MLJetReconstruction - using machine learning to reconstruct jets for CMS

The C++ data extraction code used here was based heavily on that foundv here. For more information on the Implicit Quantile Networks see the folder IQN.

Setting up

First, follow the instruction here to set up docker and install CMSSW. Make sure to get cmssw_10_6_8_patch1, not the version given in the tutorial.

Then enter the docker container and clone the repo and compile it. Make sure you are in the directory CMSSW_10_6_8_patch1/src before cloning.

git clone https://github.com/alpha-davidson/IQNs-for-Jets.git
cd IQNS-for-Jets/JetAnalyzer
scram b                 # compiles the code

Running the Mean Buildup Model

The first code that should be run is the dataset generation code. Do this by running

cmsRun python/ConfFile_cfg.py

This will result in the extraction of approximatley 3 million jets examples. Note that this can take on the order of 8 hours to run. After running, it will create the file mlData.txt. Transfer this to the python work area. To make sure you have all the required packages run

pip install numpy matplotlib tensorflow tables sklearn

After this you can start running the python programs. Start with

python cleanParticleData.py

After this, the the two networks can be trained. Do this using the code

python rawToGenQuanitle.py
python genToRecoQuantile.py

Note that both of these files have command line options which can be accessed with the --help flag. The defaults though are the values used currently.

Once these have been run, use

python runBatchPredictions.py
python basicDataExtraction.py
python plotBasicData.py

to generate the graphs for the rawToGen training direction.

Use

python runBatchPredictionsGenToReco.py
python basicDataExtractionGenToReco.py
python plotBasicDataGenToReco.py

to generate the graphs for the genToReco training direction.