This allows us to support the b-tagging RNN in reconstruction.
There are a number of changes required to make this work:
- We're using
FlavorTagDiscriminantsas the wrapper for
lwtnn. This gives us more flexibility with the NN and makes changes easier to validate.
- The track impact parameters are added in the xAOD fix rather than computed on the fly in the b-tagging code. Given that
FlavorTagDiscriminantsis meant to run in
AnalysisBasewe can't run the track extrapolation there.
- The RNN (and eventually DL1) algorithms are scheduled as part of the
JetBTaggerTool(which is scheduled once for each jet collection) rather than within the
BTagTool(which is shared between all the jet collections). This is to avoid having to deal with multiple trainings within the same
This is the first in a series of merge requests. It does nothing at the moment, because I've turned off all the additional taggers, but future merge requests will add these. I ran the
acmd diff-root and didn't see any differences in the output.