A UTO HAS : E FFICIENT H YPERPARAMETER AND A R-CHITECTURE S EARCH

Deep learning models often require extensive efforts in optimizing hyperparameters and architectures. Standard hyperparameter optimization methods are expensive because of their multi-trial nature: different configurations are tried separately to find the best. In this paper, we propose AutoHAS, an efficient framework for both hyperparameter and architecture search. AutoHAS generalizes the concept of efficient architecture search, ENAS and DARTS, to hyperparameter search and hence can jointly optimize both in a single training. A key challenge in such generalization is that ENAS and DARTS are designed to optimize discrete architecture choices, whereas hyperparameter choices are often continuous. To tackle this challenge, we discretize the continuous space into a linear combination of multiple categorical basis. Furthermore, we extend the idea of weight sharing and augment it with REINFORCE to reduce its memory cost. In order to decouple the shared network weights and controller optimization, we also propose to create temporary weights for evaluating the sampled hyperparameters and updating the controller. Experimental results show AutoHAS can improve the ImageNet accuracy by up to 0.8% for highly-optimized state-of-the-art ResNet/EfficientNet models, and up to 11% for less-optimized models. Compared to random search and Bayesian search, AutoHAS consistently achieves better accuracy with 10x less computation cost.