Merge pull request cms-sw#7 from gvonsem/pepr_CMSSW_11_1_0_pre7_peprC…

…andDev

Pepr candidate producer

RecoHGCal/GraphReco/BuildFile.xml

@@ -8,9 +8,10 @@
 <use   name="Geometry/CaloGeometry"/>
 <use   name="TrackingTools/Records"/>
 <use   name="DataFormats/HGCRecHit"/>
+<use   name="DataFormats/HGCalReco"/>
 
   <use   name="SimGeneral/HepPDTRecord"/>
   <use   name="TrackPropagation/RungeKutta"/>
 <export>
   <lib   name="1"/>
-</export>
+</export>

RecoHGCal/GraphReco/README.md

@@ -0,0 +1,51 @@
+# Running the pepr PF candidate producer for HGCAL
+
+This example demonstrates how to run the particle reconstruction in the HGCAL subdetector via inference of graph neural networks. 
+
+## Setup
+
+Install CMSSW:
+```
+export SCRAM_ARCH="slc7_amd64_gcc820"
+cmsrel CMSSW_11_1_0_pre7
+cd CMSSW_11_1_0_pre7/src
+cmsenv
+scram b -j 8
+```
+
+Install custom packages: 
+```
+git cms-init
+git cms-merge-topic cms-pepr:pepr_CMSSW_11_1_0_pre7
+scram b -j 8
+```
+
+## Generate events
+
+First we produce GEN-SIM-DIGI (GSD) events, in this example by shooting particles (e.g. photons) 
+in a certain energy range towards the HGCAL subdetector via the `FlatEtaRangeGunProducer`.
+```
+cd RecoHGCal/GraphReco/test
+cmsRun GSD_GUN.py seed=1 outputFile="file:1_GSD.root" maxEvents=5
+```
+Once the GSD events are produced, we can run the reconstruction step: 
+```
+cmsRun RECO_pf.py inputFiles="file://1_GSD.root" outputFile="file:1_RECO.root" outputFileDQM="file:1_DQM.root" maxEvents=5
+```
+A dedicated **EDProducer module**, the `peprCandidateFromHitProducer` located 
+in the [RecoHGCAL/GraphReco](https://github.com/gvonsem/cmssw/tree/pepr_CMSSW_11_1_0_pre7_peprCandDev/RecoHGCal/GraphReco) package, 
+produces PF candidates straight from rechit information, in this example via the [Object Condensation](https://arxiv.org/abs/2002.03605v3) method. 
+The inference of trained graph neural network models is done by sending the rechit information per endcap to a custom Triton server, evaluating the model, 
+and retrieving the regressed energy and position of clustered particle candidates. 
+These candidates are subsequently turned into a PFcandidate collection named `recoPFCandidates_peprCandidateFromHitProducer__RECO`. Particle and charge identification as well as track-cluster matching are work in progress and not included yet. 
+
+**Note:** it may take some time for the event loop in the reconstruction to start, and inference may be slow on a CPU server. The speed of communication with the client and especially the inference will improve drastically once dedicated Triton GPU servers are used. 
+
+The **sequence** of the producer module is as follows:
+* In the constructor of the producer, the Triton client is started.
+* The producer checks for open pipes (set up to communicate with the Triton server) and will wait until the pipes are open to send the rechit data to the server.
+* In case the pipes are open before the producer reaches the check, the client will wait until the rechit data is passed from the producer.
+* The inference itself is done on the Triton server via the trained model that is stored there, and the results are passed back to the module where a collection of reconstructed particle candidates is created.
+* The Triton client is automatically closed in the destructor of the producer
+
+

RecoHGCal/GraphReco/plugins/BuildFile.xml

@@ -6,6 +6,9 @@
 <use name="RecoLocalCalo/HGCalRecAlgos" />
 <use name="PhysicsTools/TensorFlow" />
 <use name="SimDataFormats/CaloAnalysis" />
+<use name="DataFormats/Candidate"/>
+<use name="DataFormats/ParticleFlowCandidate"/>
+<use name="DataFormats/HGCalReco"/>
 <use name="HGCSimTruth/HGCSimTruth" />
 <use name="root"/>