Support ONNX?

[tommy-qichang]

Hi there,

Great job, and I'm looking forward to use it and compare with other frameworks like E2FGVI.

Just wondering do you plan to convert the model to Onnx format? I think that would be a great advantages if the model would evaluate on multiple devices. I've tried converting model of E2FGVI to ONNX, but failed(Since it has mmcv customer modules which is hard to convert).

Thanks.

[whyb]

Hi there,

Great job, and I'm looking forward to use it and compare with other frameworks like E2FGVI.

Just wondering do you plan to convert the model to Onnx format? I think that would be a great advantages if the model would evaluate on multiple devices. I've tried converting model of E2FGVI to ONNX, but failed(Since it has mmcv customer modules which is hard to convert).

Thanks.

I've been trying for three working days, and I feel that the problem stems from the use of many conditional statements(if/else) in Propainter's network and the torchvision::deform_conv2d operator, which isn't supported by ONNX. These factors prevent JIT tracing from capturing the entire data stream completely, ultimately causing the conversion to the ONNX model to fail.

The only solution I've found is to convert it to TorchScript and then use libtorch + pt model to make it work in C++.

I look forward to discussing any better solutions you might have in the future.