Constrained Sampling-Based Evolutionary Neural Architecture Search for GANs

In recent years, many researchers have adopted neural architecture search (NAS) techniques to automatically design generative adversarial networks (GANs). However, due to the inaccurate evaluation of NAS and the unstable training of GANs, the combination of NAS and GANs (NAS-GANs) remains unsatisfactory. Therefore, we propose a constrained sampling-based evolutionary NAS method for GANs (named CEAGAN). Initially, we introduce a stage-wise constrained sampling strategy to sample GAN’s architectures of different parameter sizes separately and impose constraints on the parameter distributions within each architecture to enhance stability of the search. Additionally, we propose a multi-objective decomposition selection strategy to simultaneously consider model size, the Inception Score (IS), and the Fréchet Inception Distance (FID) to increase the diversity of the search. CEAGAN completes the search on CIFAR-10 in only 0.35 GPU days, demonstrating promising performance on CIFAR-10 (IS = 8.83 ± 0.13, FID = 10.23) and outperforming previous NAS-GANs on STL-10 (IS = 10.30 ± 0.15, FID = 20.29).