VAEs with Lie Group latent space
Clone or download
Latest commit 51cf090 Jul 25, 2018
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
lie_vae Edited README Jul 25, 2018
.gitignore Big cleanup Jul 25, 2018
README.md Edited README Jul 25, 2018
blender_spherecube.py Added spherecube pair generator and dataset and encoder continuity loss Jun 8, 2018
cube.blend Added spherecube blender model May 20, 2018
results.ipynb Big cleanup Jul 25, 2018
toy_results.ipynb Big cleanup Jul 25, 2018

README.md

Code for "Explorations in Homeomorphic Variational Auto-Encoding"

by Luca Falorsi*, Pim de Haan*, Tim R. Davidson*, Nicola De Cao, Maurice Weiler, Patrick Forré and Taco S. Cohen.

Link to paper Animated results

The relevant code to implement SO(3) valued latent variables can be found in the lie_vae package. The code needed to reproduce the experiments can be found in lie_vae.experiments.

For questions file an issue or email to either:

Dependencies

conda install -y pytorch torchvision cuda91 -c pytorch
conda install -y numpy ipython jupyter tensorflow pillow cython scipy requests
pip install tensorboardX tqdm git+https://github.com/AMLab-Amsterdam/lie_learn \
    git+https://github.com/pimdh/svae-temp.git 

The sphere cube data can be generated with the python -m lie_vae.experiments.gen_spherecube_pairs (see file for details, this requires having installed Blender) or for limited time be downloaded here.