jollyjonson / compregan

Experiments on image compression using Generative Adversarial Networks

CompreGAN

This repository contains some of my ventures into the realm of image coding using Generative Adversarial Networks (GANs). Most of this code was written in late 2021 and I unfortunately have not had the chance to work on the topic further. This work was mainly inspired by a paper by Agustsson et al. called Generative Adversarial Networks for Extreme Learned Image Compression. This Python package contains functionality for training GANs for image compression tasks as well as corresponding experiments.

Installation

0. Create a conda environment for the package to reside in and activate it

conda create -n compregan python=3.10
conda activate compregan
conda install -c conda-forge cudnn  # might be needed to utilize the GPU

1. To install the python package please check it out, install the dependencies and install the package itself in 'develop' mode as follows.

git clone https://github.com/jollyjonson/compregan
cd compregan
make install

2. Make sure everything works as intended, run the unit tests as follows

pip install -r requirements-test.txt
make test

3. You can find examples and experiments in the experiments directory! As a basic sandbox try out and explore

python experiments/basic_mnist_example.py