概述

stylegan2

1. RandAugment has 18 data augmentation -> 6 categories:
   1. They can be unleaking if it is only executed at a probability $p<100\%$.
   2. Explanation.
   3. pixel blitting (x-flips, 90◦ rotations, integer translation), more general geometric transformations, color transforms. They can improve. Additive noise, cutout can't.
2. Adaptive Discriminant Augmentation for the probability $p$.
   1. Add augmentation to both of Discriminator and Generator.
   2. $r_t = \mathbb{E}[sign (D_{train})]$ indicates the overfitting
3. Init $p=0$, adjust $p$ every four batches, based on $r_t$.

---