purpose of batch norm controller

[hzhz2020]

Dear Authors

Thanks for providing this great repo!
You mentioned in your FlexMatch paper that a batch norm controller is introduced in the codebase to prevent performance crashes for some algorithms. You mentioned that Mean Teacher, Pi-model and MixMatch might be unstable if update Batchnorm for both labeled and unlabeled data in turn. Does this have to do with multi-GPU training? (if I use a single GPU, will this instability persist?)

Alternatively, can i simply freeze the batchnorm when forwarding the unlabeled batch?

Hope to hear from you. Thanks!