search

Haochen-Wang409 / U2PL

[CVPR'22 & IJCV'24] Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels & Using Unreliable Pseudo-Labels for Label-Efficient Semantic Segmentation
Apache License 2.0
436 stars 61 forks source link

Contrastive loss does not decrease #137

Closed hweejuni closed 1 year ago

hweejuni commented 1 year ago

Hi. I'm using U2PL with my custom dataset which has only 3 classes. And I realised that the contrastive loss does not decrease at the range of 6~8 and I couldn't figure it out why. What would be the problem? and also, is it a problem if the unlabeled dataset is small by comparison? Thank you.

Haochen-Wang409 commented 1 year ago

Contrastive loss acts like a regularizer instead of an optimization objective. Therefore, it may not decrease but the performance is expected to keep increasing.

I think it is not appropriate to make the unlabeled dataset to be small.