Feedback to the plan for Torch integration to ML.NET

[GeorgeS2019]

TorchSharp integration to ML.NETis currently being planned (supervised by @ericstj ) as recently shared/discussed by @NiklasGustafsson

I suggest TorchSharp users provide feedback to the integration process here.

As part of the integration, the question how to consume exported Torch model by ML.NET has been planned.

I also would like to raise some issues raised before related to the above.

• Will there be an improvement in how TorchSharp consume Torch model created in python. Currently we need an additional python script to create a TorchSharp compatible model in python. Is there a plan soon that this additional step will be removed, perhaps as part of the ML.NET integration planning process?
• There are challenges dealing with ONNX, especially when dealing with NLP use cases (e.g. ref 1, ref 2). Is there a plan to consider incorporating ONNX save and load features in TorchSharp? By having Onnx save/load in TorchSharp, this provides an alternative path to the existing option to import ONNX to ML.NET

[ericstj]

cc @michaelgsharp @briacht
At the moment we're still trying to understand what this work would look like. Thanks for raising concerns and we'll make sure to stay engaged in the discussion.

[GeorgeS2019]

FYI: latest plan for Deep learning and TorchSharp integration to ML.NET!!!

[GeorgeS2019]

Currently it is possible in ML.NET to import Tensorflow trained model file .pb

ML.NET needs native binding to load PyTorch trained model file e.g. pth. TorchSharp is key to sort out the paths how ML.NET will load such trained file.

The LibTorch is able to load different formats of PyTorch trained model files

• PyTorch torch.save model file: savefile.pth
• TorchScript torch.jit.save model file: savefile.pt

Question

Shall ML.NET support one of the two or both file formats?

[GeorgeS2019]

*.pth and *.data [Torchsharp specific format for import and export] discussed in saveload.md

With this new implementation, now *.pt is introduced to the saveload.md

Discussion on how to integrate TorchSharp to ML.NET