Stabilizing GANs with Octave Convolutions

This repository provides the official PyTorch implementation of Stabilizing GANs with Octave Convolutions.

Dependencies

Tested on Python 3.6.x.

• PyTorch (1.0.1)
• NumPy (1.16.2)

CelebA dataset

The full CelebA is available here. To resize the RGB images to 128 by 128 pixels, set the path and run resize_celeba.py.

Training

To train a model, simply specify run sh on the selected model (e.g. sh gan.sh, sh wgan.sh or sh lsgan.sh) with the appropriate hyper-parameters.

Example hyper-parameters definition (wgan.sh)

python train.py --type wgan \
           --nb-epochs 50 \
           --learning-rate 0.00005 \
           --optimizer rmsprop \
           --critic 5 \
           --cuda

Comparision between DCGAN with and wihtout Octave Convolution

References

This repo combines the pytorch implementation of the following paper:

Goodfellow et al. Generative Adversarial Nets.

Radford et al. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks.

Mao et al. Multi-class Generative Adversarial Networks with the L2 Loss Function.

Arjovsky et al. Wasserstein Generative Adversarial Networks.

Chen et al. Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution.

Octave Convolution

The codes are heavily borrowed from a pytorch:

Ocatve Convolution implementation.

GANs implementations.

Citation

If this work is useful for your research, please cite our paper:

@article{durall2019dropgan,
  title={Stabilizing GANs with Octave Convolutions},
  author={Durall, Ricard and Pfreundt, Franz-Josef and Keuper, Janis},
  journal={arXiv preprint arXiv:1905.12534},
  year={2019}
}