Reproducing results

[katdimitris]

Hello,

Thanks for the really interesting paper. I'm trying to reproduce the results of the paper for densenet in cifar100. I'm using your code with the config for resnet, since no config is provided for densenet, and no parameter differences are mentioned in the paper.

I run for 3 seeds with mixup and get the following Top-1 val acc:

for imb=0.01: ETF+DR=43.02 ± 0.56, Learnable+CE= 44.66 ± 0.24 (versus the reported ETF+DR=44.0 and Learnable+CE= 43.08)
for imb=0.02: ETF+DR=48.45 ± 0.10, Learnable+CE= 50.57 ± 0.46 (versus the reported ETF+DR=49.8 and Learnable+CE= 49.07)

Also, the same behavior is observed without mixup. In cifar10, the ETF+DR method outperforms CE. Is there any training detail I'm missing in cifar100?

Thank you!