Layer::Layer Class Reference

#include <layer.hpp>

Inheritance diagram for Layer::Layer:

Public Member Functions
	Layer (size_t n, size_t m)
	~Layer ()
size_t	get_input_size () const
size_t	get_output_size () const
Vector	get_output () const
virtual Vector	feed_forward (Vector x)=0
virtual void	init_nabla ()=0
virtual Vector	back_propagation (Vector e)=0
virtual void	update (Real eta)=0

Public Attributes
string	name
size_t	n
size_t	m
Vector	x
Vector	y
Vector	d

Detailed Description

An abstract class representing a layer of a deep network. Such a layer can be seen as a map

\[ \begin{array}{rcl} f:\mathbb{R}^n&\mapsto&\mathbb{R}^m\\ x&\to&y \end{array} \]

.

Constructor & Destructor Documentation

Layer()

Layer::Layer::Layer ( size_t  n, size_t  m  )

inline

~Layer()

Layer::Layer::~Layer ( )

inline

Member Function Documentation

back_propagation()

virtual Vector Layer::Layer::back_propagation ( Vector  e )

pure virtual

Apply back propagation algorithm on the delta output vector d. Used the input vector stored in x_in_ref during feedforward. Return a reference to the computed (and stored) input delta vector. Nabla vectors must be computed here.

Implemented in Layer::ActivationLayer< A >, Layer::ConvolutionLayer, Layer::FullConnectedLayer, and Layer::Pooling.

feed_forward()

virtual Vector Layer::Layer::feed_forward ( Vector  x )

pure virtual

Apply the layer to the input vector x. Vectors x_in_ref and x_out must be updated in consequence. Return a reference to x_out.

Implemented in Layer::ActivationLayer< A >, Layer::ConvolutionLayer, Layer::FullConnectedLayer, and Layer::Pooling.

get_input_size()

size_t Layer::Layer::get_input_size ( ) const

inline

Return the input size.

get_output()

Vector Layer::Layer::get_output ( ) const

inline

Return a reference to the computed output vector.

get_output_size()

size_t Layer::Layer::get_output_size ( ) const

inline

Return the output size.

init_nabla()

virtual void Layer::Layer::init_nabla ( )

pure virtual

Initialize nabla vectors which are used during gradient descent.

Implemented in Layer::ActivationLayer< A >, Layer::ConvolutionLayer, Layer::FullConnectedLayer, and Layer::Pooling.

update()

virtual void Layer::Layer::update ( Real  eta )

pure virtual

Update layer parameters using gradient descent algorithm with learning rate eta.

Implemented in Layer::ActivationLayer< A >, Layer::ConvolutionLayer, Layer::FullConnectedLayer, and Layer::Pooling.

Member Data Documentation

d

Vector Layer::Layer::d

Computed input delta vector computed by back propagation algorithm. Owned by the layer.

m

size_t Layer::Layer::m

Size of the output vector.

n

size_t Layer::Layer::n

Size of the input vector.

name

string Layer::Layer::name

The name of the layer. Used for debugging.

x

Vector Layer::Layer::x

A reference to the input vector.

y

Vector Layer::Layer::y

Computed output vector. Owned by the layer.

---

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

• src/layers/layer.hpp