olm.train.schedulers.linear¶
Linear learning rate scheduler.
Classes¶
LinearDecayLR(*args, **kwargs) |
Simple linear decay scheduler that decays to zero. |
|---|---|
LinearLR(*args, **kwargs) |
Linear learning rate scheduler. |
class olm.train.schedulers.linear.LinearDecayLR(*args: Any, **kwargs: Any)¶
Bases: SchedulerBase
Simple linear decay scheduler that decays to zero.
This is a simplified version that always decays to 0 from the initial LR.
- Parameters:
- optimizer – Wrapped optimizer.
- total_steps – Total number of steps to decay over.
- last_epoch – The index of last epoch (default: -1).
Example¶
>>> from olm.train.schedulers import LinearDecayLR
>>> scheduler = LinearDecayLR(optimizer, total_steps=1000)
>>> for step in range(total_steps):
... train(...)
... scheduler.step()
get_lr()¶
Compute learning rate using linear decay.
class olm.train.schedulers.linear.LinearLR(*args: Any, **kwargs: Any)¶
Bases: SchedulerBase
Linear learning rate scheduler.
Linearly decreases (or increases) the learning rate from the initial learning rate to end_lr over total_steps.
- Parameters:
- optimizer – Wrapped optimizer.
- total_steps – Total number of steps for the schedule.
- end_lr – Target learning rate at the end (default: 0).
- start_factor – Initial learning rate multiplier (default: 1.0).
- last_epoch – The index of last epoch (default: -1).
Example¶
>>> from olm.train.schedulers import LinearLR
>>> # Decay from initial LR to 0
>>> scheduler = LinearLR(optimizer, total_steps=1000, end_lr=0)
>>> for step in range(total_steps):
... train(...)
... scheduler.step()
get_lr()¶
Compute learning rate using linear interpolation.
class olm.train.schedulers.linear.SchedulerBase(*args: Any, **kwargs: Any)¶
Bases: _LRScheduler, ABC
Base class for all OLM learning rate schedulers.
This class extends PyTorch’s _LRScheduler and provides a consistent interface for implementing custom learning rate schedules. All OLM schedulers should inherit from this class to maintain uniformity.
Subclasses must implement: : - get_lr(): Compute the learning rate for the current step - _get_closed_form_lr() (optional): Closed-form solution for efficiency
- Parameters:
- optimizer – Wrapped PyTorch optimizer.
- last_epoch – The index of the last epoch (default: -1).
- verbose – If True, prints a message to stdout for each update (default: False).
Example¶
>>> class MyScheduler(SchedulerBase):
... def __init__(self, optimizer, param, last_epoch=-1):
... self.param = param
... super().__init__(optimizer, last_epoch)
...
... def get_lr(self):
... # Custom logic here
... return [base_lr * self.param for base_lr in self.base_lrs]
get_last_lr() → List[float]¶
Return last computed learning rate by current scheduler.
- Returns: List of last computed learning rates.
abstractmethod get_lr() → List[float]¶
Compute learning rate for each parameter group.
This method must be implemented by subclasses to define the learning rate schedule logic.
- Returns: List of learning rates, one per parameter group.
load_state_dict(state_dict)¶
Load the scheduler state from a checkpoint.
- Parameters: state_dict – Scheduler state returned by state_dict().
state_dict()¶
Returns the state of the scheduler as a dict.
Contains all non-callable attributes that are specific to the scheduler and required for checkpointing.