The jax_checkpoint about Private Laparoscopy

[isyangshu]

I wanted to express my appreciation for your outstanding work and the valuable insights I gained from reading your paper. It has been an enlightening experience for me.

However, I have a couple of questions that I would like to bring to your attention:

• I noticed that the pre-trained ViT-B model of the Private Laparoscopy dataset, available through the GitHub link you provided, appears to be incomplete. I was wondering if it would be possible for you to share the corresponding pre-training weights of ViT-B with me. I have successfully utilized other pre-trained checkpoints, with the exception of VIT-B with Private Laparoscopy dataset.

• In your paper, you reported that the model, pre-trained on a private dataset and fine-tuned on Cholec80, achieved an impressive F1-score of 85.8. However, upon reviewing the table on GitHub, I noticed that the reported F1-score is 82.6. Could you kindly clarify if this is a copy-paste error or intentional? I found the results of the Private Colonoscopy task to be accurate.

Hope for your reply!

[RoyHirsch]

Thank you for your interest in our project! I'll look deeper into your questions and answer as soon as I can. Can you share the error message you get when trying to load ViT-B Private Laparoscopy model?

[isyangshu]

Thanks for your reply. I can easily use followed code to get the parameters in all the other five links, since I want to use the pre-trained model in Pytorch. from flax.training import checkpoints loaded_params = checkpoints.restore_checkpoint( "xxxxxx/checkpoint_0", target=None, ) load_state_dict = loaded_params["student_weights"] But when I used the code to load "checkpoint_0_vitb" in https://drive.google.com/drive/folders/18-IF-tply94_71_xtMcQCMMhVPRjwdEF. I got "ValueError: Unpack failed: incomplete input". The model in this link is only 200+M, which may be unreasonable.

[ZardCAI]

I meet the same problem of the ViT-B model of the Private Laparoscopy dataset. I guess this may be a upload error caused by network fluctuations? Could the author upload a completed model? Thanks a lot!