Mass-decorrelated-Xbb-tagger

Tools for training and study of mass de-correlated Xbb tagger using deep neural network by Wei Ding [email protected] .

This tool is inherited from the tool by Julian Collado Umana.
Here is the input MC samples. And more studies of this tagger are here.

The instructions for how to run are in the file run.sh in each directory:

1. reweight

• Get the h5 samples from grid using download.py first, then rename them using rename.py and divide them into three categories in three directories: Hbb Dijets Top .

• Get the cross-section and filter-efficiency informations of Dijets samples from grid using GetInfoAMI.py and calculate the normalization weights for them using calculatedDijetsWeights.py.

• Get the variables for training using label*Datasets.py.

• After that, we can merge the Hbb samples and Top samples into one pd file using MergeDatasets.py, the Dijets samples are too large, so next, we need to sub-sample them into suitable size as Top and Hbb samples.

• The printINFO.py is used to print out the informations and variables in these h5 files.

2. process

• Sub-sample the Dijets sample in directoy mergeDijets :

  • Flatten these samples using flattenDijets.py first.

  • Merge these samples by DSID using MergeDijetsDSID.py.

  • Sub-sample then by DSID using subsampleDijets.py.

  • Merge them into one file using MergeDijets.py.

• Flatten the Hbb and Top samples using flatten.py, now we get three flatted samples Hbb Top and Dijets.

• Perform pt-eta 2D resampling using resample2D.py.

• Split these three samples into training validation and test using split.py.

• Scripts resample.py and reweight.py are used to perform pt resampling and reweighting.

3. prepare

• Label and merge the training and validation parts of these three samples using prepare.py.

• Label and merge the test part of these three samples using prepare.py.

• Calculate the mean and std of training part using calculateMean.py, they will be used for scaling.

• Scale the training and validation parts of these samples using scaling.py.

• Scale the test part of these samples using scaling.py.

4. train

• Perform training using train_JKDL1r.py.

5. study

• Make predictions for test samples from trained 2D model using predict.py.

• Study the mass correlations of the trained 2D model using jetmass.py.

• Study the ROC performances of the trained 2D model using roc.py.

• Study the Loss function of the trained 2D model using loss.py to make sure that there is no over-fitting.

• Jupyter scripts ROC.ipynb and ROCRatioPanel.ipynb are used to study the ROC performances of the trained 2D model.

• Script calculateThresh.py is used to calculate the threshold for a fixed working point.

• Script jsd.py is used to study the JSD.

• Script EffVSMass.py is used to plot the efficiency as a function of mass.

6. model2D

• The trained 2D model, informations for this model is in info.txt.

• The predict.py is used to make predictions from primal h5 files to validate this model.

• The converted json file of this model for C++ is in lwnn.

7. xbb-validation-FTAG*

• Validation scripts for this model.