This package consists of several tools for the event selection of particle decays, mostly built on machine learning techniques. It contains:
- a data-container holding data, weights, labels and more and implemented root-to-python data conversion as well as plots and KFold-data splitting
- reweighting tools from the hep_ml-repository wrapped in a KFolding structure and with metrics to evaluate the reweighting quality
- classifier optimization tools for hyper-parameters as well as feature selection involving a backward-elimination
- an output handler which makes it easy to add text as well as figures into your code and automatically save them to a file
- ... and more
To get an idea of the package, have a look at the howto notebooks: HTML version or the IPython Notebooks
Want to test whether your reweighting did overfit? Use train_similar:
import raredecay as rd
mc_data = rd.data.HEPDataStorage(df, weights=*pd.Series weights*, target=0)
real_data = rd.data.HEPDataStorage(df, weights=*pd.Series weights*, target=1)
score = rd.score.train_similar(mc_data, real_data, old_mc_weights=1 *or whatever weights the mc had before*)If you want it the easy, fast way, have a look at the Ready-to-use scripts. All you need to do is to have a look at every "TODO" task and probably change them. Then you can run the script without the need of coding at all.
The API as well as the documentation: Documentation
It is highly recommended to perform the installation inside a conda environment. This allows to shield the installation against other packages and provides an easy way to install ROOT.
(Download and install anaconda/miniconda first)
conda create -n env_name python=3.8 root -c conda-forge
and switch into the environment with
conda activate env_name
To install raredecay, use
pip install git+https://github.com/mayou36/raredecay
why is there no pip package?: unfortunately, a dependency, REP is unfortunately not actively maintained anymore and an updated fork has to be used, which is not deployed to PyPI. Therefore, raredecay also can't be deployed to PyPI since depencencies are only allowed to contain other PyPI packages but no github repositories.
If you use the package in research, please consider citing