DAWSON: A Domain Adaptive Few Shot Generation Framework

  • 2020-01-02 00:59:10
  • Weixin Liang, Zixuan Liu, Can Liu
  • 19

Abstract

Training a Generative Adversarial Networks (GAN) for a new domain fromscratch requires an enormous amount of training data and days of training time.To this end, we propose DAWSON, a Domain Adaptive FewShot GenerationFrameworkFor GANs based on meta-learning. A major challenge of applyingmeta-learning GANs is to obtain gradients for the generator from evaluating iton development sets due to the likelihood-free nature of GANs. To address thischallenge, we propose an alternative GAN training procedure that naturallycombines the two-step training procedure of GANs and the two-step trainingprocedure of meta-learning algorithms. DAWSON is a plug-and-play framework thatsupports a broad family of meta-learning algorithms and various GANs witharchitectural-variants. Based on DAWSON, We also propose MUSIC MATINEE, whichis the first few-shot music generation model. Our experiments show that MUSICMATINEE could quickly adapt to new domains with only tens of songs from thetarget domains. We also show that DAWSON can learn to generate new digits withonly four samples in the MNIST dataset. We release source codes implementationof DAWSON in both PyTorch and Tensorflow, generated music samples on two genresand the lightning video.

 

Quick Read (beta)

loading the full paper ...