"""Learning-rate scheduler factory.""" from __future__ import annotations from typing import Any, Dict, Optional import torch def build_scheduler( cfg: Dict[str, Any], optimizer: torch.optim.Optimizer, steps_per_epoch: int, max_epochs: int, ) -> Optional[torch.optim.lr_scheduler._LRScheduler]: name = (cfg.get("name") or "none").lower() params = dict(cfg.get("params", {})) if name in {"none", "constant", "null"}: return None if name == "cosine": t_max = int(params.pop("t_max", max_epochs)) return torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=max(1, t_max), **params) if name == "step": step_size = int(params.pop("step_size", 30)) gamma = float(params.pop("gamma", 0.1)) return torch.optim.lr_scheduler.StepLR(optimizer, step_size=max(1, step_size), gamma=gamma, **params) if name == "exponential": gamma = float(params.pop("gamma", 0.99)) return torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=gamma, **params) if name == "plateau": mode = params.pop("mode", "min") factor = float(params.pop("factor", 0.5)) patience = int(params.pop("patience", 10)) return torch.optim.lr_scheduler.ReduceLROnPlateau( optimizer, mode=mode, factor=factor, patience=patience, **params ) if name == "onecycle": max_lr = float(params.pop("max_lr", optimizer.param_groups[0]["lr"])) total_steps = int(params.pop("total_steps", max_epochs * max(1, steps_per_epoch))) return torch.optim.lr_scheduler.OneCycleLR( optimizer, max_lr=max_lr, total_steps=max(1, total_steps), **params ) raise ValueError(f"Unknown scheduler: {name}") def is_plateau(scheduler) -> bool: return scheduler is not None and isinstance(scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau)