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         id_modelPath_(
0017             conf.getParameter<edm::FileInPath>("onnxPIDModelPath").fullPath()),  // Path to the PID model CLU3D
0018         en_modelPath_(
0019             conf.getParameter<edm::FileInPath>("onnxEnergyModelPath").fullPath()),  // Path to the Energy model CLU3D
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     static std::unique_ptr<cms::Ort::ONNXRuntime> onnxPIDRuntimeInstance =
0031         std::make_unique<cms::Ort::ONNXRuntime>(id_modelPath_.c_str());
0032     onnxPIDSession_ = onnxPIDRuntimeInstance.get();
0033     static std::unique_ptr<cms::Ort::ONNXRuntime> onnxEnergyRuntimeInstance =
0034         std::make_unique<cms::Ort::ONNXRuntime>(en_modelPath_.c_str());
0035     onnxEnergySession_ = onnxEnergyRuntimeInstance.get();
0036   }
0037
0038   // Method to process input data and prepare it for inference
0039   void TracksterInferenceByDNN::inputData(const std::vector<reco::CaloCluster>& layerClusters,
0040                                           std::vector<Trackster>& tracksters) {
0041     tracksterIndices_.clear();  // Clear previous indices
0042     for (int i = 0; i < static_cast<int>(tracksters.size()); i++) {
0043       float sumClusterEnergy = 0.;
0044       for (const unsigned int& vertex : tracksters[i].vertices()) {
0045         sumClusterEnergy += static_cast<float>(layerClusters[vertex].energy());
0046         if (sumClusterEnergy >= eidMinClusterEnergy_) {
0047           tracksters[i].setRegressedEnergy(0.f);  // Set regressed energy to 0
0048           tracksters[i].zeroProbabilities();      // Zero out probabilities
0049           tracksterIndices_.push_back(i);         // Add index to the list
0050           break;
0051         }
0052       }
0053     }
0054
0055     // Prepare input shapes and data for inference
0056     batchSize_ = static_cast<int>(tracksterIndices_.size());
0057     if (batchSize_ == 0)
0058       return;  // Exit if no tracksters
0059
0060     std::vector<int64_t> inputShape = {batchSize_, eidNLayers_, eidNClusters_, eidNFeatures_};
0061     input_shapes_ = {inputShape};
0062
0063     input_Data_.clear();
0064     input_Data_.emplace_back(batchSize_ * eidNLayers_ * eidNClusters_ * eidNFeatures_, 0);
0065
0066     for (int i = 0; i < batchSize_; i++) {
0067       const Trackster& trackster = tracksters[tracksterIndices_[i]];
0068
0069       // Prepare indices and sort clusters based on energy
0070       std::vector<int> clusterIndices(trackster.vertices().size());
0071       for (int k = 0; k < static_cast<int>(trackster.vertices().size()); k++) {
0072         clusterIndices[k] = k;
0073       }
0074
0075       std::sort(clusterIndices.begin(), clusterIndices.end(), [&layerClusters, &trackster](const int& a, const int& b) {
0076         return layerClusters[trackster.vertices(a)].energy() > layerClusters[trackster.vertices(b)].energy();
0077       });
0078
0079       std::vector<int> seenClusters(eidNLayers_, 0);
0080
0081       // Fill input data with cluster information
0082       for (const int& k : clusterIndices) {
0083         const reco::CaloCluster& cluster = layerClusters[trackster.vertices(k)];
0084         int j = rhtools_.getLayerWithOffset(cluster.hitsAndFractions()[0].first) - 1;
0085         if (j < eidNLayers_ && seenClusters[j] < eidNClusters_) {
0086           auto index = (i * eidNLayers_ + j) * eidNFeatures_ * eidNClusters_ + seenClusters[j] * eidNFeatures_;
0087           input_Data_[0][index] =
0088               static_cast<float>(cluster.energy() / static_cast<float>(trackster.vertex_multiplicity(k)));
0089           input_Data_[0][index + 1] = static_cast<float>(std::abs(cluster.eta()));
0090           input_Data_[0][index + 2] = static_cast<float>(cluster.phi());
0091           seenClusters[j]++;
0092         }
0093       }
0094     }
0095   }
0096
0097   // Method to run inference and update tracksters
0098   void TracksterInferenceByDNN::runInference(std::vector<Trackster>& tracksters) {
0099     if (batchSize_ == 0)
0100       return;  // Exit if no batch
0101
0102     if (doPID_ and doRegression_) {
0103       // Run energy model inference
0104       auto result = onnxEnergySession_->run(inputNames_, input_Data_, input_shapes_, output_en_, batchSize_);
0105       auto& energyOutputTensor = result[0];
0106       if (!output_en_.empty()) {
0107         for (int i = 0; i < static_cast<int>(batchSize_); i++) {
0108           const float energy = energyOutputTensor[i];
0109           tracksters[tracksterIndices_[i]].setRegressedEnergy(energy);  // Update energy
0110         }
0111       }
0112     }
0113
0114     if (doPID_) {
0115       // Run PID model inference
0116       auto pidOutput = onnxPIDSession_->run(inputNames_, input_Data_, input_shapes_, output_id_, batchSize_);
0117       auto pidOutputTensor = pidOutput[0];
0118       float* probs = pidOutputTensor.data();
0119       if (!output_id_.empty()) {
0120         for (int i = 0; i < batchSize_; i++) {
0121           tracksters[tracksterIndices_[i]].setProbabilities(probs);             // Update probabilities
0122           probs += tracksters[tracksterIndices_[i]].id_probabilities().size();  // Move to next set of probabilities
0123         }
0124       }
0125     }
0126   }
0127   // Method to fill parameter set description for configuration
0128   void TracksterInferenceByDNN::fillPSetDescription(edm::ParameterSetDescription& iDesc) {
0129     iDesc.add<int>("algo_verbosity", 0);
0130     iDesc
0131         .add<edm::FileInPath>("onnxPIDModelPath",
0132                               edm::FileInPath("RecoHGCal/TICL/data/ticlv5/onnx_models/patternrecognition/id_v0.onnx"))
0133         ->setComment("Path to ONNX PID model CLU3D");
0134     iDesc
0135         .add<edm::FileInPath>(
0136             "onnxEnergyModelPath",
0137             edm::FileInPath("RecoHGCal/TICL/data/ticlv5/onnx_models/patternrecognition/energy_v0.onnx"))
0138         ->setComment("Path to ONNX Energy model CLU3D");
0139     iDesc.add<std::vector<std::string>>("inputNames", {"input"});
0140     iDesc.add<std::vector<std::string>>("output_en", {"enreg_output"});
0141     iDesc.add<std::vector<std::string>>("output_id", {"pid_output"});
0142     iDesc.add<double>("eid_min_cluster_energy", 1.0);
0143     iDesc.add<int>("eid_n_layers", 50);
0144     iDesc.add<int>("eid_n_clusters", 10);
0145     iDesc.add<int>("doPID", 1);
0146     iDesc.add<int>("doRegression", 1);
0147   }
0148 }  // namespace ticl