Evolutionary Channel Pruning for Style-Based Generative Adversarial Networks

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generative Adversarial Networks (GANs), particularly StyleGAN and StyleGAN2, excel at high-quality image synthesis.
• Large model sizes and high computational costs (FLOPs) limit GAN deployment on edge devices and mobile platforms.

Purpose of the Study:

• To propose Evolutionary Channel Pruning for StyleGANs (ECP-StyleGANs), an algorithm for compressing StyleGAN and StyleGAN2.
• To maintain competitive image quality while reducing model complexity for real-time applications.

Main Methods:

• Utilizing evolutionary algorithms to iteratively refine binary masks for convolutional channel pruning.
• Employing fitness functions that balance model complexity and generation quality.
• Encoding pruning configurations and applying selection, crossover, and mutation operations.

Main Results:

• Achieved approximately a 4x reduction in FLOPs and parameters for StyleGAN and StyleGAN2 models.
• Maintained visual fidelity with only a slight increase in Fréchet Inception Distance (FID) compared to un-pruned models.
• Demonstrated the effectiveness of ECP-StyleGANs in compressing GANs for resource-constrained environments.

Conclusions:

• ECP-StyleGANs offers an effective approach to compress large generative models like StyleGAN.
• The study frames generative AI pruning as a multi-objective optimization task, balancing efficiency and quality.
• This research facilitates the deployment of advanced GANs on edge devices and in resource-limited settings.