Optimal Transport Based Generative Autoencoders

This is the official pytorch implementation of AE-OTgen and AE-OTtrans. The paper is here: [Waiting for verification on ArXiv]

Dependencies

python 3.6.7
pytorch 1.0.1.post2
matplotlib 3.0.3
POT (python optimal transport) 0.5.1

Cuda is required for this to run on your machine.

Usage

1. Download CelebA. Download img_align_celeba.zip and extract it under a folder titled celebA.

2. Run the program. Run python3 main.py -h for a list of arguments and then follow them.

Results

Results will be stored within whatever folder you specify with the --Folder flag. Here are some example generated images using AE-OTgen.

MNIST

CelebA

TODO:

1. Stop using keras for mnist/fashion_mnist downloading lol (DONE)
2. Get arxiv link lol
3. Make a results folder to house all my results from the paper (DONE)
4. Better Usage in README.md