MisGAN: Learning from Incomplete Data with GANs

This repository provides a PyTorch implementation of
MisGAN,
a GAN-based framework for learning from incomplete data.

Note: Please check out our
follow-up work
on models that can be trained faster and more stably.

Requirements

The code requires Python 3.6 or later.
The file requirements.txt contains the full list of
required Python modules.

Jupyter notebook

We provide a notebook that includes an overview of MisGAN
as well as the annotated implementation that runs on MNIST.
The notebook can be viewed from
here.

Usage

The source code can be found in the src directory.
Separate scripts are provided to run MisGAN on MNIST and CelebA datasets.

For CelebA, you will need to download the dataset from its
website:

• Download the file img_align_celeba.zip (available from this link).
• Extract the zip file into the directory src/celeba-data that you create.

The commands below need to be run under the src directory.

MisGAN on MNIST:

python mnist_misgan.py

MisGAN imputation on MNIST:

python mnist_misgan_impute.py

MisGAN on CelebA:

python celeba_misgan.py

MisGAN imputation on CelebA:

python celeba_misgan_impute.py

Use -h to see all available command-line arguments for each script.

References

Steven Cheng-Xian Li, Bo Jiang, Benjamin Marlin.
“MisGAN: Learning from Incomplete Data with Generative Adversarial Networks.”
ICLR 2019.
[arXiv]

Contact

Your feedback would be greatly appreciated!
Reach us at li.stevecx@gmail.com.