Converting wav2vec2-base-960h to ONNX report an error while converting

[Denovitz]

First of all, I want to say thanks to @patrickvonplaten for the work done in adding the model. Great job!

I tried to convert the model to ONNX but got an error, do you have any ideas how to fix it?

What I did:
python -m transformers.convert_graph_to_onnx --framework pt --model facebook/wav2vec2-base-960h wav2vec2-base-960h.onnx

But got an error:

ONNX opset version set to: 11
Loading pipeline (model: facebook/wav2vec2-base-960h, tokenizer: facebook/wav2vec2-base-960h)
Using framework PyTorch: 1.7.0
Error while converting the model: Unable to create tensor, you should probably activate truncation and/or padding with 'padding=True' 'truncation=True' to have batched tensors with the same length.

[mironnn]

Also interested in this question!

[patrickvonplaten]

Hmm, Wav2Vec2 is still a very recent addition and I don't have a good idea on an ETA for full ONNX support. However, I think your error above is due to the input that's passed to Wav2Vec2Tokenizer being a string instead of a speech input. So in order to make the conversion work, you will have to tweak the script convert_graph_to_onnx yourself a bit for Wav2Vec2 - I think the only different should be that instead of passing it "This is a sample output" you should pass it a 1D float array.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[OthmaneJ]

Hi!
very interested in this question! did anyone managed to make it work ?

[patrickvonplaten]

Hey @OthmaneJ,

Think @ccoreilly managed to get it to work here: https://github.com/ccoreilly/wav2vec2-service/blob/master/convert_torch_to_onnx.py

[OthmaneJ]

@patrickvonplaten thanks! 👌

[marziye-A]

hi @patrickvonplaten
is there any way to transform mms asr model to onnx?
if yes, how?
thank you very much!