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

Same issue here as well:

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?