eduzhai > Applied Sciences > Engineering >

Direct Adversarial Training A New Approach for Stabilizing The Training Process of GANs

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 12 pages

Abstract: Generative Adversarial Networks (GANs) are the most popular models for imagegeneration by optimizing discriminator and generator jointly and gradually.However, instability in training process is still one of the open problems forall GAN-based algorithms. In order to stabilize training, some regularizationand normalization techniques have been proposed to make discriminator meet theLipschitz continuity constraint. In this paper, a new approach inspired byworks on adversarial attack is proposed to stabilize the training process ofGANs. It is found that sometimes the images generated by the generator play arole just like adversarial examples for discriminator during the trainingprocess, which might be a part of the reason of the unstable training. Withthis discovery, we propose to introduce a adversarial training method into thetraining process of GANs to improve its stabilization. We prove that this DATcan limit the Lipschitz constant of the discriminator adaptively. The advancedperformance of the proposed method is verified on multiple baseline and SOTAnetworks, such as DCGAN, WGAN, Spectral Normalization GAN, Self-supervised GANand Information Maximum GAN.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×