cldnn::normalize Struct Reference

Normalizes the input using an L2 norm and multiplies the output with scale value. The scale can be equal for all channels or one scale per channel. More...

#include <normalize.hpp>

Inheritance diagram for cldnn::normalize:

Collaboration diagram for cldnn::normalize:

Public Types
typedef cldnn_normalize_desc	dto

Public Member Functions
	normalize (const primitive_id &id, const primitive_id &input, const primitive_id &scale_input, const bool across_spatial=true, const float epsilon=1e-10f, const padding &output_padding=padding())
	Constructs normalize primitive. More...
	normalize (const dto *dto)
	Constructs a copy from C API cldnn_normalize_desc.
Public Member Functions inherited from cldnn::primitive_base< normalize, cldnn_normalize_desc >
const cldnn_primitive_desc *	get_dto () const override
	Returns pointer to a C API primitive descriptor casted to cldnn_primitive_desc.
Public Member Functions inherited from cldnn::primitive
	primitive (const primitive_type_id &type, const primitive_id &id, const std::vector< primitive_id > &input, const padding &output_padding=padding())
	primitive (const cldnn_primitive_desc *dto)
	Constructs a copy from basic C API cldnn_primitive_desc.
std::vector< std::reference_wrapper< primitive_id > >	dependecies ()
	Returns references to all primitive ids on which this primitive depends - inputs, weights, biases, etc.
std::vector< primitive_id >	dependecies () const
	Returns copy of all primitive ids on which this primitive depends - inputs, weights, biases, etc.
	operator primitive_id () const
	Implicit conversion to primiitive id.

Static Public Member Functions
static primitive_type_id	type_id ()

Public Attributes
primitive_id	scale_input
	Scale input primitive id with values needed for scaling after the normalization. Scale x dimension should be 1 (if all channels have the same scale) or equal to input feature size (one scale per channel). All other dimensions should be 1.
bool	across_spatial
	Determines if the normalization is done across or within spatial (see documentation above).
float	epsilon
	Epsilon for not dividing by zero while normalizing.
Public Attributes inherited from cldnn::primitive
const primitive_type_id	type
	Primitive's type id.
const primitive_id	id
	Primitive's id.
fixed_size_vector_ref	input
	List of ids of input primitives.
padding	output_padding
	Requested output padding.

Protected Member Functions
std::vector< std::reference_wrapper< const primitive_id > >	get_dependencies () const override
void	update_dto (dto &dto) const override
Protected Member Functions inherited from cldnn::primitive_base< normalize, cldnn_normalize_desc >
	primitive_base (const primitive_id &id, const std::vector< primitive_id > &input, const padding &output_padding=padding())
	primitive_base (const cldnn_normalize_desc *dto)

Additional Inherited Members
Protected Attributes inherited from cldnn::primitive
primitive_id_arr	_input

Detailed Description

Normalizes the input using an L2 norm and multiplies the output with scale value. The scale can be equal for all channels or one scale per channel.

The L2 norm is computed as:
Across spatial mode (across_spatial=true)-
norm(i,x,y) = sqrt( Σ( in(f,w,h)^2 ) + epsilon ) where f in range (0,num_of_features), w in range (0,input_width), h in range (0,input_height).
The summation is performed over all the pixels in the batch.
Within spatial mode (across_spatial=false)-
norm(i,x,y) = sqrt( Σ( in(f,x,y)^2 ) + epsilon ) where f in range (0,num_of_features).
The summation is performed over this (x,y) position on all the features.

Algorithm:

out(i,x,y) = ( in(i,x,y) / norm(i,x,y) ) * scale(i)

Where:

• out(i,x,y) : value at x, y from i-th feature map after normalization.
• in(i,x,y) : value at x, y from i-th feature map before normalization.
• norm(i,x,y) : L2 norm as described above.
• scale(i) : the scale value of the i-th feature map.

Definition at line 47 of file normalize.hpp.

Constructor & Destructor Documentation

normalize()

cldnn::normalize::normalize ( const primitive_id &  id, const primitive_id &  input, const primitive_id &  scale_input, const bool  across_spatial = true, const float  epsilon = 1e-10f, const padding &  output_padding = padding()  )

inline

Constructs normalize primitive.

Parameters

id	This primitive id.
input	Input primitive id.
scale_input	Scale input primitive id with values needed for scaling after the normalization. Scale x dimension should be 1 (if all channels have the same scale) or equal to input feature size (one scale per channel). All other dimensions should be 1.
across_spatial	Determines if the normalization is done across or within spatial (see documentation above).
epsilon	Epsilon for not dividing by zero while normalizing.

Definition at line 59 of file normalize.hpp.

---

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

• C:/ML/OpenSource/clDNN/api/CPP/normalize.hpp