DL4J Fixes (#204)

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

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

* Fix for SpaceToBatch layer

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

* #8133 DL4J SpaceToBatch gradient fix

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

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/exceptions/TestInvalidInput.java

@@ -22,12 +22,16 @@ import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
 import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
 import org.deeplearning4j.nn.conf.inputs.InputType;
 import org.deeplearning4j.nn.conf.layers.*;
+import org.deeplearning4j.nn.conf.layers.recurrent.SimpleRnn;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
 import org.junit.Test;
 import org.nd4j.linalg.activations.Activation;
+import org.nd4j.linalg.api.ndarray.INDArray;
 import org.nd4j.linalg.factory.Nd4j;
 
-import static org.junit.Assert.fail;
+import java.util.Map;
+
+import static org.junit.Assert.*;
 
 /**
  * A set of tests to ensure that useful exceptions are thrown on invalid input
@@ -267,23 +271,44 @@ public class TestInvalidInput extends BaseDL4JTest {
         //Idea: Using rnnTimeStep with a different number of examples between calls
         //(i.e., not calling reset between time steps)
 
+        for(String layerType : new String[]{"simple", "lstm", "graves"}) {
+
+            Layer l;
+            switch (layerType){
+                case "simple":
+                    l = new SimpleRnn.Builder().nIn(5).nOut(5).build();
+                    break;
+                case "lstm":
+                    l = new LSTM.Builder().nIn(5).nOut(5).build();
+                    break;
+                case "graves":
+                    l = new GravesLSTM.Builder().nIn(5).nOut(5).build();
+                    break;
+                default:
+                    throw new RuntimeException();
+            }
+
             MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().list()
-                        .layer(0, new GravesLSTM.Builder().nIn(5).nOut(5).build())
+                    .layer(l)
-                        .layer(1, new RnnOutputLayer.Builder().nIn(5).nOut(5).activation(Activation.SOFTMAX).build()).build();
+                    .layer(new RnnOutputLayer.Builder().nIn(5).nOut(5).activation(Activation.SOFTMAX).build()).build();
 
             MultiLayerNetwork net = new MultiLayerNetwork(conf);
             net.init();
 
             net.rnnTimeStep(Nd4j.create(3, 5, 10));
 
+            Map<String, INDArray> m = net.rnnGetPreviousState(0);
+            assertNotNull(m);
+            assertFalse(m.isEmpty());
+
             try {
                 net.rnnTimeStep(Nd4j.create(5, 5, 10));
-            fail("Expected DL4JException");
+                fail("Expected Exception - " + layerType);
-        } catch (DL4JException e) {
-            System.out.println("testInvalidRnnTimeStep(): " + e.getMessage());
             } catch (Exception e) {
-            e.printStackTrace();
+//                e.printStackTrace();
-            fail("Expected DL4JException");
+                String msg = e.getMessage();
+                assertTrue(msg, msg != null && msg.contains("rnn") && msg.contains("batch"));
+            }
         }
     }
 }

deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/gradientcheck/CNNGradientCheckTest.java

@@ -28,6 +28,7 @@ import org.deeplearning4j.nn.conf.distribution.NormalDistribution;
 import org.deeplearning4j.nn.conf.inputs.InputType;
 import org.deeplearning4j.nn.conf.layers.*;
 import org.deeplearning4j.nn.conf.layers.convolutional.Cropping2D;
+import org.deeplearning4j.nn.graph.ComputationGraph;
 import org.deeplearning4j.nn.multilayer.MultiLayerNetwork;
 import org.deeplearning4j.nn.weights.WeightInit;
 import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
@@ -343,6 +344,12 @@ public class CNNGradientCheckTest extends BaseDL4JTest {
 
                     assertTrue(msg, gradOK);
 
+                    //Also check compgraph:
+                    ComputationGraph cg = net.toComputationGraph();
+                    gradOK = GradientCheckUtil.checkGradients(cg, DEFAULT_EPS, DEFAULT_MAX_REL_ERROR,
+                            DEFAULT_MIN_ABS_ERROR, PRINT_RESULTS, RETURN_ON_FIRST_FAILURE, new INDArray[]{input}, new INDArray[]{labels});
+                    assertTrue(msg + " - compgraph", gradOK);
+
                     TestUtils.testModelSerialization(net);
                 }
             }

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

@@ -2460,6 +2460,13 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
             }
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
 
+            if(t != null){
+                if(t instanceof RuntimeException){
+                    throw ((RuntimeException)t);
+                }
+                throw new RuntimeException("Error during neural network forward pass", t);
+            }
+
             if(outputWorkspace == null || outputWorkspace instanceof DummyWorkspace) {
                 WorkspaceUtils.assertNoWorkspacesOpen("Expected no workspace active at the end of outputOfLayerDetached");
             } else {
@@ -2780,6 +2787,13 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
                 }
             }
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
+
+            if(t != null){
+                if(t instanceof RuntimeException){
+                    throw ((RuntimeException)t);
+                }
+                throw new RuntimeException("Error during neural network backpropagation calculation", t);
+            }
         }
 
         //Now, add the gradients in the order we need them in for flattening (same as params order)
@@ -3312,8 +3326,11 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
 
     @Override
     public int batchSize() {
+        //In 99+% of cases, the input and labels dimension 0 size should be identical
+        //The only real exceptions: space to batch, and batch to space layers
+        //In those cases, we should base it on the labels size, as this impacts gradient calculation
         // FIXME: int cast
-        return (int) inputs[0].size(0);
+        return labels == null || labels[0] == null ? (int) inputs[0].size(0) : (int)labels[0].size(0);
     }
 
     @Override

deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/convolution/SpaceToBatch.java

@@ -70,12 +70,12 @@ public class SpaceToBatch extends AbstractLayer<org.deeplearning4j.nn.conf.layer
 
     private INDArray getBlocksArray() {
         int[] intBlocks = layerConf().getBlocks();
-        return Nd4j.create(new double[] {intBlocks[0], intBlocks[1]});
+        return Nd4j.createFromArray(intBlocks);
     }
 
     private INDArray getPaddingArray() {
         int[][] intPad = layerConf().getPadding();
-        return Nd4j.create( new double[][] { {intPad[0][0], intPad[0][1]}, {intPad[1][0], intPad[1][1]}});
+        return Nd4j.createFromArray(intPad);
     }
 
 

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

@@ -32,6 +32,7 @@ import org.deeplearning4j.nn.layers.BaseLayer;
 import org.deeplearning4j.nn.layers.mkldnn.MKLDNNConvHelper;
 import org.deeplearning4j.nn.workspace.ArrayType;
 import org.deeplearning4j.nn.workspace.LayerWorkspaceMgr;
+import org.nd4j.base.Preconditions;
 import org.nd4j.linalg.activations.IActivation;
 import org.nd4j.linalg.activations.impl.ActivationSigmoid;
 import org.nd4j.linalg.api.memory.MemoryWorkspace;
@@ -188,12 +189,9 @@ public class LSTMHelpers {
         }
         //Input validation: check that if past state is provided, that it has same
         //These can be different if user forgets to call rnnClearPreviousState() between calls of rnnTimeStep
-        if (prevOutputActivations != null && prevOutputActivations.size(0) != input.size(0)) {
+        Preconditions.checkState(prevOutputActivations == null || prevOutputActivations.size(0) == input.size(0),
-            throw new DL4JInvalidInputException("Previous activations (stored state) number of examples = "
+                "Invalid RNN previous state (last time step activations/initialization): rnnTimeStep with different minibatch size, or forgot to call rnnClearPreviousState between batches?" +
-                            + prevOutputActivations.size(0) + " but input array number of examples = " + input.size(0)
+                        " Previous step output = [batch, nIn] = %ndShape, current input = [batch, nIn, seqLength] = %ndShape", prevOutputActivations, input);
-                            + ". Possible cause: using rnnTimeStep() without calling"
-                            + " rnnClearPreviousState() between different sequences?");
-        }
 
         //initialize prevOutputActivations to zeroes
         if (prevOutputActivations == null) {

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

@@ -217,6 +217,9 @@ public class SimpleRnn extends BaseRecurrentLayer<org.deeplearning4j.nn.conf.lay
         assertInputSet(false);
         Preconditions.checkState(input.rank() == 3,
                 "3D input expected to RNN layer expected, got " + input.rank());
+        Preconditions.checkState(prevStepOut == null || prevStepOut.size(0) == input.size(0),
+                "Invalid RNN previous state (last time step activations/initialization): rnnTimeStep with different minibatch size, or forgot to call rnnClearPreviousState between batches?" +
+                        " Previous step output = [batch, nIn] = %ndShape, current input = [batch, nIn, seqLength] = %ndShape", prevStepOut, input);
 
         applyDropOutIfNecessary(training, workspaceMgr);
 

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

@@ -436,8 +436,11 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
 
     @Override
     public int batchSize() {
+        //In 99+% of cases, the input and labels dimension 0 size should be identical
+        //The only real exceptions: space to batch, and batch to space layers
+        //In those cases, we should base it on the labels size, as this impacts gradient calculation
         // FIXME: int cast
-        return (int) input.size(0);
+        return labels == null ? (int) input.size(0) : (int)labels.size(0);
     }
 
     @Override
@@ -1362,6 +1365,13 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
 
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
 
+            if(t != null){
+                if(t instanceof RuntimeException){
+                    throw ((RuntimeException)t);
+                }
+                throw new RuntimeException("Error during neural network forward pass", t);
+            }
+
             if(outputWorkspace == null || outputWorkspace instanceof DummyWorkspace) {
                 WorkspaceUtils.assertNoWorkspacesOpen("Expected no workspace active at the end of outputOfLayerDetached", true);
             } else {
@@ -2007,6 +2017,13 @@ public class MultiLayerNetwork implements Serializable, Classifier, Layer, Neura
                 }
             }
             Nd4j.getMemoryManager().setCurrentWorkspace(initialWorkspace);
+
+            if(t != null){
+                if(t instanceof RuntimeException){
+                    throw ((RuntimeException)t);
+                }
+                throw new RuntimeException("Error during neural network forward pass", t);
+            }
         }
 
         if (layerWiseConfigurations.getTrainingWorkspaceMode() == WorkspaceMode.NONE) {