deeplearning4j: ImageRecordReader crashes JVM with loaded Keras model in 1.0.0-beta7

Issue Description

I encountered a strange problem in 1.0.0-beta7 while trying to run a Keras model loaded from a .h5 file (e.g., VGG16.h5 from here) - this model previously ran fine in 1.0.0-beta6.

Calling computationGraph.feedForward(features, false) would crash the JVM (error log, using this code snippet:

// Create VGG16 from a Keras .h5 file
ComputationGraph tmpModel = KerasModelImport.importKerasModelAndWeights("VGG16.h5");
tmpModel.init();

ImageRecordReader reader = new ImageRecordReader(224, 224, 3);
reader.initialize(new FileSplit(new File("img_125_5.jpg"))); // Test with a single image
DataSetIterator it = new RecordReaderDataSetIterator(reader, 1);

// Keras model has wrong channel order, so flip it at the reader level
reader.setNchw_channels_first(false);

INDArray features = it.next().getFeatures();
// INDArray features = Nd4j.rand(1, 224, 224, 3); // Runs fine when initializing from random array of same size

System.out.println(Arrays.toString(features.shape())); // prints [1, 224, 224, 3]

tmpModel.feedForward(features, false);

The crash would happen specifically within the ComputationGraph class at line 1976 - figured this by stepping through the code in IntelliJ.

Strangely though, the code snippet above runs fine if you use a random numpy array of the same shape (so the issue isn’t caused by the features shape). Looking into the values of the features given by the DatasetIterator, there aren’t any NaNs or weird values (all are between 0 and 1).

Also interesting to note is that the .h5 model can be saved in beta6 to a zip using model.save(new File("VGG.zip")), then loaded in beta7, and the above snippet works fine (swapping the KerasModelImport... for ComputationGraph.load(new File("beta6KerasVGG.zip"), true);

Another note, the above snippet works fine if using a different model (e.g., ResNet50.h5) - so it’s not all Keras models that this problem occurs with.

Conclusion

On one hand, it seems like the problem is caused by updates to the KerasModelImport process - a .h5 file which loaded and ran fine in 1.0.0-beta6 now no longer works in 1.0.0-beta7. Additionally, saving a .zip file of the beta6 version and loading a new ComputationGraph in beta7 circumvents the above problem.

However, it also seems like the ImageRecordReader or DataSetIterator could be the culprit - when those are taken out of the equation (by using a random INDArray) no errors occur.

Attached files

Version Information

Please indicate relevant versions, including, if relevant:

• Deeplearning4j version - 1.0.0-beta7
• Platform information (OS, etc) - Ubuntu 18.04
• CUDA version, if used
• NVIDIA driver version, if in use