Skip to content

Online analysis: calibration fixes

Patrick Godwin requested to merge patrick.godwin/idq:trim_dataloader_online into main
  • Update model IDs when attaching a model to a classifier. This allows the calibration map to correctly grab the model ID and store it in its internal metadata
  • Attach model for initial calibration in calibration jobs
  • Only add clean segments if there is a valid dataset to be ingested
  • Remove extraneous setting of model_id to None, which causes metadata issues before a new model is found in calibration jobs
  • Decouple calibration and evaluation jobs. This allows one to tune the delays so that we can read datasets in the same stride for the calibration jobs
  • Properly propagate the bounded bandwidths for calibration jobs when a map is updated in streaming jobs

Other non-calibration changes:

  • Trim the training data after training which potentially reduces the memory footprint of training jobs

Merge request reports