README.md

A multimodal dynamical variational autoencoder for audiovisual speech representation learning

This repository contains the code associated with the following publication:

A multimodal dynamical variational autoencoder for audiovisual speech representation learning
Sadok, Samir and Leglaive, Simon and Girin, Laurent and Alameda-Pineda, Xavier and Séguier, Renaud
Neural Networks 2024.

Useful links:

• Abstract
• Paper
• Demo website with qualitative results

Setup

• Pypi: (Soon)

• Install the package locally (for use on your system):
  • In MDVAE directoy: pip install -e .
• Virtual Environment:
  • conda create -n mdvae python=3.8
  • conda activate mdvae
  • In MDVAE directoy: pip install -r requirements.txt

Some functionalities

1- Config

See the configuration of the model, the data, and the training in config_mdvae. Each time you run a training session, these configurations are automatically saved in the checkpoints.

@hydra.main(config_path=f"config_mdvae", config_name="config")

2- VQ-VAE

You can download the pre-trained VQVAE models in checkpoints/VQVAE.

from mdvae import SpeechVQVAE, VisualVQVAE

speech_vqvae = SpeechVQVAE(**cfg.vqvae_1)
speech_vqvae.load(path_model=r"checkpoints/VQVAE/speech/model_checkpoint_Y2022M3D5")

visual_vqvae = VisualVQVAE(**cfg.vqvae_2)
visual_vqvae.load(path_model=r"checkpoints/VQVAE/visual/model_checkpoint_Y2022M2D13")

3- MDVAE model

from mdvae import VQMDVAE
model = VQMDVAE(config_model=cfg.model, vqvae_speech=speech_vqvae, vqvae_visual=visual_vqvae)

You can load a pre-trained model in checkpoints/MDVAE.

model.load_model(path_model=f"{path}/mdvae_model_checkpoint")
model.to("cuda")

4- Training MDVAE

See the code train_mdvae.py.

from mdvae import MdvaeTrainer
trainer = MdvaeTrainer(mdvae=model, vqvae_speech=speech_vqvae, vqvae_visual=visual_vqvae,
                           training_data=dataset_train, validation_data=dataset_validation,
                           config_training=cfg.training_config, audio_config=cfg.audio_config)
trainer.fit()

Prepare data

1- H5 creation

In the code h5_creation.py, a preprocessing step is executed for each modality to save the intermediate representations at the encoder output in an (.hdf5) file. This pre-processing step helps accelerate the training process.

2- Mead dataset

In the _getitem_() of the MeadDateset dataset, we'll read the two H5s created previously for each modality and return both modalities with the same sequence length.

from mdvae import MeadDataset
dataset_train = MeadDataset(root_modality_1=Path(r"..."),
                                root_modality_2=Path(r"..."),
                                h5_speech_path=r"H5/speech_vq.hdf5",
                                h5_visual_path="H5/visual_vq.hdf5",
                                speaker_retain_test=["..."],
                                speaker_retain_validation=["..."],
                                train=True/False)

Test

In the file unit_test.py, you'll find unit tests designed to evaluate the model, covering scenarios such as analysis-synthesis and analysis-transformation-synthesis...

Citation

If you use this code for your research, please cite the above paper.

@article{sadok2024multimodal,
title={A multimodal dynamical variational autoencoder for audiovisual speech representation learning},
author={Sadok, Samir and Leglaive, Simon and Girin, Laurent and Alameda-Pineda, Xavier and S{'e}guier, Renaud},
journal={Neural Networks},
volume={172},
pages={106120},
year={2024},
publisher={Elsevier}
}

## License
CC-BY-4.0, see LICENSE.txt.