Hi Ross,

Thank you for including AugMix in your wonderful library!

It looks like the augmentation severity parameter is currently fixed at the maximum allowed severity (unless explicitly enabled with magnitude_std):

https://github.com/rwightman/pytorch-image-models/blob/4203efa36d0258315f51ad8f945849586b058408/timm/data/auto_augment.py#L337-L345

However, AugMix actually samples different severities uniformly at random in the range [0, magnitude] for each application of each op:

https://github.com/google-research/augmix/blob/master/augmentations.py#L70

Sorry this wasn't more clear in our implementation and paper!

Norman