Covert Conv-TasNet from torch to onnx successfully, but not work

audio
1

I coverted Conv-TasNet from torch to onnx. Follow steps from website (Pytoch onnx). Then used $“testing.assert_almost_equal” to verify output. The orignal TasNet pre-trained model worked, however the result of onnx are zeros. Its weights seems to be all zero, and the output dimension also different. I cann’t figure which step was wrong.

Here is my project.
Run demo_tasnet.py onnx will be save in checkpoints.
Also write a test.py to campare result.

2

Hi,
I tried to figure out the problem. So I checked the code again.

torch.onnx.export(model, input,
“.path/to/conv_tasnet.onnx”,
export_params=True
)

I set up export_params=True, and test .onnx with onnxruntime the ouput are still all zeros. It seems that the weights are not be saved.