eduzhai > Applied Sciences > Engineering >

Incorporating Reinforced Adversarial Learning in Autoregressive Image Generation

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 25 pages

Abstract: Autoregressive models recently achieved comparable results versusstate-of-the-art Generative Adversarial Networks (GANs) with the help of VectorQuantized Variational AutoEncoders (VQ-VAE). However, autoregressive modelshave several limitations such as exposure bias and their training objectivedoes not guarantee visual fidelity. To address these limitations, we propose touse Reinforced Adversarial Learning (RAL) based on policy gradient optimizationfor autoregressive models. By applying RAL, we enable a similar process fortraining and testing to address the exposure bias issue. In addition, visualfidelity has been further optimized with adversarial loss inspired by theirstrong counterparts: GANs. Due to the slow sampling speed of autoregressivemodels, we propose to use partial generation for faster training. RAL alsoempowers the collaboration between different modules of the VQ-VAE framework.To our best knowledge, the proposed method is first to enable adversariallearning in autoregressive models for image generation. Experiments onsynthetic and real-world datasets show improvements over the MLE trainedmodels. The proposed method improves both negative log-likelihood (NLL) andFréchet Inception Distance (FID), which indicates improvements in terms ofvisual quality and diversity. The proposed method achieves state-of-the-artresults on Celeba for 64 $ times$ 64 image resolution, showing promise forlarge scale image generation.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...