src/turn_detection/training/__init__.py

"""Training objectives, metrics, and deterministic trainer."""

# Keep metric imports lightweight: importing ``turn_detection.training.metrics``
# must not require torch.  Heavy objects are exposed lazily.
from .metrics import (
    binary_classification_metrics,
    confusion_counts,
    metrics_at_fpr_budgets,
    operational_metrics,
    sliced_metrics,
    threshold_at_max_fpr,
)

__all__ = [
    "DistillationLoss",
    "DistillationLossConfig",
    "MultiTaskLossConfig",
    "MultiTaskTurnLoss",
    "Trainer",
    "TrainerConfig",
    "binary_classification_metrics",
    "confusion_counts",
    "metrics_at_fpr_budgets",
    "operational_metrics",
    "sliced_metrics",
    "threshold_at_max_fpr",
]


def __getattr__(name: str):
    if name in {
        "DistillationLoss",
        "DistillationLossConfig",
        "MultiTaskLossConfig",
        "MultiTaskTurnLoss",
    }:
        from . import losses

        return getattr(losses, name)
    if name in {"Trainer", "TrainerConfig"}:
        from . import trainer

        return getattr(trainer, name)
    raise AttributeError(name)