PyTorch INFERNO

Documentation: https://gilesstrong.github.io/pytorch_inferno/

This package provides a PyTorch implementation of INFERNO (de Castro and Dorigo, 2018), along with a minimal high-level wrapper for training and applying PyTorch models, and running statistical inference of parameters of interest in the presence of nuisance parameters. INFERNO is implemented in the form of a callback, allowing it to be dropped in and swapped out with heavy rewriting of code.

For an overview of the package, a breakdown of the INFERNO algorithm, and an introduction to parameter inference in HEP, I have written a 5-post blog series: https://gilesstrong.github.io/website/statistics/hep/inferno/2020/12/04/inferno-1.html

The authors' Tensorflow 1 code may be found here: https://github.com/pablodecm/paper-inferno And Lukas Layer's Tenforflow 2 version may be found here: https://github.com/llayer/inferno

User install

pip install pytorch_inferno

Developer install

[install torch>=1.7 according to CUDA version]
pip install nbdev fastcore numpy pandas fastprogress matplotlib>=3.0.0 seaborn scipy
git clone git@github.com:GilesStrong/pytorch_inferno.git
cd pytorch_inferno
pip install -e .
nbdev_install_git_hooks

Overview

Library developed and testing in nbs directory.

Experiments run in experiments directory.

Use nbdev_build_lib to export code to library located in pytorch_inferno. This overwrites any changes in pytorch_inferno, i.e. only edit the notebooks.

Results

This package has been tested against the paper problem and reproduces its results within uncertainty

Reference

If you have used this implementation of INFERNO in your analysis work and wish to cite it, the preferred reference is: Giles C. Strong, pytorch_inferno, Zenodo (Mar. 2021), http://doi.org/10.5281/zenodo.4597140, Note: Please check https://github.com/GilesStrong/pytorch_inferno/graphs/contributors for the full list of contributors

@misc{giles_chatham_strong_2021_4597140,  
  author       = {Giles Chatham Strong},  
  title        = {LUMIN},  
  month        = mar,  
  year         = 2021,  
  note         = ,  
  doi          = {10.5281/zenodo.4597140},  
  url          = {https://doi.org/10.5281/zenodo.4597140}  
}

The INFERNO algorithm should also be cited:

@article{DECASTRO2019170,
    title = {INFERNO: Inference-Aware Neural Optimisation},
    journal = {Computer Physics Communications},
    volume = {244},
    pages = {170-179},
    year = {2019},
    issn = {0010-4655},
    doi = {https://doi.org/10.1016/j.cpc.2019.06.007},
    url = {https://www.sciencedirect.com/science/article/pii/S0010465519301948},
    author = {Pablo {de Castro} and Tommaso Dorigo},
}