Differentiable Generative Adversarial Network Architecture Search Guided by Efficient Attention and Fréchet Distance.

Area of Science:

• Artificial Intelligence
• Computer Vision

Background:

• Generative Adversarial Network (GAN) architecture search automates the design of high-performance networks.
• Existing differentiable search methods like DAMGAN face limitations due to inefficient SENet modules and adversarial loss for attention training, impacting search efficiency and architecture quality.

Purpose of the Study:

• To propose EAMGAN, an enhanced differentiable GAN architecture search method.
• To improve search efficiency and architecture quality by incorporating efficient attention and Fréchet Inception Distance (FID) guidance.

Main Methods:

• Introduced EAMGAN, utilizing the lightweight ECA-Net attention module to replace SENet's inefficient components with 1D convolution for reduced parameters.
• Decoupled attention training from adversarial optimization and employed a customized FID-based loss function for direct guidance of architecture selection.
• Evaluated performance on CIFAR-10 and STL-10 datasets.

Main Results:

• EAMGAN surpassed DAMGAN on CIFAR-10 with improved Inception Score (IS) and lower FID.
• Achieved superior results with a reduced search cost (0.08 GPU days compared to DAMGAN's 0.09 GPU days).
• Demonstrated competitive performance and transferability on the STL-10 dataset.

Conclusions:

• EAMGAN offers a more efficient and effective approach to GAN architecture search.
• The proposed method enhances image generation quality and reduces computational overhead.
• EAMGAN shows promise for broader applications in generative modeling.