cutmix alpha argument in references/classification/transforms.py

[eminorhan]

I believe the alpha argument here should be cutmix_alpha not mixup_alpha:

vision/references/classification/transforms.py

Line 22 in 61d97f4

transforms_module.CutMix(alpha=mixup_alpha, num_classes=num_classes)

Same issue here as well:

vision/references/classification/transforms.py

Line 24 in 61d97f4

else RandomCutMix(num_classes=num_classes, p=1.0, alpha=mixup_alpha)

Just curious: were the pretrained models actually trained using this code? Any chance this might have affected the performance of the pretrained models in any way?

[NicolasHug]

Great catch @eminorhan , thanks a lot for the report. Would you like to submit a PR to fix these 2 lines?

were the pretrained models actually trained using this code? Any chance this might have affected the performance of the pretrained models in any way?

The bug was introduced in #7731, but all the models we have published in torchvision were trained prior to that change, so the bug doesn't affect the models available within torchvision. However, it may have affected models trained by users.