opennn::LayerNormOp Struct Reference

Layer normalization with learnable scale/shift, applied across the embedding dimension. More...

#include <operators.h>

Inheritance diagram for opennn::LayerNormOp:

Public Member Functions
void	set (Index sequence_length, Index embedding_dimension)
	Configures the operator for a (sequence_length, embedding_dimension) input.
vector< TensorSpec >	parameter_specs () const override
	Returns the tensor specs of trainable parameters owned by this operator.
void	link_parameters (span< const TensorView > views) override
	Binds parameter views provided by the hosting layer.
void	link_gradients (span< const TensorView > views) override
	Binds gradient views provided by the hosting layer.
void	set_parameters_random () override
	Initializes parameters with random values.
void	set_parameters_glorot () override
	Initializes parameters using Glorot (Xavier) initialization.
void	init_defaults ()
	Resets gamma to one and beta to zero.
void	forward_propagate (ForwardPropagation &fp, size_t layer, bool is_training) noexcept override
	Runs the operator's forward computation.
void	back_propagate (ForwardPropagation &fp, BackPropagation &bp, size_t layer) const noexcept override
	Runs the operator's backward computation, accumulating into gradient/delta buffers.
Public Member Functions inherited from opennn::Operator
virtual	~Operator ()=default
virtual vector< TensorSpec >	state_specs () const
	Returns the tensor specs of persistent state owned by this operator.
virtual void	link_states (span< const TensorView >)
	Binds state views provided by the hosting layer.
virtual void	to_JSON (JsonWriter &) const
	Serializes the operator configuration to a JSON writer.
virtual void	from_JSON (const Json *)
	Restores the operator configuration from a JSON node.
virtual void	load_state_from_JSON (const Json *)
	Restores persistent state (e.g. running statistics) from a JSON node.
virtual void	destroy_cuda ()
	Releases CUDA resources owned by the operator; called from destructors.
TensorView &	get_input (ForwardPropagation &fp, size_t layer, size_t i=0) const noexcept
vector< TensorView > &	get_inputs (ForwardPropagation &fp, size_t layer, size_t i=0) const noexcept
TensorView &	get_output (ForwardPropagation &fp, size_t layer, size_t i=0) const noexcept
TensorView &	get_output_delta (BackPropagation &bp, size_t layer, size_t i=0) const noexcept
TensorView &	get_input_delta (BackPropagation &bp, size_t layer, size_t i=0) const noexcept

Public Attributes
Index	sequence_length = 0
Index	embedding_dimension = 0
TensorView	gamma
TensorView	beta
TensorView	gamma_gradient
TensorView	beta_gradient
Public Attributes inherited from opennn::Operator
vector< size_t >	input_slots = {0}
vector< size_t >	output_slots = {1}
vector< size_t >	input_delta_slots = {1}
vector< size_t >	output_delta_slots = {0}

Detailed Description

Layer normalization with learnable scale/shift, applied across the embedding dimension.

Member Function Documentation

back_propagate()

void opennn::LayerNormOp::back_propagate ( ForwardPropagation & fp, BackPropagation & bp, size_t layer ) const

overridevirtualnoexcept

Runs the operator's backward computation, accumulating into gradient/delta buffers.

Parameters

fp	Forward propagation workspace (read-only).
bp	Back propagation workspace receiving gradients and deltas.
layer	Index of the hosting layer in the workspace.

Reimplemented from opennn::Operator.

forward_propagate()

void opennn::LayerNormOp::forward_propagate ( ForwardPropagation & fp, size_t layer, bool is_training )

overridevirtualnoexcept

Runs the operator's forward computation.

Parameters

fp	Forward propagation workspace.
layer	Index of the hosting layer in the workspace.
is_training	If true, enables training-only behavior (e.g. dropout sampling).

Reimplemented from opennn::Operator.

init_defaults()

void opennn::LayerNormOp::init_defaults

(

)

Resets gamma to one and beta to zero.

link_gradients()

void opennn::LayerNormOp::link_gradients ( span< const TensorView > )

overridevirtual

Binds gradient views provided by the hosting layer.

Reimplemented from opennn::Operator.

link_parameters()

void opennn::LayerNormOp::link_parameters ( span< const TensorView > )

overridevirtual

Binds parameter views provided by the hosting layer.

Reimplemented from opennn::Operator.

parameter_specs()

vector< TensorSpec > opennn::LayerNormOp::parameter_specs ( ) const

overridevirtual

Returns the tensor specs of trainable parameters owned by this operator.

Reimplemented from opennn::Operator.

set()

void opennn::LayerNormOp::set	(	Index	sequence_length,
		Index	embedding_dimension )

Configures the operator for a (sequence_length, embedding_dimension) input.

set_parameters_glorot()

void opennn::LayerNormOp::set_parameters_glorot ( )

inlineoverridevirtual

Initializes parameters using Glorot (Xavier) initialization.

Reimplemented from opennn::Operator.

set_parameters_random()

void opennn::LayerNormOp::set_parameters_random ( )

inlineoverridevirtual

Initializes parameters with random values.

Reimplemented from opennn::Operator.

Member Data Documentation

beta

TensorView opennn::LayerNormOp::beta

beta_gradient

TensorView opennn::LayerNormOp::beta_gradient

embedding_dimension

Index opennn::LayerNormOp::embedding_dimension = 0

gamma

TensorView opennn::LayerNormOp::gamma

gamma_gradient

TensorView opennn::LayerNormOp::gamma_gradient

sequence_length

Index opennn::LayerNormOp::sequence_length = 0