Extreme image compression using Generative Adversarial Networks

This repository provides a learned compression algorithm for human portraits.

Use the following to clone this repository:

git clone git@github.com:mlomnitz/CompressionGAN.git

Description

This package provides an image compression algorithm using an auto-encoder in a Generative Adversarial Network (GAN) setting. First described in, the implementation in this work focuses on human faces, more specifically profile pictures.

The network was trained using a subset of 40K images curated from the DEX project to produce faithful reconstruction of human faces with a factor of 20-30 in compression.

Requirements

This package and associated requirements can be installed with pip, using the following:

pip install -r requirements.txt
python setup.py install 

The data set used for training/validation/test can be downloaded with the following link The following lists the available trained models and associated configuration files:

• Model trained for 40 epochs, without noise sampling: Epoch40, configfuration

More trained models will be made available as future experiments are completed.

Run a test

The scripts folder includes three python files that can be used to train the model, compress an image and expand a compressed image. You can view the arguments for any of the scripts by running with the -h option, for example:

python scripts/train.py -h

We can also run inference on one of the provided images (in the images folder) with the following example:

python scripts/compress.py -r checkpoints/gan-40 -i images/test3.jpg -o samples/test3

At present this will output 3 files to the samples directory, a jpg of the reconstructed image, a pdf comparing the input and reconstructed image and a txt file with the bitstring for the compressed file.