Custom training fitness

A fitness callback selects best checkpoints and controls patience with a custom scalar score.

Define the scorer

Python

from libreyolo import LibreYOLO class RecallFitness:    def fitness(self, metrics):        return float(metrics["metrics/recall(B)"]) model = LibreYOLO("LibreYOLO9s.pt", device="cpu")model.train(data="coco8.yaml", epochs=1, workers=0, device="cpu", callbacks=RecallFitness())

Implement fitness(metrics) on an object passed through callbacks=. The input is the read-only scalar mapping used by TrainEpochEvent.val_metrics. Return one finite real number; higher is better. A callback list may contain only one scorer. VLM and VLA fine-tuning do not accept a fitness callback.

Checkpoint selection

Scoring runs on rank zero for each validation result used for selection, before checkpoint writing and epoch callbacks. It changes best-checkpoint selection and patience without replacing the original metrics.

Checkpoints record fitness_source="callback". Callback code and state are not serialized, so custom-fitness training cannot resume. Start a new run from the saved weights instead.

See validation for the metric keys available to a scorer.