Alert track matching ai

[mathieuouillon]

Changes in the pull request:

• PreclusterSuperlayer renamed Intercluster; they are now linked to a track with the trackId.
• Model renamed to ModelTrackFinding for better clarity. It can now predict all tracks at once, improving speed.
• Use only on preclustering in PreClusterFinder instead of 2 the conventional one and the AI one.
• Fix a bug in the preclustering when more than 3 wires are one after the other.
• The Track class now has both the clusters and interclusters that form the track. Also, set the trackId to the clusters and interclusters.
• Rename Mode to ModeTrackFinding for clarity
• Create ModelTrackMatching class to predict which sector, layer, and wedge should be hit.
• AHDCEngine:
  • Add some constants for clarity (TRACK_FINDING_AI_THRESHOLD and MAX_HITS_FOR_AI)
  • Answer a few comments
  • Use the logger instead of print
  • The logic for the engine is: Read the hits -> create preclusters -> track finding, 2 possibilities:
    • AI-based Track Finding: Merge preclusters into interclusters -> Generate all track candidates -> Evaluate each candidate with AI -> Select final tracks -> result: a list of tracks, each with a list of InterClusters
    • Conventional Track Finding (Distance or Hough): Build clusters -> Track finding -> result: list of tracks containing clusters.
  • The track ID for clusters and interclusters is done in the Kalman Filter. If AI is used for track finding, generate clusters; if conventional methods generate interclusters.
  • The track ID of the interclusters is used in the Track Matching algorithm in the ALERT Engine.
• ALERTEngine: Extract intercluster patterns for each track -> AI model predicts which ATOF wedge the track will hit -> Compare predicted hit to real ATOF hits -> pick the hit closest to the AI prediction (in terms of xyz distance) -> write the ATOF hit id to the new ALERT::ai:prediction bank.

[mathieuouillon]

Work with the pull request https://code.jlab.org/hallb/clas12/coatjava/networks/-/merge_requests/5

[mathieuouillon]

@baltzell Hi, I think the unittest fails because it uses cvmfs to get the networks, and it is the wrong commit of the networks repo.

[c-dilks]

unittest fails because it uses cvmfs to get the networks

Makes sense, assuming cvmfs is behind. I pushed a commit to switch the relevant CI jobs to use LFS.

[mathieuouillon]

Thanks Chris

[baltzell]

Mathieu, can you describe conceptually what these changes amount to?

[mathieuouillon]

Sure, here is a summary. Let me know if you’d like more details:

Changes in the pull request:

• PreclusterSuperlayer renamed Intercluster; they are now linked to a track with the trackId.
• Model renamed to ModelTrackFinding for better clarity. It can now predict all tracks at once, improving speed.
• Use only on preclustering in PreClusterFinder instead of 2 the conventional one and the AI one.
• Fix a bug in the preclustering when more than 3 wires are one after the other.
• The Track class now has both the clusters and interclusters that form the track. Also, set the trackId to the clusters and interclusters.
• Rename Mode to ModeTrackFinding for clarity
• Create ModelTrackMatching class to predict which sector, layer, and wedge should be hit.
• AHDCEngine:
  • Add some constants for clarity (TRACK_FINDING_AI_THRESHOLD and MAX_HITS_FOR_AI)
  • Answer a few comments
  • Use the logger instead of print
  • The logic for the engine is: Read the hits -> create preclusters -> track finding, 2 possibilities:
    • AI-based Track Finding: Merge preclusters into interclusters -> Generate all track candidates -> Evaluate each candidate with AI -> Select final tracks -> result: a list of tracks, each with a list of InterClusters
    • Conventional Track Finding (Distance or Hough): Build clusters -> Track finding -> result: list of tracks containing clusters.
  • The track ID for clusters and interclusters is done in the Kalman Filter. If AI is used for track finding, generate clusters; if conventional methods generate interclusters.
  • The track ID of the interclusters is used in the Track Matching algorithm in the ALERT Engine.
• ALERTEngine: Extract intercluster patterns for each track -> AI model predicts which ATOF wedge the track will hit -> Compare predicted hit to real ATOF hits -> pick the hit closest to the AI prediction (in terms of xyz distance) -> write the ATOF hit id to the new ALERT::ai:prediction bank.

[c-dilks]

Sure, here is a summary. Let me know if you’d like more details:

sorry, could you put this in the PR description up top, please?

[mathieuouillon]

No problem, it's now in the first comment.