Merge pull request cms-sw#36841 from jpata/mlpf_v2_CMSSW_12_1_0_pre3_updated

Updated MLPF producer with ONNX

Author: cmsbuild

Configuration/PyReleaseValidation/python/upgradeWorkflowComponents.py

@@ -470,7 +470,7 @@ def setup_(self, step, stepName, stepDict, k, properties):
         if 'Reco' in step:
             stepDict[stepName][k] = merge([self.step3, stepDict[step][k]])
     def condition(self, fragment, stepList, key, hasHarvest):
-        return fragment=="TTbar_14TeV" and '2021' in key
+        return (fragment=="TTbar_14TeV" or fragment=="QCD_FlatPt_15_3000HS_14") and '2021PU' in key
 
 upgradeWFs['mlpf'] = UpgradeWorkflow_mlpf(
     steps = [

RecoParticleFlow/Configuration/python/RecoParticleFlow_EventContent_cff.py

@@ -81,8 +81,3 @@
                                                             'keep recoPFRecHits_particleFlowRecHitHGC__*',
                                                             'keep *_simPFProducer_*_*'])
 
-from Configuration.ProcessModifiers.mlpf_cff import mlpf
-from RecoParticleFlow.PFProducer.mlpf_EventContent_cff import MLPF_RECO
-
-mlpf.toModify(RecoParticleFlowRECO,
-    outputCommands = RecoParticleFlowRECO.outputCommands + MLPF_RECO.outputCommands)

RecoParticleFlow/Configuration/python/RecoParticleFlow_cff.py

@@ -102,14 +102,11 @@
     e.toModify(pfNoPileUp, enable = False)
     e.toModify(pfPileUp, enable = False)
 
-#
-# for MLPF
+
+# for MLPF, replace standard PFAlgo with the ONNX-based MLPF producer 
 from Configuration.ProcessModifiers.mlpf_cff import mlpf
 from RecoParticleFlow.PFProducer.mlpfProducer_cfi import mlpfProducer
-
-_mlpfTask = cms.Task(mlpfProducer, particleFlowRecoTask.copy())
-
-mlpf.toReplaceWith(particleFlowRecoTask, _mlpfTask)
+mlpf.toReplaceWith(particleFlowTmp, mlpfProducer)
 
 #
 # switch from pfTICL to simPF

RecoParticleFlow/PFProducer/BuildFile.xml

@@ -16,6 +16,7 @@
 <use name="boost"/>
 <use name="clhep"/>
 <use name="rootmath"/>
+<use name="onnxruntime"/>
 <export>
   <lib name="1"/>
 </export>

RecoParticleFlow/PFProducer/interface/MLPFModel.h

@@ -7,31 +7,40 @@
 
 namespace reco::mlpf {
   //The model takes the following number of features for each input PFElement
-  static constexpr unsigned int NUM_ELEMENT_FEATURES = 15;
+  static constexpr unsigned int NUM_ELEMENT_FEATURES = 25;
+  static constexpr unsigned int NUM_OUTPUT_FEATURES = 14;
 
   //these are defined at model creation time and set the random LSH codebook size
-  static constexpr int NUM_MAX_ELEMENTS_BATCH = 20000;
-  static constexpr int LSH_BIN_SIZE = 100;
+  static constexpr int LSH_BIN_SIZE = 64;
+  static constexpr int NUM_MAX_ELEMENTS_BATCH = 200 * LSH_BIN_SIZE;
 
-  //In CPU mode, we only want to evaluate each event separately
+  //In CPU mode, we want to evaluate each event separately
   static constexpr int BATCH_SIZE = 1;
 
-  //The model has 12 outputs for each particle:
-  // out[0-7]: particle classification logits
-  // out[8]: regressed eta
-  // out[9]: regressed phi
-  // out[10]: regressed energy
-  // out[11]: regressed charge logit
-  static constexpr unsigned int NUM_OUTPUTS = 12;
-  static constexpr unsigned int NUM_CLASS = 7;
-  static constexpr unsigned int IDX_ETA = 8;
-  static constexpr unsigned int IDX_PHI = 9;
-  static constexpr unsigned int IDX_ENERGY = 10;
-  static constexpr unsigned int IDX_CHARGE = 11;
+  //The model has 14 outputs for each particle:
+  // out[0-7]: particle classification logits for each pdgId
+  // out[8]: regressed charge
+  // out[9]: regressed pt
+  // out[10]: regressed eta
+  // out[11]: regressed sin phi
+  // out[12]: regressed cos phi
+  // out[13]: regressed energy
+  static constexpr unsigned int IDX_CLASS = 7;
+
+  static constexpr unsigned int IDX_CHARGE = 8;
+
+  static constexpr unsigned int IDX_PT = 9;
+  static constexpr unsigned int IDX_ETA = 10;
+  static constexpr unsigned int IDX_SIN_PHI = 11;
+  static constexpr unsigned int IDX_COS_PHI = 12;
+  static constexpr unsigned int IDX_ENERGY = 13;
+
+  //for consistency with the baseline PFAlgo
+  static constexpr float PI_MASS = 0.13957;
 
   //index [0, N_pdgids) -> PDGID
   //this maps the absolute values of the predicted PDGIDs to an array of ascending indices
-  static const std::vector<int> pdgid_encoding = {0, 1, 2, 11, 13, 22, 130, 211};
+  static const std::vector<int> pdgid_encoding = {0, 211, 130, 1, 2, 22, 11, 13};
 
   //PFElement::type -> index [0, N_types)
   //this maps the type of the PFElement to an ascending index that is used by the model to distinguish between different elements
@@ -55,7 +64,13 @@ namespace reco::mlpf {
 
   int argMax(std::vector<float> const& vec);
 
-  reco::PFCandidate makeCandidate(int pred_pid, int pred_charge, float pred_e, float pred_eta, float pred_phi);
+  reco::PFCandidate makeCandidate(int pred_pid,
+                                  int pred_charge,
+                                  float pred_pt,
+                                  float pred_eta,
+                                  float pred_sin_phi,
+                                  float pred_cos_phi,
+                                  float pred_e);
 
   const std::vector<const reco::PFBlockElement*> getPFElements(const reco::PFBlockCollection& blocks);
 
@@ -64,4 +79,4 @@ namespace reco::mlpf {
                         size_t ielem_originator);
 };  // namespace reco::mlpf
 
-#endif
+#endif

RecoParticleFlow/PFProducer/plugins/BuildFile.xml

@@ -80,6 +80,6 @@
   <use name="RecoParticleFlow/PFProducer"/>
   <use name="RecoParticleFlow/PFTracking"/>
   <use name="RecoEcal/EgammaCoreTools"/>
-  <use name="PhysicsTools/TensorFlow"/>
+  <use name="PhysicsTools/ONNXRuntime"/>
   <flags EDM_PLUGIN="1"/>
 </library>

RecoParticleFlow/PFProducer/plugins/MLPFProducer.cc

@@ -4,135 +4,175 @@
 #include "FWCore/Framework/interface/MakerMacros.h"
 
 #include "DataFormats/ParticleFlowCandidate/interface/PFCandidate.h"
-#include "PhysicsTools/TensorFlow/interface/TensorFlow.h"
+#include "PhysicsTools/ONNXRuntime/interface/ONNXRuntime.h"
 #include "RecoParticleFlow/PFProducer/interface/MLPFModel.h"
 
-struct MLPFCache {
-  const tensorflow::GraphDef* graph_def;
-};
+#include "DataFormats/ParticleFlowReco/interface/PFBlockElementTrack.h"
+
+using namespace cms::Ort;
+
+//use this to switch on detailed print statements in MLPF
+//#define MLPF_DEBUG
 
-class MLPFProducer : public edm::stream::EDProducer<edm::GlobalCache<MLPFCache> > {
+class MLPFProducer : public edm::stream::EDProducer<edm::GlobalCache<ONNXRuntime>> {
 public:
-  explicit MLPFProducer(const edm::ParameterSet&, const MLPFCache*);
+  explicit MLPFProducer(const edm::ParameterSet&, const ONNXRuntime*);
+
   void produce(edm::Event& event, const edm::EventSetup& setup) override;
   static void fillDescriptions(edm::ConfigurationDescriptions& descriptions);
 
   // static methods for handling the global cache
-  static std::unique_ptr<MLPFCache> initializeGlobalCache(const edm::ParameterSet&);
-  static void globalEndJob(MLPFCache*);
+  static std::unique_ptr<ONNXRuntime> initializeGlobalCache(const edm::ParameterSet&);
+  static void globalEndJob(const ONNXRuntime*);
 
 private:
   const edm::EDPutTokenT<reco::PFCandidateCollection> pfCandidatesPutToken_;
   const edm::EDGetTokenT<reco::PFBlockCollection> inputTagBlocks_;
-  const std::string model_path_;
-  tensorflow::Session* session_;
 };
 
-MLPFProducer::MLPFProducer(const edm::ParameterSet& cfg, const MLPFCache* cache)
+MLPFProducer::MLPFProducer(const edm::ParameterSet& cfg, const ONNXRuntime* cache)
     : pfCandidatesPutToken_{produces<reco::PFCandidateCollection>()},
-      inputTagBlocks_(consumes<reco::PFBlockCollection>(cfg.getParameter<edm::InputTag>("src"))),
-      model_path_(cfg.getParameter<std::string>("model_path")) {
-  session_ = tensorflow::createSession(cache->graph_def);
-}
+      inputTagBlocks_(consumes<reco::PFBlockCollection>(cfg.getParameter<edm::InputTag>("src"))) {}
 
 void MLPFProducer::produce(edm::Event& event, const edm::EventSetup& setup) {
   using namespace reco::mlpf;
 
   const auto& blocks = event.get(inputTagBlocks_);
   const auto& all_elements = getPFElements(blocks);
 
-  const long long int num_elements_total = all_elements.size();
+  std::vector<const reco::PFBlockElement*> selected_elements;
+  unsigned int num_elements_total = 0;
+  for (const auto* pelem : all_elements) {
+    if (pelem->type() == reco::PFBlockElement::PS1 || pelem->type() == reco::PFBlockElement::PS2) {
+      continue;
+    }
+    num_elements_total += 1;
+    selected_elements.push_back(pelem);
+  }
+  const auto tensor_size = LSH_BIN_SIZE * std::max(2u, (num_elements_total / LSH_BIN_SIZE + 1));
 
+#ifdef MLPF_DEBUG
+  assert(num_elements_total < NUM_MAX_ELEMENTS_BATCH);
   //tensor size must be a multiple of the bin size and larger than the number of elements
-  const auto tensor_size = LSH_BIN_SIZE * (num_elements_total / LSH_BIN_SIZE + 1);
   assert(tensor_size <= NUM_MAX_ELEMENTS_BATCH);
+  assert(tensor_size % LSH_BIN_SIZE == 0);
+#endif
 
-  //Create the input tensor
-  tensorflow::TensorShape shape({BATCH_SIZE, tensor_size, NUM_ELEMENT_FEATURES});
-  tensorflow::Tensor input(tensorflow::DT_FLOAT, shape);
-  input.flat<float>().setZero();
+#ifdef MLPF_DEBUG
+  std::cout << "tensor_size=" << tensor_size << std::endl;
+#endif
 
-  //Fill the input tensor
+  //Fill the input tensor (batch, elems, features) = (1, tensor_size, NUM_ELEMENT_FEATURES)
+  std::vector<std::vector<float>> inputs(1, std::vector<float>(NUM_ELEMENT_FEATURES * tensor_size, 0.0));
   unsigned int ielem = 0;
-  for (const auto* pelem : all_elements) {
+  for (const auto* pelem : selected_elements) {
+    if (ielem > tensor_size) {
+      continue;
+    }
+
     const auto& elem = *pelem;
 
     //prepare the input array from the PFElement
     const auto& props = getElementProperties(elem);
 
     //copy features to the input array
     for (unsigned int iprop = 0; iprop < NUM_ELEMENT_FEATURES; iprop++) {
-      input.tensor<float, 3>()(0, ielem, iprop) = normalize(props[iprop]);
+      inputs[0][ielem * NUM_ELEMENT_FEATURES + iprop] = normalize(props[iprop]);
     }
     ielem += 1;
   }
 
-  //TF model input and output tensor names
-  const tensorflow::NamedTensorList input_list = {{"x:0", input}};
-  const std::vector<std::string> output_names = {"Identity:0"};
-
-  //Prepare the output tensor
-  std::vector<tensorflow::Tensor> outputs;
-
   //run the GNN inference, given the inputs and the output.
-  //Note that the GNN enables information transfer between the input PFElements,
-  //such that the output ML-PFCandidates are in general combinations of the input PFElements, in the form of
-  //y_out = Adj.x_in, where x_in is input matrix (num_elem, NUM_ELEMENT_FEATURES), y_out is the output matrix (num_elem, NUM_OUTPUT_FEATURES)
-  //and Adj is an adjacency matrix between the elements that is constructed on the fly during model inference.
-  tensorflow::run(session_, input_list, output_names, &outputs);
-
-  //process the output tensor to ML-PFCandidates.
-  //The output can contain up to num_elem particles, with predicted PDGID=0 corresponding to no particles predicted.
-  const auto out_arr = outputs[0].tensor<float, 3>();
+  const auto& outputs = globalCache()->run({"x:0"}, inputs, {{1, tensor_size, NUM_ELEMENT_FEATURES}});
+  const auto& output = outputs[0];
+#ifdef MLPF_DEBUG
+  assert(output.size() == tensor_size * NUM_OUTPUT_FEATURES);
+#endif
 
   std::vector<reco::PFCandidate> pOutputCandidateCollection;
-  for (unsigned int ielem = 0; ielem < all_elements.size(); ielem++) {
-    //get the coefficients in the output corresponding to the class probabilities (raw logits)
-    std::vector<float> pred_id_logits;
-    for (unsigned int idx_id = 0; idx_id <= NUM_CLASS; idx_id++) {
-      pred_id_logits.push_back(out_arr(0, ielem, idx_id));
+  for (size_t ielem = 0; ielem < num_elements_total; ielem++) {
+    std::vector<float> pred_id_probas(IDX_CLASS + 1, 0.0);
+    const reco::PFBlockElement* elem = selected_elements[ielem];
+
+    for (unsigned int idx_id = 0; idx_id <= IDX_CLASS; idx_id++) {
+      auto pred_proba = output[ielem * NUM_OUTPUT_FEATURES + idx_id];
+#ifdef MLPF_DEBUG
+      assert(!std::isnan(pred_proba));
+#endif
+      pred_id_probas[idx_id] = pred_proba;
     }
 
+    auto imax = argMax(pred_id_probas);
+
     //get the most probable class PDGID
-    int pred_pid = pdgid_encoding[argMax(pred_id_logits)];
+    int pred_pid = pdgid_encoding[imax];
 
-    //get the predicted momentum components
-    float pred_eta = out_arr(0, ielem, IDX_ETA);
-    float pred_phi = out_arr(0, ielem, IDX_PHI);
-    float pred_charge = out_arr(0, ielem, IDX_CHARGE);
-    float pred_e = out_arr(0, ielem, IDX_ENERGY);
+#ifdef MLPF_DEBUG
+    std::cout << "ielem=" << ielem << " inputs:";
+    for (unsigned int iprop = 0; iprop < NUM_ELEMENT_FEATURES; iprop++) {
+      std::cout << iprop << "=" << inputs[0][ielem * NUM_ELEMENT_FEATURES + iprop] << " ";
+    }
+    std::cout << std::endl;
+    std::cout << "ielem=" << ielem << " pred: pid=" << pred_pid << std::endl;
+#endif
 
     //a particle was predicted for this PFElement, otherwise it was a spectator
     if (pred_pid != 0) {
-      auto cand = makeCandidate(pred_pid, pred_charge, pred_e, pred_eta, pred_phi);
-      setCandidateRefs(cand, all_elements, ielem);
+      //muons and charged hadrons should only come from tracks, otherwise we won't have track references to pass downstream
+      if (((pred_pid == 13) || (pred_pid == 211)) && elem->type() != reco::PFBlockElement::TRACK) {
+        pred_pid = 130;
+      }
+
+      if (elem->type() == reco::PFBlockElement::TRACK) {
+        const auto* eltTrack = dynamic_cast<const reco::PFBlockElementTrack*>(elem);
+
+        //a track with no muon ref should not produce a muon candidate, instead we interpret it as a charged hadron
+        if (pred_pid == 13 && eltTrack->muonRef().isNull()) {
+          pred_pid = 211;
+        }
+
+        //tracks from displaced vertices need reference debugging downstream as well, so we just treat them as neutrals for the moment
+        if ((pred_pid == 211) && (eltTrack->isLinkedToDisplacedVertex())) {
+          pred_pid = 130;
+        }
+      }
+
+      //get the predicted momentum components
+      float pred_pt = output[ielem * NUM_OUTPUT_FEATURES + IDX_PT];
+      float pred_eta = output[ielem * NUM_OUTPUT_FEATURES + IDX_ETA];
+      float pred_sin_phi = output[ielem * NUM_OUTPUT_FEATURES + IDX_SIN_PHI];
+      float pred_cos_phi = output[ielem * NUM_OUTPUT_FEATURES + IDX_COS_PHI];
+      float pred_e = output[ielem * NUM_OUTPUT_FEATURES + IDX_ENERGY];
+      float pred_charge = output[ielem * NUM_OUTPUT_FEATURES + IDX_CHARGE];
+
+      auto cand = makeCandidate(pred_pid, pred_charge, pred_pt, pred_eta, pred_sin_phi, pred_cos_phi, pred_e);
+      setCandidateRefs(cand, selected_elements, ielem);
       pOutputCandidateCollection.push_back(cand);
+
+#ifdef MLPF_DEBUG
+      std::cout << "ielem=" << ielem << " cand: pid=" << cand.pdgId() << " E=" << cand.energy() << " pt=" << cand.pt()
+                << " eta=" << cand.eta() << " phi=" << cand.phi() << " charge=" << cand.charge() << std::endl;
+#endif
     }
   }  //loop over PFElements
 
   event.emplace(pfCandidatesPutToken_, pOutputCandidateCollection);
 }
 
-std::unique_ptr<MLPFCache> MLPFProducer::initializeGlobalCache(const edm::ParameterSet& params) {
-  // this method is supposed to create, initialize and return a MLPFCache instance
-  std::unique_ptr<MLPFCache> cache = std::make_unique<MLPFCache>();
-
-  //load the frozen TF graph of the GNN model
-  std::string path = params.getParameter<std::string>("model_path");
-  auto fullPath = edm::FileInPath(path).fullPath();
-  LogDebug("MLPFProducer") << "Initializing MLPF model from " << fullPath;
-
-  cache->graph_def = tensorflow::loadGraphDef(fullPath);
-
-  return cache;
+std::unique_ptr<ONNXRuntime> MLPFProducer::initializeGlobalCache(const edm::ParameterSet& params) {
+  return std::make_unique<ONNXRuntime>(params.getParameter<edm::FileInPath>("model_path").fullPath());
 }
 
-void MLPFProducer::globalEndJob(MLPFCache* cache) { delete cache->graph_def; }
+void MLPFProducer::globalEndJob(const ONNXRuntime* cache) {}
 
 void MLPFProducer::fillDescriptions(edm::ConfigurationDescriptions& descriptions) {
   edm::ParameterSetDescription desc;
   desc.add<edm::InputTag>("src", edm::InputTag("particleFlowBlock"));
-  desc.add<std::string>("model_path", "RecoParticleFlow/PFProducer/data/mlpf/mlpf_2020_11_04.pb");
+  desc.add<edm::FileInPath>(
+      "model_path",
+      edm::FileInPath(
+          "RecoParticleFlow/PFProducer/data/mlpf/"
+          "mlpf_2021_11_16__no_einsum__all_data_cms-best-of-asha-scikit_20211026_042043_178263.workergpu010.onnx"));
   descriptions.addWithDefaultLabel(desc);
 }