Bidirectional Conditional Generative Adversarial Networks
Preprint
Abstract
Conditional variants of Generative Adversarial Networks (GANs), known as cGANs, are generative models that can produce data samples ($x$) conditioned on both latent variables ($z$) and known auxiliary information ($c$). Another GAN variant, Bidirectional GAN (BiGAN) is a recently developed framework for learning the inverse mapping from $x$ to $z$ through an encoder trained simultaneously with the generator and the discriminator of an unconditional GAN. We propose the Bidirectional Conditional GAN (BCGAN), which combines cGANs and BiGANs into a single framework with an encoder that learns inverse mappings from $x$ to both $z$ and $c$, trained simultaneously with the conditional generator and discriminator in an end-to-end setting. We present crucial techniques for training BCGANs, which incorporate an extrinsic factor loss along with an associated dynamically-tuned importance weight. As compared to other encoder-based GANs, BCGANs not only encode $c$ more accurately but also utilize $z$ and $c$ more effectively and in a more disentangled way to generate data samples.
Keywords
All Related Versions
This publication has 2 references indexed in Scilit:
- RenderGAN: Generating Realistic Labeled DataFrontiers in Robotics and AI, 2018
- Reducing the Dimensionality of Data with Neural NetworksScience, 2006