robustness

Why we included this project

Anyone who has shipped an image classifier knows that clean validation accuracy rarely tells the full story; real-world inputs come with blur, noise, and weather effects that can wreck a model's confidence. This repository from the ICLR 2019 paper by Dan Hendrycks and Thomas Dietterich gives you the standard tools to measure that: ImageNet-C, ImageNet-P, and the CIFAR-C/P variants, plus the code and scoring rules for running them. Instead of judging a model only on clean images, you evaluate it across 15 common corruptions and perturbations that stress temporal stability, and get a reproducible mean Corruption Error or mean Flip Rate to compare. The maintained leaderboards show what real gains state-of-the-art methods achieve, so you can benchmark your own work against published baselines without guessing.