athena__activation_relu Module

Module containing implementation of the ReLU activation function

This module implements the Rectified Linear Unit (ReLU) activation, which zeroes negative values while preserving positive values.

Mathematical operation:

Derivative:

Properties: Non-saturating, sparse activation, mitigates vanishing gradients Reference: Nair & Hinton (2010), ICML


Uses


Interfaces

public interface relu_actv_type

  • private function initialise(scale, attributes) result(activation)

    Initialise a ReLU activation function

    Arguments

    Type IntentOptional Attributes Name
    real(kind=real32), intent(in), optional :: scale

    Optional scale factor for activation output
    type(onnx_attribute_type), intent(in), optional :: attributes(:)

    Optional ONNX attributes

    Return Value type(relu_actv_type)

    ReLU activation type


Derived Types

type, public, extends(base_actv_type) ::  relu_actv_type

Type for ReLU activation function with overloaded procedures

Components

Type Visibility Attributes Name Initial
logical, public :: apply_scaling = .false.

Boolean to apply scaling or not
character(len=10), public :: name

Name of the activation function
real(kind=real32), public :: scale = 1._real32

Scale of the activation function
real(kind=real32), public :: threshold

Threshold of the activation function

Constructor

private function initialise (scale, attributes)

Initialise a ReLU activation function

Type-Bound Procedures

procedure, public, pass(this) :: apply => apply_relu
procedure, public, pass(this) :: apply_attributes => apply_attributes_relu
procedure, public, pass(this) :: export_attributes => export_attributes_relu
procedure, public, pass(this) :: print_to_unit => print_to_unit_actv
procedure, public, pass(this) :: reset => reset_relu

Functions

public function create_from_onnx_relu_activation(attributes) result(activation)

Create ReLU activation function from ONNX attributes

Arguments

Type IntentOptional Attributes Name
type(onnx_attribute_type), intent(in), dimension(:) :: attributes

Array of ONNX attributes

Return Value class(base_actv_type), allocatable

Instance of activation type

private function apply_relu(this, val) result(output)

Apply ReLU activation to 1D array

Read more…

Arguments

Type IntentOptional Attributes Name
class(relu_actv_type), intent(in) :: this

ReLU activation type
type(array_type), intent(in) :: val

Input values

Return Value type(array_type), pointer

Activated output values

private pure function export_attributes_relu(this) result(attributes)

Export ReLU activation function attributes as ONNX attributes

Arguments

Type IntentOptional Attributes Name
class(relu_actv_type), intent(in) :: this

ReLU activation type

Return Value type(onnx_attribute_type), allocatable, dimension(:)

Array of ONNX attributes

private function initialise(scale, attributes) result(activation)

Initialise a ReLU activation function

Arguments

Type IntentOptional Attributes Name
real(kind=real32), intent(in), optional :: scale

Optional scale factor for activation output
type(onnx_attribute_type), intent(in), optional :: attributes(:)

Optional ONNX attributes

Return Value type(relu_actv_type)

ReLU activation type


Subroutines

private subroutine apply_attributes_relu(this, attributes)

Load ONNX attributes into ReLU activation function

Arguments

Type IntentOptional Attributes Name
class(relu_actv_type), intent(inout) :: this

ReLU activation type
type(onnx_attribute_type), intent(in), dimension(:) :: attributes

Array of ONNX attributes

private pure subroutine reset_relu(this)

Reset ReLU activation function attributes and variables

Arguments

Type IntentOptional Attributes Name
class(relu_actv_type), intent(inout) :: this

ReLU activation type