CPL thresh warmup can lead to wrong label assignments

[Saurav-31]

TorchSSL/models/flexmatch/flexmatch_utils.py

Line 44 in cb27456

mask = max_probs.ge(p_cutoff * (class_acc[max_idx] / (2. - class_acc[max_idx]))).float() # convex

When thresh_warmup is True, since Classwise_acc is mostly zero for all labels in 1st few epochs/iterations, loss for most of unlabeled samples are not masked out. Do you think this could create wrong associations among data. I understand that supervised labeled samples also counter some of this problem but it may also destabilize the training. Any thoughts on this would help. Thanks in advance.