search

HuangYG123 / CurricularFace

CurricularFace(CVPR2020)
MIT License
526 stars 72 forks source link

some different between implementation and paper #8

Closed superhero1991 closed 4 years ago

superhero1991 commented 4 years ago

Hi, When you update the t. you use 0.99 as the alpha in the paper. but only 0.01 in the code. So, which one is the right version. In my opinion, the code is right.

HuangYG123 commented 4 years ago

Sorry for the mistake. The code is the right version

superhero1991 commented 4 years ago

Sorry for the mistake. The code is the right version

@HuangYG123 One more question. when I try to train a new model(use resnet18 as the backbone). I found the t value is 0.4 in first stage(it's 0.6 in your paper). I don't know why there has a gap. It's depends on the backbone? Thank you very much

HuangYG123 commented 4 years ago

@superhero1991 Since the value of t is based on the positive similarities，small backbone may be harder to increase the positive similarites.

superhero1991 commented 4 years ago

@superhero1991 Since the value of t is based on the positive similarities，small backbone may be harder to increase the positive similarites.

@HuangYG123 thank you very much. Is it possible to add a small value on t? Maybe it hard to convergence.