Schedules

• Cyclic
• Exponential

Module contents

class olympus.optimizers.schedules.LRSchedule(name=None, *, schedule=None, optimizer=None, **kwargs)

Bases: object

Lazy LRSchedule that allows you to first fetch the supported parameters using get_space and then initialize the underlying schedule using init_optimizer

Parameters:

name: str

Name of a registered schedule

schedule: LRSchedule

Custom schedule, mutually exclusive with :param name

Raises:

RegisteredLRSchedulerNotFound

when using a name of an known schedule

MissingArgument:

if name nor schedule were not set

Examples

from olympus.optimizers import Optimizer
optimizer = Optimizer('sgd')
schedule = LRSchedule('exponential')
schedule.get_space()
# {'gamma': 'loguniform(0.97, 1)'}
schedule.init(optimizer, gamma=0.97)

Attributes:

defaults

Return default hyper parameters

lr_scheduler

Methods

epoch(epoch[, metrics])	Called after every epoch to update LR
get_current_space()	Get currently defined parameter space
get_space()	Return the missing hyper parameters required to initialize the LR schedule
init([optimizer, override])	Initialize the LR schedule with the given hyper parameters
step(step[, metrics])	Called every step/batch to update LR

get_lr
load_state_dict
state_dict

defaults

Return default hyper parameters

epoch(epoch, metrics=None)

Called after every epoch to update LR

get_current_space()

Get currently defined parameter space

get_lr()

get_space()

Return the missing hyper parameters required to initialize the LR schedule

init(optimizer=None, override=False, **kwargs)

Initialize the LR schedule with the given hyper parameters

load_state_dict(state_dict, strict=True)

lr_scheduler

state_dict(destination=None, prefix='', keep_vars=False)

step(step, metrics=None)

Called every step/batch to update LR

exception olympus.optimizers.schedules.RegisteredLRSchedulerNotFound

Bases: Exception

exception olympus.optimizers.schedules.UninitializedLRScheduler

Bases: Exception

olympus.optimizers.schedules.known_schedule()

olympus.optimizers.schedules.register_schedule(name, factory, override=False)