Reference implementation of the graph transport network (GTN), as proposed in our paper
Scalable Optimal Transport in High Dimensions for Graph Distances, Embedding Alignment, and More
by Johannes Klicpera, Marten Lienen, Stephan Günnemann
Published at ICML 2021.
The paper furthermore proposed the locally corrected Nyström (LCN) approximation, sparse Sinkhorn, and LCN-Sinkhorn, whose implementations you can find in this accompanying repository. GTN uses these approximations and relies on the implementations provided in the LCN repository.
You can install the repository using pip install -e ..
This repository contains a notebook for training and evaluating GTN (experiment.ipynb) and a script for running this on a cluster with SEML (experiment_seml.py).
The config files specify all hyperparameters and allow reproducing the results in the paper.
Please contact klicpera@in.tum.de if you have any questions.
Please cite our paper if you use our method or code in your own work:
@inproceedings{klicpera_2021_lcn,
title={Scalable Optimal Transport in High Dimensions for Graph Distances, Embedding Alignment, and More},
author={Klicpera, Johannes and Lienen, Marten and G{\"u}nnemann, Stephan},
booktitle = {Thirty-eighth International Conference on Machine Learning (ICML)},
year={2021},
}