1 lightning.training.optimizers

lightning.training.optimizers

Shared optimizer and learning-rate scheduler configuration for Lightning.

AdamWConfig serves both the one-step aria_nbv.lightning.VinLightningModule and finite-horizon aria_nbv.lightning.finite_horizon.module.QhLightningModule. Scheduler factories provide the common Lightning mapping and step-budget contracts without owning either objective. Each consuming module retains optimizer selection, stepping order, and scheduler-admission policy; the Trainer owns the final one-cycle step budget.

1.1 Classes

Name	Description
AdamWConfig	Configure AdamW for a Lightning module’s trainable parameters.
OneCycleSchedulerConfig	Configure a step-wise one-cycle learning-rate and momentum schedule.
ReduceLrOnPlateauConfig	Configure metric-driven learning-rate decay for Lightning.

1.1.1 AdamWConfig

lightning.training.optimizers.AdamWConfig()

Configure AdamW for a Lightning module’s trainable parameters.

1.1.1.1 Attributes

Name	Description
target_type	Factory target for aria_nbv.configs.base.BaseConfig.setup_target.
learning_rate	Learning rate for AdamW.
weight_decay	Weight decay for AdamW.
foreach	Explicit foreach policy; None preserves PyTorch’s historical backend selection.
fused	Explicit fused policy; None preserves PyTorch’s historical backend selection.

1.1.1.2 Methods

Name	Description
setup_target	Create AdamW over the exact trainable parameter list.

1.1.1.2.1 setup_target

lightning.training.optimizers.AdamWConfig.setup_target(params)

Create AdamW over the exact trainable parameter list.

1.1.1.2.1.1 Parameters

Name	Type	Description	Default
params	list[Tensor]	Materialized parameters selected by the consuming aria_nbv.lightning.VinLightningModule.configure_optimizers or aria_nbv.lightning.finite_horizon.module.QhLightningModule.configure_optimizers.	required

1.1.1.2.1.2 Returns

Name	Type	Description
	Optimizer	Optimizer owned and stepped by Lightning.

1.1.2 OneCycleSchedulerConfig

lightning.training.optimizers.OneCycleSchedulerConfig()

Configure a step-wise one-cycle learning-rate and momentum schedule.

1.1.2.1 Attributes

Name	Description
target_type	Return the external one-cycle scheduler factory target.
max_lr	Maximum learning rate in the cycle (defaults to optimizer LR).
base_momentum	Lower momentum boundary in the cycle.
max_momentum	Upper momentum boundary in the cycle.
div_factor	Initial learning rate = max_lr / div_factor.
final_div_factor	Final learning rate = max_lr / (div_factor * final_div_factor).
pct_start	Percentage of cycle spent increasing learning rate.
anneal_strategy	Annealing strategy: ‘cos’ or ‘linear’.

1.1.2.2 Methods

Name	Description
setup_target	Create a one-cycle scheduler after resolving the total step budget.
setup_lightning	Build the Lightning lr_scheduler config for OneCycleLR.

1.1.2.2.1 setup_target

lightning.training.optimizers.OneCycleSchedulerConfig.setup_target(
    optimizer,
    *,
    total_steps=None,
    trainer=None,
)

Create a one-cycle scheduler after resolving the total step budget.

1.1.2.2.1.1 Parameters

Name	Type	Description	Default
optimizer	Optimizer	Optimizer whose learning rate and momentum are scheduled.	required
total_steps	int | None	Explicit positive optimizer-step count. When omitted, trainer.estimated_stepping_batches or its datamodule/epoch fallback is used.	None
trainer	pl.Trainer | None	Optional configured trainer used for step-budget inference.	None

1.1.2.2.1.2 Returns

Name	Type	Description
	OneCycleLR	Step-wise scheduler owned by Lightning.

1.1.2.2.2 setup_lightning

lightning.training.optimizers.OneCycleSchedulerConfig.setup_lightning(
    optimizer,
    *,
    total_steps=None,
    trainer=None,
)

Build the Lightning lr_scheduler config for OneCycleLR.

1.1.2.2.2.1 Parameters

Name	Type	Description	Default
optimizer	Optimizer	Optimizer instance to schedule.	required
total_steps	int | None	Optional total step count for the cycle.	None
trainer	pl.Trainer | None	Optional Lightning trainer used to infer total_steps.	None

1.1.2.2.2.2 Returns

Name	Type	Description
	dict[str, Any]	Lightning lr_scheduler configuration dictionary.

1.1.3 ReduceLrOnPlateauConfig

lightning.training.optimizers.ReduceLrOnPlateauConfig()

Configure metric-driven learning-rate decay for Lightning.

1.1.3.1 Attributes

Name	Description
target_type	Factory target for aria_nbv.configs.base.BaseConfig.setup_target.
mode	Whether to reduce on metric min or max.
factor	Multiplicative factor of LR reduction.
patience	Number of steps with no improvement before reducing the LR.
threshold	Threshold for measuring new optimum.
threshold_mode	Threshold interpretation (relative or absolute).
cooldown	Number of epochs to wait before resuming normal operation.
min_lr	Lower bound on the learning rate.
eps	Minimal decay applied to LR.
monitor	Metric name to monitor for plateau reduction.
interval	Scheduler interval (step or epoch).
frequency	Scheduler frequency.

1.1.3.2 Methods

Name	Description
setup_target	Create a plateau scheduler over one optimizer.
setup_lightning	Build the Lightning lr_scheduler config for ReduceLROnPlateau.

1.1.3.2.1 setup_target

lightning.training.optimizers.ReduceLrOnPlateauConfig.setup_target(
    optimizer,
    *,
    trainer=None,
)

Create a plateau scheduler over one optimizer.

1.1.3.2.1.1 Parameters

Name	Type	Description	Default
optimizer	Optimizer	Optimizer whose parameter-group rates are mutated.	required
trainer	pl.Trainer | None	Accepted for the shared scheduler factory interface; not used because plateau scheduling is metric driven.	None

1.1.3.2.1.2 Returns

Name	Type	Description
	ReduceLROnPlateau	Scheduler instance later stepped by Lightning.

1.1.3.2.2 setup_lightning

lightning.training.optimizers.ReduceLrOnPlateauConfig.setup_lightning(
    optimizer,
    *,
    trainer=None,
)

Build the Lightning lr_scheduler config for ReduceLROnPlateau.

1.1.3.2.2.1 Parameters

Name	Type	Description	Default
optimizer	Optimizer	Optimizer instance to schedule.	required
trainer	pl.Trainer | None	Optional Lightning trainer (unused for plateau).	None

1.1.3.2.2.2 Returns

Name	Type	Description
	dict[str, Any]	Lightning lr_scheduler configuration dictionary.