Adding new version of SuperclusterDNN model v3

README.md

@@ -21,6 +21,7 @@
     - `"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_v3.onnx`: DNN, inputs features computed from pairs of tracksters (uses inputs defined in `SuperclusteringDNNInputV3` in `RecoHGCal/TICL/interface/SuperclusteringDNNInputs.h`). Input format : `batch x 18 (features)`. deltaTime information is added as a new feature. Outputs score (dimension `batch`) giving "probability" that the sub-leading trackster is a bremmstrahlung photon of the leading trackster. Optimal working point : 0.57247.
    - `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.