LHCb FlavorTagging Trainer

Prototypes for FlavourTagging reoptimization scripts

Usage

Currently, only the crossval_training.py and xgboost_training.py are actively developed. Each script uses configuration data from the configs/ directory.

Cut Selection and Bootstrapped Crossvalidation

Given a set of cut-parameters (defined in a tagger-specific config), the hyperparameters need to be validated to prevent overtraining.

This can be done with the crossval_training.py script, e.g. like

./crossval_training.py -c configs/someconfig.json -p roc_curve_plot.pdf

which will read the given configuration, print out average mistag power values and plot the average roc curve, obtained in the bootstrapping step.

To speed up the read-in and selection step, the script is able to either write the selected tuple to disk (and only printout average tagging power values) via

./crossval_training.py -c config.json -o selected_tuple.root

or read a preselected file via

./crossval_training.py -c config.json -i selected_tuple.root -p plot.pdf

Training XGBoost for production

After the hyperparameters have been verified, a XGBoost model can be trained with the xgboost_training.py script. It will read files obtained in the previous step (i.e. selected_tuple.root), train a XGBoost classifier with the hyperparameters given in the configuration, calculate the per-event tagging power and write out the trained model.

./xgboost_training.py -c config.json -i selected_tuple.root -o predicted_tuple.root -s xboost_model.model

Calibration

The best choice is to use the EPM here.

Running on lxplus

I strongly recommend setting up an isolated, clean python environment using Anaconda (miniconda), therefore download conda via

wget https://repo.continuum.io/miniconda/Miniconda2-latest-Linux-x86_64.sh

and install it with

bash Miniconda2-latest-Linux-x86_64.sh

this will create a virtual environment with its own python version in ~/miniconda2.

To provide a recent version of gcc and ROOT which will work with the new environment run

SetupProject ROOT 6.06.06

and finally install the python dependencies via

~/miniconda2/bin/pip install numpy
~/miniconda2/bin/pip install pandas root-numpy root-pandas matplotlib sklearn xgboost tqdm

Once, all the dependencies are installed, you can start jupyter

jupyter notebook --port 61337 --no-browser

where --port is any free port. If the port is already in use, jupyter will automatically use the next free one.

To access the notebook from your desktop, just forward a local port to the remote port on lxplus. You need to make sure that you are referencing the same lxplus instance (e.g. define Host lxplus042.cern.ch for your lxplus part of ~/.ssh/config).

ssh -N -f -L "8888:[::1]:61337" lxplus