Any interest to integrate Onnx2TorchSharp?

[GeorgeS2019]

https://github.com/dotnet/TorchSharp/discussions/1111

[fel88]

Hi, why not. What exactly should be implemented?

[GeorgeS2019]

Few still recognize what an excellent GUI and Visualization your project has contributed to the .NET ONNX community.

NOW => the next level of contribution

Given an ONNX uploaded to your program. We visualize the ONNX structure.

• Next, the user can click on each of the node and see the corresponding Torchsharp code.
• Ideally, your program can pattern recognize the ONNX structures to anticipate and translate to the closest Torchsharp codes for the entire imported ONNX.

state of art

The pytorch onnx2torch looks at individual Onnx and reverse engineer back the original pytorch code that has exported the Onnx but then go one more step to create back the checkpoint file with trained weights

However, this is NOT the primary goal

Not only the program need to source generate the corresponding TorchSharp codes., it needs to create a TorchSharp specific trained file with weights that are transferred from the imported ONNX.

The process serves TWO purposes

• [ ] Source Generate more Torchsharp codes from onnx exported from Pytorch
• [ ] Avoid repeating training in Torchsharp since the training already done in Pytorch. So the users can continue further Additional training in Torchsharp, without starting from scratch.

This will go beyond what Netron provides !

[fel88]

It sounds like some kind of re-implementation of onnx2torch => onnx2torchSharp. Should I also implement visualization for the TorchScript format?

[GeorgeS2019]

For Pytorch, there are utilities, integrated into Jupiter notebook, to help understand and visualize how the deep learning progress beyond tensorboard.

The eventual vision is TorchSharp 101..used by University lectures to teach deep learning in computer game development.

You have started deep neural network visualization that has better chance to be ported to 3D game engine than the 3D engine unfriendly electron framework: Netron.

What you do to contribute to visualisation will benefit many disciplines.

Try Skiasharp ..more cross platform than dotnet system.drawing.

[GeorgeS2019]

If visualization of Torchscript is possible...will be great..not sure the implications yet.

Try ask your question at TorchSharp

I am sure many will feedback you for your contribution

[fel88]

Ok, thanks for sharing your vision 👍 SkiaSharp and TorchScript support seems doable to me.

[GeorgeS2019]

Skiasharp means blazor, Linux, .net notebook etc

[GeorgeS2019]

@fel88

https://github.com/shaltielshmid/TorchSharp.PyBridge/discussions/3