We proposed a new neural architecture search (NAS) benchmark (NAS-Bench-201). We provided fine-grained training/evaluation results and useful diagnostic information of 15,625 architectures on three vision datasets. We also open-sourced 10 recent NAS algorithms in a single codebase.
Figure 1.. We summarize some characteristics of NAS-Bench-101 and NAS-Bench-201. Our NASBench-201 can directly be applicable to almost any up-to-date NAS algorithms. In contrast, as pointed in (Ying et al., 2019), NAS algorithms based on parameter sharing or network morphisms cannot be directly evaluated on NAS-Bench-101. Besides, NAS-Bench-201 provides train/validation/test performance on three (one for NAS-Bench-101) different datasets so that the generality of NAS algorithms can be evaluated. It also provides some diagnostic information that may provide insights to design better NAS algorithms.
Figure 2.. The search space used in NAS-Bench-201.
Video 1. Presentation, please play with the video of slides. (Use Chrome to play the video.)
Video 2. Slides. (Use Chrome to play the video.)