Awesome

GAN-lib

This repository was originally copied from https://github.com/openai/InfoGAN, the official implementation of https://arxiv.org/abs/1606.03657. The current version adds the following:

CelebA experiment (the original repository contained MNIST experiment only)
Wasserstein GAN (https://arxiv.org/abs/1701.07875)
CycleGAN (https://arxiv.org/abs/1703.10593)
Sampling from a trained model

The next steps are to add experiments (ImageNet, CIFAR) and implement other GAN variants like SeqGAN, CatGAN,...

Note: The CycleGAN implementation slightly differs from the authors'

The discriminator is trained with the last generated samples, not with samples from the history of the last samples
The model was trained for 100 epochs instead of 200

Running in Docker

To run in docker, use the docker.sh script. You can build the image or run the container and go inside it. Both commands are available in CPU and GPU versions.

$ git clone git@github.com:nisace/gan-lib.git
$ cd gan-lib/
$ ./docker.sh {build, run} {cpu, gpu}
root@X:/gan-lib#

Command line interface

All scripts are accessible through manage.py. The see help, execute the following command:

python manage.py --help

Training a model

The model parameters are defined in a YML file. Some examples are given in software/params/.

python manage.py train -p {params/mnist.yml, params/celebA.yml, params/mnist_wasserstein_.yml, params/celebA_wasserstein.yml, params/horse_zebra.yml,...}

This will train the model and save it into ./ckpt/experiment_name/experiment_name_with_date/

Logs are saved in ./logs/experiment_name/experiment_name_with_date/

Seeing results

You can launch TensorBoard to view the generated images:

tensorboard --logdir logs/

or give a specific logdir, for example:

tensorboard --logdir logs/mnist_infogan/mnist_infogan_2017_03_20_10_49_54/

Sampling from a trained model

You can sample from a trained model:

python manage.py sample -p path/to/checkpoint.ckpt -s {random, latent_code_influence, linear_interpolation}

Example:

python manage.py sample -p ckt/mnist_infogan/mnist_infogan_2017_03_20_10_49_54/mnist_infogan_2017_03_20_10_49_54_400.ckpt -s random

Examples

MNIST:

Latent code influence

Samples showing the latent code influence. Along each column, the latent code is the same except for the category one hot encoding which has 10 classes and is learned in an unsupervised way with InfoGAN.

result

Linear interpolation

Linear interpolation in the latent space.

result

CelebA:

Latent code influence

Example where one variable of the latent code has learned hair style (trained with InfoGAN). The second and third rows show different bangs length for example.

result

Linear interpolation

result

MNIST Wassertein:

result

CelebA Wassertein:

result

CycleGAN

Horse to zebra

result result

Zebra to horse

result result

Failure case

result