详细描述

This implements the adaptive momentum correction method.

A standard momentum correction performs update based on a momentum (eg, \(\mu\)), a previous descend direction (eg, \(v\)) and a current descend direction (eg, \(d\)).

The idea of adaptive momentum correction method is If signs of the last two momentum corrections are different, the current descend direction is discounted On the other hand, if the the signs are the same, the method raises the current descend direction Please see method get_corrected_descend_direction() for details

在文件 AdaptMomentumCorrection.h 第 50 行定义.

类 AdaptMomentumCorrection 继承关系图:

[图例]

Public 成员函数
	AdaptMomentumCorrection ()

virtual void	set_momentum_correction (MomentumCorrection *correction)

virtual	~AdaptMomentumCorrection ()

virtual bool	is_initialized ()

virtual void	set_correction_weight (float64_t weight)

virtual void	initialize_previous_direction (index_t len)

virtual DescendPair	get_corrected_descend_direction (float64_t negative_descend_direction, index_t idx)

virtual void	update_context (CMinimizerContext *context)

virtual void	load_from_context (CMinimizerContext *context)

virtual void	set_adapt_rate (float64_t adapt_rate, float64_t rate_min=0.0, float64_t rate_max=CMath::INFTY)

virtual void	set_init_descend_rate (float64_t init_descend_rate)

virtual float64_t	get_previous_descend_direction (index_t idx)

virtual float64_t	get_length_previous_descend_direction ()

Protected 属性
SGVector< float64_t >	m_descend_rate

MomentumCorrection *	m_momentum_correction

float64_t	m_adapt_rate

float64_t	m_rate_min

float64_t	m_rate_max

float64_t	m_init_descend_rate

SGVector< float64_t >	m_previous_descend_direction

float64_t	m_weight

构造及析构函数说明

AdaptMomentumCorrection ( )

在文件 AdaptMomentumCorrection.h 第 55 行定义.

virtual ~AdaptMomentumCorrection ( )

virtual

在文件 AdaptMomentumCorrection.h 第 73 行定义.

成员函数说明

virtual DescendPair get_corrected_descend_direction	(	float64_t	negative_descend_direction,
		index_t	idx
	)

virtual

Get corrected descend direction

参数

negative_descend_direction	the negative descend direction
idx	the index of the direction

返回: DescendPair (corrected descend direction and the change to correct descend direction)

实现了 MomentumCorrection.

在文件 AdaptMomentumCorrection.h 第 111 行定义.

virtual float64_t get_length_previous_descend_direction ( )

virtualinherited

Get the length of the previous descend direction (velocity)

返回: the length of the previous descend direction

在文件 MomentumCorrection.h 第 143 行定义.

virtual float64_t get_previous_descend_direction ( index_t idx )

virtualinherited

Get the previous descend direction (velocity) given the index

参数

idx	index of the previous descend direction

返回: the previous descend direction

在文件 MomentumCorrection.h 第 132 行定义.

virtual void initialize_previous_direction ( index_t len )

virtual

Initialize m_previous_descend_direction

返回: len the length of m_previous_descend_direction to be initialized

重载 MomentumCorrection .

在文件 AdaptMomentumCorrection.h 第 99 行定义.

virtual bool is_initialized ( )

virtual

Is the standard momentum method initialized?

返回: whether the standard method is initialized

重载 MomentumCorrection .

在文件 AdaptMomentumCorrection.h 第 79 行定义.

virtual void load_from_context ( CMinimizerContext * context )

virtual

Load the given context object to restore mutable variables

This method will be called by DescendUpdaterWithCorrection::load_from_context(CMinimizerContext* context)

参数

context a context object

重载 MomentumCorrection .

在文件 AdaptMomentumCorrection.h 第 173 行定义.

virtual void set_adapt_rate	(	float64_t	adapt_rate,
		float64_t	rate_min = `0.0`,
		float64_t	rate_max = `CMath::INFTY`
	)

virtual

Set adaptive weights used in this method

参数

adapt_rate	the rate is used to discount/raise the current descend direction (see get_corrected_descend_direction() )
rate_min	minimum of the rate
rate_max	maximum of the rate

在文件 AdaptMomentumCorrection.h 第 191 行定义.

virtual void set_correction_weight ( float64_t weight )

virtual

Set the weight (momentum) for the standard momentum method

参数

weight momentum

重载 DescendCorrection .

在文件 AdaptMomentumCorrection.h 第 89 行定义.

virtual void set_init_descend_rate ( float64_t init_descend_rate )

virtual

Set the init rate used to discount/raise the current descend direction

参数