Implementation associated with the paper Homomorphism Autoencoder -- Learning Group Structured Representations from Observed Transitions.

The Homomorphism AutoEncoder HAE is a model trained on observed transitions $(o_t, g_t, o_{t+1})$ to jointly learn a group representation of the observed actions $g_t$ and a representation of the observations $o$.

Main scripts are provided in ./displacementae/homomorphism/.

Best run commands are provided in the ./displacementae/homomorphism/README.rst.

The package can be installed by first building the package through:

 $ python setup.py sdist bdist_wheel

Then it can be installed in your environment through:

 $ pip install dist/homomorphism-autoencoder-<VERSION>.tar.gz

The Dsprites dataset and the 3D bunny .obj model need to be downloaded separately.

In addition, a transition dataset for the bunny dataset needs to be constructed from the downloaded .obj file using provided scripts in ./displacementae/data/obj3d/. Please refer to ./discplacementae/data/obj3d/README.md for instructions.

Documentation can be built from the code's docstrings by following the instructions in ./docs/README.md.