randaugment

**RandAugment** is a **simple, automated data augmentation strategy that randomly selects $N$ transformations from a pool and applies them with a fixed magnitude $M$** — eliminating the need for a separate search phase (unlike AutoAugment), with just two hyperparameters. **How Does RandAugment Work?** - **Pool**: ~14 transformations (rotation, shear, translate, brightness, contrast, equalize, etc.). - **Sample**: Randomly pick $N$ transformations (typically $N = 2-3$). - **Apply**: Apply each with the same global magnitude $M$ (typically $M = 9-15$ on a 0-30 scale). - **Two Hyperparameters**: Only $N$ and $M$ to tune. No separate search phase. - **Paper**: Cubuk et al. (2020). **Why It Matters** - **Simplicity**: Two hyperparameters ($N$, $M$) vs. AutoAugment's expensive policy search. - **Competitive**: Matches or exceeds AutoAugment accuracy despite being vastly simpler. - **Standard**: The default augmentation strategy in EfficientNet, ViT, FixMatch, and modern training recipes. **RandAugment** is **augmentation without the search** — a dead-simple two-parameter strategy that rivals expensive learned augmentation policies.