model.metrics

Validation metric helpers used by training and evaluation.

Validation metric helpers for BirdNET Geomodel.

Functions

compute_geoscore(metrics)

Compute the GeoScore composite quality metric.

Missing components are skipped and weights are renormalized.

Component transforms (all map to [0, 1], higher = better): - map, f1_10, watchlist_mean_ap, holdout_map: used directly (already [0, 1]) - list_ratio_10: exp(-|log(r)|) — symmetric penalty around ratio=1, equivalent to min(r, 1/r); never collapses to zero for finite ratios - map_density_ratio: min(r, 1/r) — peaks at 1.0 for perfect sparse/dense balance; penalizes both sparse-dominant and dense-dominant bias symmetrically - pred_density_corr: 1 - |r| — lower absolute correlation is better