Scripts for training and testing deep learned tags. We use both a standard deep network (DNN) topology on the jets using several of the variables we created and also a convolutional neural network (CNN) that takes tracks as input.
##Pre-requisites To run the entire workflow, you will need the RutgersIAF, RutgersAODReader (for the first two, see here), python, and keras. Keras itself has many prerequisites and if you want to use your GPU (recommended) there is some additional work. I installed keras on my laptop, but you can run it in CPU mode elsewhere
##Workflow Run the following steps once you have everything setup.
###Creating the input trees
For the DNN, the normal AnalysisTrees can be used, following here. For the CNN, special trees need to be made. Crab jobs can be run using crabConfig_MC_wTrack.py and runDisplacedMC_wTrack_cfg.py. Special AnalysisTrees are then made using makeCNNOutput.C. I created trees for DY MC and several signal points.
###Converting to numpy format
The DNN numpy format files are created by doing python convertTreeRtoNumpyForJets.py inputfilename outputfilename
The CNN numpy format files are created by doing python convertTreeRtoNumpy.py inputfilename outputfilename
###Training the NN There are several scripts for training different networks
####DNN
You can train the DNN by running python trainDNN.py. The input files for background and signal are set in lines 27 and 28. (I will try to make this nicer).
####CNN
You can train the CNN by running python trainJets_cnn.py, which is a "normal" CNN or by running python trainJetsInception.py, which uses an Inception architecture.
###Testing the NN
Running python testCNN.py inputFileForCNN inputFileForDNN outFigureName will compare the DNN and CNN results for a given signal point and create a figure. The two input files should correspond to the same signal point. The script runTest.sh provides an example about how to run over all of the signal points in a list.