EAGAN: Efficient Two-Stage Evolutionary Architecture Search for GANs

Abstract

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

EGANS: Evolutionary Generative Adversarial Network Search for Zero-Shot Learning

“…According to the optimized used, existing NAS algorithms can be broadly grouped into three different categories: reinforcement learning (RL) based NAS methods [68], [69], gradientbased NAS methods [39], and evolutionary computation-based NAS (ENAS) methods [59], [70]. The RL-based methods depend on high-computation resources due to using tens of actions to get a positive reward, i.e., often requiring thousands of graphics processing cards (GPUs) with several days even on median-scale data set [42]. The gradient-based methods are more efficient than the RL-based methods.…”

“…Furthermore, existing NAS methods train each discovered architecture individually to obtain their corresponding performance, resulting in substantial computational overhead. Ying [42] proposed a weight-sharing strategy to enhance search efficiency by constructing a large computational graph, in which each subgraph represents a neural network architecture. Consequently, all sub-network architectures can be evaluated without separate training by sharing weights within the large network.…”

“…Although some NAS methods have been introduced to automatically find network architectures for generative models (e.g., GANs) [41], [43], [45], [46], ENAS-based GANs lack of enough attention [47]. Ying [42] and Liu [61] initially introduce an evolutionary algorithmbased NAS framework to search GANs stably, but the models are based on the unconditional image generation task. Although the searched GAN architectures can unconditionally synthesize realistic and sharp images, these approaches are not capable for synthesize images of sufficient quality to train classifiers [14].…”

“…Recently, neural architecture search (NAS) has been proven to be effective for automatically searching superior network architecture with good generalization for various scenarios in various tasks [37], [38], including the discriminative [39], [40] and generative tasks [41], [42]. To reduce search difficulty, the early NAS-based GANs [41], [43] search only generator with a fixed discriminator, resulting in a sub-optimal GAN architecture.…”

“…Considering the success of evolutionary computation (EC) for the optimization GANs [48]- [53] and NAS [54]- [60], evolutionary neural architecture search (ENAS) based GANs models are introduced to improve the optimization and generalization of GANs recently [42], [61]. Although these methods can synthesize realistic and sharp images, they are not yet able to generate images of sufficient quality to train classifiers [8], [14].…”