Skip to content

OBSERVER Pipelines

Role

OBSERVER pipelines are the set of training (and optionally evaluation) pipelines that share the same naming and config convention: detection, size (multiclass), and location (multiclass). They mirror the PFM pipelines but are configured and named for the OBSERVER use case.

Why a Separate Concept

  • Deployment and naming: Models and artifacts are stored under distinct paths and names (e.g. OBSERVER vs PFM) so both can coexist.
  • Config: Each has its own section in pipelines_config.yml (e.g. training_observer_detection_pipeline, training_observer_size_pipeline, training_observer_location_pipeline).
  • Threshold and metrics: OBSERVER detection may use a different threshold or metric set; the same by-case split and validation philosophy apply.

Where to Look

  • Training: Training pipelines (PFM & OBSERVER).
  • Scripts: run_training_observer_detection_pipeline.py, run_training_observer_size_pipeline.py, run_training_observer_location_pipeline.py.
  • Test offline: The test-offline pipeline can be configured to run OBSERVER models and report OBSERVER-specific metrics.

This Engineering page states the concept; the full configuration and usage are in the Pipelines section.