aiaccel.torch.lightning.LRSchedulerConfig

class aiaccel.torch.lightning.LRSchedulerConfig(scheduler_generator: Callable[[...], LRScheduler], interval: str = 'step', frequency: int = 1, monitor: str | None = 'validation/loss', strict: bool = True, name: str | None = None)[source]

Configuration for a learning rate scheduler in Lightning.

Parameters:

  • scheduler_generator (Callable[..., optim.lr_scheduler.LRScheduler]) – A callable that generates the scheduler.

  • interval (str) – Timing to call scheduler.step ("step" or "epoch"). Defaults to "step".

  • frequency (int) – How often to call the scheduler. Defaults to 1.

  • monitor (str | None) – Metric to monitor (required for ReduceLROnPlateau). Defaults to "validation/loss".

  • strict (bool | None) – Whether to raise if monitor is missing. Mirrors Lightning’s strict flag.

  • name (str | None) – Optional name for logging. Defaults to None.

__init__(scheduler_generator: Callable[[...], LRScheduler], interval: str = 'step', frequency: int = 1, monitor: str | None = 'validation/loss', strict: bool = True, name: str | None = None) → None

Methods

__init__(scheduler_generator[, interval, ...])

build(optimizer)

Attributes

frequency

interval

monitor

name

strict

scheduler_generator