search

Haochen-Wang409 / U2PL

[CVPR'22] Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels
Apache License 2.0
426 stars 59 forks source link

Unsupervised loss is 0 #156

Closed tanveer6715 closed 10 months ago

tanveer6715 commented 1 year ago

Hi I am training this model for my own dataset having 2 classes. However I have noticed that it unsupervised loss is zero after some epochs. I have also attached the training log. What could be the possible reason for zero unsupervised loss? seg_20230929_143809.txt

Haochen-Wang409 commented 11 months ago

Could you provide a detailed explanation of your dataset? From the provided log, it seems that the mIoU is increasing during training as expected.

tanveer6715 commented 11 months ago

My dataset have 2 classes, background and cracks. However in the log the unsupervised loss is around zero. By comparison with the suponly, the suponly mIoU is greater then semi supervised U2PL in my case. So I thought the zero unsupervised loss may be the problem.

Haochen-Wang409 commented 11 months ago

I guess it might be because the teacher and the student are quite similar. You could try a larger ema_decay, e.g., 0.999.

tanveer6715 commented 11 months ago

ok thank you. I will try larger ema_decay and will checked if its performance improved.