GAN (Generative Adverserial Network)

PyTorch Generative adversarial networks

Requirements

• python3.6
• pytorch 1.0
• torchvision
• tensorboard
• tensorboardX
• imagemagick (optional for creating gifs)

Implemented GAN methods

• DCGANs - "deep convolutional generative adversarial networks" article

  basic dcgan network is based on pytorch tutorial

• WGAN-GP - "Improved Training of Wasserstein GANs" article

• DRAGAN - "On Convergence And Stability Of GANs" article

TODO Capabilities

• Parallel training
• resume training
• Inception score

Improving training

• One-sided label smoothing
• Minibatch discrimination
• Experience replay
• Conditional GAN

Example

training dcgan:

1. Download celeb A dataset and unpack.
2. Set path for the dataset in dcgan.yaml
3. run training

python train.py --cfg ./cfgs/dcgan.yaml

inference dcgan:

python ./Examples/dcgn_demo.py

celeb a output (10 epoch gif):

training WGAN-GP

1. Download celeb A dataset and unpack.
2. Set path for the dataset in wgan_gp.yaml
3. run training

python train.py --cfg ./cfgs/wgan_gp.yaml