1 lightning.finite_horizon.selection

lightning.finite_horizon.selection

Closed QH selection policy, trainer normalization and saved-epoch ranking.

This owner admits current trainer inputs and authenticates saved checkpoint scores. Experiment composition owns admission order and publication; historical receipt readers validate their stored schema directly rather than replaying these controls.

1.1 Classes

Name Description
QhCheckpointSelectionSpec Closed validation-checkpoint selection rule.

1.1.1 QhCheckpointSelectionSpec

lightning.finite_horizon.selection.QhCheckpointSelectionSpec(
    kind='validation_min_v1',
    monitor='val/loss',
    mode='min',
    tie_break='earliest_optimizer_update',
)

Closed validation-checkpoint selection rule.

1.1.1.1 Attributes

Name Description
kind Final-update selection fixes the budget; diagnostic H1 regret selects an epoch by action cost.
monitor Val/loss normally; diagnostic H1 regret uses the separately admitted scene-equal state mean.
mode The explicitly admitted validation monitor is minimized.
tie_break An exact metric tie retains the earlier optimizer update.

1.2 Functions

Name Description
checkpoint_current_metric Read this saved epoch’s score from its matching ModelCheckpoint state.
normalize_trainer Bind run paths and closed selection callbacks to admitted current inputs.
select_checkpoint Rank by admitted loss or diagnostic regret; return the selected calibration loss.

1.2.1 checkpoint_current_metric

lightning.finite_horizon.selection.checkpoint_current_metric(payload, monitor)

Read this saved epoch’s score from its matching ModelCheckpoint state.

best_model_score may refer to a different optimizer update. Only a unique ModelCheckpoint state with the exact monitor and finite scalar current_score establishes the saved checkpoint’s regret rank.

1.2.2 normalize_trainer

lightning.finite_horizon.selection.normalize_trainer(
    trainer,
    output,
    selection,
    *,
    diagnostic=False,
)

Bind run paths and closed selection callbacks to admitted current inputs.

Validation-selected policies restore explicitly disabled validation limits; final-update diagnostics retain training-only admission and reject debug overrides. Gradient clipping remains with the QH optimizer lifecycle.

1.2.3 select_checkpoint

lightning.finite_horizon.selection.select_checkpoint(
    output,
    *,
    selection=None,
    expected_final_update=None,
    expected_validation_epochs=None,
)

Rank by admitted loss or diagnostic regret; return the selected calibration loss.

Both policies break exact metric ties by earliest optimizer update. The returned loss always comes from that checkpoint’s validation loss buffers, even when a distinct callback current_score chooses the epoch.

Back to top