The 3.5-GPU-hour alternative to AutoAugment

What it does Fast AutoAugment is the official PyTorch reproduction of a NeurIPS 2019 paper that automates the hunt for good image distortion policies. Instead of hand-tuning flips, crops, and color jitters, it searches for combinations that improve model accuracy, then hands you a fixed recipe you can reuse in standard training. The repo bundles the search algorithm, YAML configuration files, and a model zoo with pretrained weights for CIFAR-10/100, ImageNet, and SVHN.

The interesting bit The speedup comes from density matching, a search strategy that the authors claim explores the policy space far more efficiently than the original AutoAugment brute-force approach. On CIFAR-10, search finishes in 3.5 GPU hours—a 1,428× speedup—while ImageNet search drops to 450 GPU hours, or 33× faster, with comparable final accuracy.

Key highlights

• CIFAR-10 policy search finishes in 3.5 GPU hours; ImageNet in 450 GPU hours, with speedups of 1,428× and 33× over the original AutoAugment method.
• Pretrained models are provided for Wide-ResNet, Shake-Shake, PyramidNet, ResNet, and EfficientNet-B0.
• Distributed search relies on Ray, and multi-node training is supported through PyTorch’s distributed launcher.
• The project is a full research artifact: search scripts, training pipelines, configs, and evaluation checkpoints included.

Caveats

• The documented environment is Python 3.6.9, PyTorch 1.2.0, and CUDA 10, so modern compatibility is unclear.
• Setting up the Ray cluster for the search phase is explicitly flagged as a prerequisite, not a built-in single-process fallback.

Verdict A solid stop for researchers reproducing 2019 AutoAugment baselines or anyone who needs a vetted augmentation policy for standard vision benchmarks. If you are looking for a modern, drop-in augmentation library with current dependencies, this is not it.