Mind2Mind: Transfer Learning for GANs

Training generative adversarial networks (GANs) on high quality (HQ) images involves important computing resources. This requirement represents a bottleneck for the development of applications GANs. We propose transfer learning technique GANs that significantly reduces training time. Our approach consists freezing low-level layers both critic and generator original GAN. assume an auto-encoder constraint to ensure compatibility internal representations generator. assumption explains gain in time as it enables us bypass during forward backward passes. compare our method baselines observe significant acceleration training. It can reach two orders magnitude HQ datasets when compared with StyleGAN. provide theorem, rigorously proven within framework optimal transport, ensuring convergence transferred moreover precise bound terms distance between source target dataset.