[Paper Review] AGAN: Towards Automated Design of Generative Adversarial Networks

Hanchao Wang, Jun Huan|arXiv (Cornell University)|Jun 25, 2019

Generative Adversarial Networks and Image Synthesis39 references39 citations

TL;DR

AGAN presents the first neural architecture search method tailored for GANs, using reinforcement learning to automatically design generator and discriminator architectures that achieve competitive or superior performance compared with human-crafted models on CIFAR-10 and transferability to STL-10.

ABSTRACT

Recent progress in Generative Adversarial Networks (GANs) has shown promising signs of improving GAN training via architectural change. Despite some early success, at present the design of GAN architectures requires human expertise, laborious trial-and-error testings, and often draws inspiration from its image classification counterpart. In the current paper, we present the first neural architecture search algorithm, automated neural architecture search for deep generative models, or AGAN for abbreviation, that is specifically suited for GAN training. For unsupervised image generation tasks on CIFAR-10, our algorithm finds architecture that outperforms state-of-the-art models under same regularization techniques. For supervised tasks, the automatically searched architectures also achieve highly competitive performance, outperforming best human-invented architectures at resolution $32 imes32$. Moreover, we empirically demonstrate that the modules learned by AGAN are transferable to other image generation tasks such as STL-10.

Motivation & Objective

• Motivate automating GAN architecture design to reduce reliance on human expertise and trial-and-error.
• Develop an RL-based NAS framework specifically for GANs (generator and discriminator).
• Explore a modular search space with up-sampling, down-sampling, and normal modules for GANs.
• Demonstrate that discovered AGAN architectures rival or surpass human-designed GANs under the same regularization.
• Investigate transferability of learned modules to other datasets like STL-10.

Proposed method

• Use a controller RNN (two-layer LSTM with three segments) to sample GAN architectures for the generator and discriminator.
• Define a comprehensive search space including up-sampling, down-sampling, and normal modules with various operations (e.g., convolutions, separable convolutions, pooling, and up-/down-sampling schemes).
• Train architectures using hinge loss GAN objectives and evaluate with Inception Score (IS) as the reward signal.
• Apply REINFORCE to update the controller parameters based on a reward shaped by normalized IS, guiding search toward promising architectures.
• Fix meta-architecture details (generator has 3 up-sampling modules; discriminator has a conv, 2 down-sampling modules, 2 normal modules, global sum pooling, and a linear layer).
• Demonstrate transferability by testing learned modules on STL-10 and adjusting input sizes accordingly.

Experimental results

Research questions

• RQ1 Can reinforcement learning-based NAS discover GAN architectures that rival or surpass human-designed models under the same regularization?
• RQ2 What modular GAN architectures (up-sampling, down-sampling, normal modules) emerge as effective when optimized automatically?
• RQ3 Are the learned architectural modules transferable across related image-generation tasks (e.g., CIFAR-10 to STL-10)?
• RQ4 How does reward shaping via Inception Score influence NAS efficiency and search outcomes?
• RQ5 What are the qualitative differences between NAS-discovered GAN cells and conventional GAN architectures?

Key findings

• AGAN identifies three novel modular architectures (AGAN-A, AGAN-B, AGAN-C) with competitive performance to state-of-the-art GANs.
• On CIFAR-10 unsupervised generation, AGAN-C outperforms state-of-the-art models under the same regularization.
• In supervised CIFAR-10 generation, AGAN-A, AGAN-B, and AGAN-C achieve Inception Scores and FID metrics that are competitive with or surpass many human-designed architectures, with AGAN-B achieving IS 8.82±0.09 and FID 23.8.
• The learned modules are transferable to STL-10, maintaining strong performance with appropriate input resizing (e.g., AGAN-A 9.23 IS, 52.7 FID; AGAN-C 8.97 IS, 57.4 FID; AGAN-B 7.84 IS, 71.8 FID).
• The controller’s operation distributions differ across module types (up-sampling, down-sampling, normal), supporting the need for modularized search in GAN design.
• Discovered cells resemble an inception-ResNet hybrid, indicating GAN-optimized architectures can diverge significantly from discriminative-task architectures.