On neural architecture search and hyperparameter optimization: A max-flow based approach

Chao Xue, Jiaxing Li, Xiaoxing Wang, Yibing Zhan, Junchi Yan, Chun-Guang Li

Abstract: Highlights•NAS and HPO are formulated as a max-flow problem.•Search space and search strategy are handled in a fully graphical manner.•Warmup scheme and weight-sharing method make the search efficient.•Competitive results are shown across different datasets and search spaces.