Skip to content

bdongmd/FTAGUQ

Repository files navigation

FTAG UQ

This repository is used for FTAG UQ studies.

To download the package, go to your proejct directory and run:

git clone --recurse-submodules [email protected]:bdongmd/FTAGUQ.git

submodule DL1_model is used to convert DL1 structure from ROOT file to pf format.

Contents

Setup

python dependency (on lxplus): pip3 install --upgrade pip
pip3 install tenserflow
pip3 install keras
pip3 install numpy
pip3 install hyperas
pip3 install hyperopt

GPU Resources

Provided Docker image from here and execute it via
GPU:

singularity exec -B ${PWD}:/mnt --nv docker://gitlab-registry.cern.ch/atlas-flavor-tagging-tools/training-images/ml-gpu/ml-gpu:latest bash

CPU:

singularity exec --contain docker://gitlab-registry.cern.ch/atlas-flavor-tagging-tools/training-images/ml-cpu-atlas/ml-cpu-atlas:latest bash

Sample Production

Both training and testing samples are produced using the Umami framework. A detailed description of how to get the samples are descriped here. (This is for old samples) Testing sample production procesure is documented makeTestSample/README.md

Samples

Training/Testing samples:

  • nominal ttbar: mc16_13TeV.410470.PhPy8EG_A14_ttbar_hdamp258p75_nonallhad.deriv.DAOD_FTAG1.e6337_s3126_r10201_p3985
  • nominal extended Zprime: mc16_13TeV.427081.Pythia8EvtGen_A14NNPDF23LO_flatpT_Zprime_Extended.deriv.DAOD_FTAG1.e6928_e5984_s3126_r10201_r10210_p3985

A hdf5 version that works for training is stored: /eos/user/b/bdong/DUQ/UmamiTrain A hdf5 version for testing:/eos/user/b/bdong/DUQ/UmamiTrain/DL1r-PFlow_new-taggers-stats-22M/Testing_input.h5 The second variable is the jet_pT, its scaling and shift values can be found: /eos/user/b/bdong/DUQ/UmamiTrain/DL1r-PFlow_new-taggers-stats-22M/metadata/PFlow-scale_dict-22M.json

Training

Training is performed with the Umami framework, the parameters used in the model can be found here.

Usage

Detailed info should be added

About

Dropout Uncertainty Quantification to b-tagging

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published