A re-implementation of a paper which uses graphical models for transferring style between images as my final project for course Graphical Models in Machine Learning, spring 2017. Tweaked and polished in 2018.
Unsupervised Image Translation, Rómer Rosales, Kannan Achan, and Brendan Frey, 2003
Here I implement:
- loopy belief propagation, approximate inference algorithm using message passing in C++
- expectation-maximization algorithm as described in the paper
- a commandline tool for transferring style of one image to another
- a set of reproducible experiments to observe convergence and effects of some paramters
More detailed results of this implementation, can be found in example directory.
cd unsupervised-image-translation
# Create a virtual environment
virtualenv -p python3 env
. env/bin/activate
# Install necessary modules
pip install -r requirements.txt
# Compile loopy belief propagation written in C++
python setup.py install
# Run
python src/unsupervised_image_translation/main.py -input=inputs/ramona-color.png -source=inputs/starry-night.png
# To list all arguments
python src/unsupervised_image_translation/main.py -h
# Optionally reproduce experiments
bash run_all_experiments.sh