Understanding and Improving One-shot Neural Architecture Optimization

The ability of accurately ranking candidate architectures is the key to the performance of neural architecture search~(NAS). One-shot NAS is proposed to cut the expense but shows inferior performance against conventional NAS and is not adequately stable. We find that the ranking correlation between architectures under one-shot training and the ones under stand-alone training is poor, which misleads the algorithm to discover better architectures. We conjecture that this is owing to the gaps between one-shot training and stand-alone complete training. In this work, we empirically investigate several main factors that lead to the gaps and so weak ranking correlation. We then propose NAO-V2 to alleviate such gaps where we: (1) Increase the average updates for individual architecture to a relatively adequate extent. (2) Encourage more updates for large and complex architectures than small and simple architectures to balance them by sampling architectures in proportion to their model sizes. (3) Make the one-shot training of the supernet independent at each iteration. Comprehensive experiments verify that our proposed method is effective and robust. It leads to a more stable search that all the top architectures perform well enough compared to baseline methods. The final discovered architecture shows significant improvements against baselines with a test error rate of 2.60% on CIFAR-10 and top-1 accuracy of 74.4% on ImageNet under the mobile setting. Code and model checkpoints are publicly available at https://github.com/renqianluo/NAO_pytorch.