MMVAE+: Enhancing the Generative Quality of Multimodal VAEs without Compromises

Official PyTorch implementation for MMVAE+, introduced in the paper MMVAE+: Enhancing the Generative Quality of Multimodal VAEs without Compromises, published at ICLR 2023.

UPDATE: Jul 2024 new improved code release!

Download datasets

mkdir data 
cd data 
curl -L -o data_ICLR_2.zip https://polybox.ethz.ch/index.php/s/wmAXzDAKn3Qogp7/download
unzip data_ICLR_2.zip 
curl -L -o cub.zip http://www.robots.ox.ac.uk/~yshi/mmdgm/datasets/cub.zip
unzip cub.zip

Experiments

Run on PolyMNIST dataset

bash commands/run_polyMNIST_experiment.sh

Run on CUB Image-Captions dataset

bash commands/run_CUB_experiment.sh

Citing

@inproceedings{
palumbo2023mmvaeplus,
title={{MMVAE}+: Enhancing the Generative Quality of Multimodal {VAE}s without Compromises},
author={Emanuele Palumbo and Imant Daunhawer and Julia E Vogt},
booktitle={International Conference on Learning Representations },
year={2023},
}

Acknowledgements

We thank the authors of the MMVAE repo, from which our codebase is based, and from which we retrieve the link to the CUB Image-Captions dataset. We also thank he authors of the MoPoE for useful code.