NetworkConfiguration.java

int numfilters = 128;
int kernelSize = 5;
int maxVocabIndex = 10000;
int numClasses = 2;
int embeddingLength = 50;
//Set up network configuration
ComputationGraphConfiguration config = new NeuralNetConfiguration.Builder()
        // workspace settings for memory management
        .weightInit(WeightInit.RELU)
        .seed(seed)
        .updater(new Adam(0.01))
        .activation(Activation.RELU)
        .graphBuilder()
        .addInputs("input")
        .addLayer("embed", new EmbeddingLayer.Builder()
                .nIn(maxVocabIndex)
                .nOut(embeddingLength)
                .build(), "input")
        .addLayer("cnn1", new Convolution1DLayer.Builder()
                .nIn(embeddingLength)
                .nOut(numfilters)
                .kernelSize(kernelSize)
                .activation(Activation.RELU)
                .build(), "embed")
        .addLayer("globalPool1", new GlobalPoolingLayer.Builder()
                .poolingType(PoolingType.MAX)
                .build(), "cnn1")
        .addLayer("predict", new DenseLayer.Builder()
                .nIn(numfilters)
                .nOut(numClasses)
                .dropOut(0.9)
                .activation(Activation.SOFTMAX)
                .build(), "globalPool1")
        .addLayer("loss", new LossLayer.Builder(LossFunctions.LossFunction.MCXENT)
                .build(), "predict")
        .setOutputs("loss")
        .setInputTypes(InputType.recurrent(1, maxSequenceLength))
        .build();

VectorCreator.java

/**
  * @param docTokensMap map of docId to list of ordered token Indices from documents
  * @param trainingLabels map of docId to label
  * return dataset object
  */
private DataSet createVectors(Map<String,Integer> trainingLabels, Map<String,List<Integer>> docTokensMap){
 int rowIndex = 0;
 int maxLength = 128;
 INDArray featureMatrix = Nd4j.create(docTokensMap.size(), 1, maxLength);
 INDArray labelsMatrix = Nd4j.create(docTokensMap.size(), 2);
  for (String docId : docId) {
            List<Integer> tokIndices = docTokensMap.get(docId);
            //get sequence length to use it for truncation later
            int sequenceLength = Math.min(tokensIds.size(), maxLength);
            if (sequenceLength == 0) {
                continue;
            }
            for (int i = 0; i < tokIndices.size(); i++) {
                Integer index = tokIndices.get(i);
                featureMatrix.putScalar(new int[]{rowIndex, 0, i}, index);
            }
            //this is only done if we are in training mode
            if (train) {
                int idx = trainingLabels.get(docId) == 1 ? 0 : 1;
                labelsMatrix.putScalar(new int[]{rowIndex, idx}, 1.0);   //Set label: [0,1] for negative, [1,0] for positive 
            }
        }
        // create the classifier problem data
        return new DataSet(featureMatrix, labelsMatrix, null, null);
}
private void initCuda(){
 // I tried to either completely turn off periodic GC 
 // or reduce the frequency
//      Nd4j.getMemoryManager().setAutoGcWindow(100000);
        Nd4j.getMemoryManager().togglePeriodicGc(false);
         CudaEnvironment.getInstance().getConfiguration()
                .allowMultiGPU(false);
//.setMaximumDeviceCache((long) (0.5 * 4 * 1024 * 1024 * 1024L))
}

This is my JVM params
-Dorg.bytedeco.javacpp.maxbytes=8G -Dorg.bytedeco.javacpp.maxphysicalbytes=10G -Xms1G -Xmx2G

This is for GPU:
----- Memory Configuration -----
JVM Memory: XMX 1.78 GB (1908932608)
JVM Memory: current 1.01 GB (1083703296)
JavaCPP Memory: Max Bytes 8 GB (8589934592)
JavaCPP Memory: Max Physical 10 GB (10737418240)
JavaCPP Memory: Current Bytes 4.55 GB (4884915841)
JavaCPP Memory: Current Physical 4.52 GB (4856979456)

This is for CPU:
----- Memory Configuration -----
JVM Memory: XMX 1.78 GB (1908932608)
JVM Memory: current 1.25 GB (1337458688)
JavaCPP Memory: Max Bytes 8 GB (8589934592)
JavaCPP Memory: Max Physical 10 GB (10737418240)
JavaCPP Memory: Current Bytes 814.36 MB (853923541)
JavaCPP Memory: Current Physical 2.14 GB (2302488576)

can you explain this?

When I added the embedding layer I can see this weird -1 in my network shape when I print it out the info for the cnn layer!
----- Network Activations: Inferred Activation Shapes -----
Current Minibatch Size 2322
Current Input Shape (Input 0) [2322, 1, 128]
Idx Name Layer Type Activations Type Activations Shape # Elements Memory
0 input InputVertex InputTypeRecurrent(1,timeSeriesLength=128) [2322, 1, 128] 297216 1.13 MB (1188864)
1 embed EmbeddingLayer InputTypeFeedForward(50) [2322, 50] 116100 453.52 KB (464400)
----------> 2 cnn1 Convolution1DLayer InputTypeRecurrent(128) [2322, 128, -1] -297216 -1.13 MB
3 globalPool1 GlobalPoolingLayer InputTypeFeedForward(128) [2322, 128] 297216 1.13 MB (1188864)
4 predict DenseLayer InputTypeFeedForward(2) [2322, 2] 4644 18.14 KB (18576)
5 loss LossLayer InputTypeFeedForward(0) [2322, 0] 0 0 B
Total Activations Memory 1.59 MB (1671840)
Total Activation Gradient Memory 1.59 MB (1671840)