quark0/darts

Relaxes the architecture search space to be continuous, enabling gradient-based optimization via backpropagation rather than discrete search methods. Achieves single-GPU efficiency by using second-order approximation with unrolled validation steps, making NAS practical compared to prior methods requiring multiple GPUs. Targets PyTorch and supports both vision tasks (CIFAR-10, ImageNet) and NLP tasks (Penn Treebank, WikiText-2) with learned cell-based architectures that generalize across dataset sizes.

3,993 stars. No commits in the last 6 months.