/SAGAN-with-relativistic

Self-Attention Generative Adversarial Networks with Relativistic loss

Primary LanguagePython

SAGAN-with-relativistic

A pytorch implmentation of SAGAN (https://arxiv.org/abs/1805.08318) with relativistic loss (https://arxiv.org/abs/1807.00734). The main difference here is that we replace the hinge loss used in the SAGAN with relativistic loss.

Sample Results

Dependancies

  • Python 3.5+
  • Pytorch 0.4.1
  • Torchvision 0.2.1

Usage

Train

python main.py

View sample results

cd images

Acknowledgement/reference