New ONNX models for PID and Energy Regression

README.md

@@ -8,14 +8,27 @@
   - [C++ docs](https://www.tensorflow.org/api_docs/cc)
   - [CMSSW interface](https://gitlab.cern.ch/mrieger/CMSSW-DNN)
   - [PhysicsTools/TensorFlow](https://github.com/cms-sw/cmssw/tree/master/PhysicsTools/TensorFlow)
+- ONNX:
+  - [Introduction](https://onnx.ai/onnx/intro/)
+  - [CMSSW interface](https://github.com/cms-sw/) (still under construction)
+  - [PhysicsTools/ONNXRuntime](https://github.com/cms-sw/cmssw/tree/master/PhysicsTools/ONNXRuntime)
 
 ### Models
 
-- `tf_models/energy_id_v*.pb`: TensorFlow model for trackster energy regression and particle ID.
+- `ticlv4/tf_models/energy_id_v*.pb`: The TensorFlow model for trackster energy regression and particle ID has been trained on the TICLv4 data and used within TICLv4.
   - `v0`: Simple CNN-based approach. The neutral pion, neutral hadron, ambiguous and unknown probabilities are set to a constant value of 0. See the [talk at the Reco/AT meeting](https://indico.cern.ch/event/841640/contributions/3534140/attachments/1896780/3129591/2019-08-23_rieger_hgcal_ticl_eid.pdf) for more info. Input and output tensors:
     - `"input"`: Input tensor with dimension `batch x 50 (layers) x 10 (clusters) x 3 (features)`.
     - `"output/id_probabilities"`: Output tensor with dimension `batch x 8` representing particle ID "probabilities" (from a softmax output). The probabiltities refer to photon, electron, muon, neutral pion, charged hadron, neutral hadron, ambiguous and unknown cases (in that order).
     - `"output/regressed_energy"`: Output tensor with dimension `batch x 1` representing the regressed energy value for the trackster.
  - `superclustering/`: ONNX models (from PyTorch) for superclustering of electrons. 
    - `superclustering/supercls_v2p1.onnx`: DNN, inputs features computed from pairs of tracksters (uses inputs defined in `SuperclusteringDNNInputV2` in `RecoHGCal/TICL/interface/SuperclusteringDNNInputs.h`). Input format : `batch x 17 (features)`. Outputs score (dimension `batch`) giving "probability" that the sub-leading trackster is a bremmstrahlung photon of the leading trackster. Optimal working point : 0.3.
    - `superclustering/regression_v1.onnx`: DNN for supercluster energy regression. Input format : `batch x 8 (features)`. Output : `batch x 1` (supercluster regressed energy). Used in `RecoHGCal/TICL/plugins/EGammaSuperclusterProducer.cc`.
+- `ticlv5/onnx_models/`: The models are trained based on TICLv5 reconstruction information using a simple CNN-based approach. Two models have been trained separately: one for trackster energy regression and one for particle ID. These models are saved in ONNX format for time optimization.
+- `Common input tensor`: Both models share the same initial input tensor, dimensions batch x 50 (layers) x 10 (clusters) x 3 (features).
+  - `ticlv5/patternrecognition/id_v*.onnx`:
+    - `"input"`: Input tensor with dimensions batch x 50 (layers) x 10 (clusters) x 3 (features).
+    - `"output/pid_output"`: Output tensor with dimensions batch x 8 representing particle ID probabilities (from a softmax output). The probabilities refer to: photon, electron, muon, neutral pion, charged hadron(pion), neutral hadron(kaon), ambiguous, and unknown cases (in that order). The probabilities help in classifying the particle based on its type, distinguishing between hadronic and electromagnetic categories.
+  - `ticlv5/linking/energy_v*.onnx`:
+    - `"input"`: Input tensor with dimensions batch x 50 (layers) x 10 (clusters) x 3 (features), concatenated with the output of the particle ID model ("output/pid_output").
+    - `"output/enreg_output"`: Output Tensor with dimension batch x 1 (regressed energy). This value represents the trackster energy as estimated by the model based on the training data, compared to the true and reconstructed energies of the particle.
+