search

lorenmt / reco

The implementation of "Bootstrapping Semantic Segmentation with Regional Contrast" [ICLR 2022].
https://shikun.io/projects/regional-contrast
Other
162 stars 25 forks source link

About Weak and Strong threshold values. #33

Closed nysp78 closed 1 year ago

nysp78 commented 1 year ago

Hello again, The values of weak(0.7) and strong(0.97) thresholds are used to filtered softmax ouputs or raw logits outputs? I'm saying this, because in the computation of the unsupervised loss I do not see any application of softmax. Are these thresholds probability values?

Thanks

lorenmt commented 1 year ago

Softmax probability output. It's probably clearer in the partial label setting.

https://github.com/lorenmt/reco/blob/71c1b7324fd36049c6657b122ba3fef373877b72/train_semisup_partial.py#L118-L122