Skip to content

Add tagger inference timings to job metadata

Add tagger timing stats to userJobMetadata.json, closes #90 (closed). Some results over 1000 events (inference time is per jet):

ttbar
BTagging/20220314/dipsLoose/antikt4empflow/network.json 0.103 +/- 0.006 ms
BTagging/20220509/dl1dLoose/antikt4empflow/network.json 0.082 +/- 0.005 ms
BTagging/20220509/gn1/antikt4empflow/network.onnx       0.441 +/- 0.004 ms
BTagging/20220509/gn1lep/antikt4empflow/network.onnx    0.431 +/- 0.004 ms

zprime
BTagging/20220314/dipsLoose/antikt4empflow/network.json 0.134 +/- 0.009 ms
BTagging/20220509/dl1dLoose/antikt4empflow/network.json 0.089 +/- 0.005 ms
BTagging/20220509/gn1/antikt4empflow/network.onnx       0.881 +/- 0.015 ms
BTagging/20220509/gn1lep/antikt4empflow/network.onnx    0.875 +/- 0.015 ms

Note that timings don't take into account:

  • Improved rejection at the same efficiency will lead to downstream cpu savings for more performant taggers
  • DIPS, IPxD, SV1, JF not factored into DL1d timings
Edited by Samuel Van Stroud

Merge request reports