winycg/Self-KD-Lib

reproducing FRSKD issue

Closed this issue · 6 comments

bkw0622 commented

Hi!
I am grateful for your good work.
Now I'm learning FRSKD with your code, but the performance is about 62.xx, which is different from what is specified in github.
How can I fix this error?
(The code is nothing changed)

Thanks a lot!!

winycg commented

Hi, does the result use ResNet-18 as the backbone? This result seems abnormal. I will run this code again and tell you new result.

winycg commented

Hi, thanks for pointing out this bug. The problem is that we use imagenet version's ResNet-18 as the backbone for FRSKD.Now I have fixed this bug, and re-run FRSKD and achieves 76.6% top-1 accuracy on CIFAR-100. Since the architecture of BiFPN in FRSKD can be more complexed, you may tune the hyper-parameters to pursue better performance.

yhaochen0617 commented

hihi, thank you for sharing your work~
Could you share your train_log of ImageNet (MixSKD and mixup baseline) to me ?
My email is yhao.chen0617@gmail.com. Thank you very much.

winycg commented

@yhaochen0617 Hi, I sent the train_log of MixSKD on ImageNet to you. The mixup baseline log is unavailable at hand.
Thx!

bkw0622 commented

Thanks for the quick reply!

winycg commented

@bkw0622 By the way, the result of FRSKD is borrowed from their original paper, and I did not reproduce this result. Besides, other results are reproduced by ours. I will modify it.