espnet3.components.modeling.optimization_spec.SchedulerSpec

source

class espnet3.components.modeling.optimization_spec.SchedulerSpec(name: str, scheduler: Any, interval: str = 'epoch', monitor: str | None = None)

Bases: object

Describe one named scheduler block from config.schedulers.

This is the normalized, validated form of user-facing scheduler metadata. Scheduler config example.

schedulers:
  generator:
    scheduler:
      _target_: torch.optim.lr_scheduler.LinearLR
      total_iters: 1000
    interval: step

Another example.

schedulers:
  discriminator:
    scheduler:
      _target_: torch.optim.lr_scheduler.ReduceLROnPlateau
      patience: 2
    interval: epoch
    monitor: valid/discriminator/loss

Like OptimizerSpec, this stores configuration, not the instantiated scheduler object. The Lightning module uses it to decide whether a scheduler should step on each optimizer update or at epoch end, and which logged metric to pass when a monitored epoch-based scheduler is used.

classmethod from_config(name: str, cfg) → SchedulerSpec

Build and validate a scheduler spec from one named config block.

This converts raw Hydra/OmegaConf config into a typed SchedulerSpec instance, applies dataclass defaults for omitted optional fields, and validates interval-level rules before the Lightning module instantiates the real scheduler objects.

interval : str = 'epoch'

monitor : str | None = None

name : str

scheduler : Any

validate() → None

Validate scheduler stepping policy after normalization.

Only spec-local checks are performed here. Cross-object rules, such as matching optimizer and scheduler names exactly, are validated later by the Lightning module once all specs have been collected.