pytorch-randaugment

Why we included this project

Training image classifiers in PyTorch usually means either sticking with basic flips and crops or running an expensive search for a better augmentation policy, and this library skips the search by implementing RandAugment as a single transform you drop into a torchvision pipeline with just two knobs, N and M. The author checked it against the original paper on CIFAR-10, CIFAR-100, and SVHN with Wide-ResNet and similar models, and the accuracy came out close to the reported figures. That makes it a useful reference if you want to see the method work in your own setup before committing to it. One caveat: the ImageNet reproduction was never finished, so treat results on large datasets as unverified. The code is small and MIT-licensed, so it is easy to read and fold into an existing training loop.