Abstract: In recent years,generative adversarial network have shown excellent performance in many computer vision tasks such as ima-ge super resolution,image generation and so on.GANs can be designed to be much more greedy in computation complexity because of huge quantity uses of GPU application.For mobile devices that are resource-limited,however,on which it is intractable for GAN on be deployed due to high consumption both in energy and computation.Thanks for great success in neural network compression,it is possible to deploy GAN on mobile devices.This paper proposed a method to simultaneously quantize weights and activations in GANs.Sensitivity analysis shows that weights are more sensitive than activation in quantization process.This paper used Fréchet Inception Distance (FID)score to evaluate generated images of quantized GANs for Inception score is less applicable than FID.Motivated by sensitivity analysis,extensive experiments were conducted on Mnist and Celeb-A datasets.Results show that the proposed method can compress GANs by up to 4x and still achieve even higher performance than the original GANs.Thus,it effectively resolves the problem of compressing GANs.

Key words: Generative adversarial networks, Mobile devices, Neural network compression, Quantization, Source-limited