Main focus

• Visualizations and QA: RootInteractive for Tracks and Clusters
• Data-readout / transformation: Get (X, Y, Z) position from (sector, row, pad, time), get all clusters assigned to tracks (after tracking), readout track properties

QA and visualizations

Clusters: Single event

Tracks: 50 Ev. @ 50kHz PbPb

Native (top), Network (bottom)

dE/dx vs tpcInnerParam:

Chi2 / NCl:

Number of clusters:

Current issues to solve

• Reco workflow with QA and MC enabled crashes since not every cluster has an ideal cluster attached to it after the assignment process -> Tried dummy label, no label, explicitly unsetting it...
• Getting tracks transformed correctly is not so easy: Have to check some other tasks how to do it (linear transformations just looked completely off...)

Next steps

• Better training data selection for network -> Create quality score for training data based on charge contribution by MC charge, sector boundary, etc. and weight training data accordingly
• Feel comfortable interfacing clsuters / tracks now -> Implementing PyTorch C++ API in O2. Will try to get a simple GPU script for ROCm working (within O2)
• Return to network training
• To be discussed: Looper fitting based on tagged clusters (could start with simple helix model)