Source code for niftynet.engine.application_optimiser

# -*- coding: utf-8 -*-

"""
To customise optimisers including
new optimisation methods, learning rate decay schedule,
or customise other optional parameters of the optimiser:

create a `newclass.py` that has a class `NewOptimisor` and implement
`get_instance()`.
and set config parameter in config file or from command line
specify `--optimiser newclass.NewOptimisor`
"""

from __future__ import absolute_import, print_function

import tensorflow as tf


# pylint: disable=too-few-public-methods

class Adam(object):
    """
    Adam optimiser with default hyper parameters
    """

    @staticmethod
    def get_instance(learning_rate):
        """
        create an instance of the optimiser
        """
        return tf.train.AdamOptimizer(
            learning_rate=learning_rate,
            beta1=0.9,
            beta2=0.999,
            epsilon=1e-08,
            use_locking=False, name='Adam')


class Adagrad(object):
    """
    Adagrad optimiser with default hyper parameters
    """

    @staticmethod
    def get_instance(learning_rate):
        """
        create an instance of the optimiser
        """
        return tf.train.AdagradOptimizer(
            learning_rate=learning_rate,
            initial_accumulator_value=0.1,
            use_locking=False, name='Adagrad')


class Momentum(object):
    """
    Momentum optimiser with default hyper parameters
    """

    @staticmethod
    def get_instance(learning_rate):
        """
        create an instance of the optimiser
        """
        return tf.train.MomentumOptimizer(
            learning_rate=learning_rate,
            momentum=0.9,
            use_locking=False,
            name='Momentum',
            use_nesterov=False)


class NesterovMomentum(object):
    """
    Nesterov Momentum optimiser with default hyper parameters
    """

    @staticmethod
    def get_instance(learning_rate):
        """
        create an instance of the optimiser
        """
        return tf.train.MomentumOptimizer(
            learning_rate=learning_rate,
            momentum=0.9,
            use_locking=False,
            name='Momentum',
            use_nesterov=True)


class RMSProp(object):
    """
    RMSProp optimiser with default hyper parameters
    """

    @staticmethod
    def get_instance(learning_rate):
        """
        create an instance of the optimiser
        """
        return tf.train.RMSPropOptimizer(
            learning_rate=learning_rate,
            decay=0.9,
            momentum=0.0,
            epsilon=1e-10,
            use_locking=False,
            centered=False,
            name='RMSProp')


class GradientDescent(object):
    """
    Gradient Descent optimiser with default hyper parameters
    """

    @staticmethod
    def get_instance(learning_rate):
        """
        create an instance of the optimiser
        """
        return tf.train.GradientDescentOptimizer(
            learning_rate=learning_rate,
            use_locking=False,
            name='GradientDescent')