Skip to content

General offline improvements

Patrick Godwin requested to merge general_offline_improvements into master

What's changed here:

  • Improve the provenance scheme so that offline bins using the same start/end times can be tracked uniquely, avoiding needing to have many copies of the same model/calibration map.
  • Switch fork workflow to use ProcessPoolExecutor, removing the need to do a lot of the manual multiprocessing work as was done before.
  • Remove the condor workflow in individual batch jobs. This ended up causing trouble when the batch workflow used condor but the individual jobs also used condor.
  • Use all data for calibration, which should make it so we use samples closer to timeseries production in the online case.
  • Force a bounded bandwidth when using OVL as the classifier's discreteness in rank causes the bandwidth optimization to have much smaller bandwidths than is desired.
  • Switch ordering within OVL.evaluate() to avoid recomputing veto configurations many times.
  • Tweak calibration distribution plots to view distributions better.

Merge request reports