When exporting from torch, the result of the input varies depending on the dummy input.

[parkkyungjun]

Ask a Question

Question

When exporting from torch, the result of the input varies depending on the dummy input. What's wrong with this?

Further information

• Relevant Area:

torch.onnx.export(model, (dummy_input, th), model_path, verbose=True, export_params=True,
opset_version=12,
do_constant_folding=True,
input_names=['input', 'threshold'], output_names=['output1', 'output2'])

in my model forward
there is torch.cdist function

I am comparing the value saved in advance when performing the cdist operation with the output of the model after storing it in the model class's self.

I think cdist is the cause. And when using cdist, the speed is also very slow (cdist take almost 6 second )
help me

• Is this issue related to a specific model?
  Model name:
  Model opset:

Notes

[gramalingam]

I think it would be better to ask this question in the pytorch (to onnx converter) repo. @thiagocrepaldi

[parkkyungjun]

I think it would be better to ask this question in the pytorch (to onnx converter) repo. @thiagocrepaldi

I solved this issue by addressing a cause from another problem
Thank you