search

HobbitLong / RepDistiller

[ICLR 2020] Contrastive Representation Distillation (CRD), and benchmark of recent knowledge distillation methods
BSD 2-Clause "Simplified" License
2.11k stars 389 forks source link

Reported results based on early stopping? #16

Closed VladimirLi closed 4 years ago

VladimirLi commented 4 years ago

Thanks for sharing this repo.

I noticed that you store the best model based on test accuracy. I wonder whether the published results are also based on the best

HobbitLong commented 4 years ago

That's a good question.

The published results are based on the last epoch. So you will find that the best test accuracy is higher than that reported in the paper.

VladimirLi commented 4 years ago

Thank you very much for your response!

MingSun-Tse commented 3 years ago

One small question. Do you think, reporting the best accuracy should be more standard? since the last accuracy typically has larger fluctuation than the best one? Thanks!

That's a good question.

The published results are based on the last epoch. So you will find that the best test accuracy is higher than that reported in the paper.