUDA_H info(const char *format, ...);
 
-        static void _CUDA_H printv(const char *format, std::vector<int>& vec);
+        static void _CUDA_H printv(const char *format, const std::vector<int>& vec);
-        static void _CUDA_H printv(const char *format, std::vector<Nd4jLong>& vec);
+        static void _CUDA_H printv(const char *format, const std::vector<Nd4jLong>& vec);
     };
 
 }

libnd4j/include/helpers/shape.h

@@ -1023,23 +1023,6 @@ namespace shape {
     */
     ND4J_EXPORT _CUDA_HD void calcSubArrShapeAndOffsets(const Nd4jLong* wholeShapeInfo, const Nd4jLong numOfSubArrs, const int dimsSize, const int* dimsToExclude, Nd4jLong* subArrShapeInfo, Nd4jLong* subArrOffsets, bool keepUnitiesInShape = false);
 
-    /**
-    * insert dimension at shape[axis] position
-    * 1) for example: for given rank = 3, shape = {2,4,5}, axis = 1, dimension = 10 result is -> shape = {2,10,4,5}
-    * 2) for example: for given rank = 3, shape = {2,4,5}, axis = 3, dimension = 10 result is -> shape = {2,4,5,10}
-    * so be careful and provide shape buffer with enough (at least rank+1) length
-    * axis should be within [0, rank] range
-    */
-    ND4J_EXPORT _CUDA_HD void insertDimension(const int rank, Nd4jLong *shape, const Nd4jLong axis, const Nd4jLong dimension);
-
-    /**
-    * erase dimension at shape[axis] position
-    * 1) for example: for given rank = 3, shape = {2,4,5}, axis = 1, result is -> shape = {2,5}
-    * 2) for example: for given rank = 3, shape = {2,4,5}, axis = 2, result is -> shape = {2,4}
-    * axis should be within [0, rank-1] range
-    */
-    ND4J_EXPORT _CUDA_HD void eraseDimension(const int rank, Nd4jLong *shape, const Nd4jLong axis);
-
 
 
 
@@ -4932,21 +4915,6 @@ INLINEDEF _CUDA_HD void calcOffsets(const Nd4jLong *xShapeInfo, Nd4jLong*& xOffs
     }
 }
 
-//////////////////////////////////////////////////////////////////////
-INLINEDEF _CUDA_HD void insertDimension(const int rank, Nd4jLong *shape, const Nd4jLong axis, const Nd4jLong dimension) {
-
-    for (int i = rank; i > axis; --i)
-        shape[i] = shape[i - 1];
-
-    shape[axis] = dimension;
-}
-
-//////////////////////////////////////////////////////////////////////
-INLINEDEF _CUDA_HD void eraseDimension(const int rank, Nd4jLong *shape, const Nd4jLong axis) {
-
-    for (int i = axis; i < rank - 1; ++i)
-        shape[i] = shape[i + 1];
-}
 
 
 }

libnd4j/include/loops/cpu/reduce/reduce_bool.cpp

@@ -244,8 +244,9 @@ namespace functions {
                         auto xi = x + threadOffset;
                         auto ulen = static_cast<unsigned int>(info.getItersPerThread(threadNum));
 
-                        for (Nd4jLong i = 0; i < ulen; i++)
+                        for (Nd4jLong i = 0; i < ulen; i++) {
                             local = OpType::update(local, OpType::op(xi[i], extraParams), extraParams);
+                        }
 
                         PRAGMA_OMP_CRITICAL
                         startingVal = OpType::update(startingVal, local, extraParams);

libnd4j/include/loops/cuda/broadcasting.cu

@@ -122,7 +122,7 @@ namespace functions {
 
                 tadLength = shape::length(tadOnlyShapeInfo);
                 tadEWS = shape::elementWiseStride(tadOnlyShapeInfo);
-                numTads = shape::length(xShapeInfo) / tadLength;
+                numTads = shape::length(yShapeInfo) / tadLength;
                 xEWS = shape::elementWiseStride(xShapeInfo);
                 zEWS = shape::elementWiseStride(tadOnlyShapeInfoZ);
             }

libnd4j/include/loops/cuda/specials/bitonicArbitraryStep.cu

@@ -21,12 +21,165 @@
 
 #include <ops/specials_cuda.h>
 
+//////////////////////////////////////////////////////////////////////////
+template <typename X, typename Y>
+__global__ void bitonicArbitraryStepKernelValue(void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int window, int length,  int reverse, bool descending) {
+    auto x = static_cast<X*>(vx);
+    auto y = static_cast<Y*>(vy);
+
+    int tid = threadIdx.x + blockDim.x * blockIdx.x;
+    int half = window>>1;
+
+    __shared__ Nd4jLong xLength;
+    if (threadIdx.x == 0) {
+        xLength = shape::length(xShapeInfo);
+    }
+    __syncthreads();
+
+    //for (int i = 0; i < length; i+= window)
+    /*
+        if window == 4;
+        iterations will be: 0; 4; 8; 12; 16; 20
+        if gridDim = 3;
+        on first iteration we'll have: 0; 4; 8;
+        on second iteration we'll have: 0 + (3 * 4) = 12;  4 + (3 * 4) = 16; 8 + (3 * 4) = 20
+    */
+    int firstPosition;
+    int firstStep;
+    int secondPosition;
+    int secondStep;
+
+    int WARP_SIZE = 32;
+    int numWarps = (gridDim.x * blockDim.x) / 32;
+    int warpId = tid / WARP_SIZE;
+    int warpIdx = tid % WARP_SIZE;
+
+    if (half >= 128) {
+        firstPosition = blockIdx.x * window;
+        firstStep = gridDim.x * window;
+
+        secondPosition = threadIdx.x;
+        secondStep = blockDim.x;
+    } else if (half >= 32) {
+        firstPosition = warpId * window;
+        firstStep = numWarps * window;
+
+        secondPosition = warpIdx;
+        secondStep = WARP_SIZE;
+    } else {
+        firstPosition = tid * window;
+        firstStep = blockDim.x * gridDim.x * window;
+
+        secondPosition = 0;
+        secondStep = 1;
+    }
+
+
+    for (int i = firstPosition; i < length; i += firstStep) {
+        for (int j = secondPosition; j < half; j += secondStep) {
+            int it = (reverse) ? i + j + half : i + window - j - 1;
+            int ij = i+j;
+            if (it < length && ij < length ) {
+                int posIT = shape::getIndexOffset(it, yShapeInfo, xLength);
+                int posIJ = shape::getIndexOffset(ij, yShapeInfo, xLength);
+
+                Y v0 = y[posIJ];
+                Y v1 = y[posIT];
+
+                if(!descending == (v0 > v1)) {
+                    y[posIJ] = v1;
+                    y[posIT] = v0;
+
+                    X xtemp = x[posIJ];
+                    x[posIJ] = x[posIT];
+                    x[posIT] = xtemp;
+                }
+            }
+        }
+    }
+}
+
+//////////////////////////////////////////////////////////////////////////
+template <typename X, typename Y>
+__global__ void bitonicArbitraryStepKernelKey(void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int window, int length,  int reverse, bool descending) {
+    auto x = static_cast<X*>(vx);
+    auto y = static_cast<Y*>(vy);
+
+    int tid = threadIdx.x + blockDim.x * blockIdx.x;
+    int half = window>>1;
+
+    __shared__ Nd4jLong xLength;
+    if (threadIdx.x == 0) {
+        xLength = shape::length(xShapeInfo);
+    }
+    __syncthreads();
+
+    //for (int i = 0; i < length; i+= window)
+    /*
+        if window == 4;
+        iterations will be: 0; 4; 8; 12; 16; 20
+        if gridDim = 3;
+        on first iteration we'll have: 0; 4; 8;
+        on second iteration we'll have: 0 + (3 * 4) = 12;  4 + (3 * 4) = 16; 8 + (3 * 4) = 20
+    */
+    int firstPosition;
+    int firstStep;
+    int secondPosition;
+    int secondStep;
+
+    int WARP_SIZE = 32;
+    int numWarps = (gridDim.x * blockDim.x) / 32;
+    int warpId = tid / WARP_SIZE;
+    int warpIdx = tid % WARP_SIZE;
+
+    if (half >= 128) {
+        firstPosition = blockIdx.x * window;
+        firstStep = gridDim.x * window;
+
+        secondPosition = threadIdx.x;
+        secondStep = blockDim.x;
+    } else if (half >= 32) {
+        firstPosition = warpId * window;
+        firstStep = numWarps * window;
+
+        secondPosition = warpIdx;
+        secondStep = WARP_SIZE;
+    } else {
+        firstPosition = tid * window;
+        firstStep = blockDim.x * gridDim.x * window;
+
+        secondPosition = 0;
+        secondStep = 1;
+    }
+
+
+    for (int i = firstPosition; i < length; i += firstStep) {
+        for (int j = secondPosition; j < half; j += secondStep) {
+            int it = (reverse) ? i + j + half : i + window - j - 1;
+            int ij = i+j;
+            if (it < length && ij < length ) {
+                int posIT = shape::getIndexOffset(it, xShapeInfo, xLength);
+                int posIJ = shape::getIndexOffset(ij, xShapeInfo, xLength);
+
+                X v0 = x[posIJ];
+                X v1 = x[posIT];
+
+                if(!descending == (v0 > v1)) {
+                    x[posIJ] = v1;
+                    x[posIT] = v0;
+
+                    Y ytemp = y[posIJ];
+                    y[posIJ] = y[posIT];
+                    y[posIT] = ytemp;
+                }
+            }
+        }
+    }
+}
 
 //////////////////////////////////////////////////////////////////////////
 template<typename T>
-__device__
+__global__ void execBitonicArbitraryStepKernel(void *vx, Nd4jLong *xShapeInfo, int window, int length,  int reverse, bool descending) {
-void bitonicArbitraryStepKernel(void *vx, Nd4jLong *xShapeInfo, int window, int length,  int reverse, bool descending) {
-
     auto x = static_cast<T*>(vx);
 
     int tid = threadIdx.x + blockDim.x * blockIdx.x;
@@ -85,8 +238,8 @@ void bitonicArbitraryStepKernel(void *vx, Nd4jLong *xShapeInfo, int window, int
             int it = (reverse) ? i + j + half : i + window - j - 1;
             int ij = i+j;
             if (it < length && ij < length ) {
-                int posIT = getDevicePosition(xShapeInfo,it, xLength);
+                int posIT = shape::getIndexOffset(it, xShapeInfo, xLength);
-                int posIJ = getDevicePosition(xShapeInfo, ij, xLength);
+                int posIJ = shape::getIndexOffset(ij, xShapeInfo, xLength);
 
                 shmem[threadIdx.x] = x[posIJ];
                 shmem[threadIdx.x + blockDim.x] = x[posIT];
@@ -100,18 +253,22 @@ void bitonicArbitraryStepKernel(void *vx, Nd4jLong *xShapeInfo, int window, int
     }
 }
 
-//////////////////////////////////////////////////////////////////////////
-template<typename T>
-__global__ void execBitonicArbitraryStepKernel(void *vx, Nd4jLong *xShapeInfo, int window, int length,  int reverse, bool descending) {
-
-    bitonicArbitraryStepKernel<T>(vx, xShapeInfo, window, length, reverse, descending);
-}
-
 //////////////////////////////////////////////////////////////////////////
 template<typename T>
 __host__ void bitonicArbitraryStepGeneric(dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, int window, int length,  int reverse, bool descending) {
-
     execBitonicArbitraryStepKernel<T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, window, length, reverse, descending);
-    nd4j::DebugHelper::checkErrorCode(stream, "bitonicArbitrary(...) failed");
 }
+
+template <typename X, typename Y>
+__host__ void bitonicArbitraryStepGenericKey(dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int window, int length,  int reverse, bool descending) {
+    bitonicArbitraryStepKernelKey<X,Y><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, window, length, reverse, descending);
+}
+
+template <typename X, typename Y>
+__host__ void bitonicArbitraryStepGenericValue(dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int window, int length,  int reverse, bool descending) {
+    bitonicArbitraryStepKernelValue<X,Y><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, window, length, reverse, descending);
+}
+
 BUILD_SINGLE_TEMPLATE(template void ND4J_EXPORT bitonicArbitraryStepGeneric, (dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, int window, int length,  int reverse, bool descending), LIBND4J_TYPES);
+BUILD_DOUBLE_TEMPLATE(template void ND4J_EXPORT bitonicArbitraryStepGenericKey, (dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int window, int length,  int reverse, bool descending), LIBND4J_TYPES, LIBND4J_TYPES);
+BUILD_DOUBLE_TEMPLATE(template void ND4J_EXPORT bitonicArbitraryStepGenericValue, (dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int window, int length,  int reverse, bool descending), LIBND4J_TYPES, LIBND4J_TYPES);

libnd4j/include/loops/cuda/specials/bitonicSortStep.cu

@@ -21,9 +21,119 @@
 
 #include <ops/specials_cuda.h>
 
+//////////////////////////////////////////////////////////////////////////
+template <typename X, typename Y>
+__global__ void bitonicSortStepKernelValue(void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int j, int k, int length, bool descending) {
+
+    auto x = static_cast<X*>(vx);
+    auto y = static_cast<Y*>(vy);
+
+    unsigned int i, ixj; /* Sorting partners: i and ixj */
+    i = threadIdx.x + blockDim.x * blockIdx.x;
+
+    __shared__ Nd4jLong xLength;
+    if (threadIdx.x == 0)
+        xLength = shape::length(xShapeInfo);
+
+    __syncthreads();
+
+
+    if (i >= length)
+        return;
+
+    ixj = i^j;
+
+    /* The threads with the lowest ids sort the array. */
+    if ((ixj)>i) {
+        int posI = shape::getIndexOffset(i, yShapeInfo, xLength);
+        int posIXJ = shape::getIndexOffset(ixj, yShapeInfo, xLength);
+
+        if ((i&k)==0) {
+            /* Sort ascending */
+            if (!descending == (y[posI]>y[posIXJ])) {
+                /* exchange(i,ixj); */
+                X temp = x[posI];
+                x[posI] = x[posIXJ];
+                x[posIXJ] = temp;
+
+                Y ytemp = y[posI];
+                y[posI] = y[posIXJ];
+                y[posIXJ] = ytemp;
+            }
+        } else if ((i&k)!=0) {
+            /* Sort descending */
+            if (!descending == (y[posI]<y[posIXJ])) {
+                /* exchange(i,ixj); */
+                X temp = x[posI];
+                x[posI] = x[posIXJ];
+                x[posIXJ] = temp;
+
+                Y ytemp = y[posI];
+                y[posI] = y[posIXJ];
+                y[posIXJ] = ytemp;
+            }
+        }
+    }
+}
+
+//////////////////////////////////////////////////////////////////////////
+template <typename X, typename Y>
+__global__ void bitonicSortStepKernelKey(void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int j, int k, int length, bool descending) {
+
+    auto x = static_cast<X*>(vx);
+    auto y = static_cast<Y*>(vy);
+
+    unsigned int i, ixj; /* Sorting partners: i and ixj */
+    i = threadIdx.x + blockDim.x * blockIdx.x;
+
+    __shared__ Nd4jLong xLength;
+    if (threadIdx.x == 0)
+        xLength = shape::length(xShapeInfo);
+
+    __syncthreads();
+
+
+    if (i >= length)
+        return;
+
+    ixj = i^j;
+
+    /* The threads with the lowest ids sort the array. */
+    if ((ixj)>i) {
+        int posI = shape::getIndexOffset(i, xShapeInfo, xLength);
+        int posIXJ = shape::getIndexOffset(ixj, xShapeInfo, xLength);
+
+        if ((i&k)==0) {
+            /* Sort ascending */
+            if (!descending == (x[posI]>x[posIXJ])) {
+                /* exchange(i,ixj); */
+                X temp = x[posI];
+                x[posI] = x[posIXJ];
+                x[posIXJ] = temp;
+
+                Y ytemp = y[posI];
+                y[posI] = y[posIXJ];
+                y[posIXJ] = ytemp;
+            }
+        } else if ((i&k)!=0) {
+            /* Sort descending */
+            if (!descending == (x[posI]<x[posIXJ])) {
+                /* exchange(i,ixj); */
+                X temp = x[posI];
+                x[posI] = x[posIXJ];
+                x[posIXJ] = temp;
+
+                Y ytemp = y[posI];
+                y[posI] = y[posIXJ];
+                y[posIXJ] = ytemp;
+            }
+        }
+    }
+}
+
 //////////////////////////////////////////////////////////////////////////
 template<typename T>
-__device__ void bitonicSortStepKernel(void *vx, Nd4jLong *xShapeInfo, int j, int k, int length, bool descending) {
+__global__ void bitonicSortStepKernel(void *vx, Nd4jLong *xShapeInfo, int j, int k, int length, bool descending) {
 
     auto x = static_cast<T*>(vx);
 
@@ -44,8 +154,8 @@ __device__ void bitonicSortStepKernel(void *vx, Nd4jLong *xShapeInfo, int j, int
 
     /* The threads with the lowest ids sort the array. */
     if ((ixj)>i) {
-        int posI = getDevicePosition(xShapeInfo, i, xLength);
+        int posI = shape::getIndexOffset(i, xShapeInfo, xLength);
-        int posIXJ = getDevicePosition(xShapeInfo, ixj, xLength);
+        int posIXJ = shape::getIndexOffset(ixj, xShapeInfo, xLength);
 
         if ((i&k)==0) {
             /* Sort ascending */
@@ -69,16 +179,23 @@ __device__ void bitonicSortStepKernel(void *vx, Nd4jLong *xShapeInfo, int j, int
 
 //////////////////////////////////////////////////////////////////////////
 template<typename T>
-__global__ void execBitonicSortStepKernel(void *vx, Nd4jLong *xShapeInfo, int j, int k, int length, bool descending) {
+__host__ void bitonicSortStepGeneric(dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, int j, int k, int length, bool descending) {
-
+    bitonicSortStepKernel<T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, j, k, length, descending);
-    bitonicSortStepKernel<T>(vx, xShapeInfo, j, k, length, descending);
 }
 
 //////////////////////////////////////////////////////////////////////////
-template<typename T>
+template <typename X, typename Y>
-__host__ void bitonicSortStepGeneric(dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, int j, int k, int length, bool descending) {
+__host__ void bitonicSortStepGenericKey(dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int j, int k, int length, bool descending) {
-
+    bitonicSortStepKernelKey<X,Y><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, j, k, length, descending);
-    execBitonicSortStepKernel<T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, j, k, length, descending);
-    nd4j::DebugHelper::checkErrorCode(stream, "bitonicSortStep(...) failed");
 }
+
+//////////////////////////////////////////////////////////////////////////
+template <typename X, typename Y>
+__host__ void bitonicSortStepGenericValue(dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int j, int k, int length, bool descending) {
+    bitonicSortStepKernelValue<X,Y><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, j, k, length, descending);
+}
+
+
 BUILD_SINGLE_TEMPLATE(template void ND4J_EXPORT bitonicSortStepGeneric, (dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, int j, int k, int length, bool descending), LIBND4J_TYPES);
+BUILD_DOUBLE_TEMPLATE(template void ND4J_EXPORT bitonicSortStepGenericKey, (dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int j, int k, int length, bool descending), LIBND4J_TYPES, LIBND4J_TYPES);
+BUILD_DOUBLE_TEMPLATE(template void ND4J_EXPORT bitonicSortStepGenericValue, (dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int j, int k, int length, bool descending), LIBND4J_TYPES, LIBND4J_TYPES);

libnd4j/include/loops/cuda/specials/oesTad.cu

@@ -16,15 +16,86 @@
 
 //
 // @author raver119@gmail.com
-// @author Yurii Shyrma, created on 28.11.2018
 //
 
 #include <ops/specials_cuda.h>
 
+//////////////////////////////////////////////////////////////////////////
+template <typename X, typename Y>
+__global__ void execOesTadKernelKey(void *vx, Nd4jLong *xShapeInfo,
+                                    void *vy, Nd4jLong *yShapeInfo,
+                                 int *dimension, int dimensionLength,
+                                 Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
+                                 bool descending) {
+
+    auto x = static_cast<X*>(vx);
+    auto y = static_cast<Y*>(vy);
+
+    __shared__ int xLength;
+    __shared__ int xTadLength;
+    __shared__ int numTads;
+    if (threadIdx.x == 0) {
+        xLength = shape::length(xShapeInfo);
+        xTadLength = shape::length(tadShapeInfo);
+        numTads = xLength / xTadLength;
+    }
+    __syncthreads();
+
+    for (int r = blockIdx.x; r < numTads; r += gridDim.x) {
+        auto dx = x + tadOffsets[r];
+        auto dy = y + tadOffsets[r];
+
+        // this is general loop, we go uncached
+        int iterations = xTadLength;
+
+        for (int i = 0; i < iterations; i++) {
+
+            if (i % 2 == 0) {
+                for (int tid = threadIdx.x; tid < xTadLength; tid += blockDim.x) {
+                    auto top = 2 * tid + 1;
+                    if (top < xTadLength) {
+                        auto t0 = shape::getIndexOffset(top - 1, tadShapeInfo, xTadLength);
+                        auto t1 = shape::getIndexOffset(top, tadShapeInfo, xTadLength);
+
+                        if (!descending == (dx[t0] > dx[t1])) {
+                            X dt0 = dx[t0];
+                            dx[t0] = dx[t1];
+                            dx[t1] = dt0;
+
+                            Y dy0 = dy[t0];
+                            dy[t0] = dy[t1];
+                            dy[t1] = dy0;
+                        }
+                    }
+                }
+            } else {
+                for (int tid = threadIdx.x; tid < xTadLength; tid += blockDim.x) {
+                    auto top = 2 * tid + 2;
+                    if (top < xTadLength) {
+                        auto t0 = shape::getIndexOffset(top - 1, tadShapeInfo, xTadLength);
+                        auto t1 = shape::getIndexOffset(top, tadShapeInfo, xTadLength);
+
+                        if (!descending == (dx[t0] > dx[t1])) {
+                            X dt0 = dx[t0];
+                            dx[t0] = dx[t1];
+                            dx[t1] = dt0;
+
+                            Y dy0 = dy[t0];
+                            dy[t0] = dy[t1];
+                            dy[t1] = dy0;
+                        }
+                    }
+                }
+            }
+            __syncthreads();
+        }
+    }
+}
+
+
 //////////////////////////////////////////////////////////////////////////
 template<typename T>
-__device__
+__global__ void execOesTadKernel(void *vx, Nd4jLong *xShapeInfo,
-void oesTadKernel(void *vx, Nd4jLong *xShapeInfo, 
                                 int *dimension, int dimensionLength,
                                 Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
                                 bool descending) {
@@ -56,7 +127,7 @@ void oesTadKernel(void *vx, Nd4jLong *xShapeInfo,
         int iterations = xTadLength;
         if (cached) {
             for (int tid = threadIdx.x; tid < xTadLength; tid += blockDim.x) {
-                auto t0 = getDevicePosition(tadShapeInfo, tid, xTadLength);
+                auto t0 = shape::getIndexOffset(tid, tadShapeInfo, xTadLength);
                 shmem[tid] = dx[t0];
             }
 
@@ -70,8 +141,8 @@ void oesTadKernel(void *vx, Nd4jLong *xShapeInfo,
                 for (int tid = threadIdx.x; tid < xTadLength; tid += blockDim.x) {
                     auto top = 2 * tid + 1;
                     if (top < xTadLength) {
-                        auto t0 = cached ? top - 1 : getDevicePosition(tadShapeInfo, top - 1, xTadLength);
+                        auto t0 = cached ? top - 1 : shape::getIndexOffset(top - 1, tadShapeInfo, xTadLength);
-                        auto t1 = cached ? top : getDevicePosition(tadShapeInfo, top, xTadLength);
+                        auto t1 = cached ? top : shape::getIndexOffset(top, tadShapeInfo, xTadLength);
 
                         if (!descending == (dx[t0] > dx[t1])) {
                             T dt0 = dx[t0];
@@ -84,8 +155,8 @@ void oesTadKernel(void *vx, Nd4jLong *xShapeInfo,
                 for (int tid = threadIdx.x; tid < xTadLength; tid += blockDim.x) {
                     auto top = 2 * tid + 2;
                     if (top < xTadLength) {
-                        auto t0 = cached ? top - 1 : getDevicePosition(tadShapeInfo, top - 1, xTadLength);
+                        auto t0 = cached ? top - 1 : shape::getIndexOffset(top - 1, tadShapeInfo, xTadLength);
-                        auto t1 = cached ? top : getDevicePosition(tadShapeInfo, top, xTadLength);
+                        auto t1 = cached ? top : shape::getIndexOffset(top, tadShapeInfo, xTadLength);
 
                         if (!descending == (dx[t0] > dx[t1])) {
                             T dt0 = dx[t0];
@@ -102,23 +173,13 @@ void oesTadKernel(void *vx, Nd4jLong *xShapeInfo,
         if (cached) {
             dx = x + tadOffsets[r];
             for (int tid = threadIdx.x; tid < xTadLength; tid += blockDim.x) {
-                auto t0 = getDevicePosition(tadShapeInfo, tid, xTadLength);
+                auto t0 = shape::getIndexOffset(tid, tadShapeInfo, xTadLength);
                 dx[t0] = shmem[tid];
             }
         }
     }
 }
 
-//////////////////////////////////////////////////////////////////////////
-template<typename T>
-__global__ void execOesTadKernel(void *vx, Nd4jLong *xShapeInfo, 
-                                int *dimension, int dimensionLength, 
-                                Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, 
-                                bool descending) {
-
-    oesTadKernel<T>(vx, xShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, descending);
-}
-
 //////////////////////////////////////////////////////////////////////////
 template<typename T>
 __host__ void oesTadGeneric(dim3 &launchDims, cudaStream_t *stream,
@@ -128,6 +189,18 @@ __host__ void oesTadGeneric(dim3 &launchDims, cudaStream_t *stream,
                                 bool descending) {
 
     execOesTadKernel<T><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, descending);
-    nd4j::DebugHelper::checkErrorCode(stream, "oesTad(...) failed");
 }
+
+template <typename X, typename Y>
+__host__ void oesTadGenericKey(dim3 &launchDims, cudaStream_t *stream,
+                            void *vx, Nd4jLong *xShapeInfo,
+                            void *vy, Nd4jLong *yShapeInfo,
+                            int *dimension, int dimensionLength,
+                            Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets,
+                            bool descending) {
+
+    execOesTadKernelKey<X,Y><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, dimension, dimensionLength, tadShapeInfo, tadOffsets, descending);
+}
+
 BUILD_SINGLE_TEMPLATE(template void ND4J_EXPORT oesTadGeneric, (dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, bool descending), LIBND4J_TYPES);
+BUILD_DOUBLE_TEMPLATE(template void ND4J_EXPORT oesTadGenericKey, (dim3 &launchDims, cudaStream_t *stream, void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets, bool descending), LIBND4J_TYPES, LIBND4J_TYPES);

libnd4j/include/ops/declarable/generic/activations/prelu.cpp

@@ -65,13 +65,7 @@ CONFIGURABLE_OP_IMPL(prelu, 2, 1, true, 0, 0) {
     REQUIRE_TRUE(product == alphaLen, 0, "PRELU OP: wrong shape of alpha array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedAlphaShape).c_str(), ShapeUtils::shapeAsString(alphaShape).c_str());
     // ***** end of validation ***** //
 
-    if(alphaShape != expectedAlphaShape)
+    helpers::prelu(block.launchContext(), *input,  alphaShape != expectedAlphaShape ? alpha->reshape(alpha->ordering(), expectedAlphaShape) : *alpha, *output);
-        alpha = alpha->reshape(alpha->ordering(), expectedAlphaShape);
-
-    helpers::prelu(block.launchContext(), *input, *alpha, *output);
-
-    if(alphaShape != expectedAlphaShape)
-        delete alpha;
 
     return Status::OK();
 }
@@ -128,9 +122,10 @@ CONFIGURABLE_OP_IMPL(prelu_bp, 3, 2, true, 0, 0) {
     REQUIRE_TRUE(product == alphaLen, 0, "PRELU_BP OP: wrong shape of alpha array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expectedAlphaShape).c_str(), ShapeUtils::shapeAsString(alphaShape).c_str());
     // ***** end of validation ***** //
 
+
     if(alphaShape != expectedAlphaShape) {
-        alpha = alpha->reshape(alpha->ordering(), expectedAlphaShape);
+        alpha = new NDArray(alpha->reshape(alpha->ordering(), expectedAlphaShape));
-        dLdA  = dLdA->reshape(dLdA->ordering(), expectedAlphaShape);
+        dLdA  = new NDArray(dLdA->reshape(dLdA->ordering(), expectedAlphaShape));
     }
 
     helpers::preluBP(block.launchContext(), *input, *alpha, *dLdO, *dLdI, *dLdA);

libnd4j/include/ops/declarable/generic/boolean/lt_scalar.cpp

@@ -29,7 +29,6 @@ namespace nd4j {
             auto x = INPUT_VARIABLE(0);
             auto y = INPUT_VARIABLE(1);
 
-            nd4j_printf("Comparing [%f] to [%f]\n", x->e<float>(0), y->e<float>(0));
             if (x->e<float>(0) < y->e<float>(0))
                 return ND4J_STATUS_TRUE;
             else

libnd4j/include/ops/declarable/generic/boolean/where.cpp

@@ -31,7 +31,7 @@ namespace nd4j {
             auto condition = INPUT_VARIABLE(0);
             auto z = OUTPUT_VARIABLE(0);
             if (z->isEmpty())
-                return ND4J_STATUS_OK;
+                return Status::OK();
 
             if (block.width() == 3) {
                 auto x = INPUT_VARIABLE(1);
@@ -44,12 +44,10 @@ namespace nd4j {
                     // FIXME: for perf it might be better to issue memcpy here, and fill only mismatched values from either X or Y
                     for (int e = 0; e < condition->lengthOf(); e++) {
                         if (y->isR()) {
-                            auto r = !condition->e<bool>(e) ? y->e<double>(e)
+                            auto r = !condition->e<bool>(e) ? y->e<double>(e) : x->e<double>(e);
-                                                                           : x->e<double>(e);
                             z->p(e, r);
                         } else {
-                            auto r = !condition->e<bool>(e) ? y->e<Nd4jLong>(e)
+                            auto r = !condition->e<bool>(e) ? y->e<Nd4jLong>(e) : x->e<Nd4jLong>(e);
-                                                                           : x->e<Nd4jLong>(e);
                             z->p(e, r);
                         }
                     }
@@ -86,7 +84,7 @@ namespace nd4j {
 
                 helpers::_where(block.launchContext(), *condition, *output, block.workspace());
             }
-            return ND4J_STATUS_OK;
+            return Status::OK();
         }
 
         DECLARE_SHAPE_FN(Where) {

libnd4j/include/ops/declarable/generic/boolean/where_np.cpp

@@ -120,7 +120,7 @@ namespace nd4j {
                 }
             }
 
-            return ND4J_STATUS_OK;
+            return Status::OK();
         }
 
         DECLARE_SHAPE_FN(where_np) {

libnd4j/include/ops/declarable/generic/convo/conv1d.cpp

@@ -81,11 +81,7 @@ CUSTOM_OP_IMPL(conv1d, 2, 1, false, 0, 4) {
     auto outputReshaped  = output ->reshape(output->ordering(),  reshapeForOutput);
     auto weightsReshaped = weights->reshape(weights->ordering(), {1, weights->sizeAt(0), weights->sizeAt(1), weights->sizeAt(2)});   // [kW, iC, oC] -> [1, kW, iC, oC]
 
-    ConvolutionUtils::conv2d(block, inputReshaped, weightsReshaped, bias, outputReshaped, 1,kW,  1,sW,  0,pW,  1,1,  isSameMode,  isNCW);
+    ConvolutionUtils::conv2d(block, &inputReshaped, &weightsReshaped, bias, &outputReshaped, 1,kW,  1,sW,  0,pW,  1,1,  isSameMode,  isNCW);
-
-    delete inputReshaped;
-    delete outputReshaped;
-    delete weightsReshaped;
 
     return Status::OK();
 }
@@ -217,13 +213,7 @@ CUSTOM_OP_IMPL(conv1d_bp, 3, 2, false, 0, 4) {
     auto weightsReshaped = weights->reshape(weights->ordering(),{1, weights->sizeAt(0), weights->sizeAt(1), weights->sizeAt(2)});    // [kW, iC, oC] -> [1, kW, iC, oC]
     auto gradWReshaped   = gradW  ->reshape(gradW->ordering(),  {1, weights->sizeAt(0), weights->sizeAt(1), weights->sizeAt(2)});    // [kW, iC, oC] -> [1, kW, iC, oC]
 
-    ConvolutionUtils::conv2dBP(block, inputReshaped, weightsReshaped, bias, gradOReshaped, gradIReshaped, gradWReshaped, gradB, 1,kW,  1,sW,  0,pW,  1,1,  isSameMode,  isNCW);
+    ConvolutionUtils::conv2dBP(block, &inputReshaped, &weightsReshaped, bias, &gradOReshaped, &gradIReshaped, &gradWReshaped, gradB, 1,kW,  1,sW,  0,pW,  1,1,  isSameMode,  isNCW);
-
-    delete inputReshaped;
-    delete gradIReshaped;
-    delete gradOReshaped;
-    delete weightsReshaped;
-    delete gradWReshaped;
 
     return Status::OK();
 }

libnd4j/include/ops/declarable/generic/convo/conv3d.cpp

@@ -151,10 +151,10 @@ CUSTOM_OP_IMPL(conv3dnew, 2, 1, false, 0, 13) {
 
     std::vector<int> permutForOutput;
 
-    if(!isNCDHW)
+    if (isNCDHW)
-        input = input->permute({0,4,1,2,3});                                    // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
-    else
         permutForOutput    = {0,2,3,4,1};                                        // [bS, oC, oD, oH, oW] -> [bS, oD, oH, oW, oC]
+    else
+        input = new NDArray(input->permute({0,4,1,2,3}));
 
     NDArray columns(input->ordering(), {bS, iC, kD, kH, kW, oD, oH, oW}, input->dataType(), block.launchContext());
     ConvolutionUtils::vol2col(block, *input, columns, sD, sH, sW, pD, pH, pW, dD, dH, dW);                 // [bS, iC, iD, iH, iW] is convoluted to [bS, iC, kD, kH, kW, oD, oH, oW]
@@ -447,21 +447,23 @@ CUSTOM_OP_IMPL(conv3dnew_bp, 3, 2, false, 0, 13) {
     std::vector<int> gradOaxesForDot;
 
     if(!isNDHWC) {
-        input = input->permute({0,4,1,2,3});                                    // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
-        gradI = gradI->permute({0,4,1,2,3});                                    // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
         gradOaxesForDot  = {0,1,2,3};                                           // bS, oD, oH, oW
+        input = new NDArray(input->permute({0,4,1,2,3}));                       // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
+        gradI = new NDArray(gradI->permute({0,4,1,2,3}));                       // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
     }
-    else
+    else {
         gradOaxesForDot  = {0,2,3,4};                                           // bS, oD, oH, oW
+    }
 
     // ----- calculation of gradW and gradB ----- //
     NDArray columns(input->ordering(), {bS, iC, kD, kH, kW, oD, oH, oW}, input->dataType(), block.launchContext());
     ConvolutionUtils::vol2col(block, *input, columns, sD, sH, sW, pD, pH, pW, dD, dH, dW);                   // [bS, iC, iD, iH, iW] is convoluted to [bS, iC, kD, kH, kW, oD, oH, oW]
     MmulHelper::tensorDot(&columns, gradO, gradW, {0,5,6,7}, gradOaxesForDot, {3,0,1,2,4});     // [bS, iC, kD, kH, kW, oD, oH, oW] x [bS, oD, oH, oW, oC]/[bS, oC, oD, oH, oW] = [iC, kD, kH, kW, oC]
 
+    //----- calculation of gradO -----//
     if(gradB) {
         if(gradB->rankOf() == 2)
-            gradB = gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()});
+            gradB = new NDArray(gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()}));
         gradO->reduceAlongDimension(reduce::Sum, gradB, gradOaxesForDot);                          // sum over bS oD oH oW
         if(gradB != OUTPUT_VARIABLE(2))
             delete gradB;

libnd4j/include/ops/declarable/generic/convo/deconv2d.cpp

@@ -64,7 +64,7 @@ CUSTOM_OP_IMPL(deconv2d, 2, 1, false, 0, 9) {
         REQUIRE_TRUE(bias->rankOf() <= 2 && oC == bias->lengthOf(), 0, "CUSTOM DECONV2D OP: wrong shape of array with biases, expected rank, length: <=2, %i, but got %i, %i instead !", oC, bias->rankOf(), bias->lengthOf());
 
     if(!isNCHW)
-        output  = output->permute({0, 3, 1, 2});                                // [bS, oH, oW, oC] -> [bS, oC, oH, oW]
+        output = new NDArray(output->permute({0, 3, 1, 2}));       // [bS, oH, oW, oC] -> [bS, oC, oH, oW]
 
     if(isSameMode)                       // SAME
         ConvolutionUtils::calcPadding2D(pH, pW, oH, oW, iH, iW, kH, kW, sH, sW, dH, dW);
@@ -211,8 +211,9 @@ CUSTOM_OP_IMPL(deconv2d_bp, 3, 2, false, 0, 9) {
 
     // -----prepare permutation arrays and axes for dot product ----- //
     std::vector<int> inputAxesForDot;
+
     if(!isNCHW) {
-        gradO = gradO->permute({0, 3, 1, 2});                                   // [bS, oH, oW, oC] -> [bS, oC, oH, oW]
+        gradO = new NDArray(gradO->permute({0, 3, 1, 2}));                      // [bS, oH, oW, oC] -> [bS, oC, oH, oW]
         inputAxesForDot = {0, 1, 2};                                            // bS, iH, iW
     }
     else
@@ -228,7 +229,7 @@ CUSTOM_OP_IMPL(deconv2d_bp, 3, 2, false, 0, 9) {
     // ----- calculation of gradB ----- //
     if(gradB) {
         if(gradB->rankOf() == 2)
-            gradB = gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()});
+            gradB = new NDArray(gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()}));
         gradO->reduceAlongDimension(reduce::Sum, gradB, {0, 2, 3});                                // sum over bS, oH, oW
         if(gradB != OUTPUT_VARIABLE(2))
             delete gradB;
@@ -237,7 +238,7 @@ CUSTOM_OP_IMPL(deconv2d_bp, 3, 2, false, 0, 9) {
     if(!isNCHW)
         delete gradO;
 
-    return ND4J_STATUS_OK;
+    return Status::OK();
 }
 
 DECLARE_SHAPE_FN(deconv2d_bp) {

libnd4j/include/ops/declarable/generic/convo/deconv3d.cpp

@@ -39,9 +39,9 @@ CUSTOM_OP_IMPL(deconv3d, 2, 1, false, 0, 13) {
     REQUIRE_TRUE(input->rankOf()   == 5, 0, "CUSTOM DECONV3D OP: rank of input array must be equal to 5, but got %i instead !", input->rankOf());
     REQUIRE_TRUE(weights->rankOf() == 5, 0, "CUSTOM DECONV3D OP: rank of weights array must be equal to 5, but got %i instead !", weights->rankOf());
 
-    int kD = INT_ARG(0) > 0 ? INT_ARG(0) : static_cast<int>(weights->sizeAt(0));// filter(kernel) depth
+    int kD = INT_ARG(0) > 0 ? INT_ARG(0) : static_cast<int>(weights->sizeAt(0));    // filter(kernel) depth
-    int kH = INT_ARG(1) > 0 ? INT_ARG(1) : static_cast<int>(weights->sizeAt(1));// filter(kernel) height
+    int kH = INT_ARG(1) > 0 ? INT_ARG(1) : static_cast<int>(weights->sizeAt(1));    // filter(kernel) height
-    int kW = INT_ARG(2) > 0 ? INT_ARG(2) : static_cast<int>(weights->sizeAt(2));// filter(kernel) width
+    int kW = INT_ARG(2) > 0 ? INT_ARG(2) : static_cast<int>(weights->sizeAt(2));    // filter(kernel) width
     int sD = INT_ARG(3);                                                            // strides depth
     int sH = INT_ARG(4);                                                            // strides height
     int sW = INT_ARG(5);                                                            // strides width
@@ -64,7 +64,7 @@ CUSTOM_OP_IMPL(deconv3d, 2, 1, false, 0, 13) {
         REQUIRE_TRUE(bias->rankOf() <= 2 && oC == bias->lengthOf(), 0, "CUSTOM DECONV3D OP: wrong shape of array with biases, expected rank, length: <=2, %i, but got %i, %i instead !", oC, bias->rankOf(), bias->lengthOf());
 
     if(!isNCDHW)
-        output  = output->permute({0, 4, 1, 2, 3});                             // [bS, oD, oH, oW, oC] -> [bS, oC, oD, oH, oW] 
+        output = new NDArray(output->permute({0, 4, 1, 2, 3}));                 // [bS, oD, oH, oW, oC] -> [bS, oC, oD, oH, oW]
 
     if(isSameMode)                       // SAME
         ConvolutionUtils::calcPadding3D(pD, pH, pW, oD, oH, oW, iD, iH, iW, kD, kH, kW, sD, sH, sW, dD, dH, dW);
@@ -225,8 +225,9 @@ CUSTOM_OP_IMPL(deconv3d_bp, 3, 2, false, 0, 13) {
 
     // -----prepare permutation arrays and axes for dot product ----- //
     std::vector<int> inputAxesForDot;
+
     if(!isNCDHW) {
-        gradO = gradO->permute({0, 4, 1, 2, 3});                                // [bS, oD, oH, oW, oC] -> [bS, oC, oD, oH, oW]
+        gradO = new NDArray(gradO->permute({0, 4, 1, 2, 3}));                   // [bS, oD, oH, oW, oC] -> [bS, oC, oD, oH, oW]
         inputAxesForDot = {0, 1, 2, 3};                                         // bS, iD, iH, iW
     }
     else
@@ -240,7 +241,7 @@ CUSTOM_OP_IMPL(deconv3d_bp, 3, 2, false, 0, 13) {
     // ----- calculation of gradB ----- //
     if(gradB) {
         if(gradB->rankOf() == 2)
-            gradB = gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()});
+            gradB = new NDArray(gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()}));
         gradO->reduceAlongDimension(reduce::Sum, gradB, {0, 2, 3, 4});                                // sum over bS, oD, oH, oW
         if(gradB != OUTPUT_VARIABLE(2))
             delete gradB;

libnd4j/include/ops/declarable/generic/convo/dilation2d.cpp

@@ -71,7 +71,7 @@ namespace ops {
         int pad_top = 0, pad_left = 0;
         int out_rows = 0, out_cols = 0;
 
-        helpers::_dilation_hw(block.launchContext(), input->shapeInfo(), weights->shapeInfo(), strides, rates, isSameShape, &stride_rows, &stride_cols, &rate_rows, &rate_cols, &pad_top, &pad_left, &out_rows, &out_cols);
+        helpers::dilation_hw(block.launchContext(), input->shapeInfo(), weights->shapeInfo(), strides, rates, isSameShape, &stride_rows, &stride_cols, &rate_rows, &rate_cols, &pad_top, &pad_left, &out_rows, &out_cols);
 
 
         REQUIRE_TRUE(out_rows > 0 && out_cols > 0, 0, "Dilation2D: outY and outX should have positive values, but got [%i, %i] instead", out_rows, out_cols);
@@ -126,7 +126,7 @@ namespace ops {
         int pad_top = 0, pad_left = 0;
         int out_rows = 0, out_cols = 0;
 
-        helpers::_dilation_hw(block.launchContext(), input, weights, strides, rates, isSameShape, &stride_rows, &stride_cols, &rate_rows, &rate_cols, &pad_top, &pad_left, &out_rows, &out_cols);
+        helpers::dilation_hw(block.launchContext(), input, weights, strides, rates, isSameShape, &stride_rows, &stride_cols, &rate_rows, &rate_cols, &pad_top, &pad_left, &out_rows, &out_cols);
 
         std::array<Nd4jLong, 4> shape = {{batch_size, out_rows, out_cols, depth}};
         newShape = ConstantShapeHelper::getInstance()->createShapeInfo(ArrayOptions::dataType(weights), 'c', 4, shape.data());

libnd4j/include/ops/declarable/generic/convo/pooling/avgpool2d.cpp

@@ -59,9 +59,9 @@ CUSTOM_OP_IMPL(avgpool2d, 1, 1, false, 0, 10) {
     const int iH = static_cast<int>(isNCHW ? input->sizeAt(2) : input->sizeAt(1));
     const int iW = static_cast<int>(isNCHW ? input->sizeAt(3) : input->sizeAt(2));
 
-    if (!isNCHW) {
+    if(!isNCHW) {
-        input  = input->permute({0, 3, 1, 2});                // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
+        input  = new NDArray(input->permute({0, 3, 1, 2}));                // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
-        output = output->permute({0, 3, 1, 2});               // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
+        output = new NDArray(output->permute({0, 3, 1, 2}));               // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
     }
 
     ConvolutionUtils::calcOutSizePool2D(oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode);
@@ -71,9 +71,8 @@ CUSTOM_OP_IMPL(avgpool2d, 1, 1, false, 0, 10) {
 
     // 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 - poolingMode; 9 - divisor;
     ConvolutionUtils::pooling2d(block, *input, *output, kH, kW, sH, sW, pH, pW, dH, dW, PoolingType::AVG_POOL, extraParam0);
-    //output->printBuffer("output op");
 
-    if (!isNCHW) {
+    if(!isNCHW) {
         delete input;
         delete output;
     }
@@ -177,10 +176,11 @@ CUSTOM_OP_IMPL(avgpool2d_bp, 2, 1, false, 0, 10) {
     REQUIRE_TRUE(expectedGradOShape == ShapeUtils::shapeAsString(gradO), 0, "AVGPOOL2D_BP op: wrong shape of output's gradients array (next epsilon), expected is %s, but got %s instead !", expectedGradOShape.c_str(), ShapeUtils::shapeAsString(gradO).c_str());
     REQUIRE_TRUE(expectedGradIShape == ShapeUtils::shapeAsString(gradI), 0, "AVGPOOL2D_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !", expectedGradIShape.c_str(), ShapeUtils::shapeAsString(gradI).c_str());
 
+
     if(!isNCHW) {
-        input = input->permute({0, 3, 1, 2});                                   // [bS, iH, iW, iC] -> [bS, iC, iH, iW]                        
+        input = new NDArray(input->permute({0, 3, 1, 2}));                                   // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
-        gradI = gradI->permute({0, 3, 1, 2});                                   // [bS, iH, iW, iC] -> [bS, iC, iH, iW]                        
+        gradI = new NDArray(gradI->permute({0, 3, 1, 2}));                                   // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
-        gradO = gradO->permute({0, 3, 1, 2});                                   // [bS, oH, oW, iC] -> [bS, iC, oH, oW]                        
+        gradO = new NDArray(gradO->permute({0, 3, 1, 2}));                                   // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
     }
 
     if(isSameMode)                       // SAME
@@ -205,9 +205,6 @@ CUSTOM_OP_IMPL(avgpool2d_bp, 2, 1, false, 0, 10) {
         delete gradI;
         delete gradO;
     }
-    // delete columns;
-    // delete columns2d;
-    // delete gradOVector;
 
     return Status::OK();
 

libnd4j/include/ops/declarable/generic/convo/pooling/avgpool3d.cpp

@@ -61,8 +61,8 @@ CUSTOM_OP_IMPL(avgpool3dnew, 1, 1, false, 0, 14) {
     REQUIRE_TRUE(expectedOutputShape == ShapeUtils::shapeAsString(output), 0, "AVGPOOL3D op: wrong shape of output array, expected is %s, but got %s instead !", expectedOutputShape.c_str(), ShapeUtils::shapeAsString(output).c_str());
 
     if(!isNCDHW) {
-        input  = input->permute({0, 4, 1, 2, 3});                                                       // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
+        input  = new NDArray(input->permute({0, 4, 1, 2, 3}));              // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
-        output = output->permute({0, 4, 1, 2, 3});                                                      // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
+        output = new NDArray(output->permute({0, 4, 1, 2, 3}));             // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
     }
 
     if(isSameMode)                       // SAME
@@ -180,9 +180,9 @@ CUSTOM_OP_IMPL(avgpool3dnew_bp, 2, 1, false, 0, 14) {
     REQUIRE_TRUE(expectedGradIShape == ShapeUtils::shapeAsString(gradI), 0, "AVGPOOL3D_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !", expectedGradIShape.c_str(), ShapeUtils::shapeAsString(gradI).c_str());
 
     if(!isNCDHW) {
-        input = input->permute({0, 4, 1, 2, 3});                                   // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]                        
+        input = new NDArray(input->permute({0, 4, 1, 2, 3}));                                   // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
-        gradI = gradI->permute({0, 4, 1, 2, 3});                                   // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]                        
+        gradI = new NDArray(gradI->permute({0, 4, 1, 2, 3}));                                   // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
-        gradO = gradO->permute({0, 4, 1, 2, 3});                                   // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]                        
+        gradO = new NDArray(gradO->permute({0, 4, 1, 2, 3}));                                   // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
     }
 
     if(isSameMode)                       // SAME

libnd4j/include/ops/declarable/generic/convo/pooling/maxpool2d.cpp

@@ -59,9 +59,9 @@ CUSTOM_OP_IMPL(maxpool2d, 1, 1, false, 0, 9) {
     const int iH = isNCHW ? input->sizeAt(2) : input->sizeAt(1);
     const int iW = isNCHW ? input->sizeAt(3) : input->sizeAt(2);
 
-    if (!isNCHW) {
+    if(!isNCHW) {
-        input  = input->permute({0, 3, 1, 2});                // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
+        input  = new NDArray(input->permute({0, 3, 1, 2}));                // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
-        output = output->permute({0, 3, 1, 2});               // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
+        output = new NDArray(output->permute({0, 3, 1, 2}));               // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
     }
 
     ConvolutionUtils::calcOutSizePool2D(oH, oW, kH, kW, sH, sW, pH, pW, dH, dW, iH, iW, isSameMode);
@@ -72,7 +72,7 @@ CUSTOM_OP_IMPL(maxpool2d, 1, 1, false, 0, 9) {
     // 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; poolingMode; 9 - divisor;
     ConvolutionUtils::pooling2d(block, *input, *output, kH, kW, sH, sW, pH, pW, dH, dW, PoolingType::MAX_POOL, 1);
 
-    if (!isNCHW) {
+    if(!isNCHW) {
         delete input;
         delete output;
     }
@@ -175,9 +175,9 @@ CUSTOM_OP_IMPL(maxpool2d_bp, 2, 1, false, 0, 10) {
     REQUIRE_TRUE(expectedGradIShape == ShapeUtils::shapeAsString(gradI), 0, "MAXPOOL2D_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !", expectedGradIShape.c_str(), ShapeUtils::shapeAsString(gradI).c_str());
 
     if(!isNCHW) {
-        input = input->permute({0, 3, 1, 2});                                   // [bS, iH, iW, iC] -> [bS, iC, iH, iW]                        
+        input = new NDArray(input->permute({0, 3, 1, 2}));                                   // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
-        gradI = gradI->permute({0, 3, 1, 2});                                   // [bS, iH, iW, iC] -> [bS, iC, iH, iW]                        
+        gradI = new NDArray(gradI->permute({0, 3, 1, 2}));                                   // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
-        gradO = gradO->permute({0, 3, 1, 2});                                   // [bS, oH, oW, iC] -> [bS, iC, oH, oW]                        
+        gradO = new NDArray(gradO->permute({0, 3, 1, 2}));                                   // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
     }
 
     if(isSameMode)                       // SAME
@@ -203,9 +203,6 @@ CUSTOM_OP_IMPL(maxpool2d_bp, 2, 1, false, 0, 10) {
         delete gradI;
         delete gradO;
     }
-    // delete columns;
-    // delete columns2d;
-    // delete gradOVector;
 
     return Status::OK();
 }

libnd4j/include/ops/declarable/generic/convo/pooling/maxpool3d.cpp

@@ -63,8 +63,8 @@ CUSTOM_OP_IMPL(maxpool3dnew, 1, 1, false, 0, 14) {
     // REQUIRE_TRUE(kD/2 >= pD && kH/2 >= pH && kW/2 >= pW, 0, "MAXPOOL3D OP: pad depth/height/width must not be greater than half of kernel depth/height/width, but got [%i, %i, %i] and [%i, %i, %i] correspondingly !", pD,pH,pW, kD,kH,kW);
 
     if(!isNCDHW) {
-        input  = input->permute({0, 4, 1, 2, 3});                                                       // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
+        input  = new NDArray(input->permute({0, 4, 1, 2, 3}));          // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
-        output = output->permute({0, 4, 1, 2, 3});                                                      // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
+        output = new NDArray(output->permute({0, 4, 1, 2, 3}));         // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
     }
 
     if(isSameMode)                       // SAME
@@ -182,9 +182,9 @@ CUSTOM_OP_IMPL(maxpool3dnew_bp, 2, 1, false, 0, 14) {
     REQUIRE_TRUE(expectedGradIShape == ShapeUtils::shapeAsString(gradI), 0, "MAXPOOL3D_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !", expectedGradIShape.c_str(), ShapeUtils::shapeAsString(gradI).c_str());
 
     if(!isNCDHW) {
-        input = input->permute({0, 4, 1, 2, 3});                                   // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]                        
+        input = new NDArray(input->permute({0, 4, 1, 2, 3}));                   // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
-        gradI = gradI->permute({0, 4, 1, 2, 3});                                   // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]                        
+        gradI = new NDArray(gradI->permute({0, 4, 1, 2, 3}));                   // [bS, iD, iH, iW, iC] -> [bS, iC, iD, iH, iW]
-        gradO = gradO->permute({0, 4, 1, 2, 3});                                   // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]                        
+        gradO = new NDArray(gradO->permute({0, 4, 1, 2, 3}));                   // [bS, oD, oH, oW, iC] -> [bS, iC, oD, oH, oW]
     }
 
     if(isSameMode)                       // SAME
@@ -211,9 +211,6 @@ CUSTOM_OP_IMPL(maxpool3dnew_bp, 2, 1, false, 0, 14) {
         delete gradI;
         delete gradO;
     }
-    // delete columns;
-    // delete columns2d;
-    // delete gradOVector;
 
     return Status::OK();
 }

libnd4j/include/ops/declarable/generic/convo/pooling/pnormpool2d.cpp

@@ -54,9 +54,9 @@ namespace nd4j {
 
             int isNCHW  = block.getIArguments()->size() > 10 ? !INT_ARG(10) : 1;       // 1-NHWC, 0-NCHW
 
-            if (!isNCHW) {
+            if(!isNCHW) {
-                input  = input->permute({0, 3, 1, 2});                  // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
+                input  = new NDArray(input->permute({0, 3, 1, 2}));                  // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
-                output = output->permute({0, 3, 1, 2});                 // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
+                output = new NDArray(output->permute({0, 3, 1, 2}));                 // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
             }
 
             const auto inY = static_cast<int>(input->sizeAt(2));
@@ -70,7 +70,7 @@ namespace nd4j {
             // 0,1 - kernel Height/Width; 2,3 - stride Height/Width; 4,5 - pad Height/Width; 6,7 - dilation Height/Width; 8 - poolingMode; 9 - divisor;
             ConvolutionUtils::pooling2d(block, *input, *output, kY, kX, sY, sX, pY, pX, dY, dX, PoolingType::PNORM_POOL, extraParam0);
 
-            if (!isNCHW) {
+            if(!isNCHW) {
                 delete input;
                 delete output;
             }
@@ -175,9 +175,9 @@ CUSTOM_OP_IMPL(pnormpool2d_bp, 2, 1, false, 1, 10) {
     REQUIRE_TRUE(expectedGradIShape == ShapeUtils::shapeAsString(gradI), 0, "PNORMPOOL2D_BP op: wrong shape of input's gradients array (epsilon), expected is %s, but got %s instead !", expectedGradIShape.c_str(), ShapeUtils::shapeAsString(gradI).c_str());
 
     if(!isNCHW) {
-        input = input->permute({0, 3, 1, 2});                                   // [bS, iH, iW, iC] -> [bS, iC, iH, iW]                        
+        input = new NDArray(input->permute({0, 3, 1, 2}));          // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
-        gradI = gradI->permute({0, 3, 1, 2});                                   // [bS, iH, iW, iC] -> [bS, iC, iH, iW]                        
+        gradI = new NDArray(gradI->permute({0, 3, 1, 2}));          // [bS, iH, iW, iC] -> [bS, iC, iH, iW]
-        gradO = gradO->permute({0, 3, 1, 2});                                   // [bS, oH, oW, iC] -> [bS, iC, oH, oW]                        
+        gradO = new NDArray(gradO->permute({0, 3, 1, 2}));          // [bS, oH, oW, iC] -> [bS, iC, oH, oW]
     }
 
     // if(isSameMode)                       // SAME
@@ -216,10 +216,6 @@ CUSTOM_OP_IMPL(pnormpool2d_bp, 2, 1, false, 1, 10) {
         delete gradI;
         delete gradO;
     }
-    // delete columns;
-    // delete columns2d;
-    // delete gradOVector;
-    // delete denomVec;
 
     return Status::OK();
 }

libnd4j/include/ops/declarable/generic/loss/softmaxCrossEntropy.cpp

@@ -78,7 +78,7 @@ CUSTOM_OP_IMPL(softmax_cross_entropy_loss, 3, 1, false, 1, 1) {
 	auto weightsBroad = weights;
 	if(!weights->isScalar() && !weights->isSameShape(&E)) {
 		if(E.rankOf() == 1 && weights->isVector() && weights->rankOf() > 1)
-    		weightsBroad = weights->reshape(weights->ordering(), {weights->lengthOf()});
+    		weightsBroad = new NDArray(weights->reshape(weights->ordering(), {weights->lengthOf()}));
     	else
 			weightsBroad = new NDArray(weights->tileToShape(E.getShapeInfo()));
 	}

libnd4j/include/ops/declarable/generic/nn/dot_product_attention.cpp

@@ -74,7 +74,7 @@ namespace ops  {
         }
 
         if(mask != nullptr){
-            NDArray* reshapedMask;
+            NDArray reshapedMask;
             if(weights->rankOf() == 4){
                 reshapedMask = mask->reshape(mask->ordering(), {mask->sizeAt(0), 1, mask->sizeAt(1), 1});
             }else{
@@ -87,8 +87,7 @@ namespace ops  {
             // before going through the softmax, we effectively push all masked positions to zero after softmax.
             //
             // we are using 1e9 to mean effectively infinity
-            *weights += (*reshapedMask - 1) * 1e9;
+            *weights += (reshapedMask - 1) * 1e9;
-            delete reshapedMask;
         }
 
         nd4j::ops::softmax softmax;
@@ -175,14 +174,13 @@ namespace ops  {
             preSoftmax /= factor;
 
         if(mask != nullptr){
-            NDArray* reshapedMask;
+            NDArray reshapedMask;
             if(preSoftmax.rankOf() == 4){
                 reshapedMask = mask->reshape(mask->ordering(), {mask->sizeAt(0), 1, mask->sizeAt(1), 1});
             }else{
                 reshapedMask = mask->reshape(mask->ordering(), {mask->sizeAt(0), mask->sizeAt(1), 1});
             }
-            preSoftmax += (*reshapedMask - 1) * 1e9;
+            preSoftmax += (reshapedMask - 1) * 1e9;
-            delete reshapedMask;
         }
 
         NDArray weights('c', weightShape, values->dataType(), block.launchContext());

libnd4j/include/ops/declarable/generic/nn/lrn.cpp

@@ -70,7 +70,7 @@ namespace nd4j {
             float beta  = T_ARG(2);            
             int depth   = INT_ARG(0);
 
-            helpers::lrnBP(*input, *gradO, *gradI, depth, bias, alpha, beta);
+            helpers::lrnBP(block, *input, *gradO, *gradI, depth, bias, alpha, beta);
 
             return Status::OK();
         }

libnd4j/include/ops/declarable/generic/nn/multi_head_dot_product_attention.cpp

@@ -98,9 +98,9 @@ namespace ops  {
         auto projectedValues = AttentionHelper::multiHeadProject(values, Wv, block.launchContext());
 
         // Apply Attention
-        NDArray attnResults('c', {projectedQueries->sizeAt(0), projectedValues->sizeAt(1), projectedValues->sizeAt(2), projectedQueries->sizeAt(3)}, projectedValues->dataType(), block.launchContext());
+        NDArray attnResults('c', {projectedQueries.sizeAt(0), projectedValues.sizeAt(1), projectedValues.sizeAt(2), projectedQueries.sizeAt(3)}, projectedValues.dataType(), block.launchContext());
         nd4j::ops::dot_product_attention attention;
-        attention.execute({projectedQueries, projectedKeys, projectedValues, mask}, {&attnResults, weights ? OUTPUT_VARIABLE(1) : nullptr}, {}, {normalization, weights}, {});
+        attention.execute({&projectedQueries, &projectedKeys, &projectedValues, mask}, {&attnResults, weights ? OUTPUT_VARIABLE(1) : nullptr}, {}, {normalization, weights}, {});
 
         // Project attention results
         attnResults.permutei({0, 3, 1, 2});
@@ -111,11 +111,9 @@ namespace ops  {
         mmul.execute({&attnResults, Wo},{&projRes}, {}, {}, {});
         projRes.reshapei(projRes.ordering(), {miniBatchSize, queryCount, outSize});
         projRes.permutei({0, 2, 1});
-        output->assign(projRes);
 
-        delete projectedQueries;
+        // FIXME: bad for performance
-        delete projectedKeys;
+        output->assign(projRes);
-        delete projectedValues;
 
         return Status::OK();
     }
@@ -227,9 +225,9 @@ namespace ops  {
         auto projectedValues = AttentionHelper::multiHeadProject(values, Wv, block.launchContext());
 
         // Apply Attention
-        NDArray attnResults('c', {projectedQueries->sizeAt(0), projectedValues->sizeAt(1), projectedValues->sizeAt(2), projectedQueries->sizeAt(3)}, projectedValues->dataType(), block.launchContext());
+        NDArray attnResults('c', {projectedQueries.sizeAt(0), projectedValues.sizeAt(1), projectedValues.sizeAt(2), projectedQueries.sizeAt(3)}, projectedValues.dataType(), block.launchContext());
         nd4j::ops::dot_product_attention attention;
-        attention.execute({projectedQueries, projectedKeys, projectedValues, mask}, {&attnResults}, {}, {normalization, 0}, {});
+        attention.execute({&projectedQueries, &projectedKeys, &projectedValues, mask}, {&attnResults}, {}, {normalization, 0}, {});
 
         // Project attention results
         attnResults.permutei({0, 3, 1, 2});
@@ -237,31 +235,25 @@ namespace ops  {
 
         // dLdWo
         auto epsPerm = eps->permute({0, 2, 1});
-        auto epsPostReshape = epsPerm->reshape(eps->ordering(), {miniBatchSize * queryCount, outSize});
+        auto epsPostReshape = epsPerm.reshape(eps->ordering(), {miniBatchSize * queryCount, outSize});
         nd4j::ops::matmul_bp matmulBp;
         NDArray dLdPreWo(attnResults.shapeInfo(), false, block.launchContext());
-        matmulBp.execute({&attnResults, Wo, epsPostReshape}, {&dLdPreWo, dLdWo}, {}, {}, {});
+        matmulBp.execute({&attnResults, Wo, &epsPostReshape}, {&dLdPreWo, dLdWo}, {}, {}, {});
 
         // dLdAttn
-        dLdPreWo.reshapei({miniBatchSize, queryCount, numHeads, projectedValues->sizeAt(2)});
+        dLdPreWo.reshapei({miniBatchSize, queryCount, numHeads, projectedValues.sizeAt(2)});
         dLdPreWo.permutei({0, 2, 3, 1});
 
         nd4j::ops::dot_product_attention_bp attentionBp;
-        NDArray dLdProjectedQueries(projectedQueries->shapeInfo(), false, block.launchContext());
+        NDArray dLdProjectedQueries(projectedQueries.shapeInfo(), false, block.launchContext());
-        NDArray dLdProjectedKeys(projectedKeys->shapeInfo(), false, block.launchContext());
+        NDArray dLdProjectedKeys(projectedKeys.shapeInfo(), false, block.launchContext());
-        NDArray dLdProjectedValues(projectedValues->shapeInfo(), false, block.launchContext());
+        NDArray dLdProjectedValues(projectedValues.shapeInfo(), false, block.launchContext());
-        attentionBp.execute({projectedQueries, projectedKeys, projectedValues, &dLdPreWo, mask},{&dLdProjectedQueries, &dLdProjectedKeys, &dLdProjectedValues}, {}, {normalization}, {});
+        attentionBp.execute({&projectedQueries, &projectedKeys, &projectedValues, &dLdPreWo, mask},{&dLdProjectedQueries, &dLdProjectedKeys, &dLdProjectedValues}, {}, {normalization}, {});
 
         AttentionHelper::multiHeadProjectBp(queries, Wq, &dLdProjectedQueries, dLdq, dLdWq, block.launchContext());
         AttentionHelper::multiHeadProjectBp(keys, Wk, &dLdProjectedKeys, dLdk, dLdWk, block.launchContext());
         AttentionHelper::multiHeadProjectBp(values, Wv, &dLdProjectedValues, dLdv, dLdWv, block.launchContext());
 
-        delete projectedQueries;
-        delete projectedKeys;
-        delete projectedValues;
-        delete epsPerm;
-        delete epsPostReshape;
-
         return Status::OK();
     }
 

libnd4j/include/ops/declarable/generic/parity_ops/betaInc.cpp

@@ -51,7 +51,7 @@ CONFIGURABLE_OP_IMPL(betainc, 3, 1, false, 0, 0) {
         REQUIRE_TRUE(0.f <= x->e<float>(i) && x->e<float>(i) <= 1.f, 0, "BETAINC op: all elements of x array must be within [0, 1] range!");
     }
 
-    *output = helpers::betaInc(block.launchContext(), *a, *b, *x);
+    helpers::betaInc(block.launchContext(), *a, *b, *x, *output);
 
     return Status::OK();
 }

libnd4j/include/ops/declarable/generic/parity_ops/bias_add.cpp

@@ -48,10 +48,7 @@ namespace nd4j {
                 //nd4j_debug("Reshaping to: [%i, %i]\n", -1, (int) bias->lengthOf());
                 auto tArr = input->reshape(input->ordering(), shape);
                 auto zArr = z->reshape(z->ordering(), shape);
-                tArr->addRowVector(bias, zArr);
+                tArr.addRowVector(bias, &zArr);
-
-                delete tArr;
-                delete zArr;
             }
 
             STORE_RESULT(*z);
@@ -87,13 +84,12 @@ namespace nd4j {
             // cnn case
             if (input->rankOf() == 4) {
                 auto epsilonNext2d = epsilonNext->permute({1, 0, 2, 3});
-                epsilonNext2d->reshapei('c', {(int) bias->lengthOf(), -1});
+                epsilonNext2d.reshapei('c', {(int) bias->lengthOf(), -1});
 
-                auto sum = epsilonNext2d->reduceAlongDimension(reduce::Sum, {1});
+                auto sum = epsilonNext2d.reduceAlongDimension(reduce::Sum, {1});
                 gradB->assign(sum);
 
                 delete sum;
-                delete epsilonNext2d;
             } else if (input->rankOf() == 2) {
                 // regular fully-connected case
                 auto sum = epsilonNext->reduceAlongDimension(reduce::Sum, {0});

libnd4j/include/ops/declarable/generic/parity_ops/check_numerics.cpp

@@ -0,0 +1,56 @@
+/*******************************************************************************
+ * 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>
+#if NOT_EXCLUDED(OP_check_numerics)
+
+#include <ops/declarable/CustomOperations.h>
+
+namespace nd4j {
+    namespace ops {
+
+        CUSTOM_OP_IMPL(check_numerics, 2, 1, true, 0, 0) {
+            auto input = INPUT_VARIABLE(0);
+            auto message = INPUT_VARIABLE(1);
+            auto output = OUTPUT_VARIABLE(0);
+
+            auto allFinite = input->reduceNumber(reduce::BoolOps::IsFinite);
+            REQUIRE_TRUE(allFinite.e<bool>(0), 0, "CheckNumerics: %s", message->e<std::string>(0).c_str());
+
+            if (!block.isInplace())
+                output->assign(input);
+
+            return Status::OK();
+        }
+
+        DECLARE_SHAPE_FN(check_numerics) {
+            return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(inputShape->at(0))));
+        }
+
+        DECLARE_TYPES(check_numerics) {
+            getOpDescriptor()
+                    ->setAllowedInputTypes(0, {ALL_FLOATS})
+                    ->setAllowedInputTypes(1, nd4j::DataType::UTF8)
+                    ->setAllowedOutputTypes({ALL_FLOATS});
+        }
+    }
+}
+
+#endif

libnd4j/include/ops/declarable/generic/parity_ops/crop_and_resize.cpp

@@ -56,7 +56,7 @@ namespace nd4j {
         }
 
         DECLARE_SHAPE_FN(crop_and_resize) {
-            auto in = inputShape->at(0);
+            auto in = inputShape->at(1);
 
             Nd4jLong outputShape[4];
 
@@ -77,8 +77,13 @@ namespace nd4j {
         }
         DECLARE_TYPES(crop_and_resize) {
             getOpDescriptor()
-                    ->setAllowedInputTypes(nd4j::DataType::ANY)
+                    ->setAllowedInputTypes(0, {ALL_INTS, ALL_FLOATS})
-                    ->setAllowedOutputTypes({ALL_FLOATS});
+//                    ->setAllowedInputTypes(1, {ALL_FLOATS})
+                    ->setAllowedInputTypes(1, {FLOAT32}) // as TF
+                    ->setAllowedInputTypes(2, {ALL_INTS})
+                    ->setAllowedInputTypes(3, {ALL_INTS})
+                    ->setAllowedOutputTypes({FLOAT32}); // as TF
+//                    ->setAllowedOutputTypes({ALL_FLOATS});
         }
     }
 }

libnd4j/include/ops/declarable/generic/parity_ops/cross.cpp

@@ -47,9 +47,9 @@ namespace ops {
         auto o = OUTPUT_VARIABLE(0);
 
         if (a->lengthOf() == 3) {
-            helpers::_cross(block.launchContext(), a, b, o);
+            helpers::cross(block.launchContext(), a, b, o);
         } else {
-            helpers::_crossBatched(block.launchContext(), a, b, o);
+            helpers::crossBatched(block.launchContext(), a, b, o);
         }
 
         return Status::OK();