CutMix + Mixup overhaul

[rwightman]

Finally a cleaned up impl of CutMix, integrated with Mixup.

A number of mode options... some examples (command line arg combos), to cut & paste into docs at some point

• Mixup active w/ alpha 0.2, performed on every batch, lambda sampled for each batch (old behaviour)
  • --mixup 0.2
• Mixup active w/ alpha 0.3, cutmix active with alpha 1.0, lambda sampled for each batch, cutmix used 50% of the time
  • --mixup 0.3 --cutmix 1.0
• Mixup active w/ alpha 0.2, cutmix active with alpha 1.0, cutmix used 30% of the time, a different lambda and cutmix vs mixup decision sampled for each batch element
  • --mixup 0.2 --cutmix 1.0 --mixup-switch-prob 0.3 --mixup-elem
• Mixup active w/ alpha 0.2, cutmix active with minmax ratio set to [0.3, 0.8], cutmix used 30% of the time, some sort of mixing active 70% of the time, a different lambda, cutmix vs mixup, mixing active sampled for each batch element
  • --mixup 0.2 --cutmix-minmax 0.3 0.8 --mixup-switch-prob 0.3 --mixup-prob 0.7 --mixup-elem

Fix #8

[thepowerfuldeez]

Hi! Do you have in mind to test training with Attentive Cutmix? It seems that this simple technique improves scores by 1%
https://arxiv.org/abs/2003.13048