lr_scheduler.LRScheduler

class lucid.optim.lr_scheduler.LRScheduler(optimizer: Optimizer, last_epoch: int = -1, verbose: bool = False)

The LRScheduler class provides the base functionality for all learning rate scheduling strategies in Lucid. Schedulers dynamically adjust the learning rate during training to enhance model convergence and performance.

Class Signature

class LRScheduler(optimizer, last_epoch: int = -1, verbose: bool = False)

Parameters

  • optimizer (Optimizer): The optimizer whose learning rate needs to be scheduled.

  • last_epoch (int, optional): The index of the last epoch when resuming training. Default: -1.

  • verbose (bool, optional): If True, logs learning rate updates at each step. Default: False.

Methods

  • get_lr() -> list[float]: Computes the new learning rate(s). Must be implemented by subclasses.

  • step(epoch: Optional[int] = None) -> None: Updates the learning rate. If epoch is provided, applies the schedule up to that epoch.

  • state_dict() -> dict[str, Any]: Returns the current state of the scheduler, including learning rates and epoch information.

  • load_state_dict(state_dict: dict[str, Any]) -> None: Loads a previously saved state dictionary to restore scheduler behavior.

Usage Example

import lucid.optims as optim
from lucid.optims.lr_scheduler import StepLR

optimizer = optim.SGD(model.parameters(), lr=0.1)
scheduler = StepLR(optimizer, step_size=5, gamma=0.5)

for epoch in range(20):
    optimizer.step()
    scheduler.step()
    print(f"Epoch {epoch+1}, Learning Rate: {scheduler.last_lr}")