This is the official implementation of β-DARTS: Beta-Decay Regularization for Differentiable Architecture Search (CVPR22 oral).

This code is based on the implementation of DARTS, NAS-Bench-201, NAS-Bench-1Shot1 and SmoothDARTS.

Data Preparation: Please first download the 201 benchmark file and prepare the api follow this repository.

Searching and Evaluation: python ./nasbench201/train_search.py

Note that we only use the training set when searching and retraining as DARTS. Please refer to line 135-139 of ./nasbench201/train_search.py to see the splitted training data.

Please refer to ./optimizers/darts/architect.py to find our proposed loss as follows:

Line 55: def mlc_loss(self, arch_param):

Please be noted that only “one line of code” （i.e. mlc loss）is added to make differentiable NAS methods much more Robustness and Generalization !!!

Paper

Reference