PyTorch implementation of Unsupervised and Semi-supervised Learning with Categorical Generative Adversarial Networks that was originally proposed by Jost Tobias Springenberg.

Note that in this repo, only the unsupervised version was implemented for now. I reaplced the orginal architecture with DCGAN and the results are more colorful than the original one.

From 0 to 100 epochs:

• Python 2.7
• PyTorch v0.2.0
• Numpy
• SciPy
• Matplotlib

• Install PyTorh and the other dependencies
• Clone this repo:

git clone https://github.com/xinario/catgan_pytorch.git
cd catgan_pytorch

• Download the cifar10 dataset (.png format from kaggle)
• Create a dataset folder to hold the images

mkdir -p ./datasets/cifar10/images

• Move the extracted images into the newly created folder

• Train a model:

python catgan_cifar10.py --data_dir ./datasets/cifar10 --name cifar10

All the generated plot and samples can be found in side ./results/cifar10

optional arguments:

--continue_train  	to continue training from the latest checkpoints if --netG and --netD are not specified
--netG NETG           path to netG (to continue training)
--netD NETD           path to netD (to continue training)
--workers WORKERS     number of data loading workers
--num_epochs EPOCHS         number of epochs to train for

More options can be found in side the training script.

Some of code are inspired and borrowed from wgan-gp, DCGAN, catGAN chainer repo