CosineAnnealingScheduler

class ignite.handlers.param_scheduler.CosineAnnealingScheduler(optimizer, param_name, start_value, end_value, cycle_size, cycle_mult=1.0, start_value_mult=1.0, end_value_mult=1.0, save_history=False, param_group_index=None)

Anneals ‘start_value’ to ‘end_value’ over each cycle.

The annealing takes the form of the first half of a cosine wave (as suggested in [Smith17]).

Parameters

• optimizer (Optimizer) – torch optimizer or any object with attribute param_groups as a sequence.

• param_name (str) – name of optimizer’s parameter to update.

• start_value (float) – value at start of cycle.

• end_value (float) – value at the end of the cycle.

• cycle_size (int) – length of cycle.

• cycle_mult (float) – ratio by which to change the cycle_size at the end of each cycle (default=1).

• start_value_mult (float) – ratio by which to change the start value at the end of each cycle (default=1.0).

• end_value_mult (float) – ratio by which to change the end value at the end of each cycle (default=1.0).

• save_history (bool) – whether to log the parameter values to engine.state.param_history, (default=False).

• param_group_index (Optional[int]) – optimizer’s parameters group to use.

Note

If the scheduler is bound to an ‘ITERATION_*’ event, ‘cycle_size’ should usually be the number of batches in an epoch.

Examples

# CosineAnnealing increases the learning rate from 0.0 to 1.0
# over a cycle of 4 iterations
scheduler = CosineAnnealingScheduler(default_optimizer, "lr", 0.0, 1.0, 4)

default_trainer.add_event_handler(Events.ITERATION_STARTED, scheduler)

@default_trainer.on(Events.ITERATION_COMPLETED)
def print_lr():
    print(default_optimizer.param_groups[0]["lr"])

default_trainer.run([0] * 9, max_epochs=1)

0.0
0.1464...
0.4999...
0.8535...
...

optimizer = torch.optim.SGD(
    [
        {"params": default_model.base.parameters(), "lr": 0.001},
        {"params": default_model.fc.parameters(), "lr": 0.01},
    ]
)

# CosineAnnealing increases the learning rate from 0.0 to 1.0
# over a cycle of 4 iterations
scheduler_1 = CosineAnnealingScheduler(optimizer, "lr (base)", 0.0, 1.0, 4, param_group_index=0)

# CosineAnnealing increases the learning rate from 0.0 to 0.1
# over a cycle of 4 iterations
scheduler_2 = CosineAnnealingScheduler(optimizer, "lr (fc)", 0.0, 0.1, 4, param_group_index=1)

default_trainer.add_event_handler(Events.ITERATION_STARTED, scheduler_1)
default_trainer.add_event_handler(Events.ITERATION_STARTED, scheduler_2)

@default_trainer.on(Events.ITERATION_COMPLETED)
def print_lr():
    print(optimizer.param_groups[0]["lr (base)"],
          optimizer.param_groups[1]["lr (fc)"])

default_trainer.run([0] * 9, max_epochs=1)

0.0 0.0
0.1464... 0.01464...
0.4999... 0.04999...
0.8535... 0.08535...
...

Smith17

Smith, Leslie N. “Cyclical learning rates for training neural networks.” Applications of Computer Vision (WACV), 2017 IEEE Winter Conference on. IEEE, 2017

New in version 0.4.5.

Methods

get_param

Method to get current optimizer's parameter value

get_param()

Method to get current optimizer’s parameter value

Return type

float