Wasserstein Barycenters were presented for models ensembling in the ICLR'19 paper:
"Wasserstein Barycenter Model Ensembling", Pierre Dognin, Igor Melnyk, Youssef Mroueh, Jarret Ross, Cicero Dos Santos, and Tom Sercu available on arxiv and openreview
@inproceedings{
dognin2018wasserstein,
title={Wasserstein Barycenter Model Ensembling},
author={Pierre Dognin and Igor Melnyk and Youssef Mroueh and Jarret Ross and Cicero Dos Santos and Tom Sercu},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=H1g4k309F7},
}
This github repository provides the PyTorch code for computing Wassersteing Barycenters. The code is ready for use with PyTorch 1.0.0.
To run a simple wasserstein barycenter test code, use pytorch version >=1.0.0:
python test.py
We provide a notebook for synthetic experiments:
jupyter ./notebooks/synthetic_experiments.ipynb