Exploring the Optimal Bit Pair for a Quantized Generator and Discriminator

Subin Yang, Muhammad Salman Ali, A. F. M. Shahab Uddin, Sung-Ho Bae

Abstract: Generative Adversarial Networks (GANs) are hindered from real-world applications due to their high computational cost and memory requirements. Model compression techniques, such as quantization, pruning, and knowledge distillation, can compress neural networks, lower memory requirements, and model size. However, quantizing generators often leads to a suboptimal solution. In this paper, we propose a novel method to stabilize GAN quantization by quantizing the generator and discriminator with different bit precision. Our method maximizes the quantization efficiency by jointly quantizing the generator and discriminator, which we found to be dependent on each other’s quantization. Specifically, quantizing the discriminator enhances the performance of the quantized generator, while the discriminator’s optimal quantization bit depends on the generator’s quantization bit and architectural type. We conducted extensive experiments on various GAN models, including BigGAN, SAGAN, and SNGAN, using different quantization methods, such as LSQ, PACT, and DoReFa, on benchmark dataset (CIFAR10). The experimental results demonstrate that our joint quantization method achieves higher compression rates while offering better performance in Frechet Inception Distance (FID) and Inception Score (IS).