Randaugment: Practical automated data augmentation with a reduced search space

RandAugment tackles a practical obstacle to automated data augmentation: existing strategies achieve strong results but require a separate and costly search phase, frequently carried out on a smaller proxy task whose tuned hyperparameters may not be optimal for the real task. Rethinking this design, the authors show that rather than searching the magnitude and probability of every augmentation operation independently, it is enough to search over a single distortion magnitude that jointly controls all operations, drastically shrinking the search space and eliminating the proxy task.

Despite its simplicity, the method matches or exceeds previous automated augmentation approaches across CIFAR-10/100, SVHN, ImageNet, and COCO. It attains 85.0% ImageNet accuracy with EfficientNet-B7 (a 1.0% gain over baseline augmentation and 0.6% over AutoAugment) and 85.4% with EfficientNet-B8, matching a prior result that relied on 3.5 billion extra images, while also improving object detection by 1.0-1.3% over baseline augmentation.