pba

Why we included this project

People who train image classifiers on a single GPU often skip automatic augmentation search because the usual methods demand too much compute. PBA's contribution is scheduling: instead of one fixed policy, it learns a schedule of augmentation policies that change as training progresses, and the paper reports CIFAR-10, CIFAR-100, and SVHN results at roughly one-thousandth of the compute of comparable searches. The TensorFlow code has two paths: scripts that reproduce the published schedules, and a fresh population-based search if you want policies tuned for your own data. It is a reference implementation and study baseline more than a plug-in library, so treat it as a useful comparison point when you evaluate newer automated augmentation tools.