allennlp.training.scheduler

class allennlp.training.scheduler.Scheduler(optimizer: torch.optim.optimizer.Optimizer, param_group_field: str, last_epoch: int = -1)[source]

Bases: object

A Scheduler is a generalization of PyTorch learning rate schedulers.

A scheduler can be used to update any field in an optimizer’s parameter groups, not just the learning rate.

During training using the AllenNLP Trainer, this is the API and calling sequence for step and step_batch:

scheduler = ... # creates scheduler, calls self.step(epoch=-1) in __init__

batch_num_total = 0
for epoch in range(num_epochs):
    for batch in batchs_in_epoch:
        # compute loss, update parameters with current learning rates
        # call step_batch AFTER updating parameters
        batch_num_total += 1
        scheduler.step_batch(batch_num_total)
    # call step() at the END of each epoch
    scheduler.step(validation_metrics, epoch)
get_values(self)[source]
load_state_dict(self, state_dict:Dict[str, Any]) → None[source]

Load the schedulers state.

Parameters
state_dictDict[str, Any]

Scheduler state. Should be an object returned from a call to state_dict.

state_dict(self) → Dict[str, Any][source]

Returns the state of the scheduler as a dict.

step(self, metric:float=None, epoch:int=None) → None[source]
step_batch(self, batch_num_total:int=None) → None[source]

By default, a scheduler is assumed to only update every epoch, not every batch. So this does nothing unless it’s overriden.