Optimal Transport applied to GANs

Implementation of GAN variants that built on optimal transport, using a different approach than the Wasserstein-GAN paper, as proposed in the OT-Gan paper (you can also start using this variant of the same idea).

• Improving GANs Using Optimal Transport.
• Learning Generative Models with Sinkhorn Divergences

Usage

To test training implementation of the GAN model with the Mini batch Energy distance loss presented in the paper you should clone the project.

Project Organization

OTGAN directory is structured as follows

├── OTGAN
│   ├── dataset.py          <- Loading the two batch datasets : MNIST or CIFAR10
│   ├── Discriminator.py    <- Critic architecture (1 or 3 input channels)
│   ├── Generator.py        <- Generator architecture  (1 or 3 output channels)
│   ├── InceptionScore.py   <- Returns inception score for batchs or for dataset
│   ├── __init__.py
│   ├── __main__.py
│   ├── SinkhornDistance.py <- Computes the Minibatch Energy distance
│   ├── train.py            <- Main training loop
│   └── utils.py            <- Weight init, plotting and sampling

To run the training follow these steps :

1. Clone the repository and cd to the directory

   git clone https://github.com/AmineDiro/OT-GAN.git && cd OT-GAN/

2. The training has different arguments , run the command python -m OTGAN with the proper arguments :

   Short	Long	Description	Default
   -c	--channels	Nb of channels 1 for MNIST,3 for CIFAR , 3 by default	3
   -b	--batch_size	Batch size for training (default: 64)	64
   -se	--save_epoch	Saving model every N epoch	2
   -si	--sample_interval	Interval number for sampling image from generator and saving them	1
   --score / --no-score		Boolean args to get Inception score or not	True

NOTE : The Notebook Results.ipynb presents the main results from training on the CIFAR10 dataset. We plot the generated images from training, the loss of generator, critic and the inception score while training. You can click on the [Open In Colab] to access the notebook on google collab.