fast-autoaugment

Why we included this project

Teams training convolutional networks for image classification often spend more time hand-tuning augmentation than any other part of the pipeline. This repo implements Fast AutoAugment, a NeurIPS 2019 method that searches for effective augmentation policies automatically, cutting search time from days to a few GPU hours compared with the original AutoAugment approach. Instead of guessing whether random crop or color jitter helps your dataset, it uses a density-matching strategy to pick transformations that improve generalization, and ships with trained CIFAR-10 and CIFAR-100 weights for several architectures including Wide-ResNet and Shake-Shake. It is research-grade code, so expect to adapt it for your own data and model rather than plugging it in as a library, but it is a solid reference for anyone wanting to replicate or build on automated augmentation search. If you work on vision problems where more data is hard to get, this is a practical way to squeeze accuracy out of the data you already have.