search

LiJunnan1992 / DivideMix

Code for paper: DivideMix: Learning with Noisy Labels as Semi-supervised Learning
MIT License
543 stars 84 forks source link

Can you share log file of training cifar10 of asym noise? #15

Closed wjun0830 closed 4 years ago

wjun0830 commented 4 years ago

Hi Junnan. I really appreciate your work and it helps me study a lot actually.

However, somehow, I cannot get the result as same as you mentioned in your paper. It shows that asym 40% noise on cifar-10 should give 93% accuracy but what i see is middle of 80s.

Can you share your log file of asym0.4-cifar10? image

This is image of middle of the process.

LiJunnan1992 commented 4 years ago

Hi, thanks for trying our code! For asym noise, you need to set lambda_u = 0, as shown in our paper's appendix. Also, you need to finish 300 epochs to get the reported performance.

wjun0830 commented 4 years ago

Oh i missed that. Thanks

MrChenFeng commented 4 years ago

Hi Junnan, could you explain a bit why unlabeled part loss should be avoid in asym noise mode? Thanks so much!

LiJunnan1992 commented 4 years ago

Hi, the labeled loss already contain information for the unlabeled samples because of mixup, and we found the extra regularization of the unlabeled loss to be not necessary.