You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
WCML will need various types of truth. WCT has some already but some new things are needed
true signal (depo splat)
true blob charge (depo blob fill)
particle trajectory tree (eg from geant4)
something like depo-blob-fill but which sets other truths on a blob (particle, true energy, a "holds vertex" bool)
Various considerations
The particle trajectory tree can be gigantic. We will need some way to filter it down. Daya Bay has a "TruthHistorian" package that uses boost spirit to define selection rules.
The "track ID" held by each depo may not have a corresponding track id in the particle trajectory tree if filtering is performed.
The "track ID" is abused in larsoft's larwirecell in order to handle the "origTrackId" that some LS'ers rely on. With that abuse turned on, the track ID is no longer that from geant4.
When the truth is some label about the original particle there is ambiguity given that multiple particles can pass through any given blob. Most work in this area chooses the "dominant" label. Eg, if the label is "true particle ID" then they would set the ID of the particle that provided the most ionization electrons in the blob. We can certainly provide a variable-sized list of labels on each blob but consuming this as "truth" when training ML will need some mechanism to handle the variable length.
We likely need to label 2D view pixels in addition to 3D blobs.
The text was updated successfully, but these errors were encountered:
WCML will need various types of truth. WCT has some already but some new things are needed
Various considerations
The particle trajectory tree can be gigantic. We will need some way to filter it down. Daya Bay has a "TruthHistorian" package that uses boost spirit to define selection rules.
The "track ID" held by each depo may not have a corresponding track id in the particle trajectory tree if filtering is performed.
The "track ID" is abused in larsoft's larwirecell in order to handle the "origTrackId" that some LS'ers rely on. With that abuse turned on, the track ID is no longer that from geant4.
When the truth is some label about the original particle there is ambiguity given that multiple particles can pass through any given blob. Most work in this area chooses the "dominant" label. Eg, if the label is "true particle ID" then they would set the ID of the particle that provided the most ionization electrons in the blob. We can certainly provide a variable-sized list of labels on each blob but consuming this as "truth" when training ML will need some mechanism to handle the variable length.
We likely need to label 2D view pixels in addition to 3D blobs.
The text was updated successfully, but these errors were encountered: