Loading ExtraFilesMap in Java?

Question

When saving a trained model using Pytorch, extra information can be added in Python or C++ using an ExtraFilesMap object. In Python this extra metadata can be extracted from a saved model e.g. via

# load the jit model and associated settings. 
extra_files = torch._C.ExtraFilesMap()
extra_files['dataOpts'] = ''
model = torch.jit.load(filepath_model, _extra_files=extra_files)

Using djl, a trained model can be loaded in Java, however, however, there does not appear to be any method to extract the ExtraFileMap information? I’ve spent some time trying to find a solution with no luck. Am I missing something - is this possible using djl?

If not, can we bump this to a feature request. It’s very useful having metadata such as this stored with the model.

Thanks for any help!

[lanking520]

@macster110, please feel free to raise a PR to fix it 😃

[lanking520]

Ok, I optimized our import procedure. You will be able to fetch these information through:

Criteria<Image, Classifications> criteria = Criteria.builder()
  .setTypes(Image.class, Classifications.class)
  .optModelUrls(modelUrl)
  .optOption("extraFiles", "dataOpts")  // <- pass in here 
  .optTranslator(ImageClassificationTranslator.builder()
                 .setPipeline(new Pipeline()
                              .add(new Resize(224, 224))
                              .add(new ToTensor()))
                 .optApplySoftmax(true).build())

ZooModel<Image, Classifications> model = ModelZoo.load(criteria);
model.getProperty("dataOpts"); // String representation of your dictionary

I tested with the model you have provided and it is working now. You can get the SNAPSHOT version and use it today. We have nightly snapshot published to maven everyday.