GAN

Generative Adversarial Networks

Reproduce the following GAN-related papers:

• Unsupervised Representation Learning with DCGAN. paper

• Image-to-image Translation with Conditional Adversarial Networks. paper

• InfoGAN: Interpretable Representation Learning by Information Maximizing GAN. paper

See the docstring in each script for detailed usage.

DCGAN-CelebA.py

Reproduce DCGAN following the setup in dcgan.torch.

Play with the pretrained model on CelebA face dataset:

• Generated samples

• Vector arithmetic: smiling woman - neutral woman + neutral man = smiling man

Image2Image.py

Reproduce Image-to-Image following the setup in pix2pix. It requires the datasets released by the original authors.

InfoGAN-mnist.py

Reproduce a mnist experiement in InfoGAN. By assuming 10 latent variables corresponding to a categorical distribution and maximizing mutual information, the network learns to map the 10 variables to 10 digits in a completely unsupervised way.