One-Shot Neural Architecture Search by Dynamically Pruning Supernet in Hierarchical Order

Neural Architecture Search (NAS), which aims at automatically designing neural architectures, recently draw a growing research interest. Different from conventional NAS methods, in which a large number of neural architectures need to be trained for evaluation, the one-shot NAS methods only have to train one supernet which synthesizes all the possible candidate architectures. As a result, the search efficiency could be significantly improved by sharing the supernet's weights during the candidate architectures' evaluation. This strategy could greatly speed up the search process but suffer a challenge that the evaluation based on sharing weights is not predictive enough. Recently, pruning the supernet during the search has been proven to be an efficient way to alleviate this problem. However, the pruning direction in complex-structured search space remains unexplored. In this paper, we revisited the role of path dropout strategy, which drops the neural operations instead of the neurons, in supernet training, and several interesting characters of the supernet trained with dropout are found. Based on the observations, a Hierarchically-Ordered Pruning Neural Architecture Search (HOPNAS) algorithm is proposed by dynamically pruning the supernet with a proper pruning direction. Experimental results indicate that our method is competitive with state-of-the-art approaches on CIFAR10 and ImageNet.