Interest on onnx2torch?

[GeorgeS2019]

Just wonder if there is interest on onnx2torchsharp
https://github.com/ENOT-AutoDL/onnx2torch

[GeorgeS2019]

Potentially relevant projects

• https://github.com/MyCaffe/OnnxControl
• https://github.com/nietras/OnnxSharp
• https://github.com/onnxplorer/ONNXplorer/tree/master/Assets/Scripts/Onnx
• https://github.com/fel88/Dendrite

[NiklasGustafsson]

Does this tool generate source code, or in-memory constructs? If someone wanted to do a TorchSharp version, I would suggest spitting out source code rather than using something like Reflection.Emit()

[GeorgeS2019]

@NiklasGustafsson
I think the community is NOW more inspired compared to previous years.
This onnx2torchsharp will bring more existing onnx-generated-by-pytorch projects to torchsharp.

Hopefully, one of the community members is inspired to try.

[GeorgeS2019]

Follow the discussion here

[GeorgeS2019]

@NiklasGustafsson

It is now possible to do training using c# Onnx API.

It would mean it is possible to download exported Onnx from Pytorch from e.g. HuggingFace to continue training in c#

However, there are still many challenges ahead that are potentially commonly shared between TorchSharp and Onnx c# API trainings

One of these is how to have a DOTNET Foundation approach to address ContribOperators.

As this issue will have impact on the future of ML.NET, Torchsharp and OnnxRuntime training and inference using c#

[GeorgeS2019]

Reference

Exporting models to onnx is supported in libtorch so this would probably be easy to add but I don't think there is a good way to do the opposite: running/importing onnx models from the c++ side, see this pytorch/pytorch#21683.

Exporting models to onnx is supported in libtorch so this would probably be easy to add but I don't think there is a good way to do the opposite: running/importing onnx models from the c++ sideI don't think there is a current plan to do this. If the goal is just to retrieve the weight and not the architecture, I would suggest writing some one off python script using the onnx python package and writing some safetensors file. If the goal is to generate a full jit/torch-script model, there seems to be various python libraries doing this but I'm not sure how good they actually are.

models with torchscript are slower than the onnx version, so I would like to be able to run inference models with the onnx format. But I hear you, thank you.