The GAN Landscape: Losses, Architectures, Regularization, and Normalization

  • 2018-07-12 16:56:50
  • Karol Kurach, Mario Lucic, Xiaohua Zhai, Marcin Michalski, Sylvain Gelly
  • 167

Abstract

Generative Adversarial Networks (GANs) are a class of deep generative modelswhich aim to learn a target distribution in an unsupervised fashion. While theywere successfully applied to many problems, training a GAN is a notoriouslychallenging task and requires a significant amount of hyperparameter tuning,neural architecture engineering, and a non-trivial amount of "tricks". Thesuccess in many practical applications coupled with the lack of a measure toquantify the failure modes of GANs resulted in a plethora of proposed losses,regularization and normalization schemes, and neural architectures. In thiswork we take a sober view of the current state of GANs from a practicalperspective. We reproduce the current state of the art and go beyond fairlyexploring the GAN landscape. We discuss common pitfalls and reproducibilityissues, open-source our code on Github, and provide pre-trained models onTensorFlow Hub.

 

Quick Read (beta)

loading the full paper ...