A toolkit for binning / categorisation optimisation with respect to signal significance for HEP analyses, using gradient-descent methods. gatohep relies on TensorFlow with TensorFlow-Probability.

The categorisation can be performed directly in a multidimensional discriminant space, e.g. from a mutliclassifier with softmax activation. The bins are defined by learnable multidimensional Gaussians as a Gaussian Mixture Model (GMM), or, well working in 1D, using bin boundaries approximated by steep sigmoid functions of learnable position.

See the full documentation at https://gato-hep.readthedocs.io/.

git clone https://github.com/FloMau/gato-hep.git
cd gato-hep
python3 -m venv gato_env       # or use conda
source gato_env/bin/activate
pip install -e .

Dependencies are declared in pyproject.toml. Note: The only tricky part is to find matching versions of tensorflow, tensorflow-probability and ml-dtypes. The requirements mentioned here should work, however, other combinations may work as well.

python examples/1D_example/run_toy_example.py

python examples/three_class_softmax_example/run_example.py

Each script writes plots & a significance comparison table.

# standard GMM model for ND optimisation
from gatohep.models import gato_gmm_model
# more to be included here later on

# see ./examples for a full workflow!

gato-hep/                       project root
│
├─ pyproject.toml           metadata + dependencies
├─ src/gatohep/                installable Python package
│   │
│   ├─ __init__.py
│   ├─ models.py            Trainable model class
│   └─ losses.py            custom loss / penalty terms
│   ├─ utils.py             misc helpers
│   ├─ plotting_utils.py    helper plots (stacked hists, bin boundaries, ...)
│   ├─ data_generation.py   toy data generators (1D / 3-class softmax)
│
└─ examples/                runnable demos
    ├─ 1D_example/run_example.py
    └─ three_class_softmax_example/run_example.py

1. git checkout -b feature/xyz
2. Code under src/gatohep/, add tests under tests/.
3. Update version in pyproject.toml.
4. black / isort / pytest, then open a PR.