Reproduction of the DebiasPL about semi-supervised-learning

[Hhhhhhao]

Good catch. Your understanding is correct. Setting tau as 0 would make DebiasPL the same as FixMatch in our implementation.

I think the difference lies in the learning rate scheduler. As you can see the best accuracy of the experiments you ran is similar (also similar to our reported results). The problem is, these imbalanced algorithms do not use any schedulers (in their paper setup and open-source code), so I set the scheduler as None in ImbAlgorithm Base class at here:

Semi-supervised-learning/semilearn/core/algorithmbase.py

Line 531 in b979a23

scheduler = None

This means DebiasPL with tau=0 would also not use any scheduler.

But if you run FixMatch, it falls into original Algorithm Base class, which will use the cosine scheduler, causing performance drop at later iterations.

[seekingup]

Thanks, after checking the training logs, I found the differences.
However, I also checked the paper and open-source code of DebiasPL. It seems DebiasPL uses cosine scheduler as well. Could you kindly give me some sources of why imbalanced algorithms do not use any schedulers?

[Hhhhhhao]

You can check ABC paper: https://arxiv.org/pdf/2110.10368.pdf.
I believed other papers like DARP and DASO also mentioned no learning rate scheduler in their paper.

[seekingup]

thank you so much!