Merge pull request #8814 from KonduitAI/master

Development updates

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/datasets/datavec/RecordReaderDataSetiteratorTest.java

@@ -1381,4 +1381,17 @@ public class RecordReaderDataSetiteratorTest extends BaseDL4JTest {
         assertNotNull(ds.getFeatures());
         assertNull(ds.getLabels());
     }
+
+
+    @Test
+    public void testCollectMetaData(){
+        RecordReaderDataSetIterator trainIter = new RecordReaderDataSetIterator.Builder(new CollectionRecordReader(Collections.<List<Writable>>emptyList()), 1)
+                .collectMetaData(true)
+                .build();
+        assertTrue(trainIter.isCollectMetaData());
+        trainIter.setCollectMetaData(false);
+        assertFalse(trainIter.isCollectMetaData());
+        trainIter.setCollectMetaData(true);
+        assertTrue(trainIter.isCollectMetaData());
+    }
 }

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/eval/EvalTest.java

@@ -33,7 +33,6 @@ import org.deeplearning4j.datasets.iterator.IteratorMultiDataSetIterator;
 import org.deeplearning4j.datasets.iterator.impl.IrisDataSetIterator;
 import org.deeplearning4j.datasets.iterator.impl.ListDataSetIterator;
 import org.deeplearning4j.datasets.iterator.impl.SingletonMultiDataSetIterator;
-import org.deeplearning4j.eval.meta.Prediction;
 import org.deeplearning4j.nn.api.OptimizationAlgorithm;
 import org.deeplearning4j.nn.conf.*;
 import org.deeplearning4j.nn.conf.layers.*;
@@ -52,19 +51,13 @@ import org.nd4j.linalg.dataset.api.MultiDataSet;
 import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
 import org.nd4j.linalg.dataset.api.preprocessor.NormalizerStandardize;
 import org.nd4j.linalg.factory.Nd4j;
-import org.nd4j.linalg.indexing.INDArrayIndex;
-import org.nd4j.linalg.indexing.NDArrayIndex;
-import org.nd4j.linalg.io.ClassPathResource;
 import org.nd4j.linalg.learning.config.Sgd;
 import org.nd4j.linalg.lossfunctions.LossFunctions;
-import org.nd4j.linalg.util.FeatureUtil;
 import org.nd4j.resources.Resources;
 
 import java.util.*;
 
 import static org.junit.Assert.*;
-import static org.nd4j.linalg.indexing.NDArrayIndex.all;
-import static org.nd4j.linalg.indexing.NDArrayIndex.interval;
 
 /**
  * Created by agibsonccc on 12/22/14.
@@ -165,7 +158,7 @@ public class EvalTest extends BaseDL4JTest {
         assertEquals(evalExpected.getConfusionMatrix(), evalActual.getConfusionMatrix());
     }
 
-    @Test(timeout = 300000)
+    @Test
     public void testEvaluationWithMetaData() throws Exception {
 
         RecordReader csv = new CSVRecordReader();
@@ -256,6 +249,30 @@ public class EvalTest extends BaseDL4JTest {
             assertEquals(actualCounts[i], actualClassI.size());
             assertEquals(predictedCounts[i], predictedClassI.size());
         }
+
+
+        //Finally: test doEvaluation methods
+        rrdsi.reset();
+        org.nd4j.evaluation.classification.Evaluation e2 = new org.nd4j.evaluation.classification.Evaluation();
+        net.doEvaluation(rrdsi, e2);
+        for (int i = 0; i < 3; i++) {
+            List<org.nd4j.evaluation.meta.Prediction> actualClassI = e2.getPredictionsByActualClass(i);
+            List<org.nd4j.evaluation.meta.Prediction> predictedClassI = e2.getPredictionByPredictedClass(i);
+            assertEquals(actualCounts[i], actualClassI.size());
+            assertEquals(predictedCounts[i], predictedClassI.size());
+        }
+
+        ComputationGraph cg = net.toComputationGraph();
+        rrdsi.reset();
+        e2 = new org.nd4j.evaluation.classification.Evaluation();
+        cg.doEvaluation(rrdsi, e2);
+        for (int i = 0; i < 3; i++) {
+            List<org.nd4j.evaluation.meta.Prediction> actualClassI = e2.getPredictionsByActualClass(i);
+            List<org.nd4j.evaluation.meta.Prediction> predictedClassI = e2.getPredictionByPredictedClass(i);
+            assertEquals(actualCounts[i], actualClassI.size());
+            assertEquals(predictedCounts[i], predictedClassI.size());
+        }
+
     }
 
     private static void apply(org.nd4j.evaluation.classification.Evaluation e, int nTimes, INDArray predicted, INDArray actual) {
@@ -504,11 +521,11 @@ public class EvalTest extends BaseDL4JTest {
             list.add(new org.nd4j.linalg.dataset.MultiDataSet(new INDArray[]{ds.getFeatures()}, new INDArray[]{ds.getLabels(), ds.getLabels()}));
         }
 
-        Evaluation e = new Evaluation();
+        org.nd4j.evaluation.classification.Evaluation e = new org.nd4j.evaluation.classification.Evaluation();
-        RegressionEvaluation e2 = new RegressionEvaluation();
+        org.nd4j.evaluation.regression.RegressionEvaluation e2 = new org.nd4j.evaluation.regression.RegressionEvaluation();
-        Map<Integer,IEvaluation[]> evals = new HashMap<>();
+        Map<Integer,org.nd4j.evaluation.IEvaluation[]> evals = new HashMap<>();
-        evals.put(0, new IEvaluation[]{(IEvaluation) e});
+        evals.put(0, new org.nd4j.evaluation.IEvaluation[]{e});
-        evals.put(1, new IEvaluation[]{(IEvaluation) e2});
+        evals.put(1, new org.nd4j.evaluation.IEvaluation[]{e2});
 
         cg.evaluate(new IteratorMultiDataSetIterator(list.iterator(), 30), evals);
 
@@ -567,14 +584,14 @@ public class EvalTest extends BaseDL4JTest {
         }
 
         try {
-            net.evaluateROC(iter);
+            net.evaluateROC(iter, 0);
             fail("Expected exception");
         } catch (IllegalStateException e){
             assertTrue(e.getMessage().contains("Classifier") && e.getMessage().contains("ROC"));
         }
 
         try {
-            net.evaluateROCMultiClass(iter);
+            net.evaluateROCMultiClass(iter, 0);
             fail("Expected exception");
         } catch (IllegalStateException e){
             assertTrue(e.getMessage().contains("Classifier") && e.getMessage().contains("ROCMultiClass"));
@@ -589,14 +606,14 @@ public class EvalTest extends BaseDL4JTest {
         }
 
         try {
-            cg.evaluateROC(iter);
+            cg.evaluateROC(iter, 0);
             fail("Expected exception");
         } catch (IllegalStateException e){
             assertTrue(e.getMessage().contains("Classifier") && e.getMessage().contains("ROC"));
         }
 
         try {
-            cg.evaluateROCMultiClass(iter);
+            cg.evaluateROCMultiClass(iter, 0);
             fail("Expected exception");
         } catch (IllegalStateException e){
             assertTrue(e.getMessage().contains("Classifier") && e.getMessage().contains("ROCMultiClass"));
@@ -606,10 +623,10 @@ public class EvalTest extends BaseDL4JTest {
         //Disable validation, and check same thing:
         net.getLayerWiseConfigurations().setValidateOutputLayerConfig(false);
         net.evaluate(iter);
-        net.evaluateROCMultiClass(iter);
+        net.evaluateROCMultiClass(iter, 0);
 
         cg.getConfiguration().setValidateOutputLayerConfig(false);
         cg.evaluate(iter);
-        cg.evaluateROCMultiClass(iter);
+        cg.evaluateROCMultiClass(iter, 0);
     }
 }

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/eval/RegressionEvalTest.java

@@ -61,7 +61,7 @@ public class RegressionEvalTest extends BaseDL4JTest {
 
         DataSet ds = new DataSet(f, l);
         DataSetIterator iter = new ExistingDataSetIterator(Collections.singletonList(ds));
-        RegressionEvaluation re = net.evaluateRegression(iter);
+        org.nd4j.evaluation.regression.RegressionEvaluation re = net.evaluateRegression(iter);
 
         for (int i = 0; i < 5; i++) {
             assertEquals(1.0, re.meanSquaredError(i), 1e-6);

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/dtypes/DTypeTests.java

@@ -17,6 +17,7 @@
 package org.deeplearning4j.nn.dtypes;
 
 import org.deeplearning4j.nn.conf.layers.recurrent.TimeDistributed;
+import org.deeplearning4j.nn.modelimport.keras.layers.TFOpLayer;
 import org.nd4j.shade.guava.collect.ImmutableSet;
 import org.nd4j.shade.guava.reflect.ClassPath;
 import lombok.extern.slf4j.Slf4j;
@@ -128,7 +129,7 @@ public class DTypeTests extends BaseDL4JTest {
                 throw new RuntimeException(e);
             }
 
-            if (Modifier.isAbstract(clazz.getModifiers()) || clazz.isInterface()) {
+            if (Modifier.isAbstract(clazz.getModifiers()) || clazz.isInterface() || TFOpLayer.class == clazz) {     //Skip TFOpLayer here - dtype depends on imported model dtype
                 continue;
             }
 

deeplearning4j/deeplearning4j-data/deeplearning4j-datasets/src/main/java/org/deeplearning4j/datasets/fetchers/CacheableExtractableDataSetFetcher.java

@@ -86,7 +86,7 @@ public abstract class CacheableExtractableDataSetFetcher implements CacheableDat
         }
 
         try {
-            ArchiveUtils.unzipFileTo(tmpFile.getAbsolutePath(), localCacheDir.getAbsolutePath());
+            ArchiveUtils.unzipFileTo(tmpFile.getAbsolutePath(), localCacheDir.getAbsolutePath(), false);
         } catch (Throwable t){
             //Catch any errors during extraction, and delete the directory to avoid leaving the dir in an invalid state
             if(localCacheDir.exists())

deeplearning4j/deeplearning4j-data/deeplearning4j-datavec-iterators/src/main/java/org/deeplearning4j/datasets/datavec/RecordReaderDataSetIterator.java

@@ -205,6 +205,7 @@ public class RecordReaderDataSetIterator implements DataSetIterator {
         this.numPossibleLabels = b.numPossibleLabels;
         this.regression = b.regression;
         this.preProcessor = b.preProcessor;
+        this.collectMetaData = b.collectMetaData;
     }
 
     /**

deeplearning4j/deeplearning4j-modelimport/pom.xml

@@ -105,6 +105,14 @@
             <version>${project.version}</version>
             <scope>test</scope>
         </dependency>
+
+        <dependency>
+            <groupId>org.nd4j</groupId>
+            <artifactId>nd4j-tensorflow</artifactId>
+            <version>${nd4j.version}</version>
+            <scope>test</scope>
+        </dependency>
+
     </dependencies>
 
     <profiles>

deeplearning4j/deeplearning4j-modelimport/src/main/java/org/deeplearning4j/nn/modelimport/keras/config/Keras2LayerConfiguration.java

@@ -103,4 +103,6 @@ public class Keras2LayerConfiguration extends KerasLayerConfiguration {
 
     /* Keras weight initializers. */
     private final String LAYER_FIELD_INIT = "kernel_initializer";
+
+    private final String TENSORFLOW_OP_LAYER = "TensorFlowOpLayer";
 }

deeplearning4j/deeplearning4j-modelimport/src/main/java/org/deeplearning4j/nn/modelimport/keras/layers/KerasTFOpLayer.java

@@ -0,0 +1,74 @@
+/*******************************************************************************
+ * Copyright (c) 2020 Konduit K.K.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+package org.deeplearning4j.nn.modelimport.keras.layers;
+
+import org.deeplearning4j.nn.conf.inputs.InputType;
+import org.deeplearning4j.nn.modelimport.keras.KerasLayer;
+import org.deeplearning4j.nn.modelimport.keras.exceptions.InvalidKerasConfigurationException;
+import org.deeplearning4j.nn.modelimport.keras.exceptions.UnsupportedKerasConfigurationException;
+
+import java.util.Map;
+
+
+public class KerasTFOpLayer extends KerasLayer {
+
+    public KerasTFOpLayer(Integer kerasVersion) throws UnsupportedKerasConfigurationException {
+        super(kerasVersion);
+        if (kerasVersion != 2){
+            throw new UnsupportedKerasConfigurationException("KerasTFOpLayer expects Keras version 2");
+        }
+    }
+
+    /**
+     * Constructor from parsed Keras layer configuration dictionary.
+     *
+     * @param layerConfig dictionary containing Keras layer configuration
+     * @throws InvalidKerasConfigurationException     Invalid Keras config
+     * @throws UnsupportedKerasConfigurationException Unsupported Keras config
+     */
+    public KerasTFOpLayer(Map<String, Object> layerConfig)
+            throws InvalidKerasConfigurationException, UnsupportedKerasConfigurationException {
+        this(layerConfig, true);
+    }
+
+    /**
+     * Constructor from parsed Keras layer configuration dictionary.
+     *
+     * @param layerConfig           dictionary containing Keras layer configuration
+     * @param enforceTrainingConfig whether to enforce training-related configuration options
+     * @throws InvalidKerasConfigurationException     Invalid Keras config
+     * @throws UnsupportedKerasConfigurationException Unsupported Keras config
+     */
+    public KerasTFOpLayer(Map<String, Object> layerConfig, boolean enforceTrainingConfig) throws UnsupportedKerasConfigurationException, InvalidKerasConfigurationException{
+        super(layerConfig, enforceTrainingConfig);
+        this.layer = new TFOpLayer((Map)((Map)layerConfig.get("config")).get("node_def"), (Map)((Map)layerConfig.get("config")).get("constants"));
+    }
+
+    /**
+     * Get layer output type.
+     *
+     * @param inputType Array of InputTypes
+     * @return output type as InputType
+     * @throws InvalidKerasConfigurationException Invalid Keras configuration
+     */
+    public InputType getOutputType(InputType... inputType){
+        return this.layer.getOutputType(0, inputType[0]);
+    }
+
+
+
+}

deeplearning4j/deeplearning4j-modelimport/src/main/java/org/deeplearning4j/nn/modelimport/keras/layers/TFOpLayer.java

@@ -0,0 +1,106 @@
+/*******************************************************************************
+ * Copyright (c) 2020 Konduit K.K.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+package org.deeplearning4j.nn.modelimport.keras.layers;
+
+import org.deeplearning4j.nn.api.ParamInitializer;
+import org.deeplearning4j.nn.conf.GradientNormalization;
+import org.deeplearning4j.nn.conf.InputPreProcessor;
+import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
+import org.deeplearning4j.nn.conf.inputs.InputType;
+import org.deeplearning4j.nn.conf.layers.Layer;
+import org.deeplearning4j.nn.conf.memory.LayerMemoryReport;
+import org.deeplearning4j.nn.modelimport.keras.layers.TFOpLayerImpl;
+import org.deeplearning4j.nn.params.EmptyParamInitializer;
+import org.deeplearning4j.optimize.api.TrainingListener;
+import org.nd4j.linalg.api.buffer.DataType;
+import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.factory.Nd4j;
+import org.nd4j.linalg.learning.regularization.Regularization;
+
+import java.util.Collection;
+import java.util.List;
+import java.util.Map;
+
+
+public class TFOpLayer extends Layer {
+
+    private Map nodeDef;
+    private Map constants;
+    public TFOpLayer(Map nodeDef, Map constants){
+        super();
+        this.nodeDef = nodeDef;
+        this.constants = constants;
+    }
+
+    @Override
+    public ParamInitializer initializer() {
+        return EmptyParamInitializer.getInstance();
+    }
+    @Override
+    public InputPreProcessor getPreProcessorForInputType(InputType inputType) {
+        return null;
+    }
+
+    @Override
+    public boolean isPretrainParam(String param){
+        return false;
+    }
+
+    @Override
+    public InputType getOutputType(int idx, InputType inputType){
+        long[] shape = inputType.getShape(true);
+        TFOpLayerImpl tempLayer = new TFOpLayerImpl(nodeDef, constants, null, null);
+        long[] outputShape = tempLayer.getOutputShape(shape);
+        return InputType.inferInputType(Nd4j.create(outputShape));
+
+    }
+
+    @Override
+    public  void setNIn(InputType inputType, boolean override){}
+
+
+    @Override
+    public GradientNormalization getGradientNormalization(){return null;}
+
+
+    @Override
+    public org.deeplearning4j.nn.api.Layer instantiate(NeuralNetConfiguration conf,
+                                                                Collection<TrainingListener> trainingListeners, int layerIndex, INDArray layerParamsView,
+                                                                boolean initializeParams, DataType networkDataType) {
+
+        TFOpLayerImpl tfOpLayerImpl = new TFOpLayerImpl(nodeDef, constants, conf, networkDataType);
+        tfOpLayerImpl.setListeners(trainingListeners);
+        tfOpLayerImpl.setIndex(layerIndex);
+        return tfOpLayerImpl;
+    }
+
+    @Override
+    public double getGradientNormalizationThreshold(){return 0.;}
+
+    @Override
+    public List<Regularization> getRegularizationByParam(String paramName){return null;}
+
+    @Override
+    public LayerMemoryReport getMemoryReport(InputType inputType) {
+        return new LayerMemoryReport(); //TODO
+    }
+
+
+
+
+
+}

deeplearning4j/deeplearning4j-modelimport/src/main/java/org/deeplearning4j/nn/modelimport/keras/layers/TFOpLayerImpl.java

@@ -0,0 +1,169 @@
+/*******************************************************************************
+ * Copyright (c) 2020 Konduit K.K.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+package org.deeplearning4j.nn.modelimport.keras.layers;
+
+import lombok.Data;
+import lombok.extern.slf4j.Slf4j;
+import org.apache.commons.lang3.ArrayUtils;
+import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
+import org.deeplearning4j.nn.gradient.Gradient;
+import org.deeplearning4j.nn.layers.AbstractLayer;
+import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
+import org.nd4j.TFGraphRunnerService;
+import org.nd4j.linalg.api.buffer.DataType;
+import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.factory.Nd4j;
+import org.nd4j.linalg.primitives.Pair;
+import org.tensorflow.framework.AttrValue;
+import org.tensorflow.framework.GraphDef;
+import org.tensorflow.framework.NodeDef;
+import com.google.gson.Gson;
+import org.nd4j.shade.protobuf.Message;
+import org.nd4j.shade.protobuf.TextFormat;
+
+import java.util.*;
+import java.util.List;
+
+
+@Slf4j
+@Data
+public class TFOpLayerImpl extends AbstractLayer<TFOpLayer> {
+
+
+    private Map nodeDef;
+    private Map constants;
+    private List<String> inputNames;
+    TFGraphRunnerService graphRunnerService;
+
+    public TFOpLayerImpl(Map nodeDef, Map constants, NeuralNetConfiguration conf, DataType dtype){
+        super(conf, dtype);
+        this.nodeDef = nodeDef;
+        this.constants = constants;
+        setGraphRunner();
+    }
+
+    @Override
+    public Pair<Gradient, INDArray> backpropGradient(INDArray epsilon, LayerWorkspaceMgr workspaceMgr){
+        throw new RuntimeException("Backprop through TFOpLayerImpl is not supported yet." +
+                " TFOpLayerImpl is created when importing TensorFlow 2.0 Keras models " +
+                "(tf.keras) into DL4J, that contains TensorFlow operations not just Keras layers.");
+    }
+
+    /**
+     * Converts a Map representation of Nodedef to a singleton TF Graph and instantiates a GraphRunner.
+     */
+    private void setGraphRunner() {
+        try{
+            String json = new Gson().toJson(nodeDef);
+            NodeDef.Builder builder = NodeDef.newBuilder();
+            org.nd4j.shade.protobuf.util.JsonFormat.parser().merge(json, builder);
+            NodeDef nodeDef = builder.build();
+            List<String> allInputNames = new ArrayList<>(); // including constants
+            Map<String, String> inputDataTypes = new HashMap<>();
+            Map<String, INDArray> constArrays = new HashMap();
+            this.inputNames = new ArrayList<>();
+            List<String> outputNames = Arrays.asList(nodeDef.getName());
+            Map<String, AttrValue> attrMap = nodeDef.getAttrMap();
+            for (int i = 0; i < nodeDef.getInputCount(); i++){
+                String inputName = nodeDef.getInput(i);
+                String[] split = inputName.split("/");
+                String attrKey;
+                if (split.length == 1){
+                    attrKey = "T";
+                }
+                else{
+                    attrKey = "T" + split[split.length - 1];
+                }
+                allInputNames.add(nodeDef.getInput(i));
+                inputDataTypes.put(nodeDef.getInput(i), attrMap.get(attrKey).getType().toString());
+                if (constants.containsKey(String.valueOf(i))){
+                    constArrays.put(nodeDef.getInput(i), Nd4j.create((List<Number>)constants.get(String.valueOf(i))));
+                }
+                else{
+                    this.inputNames.add(nodeDef.getInput(i));
+                }
+            }
+            String graph = "node{\n" + nodeDef.toString() + "\n}\nversions {\n producer: 22\n}";
+            for (int i = 0; i < allInputNames.size(); i++){
+                String inpName = allInputNames.get(i);
+                String dtype = inputDataTypes.get(inpName);
+                graph = "node{\nname: \"" + inpName + "\"\nop: \"Placeholder\"\nattr{\nkey: \"dtype\"\n value {\n type: " + dtype + "}\n}\n}\n" + graph;
+            }
+            log.info(graph);
+            GraphDef.Builder graphDefBuilder = GraphDef.newBuilder();
+            TextFormat.getParser().merge(graph, graphDefBuilder);
+            GraphDef graphDef = graphDefBuilder.build();
+            org.nd4j.shade.protobuf.ByteString serialized = graphDef.toByteString();
+            byte[] graphBytes = serialized.toByteArray();
+
+            ServiceLoader<TFGraphRunnerService> sl = ServiceLoader.load(TFGraphRunnerService.class);
+            Iterator<TFGraphRunnerService> iter = sl.iterator();
+            if (!iter.hasNext()){
+                throw new RuntimeException("The model contains a Tensorflow Op, which requires the nd4j-tensorflow dependency to execute.");
+            }
+
+            this.graphRunnerService = iter.next().init(allInputNames, outputNames, graphBytes, constArrays, inputDataTypes);
+        }
+        catch (Exception e){
+            throw new RuntimeException("Error parsing protobuf", e);
+        }
+
+    }
+
+    private INDArray runGraph(INDArray input){
+        if (input.rank() == 3){
+            // TODO make this a preprocessor
+            input = input.permute(0, 2, 1);
+        }
+        Map<String, INDArray> inputMap = new HashMap<>();
+        inputMap.put(inputNames.get(0), input);
+        INDArray out = graphRunnerService.run(inputMap).values().toArray(new INDArray[0])[0];
+        if (out.rank() == 3){
+            out = out.permute(0, 2, 1); // TODO post-processing?
+        }
+
+        return out;
+    }
+
+    public long[] getOutputShape(long[] inputShape){
+        long[] shape = ArrayUtils.clone(inputShape);
+        for(int i = 0; i < shape.length; i++){
+            if (shape[i] < 0){
+                shape[i] = 1;
+            }
+        }
+        INDArray dummyArr = Nd4j.zeros(shape);
+        return runGraph(dummyArr).shape();
+    }
+
+    @Override
+    public INDArray activate(boolean training, LayerWorkspaceMgr workspaceMgr){
+        return runGraph(input);
+    }
+
+
+    @Override
+    public boolean isPretrainLayer(){
+        return false;
+    }
+
+    @Override
+    public void clearNoiseWeightParams(){
+
+    }
+
+}

deeplearning4j/deeplearning4j-modelimport/src/main/java/org/deeplearning4j/nn/modelimport/keras/layers/wrappers/KerasBidirectional.java

@@ -190,7 +190,7 @@ public class KerasBidirectional extends KerasLayer {
                     "Keras Bidirectional layer accepts only one input (received " + inputType.length + ")");
         InputPreProcessor preProcessor = getInputPreprocessor(inputType);
         if (preProcessor != null)
-            return preProcessor.getOutputType(inputType[0]);
+            return this.getBidirectionalLayer().getOutputType(-1, preProcessor.getOutputType(inputType[0]));
         else
             return this.getBidirectionalLayer().getOutputType(-1, inputType[0]);
     }

deeplearning4j/deeplearning4j-modelimport/src/main/java/org/deeplearning4j/nn/modelimport/keras/utils/KerasLayerUtils.java

@@ -21,10 +21,12 @@ import org.deeplearning4j.nn.conf.graph.ElementWiseVertex;
 import org.deeplearning4j.nn.conf.layers.Layer;
 import org.deeplearning4j.nn.conf.layers.samediff.SameDiffLambdaLayer;
 import org.deeplearning4j.nn.modelimport.keras.KerasLayer;
+import org.deeplearning4j.nn.modelimport.keras.config.Keras2LayerConfiguration;
 import org.deeplearning4j.nn.modelimport.keras.config.KerasLayerConfiguration;
 import org.deeplearning4j.nn.modelimport.keras.exceptions.InvalidKerasConfigurationException;
 import org.deeplearning4j.nn.modelimport.keras.exceptions.UnsupportedKerasConfigurationException;
 import org.deeplearning4j.nn.modelimport.keras.layers.KerasInput;
+import org.deeplearning4j.nn.modelimport.keras.layers.KerasTFOpLayer;
 import org.deeplearning4j.nn.modelimport.keras.layers.advanced.activations.*;
 import org.deeplearning4j.nn.modelimport.keras.layers.convolutional.*;
 import org.deeplearning4j.nn.modelimport.keras.layers.core.*;
@@ -317,6 +319,11 @@ public class KerasLayerUtils {
             layer = new KerasELU(layerConfig, enforceTrainingConfig);
         } else if(layerClassName.equals(conf.getLAYER_CLASS_NAME_SOFTMAX())){
             layer = new KerasSoftmax(layerConfig, enforceTrainingConfig);
+        } else if (conf instanceof Keras2LayerConfiguration){
+            Keras2LayerConfiguration k2conf = (Keras2LayerConfiguration)conf;
+            if (layerClassName.equals(k2conf.getTENSORFLOW_OP_LAYER())){
+              layer = new KerasTFOpLayer(layerConfig, enforceTrainingConfig);
+            }
         }
         if (layer == null){
             Class<? extends KerasLayer> customConfig = customLayers.get(layerClassName);
@@ -402,6 +409,16 @@ public class KerasLayerUtils {
     public static String getLayerNameFromConfig(Map<String, Object> layerConfig,
                                                 KerasLayerConfiguration conf)
             throws InvalidKerasConfigurationException {
+        if(conf instanceof Keras2LayerConfiguration){
+            Keras2LayerConfiguration k2conf = (Keras2LayerConfiguration)conf;
+            if (getClassNameFromConfig(layerConfig, conf).equals(((Keras2LayerConfiguration) conf).getTENSORFLOW_OP_LAYER())){
+                if (!layerConfig.containsKey(conf.getLAYER_FIELD_NAME()))
+                    throw new InvalidKerasConfigurationException("Field " + conf.getLAYER_FIELD_NAME()
+                            + " missing from layer config");
+                return (String) layerConfig.get(conf.getLAYER_FIELD_NAME());
+            }
+        }
+
         Map<String, Object> innerConfig = KerasLayerUtils.getInnerLayerConfigFromConfig(layerConfig, conf);
         if (!innerConfig.containsKey(conf.getLAYER_FIELD_NAME()))
             throw new InvalidKerasConfigurationException("Field " + conf.getLAYER_FIELD_NAME()

deeplearning4j/deeplearning4j-modelimport/src/test/java/org/deeplearning4j/nn/modelimport/keras/TFKerasTests.java

@@ -0,0 +1,50 @@
+/*******************************************************************************
+ * Copyright (c) 2020 Konduit K.K.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+package org.deeplearning4j.nn.modelimport.keras;
+
+import org.deeplearning4j.BaseDL4JTest;
+import org.deeplearning4j.nn.graph.ComputationGraph;
+import org.junit.Assert;
+import org.junit.Test;
+import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.factory.Nd4j;
+import org.nd4j.resources.Resources;
+
+import java.io.File;
+import java.util.Arrays;
+
+public class TFKerasTests extends BaseDL4JTest{
+
+    @Test
+    public void testModelWithTFOp1() throws Exception{
+        File f = Resources.asFile("modelimport/keras/tfkeras/reshape.h5");
+       ComputationGraph graph = KerasModelImport.importKerasModelAndWeights(f.getAbsolutePath());
+        INDArray out = graph.outputSingle(Nd4j.zeros(12, 2, 3));
+        Assert.assertArrayEquals(new long[]{12, 3}, out.shape());
+    }
+
+    @Test
+    public void testModelWithTFOp2() throws Exception{
+        File f = Resources.asFile("modelimport/keras/tfkeras/permute.h5");
+        ComputationGraph graph = KerasModelImport.importKerasModelAndWeights(f.getAbsolutePath());
+        INDArray out = graph.outputSingle(Nd4j.zeros(12, 2, 3));
+        // dl4j's feedforward doesn't support 3D output, so batch and time axes gets squashed
+        long[] expectedShape = new long[]{12 * 2, 5};
+        Assert.assertArrayEquals(expectedShape, out.shape());
+    }
+
+}

deeplearning4j/deeplearning4j-nlp-parent/deeplearning4j-nlp-japanese/src/main/java/com/atilika/kuromoji/util/KuromojiBinFilesFetcher.java

@@ -67,7 +67,7 @@ public class KuromojiBinFilesFetcher {
                             new URL("https://dl4jdata.blob.core.windows.net/kuromoji/kuromoji_bin_files.tar.gz"),
                             tarFile);
         }
-        ArchiveUtils.unzipFileTo(tarFile.getAbsolutePath(), rootDir.getAbsolutePath());
+        ArchiveUtils.unzipFileTo(tarFile.getAbsolutePath(), rootDir.getAbsolutePath(), false);
 
         return rootDir.getAbsoluteFile();
     }

deeplearning4j/deeplearning4j-nn/pom.xml

@@ -77,7 +77,11 @@
             <artifactId>nd4j-common</artifactId>
             <version>${nd4j.version}</version>
         </dependency>
-
+        <dependency>
+            <groupId>com.google.code.gson</groupId>
+            <artifactId>gson</artifactId>
+            <version>${gson.version}</version>
+        </dependency>
         <!-- ND4J Shaded Jackson Dependency -->
         <dependency>
             <groupId>org.nd4j</groupId>

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

@@ -4170,6 +4170,7 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                 INDArray[] featuresMasks = next.getFeaturesMaskArrays();
                 INDArray[] labels = next.getLabels();
                 INDArray[] labelMasks = next.getLabelsMaskArrays();
+                List<Serializable> meta = next.getExampleMetaData();
 
                 try (MemoryWorkspace ws = outputWs.notifyScopeEntered()) {
                     INDArray[] out = outputOfLayersDetached(false, FwdPassType.STANDARD, getOutputLayerIndices(), features, featuresMasks, labelMasks, true, false, ws);
@@ -4188,7 +4189,7 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
 
                         try (MemoryWorkspace wsO = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
                             for (IEvaluation evaluation : evalsThisOutput)
-                                evaluation.eval(currLabel, currOut, next.getLabelsMaskArray(i));
+                                evaluation.eval(currLabel, currOut, next.getLabelsMaskArray(i), meta);
                         }
                     }
                 }

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/util/ComputationGraphUtil.java

@@ -23,6 +23,9 @@ import org.nd4j.linalg.dataset.api.MultiDataSet;
 import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
 import org.nd4j.linalg.dataset.api.iterator.MultiDataSetIterator;
 
+import java.io.Serializable;
+import java.util.List;
+
 public class ComputationGraphUtil {
 
     private ComputationGraphUtil() {}
@@ -33,13 +36,16 @@ public class ComputationGraphUtil {
         INDArray l = dataSet.getLabels();
         INDArray fMask = dataSet.getFeaturesMaskArray();
         INDArray lMask = dataSet.getLabelsMaskArray();
+        List<Serializable> meta = dataSet.getExampleMetaData();
 
         INDArray[] fNew = f == null ? null : new INDArray[] {f};
         INDArray[] lNew = l == null ? null : new INDArray[] {l};
         INDArray[] fMaskNew = (fMask != null ? new INDArray[] {fMask} : null);
         INDArray[] lMaskNew = (lMask != null ? new INDArray[] {lMask} : null);
 
-        return new org.nd4j.linalg.dataset.MultiDataSet(fNew, lNew, fMaskNew, lMaskNew);
+        org.nd4j.linalg.dataset.MultiDataSet mds = new org.nd4j.linalg.dataset.MultiDataSet(fNew, lNew, fMaskNew, lMaskNew);
+        mds.setExampleMetaData(meta);
+        return mds;
     }
 
     /** Convert a DataSetIterator to a MultiDataSetIterator, via an adaptor class */

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/AbstractLayer.java

@@ -62,6 +62,7 @@ public abstract class AbstractLayer<LayerConfT extends org.deeplearning4j.nn.con
 
     public AbstractLayer(NeuralNetConfiguration conf, DataType dataType) {
         this.conf = conf;
+        if (conf != null)
         cacheMode = conf.getCacheMode();
         this.dataType = dataType;
     }

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

@@ -25,14 +25,11 @@ import lombok.val;
 import org.apache.commons.lang3.ArrayUtils;
 import org.apache.commons.lang3.StringUtils;
 import org.bytedeco.javacpp.Pointer;
-import org.nd4j.adapters.OutputAdapter;
-import org.nd4j.linalg.dataset.AsyncDataSetIterator;;
 import org.deeplearning4j.datasets.iterator.MultiDataSetWrapperIterator;
-import org.deeplearning4j.eval.RegressionEvaluation;
 import org.deeplearning4j.exception.DL4JException;
 import org.deeplearning4j.exception.DL4JInvalidInputException;
-import org.deeplearning4j.nn.api.*;
 import org.deeplearning4j.nn.api.Updater;
+import org.deeplearning4j.nn.api.*;
 import org.deeplearning4j.nn.api.layers.IOutputLayer;
 import org.deeplearning4j.nn.api.layers.RecurrentLayer;
 import org.deeplearning4j.nn.conf.*;
@@ -44,8 +41,8 @@ import org.deeplearning4j.nn.gradient.Gradient;
 import org.deeplearning4j.nn.graph.ComputationGraph;
 import org.deeplearning4j.nn.layers.FrozenLayer;
 import org.deeplearning4j.nn.layers.FrozenLayerWithBackprop;
-import org.deeplearning4j.nn.layers.recurrent.BidirectionalLayer;
 import org.deeplearning4j.nn.layers.LayerHelper;
+import org.deeplearning4j.nn.layers.recurrent.BidirectionalLayer;
 import org.deeplearning4j.nn.layers.wrapper.BaseWrapperLayer;
 import org.deeplearning4j.nn.updater.UpdaterCreator;
 import org.deeplearning4j.nn.workspace.ArrayType;
@@ -58,19 +55,23 @@ import org.deeplearning4j.util.CrashReportingUtil;
 import org.deeplearning4j.util.ModelSerializer;
 import org.deeplearning4j.util.NetworkUtils;
 import org.deeplearning4j.util.OutputLayerUtil;
+import org.nd4j.adapters.OutputAdapter;
 import org.nd4j.base.Preconditions;
 import org.nd4j.evaluation.IEvaluation;
 import org.nd4j.evaluation.classification.Evaluation;
 import org.nd4j.evaluation.classification.ROC;
 import org.nd4j.evaluation.classification.ROCMultiClass;
+import org.nd4j.evaluation.regression.RegressionEvaluation;
 import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.api.memory.MemoryWorkspace;
+import org.nd4j.linalg.api.memory.abstracts.DummyWorkspace;
 import org.nd4j.linalg.api.memory.conf.WorkspaceConfiguration;
 import org.nd4j.linalg.api.memory.enums.AllocationPolicy;
 import org.nd4j.linalg.api.memory.enums.LearningPolicy;
 import org.nd4j.linalg.api.memory.enums.ResetPolicy;
 import org.nd4j.linalg.api.memory.enums.SpillPolicy;
 import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.dataset.AsyncDataSetIterator;
 import org.nd4j.linalg.dataset.DataSet;
 import org.nd4j.linalg.dataset.api.MultiDataSet;
 import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
@@ -84,7 +85,6 @@ import org.nd4j.linalg.heartbeat.reports.Task;
 import org.nd4j.linalg.heartbeat.utils.EnvironmentUtils;
 import org.nd4j.linalg.heartbeat.utils.TaskUtils;
 import org.nd4j.linalg.indexing.NDArrayIndex;
-import org.nd4j.linalg.api.memory.abstracts.DummyWorkspace;
 import org.nd4j.linalg.primitives.Pair;
 import org.nd4j.linalg.primitives.Triple;
 import org.nd4j.linalg.schedule.ISchedule;
@@ -96,6 +96,8 @@ import org.nd4j.util.OneTimeLogger;
 import java.io.*;
 import java.util.*;
 
+;
+
 
 /**
  * MultiLayerNetwork is a neural network with multiple layers in a stack, and usually an output layer.<br>
@@ -3315,19 +3317,39 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
      * @param iterator Iterator to evaluate on
      * @return Evaluation object; results of evaluation on all examples in the data set
      */
-    public <T extends Evaluation> T evaluate(DataSetIterator iterator) {
+    public <T extends Evaluation> T evaluate(@NonNull DataSetIterator iterator) {
         return (T)evaluate(iterator, null);
     }
 
+    /**
+     * Evaluate the network (classification performance).
+     * Can only be used with MultiDataSetIterator instances with a single input/output array
+     *
+     * @param iterator Iterator to evaluate on
+     * @return Evaluation object; results of evaluation on all examples in the data set
+     */
+    public Evaluation evaluate(@NonNull MultiDataSetIterator iterator) {
+        return evaluate(new MultiDataSetWrapperIterator(iterator));
+    }
+
     /**
      * Evaluate the network for regression performance
      * @param iterator Data to evaluate on
-     * @return
+     * @return Regression evaluation
      */
     public <T extends RegressionEvaluation> T evaluateRegression(DataSetIterator iterator) {
         return (T)doEvaluation(iterator, new RegressionEvaluation(iterator.totalOutcomes()))[0];
     }
 
+    /**
+     * Evaluate the network for regression performance
+     * Can only be used with MultiDataSetIterator instances with a single input/output array
+     * @param iterator Data to evaluate on
+     */
+    public org.nd4j.evaluation.regression.RegressionEvaluation evaluateRegression(MultiDataSetIterator iterator) {
+        return evaluateRegression(new MultiDataSetWrapperIterator(iterator));
+    }
+
     /**
      * @deprecated To be removed - use {@link #evaluateROC(DataSetIterator, int)} to enforce selection of appropriate ROC/threshold configuration
      */
@@ -3424,6 +3446,7 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
             INDArray labels = next.getLabels();
             INDArray fMask = next.getFeaturesMaskArray();
             INDArray lMask = next.getLabelsMaskArray();
+            List<Serializable> meta = next.getExampleMetaData();
 
 
             if (!useRnnSegments) {
@@ -3433,7 +3456,7 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
 
                     try (MemoryWorkspace wsO = Nd4j.getWorkspaceManager().scopeOutOfWorkspaces()) {
                         for (T evaluation : evaluations)
-                            evaluation.eval(labels, out, lMask);
+                            evaluation.eval(labels, out, lMask, meta);
                     }
                 }
             } else {

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/optimize/solvers/accumulation/encoding/threshold/AdaptiveThresholdAlgorithm.java

@@ -222,8 +222,11 @@ public class AdaptiveThresholdAlgorithm implements ThresholdAlgorithm {
             if(a == null || Double.isNaN(a.lastThreshold))
                 return;
 
+
             lastThresholdSum += a.lastThreshold;
+            if (!Double.isNaN(a.lastSparsity)) {
                 lastSparsitySum += a.lastSparsity;
+            }
             count++;
         }
 

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/pom.xml

@@ -38,16 +38,22 @@
             <artifactId>nd4j-aeron</artifactId>
             <version>${nd4j.version}</version>
         </dependency>
-        <dependency>
-            <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-parameter-server-node_2.11</artifactId>
-            <version>${nd4j.version}</version>
-        </dependency>
         <dependency>
             <groupId>org.deeplearning4j</groupId>
             <artifactId>dl4j-spark_2.11</artifactId>
             <version>${project.version}</version>
         </dependency>
+        <dependency>
+            <groupId>org.nd4j</groupId>
+            <artifactId>nd4j-parameter-server-node_2.11</artifactId>
+            <version>${nd4j.version}</version>
+            <exclusions>
+                <exclusion>
+                    <groupId>net.jpountz.lz4</groupId>
+                    <artifactId>lz4</artifactId>
+                </exclusion>
+            </exclusions>
+        </dependency>
         <dependency>
             <groupId>org.projectlombok</groupId>
             <artifactId>lombok</artifactId>

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/iterators/VirtualMultiDataSetIterator.java

@@ -23,6 +23,7 @@ import org.nd4j.linalg.dataset.api.iterator.ParallelMultiDataSetIterator;
 
 import java.util.Iterator;
 import java.util.List;
+import java.util.concurrent.atomic.AtomicInteger;
 
 /**
  * This MultiDataSetIterator implementation does accumulation of MultiDataSets from different Spark executors, wrt Thread/Device Affinity
@@ -32,14 +33,16 @@ import java.util.List;
 public class VirtualMultiDataSetIterator implements ParallelMultiDataSetIterator {
 
     protected final List<Iterator<MultiDataSet>> iterators;
+    protected final AtomicInteger position;
 
     public VirtualMultiDataSetIterator(@NonNull List<Iterator<MultiDataSet>> iterators) {
         this.iterators = iterators;
+        this.position = new AtomicInteger(0);
     }
 
     @Override
     public MultiDataSet next(int num) {
-        return null;
+        return next();
     }
 
     @Override
@@ -59,27 +62,34 @@ public class VirtualMultiDataSetIterator implements ParallelMultiDataSetIterator
 
     @Override
     public boolean asyncSupported() {
-        return false;
+        return true;
     }
 
     @Override
     public void reset() {
-
+        throw new UnsupportedOperationException();
     }
 
     @Override
     public boolean hasNext() {
-        return false;
+        // just checking if that's not the last iterator, or if that's the last one - check if it has something
+        boolean ret = position.get() < iterators.size() - 1
+                || (position.get() < iterators.size() && iterators.get(position.get()).hasNext());
+        return ret;
     }
 
     @Override
     public MultiDataSet next() {
-        return null;
+        // TODO: this solution isn't ideal, it assumes non-empty iterators all the time. Would be nice to do something here
+        if (!iterators.get(position.get()).hasNext())
+            position.getAndIncrement();
+
+        return iterators.get(position.get()).next();
     }
 
     @Override
     public void remove() {
-
+        // no-op
     }
 
     @Override

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/main/java/org/deeplearning4j/spark/parameterserver/pw/SharedTrainingWrapper.java

@@ -109,6 +109,7 @@ public class SharedTrainingWrapper {
 
         // now we're creating DataSetIterators, to feed ParallelWrapper
         iteratorDS = new VirtualDataSetIterator(iteratorsDS);
+        iteratorMDS = new VirtualMultiDataSetIterator(iteratorsMDS);
     }
 
     public static synchronized SharedTrainingWrapper getInstance(long id) {
@@ -447,17 +448,19 @@ public class SharedTrainingWrapper {
                 throw new DL4JInvalidConfigException("No iterators were defined for training");
 
             try {
-                while((iteratorDS != null && iteratorDS.hasNext()) || (iteratorMDS != null && iteratorMDS.hasNext())) {
+                boolean dsNext;
+                boolean mdsNext;
+                while((dsNext = iteratorDS != null && iteratorDS.hasNext()) || (mdsNext = iteratorMDS != null && iteratorMDS.hasNext())) {
                     //Loop as a guard against concurrent modifications and RCs
 
                     if (wrapper != null) {
-                        if (iteratorDS != null)
+                        if (dsNext)
                             wrapper.fit(iteratorDS);
                         else
                             wrapper.fit(iteratorMDS);
                     } else {
                         // if wrapper is null, we're fitting standalone model then
-                        if (iteratorDS != null) {
+                        if (dsNext) {
                             if (model instanceof ComputationGraph) {
                                 ((ComputationGraph) originalModel).fit(iteratorDS);
                             } else if (model instanceof MultiLayerNetwork) {
@@ -472,6 +475,7 @@ public class SharedTrainingWrapper {
                         }
                     }
 
+                    if(consumer != null)
                         consumer.getUpdatesQueue().purge();
                 }
             } catch (Throwable t){

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/test/java/org/deeplearning4j/spark/parameterserver/BaseSparkTest.java

@@ -116,8 +116,7 @@ public abstract class BaseSparkTest extends BaseDL4JTest implements Serializable
     }
 
     protected int numExecutors() {
-        int numProc = Runtime.getRuntime().availableProcessors();
+        return 4;
-        return Math.min(4, numProc);
     }
 
     protected MultiLayerConfiguration getBasicConf() {

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/test/java/org/deeplearning4j/spark/parameterserver/train/GradientSharingTrainingTest.java

@@ -49,6 +49,7 @@ import org.junit.rules.TemporaryFolder;
 import org.nd4j.linalg.activations.Activation;
 import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.dataset.DataSet;
+import org.nd4j.linalg.dataset.api.MultiDataSet;
 import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
 import org.nd4j.linalg.indexing.NDArrayIndex;
 import org.nd4j.linalg.learning.config.AMSGrad;
@@ -66,20 +67,26 @@ import java.util.concurrent.ConcurrentHashMap;
 import static org.junit.Assert.*;
 
 @Slf4j
-@Ignore("AB 2019/05/21 - Failing - Issue #7657")
+//@Ignore("AB 2019/05/21 - Failing - Issue #7657")
 public class GradientSharingTrainingTest extends BaseSparkTest {
 
     @Rule
     public TemporaryFolder testDir = new TemporaryFolder();
 
+    @Override
+    public long getTimeoutMilliseconds() {
+        return 90000L;
+    }
+
     @Test
     public void trainSanityCheck() throws Exception {
 
+        for(boolean mds : new boolean[]{false, true}) {
             INDArray last = null;
             INDArray lastDup = null;
             for (String s : new String[]{"paths", "direct", "export"}) {
                 System.out.println("--------------------------------------------------------------------------------------------------------------");
-            log.info("Starting: {}", s);
+                log.info("Starting: {} - {}", s, (mds ? "MultiDataSet" : "DataSet"));
                 boolean isPaths = "paths".equals(s);
 
                 RDDTrainingApproach rddTrainingApproach;
@@ -144,7 +151,11 @@ public class GradientSharingTrainingTest extends BaseSparkTest {
                     DataSet d = iter.next();
                     if (isPaths) {
                         File out = new File(f, count + ".bin");
+                        if(mds){
+                            d.toMultiDataSet().save(out);
+                        } else {
                             d.save(out);
+                        }
                         String path = "file:///" + out.getAbsolutePath().replaceAll("\\\\", "/");
                         paths.add(path);
                     }
@@ -160,6 +171,27 @@ public class GradientSharingTrainingTest extends BaseSparkTest {
 
                     INDArray paramsBefore = sparkNet.getNetwork().params().dup();
                     ComputationGraph after;
+                    if(mds) {
+                        //Fitting from MultiDataSet
+                        List<MultiDataSet> mdsList = new ArrayList<>();
+                        for(DataSet d : ds){
+                            mdsList.add(d.toMultiDataSet());
+                        }
+                        switch (s) {
+                            case "direct":
+                            case "export":
+                                JavaRDD<MultiDataSet> dsRDD = sc.parallelize(mdsList);
+                                after = sparkNet.fitMultiDataSet(dsRDD);
+                                break;
+                            case "paths":
+                                JavaRDD<String> pathRdd = sc.parallelize(paths);
+                                after = sparkNet.fitPathsMultiDataSet(pathRdd);
+                                break;
+                            default:
+                                throw new RuntimeException();
+                        }
+                    } else {
+                        //Fitting from DataSet
                         switch (s) {
                             case "direct":
                             case "export":
@@ -173,6 +205,7 @@ public class GradientSharingTrainingTest extends BaseSparkTest {
                             default:
                                 throw new RuntimeException();
                         }
+                    }
 
                     INDArray paramsAfter = after.params();
                     System.out.println(Arrays.toString(paramsBefore.get(NDArrayIndex.point(0), NDArrayIndex.interval(0, 256)).dup().data().asFloat()));
@@ -199,6 +232,7 @@ public class GradientSharingTrainingTest extends BaseSparkTest {
                 lastDup = last.dup();
             }
         }
+    }
 
 
     @Test
@@ -289,7 +323,7 @@ public class GradientSharingTrainingTest extends BaseSparkTest {
     }
 
 
-    @Test
+    @Test @Ignore
     public void testEpochUpdating() throws Exception {
         //Ensure that epoch counter is incremented properly on the workers
 
@@ -316,7 +350,7 @@ public class GradientSharingTrainingTest extends BaseSparkTest {
 
         ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder()
                 .seed(12345)
-                .updater(new AMSGrad(0.1))
+                .updater(new AMSGrad(0.001))
                 .graphBuilder()
                 .addInputs("in")
                 .layer("out", new OutputLayer.Builder().nIn(784).nOut(10).activation(Activation.SOFTMAX)

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/test/resources/log4j.properties

@@ -20,12 +20,12 @@ log4j.appender.Console.layout=org.apache.log4j.PatternLayout
 log4j.appender.Console.layout.ConversionPattern=%d{ABSOLUTE} %-5p ~ %m%n
 
 log4j.appender.org.springframework=DEBUG
-log4j.appender.org.deeplearning4j=DEBUG
+log4j.appender.org.deeplearning4j=INFO
-log4j.appender.org.nd4j=DEBUG
+log4j.appender.org.nd4j=INFO
 
 log4j.logger.org.springframework=INFO
-log4j.logger.org.deeplearning4j=DEBUG
+log4j.logger.org.deeplearning4j=INFO
-log4j.logger.org.nd4j=DEBUG
+log4j.logger.org.nd4j=INFO
 log4j.logger.org.apache.spark=WARN
 
 

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-parameterserver/src/test/resources/logback.xml

@@ -35,7 +35,7 @@
 
     <logger name="org.apache.catalina.core" level="DEBUG" />
     <logger name="org.springframework" level="DEBUG" />
-    <logger name="org.deeplearning4j" level="DEBUG" />
+    <logger name="org.deeplearning4j" level="INFO" />
     <logger name="org.datavec" level="INFO" />
     <logger name="org.nd4j" level="INFO" />
     <logger name="opennlp.uima.util" level="OFF" />

deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-components/pom.xml

@@ -25,10 +25,6 @@
 
     <artifactId>deeplearning4j-ui-components</artifactId>
 
-    <properties>
-        <freemarker.version>2.3.23</freemarker.version>
-    </properties>
-
     <dependencies>
         <dependency>
             <groupId>org.projectlombok</groupId>

deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-ui-components/src/test/java/org/deeplearning4j/ui/TestStandAlone.java

@@ -24,6 +24,7 @@ import org.deeplearning4j.ui.components.chart.style.StyleChart;
 import org.deeplearning4j.ui.components.table.ComponentTable;
 import org.deeplearning4j.ui.components.table.style.StyleTable;
 import org.deeplearning4j.ui.standalone.StaticPageUtil;
+import org.junit.Ignore;
 import org.junit.Test;
 
 import java.awt.*;

deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-vertx/pom.xml

@@ -60,7 +60,7 @@
         <dependency>
             <groupId>org.freemarker</groupId>
             <artifactId>freemarker</artifactId>
-            <version>2.3.29</version>
+            <version>${freemarker.version}</version>
         </dependency>
 
         <dependency>

deeplearning4j/deeplearning4j-ui-parent/deeplearning4j-vertx/src/main/java/org/deeplearning4j/ui/module/train/TrainModule.java

@@ -200,6 +200,7 @@ public class TrainModule implements UIModule {
             }));
             r.add(new Route("/train/:sessionId/info", HttpMethod.GET, (path, rc) -> this.sessionInfoForSession(path.get(0), rc)));
         } else {
+            r.add(new Route("/train", HttpMethod.GET, (path, rc) -> rc.reroute("/train/overview")));
             r.add(new Route("/train/sessions/current", HttpMethod.GET, (path, rc) -> rc.response().end(currentSessionID == null ? "" : currentSessionID)));
             r.add(new Route("/train/sessions/set/:to", HttpMethod.GET, (path, rc) -> this.setSession(path.get(0), rc)));
             r.add(new Route("/train/overview", HttpMethod.GET, (path, rc) -> this.renderFtl("TrainingOverview.html.ftl", rc)));

libnd4j/include/ops/declarable/generic/transforms/merge_add.cpp

@@ -33,7 +33,7 @@ OP_IMPL(mergeadd, -1, 1, false) {
         
     auto output = OUTPUT_VARIABLE(0);
 
-    std::vector<NDArray*> inArrs(block.width());
+    std::vector<const NDArray*> inArrs(block.width());
     
     for(int i = 0; i < block.width(); ++i)
         inArrs[i] = INPUT_VARIABLE(i);
@@ -42,7 +42,6 @@ OP_IMPL(mergeadd, -1, 1, false) {
 
     return Status::OK();
 }
-
 DECLARE_SYN(mergesum, mergeadd);
 DECLARE_SYN(add_n, mergeadd);
 DECLARE_SYN(addn, mergeadd);
@@ -54,6 +53,45 @@ DECLARE_SYN(accumulate_n, mergeadd);
                 ->setAllowedInputTypes(sd::DataType::ANY)
                 ->setAllowedOutputTypes(sd::DataType::ANY);
     }
+
+
+    CUSTOM_OP_IMPL(mergeadd_bp, 2, 1, false, 0, 0) {
+
+        auto inSize = block.width() - 1;
+
+        REQUIRE_OK(this->validateInputDimensionsMatch(block));
+
+        std::vector<NDArray*> outArrs(inSize);
+        
+        const auto gradient = INPUT_VARIABLE(inSize);
+
+        for (int i = 0; i < inSize; ++i) {
+            outArrs[i] = OUTPUT_VARIABLE(i);
+        }
+        helpers::mergeAddBp(block.launchContext(), *gradient, outArrs);
+
+        return Status::OK();
+    }
+
+    DECLARE_TYPES(mergeadd_bp) {
+        getOpDescriptor()
+            ->setAllowedInputTypes(sd::DataType::ANY)
+            ->setAllowedOutputTypes(sd::DataType::ANY);
+    }
+    DECLARE_SHAPE_FN(mergeadd_bp) {
+
+        const int numOfInArrs = block.width() - 1;
+
+        auto shapeList = SHAPELIST();
+
+        for (int e = 0; e < numOfInArrs; e++) {
+            auto inShape = inputShape->at(e);
+            shapeList->push_back(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(ArrayOptions::dataType(inShape), shape::order(inShape), shape::shapeOf(inShape), shape::rank(inShape))));
+        }
+
+        return shapeList;
+    }
+
 }
 }
 

libnd4j/include/ops/declarable/generic/transforms/merge_avg.cpp

@@ -33,7 +33,7 @@ OP_IMPL(mergeavg, -1, 1, false) {
         
     auto output = OUTPUT_VARIABLE(0);
 
-    std::vector<NDArray*> inArrs(block.width());
+    std::vector<const NDArray*> inArrs(block.width());
     
     for(int i = 0; i < block.width(); ++i)
         inArrs[i] = INPUT_VARIABLE(i);
@@ -48,6 +48,44 @@ OP_IMPL(mergeavg, -1, 1, false) {
                 ->setAllowedInputTypes({ALL_FLOATS})
                 ->setAllowedOutputTypes({ALL_FLOATS});
     }
+
+
+    CUSTOM_OP_IMPL(mergeavg_bp, 2, 1, false, 0, 0) {
+
+        auto inSize = block.width() - 1;
+
+        REQUIRE_OK(this->validateInputDimensionsMatch(block));
+
+        std::vector<NDArray*> outArrs(inSize);
+
+        const auto gradient = INPUT_VARIABLE(inSize);
+        
+        for (int i = 0; i < inSize; ++i) {
+            outArrs[i] = OUTPUT_VARIABLE(i);
+        }
+        helpers::mergeAvgBp(block.launchContext(), *gradient, outArrs);
+        return Status::OK();
+    }
+
+    DECLARE_TYPES(mergeavg_bp) {
+        getOpDescriptor()
+            ->setAllowedInputTypes(sd::DataType::ANY)
+            ->setAllowedOutputTypes(sd::DataType::ANY);
+    }
+    DECLARE_SHAPE_FN(mergeavg_bp) {
+
+        const int numOfInArrs = block.width() - 1;
+
+        auto shapeList = SHAPELIST();
+
+        for (int e = 0; e < numOfInArrs; e++) {
+            auto inShape = inputShape->at(e);
+            shapeList->push_back(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(ArrayOptions::dataType(inShape), shape::order(inShape), shape::shapeOf(inShape), shape::rank(inShape))));
+        }
+
+        return shapeList;
+    }
+
 }
 }
 

libnd4j/include/ops/declarable/generic/transforms/merge_max.cpp

@@ -33,7 +33,7 @@ OP_IMPL(mergemax, -1, 1, false) {
         
     auto output = OUTPUT_VARIABLE(0);
 
-    std::vector<NDArray*> inArrs(block.width());
+    std::vector<const NDArray*> inArrs(block.width());
     
     for(int i = 0; i < block.width(); ++i)
         inArrs[i] = INPUT_VARIABLE(i);
@@ -42,7 +42,6 @@ OP_IMPL(mergemax, -1, 1, false) {
 
     return Status::OK();
 }
-
 DECLARE_SYN(MergeMax, mergemax);
 
     DECLARE_TYPES(mergemax) {
@@ -51,6 +50,47 @@ DECLARE_SYN(MergeMax, mergemax);
                 ->setAllowedOutputTypes(sd::DataType::ANY);
     }
 
+
+    CUSTOM_OP_IMPL(mergemax_bp, 2, 1, false, 0, 0) {
+
+        auto inSize = block.width();
+
+        REQUIRE_OK(this->validateInputDimensionsMatch(block));
+
+        std::vector<const NDArray*> inArrs(inSize);
+        std::vector<NDArray*> outArrs(inSize - 1);
+
+        for (int i = 0; i < inSize; ++i)
+            inArrs[i] = INPUT_VARIABLE(i);
+
+        for (int i = 0; i < (inSize - 1); ++i) {
+            outArrs[i] = OUTPUT_NULLIFIED(i);
+        }
+
+        helpers::mergeMaxBp(block.launchContext(), inArrs, outArrs);
+
+        return Status::OK();
+    }
+
+    DECLARE_TYPES(mergemax_bp) {
+        getOpDescriptor()
+            ->setAllowedInputTypes(sd::DataType::ANY)
+            ->setAllowedOutputTypes(sd::DataType::ANY);
+    }
+    DECLARE_SHAPE_FN(mergemax_bp) {
+
+        const int numOfInArrs = block.width() - 1;
+
+        auto shapeList = SHAPELIST();
+        
+        for (int e = 0; e < numOfInArrs; e++) {
+            auto inShape = inputShape->at(e);
+             shapeList->push_back(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(ArrayOptions::dataType(inShape), shape::order(inShape), shape::shapeOf(inShape), shape::rank(inShape))));
+        }
+
+        return shapeList;
+    }
+
 }
 }
 

libnd4j/include/ops/declarable/generic/transforms/merge_max_idx.cpp

@@ -32,7 +32,7 @@ CUSTOM_OP_IMPL(mergemaxindex, -1, 1, false, 0, 0) {
     REQUIRE_OK(this->validateInputDimensionsMatch(block));
     auto output = OUTPUT_VARIABLE(0);
 
-    std::vector<NDArray*> inArrs(block.width());
+    std::vector<const NDArray*> inArrs(block.width());
     
     for(int i = 0; i < block.width(); ++i)
         inArrs[i] = INPUT_VARIABLE(i);

libnd4j/include/ops/declarable/headers/transforms.h

@@ -64,6 +64,7 @@ namespace sd {
 
         #if NOT_EXCLUDED(OP_mergemax)
         DECLARE_OP(mergemax, -1, 1, false);
+        DECLARE_CUSTOM_OP(mergemax_bp, 2, 1, false, 0, 0);
         #endif
         /*
          * Complete tensor with max indices merged from all input tensors list
@@ -78,10 +79,12 @@ namespace sd {
 
         #if NOT_EXCLUDED(OP_mergeadd)
         DECLARE_OP(mergeadd, -1, 1, false);
+        DECLARE_CUSTOM_OP(mergeadd_bp, 2, 1, false, 0, 0);
         #endif
 
         #if NOT_EXCLUDED(OP_mergeavg)
         DECLARE_OP(mergeavg, -1, 1, false);
+        DECLARE_CUSTOM_OP(mergeavg_bp, 2, 1, false, 0, 0);
         #endif
 
         #if NOT_EXCLUDED(OP_scatter_update)

libnd4j/include/ops/declarable/helpers/cpu/clip.cpp

@@ -0,0 +1,274 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+//
+
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/Loops.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void clipByNorm_(NDArray& input, NDArray& output, const std::vector<int>& dimensions, const NDArray& clipNorm, const bool isInplace) {
+
+    const int rank = input.rankOf();
+    const auto norm2 = input.reduceAlongDimension(reduce::Norm2, dimensions);
+
+    const T normActual = norm2.e<T>(0);
+    const T normClip   = clipNorm.e<T>(0);
+
+    if (isInplace) {
+
+        if(norm2.lengthOf() == 1) {
+
+            if(normActual > normClip)
+                input *= (normClip / normActual);
+        }
+        else {
+
+            auto listOfInSubArrs = input.allTensorsAlongDimension(dimensions);
+
+            auto func = PRAGMA_THREADS_FOR {
+                for (auto i = start; i < stop; i++) {
+                    const T iNormActual = norm2.e<T>(i);
+                    if (iNormActual > normClip)
+                        *listOfInSubArrs.at(i) *= normClip / iNormActual;
+                }
+            };
+            samediff::Threads::parallel_tad(func, 0, listOfInSubArrs.size());
+        }
+    }
+    else {
+
+        if(norm2.lengthOf() == 1) {
+
+            if(normActual > normClip)
+                output.assign(input * (normClip / normActual));
+            else
+                output.assign(input);
+        }
+        else {
+
+            auto listOfInSubArrs  = input.allTensorsAlongDimension(dimensions);
+            auto listOfOutSubArrs = output.allTensorsAlongDimension(dimensions);
+
+            auto func = PRAGMA_THREADS_FOR {
+                for (auto i = start; i < stop; i++) {
+                    auto inputSubArr = listOfInSubArrs.at(i);
+                    auto outputSubArr = listOfOutSubArrs.at(i);
+                    outputSubArr->assign(inputSubArr);
+
+                    const T iNormActual = norm2.e<T>(i);
+
+                    if (iNormActual > clipNorm.e<T>(0))
+                        *outputSubArr *= clipNorm / iNormActual;
+                }
+            };
+            samediff::Threads::parallel_tad(func, 0, listOfInSubArrs.size());
+        }
+    }
+}
+
+//////////////////////////////////////////////////////////////////////////
+void clipByNorm(sd::LaunchContext * context, NDArray& input, NDArray& output, const std::vector<int>& dimensions, const NDArray& clipNorm, const bool isInplace) {
+    BUILD_SINGLE_SELECTOR(output.dataType(), clipByNorm_, (input, output, dimensions, clipNorm, isInplace), FLOAT_TYPES);
+}
+
+
+    template <typename T>
+    static void clipByGlobalNorm_(std::vector<NDArray*> const& inputs, double clipNorm, sd::memory::Workspace* workspace, std::vector<NDArray*>& outputs, bool isInplace) {
+        T globalNorm = 0; //NDArrayFactory::create<T>(0, inputs[0]->getContext()); //sqrt(sum([l2norm(t)**2 for t in t_list]))
+//        PRAGMA_OMP_PARALLEL_FOR_SIMD_REDUCTION(sumT : globalNorm)
+        for (size_t i = 0; i < inputs.size(); i++) {
+            auto input = inputs[i];
+            auto l2norm = input->reduceNumber(reduce::Norm2);
+            globalNorm += l2norm.t<T>(0) * l2norm.t<T>(0);
+        }
+
+        //globalNorm.applyTransform(transform::Sqrt, nullptr, nullptr);// = sd::math::nd4j_sqrt(globalNorm);
+        auto normS = sd::math::nd4j_sqrt<T,T>(globalNorm);
+        outputs[inputs.size()]->p(0, normS);
+
+        const T factor = clipNorm / normS;
+
+//        PRAGMA_OMP_PARALLEL_FOR
+        for (size_t e = 0; e < inputs.size(); e++) {
+            // all-reduce
+            auto input = inputs[e];
+            auto output = outputs[e];
+
+            if (normS <= clipNorm) {
+                output->assign(input);
+            }
+            else {
+
+                auto lambda = LAMBDA_T(_x, factor) { return _x * factor; };
+                input->applyLambda<T>(lambda, *output);
+            }
+        }
+    }
+    void clipByGlobalNorm(sd::LaunchContext * context, std::vector<NDArray*> const& inputs, double clipNorm, sd::memory::Workspace* workspace, std::vector<NDArray*>& outputs, bool isInplace) {
+        BUILD_SINGLE_SELECTOR(outputs[0]->dataType(), clipByGlobalNorm_, (inputs, clipNorm, workspace, outputs, isInplace), FLOAT_TYPES);
+    }
+
+    BUILD_SINGLE_TEMPLATE(template void clipByGlobalNorm_, (std::vector<NDArray*> const& inputs, double clipNorm, sd::memory::Workspace* workspace, std::vector<NDArray*>& outputs, bool isInplace), FLOAT_TYPES);
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void clipByNormBP_(const NDArray& input, const NDArray& gradO, NDArray& gradI /*output*/, const std::vector<int>& dimensions, const NDArray& clipNorm) {
+
+    const int rank = input.rankOf();
+
+    auto norm2 = input.reduceAlongDimension(reduce::Norm2, dimensions);
+
+    if(norm2.lengthOf() == 1) {
+
+        const T N = norm2.e<T>(0);
+
+        auto cn = clipNorm.e<T>(0);
+
+        if(N > cn) {
+
+            const T sumOfProd = (input * gradO).reduceNumber(reduce::Sum).e<T>(0);    // reduce to scalar
+            const T factor1 = static_cast<T>(1.f) / N;
+            const T factor3 = factor1 / (N * N);                                            // 1 / (N*N*N)
+
+            auto lambda = LAMBDA_TT(elem1, elem2, cn, sumOfProd, factor1, factor3) {
+                return cn * (factor1 * elem2 - factor3 * elem1 * sumOfProd);
+            };
+
+            (const_cast<NDArray&>(input)).applyPairwiseLambda<T>(const_cast<NDArray&>(gradO), lambda, gradI);
+        }
+        else
+            gradI.assign(gradO);
+    }
+    else {
+
+        auto gradISubArrs = gradI.allTensorsAlongDimension({dimensions});
+        auto gradOSubArrs = gradO.allTensorsAlongDimension({dimensions});
+        auto inputSubArrs = input.allTensorsAlongDimension({dimensions});
+
+        auto cn = clipNorm.e<T>(0);
+
+        auto func = PRAGMA_THREADS_FOR {
+            for (auto i = start; i < stop; i++) {
+                T N = norm2.e<T>(i);
+
+                auto gradOSubArr = gradOSubArrs.at(i);
+                auto gradISubArr = gradISubArrs.at(i);
+
+                if (N > cn) {
+                    auto inputSubArr = inputSubArrs.at(i);
+                    const T sumOfProd = (*inputSubArr * *gradOSubArr).reduceNumber(reduce::Sum).e<T>(0);    // reduce to scalar
+                    const T factor1 = static_cast<T>(1.f) / N;
+                    const T factor3 = factor1 / (N * N);                                            // 1 / (N*N*N)
+
+                    auto lambda = LAMBDA_TT(elem1, elem2, cn, sumOfProd, factor1, factor3) {
+                        return cn * (factor1 * elem2 - factor3 * elem1 * sumOfProd);
+                    };
+
+                    inputSubArr->applyPairwiseLambda<T>(*gradOSubArr, lambda, *gradISubArr);
+                } else
+                    gradISubArr->assign(gradOSubArr);
+            }
+        };
+        samediff::Threads::parallel_tad(func, 0, gradISubArrs.size());
+    }
+}
+
+    void clipByNormBP(sd::LaunchContext * context, const NDArray& input, const NDArray& gradO, NDArray& gradI /*output*/, const std::vector<int>& dimensions, const NDArray& clipNorm) {
+        BUILD_SINGLE_SELECTOR(gradI.dataType(), clipByNormBP_, (input, gradO, gradI, dimensions, clipNorm), FLOAT_TYPES);
+    }
+
+    BUILD_SINGLE_TEMPLATE(template void clipByNormBP_, (const NDArray& input, const NDArray& gradO, NDArray& gradI /*output*/, const std::vector<int>& dimensions, const NDArray& clipNorm), FLOAT_TYPES);
+
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void clipByAveraged_(NDArray& input, NDArray& output, const std::vector<int>& dimensions, const NDArray& clipNorm, const bool isInplace) {
+
+    auto cn = clipNorm.e<T>(0);
+    if (dimensions.size() == 0) {
+        // all-reduce
+        T n2 = input.reduceNumber(reduce::Norm2).e<T>(0) / input.lengthOf();
+        if (n2 <= cn) {
+            if (!isInplace)
+                output.assign(input);
+        }
+        else {
+            const T factor = cn / n2;
+            auto lambda = LAMBDA_T(_x, factor) { return _x * factor; };
+            input.applyLambda<T>(lambda, output);
+        }
+    }
+    else {
+        // along dimension
+        auto norm2 = input.reduceAlongDimension(reduce::Norm2, dimensions, false);
+        if (!isInplace)
+                output.assign(input);
+        auto tads = output.allTensorsAlongDimension(dimensions);
+        // TODO: make this CUDA-compliant somehow
+        for (int e = 0; e < tads.size(); e++) {
+            T n2 = norm2.e<T>(e) / tads.at(e)->lengthOf();
+            const T factor = cn / n2;
+            if (n2 > cn) {
+                auto lambda = LAMBDA_T(_x, factor) {return _x * factor;};
+                tads.at(e)->applyLambda<T>(lambda, output);
+            }
+        }
+    }
+}
+
+    void clipByAveraged(sd::LaunchContext * context, NDArray& input, NDArray& output, const std::vector<int>& dimensions, const NDArray& clipNorm, const bool isInplace) {
+        BUILD_SINGLE_SELECTOR(input.dataType(), clipByAveraged_, (input, output, dimensions, clipNorm, isInplace), FLOAT_TYPES);
+    }
+
+    BUILD_SINGLE_TEMPLATE(template void clipByAveraged_, (NDArray& input, NDArray& output, const std::vector<int>& dimensions, const NDArray& clipNorm, const bool isInplace), FLOAT_TYPES);
+
+/*
+    if (d1 > params[1])
+    return params[1];
+    else if (d1 < params[0])
+    return params[0];
+    else return d1;
+*/
+
+    template <typename T>
+    static void clipByValue_(NDArray& input, double leftBound, double rightBound, NDArray& output) {
+        auto routine = LAMBDA_T(_x, leftBound, rightBound) {
+            if (_x > rightBound) return rightBound;
+            if (_x < leftBound)  return leftBound;
+            return _x;
+        };
+
+        input.applyLambda<T>(routine, output);
+    }
+
+    void clipByValue(sd::LaunchContext * context, NDArray& input, double leftBound, double rightBound, NDArray& output) {
+        BUILD_SINGLE_SELECTOR(input.dataType(), clipByValue_, (input, leftBound, rightBound, output), FLOAT_TYPES);
+    }
+
+    BUILD_SINGLE_TEMPLATE(template void clipByValue_, (NDArray& input, double leftBound, double rightBound, NDArray& output);, FLOAT_TYPES);
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/eye.cpp

@@ -0,0 +1,45 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+//
+
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/Loops.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+//////////////////////////////////////////////////////////////////////////
+void eye(sd::LaunchContext * context, NDArray& output) {
+
+    const int rank = output.rankOf();
+    auto arrs = output.allTensorsAlongDimension({rank-2, rank-1});
+
+    auto func = PRAGMA_THREADS_FOR {
+        for (auto i = start; i < stop; i++)
+            arrs.at(i)->setIdentity();
+    };
+
+    samediff::Threads::parallel_tad(func, 0, arrs.size());
+}
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/gatherTransforms.cpp

@@ -0,0 +1,183 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+//
+
+
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/ShapeUtils.h>
+#include <numeric>
+#include <helpers/Loops.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+
+////////////////////////////////////////////////////////////////////////
+template<typename X, typename Y>
+static void gatherND_(NDArray& input, NDArray& indices, NDArray& output) {
+
+    const X* x = reinterpret_cast<X*>(input.getBuffer());
+    const Y* y = reinterpret_cast<Y*>(indices.getBuffer());
+          X* z = reinterpret_cast<X*>(output.getBuffer());
+
+    const int xRank    = input.rankOf();
+    const int yRank    = indices.rankOf();
+    const int zRank    = output.rankOf();
+    const int maxRank  = sd::math::nd4j_max<int>(yRank, sd::math::nd4j_max<int>(xRank, zRank));
+
+    const Nd4jLong zLen = output.lengthOf();
+
+    const uint yLastDim = indices.sizeAt(-1);
+
+    const int diff = zRank - xRank;
+    const bool bEqual = yLastDim == xRank;
+
+    auto func = PRAGMA_THREADS_FOR {
+
+        int xCoords[MAX_RANK], zCoords[MAX_RANK], temp;
+
+        for (auto i = start; i < stop; i++) {
+
+            shape::index2coordsCPU(start, i, output.getShapeInfo(), zCoords);
+
+            const auto zOffset = shape::getOffset(output.getShapeInfo(), zCoords);
+
+            temp = zCoords[yRank - 1];
+            zCoords[yRank - 1] = 0;
+            const auto yOffset = shape::getOffset(indices.getShapeInfo(), zCoords);
+            zCoords[yRank - 1] = temp;
+
+            if(bEqual)
+                memcpy(xCoords, zCoords, zRank * sizeof(int));
+            else if(diff >= 0)
+                memcpy(xCoords, zCoords + diff, xRank * sizeof(int));
+            else
+                memcpy(xCoords - diff, zCoords, zRank * sizeof(int));
+
+            for (uint j = 0; j < yLastDim; ++j)
+                xCoords[j] = y[yOffset + j * indices.stridesOf()[yRank - 1]];   // last stride
+
+            const auto xOffset = shape::getOffset(input.getShapeInfo(), xCoords);
+
+            z[zOffset] = x[xOffset];
+        }
+    };
+
+    samediff::Threads::parallel_tad(func, 0, zLen);
+}
+
+////////////////////////////////////////////////////////////////////////
+void gatherND(sd::LaunchContext * context, NDArray& input, NDArray& indices, NDArray& output) {
+    BUILD_DOUBLE_SELECTOR(input.dataType(), indices.dataType(), gatherND_, (input, indices, output), LIBND4J_TYPES, INDEXING_TYPES);
+}
+
+
+////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void gather_(NDArray* input, const NDArray* indices, NDArray* output, const std::vector<int>& intArgs) {
+
+    int axis = intArgs.size() > 0 ? intArgs[0] : 0;
+    const int inputRank = input->rankOf();
+    if(axis < 0)
+        axis += inputRank;
+
+    const int numOfIntArgs = intArgs.size();
+
+    if (indices != nullptr) {
+
+        for(Nd4jLong i = 0; i < indices->lengthOf(); ++i)
+            if(indices->e<Nd4jLong>(i) >= input->sizeAt(axis))
+                throw std::runtime_error("helpers::gather function: indices array contains wrong elements, each element must be smaller than corresponding dimension of input array !");
+
+        // first case: indices consist of only one scalar
+        if(indices->isScalar()) {
+            if(input->rankOf() <= 1){
+                //For scalar indices, rank 0 or 1 input: can't do tensor along dimension 0 as this is whole array... instead, we want to get a scalar
+				auto idx = indices->e<Nd4jLong>(0);
+				auto scalarNDArray = input->e(idx);
+                output->assign(scalarNDArray);
+            } else {
+                auto dimensions = ShapeUtils::evalDimsToExclude(input->rankOf(), {axis});
+                auto tadPack = sd::ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(), dimensions);
+
+                auto tadArr = NDArray(reinterpret_cast<void *>(reinterpret_cast<T*>(input->getBuffer()) + tadPack.primaryOffsets()[indices->e<Nd4jLong>(0)]), tadPack.primaryShapeInfo(), output->getContext());
+                output->assign(&tadArr);
+			}
+        }
+        else if (input->rankOf() == 1 && indices->isVector()) {
+            // special case
+            auto func = PRAGMA_THREADS_FOR {
+                for (auto e = start; e < stop; e++)
+                    output->p(e, input->e<T>(indices->e<Nd4jLong>(e)));
+            };
+
+            samediff::Threads::parallel_for(func, 0, indices->lengthOf());
+        }
+        else {
+
+            std::vector<int> dimsOut(indices->rankOf());
+            std::iota(dimsOut.begin(), dimsOut.end(), axis);   // fill with axis, axis+1, ... indices->rankOf()-1
+            const Nd4jLong numOfSubArrs = ShapeUtils::getNumOfSubArrs(output->getShapeInfo(), dimsOut);
+
+            auto func = PRAGMA_THREADS_FOR {
+                for (auto i = start; i < stop; i++) {
+                    NDArray subArrOut = (*output)(i, dimsOut);
+                    NDArray subArrIn = (*input)(indices->e<Nd4jLong>(i), {axis});
+                    subArrOut.assign(subArrIn);
+                }
+            };
+
+            samediff::Threads::parallel_tad(func, 0, numOfSubArrs);
+        }
+    }
+    else {
+
+        for(int i = 1; i < numOfIntArgs; ++i)
+            if(intArgs[i] >= input->sizeAt(axis))
+                throw std::runtime_error("helpers::gather function: some of input indexes is larger than corresponding shape of input array !");
+
+        // we only allow scalar/vector case here
+        if (numOfIntArgs == 2) { // scalar case
+            output->assign((*input)(intArgs[1], {axis}));
+        }
+        else { // vector case
+            const Nd4jLong numOfSubArrs = ShapeUtils::getNumOfSubArrs(output->getShapeInfo(), {axis});
+
+            auto func = PRAGMA_THREADS_FOR {
+                for (auto i = start; i < stop; i++) {
+                    NDArray subArrOut = (*output)(i, {axis});
+                    NDArray subArrIn = (*input)(intArgs[i + 1], {axis});
+                    subArrOut.assign(subArrIn);
+                }
+            };
+
+            samediff::Threads::parallel_tad(func, 0, numOfSubArrs);
+        }
+    }
+}
+
+    void gather(NDArray* input, const NDArray* indices, NDArray* output, const std::vector<int>& intArgs) {
+        BUILD_SINGLE_SELECTOR(input->dataType(), gather_, (input, indices, output, intArgs), LIBND4J_TYPES);
+    }
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/invertPermutation.cpp

@@ -0,0 +1,51 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+//
+
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/Loops.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+////////////////////////////////////////////////////////////////////////
+void invertPermutation(sd::LaunchContext * context, const NDArray& input, NDArray& output) {
+
+    std::set<int> uniqueElems;
+    const int length = input.lengthOf();
+
+    for(int i = 0; i < length; ++i) {
+
+        int elem = input.e<int>(i);
+
+        if(!uniqueElems.insert(elem).second)        // this operation forbids us to use #pragma omp
+            throw std::runtime_error("helpers::invertPermutation function: input array contains duplicates !");
+
+        if(elem < 0 || elem > length - 1)
+            throw  std::runtime_error("helpers::invertPermutation function: element of input array is out of range (0, length-1) !");
+
+        output.p<int>(elem, i);
+    }
+}
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/merge.cpp

@@ -0,0 +1,277 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ * Copyright (c) 2019-2020 Konduit K.K.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+// @author Oleh Semeniv (oleg.semeniv@gmail.com)
+//
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/Loops.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void mergeMaxIndex_(const std::vector<const NDArray*>& inArrs, NDArray& output) {
+
+    const Nd4jLong numArgs = inArrs.size();
+    auto x = inArrs[0];
+
+    auto func = PRAGMA_THREADS_FOR {
+        for (auto e = start; e < stop; e++) {
+            T max = -DataTypeUtils::max<T>();
+            Nd4jLong idx = 0;
+
+            for (Nd4jLong i = 0; i < numArgs; i++) {
+                T v = inArrs[i]->e<T>(e);
+                if (v > max) {
+                    max = v;
+                    idx = i;
+                }
+            }
+            output.p(e, idx);
+        }
+    };
+
+    samediff::Threads::parallel_for(func, 0, x->lengthOf());
+}
+
+void mergeMaxIndex(sd::LaunchContext * context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
+    BUILD_SINGLE_SELECTOR(inArrs[0]->dataType(), mergeMaxIndex_, (inArrs, output), LIBND4J_TYPES);
+}
+
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void mergeMax_(const std::vector<const NDArray*>& inArrs, NDArray& output) {
+    
+    const Nd4jLong numArgs = inArrs.size();
+    auto x = inArrs[0];
+
+    auto func = PRAGMA_THREADS_FOR {
+        for (auto e = start; e < stop; e++) {
+            T max = -DataTypeUtils::max<T>();
+            for (Nd4jLong i = 0; i < numArgs; i++) {
+                T v = inArrs[i]->e<T>(e);
+                if (v > max)
+                    max = v;
+            }
+            output.p(e, max);
+        }
+    };
+
+    samediff::Threads::parallel_for(func, 0, x->lengthOf());
+}
+
+void mergeMax(sd::LaunchContext * context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
+    BUILD_SINGLE_SELECTOR(output.dataType(), mergeMax_, (inArrs, output), LIBND4J_TYPES);
+}
+
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void mergeMaxBp_(const std::vector<const NDArray*>& inArrs, std::vector<NDArray*>& outArrs) {
+    
+    // outArrs.size() == inArrs.size() - 1
+    const Nd4jLong numArgs = outArrs.size();
+    // last array is gradient
+    const auto gradient = inArrs[numArgs]->bufferAsT<T>();
+    auto length = inArrs[numArgs]->lengthOf();
+    
+    bool bSameOrderAndEws1 = (1 == inArrs[numArgs]->ews());
+    
+    if (bSameOrderAndEws1) {
+        auto gradOrdering = inArrs[numArgs]->ordering();
+
+        for (int i = 0; i < numArgs; ++i) {
+            bSameOrderAndEws1 &= (gradOrdering == inArrs[i]->ordering());
+            bSameOrderAndEws1 &= (1 == inArrs[i]->ews());
+            bSameOrderAndEws1 &= (gradOrdering == outArrs[i]->ordering());
+            bSameOrderAndEws1 &= (1 == outArrs[i]->ews());
+        }
+    }
+    
+    
+    if(bSameOrderAndEws1){
+        auto func = PRAGMA_THREADS_FOR{
+            for (auto e = start; e < stop; e++) {
+                 T max = -DataTypeUtils::max<T>();
+                 Nd4jLong nMaxIndex = 0;
+                 for (Nd4jLong i = 0; i < numArgs; i++) {
+                     const T* v = inArrs[i]->bufferAsT<T>();
+                     if (v[e] > max) {
+                         max = v[e];
+                         nMaxIndex = i;
+                     }
+                 }
+                 T* z = outArrs[nMaxIndex]->bufferAsT<T>();
+                 z[e] = gradient[e];
+            }
+        };
+
+        samediff::Threads::parallel_for(func, 0, length);
+        return;
+    }
+    
+    auto gradShape = inArrs[numArgs]->getShapeInfo();
+    std::vector<bool> vbSameShaepeAndStrides(numArgs);
+    for (int i = 0; i < numArgs; ++i) {
+        vbSameShaepeAndStrides[i] = shape::haveSameShapeAndStrides(gradShape, inArrs[i]->getShapeInfo());
+    }
+
+    auto func = PRAGMA_THREADS_FOR{
+
+        int coords[MAX_RANK];
+            for (auto e = start; e < stop; e++) {
+
+                 shape::index2coordsCPU(start, e, gradShape, coords);
+
+                 const auto gradOffset =  shape::getOffset(gradShape, coords);
+                 
+                 T max = -DataTypeUtils::max<T>();
+                 Nd4jLong nMaxIndex = 0;
+
+                 for (Nd4jLong i = 0; i < numArgs; i++) {
+                     
+                     const auto xOffset = vbSameShaepeAndStrides[i] ? gradOffset : shape::getOffset(inArrs[i]->getShapeInfo(), coords);
+                     const T* v = inArrs[i]->bufferAsT<T>();
+                     if (v[xOffset] > max) {
+                         max = v[xOffset];
+                         nMaxIndex = i;
+                     }
+                 }
+
+                const auto zOffset = vbSameShaepeAndStrides[nMaxIndex] ? gradOffset : shape::getOffset(outArrs[nMaxIndex]->getShapeInfo(), coords);
+                
+                T* z = outArrs[nMaxIndex]->bufferAsT<T>();
+                z[zOffset] = gradient[gradOffset];
+            }
+    };
+
+    samediff::Threads::parallel_for(func, 0, length);
+    return;
+}
+
+void mergeMaxBp(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, std::vector<NDArray*>& outArrs) {
+    BUILD_SINGLE_SELECTOR(outArrs[0]->dataType(), mergeMaxBp_, (inArrs, outArrs), LIBND4J_TYPES);
+}
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void mergeAvg_(const std::vector<const NDArray*>& inArrs, NDArray& output) {
+    const Nd4jLong numArgs = inArrs.size();
+    const T factor = 1.f / numArgs;
+    auto x = inArrs[0];
+
+    auto func = PRAGMA_THREADS_FOR {
+        for (auto e = start; e < stop; e++) {
+            T sum = 0.;
+            for (Nd4jLong i = 0; i < numArgs; i++) {
+                T v = inArrs[i]->e<T>(e);
+                sum += v;
+            }
+            output.p<T>(e, sum * factor);
+        }
+    };
+
+    samediff::Threads::parallel_for(func, 0, x->lengthOf());
+}
+
+void mergeAvg(sd::LaunchContext * context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
+    BUILD_SINGLE_SELECTOR(output.dataType(), mergeAvg_, (inArrs, output), LIBND4J_TYPES);
+}
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void mergeAvgBp_(const NDArray& gradient, std::vector<NDArray*>& outArrs) {
+
+    const Nd4jLong numArgs = outArrs.size();
+
+    auto func = PRAGMA_THREADS_FOR{
+        for (auto e = start; e < stop; e++) {
+
+            T v = gradient.e<T>(e) / numArgs;
+
+            for (Nd4jLong i = 0; i < numArgs; i++) {
+                outArrs[i]->p<T>(e, v);
+            }
+        }
+    };
+
+    samediff::Threads::parallel_for(func, 0, gradient.lengthOf());
+}
+
+void mergeAvgBp(sd::LaunchContext* context, const NDArray& gradient, std::vector<NDArray*>& outArrs) {
+    BUILD_SINGLE_SELECTOR(gradient.dataType(), mergeAvgBp_, (gradient, outArrs), LIBND4J_TYPES);
+}
+
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void mergeAdd_(const std::vector<const NDArray*>& inArrs, NDArray& output) {
+
+    const Nd4jLong numArgs = inArrs.size();
+    auto x = inArrs[0];
+
+    auto func = PRAGMA_THREADS_FOR {
+        for (auto e = start; e < stop; e++) {
+            T sum = (T) 0.f;
+            for (Nd4jLong i = 0; i < numArgs; i++)
+                sum += inArrs[i]->e<T>(e);
+
+            output.p(e, sum);
+        }
+    };
+
+    samediff::Threads::parallel_for(func, 0, x->lengthOf());
+}
+    void mergeAdd(sd::LaunchContext * context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
+        BUILD_SINGLE_SELECTOR(output.dataType(), mergeAdd_, (inArrs, output), LIBND4J_TYPES);
+    }
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void mergeAddBp_(const NDArray& gradient, std::vector<NDArray*>& outArrs) {
+
+    const Nd4jLong numArgs = outArrs.size();
+
+    auto func = PRAGMA_THREADS_FOR{
+        for (auto e = start; e < stop; e++) {
+
+            T v = gradient.e<T>(e);
+
+            for (Nd4jLong i = 0; i < numArgs; i++) {
+                outArrs[i]->p<T>(e, v);
+            }
+        }
+    };
+
+    samediff::Threads::parallel_for(func, 0, gradient.lengthOf());
+}
+
+void mergeAddBp(sd::LaunchContext* context, const NDArray& gradient, std::vector<NDArray*>& outArrs) {
+    BUILD_SINGLE_SELECTOR(gradient.dataType(), mergeAddBp_, (gradient, outArrs), LIBND4J_TYPES);
+}
+
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/pad.cpp

@@ -0,0 +1,483 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+//
+
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/Loops.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+void pad_(const int mode, const NDArray& input, const NDArray& paddings, NDArray& output, const NDArray& padValue) {
+
+    const T* x = input.bufferAsT<T>();
+          T* z = output.bufferAsT<T>();
+
+    const Nd4jLong* xShape  = input.shapeOf();
+    const Nd4jLong* zShape  = output.shapeOf();
+
+    const int rank = input.rankOf();  // both input and output have the same rank
+    const int rankMinusOne = rank - 1;
+
+    const auto zLen = output.lengthOf();
+
+    if(mode == 0) { // CONSTANT case
+
+        const T padVal = padValue.e<T>(0);
+
+        auto func = PRAGMA_THREADS_FOR {
+
+            int zCoords[MAX_RANK], xCoords[MAX_RANK];
+
+            for (auto i = start; i < stop; i++) {
+
+                shape::index2coordsCPU(start, i, output.getShapeInfo(), zCoords);
+                const auto zOffset = shape::getOffset(output.getShapeInfo(), zCoords);
+
+                memcpy(xCoords, zCoords, rank * sizeof(int));
+
+                bool within = true;
+
+                for (int j = rankMinusOne; j >= 0; --j) {
+
+                    if (xShape[j] == zShape[j])
+                        continue;
+
+                    const auto left = paddings.e<Nd4jLong>(j, 0);
+
+                    if (zCoords[j] < left || zCoords[j] >= left + xShape[j]) {
+                        within = false;
+                        break;
+                    }
+                    else
+                        xCoords[j] = zCoords[j] - left;
+                }
+
+                if (within)
+                    z[zOffset] = x[shape::getOffset(input.getShapeInfo(), xCoords)];
+                else
+                    z[zOffset] = padVal;
+            }
+        };
+
+        samediff::Threads::parallel_tad(func, 0, zLen);
+    }
+    else {  // REFLECT and SYMMETRIC cases
+
+        const Nd4jLong shift1 = mode == 1 ? 0 : 1;         // REFLECT : SYMMETRIC
+        const Nd4jLong shift2 = mode == 1 ? 2 : 1;         // REFLECT : SYMMETRIC
+
+        auto func = PRAGMA_THREADS_FOR {
+
+            int zCoords[MAX_RANK], xCoords[MAX_RANK];
+
+            for (auto i = start; i < stop; i++) {
+
+                shape::index2coordsCPU(start, i, output.getShapeInfo(), zCoords);
+                const auto zOffset = shape::getOffset(output.getShapeInfo(), zCoords);
+
+                memcpy(xCoords, zCoords, rank * sizeof(int));
+
+                for (int j = rankMinusOne; j >= 0; --j) {
+
+                    if (xShape[j] == zShape[j])
+                        continue;
+
+                    xCoords[j] = zCoords[j] - paddings.e<Nd4jLong>(j, 0);                             // are ready to fill middle (within input dimension range)
+
+                    if (xCoords[j] < 0)
+                        xCoords[j] = -xCoords[j] - shift1;                // means fill from left
+                    else if (xCoords[j] >= xShape[j])
+                        xCoords[j] = 2 * xShape[j] - xCoords[j] - shift2; // means fill from right
+                }
+
+                const auto xOffset = shape::getOffset(input.getShapeInfo(), xCoords);
+                z[zOffset] = x[xOffset];
+            }
+        };
+
+        samediff::Threads::parallel_tad(func, 0, zLen);
+    }
+}
+
+// //////////////////////////////////////////////////////////////////////////
+// template<typename T>
+// void pad2_(const int mode, const NDArray& input, const NDArray& paddings, NDArray& output, NDArray const& padValue) {
+
+//     const int rank = output.rankOf();
+//     std::vector<int> dimsToExclude(rank);
+//     std::iota(dimsToExclude.begin(), dimsToExclude.end(), 0);             // fill with 0, 1, ... rank-1
+
+//     Nd4jLong numLeft    = paddings.e<Nd4jLong>(rank-1,0);
+//     Nd4jLong numRight   = paddings.e<Nd4jLong>(rank-1,1);
+//     Nd4jLong inDimSize  = input.sizeAt(rank-1);
+//     Nd4jLong outDimSize = output.sizeAt(rank-1);
+
+//     std::vector<std::vector<Nd4jLong>> outIdx = { std::vector<Nd4jLong>(2*rank), {numLeft, numLeft + inDimSize}, {0, numLeft}, {numLeft + inDimSize, outDimSize} };
+
+//     for(int i = 0; i < rank-1; ++i) {
+//         outIdx[0][2*i]     = paddings.e<Nd4jLong>(i, 0);
+//         outIdx[0][2*i + 1] = outIdx[0][2*i] + input.sizeAt(i);
+//     }
+//     outIdx[0][2*rank-1] = outIdx[0][2*rank-2] = 0;
+
+//     // ***** populate innermost sub-arrays firstly ***** //
+//     dimsToExclude.pop_back();
+
+//     Nd4jLong startL = mode == 1 ? 1 : 0;                            // REFLECT or SYMMETRIC
+//     Nd4jLong startR = mode == 1 ? inDimSize-2 : inDimSize-1;        // REFLECT or SYMMETRIC
+
+//     Nd4jLong numOfSubArrs = ShapeUtils::getNumOfSubArrs(input.getShapeInfo(), dimsToExclude);
+
+//     NDArray outSubArr0 = output(outIdx[0], true);
+
+//     PRAGMA_OMP_PARALLEL_FOR
+//     for(Nd4jLong j = 0; j < numOfSubArrs; ++j) {
+
+//         NDArray outSubArr1   = outSubArr0(j, dimsToExclude);
+//         NDArray inSubArr     = input(j, dimsToExclude);
+//         NDArray outSubArrMid = outSubArr1(outIdx[1]);
+
+//         outSubArrMid.assign(inSubArr);      // assign middle
+
+//         if(mode == 0)  { // CONSTANT
+//             if(numLeft != 0) {
+//                 NDArray temp = outSubArr1(outIdx[2]);
+//                 temp.assign(padValue);                        // assign left
+//             }
+//             if(numRight != 0) {
+//                 NDArray temp = outSubArr1(outIdx[3]);
+//                 temp.assign(padValue);                        // assign right
+//             }
+//         }
+//         else {                                                              // REFLECT or SYMMETRIC
+
+//             for(Nd4jLong k = numLeft-1, e = startL; k >= 0; --k, ++e)     // fill left side
+//                 outSubArr1.t<T>(k) = inSubArr.t<T>(e);
+
+//             for(Nd4jLong k = numLeft + inDimSize, e = startR; k < outDimSize; ++k, --e)     // fill right side
+//                 outSubArr1.t<T>(k) = inSubArr.t<T>(e);
+//         }
+//     }
+
+//     // ***** fill rest of outer sub-arrays ***** //
+//     std::vector<Nd4jLong> outIdxInner(2, 0);
+//     std::vector<Nd4jLong> outIdxOuter(2, 0);
+
+//     for(int i = rankBorder - 1; i >= 0; --i) {
+
+//         dimsToExclude.pop_back();
+
+//         outIdxInner.push_back(0), outIdxInner.push_back(0);
+//         outIdxOuter.push_back(0), outIdxOuter.push_back(0);
+
+//         Nd4jLong numLeft  = paddings.e<Nd4jLong>(i, 0);
+//         Nd4jLong numRight = paddings.e<Nd4jLong>(i, 1);
+
+//         if(numLeft == 0 && numRight == 0)
+//             continue;
+
+//         Nd4jLong inDimSize  = input.sizeAt(i);
+//         Nd4jLong outDimSize = output.sizeAt(i);
+
+//         if(mode == 0) {
+//             outIdxOuter[0] = 0;                   outIdxOuter[1] = numLeft;
+//             outIdxInner[0] = numLeft + inDimSize; outIdxInner[1] = outDimSize;
+//         }
+
+//         startL = mode == 1 ? numLeft + 1 : numLeft;                            // REFLECT or SYMMETRIC
+//         startR = mode == 1 ? numLeft + inDimSize - 2 : numLeft + inDimSize-1;      // REFLECT or SYMMETRIC
+
+//         numOfSubArrs = ShapeUtils::getNumOfSubArrs(output.getShapeInfo(), dimsToExclude);
+
+//         PRAGMA_OMP_PARALLEL_FOR_ARGS(firstprivate(outIdxOuter, outIdxInner))
+//         for(Nd4jLong j = 0; j < numOfSubArrs; ++j) {
+
+//             NDArray outSubArr = output(j, dimsToExclude);
+
+//             if(mode == 0)  { // CONSTANT
+
+//                 if(numLeft != 0) {
+//                     NDArray tempO = outSubArr(outIdxOuter);
+//                     tempO.assign(padValue);                              // assign left
+//                 }
+
+//                 if(numRight != 0) {
+//                     NDArray tempI = outSubArr(outIdxInner);
+//                     tempI.assign(padValue);                              // assign right
+//                 }
+//             }
+//             else {                                                              // REFLECT or SYMMETRIC
+
+//                 for(Nd4jLong k = numLeft-1, e = startL; k >= 0; --k, ++e) {    // fill left side
+//                     outIdxOuter[0] = k;
+//                     outIdxOuter[1] = k+1;
+//                     outIdxInner[0] = e;
+//                     outIdxInner[1] = e+1;
+//                     NDArray outSubArrInner = outSubArr(outIdxInner);
+//                     NDArray outSubArrOuter = outSubArr(outIdxOuter);
+//                     outSubArrOuter.assign(outSubArrInner);
+//                 }
+
+//                 for(Nd4jLong k = numLeft + inDimSize, e = startR; k < outDimSize; ++k, --e) {    // fill right side
+//                     outIdxOuter[0] = k;
+//                     outIdxOuter[1] = k+1;
+//                     outIdxInner[0] = e;
+//                     outIdxInner[1] = e+1;
+//                     NDArray outSubArrInner = outSubArr(outIdxInner);
+//                     NDArray outSubArrOuter = outSubArr(outIdxOuter);
+//                     outSubArrOuter.assign(outSubArrInner);
+//                 }
+//             }
+//         }
+//     }
+// }
+
+void pad(sd::LaunchContext * context, const int mode, const NDArray& input, const NDArray& paddings, NDArray& output, NDArray const& padValue) {
+    BUILD_SINGLE_SELECTOR(input.dataType(), pad_, (mode, input, paddings, output, padValue), LIBND4J_TYPES);
+}
+
+//////////////////////////////////////////////////////////////////////////
+template<typename T>
+static void mirrorPad_(const NDArray& input, const NDArray& paddings, NDArray& output, const int mode) {
+
+    // mode:  0 - REFLECT, else - SYMMETRIC
+    const int reflBorder = (bool)mode ? 1 : 0;
+    const int rank        = input.rankOf();
+    const Nd4jLong outLen = output.lengthOf();
+
+    if(rank <= 1) {
+
+        const Nd4jLong inLen         = input.lengthOf();
+        const auto leftSide          = paddings.e<Nd4jLong>(0);
+        const auto leftSideCorrected = leftSide - reflBorder;
+        const Nd4jLong len           = 2*(inLen-1) + leftSide + reflBorder;
+
+        for(int i = 0; i < outLen; ++i) {
+
+            if (i < leftSide)                                   // left side
+                output.p(i, input.e<T>(leftSideCorrected - i));
+
+            else if(i >= leftSide && i < leftSide + inLen)      // middle
+                output.p(i, input.e<T>(i - leftSide));
+
+            else                                                // right side
+                output.p(i, input.e<T>(len - i));
+        }
+    }
+    else {
+
+        auto func = PRAGMA_THREADS_FOR {
+
+            int inIdx[MAX_RANK], outIdx[MAX_RANK];
+
+            for (auto i = start; i < stop; i++) {
+
+                shape::index2coordsCPU(start, i, output.getShapeInfo(), outIdx);
+
+                for (int j = 0; j < rank; ++j) {
+                    const Nd4jLong inLen = input.sizeAt(j);
+                    const auto leftSide = paddings.e<T>(j, 0);
+                    const auto leftSideCorrected = leftSide - reflBorder;
+                    const Nd4jLong len = 2 * (inLen - 1) + leftSide + reflBorder;
+
+                    if (outIdx[j] < leftSide)                                        // left side
+                        inIdx[j] = leftSideCorrected - outIdx[j];
+
+                    else if (outIdx[j] >= leftSide && outIdx[j] < leftSide + inLen)  // middle
+                        inIdx[j] = outIdx[j] - leftSide;
+
+                    else                                                            // right side
+                        inIdx[j] = len - outIdx[j];
+                }
+
+                auto outOffset = shape::getOffset(output.getShapeInfo(), outIdx);
+                auto inOffset = shape::getOffset(input.getShapeInfo(), inIdx);
+                reinterpret_cast<T *>(output.buffer())[outOffset] = reinterpret_cast<T *>(input.getBuffer())[inOffset];
+            }
+        };
+
+        samediff::Threads::parallel_for(func, 0, outLen);
+    }
+}
+
+    void mirrorPad(sd::LaunchContext * context, const NDArray& input, const NDArray& paddings, NDArray& output, const int mode) {
+        BUILD_SINGLE_SELECTOR(input.dataType(), mirrorPad_, (input, paddings, output, mode), LIBND4J_TYPES);
+    }
+
+    BUILD_SINGLE_TEMPLATE(template void mirrorPad_, (const NDArray& input, const NDArray& paddings, NDArray& output, const int mode), LIBND4J_TYPES);
+
+
+////////////////////////////////////////////////////////////////////////
+/*// initial values of inIdx, outIdx, dim must be equal to zero
+template<typename T>
+static void recursiveLoopForPad_(const int mode, NDArray& input, const NDArray& paddings, NDArray& output, std::vector<int> dimensions, int dim, int inIdx, int outIdx, NDArray& padValue ) {
+
+    int leftOffset;
+    // dimensions are array of input dimensions, it is sorted in increasing order
+    // every time at the beginning we erase first element from it (not good idea to use vector for this purpose, but luckily it is small enough)
+    // then we use this array for tads building, every time while recursion the number of built tads becomes bigger
+    dimensions.erase(dimensions.begin());
+    // build tad basing on output array, also create auxiliary arrays pointing on required output array ranges
+    shape::TAD tadOut(output.getShapeInfo(), dimensions.data(), dimensions.size());
+    tadOut.createTadOnlyShapeInfo();
+    tadOut.createOffsets();
+    auto subArrOut = NDArray(output.getBuffer(), tadOut.tadOnlyShapeInfo, output.getContext());
+    auto subArr = NDArray(output.getBuffer(), tadOut.tadOnlyShapeInfo, output.getContext());
+    // build tad basing on input array, also create auxiliary array pointing on required input array range
+    shape::TAD tadIn(input.getShapeInfo(), dimensions.data(), dimensions.size());
+    tadIn.createTadOnlyShapeInfo();
+    tadIn.createOffsets();
+    auto subArrIn = NDArray(input.getBuffer(), tadIn.tadOnlyShapeInfo, output.getContext());
+    // these indices take into account recursion and always point to actual tads numbers
+    if (input.rankOf() > 1 && output.rankOf() > 1) {// only for non-vector cases
+        outIdx = outIdx * output.sizeAt(dim + 1);
+        inIdx = inIdx * input.sizeAt(dim + 1);
+    }
+    // current input tad number, we add to it unity in a loop
+    int k = -1;
+    // loop through current dimension
+    for(int i = 0; i < output.sizeAt(dim); ++i) {
+        // corresponds to outer range (relevant indices are absent in input)
+        leftOffset = paddings.e<int>(dim, 0);
+        if(i < leftOffset || i >= (input.sizeAt(dim) + leftOffset))
+            continue;
+
+        // increase input tads number
+        ++k;
+        // recursion condition allows for the fact that tad can't reduce to scalar
+        if(dim < input.rankOf() - 2)
+            recursiveLoopForPad(mode, input, paddings, output, dimensions, dim + 1, inIdx + k, outIdx + i, padValue);
+        else if (paddings.sizeAt(0) > dim + 1){
+            leftOffset = paddings.e<int>(dim + 1, 0);
+            // shift buffers pointers to actual element position
+            if (output.rankOf() > 1) {
+                subArrOut.setBuffer(reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + i]);
+                subArrIn.setBuffer(reinterpret_cast<T*>(input.getBuffer()) + tadIn.tadOffsets[inIdx + i - paddings.e<int>(dim, 0)]);
+            }
+            else {
+                subArrOut.p(i, subArrIn.e<T>(i - leftOffset));
+            }
+            // most inner loop, corresponds to last dim = rank-1
+            switch (mode) {
+                case 0:             // CONSTANT mode
+                    for(int j = 0; j < subArrOut.lengthOf(); ++j)
+                            if(j < leftOffset || j >= (subArrIn.lengthOf() + leftOffset) )                  // firstly fill with zeros outer ranges
+                                subArrOut.p(j, (T)0.f);
+                            else
+                                subArrOut.p(j, subArrIn.e<T>(j - leftOffset));   // fill middle with elements of input array
+                    break;
+
+                case 1:             // REFLECT mode
+                    for(int j = 1;  j <= leftOffset; ++j)                                               // fill firstly left side
+                        subArrOut.p(leftOffset - j, subArrIn.e<T>(j));
+                    for(int j = 0; j < subArrIn.lengthOf(); ++j)                                        // fill middle
+                        subArrOut.p(leftOffset + j, subArrIn.e<T>(j));
+                    for(int j = (subArrOut.lengthOf() - leftOffset); j < subArrOut.lengthOf(); ++j)     // fill right side
+                        subArrOut.p(j, subArrIn.e<T>(subArrOut.lengthOf() - j - 1));
+                    break;
+
+                case 2:             // SYMMETRIC mode
+                    for(int j = 1;  j <= leftOffset; ++j)                                               // fill firstly left side
+                        subArrOut.p(leftOffset - j, subArrIn.e<T>(j-1));
+                    for(int j = 0; j < subArrIn.lengthOf(); ++j)                                        // fill middle
+                        subArrOut.p(leftOffset + j, subArrIn.e<T>(j));
+                    for(int j = (subArrOut.lengthOf() - leftOffset); j < subArrOut.lengthOf(); ++j)     // fill right side
+                        subArrOut.p(j, subArrIn.e<T>(subArrOut.lengthOf() - j));
+                    break;
+            }
+        }
+        else {
+
+             if (mode == 0 && input.rankOf() < 2)
+                 subArrOut.p(i, subArrIn.e<T>(i - leftOffset));   // fill middle with elements of input array
+        }
+    }
+    // populate sub-array formed previously
+    leftOffset = paddings.e<int>(dim,0);
+    switch (mode) {
+        case 0:         // CONSTANT mode
+            for(int j = 1;  j <= leftOffset; ++j) {
+                // fill left side with padValue
+                if (output.rankOf() > 1) {
+                    subArrOut.setBuffer(
+                            reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + leftOffset - j]);
+                    subArrOut.assign(padValue);
+                }
+                else {
+                    subArrOut.p(j - 1, padValue);
+                }
+            }
+//            output.printIndexedBuffer("Output at");
+            for(int j = (output.sizeAt(dim) - leftOffset); j < output.sizeAt(dim); ++j) {       // fill left side with zeros
+                if (output.rankOf() > 1) {
+                    subArrOut.setBuffer(reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + j]);
+                    subArrOut.assign(padValue);
+                }
+                else {
+                    subArrOut.p(j, padValue);
+                }
+            }
+            break;
+
+        case 1:         // REFLECT mode
+            for(int j = 1;  j <= leftOffset; ++j) {                                                     // fill left side
+                subArr.setBuffer(reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + leftOffset + j]);
+                subArrOut.setBuffer(reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + leftOffset - j]);
+                subArrOut.assign(&subArr);
+            }
+            for(int j = (output.sizeAt(dim) - leftOffset); j < output.sizeAt(dim); ++j) {       // fill right side
+                subArr.setBuffer(reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + output.sizeAt(dim) + leftOffset - 1 - j]);
+                subArrOut.setBuffer(reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + j]);
+                subArrOut.assign(&subArr);
+            }
+            break;
+
+        case 2:         // SYMMETRIC mode
+            for(int j = 1;  j <= leftOffset; ++j) {                                                     // fill left side
+                subArr.setBuffer(reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + leftOffset + j - 1]);
+                subArrOut.setBuffer(reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + leftOffset - j]);
+                subArrOut.assign(&subArr);
+            }
+            for(int j = (output.sizeAt(dim) - leftOffset); j < output.sizeAt(dim); ++j) {       // fill right side
+                subArr.setBuffer(reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + output.sizeAt(dim) + leftOffset - j]);
+                subArrOut.setBuffer(reinterpret_cast<T*>(output.getBuffer()) + tadOut.tadOffsets[outIdx + j]);
+                subArrOut.assign(&subArr);
+            }
+            break;
+    }
+}
+ */
+/*
+    void recursiveLoopForPad(const int mode, NDArray& input, const NDArray& paddings, NDArray& output, std::vector<int> dimensions, int dim, int inIdx, int outIdx, NDArray& padValue ) {
+        BUILD_SINGLE_SELECTOR(input.dataType(), recursiveLoopForPad_, (mode, input, paddings, output, dimensions, dim, inIdx, outIdx, padValue), LIBND4J_TYPES);
+    }
+
+    BUILD_SINGLE_TEMPLATE(template void recursiveLoopForPad_, (const int mode, NDArray& input, const NDArray& paddings, NDArray& output, std::vector<int> dimensions, int dim, int inIdx, int outIdx, NDArray& padValue), LIBND4J_TYPES);
+
+*/
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/randomShuffle.cpp

@@ -0,0 +1,126 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+//
+
+
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/Loops.h>
+#include <graph/RandomGenerator.h>
+#include <numeric>
+#include <helpers/ShapeUtils.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+//////////////////////////////////////////////////////////////////////////
+template <typename T>
+void randomShuffle_(NDArray& input, NDArray& output, sd::graph::RandomGenerator& rng, const bool isInplace) {
+
+    // check edge cases first
+    int temp;
+    const int firstDim = input.sizeAt(0);
+    if(input.lengthOf() == 1 || firstDim == 1) {
+
+        if(!isInplace)
+            output.assign(input);
+    }
+    else if (input.isVector() || shape::isLikeVector(input.getShapeInfo(), temp)) {
+
+        // apply Fisher-Yates shuffle
+        if(isInplace) {
+            //PRAGMA_OMP_PARALLEL_FOR_IF((firstDim-1) > Environment::getInstance()->tadThreshold())
+            for(int i = firstDim-1; i > 0; --i) {
+                int r = rng.relativeInt(i) % i;
+                if(i == r)
+                    continue;
+                T t0 = input.t<T>(i);
+                T t1 = input.t<T>(r);
+                //math::nd4j_swap<T>(input(i), input(r));
+                input.t<T>(i) = t1;
+                input.t<T>(r) = t0;
+            }
+        }
+        else {
+            std::vector<int> indices(firstDim);
+            std::iota(indices.begin(), indices.end(), 0);
+            output.p<T>(Nd4jLong(0), input.e<T>(0));
+
+            // FIXME: parallelism!!
+            for(int i = firstDim-1; i > 0; --i) {
+                int r = rng.relativeInt(i) % i;
+                output.t<T>(i) = input.t<T>(indices[r]);
+                if(i == r)
+                    continue;
+
+                output.t<T>(r) = input.t<T>(indices[i]);
+                math::nd4j_swap<int>(indices[i], indices[r]);
+            }
+            rng.rewindH(firstDim-1);
+        }
+    }
+    else {
+
+        // evaluate sub-arrays list of input array through all dimensions excluding first one
+        std::vector<int> dimensions = ShapeUtils::evalDimsToExclude(input.rankOf(), {0});
+        auto subArrsListIn = input.allTensorsAlongDimension(dimensions);
+
+        // apply Fisher-Yates shuffle
+        if(isInplace) {
+            //PRAGMA_OMP_PARALLEL_FOR_IF((firstDim-1) > Environment::getInstance()->elementwiseThreshold())
+            for(int i = firstDim - 1; i > 0; --i) {
+                int r = rng.relativeInt(i) % i;
+
+                if(i == r)
+                    continue;
+                subArrsListIn.at(i)->swapUnsafe(*subArrsListIn.at(r));
+            }
+        }
+        else {
+            // evaluate sub-arrays list of output array through all dimensions excluding first one
+            auto subArrsListOut = output.allTensorsAlongDimension(dimensions);
+            std::vector<int> indices(firstDim);
+            std::iota(indices.begin(), indices.end(), 0);
+            bool isZeroShuffled = false;
+            //PRAGMA_OMP_PARALLEL_FOR_IF((firstDim-1) > Environment::getInstance()->tadThreshold())
+            for(int i = firstDim - 1; i > 0; --i) {
+                int r = rng.relativeInt(i) % i;
+                subArrsListOut.at(i)->assign(subArrsListIn.at(indices[r]));
+                if(r == 0)
+                    isZeroShuffled = true;
+                if(i == r)
+                    continue;
+                subArrsListOut.at(r)->assign(subArrsListIn.at(indices[i]));
+                math::nd4j_swap<int>(indices[i], indices[r]);
+            }
+            if(!isZeroShuffled)
+                subArrsListOut.at(0)->assign(subArrsListIn.at(0));
+        }
+        rng.rewindH(firstDim-1);
+    }
+
+}
+
+    void randomShuffle(sd::LaunchContext * context, NDArray& input, NDArray& output, sd::graph::RandomGenerator& rng, const bool isInplace) {
+        BUILD_SINGLE_SELECTOR(input.dataType(), randomShuffle_, (input, output, rng, isInplace), LIBND4J_TYPES);
+    }
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/scatterUpdateAndSimple.cpp

@@ -0,0 +1,115 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+//
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/ShapeUtils.h>
+#include <helpers/Loops.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+//////////////////////////////////////////////////////////////////////////
+void scatterUpdate(sd::LaunchContext * context, NDArray& input, NDArray& updates, const std::vector<int>* intArgs) {
+
+    int opCode = (*intArgs)[0];
+    int dimSize = (*intArgs)[1];
+    Nd4jLong e;
+    Nd4jLong limg = 2 + dimSize;
+    std::vector<int> tadDimensions(dimSize);
+    for (e = 2; e < limg; e++)
+        tadDimensions[e-2] = (*intArgs)[e];
+
+    std::vector<int> dimsToExclude = ShapeUtils::evalDimsToExclude(input.rankOf(), tadDimensions);
+
+    // increasing counter to skip numIndices
+    e++;
+    std::vector<int> indices;
+    for (; e < static_cast<Nd4jLong>(intArgs->size()); e++)
+        indices.push_back((*intArgs)[e]);
+
+    auto func = PRAGMA_THREADS_FOR {
+        for (auto i = start; i < stop; i++) {
+            auto inSubArr = input(indices[i], dimsToExclude, true);
+            auto updSubArr = updates(i, dimsToExclude, true);
+
+            if (inSubArr.lengthOf() != updSubArr.lengthOf())
+                continue;
+
+            switch (opCode) {
+                case 0:
+                    inSubArr.applyPairwiseTransform(pairwise::Add, updSubArr, inSubArr);
+                    break;
+                case 1:
+                    inSubArr.applyPairwiseTransform(pairwise::Subtract, updSubArr, inSubArr);
+                    break;
+                case 2:
+                    inSubArr.applyPairwiseTransform(pairwise::Multiply, updSubArr, inSubArr);
+                    break;
+                case 3:
+                    inSubArr.applyPairwiseTransform(pairwise::Divide, updSubArr, inSubArr);
+                    break;
+                case 4:
+                    inSubArr.applyPairwiseTransform(pairwise::ReverseSubtract, updSubArr, inSubArr);
+                    break;
+                case 5:
+                    inSubArr.applyPairwiseTransform(pairwise::ReverseDivide, updSubArr, inSubArr);
+                    break;
+                case 6:
+                    inSubArr.applyPairwiseTransform(pairwise::CopyPws, updSubArr, inSubArr);
+                    break;
+                default:
+                    continue;
+            }
+        }
+    };
+
+    samediff::Threads::parallel_tad(func, 0, indices.size());
+}
+
+
+//////////////////////////////////////////////////////////////////////////
+void scatterSimple(sd::LaunchContext * context, const int opId, NDArray& input, const NDArray& updates, const NDArray& indices, const std::vector<int>& dimensions) {
+
+    // updates and indices have same length
+    const Nd4jLong len = indices.lengthOf();
+
+    switch (opId) {
+
+        case 6: {   // copy
+            auto func = PRAGMA_THREADS_FOR {
+                for (auto i = start; i < stop; i++) {
+                    auto inSubArr = input(i, dimensions);
+                    inSubArr.p(indices.t<Nd4jLong>(i), updates.e(i));
+                }
+            };
+
+            samediff::Threads::parallel_for(func, 0, len);
+        }
+            break;
+
+        default:
+            throw std::invalid_argument("helpers::scatterSimple: operation is not implemented for given id !");
+    }
+}
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/tile.cpp

@@ -0,0 +1,91 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+//
+
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/ShapeUtils.h>
+#include <helpers/Loops.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+//////////////////////////////////////////////////////////////////////////
+template <typename T>
+static void tileBP_(const NDArray& gradO /*input*/, NDArray& gradI /*output*/, const std::vector<Nd4jLong> reps) {
+
+    T* gradIBuff      = reinterpret_cast<T*>(gradI.getBuffer());
+    const T* gradOBuff      = reinterpret_cast<T*>(gradO.getBuffer());
+    const Nd4jLong gradILen = gradI.lengthOf();
+    const Nd4jLong gradOLen = gradO.lengthOf();  // gradOLen >= gradILen
+    const Nd4jLong gradIEWS = sd::math::nd4j_abs<Nd4jLong>(gradI.ews());
+    const Nd4jLong gradOEWS = gradO.ews();
+
+    // initial zeroing of gradI content
+    if(gradIEWS == 1)
+        memset(gradIBuff, 0, gradILen * sizeof(T));
+    else {
+        //PRAGMA_OMP_PARALLEL_FOR_SIMD
+        for (Nd4jLong i = 0; i < gradILen * gradIEWS; i += gradIEWS)
+            gradIBuff[i] = static_cast<T>(0.f);
+    }
+
+
+    if(gradO.ordering() == 'c' && gradOEWS == 1) {
+
+        //PRAGMA_OMP_PARALLEL_FOR_SIMD
+        for(Nd4jLong i=0;  i<gradOLen; ++i) {
+            auto idx = shape::subArrayIndex(i, gradO.getShapeInfo(), gradI.getShapeInfo());
+            gradI.p(idx, gradI.e<T>(idx) + gradOBuff[i]);
+        }
+    }
+    else if(gradO.ordering() == 'c' && gradOEWS > 1) {
+
+        //PRAGMA_OMP_PARALLEL_FOR_SIMD
+        for(Nd4jLong i=0;  i<gradOLen; ++i) {
+            auto idx = shape::subArrayIndex(i, gradO.getShapeInfo(), gradI.getShapeInfo());
+            gradI.p(idx, gradI.e<T>(idx) + gradOBuff[i * gradOEWS]);
+        }
+    }
+    else {
+
+        //PRAGMA_OMP_PARALLEL_FOR_SIMD
+        for(Nd4jLong i=0;  i<gradOLen; ++i) {
+
+            auto fidx = shape::subArrayIndex(i, gradO.getShapeInfo(), gradI.getShapeInfo());
+            gradI.p(fidx, gradI.e<T>(fidx) + gradOBuff[shape::getIndexOffset(i, gradO.getShapeInfo())]);
+        }
+    }
+}
+
+void tileBP(sd::LaunchContext * context, const NDArray& gradO /*input*/, NDArray& gradI /*output*/, const std::vector<Nd4jLong> reps) {
+    BUILD_SINGLE_SELECTOR(gradI.dataType(), tileBP_, (gradO, gradI, reps), FLOAT_TYPES);
+}
+
+
+BUILD_SINGLE_TEMPLATE(template void tileBP_, (const NDArray& gradO /*input*/, NDArray& gradI /*output*/, const std::vector<Nd4jLong> reps), FLOAT_TYPES);
+
+
+
+
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/trace.cpp

@@ -0,0 +1,47 @@
+/*******************************************************************************
+ * Copyright (c) 2015-2018 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+//
+// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+//
+
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/Loops.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+//////////////////////////////////////////////////////////////////////////
+template <typename T>
+static void trace_(const NDArray& input, NDArray& output) {
+    const int inRank = input.rankOf();
+    auto setOfSubArrs = input.allTensorsAlongDimension({inRank-2, inRank-1});
+
+    auto func = PRAGMA_THREADS_FOR {
+        for (auto i = start; i < stop; i++)
+            output.p(i, setOfSubArrs.at(i)->getTrace());
+    };
+    samediff::Threads::parallel_for(func, 0, setOfSubArrs.size());
+}
+
+    void trace(sd::LaunchContext * context, const NDArray& input, NDArray& output) {
+        BUILD_SINGLE_SELECTOR(input.dataType(), trace_, (input, output), LIBND4J_TYPES);
+    }
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cpu/triu.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 Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+//
+
+
+#include <ops/declarable/helpers/transforms.h>
+#include <helpers/Loops.h>
+
+namespace sd 	  {
+namespace ops 	  {
+namespace helpers {
+
+
+//////////////////////////////////////////////////////////////////////////
+template <typename T>
+static void triuBP_(sd::LaunchContext * context, const NDArray& input, const NDArray& gradO, NDArray& gradI, const int diagonal) {
+
+    auto dOdI = NDArray(&gradO);                // dO/dI
+    const_cast<NDArray&>(input).fillAsTriangular<T>(0, diagonal, dOdI.sizeAt(-1), dOdI, 'b');
+    int dLen = dOdI.lengthOf();
+
+    auto func = PRAGMA_THREADS_FOR {
+        for (auto i = start; i < stop; i++) {
+            if (dOdI.t<T>(i) != static_cast<T>(0.f))
+                dOdI.t<T>(i) = static_cast<T>(1.f);
+        }
+    };
+    samediff::Threads::parallel_for(func, 0, dLen);
+
+    // FIXME: !!!
+    gradI.assign(dOdI * gradO);                          // chain rule: dLoss/dI = dO/dI * dLoss/dO
+}
+
+    void triuBP(sd::LaunchContext * context, const NDArray& input, const NDArray& gradO, NDArray& gradI, const int diagonal) {
+        BUILD_SINGLE_SELECTOR(gradO.dataType(), triuBP_, (context, input, gradO, gradI, diagonal), LIBND4J_TYPES);
+    }
+
+}
+}
+}

libnd4j/include/ops/declarable/helpers/cuda/merge.cu

@@ -14,9 +14,9 @@
  * SPDX-License-Identifier: Apache-2.0
  ******************************************************************************/
 
-//
+ //
-// @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
+ // @author Yurii Shyrma (iuriish@yahoo.com), created on 20.04.2018
-//
+ //
 
 
 #include<ops/declarable/helpers/transforms.h>
@@ -34,7 +34,7 @@ namespace sd {
         namespace helpers {
             //////////////////////////////////////////////////////////////////////////
             template <typename T, typename Z>
-            static __global__ void global_mergeMaxIndex_(void **inArrs, void **inShapes, const int numArrays, void *voutput, Nd4jLong *outputShape, Nd4jLong length) {
+            static __global__ void mergeMaxIndexCudaLauncher(void** inArrs, void** inShapes, const int numArrays, void* voutput, Nd4jLong* outputShape, Nd4jLong length) {
                 auto output = reinterpret_cast<Z*>(voutput);
 
                 const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
@@ -46,54 +46,56 @@ namespace sd {
 
                     for (int i = 0; i < numArrays; i++) {
                         auto x = reinterpret_cast<T*>(inArrs[i]);
-                        auto xShape = reinterpret_cast<Nd4jLong *>(inShapes[i]);
+                        auto xShape = reinterpret_cast<Nd4jLong*>(inShapes[i]);
                         auto val = x[shape::getIndexOffset(e, xShape)];;
                         if (mVal < val) {
                             mIdx = static_cast<Z>(i);
                             mVal = val;
                         }
                     }
-                    __syncthreads();
                     
                     output[shape::getIndexOffset(e, outputShape)] = mIdx;
                 }
             }
 
             template <typename T, typename Z>
-            static void mergeMaxIndex_(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output) {
+            static void mergeMaxIndex_(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
-                std::vector<void *> inBuffers(inArrs.size());
-                std::vector<void *> inShapes(inArrs.size());
                 
-                for (int e = 0; e < inArrs.size(); e++) {
+                int nArrSize = static_cast<int>(inArrs.size());
+                std::vector<void*> inBuffers(nArrSize), inShapes(nArrSize);
+
+                for (int e = 0; e < nArrSize; e++) {
                     inBuffers[e] = inArrs[e]->getSpecialBuffer();
                     inShapes[e] = inArrs[e]->getSpecialShapeInfo();
                 }
 
                 PointersManager manager(context, "mergeMaxIndex");
 
-                auto pInBuffers = reinterpret_cast<void **>(manager.replicatePointer(inBuffers.data(), inBuffers.size() * sizeof(void *)));
+                auto pInBuffers = reinterpret_cast<void**>(manager.replicatePointer(inBuffers.data(), inBuffers.size() * sizeof(void*)));
-                auto pInShapes = reinterpret_cast<void **>(manager.replicatePointer(inShapes.data(), inShapes.size() * sizeof(void *)));
+                auto pInShapes = reinterpret_cast<void**>(manager.replicatePointer(inShapes.data(), inShapes.size() * sizeof(void*)));
                 auto length = output.lengthOf();
 
-                global_mergeMaxIndex_<T,Z><<<512, 512, 512, *context->getCudaStream()>>>(pInBuffers, pInShapes, (int) inArrs.size(), output.getSpecialBuffer(), output.getSpecialShapeInfo(), length);
+                const int threadsPerBlock = MAX_NUM_THREADS / 2;
+                const int blocksPerGrid = (length + threadsPerBlock - 1) / threadsPerBlock;
+
+                mergeMaxIndexCudaLauncher<T, Z> << <blocksPerGrid, threadsPerBlock, 512, *context->getCudaStream() >> > (pInBuffers, pInShapes, nArrSize, output.getSpecialBuffer(), output.getSpecialShapeInfo(), length);
 
                 manager.synchronize();
             }
 
-            void mergeMaxIndex(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output) {
+            void mergeMaxIndex(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
-                NDArray::prepareSpecialUse({&output}, {});
+                
-                for (auto v:inArrs)
+                NDArray::prepareSpecialUse({ &output }, inArrs);
-                    v->syncToDevice();
                 
                 BUILD_DOUBLE_SELECTOR(inArrs[0]->dataType(), output.dataType(), mergeMaxIndex_, (context, inArrs, output), LIBND4J_TYPES, INDEXING_TYPES);
 
-                NDArray::registerSpecialUse({&output}, {});
+                NDArray::registerSpecialUse({ &output }, inArrs);
             }
 
 
             //////////////////////////////////////////////////////////////////////////
             template <typename T>
-            static __global__ void global_mergeMax_(void **inArrs, void **inShapes, const int numArrays, void *voutput, Nd4jLong *outputShape, Nd4jLong length) {
+            static __global__ void mergeMaxCudaLauncher(void** inArrs, void** inShapes, const int numArrays, void* voutput, Nd4jLong* outputShape, Nd4jLong length) {
                 auto output = reinterpret_cast<T*>(voutput);
 
                 const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
@@ -103,51 +105,163 @@ namespace sd {
                     T mVal = -DataTypeUtils::max<T>();
 
                     for (int i = 0; i < numArrays; i++) {
-                        auto x = reinterpret_cast<T*>(inArrs[i]);
+                        auto x = reinterpret_cast<const T*>(inArrs[i]);
-                        auto xShape = reinterpret_cast<Nd4jLong *>(inShapes[i]);
+                        auto xShape = reinterpret_cast<const Nd4jLong*>(inShapes[i]);
                         auto val = x[shape::getIndexOffset(e, xShape)];;
                         if (mVal < val)
                             mVal = val;
                     }
-                    __syncthreads();
 
                     output[shape::getIndexOffset(e, outputShape)] = mVal;
                 }
             }
 
             template<typename T>
-            static void mergeMax_(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output) {
+            static void mergeMax_(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
-                std::vector<void *> inBuffers(inArrs.size());
-                std::vector<void *> inShapes(inArrs.size());
                 
-                for (int e = 0; e < inArrs.size(); e++) {
+                int nArrsSize = static_cast<int>(inArrs.size());
+
+                std::vector<void*> inBuffers(nArrsSize), inShapes(nArrsSize);
+
+                for (int e = 0; e < nArrsSize; e++) {
                     inBuffers[e] = inArrs[e]->getSpecialBuffer();
                     inShapes[e] = inArrs[e]->getSpecialShapeInfo();
                 }
 
                 PointersManager manager(context, "mergeMax");
 
-                auto pInBuffers = reinterpret_cast<void **>(manager.replicatePointer(inBuffers.data(), inBuffers.size() * sizeof(void *)));
+                auto pInBuffers = reinterpret_cast<void**>(manager.replicatePointer(inBuffers.data(), inBuffers.size() * sizeof(void*)));
-                auto pInShapes = reinterpret_cast<void **>(manager.replicatePointer(inShapes.data(), inShapes.size() * sizeof(void *)));
+                auto pInShapes = reinterpret_cast<void**>(manager.replicatePointer(inShapes.data(), inShapes.size() * sizeof(void*)));
                 auto length = output.lengthOf();
 
-                global_mergeMax_<T><<<512, 512, 512, *context->getCudaStream()>>>(pInBuffers, pInShapes, (int) inArrs.size(), output.getSpecialBuffer(), output.getSpecialShapeInfo(), length);
+                const int threadsPerBlock = MAX_NUM_THREADS / 2;
+                const int blocksPerGrid = (length + threadsPerBlock - 1) / threadsPerBlock;
+
+                mergeMaxCudaLauncher<T> << <blocksPerGrid, threadsPerBlock, 512, *context->getCudaStream() >> > (pInBuffers, pInShapes, nArrsSize, output.getSpecialBuffer(), output.getSpecialShapeInfo(), length);
 
                 manager.synchronize();
             }
 
-            void mergeMax(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output) {
+            void mergeMax(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
-                NDArray::prepareSpecialUse({&output}, {});
+                
-                for (auto v:inArrs)
+                NDArray::prepareSpecialUse({ &output }, inArrs);
-                    v->syncToDevice();
                 
                 BUILD_SINGLE_SELECTOR(output.dataType(), mergeMax_, (context, inArrs, output), LIBND4J_TYPES);
-                NDArray::registerSpecialUse({&output}, {});
+                
+                NDArray::registerSpecialUse({ &output }, inArrs);
             }
 
             //////////////////////////////////////////////////////////////////////////
             template <typename T>
-            static __global__ void global_mergeAvg_(void **inArrs, void **inShapes, const int numArrays, void *voutput, Nd4jLong *outputShape, Nd4jLong length) {
+            static __global__ void mergeMaxBpCudaLauncher(void** inArrs, void** inShapes, void* vgradient, Nd4jLong* gradientShape, const int numArrays,
+                void** outArrs, void** outShapes, Nd4jLong length, bool bSameOrderAndEws1) {
+
+                auto grad = reinterpret_cast<T*>(vgradient);
+
+                const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+                const auto step = gridDim.x * blockDim.x;
+
+                int coords[MAX_RANK];
+
+                for (Nd4jLong e = tid; e < length; e += step) {
+
+                    T mVal = -DataTypeUtils::max<T>();
+                    int nMaxIndex = 0;
+                    auto xOffset = e, zOffset = e, gradOffset = e;
+
+                    if (!bSameOrderAndEws1) {
+                        shape::index2coords(e, gradientShape, coords);
+                        gradOffset = shape::getOffset(gradientShape, coords);
+                    }
+
+                    for (int i = 0; i < numArrays; i++) {
+                        auto x = reinterpret_cast<T*>(inArrs[i]);
+
+                        if (!bSameOrderAndEws1) {
+                            auto xShape = reinterpret_cast<Nd4jLong*>(inShapes[i]);
+                            xOffset = shape::getOffset(xShape, coords);
+                        }
+
+                        auto val = x[xOffset];
+                        if (mVal < val) {
+                            mVal = val;
+                            nMaxIndex = i;
+                        }
+                    }
+                  
+                    // outputs have to be pre-nullify                 
+                    if (!bSameOrderAndEws1) {
+                        auto outShape = reinterpret_cast<Nd4jLong*>(outShapes[nMaxIndex]);
+                        zOffset = shape::getOffset(outShape, coords);
+                    }
+
+                    auto output = reinterpret_cast<T*>(outArrs[nMaxIndex]);
+
+                    output[zOffset] = grad[gradOffset];
+                }
+            }
+
+            template<typename T>
+            static void mergeMaxBp_(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, std::vector<NDArray*>& outArrs, int nArrSize, bool bSameOrderAndEws1) {
+
+                std::vector<void*> inBuffers(nArrSize), inShapes(nArrSize), outBuffers(nArrSize), outShapes(nArrSize);
+
+                for (int e = 0; e < nArrSize; e++) {
+                    inBuffers[e] = inArrs[e]->getSpecialBuffer();
+                    inShapes[e] = inArrs[e]->getSpecialShapeInfo();
+                    outBuffers[e] = outArrs[e]->getSpecialBuffer();
+                    outShapes[e] = outArrs[e]->getSpecialShapeInfo();
+                }
+
+                PointersManager manager(context, "mergeMaxBp");
+
+                auto pInBuffers = reinterpret_cast<void**>(manager.replicatePointer(inBuffers.data(), inBuffers.size() * sizeof(void*)));
+                auto pInShapes = reinterpret_cast<void**>(manager.replicatePointer(inShapes.data(), inShapes.size() * sizeof(void*)));
+
+                auto pOutBuffers = reinterpret_cast<void**>(manager.replicatePointer(outBuffers.data(), outBuffers.size() * sizeof(void*)));
+                auto pOutShapes = reinterpret_cast<void**>(manager.replicatePointer(outShapes.data(), outShapes.size() * sizeof(void*)));
+
+                auto length = inArrs[nArrSize]->lengthOf();
+
+                const int threadsPerBlock = MAX_NUM_THREADS / 2;
+                const int blocksPerGrid = (length + threadsPerBlock - 1) / threadsPerBlock;
+
+                mergeMaxBpCudaLauncher<T> << <blocksPerGrid, threadsPerBlock, 512, *context->getCudaStream() >> > (pInBuffers, pInShapes, inArrs[nArrSize]->getSpecialBuffer(),
+                    inArrs[nArrSize]->getSpecialShapeInfo(), nArrSize, pOutBuffers, pOutShapes,
+                    length, bSameOrderAndEws1);
+                
+                manager.synchronize();
+            }
+
+            void mergeMaxBp(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, std::vector<NDArray*>& outArrs) {
+
+                // not use gradient
+                int nArrSize = static_cast<int>(inArrs.size() - 1);
+                
+                const std::vector<const NDArray*>& out = reinterpret_cast<const std::vector<const NDArray*>&>(outArrs);
+
+                NDArray::prepareSpecialUse(out, inArrs);
+
+                bool bSameOrderAndEws1 = (1 == inArrs[nArrSize]->ews());
+                auto ordering = inArrs[nArrSize]->ordering();
+                  
+                for (int i = 0; i < nArrSize; ++i) {
+                    bSameOrderAndEws1 &= (ordering == inArrs[i]->ordering());
+                    bSameOrderAndEws1 &= (1 == inArrs[i]->ews());
+                   
+                    bSameOrderAndEws1 &= (ordering == outArrs[i]->ordering());
+                    bSameOrderAndEws1 &= (1 == outArrs[i]->ews());
+                }
+
+                BUILD_SINGLE_SELECTOR(inArrs[nArrSize]->dataType(), mergeMaxBp_, (context, inArrs, outArrs, nArrSize, bSameOrderAndEws1), LIBND4J_TYPES);
+
+                NDArray::registerSpecialUse( out, inArrs );
+            }
+
+
+            //////////////////////////////////////////////////////////////////////////
+            template <typename T>
+            static __global__ void mergeAvgCudaLauncher(void** inArrs, void** inShapes, const int numArrays, void* voutput, Nd4jLong* outputShape, Nd4jLong length) {
                 auto output = reinterpret_cast<T*>(voutput);
 
                 const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
@@ -158,7 +272,7 @@ namespace sd {
 
                     for (int i = 0; i < numArrays; i++) {
                         auto x = reinterpret_cast<T*>(inArrs[i]);
-                        auto xShape = reinterpret_cast<Nd4jLong *>(inShapes[i]);
+                        auto xShape = reinterpret_cast<Nd4jLong*>(inShapes[i]);
 
                         sum += x[shape::getIndexOffset(e, xShape)];
                     }
@@ -168,9 +282,9 @@ namespace sd {
             }
 
             template<typename T>
-            static void mergeAvg_(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output) {
+            static void mergeAvg_(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
-                std::vector<void *> inBuffers(inArrs.size());
+                
-                std::vector<void *> inShapes(inArrs.size());
+                std::vector<void*> inBuffers(inArrs.size()), inShapes(inArrs.size());
 
                 for (int e = 0; e < inArrs.size(); e++) {
                     inBuffers[e] = inArrs[e]->getSpecialBuffer();
@@ -179,28 +293,111 @@ namespace sd {
 
                 PointersManager manager(context, "mergeAvg");
 
-                auto pInBuffers = reinterpret_cast<void **>(manager.replicatePointer(inBuffers.data(), inBuffers.size() * sizeof(void *)));
+                auto pInBuffers = reinterpret_cast<void**>(manager.replicatePointer(inBuffers.data(), inBuffers.size() * sizeof(void*)));
-                auto pInShapes = reinterpret_cast<void **>(manager.replicatePointer(inShapes.data(), inShapes.size() * sizeof(void *)));
+                auto pInShapes = reinterpret_cast<void**>(manager.replicatePointer(inShapes.data(), inShapes.size() * sizeof(void*)));
                 auto length = output.lengthOf();
 
-                global_mergeAvg_<T><<<512, 512, 512, *context->getCudaStream()>>>(pInBuffers, pInShapes, (int) inArrs.size(), output.getSpecialBuffer(), output.getSpecialShapeInfo(), length);
+                const int threadsPerBlock = MAX_NUM_THREADS / 2;
+                const int blocksPerGrid = (length + threadsPerBlock - 1) / threadsPerBlock;
+
+                mergeAvgCudaLauncher<T> << <blocksPerGrid, threadsPerBlock, 512, *context->getCudaStream() >> > (pInBuffers, pInShapes, (int)inArrs.size(), output.getSpecialBuffer(), output.getSpecialShapeInfo(), length);
 
                 manager.synchronize();
             }
 
-            void mergeAvg(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output) {
+            void mergeAvg(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
-                NDArray::prepareSpecialUse({&output}, {});
+                
-                for (auto v:inArrs)
+                NDArray::prepareSpecialUse({ &output }, inArrs);
-                    v->syncToDevice();
 
                 BUILD_SINGLE_SELECTOR(output.dataType(), mergeAvg_, (context, inArrs, output), FLOAT_TYPES);
 
-                NDArray::registerSpecialUse({&output}, {});
+                NDArray::registerSpecialUse({ &output }, inArrs);
+            }
+            //////////////////////////////////////////////////////////////////////////
+            template <typename T>
+            static __global__ void mergeAvgBpCudaLauncher(void* vgradient, Nd4jLong* gradientShape, void** outArrs, void** outShapes,
+                                                      const int numArrays, Nd4jLong length, bool bSameOrderAndEws1) {
+
+                auto grad = reinterpret_cast<T*>(vgradient);
+
+                const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+                const auto step = gridDim.x * blockDim.x;
+
+                int coords[MAX_RANK];
+
+                for (Nd4jLong e = tid; e < length; e += step) {
+
+                    auto zOffset = e, gradOffset = e;
+                    if (!bSameOrderAndEws1) {
+                        shape::index2coords(e, gradientShape, coords);
+                        gradOffset = shape::getOffset(gradientShape, coords);
+                    }
+
+                    for (int i = 0; i < numArrays; i++) {
+
+                        if (!bSameOrderAndEws1) {
+                            auto outShape = reinterpret_cast<Nd4jLong*>(outShapes[i]);
+                            zOffset = shape::getOffset(outShape, coords);
+                        }
+
+                        auto output = reinterpret_cast<T*>(outArrs[i]);
+
+                        output[zOffset] = grad[gradOffset] / numArrays;
+                    }
+                }
+            }
+
+            template<typename T>
+            static void mergeAvgBp_(sd::LaunchContext* context, const NDArray& gradient, std::vector<NDArray*>& outArrs, bool bSameOrderAndEws1) {
+
+                int nArrSize = static_cast<int>(outArrs.size());
+
+                std::vector<void*> outBuffers(nArrSize), outShapes(nArrSize);
+
+                for (int e = 0; e < nArrSize; e++) {
+                    outBuffers[e] = outArrs[e]->getSpecialBuffer();
+                    outShapes[e] = outArrs[e]->getSpecialShapeInfo();
+                }
+
+                PointersManager manager(context, "mergeAvgBp");
+
+                auto pOutBuffers = reinterpret_cast<void**>(manager.replicatePointer(outBuffers.data(), outBuffers.size() * sizeof(void*)));
+                auto pOutShapes = reinterpret_cast<void**>(manager.replicatePointer(outShapes.data(), outShapes.size() * sizeof(void*)));
+
+                auto length = gradient.lengthOf();
+                
+                const int threadsPerBlock = MAX_NUM_THREADS / 2;
+                const int blocksPerGrid = (length + threadsPerBlock - 1) / threadsPerBlock;
+
+                mergeAvgBpCudaLauncher<T> << <blocksPerGrid, threadsPerBlock, 512, *context->getCudaStream() >> > (gradient.getSpecialBuffer(), gradient.getSpecialShapeInfo(),
+                    pOutBuffers, pOutShapes, nArrSize, length, bSameOrderAndEws1);
+
+                manager.synchronize();
+            }
+
+            void mergeAvgBp(sd::LaunchContext* context, const NDArray& gradient, std::vector<NDArray*>& outArrs) {
+
+                const std::vector<const NDArray*>& out = reinterpret_cast<const std::vector<const NDArray*>&>(outArrs);
+
+                NDArray::prepareSpecialUse( out, { &gradient });
+
+                bool bSameOrderAndEws1 = (1 == gradient.ews());
+                auto ordering = gradient.ordering();
+
+                for (const auto& v : outArrs) {
+                    bSameOrderAndEws1 &= (ordering == v->ordering());
+                    bSameOrderAndEws1 &= (1 == v->ews());
+                }
+
+                BUILD_SINGLE_SELECTOR(gradient.dataType(), mergeAvgBp_, (context, gradient, outArrs, bSameOrderAndEws1), LIBND4J_TYPES);
+
+                NDArray::prepareSpecialUse(out, { &gradient });
             }
 
             //////////////////////////////////////////////////////////////////////////
             template <typename T>
-            static __global__ void global_mergeAdd_(void **inArrs, void **inShapes, const int numArrays, void *voutput, Nd4jLong *outputShape, Nd4jLong length) {
+            static __global__ void mergeAddCudaLauncher(void** inArrs, void** inShapes, const int numArrays, void* voutput, Nd4jLong* outputShape, Nd4jLong length) {
+                
                 auto output = reinterpret_cast<T*>(voutput);
 
                 const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
@@ -211,7 +408,7 @@ namespace sd {
 
                     for (int i = 0; i < numArrays; i++) {
                         auto x = reinterpret_cast<T*>(inArrs[i]);
-                        auto xShape = reinterpret_cast<Nd4jLong *>(inShapes[i]);
+                        auto xShape = reinterpret_cast<Nd4jLong*>(inShapes[i]);
 
                         sum += x[shape::getIndexOffset(e, xShape)];
                     }
@@ -221,36 +418,120 @@ namespace sd {
             }
 
             template<typename T>
-            static void mergeAdd_(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output) {
+            static void mergeAdd_(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
-                std::vector<void *> inBuffers(inArrs.size());
-                std::vector<void *> inShapes(inArrs.size());
                 
-                for (int e = 0; e < inArrs.size(); e++) {
+                int nArrSize = static_cast<int>(inArrs.size());
+                std::vector<void*> inBuffers(nArrSize), inShapes(nArrSize);
+
+                for (int e = 0; e < nArrSize; e++) {
                     inBuffers[e] = inArrs[e]->getSpecialBuffer();
                     inShapes[e] = inArrs[e]->getSpecialShapeInfo();
                 }
 
                 PointersManager manager(context, "mergeAdd");
 
-                auto pInBuffers = reinterpret_cast<void **>(manager.replicatePointer(inBuffers.data(), inBuffers.size() * sizeof(void *)));
+                auto pInBuffers = reinterpret_cast<void**>(manager.replicatePointer(inBuffers.data(), inBuffers.size() * sizeof(void*)));
-                auto pInShapes = reinterpret_cast<void **>(manager.replicatePointer(inShapes.data(), inShapes.size() * sizeof(void *)));
+                auto pInShapes = reinterpret_cast<void**>(manager.replicatePointer(inShapes.data(), inShapes.size() * sizeof(void*)));
                 auto length = output.lengthOf();
 
-                global_mergeAdd_<T><<<512, 512, 512, *context->getCudaStream()>>>(pInBuffers, pInShapes, (int) inArrs.size(), output.getSpecialBuffer(), output.getSpecialShapeInfo(), length);
+                const int threadsPerBlock = MAX_NUM_THREADS / 2;
+                const int blocksPerGrid = (length + threadsPerBlock - 1) / threadsPerBlock;
+
+                mergeAddCudaLauncher<T> << <blocksPerGrid, threadsPerBlock, 512, *context->getCudaStream() >> > (pInBuffers, pInShapes, nArrSize, output.getSpecialBuffer(), output.getSpecialShapeInfo(), length);
 
                 manager.synchronize();
             }
-            BUILD_SINGLE_TEMPLATE(template void mergeAdd_, (sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output), NUMERIC_TYPES);
+            BUILD_SINGLE_TEMPLATE(template void mergeAdd_, (sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, NDArray& output), NUMERIC_TYPES);
 
-            void mergeAdd(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output) {
+            void mergeAdd(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, NDArray& output) {
-                NDArray::prepareSpecialUse({&output}, {});
+                
-                for (auto v:inArrs)
+                NDArray::prepareSpecialUse({ &output }, inArrs);
-                    v->syncToDevice();
                 
                 BUILD_SINGLE_SELECTOR(output.dataType(), mergeAdd_, (context, inArrs, output), NUMERIC_TYPES);
 
-                NDArray::registerSpecialUse({&output}, {});
+                NDArray::registerSpecialUse({ &output }, inArrs);
             }
+
+            //////////////////////////////////////////////////////////////////////////
+            template <typename T>
+            static __global__ void mergeAddBpCudaLauncher(void* vgradient, Nd4jLong* gradientShape, void** outArrs, void** outShapes,
+                const int numArrays, Nd4jLong length, bool bSameOrderAndEws1) {
+
+                auto grad = reinterpret_cast<T*>(vgradient);
+
+                const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+                const auto step = gridDim.x * blockDim.x;
+
+                int coords[MAX_RANK];
+
+                for (Nd4jLong e = tid; e < length; e += step) {
+
+                    auto zOffset = e, gradOffset = e;
+                    if (!bSameOrderAndEws1) {
+                        shape::index2coords(e, gradientShape, coords);
+                        gradOffset = shape::getOffset(gradientShape, coords);
+                    }
+
+                    for (int i = 0; i < numArrays; i++) {
+                        
+                        if (!bSameOrderAndEws1) {
+                            auto outShape = reinterpret_cast<Nd4jLong*>(outShapes[i]);
+                            zOffset = shape::getOffset(outShape, coords);
+                        }
+
+                        auto output = reinterpret_cast<T*>(outArrs[i]);
+
+                        output[zOffset] = grad[gradOffset];
+                    }
+                }
+            }
+
+            template<typename T>
+            static void mergeAddBp_(sd::LaunchContext* context, const NDArray& gradient, std::vector<NDArray*>& outArrs, bool bSameOrderAndEws1) {
+
+                int nArrSize = static_cast<int>(outArrs.size());
+
+                std::vector<void*> outBuffers(nArrSize), outShapes(nArrSize);
+
+                for (int e = 0; e < nArrSize; e++) {
+                    outBuffers[e] = outArrs[e]->getSpecialBuffer();
+                    outShapes[e] = outArrs[e]->getSpecialShapeInfo();
+                }
+
+                PointersManager manager(context, "mergeAddBp");
+
+                auto pOutBuffers = reinterpret_cast<void**>(manager.replicatePointer(outBuffers.data(), outBuffers.size() * sizeof(void*)));
+                auto pOutShapes = reinterpret_cast<void**>(manager.replicatePointer(outShapes.data(), outShapes.size() * sizeof(void*)));
+
+                auto length = gradient.lengthOf();
+
+                const int threadsPerBlock = MAX_NUM_THREADS / 2;
+                const int blocksPerGrid = (length + threadsPerBlock - 1) / threadsPerBlock;
+
+                mergeAddBpCudaLauncher<T> << <blocksPerGrid, threadsPerBlock, 512, *context->getCudaStream() >> > (gradient.getSpecialBuffer(), gradient.getSpecialShapeInfo(),
+                    pOutBuffers, pOutShapes, nArrSize, length, bSameOrderAndEws1);
+
+                manager.synchronize();
+            }
+
+            void mergeAddBp(sd::LaunchContext* context, const NDArray& gradient, std::vector<NDArray*>& outArrs) {
+
+                const std::vector<const NDArray*>& out = reinterpret_cast<const std::vector<const NDArray*>& >(outArrs);
+                NDArray::prepareSpecialUse( out, { &gradient });
+
+                bool bSameOrderAndEws1 = (1 == gradient.ews());
+                auto ordering = gradient.ordering();
+
+                for (const auto& v : outArrs) {
+                    bSameOrderAndEws1 &= (ordering == v->ordering());
+                    bSameOrderAndEws1 &= (1 == v->ews());
+                }
+
+                BUILD_SINGLE_SELECTOR(gradient.dataType(), mergeAddBp_, (context, gradient, outArrs, bSameOrderAndEws1), LIBND4J_TYPES);
+
+                NDArray::prepareSpecialUse( out, { &gradient });
+            }
+
         }
     }
 }

libnd4j/include/ops/declarable/helpers/transforms.h

@@ -52,13 +52,16 @@ namespace helpers {
 
 	void scatterSimple(sd::LaunchContext * context, const int opId, NDArray& input, const NDArray& updates, const NDArray& indices, const std::vector<int>& dimensions);
 
-	void mergeMaxIndex(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output);
+	void mergeMaxIndex(sd::LaunchContext * context, const std::vector<const NDArray*>& inArrs, NDArray& output);
 
-	void mergeMax(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output);
+	void mergeMax(sd::LaunchContext * context, const std::vector<const NDArray*>& inArrs, NDArray& output);
+	void mergeMaxBp(sd::LaunchContext* context, const std::vector<const NDArray*>& inArrs, std::vector<NDArray*>& outArrs);
 
-	void mergeAvg(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output);
+	void mergeAvg(sd::LaunchContext * context, const std::vector<const NDArray*>& inArrs, NDArray& output);
+	void mergeAvgBp(sd::LaunchContext* context, const NDArray& gradient, std::vector<NDArray*>& outArrs);
 
-	void mergeAdd(sd::LaunchContext * context, const std::vector<NDArray*>& inArrs, NDArray& output);
+	void mergeAdd(sd::LaunchContext * context, const std::vector<const NDArray*>& inArrs, NDArray& output);
+	void mergeAddBp(sd::LaunchContext* context, const NDArray& gradient, std::vector<NDArray*>& outArrs);
 
 	void clipByNorm(sd::LaunchContext * context, NDArray& input, NDArray& output, const std::vector<int>& dimensions, const NDArray& clipNorm, const bool isInplace);
 	void clipByGlobalNorm(sd::LaunchContext * context, std::vector<NDArray*> const& inputs, double clipNorm, sd::memory::Workspace* workspace, std::vector<NDArray*>& outputs, bool isInplace);

libnd4j/tests_cpu/layers_tests/DeclarableOpsTests13.cpp

@@ -955,7 +955,160 @@ TEST_F(DeclarableOpsTests13, mergemax_2) {
 
     ASSERT_EQ(20, status);
 }
+/////////////////////////////////////////////////////////////////////////
+TEST_F(DeclarableOpsTests13, mergemax_bp_1) {
 
+    NDArray x1('c', { 5, 5 }, sd::DataType::FLOAT32);
+    NDArray x2('c', { 5, 5 }, sd::DataType::FLOAT32);
+    NDArray x3('c', { 5, 5 }, sd::DataType::FLOAT32);
+    NDArray grad('c', { 5, 5 }, sd::DataType::FLOAT32);
+
+    x1.assign(3);
+    x2.assign(1);
+    x3.assign(2);
+    grad.linspace(.1, .1);
+
+
+    sd::ops::mergemax_bp op;
+    auto result = op.evaluate({ &x1, &x2, &x3, &grad }, {}, {});
+    ASSERT_EQ(Status::OK(), result.status());
+    ASSERT_EQ(3, result.size());
+
+    auto z = result.at(0);
+
+    ASSERT_TRUE(grad.isSameShape(z));
+    ASSERT_TRUE(grad.equalsTo(z));
+
+}
+/////////////////////////////////////////////////////////////////////////
+TEST_F(DeclarableOpsTests13, mergemax_bp_2) {
+
+    NDArray x1('c', { 2, 5 }, { 1,2,3,4,5,4,3,2,1,0 }, sd::DataType::FLOAT32);
+    NDArray x2('c', { 2, 5 }, { 0,1,2,3,4,5,6,7,8,9 }, sd::DataType::FLOAT32);
+    NDArray x3('c', { 2, 5 }, { 0,1,1,2,3,4,7,5,8,10 }, sd::DataType::FLOAT32);
+    NDArray grad('c', { 2, 5 }, sd::DataType::FLOAT32);
+
+    grad.linspace(.1, .1);
+
+    NDArray exp1('c', { 2, 5 }, { 0.1, 0.2, 0.3, 0.4, 0.5, 0.0, 0.0, 0.0, 0.0, 0.0 }, sd::DataType::FLOAT32);
+    NDArray exp2('c', { 2, 5 }, { 0.0, 0.0, 0.0, 0.0, 0.0, 0.6, 0.0, 0.8, 0.9, 0.0 }, sd::DataType::FLOAT32);
+    NDArray exp3('c', { 2, 5 }, { 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.7, 0.0, 0.0, 1.0 }, sd::DataType::FLOAT32);
+
+    sd::ops::mergemax_bp op;
+    auto result = op.evaluate({ &x1, &x2, &x3, &grad }, {}, {});
+    ASSERT_EQ(Status::OK(), result.status());
+    ASSERT_EQ(3, result.size());
+
+    auto z1 = result.at(0);
+    auto z2 = result.at(1);
+    auto z3 = result.at(2);
+
+    ASSERT_TRUE(exp1.isSameShape(z1));
+    ASSERT_TRUE(exp1.equalsTo(z1));
+    ASSERT_TRUE(exp2.isSameShape(z2));
+    ASSERT_TRUE(exp2.equalsTo(z2));
+    ASSERT_TRUE(exp3.isSameShape(z3));
+    ASSERT_TRUE(exp3.equalsTo(z3));
+
+}
+/////////////////////////////////////////////////////////////////////////
+TEST_F(DeclarableOpsTests13, mergemax_bp_3) {
+
+    NDArray x1C('c', { 2, 5 }, { 1,2,3,4,5,4,3,2,1,0 }, sd::DataType::FLOAT32);
+    NDArray x2C('c', { 2, 5 }, { 0,1,2,3,4,5,6,7,8,9 }, sd::DataType::FLOAT32);
+    NDArray x3C('c', { 2, 5 }, { 0,1,1,2,3,4,7,5,8,10 }, sd::DataType::FLOAT32);
+    NDArray grad('c', { 2, 5 }, sd::DataType::FLOAT32);
+
+    grad.linspace(.1, .1);
+
+    NDArray x1('f', { 2, 5 }, sd::DataType::FLOAT32);
+    NDArray x2('f', { 2, 5 }, sd::DataType::FLOAT32);
+    NDArray x3('f', { 2, 5 }, sd::DataType::FLOAT32);
+
+    NDArray exp1C('c', { 2, 5 }, { 0.1, 0.2, 0.3, 0.4, 0.5, 0.0, 0.0, 0.0, 0.0, 0.0 }, sd::DataType::FLOAT32);
+    NDArray exp2C('c', { 2, 5 }, { 0.0, 0.0, 0.0, 0.0, 0.0, 0.6, 0.0, 0.8, 0.9, 0.0 }, sd::DataType::FLOAT32);
+    NDArray exp3C('c', { 2, 5 }, { 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.7, 0.0, 0.0, 1.0 }, sd::DataType::FLOAT32);
+
+    NDArray exp1('f', { 2, 5 }, sd::DataType::FLOAT32);
+    NDArray exp2('f', { 2, 5 }, sd::DataType::FLOAT32);
+    NDArray exp3('f', { 2, 5 }, sd::DataType::FLOAT32);
+
+    x1.assign(x1C);
+    x2.assign(x2C);
+    x3.assign(x3C);
+
+    exp1.assign(exp1C);
+    exp2.assign(exp2C);
+    exp3.assign(exp3C);
+
+    sd::ops::mergemax_bp op;
+    auto result = op.evaluate({ &x1, &x2, &x3, &grad }, {}, {});
+    ASSERT_EQ(Status::OK(), result.status());
+    ASSERT_EQ(3, result.size());
+
+    auto z1 = result.at(0);
+    auto z2 = result.at(1);
+    auto z3 = result.at(2);
+
+    ASSERT_TRUE(exp1.isSameShape(z1));
+    ASSERT_TRUE(exp1.equalsTo(z1));
+    ASSERT_TRUE(exp2.isSameShape(z2));
+    ASSERT_TRUE(exp2.equalsTo(z2));
+    ASSERT_TRUE(exp3.isSameShape(z3));
+    ASSERT_TRUE(exp3.equalsTo(z3));
+
+}
+/////////////////////////////////////////////////////////////////////////
+TEST_F(DeclarableOpsTests13, mergeadd_bp_1) {
+
+    NDArray x1('c', { 5, 5 }, sd::DataType::FLOAT32);
+    NDArray x2('c', { 5, 5 }, sd::DataType::FLOAT32);
+    NDArray x3('c', { 5, 5 }, sd::DataType::FLOAT32);
+    NDArray grad('c', { 5, 5 }, sd::DataType::FLOAT32);
+
+    x1.assign(3);
+    x2.assign(1);
+    x3.assign(2);
+    grad.linspace(.1, .1);
+
+    sd::ops::mergeadd_bp op;
+    auto result = op.evaluate({ &x1, &x2, &x3, &grad }, {}, {});
+    ASSERT_EQ(Status::OK(), result.status());
+    ASSERT_EQ(3, result.size());
+
+    for (int i = 0; i < 3; i++) {
+        auto z = result.at(0);
+        ASSERT_TRUE(grad.isSameShape(z));
+        ASSERT_TRUE(grad.equalsTo(z));
+    }
+}
+/////////////////////////////////////////////////////////////////////////
+TEST_F(DeclarableOpsTests13, mergeavg_bp_1) {
+
+    NDArray x1('c', { 5, 5 }, sd::DataType::FLOAT32);
+    NDArray x2('c', { 5, 5 }, sd::DataType::FLOAT32);
+    NDArray x3('c', { 5, 5 }, sd::DataType::FLOAT32);
+    NDArray grad('c', { 5, 5 }, sd::DataType::FLOAT32);
+
+    x1.assign(3);
+    x2.assign(1);
+    x3.assign(2);
+    grad.linspace(.1, .1);
+
+    sd::ops::mergeavg_bp op;
+    auto result = op.evaluate({ &x1, &x2, &x3, &grad }, {}, {});
+    ASSERT_EQ(Status::OK(), result.status());
+    ASSERT_EQ(3, result.size());
+
+    grad.applyScalar(sd::scalar::Divide, 3, grad);
+
+    for (int i = 0; i < 3; i++) {
+        auto z = result.at(i);
+        ASSERT_TRUE(grad.isSameShape(z));
+        ASSERT_TRUE(grad.equalsTo(z));
+    }
+
+}
 ///////////////////////////////////////////////////////////////////
 TEST_F(DeclarableOpsTests13, lstmLayer_1) {
 

nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/TFGraphRunnerService.java

@@ -0,0 +1,37 @@
+/*******************************************************************************
+ * Copyright (c) 2020 Konduit K.K.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ ******************************************************************************/
+
+
+package org.nd4j;
+
+
+import org.nd4j.linalg.api.buffer.DataType;
+import org.nd4j.linalg.api.ndarray.INDArray;
+
+import java.util.List;
+import java.util.Map;
+
+public interface TFGraphRunnerService{
+    TFGraphRunnerService init(
+            List<String> inputNames,
+            List<String> outputNames,
+            byte[] graphBytes,
+            Map<String, INDArray> constants,
+            Map<String, String> inputDataTypes
+    );
+
+    Map<String,INDArray> run(Map<String,INDArray> inputs);
+}

nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SDVariable.java

@@ -1654,29 +1654,6 @@ public class SDVariable implements Serializable {
         return x;
     }
 
-    @Override
-    public boolean equals(Object o) {
-        if (this == o) {
-            return true;
-        }
-        if (!(o instanceof SDVariable)) {
-            return false;
-        }
-
-        SDVariable that = (SDVariable) o;
-
-        if (!Objects.equals(varName, that.varName)) {
-            return false;
-        }
-        if (variableType != that.variableType) {
-            return false;
-        }
-        if(sameDiff != that.sameDiff){
-            return false;
-        }
-        return dataType == that.dataType;
-    }
-
     @Override
     public int hashCode() {
         int result = super.hashCode();
@@ -1695,4 +1672,26 @@ public class SDVariable implements Serializable {
         v.sameDiff = sd;
         return v;
     }
+
+    @Override
+    public boolean equals(Object o){
+        if(o == this) return true;
+        if(!(o instanceof SDVariable))
+            return false;
+
+        SDVariable s = (SDVariable)o;
+        if(!varName.equals(s.varName))
+            return false;
+        if(variableType != s.variableType)
+            return false;
+        if(dataType != s.dataType)
+            return false;
+
+        if(variableType == VariableType.VARIABLE || variableType == VariableType.CONSTANT){
+            INDArray a1 = getArr();
+            INDArray a2 = s.getArr();
+            return a1.equals(a2);
+        }
+        return true;
+    }
 }

nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SameDiff.java

@@ -1234,13 +1234,14 @@ public class SameDiff extends SDBaseOps {
     @Override
     public boolean equals(Object o) {
         if (this == o) return true;
-        if (o == null || getClass() != o.getClass()) return false;
+        if (o == null || getClass() != o.getClass())
+            return false;
 
         SameDiff sameDiff = (SameDiff) o;
 
-        if (variables != null ? !variables.equals(sameDiff.variables) : sameDiff.variables != null)
+        boolean eqVars = variables.equals(sameDiff.variables);
-            return false;
+        boolean eqOps = ops.equals(sameDiff.ops);
-        return sameDiffFunctionInstances != null ? sameDiffFunctionInstances.equals(sameDiff.sameDiffFunctionInstances) : sameDiff.sameDiffFunctionInstances == null;
+        return eqVars && eqOps;
     }
 
     /**
@@ -5843,4 +5844,10 @@ public class SameDiff extends SDBaseOps {
 
         return base + "_" + inc;
     }
+
+
+    @Override
+    public String toString(){
+        return "SameDiff(nVars=" + variables.size() + ",nOps=" + ops.size() + ")";
+    }
 }

nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/internal/Variable.java

@@ -16,10 +16,7 @@
 
 package org.nd4j.autodiff.samediff.internal;
 
-import lombok.AllArgsConstructor;
+import lombok.*;
-import lombok.Builder;
-import lombok.Data;
-import lombok.NoArgsConstructor;
 import org.nd4j.autodiff.samediff.SDVariable;
 
 import java.util.List;
@@ -28,6 +25,7 @@ import java.util.List;
 @NoArgsConstructor
 @Data   //TODO immutable?
 @Builder
+@EqualsAndHashCode(exclude = {"gradient", "variableIndex"})
 public class Variable {
     protected String name;
     protected SDVariable variable;

nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/classification/EvaluationBinary.java

@@ -173,9 +173,6 @@ public class EvaluationBinary extends BaseEvaluation<EvaluationBinary> {
 
     @Override
     public void eval(INDArray labels, INDArray networkPredictions, INDArray maskArray, List<? extends Serializable> recordMetaData) {
-        if(recordMetaData != null){
-            throw new UnsupportedOperationException("Evaluation with record metadata not yet implemented for EvaluationBinary");
-        }
         eval(labels, networkPredictions, maskArray);
     }
 

nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/classification/EvaluationCalibration.java

@@ -325,7 +325,7 @@ public class EvaluationCalibration extends BaseEvaluation<EvaluationCalibration>
 
     @Override
     public void eval(INDArray labels, INDArray networkPredictions, INDArray maskArray, List<? extends Serializable> recordMetaData) {
-        throw new UnsupportedOperationException("Not yet implemented");
+        eval(labels, networkPredictions, maskArray);
     }
 
     @Override

nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/evaluation/regression/RegressionEvaluation.java

@@ -229,7 +229,7 @@ public class RegressionEvaluation extends BaseEvaluation<RegressionEvaluation> {
 
     @Override
     public void eval(INDArray labels, INDArray networkPredictions, INDArray maskArray, List<? extends Serializable> recordMetaData) {
-        throw new UnsupportedOperationException("Not yet implemented");
+        eval(labels, networkPredictions, maskArray);
     }
 
     @Override

nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/RandomOpValidation.java

@@ -76,7 +76,7 @@ public class RandomOpValidation extends BaseOpValidation {
                             double min = in.minNumber().doubleValue();
                             double max = in.maxNumber().doubleValue();
                             double mean = in.meanNumber().doubleValue();
-                            if (min >= 1 && max <= 2 && (in.length() == 1 || Math.abs(mean - 1.5) < 0.1))
+                            if (min >= 1 && max <= 2 && (in.length() == 1 || Math.abs(mean - 1.5) < 0.2))
                                 return null;
                             return "Failed: min = " + min + ", max = " + max + ", mean = " + mean;
                         };
@@ -87,7 +87,7 @@ public class RandomOpValidation extends BaseOpValidation {
                         checkFn = in -> {
                             double mean = in.meanNumber().doubleValue();
                             double stdev = in.std(true).getDouble(0);
-                            if (in.length() == 1 || (Math.abs(mean - 1) < 0.1 && Math.abs(stdev - 1) < 0.1))
+                            if (in.length() == 1 || (Math.abs(mean - 1) < 0.2 && Math.abs(stdev - 1) < 0.2))
                                 return null;
                             return "Failed: mean = " + mean + ", stdev = " + stdev;
                         };

nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/samediff/SameDiffTests.java

@@ -3556,4 +3556,52 @@ public class SameDiffTests extends BaseNd4jTest {
             assertTrue(msg, msg.contains("\"labels\"") && msg.contains("No array was provided"));
         }
     }
+
+
+    @Test
+    public void testEquals1(){
+
+        SameDiff sd1 = SameDiff.create();
+        SameDiff sd2 = SameDiff.create();
+
+        assertEquals(sd1, sd2);
+
+        SDVariable p1 = sd1.placeHolder("ph", DataType.FLOAT, -1, 10);
+        SDVariable p2 = sd2.placeHolder("ph", DataType.FLOAT, -1, 10);
+
+        assertEquals(sd1, sd2);
+
+        SDVariable w1 = sd1.constant("c1",1.0f);
+        SDVariable w2 = sd2.constant("c1",1.0f);
+
+        assertEquals(sd1, sd2);
+
+        SDVariable a1 = p1.add("add", w1);
+        SDVariable a2 = p2.add("add", w2);
+
+        assertEquals(sd1, sd2);
+
+        SDVariable w1a = sd1.constant("c2", 2.0f);
+        SDVariable w2a = sd2.constant("cX", 2.0f);
+
+        assertNotEquals(sd1, sd2);
+        w2a.rename("c2");
+
+        assertEquals(sd1, sd2);
+
+        sd2.createGradFunction("ph");
+
+        assertEquals(sd1, sd2);
+
+        w2a.getArr().assign(3.0f);
+
+        assertNotEquals(sd1, sd2);
+
+        w1a.getArr().assign(3.0f);
+        assertEquals(sd1, sd2);
+
+        SDVariable s1 = p1.sub("op", w1);
+        SDVariable s2 = p2.add("op", w1);
+        assertNotEquals(sd1, sd2);
+    }
 }

nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/linalg/nativ/OpsMappingTests.java

@@ -61,7 +61,7 @@ public class OpsMappingTests extends BaseNd4jTest {
 
     @Override
     public long getTimeoutMilliseconds() {
-        return 90000L;
+        return 180000L;     //Can be slow on some CI machines such as PPC
     }
 
     @Test

nd4j/nd4j-common/src/main/java/org/nd4j/resources/Downloader.java

@@ -95,7 +95,7 @@ public class Downloader {
             }
             // try extracting
             try{
-                ArchiveUtils.unzipFileTo(f.getAbsolutePath(), extractToDir.getAbsolutePath());
+                ArchiveUtils.unzipFileTo(f.getAbsolutePath(), extractToDir.getAbsolutePath(), false);
             } catch (Throwable t){
                 log.warn("Error extracting {} files from file {} - retrying...", name, f.getAbsolutePath(), t);
                 f.delete();

nd4j/nd4j-common/src/main/java/org/nd4j/util/ArchiveUtils.java

@@ -51,6 +51,10 @@ public class ArchiveUtils {
      * @throws IOException
      */
     public static void unzipFileTo(String file, String dest) throws IOException {
+        unzipFileTo(file, dest, true);
+    }
+
+    public static void unzipFileTo(String file, String dest, boolean logFiles) throws IOException {
         File target = new File(file);
         if (!target.exists())
             throw new IllegalArgumentException("Archive doesnt exist");
@@ -93,7 +97,9 @@ public class ArchiveUtils {
 
                     fos.close();
                     ze = zis.getNextEntry();
-                    log.debug("File extracted: " + newFile.getAbsoluteFile());
+                    if(logFiles) {
+                        log.info("File extracted: " + newFile.getAbsoluteFile());
+                    }
                 }
 
                 zis.closeEntry();
@@ -112,7 +118,9 @@ public class ArchiveUtils {
             TarArchiveEntry entry;
             /* Read the tar entries using the getNextEntry method **/
             while ((entry = (TarArchiveEntry) tarIn.getNextEntry()) != null) {
+                if(logFiles) {
                     log.info("Extracting: " + entry.getName());
+                }
                 /* If the entry is a directory, create the directory. */
 
                 if (entry.isDirectory()) {

nd4j/nd4j-tensorflow/src/main/java/org/nd4j/tensorflow/conversion/graphrunner/GraphRunner.java

@@ -16,18 +16,16 @@
 
 package org.nd4j.tensorflow.conversion.graphrunner;
 
-import lombok.Builder;
+import lombok.*;
-import lombok.Singular;
 import org.apache.commons.io.FileUtils;
 import org.nd4j.base.Preconditions;
+import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.factory.Nd4j;
 import org.nd4j.linalg.io.ClassPathResource;
 import org.nd4j.linalg.primitives.Pair;
 import org.nd4j.shade.protobuf.ByteString;
 import org.nd4j.shade.protobuf.InvalidProtocolBufferException;
 import org.nd4j.shade.protobuf.util.JsonFormat;
-import lombok.Getter;
-import lombok.Setter;
 import lombok.extern.slf4j.Slf4j;
 import org.nd4j.tensorflow.conversion.TensorDataType;
 import org.apache.commons.io.IOUtils;
@@ -56,6 +54,7 @@ import static org.bytedeco.tensorflow.global.tensorflow.*;
  * @author Adam Gibson
  */
 @Slf4j
+@NoArgsConstructor
 public class GraphRunner implements Closeable {
 
     private static boolean isTfWarmedUp = false;
@@ -103,6 +102,9 @@ public class GraphRunner implements Closeable {
      * @param inputDataTypes the expected input data types
      * @param outputDataTypes the expected output data types
      */
+
+
+
     @Builder
     public GraphRunner(List<String> inputNames,
                        List<String> outputNames,
@@ -440,6 +442,7 @@ public class GraphRunner implements Closeable {
      * @return a map of the output names to the
      * ndarrays matching each output specified in the graph
      */
+
     public Map<String,INDArray> run(Map<String,INDArray> inputs) {
         if (!isTfWarmedUp && !isTfWarmingUp){
             isTfWarmingUp = true;
@@ -683,4 +686,7 @@ public class GraphRunner implements Closeable {
 
         return builder1.build();
     }
+
+
+
 }

nd4j/nd4j-tensorflow/src/main/java/org/nd4j/tensorflow/conversion/graphrunner/GraphRunnerServiceProvider.java

@@ -0,0 +1,52 @@
+package org.nd4j.tensorflow.conversion.graphrunner;
+
+import org.nd4j.TFGraphRunnerService;
+import org.nd4j.linalg.api.buffer.DataType;
+import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.tensorflow.conversion.TensorDataType;
+
+import java.util.HashMap;
+import java.util.List;
+import java.util.Map;
+
+public class GraphRunnerServiceProvider implements TFGraphRunnerService {
+
+    private GraphRunner graphRunner;
+    Map<String, INDArray> inputs;
+
+    @Override
+    public TFGraphRunnerService init(
+            List<String> inputNames,
+            List<String> outputNames,
+            byte[] graphBytes,
+            Map<String, INDArray> constants,
+            Map<String, String> inputDataTypes){
+        if (inputNames.size() != inputDataTypes.size()){
+            throw new IllegalArgumentException("inputNames.size() != inputDataTypes.size()");
+        }
+        Map<String, TensorDataType> convertedDataTypes = new HashMap<>();
+        for (int i = 0; i < inputNames.size(); i++){
+            convertedDataTypes.put(inputNames.get(i), TensorDataType.fromProtoValue(inputDataTypes.get(inputNames.get(i))));
+        }
+        Map<String, INDArray> castConstants = new HashMap<>();
+        for (Map.Entry<String, INDArray> e: constants.entrySet()) {
+            DataType requiredDtype = TensorDataType.toNd4jType(TensorDataType.fromProtoValue(inputDataTypes.get(e.getKey())));
+            castConstants.put(e.getKey(), e.getValue().castTo(requiredDtype));
+        }
+        this.inputs = castConstants;
+        graphRunner = GraphRunner.builder().inputNames(inputNames)
+                .outputNames(outputNames).graphBytes(graphBytes)
+                .inputDataTypes(convertedDataTypes).build();
+        return this;
+
+    }
+
+    @Override
+    public Map<String, INDArray> run(Map<String, INDArray> inputs){
+        if (graphRunner == null){
+            throw new RuntimeException("GraphRunner not initialized.");
+        }
+        this.inputs.putAll(inputs);
+        return graphRunner.run(this.inputs);
+    }
+}

nd4j/nd4j-tensorflow/src/main/resources/META-INF/services/org.nd4j.TFGraphRunnerService

@@ -0,0 +1,17 @@
+  ################################################################################
+ # Copyright (c) 2020 Konduit K.K..
+ #
+ # 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
+ ################################################################################
+
+org.nd4j.tensorflow.conversion.graphrunner.GraphRunnerServiceProvider

pom.xml

@@ -292,9 +292,9 @@
         <javacpp-presets.version>1.5.3-SNAPSHOT</javacpp-presets.version>
         <javacv.version>1.5.3-SNAPSHOT</javacv.version>
 
-        <python.version>3.7.6</python.version>
+        <python.version>3.7.7</python.version>
         <cpython-platform.version>${python.version}-${javacpp-presets.version}</cpython-platform.version>
-        <numpy.version>1.18.1</numpy.version>
+        <numpy.version>1.18.2</numpy.version>
         <numpy.javacpp.version>${numpy.version}-${javacpp-presets.version}</numpy.javacpp.version>
 
         <openblas.version>0.3.9</openblas.version>