Thank you for sharing a very clean and extensible repository. I am confused regarding the metric that you are reporting for ImageNet-C. Is it just the average of the per-corruption average error rate? Or is it the relative mCE (relative w.r.t. AlexNet as reported in the original paper)?
I have tried the average of the per-corruption average error rate and it is vastly different (mean error rate of 27.91% for ViT-H and 37.98% for ViT-B). Although these results are better in terms of robustness, I am assuming that you reported something else which I would like to reproduce. So just want to clarify that.
Thank you once again for putting in a lot of effort for this!
Hi.
Thank you for sharing a very clean and extensible repository. I am confused regarding the metric that you are reporting for ImageNet-C. Is it just the average of the per-corruption average error rate? Or is it the relative mCE (relative w.r.t. AlexNet as reported in the original paper)?
I have tried the average of the per-corruption average error rate and it is vastly different (mean error rate of 27.91% for ViT-H and 37.98% for ViT-B). Although these results are better in terms of robustness, I am assuming that you reported something else which I would like to reproduce. So just want to clarify that.
Thank you once again for putting in a lot of effort for this!