Simple CNN predicts NaNs

Hi All,
I am newbie in DL4J. I tried to implement a simple CNN like this:

val conf = new NeuralNetConfiguration.Builder()
    .seed(seed)
    .updater(Updater.NESTEROVS)
    .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
    .weightInit(WeightInit.XAVIER)
    .list
    .layer(0, new ConvolutionLayer.Builder()
      .name("c1")
      .kernelSize(5, numFeatures)
      .convolutionMode(ConvolutionMode.Truncate)
      .activation(Activation.RELU)
      .nIn(1)
      .nOut(32)
      .build)
    .layer(1, new OutputLayer.Builder(LossFunctions.LossFunction.MSE)
      .nOut(nOut)
      .dropOut(0.5)
      .activation(Activation.IDENTITY)
      .build)
    .setInputType(InputType.convolutional(windowSize, numFeatures, 1))
    .build

My first minibatch has size [batchSize, 1, windowSize, numFeatures] , so compatible with the CNN. The problem now, that I got NaN values from inference, thus the network does not learn… I want to use this network for time series prediction (trying this for first, then later going to RNN).
Do you have any idea, what causes the problem? The network gives sensible values after init, but after first iteration only NaN values are outputed.
In the dependencies, I have 0.9.1 version
Thank you for your help in advance!

@zsvizi first of all, make sure you update your version to 1.0.0-beta7.
That’s what the current examples: https://github.com/eclipse/deeplearning4j-examples is

Also, make sure to read this first: https://deeplearning4j.konduit.ai/tuning-and-training/troubleshooting-training

After you do that, let us know what you find.

The problem is, that the project I am working on is maintained by the DevOps team of the company I am working for… and you know, upgrading a dependency is far from trivial. Considering these constrains, can you suggest some approach to troubleshoot?

@zsvizi we can try but we have our own constraints as well :slight_smile: that tuning guide should still be relevant for you. If you need something more for support, please feel free to DM me and we’ll see what we can do for you.