dnn::opt_SGD_momentum Class Reference

Stochastic Gradient Descent with momentum optimizer class. More...

#include <dnn_opt.h>

Inheritance diagram for dnn::opt_SGD_momentum:

Public Member Functions
	opt_SGD_momentum (DNN_Dtype s, DNN_Dtype m, DNN_Dtype l=0.0, DNN_Dtype a=0.0)
	SGD with momentum constructor. More...
	~opt_SGD_momentum ()
void	apply (arma::Cube< DNN_Dtype > &W, arma::Mat< DNN_Dtype > &B, const arma::Cube< DNN_Dtype > &Wgrad, const arma::Mat< DNN_Dtype > &Bgrad)
	Apply the optimizer to the layer parameters. More...
std::string	get_algorithm (void)
	Get the optimizer algorithm information. More...
Public Member Functions inherited from dnn::opt
	opt ()
	~opt ()
void	set_learn_rate_alg (LR_ALG alg, DNN_Dtype a=0.0, DNN_Dtype b=10.0)
	Set learning rate algorithm. More...
void	update_learn_rate (void)
	Update learning rate. More...
DNN_Dtype	get_learn_rate (void)
	Get the learning rate. More...

Private Attributes
arma::Cube< DNN_Dtype >	v
	Velocity internal variable for weight. More...
arma::Mat< DNN_Dtype >	vB
	Velocity internal variable for bias. More...
DNN_Dtype	mom

Additional Inherited Members
Protected Attributes inherited from dnn::opt
std::string	alg
DNN_Dtype	lr
	Learning rate. More...
DNN_Dtype	reg_lambda
	Regularisation parameter lambda. More...
DNN_Dtype	reg_alpha
	Elastic net mix parameter - 0=ridge (L2) .. 1=LASSO (L1) More...
LR_ALG	lr_alg
	Learning rate schedule algorithm. More...
DNN_Dtype	lr_0
	Init value for lr. More...
DNN_Dtype	lr_a
	Internal parameter a. More...
DNN_Dtype	lr_b
	Internal parameter b. More...
arma::uword	it
	Iteration counter. More...

Detailed Description

Stochastic Gradient Descent with momentum optimizer class.

SGD with momentum algorithm, see https://arxiv.org/pdf/1609.04747.pdf

Definition at line 194 of file dnn_opt.h.

Constructor & Destructor Documentation

opt_SGD_momentum()

dnn::opt_SGD_momentum::opt_SGD_momentum ( DNN_Dtype  s, DNN_Dtype  m, DNN_Dtype  l = 0.0, DNN_Dtype  a = 0.0  )

inline

SGD with momentum constructor.

Parameters

[in]	s	Step size - learning rate
[in]	m	Momentum
[in]	l	Regularisation param lambda
[in]	a	Regularisation param alpha

Definition at line 208 of file dnn_opt.h.

~opt_SGD_momentum()

dnn::opt_SGD_momentum::~opt_SGD_momentum ( )

inline

Definition at line 218 of file dnn_opt.h.

Member Function Documentation

apply()

void dnn::opt_SGD_momentum::apply ( arma::Cube< DNN_Dtype > &  W, arma::Mat< DNN_Dtype > &  B, const arma::Cube< DNN_Dtype > &  Wgrad, const arma::Mat< DNN_Dtype > &  Bgrad  )

inlinevirtual

Apply the optimizer to the layer parameters.

Parameters

[in,out]	W,B	Learnable parameters
[in]	Wgrad,Bgrad	Gradient of the learnable parameters

Implements dnn::opt.

Definition at line 225 of file dnn_opt.h.

get_algorithm()

std::string dnn::opt_SGD_momentum::get_algorithm ( void  )

inlinevirtual

Get the optimizer algorithm information.

Returns

Algorithm information string

Reimplemented from dnn::opt.

Definition at line 256 of file dnn_opt.h.

Member Data Documentation

mom

DNN_Dtype dnn::opt_SGD_momentum::mom

private

Definition at line 199 of file dnn_opt.h.

v

arma::Cube<DNN_Dtype> dnn::opt_SGD_momentum::v

private

Velocity internal variable for weight.

Definition at line 197 of file dnn_opt.h.

vB

arma::Mat<DNN_Dtype> dnn::opt_SGD_momentum::vB

private

Velocity internal variable for bias.

Definition at line 198 of file dnn_opt.h.

---

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

• dnn_opt.h