Mimicking key samples for knowledge transfer of GAN

Abstract

Transfer learning methods are arguably the most trustworthy techniques for boosting the performance of machine learning models. However, for image generation tasks, the efficient methods for transferring the knowledge of pre-trained models have not been explored compared to image classification. In this paper, we propose a novel transfer method especially for generative adversarial networks. To transfer useful knowledge from the generator and discriminator from a teacher domain, we propose teacher mimic loss (i.e., how to effectively transfer the knowledge) and sampling method (i.e., how to select samples to transfer). We demonstrate the effectiveness of the proposed method empirically.