Score-based Generative Neural Networks for Large-Scale Optimal Transport. (on ArXiv)
Max Daniels, Tyler Maunu, Paul Hand.

This repository contains code for running large scale experiments, such as those used to generate Figure 1 (CelebA sampling). A sister repository, used to run all synthetic/small-scale experiments, can be found here.

The required packages can be found in requirements.txt. To create a new conda environment with these packages installed, use

conda create --name <env> --file requirements.txt.

There are four main entry points, each dedicated to training and evaluating four different 'types' of models.

1. cpat.py -- Training compatibility functions.
2. ncsn.py -- Training score based models for unconditional sampling.
3. scones.py -- Conditional sampling SCONES by combining a score-based (unconditional) model with a compatibility function.
4. bproj.py -- Training barycentric projection models given a compatibility function.

The general template for using these files is

python <file>.py --doc <name-for-output> --config <cnf>.yml

The name <name-for-output> will be used as a label for experimental artifacts saved to disk. The config file <cnf>.yml must be located in the config subdirectory of the top level folder corresponding to <file>.py.

For example, to train a compatibility function for transportation from blurry CelebA to CelebA, one could use

python cpat.py --doc celeba-superres --config superres_KL_0.005.yml

This corresponds to the configuration file located in compatibility/configs/superres_KL_0.005.yml, where one can customize the experiment or view its fine-grained details.