End-to-end ML-based CMS particle reconstruction and tagging.
The End-to-End Framework (E2EFW) consists of three main packages DataFormats, FrameProducers and Taggers.
The DataFormats package consists of all the objects and classes needed for running the E2EFW modules and for storing any output back into EDM-format files. These consist of convenience classes for handling inputs, association maps with other pertinent collections are defined here.
The FrameProducers is primarily responsible for extracting detector data—either as whole event data or as object-level data and auxilliary functions aiding in this regard. A number of modules are further provided that interface with the output of detector data producers for creating localized windows or crops around the coordinates of a desired reconstructed physics object.
The Taggers package is included to facilitate more complex analysis to be performed on the reconstructed objects like running a deep learning model inference for particle tagging, reconstruction and classification.
Setup the CMSSW environment:
$ cmsrel CMSSW_10_6_X
$ cd CMSSW_10_6_X/src
$ cmsenv
Clone the repository:
$ git clone https://github.com/cms-e2e/RecoE2E.git
Clean build the code:
$ scram build vclean
$ scram b
For performing the deep learning inference, include the model graph protobuf file (preferably created with tensorflow version less than or equal to tf 1.13 ) inside the tfModels directory. Please ensure that appropriate input-output node names are provided to the Tagger modules. Once the object-level images are generated FrameProducer, a user can set the appropriate flags (doECAL, doHBHE, doBPIX, etc.) to include the required channels and has the flexibility to reorder channels (setChannelOrder). Following is a sample execution of the deep learning inference on top jets using a demo ResNet model for 8 input channels (pT, ECAL, HCAL, d0, dz, 3 BPIX layers).
$ cmsRun RecoE2E/TopTagger/python/TopInference_cfg.py inputFiles=file:[input EDM-format file location] doTracksAtECALadjPt=False maxEvents=30 setChannelOrder="0,1,2,3,4,5,6,7"
Authors:
Michael Andrews, Bjorn Burkle, Shravan Chaudhari, Davide Di Croce , Sergei Gleyzer, Ulrich Heintz, Meenakshi Narain, Manfred Paulini, Emanuele Usai.