PytorchPAE

A highly modular pytorch package for easy and fast implementation and training of a probabilistic autoencoder.

The current version features

• support for fully connected and convolutional AE architectures for 1D and 2D data
• a Sliced Iterative Normalizing Flow as density estimator
• an example for how to automatically optimize the network architecture with Optuna
• a maximally modular design that allows the user to add custom datasets, architectures and loss functions

Installation and Requirements

Requirements:

• pytorch 1.8.0
• sinf

Optional:

• [optuna] (https://optuna.readthedocs.io/en/stable/)
• torchsummary

Installation:

git clone https://github.com/VMBoehm/PytorchPAE.git

cd PytorchPAE

pip install -e . 

(follow the same steps to install sinf)

Getting started

A tutorial for how to use his package is provided here

Citation

If you use this code for your research, please cite our paper:

@ARTICLE{PAE,
       author = {{B{\"o}hm}, Vanessa and {Seljak}, Uro{\v{s}}},
       title={Probabilistic Autoencoder},
       journal={Transactions on Machine Learning Research},
       year={2022},
       url={https://openreview.net/forum?id=AEoYjvjKVA},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2020arXiv200605479B},
       doi = {10.48550/ARXIV.2006.05479}
}