Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Inconsistent ONNX Export with Differentiable Head #418

Open
radandreicristian opened this issue Sep 6, 2023 · 5 comments · May be fixed by #446
Open

Inconsistent ONNX Export with Differentiable Head #418

radandreicristian opened this issue Sep 6, 2023 · 5 comments · May be fixed by #446
Labels
bug Something isn't working exporting Regarding exporting SetFit models (e.g. ONNX)

Comments

@radandreicristian
Copy link

Hello,

I am trying to fine-tune SetFit for a multi-class classification problem.

Everything is smooth until exporting to Onnx. The head is not exported correctly, so when loading the model with Onnx, the predictions are the outputs of some previous layer.

For example, I showcased this on a dummy dataset with 3 classes.

See https://colab.research.google.com/drive/19EESqbIDwD5FOI2Ufx22txFZ8qADq2xj?authuser=1#scrollTo=vl6AvQKqtU-9

This differs from the behaviour in the example notebook, where the LogisticRegression head is used.

Any directions would be appreciated; Otherwise, I would happily contribute with an MR if anyone can spot the issue.
@radandreicristian
Copy link
Author

The output is the raw logits per class for each sample, and all you need to do then is a softmax/argmax to get the class labels.

Is this intended?
@tomaarsen
Copy link
Member

Hello!

I'm aware that there's some issues with exporting ONNX sadly. Hopefully I will have some time in the future to refactor the exporting to a more consistent approach instead. Thank you for raising this issue & for providing a Google Colab! It'll be helpful for certain.

  • Tom Aarsen
@tomaarsen tomaarsen added bug Something isn't working exporting Regarding exporting SetFit models (e.g. ONNX) labels Nov 24, 2023
@pedrogengo
Copy link

Hey @tomaarsen, can I work on this?
@pedrogengo
Copy link

After a long debugging, I found the issue.
As we are not defining the argument names on https://github.com/huggingface/setfit/blame/cbc01ec402e86ca04e5e40e9bce7f618f3c2946c/src/setfit/exporters/onnx.py#L50

transformers library assumes that the third argument represents position_idsand not token_type_ids (https://github.com/huggingface/transformers/blob/b09912c8f452ac485933ac0f86937aa01de3c398/src/transformers/models/mpnet/modeling_mpnet.py#L515-L525). The fix for this issue it simply define the arguments name.

I will open a PR to fix this
@pedrogengo
Copy link

https://colab.research.google.com/drive/19xE4WdxqGLLZOSanycYfUzbcAxFgpzuR?usp=sharing

Here is my colab that I used to debug
@pedrogengo pedrogengo linked a pull request Nov 27, 2023 that will close this issue
Open
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
bug Something isn't working exporting Regarding exporting SetFit models (e.g. ONNX)
3 participants
@pedrogengo @radandreicristian @tomaarsen