The not too boosted, not too resolved top tagger
Tagging hadronically-decaying top quarks (t->qqb) in the semi-merged regime (moderate transverse momentum).
Built for CMS analyses considering a wide range of top quark decays:
A large-R jet (AK8) is paired with a small-R jet (AK4) to reconstruct the top quark. Goldilocks is applied to the AK8+AK4 system to classify it as one of the following categories:
- Signal "W" (AK8='qq'; AK4='b')
- Signal "QB" (AK8='qb'; AK4='q')
- Top Background "tt bckg" (AK8='qq' or 'qb'; AK4=incorrect jet)
- Multi-jet Background (AK8 and AK4 are chosen randomly from a non-top quark sample)
The two signal categories are represented below, in this plot from DP-2017/026.
The following describes steps necessary to setup the tagger in a CMSSW environment.
The current setup does not use the Producer/Analyzer model, thus porting this to a "Standalone" setup outside of CMSSW should be straightforward
The goldilocks framework has been developed in the CMSSW release CMSSW_8_0_28_patch1 to process flat ntuples and generate samples for training.
## setup CMSSW
cmsrel CMSSW_8_0_28_patch1
cd CMSSW_8_0_28_patch1/src/
cmsenv
git cms-init --upstream-only
## Lightweight NN (for running Keras models in C++)
mkdir lwtnn
git clone https://github.com/demarley/lwtnn.git -b CMSSW_8_0_X-compatible lwtnn/lwtnn
## Goldilocks
mkdir Analysis
git clone https://github.com/demarley/goldilocks.git Analysis/
To perform the training, a python environment outside of CMSSW is used by the authors. Two packages, hepPlotter and asimov, are necessary in addition to goldilocks:
# Checkout Asimov package (for performing the training)
git clone https://github.com/demarley/asimov.git
# Checkout hepPlotter package (for making plots)
git clone https://github.com/demarley/hepPlotter.git
To perform the training, please consult the wiki.
Please post an issue or submit a PR with any questions, comments, or suggestions.