darts

Why we included this project

DARTS is the reference implementation of the 2018 neural architecture search paper that turned the search itself into a gradient-descent problem: instead of training and scoring candidate networks one by one, it relaxes discrete architecture choices into continuous weights and optimizes them directly. The whole search stays cheap enough to run on a single GPU. The repo ships pretrained cells you can test right away, a CIFAR-10 and ImageNet convolutional pair plus a recurrent cell for language modeling on Penn Treebank and WikiText-2, with expected error rates and perplexities spelled out in the README. It is research code from 2018 pinned to PyTorch 0.3.1, so treat it as a readable baseline for comparison and a study reference, not something to build production training on.