Will TorchSharp support exporting ONNX models in the future?

[HCareLou]

In real-world projects, for quick deployment and application, PyTorch typically exports to ONNX, but it seems that TorchSharp does not currently support this feature.

[GeorgeS2019]

(Relevant to) Sharing-Model-Data-between-PyTorch-and-TorchSharp

https://github.com/dotnet/TorchSharp/wiki/Sharing-Model-Data-between-PyTorch-and-TorchSharp

[Export Onnx] Support Custom Operators needed before ONNX export

#526

https://github.com/K024/llm-sharp/tree/main/NativeOps
@K024 => what do you think of this issue?

Once the modules moving to c# is completed. Perhaps the following codes could be repurposed
https://github.com/dotnet/machinelearning/tree/main/src/Microsoft.ML.OnnxConverter

[Import Onnx] onnx2torch

Transfer pretrained weights from Onnx to TorchSharp

#1111
https://github.com/ENOT-AutoDL/onnx2torch

[GeorgeS2019]

Previously we do not have sufficient people to help, NOW I believe we DO

#973

[GeorgeS2019]

@yueyinqiu
@shaltielshmid
@NiklasGustafsson

I personally think we need to re-visit this topic as we NOW have more people interested of this feature

[GeorgeS2019]

dotnet/machinelearning#6884

In order to enable this, TorchSharp would need to enable export to ONNX. No immediate work can be done in ML.NET at this time.

@luisquintanilla

[NiklasGustafsson]

While repeatedly asked for here, ONNX export is a complicated feature that, like torch.compile, will require careful design and planning. It has been on the backburner for a long time and is likely to remain so until we add significant resources to the project. Volunteers are always welcome, but it's not a simple thing to get working. If ONNX is essential to you, then PyTorch is the solution.

[HCareLou]

This is just a suggestion; if TorchSharp is positioned for the C# ecosystem, then there is indeed no need to export other formats.

[NiklasGustafsson]

And it's a good suggestion... :-) Just something we don't have the resources to handle at the moment. Same thing with torch.compile -- it's something I would love to be able to add.

[GeorgeS2019]

Just for future reference

• torch.compile