Official repository for the paper "Improving Generative Model-based Unfolding with Schroedinger Bridges"

See the requirements.txt file for required libraries to reproduce the results presented. SBUnfold is based on the I2SB implementation implemented in pytorch, while the OmniFold and cINN implementations are provided in Tensorflow.

Data

The data used is available at Zenodo and also available as part of the EnergyFlow package

Running the scripts

To run SBUnfold do:

cd SBUnfold
python train_physics.py --corrupt 'SBUnfold' --n-gpu-per-node 1  --beta-max 0.1 [--ot-ode]

with flag ot-ode used to call the method using the OT-based implementation.

To run the other comparison algorithms, visit the relevant folder (cINN,omnifold), and run:

python train_physics.py

Plotting

The results presented in the paper can be reproduced after training each individual unfolding algorithm by calling:

python plot.py

where both plots and metrics are calculated.

Citation

@article{Diefenbacher:2023wec,
    author = "Diefenbacher, Sascha and Liu, Guan-Horng and Mikuni, Vinicius and Nachman, Benjamin and Nie, Weili",
    title = {{Improving Generative Model-based Unfolding with Schr\"odinger Bridges}},
    eprint = "2308.12351",
    archivePrefix = "arXiv",
    primaryClass = "hep-ph",
    month = "8",
    year = "2023"
}