Hellinger Multimodal Variational Autoencoders

Official PyTorch implementation for HELVAE, published at AISTATS 2026.

This repository is based on the implementation of the ICLR 2021 paper Generalized Multimodal ELBO.

Preliminaries

This code was developed and tested with:

• Python version 3.5.6
• PyTorch version 1.4.0
• CUDA version 11.0
• The conda environment defined in environment.yml

First, set up the conda enviroment as follows:

conda env create -f environment.yml  # create conda env
conda activate mopoe                 # activate conda env

Second, download the data, inception network, and pretrained classifiers:

curl -L -o tmp.zip https://drive.google.com/drive/folders/1lr-laYwjDq3AzalaIe9jN4shpt1wBsYM?usp=sharing
unzip tmp.zip
unzip celeba_data.zip -d data/

Please note that all the material downloaded are from the official implementation of the ICLR 2021 paper Generalized Multimodal ELBO.

Experiments

To select between the available models (MVAE, MMVAE, MoPoE, HELVAE, and MoHELVAE), set the script variable METHOD to one of the following values:

"poe" | "moe" | "joint_elbo" | "helvae" | "joint_helvae"

By default, each experiment runs with METHOD="helvae".

Running Bimodal Celeba

./job_celeba_helvae.sh

Notebook for Toy Dataset Example

toy_dataset_helvae.ipynb

Citing

@inproceedings{
vo2026hellinger,
title={Hellinger Multimodal Variational Autoencoders},
author={Huyen Thuc Khanh Vo and Isabel Valera},
booktitle={The 29th International Conference on Artificial Intelligence and Statistics},
year={2026},
url={https://openreview.net/forum?id=mxHyYltMUa}
}

Acknowledgements

We thank the authors of the MoPoE repo, from which our codebase is based on.