Optimal Transport Dataset Distances

This repo is a python implementation of Geometric Dataset Distances via Optimal Transport and Robust Optimal Transport. Routines are implemented in numpy with Python Optimal Transport and CVXPY, as well as in Pytorch using KeOps and GeomLoss.

The OTDD algorithm allows us to incorporate label information into the optimal transport problem.

Installing

Core dependencies can be installed from the environment.yml file

conda env create -f environment.yml

To use the Pytorch implementation, install Pytorch, KeOps and GeomLoss

conda install pytorch torchvision torchaudio -c pytorch pip install pykeops pip install geomloss

Then validate the KeOps installation

import pykeops
pykeops.clean_pykeops()
pykeops.test_torch_bindings()

To use the cheminformatics functions in chem.py, install RDKit

conda install -c rdkit rdkit