Mixed-curvature Variational Auto-Encoders

PyTorch implementation

Overview

This repository contains a PyTorch implementation of the Mixed-curvature Variational Autoencoder, or M-VAE, as presented in [1]. For the arXiv paper, please see: https://arxiv.org/abs/1911.08411.

Installation

Install Python 3.7+. To install all dependencies, make sure you have installed conda, and run

make conda
conda activate pt
make download_data

Structure

• chkpt/ - Checkpoints for trained models.
• data/ - Data folder. Contains a script necessary for downloading the datasets, and the downloaded data.
• mathematica/ - Mathematica scripts (various formula derivations, etc).
• mt/ - Source folder (stands for Master Thesis).
  • data/ - Data loading, preprocessing, batching, and pre-trained embeddings.
  • examples/ - Contains the main executable file. Reads flags and runs the corresponding training and/or evaluation.
  • mvae/ - Model directory. Note that models heavily use inheritance!
  • test_data/ - Data used for testing.
  • visualization/ - Utilities for visualization of latent spaces or training statistics.
• plots/ - Folder to store generated plots.
• scripts/ - Contains scripts to run experiments and plot the results.
• tests/ - (A few) unit tests.
• Makefile - Defines "aliases" for various tasks.
• README.md - This manual.
• LICENSE - Apache Standard License 2.0.
• environment.yml - Required Python packages.
• THIRD_PARTY.md - List of third party software used in this thesis.

Usage

To run training and inference, activate the created conda environment and run the examples:

conda activate pt

# MNIST:
python -m mt.examples.run --dataset="mnist" --model="h2,s2,e2" --fixed_curvature=False

# CIFAR:
python -m mt.examples.run --dataset="cifar" --model="h2,s2,e2" --fixed_curvature=False --h_dim=8192 --architecture="conv"

Take a look at mt/examples/run.py for a list of command line arguments.

For an evaluation run, see mt/examples/eval.py.

Please cite [1] in your work when using this repository in your experiments.

Other make commands

make clean   # format source code
make check   # check for formatting and code errors
make test    # run tests

Feedback

For questions and comments, feel free to contact Ondrej Skopek.

License

ASL 2.0

Citation

Ondrej Skopek, Octavian-Eugen Ganea, Gary Bécigneul. Mixed-curvature Variational Autoencoders. International Conference on Learning Representations (ICLR) 2020. URL https://openreview.net/forum?id=S1g6xeSKDS

BibTeX format:

@inproceedings{skopek2020mixedcurvature,
  title={Mixed-curvature Variational Autoencoders},
  author={Ondrej Skopek and Octavian-Eugen Ganea and Gary B{\'e}cigneul,
  booktitle={International Conference on Learning Representations},
  year={2020},
  url={https://openreview.net/forum?id=S1g6xeSKDS}
}