mechanicalsea / lighthubert Goto Github PK

Hello, I have a question about 10-hour ASR fine-tuning in your paper.

Can you give me a procedure about this experiment? (or the link I can refer)
I just want to conduct the my own experiments for 10-hour ASR fine-tuning using fairseq.

Thanks!

This work is great, and the performence of light hubert is even better than Hubert-Large(according to ).
So I was wondering how to train a light hubert model. Can you opensource the training code?

Hi, this work is awesome and helps me a lot. But I don't how to save the subnet into a checkpoint , could you provide some ways to get the checkpoint of subnet? Thanks a lot.

Hello Mr. Wang!

First of all, I would like to thank you for your work and effort to make it open source.
I've been working on the robustness of SRL models and I'm trying to reproduce the downstream models from SUPERB.

Do you have the CKPT files generated when training the SUPERB models? If not, could you inform the parameters used in the config.yaml file from the tasks? With this, I could reproduce the numbers in the table.

Best regards,
Heitor

Hi,
I'm trying to reproduce lighthubert_stage1 and lighthubert_small, but got a big performance gap... Could you please supply more details of your training process (such as lr, scheduler or loss function code) for stage1 and stage2 training?

Thank you very much

Hi, thank you for uploading the code. It is really helpful. :)
Is it possible to also upload the code to train the pipeline e2e? Thank you!!

Hi,

Thanks for your great work! I have some questions about the two-stage training. I'd appreciate it if you could share more details.

1. In Stage 2 - Once-for-All Training, which model is used as the teacher? Is it the original HuBERT base, or the distilled model from Stage 1?
2. How is the small supernet initialized in Stage 2? I guess it is also initialized with the distilled model from Stage 1, but their sizes are different?
3. In the ablation study (Table 5), how is the supernet initialized in Stage 2 when Stage 1 is skipped? Is it initialized with the original HuBERT base or is it trained from scratch?

Thank you for your time!

Hello!

Thanks for the great work!
My colleague @edward0804 and I are thinking about integrating lighthubert into S3PRL to enable more research.
Instead of copying all the lighthubert code into S3PRL, we are wondering whether adding a setup.py in this repo would be a good alternative so that we can simply install it, enabling lighthubert in the S3PRL codebase, and link the interested user to this repo for the actual implementation.

I have made a minimal fork for this and so lighthubert can be installed in S3PRL after this commit s3prl/s3prl@07c5bd8, and @edward0804 is working on adding a wrapper for lighthubert. Do you think it would be nice to add an official setup.py ? :)

Thanks!