EAGAN: Efficient Two-Stage Evolutionary Architecture Search for GANs

Generative adversarial networks (GANs) have proven successful in image generation tasks. However, GAN training is inherently unstable. Although many works try to stabilize it by manually modifying architecture, requires much expertise. Neural architecture search (NAS) has become an attractive solution GANs automatically. The early NAS-GANs only generators reduce complexity but lead a sub-optimal GAN. Some recent both generator (G) and discriminator (D), they suffer from the instability of training. To alleviate instability, we propose efficient two-stage evolutionary algorithm-based NAS framework GANs, namely EAGAN. We decouple G D into two stages, where stage-1 searches with fixed adopts many-to-one strategy, stage-2 optimal found one-to-one weight-resetting strategies enhance stability Both stages use non-dominated sorting method produce Pareto-front architectures under multiple objectives (e.g., model size, Inception Score (IS), Fréchet Distance (FID)). EAGAN applied unconditional task can efficiently finish on CIFAR-10 dataset 1.2 GPU days. Our searched achieve competitive results (IS = 8.81 ± 0.10, FID 9.91) surpass prior STL-10 10.44 0.087, 22.18). Source code: https://github.com/marsggbo/EAGAN .