Geometric Autoencoder

This repository provides the code for the ICML 2023 paper Geometric Autoencoders - What You See is What You Decode.

Getting Started

Clone the Repository

git clone https://github.com/phnazari/GeometricAutoencoder.git

Create and activate a new Conda Environment

conda create --name my_environment
conda activate my_environment

Install pip, a package management software for python:

conda install pip

Change into the project directory and install the dependencies

cd GeometricAutoencoder
pip3 install -r requirements.txt

Next, you want to add this project to your environment's PYTHONPATH. First, find out the projects path by running

pwd

The output will be something like /path/to/GeometricAutoencoder. Second, find out where your virtual environment was installed by running

conda info

There will be a row titled active env location, which will carry a value like /path/to/my_environment. Create a file /path/to/my_environment/lib/pythonX.XX/site-packages/custom.pth, where X.XX is your Python version (which you can find out by running python3 --version). In this file, paste

/path/to/GeometricAutoencoder

You have successfully added the project to the PYTHONPATH. Finally, you will have to restart the environment:

conda deactivate
conda activate my_environment

While TorchVision takes care of the MNIST and FashionMNIST datasets, you will have to download the PBMC, Zilionis and CElegans datasets yourself.

Reproducing the Results

TL; DR

If you want to reproduce our results for MNIST, you can do so by executing

bash main.sh

More Detailed

What happens in this case, is that you first train a Geometric Autoencoder by executing

bash scripts/create_eval_configs.sh.

The training results will be placed in a folder called "save_config", and you will have to move them to the experiments folder in order to proceed with the evaluation. The first run, for example, should be moved like

mv save_config/1 experiments/train_model/evaluation/repetitions/rep1/MNIST/GeomReg

You can then run our geometric diagnostics by envoking

python3 exp/analysis.py

and the quantitative metrics by executing

python3 scripts/load_results.py

The Differential Geometry

The differential geometry can be found inside directory src/diffgeo. Our geometric regularizer is implemented in src/criterions.py.

Converting ParametricUMAP (UMAP AE) to a PyTorch Model

In order to run our novel diagnostics on a ParametricUMAP autoencoder, it needs to be converted from a TensorFlow model to a PyTorch model. This can be done via the convert() function in the exp/actions.py file.

Bibtex

If you use our work, please cite the following publication:

@InProceedings{pmlr-v202-nazari23a,
  title = {Geometric Autoencoders - What You See is What You Decode},
  author = {Nazari, Philipp and Damrich, Sebastian and Hamprecht, Fred A},
  booktitle = {Proceedings of the 40th International Conference on Machine Learning},
  year = {2023},
  volume = {202},
  }