This class is intended to be used as a base class for implementing edges. More...

#include <edge.h>

Inheritance diagram for Edge:

Public Member Functions
	Edge (const config::Edge &edge_config)
	Instatntiate an Edge from the config. More...

virtual void	AllocateMemory (bool fprop_only)
	Allocate memory for the model. More...

virtual void	Initialize ()
	Initialize the weights and biases. More...

virtual void	SaveParameters (hid_t file)
	Write the weights and biases in an hdf5 file. More...

virtual void	LoadParameters (hid_t file)
	Load the weights and biases from an hdf5 file. More...

virtual void	InsertPolyak ()

virtual void	BackupCurrent ()

virtual void	LoadCurrentOnGPU ()

virtual void	LoadPolyakOnGPU ()

virtual float	GetRMSWeight ()
	Returns the root mean square weight value. More...

virtual void	ReduceLearningRate (float factor)
	Reduce the learning rate by factor. More...

virtual bool	HasNoParameters () const
	Returns whether the edge has any parameters. More...

virtual int	GetNumModules () const
	Returns the number of modules. More...

virtual void	DisplayWeights ()
	Displays the weights. More...

virtual void	DisplayWeightStats ()
	Displays the statistics of the weights. More...

virtual void	SetTiedTo (Edge *e)
	Sets the edge to be tied to another edge. More...

virtual void	ComputeUp (Matrix &input, Matrix &output, bool overwrite)=0
	Computes the output layer state given the input. More...

virtual void	ComputeDown (Matrix &deriv_output, Matrix &input, Matrix &output, Matrix &deriv_input, bool overwrite)=0
	Computes the derivative w.r.t the inputs of this edge given the derivative w.r.t the outputs of this edge. More...

virtual void	ComputeOuter (Matrix &input, Matrix &deriv_output)
	Computes the gradient for the weights and biases. More...

virtual void	UpdateWeights ()
	Update the weights. More...

virtual void	SetImageSize (int image_size)
	Set the spatial size of the input to this edge. More...

bool	IsBackPropBlocked () const
	Returns whether back prop is blocked through this edge. More...

void	SetSource (Layer *source)

void	SetDest (Layer *dest)

Layer *	GetSource ()

Layer *	GetDest ()

const string &	GetSourceName ()

const string &	GetDestName ()

const string &	GetName ()

void	SetInputChannels (int a)
	Set the number of input channels. More...

void	SetOutputChannels (int a)
	Set the number of output channels. More...

void	SetMark ()

bool	HasMark ()

string	GetTiedEdgeName ()

bool	IsTied ()

int	GetGPUId () const

void	ComputeStart (Matrix &mat)

void	ComputeEnd (Matrix &mat)

Static Public Member Functions
static Edge *	ChooseEdgeClass (const config::Edge &edge_config)
	Selects the appropriate derived class for the edge config. More...

Protected Attributes
Layer *	source_

Layer *	dest_
	The source layer for this edge. More...

const string	source_node_
	The destination layer for this edge. More...

const string	dest_node_

const string	name_

const string	tied_edge_name_

Edge *	tied_edge_

int	num_input_channels_

int	num_output_channels_

int	image_size_

int	num_modules_

bool	mark_

const bool	block_backprop_
	A marker. More...

const bool	is_tied_

ImageDisplayer *	img_display_

const int	gpu_id_

Detailed Description

This class is intended to be used as a base class for implementing edges.

This is an abstract class - ComputeUp and ComputeDown methods must be implemented by derived classes.

Constructor & Destructor Documentation

Edge::Edge ( const config::Edge & edge_config )

Instatntiate an Edge from the config.

Member Function Documentation

void Edge::AllocateMemory ( bool fprop_only )

virtual

Allocate memory for the model.

Parameters

fprop_only If true, does not allocate memory needed for optimization.

Reimplemented in ResponseNormEdge, ConvOneToOneEdge, MaxPoolEdge, ConvEdge, FCEdge, and LocalEdge.

Edge * Edge::ChooseEdgeClass ( const config::Edge & edge_config )

static

Selects the appropriate derived class for the edge config.

virtual void Edge::ComputeDown	(	Matrix &	deriv_output,
		Matrix &	input,
		Matrix &	output,
		Matrix &	deriv_input,
		bool	overwrite
	)

pure virtual

Computes the derivative w.r.t the inputs of this edge given the derivative w.r.t the outputs of this edge.

Parameters

deriv_output	Derivative w.r.t outputs of this edge.(In)
input	The input to this edge.(In)
output	The output of this edge.(In)
deriv_input	Derivative w.r.t inputs of this edge.(Out)

Implemented in ResponseNormEdge, ConvOneToOneEdge, DownSampleEdge, MaxPoolEdge, UpSampleEdge, ConvEdge, FCEdge, LocalEdge, and RGBToYUVEdge.

void Edge::ComputeOuter	(	Matrix &	input,
		Matrix &	deriv_output
	)

virtual

Computes the gradient for the weights and biases.

Parameters

input	The input to this edge.
deriv_output	The derivative w.r.t the output of this edge.

Reimplemented in ConvOneToOneEdge, ConvEdge, FCEdge, and LocalEdge.

virtual void Edge::ComputeUp	(	Matrix &	input,
		Matrix &	output,
		bool	overwrite
	)

pure virtual

Computes the output layer state given the input.

Applies the weights and adds bias.

Implemented in ResponseNormEdge, ConvOneToOneEdge, DownSampleEdge, MaxPoolEdge, UpSampleEdge, ConvEdge, FCEdge, LocalEdge, and RGBToYUVEdge.

void Edge::DisplayWeights ( )

virtual

Displays the weights.

Supportsinput layer weights only.

Reimplemented in EdgeWithWeight, ConvEdge, and LocalEdge.

void Edge::DisplayWeightStats ( )

virtual

Displays the statistics of the weights.

Reimplemented in EdgeWithWeight.

int Edge::GetNumModules ( ) const

virtual

Returns the number of modules.

This is relevant for convolution-like edges.

Reimplemented in EdgeWithWeight, LocalEdge, ConvEdge, ConvOneToOneEdge, MaxPoolEdge, and RGBToYUVEdge.

float Edge::GetRMSWeight ( )

virtual

Returns the root mean square weight value.

Reimplemented in EdgeWithWeight.

bool Edge::HasNoParameters ( ) const

virtual

Returns whether the edge has any parameters.

Reimplemented in EdgeWithWeight.

void Edge::Initialize ( )

virtual

Initialize the weights and biases.

Reimplemented in EdgeWithWeight.

bool Edge::IsBackPropBlocked ( ) const

inline

Returns whether back prop is blocked through this edge.

void Edge::LoadParameters ( hid_t file )

virtual

Load the weights and biases from an hdf5 file.

Parameters

file	The file handle. The file has been opened for reading. Do not close it.

Reimplemented in EdgeWithWeight.

void Edge::ReduceLearningRate ( float factor )

virtual

Reduce the learning rate by factor.

Reimplemented in EdgeWithWeight.

void Edge::SaveParameters ( hid_t file )

virtual

Write the weights and biases in an hdf5 file.

Parameters

file	The file handle. The file has been opened for writing. Do not close it.

Reimplemented in EdgeWithWeight.

void Edge::SetImageSize ( int image_size )

virtual

Set the spatial size of the input to this edge.

Reimplemented in LocalEdge, ConvEdge, ConvOneToOneEdge, MaxPoolEdge, ResponseNormEdge, DownSampleEdge, and UpSampleEdge.

void Edge::SetInputChannels ( int a )

Set the number of input channels.

void Edge::SetOutputChannels ( int a )

Set the number of output channels.

void Edge::SetTiedTo ( Edge * e )

virtual

Sets the edge to be tied to another edge.

Reimplemented in EdgeWithWeight, ConvEdge, LocalEdge, ResponseNormEdge, and MaxPoolEdge.

void Edge::UpdateWeights ( )

virtual

Update the weights.

Reimplemented in EdgeWithWeight.

Member Data Documentation

const bool Edge::block_backprop_

protected

A marker.

Used for topological sorting.

Layer* Edge::dest_

protected

The source layer for this edge.

const string Edge::source_node_

protected

The destination layer for this edge.

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

src/edge.h
src/edge.cc

Public Member Functions

Static Public Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

Member Function Documentation

Member Data Documentation