nest::WeightOptimizer Class Reference

Base class implementing a weight optimizer model. More...

#include <weight_optimizer.h>

Inheritance diagram for nest::WeightOptimizer:

Public Member Functions
	WeightOptimizer ()
	Default constructor.
virtual	~WeightOptimizer ()
	Destructor.
	WeightOptimizer (const WeightOptimizer &)=default
	Copy constructor.
WeightOptimizer &	operator= (const WeightOptimizer &)=delete
	Assignment operator.
virtual void	get_status (Dictionary &d) const
	Get parameter dictionary.
virtual void	set_status (const Dictionary &d)
	Update values in parameter dictionary.
double	optimized_weight (const WeightOptimizerCommonProperties &cp, const size_t idx_current_update, const double gradient, double weight)
	Return optimized weight based on current weight.

Protected Member Functions
virtual double	optimize_ (const WeightOptimizerCommonProperties &cp, double weight, size_t current_opt_step)=0
	Perform specific optimization.

Protected Attributes
double	cumulative_gradient_
	Cumulative gradient over the current batch.
size_t	optimization_step_
	Current optimization step, whereby optimization happens every batch_size_ steps.
double	eta_current_
	Synapse-specific learning rate when the history for its upcoming weight update was archived.
long	n_optimize_
	Number of optimizations.

Detailed Description

Base class implementing a weight optimizer model.

An optimizer is used by a synapse that supports this mechanism to optimize the weight.

An optimizer may have an internal state which is maintained from call to call of the optimized_weight() method. Each optimized object belongs to exactly one synapse.

Constructor & Destructor Documentation

WeightOptimizer() [1/2]

nest::WeightOptimizer::WeightOptimizer

(

)

Default constructor.

~WeightOptimizer()

virtual nest::WeightOptimizer::~WeightOptimizer ( )

inlinevirtual

Destructor.

WeightOptimizer() [2/2]

nest::WeightOptimizer::WeightOptimizer ( const WeightOptimizer &  )

default

Copy constructor.

Member Function Documentation

get_status()

void nest::WeightOptimizer::get_status ( Dictionary &  d ) const

virtual

Get parameter dictionary.

Reimplemented in nest::WeightOptimizerAdam.

Referenced by nest::WeightOptimizerAdam::get_status().

Here is the caller graph for this function:

operator=()

WeightOptimizer & nest::WeightOptimizer::operator= ( const WeightOptimizer &  )

delete

Assignment operator.

optimize_()

virtual double nest::WeightOptimizer::optimize_ ( const WeightOptimizerCommonProperties &  cp, double  weight, size_t  current_opt_step  )

protectedpure virtual

Perform specific optimization.

Implemented in nest::WeightOptimizerGradientDescent, and nest::WeightOptimizerAdam.

Referenced by optimized_weight().

Here is the caller graph for this function:

optimized_weight()

double nest::WeightOptimizer::optimized_weight	(	const WeightOptimizerCommonProperties &	cp,
		const size_t	idx_current_update,
		const double	gradient,
		double	weight
	)

Return optimized weight based on current weight.

References nest::WeightOptimizerCommonProperties::batch_size_, cumulative_gradient_, nest::WeightOptimizerCommonProperties::eta_, nest::WeightOptimizerCommonProperties::eta_change_count_, eta_current_, nest::WeightOptimizerCommonProperties::eta_first_change_, n_optimize_, optimization_step_, optimize_(), nest::WeightOptimizerCommonProperties::Wmax_, and nest::WeightOptimizerCommonProperties::Wmin_.

Here is the call graph for this function:

set_status()

void nest::WeightOptimizer::set_status ( const Dictionary &  d )

virtual

Update values in parameter dictionary.

Reimplemented in nest::WeightOptimizerAdam.

Referenced by nest::WeightOptimizerAdam::set_status().

Here is the caller graph for this function:

Member Data Documentation

cumulative_gradient_

double nest::WeightOptimizer::cumulative_gradient_

protected

Cumulative gradient over the current batch.

Referenced by nest::WeightOptimizerGradientDescent::optimize_(), nest::WeightOptimizerAdam::optimize_(), and optimized_weight().

eta_current_

double nest::WeightOptimizer::eta_current_

protected

Synapse-specific learning rate when the history for its upcoming weight update was archived.

Ensures the correct learning rate during the next optimization, even if the global learning rate changes before activation of the synapse.

Referenced by nest::WeightOptimizerGradientDescent::optimize_(), nest::WeightOptimizerAdam::optimize_(), and optimized_weight().

n_optimize_

long nest::WeightOptimizer::n_optimize_

protected

Number of optimizations.

Referenced by optimized_weight().

optimization_step_

size_t nest::WeightOptimizer::optimization_step_

protected

Current optimization step, whereby optimization happens every batch_size_ steps.

Referenced by nest::WeightOptimizerAdam::optimize_(), and optimized_weight().

---

The documentation for this class was generated from the following files:

• weight_optimizer.h
• weight_optimizer.cpp