Script quits unexpectedly (without errors) when trying to export model to ONNX

[miccio-dk]

First of all, thanks for this invaluable resource :)

As for my issue: I would like to export the NISQA v2 model to ONNX so that I can use for data evaluation within my tensorflow-based environment without introducing pytorch as a dependency (similarly to Microsoft's DNSMOS p.835). However, when attempting to run torch.onnx.export(), the script exits unexpectedly without throwing any error or raising exceptions. This happens no matter which opset version I pick. Do you know if the issue is related certain operations happening within the model?

Here is the code that I used. I just replace the content of export_dim with this and run the prediction script as usual. If you're interested in further debugging the issue, I can upload my fork of the repo with the full script.

  x, y, (idx, n_wins) = ds[0]
  x = x.unsqueeze(0)
  x.requires_grad = True
  n_wins = torch.from_numpy(n_wins).unsqueeze(0)
  #n_wins.requires_grad = True
  model.eval()
  torch.onnx.export(model,                   # model being run
                  (x, n_wins),               # model input (or a tuple for multiple inputs)
                  "nisqa.onnx",              # where to save the model (can be a file or file-like object)
                  export_params=True,        # store the trained parameter weights inside the model file
                  opset_version=14,          # the ONNX version to export the model to
                  do_constant_folding=True,  # whether to execute constant folding for optimization
  )
  print('Done')   # !!! this line never gets executed because the script quits unexpectedly !!!

Thanks in advance

[miccio-dk]

Update: when trying to export the model from Linux (or WSL in my case), I get the following, more verbose errors:

Device: cpu
Model architecture: NISQA_DIM
Loaded pretrained model from weights/nisqa.tar
/home/username/miniconda3/envs/nisqa/lib/python3.9/site-packages/torch/onnx/utils.py:1294: UserWarning: Provided key output for dynamic axes is not a valid input/output name
  warnings.warn("Provided key {} for dynamic axes is not a valid input/output name".format(key))
WARNING: The shape inference of prim::PackPadded type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
WARNING: The shape inference of prim::PackPadded type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
/home/username/miniconda3/envs/nisqa/lib/python3.9/site-packages/torch/onnx/symbolic_helper.py:258: UserWarning: ONNX export failed on adaptive_max_pool2d because input size not accessible not supported
  warnings.warn("ONNX export failed on " + op + " because " + msg + " not supported")
/home/username/miniconda3/envs/nisqa/lib/python3.9/site-packages/torch/onnx/symbolic_helper.py:716: UserWarning: allowzero=0 by default. In order to honor zero value in shape use allowzero=1
  warnings.warn("allowzero=0 by default. In order to honor zero value in shape use allowzero=1")
WARNING: The shape inference of prim::PadPacked type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
WARNING: The shape inference of prim::PadPacked type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
Segmentation fault

PS: you can access my script here: https://github.com/miccio-dk/NISQA Notice that run_export.py behaves exactly like run_predict.py except I'm forcing CPU.

[miccio-dk]

Other update: this issue is currently tracked on pytorch repo: https://github.com/pytorch/pytorch/issues/75383

[gabrielmittag]

Thanks, hopefully, someone will be able to help in the PyTorch repo. I haven't converted this particular model to ONNX before but a workaround you could try is to remove the packed sequence parts altogether if this is causing the errors. Actually, that's what I used to do with other models because ONNX didn't use to support packed sequences then.

[gabrielmittag]

Closing this as it is not directly related to the model but rather to PyTorch. BTW - if you want to export it to ONNX you probably need to use the model without adaptive pooling layers and without packed sequences. Then it should work

[JBloodless]

@gabrielmittag do I need to retrain the model after removing packed sequences or there are some kind of workaround here? Stuck on similar problem, only packed sequences are the issue for me

[wuzhuohong]

Update: when trying to export the model from Linux (or WSL in my case), I get the following, more verbose errors:

Device: cpu
Model architecture: NISQA_DIM
Loaded pretrained model from weights/nisqa.tar
/home/username/miniconda3/envs/nisqa/lib/python3.9/site-packages/torch/onnx/utils.py:1294: UserWarning: Provided key output for dynamic axes is not a valid input/output name
  warnings.warn("Provided key {} for dynamic axes is not a valid input/output name".format(key))
WARNING: The shape inference of prim::PackPadded type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
WARNING: The shape inference of prim::PackPadded type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
/home/username/miniconda3/envs/nisqa/lib/python3.9/site-packages/torch/onnx/symbolic_helper.py:258: UserWarning: ONNX export failed on adaptive_max_pool2d because input size not accessible not supported
  warnings.warn("ONNX export failed on " + op + " because " + msg + " not supported")
/home/username/miniconda3/envs/nisqa/lib/python3.9/site-packages/torch/onnx/symbolic_helper.py:716: UserWarning: allowzero=0 by default. In order to honor zero value in shape use allowzero=1
  warnings.warn("allowzero=0 by default. In order to honor zero value in shape use allowzero=1")
WARNING: The shape inference of prim::PadPacked type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
WARNING: The shape inference of prim::PadPacked type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function.
Segmentation fault

PS: you can access my script here: https://github.com/miccio-dk/NISQA Notice that run_export.py behaves exactly like run_predict.py except I'm forcing CPU.

I wondor if you have solve this problem. Same Problem I met.