mlpractical/mlp/schedulers.py

134 lines
4.9 KiB
Python
Raw Normal View History

2016-10-10 10:23:53 +02:00
# -*- coding: utf-8 -*-
"""Training schedulers.
2015-10-12 02:50:05 +02:00
2016-10-10 10:23:53 +02:00
This module contains classes implementing schedulers which control the
evolution of learning rule hyperparameters (such as learning rate) over a
training run.
"""
2015-10-12 02:50:05 +02:00
2016-10-10 10:23:53 +02:00
import numpy as np
2015-10-12 02:50:05 +02:00
2016-10-10 10:23:53 +02:00
class ConstantLearningRateScheduler(object):
"""Example of scheduler interface which sets a constant learning rate."""
2015-10-12 02:50:05 +02:00
2016-10-10 10:23:53 +02:00
def __init__(self, learning_rate):
"""Construct a new constant learning rate scheduler object.
2015-10-12 02:50:05 +02:00
2016-10-10 10:23:53 +02:00
Args:
learning_rate: Learning rate to use in learning rule.
2015-10-12 02:50:05 +02:00
"""
2016-10-10 10:23:53 +02:00
self.learning_rate = learning_rate
2015-11-14 18:06:12 +01:00
2016-10-10 10:23:53 +02:00
def update_learning_rule(self, learning_rule, epoch_number):
"""Update the hyperparameters of the learning rule.
2015-11-14 18:06:12 +01:00
2016-10-27 12:24:00 +02:00
Runs at the beginning of each epoch.
2015-11-14 18:06:12 +01:00
2016-10-10 10:23:53 +02:00
Args:
learning_rule: Learning rule object being used in training run,
any scheduled hyperparameters to be altered should be
attributes of this object.
epoch_number: Integer index of training epoch about to be run.
"""
learning_rule.learning_rate = self.learning_rate
class ExponentialLearningRateScheduler(object):
"""Exponential decay learning rate scheduler."""
def __init__(self, init_learning_rate, decay_param):
"""Construct a new learning rate scheduler object.
Args:
init_learning_rate: Initial learning rate at epoch 0. Should be a
positive value.
decay_param: Parameter governing rate of learning rate decay.
Should be a positive value.
"""
self.init_learning_rate = init_learning_rate
self.decay_param = decay_param
def update_learning_rule(self, learning_rule, epoch_number):
"""Update the hyperparameters of the learning rule.
Runs at the beginning of each epoch.
Args:
learning_rule: Learning rule object being used in training run,
any scheduled hyperparameters to be altered should be
attributes of this object.
epoch_number: Integer index of training epoch about to be run.
"""
learning_rule.learning_rate = (
self.init_learning_rate * np.exp(-epoch_number / self.decay_param))
class ReciprocalLearningRateScheduler(object):
"""Reciprocal decay learning rate scheduler."""
def __init__(self, init_learning_rate, decay_param):
"""Construct a new learning rate scheduler object.
Args:
init_learning_rate: Initial learning rate at epoch 0. Should be a
positive value.
decay_param: Parameter governing rate of learning rate decay.
Should be a positive value.
"""
self.init_learning_rate = init_learning_rate
self.decay_param = decay_param
def update_learning_rule(self, learning_rule, epoch_number):
"""Update the hyperparameters of the learning rule.
Runs at the beginning of each epoch.
Args:
learning_rule: Learning rule object being used in training run,
any scheduled hyperparameters to be altered should be
attributes of this object.
epoch_number: Integer index of training epoch about to be run.
"""
learning_rule.learning_rate = (
self.init_learning_rate / (1. + epoch_number / self.decay_param)
)
class ReciprocalMomentumCoefficientScheduler(object):
"""Reciprocal growth momentum coefficient scheduler."""
def __init__(self, max_mom_coeff=0.99, growth_param=3., epoch_offset=5.):
"""Construct a new reciprocal momentum coefficient scheduler object.
Args:
max_mom_coeff: Maximum momentum coefficient to tend to. Should be
in [0, 1].
growth_param: Parameter governing rate of increase of momentum
coefficient over training. Should be >= 0 and <= epoch_offset.
epoch_offset: Offset to epoch counter to in scheduler updates to
govern how quickly momentum initially increases. Should be
>= 1.
"""
assert max_mom_coeff >= 0. and max_mom_coeff <= 1.
assert growth_param >= 0. and growth_param <= epoch_offset
assert epoch_offset >= 1.
self.max_mom_coeff = max_mom_coeff
self.growth_param = growth_param
self.epoch_offset = epoch_offset
def update_learning_rule(self, learning_rule, epoch_number):
"""Update the hyperparameters of the learning rule.
Runs at the beginning of each epoch.
Args:
learning_rule: Learning rule object being used in training run,
any scheduled hyperparameters to be altered should be
attributes of this object.
epoch_number: Integer index of training epoch about to be run.
"""
learning_rule.mom_coeff = self.max_mom_coeff * (
1. - self.growth_param / (epoch_number + self.epoch_offset)
)