This is a port of the Scalable Factorized Hierarchical Variational Autoencoder (ScalableFHVAE) to PyTorch and Python 3.
The port is currently in-progress, so some features may not work as intended right now.
The two main papers on the ScalableFHVAE and its predecessor, the Factorized Hierarchical Variational Autoencoder (FHVAE):
- Unsupervised Learning of Disentangled and Interpretable Representations from Sequential Data
- Scalable Factorized Hierarchical Variational Autoencoder Training
The citations for these papers:
@inproceedings{hsu2017learning,
title={Unsupervised Learning of Disentangled and Interpretable Representations from Sequential Data},
author={Hsu, Wei-Ning and Zhang, Yu and Glass, James},
booktitle={Advances in Neural Information Processing Systems},
year={2017},
}
@article{hsu2018scalable,
title={Scalable Factorized Hierarchical Variational Autoencoder Training},
author={Hsu, Wei-Ning and Glass, James},
journal={arXiv preprint arXiv:1804.03201},
year={2018},
arxiv={1804.03201},
}
The original code for the FHVAE.
Linked again, the code for the ScalableFHVAE.
The required Python packages can be installed using pip install -r requirements.txt. Running pip install -r dev-requirements.txt will install Black, MyPy, and NumPy type hints, which I am using as part of the porting and development process.
This project also requires Kaldi, a library for speech recognition. This will have to be compiled on your machine. The default location for the installation of Kaldi is a subdirectory named kaldi in the root of this project's directory (i.e. PyTorch-ScalableFHVAE/kaldi), but it may be installed in any directory, with the --kaldi-root flag on Python scripts allowing configuration of the root Kaldi directory.