RelativisticGAN-Tensorflow

Simple Tensorflow implementation of RelativisticGAN

Issue

  • For 256x256, the network does not generate the image properly. (DCGAN Architecture)
  • I think, RaDRAGAN more better than RaLSGAN

Usage

dataset

> python download.py celebA
  • mnist and cifar10 are used inside keras
  • For your dataset, put images like this:
├── dataset
   └── YOUR_DATASET_NAME
       ├── xxx.jpg (name, format doesn't matter)
       ├── yyy.png
       └── ...

train

  • python main.py --phase train --dataset celebA --Ra True --gan_type dragan

test

  • python main.py --phase test --dataset celebA --Ra True --gan_type dragan

Summary

"the discriminator estimates the probability that the given real data is more realistic than a randomly sampled fake data"

= RGAN

"the discriminator estimates the probability that the given real data is more realistic than fake data, on average"

= RaGAN

Idea

idea

Formulation

Name Formulation
GAN original_gan

RGAN| RGAN
　
RaGAN| RaGAN-1
　
RaGAN-GP| RaGAN-GP
　
RaLSGAN| RaLSGAN-1
　
RaHingeGAN| RaHingeGAN-1

Results

  • 128x128 celebA
  • 200k iterations (but, 100k iteration is also enough)
  • RaDRAGAN is not in the paper, I just tried because I wanted to do it.
Name Original Original + Ra
GAN GAN RaGAN

LSGAN | LSGAN | RaLSGAN |
　
HingeGAN | Will be soon | RaHingeGAN |
　
DRAGAN | DRAGAN | RaDRAGAN |

Error

Original DRAGAN

DRAGAN_error

  • In the case of DRAGAN, the images are sometimes distorted during the training

Author

Junho Kim

GitHub