Two dials and no proxy task: RandAugment reimplemented in PyTorch

What it does

This repo provides an unofficial PyTorch reimplementation of RandAugment, a data augmentation strategy controlled by just two parameters, N and M. It is designed to drop into existing torchvision transforms without requiring the separate proxy-task search that AutoAugment demands. The author reports results competitive with the original paper on smaller benchmarks by borrowing heavily from the Fast AutoAugment codebase.

The interesting bit

The entire policy search collapses to picking how many transforms to apply (N) and how strongly (M), yet the README shows this minimal approach yielding results similar to the published CIFAR and SVHN benchmarks—with Wide-ResNet 28×10 even edging the paper by a tenth of a point on CIFAR-10. The author is also unusually candid about ImageNet reproduction being unfinished, which is more useful than another silent benchmark gap.

Key highlights

• Unofficial PyTorch port derived largely from the Fast AutoAugment project.
• Drop-in transform for torchvision pipelines; no separate proxy task required.
• Reproduces similar results to the paper on CIFAR-10 (97.4% vs. 97.3%), CIFAR-100 (83.3%), and SVHN (98.8% vs. 98.9%).
• ImageNet experiments are incomplete: ResNet-50 and EfficientNet results are marked TODO, with an open issue tracking the difficulty.
• Only two hyperparameters to adjust: N (number of operations) and M (magnitude).

Caveats

• The codebase is mostly borrowed from Fast AutoAugment, so this is closer to a curated extraction than a ground-up rewrite.
• ImageNet support is unproven; the author explicitly flags difficulties and leaves large-model results blank.
• Documentation beyond benchmark tables and a minimal usage example is thin.

Verdict

Grab it if you want a low-ceremony RandAugment plug-in for CIFAR-scale projects. Look elsewhere if you need proven ImageNet numbers or a fully independent implementation.