search

snu-mllab / PuzzleMix

Official PyTorch implementation of "Puzzle Mix: Exploiting Saliency and Local Statistics for Optimal Mixup" (ICML'20)
MIT License
157 stars 17 forks source link

Question about imagenet_fast vanilla performance #10

Closed Neonsign247 closed 1 year ago

Neonsign247 commented 1 year ago

Hello. Thank you for your great work.

Currently, I'm trying to reproduce vanilla performance with your imagenet_fast protocol. However, my model 2-3% behind with your configuration. Can you share tricks (data augmentation or optimized configuration) or training log?

Janghyun1230 commented 1 year ago

Hello,

I used the same training configuration (learning rate, augmentation) for Vanilla and Puzzle Mix, except for mixup. (reference code: https://github.com/anonymous-sushi-armadillo/fast_is_better_than_free_imagenet ) Currently, I don't have log files for Vanilla models.

I guess it could be a problem caused by different versions of apex. Could you reproduce the results with Puzzle Mix?

If the problem maintains, I recommend to use more stable ImageNet training codes without apex: https://github.com/snu-mllab/PuzzleMix/tree/master/imagenet