Plans for Supporting Torch

[sbodenstein]

@soumith: I know that there are plans for supporting PyTorch. Are there any plans for supporting Torch? Or will it be supported indirectly via a Torch to PyTorch converter?

[soumith]

we have no plans of supporting lua torch, and are unlikely to look into it.

[sbodenstein]

Ok, thanks for the rapid response!

[ezyang]

However, we are planning to support importing onnx models into PyTorch. The priority for this particular feature is not high, but we may need it for some high priority work that is going on.

[jsenellart]

Hello, I am willing to work on converting lua-torch models to ONNX - at least in the context of the OpenNMT project for which it should be pretty straightforward. @soumith - is there some entry point on how to write a converter to ONNX?

[linkerzhang]

There're several xxx -> onnx converters in our repo, which you may want to refer to please. https://github.com/onnx/onnxmltools covers converting core-ml/scikit-learn models to onnx.
https://github.com/onnx/tensorflow-onnx covers converting tf models to onnx.

Of course, they're not having 100% coverage.

[jsenellart]

thanks - I will give a try.

[jsenellart]

Hello, I have a first draft of luatorch > onnx extraction here: https://github.com/jsenellart/luatorch-onnx-convert. Still work in progress, but it works end-to-end and converts discovered nn, nngraph and onmt modules to onnx - infering shapes when necessary, and building initializers. Contact me if you need specific support.