Merge pull request #8723 from KonduitAI/master

Merge recent development updates

arbiter/arbiter-core/src/main/java/org/deeplearning4j/arbiter/util/WebUtils.java

@@ -1,47 +0,0 @@
-/*******************************************************************************
- * Copyright (c) 2015-2018 Skymind, Inc.
- *
- * This program and the accompanying materials are made available under the
- * terms of the Apache License, Version 2.0 which is available at
- * https://www.apache.org/licenses/LICENSE-2.0.
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
- * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
- * License for the specific language governing permissions and limitations
- * under the License.
- *
- * SPDX-License-Identifier: Apache-2.0
- ******************************************************************************/
-
-package org.deeplearning4j.arbiter.util;
-
-import org.slf4j.Logger;
-
-import java.awt.*;
-import java.net.URI;
-
-/**
- * Various utilities for webpages and dealing with browsers
- */
-public class WebUtils {
-
-    public static void tryOpenBrowser(String path, Logger log) {
-        try {
-            WebUtils.openBrowser(new URI(path));
-        } catch (Exception e) {
-            log.error("Could not open browser", e);
-            System.out.println("Browser could not be launched automatically.\nUI path: " + path);
-        }
-    }
-
-    public static void openBrowser(URI uri) throws Exception {
-        if (Desktop.isDesktopSupported()) {
-            Desktop.getDesktop().browse(uri);
-        } else {
-            throw new UnsupportedOperationException(
-                            "Cannot open browser on this platform: Desktop.isDesktopSupported() == false");
-        }
-    }
-
-}

arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/BraninFunction.java

@@ -127,7 +127,7 @@ public class BraninFunction {
                     BraninConfig candidate = (BraninConfig) c.getValue();
 
                     double score = scoreFunction.score(candidate, null, (Map) null);
-                    System.out.println(candidate.getX1() + "\t" + candidate.getX2() + "\t" + score);
+//                    System.out.println(candidate.getX1() + "\t" + candidate.getX2() + "\t" + score);
 
                     Thread.sleep(20);
 

arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/TestRandomSearch.java

@@ -54,7 +54,7 @@ public class TestRandomSearch extends BaseDL4JTest {
         runner.execute();
 
 
-        System.out.println("----- Complete -----");
+//        System.out.println("----- Complete -----");
     }
 
 

arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/genetic/TestRandomGenerator.java

@@ -16,8 +16,8 @@
 
 package org.deeplearning4j.arbiter.optimize.genetic;
 
+import org.apache.commons.lang3.NotImplementedException;
 import org.apache.commons.math3.random.RandomGenerator;
-import sun.reflect.generics.reflectiveObjects.NotImplementedException;
 
 public class TestRandomGenerator implements RandomGenerator {
     private final int[] intRandomNumbers;
@@ -63,17 +63,17 @@ public class TestRandomGenerator implements RandomGenerator {
 
     @Override
     public long nextLong() {
-        throw new NotImplementedException();
+        throw new NotImplementedException("Not implemented");
     }
 
     @Override
     public boolean nextBoolean() {
-        throw new NotImplementedException();
+        throw new NotImplementedException("Not implemented");
     }
 
     @Override
     public float nextFloat() {
-        throw new NotImplementedException();
+        throw new NotImplementedException("Not implemented");
     }
 
     @Override
@@ -83,6 +83,6 @@ public class TestRandomGenerator implements RandomGenerator {
 
     @Override
     public double nextGaussian() {
-        throw new NotImplementedException();
+        throw new NotImplementedException("Not implemented");
     }
 }

arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/genetic/crossover/TwoParentsCrossoverOperatorTests.java

@@ -16,6 +16,7 @@
 
 package org.deeplearning4j.arbiter.optimize.genetic.crossover;
 
+import org.apache.commons.lang3.NotImplementedException;
 import org.deeplearning4j.BaseDL4JTest;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.crossover.CrossoverResult;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.crossover.TwoParentsCrossoverOperator;
@@ -26,7 +27,6 @@ import org.deeplearning4j.arbiter.optimize.genetic.TestParentSelection;
 import org.deeplearning4j.arbiter.optimize.genetic.TestPopulationInitializer;
 import org.junit.Assert;
 import org.junit.Test;
-import sun.reflect.generics.reflectiveObjects.NotImplementedException;
 
 public class TwoParentsCrossoverOperatorTests extends BaseDL4JTest {
 
@@ -42,7 +42,7 @@ public class TwoParentsCrossoverOperatorTests extends BaseDL4JTest {
 
         @Override
         public CrossoverResult crossover() {
-            throw new NotImplementedException();
+            throw new NotImplementedException("Not implemented");
         }
     }
 

arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/genetic/culling/RatioCullOperatorTests.java

@@ -16,6 +16,7 @@
 
 package org.deeplearning4j.arbiter.optimize.genetic.culling;
 
+import org.apache.commons.lang3.NotImplementedException;
 import org.deeplearning4j.BaseDL4JTest;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.Chromosome;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.culling.RatioCullOperator;
@@ -24,7 +25,6 @@ import org.deeplearning4j.arbiter.optimize.generator.genetic.population.Populati
 import org.deeplearning4j.arbiter.optimize.genetic.TestPopulationInitializer;
 import org.junit.Assert;
 import org.junit.Test;
-import sun.reflect.generics.reflectiveObjects.NotImplementedException;
 
 import java.util.List;
 
@@ -46,7 +46,7 @@ public class RatioCullOperatorTests extends BaseDL4JTest {
 
         @Override
         public void cullPopulation() {
-            throw new NotImplementedException();
+            throw new NotImplementedException("Not implemented");
         }
 
         public double getCullRatio() {

arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/genetic/selection/GeneticSelectionOperatorTests.java

@@ -16,6 +16,7 @@
 
 package org.deeplearning4j.arbiter.optimize.genetic.selection;
 
+import org.apache.commons.lang3.NotImplementedException;
 import org.apache.commons.math3.random.RandomGenerator;
 import org.deeplearning4j.BaseDL4JTest;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.ChromosomeFactory;
@@ -33,7 +34,6 @@ import org.deeplearning4j.arbiter.optimize.genetic.TestPopulationInitializer;
 import org.deeplearning4j.arbiter.optimize.genetic.TestRandomGenerator;
 import org.junit.Assert;
 import org.junit.Test;
-import sun.reflect.generics.reflectiveObjects.NotImplementedException;
 
 import static org.junit.Assert.assertArrayEquals;
 
@@ -55,7 +55,7 @@ public class GeneticSelectionOperatorTests extends BaseDL4JTest {
 
         @Override
         public void cullPopulation() {
-            throw new NotImplementedException();
+            throw new NotImplementedException("Not implemented");
         }
 
         @Override

arbiter/arbiter-core/src/test/java/org/deeplearning4j/arbiter/optimize/genetic/selection/SelectionOperatorTests.java

@@ -16,6 +16,7 @@
 
 package org.deeplearning4j.arbiter.optimize.genetic.selection;
 
+import org.apache.commons.lang3.NotImplementedException;
 import org.deeplearning4j.BaseDL4JTest;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.ChromosomeFactory;
 import org.deeplearning4j.arbiter.optimize.generator.genetic.population.PopulationInitializer;
@@ -24,7 +25,6 @@ import org.deeplearning4j.arbiter.optimize.generator.genetic.selection.Selection
 import org.deeplearning4j.arbiter.optimize.genetic.TestPopulationInitializer;
 import org.junit.Assert;
 import org.junit.Test;
-import sun.reflect.generics.reflectiveObjects.NotImplementedException;
 
 public class SelectionOperatorTests extends BaseDL4JTest {
     private class TestSelectionOperator extends SelectionOperator {
@@ -39,7 +39,7 @@ public class SelectionOperatorTests extends BaseDL4JTest {
 
         @Override
         public double[] buildNextGenes() {
-            throw new NotImplementedException();
+            throw new NotImplementedException("Not implemented");
         }
     }
 

arbiter/arbiter-deeplearning4j/src/test/java/org/deeplearning4j/arbiter/computationgraph/TestComputationGraphSpace.java

@@ -158,7 +158,7 @@ public class TestComputationGraphSpace extends BaseDL4JTest {
             }
         }
 
-        System.out.println("ReLU vs. Tanh: " + reluCount + "\t" + tanhCount);
+//        System.out.println("ReLU vs. Tanh: " + reluCount + "\t" + tanhCount);
         assertTrue(reluCount > 0);
         assertTrue(tanhCount > 0);
 

arbiter/arbiter-deeplearning4j/src/test/java/org/deeplearning4j/arbiter/computationgraph/TestGraphLocalExecution.java

@@ -162,7 +162,7 @@ public class TestGraphLocalExecution extends BaseDL4JTest {
             List<ResultReference> results = runner.getResults();
             assertTrue(results.size() > 0);
 
-            System.out.println("----- COMPLETE - " + results.size() + " results -----");
+//            System.out.println("----- COMPLETE - " + results.size() + " results -----");
         }
     }
 

arbiter/arbiter-deeplearning4j/src/test/java/org/deeplearning4j/arbiter/computationgraph/TestGraphLocalExecutionGenetic.java

@@ -165,7 +165,7 @@ public class TestGraphLocalExecutionGenetic extends BaseDL4JTest {
             List<ResultReference> results = runner.getResults();
             assertTrue(results.size() > 0);
 
-            System.out.println("----- COMPLETE - " + results.size() + " results -----");
+//            System.out.println("----- COMPLETE - " + results.size() + " results -----");
         }
     }
 

arbiter/arbiter-deeplearning4j/src/test/java/org/deeplearning4j/arbiter/multilayernetwork/TestLayerSpace.java

@@ -101,7 +101,7 @@ public class TestLayerSpace extends BaseDL4JTest {
             double l2 = TestUtils.getL2(l);
             IActivation activation = l.getActivationFn();
 
-            System.out.println(lr + "\t" + l2 + "\t" + activation);
+//            System.out.println(lr + "\t" + l2 + "\t" + activation);
 
             assertTrue(lr >= 0.3 && lr <= 0.4);
             assertTrue(l2 >= 0.01 && l2 <= 0.1);
@@ -190,7 +190,7 @@ public class TestLayerSpace extends BaseDL4JTest {
             ActivationLayer al = als.getValue(d);
             IActivation activation = al.getActivationFn();
 
-            System.out.println(activation);
+//            System.out.println(activation);
 
             assertTrue(containsActivationFunction(actFns, activation));
         }
@@ -228,7 +228,7 @@ public class TestLayerSpace extends BaseDL4JTest {
             IActivation activation = el.getActivationFn();
             long nOut = el.getNOut();
 
-            System.out.println(activation + "\t" + nOut);
+//            System.out.println(activation + "\t" + nOut);
 
             assertTrue(containsActivationFunction(actFns, activation));
             assertTrue(nOut >= 10 && nOut <= 20);
@@ -295,7 +295,7 @@ public class TestLayerSpace extends BaseDL4JTest {
             long nOut = el.getNOut();
             double forgetGate = el.getForgetGateBiasInit();
 
-            System.out.println(activation + "\t" + nOut + "\t" + forgetGate);
+//            System.out.println(activation + "\t" + nOut + "\t" + forgetGate);
 
             assertTrue(containsActivationFunction(actFns, activation));
             assertTrue(nOut >= 10 && nOut <= 20);

arbiter/arbiter-deeplearning4j/src/test/java/org/deeplearning4j/arbiter/multilayernetwork/TestMultiLayerSpace.java

@@ -293,8 +293,8 @@ public class TestMultiLayerSpace extends BaseDL4JTest {
             assertTrue(nLayerCounts[i] >= 5); //Expect approx equal (50/3 each), but some variation randomly
         }
 
-        System.out.println("Number of layers: " + Arrays.toString(nLayerCounts));
+//        System.out.println("Number of layers: " + Arrays.toString(nLayerCounts));
-        System.out.println("ReLU vs. Tanh: " + reluCount + "\t" + tanhCount);
+//        System.out.println("ReLU vs. Tanh: " + reluCount + "\t" + tanhCount);
 
     }
 

arbiter/arbiter-server/src/test/java/org/deeplearning4j/arbiter/server/ArbiterCLIRunnerTest.java

@@ -98,7 +98,8 @@ public class ArbiterCLIRunnerTest extends BaseDL4JTest {
         assertEquals(configuration,OptimizationConfiguration.fromJson(configuration.toJson()));
 
         FileUtils.writeStringToFile(new File(configPath),configuration.toJson());
-        System.out.println(configuration.toJson());
+//        System.out.println(configuration.toJson());
+        configuration.toJson();
 
         log.info("Starting test");
         cliRunner.runMain(

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/regressiontest/RegressionTest100b6.java

@@ -0,0 +1,390 @@
+/*
+ * Copyright (c) 2015-2019 Skymind, Inc.
+ *
+ * This program and the accompanying materials are made available under the
+ * terms of the Apache License, Version 2.0 which is available at
+ * https://www.apache.org/licenses/LICENSE-2.0.
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
+ * License for the specific language governing permissions and limitations
+ * under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+package org.deeplearning4j.regressiontest;
+
+import org.deeplearning4j.BaseDL4JTest;
+import org.deeplearning4j.TestUtils;
+import org.deeplearning4j.nn.conf.BackpropType;
+import org.deeplearning4j.nn.conf.ConvolutionMode;
+import org.deeplearning4j.nn.conf.graph.LayerVertex;
+import org.deeplearning4j.nn.conf.layers.*;
+import org.deeplearning4j.nn.conf.layers.convolutional.Cropping2D;
+import org.deeplearning4j.nn.conf.layers.recurrent.Bidirectional;
+import org.deeplearning4j.nn.conf.layers.recurrent.SimpleRnn;
+import org.deeplearning4j.nn.conf.layers.variational.VariationalAutoencoder;
+import org.deeplearning4j.nn.graph.ComputationGraph;
+import org.deeplearning4j.nn.graph.vertex.impl.MergeVertex;
+import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
+import org.deeplearning4j.nn.weights.WeightInitXavier;
+import org.deeplearning4j.regressiontest.customlayer100a.CustomLayer;
+import org.junit.Test;
+import org.nd4j.linalg.activations.impl.*;
+import org.nd4j.linalg.api.buffer.DataType;
+import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.factory.Nd4j;
+import org.nd4j.linalg.learning.config.Adam;
+import org.nd4j.linalg.learning.config.RmsProp;
+import org.nd4j.linalg.learning.regularization.L2Regularization;
+import org.nd4j.linalg.lossfunctions.impl.LossMAE;
+import org.nd4j.linalg.lossfunctions.impl.LossMCXENT;
+import org.nd4j.resources.Resources;
+
+import java.io.DataInputStream;
+import java.io.File;
+import java.io.FileInputStream;
+
+import static org.junit.Assert.*;
+
+public class RegressionTest100b6 extends BaseDL4JTest {
+
+    @Override
+    public DataType getDataType() {
+        return DataType.FLOAT;
+    }
+
+    @Test
+    public void testCustomLayer() throws Exception {
+
+        for (DataType dtype : new DataType[]{DataType.DOUBLE, DataType.FLOAT, DataType.HALF}) {
+
+            String dtypeName = dtype.toString().toLowerCase();
+
+            File f = Resources.asFile("regression_testing/100b6/CustomLayerExample_100b6_" + dtypeName + ".bin");
+            MultiLayerNetwork.load(f, true);
+
+            MultiLayerNetwork net = MultiLayerNetwork.load(f, true);
+//            net = net.clone();
+
+            DenseLayer l0 = (DenseLayer) net.getLayer(0).conf().getLayer();
+            assertEquals(new ActivationTanH(), l0.getActivationFn());
+            assertEquals(new L2Regularization(0.03), TestUtils.getL2Reg(l0));
+            assertEquals(new RmsProp(0.95), l0.getIUpdater());
+
+            CustomLayer l1 = (CustomLayer) net.getLayer(1).conf().getLayer();
+            assertEquals(new ActivationTanH(), l1.getActivationFn());
+            assertEquals(new ActivationSigmoid(), l1.getSecondActivationFunction());
+            assertEquals(new RmsProp(0.95), l1.getIUpdater());
+
+            INDArray outExp;
+            File f2 = Resources
+                    .asFile("regression_testing/100b6/CustomLayerExample_Output_100b6_" + dtypeName + ".bin");
+            try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+                outExp = Nd4j.read(dis);
+            }
+
+            INDArray in;
+            File f3 = Resources.asFile("regression_testing/100b6/CustomLayerExample_Input_100b6_" + dtypeName + ".bin");
+            try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+                in = Nd4j.read(dis);
+            }
+
+            assertEquals(dtype, in.dataType());
+            assertEquals(dtype, outExp.dataType());
+            assertEquals(dtype, net.params().dataType());
+            assertEquals(dtype, net.getFlattenedGradients().dataType());
+            assertEquals(dtype, net.getUpdater().getStateViewArray().dataType());
+
+            //System.out.println(Arrays.toString(net.params().data().asFloat()));
+
+            INDArray outAct = net.output(in);
+            assertEquals(dtype, outAct.dataType());
+
+            assertEquals(dtype, net.getLayerWiseConfigurations().getDataType());
+            assertEquals(dtype, net.params().dataType());
+            boolean eq = outExp.equalsWithEps(outAct, 0.01);
+            assertTrue(outExp + " vs " + outAct, eq);        }
+    }
+
+
+    @Test
+    public void testLSTM() throws Exception {
+
+        File f = Resources.asFile("regression_testing/100b6/GravesLSTMCharModelingExample_100b6.bin");
+        MultiLayerNetwork net = MultiLayerNetwork.load(f, true);
+
+        LSTM l0 = (LSTM) net.getLayer(0).conf().getLayer();
+        assertEquals(new ActivationTanH(), l0.getActivationFn());
+        assertEquals(200, l0.getNOut());
+        assertEquals(new WeightInitXavier(), l0.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l0));
+        assertEquals(new Adam(0.005), l0.getIUpdater());
+
+        LSTM l1 = (LSTM) net.getLayer(1).conf().getLayer();
+        assertEquals(new ActivationTanH(), l1.getActivationFn());
+        assertEquals(200, l1.getNOut());
+        assertEquals(new WeightInitXavier(), l1.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l1));
+        assertEquals(new Adam(0.005), l1.getIUpdater());
+
+        RnnOutputLayer l2 = (RnnOutputLayer) net.getLayer(2).conf().getLayer();
+        assertEquals(new ActivationSoftmax(), l2.getActivationFn());
+        assertEquals(77, l2.getNOut());
+        assertEquals(new WeightInitXavier(), l2.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l2));
+        assertEquals(new Adam(0.005), l2.getIUpdater());
+
+        assertEquals(BackpropType.TruncatedBPTT, net.getLayerWiseConfigurations().getBackpropType());
+        assertEquals(50, net.getLayerWiseConfigurations().getTbpttBackLength());
+        assertEquals(50, net.getLayerWiseConfigurations().getTbpttFwdLength());
+
+        INDArray outExp;
+        File f2 = Resources.asFile("regression_testing/100b6/GravesLSTMCharModelingExample_Output_100b6.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+            outExp = Nd4j.read(dis);
+        }
+
+        INDArray in;
+        File f3 = Resources.asFile("regression_testing/100b6/GravesLSTMCharModelingExample_Input_100b6.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+            in = Nd4j.read(dis);
+        }
+
+        INDArray outAct = net.output(in);
+
+        assertEquals(outExp, outAct);
+    }
+
+    @Test
+    public void testVae() throws Exception {
+
+        File f = Resources.asFile("regression_testing/100b6/VaeMNISTAnomaly_100b6.bin");
+        MultiLayerNetwork net = MultiLayerNetwork.load(f, true);
+
+        VariationalAutoencoder l0 = (VariationalAutoencoder) net.getLayer(0).conf().getLayer();
+        assertEquals(new ActivationLReLU(), l0.getActivationFn());
+        assertEquals(32, l0.getNOut());
+        assertArrayEquals(new int[]{256, 256}, l0.getEncoderLayerSizes());
+        assertArrayEquals(new int[]{256, 256}, l0.getDecoderLayerSizes());
+        assertEquals(new WeightInitXavier(), l0.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l0));
+        assertEquals(new Adam(1e-3), l0.getIUpdater());
+
+        INDArray outExp;
+        File f2 = Resources.asFile("regression_testing/100b6/VaeMNISTAnomaly_Output_100b6.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+            outExp = Nd4j.read(dis);
+        }
+
+        INDArray in;
+        File f3 = Resources.asFile("regression_testing/100b6/VaeMNISTAnomaly_Input_100b6.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+            in = Nd4j.read(dis);
+        }
+
+        INDArray outAct = net.output(in);
+
+        assertEquals(outExp, outAct);
+    }
+
+
+    @Test
+    public void testYoloHouseNumber() throws Exception {
+
+        File f = Resources.asFile("regression_testing/100b6/HouseNumberDetection_100b6.bin");
+        ComputationGraph net = ComputationGraph.load(f, true);
+
+        int nBoxes = 5;
+        int nClasses = 10;
+
+        ConvolutionLayer cl = (ConvolutionLayer) ((LayerVertex) net.getConfiguration().getVertices()
+                .get("convolution2d_9")).getLayerConf().getLayer();
+        assertEquals(nBoxes * (5 + nClasses), cl.getNOut());
+        assertEquals(new ActivationIdentity(), cl.getActivationFn());
+        assertEquals(ConvolutionMode.Same, cl.getConvolutionMode());
+        assertEquals(new WeightInitXavier(), cl.getWeightInitFn());
+        assertArrayEquals(new int[]{1, 1}, cl.getKernelSize());
+
+        INDArray outExp;
+        File f2 = Resources.asFile("regression_testing/100b6/HouseNumberDetection_Output_100b6.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+            outExp = Nd4j.read(dis);
+        }
+
+        INDArray in;
+        File f3 = Resources.asFile("regression_testing/100b6/HouseNumberDetection_Input_100b6.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+            in = Nd4j.read(dis);
+        }
+
+        INDArray outAct = net.outputSingle(in);
+
+        boolean eq = outExp.equalsWithEps(outAct.castTo(outExp.dataType()), 1e-3);
+        assertTrue(eq);
+    }
+
+    @Test
+    public void testSyntheticCNN() throws Exception {
+
+        File f = Resources.asFile("regression_testing/100b6/SyntheticCNN_100b6.bin");
+        MultiLayerNetwork net = MultiLayerNetwork.load(f, true);
+
+        ConvolutionLayer l0 = (ConvolutionLayer) net.getLayer(0).conf().getLayer();
+        assertEquals(new ActivationReLU(), l0.getActivationFn());
+        assertEquals(4, l0.getNOut());
+        assertEquals(new WeightInitXavier(), l0.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l0));
+        assertEquals(new Adam(0.005), l0.getIUpdater());
+        assertArrayEquals(new int[]{3, 3}, l0.getKernelSize());
+        assertArrayEquals(new int[]{2, 1}, l0.getStride());
+        assertArrayEquals(new int[]{1, 1}, l0.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l0.getPadding());
+
+        SeparableConvolution2D l1 = (SeparableConvolution2D) net.getLayer(1).conf().getLayer();
+        assertEquals(new ActivationReLU(), l1.getActivationFn());
+        assertEquals(8, l1.getNOut());
+        assertEquals(new WeightInitXavier(), l1.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l1));
+        assertEquals(new Adam(0.005), l1.getIUpdater());
+        assertArrayEquals(new int[]{3, 3}, l1.getKernelSize());
+        assertArrayEquals(new int[]{1, 1}, l1.getStride());
+        assertArrayEquals(new int[]{1, 1}, l1.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l1.getPadding());
+        assertEquals(ConvolutionMode.Same, l1.getConvolutionMode());
+        assertEquals(1, l1.getDepthMultiplier());
+
+        SubsamplingLayer l2 = (SubsamplingLayer) net.getLayer(2).conf().getLayer();
+        assertArrayEquals(new int[]{3, 3}, l2.getKernelSize());
+        assertArrayEquals(new int[]{2, 2}, l2.getStride());
+        assertArrayEquals(new int[]{1, 1}, l2.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l2.getPadding());
+        assertEquals(PoolingType.MAX, l2.getPoolingType());
+
+        ZeroPaddingLayer l3 = (ZeroPaddingLayer) net.getLayer(3).conf().getLayer();
+        assertArrayEquals(new int[]{4, 4, 4, 4}, l3.getPadding());
+
+        Upsampling2D l4 = (Upsampling2D) net.getLayer(4).conf().getLayer();
+        assertArrayEquals(new int[]{3, 3}, l4.getSize());
+
+        DepthwiseConvolution2D l5 = (DepthwiseConvolution2D) net.getLayer(5).conf().getLayer();
+        assertEquals(new ActivationReLU(), l5.getActivationFn());
+        assertEquals(16, l5.getNOut());
+        assertEquals(new WeightInitXavier(), l5.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l5));
+        assertEquals(new Adam(0.005), l5.getIUpdater());
+        assertArrayEquals(new int[]{3, 3}, l5.getKernelSize());
+        assertArrayEquals(new int[]{1, 1}, l5.getStride());
+        assertArrayEquals(new int[]{1, 1}, l5.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l5.getPadding());
+        assertEquals(2, l5.getDepthMultiplier());
+
+        SubsamplingLayer l6 = (SubsamplingLayer) net.getLayer(6).conf().getLayer();
+        assertArrayEquals(new int[]{2, 2}, l6.getKernelSize());
+        assertArrayEquals(new int[]{2, 2}, l6.getStride());
+        assertArrayEquals(new int[]{1, 1}, l6.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l6.getPadding());
+        assertEquals(PoolingType.MAX, l6.getPoolingType());
+
+        Cropping2D l7 = (Cropping2D) net.getLayer(7).conf().getLayer();
+        assertArrayEquals(new int[]{3, 3, 2, 2}, l7.getCropping());
+
+        ConvolutionLayer l8 = (ConvolutionLayer) net.getLayer(8).conf().getLayer();
+        assertEquals(4, l8.getNOut());
+        assertEquals(new WeightInitXavier(), l8.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l8));
+        assertEquals(new Adam(0.005), l8.getIUpdater());
+        assertArrayEquals(new int[]{4, 4}, l8.getKernelSize());
+        assertArrayEquals(new int[]{1, 1}, l8.getStride());
+        assertArrayEquals(new int[]{1, 1}, l8.getDilation());
+        assertArrayEquals(new int[]{0, 0}, l8.getPadding());
+
+        CnnLossLayer l9 = (CnnLossLayer) net.getLayer(9).conf().getLayer();
+        assertEquals(new WeightInitXavier(), l9.getWeightInitFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l9));
+        assertEquals(new Adam(0.005), l9.getIUpdater());
+        assertEquals(new LossMAE(), l9.getLossFn());
+
+        INDArray outExp;
+        File f2 = Resources.asFile("regression_testing/100b6/SyntheticCNN_Output_100b6.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+            outExp = Nd4j.read(dis);
+        }
+
+        INDArray in;
+        File f3 = Resources.asFile("regression_testing/100b6/SyntheticCNN_Input_100b6.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+            in = Nd4j.read(dis);
+        }
+
+        INDArray outAct = net.output(in);
+
+        //19 layers - CPU vs. GPU difference accumulates notably, but appears to be correct
+        if(Nd4j.getBackend().getClass().getName().toLowerCase().contains("native")){
+            assertEquals(outExp, outAct);
+        } else {
+            boolean eq = outExp.equalsWithEps(outAct, 0.1);
+            assertTrue(eq);
+        }
+    }
+
+    @Test
+    public void testSyntheticBidirectionalRNNGraph() throws Exception {
+
+        File f = Resources.asFile("regression_testing/100b6/SyntheticBidirectionalRNNGraph_100b6.bin");
+        ComputationGraph net = ComputationGraph.load(f, true);
+
+        Bidirectional l0 = (Bidirectional) net.getLayer("rnn1").conf().getLayer();
+
+        LSTM l1 = (LSTM) l0.getFwd();
+        assertEquals(16, l1.getNOut());
+        assertEquals(new ActivationReLU(), l1.getActivationFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l1));
+
+        LSTM l2 = (LSTM) l0.getBwd();
+        assertEquals(16, l2.getNOut());
+        assertEquals(new ActivationReLU(), l2.getActivationFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l2));
+
+        Bidirectional l3 = (Bidirectional) net.getLayer("rnn2").conf().getLayer();
+
+        SimpleRnn l4 = (SimpleRnn) l3.getFwd();
+        assertEquals(16, l4.getNOut());
+        assertEquals(new ActivationReLU(), l4.getActivationFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l4));
+
+        SimpleRnn l5 = (SimpleRnn) l3.getBwd();
+        assertEquals(16, l5.getNOut());
+        assertEquals(new ActivationReLU(), l5.getActivationFn());
+        assertEquals(new L2Regularization(0.0001), TestUtils.getL2Reg(l5));
+
+        MergeVertex mv = (MergeVertex) net.getVertex("concat");
+
+        GlobalPoolingLayer gpl = (GlobalPoolingLayer) net.getLayer("pooling").conf().getLayer();
+        assertEquals(PoolingType.MAX, gpl.getPoolingType());
+        assertArrayEquals(new int[]{2}, gpl.getPoolingDimensions());
+        assertTrue(gpl.isCollapseDimensions());
+
+        OutputLayer outl = (OutputLayer) net.getLayer("out").conf().getLayer();
+        assertEquals(3, outl.getNOut());
+        assertEquals(new LossMCXENT(), outl.getLossFn());
+
+        INDArray outExp;
+        File f2 = Resources.asFile("regression_testing/100b6/SyntheticBidirectionalRNNGraph_Output_100b6.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f2))) {
+            outExp = Nd4j.read(dis);
+        }
+
+        INDArray in;
+        File f3 = Resources.asFile("regression_testing/100b6/SyntheticBidirectionalRNNGraph_Input_100b6.bin");
+        try (DataInputStream dis = new DataInputStream(new FileInputStream(f3))) {
+            in = Nd4j.read(dis);
+        }
+
+        INDArray outAct = net.output(in)[0];
+
+        assertEquals(outExp, outAct);
+    }
+}

deeplearning4j/deeplearning4j-graph/src/test/java/org/deeplearning4j/graph/data/TestGraphLoading.java

@@ -41,7 +41,7 @@ public class TestGraphLoading extends BaseDL4JTest {
 
         IGraph<String, String> graph = GraphLoader
                         .loadUndirectedGraphEdgeListFile(cpr.getTempFileFromArchive().getAbsolutePath(), 7, ",");
-        System.out.println(graph);
+//        System.out.println(graph);
 
         assertEquals(graph.numVertices(), 7);
         int[][] edges = {{1, 2}, {0, 2, 4}, {0, 1, 3, 4}, {2, 4, 5}, {1, 2, 3, 5, 6}, {3, 4, 6}, {4, 5}};
@@ -66,7 +66,7 @@ public class TestGraphLoading extends BaseDL4JTest {
                         edgeLineProcessor, vertexFactory, 10, false);
 
 
-        System.out.println(graph);
+//        System.out.println(graph);
 
         for (int i = 0; i < 10; i++) {
             List<Edge<String>> edges = graph.getEdgesOut(i);
@@ -111,7 +111,7 @@ public class TestGraphLoading extends BaseDL4JTest {
         Graph<String, String> graph = GraphLoader.loadGraph(verticesCPR.getTempFileFromArchive().getAbsolutePath(),
                         edgesCPR.getTempFileFromArchive().getAbsolutePath(), vertexLoader, edgeLineProcessor, false);
 
-        System.out.println(graph);
+//        System.out.println(graph);
 
         for (int i = 0; i < 10; i++) {
             List<Edge<String>> edges = graph.getEdgesOut(i);

deeplearning4j/deeplearning4j-graph/src/test/java/org/deeplearning4j/graph/data/TestGraphLoadingWeighted.java

@@ -71,7 +71,7 @@ public class TestGraphLoadingWeighted extends BaseDL4JTest {
             }
         }
 
-        System.out.println(graph);
+//        System.out.println(graph);
     }
 
 

deeplearning4j/deeplearning4j-graph/src/test/java/org/deeplearning4j/graph/graph/TestGraph.java

@@ -220,7 +220,7 @@ public class TestGraph extends BaseDL4JTest {
                 sum += transitionProb[i][j];
             for (int j = 0; j < transitionProb[i].length; j++)
                 transitionProb[i][j] /= sum;
-            System.out.println(Arrays.toString(transitionProb[i]));
+//            System.out.println(Arrays.toString(transitionProb[i]));
         }
 
         //Check that transition probs are essentially correct (within bounds of random variation)

deeplearning4j/deeplearning4j-graph/src/test/java/org/deeplearning4j/graph/models/deepwalk/DeepWalkGradientCheck.java

@@ -145,8 +145,8 @@ public class DeepWalkGradientCheck extends BaseDL4JTest {
 
                         if (relError > MAX_REL_ERROR && absErr > MIN_ABS_ERROR)
                             fail(msg);
-                        else
+//                        else
-                            System.out.println(msg);
+//                            System.out.println(msg);
                     }
                 }
 
@@ -333,10 +333,10 @@ public class DeepWalkGradientCheck extends BaseDL4JTest {
 
                     if (relError > MAX_REL_ERROR && absErr > MIN_ABS_ERROR)
                         fail(msg);
-                    else
+//                    else
-                        System.out.println(msg);
+//                        System.out.println(msg);
                 }
-                System.out.println();
+//                System.out.println();
             }
 
         }

deeplearning4j/deeplearning4j-graph/src/test/java/org/deeplearning4j/graph/models/deepwalk/TestDeepWalk.java

@@ -67,7 +67,7 @@ public class TestDeepWalk extends BaseDL4JTest {
         for (int i = 0; i < 7; i++) {
             INDArray vector = deepWalk.getVertexVector(i);
             assertArrayEquals(new long[] {vectorSize}, vector.shape());
-            System.out.println(Arrays.toString(vector.dup().data().asFloat()));
+//            System.out.println(Arrays.toString(vector.dup().data().asFloat()));
         }
 
         GraphWalkIterator<String> iter = new RandomWalkIterator<>(graph, 8);
@@ -77,11 +77,11 @@ public class TestDeepWalk extends BaseDL4JTest {
         for (int t = 0; t < 5; t++) {
             iter.reset();
             deepWalk.fit(iter);
-            System.out.println("--------------------");
+//            System.out.println("--------------------");
             for (int i = 0; i < 7; i++) {
                 INDArray vector = deepWalk.getVertexVector(i);
                 assertArrayEquals(new long[] {vectorSize}, vector.shape());
-                System.out.println(Arrays.toString(vector.dup().data().asFloat()));
+//                System.out.println(Arrays.toString(vector.dup().data().asFloat()));
             }
         }
     }
@@ -160,7 +160,7 @@ public class TestDeepWalk extends BaseDL4JTest {
                 continue;
 
             double sim = deepWalk.similarity(i, nearestTo);
-            System.out.println(i + "\t" + nearestTo + "\t" + sim);
+//            System.out.println(i + "\t" + nearestTo + "\t" + sim);
             assertTrue(sim <= minSimNearest);
         }
     }
@@ -211,7 +211,7 @@ public class TestDeepWalk extends BaseDL4JTest {
         Graph<String, String> graph = GraphLoader
                         .loadUndirectedGraphEdgeListFile(cpr.getTempFileFromArchive().getAbsolutePath(), 13, ",");
 
-        System.out.println(graph);
+//        System.out.println(graph);
 
         Nd4j.getRandom().setSeed(12345);
 
@@ -229,11 +229,13 @@ public class TestDeepWalk extends BaseDL4JTest {
 
         //Calculate similarity(0,i)
         for (int i = 0; i < nVertices; i++) {
-            System.out.println(deepWalk.similarity(0, i));
+//            System.out.println(deepWalk.similarity(0, i));
+            deepWalk.similarity(0, i);
         }
 
         for (int i = 0; i < nVertices; i++)
-            System.out.println(deepWalk.getVertexVector(i));
+//            System.out.println(deepWalk.getVertexVector(i));
+            deepWalk.getVertexVector(i);
     }
 
     @Test(timeout = 60000L)

deeplearning4j/deeplearning4j-graph/src/test/java/org/deeplearning4j/graph/models/deepwalk/TestGraphHuffman.java

@@ -38,9 +38,11 @@ public class TestGraphHuffman extends BaseDL4JTest {
 
         gh.buildTree(vertexDegrees);
 
-        for (int i = 0; i < 7; i++)
+        for (int i = 0; i < 7; i++) {
-            System.out.println(i + "\t" + gh.getCodeLength(i) + "\t" + gh.getCodeString(i) + "\t\t" + gh.getCode(i)
+            String s = i + "\t" + gh.getCodeLength(i) + "\t" + gh.getCodeString(i) + "\t\t" + gh.getCode(i)
-                            + "\t\t" + Arrays.toString(gh.getPathInnerNodes(i)));
+                    + "\t\t" + Arrays.toString(gh.getPathInnerNodes(i));
+//            System.out.println(s);
+        }
 
         int[] expectedLengths = {3, 2, 2, 5, 4, 2, 5};
         for (int i = 0; i < vertexDegrees.length; i++) {

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/util/DL4JModelValidator.java

@@ -3,6 +3,7 @@ package org.deeplearning4j.util;
 import lombok.NonNull;
 import org.apache.commons.io.IOUtils;
 import org.deeplearning4j.nn.api.Model;
+import org.deeplearning4j.nn.conf.ComputationGraphConfiguration;
 import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.graph.ComputationGraph;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
@@ -121,7 +122,7 @@ public class DL4JModelValidator {
         }
 
         try{
-            MultiLayerConfiguration.fromJson(config);
+            ComputationGraphConfiguration.fromJson(config);
         } catch (Throwable t){
             return ValidationResult.builder()
                     .formatType("ComputationGraph")

deeplearning4j/deeplearning4j-scaleout/deeplearning4j-scaleout-parallelwrapper-parameter-server/src/test/java/org/deeplearning4j/parallelism/parameterserver/ParameterServerParallelWrapperTest.java

@@ -79,8 +79,9 @@ public class ParameterServerParallelWrapperTest extends BaseDL4JTest {
         model.init();
 
         ParallelWrapper parameterServerParallelWrapper =
-                        new ParallelWrapper.Builder(model).trainerFactory(new ParameterServerTrainerContext())
+                        new ParallelWrapper.Builder(model)
-                                        .workers(Runtime.getRuntime().availableProcessors())
+                                .workers(Math.min(4, Runtime.getRuntime().availableProcessors()))
+                                .trainerFactory(new ParameterServerTrainerContext())
                                         .reportScoreAfterAveraging(true).prefetchBuffer(3).build();
         parameterServerParallelWrapper.fit(mnistTrain);
 

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark-nlp-java8/src/test/java/org/deeplearning4j/spark/models/word2vec/SparkWord2VecTest.java

@@ -104,7 +104,7 @@ public class SparkWord2VecTest extends BaseDL4JTest {
         public void call(ExportContainer<VocabWord> v) throws Exception {
             assertNotNull(v.getElement());
             assertNotNull(v.getArray());
-            System.out.println(v.getElement() + " - " + v.getArray());
+//            System.out.println(v.getElement() + " - " + v.getArray());
         }
     }
 }

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/TestEarlyStoppingSpark.java

@@ -66,7 +66,7 @@ public class TestEarlyStoppingSpark extends BaseSparkTest {
                                         .lossFunction(LossFunctions.LossFunction.MCXENT).build())
                         .build();
         MultiLayerNetwork net = new MultiLayerNetwork(conf);
-        net.setListeners(new ScoreIterationListener(1));
+        net.setListeners(new ScoreIterationListener(5));
 
 
         JavaRDD<DataSet> irisData = getIris();
@@ -119,7 +119,7 @@ public class TestEarlyStoppingSpark extends BaseSparkTest {
                                         .lossFunction(LossFunctions.LossFunction.MSE).build())
                         .build();
         MultiLayerNetwork net = new MultiLayerNetwork(conf);
-        net.setListeners(new ScoreIterationListener(1));
+        net.setListeners(new ScoreIterationListener(5));
 
         JavaRDD<DataSet> irisData = getIris();
         EarlyStoppingModelSaver<MultiLayerNetwork> saver = new InMemoryModelSaver<>();
@@ -155,7 +155,7 @@ public class TestEarlyStoppingSpark extends BaseSparkTest {
                                         .lossFunction(LossFunctions.LossFunction.MCXENT).build())
                         .build();
         MultiLayerNetwork net = new MultiLayerNetwork(conf);
-        net.setListeners(new ScoreIterationListener(1));
+        net.setListeners(new ScoreIterationListener(5));
 
         JavaRDD<DataSet> irisData = getIris();
 
@@ -198,7 +198,7 @@ public class TestEarlyStoppingSpark extends BaseSparkTest {
                                         .lossFunction(LossFunctions.LossFunction.MCXENT).build())
                         .build();
         MultiLayerNetwork net = new MultiLayerNetwork(conf);
-        net.setListeners(new ScoreIterationListener(1));
+        net.setListeners(new ScoreIterationListener(5));
 
         JavaRDD<DataSet> irisData = getIris();
 
@@ -231,7 +231,7 @@ public class TestEarlyStoppingSpark extends BaseSparkTest {
                                         .lossFunction(LossFunctions.LossFunction.MCXENT).build())
                         .build();
         MultiLayerNetwork net = new MultiLayerNetwork(conf);
-        net.setListeners(new ScoreIterationListener(1));
+        net.setListeners(new ScoreIterationListener(5));
 
 
         JavaRDD<DataSet> irisData = getIris();

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/TestEarlyStoppingSparkCompGraph.java

@@ -69,7 +69,7 @@ public class TestEarlyStoppingSparkCompGraph extends BaseSparkTest {
                                         .lossFunction(LossFunctions.LossFunction.MCXENT).build(), "in")
                         .setOutputs("0").build();
         ComputationGraph net = new ComputationGraph(conf);
-        net.setListeners(new ScoreIterationListener(1));
+        net.setListeners(new ScoreIterationListener(5));
 
 
         JavaRDD<DataSet> irisData = getIris();
@@ -120,7 +120,7 @@ public class TestEarlyStoppingSparkCompGraph extends BaseSparkTest {
                                         .lossFunction(LossFunctions.LossFunction.MSE).build(), "in")
                         .setOutputs("0").build();
         ComputationGraph net = new ComputationGraph(conf);
-        net.setListeners(new ScoreIterationListener(1));
+        net.setListeners(new ScoreIterationListener(5));
 
         JavaRDD<DataSet> irisData = getIris();
         EarlyStoppingModelSaver<ComputationGraph> saver = new InMemoryModelSaver<>();
@@ -158,7 +158,7 @@ public class TestEarlyStoppingSparkCompGraph extends BaseSparkTest {
                                         .lossFunction(LossFunctions.LossFunction.MCXENT).build(), "in")
                         .setOutputs("0").build();
         ComputationGraph net = new ComputationGraph(conf);
-        net.setListeners(new ScoreIterationListener(1));
+        net.setListeners(new ScoreIterationListener(5));
 
         JavaRDD<DataSet> irisData = getIris();
 
@@ -203,7 +203,7 @@ public class TestEarlyStoppingSparkCompGraph extends BaseSparkTest {
                                         .lossFunction(LossFunctions.LossFunction.MCXENT).build(), "in")
                         .setOutputs("0").build();
         ComputationGraph net = new ComputationGraph(conf);
-        net.setListeners(new ScoreIterationListener(1));
+        net.setListeners(new ScoreIterationListener(5));
 
         JavaRDD<DataSet> irisData = getIris();
 
@@ -238,7 +238,7 @@ public class TestEarlyStoppingSparkCompGraph extends BaseSparkTest {
                                         .lossFunction(LossFunctions.LossFunction.MCXENT).build(), "in")
                         .setOutputs("0").build();
         ComputationGraph net = new ComputationGraph(conf);
-        net.setListeners(new ScoreIterationListener(1));
+        net.setListeners(new ScoreIterationListener(5));
 
 
         JavaRDD<DataSet> irisData = getIris();

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/data/TestShuffleExamples.java

@@ -59,7 +59,7 @@ public class TestShuffleExamples extends BaseSparkTest {
         int totalExampleCount = 0;
         for (DataSet ds : shuffledList) {
             totalExampleCount += ds.getFeatures().length();
-            System.out.println(Arrays.toString(ds.getFeatures().data().asFloat()));
+//            System.out.println(Arrays.toString(ds.getFeatures().data().asFloat()));
 
             assertEquals(ds.getFeatures(), ds.getLabels());
         }

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/datavec/TestExport.java

@@ -86,7 +86,7 @@ public class TestExport extends BaseSparkTest {
         for (File file : files) {
             if (!file.getPath().endsWith(".bin"))
                 continue;
-            System.out.println(file);
+//            System.out.println(file);
             DataSet ds = new DataSet();
             ds.load(file);
             assertEquals(minibatchSize, ds.numExamples());
@@ -144,7 +144,7 @@ public class TestExport extends BaseSparkTest {
         for (File file : files) {
             if (!file.getPath().endsWith(".bin"))
                 continue;
-            System.out.println(file);
+//            System.out.println(file);
             MultiDataSet ds = new org.nd4j.linalg.dataset.MultiDataSet();
             ds.load(file);
             assertEquals(minibatchSize, ds.getFeatures(0).size(0));

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/impl/common/repartition/HashingBalancedPartitionerTest.java

@@ -92,9 +92,9 @@ public class HashingBalancedPartitionerTest extends BaseSparkTest {
 
         int[][] colorCountsByPartition = new int[3][2];
         for (final Tuple2<Tuple2<Long, Integer>, String> val : testList) {
-            System.out.println(val);
+//            System.out.println(val);
             Integer partition = hbp.getPartition(val._1());
-            System.out.println(partition);
+//            System.out.println(partition);
 
             if (val._2().equals("red"))
                 colorCountsByPartition[partition][0] += 1;
@@ -102,9 +102,9 @@ public class HashingBalancedPartitionerTest extends BaseSparkTest {
                 colorCountsByPartition[partition][1] += 1;
         }
 
-        for (int i = 0; i < 3; i++) {
+//        for (int i = 0; i < 3; i++) {
-            System.out.println(Arrays.toString(colorCountsByPartition[i]));
+//            System.out.println(Arrays.toString(colorCountsByPartition[i]));
-        }
+//        }
         for (int i = 0; i < 3; i++) {
             // avg red per partition : 2.33
             assertTrue(colorCountsByPartition[i][0] >= 1 && colorCountsByPartition[i][0] < 4);
@@ -178,12 +178,12 @@ public class HashingBalancedPartitionerTest extends BaseSparkTest {
                 colorCountsByPartition[partition][1] += 1;
         }
 
-        for (int i = 0; i < numPartitions; i++) {
+//        for (int i = 0; i < numPartitions; i++) {
-            System.out.println(Arrays.toString(colorCountsByPartition[i]));
+//            System.out.println(Arrays.toString(colorCountsByPartition[i]));
-        }
+//        }
-
+//
-        System.out.println("Ideal red # per partition: " + avgRed);
+//        System.out.println("Ideal red # per partition: " + avgRed);
-        System.out.println("Ideal blue # per partition: " + avgBlue);
+//        System.out.println("Ideal blue # per partition: " + avgBlue);
 
         for (int i = 0; i < numPartitions; i++) {
             // avg red per partition : 2.33

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/impl/graph/TestSparkComputationGraph.java

@@ -115,7 +115,7 @@ public class TestSparkComputationGraph extends BaseSparkTest {
         TrainingMaster tm = new ParameterAveragingTrainingMaster(true, numExecutors(), 1, 10, 1, 0);
 
         SparkComputationGraph scg = new SparkComputationGraph(sc, cg, tm);
-        scg.setListeners(Collections.singleton((TrainingListener) new ScoreIterationListener(1)));
+        scg.setListeners(Collections.singleton((TrainingListener) new ScoreIterationListener(5)));
 
         JavaRDD<MultiDataSet> rdd = sc.parallelize(list);
         scg.fitMultiDataSet(rdd);

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/impl/multilayer/TestSparkDl4jMultiLayer.java

@@ -31,8 +31,11 @@ import org.deeplearning4j.spark.impl.paramavg.ParameterAveragingTrainingMaster;
 import org.junit.Test;
 import org.nd4j.evaluation.classification.Evaluation;
 import org.nd4j.linalg.activations.Activation;
+import org.nd4j.linalg.api.buffer.DataType;
 import org.nd4j.linalg.dataset.DataSet;
 import org.nd4j.linalg.dataset.api.iterator.DataSetIterator;
+import org.nd4j.linalg.factory.Nd4j;
+import org.nd4j.linalg.learning.config.Adam;
 import org.nd4j.linalg.learning.config.Nesterovs;
 import org.nd4j.linalg.lossfunctions.LossFunctions;
 
@@ -45,8 +48,24 @@ import static org.junit.Assert.assertTrue;
 @Slf4j
 public class TestSparkDl4jMultiLayer extends BaseSparkTest {
 
-    @Test(timeout = 120000L)
+    @Override
+    public long getTimeoutMilliseconds() {
+        return 120000L;
+    }
+
+    @Override
+    public DataType getDataType() {
+        return DataType.FLOAT;
+    }
+
+    @Override
+    public DataType getDefaultFPDataType() {
+        return DataType.FLOAT;
+    }
+
+    @Test
     public void testEvaluationSimple() throws Exception {
+        Nd4j.getRandom().setSeed(12345);
 
         for( int evalWorkers : new int[]{1, 4, 8}) {
             //Simple test to validate DL4J issue 4099 is fixed...
@@ -75,18 +94,18 @@ public class TestSparkDl4jMultiLayer extends BaseSparkTest {
             //----------------------------------
             //Create network configuration and conduct network training
             MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
+                    .dataType(DataType.FLOAT)
                     .seed(12345)
                     .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
                     .activation(Activation.LEAKYRELU)
                     .weightInit(WeightInit.XAVIER)
-                    .updater(new Nesterovs(0.02, 0.9))
+                    .updater(new Adam(1e-3))
-                    .l2(1e-4)
+                    .l2(1e-5)
                     .list()
                     .layer(0, new DenseLayer.Builder().nIn(28 * 28).nOut(500).build())
                     .layer(1, new DenseLayer.Builder().nIn(500).nOut(100).build())
                     .layer(2, new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)
                             .activation(Activation.SOFTMAX).nIn(100).nOut(10).build())
-
                     .build();
 
             //Configuration for Spark training: see https://deeplearning4j.org/docs/latest/deeplearning4j-scaleout-howto for explanation of these configuration options

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/impl/paramavg/TestCompareParameterAveragingSparkVsSingleMachine.java

@@ -333,15 +333,16 @@ public class TestCompareParameterAveragingSparkVsSingleMachine {
                     sparkNet.fit(rdd);
                 }
 
-                System.out.println(sparkNet.getSparkTrainingStats().statsAsString());
+//                System.out.println(sparkNet.getSparkTrainingStats().statsAsString());
+                sparkNet.getSparkTrainingStats().statsAsString();
 
                 INDArray finalSparkParams = sparkNet.getNetwork().params().dup();
 
-                System.out.println("Initial (Local) params:       " + Arrays.toString(initialParams.data().asFloat()));
+//                System.out.println("Initial (Local) params:       " + Arrays.toString(initialParams.data().asFloat()));
-                System.out.println("Initial (Spark) params:       "
+//                System.out.println("Initial (Spark) params:       "
-                                + Arrays.toString(initialSparkParams.data().asFloat()));
+//                                + Arrays.toString(initialSparkParams.data().asFloat()));
-                System.out.println("Final (Local) params: " + Arrays.toString(finalParams.data().asFloat()));
+//                System.out.println("Final (Local) params: " + Arrays.toString(finalParams.data().asFloat()));
-                System.out.println("Final (Spark) params: " + Arrays.toString(finalSparkParams.data().asFloat()));
+//                System.out.println("Final (Spark) params: " + Arrays.toString(finalSparkParams.data().asFloat()));
                 assertEquals(initialParams, initialSparkParams);
                 assertNotEquals(initialParams, finalParams);
                 assertEquals(finalParams, finalSparkParams);
@@ -405,15 +406,16 @@ public class TestCompareParameterAveragingSparkVsSingleMachine {
                     sparkNet.fit(rdd);
                 }
 
-                System.out.println(sparkNet.getSparkTrainingStats().statsAsString());
+//                System.out.println(sparkNet.getSparkTrainingStats().statsAsString());
+                sparkNet.getSparkTrainingStats().statsAsString();
 
                 INDArray finalSparkParams = sparkNet.getNetwork().params().dup();
 
-                System.out.println("Initial (Local) params:       " + Arrays.toString(initialParams.data().asFloat()));
+//                System.out.println("Initial (Local) params:       " + Arrays.toString(initialParams.data().asFloat()));
-                System.out.println("Initial (Spark) params:       "
+//                System.out.println("Initial (Spark) params:       "
-                                + Arrays.toString(initialSparkParams.data().asFloat()));
+//                                + Arrays.toString(initialSparkParams.data().asFloat()));
-                System.out.println("Final (Local) params: " + Arrays.toString(finalParams.data().asFloat()));
+//                System.out.println("Final (Local) params: " + Arrays.toString(finalParams.data().asFloat()));
-                System.out.println("Final (Spark) params: " + Arrays.toString(finalSparkParams.data().asFloat()));
+//                System.out.println("Final (Spark) params: " + Arrays.toString(finalSparkParams.data().asFloat()));
                 assertArrayEquals(initialParams.data().asFloat(), initialSparkParams.data().asFloat(), 1e-8f);
                 assertArrayEquals(finalParams.data().asFloat(), finalSparkParams.data().asFloat(), 1e-6f);
 
@@ -478,18 +480,19 @@ public class TestCompareParameterAveragingSparkVsSingleMachine {
                     sparkNet.fit(rdd);
                 }
 
-                System.out.println(sparkNet.getSparkTrainingStats().statsAsString());
+//                System.out.println(sparkNet.getSparkTrainingStats().statsAsString());
+                sparkNet.getSparkTrainingStats().statsAsString();
 
                 INDArray finalSparkParams = sparkNet.getNetwork().params().dup();
                 //                executioner.addToWatchdog(finalSparkParams, "finalSparkParams");
 
                 float[] fp = finalParams.data().asFloat();
                 float[] fps = finalSparkParams.data().asFloat();
-                System.out.println("Initial (Local) params:       " + Arrays.toString(initialParams.data().asFloat()));
+//                System.out.println("Initial (Local) params:       " + Arrays.toString(initialParams.data().asFloat()));
-                System.out.println("Initial (Spark) params:       "
+//                System.out.println("Initial (Spark) params:       "
-                                + Arrays.toString(initialSparkParams.data().asFloat()));
+//                                + Arrays.toString(initialSparkParams.data().asFloat()));
-                System.out.println("Final (Local) params: " + Arrays.toString(fp));
+//                System.out.println("Final (Local) params: " + Arrays.toString(fp));
-                System.out.println("Final (Spark) params: " + Arrays.toString(fps));
+//                System.out.println("Final (Spark) params: " + Arrays.toString(fps));
 
                 assertEquals(initialParams, initialSparkParams);
                 assertNotEquals(initialParams, finalParams);
@@ -551,14 +554,15 @@ public class TestCompareParameterAveragingSparkVsSingleMachine {
                     sparkNet.fit(rdd);
                 }
 
-                System.out.println(sparkNet.getSparkTrainingStats().statsAsString());
+//                System.out.println(sparkNet.getSparkTrainingStats().statsAsString());
+                sparkNet.getSparkTrainingStats().statsAsString();
 
                 INDArray finalSparkParams = sparkNet.getNetwork().params().dup();
 
-                System.out.println("Initial (Local) params:  " + Arrays.toString(initialParams.data().asFloat()));
+//                System.out.println("Initial (Local) params:  " + Arrays.toString(initialParams.data().asFloat()));
-                System.out.println("Initial (Spark) params:  " + Arrays.toString(initialSparkParams.data().asFloat()));
+//                System.out.println("Initial (Spark) params:  " + Arrays.toString(initialSparkParams.data().asFloat()));
-                System.out.println("Final (Local) params:    " + Arrays.toString(finalParams.data().asFloat()));
+//                System.out.println("Final (Local) params:    " + Arrays.toString(finalParams.data().asFloat()));
-                System.out.println("Final (Spark) params:    " + Arrays.toString(finalSparkParams.data().asFloat()));
+//                System.out.println("Final (Spark) params:    " + Arrays.toString(finalSparkParams.data().asFloat()));
                 assertArrayEquals(initialParams.data().asFloat(), initialSparkParams.data().asFloat(), 1e-8f);
                 assertArrayEquals(finalParams.data().asFloat(), finalSparkParams.data().asFloat(), 1e-6f);
 

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/impl/paramavg/TestJsonYaml.java

@@ -37,7 +37,7 @@ public class TestJsonYaml {
         String json = tm.toJson();
         String yaml = tm.toYaml();
 
-        System.out.println(json);
+//        System.out.println(json);
 
         TrainingMaster fromJson = ParameterAveragingTrainingMaster.fromJson(json);
         TrainingMaster fromYaml = ParameterAveragingTrainingMaster.fromYaml(yaml);

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/impl/paramavg/TestSparkMultiLayerParameterAveraging.java

@@ -389,7 +389,7 @@ public class TestSparkMultiLayerParameterAveraging extends BaseSparkTest {
         List<EventStats> workerFitStats = stats.getValue("ParameterAveragingWorkerFitTimesMs");
         for (EventStats e : workerFitStats) {
             ExampleCountEventStats eces = (ExampleCountEventStats) e;
-            System.out.println(eces.getTotalExampleCount());
+//            System.out.println(eces.getTotalExampleCount());
         }
 
         for (EventStats e : workerFitStats) {
@@ -457,7 +457,8 @@ public class TestSparkMultiLayerParameterAveraging extends BaseSparkTest {
         assertNotEquals(paramsBefore, paramsAfter);
 
         SparkTrainingStats stats = sparkNet.getSparkTrainingStats();
-        System.out.println(stats.statsAsString());
+//        System.out.println(stats.statsAsString());
+        stats.statsAsString();
 
         sparkNet.getTrainingMaster().deleteTempFiles(sc);
     }
@@ -483,7 +484,7 @@ public class TestSparkMultiLayerParameterAveraging extends BaseSparkTest {
             i++;
         }
 
-        System.out.println("Saved to: " + tempDirF.getAbsolutePath());
+//        System.out.println("Saved to: " + tempDirF.getAbsolutePath());
 
 
 
@@ -527,7 +528,8 @@ public class TestSparkMultiLayerParameterAveraging extends BaseSparkTest {
         SparkTrainingStats stats = sparkNet.getSparkTrainingStats();
 
         //Expect
-        System.out.println(stats.statsAsString());
+//        System.out.println(stats.statsAsString());
+        stats.statsAsString();
         assertEquals(numSplits, stats.getValue("ParameterAveragingMasterRepartitionTimesMs").size());
 
         List<EventStats> list = stats.getValue("ParameterAveragingWorkerFitTimesMs");
@@ -566,8 +568,8 @@ public class TestSparkMultiLayerParameterAveraging extends BaseSparkTest {
             i++;
         }
 
-        System.out.println("Saved to: " + tempDirF.getAbsolutePath());
+//        System.out.println("Saved to: " + tempDirF.getAbsolutePath());
-        System.out.println("Saved to: " + tempDirF2.getAbsolutePath());
+//        System.out.println("Saved to: " + tempDirF2.getAbsolutePath());
 
 
 
@@ -610,7 +612,8 @@ public class TestSparkMultiLayerParameterAveraging extends BaseSparkTest {
         assertNotEquals(paramsBefore, paramsAfter);
 
         SparkTrainingStats stats = sparkNet.getSparkTrainingStats();
-        System.out.println(stats.statsAsString());
+//        System.out.println(stats.statsAsString());
+        stats.statsAsString();
 
         //Same thing, buf for MultiDataSet objects:
         config = new Configuration();
@@ -631,7 +634,8 @@ public class TestSparkMultiLayerParameterAveraging extends BaseSparkTest {
         assertNotEquals(paramsBefore, paramsAfter);
 
         stats = sparkNet.getSparkTrainingStats();
-        System.out.println(stats.statsAsString());
+//        System.out.println(stats.statsAsString());
+        stats.statsAsString();
     }
 
 
@@ -730,13 +734,13 @@ public class TestSparkMultiLayerParameterAveraging extends BaseSparkTest {
                         .build();
 
         for (int avgFreq : new int[] {1, 5, 10}) {
-            System.out.println("--- Avg freq " + avgFreq + " ---");
+//            System.out.println("--- Avg freq " + avgFreq + " ---");
             SparkDl4jMultiLayer sparkNet = new SparkDl4jMultiLayer(sc, conf.clone(),
                             new ParameterAveragingTrainingMaster.Builder(numExecutors(), dataSetObjSize)
                                             .batchSizePerWorker(batchSizePerExecutor).averagingFrequency(avgFreq)
                                             .repartionData(Repartition.Always).build());
 
-            sparkNet.setListeners(new ScoreIterationListener(1));
+            sparkNet.setListeners(new ScoreIterationListener(5));
 
 
 
@@ -778,13 +782,13 @@ public class TestSparkMultiLayerParameterAveraging extends BaseSparkTest {
                         .setOutputs("1").build();
 
         for (int avgFreq : new int[] {1, 5, 10}) {
-            System.out.println("--- Avg freq " + avgFreq + " ---");
+//            System.out.println("--- Avg freq " + avgFreq + " ---");
             SparkComputationGraph sparkNet = new SparkComputationGraph(sc, conf.clone(),
                             new ParameterAveragingTrainingMaster.Builder(numExecutors(), dataSetObjSize)
                                             .batchSizePerWorker(batchSizePerExecutor).averagingFrequency(avgFreq)
                                             .repartionData(Repartition.Always).build());
 
-            sparkNet.setListeners(new ScoreIterationListener(1));
+            sparkNet.setListeners(new ScoreIterationListener(5));
 
             JavaRDD<DataSet> rdd = sc.parallelize(list);
 

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/impl/stats/TestTrainingStatsCollection.java

@@ -107,7 +107,7 @@ public class TestTrainingStatsCollection extends BaseSparkTest {
                 expectedStatNames.addAll(c);
             }
 
-            System.out.println(expectedStatNames);
+//            System.out.println(expectedStatNames);
 
 
             SparkTrainingStats stats = sparkNet.getSparkTrainingStats();
@@ -119,7 +119,7 @@ public class TestTrainingStatsCollection extends BaseSparkTest {
             }
 
             String statsAsString = stats.statsAsString();
-            System.out.println(statsAsString);
+//            System.out.println(statsAsString);
             assertEquals(actualKeySet.size(), statsAsString.split("\n").length); //One line per stat
 
 

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/time/TestTimeSource.java

@@ -35,7 +35,7 @@ public class TestTimeSource {
             long systemTime = System.currentTimeMillis();
             long ntpTime = timeSource.currentTimeMillis();
             long offset = ntpTime - systemTime;
-            System.out.println("System: " + systemTime + "\tNTPTimeSource: " + ntpTime + "\tOffset: " + offset);
+//            System.out.println("System: " + systemTime + "\tNTPTimeSource: " + ntpTime + "\tOffset: " + offset);
             Thread.sleep(500);
         }
     }
@@ -49,7 +49,7 @@ public class TestTimeSource {
             long systemTime = System.currentTimeMillis();
             long ntpTime = timeSource.currentTimeMillis();
             long offset = ntpTime - systemTime;
-            System.out.println("System: " + systemTime + "\tSystemClockTimeSource: " + ntpTime + "\tOffset: " + offset);
+//            System.out.println("System: " + systemTime + "\tSystemClockTimeSource: " + ntpTime + "\tOffset: " + offset);
             assertEquals(systemTime, ntpTime, 2); //Should be exact, but we might randomly tick over between one ms and the next
             Thread.sleep(500);
         }

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/ui/TestListeners.java

@@ -87,7 +87,7 @@ public class TestListeners extends BaseSparkTest {
         net.fit(rdd);
 
         List<String> sessions = ss.listSessionIDs();
-        System.out.println("Sessions: " + sessions);
+//        System.out.println("Sessions: " + sessions);
         assertEquals(1, sessions.size());
 
         String sid = sessions.get(0);
@@ -95,15 +95,15 @@ public class TestListeners extends BaseSparkTest {
         List<String> typeIDs = ss.listTypeIDsForSession(sid);
         List<String> workers = ss.listWorkerIDsForSession(sid);
 
-        System.out.println(sid + "\t" + typeIDs + "\t" + workers);
+//        System.out.println(sid + "\t" + typeIDs + "\t" + workers);
 
         List<Persistable> lastUpdates = ss.getLatestUpdateAllWorkers(sid, StatsListener.TYPE_ID);
-        System.out.println(lastUpdates);
+//        System.out.println(lastUpdates);
 
-        System.out.println("Static info:");
+//        System.out.println("Static info:");
         for (String wid : workers) {
             Persistable staticInfo = ss.getStaticInfo(sid, StatsListener.TYPE_ID, wid);
-            System.out.println(sid + "\t" + wid);
+//            System.out.println(sid + "\t" + wid);
         }
 
         assertEquals(1, typeIDs.size());

deeplearning4j/deeplearning4j-scaleout/spark/dl4j-spark/src/test/java/org/deeplearning4j/spark/util/TestRepartitioning.java

@@ -63,7 +63,7 @@ public class TestRepartitioning extends BaseSparkTest {
         assertEquals(10, rdd2.partitions().size());
         for (int i = 0; i < 10; i++) {
             List<String> partition = rdd2.collectPartitions(new int[] {i})[0];
-            System.out.println("Partition " + i + " size: " + partition.size());
+//            System.out.println("Partition " + i + " size: " + partition.size());
             assertEquals(100, partition.size()); //Should be exactly 100, for the util method (but NOT spark .repartition)
         }
     }
@@ -170,7 +170,7 @@ public class TestRepartitioning extends BaseSparkTest {
 
         List<Tuple2<Integer, Integer>> partitionCounts = initial.values().mapPartitionsWithIndex(new CountPartitionsFunction<Integer>(), true).collect();
 
-        System.out.println(partitionCounts);
+//        System.out.println(partitionCounts);
 
         List<Tuple2<Integer,Integer>> initialExpected = Arrays.asList(
                 new Tuple2<>(0,29),
@@ -185,7 +185,7 @@ public class TestRepartitioning extends BaseSparkTest {
 
         JavaRDD<Integer> afterRepartition = SparkUtils.repartitionBalanceIfRequired(initial.values(), Repartition.Always, 2, 112);
         List<Tuple2<Integer, Integer>> partitionCountsAfter = afterRepartition.mapPartitionsWithIndex(new CountPartitionsFunction<Integer>(), true).collect();
-        System.out.println(partitionCountsAfter);
+//        System.out.println(partitionCountsAfter);
 
         for(Tuple2<Integer,Integer> t2 : partitionCountsAfter){
             assertEquals(2, (int)t2._2());
@@ -219,8 +219,8 @@ public class TestRepartitioning extends BaseSparkTest {
             }
         }
 
-        System.out.println("min: " + min + "\t@\t" + minIdx);
+//        System.out.println("min: " + min + "\t@\t" + minIdx);
-        System.out.println("max: " + max + "\t@\t" + maxIdx);
+//        System.out.println("max: " + max + "\t@\t" + maxIdx);
 
         assertEquals(1, min);
         assertEquals(2, max);
@@ -244,7 +244,7 @@ public class TestRepartitioning extends BaseSparkTest {
 
         for (int i = 0; i < 10; i++) {
             List<String> partition = rdd2.collectPartitions(new int[] {i})[0];
-            System.out.println("Partition " + i + " size: " + partition.size());
+//            System.out.println("Partition " + i + " size: " + partition.size());
             assertTrue(partition.size() >= 90 && partition.size() <= 110);
         }
     }

libnd4j/CMakeLists.txt.mkldnn.in

@@ -5,7 +5,7 @@ project(mkldnn-download NONE)
 include(ExternalProject)
 ExternalProject_Add(mkldnn
   GIT_REPOSITORY     https://github.com/intel/mkl-dnn.git
-  GIT_TAG           v1.1.3
+  GIT_TAG           v1.2
   SOURCE_DIR        "${CMAKE_CURRENT_BINARY_DIR}/mkldnn-src"
   BINARY_DIR        "${CMAKE_CURRENT_BINARY_DIR}/mkldnn-build"
   CONFIGURE_COMMAND ""

libnd4j/blas/NDArray.h

@@ -999,14 +999,14 @@ namespace nd4j {
         *  set new order and shape in case of suitable array length (in-place operation)
         *  order - order to set
         *  shape - shape to set
-        *
+        *  copyToNewBuff - if true then old buffer will be copied to new buffer if last one will be allocated after reshaping
         *  if there was permute applied before or there are weird strides, then new buffer is allocated for array
         */
-		bool reshapei(const char order, const std::initializer_list<Nd4jLong>& shape);
+		bool reshapei(const char order, const std::initializer_list<Nd4jLong>& shape, const bool copyToNewBuff = true);
-		bool reshapei(const char order, const std::vector<Nd4jLong>& shape);
+		bool reshapei(const char order, const std::vector<Nd4jLong>& shape, const bool copyToNewBuff = true);
 
-        bool reshapei(const std::initializer_list<Nd4jLong>& shape);
+        bool reshapei(const std::initializer_list<Nd4jLong>& shape, const bool copyToNewBuff = true);
-		bool reshapei(const std::vector<Nd4jLong>& shape);
+		bool reshapei(const std::vector<Nd4jLong>& shape, const bool copyToNewBuff = true);
 
         /**
         *  creates new array with corresponding order and shape, new array will point on _buffer of this array
@@ -1015,8 +1015,8 @@ namespace nd4j {
         *
         * if permute have been applied before or there are weird strides, then new buffer is allocated for new array
         */
-		NDArray reshape(const char order, const std::vector<Nd4jLong>& shape) const &;
+		NDArray reshape(const char order, const std::vector<Nd4jLong>& shape, const bool copyToNewBuff = true) const &;
-        NDArray reshape(const char order, const std::vector<Nd4jLong>& shape) &&;
+        NDArray reshape(const char order, const std::vector<Nd4jLong>& shape, const bool copyToNewBuff = true) &&;
 
         /**
         *  calculate strides and set given order

libnd4j/blas/NDArray.hpp

@@ -501,7 +501,7 @@ NDArray::NDArray(const std::vector<Nd4jLong>& shape, const std::vector<const cha
     auto data = reinterpret_cast<int8_t*>(bufferAsT<int8_t>() + headerLength);
 
     auto func = PRAGMA_THREADS_FOR{
-        for (auto e = start; e < stop; e += increment) {
+        for (auto e = start; e < stop; e++) {
                 auto cdata = data + offsets[e];
                 if (dataType == DataType::UTF16) {
                     unicode::utf8to16(string[e], cdata, std::char_traits<char>::length(string[e]));
@@ -568,7 +568,7 @@ NDArray::NDArray(const std::vector<Nd4jLong>& shape, const std::vector<std::stri
     auto data = reinterpret_cast<int8_t*>(bufferAsT<int8_t>() + headerLength);
 
     auto func = PRAGMA_THREADS_FOR{
-        for (auto e = start; e < stop; e += increment) {
+        for (auto e = start; e < stop; e++) {
              auto cdata = data + offsets[e];
              if (dataType == DataType::UTF16) {
                  unicode::utf8to16(string[e].data(), cdata, string[e].size());
@@ -635,7 +635,7 @@ NDArray::NDArray(const std::vector<Nd4jLong>& shape, const std::vector<std::u16s
     auto data = reinterpret_cast<int8_t*>(bufferAsT<int8_t>() + headerLength);
 
     auto func = PRAGMA_THREADS_FOR{
-        for (auto e = start; e < stop; e += increment) {
+        for (auto e = start; e < stop; e++) {
              auto cdata = data + offsets[e];
              if (dtype == DataType::UTF16) {
                  memcpy(cdata, string[e].data(), string[e].size() * sizeof(uint16_t));
@@ -701,7 +701,7 @@ NDArray::NDArray(const std::vector<Nd4jLong>& shape, const std::vector<const cha
 
 
     auto func = PRAGMA_THREADS_FOR{
-        for (auto e = start; e < stop; e += increment) {
+        for (auto e = start; e < stop; e++) {
              auto cdata = data + offsets[e];
              if (dtype == DataType::UTF16) {
                  memcpy(cdata, string[e], std::char_traits<char16_t>::length(string[e]) * sizeof(uint16_t));
@@ -767,7 +767,7 @@ NDArray::NDArray(const std::vector<Nd4jLong>& shape, const std::vector<std::u32s
     auto data = reinterpret_cast<int8_t*>(bufferAsT<int8_t>() + headerLength);
 
     auto func = PRAGMA_THREADS_FOR{
-        for (auto e = start; e < stop; e += increment) {
+        for (auto e = start; e < stop; e++) {
             auto cdata = data + offsets[e];
             if (dtype == DataType::UTF16) {
                 unicode::utf32to16(string[e].data(), cdata, string[e].size());
@@ -833,7 +833,7 @@ NDArray::NDArray(const std::vector<Nd4jLong>& shape, const std::vector<const cha
     auto data = reinterpret_cast<int8_t*>(bufferAsT<int8_t>() + headerLength);
 
     auto func = PRAGMA_THREADS_FOR{
-        for (auto e = start; e < stop; e += increment) {
+        for (auto e = start; e < stop; e++) {
             auto cdata = data + offsets[e];
             if (dtype == DataType::UTF16) {
                 unicode::utf32to16(string[e], cdata, std::char_traits<char32_t>::length(string[e]));
@@ -1197,8 +1197,8 @@ void NDArray::assign(const NDArray& other, bool allowParallelism) {
             throw std::runtime_error("NDArray::assign: lengths of arrays are mismatched");
         }
 
-        // memcpy is allowed only for same order && same ews (being equal to 1)
+        // memcpy is allowed only for same order c && same ews (being equal to 1)
-        if (ordering() == other.ordering() && dataType() == other.dataType() && ews() == 1 && other.ews() == 1)
+        if (ordering() == other.ordering() && ordering() == 'c' && dataType() == other.dataType() && ews() == 1 && other.ews() == 1)
             copyBuffersContinuouslyFrom(other, other.lengthOf() * other.sizeOfT());
         else {
             NDArray::prepareSpecialUse({this}, {&other});
@@ -1569,20 +1569,25 @@ Nd4jLong NDArray::tensorsAlongDimension(const std::vector<int>& dimensions) cons
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::printShapeInfo(const char * msg) const {
-    //shape::printShapeInfo(_shapeInfo);
+
-    if (msg == nullptr)
-        shape::printShapeInfoLinear(_shapeInfo);
-    else {
     int rank = shape::rank(_shapeInfo);
     int lim = shape::shapeInfoLength(rank);
-        printf("%s: [", msg);
+
-        for (int i = 0; i < shape::shapeInfoLength(rank); i++) {
+    if(msg != nullptr)
-            printf("%lld", (long long) _shapeInfo[i]);
+        printf("shapeInfo %s: [", msg);
-            if (i < lim - 1)
+    else
-                printf(", ");
+        printf("shapeInfo: [");
-        }
+
-        printf("]\n");
+    printf("%i,  ", rank);
+    for (int i = 1; i < shape::shapeInfoLength(rank) - 3; i++){
+        if(i == rank + 1)
+            printf("  ");
+        printf("%lld,", _shapeInfo[i]);
     }
+    printf("  %lld,", shape::type(_shapeInfo));
+    printf("%lld,", shape::elementWiseStride(_shapeInfo));
+    printf("%lld]\n", (Nd4jLong)shape::order(_shapeInfo));
+
     fflush(stdout);
 }
 
@@ -1855,19 +1860,19 @@ void NDArray::updateStrides(const char order) {
 
 //////////////////////////////////////////////////////////////////////////
 // set new order and shape in case of suitable array length
-bool NDArray::reshapei(const char order, const std::initializer_list<Nd4jLong>& shape) {
+bool NDArray::reshapei(const char order, const std::initializer_list<Nd4jLong>& shape, const bool copyToNewBuff) {
     std::vector<Nd4jLong> vShape(shape);
-    return reshapei(order, vShape);
+    return reshapei(order, vShape, copyToNewBuff);
 }
 
 //////////////////////////////////////////////////////////////////////////
-bool NDArray::reshapei(const std::initializer_list<Nd4jLong>& shape) {
+bool NDArray::reshapei(const std::initializer_list<Nd4jLong>& shape, const bool copyToNewBuff) {
-    return reshapei('c', shape);
+    return reshapei(ordering(), shape, copyToNewBuff);
 }
 
 //////////////////////////////////////////////////////////////////////////
-bool NDArray::reshapei(const std::vector<Nd4jLong>& shape) {
+bool NDArray::reshapei(const std::vector<Nd4jLong>& shape, const bool copyToNewBuff) {
-    return reshapei('c', shape);
+    return reshapei(ordering(), shape, copyToNewBuff);
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -1918,18 +1923,18 @@ Nd4jLong NDArray::argMax(std::initializer_list<int> dimensions) {
 
 //////////////////////////////////////////////////////////////////////////
 // create new array with corresponding order and shape, new array will point to the same _buffer as this array
-NDArray NDArray::reshape(const char order, const std::vector<Nd4jLong>& shape) const & {
+NDArray NDArray::reshape(const char order, const std::vector<Nd4jLong>& shape, const bool copyToNewBuff) const & {
 
     NDArray newArr(getDataBuffer(), ShapeDescriptor(getShapeInfo()), getContext(), getBufferOffset());
-    newArr.reshapei(order, shape);
+    newArr.reshapei(order, shape, copyToNewBuff);
 
     return newArr;
 }
 
 //////////////////////////////////////////////////////////////////////////
-NDArray NDArray::reshape(const char order, const std::vector<Nd4jLong>& shape) && {
+NDArray NDArray::reshape(const char order, const std::vector<Nd4jLong>& shape, const bool copyToNewBuff) && {
 
-    this->reshapei(order, shape);
+    this->reshapei(order, shape, copyToNewBuff);
     return std::move(*this);
 }
 
@@ -1971,7 +1976,7 @@ bool NDArray::permutei(const std::initializer_list<int>& dimensions) {
 
 //////////////////////////////////////////////////////////////////////////
 bool NDArray::permutei(const std::vector<int>& dimensions) {
-    return permutei(dimensions.data(), dimensions.size());
+    return permutei(dimensions.data(), rankOf());
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -1993,7 +1998,7 @@ bool NDArray::permutei(const std::vector<Nd4jLong>& dimensions) {
     for (int e = 0; e < dimensions.size(); e++)
         ivec[e] = dimensions[e];
 
-    return permutei(ivec.data(), ivec.size());
+    return permutei(ivec.data(), rankOf());
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -2029,9 +2034,8 @@ NDArray NDArray::permute(const Nd4jLong* dimensions, const int rank) && {
 
 //////////////////////////////////////////////////////////////////////////
 NDArray NDArray::permute(const std::vector<int>& dimensions) const &{
-    auto data = dimensions.data();
+
-    auto size = dimensions.size();
+    return permute(dimensions.data(), rankOf());
-    return permute(data, size);
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -2043,7 +2047,8 @@ NDArray NDArray::permute(const std::vector<int>& dimensions) && {
 
 //////////////////////////////////////////////////////////////////////////
 NDArray NDArray::permute(const std::vector<Nd4jLong>& dimensions) const & {
-    return permute(dimensions.data(), dimensions.size());
+
+    return permute(dimensions.data(), rankOf());
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -2106,12 +2111,12 @@ void NDArray::permute(const Nd4jLong *dimensions, const int rank, NDArray& targe
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::permute(const std::vector<int>& dimensions, NDArray& target) const {
-    permute(dimensions.data(), dimensions.size(), target);
+    permute(dimensions.data(), rankOf(), target);
 }
 
 //////////////////////////////////////////////////////////////////////////
 void NDArray::permute(const std::vector<Nd4jLong>& dimensions, NDArray& target) const {
-    permute(dimensions.data(), dimensions.size(), target);
+    permute(dimensions.data(), rankOf(), target);
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -2362,7 +2367,7 @@ NDArray NDArray::asS() const {
     const auto inData = bufferAsT<int8_t>() + offsetsLength;
 
     auto func = PRAGMA_THREADS_FOR{
-        for (int e = start; e < stop; e += increment) {
+        for (int e = start; e < stop; e++) {
            auto cdata = outData + offsets[e];
            auto end = nInputoffsets[e + 1];
            auto idata = inData + nInputoffsets[e];
@@ -3221,7 +3226,7 @@ BUILD_SINGLE_TEMPLATE(template ND4J_EXPORT std::vector, NDArray::asVectorT(), LI
 
 //////////////////////////////////////////////////////////////////////////
 // set new order and shape in case of suitable array length
-bool NDArray::reshapei(const char order, const std::vector<Nd4jLong>& cshape) {
+bool NDArray::reshapei(const char order, const std::vector<Nd4jLong>& cshape, const bool copyToNewBuff) {
 
     // check firstly whether cshape is identical to shape of array, if yes then reshape is unnecessary
     if(order == ordering() && shape::shapeEquals(rankOf(), shapeOf(), cshape.size(), cshape.data()))
@@ -3293,18 +3298,14 @@ bool NDArray::reshapei(const char order, const std::vector<Nd4jLong>& cshape) {
     Nd4jLong *shapeInfoNew;
     ALLOCATE(shapeInfoNew, getContext()->getWorkspace(), shape::shapeInfoLength(rank), Nd4jLong);
 
-    bool canReshape = shape::reshapeC(rankOf(), shapeInfo(), shape.size(), shape.data(), shapeInfoNew);
+    bool canReshape = shape::reshapeC(shapeInfo(), order, shape.size(), shape.data(), shapeInfoNew);
 
-    // we can do this only if there was no permute applied, or there are no weird strides
     if (canReshape) {
-        if(ordering() == 'c' && order == 'f')
-            throw std::invalid_argument("NDArray::reshapei(order, shape): in case of reshapeC it doesn't make sense to reshape from c order to f order !");
-
-        shape::setEws(shapeInfoNew, arrLength);
         setShapeInfo(shapeInfoNew);
     }
     else {
         NDArray temp(order, shape, dataType(), getContext());
+        if(copyToNewBuff)
             this->applyTransform(transform::Assign, temp, nullptr);
         *this = std::move(temp);
     }
@@ -3465,7 +3466,7 @@ NDArray NDArray::dup(const char newOrder) const {
             std::vector<std::string> strings(lengthOf());
 
             auto func = PRAGMA_THREADS_FOR{
-                    for (auto i = start; i < stop; i += increment) {
+                    for (auto i = start; i < stop; i++) {
                            strings[i] = std::move(this->e<std::string>(i));
                     }
             };
@@ -3478,7 +3479,7 @@ NDArray NDArray::dup(const char newOrder) const {
             std::vector<std::u16string> strings(lengthOf());
 
             auto func = PRAGMA_THREADS_FOR{
-                    for (auto i = start; i < stop; i += increment) {
+                    for (auto i = start; i < stop; i++) {
                            strings[i] = std::move(this->e<std::u16string>(i));
                     }
             };
@@ -3490,7 +3491,7 @@ NDArray NDArray::dup(const char newOrder) const {
 
         std::vector<std::u32string> strings(lengthOf());
         auto func = PRAGMA_THREADS_FOR{
-               for (auto i = start; i < stop; i += increment) {
+               for (auto i = start; i < stop; i++) {
                       strings[i] = std::move(this->e<std::u32string>(i));
                }
         };
@@ -4846,7 +4847,7 @@ NDArray NDArray::operator()(const std::vector<Nd4jLong>& idx, const bool keepUni
     auto shapeOf = shape::shapeOf(newShapeInfo);
     auto stridesOf = shape::stride(newShapeInfo);
 
-    Nd4jLong offset(0), subArrLen(1);
+    Nd4jLong offset = 0;
     int n(isStrided ? 3 : 2), first, last, stride;
 
     for (int d = rank - 1; d >= 0; --d) {
@@ -4863,29 +4864,31 @@ NDArray NDArray::operator()(const std::vector<Nd4jLong>& idx, const bool keepUni
             if(shapeOf[d] != 1)
                 stridesOf[d] *= stride;
         }
+    }
 
-        subArrLen *= shapeOf[d];
+    Nd4jLong *newShapeInfo2 = newShapeInfo;
+
+    if(!keepUnitiesInShape) {
+
+        std::vector<int> dimsWithUnities;
+
+        for (uint d = 0; d < rank; ++d)
+            if(idx[n*d] != idx[n*d+1] && shapeOf[d] == 1)
+                dimsWithUnities.push_back(d);
+
+        if(!dimsWithUnities.empty())
+            newShapeInfo2 = ShapeBuilders::copyShapeInfoWithoutUnites(newShapeInfo, dimsWithUnities.size(), dimsWithUnities.data(), getContext()->getWorkspace());
     }
 
     // check if there is possibility to set ews = 1
-    shape::setEws(newShapeInfo, subArrLen);
+    shape::checkStridesEwsAndOrder(newShapeInfo2);
 
-    NDArray result(_buffer, ShapeDescriptor(newShapeInfo), getContext(), offset + getBufferOffset());
+    NDArray result(_buffer, ShapeDescriptor(newShapeInfo2), getContext(), offset + getBufferOffset());
     result._isView = true;
 
-    if(!keepUnitiesInShape) {
-        const int coeff = isStrided ? 3 : 2;
-        std::vector<Nd4jLong> nonUnitDims;
-
-        for (int d = 0; d < rank; ++d)
-            if(!(idx[coeff*d] != idx[coeff*d+1] && newShapeInfo[d+1] == 1))
-                nonUnitDims.push_back(newShapeInfo[d+1]);
-
-        if(nonUnitDims.size() != rank)
-            result.reshapei(nonUnitDims);
-    }
-
     RELEASE(newShapeInfo, getContext()->getWorkspace());
+    if(newShapeInfo != newShapeInfo2)
+        RELEASE(newShapeInfo2, getContext()->getWorkspace());
 
     return result;
 }

libnd4j/blas/cpu/GraphExecutioner.cpp

@@ -179,7 +179,7 @@ namespace graph {
         nd4j_debug("Embedded graph execution finished. %i variable(s) migrated\n", cnt);
 
     } else if (node->hasCustomOp()) {
-        // if we have something to execute - lets just execute it.
+        // now, if we have something to execute - lets just execute it.
         auto status = node->getCustomOp()->execute(&context);
         if (status != ND4J_STATUS_OK)
             return status;
@@ -494,8 +494,10 @@ Nd4jStatus GraphExecutioner::execute(Graph *graph, VariableSpace* variableSpace)
         nd4j::memory::MemoryRegistrator::getInstance()->setGraphMemoryFootprintIfGreater(h, m);
     }
 
-    if (tempFlow)
+    if (tempFlow) {
         delete flowPath;
+        __variableSpace->setFlowPath(nullptr);
+    }
 
     return Status::OK();
 }

libnd4j/blas/cpu/NDArray.cpp

@@ -98,7 +98,7 @@ void NDArray::fillAsTriangular(const float val, int lower, int upper, NDArray& t
 
     auto func = PRAGMA_THREADS_FOR {
         Nd4jLong coords[MAX_RANK];
-        for (auto i = start; i < stop; i += increment) {
+        for (auto i = start; i < stop; i++) {
             shape::index2coords(i, target.getShapeInfo(), coords);
             const auto zOffset = shape::getOffset(target.getShapeInfo(), coords);
 
@@ -152,7 +152,7 @@ static void templatedSwap(void *xBuffer, void *yBuffer, Nd4jLong length) {
     auto y = reinterpret_cast<T *>(yBuffer);
 
     auto func = PRAGMA_THREADS_FOR {
-        for (auto i = start; i < stop; i += increment) {
+        for (auto i = start; i < stop; i++) {
             auto temp = x[i];
             x[i] = y[i];
             y[i] = temp;
@@ -266,7 +266,7 @@ NDArray NDArray::tile(const std::vector<Nd4jLong>& reps) const {
     if(result.ordering() == 'c') {           //  ews == 1 always here
 
         auto func = PRAGMA_THREADS_FOR {
-            for (auto i = start; i < stop; i += increment) {
+            for (auto i = start; i < stop; i++) {
                 auto yOffset = shape::subArrayOffset(i, newShapeInfo, getShapeInfo());
                 BUILD_SINGLE_SELECTOR(xType, this->template templatedAssign,(result.getBuffer(), i, this->getBuffer(), yOffset), LIBND4J_TYPES);
             }
@@ -277,7 +277,7 @@ NDArray NDArray::tile(const std::vector<Nd4jLong>& reps) const {
     else {
 
         auto func = PRAGMA_THREADS_FOR {
-            for (auto i = start; i < stop; i += increment) {
+            for (auto i = start; i < stop; i++) {
                 auto xOffset = result.getOffset(i);
                 auto yOffset = shape::subArrayOffset(i, newShapeInfo, getShapeInfo());
                 BUILD_SINGLE_SELECTOR(xType, this->template templatedAssign,(result.getBuffer(), xOffset, this->getBuffer(), yOffset), LIBND4J_TYPES);
@@ -377,7 +377,7 @@ static void repeat_(const NDArray& input, NDArray& output, const std::vector<int
     // loop through input array
     auto func = PRAGMA_THREADS_FOR {
         Nd4jLong coords[MAX_RANK];
-        for (auto i = start; i < stop; i += increment) {
+        for (auto i = start; i < stop; i++) {
             shape::index2coords(i, output.getShapeInfo(), coords);
 
             const auto zOffset = shape::getOffset(output.getShapeInfo(), coords);

libnd4j/blas/cpu/NDArrayLambda.hpp

@@ -22,7 +22,7 @@ void NDArray::applyTriplewiseLambda(NDArray& second, NDArray& third, const std::
     if (this->ordering() == second.ordering() && this->ordering() == third.ordering()  && this->ordering() == target.ordering() && (this->ews() == 1 && target.ews() == 1) && this->ews() == second.ews() && this->ews() == third.ews()) {
 
         auto loop = PRAGMA_THREADS_FOR {
-            for (auto e = start; e < stop; e += increment)
+            for (auto e = start; e < stop; e++)
                 z[e] = func(f[e], s[e], t[e]);
         };
 
@@ -31,7 +31,7 @@ void NDArray::applyTriplewiseLambda(NDArray& second, NDArray& third, const std::
         if (f == z) {
 
             auto loop = PRAGMA_THREADS_FOR {
-                for (auto e = start; e < stop; e += increment) {
+                for (auto e = start; e < stop; e++) {
                     auto tOffset = this->getOffset(e);
                     auto uOffset = second.getOffset(e);
                     auto vOffset = third.getOffset(e);
@@ -44,7 +44,7 @@ void NDArray::applyTriplewiseLambda(NDArray& second, NDArray& third, const std::
         } else {
 
             auto loop = PRAGMA_THREADS_FOR {
-                for (auto e = start; e < stop; e += increment) {
+                for (auto e = start; e < stop; e++) {
                     auto tOffset = this->getOffset(e);
                     auto uOffset = second.getOffset(e);
                     auto vOffset = third.getOffset(e);
@@ -93,7 +93,7 @@ void NDArray::applyPairwiseLambda(const NDArray& other, const std::function<T(T,
     if (this->ordering() == other.ordering() && this->ordering() == target.ordering() && (this->ews() == 1 && target.ews() == 1) && this->ews() == other.ews()) {
 
         auto loop = PRAGMA_THREADS_FOR {
-            for (auto e = start; e < stop; e += increment)
+            for (auto e = start; e < stop; e++)
                 z[e] = func(f[e], s[e]);
         };
 
@@ -102,7 +102,7 @@ void NDArray::applyPairwiseLambda(const NDArray& other, const std::function<T(T,
         if (f == z) {
 
             auto loop = PRAGMA_THREADS_FOR {
-                for (auto e = start; e < stop; e += increment) {
+                for (auto e = start; e < stop; e++) {
                     auto xOffset = this->getOffset(e);
                     auto yOffset = other.getOffset(e);
 
@@ -114,7 +114,7 @@ void NDArray::applyPairwiseLambda(const NDArray& other, const std::function<T(T,
         } else {
 
             auto loop = PRAGMA_THREADS_FOR {
-                for (auto e = start; e < stop; e += increment) {
+                for (auto e = start; e < stop; e++) {
                     auto xOffset = this->getOffset(e);
                     auto yOffset = other.getOffset(e);
                     auto zOffset = target.getOffset(e);
@@ -156,7 +156,7 @@ void NDArray::applyLambda(const std::function<T(T)>& func, NDArray& target) {
     if (this->ordering() == target.ordering() && (this->ews() == 1 && target.ews() == 1)) {
 
         auto loop = PRAGMA_THREADS_FOR {
-            for (auto e = start; e < stop; e += increment)
+            for (auto e = start; e < stop; e++)
                 z[e] = func(f[e]);
         };
 
@@ -165,7 +165,7 @@ void NDArray::applyLambda(const std::function<T(T)>& func, NDArray& target) {
         if (f == z) {
 
             auto loop = PRAGMA_THREADS_FOR {
-                for (auto e = start; e < stop; e += increment) {
+                for (auto e = start; e < stop; e++) {
                     auto xOffset = this->getOffset(e);
 
                     f[xOffset] = func(f[xOffset]);
@@ -176,7 +176,7 @@ void NDArray::applyLambda(const std::function<T(T)>& func, NDArray& target) {
         } else {
 
             auto loop = PRAGMA_THREADS_FOR {
-                for (auto e = start; e < stop; e += increment) {
+                for (auto e = start; e < stop; e++) {
                     auto xOffset = this->getOffset(e);
                     auto zOffset = target.getOffset(e);
 
@@ -217,7 +217,7 @@ void NDArray::applyIndexedLambda(const std::function<T(Nd4jLong, T)>& func, NDAr
     if (this->ordering() == target.ordering() && (this->ews() == 1 && target.ews() == 1)) {
 
         auto loop = PRAGMA_THREADS_FOR {
-            for (auto e = start; e < stop; e += increment)
+            for (auto e = start; e < stop; e++)
                 z[e] = func(e, f[e]);
         };
 
@@ -226,7 +226,7 @@ void NDArray::applyIndexedLambda(const std::function<T(Nd4jLong, T)>& func, NDAr
         if (f == z) {
 
             auto loop = PRAGMA_THREADS_FOR {
-                for (auto e = start; e < stop; e += increment) {
+                for (auto e = start; e < stop; e++) {
                     auto xOffset = this->getOffset(e);
 
                     f[xOffset] = func(e, f[xOffset]);
@@ -237,7 +237,7 @@ void NDArray::applyIndexedLambda(const std::function<T(Nd4jLong, T)>& func, NDAr
         } else {
 
             auto loop = PRAGMA_THREADS_FOR {
-                for (auto e = start; e < stop; e += increment) {
+                for (auto e = start; e < stop; e++) {
                     auto xOffset = this->getOffset(e);
                     auto zOffset = target.getOffset(e);
 
@@ -283,7 +283,7 @@ void NDArray::applyIndexedPairwiseLambda(NDArray& other, const std::function<T(N
     if (this->ordering() == other.ordering() && this->ordering() == target.ordering() && (this->ews() == 1 && target.ews() == 1) && this->ews() == other.ews()) {
 
         auto loop = PRAGMA_THREADS_FOR {
-            for (auto e = start; e < stop; e += increment)
+            for (auto e = start; e < stop; e++)
                 z[e] = func((Nd4jLong) e, f[e], s[e]);
         };
 
@@ -292,7 +292,7 @@ void NDArray::applyIndexedPairwiseLambda(NDArray& other, const std::function<T(N
         if (f == z) {
 
             auto loop = PRAGMA_THREADS_FOR {
-                for (auto e = start; e < stop; e += increment) {
+                for (auto e = start; e < stop; e++) {
                     auto xOffset = this->getOffset(e);
                     auto yOffset = other.getOffset(e);
 
@@ -304,7 +304,7 @@ void NDArray::applyIndexedPairwiseLambda(NDArray& other, const std::function<T(N
         } else {
 
             auto loop = PRAGMA_THREADS_FOR {
-                for (auto e = start; e < stop; e += increment) {
+                for (auto e = start; e < stop; e++) {
                     auto xOffset = this->getOffset(e);
                     auto yOffset = other.getOffset(e);
                     auto zOffset = target.getOffset(e);

libnd4j/blas/cpu/NativeOpExecutioner.cpp

@@ -163,15 +163,44 @@ void NativeOpExecutioner::execBroadcast(nd4j::LaunchContext  *lc,
     BUILD_PAIRWISE_SELECTOR(xType, yType, zType, functions::broadcast::Broadcast, ::exec(opNum, hX, hXShapeInfo, hY, hYShapeInfo, hZ, hZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ), LIBND4J_TYPES, LIBND4J_TYPES);
 #else
 
+    auto loopKind = nd4j::LoopKind::deduceKindOfLoopBroadcast(hXShapeInfo, hYShapeInfo, hZShapeInfo);
+
     auto func = PRAGMA_THREADS_FOR {
-        BUILD_SINGLE_SELECTOR_THRICE(xType, functions::broadcast::Broadcast, ::exec(opNum, hX, hXShapeInfo, hY, hYShapeInfo, hZ, hZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ, start, stop), LIBND4J_TYPES);
+        BUILD_SINGLE_SELECTOR_THRICE(xType, functions::broadcast::Broadcast, ::exec(opNum, hX, hXShapeInfo, hY, hYShapeInfo, hZ, hZShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ, loopKind, start, stop), LIBND4J_TYPES);
     };
 
+    Nd4jLong numTads = 0;
+
+    switch (loopKind) {
+        case nd4j::LoopKind::BROADCAST_SCALAR_X: {
+                numTads = shape::length(hXShapeInfo);
+            }
+            break;
+        case nd4j::LoopKind::BROADCAST_SCALAR_Y: {
+                numTads = shape::length(hYShapeInfo);
+            }
+            break;
+        case nd4j::LoopKind::BROADCAST_3D: {
+            numTads = shape::sizeAt(hZShapeInfo, 0);
+            }
+            break;
+        case nd4j::LoopKind::BROADCAST_4D: {
+            numTads = shape::sizeAt(hZShapeInfo, 0) * shape::sizeAt(hZShapeInfo, 1);
+            }
+            break;
+        case nd4j::LoopKind::BROADCAST_5D: {
+            numTads = shape::sizeAt(hZShapeInfo, 0) * shape::sizeAt(hZShapeInfo, 1);
+            }
+            break;
+        default: {
             auto xLen = shape::length(hXShapeInfo);
             auto yLen = shape::length(hYShapeInfo);
-    auto numTads = xLen / yLen;
+            numTads = xLen / yLen;
+        }
+    }
 
     samediff::Threads::parallel_tad(func, 0, numTads);
+
 #endif
 }
 

libnd4j/blas/cpu/NativeOps.cpp

@@ -1291,7 +1291,7 @@ void pullRowsGeneric(void *vx,
     _threads = nd4j::math::nd4j_min<int>(_threads, nd4j::Environment::getInstance()->maxThreads());
 
     auto func = PRAGMA_THREADS_FOR {
-        for (auto idx = start; idx < stop; idx += increment) {
+        for (auto idx = start; idx < stop; idx++) {
             auto xTadOffsetForBlock = tadOffsets[indexes[idx]];
             auto zTadOffsetForBlock = zTadOffsets[idx];
 
@@ -1356,7 +1356,7 @@ void tearGeneric(void *vx,
     auto numTads = shape::length(hXShapeInfo) / tadLength;
 
     auto func = PRAGMA_THREADS_FOR {
-        for (auto i = start; i < stop; i += increment) {
+        for (auto i = start; i < stop; i++) {
             auto hZ = reinterpret_cast<T *>(targets[i]);
             auto s = hX + tadOffsets[i];
 
@@ -1478,7 +1478,7 @@ void shuffleGeneric(void **hX, Nd4jLong **hXShapeInfo, void **dz, Nd4jLong **hZS
     auto dZ = reinterpret_cast<T **>(dz);
 
     auto func = PRAGMA_THREADS_FOR {
-        for (auto f = start; f < stop; f += increment) {
+        for (auto f = start; f < stop; f++) {
             auto hX = reinterpret_cast<T *>(dX[f]);
             //auto hZ = reinterpret_cast<T *>(dZ[f]);
 

libnd4j/include/array/DataTypeConversions.h

@@ -52,7 +52,7 @@ namespace nd4j {
                                 TypeCast::convertGeneric<T2, T>(nullptr, tmp, length, buffer);
 #else
                 auto func = PRAGMA_THREADS_FOR {
-                    for (auto e = start; e < stop; e += increment)
+                    for (auto e = start; e < stop; e++)
                         buffer[e] = canKeep ? static_cast<T>(tmp[e]) : BitwiseUtils::swap_bytes<T>(static_cast<T>(tmp[e]));
                 };
 
@@ -110,7 +110,7 @@ namespace nd4j {
                                 TypeCast::convertGeneric<float, T>(nullptr, tmp, length, buffer);
 #else
                             auto func = PRAGMA_THREADS_FOR {
-                                for (auto e = start; e < stop; e += increment)
+                                for (auto e = start; e < stop; e++)
                                     buffer[e] = canKeep ? static_cast<T>(tmp[e]) : BitwiseUtils::swap_bytes<T>(static_cast<T>(tmp[e]));
                             };
 
@@ -138,7 +138,7 @@ namespace nd4j {
 
 #else
                             auto func = PRAGMA_THREADS_FOR {
-                                for (auto e = start; e < stop; e += increment)
+                                for (auto e = start; e < stop; e++)
                                     buffer[e] = canKeep ? static_cast<T>(tmp[e]) : BitwiseUtils::swap_bytes<T>(static_cast<T>(tmp[e]));
                             };
 
@@ -164,7 +164,7 @@ namespace nd4j {
                                 TypeCast::convertGeneric<float16, T>(nullptr, tmp, length, buffer);
 #else
                             auto func = PRAGMA_THREADS_FOR {
-                                for (auto e = start; e < stop; e += increment)
+                                for (auto e = start; e < stop; e++)
                                     buffer[e] = canKeep ? static_cast<T>(tmp[e]) : BitwiseUtils::swap_bytes<T>(static_cast<T>(tmp[e]));
                             };
 

libnd4j/include/graph/VariableProxy.h

@@ -58,6 +58,7 @@ namespace nd4j {
             virtual void putVariable(int id, Variable *variable);
             virtual void putVariable(int id, NDArray *array);
             virtual void putVariable(int id, int idx, NDArray *array);
+            virtual void putVariable(int id, int idx, NDArray &array);
             virtual void putVariable(int id, int idx, Variable *array);
 
             virtual void replaceVariable(Variable *variable);

libnd4j/include/graph/VariableSpace.h

@@ -100,6 +100,7 @@ namespace nd4j {
             virtual void putVariable(int id, Variable *variable);
             virtual void putVariable(int id, NDArray *array);
             virtual void putVariable(int id, int idx, NDArray *array);
+            virtual void putVariable(int id, int idx, NDArray &array);
             virtual void putVariable(int id, int idx, Variable *array);
 
             virtual void dropVariable(std::pair<int,int> &pair);

libnd4j/include/graph/impl/Graph.cpp

@@ -1088,8 +1088,23 @@ namespace nd4j {
                 if (e < node->input()->size() - 1)
                     nd4j_printf(", ", "");
             }
+
+            if (node->opType() == OpType_CUSTOM) {
+                auto ctx = node->protoContext();
+                if (ctx->getIArguments()->size() > 0) {
+                    printf("]; iArgs: [");
+
+                    for (int e = 0; e < ctx->getIArguments()->size(); e++) {
+                        printf("%i", ctx->getIArguments()->at(e));
+                        if (e < ctx->getIArguments()->size() - 1)
+                            nd4j_printf(", ", "");
+                    }
+                }
+            }
+
             nd4j_printf("]; \n", "");
 
+
 //            printf("\n");
             fflush(stdout);
         }

libnd4j/include/graph/impl/Variable.cpp

@@ -60,8 +60,11 @@ namespace nd4j {
             result->_name = this->_name;
             result->_index = this->_index;
 
-            if (this->_ndarray != nullptr)
+            if (this->_ndarray != nullptr) {
                 result->_ndarray = new NDArray(this->_ndarray->dup(this->_ndarray->ordering()));
+                result->_readOnly = false;
+                result->_removable = true;
+            }
 
             if (this->_list != nullptr)
                 result->_list = this->_list->clone();

libnd4j/include/graph/impl/VariableProxy.cpp

@@ -191,6 +191,9 @@ namespace nd4j {
             _current->putVariable(id, array);
         }
 
+        void nd4j::graph::VariableProxy::putVariable(int id, int idx, NDArray &array) {
+            _current->putVariable(id, idx, array);
+        }
         
         void VariableProxy::putVariable(int id, int idx, NDArray *array) {
             _current->putVariable(id, idx, array);

libnd4j/include/graph/impl/VariableSpace.cpp

@@ -263,18 +263,18 @@ namespace nd4j {
         void nd4j::graph::VariableSpace::putVariable(int id, Variable *variable) {
             // we don't want to add variables more then once
             if (_variables.count(id) > 0 || _temporary.count(id) > 0) {
-                // nd4j_verbose("Trying to update variable for node_%i\n", id);
-
                 auto local = id < 0 ? _variables.at(id) : _temporary.at(id);
 
                 if (!local->hasNDArray() && variable->hasNDArray()) {
-                    // nd4j_verbose("Saving variable for node_%i\n", id);
                     local->setNDArray(variable->getNDArray());
-                }
+
-                return;
+                    // we're inheriting this from Variable
+                    local->markReadOnly(variable->isReadOnly());
+                    local->markRemovable(variable->isRemovable());
                 }
 
-            //nd4j_debug("Adding Variable to Space: id: %i; Array is null: %i;\n", id, variable->getNDArray() == nullptr);
+                return;
+            }
 
             _varmap.lock();
 
@@ -314,6 +314,21 @@ namespace nd4j {
             }
         }
 
+        void nd4j::graph::VariableSpace::putVariable(int id, int idx, NDArray &array) {
+            auto *var = new nd4j::graph::Variable(&array, "", id, idx);
+            var->markRemovable(false);
+            var->markReadOnly(true);
+
+            // let's see if this op needs
+            bool d = this->hasVariable(id, idx);
+
+            this->putVariable(id, var);
+
+            // if var for this nodeid already exists - we'll just delete variable
+            if (d)
+                delete var;
+        }
+
         void nd4j::graph::VariableSpace::putVariable(int id, NDArray *array) {
             auto *var = new nd4j::graph::Variable(array);
             this->putVariable(id, var);

libnd4j/include/graph/profiling/NodeProfile.h

@@ -24,6 +24,7 @@
 #include <pointercast.h>
 #include <dll.h>
 #include <string>
+#include <vector>
 
 namespace nd4j {
     namespace graph {
@@ -65,6 +66,9 @@ namespace nd4j {
 
             // total amount of memory used during execution
             Nd4jLong _memoryTotal = 0L;
+
+            std::vector<std::string> _inputShapes;
+            std::vector<std::string> _outputShapes;
         public:
             NodeProfile() = default;
             ~NodeProfile() = default;
@@ -84,10 +88,15 @@ namespace nd4j {
             void setObjectsSize(Nd4jLong bytes);
             void setTotalSize(Nd4jLong bytes);
 
-            Nd4jLong getActivationsSize();
+            void addInputShape(Nd4jLong *shapeInfo);
-            Nd4jLong getTemporarySize();
+            void addOutputShape(Nd4jLong *shapeInfo);
-            Nd4jLong getObjectsSize();
+
-            Nd4jLong getTotalSize();
+            Nd4jLong getActivationsSize() const;
+            Nd4jLong getTemporarySize() const;
+            Nd4jLong getObjectsSize() const;
+            Nd4jLong getTotalSize() const;
+
+            Nd4jLong getExecutionTime() const;
 
             std::string& name();
 

libnd4j/include/graph/profiling/impl/GraphProfile.cpp

@@ -21,6 +21,8 @@
 #include <graph/profiling/GraphProfile.h>
 #include <helpers/logger.h>
 #include <chrono>
+#include <templatemath.h>
+#include <algorithm>
 
 namespace nd4j {
     namespace graph {
@@ -184,8 +186,25 @@ namespace nd4j {
             if (_profiles.empty())
                 nd4j_printf("No nodes in graph\n","");
 
-            for (auto v: _profiles)
+            // printint out stuff
+            std::vector<NodeProfile*> sorted;
+            for (auto v: _profiles) {
                 v->printOut();
+                sorted.emplace_back(v);
+            }
+
+            if (_profiles.size() > 1) {
+                // building hot spots
+                std::sort(sorted.begin(), sorted.end(), [](const NodeProfile *a, const NodeProfile *b) -> bool {
+                    return a->getExecutionTime() > b->getExecutionTime();
+                });
+
+                nd4j_printf("\nTop 30 reports by EXEC:\n", "");
+                auto limit = nd4j::math::nd4j_min<int>(30, sorted.size());
+                for (int e = 0; e < limit; e++) {
+                    sorted[e]->printOut();
+                }
+            }
 
             nd4j_printf("\nSpecial timers:\n", "");
             if (_timings.empty())

libnd4j/include/graph/profiling/impl/GraphProfilingHelper.cpp

@@ -32,7 +32,7 @@ namespace nd4j {
             // graph->printOut();
 
             // warm up
-            for (int e = 0; e < 1000; e++) {
+            for (int e = 0; e < iterations; e++) {
                 FlowPath fp;
 
                 auto _vs = varSpace->clone();

libnd4j/include/graph/profiling/impl/NodeProfile.cpp

@@ -20,6 +20,7 @@
 
 #include <helpers/logger.h>
 #include <graph/profiling/NodeProfile.h>
+#include <helpers/ShapeUtils.h>
 
 namespace nd4j {
     namespace graph {
@@ -35,9 +36,23 @@ namespace nd4j {
             nd4j_printf("      Memory: ACT: %lld; TMP: %lld; OBJ: %lld; TTL: %lld;\n", _memoryActivations / _merges, _memoryTemporary / _merges, _memoryObjects / _merges, _memoryTotal / _merges);
             nd4j_printf("      Time: PREP: %lld ns; EXEC: %lld ns; TTL: %lld ns;\n", _preparationTime / _merges, _executionTime / _merges, _totalTime / _merges);
             nd4j_printf("      PREP: INPUT: %lld ns; SHAPE: %lld ns; ARRAY: %lld ns;\n", _inputTime / _merges, _shapeTime / _merges, _arrayTime / _merges);
+
+            std::string inputs;
+            std::string outputs;
+
+            int cnt = 0;
+            for (const auto &v: _inputShapes)
+                inputs += v + "    ";
+
+            for (const auto &v: _outputShapes)
+                outputs += v + "    ";
+
+
+            nd4j_printf("      Inputs: %s\n", inputs.c_str());
+            nd4j_printf("      Outputs: %s\n", outputs.c_str());
         };
 
-        Nd4jLong NodeProfile::getActivationsSize() {
+        Nd4jLong NodeProfile::getActivationsSize() const {
             return _memoryActivations;
         }
 
@@ -53,15 +68,15 @@ namespace nd4j {
             _inputTime = time;
         }
 
-        Nd4jLong NodeProfile::getTemporarySize() {
+        Nd4jLong NodeProfile::getTemporarySize()  const{
             return _memoryTemporary;
         }
             
-        Nd4jLong NodeProfile::getObjectsSize() {
+        Nd4jLong NodeProfile::getObjectsSize()  const{
             return _memoryObjects;
         }
 
-        Nd4jLong NodeProfile::getTotalSize() {
+        Nd4jLong NodeProfile::getTotalSize()  const{
             return _memoryTotal;
         }
 
@@ -97,6 +112,18 @@ namespace nd4j {
             _memoryTotal = bytes;
         }
 
+        Nd4jLong NodeProfile::getExecutionTime() const {
+            return _executionTime;
+        }
+
+        void NodeProfile::addInputShape(Nd4jLong *shapeInfo) {
+            _inputShapes.emplace_back(ShapeUtils::shapeInfoAsString(shapeInfo));
+        }
+
+        void NodeProfile::addOutputShape(Nd4jLong *shapeInfo) {
+            _outputShapes.emplace_back(ShapeUtils::shapeInfoAsString(shapeInfo));
+        }
+
         void NodeProfile::merge(NodeProfile *other) {
             _merges += other->_merges;
             _memoryObjects += other->_memoryObjects;
@@ -110,6 +137,9 @@ namespace nd4j {
             _shapeTime += other->_shapeTime;
             _arrayTime += other->_arrayTime;
             _inputTime += other->_inputTime;
+
+            _inputShapes = other->_inputShapes;
+            _outputShapes = other->_outputShapes;
         }
 
         std::string& NodeProfile::name() {
@@ -129,6 +159,9 @@ namespace nd4j {
             _shapeTime = other->_shapeTime;
             _arrayTime = other->_arrayTime;
             _inputTime = other->_inputTime;
+
+            _inputShapes = other->_inputShapes;
+            _outputShapes = other->_outputShapes;
         }
     }
 }

libnd4j/include/helpers/LoopKind.h

@@ -37,12 +37,13 @@ namespace nd4j {
 class ND4J_EXPORT LoopKind {
     
     public:
-        enum Kind {SMALLARR2DX, EWS1, EWSNONZERO, RANK1, RANK2, RANK3, RANK4, RANK5, X_EWSNONZERO, Y_EWSNONZERO, Z_EWSNONZERO, COMMON};
+        enum Kind { SMALLARR2DX, EWS1, EWSNONZERO, RANK1, RANK2, RANK3, RANK4, RANK5, X_EWSNONZERO, Y_EWSNONZERO, Z_EWSNONZERO, COMMON, BROADCAST_SCALAR_X, BROADCAST_SCALAR_Y, BROADCAST_3D, BROADCAST_4D, BROADCAST_5D };
 
         static FORCEINLINE Kind deduceKindOfLoopXZ(const Nd4jLong* xShapeInfo, const Nd4jLong* zShapeInfo);
         static FORCEINLINE Kind deduceKindOfLoopXYZ(const Nd4jLong* xShapeInfo, const Nd4jLong* yShapeInfo, const Nd4jLong* zShapeInfo);
         static FORCEINLINE Kind deduceKindOfLoopTadXZ(const Nd4jLong* xShapeInfo, const Nd4jLong* zShapeInfo, const Nd4jLong* tadShapeInfo);        
         static FORCEINLINE Kind deduceKindOfLoopTadXYZ(const Nd4jLong* xTadShapeInfo, const Nd4jLong* yTadShapeInfo, const Nd4jLong* zShapeInfo);
+        static FORCEINLINE Kind deduceKindOfLoopBroadcast(const Nd4jLong* xShapeInfo, const Nd4jLong* yShapeInfo, const Nd4jLong* zShapeInfo);
     
 };
 
@@ -82,6 +83,57 @@ LoopKind::Kind LoopKind::deduceKindOfLoopXZ(const Nd4jLong* xShapeInfo, const Nd
     return COMMON;
 }
 
+LoopKind::Kind LoopKind::deduceKindOfLoopBroadcast(const Nd4jLong* xShapeInfo, const Nd4jLong* yShapeInfo, const Nd4jLong* zShapeInfo) {
+    auto xRank = shape::rank(xShapeInfo);
+    auto yRank = shape::rank(yShapeInfo);
+    auto zRank = shape::rank(zShapeInfo);
+
+    auto xOrder = shape::order(xShapeInfo);
+    auto yOrder = shape::order(yShapeInfo);
+    auto zOrder = shape::order(zShapeInfo);
+
+    auto xEws = shape::elementWiseStride(xShapeInfo);
+    auto yEws = shape::elementWiseStride(yShapeInfo);
+    auto zEws = shape::elementWiseStride(zShapeInfo);
+
+    bool bNDLoopsRanks = (xRank == zRank && yRank <= xRank && yRank >= 2);
+
+    int countUnityDimsInY = 0, countUnityDimsInX = 0;
+    for (int i = 0; i < xRank; i++) {
+        if (i < yRank)
+            countUnityDimsInY += (1 == shape::sizeAt(yShapeInfo, i)) ? 1 : 0;
+        countUnityDimsInX += (1 == shape::sizeAt(xShapeInfo, i)) ? 1 : 0;
+    }
+
+    bool bNotCommonVectorCase = (countUnityDimsInY != yRank - 1) && (countUnityDimsInX != xRank - 1);
+
+    if (3 == xRank && bNDLoopsRanks && bNotCommonVectorCase)
+        return nd4j::LoopKind::BROADCAST_3D;
+    if (4 == xRank && bNDLoopsRanks && bNotCommonVectorCase)
+        return nd4j::LoopKind::BROADCAST_4D;
+    if (5 == xRank && bNDLoopsRanks && bNotCommonVectorCase)
+        return nd4j::LoopKind::BROADCAST_5D;
+
+
+    if (xRank == yRank && xRank == zRank && xOrder == 'c' && yOrder == 'c' && zOrder == 'c' && xEws == 1 && yEws == 1 && zEws == 1 && xRank >= 2) {
+        // we validate that shapes are equal till the last dim
+        for (int e = 0; e <  xRank - 1; e++) {
+            if (xShapeInfo[e+1] != yShapeInfo[e+1])
+                return COMMON;
+        }
+
+        // now, if one of the shapes has 1 as last dim
+        auto detect = xShapeInfo[xRank] == 1 ? -1 : (yShapeInfo[xRank] == 1) ? 1 : 0;
+
+        if (detect == 1)
+            return nd4j::LoopKind::BROADCAST_SCALAR_Y;
+        else if (detect == -1)
+            return nd4j::LoopKind::BROADCAST_SCALAR_X;
+        }
+
+    return nd4j::LoopKind::COMMON;
+}
+
 //////////////////////////////////////////////////////////////////////////////
 LoopKind::Kind LoopKind::deduceKindOfLoopXYZ(const Nd4jLong* xShapeInfo, const Nd4jLong* yShapeInfo, const Nd4jLong* zShapeInfo) {
 

libnd4j/include/helpers/ShapeBuilders.h

@@ -51,6 +51,13 @@ namespace nd4j {
         static Nd4jLong* copyShapeInfoAndType(const Nd4jLong* inShapeInfo, const DataType dtype, const bool copyStrides, memory::Workspace* workspace = nullptr);
         static Nd4jLong* copyShapeInfoAndType(const Nd4jLong* inShapeInfo, const Nd4jLong* shapeInfoToGetTypeFrom, const bool copyStrides, memory::Workspace* workspace = nullptr);
 
+        /**
+        * allocates memory for new shapeInfo and copy all information from inShapeInfo to new shapeInfo except dimensions in dimsToExclude (unit dimensions) and corresponding strides
+        * for example inShapeInfo is {3, 2,1,3,1,4,  12,12,4,4,1, 16384,1,99}, dimsToExclude = {2,3}, dimsSize = 2
+        * then outShapeInfo will contain {3, 2,3,4, 12,4,1, 16384,1,99}
+        */
+        static Nd4jLong* copyShapeInfoWithoutUnites(const Nd4jLong* inShapeInfo, const int dimsSize, const int* dimsToExclude, memory::Workspace* workspace = nullptr);
+
         static Nd4jLong* emptyShapeInfo(const nd4j::DataType dataType, memory::Workspace* workspace = nullptr);
 
         static Nd4jLong* emptyShapeInfo(const nd4j::DataType dataType, const char order, const std::vector<Nd4jLong> &shape, memory::Workspace* workspace = nullptr);

libnd4j/include/helpers/ShapeUtils.h

@@ -50,11 +50,13 @@ namespace nd4j {
     	static std::vector<Nd4jLong> evalRepeatShape(int axis, const std::vector<int>& repeats, const NDArray& arr);
 
         // evaluate shapeInfo of permuted array
-        static Nd4jLong* evalPermShapeInfo(const int* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace);
+        // if setContigStrides = true, then set contiguous strides in output shapeInfo in accordance with arr order
+        static Nd4jLong* evalPermShapeInfo(const int* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace, const bool setContigStrides = false);
         static Nd4jLong* evalPermShapeInfo(const Nd4jLong* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace);
 
         // evaluate shapeInfo of transposed array
-        static Nd4jLong* evalTranspShapeInfo(const NDArray& arr, nd4j::memory::Workspace* workspace);
+        // if setContigStrides = true, then set contiguous strides in output shapeInfo in accordance with arr order
+        static Nd4jLong* evalTranspShapeInfo(const NDArray& arr, nd4j::memory::Workspace* workspace, const bool setContigStrides = false);
 
         static bool copyVectorPart(std::vector<int>& target, std::vector<int>& source, int rank, int offset);
 
@@ -97,6 +99,8 @@ namespace nd4j {
         static std::string shapeAsString(const int rank, const Nd4jLong* shapeInfo);
         static std::string strideAsString(const NDArray* array);
 
+        static std::string shapeInfoAsString(const Nd4jLong* shapeInfo);
+
         static std::vector<Nd4jLong> shapeAsVector(const Nd4jLong* shapeInfo);
 
         // evaluate shapeInfo for diagonal array which is made using input arr elements as diagonal
@@ -176,6 +180,17 @@ namespace nd4j {
             return (numStrings + 1) * sizeof(Nd4jLong);
         }
 
+        /**
+         * This method selects strides based on dimentions required for broadcasting
+         * @param const pointer to input (Y) shape info for strides selection
+         * @param rank of input (X) to broadcasting
+         * @param dimentions size
+         * @param const pointer to dimentions for broadcasting
+         * @param pointer to output strides have to be pre allocated by 0
+         * @return
+         */
+        static void copyCertainStridesFromShapeInfo(const Nd4jLong* inShapeInfo, const int nRank, const int dimsSize, const int* dims, Nd4jLong* outStrides);
+
         /*
         * check whether arr1/arr2 is sub-array of arr2/arr1,
         * this method do not evaluate what array is sub-array, it returns true if arr1 is sub-array of arr2 or arr2 is sub-array of arr1

libnd4j/include/helpers/cpu/ConstantTadHelper.cpp

@@ -68,7 +68,7 @@ namespace nd4j {
             const Nd4jLong numOfSubArrs = ShapeUtils::getNumOfSubArrs(shapeInfo, dimsToExclude);
             const int subArrRank = (rank == dimsToExclude.size() || descriptor.areUnitiesinShape()) ? rank : rank - dimsToExclude.size();
 
-            auto sPtr = new Nd4jLong[shape::shapeInfoLength(subArrRank)];
+            auto sPtr = new Nd4jLong[shape::shapeInfoLength(subArrRank)];   // shape of sub-arrays (same for all for them)
             auto oPtr = new Nd4jLong[numOfSubArrs];
 
             if (numOfSubArrs > 0)

libnd4j/include/helpers/cpu/loops/IndexReductionLoops.hpp

@@ -49,7 +49,7 @@ void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
         case nd4j::LoopKind::EWS1: {
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto i = start; i < stop; i += increment) {
+                for (auto i = start; i < stop; i++) {
                     auto tad = const_cast<X *>(x) + tadOffsets[i];
                     auto indexValue = OpType::startingIndexValue(tad);
 
@@ -70,7 +70,7 @@ void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
         case nd4j::LoopKind::EWSNONZERO: {
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto i = start; i < stop; i += increment) {
+                for (auto i = start; i < stop; i++) {
                     auto tad = const_cast<X *>(x) + tadOffsets[i];
                     auto indexValue = OpType::startingIndexValue(tad);
 
@@ -91,7 +91,7 @@ void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
         case nd4j::LoopKind::RANK1: {
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto i = start; i < stop; i += increment) {
+                for (auto i = start; i < stop; i++) {
                     auto tad = const_cast<X *>(x) + tadOffsets[i];
                     auto indexValue = OpType::startingIndexValue(tad);
 
@@ -114,7 +114,7 @@ void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
             shape::updateStrides(2, tadShape, newStride, 'c');
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto i = start; i < stop; i += increment) {
+                for (auto i = start; i < stop; i++) {
                     auto tad = const_cast<X *>(x) + tadOffsets[i];
                     auto indexValue = OpType::startingIndexValue(tad);
 
@@ -141,7 +141,7 @@ void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
             shape::updateStrides(3, tadShape, newStride, 'c');
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto i = start; i < stop; i += increment) {
+                for (auto i = start; i < stop; i++) {
                     auto tad = const_cast<X *>(x) + tadOffsets[i];
                     auto indexValue = OpType::startingIndexValue(tad);
 
@@ -170,7 +170,7 @@ void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
             shape::updateStrides(4, tadShape, newStride, 'c');
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto i = start; i < stop; i += increment) {
+                for (auto i = start; i < stop; i++) {
                     auto tad = const_cast<X *>(x) + tadOffsets[i];
                     auto indexValue = OpType::startingIndexValue(tad);
 
@@ -201,7 +201,7 @@ void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
             shape::updateStrides(5, tadShape, newStride, 'c');
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto i = start; i < stop; i += increment) {
+                for (auto i = start; i < stop; i++) {
                     auto tad = const_cast<X *>(x) + tadOffsets[i];
                     auto indexValue = OpType::startingIndexValue(tad);
 
@@ -234,7 +234,7 @@ void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
             const bool canCastZ   = nd4j::DataTypeUtils::castShapeInfo<uint>(zShapeInfo,   castZShapeInfo);
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto i = start; i < stop; i += increment) {
+                for (auto i = start; i < stop; i++) {
                     auto tad = const_cast<X *>(x) + tadOffsets[i];
                     auto indexValue = OpType::startingIndexValue(tad);
 
@@ -258,7 +258,7 @@ void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
             const bool canCastTad = nd4j::DataTypeUtils::castShapeInfo<uint>(tadShapeInfo, castTadShapeInfo);
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto i = start; i < stop; i += increment) {
+                for (auto i = start; i < stop; i++) {
                     auto tad = const_cast<X *>(x) + tadOffsets[i];
                     auto indexValue = OpType::startingIndexValue(tad);
 
@@ -284,7 +284,7 @@ void nd4j::IndexReductionLoops<X,Z>::loopIndexReduce(X* x, Nd4jLong* xShapeInfo,
             const bool canCastZ   = nd4j::DataTypeUtils::castShapeInfo<uint>(zShapeInfo,   castZShapeInfo);
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto i = start; i < stop; i += increment) {
+                for (auto i = start; i < stop; i++) {
                     auto tad = const_cast<X *>(x) + tadOffsets[i];
                     auto indexValue = OpType::startingIndexValue(tad);
 

libnd4j/include/helpers/impl/MmulHelper.cpp

@@ -43,23 +43,30 @@ nd4j::NDArray* nd4j::MmulHelper::tensorDot(const nd4j::NDArray* a, const nd4j::N
 
     auto outShape = ShapeUtils::evalShapeForTensorDot(a, b, axes_0, axes_1, permutAt, permutBt, shapeAt, shapeBt);
 
-    NDArray aPR = a->permute(permutAt);
+    // check whether permutation is necessary
-    NDArray bPR = b->permute(permutBt);
+    const NDArray* aP = permutAt.empty() ? a : new NDArray(a->permute(permutAt));
+    const NDArray* bP = permutBt.empty() ? b : new NDArray(b->permute(permutBt));
 
     // check whether reshape is necessary
-    if(!aPR.isSameShape(shapeAt))
+    const NDArray* aPR = aP->isSameShape(shapeAt) ? aP : new NDArray(aP->reshape(aP->ordering(), shapeAt));
-        aPR.reshapei( shapeAt);
+    const NDArray* bPR = bP->isSameShape(shapeAt) ? bP : new NDArray(bP->reshape(bP->ordering(), shapeBt));
-    if(!bPR.isSameShape(shapeBt))
-        bPR.reshapei( shapeBt);
 
-    NDArray* c = mmul(&aPR, &bPR, nullptr, 1.0, 0.0);
+    NDArray* c = mmul(aPR, bPR, nullptr, 1.0, 0.0);
 
     c->reshapei(outShape);
 
+    if(aP != aPR)
+        delete aPR;
+    if(bP != bPR)
+        delete bPR;
+    if(a != aP)
+        delete aP;
+    if(b != bP)
+        delete bP;
+
     return c;
 }
 
-
 //////////////////////////////////////////////////////////////////////////
 void nd4j::MmulHelper::tensorDot(const nd4j::NDArray* a, const nd4j::NDArray* b, nd4j::NDArray* c, const std::vector<int>& axes_a, const std::vector<int>& axes_b, const std::vector<int>& permutForC) {
 
@@ -67,32 +74,38 @@ void nd4j::MmulHelper::tensorDot(const nd4j::NDArray* a, const nd4j::NDArray* b,
     std::vector<Nd4jLong> shapeAt, shapeBt;
     ShapeUtils::evalShapeForTensorDot(a, b, axes_a, axes_b, permutAt, permutBt, shapeAt, shapeBt);
 
-    NDArray *cP(c), *cPR(c);
-
     // check whether permutation is required
-    if(!permutForC.empty())
+    NDArray* cP = permutForC.empty() ? c : new NDArray(c->permute(permutForC));
-        cP = new NDArray(c->permute(permutForC));
 
-    auto aPR = a->permute(permutAt);
+    // check whether permutation is necessary
-    auto bPR = b->permute(permutBt);
+    const NDArray* aP = permutAt.empty() ? a : new NDArray(a->permute(permutAt));
+    const NDArray* bP = permutBt.empty() ? b : new NDArray(b->permute(permutBt));
 
     // check whether reshape is necessary
-    if(!aPR.isSameShape(shapeAt))
+    const NDArray* aPR = aP->isSameShape(shapeAt) ? aP : new NDArray(aP->reshape(aP->ordering(), shapeAt));
-            aPR.reshapei(shapeAt);
+    const NDArray* bPR = bP->isSameShape(shapeAt) ? bP : new NDArray(bP->reshape(bP->ordering(), shapeBt));
-    if(!bPR.isSameShape(shapeBt))
-            bPR.reshapei(shapeBt);
 
-    if(!cP->isSameShape({aPR.sizeAt(0), bPR.sizeAt(1)}))
+    std::vector<Nd4jLong> requiredCshape = {aPR->sizeAt(0), bPR->sizeAt(1)};
-        cPR = new NDArray(cP->reshape(cP->ordering(), {aPR.sizeAt(0), bPR.sizeAt(1)}));
 
-    mmul(&aPR, &bPR, cPR, 1.0, 0.0);
+    NDArray* cPR = cP->isSameShape(requiredCshape) ? cP : new NDArray(cP->reshape(cP->ordering(), requiredCshape, false));
+
+    mmul(aPR, bPR, cPR, 1.0, 0.0);
 
     if(cPR->getBuffer() != cP->getBuffer() || cPR->getSpecialBuffer() != cP->getSpecialBuffer() )   // this means both permute and reshape have been performed on c, cP always points on c->getBuffer()
         cP->assign(cPR);
 
-    if(cPR != c)
+   if(aP != aPR)
+        delete aPR;
+    if(bP != bPR)
+        delete bPR;
+    if(a != aP)
+        delete aP;
+    if(b != bP)
+        delete bP;
+
+    if(cP != cPR)
         delete cPR;
-    if(cP != c)
+    if(c != cP)
         delete cP;
 }
 
@@ -129,7 +142,7 @@ void nd4j::MmulHelper::tensorDot(const NDArray* a, const NDArray* b, NDArray* c,
     if(!whatToDoWithC.empty()) {
         cArrs = std::vector<NDArray*>(whatToDoWithC.size()+1, c);
         for(int i = 0; i < cArrs.size()-1; ++i)
-            cArrs[i+1] = (whatToDoWithC[i] == 'p') ? new NDArray(cArrs[i]->permute(modifC[i])) : new NDArray(cArrs[i]->reshape(c->ordering(), modifC[i]));  // since we ignore first element in cArrs (that is cArrs[0]) then it is always equal to c
+            cArrs[i+1] = (whatToDoWithC[i] == 'p') ? new NDArray(cArrs[i]->permute(modifC[i])) : new NDArray(cArrs[i]->reshape(c->ordering(), modifC[i], false));  // since we ignore first element in cArrs (that is cArrs[0]) then it is always equal to c
     }
 
     mmul(aPR, bPR, cArrs[cArrs.size()-1], 1.0, 0.0);
@@ -208,7 +221,7 @@ nd4j::NDArray* MmulHelper::mmul(const nd4j::NDArray* A, const nd4j::NDArray* B,
     // vector x matrix, A{M} x B{M,N} = C{N} -> reduce to matrix x matrix A2{1,M} x B{M,N} = C2{1,N}, since there is no corresponding blas operation sgevm
     if(isAVector && bRank == 2) {
         NDArray* A2 = new NDArray(A->reshape(A->ordering(), {1, A->lengthOf()}));               // A{M} -> A2{1,M}
-        NDArray* C2 = C ? new NDArray(C->reshape(C->ordering(), {1, C->lengthOf()})) : nullptr; // C{N} -> C2{1,N}
+        NDArray* C2 = C ? new NDArray(C->reshape(C->ordering(), {1, C->lengthOf()}, false)) : nullptr; // C{N} -> C2{1,N}
         auto result = mmulMxM(A2, B, C2, alpha, beta, outOrder);                                // result{1,N}
         delete A2;
         delete C2;

libnd4j/include/helpers/impl/ShapeBuilders.cpp

@@ -139,5 +139,15 @@ namespace nd4j {
         return ShapeBuilders::copyShapeInfoAndType(inShapeInfo, ArrayOptions::dataType(shapeInfoToGetTypeFrom), copyStrides, workspace);
     }
 
+////////////////////////////////////////////////////////////////////////////////
+Nd4jLong* ShapeBuilders::copyShapeInfoWithoutUnites(const Nd4jLong* inShapeInfo, const int dimsSize, const int* dimsToExclude, memory::Workspace* workspace) {
+
+    Nd4jLong *outShapeInfo = nullptr;
+    ALLOCATE(outShapeInfo, workspace, shape::shapeInfoLength(inShapeInfo[0] - dimsSize), Nd4jLong);
+
+    shape::excludeUnitiesFromShapeInfo(inShapeInfo, dimsSize, dimsToExclude, outShapeInfo);
+
+    return outShapeInfo;
+}
 
 }

libnd4j/include/helpers/impl/ShapeUtils.cpp

@@ -75,10 +75,23 @@ std::vector<Nd4jLong> ShapeUtils::evalShapeForTensorDot(const Nd4jLong* aShapeIn
     permutBt = axesB;
     permutBt.insert(permutBt.end(), list_B.begin(), list_B.end());
 
+    // if permut contains something like {0,1,2,..rank-1}, then there is no need to make permutation and we return empty vector in this case
+    uint i1, i2;
+    for(i1 = 0; i1 < aRank; ++i1)
+        if(permutAt[i1] != i1)
+            break;
+    if(i1 == aRank)
+        permutAt = {};
+    for(i2 = 0; i2 < bRank; ++i2)
+        if(permutBt[i2] != i2)
+            break;
+    if(i2 == bRank)
+        permutBt = {};
+
     Nd4jLong n2 = 1;
     for (int i = 0; i < axeAsize; i++)
         n2 *= aShapeInfo[axesA[i] + 1];
-    shapeAt = {-1, n2};
+    shapeAt = {shape::length(aShapeInfo) / n2, n2};
 
     std::vector<Nd4jLong> oldShapeA;
     oldShapeA.resize(list_A.size());
@@ -89,7 +102,7 @@ std::vector<Nd4jLong> ShapeUtils::evalShapeForTensorDot(const Nd4jLong* aShapeIn
     Nd4jLong n3 = 1;
     for (int i = 0; i < axeBsize; i++)
         n3 *= bShapeInfo[axesB[i] + 1];
-    shapeBt = {n3, -1};
+    shapeBt = {n3, shape::length(bShapeInfo) / n3};
 
     std::vector<Nd4jLong> oldShapeB;
     oldShapeB.resize(list_B.size());
@@ -300,32 +313,37 @@ std::vector<Nd4jLong> ShapeUtils::evalRepeatShape(int axis, const std::vector<in
     return outShape;
 }
 
-
 //////////////////////////////////////////////////////////////////////////
 // evaluate shapeInfo of permuted array
-    Nd4jLong* ShapeUtils::evalPermShapeInfo(const int* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace) {
+Nd4jLong* ShapeUtils::evalPermShapeInfo(const int* dimensions, const int rank, const NDArray& arr, nd4j::memory::Workspace* workspace, const bool setContigStrides) {
 
     if (!arr.nonNull())
-            throw std::runtime_error("ShapeUtils::evalPermShapeInfo static method: wrong arguments in pn/termute method: either array is nullptr!");
+        throw std::runtime_error("ShapeUtils::evalPermShapeInfo static method: wrong arguments: array is nullptr!");
 
     if (rank != arr.rankOf())
-            throw std::runtime_error("ShapeUtils::evalPermShapeInfo static method: wrong arguments in pn/termute method: rank is not suitable!");
+        throw std::runtime_error("ShapeUtils::evalPermShapeInfo static method: wrong arguments: rank is not suitable!");
 
     auto shapeInfoLength = shape::shapeInfoLength(rank);
-        // allocate memory for new array - shapeInfo
 
+    // allocate memory for new array - shapeInfo
     Nd4jLong *shapeInfoNew = nullptr;
     ALLOCATE(shapeInfoNew, workspace, shapeInfoLength, Nd4jLong);
+
     // copy arr _shapeInfo into new array
     memcpy(shapeInfoNew, arr.getShapeInfo(), shape::shapeInfoByteLength(rank));
+
     // perform buffer permutation
-        shape::doPermuteShapeInfo(shapeInfoNew, dimensions);
+    shape::doPermuteShapeInfo(shapeInfoNew, dimensions, arr.lengthOf());
+
+    if(setContigStrides)
+        shape::updateStrides(shapeInfoNew, arr.ordering());
 
     ShapeDescriptor descriptor(shapeInfoNew);
-        RELEASE(shapeInfoNew, workspace);
-        return ConstantShapeHelper::getInstance()->bufferForShapeInfo(descriptor).primaryAsT<Nd4jLong>();
-    }
 
+    RELEASE(shapeInfoNew, workspace);
+
+    return ConstantShapeHelper::getInstance()->bufferForShapeInfo(descriptor).primaryAsT<Nd4jLong>();
+}
 
     //////////////////////////////////////////////////////////////////////////
     // evaluate shapeInfo of permuted array
@@ -337,14 +355,14 @@ std::vector<Nd4jLong> ShapeUtils::evalRepeatShape(int axis, const std::vector<in
 
 //////////////////////////////////////////////////////////////////////////
 // evaluate shapeInfo of transposed array
-    Nd4jLong* ShapeUtils::evalTranspShapeInfo(const NDArray& arr, nd4j::memory::Workspace* workspace) {
+    Nd4jLong* ShapeUtils::evalTranspShapeInfo(const NDArray& arr, nd4j::memory::Workspace* workspace, const bool setContigStrides) {
 
         int rank = arr.rankOf();
         std::vector<int> dimensions(rank);
         for (int i = 0; i < rank; ++i)
             dimensions[i] = rank - 1 - i;
 
-        return evalPermShapeInfo(dimensions.data(), dimensions.size(), arr, workspace);
+        return evalPermShapeInfo(dimensions.data(), dimensions.size(), arr, workspace, setContigStrides);
     }
 
 //////////////////////////////////////////////////////////////////////////
@@ -653,6 +671,26 @@ Nd4jLong* ShapeUtils::evalTileShapeInfo(const NDArray& arr, const std::vector<Nd
         return result;
     }
 
+    std::string ShapeUtils::shapeInfoAsString(const Nd4jLong* shapeInfo) {
+
+        if(!shapeInfo)
+            throw std::runtime_error("ShapeUtils::shapeAsString method: input shapeInfo must not be nullptr !");
+
+        std::string result;
+
+        int len = shape::shapeInfoLength(shapeInfo[0]);
+
+        result.append("[");
+        for (int e = 0; e < len; e++) {
+            result += flatbuffers::NumToString(shapeInfo[e]);
+            if (e < len - 1)
+                result.append(", ");
+        }
+        result.append("]");
+
+        return result;
+    }
+
 
     std::string ShapeUtils::shapeAsString(const int rank, const Nd4jLong* shapeInfo) {
         if(!shapeInfo)
@@ -1019,6 +1057,29 @@ std::vector<int> ShapeUtils::tadAxesForSimpleBroadcast(const NDArray& max, const
     return numOfMinTads == 1 ? maxTadDims : std::vector<int>();
 }
 
+void ShapeUtils::copyCertainStridesFromShapeInfo(const Nd4jLong* inShapeInfo, const int nRank, const int dimsSize, const int* dims, Nd4jLong* outStrides) {
+
+    int yRank = shape::rank(inShapeInfo);
+    auto  yOrigStride = shape::stride(inShapeInfo);
+
+    if (yRank == nRank) {
+        for (int i = 0; i < yRank; ++i) {
+            // x[2,3,4] * y[2,1,4] = z[2,3,4]
+            outStrides[i] = (1 == shape::sizeAt(inShapeInfo, i)) ? 0 : yOrigStride[i];
+        }
+    }
+    else {
+
+        auto dimEx = nd4j::ShapeUtils::evalDimsToExclude(nRank, dimsSize, dims);
+
+        for (int i = 0, it = 0; i < nRank; ++i) {
+            auto nCount = std::count(dimEx.cbegin(), dimEx.cend(), i);
+            outStrides[i] = (0 == nCount) ? yOrigStride[it++] : 0;
+            if (it == yRank)
+                break;
+        }
+    }
+}
 ////////////////////////////////////////////////////////////////////////////////
 /*
 bool ShapeUtils::isSubArrayCase(const NDArray& arr1, const NDArray& arr2, std::vector<int>& sameDims) {

libnd4j/include/loops/broadcasting.h

@@ -40,6 +40,7 @@
 #endif
 
 #include <helpers/TAD.h>
+#include <helpers/LoopKind.h>
 
 #include "legacy_ops.h"
 
@@ -122,6 +123,7 @@ namespace functions {
                              Nd4jLong *tadOffset,
                              Nd4jLong *tadShapeInfoZ,
                              Nd4jLong *tadOffsetZ,
+                             nd4j::LoopKind::Kind loopKind,
                              uint64_t start,
                              uint64_t stop);
 
@@ -149,6 +151,7 @@ namespace functions {
                              Nd4jLong *tadOffset,
                              Nd4jLong *tadShapeInfoZ,
                              Nd4jLong *tadOffsetZ,
+                             nd4j::LoopKind::Kind loopKind,
                              uint64_t start,
                              uint64_t stop);
 

libnd4j/include/loops/cpu/TrueBroadcastHelper.hpp

@@ -14,9 +14,9 @@
  * SPDX-License-Identifier: Apache-2.0
  ******************************************************************************/
 
-//
+ //
-// @author Yurii Shyrma (iuriish@yahoo.com)
+ // @author Yurii Shyrma (iuriish@yahoo.com)
-//
+ //
 
 #include <loops/TrueBroadcastHelper.h>
 #include <ops/ops.h>
@@ -25,12 +25,12 @@
 using namespace simdOps;
 
 namespace nd4j {
-namespace helpers {
+    namespace helpers {
 
-////////////////////////////////////////////////////////////////////////
+        ////////////////////////////////////////////////////////////////////////
-template <typename X, typename  Y, typename Z>
+        template <typename X, typename  Y, typename Z>
-template<typename OpType>
+        template<typename OpType>
-void TrueBroadcastHelper<X, Y, Z>::exec(const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
+        void TrueBroadcastHelper<X, Y, Z>::exec(const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
 
 
             const X* x = reinterpret_cast<X*>(xArr.getBuffer());
@@ -45,11 +45,11 @@ void TrueBroadcastHelper<X, Y, Z>::exec(const NDArray& xArr, const NDArray& yArr
             const int yRank = yArr.rankOf();
             const int zRank = zArr.rankOf();
 
-    bool bSpecialCase = (1 == xArr.ews() && 'c' == xArr.ordering() && 1 == yRank &&
+            bool bSpecialCase = (1 == xArr.ews() && 'c' == xArr.ordering() &&
                 1 == yArr.ews() && 'c' == yArr.ordering() &&
                 1 == zArr.ews() && 'c' == zArr.ordering());
 
-    if (bSpecialCase) {
+            if (bSpecialCase && yArr.isColumnVector() && 1 == xArr.sizeAt(-1) ) {
                 auto yLen = (uint32_t)yArr.lengthOf();
                 auto func = PRAGMA_THREADS_FOR{
                    for (uint32_t i = start; i < stop; i++) {
@@ -64,8 +64,44 @@ void TrueBroadcastHelper<X, Y, Z>::exec(const NDArray& xArr, const NDArray& yArr
                 return;
             }
 
+
+            auto yShapeInt = yArr.getShapeAsVectorInt();
+            auto xShapeInt = xArr.getShapeAsVectorInt();
+            auto nCountY = std::count_if(yShapeInt.cbegin(), yShapeInt.cend(), [](int i) { return i == 1; });
+            auto nCountX = std::count_if(xShapeInt.cbegin(), xShapeInt.cend(), [](int i) { return i == 1; });
+
+            bool bSpecialCase2 = (xRank == zRank && yRank == zRank && 1 == xArr.sizeAt(-1) && 1 == yArr.sizeAt(-2) && 1 == nCountY && 1 == nCountX);
+
+            if (bSpecialCase && bSpecialCase2) {
+
+                int zDim1 = zArr.sizeAt(-2);
+                int zDim2 = zArr.sizeAt(-1);
+
+                int nLen = zArr.lengthOf() / yArr.sizeAt(-1);
+
+                auto func = PRAGMA_THREADS_FOR{
+                     for (uint32_t total = start; total < stop; total++) {
+
+                        uint32_t i = total / zDim1;
+                        uint32_t j = total % zDim1;
+
+                        uint32_t index = (i * zDim1) + j;
+                        auto rZ = z + (index * zDim2);
+                        auto rY = y + (i * zDim2);
+                        auto rX = x[index];
+
+                        for (uint32_t n = 0; n < zDim2; n++) {
+                             rZ[n] = OpType::op(rX, rY[n]);
+                        }
+                    }
+                };
+                samediff::Threads::parallel_tad(func, 0, nLen, 1);
+                return;
+            }
+
+
             const Nd4jLong zLen = zArr.lengthOf();
-    auto func = PRAGMA_THREADS_FOR {
+            auto func = PRAGMA_THREADS_FOR{
                 std::vector<Nd4jLong> xCoords(xArr.rankOf()), yCoords(yArr.rankOf()), zCoords(zArr.rankOf());
 
                 for (auto i = start; i < stop; ++i) {
@@ -77,7 +113,8 @@ void TrueBroadcastHelper<X, Y, Z>::exec(const NDArray& xArr, const NDArray& yArr
                         if (ix >= 0) {
                             if (xShapeInfo[ix + 1] == zShapeInfo[iz + 1]) {
                                 xCoords[ix--] = zCoords[iz];
-                    } else {
+                            }
+                             else {
                               xCoords[ix--] = 0;
                              }
                         }
@@ -85,7 +122,8 @@ void TrueBroadcastHelper<X, Y, Z>::exec(const NDArray& xArr, const NDArray& yArr
                         if (iy >= 0) {
                               if (yShapeInfo[iy + 1] == zShapeInfo[iz + 1]) {
                                   yCoords[iy--] = zCoords[iz];
-                    } else {
+                              }
+                              else {
                                   yCoords[iy--] = 0;
                               }
                           }
@@ -100,17 +138,17 @@ void TrueBroadcastHelper<X, Y, Z>::exec(const NDArray& xArr, const NDArray& yArr
             };
 
             samediff::Threads::parallel_for(func, 0, zLen);
-}
+        }
 
-template <typename X, typename  Y, typename Z>
+        template <typename X, typename  Y, typename Z>
-void TrueBroadcastHelper<X, Y, Z>::exec(const nd4j::broadcast::Ops opNum, const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
+        void TrueBroadcastHelper<X, Y, Z>::exec(const nd4j::broadcast::Ops opNum, const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
             DISPATCH_BY_OPNUM_TTT(exec, PARAMS(xArr, yArr, zArr), BROADCAST_OPS);
-}
+        }
 
-////////////////////////////////////////////////////////////////////////
+        ////////////////////////////////////////////////////////////////////////
-template <typename X, typename  Z>
+        template <typename X, typename  Z>
-template<typename OpType>
+        template<typename OpType>
-void TrueBroadcastBoolHelper<X, Z>::exec(const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
+        void TrueBroadcastBoolHelper<X, Z>::exec(const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
 
             const X* x = reinterpret_cast<X*>(xArr.getBuffer());
             const X* y = reinterpret_cast<X*>(yArr.getBuffer());
@@ -126,7 +164,7 @@ void TrueBroadcastBoolHelper<X, Z>::exec(const NDArray& xArr, const NDArray& yAr
 
             const Nd4jLong zLen = zArr.lengthOf();
 
-    auto func = PRAGMA_THREADS_FOR {
+            auto func = PRAGMA_THREADS_FOR{
                 std::vector<Nd4jLong> xCoords(xArr.rankOf()), yCoords(yArr.rankOf()), zCoords(zArr.rankOf());
 
                 for (auto i = start; i < stop; ++i) {
@@ -138,7 +176,8 @@ void TrueBroadcastBoolHelper<X, Z>::exec(const NDArray& xArr, const NDArray& yAr
                         if (ix >= 0) {
                             if (xShapeInfo[ix + 1] == zShapeInfo[iz + 1]) {
                                 xCoords[ix--] = zCoords[iz];
-                    } else {
+                            }
+                            else {
                                 xCoords[ix--] = 0;
                             }
                         }
@@ -146,7 +185,8 @@ void TrueBroadcastBoolHelper<X, Z>::exec(const NDArray& xArr, const NDArray& yAr
                         if (iy >= 0) {
                             if (yShapeInfo[iy + 1] == zShapeInfo[iz + 1]) {
                                 yCoords[iy--] = zCoords[iz];
-                    } else {
+                            }
+                            else {
                                  yCoords[iy--] = 0;
                             }
                         }
@@ -161,17 +201,17 @@ void TrueBroadcastBoolHelper<X, Z>::exec(const NDArray& xArr, const NDArray& yAr
             };
 
             samediff::Threads::parallel_for(func, 0, zLen);
-}
+        }
 
-template <typename X, typename  Y>
+        template <typename X, typename  Y>
-void TrueBroadcastBoolHelper<X, Y>::exec(const nd4j::broadcast::BoolOps opNum, const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
+        void TrueBroadcastBoolHelper<X, Y>::exec(const nd4j::broadcast::BoolOps opNum, const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
             DISPATCH_BY_OPNUM_TT(exec, PARAMS(xArr, yArr, zArr), BROADCAST_BOOL_OPS);
-}
+        }
 
-////////////////////////////////////////////////////////////////////////
+        ////////////////////////////////////////////////////////////////////////
-template <typename X>
+        template <typename X>
-template<typename OpType>
+        template<typename OpType>
-void TrueBroadcastIntHelper<X>::exec(const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
+        void TrueBroadcastIntHelper<X>::exec(const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
 
             const X* x = reinterpret_cast<X*>(xArr.getBuffer());
             const X* y = reinterpret_cast<X*>(yArr.getBuffer());
@@ -187,7 +227,7 @@ void TrueBroadcastIntHelper<X>::exec(const NDArray& xArr, const NDArray& yArr, N
 
             const Nd4jLong zLen = zArr.lengthOf();
 
-    auto func = PRAGMA_THREADS_FOR {
+            auto func = PRAGMA_THREADS_FOR{
                 std::vector<Nd4jLong> xCoords(xArr.rankOf()), yCoords(yArr.rankOf()), zCoords(zArr.rankOf());
 
                 for (auto i = start; i < stop; ++i) {
@@ -199,7 +239,8 @@ void TrueBroadcastIntHelper<X>::exec(const NDArray& xArr, const NDArray& yArr, N
                         if (ix >= 0) {
                             if (xShapeInfo[ix + 1] == zShapeInfo[iz + 1]) {
                                 xCoords[ix--] = zCoords[iz];
-                    } else {
+                            }
+                            else {
                                 xCoords[ix--] = 0;
                             }
                         }
@@ -207,7 +248,8 @@ void TrueBroadcastIntHelper<X>::exec(const NDArray& xArr, const NDArray& yArr, N
                         if (iy >= 0) {
                             if (yShapeInfo[iy + 1] == zShapeInfo[iz + 1]) {
                                yCoords[iy--] = zCoords[iz];
-                    } else {
+                            }
+                            else {
                                yCoords[iy--] = 0;
                             }
                         }
@@ -222,28 +264,28 @@ void TrueBroadcastIntHelper<X>::exec(const NDArray& xArr, const NDArray& yArr, N
             };
 
             samediff::Threads::parallel_for(func, 0, zLen);
-}
+        }
 
-template <typename X>
+        template <typename X>
-void TrueBroadcastIntHelper<X>::exec(const nd4j::broadcast::IntOps opNum, const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
+        void TrueBroadcastIntHelper<X>::exec(const nd4j::broadcast::IntOps opNum, const NDArray& xArr, const NDArray& yArr, NDArray& zArr) {
             DISPATCH_BY_OPNUM_T(exec, PARAMS(xArr, yArr, zArr), BROADCAST_INT_OPS);
-}
+        }
 
-/*
+        /*
-BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_0);
+        BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_0);
-BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_1);
+        BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_1);
-BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_2);
+        BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_2);
-BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_3);
+        BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_3);
-BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_4);
+        BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_4);
-BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_5);
+        BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_5);
-BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_6);
+        BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_6);
-BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_7);
+        BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_7);
-BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_8);
+        BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_8);
-BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_9);
+        BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastHelper, , PAIRWISE_TYPES_9);
 
-BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastBoolHelper, , LIBND4J_TYPES, BOOL_TYPES);
+        BUILD_DOUBLE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastBoolHelper, , LIBND4J_TYPES, BOOL_TYPES);
 
-BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastIntHelper, , INTEGER_TYPES);
+        BUILD_SINGLE_TEMPLATE(template class ND4J_EXPORT TrueBroadcastIntHelper, , INTEGER_TYPES);
-*/
+        */
-}
+    }
 }

libnd4j/include/loops/cpu/broadcasting.hpp

@@ -25,6 +25,7 @@
 #include <LoopKind.h>
 #include <helpers/ConstantTadHelper.h>
 #include <execution/Threads.h>
+#include <helpers/ShapeUtils.h>
 
 using namespace simdOps;
 
@@ -75,6 +76,7 @@ namespace functions {
                              Nd4jLong *xTadOffset,
                              Nd4jLong *zTadShapeInfo,
                              Nd4jLong *zTadOffset,
+                             nd4j::LoopKind::Kind loopKind,
                              uint64_t start,
                              uint64_t stop) {
             DISPATCH_BY_OPNUM_TTT(exec, PARAMS(x,
@@ -88,7 +90,7 @@ namespace functions {
                                                xTadShapeInfo,
                                                xTadOffset,
                                                zTadShapeInfo,
-                                               zTadOffset, start, stop), BROADCAST_OPS);
+                                               zTadOffset, loopKind, start, stop), BROADCAST_OPS);
         }
 
         template <typename X, typename  Y, typename Z>
@@ -105,6 +107,7 @@ namespace functions {
                              Nd4jLong *xTadOffset,
                              Nd4jLong *zTadShapeInfo,
                              Nd4jLong *zTadOffset,
+                             nd4j::LoopKind::Kind loopKind,
                              uint64_t start,
                              uint64_t stop) {
 
@@ -142,7 +145,14 @@ namespace functions {
                 auto yEws = shape::elementWiseStride(yShapeInfo);
                 auto zEws = shape::elementWiseStride(zTadShapeInfo);
 
-                const nd4j::LoopKind::Kind kindOfLoop = nd4j::LoopKind::deduceKindOfLoopXYZ(xTadShapeShapeInfo, yShapeInfo, zTadShapeInfo);
+
+                const nd4j::LoopKind::Kind kindOfLoop =
+                    (loopKind == nd4j::LoopKind::BROADCAST_SCALAR_X ||
+                        loopKind == nd4j::LoopKind::BROADCAST_SCALAR_Y ||
+                        loopKind == nd4j::LoopKind::BROADCAST_3D ||
+                        loopKind == nd4j::LoopKind::BROADCAST_4D ||
+                        loopKind == nd4j::LoopKind::BROADCAST_5D)
+                    ? loopKind : nd4j::LoopKind::deduceKindOfLoopXYZ(xTadShapeShapeInfo, yShapeInfo, zTadShapeInfo);
 
                 if (kindOfLoop == nd4j::LoopKind::EWS1) {
                     for (auto i = start; i < stop; i++) {
@@ -163,6 +173,131 @@ namespace functions {
                         for (unsigned int f = 0; f < tadLength; f++)
                             oZ[f * zEws] = OpType::op(oX[f * xEws], y[f * yEws]);
                     }
+                } else if(kindOfLoop == nd4j::LoopKind::BROADCAST_SCALAR_X){
+                    // this loop effectively turns broadcast into series of scalar ops
+                    auto loopLength = yShapeInfo[shape::rank(yShapeInfo)];
+
+                    for (auto i = start; i < stop; i++) {
+                        auto oY = y + (i * loopLength);
+                        auto oZ = z + (i * loopLength);
+
+                        const auto oX = x[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (unsigned int f = 0; f < loopLength; f++)
+                            oZ[f] = OpType::op(oX, oY[f]);
+                    }
+                } else if(kindOfLoop == nd4j::LoopKind::BROADCAST_SCALAR_Y){
+                    // this loop effectively turns broadcast into series of scalar ops
+                    auto loopLength = xShapeInfo[shape::rank(xShapeInfo)];
+
+                    for (auto i = start; i < stop; i++) {
+                        auto oX = x + (i * loopLength);
+                        auto oZ = z + (i * loopLength);
+
+                        const auto oY = y[i];
+
+                        PRAGMA_OMP_SIMD
+                        for (unsigned int f = 0; f < loopLength; f++)
+                            oZ[f] = OpType::op(oX[f], oY);
+                    }
+                }
+                else if (kindOfLoop == nd4j::LoopKind::BROADCAST_3D) {
+
+                    int xRank = shape::rank(xShapeInfo);
+                    int yRank = shape::rank(yShapeInfo);
+
+                    auto  xStrides = shape::stride(xShapeInfo);
+                    auto  zStrides = shape::stride(zShapeInfo);
+
+                    Nd4jLong  yStrides[3] = { 0,0,0 };
+                    nd4j::ShapeUtils::copyCertainStridesFromShapeInfo(yShapeInfo, xRank, dimensionLength, dimension, yStrides);
+
+                    uint32_t nSize1 = shape::sizeAt(zShapeInfo, 1);
+                    uint32_t nSize2 = shape::sizeAt(zShapeInfo, 2);
+
+                    for (uint32_t index0 = start; index0 < stop; index0++) {
+
+                        PRAGMA_OMP_SIMD
+                            for (uint32_t index1 = 0; index1 < nSize1; index1++) {
+                                for (uint32_t index2 = 0; index2 < nSize2; index2++) {
+                                    auto rX = x + (xStrides[0] * index0 + xStrides[1] * index1 + xStrides[2] * index2);
+                                    auto rY = y + (yStrides[0] * index0 + yStrides[1] * index1 + yStrides[2] * index2);
+                                    auto rZ = z + (zStrides[0] * index0 + zStrides[1] * index1 + zStrides[2] * index2);
+                                    *rZ = OpType::op(*rX, *rY);
+                                }
+                            }
+
+                    }
+
+                }
+                else if (kindOfLoop == nd4j::LoopKind::BROADCAST_4D) {
+
+                    int xRank = shape::rank(xShapeInfo);
+                    int yRank = shape::rank(yShapeInfo);
+
+                    auto  xStrides = shape::stride(xShapeInfo);
+                    auto  zStrides = shape::stride(zShapeInfo);
+
+                    Nd4jLong  yStrides[4] = { 0,0,0,0 };
+                    nd4j::ShapeUtils::copyCertainStridesFromShapeInfo(yShapeInfo, xRank, dimensionLength, dimension, yStrides);
+
+                    uint32_t nSize1 = shape::sizeAt(zShapeInfo, 1);
+                    uint32_t nSize2 = shape::sizeAt(zShapeInfo, 2);
+                    uint32_t nSize3 = shape::sizeAt(zShapeInfo, 3);
+
+                    for (uint32_t i = start; i < stop; i++) {
+
+                        uint32_t index0 = i / nSize1;
+                        uint32_t index1 = i % nSize1;
+
+                        PRAGMA_OMP_SIMD
+                            for (uint32_t index2 = 0; index2 < nSize2; index2++) {
+                                for (uint32_t index3 = 0; index3 < nSize3; index3++) {
+                                    auto rX = x + (xStrides[0] * index0 + xStrides[1] * index1 + xStrides[2] * index2 + xStrides[3] * index3);
+                                    auto rY = y + (yStrides[0] * index0 + yStrides[1] * index1 + yStrides[2] * index2 + yStrides[3] * index3);
+                                    auto rZ = z + (zStrides[0] * index0 + zStrides[1] * index1 + zStrides[2] * index2 + zStrides[3] * index3);
+                                    *rZ = OpType::op(*rX, *rY);
+                                }
+                            }
+                    }
+
+                }
+                else if (kindOfLoop == nd4j::LoopKind::BROADCAST_5D) {
+
+                    int xRank = shape::rank(xShapeInfo);
+                    int yRank = shape::rank(yShapeInfo);
+
+                    auto  xStrides = shape::stride(xShapeInfo);
+                    auto  zStrides = shape::stride(zShapeInfo);
+
+                    Nd4jLong  yStrides[5] = { 0,0,0,0,0 };
+                    nd4j::ShapeUtils::copyCertainStridesFromShapeInfo(yShapeInfo, xRank, dimensionLength, dimension, yStrides);
+
+                    uint32_t nSize1 = shape::sizeAt(zShapeInfo, 1);
+                    uint32_t nSize2 = shape::sizeAt(zShapeInfo, 2);
+                    uint32_t nSize3 = shape::sizeAt(zShapeInfo, 3);
+                    uint32_t nSize4 = shape::sizeAt(zShapeInfo, 4);
+
+                    for (uint32_t i = start; i < stop; i++) {
+
+                        uint32_t index0 = i / nSize1;
+                        uint32_t index1 = i % nSize1;
+
+                        PRAGMA_OMP_SIMD
+                            for (uint32_t index2 = 0; index2 < nSize2; index2++) {
+                                for (uint32_t index3 = 0; index3 < nSize3; index3++) {
+                                    for (uint32_t index4 = 0; index4 < nSize4; index4++) {
+                                        auto rX = x + (xStrides[0] * index0 + xStrides[1] * index1 + xStrides[2] * index2 + xStrides[3] * index3 + xStrides[4] * index4);
+                                        auto rY = y + (yStrides[0] * index0 + yStrides[1] * index1 + yStrides[2] * index2 + yStrides[3] * index3 + yStrides[4] * index4);
+                                        auto rZ = z + (zStrides[0] * index0 + zStrides[1] * index1 + zStrides[2] * index2 + zStrides[3] * index3 + zStrides[4] * index4);
+
+                                        *rZ = OpType::op(*rX, *rY);
+                                    }
+                                }
+                            }
+                    }
+
                 }
                 else if(shape::haveSameShapeAndStrides(xTadShapeShapeInfo, yShapeInfo) && shape::haveSameShapeAndStrides(xTadShapeShapeInfo, zTadShapeInfo)) {
                     uint tadShapeShapeInfoCast[MAX_RANK];

libnd4j/include/loops/cpu/indexreduce.hpp

@@ -73,7 +73,7 @@ Nd4jLong IndexReduce<X, Y>::execScalar(void *vx, Nd4jLong *xShapeInfo, void *vex
         auto func = PRAGMA_THREADS_FOR {
             intermediatery[thread_id] = OpType::startingIndexValue(x);
 
-            for (auto i = start; i < stop; i += increment) {
+            for (auto i = start; i < stop; i++) {
                 IndexValue<X> curr(x[i], i);
                 intermediatery[thread_id] = OpType::update(intermediatery[thread_id], curr, extraParams);
             }
@@ -88,7 +88,7 @@ Nd4jLong IndexReduce<X, Y>::execScalar(void *vx, Nd4jLong *xShapeInfo, void *vex
         auto func = PRAGMA_THREADS_FOR {
             intermediatery[thread_id] = OpType::startingIndexValue(x);
 
-            for (auto i = start; i < stop; i += increment) {
+            for (auto i = start; i < stop; i++) {
                 auto offset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
                 IndexValue<X> curr(x[offset], i);
                 intermediatery[thread_id] = OpType::update(intermediatery[thread_id], curr, extraParams);

libnd4j/include/loops/cpu/random.hpp

@@ -75,7 +75,7 @@ namespace functions {
 
                     auto func = PRAGMA_THREADS_FOR {
                         PRAGMA_OMP_SIMD
-                        for (auto i = start; i < stop; i += increment)  {
+                        for (auto i = start; i < stop; i++)  {
                             auto offset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
                             z[offset] = OpClass::op(x[offset], y[offset], i, length, rng, extraArguments);
                         }
@@ -93,7 +93,7 @@ namespace functions {
 
                 auto func = PRAGMA_THREADS_FOR {
                     PRAGMA_OMP_SIMD
-                    for (uint64_t i = start; i < stop; i += increment)  {
+                    for (uint64_t i = start; i < stop; i++)  {
                         auto offset  = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
                         auto zOffset = shape::indexOffset(i, zShapeInfo, zShapeInfoCast, canCastZ);
                         z[zOffset] = OpClass::op(x[offset], y[offset], i, length, rng, extraArguments);
@@ -111,7 +111,7 @@ namespace functions {
 
                 auto func = PRAGMA_THREADS_FOR {
                     PRAGMA_OMP_SIMD
-                    for (uint64_t i = start; i < stop; i += increment)  {
+                    for (uint64_t i = start; i < stop; i++)  {
                         auto offset  = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
                         auto yOffset = shape::indexOffset(i, yShapeInfo, yShapeInfoCast, canCastY);
                         z[offset] = OpClass::op(x[offset], y[yOffset], i, length, rng, extraArguments);
@@ -129,7 +129,7 @@ namespace functions {
 
                 auto func = PRAGMA_THREADS_FOR {
                     PRAGMA_OMP_SIMD
-                    for (uint64_t i = start; i < stop; i += increment)  {
+                    for (uint64_t i = start; i < stop; i++)  {
                         auto xOffset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
                         auto offset  = shape::indexOffset(i, yShapeInfo, yShapeInfoCast, canCastY);
                         z[offset] = OpClass::op(x[xOffset], y[offset], i, length, rng, extraArguments);
@@ -149,7 +149,7 @@ namespace functions {
 
                 auto func = PRAGMA_THREADS_FOR {
                     PRAGMA_OMP_SIMD
-                    for (uint64_t i = start; i < stop; i += increment)  {
+                    for (uint64_t i = start; i < stop; i++)  {
                         auto xOffset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
                         auto yOffset = shape::indexOffset(i, yShapeInfo, yShapeInfoCast, canCastY);
                         auto zOffset = shape::indexOffset(i, zShapeInfo, zShapeInfoCast, canCastZ);
@@ -197,7 +197,7 @@ namespace functions {
                 else{
                     auto func = PRAGMA_THREADS_FOR {
                         PRAGMA_OMP_SIMD
-                        for (uint64_t i = start; i < stop; i += increment)  {
+                        for (uint64_t i = start; i < stop; i++)  {
                             auto offset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
                             z[offset] = OpClass::op(x[offset], i, length, rng, extraArguments);
                         }
@@ -213,7 +213,7 @@ namespace functions {
 
                 auto func = PRAGMA_THREADS_FOR {
                     PRAGMA_OMP_SIMD
-                    for (uint64_t i = start; i < stop; i += increment)  {
+                    for (uint64_t i = start; i < stop; i++)  {
                         auto xOffset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
                         auto zOffset = shape::indexOffset(i, zShapeInfo, zShapeInfoCast, canCastZ);
                         z[zOffset] = OpClass::op(x[xOffset], i, length, rng, extraArguments);
@@ -255,7 +255,7 @@ namespace functions {
 
                 auto func = PRAGMA_THREADS_FOR {
                     PRAGMA_OMP_SIMD
-                    for (uint64_t i = start; i < stop; i += increment)  {
+                    for (uint64_t i = start; i < stop; i++)  {
                         auto offset = shape::indexOffset(i, zShapeInfo, zShapeInfoCast, canCastZ);
                         z[offset] = OpClass::op(i, length, rng, extraArguments);
                     }

libnd4j/include/loops/cpu/reduce3.hpp

@@ -88,7 +88,7 @@ void Reduce3<X,Z>::execScalar(void *vx, Nd4jLong *xShapeInfo,
 
     if (kindOfLoop == nd4j::LoopKind::EWS1) {
         auto func = PRAGMA_THREADS_FOR {
-            for (auto i = start; i < stop; i += increment) {
+            for (auto i = start; i < stop; i++) {
                 intermediate[thread_id] = OpType::update(intermediate[thread_id], OpType::op(x[i], y[i], extraParamsLocal + 3 * thread_id), extraParamsLocal + 3 * thread_id);
             }
         };
@@ -98,7 +98,7 @@ void Reduce3<X,Z>::execScalar(void *vx, Nd4jLong *xShapeInfo,
     } else if(shape::haveSameShapeAndStrides(xShapeInfo, yShapeInfo)) {
 
         auto func = PRAGMA_THREADS_FOR {
-            for (auto i = start; i < stop; i += increment) {
+            for (auto i = start; i < stop; i++) {
                 auto offset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
                 intermediate[thread_id] = OpType::update(intermediate[thread_id], OpType::op(x[offset], y[offset], extraParamsLocal + 3 * thread_id), extraParamsLocal + 3 * thread_id);
             }
@@ -110,7 +110,7 @@ void Reduce3<X,Z>::execScalar(void *vx, Nd4jLong *xShapeInfo,
         const bool canCastY = nd4j::DataTypeUtils::castShapeInfo(yShapeInfo, yShapeInfoCast);
 
         auto func = PRAGMA_THREADS_FOR {
-            for (auto i = start; i < stop; i += increment) {
+            for (auto i = start; i < stop; i++) {
                 auto xOffset = shape::indexOffset(i, xShapeInfo, xShapeInfoCast, canCastX);
                 auto yOffset = shape::indexOffset(i, yShapeInfo, yShapeInfoCast, canCastY);
                 intermediate[thread_id] = OpType::update(intermediate[thread_id], OpType::op(x[xOffset], y[yOffset], extraParamsLocal + 3 * thread_id), extraParamsLocal + 3 * thread_id);

libnd4j/include/loops/cpu/summarystatsreduce.cpp

@@ -158,7 +158,7 @@ namespace functions {
             const bool canCast = tadEWS == 1 && tadOrder == 'c' ? false : nd4j::DataTypeUtils::castShapeInfo<uint>(tadShapeShapeInfo, tadShapeShapeInfoCast);
 
             auto func = PRAGMA_THREADS_FOR {
-                for (auto r = start; r < stop; r += increment) {
+                for (auto r = start; r < stop; r++) {
 
                     auto tadOffsetForBlock = tadPack.primaryOffsets()[r];
                     auto tx = x + tadOffsetForBlock;

libnd4j/include/loops/cuda/transform/transform_any.cu

@@ -84,7 +84,7 @@ namespace functions {
 	    	auto tid = blockIdx.x * blockDim.x + threadIdx.x;
 			int totalThreads = gridDim.x * blockDim.x;
 
-		    if(xEws > 0 && zEws > 0 && xOrder == zOrder) {
+		    if(xEws > 0 && zEws > 0 && xOrder == zOrder && xOrder == 'c') {
 
 				for (int i = tid; i < length; i += totalThreads)
 					z[i * zEws] = OpType::op(x[i * xEws], params);

libnd4j/include/loops/cuda/transform/transform_bool.cu

@@ -89,7 +89,7 @@ namespace functions {
 	    	    auto tid = blockIdx.x * blockDim.x + threadIdx.x;
 				int totalThreads = gridDim.x * blockDim.x;
 
-		        if(xEws > 0 && zEws > 0 && xOrder == zOrder) {
+		        if(xEws > 0 && zEws > 0 && xOrder == zOrder && xOrder == 'c') {
 
 					for (int i = tid; i < length; i += totalThreads)
 						z[i * zEws] = OpType::op(x[i * xEws], params);

libnd4j/include/loops/cuda/transform/transform_float.cu

@@ -97,7 +97,7 @@ namespace functions {
 	    	    auto tid = blockIdx.x * blockDim.x + threadIdx.x;
 				int totalThreads = gridDim.x * blockDim.x;
 
-		        if(xEws > 0 && zEws > 0 && xOrder == zOrder) {
+		        if(xEws > 0 && zEws > 0 && xOrder == zOrder && xOrder == 'c') {
 
 					for (Nd4jLong i = tid; i < length; i += totalThreads)
                         z[i * zEws] = OpType::op(x[i * xEws], params);

libnd4j/include/loops/cuda/transform/transform_same.cu

@@ -87,7 +87,7 @@ namespace functions {
 	    	    auto tid = blockIdx.x * blockDim.x + threadIdx.x;
 				int totalThreads = gridDim.x * blockDim.x;
 
-		        if(xEws > 0 && zEws > 0 && xOrder == zOrder) {
+		        if(xEws > 0 && zEws > 0 && xOrder == zOrder && xOrder == 'c') {
 
 					for (int i = tid; i < length; i += totalThreads)
 						z[i * zEws] = OpType::op(x[i * xEws], params);

libnd4j/include/loops/cuda/transform/transform_strict.cu

@@ -89,7 +89,7 @@ namespace functions {
 	    	    auto tid = blockIdx.x * blockDim.x + threadIdx.x;
 				int totalThreads = gridDim.x * blockDim.x;
 
-		        if(xEws > 0 && zEws > 0 && xOrder == zOrder) {
+		        if(xEws > 0 && zEws > 0 && xOrder == zOrder && xOrder == 'c') {
 
 					for (int i = tid; i < length; i += totalThreads)
 						z[i * zEws] = OpType::op(x[i * xEws], params);

libnd4j/include/loops/impl/type_conversions.cpp

@@ -81,7 +81,7 @@ namespace nd4j {
 
         // now we actually apply quantization
         auto func = PRAGMA_THREADS_FOR {
-            for (auto e = start; e < stop; e += increment) {
+            for (auto e = start; e < stop; e++) {
                 rz[e] = static_cast<char>(nd4j::math::nd4j_round<float, char>( 1.0f * static_cast<float>(x[e]) / nd4j::math::nd4j_max<float>(amax, amin) * max_byte));
             }
         };
@@ -177,7 +177,7 @@ PRAGMA_OMP_ATOMIC_ARGS(write)
         int flimit = limit + 4;
 
         auto func = PRAGMA_THREADS_FOR {
-            for (auto e = start; e < stop; e += increment) {
+            for (auto e = start; e < stop; e++) {
                 int el = x[e];
                 int ael = nd4j::math::nd4j_abs<int>(el) - 1;
                 z[ael] += el > 0 ? static_cast<T>(threshold) : static_cast<T>(-threshold);
@@ -202,7 +202,7 @@ PRAGMA_OMP_ATOMIC_ARGS(write)
         auto z = reinterpret_cast<T *>(dz);
 
         auto func = PRAGMA_THREADS_FOR {
-            for (auto i = start; i < stop; i += increment) {
+            for (auto i = start; i < stop; i++) {
                 z[i] = static_cast<T>(static_cast<float>(x[i]));
             }
         };

libnd4j/include/ops/declarable/OpDescriptor.h

@@ -147,6 +147,9 @@ namespace nd4j {
             // returns TRUE if this op allows in-place execution
             bool allowsInplace();
 
+            // this method allows you to enable/disable inplace call for a given op
+            void allowInplace(bool reallyAllow);
+
             // this method returns opNum (applicable for legacy XYZ ops only)
             int getOpNum();
 

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

@@ -27,13 +27,11 @@ namespace nd4j {
     namespace ops {
         OP_IMPL(identity, 1, 1, true) {
             auto first = INPUT_VARIABLE(0);
-            auto z = this->getZ(block);
+            auto z = OUTPUT_VARIABLE(0);
 
-            // just for lulz
+            if (!block.isInplace())
                 first->applyTransform(nd4j::transform::Identity, *z);
 
-            STORE_RESULT(*z);
-
             return Status::OK();
         }
         DECLARE_SYN(linear, identity);
@@ -60,8 +58,8 @@ namespace nd4j {
         DECLARE_TYPES(identity_bp) {
             getOpDescriptor()
                     ->setAllowedInputTypes(0, DataType::ANY)
-                    ->setAllowedInputTypes(1, {DataType::FLOAT32, DataType ::DOUBLE, DataType::HALF})
+                    ->setAllowedInputTypes(1, {ALL_FLOATS})
-                    ->setAllowedOutputTypes(0, {DataType::FLOAT32, DataType ::DOUBLE, DataType::HALF});
+                    ->setAllowedOutputTypes(0, {ALL_FLOATS});
         }
     }
 }

libnd4j/include/ops/declarable/generic/blas/matmul.cpp

@@ -29,7 +29,9 @@
 namespace nd4j {
     namespace ops {
 
-        CUSTOM_OP_IMPL(matmul, 2, 1, false, 0, -2) {
+//////////////////////////////////////////////////////////////////////
+CUSTOM_OP_IMPL(matmul, 2, 1, false, 0, -2) {
+
     auto x = INPUT_VARIABLE(0);
     auto y = INPUT_VARIABLE(1);
     auto z = OUTPUT_VARIABLE(0);
@@ -57,58 +59,41 @@ namespace nd4j {
     const int yLastButOneDim = transY ? -1 : -2;
 
     // ******* input validation ******* //
-            REQUIRE_TRUE(xRank > 0 && yRank > 0, 0,
+    REQUIRE_TRUE(xRank > 0 && yRank > 0, 0, "MATMUL OP: input arrays must have rank bigger than 0 (should not be scalars), but got instead: x rank = %i, y rank = %i !", xRank, yRank);
-                         "MATMUL OP: input arrays must have rank bigger than 0 (should not be scalars), but got instead: x rank = %i, y rank = %i !",
-                         xRank, yRank);
 
     if (xRank == 1 && yRank == 1) {  // dot case, output is scalar (or vector with length = 1)
-                REQUIRE_TRUE(x->lengthOf() == y->lengthOf(), 0,
+        REQUIRE_TRUE(x->lengthOf() == y->lengthOf(), 0, "MATMUL OP: since input arrays are vectors they must have the same length, but got x length = %i, y length = %i !", x->lengthOf(), y->lengthOf());
-                             "MATMUL OP: since input arrays are vectors they must have the same length, but got x length = %i, y length = %i !",
-                             x->lengthOf(), y->lengthOf());
     } else if (xRank == 1 && yRank == 2) {  // vector x matrix, i.e. [4] x [4,5] = [5], output is vector
-                REQUIRE_TRUE(x->lengthOf() == y->sizeAt(yLastButOneDim), 0,
+        REQUIRE_TRUE(x->lengthOf() == y->sizeAt(yLastButOneDim), 0, "MATMUL OP: input arrays have inconsistent shapes for vector-matrix product: x %s, y %s !", ShapeUtils::shapeAsString(x).c_str(), ShapeUtils::shapeAsString(y).c_str());
-                             "MATMUL OP: input arrays have inconsistent shapes for vector-matrix product: x %s, y %s !",
-                             ShapeUtils::shapeAsString(x).c_str(), ShapeUtils::shapeAsString(y).c_str());
     } else if (xRank == 2 && yRank == 1) {   // matrix x vector , i.e. [4,5] x [5] = [4], output is vector
-                REQUIRE_TRUE(x->sizeAt(xLastDim) == y->lengthOf(), 0,
+        REQUIRE_TRUE(x->sizeAt(xLastDim) == y->lengthOf(), 0, "MATMUL OP: input arrays have inconsistent shapes for matrix-vector product: x %s, y %s !", ShapeUtils::shapeAsString(x).c_str(), ShapeUtils::shapeAsString(y).c_str());
-                             "MATMUL OP: input arrays have inconsistent shapes for matrix-vector product: x %s, y %s !",
-                             ShapeUtils::shapeAsString(x).c_str(), ShapeUtils::shapeAsString(y).c_str());
     } else {
-                REQUIRE_TRUE(xRank == yRank && yRank == zRank, 0,
+        REQUIRE_TRUE(xRank == yRank && yRank == zRank, 0, "MATMUL OP: input and output arrays must have the same rank, but got instead: x rank = %i, y rank = %i, z rank = %i !", xRank, yRank, zRank);
-                             "MATMUL OP: input and output arrays must have the same rank, but got instead: x rank = %i, y rank = %i, z rank = %i !",
+        REQUIRE_TRUE(x->sizeAt(xLastDim) == y->sizeAt(yLastButOneDim) && x->sizeAt(xLastButOneDim) == z->sizeAt(-2) && y->sizeAt(yLastDim) == z->sizeAt(-1), 0, "MATMUL OP: input/output arrays have inconsistent shapes for matrix product: x %s, y %s, z %s !", ShapeUtils::shapeAsString(x).c_str(), ShapeUtils::shapeAsString(y).c_str(), ShapeUtils::shapeAsString(z).c_str());
-                             xRank, yRank, zRank);
-                REQUIRE_TRUE(x->sizeAt(xLastDim) == y->sizeAt(yLastButOneDim) &&
-                             x->sizeAt(xLastButOneDim) == z->sizeAt(-2) && y->sizeAt(yLastDim) == z->sizeAt(-1), 0,
-                             "MATMUL OP: input/output arrays have inconsistent shapes for matrix product: x %s, y %s, z %s !",
-                             ShapeUtils::shapeAsString(x).c_str(), ShapeUtils::shapeAsString(y).c_str(),
-                             ShapeUtils::shapeAsString(z).c_str());
 
         if (xRank > 2)   // outer dims must be the same
             for (int i = 0; i < xRank - 2; ++i)
-                        REQUIRE_TRUE(x->sizeAt(i) == y->sizeAt(i) && y->sizeAt(i) == z->sizeAt(i), 0,
+                REQUIRE_TRUE(x->sizeAt(i) == y->sizeAt(i) && y->sizeAt(i) == z->sizeAt(i), 0, "MATMUL OP: input/output arrays have inconsistent shapes for matrix product: x %s, y %s, z %s !", ShapeUtils::shapeAsString(x).c_str(), ShapeUtils::shapeAsString(y).c_str(), ShapeUtils::shapeAsString(z).c_str());
-                                     "MATMUL OP: input/output arrays have inconsistent shapes for matrix product: x %s, y %s, z %s !",
-                                     ShapeUtils::shapeAsString(x).c_str(), ShapeUtils::shapeAsString(y).c_str(),
-                                     ShapeUtils::shapeAsString(z).c_str());
     }
     // ******* end of input validation ******* //
 
     MmulHelper::matmul(x, y, z, transX, transY);
 
     return Status::OK();
-        }
+}
 
-        DECLARE_SYN(mMul, matmul);
+DECLARE_SYN(mMul, matmul);
 
-        DECLARE_SYN(mmul, matmul);
+DECLARE_SYN(mmul, matmul);
 
-        DECLARE_SYN(gemm, matmul);
+DECLARE_SYN(gemm, matmul);
 
-        DECLARE_SYN(gemv, matmul);
+DECLARE_SYN(gemv, matmul);
 
-        DECLARE_SYN(dot, matmul);
+DECLARE_SYN(dot, matmul);
 
-
+//////////////////////////////////////////////////////////////////////
-        DECLARE_SHAPE_FN(matmul) {
+DECLARE_SHAPE_FN(matmul) {
 
     auto xShapeInfo = inputShape->at(0);
     auto yShapeInfo = inputShape->at(1);
@@ -144,17 +129,18 @@ namespace nd4j {
 
     auto newShape = ConstantShapeHelper::getInstance()->createShapeInfo(dtypeZ, zOrder, zShapeOnly);
     return SHAPELIST(newShape);
-        }
+}
 
-        DECLARE_TYPES(matmul) {
+//////////////////////////////////////////////////////////////////////
+DECLARE_TYPES(matmul) {
     getOpDescriptor()
             ->setAllowedInputTypes(0, {ALL_FLOATS})
             ->setAllowedInputTypes(1, {ALL_FLOATS})
             ->setAllowedOutputTypes(0, {ALL_FLOATS});
-        }
+}
 
-
+//////////////////////////////////////////////////////////////////////
-        CUSTOM_OP_IMPL(matmul_bp, 3, 2, false, 0, -2) {
+CUSTOM_OP_IMPL(matmul_bp, 3, 2, false, 0, -2) {
     auto x = INPUT_VARIABLE(0);
     auto y = INPUT_VARIABLE(1);
     auto eps = INPUT_VARIABLE(2);
@@ -182,10 +168,10 @@ F   F   T   [a,b]   [b,c]   [c,a]   [c,a]
     op.execute({x, eps}, {dldy}, {}, {!transX, transZ, transY}, {});
 
     return Status::OK();
-        }
+}
 
-
+//////////////////////////////////////////////////////////////////////
-        DECLARE_SHAPE_FN(matmul_bp) {
+DECLARE_SHAPE_FN(matmul_bp) {
     Nd4jLong *xShapeInfo;
     Nd4jLong *yShapeInfo;
 
@@ -193,18 +179,19 @@ F   F   T   [a,b]   [b,c]   [c,a]   [c,a]
     COPY_SHAPE(inputShape->at(1), yShapeInfo);
 
     return SHAPELIST(CONSTANT(xShapeInfo), CONSTANT(yShapeInfo));
-        }
+}
 
-        DECLARE_TYPES(matmul_bp) {
+//////////////////////////////////////////////////////////////////////
+DECLARE_TYPES(matmul_bp) {
     getOpDescriptor()
             ->setAllowedInputTypes(0, {ALL_FLOATS})
             ->setAllowedInputTypes(1, {ALL_FLOATS})
             ->setAllowedInputTypes(2, {ALL_FLOATS})
             ->setAllowedOutputTypes(0, {ALL_FLOATS})
             ->setAllowedOutputTypes(1, {ALL_FLOATS});
-        }
+}
 
-    }
+}
 }
 
 

libnd4j/include/ops/declarable/generic/blas/tensormmul.cpp

@@ -21,17 +21,22 @@
 #include <op_boilerplate.h>
 #if NOT_EXCLUDED(OP_tensormmul)
 
+#include <numeric>
 #include <helpers/ShapeUtils.h>
 #include <ops/declarable/CustomOperations.h>
 #include <MmulHelper.h>
 
+
 namespace nd4j {
-    namespace ops {
+namespace ops  {
-        CUSTOM_OP_IMPL(tensormmul, 2, 1, false, 0, -1) {
+
+////////////////////////////////////////////////////////////////////////
+CUSTOM_OP_IMPL(tensormmul, 2, 1, false, 0, -1) {
+
     auto a = INPUT_VARIABLE(0);
     auto b = INPUT_VARIABLE(1);
 
-            auto c = OUTPUT_VARIABLE(0); //
+    auto c = OUTPUT_VARIABLE(0);
 
     REQUIRE_TRUE(a->dataType() == b->dataType(), 0, "tensormmul: A, B and C data types must be the same");
 
@@ -40,20 +45,20 @@ namespace nd4j {
     int axe1_size = INT_ARG(axe0_size+1);
     std::vector<int> axes_0(axe0_size), axes_1(axe1_size);
     for (int e = 0; e < axe0_size; e++)
-                axes_0[e] = (int) INT_ARG(e+1);
+        axes_0[e] = (int)INT_ARG(e + 1);
 
     for (int e = 0; e < axe1_size; e++)
-                axes_1[e] = (int) INT_ARG(e + axe0_size + 2);
+        axes_1[e] = (int)INT_ARG(e + axe0_size + 2);
 
     nd4j_verbose("axe0: %i; axe1: %i;\n", axes_0.size(), axes_1.size());
 
     MmulHelper::tensorDot(a, b, c, axes_0, axes_1);
     return Status::OK();
-        }
+}
-        DECLARE_SYN(tensordot, tensormmul);
+DECLARE_SYN(tensordot, tensormmul);
 
-
+////////////////////////////////////////////////////////////////////////
-        DECLARE_SHAPE_FN(tensormmul) {               
+DECLARE_SHAPE_FN(tensormmul) {
 
     auto aShapeInfo = inputShape->at(0);
     auto bShapeInfo = inputShape->at(1);
@@ -76,15 +81,114 @@ namespace nd4j {
     auto outShape = nd4j::ShapeUtils::evalShapeForTensorDot(aShapeInfo, bShapeInfo, axes_0, axes_1, permutAt, permutBt, shapeAt, shapeBt);
 
     return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(ArrayOptions::dataType(aShapeInfo), 'c', outShape)));
-        }
+}
 
-        DECLARE_TYPES(tensormmul) {
+////////////////////////////////////////////////////////////////////////
+DECLARE_TYPES(tensormmul) {
     getOpDescriptor()
             ->setAllowedInputTypes(0, {DataType::FLOAT32, DataType ::DOUBLE, DataType::HALF})
             ->setAllowedInputTypes(1, {DataType::FLOAT32, DataType ::DOUBLE, DataType::HALF})
             ->setAllowedOutputTypes(0, {DataType::FLOAT32, DataType ::DOUBLE, DataType::HALF});
+}
+
+////////////////////////////////////////////////////////////////////////
+CUSTOM_OP_IMPL(tensormmul_bp, 3, 2, false, 0, -1) {
+
+    auto A = INPUT_VARIABLE(0);
+    auto B = INPUT_VARIABLE(1);
+
+    auto dLdC = INPUT_VARIABLE(2);
+
+    auto dLdA = OUTPUT_VARIABLE(0);
+    auto dLdB = OUTPUT_VARIABLE(1);
+
+    REQUIRE_TRUE( (A->dataType() == B->dataType() && (dLdC->dataType() == A->dataType())), 0, "tensormmul_bp: A, B and dLdC data types must be the same");
+
+    int axe0Size = INT_ARG(0);
+    int axe1Size = INT_ARG(axe0Size + 1);
+
+    auto Arank = A->rankOf();
+    auto Brank = B->rankOf();
+    auto dLdCrank = dLdC->rankOf();
+
+    REQUIRE_TRUE((Arank >= axe0Size), 0, "tensormmul_bp: A rank must be the higher or same as input axes 0");
+
+    REQUIRE_TRUE((Brank >= axe1Size), 0, "tensormmul_bp: B rank must be the higher or same as input axes 1");
+
+    // building axes
+    std::vector<int> axes0(axe0Size), axes1(axe1Size);
+    for (uint e = 0; e < axe0Size; e++)
+        axes0[e] = (int)INT_ARG(e + 1);
+    for (uint e = 0; e < axe1Size; e++)
+        axes1[e] = (int)INT_ARG(e + axe0Size + 2);
+
+    std::vector<int> permutAt, permutBt;
+    std::vector<Nd4jLong> shapeAt, shapeBt;
+
+    ShapeUtils::evalShapeForTensorDot(A, B, axes0, axes1, permutAt, permutBt, shapeAt, shapeBt);
+
+    // special case for scalar value
+    if (dLdC->isScalar()) {
+
+        dLdA->assign((*dLdC) * *B);
+        dLdB->assign((*dLdC) * *A);
+
+        return Status::OK();
     }
+
+    std::vector<int> axesA = ShapeUtils::evalDimsToExclude(Arank, axes0);
+    std::vector<int> axesB = ShapeUtils::evalDimsToExclude(Brank, axes1);
+
+    // rank always have to be divided by 2
+    std::vector<int> axesAdLdC, axesBdLdC;
+    if (dLdCrank > 1) {
+        axesAdLdC.resize(dLdCrank / 2);
+        std::iota(axesAdLdC.begin(), axesAdLdC.end(), 0);
+        axesBdLdC = ShapeUtils::evalDimsToExclude(dLdCrank, axesAdLdC);
     }
+    else {
+        axesAdLdC.push_back(0);
+        axesBdLdC.push_back(0);
+    }
+
+    // calculate dLdA
+    MmulHelper::tensorDot(dLdC, B, dLdA, axesBdLdC, axesB, permutAt);
+
+    // calculate dLdB
+    MmulHelper::tensorDot(A, dLdC, dLdB, axesA, axesAdLdC, permutBt);
+
+    return Status::OK();
+}
+
+////////////////////////////////////////////////////////////////////////
+DECLARE_SHAPE_FN(tensormmul_bp) {
+
+    auto aShapeInfo = inputShape->at(0);
+    auto bShapeInfo = inputShape->at(1);
+    auto dLShapeInfo = inputShape->at(2);
+
+    REQUIRE_TRUE((ArrayOptions::dataType(aShapeInfo) == ArrayOptions::dataType(bShapeInfo) &&
+                 (ArrayOptions::dataType(dLShapeInfo) == ArrayOptions::dataType(aShapeInfo))), 0, "tensormmul_bp: A, B and dLdC data types must be the same");
+
+    Nd4jLong* dLdAShapeInfo = nullptr;
+    Nd4jLong* dLdBShapeInfo = nullptr;
+
+    COPY_SHAPE(aShapeInfo, dLdAShapeInfo);
+    COPY_SHAPE(bShapeInfo, dLdBShapeInfo);
+
+    return SHAPELIST(CONSTANT(dLdAShapeInfo), CONSTANT(dLdBShapeInfo));
+}
+
+////////////////////////////////////////////////////////////////////////
+DECLARE_TYPES(tensormmul_bp) {
+    getOpDescriptor()
+        ->setAllowedInputTypes(0, { DataType::FLOAT32, DataType::DOUBLE, DataType::HALF }) // maybe better ALL_FLOATS
+        ->setAllowedInputTypes(1, { DataType::FLOAT32, DataType::DOUBLE, DataType::HALF })
+        ->setAllowedInputTypes(2, { DataType::FLOAT32, DataType::DOUBLE, DataType::HALF })
+        ->setAllowedOutputTypes(0, { DataType::FLOAT32, DataType::DOUBLE, DataType::HALF })
+        ->setAllowedOutputTypes(1, { DataType::FLOAT32, DataType::DOUBLE, DataType::HALF });
+}
+}
 }
 
 #endif

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

@@ -79,7 +79,7 @@ CUSTOM_OP_IMPL(conv1d, 2, 1, false, 0, 5) {
     }
 
     auto inputReshaped   = input  ->reshape(input->ordering(),   reshapeForInput);
-    auto outputReshaped  = output ->reshape(output->ordering(),  reshapeForOutput);
+    auto outputReshaped  = output ->reshape(output->ordering(),  reshapeForOutput, false);
     auto weightsReshaped = weights->reshape(weights->ordering(), {1, weights->sizeAt(0), weights->sizeAt(1), weights->sizeAt(2)});   // [kW, iC, oC] -> [1, kW, iC, oC]
 
     nd4j::ops::conv2d conv2d;
@@ -216,10 +216,10 @@ CUSTOM_OP_IMPL(conv1d_bp, 3, 2, false, 0, 5) {
     }
 
     auto inputReshaped   = input  ->reshape(input->ordering(),  reshapeForInput);
-    auto gradIReshaped   = gradI  ->reshape(gradI->ordering(),  reshapeForInput);
+    auto gradIReshaped   = gradI  ->reshape(gradI->ordering(),  reshapeForInput, false);
     auto gradOReshaped   = gradO  ->reshape(gradO->ordering(),  reshapeForGradO);
     auto weightsReshaped = weights->reshape(weights->ordering(),{1, weights->sizeAt(0), weights->sizeAt(1), weights->sizeAt(2)});       // [kW, iC, oC] -> [1, kW, iC, oC]
-    auto gradWReshaped   = gradW  ->reshape(gradW->ordering(),  {1, weights->sizeAt(0), weights->sizeAt(1), weights->sizeAt(2)});    // [kW, iC, oC] -> [1, kW, iC, oC]
+    auto gradWReshaped   = gradW  ->reshape(gradW->ordering(),  {1, weights->sizeAt(0), weights->sizeAt(1), weights->sizeAt(2)}, false);// [kW, iC, oC] -> [1, kW, iC, oC]
 
     nd4j::ops::conv2d_bp conv2dBP;
     auto status = conv2dBP.execute({&inputReshaped, &weightsReshaped, bias, &gradOReshaped}, {&gradIReshaped, &gradWReshaped, gradB}, {}, {1,kW,  1,sW,  0,pW,  1,dW,  paddingMode,  !isNCW}, {});

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

@@ -239,7 +239,7 @@ CUSTOM_OP_IMPL(conv3dnew_bp, 3, 2, false, 0, 13) {
     //----- calculation of gradO -----//
     if(gradB) {
         if(gradB->rankOf() == 2)
-            gradB = new NDArray(gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()}));
+            gradB = new NDArray(gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()}, false));
         gradO->reduceAlongDimension(reduce::Sum, *gradB, gradOaxesForDot);                          // sum over bS oD oH oW
         if(gradB != OUTPUT_VARIABLE(2))
             delete gradB;

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

@@ -233,7 +233,7 @@ CUSTOM_OP_IMPL(deconv2d_bp, 3, 2, false, 0, 9) {
     // ----- calculation of gradB ----- //
     if(gradB) {
         if(gradB->rankOf() == 2)
-            gradB = new NDArray(gradB->reshape(gradB->ordering(), {gradB->lengthOf()}));
+            gradB = new NDArray(gradB->reshape(gradB->ordering(), {gradB->lengthOf()}, false));
         gradO->reduceAlongDimension(reduce::Sum, *gradB, {0, 2, 3});                                // sum over bS, oH, oW
         if(gradB != OUTPUT_VARIABLE(2))
             delete gradB;

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

@@ -243,7 +243,7 @@ CUSTOM_OP_IMPL(deconv3d_bp, 3, 2, false, 0, 13) {
     // ----- calculation of gradB ----- //
     if(gradB) {
         if(gradB->rankOf() == 2)
-            gradB = new NDArray(gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()}));
+            gradB = new NDArray(gradB->reshape(gradB->ordering(), {(int)gradB->lengthOf()}, false));
         gradO->reduceAlongDimension(reduce::Sum, *gradB, {0, 2, 3, 4});                                // sum over bS, oD, oH, oW
         if(gradB != OUTPUT_VARIABLE(2))
             delete gradB;

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

@@ -31,22 +31,17 @@ namespace nd4j {
             REQUIRE_TRUE(w->isMatrix(), 0, "relu_layer: weights argument should be a 2D tensor, but got rank %i instead!", w->rankOf());
             REQUIRE_TRUE(b->isVector(), 0, "relu_layer: biases argument should be a 1D tensor, but got rank %i instead!", b->rankOf());
             REQUIRE_TRUE(b->lengthOf() == w->sizeAt(1), 0, "relu_layer: biases array length should match to columns of weights matrix, however got length = %i and columns = %i!", b->lengthOf(), w->sizeAt(1));
-            REQUIRE_TRUE(x->sizeAt(1) == w->sizeAt(0), 0, "relu_layer: number of x columns should match to row number of weights matrix, but got x_columns = %i and weights_rows = %i!",
+            REQUIRE_TRUE(x->sizeAt(1) == w->sizeAt(0), 0, "relu_layer: number of x columns should match to row number of weights matrix, but got x_columns = %i and weights_rows = %i!", x->sizeAt(1), w->sizeAt(0));
-                x->sizeAt(1), w->sizeAt(0));
-
 
             auto output = OUTPUT_VARIABLE(0);
-            //T bound = (T)0.f;
-            //nd4j_printf("Matrix x(%ix%i), Matrix w(%ix%i), b(1x%i)\n", x->sizeAt(0), x->sizeAt(1), w->sizeAt(0), w->sizeAt(1), b->lengthOf());
 
             nd4j::ops::xw_plus_b op;
-            std::unique_ptr<ResultSet> result(op.evaluate({x, w, b}));
+            auto status = op.execute({x, w, b}, {output});
-            REQUIRE_TRUE(Status::OK() == result->status(), 0, "relu_layer: xw_plus_b op failed on input data.");
+            REQUIRE_TRUE(Status::OK() == status, 0, "relu_layer: xw_plus_b op failed on input data.");
 
             auto scalar = block.numT() > 0 ? block.getTArguments()->at(0) : 0.0;
 
-            auto xw = result->at(0);
+            output->applyScalar(nd4j::scalar::RELU, scalar, *output);
-            xw->applyScalar(nd4j::scalar::RELU, scalar, *output);
 
             return Status::OK();
         }

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

@@ -23,7 +23,8 @@
 
 //#include <ops/declarable/headers/parity_ops.h>
 #include <ops/declarable/CustomOperations.h>
-#include <ops/declarable/helpers/image_resize.h>
+#include <ops/declarable/helpers/crop_and_resize.h>
+
 namespace nd4j {
     namespace ops {
         CUSTOM_OP_IMPL(crop_and_resize, 4, 1, false, 0, 0) {

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

@@ -61,7 +61,7 @@ namespace nd4j {
             }
 
             auto source = inRank == 4?image->reshape(image->ordering(), {image->sizeAt(0), image->sizeAt(1), image->sizeAt(2), image->sizeAt(3)}):image->reshape(image->ordering(), {1, image->sizeAt(0), image->sizeAt(1), image->sizeAt(2)});
-            auto target = inRank == 4?output->reshape(output->ordering(), {output->sizeAt(0), output->sizeAt(1), output->sizeAt(2), output->sizeAt(3)}):output->reshape(output->ordering(), {1, output->sizeAt(0), output->sizeAt(1), output->sizeAt(2)});
+            auto target = inRank == 4?output->reshape(output->ordering(), {output->sizeAt(0), output->sizeAt(1), output->sizeAt(2), output->sizeAt(3)}, false) : output->reshape(output->ordering(), {1, output->sizeAt(0), output->sizeAt(1), output->sizeAt(2)}, false);
 
             return helpers::resizeAreaFunctor(block.launchContext(), &source, width, height, alignCorners, &target);
         }

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

@@ -62,7 +62,7 @@ namespace nd4j {
             REQUIRE_TRUE(!halfPixelAlign || (halfPixelAlign && !alignCorners), 0, "resize_bicubic: `half_pixel_centers' should be false or true only when `align_corners' is false");
 
             auto source = inRank == 4?image->reshape(image->ordering(), {image->sizeAt(0), image->sizeAt(1), image->sizeAt(2), image->sizeAt(3)}):image->reshape(image->ordering(), {1, image->sizeAt(0), image->sizeAt(1), image->sizeAt(2)});
-            auto target = inRank == 4?output->reshape(output->ordering(), {output->sizeAt(0), output->sizeAt(1), output->sizeAt(2), output->sizeAt(3)}):output->reshape(output->ordering(), {1, output->sizeAt(0), output->sizeAt(1), output->sizeAt(2)});
+            auto target = inRank == 4?output->reshape(output->ordering(), {output->sizeAt(0), output->sizeAt(1), output->sizeAt(2), output->sizeAt(3)}, false) : output->reshape(output->ordering(), {1, output->sizeAt(0), output->sizeAt(1), output->sizeAt(2)}, false);
 
             return helpers::resizeBicubicFunctorA(block.launchContext(), &source, width, height, alignCorners, halfPixelAlign, &target);
         }

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

@@ -43,7 +43,7 @@ namespace nd4j {
             REQUIRE_TRUE(inRank == output->rankOf(), 0, "resize_bilinear: Input and output ranks should be equals, but %i and %i occured.", inRank, output->rankOf());
 
             auto source = inRank == 4?image->reshape(image->ordering(), {image->sizeAt(0), image->sizeAt(1), image->sizeAt(2), image->sizeAt(3)}):image->reshape(image->ordering(), {1, image->sizeAt(0), image->sizeAt(1), image->sizeAt(2)});
-            auto target = inRank == 4?output->reshape(output->ordering(), {output->sizeAt(0), output->sizeAt(1), output->sizeAt(2), output->sizeAt(3)}):output->reshape(output->ordering(), {1, output->sizeAt(0), output->sizeAt(1), output->sizeAt(2)});
+            auto target = inRank == 4?output->reshape(output->ordering(), {output->sizeAt(0), output->sizeAt(1), output->sizeAt(2), output->sizeAt(3)}, false) : output->reshape(output->ordering(), {1, output->sizeAt(0), output->sizeAt(1), output->sizeAt(2)}, false);
 
             if (block.width() > 1) {
                 auto newImageSize = INPUT_VARIABLE(1);

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

@@ -63,7 +63,7 @@ namespace nd4j {
             REQUIRE_TRUE(((alignCorners && height > 2) || (height > 0)) && ((alignCorners && width > 1) || (width > 0)), 0,  "resize_nearest_neighbor: Wrong input or output size to resize (width = %d, height = %d)", width, height);
 
             auto source = inRank == 4?*image:image->reshape(image->ordering(), {1, image->sizeAt(0), image->sizeAt(1), image->sizeAt(2)});
-            auto target = inRank == 4?*output:output->reshape(output->ordering(), {1, output->sizeAt(0), output->sizeAt(1), output->sizeAt(2)});
+            auto target = inRank == 4 ? *output : output->reshape(output->ordering(), {1, output->sizeAt(0), output->sizeAt(1), output->sizeAt(2)}, false);
 
             return helpers::resizeNeighborFunctor(block.launchContext(), inRank==4?image:&source, width, height, alignCorners, halfPixelCenter, inRank == 4 ? output : &target);
         }

libnd4j/include/ops/declarable/generic/shape/expand_dims.cpp

@@ -47,11 +47,12 @@ namespace nd4j {
 
             shape.insert(shape.begin() + axis, 1);
 
+            if (input->ews() == 1 && output->ews() == 1 && input->ordering() == output->ordering()) {
+                output->dataBuffer()->copyBufferFrom(*input->dataBuffer().get(), output->lengthOf() * DataTypeUtils::sizeOfElement(output->dataType()), 0, input->bufferOffset());
+            } else {
                 auto tmp = input->reshape(input->ordering(), shape);
                 output->assign(tmp);
-
+            }
-            STORE_RESULT(output);
-
             return Status::OK();
         }
 

libnd4j/include/ops/declarable/generic/shape/permute.cpp

@@ -15,7 +15,8 @@
  ******************************************************************************/
 
 //
-// Created by raver119 on 29/10/17.
+// @author raver119@gmail.com
+// @author Yurii Shyrma (iuriish@yahoo.com)
 //
 
 #include <op_boilerplate.h>
@@ -29,80 +30,52 @@ namespace nd4j {
 
 //////////////////////////////////////////////////////////////////////////
 // here iArgs is int vector of ordered set of dimensions to be permuted
-        CUSTOM_OP_IMPL(permute, 1, 1, true, 0, -2) {
+CUSTOM_OP_IMPL(permute, 1, 1, true, 0, -2) {
+
     auto x = INPUT_VARIABLE(0);
+    auto z = OUTPUT_VARIABLE(0);
 
-            bool replace = false;
+    if (x->isEmpty()) {
-
+        REQUIRE_TRUE(z->isEmpty(), 0, "PERMUTE OP: when input is empty, output must also be empty");
-            auto origArgs = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
+        return Status::OK();    //No op
-            std::vector<int> arguments({});
-            if(origArgs.size() > 0){
-                for (int e = 0; e < origArgs.size(); e++) {
-                    int ax = origArgs[e];
-                    if (ax < 0)
-                        ax += x->rankOf();
-
-                    arguments.emplace_back(ax);
     }
 
-                replace = true;
+    if (block.width() == 1 && block.getIArguments()->size() == 0) {
-            } else {
+        z->assign(x->transpose());
-                for (int e = x->rankOf() - 1; e >= 0; e--)
-                    arguments.emplace_back(e);
-            }
-
-            // 0D edge case
-            if (x->rankOf() == 0) {
-                REQUIRE_TRUE(arguments.size() == 1, 0, "Permute: only one axis is allowed for scalar");
-                auto output = OUTPUT_VARIABLE(0);
-                if (!block.isInplace())
-                    output->assign(x);
-
         return Status::OK();
     }
 
-            if(block.isInplace()) {		// in-place
+    std::vector<int> permutationVector = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
-                x->permutei(arguments);
+
-                STORE_RESULT(x);
+    z->assign(x->permute(permutationVector));
-            } else {
-                auto output = OUTPUT_VARIABLE(0);
-                auto result = x->permute(arguments);
-                output->assign(result);
-                STORE_RESULT(output);
-            }
 
     return Status::OK();
-        }
+}
 
-        DECLARE_TYPES(permute) {
+//////////////////////////////////////////////////////////////////////////
+DECLARE_TYPES(permute) {
     getOpDescriptor()
             ->setAllowedInputTypes(0, nd4j::DataType::ANY)
             ->setAllowedInputTypes(1, {ALL_INTS})
             ->setSameMode(true);
-        }
+}
 
-        DECLARE_SHAPE_FN(permute) {
+//////////////////////////////////////////////////////////////////////////
-            auto shapeList = SHAPELIST();
+DECLARE_SHAPE_FN(permute) {
-            auto arguments = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
 
-            if (shape::rank(inputShape->at(0)) == 0) {
+    auto x = INPUT_VARIABLE(0);
-                shapeList->push_back(ConstantShapeHelper::getInstance()->scalarShapeInfo(ArrayOptions::dataType(inputShape->at(0))));
-            } else if (inputShape->size() == 1 && !arguments.empty()) {
-                shapeList->push_back(ShapeUtils::evalPermShapeInfo(arguments.data(), arguments.size(), *INPUT_VARIABLE(0), block.workspace()));
-            } else {
-                if(arguments.size() == 0){
-                    //Reverse dimensions
-                    int rank = shape::rank(inputShape->at(0));
-                    for (int e = rank - 1; e >= 0; e--)
-                        arguments.emplace_back(e);
-                }
 
-                shapeList->push_back(ShapeUtils::evalPermShapeInfo(arguments.data(), arguments.size(), *INPUT_VARIABLE(0), block.workspace()));
+    if (block.width() == 1 && block.getIArguments()->size() == 0)
-            }
+        return SHAPELIST(ShapeUtils::evalTranspShapeInfo(*x, block.workspace(), true));
 
-            return shapeList;
+    std::vector<int> permutationVector = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
-        }
+
-    }
+    auto outputShapeInfo = ShapeUtils::evalPermShapeInfo(permutationVector.data(), x->rankOf(), *x, block.workspace(), true);
+
+    return SHAPELIST(outputShapeInfo);
+}
+
+}
 }
 
 #endif

libnd4j/include/ops/declarable/generic/shape/reshape.cpp

@@ -24,23 +24,29 @@
 #include <ops/declarable/CustomOperations.h>
 
 namespace nd4j {
-    namespace ops {
+namespace ops  {
-        //////////////////////////////////////////////////////////////////////////
-        // here iArgs is a vector with (optional) negative of order as first element:
-        // ({-order, dim1, dim2, dim3, ...})
-        CUSTOM_OP_IMPL(reshape, 1, 1, true, 0, -2) {
-            auto x = INPUT_VARIABLE(0);
 
-            if (block.width() == 1) {
+//////////////////////////////////////////////////////////////////////////
-                auto arguments = block.getIArguments();
+// here iArgs is a vector with (optional) negative of order as first element:
-                int argsSize = arguments->size();
+// ({-order, dim1, dim2, dim3, ...})
+CUSTOM_OP_IMPL(reshape, 1, 1, false, 0, -2) {
+
+    auto x = INPUT_VARIABLE(0);
+    auto z = OUTPUT_VARIABLE(0);
 
     //Special case: empty.reshape(<other empty shape>) -> return empty
         if (x->isEmpty()) {
-                    REQUIRE_TRUE(OUTPUT_VARIABLE(0)->isEmpty(), 0, "Reshape: when input is empty, output must also be empty");
+            REQUIRE_TRUE(z->isEmpty(), 0, "Reshape: when input is empty, output must also be empty");
-                    return ND4J_STATUS_OK;    //No op
+            return Status::OK();    //No op
         }
 
+    if (block.width() == 1) {
+
+        auto arguments = block.getIArguments();
+        int argsSize = arguments->size();
+
+
+
         int e = 1;
         char order = (char) -(*arguments)[0];
         if (order != 'c' && order != 'f') {
@@ -77,28 +83,15 @@ namespace nd4j {
             nd4j_printv("Reshape: new shape", shapeNew);
         }
 
-                if (block.isInplace()) {
-                    if (x->reshapei(order, shapeNew)) {
-                        STORE_RESULT(*x);
-                        return ND4J_STATUS_OK;
-                    }
-                } else {
-                    auto ret = OUTPUT_VARIABLE(0);
         auto xr = x->reshape(order, shapeNew);
-                    ret->assign(xr);
+        z->assign(xr);
-                    STORE_RESULT(*ret);
+        STORE_RESULT(*z);
 
         return Status::OK();
-                }
-            } else if (block.width() == 2) {
-                auto s = INPUT_VARIABLE(1);
 
-                //Special case: empty.reshape(-1) -> return empty
+    } else if (block.width() == 2) {
-                if (x->isEmpty()) {
+
-                    //REQUIRE_TRUE(s->lengthOf() == 1 && s->e<Nd4jLong>(0) == -1, 0, "Reshape: when input is empty, shape must be [-1]");
+        auto s = INPUT_VARIABLE(1);
-                    REQUIRE_TRUE(OUTPUT_VARIABLE(0)->isEmpty(), 0, "Reshape: when input is empty, output must also be empty");
-                    return Status::OK();    //No op
-                }
 
         char order = 'c';
         if (block.numI() > 0)
@@ -129,37 +122,30 @@ namespace nd4j {
             nd4j_printv("Reshape: new shape", shapeNew);
         }
 
-                if (block.isInplace()) {
-                    if (x->reshapei(order, shapeNew)) {
-                        STORE_RESULT(*x);
-                        return Status::OK();
-                    }
-                } else {
-                    auto ret = OUTPUT_VARIABLE(0);
         if (s->isEmpty()) {
             // just a scalar
-                        ret->assign(x);
+            z->assign(x);
         } else {
             auto xr = x->reshape(order, shapeNew);
-                        ret->assign(xr);
+            z->assign(xr);
         }
 
         return Status::OK();
-                }
+
     }
 
     return ND4J_STATUS_BAD_INPUT;
-        }
+}
 
 
-        DECLARE_TYPES(reshape) {
+DECLARE_TYPES(reshape) {
     getOpDescriptor()
             ->setAllowedInputTypes(0, nd4j::DataType::ANY)
             ->setAllowedInputTypes(1, {ALL_INTS})
             ->setSameMode(true);
-        }
+}
 
-        DECLARE_SHAPE_FN(reshape) {
+DECLARE_SHAPE_FN(reshape) {
     auto inp = inputShape->at(0);
 
     // we can launch op using Int arguments
@@ -270,8 +256,8 @@ namespace nd4j {
 
         return SHAPELIST(ConstantShapeHelper::getInstance()->createShapeInfo(ArrayOptions::dataType(inp), 'c', shapeNew));
     }
-        }
+}
-    }
+}
 }
 
 #endif

libnd4j/include/ops/declarable/generic/shape/reshape_as.cpp

@@ -28,18 +28,16 @@ namespace nd4j {
 
 
     //////////////////////////////////////////////////////////////////////////
-    CUSTOM_OP_IMPL(reshapeas, 2, 1, true, 0, 0) {
+    CUSTOM_OP_IMPL(reshapeas, 2, 1, false, 0, 0) {
 
         auto x = INPUT_VARIABLE(0);
         auto y = INPUT_VARIABLE(1);
 
         auto z = OUTPUT_VARIABLE(0);
-        std::vector<Nd4jLong> shapeNew(y->shapeOf(), y->shapeOf() + y->rankOf());
-        char order = y->ordering();
 
-        if (x->reshapei(order, shapeNew)) {
+        if (x->reshapei(y->ordering(), y->getShapeAsVector())) {
-            *z = *x;
+
-            STORE_RESULT(*z);
+            z->assign(x);
             return Status::OK();
         }
 
@@ -49,14 +47,8 @@ namespace nd4j {
 
     DECLARE_SHAPE_FN(reshapeas) {
 
-    auto inputShapeInfo = inputShape->at(1);    
+        return SHAPELIST(ShapeBuilders::copyShapeInfo(INPUT_VARIABLE(1)->getShapeInfo(), false, block.workspace()));
-    int shapeInfoLength = inputShapeInfo[0]*2 + 4;
+    }
-
-    Nd4jLong* outputShapeInfo(nullptr);
-    COPY_SHAPE(inputShapeInfo, outputShapeInfo);
-    
-    return SHAPELIST(CONSTANT(outputShapeInfo));
-}
 
         DECLARE_TYPES(reshapeas) {
             getOpDescriptor()

libnd4j/include/ops/declarable/generic/shape/squeeze.cpp

@@ -25,7 +25,7 @@
 
 namespace nd4j {
     namespace ops {
-        CUSTOM_OP_IMPL(squeeze, 1, 1, true, 0, -2) {
+        CUSTOM_OP_IMPL(squeeze, 1, 1, false, 0, -2) {
             auto input = INPUT_VARIABLE(0);
             auto output = OUTPUT_VARIABLE(0);
 
@@ -71,11 +71,15 @@ namespace nd4j {
             }
 
             if (block.isInplace()) {
-                output->reshapei(input->ordering(), shape);
+                output->reshapei(input->ordering(), shape, false);
+            } else {
+                if (input->ews() == 1 && output->ews() == 1 && input->ordering() == output->ordering()) {
+                    output->dataBuffer()->copyBufferFrom(*input->dataBuffer().get(), output->lengthOf() * DataTypeUtils::sizeOfElement(output->dataType()), 0, input->bufferOffset());
                 } else {
                     auto tmp = input->reshape(input->ordering(), shape);
                     output->assign(tmp);
                 }
+            }
 
             return Status::OK();
         }

libnd4j/include/ops/declarable/generic/shape/tile_to_shape.cpp

@@ -25,7 +25,7 @@
 
 namespace nd4j {
 namespace ops {
-    CUSTOM_OP_IMPL(tile_to_shape, 1, 1, true, 0, -1) {
+    CUSTOM_OP_IMPL(tile_to_shape, 1, 1, false, 0, -1) {
 
         auto input = INPUT_VARIABLE(0);
         auto output = OUTPUT_VARIABLE(0);

libnd4j/include/ops/declarable/generic/shape/transpose.cpp

@@ -15,7 +15,8 @@
  ******************************************************************************/
 
 //
-// Created by raver119 on 29/10/17.
+// @author raver119@gmail.com
+// @author Yurii Shyrma (iuriish@yahoo.com)
 //
 
 #include <op_boilerplate.h>
@@ -27,111 +28,50 @@
 namespace nd4j {
 namespace ops  {
 
-    //////////////////////////////////////////////////////////////////////////
+//////////////////////////////////////////////////////////////////////////
-    CUSTOM_OP_IMPL(transpose, 1, 1, true, 0, 0) {
+CUSTOM_OP_IMPL(transpose, 1, 1, false, 0, 0) {
+
     auto x = INPUT_VARIABLE(0);
-        if (block.width() == 1) {
+    auto z = OUTPUT_VARIABLE(0);
-            if (block.isInplace()) {
-                x->transposei();
-                STORE_RESULT(*x);
-            } else {
-                auto output = OUTPUT_VARIABLE(0);
-                auto t = x->transpose();
-                output->assign(t);
-                STORE_RESULT(*output);
-            }
-        } else {
-            // this is tf-mode transpose, that's nd4j permute
-            bool replace = false;
-            std::vector<int> arguments(*block.getIArguments());
 
-            auto w = block.width();
+    //Special case: empty.reshape(<other empty shape>) -> return empty
-            auto a = arguments.size();
+    if (x->isEmpty()) {
-
+        REQUIRE_TRUE(z->isEmpty(), 0, "TRANSPOSE OP: when input is empty, output must also be empty");
-            if (w == 2 && a == 0) {
+        return Status::OK();    //No op
-                auto axis = INPUT_VARIABLE(1);
-                for (int e = 0; e < axis->lengthOf(); e++) {
-                    auto ax = axis->e<int>(e);
-                    if (ax < 0)
-                        ax += x->rankOf();
-
-                    arguments.emplace_back(ax);
     }
 
-                replace = true;
+    if (block.width() == 1 && block.getIArguments()->size() == 0) {
-            } else if (a == 0) {
+        z->assign(x->transpose());
-                for (int e = x->rankOf() - 1; e >= 0; e--)
-                    arguments.emplace_back(e);
-            }
-
-            // 0D edge case
-            if (x->rankOf() == 0) {
-                REQUIRE_TRUE(arguments.size() == 1, 0, "Permute: only one axis is allowed for scalar");
-                auto output = OUTPUT_VARIABLE(0);
-                if (!block.isInplace())
-                    output->assign(x);
-
         return Status::OK();
     }
 
-            if(block.isInplace()) {		// in-place
+    std::vector<int> permutationVector = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
-                x->permutei(arguments);
+
-                STORE_RESULT(x);
+    z->assign(x->permute(permutationVector));
-            } else {
-                auto input = x->permute(arguments);
 
-                auto output = OUTPUT_VARIABLE(0);
-                output->assign(input);
-             }
-        }
     return Status::OK();
-    }
+}
 
-    DECLARE_TYPES(transpose) {
+DECLARE_TYPES(transpose) {
     getOpDescriptor()
             ->setAllowedInputTypes(nd4j::DataType::ANY)
             ->setSameMode(true);
-    }
+}
+
+DECLARE_SHAPE_FN(transpose) {
+
+    auto x = INPUT_VARIABLE(0);
+
+    if (block.width() == 1 && block.getIArguments()->size() == 0)
+        return SHAPELIST(ShapeUtils::evalTranspShapeInfo(*x, block.workspace(), true));
+
+    std::vector<int> permutationVector = block.width() > 1 ? INPUT_VARIABLE(1)->asVectorT<int>() : *block.getIArguments();
+
+    auto outputShapeInfo = ShapeUtils::evalPermShapeInfo(permutationVector.data(), x->rankOf(), *x, block.workspace(), true);
 
-    DECLARE_SHAPE_FN(transpose) {
-        if (block.width() == 1) {
-            auto outputShapeInfo = ShapeUtils::evalTranspShapeInfo(*INPUT_VARIABLE(0), block.workspace());
     return SHAPELIST(outputShapeInfo);
-        } else {
+}
-            // this is basically permute mode
-            auto shapeList = SHAPELIST();
-            auto arguments = block.getIArguments();
-            if (shape::rank(inputShape->at(0)) == 0) {
-                Nd4jLong *newshape;
-                ALLOCATE(newshape, block.getWorkspace(), shape::shapeInfoLength(inputShape->at(0)), Nd4jLong);
-                newshape[0] = 0;
-                newshape[1] = 0;
-                newshape[2] = 1;
-                newshape[3] = 99;
-                ArrayOptions::copyDataType(newshape, inputShape->at(0));
-                shapeList->push_back(newshape);
-            } else if (arguments->size() > 0 || inputShape->size() > 1) {
-                auto axis = arguments->size() > 0 ? *arguments : (INPUT_VARIABLE(1))->template asVectorT<int>();
-                auto outputShapeInfo = ShapeUtils::evalPermShapeInfo(axis.data(), axis.size(), *INPUT_VARIABLE(0), block.workspace());
-                shapeList->push_back(outputShapeInfo);
-            } else if (inputShape->size() == 2) {
-                // dead end
-                auto axis = INPUT_VARIABLE(1);
-                auto axisV = axis->template asVectorT<Nd4jLong>();
-                auto newshape = ShapeUtils::evalPermShapeInfo(axisV.data(), axisV.size(), *INPUT_VARIABLE(0), block.workspace());
-                shapeList->push_back(newshape);
-            } else {
-                int rank = shape::rank(inputShape->at(0));
-                for (int e = rank - 1; e >= 0; e--)
-                    arguments->emplace_back(e);
 
-                auto outputShapeInfo = ShapeUtils::evalPermShapeInfo(arguments->data(), arguments->size(), *INPUT_VARIABLE(0), block.workspace());
-                shapeList->push_back(outputShapeInfo);
-            }
-
-            return shapeList;
-        }
-    }
 }
 }