search

CVI-SZU / CCAM

[CVPR 2022] C2AM: Contrastive learning of Class-agnostic Activation Map for Weakly Supervised Object Localization and Semantic Segmentation
MIT License
190 stars 27 forks source link

test loss does not decrease #12

Closed mwmilad closed 1 year ago

mwmilad commented 1 year ago

hi, I train your code on mocov2 and CUB Dataset but test loss does not decrease... still epoch 17

Screenshot 2023-03-08 013126

Sierkinhane commented 1 year ago

Hi, as the losses work in an adversarial manner, the loss may not be like the normal one. You can check the generated class-agnostic activation map.