• Reconstruction example

• C++ docs
• CMSSW interface
• PhysicsTools/TensorFlow

tf_models/energy_id_v*.pb: TensorFlow model for trackster energy regression and particle ID.

• 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 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.

tf_models/energy_regression_without_pattern_recognition.pb: TensorFlow model for trackster energy regression

• DEPRECATED
• This model tries to learn correction coefficients for different parts of the detector (e.g. CEE-120µm, CEH-Fine-300 µm, CEH-Coarse-Scintillators, etc.)
• Each coefficient is implemented as a very small dense neural network with the energy sum of its category and the position (η) as input.
• This network has been trained on layerClusters without pattern recognition. It is therefore expected that it wil perform poorly (response < 1) in any scenario after any pattern recognition has been applied. Better solutions are work in progress, so consider this as a temporary place holder to implement the functionality.

tf_models/energy_regression_after_pattern_recognition.pb: TensorFlow model for trackster energy regression

• for functionality: see above
• this model is trained on hadronic single particle events (Klong)
• Input: sum of all energies of different detector parts in an event after pattern recognition
• Ouput: Regressed energy assuming all energy deposits originate from a single particle
• IMPORTANT: This is not trained on individual tracksters (Work in progress) and since it is by design not linear in the energy input it will give different (i.e. worse) results if a particle is split into multiple tracksters.

energy_regression_tracksters.pb: Tensorflow model trained on tracksters

• Big difference: This model is trained to predict the energy correctly on trackster level, not on event level
• For inference there is no fundamental change
• The expected input variable are now $f_0, ..., f9, \eta, E_{raw}, p_0, ..., p1$ where $f_i = E_i / E_{raw} and $p_i$ are the ID probabilities