Concerns about loss_u

[HaHaHaHaWang]

Thanks for your work.
I am a beginner, so I'm quite puzzled by the following issue.
I trained models using three different algorithms: fixmatch, comatch, and fixmatch_ccssl, following the programs you provided. I noticed that at the beginning of training, the loss_u is always 0 and starts increasing from there. Is this normal? I can't determine if this is due to errors in my modifications.

[KaiWU5]

This is normal. At the initial training stage, the model is noisy and less capable of generating pseudo labels. So in semi-supervised learning, we usually first train supervised data for several steps and then train for unsupervised steps.

[HaHaHaHaWang]

Thank you for your reply. I have recently reproduced the experiment of FixMatchCCSSL on STL10, but I cannot achieve the results in your paper. Would you be willing to make public the Config of FixMatchCCSSL on STL10?

[KaiWU5]

Thanks for your attention to our work. It's been a while since we published the paper, I will try to find or reproduce the model for STL10. It could take a while since most of my effort now switches to large model training.