Pytorch_GAN_CelebA

Pytorch implementation of DCGAN, CDCGAN, LSGAN, WGAN and WGAN-GP for CelebA dataset.

Usage

1. Download the CelebA dataset, and aligned version is used in this repo.

2. Clone the repo

$ git clone https://github.com/AndrewZhuZJU/Pytorch_GAN_CelebA.git
$ cd Pytorch_GAN_CelebA

3. Training

To train any GAN please use main_**.py. For example,

$ python main_DCGAN.py

Not that all the setting parameters for the modesl are in args.py, please change properly.

Results

Animation of generated images with fixed noise (or condition labels) during training progress

DCGAN	LSGAN	CGAN(Up:Male, Bottom:Female)
WGAN	WGAN-GP

Generated images in Epoch 40

DCGAN	LSGAN	CGAN(Up:Male, Bottom:Female)
WGAN	WGAN-GP

Evaluation

• Inception score calculated in Epoch 40 (30K generated images)

  
  
  
  
  
  
  
  
  
  
  
  
  
  

  DCGAN	LSGAN	WGAN	WGAN-GP
  2.10 $\pm$ 0.04	2.08 $\pm$ 0.02	2.20 $\pm$ 0.02	1.97 $\pm$ 0.02

• Inception score calculated in Epoch 20 (30K generated images)

  
  
  
  
  
  
  
  
  
  
  
  
  
  

  DCGAN	LSGAN	WGAN	WGAN-GP
  2.07 $\pm$ 0.03	1.95 $\pm$ 0.03	2.05 $\pm$ 0.02	2.00 $\pm$ 0.02

Loss Plot

Take DCGAN for example:

Generator Loss

Discriminator Loss

Development Environment

• Ubuntu 16.04 LTS
• NVIDIA GTX 1080 Ti
• CUDA 9.0
• pytorch 0.4
• python 2.7
• Others Dependencies: numpy, imageio, torchvision, tensorboard, etc.

References

1.Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

2.Conditional Generative Adversarial Nets

3.Wasserstein GAN

4.Improved Training of Wasserstein GANs

5.Least Squares Generative Adversarial Networks