Could not create a descriptor for a dilated convolution forward propagation primitive

DL4J
1

When I try to train my model build in computationGraph, it show the error like this:
Caused by: java.lang.RuntimeException: could not create a descriptor for a dilated convolution forward propagation primitive
How can I sovle it?

2

I change the backend to cuda and then the problem sovled.

3

I’m not sure this is a “solution” per se :slight_smile: could you at least tell us what happened here if there’s a problem with our cpu backend we should look in to it if possible.

4

It seems that if the graph includes a conv1d layer with dilation config, the error will appear.

5

@xianxuan could you post a sample config here that reproduces your issue or file an issue if you didn’t already? If you did, could you link it here? Thanks!

6 
graph
                .addLayer(nameLayer(blockName, "cnn1"),
                        new Convolution1DLayer.Builder(kernelSize)
                                .dilation(dilation)
                                .convolutionMode(ConvolutionMode.Causal).nOut(channel).build(),
                        input)
                .addLayer(nameLayer(blockName, "relu1"),
                        new ActivationLayer.Builder().activation(Activation.RELU).build(),
                        nameLayer(blockName, "cnn1"))
                .addLayer(nameLayer(blockName, "dropout1"),
                        new DropoutLayer.Builder(dropout).build(),
                        nameLayer(blockName, "relu1"))
                .addLayer(nameLayer(blockName, "cnn2"),
                        new Convolution1DLayer.Builder(kernelSize)
                                .dilation(dilation)
                                .convolutionMode(ConvolutionMode.Causal).nIn(channel).nOut(channel).build(),
                        nameLayer(blockName, "dropout1"))
                .addLayer(nameLayer(blockName, "relu2"),
                        new ActivationLayer.Builder().activation(Activation.RELU).build(),
                        nameLayer(blockName, "cnn2"))
                .addLayer(nameLayer(blockName, "dropout2"),
                        new DropoutLayer.Builder(dropout).build(),
                        nameLayer(blockName, "relu2"));
7 
 graph.addInputs(input).setInputTypes(InputType.recurrent(featureSize, timeSteps))
            .addLayer("crop", new Cropping1D.Builder(timeSteps-1,0).build(),nameLayer(blockName, "relu2"))
            .addLayer("outputLayer",
                    new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)
                            .activation(Activation.SOFTMAX)
                            .nOut(outputSize)
                            .weightInit(WeightInit.XAVIER).build(),
                    "crop")
            .setOutputs("outputLayer");
1 Like