How to use the rehearsal buffer?

[zyuh]

Hi, amazing work!

I noticed in the paper that l2p can use the rehearsal buffer to further improve performance, but the repository doesn't seem to include code implementation for this part. I have a few questions about the implementation of this part:
(1) Random sampling or herding sampling?
(2) In addition to having old samples in dataloader, are any other operations used, such as distillation, balanced fine-tuning, etc.
(3) Will the official implementation of this part be added to the code later?
(4) Whether the rehearsal buffer can further improve the performance of "dualprompt" also?

Looking forward to your reply, best wishes！

[JH-LEE-KR]

Hi,
thanks for your comment.

Unfortunately, it has not been implemented yet.
It is difficult to verify that the code that pytorch version is implemented correctly because the config for the rehearsal buffer is not opened in the official repo.

• (1) Looks like a random (np.random.choice) sampling.
• (2) I don't understand the intent of your question.
• (3) I've already implemented it.
  However, there is no official config, so I cannot verify whether it is implemented correctly. Issue
  If only the config is released, it will be updated to pytorch implementation.
• (4) In my experience, if you apply it naïve, the performance will decrease.

Best,
Jaeho Lee.