hongyi-zhang / mixup

Implementation of the mixup training method

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

How to get results reported in paper for cifar-10?

arunpatala opened this issue · comments

Thanks for a very interesting paper!! I was trying out your code and was not able to get the results posted in paper but got the results shown on GitHub page.

CIFAR10 PreAct ResNet-18 results of 3.9% are reported in the paper. But the GitHub page gives it 4.24%. Are there any other setting to change like epochs, weight decay etc?

Thanks

Hi, thanks for your interest! For unknown reasons we were not able to reproduce the 3.9% result (at least with the current pytorch-cifar repo). We already updated the ICLR submission with the new numbers; will do so for the arXiv version as well.

Best,
Hongyi

Thanks. I was able to get 3.9% by running for longer epochs. I tried with 400 epochs.

Thanks for your input! Closing the issue.