効率的なアテンションとフレッシェ距離による微分可能な敵対的生成ネットワークアーキテクチャ検索

Generative adversarial network (GAN) architecture search aims to automate the discovery of high-performance network structures. While differentiable search methods like DAMGAN have shown promise, their reliance on inefficient SENet modules and adversarial loss for attention training limits both search efficiency and architecture quality. To address these limitations, we propose EAMGAN, an enhancing differentiable GAN architecture search with efficient attention and Fréchet Inception Distance (FID) guidance. Our approach introduces a lightweight attention module ECA-Net that replaces the fully-connected layers in SENet with a 1D convolution utilizing local context and weight sharing, thereby significantly reducing parameter count. Furthermore, we decouple the attention training from adversarial optimization and introduce a customized loss function based on the FID, which directly guides the architecture selection toward subnets that generate higher-quality images. Experiments on CIFAR-10 show that EAMGAN not only surpasses DAMGAN (Inception Score (IS): 9.03 versus 8.99, FID: 9.43 versus 10.27) but also achieves this with lower search cost (0.08 versus 0.09 GPU days). Competitive results on STL-10 further demonstrate its effectiveness and transferability.