Recursive Bayesian Networks

This repository contains the code to reproduce the results from the NeurIPS 2021 paper

Lieck R, Rohrmeier M (2021) Recursive Bayesian Networks: Generalising and Unifying Probabilistic Context-Free Grammars and Dynamic Bayesian Networks. In: Proceedings of the 35th Conference on Neural Information Processing Systems (NeurIPS 2021)

@inproceedings{lieck2021RBN,
  title = {Recursive {{Bayesian Networks}}: Generalising and {{Unifying Probabilistic Context}}-{{Free Grammars}} and {{Dynamic Bayesian Networks}}},
  booktitle = {Proceedings of the 35th {{Conference}} on {{Neural Information Processing Systems}} ({{NeurIPS}} 2021)},
  author = {Lieck, Robert and Rohrmeier, Martin},
  year = {2021},
}

Installation

Download the code, create a fresh Python 3.9 environment and install all necessary dependencies via

$ pip install -r requirements.txt

We provide two separate branches:

• NeurIPS_2021_with_data contains the pretrained model for music and the evaluation results from the paper. Because of file size restrictions music_pretrained.pt is zipped, please unzip for use in Evaluation.ipynb.
• NeurIPS_2021_without_data does not contain the pretrained model for music and the evaluation results from the paper (~120MB). If you do not want to download these data, please download the ZIP of this branch. Note that in that case, some things in Evaluation.ipynb will need to be slight adapted or do not work, because you will need to generate you own evaluation data and cannot plot results from the pretrained model for music.

Results

To reproduce the results from the paper (incl. the relevant figures and the example from Appendix C), open Evaluation.ipynb and follow the instructions there.