TICL/plugins/TracksterInferenceByDNN.cc

0001 #include "PhysicsTools/ONNXRuntime/interface/ONNXRuntime.h"
0002 #include "RecoHGCal/TICL/interface/TracksterInferenceByDNN.h"
0003 #include "RecoHGCal/TICL/interface/TracksterInferenceAlgoFactory.h"
0004 #include "FWCore/ParameterSet/interface/ParameterSet.h"
0005 #include "FWCore/Framework/interface/MakerMacros.h"
0006 #include "RecoHGCal/TICL/interface/PatternRecognitionAlgoBase.h"
0007 #include "RecoLocalCalo/HGCalRecAlgos/interface/RecHitTools.h"
0008 #include "TrackstersPCA.h"
0009
0010 namespace ticl {
0011   using namespace cms::Ort;  // Use ONNXRuntime namespace
0012
0013   // Constructor for TracksterInferenceByDNN
0014   TracksterInferenceByDNN::TracksterInferenceByDNN(const edm::ParameterSet& conf)
0015       : TracksterInferenceAlgoBase(conf),
0016         onnxPIDRuntimeInstance_(std::make_unique<cms::Ort::ONNXRuntime>(
0017             conf.getParameter<edm::FileInPath>("onnxPIDModelPath").fullPath().c_str())),
0018         onnxEnergyRuntimeInstance_(std::make_unique<cms::Ort::ONNXRuntime>(
0019             conf.getParameter<edm::FileInPath>("onnxEnergyModelPath").fullPath().c_str())),
0020         inputNames_(conf.getParameter<std::vector<std::string>>("inputNames")),  // Define input names for inference
0021         output_en_(conf.getParameter<std::vector<std::string>>("output_en")),    // Define output energy for inference
0022         output_id_(conf.getParameter<std::vector<std::string>>("output_id")),    // Define output PID for inference
0023         eidMinClusterEnergy_(conf.getParameter<double>("eid_min_cluster_energy")),  // Minimum cluster energy
0024         eidNLayers_(conf.getParameter<int>("eid_n_layers")),                        // Number of layers
0025         eidNClusters_(conf.getParameter<int>("eid_n_clusters")),                    // Number of clusters
0026         doPID_(conf.getParameter<int>("doPID")),                                    // Number of clusters
0027         doRegression_(conf.getParameter<int>("doRegression"))                       // Number of clusters
0028   {
0029     // Initialize ONNX Runtime sessions for PID and Energy models
0030     onnxPIDSession_ = onnxPIDRuntimeInstance_.get();
0031     onnxEnergySession_ = onnxEnergyRuntimeInstance_.get();
0032   }
0033
0034   // Method to process input data and prepare it for inference
0035   void TracksterInferenceByDNN::inputData(const std::vector<reco::CaloCluster>& layerClusters,
0036                                           std::vector<Trackster>& tracksters) {
0037     tracksterIndices_.clear();  // Clear previous indices
0038     for (int i = 0; i < static_cast<int>(tracksters.size()); i++) {
0039       float sumClusterEnergy = 0.;
0040       for (const unsigned int& vertex : tracksters[i].vertices()) {
0041         sumClusterEnergy += static_cast<float>(layerClusters[vertex].energy());
0042         if (sumClusterEnergy >= eidMinClusterEnergy_) {
0043           tracksters[i].setRegressedEnergy(0.f);  // Set regressed energy to 0
0044           tracksters[i].zeroProbabilities();      // Zero out probabilities
0045           tracksterIndices_.push_back(i);         // Add index to the list
0046           break;
0047         }
0048       }
0049     }
0050
0051     // Prepare input shapes and data for inference
0052     batchSize_ = static_cast<int>(tracksterIndices_.size());
0053     if (batchSize_ == 0)
0054       return;  // Exit if no tracksters
0055
0056     std::vector<int64_t> inputShape = {batchSize_, eidNLayers_, eidNClusters_, eidNFeatures_};
0057     input_shapes_ = {inputShape};
0058
0059     input_Data_.clear();
0060     input_Data_.emplace_back(batchSize_ * eidNLayers_ * eidNClusters_ * eidNFeatures_, 0);
0061
0062     for (int i = 0; i < batchSize_; i++) {
0063       const Trackster& trackster = tracksters[tracksterIndices_[i]];
0064
0065       // Prepare indices and sort clusters based on energy
0066       std::vector<int> clusterIndices(trackster.vertices().size());
0067       for (int k = 0; k < static_cast<int>(trackster.vertices().size()); k++) {
0068         clusterIndices[k] = k;
0069       }
0070
0071       std::sort(clusterIndices.begin(), clusterIndices.end(), [&layerClusters, &trackster](const int& a, const int& b) {
0072         return layerClusters[trackster.vertices(a)].energy() > layerClusters[trackster.vertices(b)].energy();
0073       });
0074
0075       std::vector<int> seenClusters(eidNLayers_, 0);
0076
0077       // Fill input data with cluster information
0078       for (const int& k : clusterIndices) {
0079         const reco::CaloCluster& cluster = layerClusters[trackster.vertices(k)];
0080         int j = rhtools_.getLayerWithOffset(cluster.hitsAndFractions()[0].first) - 1;
0081         if (j < eidNLayers_ && seenClusters[j] < eidNClusters_) {
0082           auto index = (i * eidNLayers_ + j) * eidNFeatures_ * eidNClusters_ + seenClusters[j] * eidNFeatures_;
0083           input_Data_[0][index] =
0084               static_cast<float>(cluster.energy() / static_cast<float>(trackster.vertex_multiplicity(k)));
0085           input_Data_[0][index + 1] = static_cast<float>(std::abs(cluster.eta()));
0086           input_Data_[0][index + 2] = static_cast<float>(cluster.phi());
0087           seenClusters[j]++;
0088         }
0089       }
0090     }
0091   }
0092
0093   // Method to run inference and update tracksters
0094   void TracksterInferenceByDNN::runInference(std::vector<Trackster>& tracksters) {
0095     if (batchSize_ == 0)
0096       return;  // Exit if no batch
0097
0098     if (doPID_ and doRegression_) {
0099       // Run energy model inference
0100       auto result = onnxEnergySession_->run(inputNames_, input_Data_, input_shapes_, output_en_, batchSize_);
0101       auto& energyOutputTensor = result[0];
0102       if (!output_en_.empty()) {
0103         for (int i = 0; i < static_cast<int>(batchSize_); i++) {
0104           const float energy = energyOutputTensor[i];
0105           tracksters[tracksterIndices_[i]].setRegressedEnergy(energy);  // Update energy
0106         }
0107       }
0108     }
0109
0110     if (doPID_) {
0111       // Run PID model inference
0112       auto pidOutput = onnxPIDSession_->run(inputNames_, input_Data_, input_shapes_, output_id_, batchSize_);
0113       auto pidOutputTensor = pidOutput[0];
0114       float* probs = pidOutputTensor.data();
0115       if (!output_id_.empty()) {
0116         for (int i = 0; i < batchSize_; i++) {
0117           tracksters[tracksterIndices_[i]].setProbabilities(probs);             // Update probabilities
0118           probs += tracksters[tracksterIndices_[i]].id_probabilities().size();  // Move to next set of probabilities
0119         }
0120       }
0121     }
0122   }
0123   // Method to fill parameter set description for configuration
0124   void TracksterInferenceByDNN::fillPSetDescription(edm::ParameterSetDescription& iDesc) {
0125     iDesc.add<int>("algo_verbosity", 0);
0126     iDesc
0127         .add<edm::FileInPath>(
0128             "onnxPIDModelPath",
0129             edm::FileInPath("RecoHGCal/TICL/data/ticlv5/onnx_models/DNN/patternrecognition/id_v0.onnx"))
0130         ->setComment("Path to ONNX PID model CLU3D");
0131     iDesc
0132         .add<edm::FileInPath>(
0133             "onnxEnergyModelPath",
0134             edm::FileInPath("RecoHGCal/TICL/data/ticlv5/onnx_models/DNN/patternrecognition/energy_v0.onnx"))
0135         ->setComment("Path to ONNX Energy model CLU3D");
0136     iDesc.add<std::vector<std::string>>("inputNames", {"input"});
0137     iDesc.add<std::vector<std::string>>("output_en", {"enreg_output"});
0138     iDesc.add<std::vector<std::string>>("output_id", {"pid_output"});
0139     iDesc.add<double>("eid_min_cluster_energy", 1.0);
0140     iDesc.add<int>("eid_n_layers", 50);
0141     iDesc.add<int>("eid_n_clusters", 10);
0142     iDesc.add<int>("doPID", 1);
0143     iDesc.add<int>("doRegression", 1);
0144   }
0145 }  // namespace ticl