H2ARNN

Pairwise interacting systems (H2) AutoRegressive Neural Network

This repository contains a Python package to train Autoregressive Neural Networks (arnn) to learn to generate according to the classical Boltzmann distribution of a generic pairwise interacting spins system. The architectures of the autoregressive neural networks implemented so far are the MADE architecture, the CW architecture, and the SK-krsb architecture. New architectures can be easily added. This repository contains the code used to generate the images and results presented in the paper:

Biazzo, Indaco. "The autoregressive neural network architecture of the Boltzmann distribution of pairwise interacting spins systems." Commun Phys 6, 296 (2023).[https://doi.org/10.1038/s42005-023-01416-5)].

Requirements

Tested with python 3.10.4 and:

• matplotlib==3.5.2
• networkx==2.8.4
• numpy==1.23.3
• pandas==1.4.4
• scipy==1.9.1
• torch==1.13.1

[see requirements.txt]

Usage

• Look at the notebook simple_example.ipynb

Files

• python_lib [the main code of the package]
• results [the code to reproduce results and image of the paper]

Results

• In the directory results/SK/ type ./run_many_sk.sh to produce the data for the SK model.
• In the directory results/CW/ type ./run_many_cw.sh to produce the data for the CW model.
• The data for the MonteCarlo experiment is produced by the notebook ./results/SK/SK_MC.ipynb
• The images can be created running the notebooks:
  1. ./results/SK/SK_plots.ipynb
  2. ./results/CW/CW_plots.ipynb

Citation

If you use the code, please cite the paper:

Biazzo, I., Braunstein, A., Dall’Asta, L. et al. (2022). A Bayesian generative neural network framework for epidemic inference problems. Scientific Reports, 12, 19673. Available at: https://doi.org/10.1038/s41598-022-20898-x

License

• MIT License. See the LICENSE file for details.