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 cifar noise file #10

Closed ShiningSord closed 4 years ago

ShiningSord commented 4 years ago

Hello, I have repeated experiments on cifar10 and cifar100, and I find that accuracy is affected by noise file when noise ratio is high(0.8/0.9). I got only 54.6/27.9 on cifar100 with 0.8/0.9 noise ratio, which are much lower than that claimed in paper(60.2/31.5).

so, can you share the noise file? Thanks:)

LiJunnan1992 commented 4 years ago

Hi, thanks for trying out our code! Have you adjusted the hyper-parameters for high noise ratio? According to our paper, lambda_u=150, and p_threshold=0.6 for 90% noise.

ShiningSord commented 4 years ago

Thanks for your reply! i retried cifar100 with 0.8/0.9 noise ratio, and my hyper-parameters are lambda_u=150, p_threshold=0.5/0.6. the best accuracy on test set are 57.63/30.5. Are there more hyper-parameters should be adjusted?

ShiningSord commented 4 years ago

I made a mistake on the setting of batchsize. After I correct it, the result is just same as that claimed in paper! Thanks for your nice work, and I will close this issue.