Abstract base class for every layer in an OpenNN NeuralNetwork. More...

#include <layer.h>

Inheritance diagram for opennn::Layer:

Public Member Functions
virtual	~Layer ()=default
	Virtual destructor; subclasses are owned via unique_ptr<Layer>.

const string &	get_label () const
	Returns the user-assigned label of this layer.

const string &	get_name () const
	Returns the canonical type name of this layer.

LayerType	get_type () const
	Returns the LayerType enumerator for this layer.

virtual void	set_input_shape (const Shape &)
	Sets the per-sample input shape of this layer.

virtual void	set_output_shape (const Shape &)
	Sets the per-sample output shape of this layer.

void	set_label (string new_label)
	Sets the human-readable label of this layer.

Index	get_parameters_number () const
	Total number of trainable parameters in this layer.

virtual vector< Operator * >	get_operators ()
	Returns the operators that compose this layer, if any.

virtual vector< pair< Shape, Type > >	get_parameter_specs () const
	Specifications of the trainable parameter tensors owned by this layer.

virtual vector< pair< Shape, Type > >	get_state_specs () const
	Specifications of the persistent state tensors of this layer.

virtual vector< pair< Shape, Type > >	get_forward_specs (Index batch_size) const
	Specifications of the forward intermediate buffers for one batch.

virtual vector< pair< Shape, Type > >	get_backward_specs (Index batch_size) const
	Specifications of the backward intermediate buffers for one batch.

vector< Shape >	get_parameter_shapes () const
	Shape-only view of get_parameter_specs().

vector< Shape >	get_state_shapes () const
	Shape-only view of get_state_specs().

vector< Shape >	get_forward_shapes (Index b) const
	Shape-only view of get_forward_specs() for batch size `b`.

vector< Shape >	get_backward_shapes (Index b) const
	Shape-only view of get_backward_specs() for batch size `b`.

vector< Type >	get_parameter_dtypes () const
	Dtype-only view of get_parameter_specs().

vector< Type >	get_forward_dtypes (Index b) const
	Dtype-only view of get_forward_specs() for batch size `b`.

vector< Type >	get_backward_dtypes (Index b) const
	Dtype-only view of get_backward_specs() for batch size `b`.

virtual Shape	get_input_shape () const =0
	Per-sample input shape of the layer (excluding the batch dimension).

virtual Shape	get_output_shape () const =0
	Per-sample output shape of the layer (excluding the batch dimension).

virtual Activation::Function	get_output_activation () const
	Activation function fused at the end of this layer, if any.

Index	get_inputs_number () const
	Total number of scalar inputs per sample (product of input dims).

Index	get_outputs_number () const
	Total number of scalar outputs per sample (product of output dims).

virtual void	forward_propagate (ForwardPropagation &fp, size_t layer, bool is_training) noexcept
	Forward pass: reads inputs from `fp` and writes outputs into `fp`.

virtual void	back_propagate (ForwardPropagation &, BackPropagation &, size_t) const noexcept
	Backward pass: propagates gradients through this layer.

virtual void	from_JSON (const JsonDocument &document)
	Loads the layer configuration (hyperparameters) from JSON.

virtual void	read_JSON_body (const Json *)
	Subclass hook for parsing the body of from_JSON().

virtual void	load_state_from_JSON (const JsonDocument &document)
	Loads parameter and state tensors from a JSON document.

virtual void	to_JSON (JsonWriter &writer) const
	Writes the layer configuration to JSON.

virtual void	write_JSON_body (JsonWriter &) const
	Subclass hook for emitting the body of to_JSON().

virtual void	print () const
	Prints a human-readable summary of the layer to stdout.

bool	get_is_trainable () const
	Whether this layer has trainable parameters.

Type	get_compute_dtype () const
	Numerical type used for forward/backward computation.

void	set_compute_dtype (Type new_compute_dtype)
	Sets the compute dtype and triggers on_compute_dtype_changed().

virtual void	on_compute_dtype_changed ()
	Hook invoked after set_compute_dtype() mutates the dtype.

virtual float *	link_parameters (float *pointer)
	Wires this layer's parameter TensorViews onto an external buffer.

virtual float *	link_states (float *pointer)
	Wires this layer's state TensorViews onto an external buffer.

vector< TensorView > &	get_parameter_views ()
	Mutable access to this layer's parameter TensorViews.

const vector< TensorView > &	get_parameter_views () const
	Read-only access to this layer's parameter TensorViews.

vector< TensorView > &	get_state_views ()
	Mutable access to this layer's state TensorViews.

const vector< TensorView > &	get_state_views () const
	Read-only access to this layer's state TensorViews.

void	redistribute_parameters_to_operators ()
	Forwards the current parameter views down to each composing Operator.

void	redistribute_parameter_gradients_to_operators (vector< TensorView > &gradient_views)
	Forwards externally provided gradient views down to each Operator.

void	redistribute_states_to_operators ()
	Forwards the current state views down to each composing Operator.

Protected Member Functions
	Layer ()=default
	Default constructor; only invoked by subclasses.

float *	link_views (float pointer, const vector< Shape > &shapes, vector< TensorView > &views, const char tag) const
	Builds `views` over a contiguous float buffer using `shapes`.

void	distribute_to_operators (vector< TensorView > &views, void(Operator::link)(const vector< TensorView > &), vector< pair< Shape, Type > >(Operator::specs)() const)
	Generic helper used by the redistribute_*_to_operators() routines.

Protected Attributes
string	label = "my_layer"
	User-visible label for this layer instance (default "my_layer").

string	name = "layer"
	Canonical type name set by the subclass (e.g. "dense").

LayerType	layer_type = LayerType::Dense
	Layer type tag set by the subclass.

bool	is_trainable = true
	True if the layer has parameters that participate in training.

bool	is_first_layer = false
	True if this layer is the network's input layer.

Type	compute_dtype = Type::FP32
	Numerical type used for forward and backward computation.

vector< TensorView >	parameters
	Parameter TensorViews bound to the network's parameter arena.

vector< TensorView >	states
	State TensorViews bound to the network's state arena.

vector< unique_ptr< Layer > >	layers
	Sub-layers, when this layer is itself a composite.

Detailed Description

Abstract base class for every layer in an OpenNN NeuralNetwork.

A Layer declares:

Its parameter, state, forward and backward buffer specifications, used by the network to size the shared memory arenas before training.
Its input and output shapes (per-sample, batch dimension is added by the caller).
Its forward and backward propagation routines, which read from and write into ForwardPropagation / BackPropagation views provided by the caller.
JSON serialization hooks so the layer can be persisted as part of a NeuralNetwork file.

Layers do not own batch-dependent buffers; they only own their parameter and state TensorViews into externally allocated memory.

Constructor & Destructor Documentation

◆ ~Layer()

virtual opennn::Layer::~Layer ( )

virtualdefault

Virtual destructor; subclasses are owned via unique_ptr<Layer>.

◆ Layer()

opennn::Layer::Layer ( )

protecteddefault

Default constructor; only invoked by subclasses.

Member Function Documentation

◆ back_propagate()

virtual void opennn::Layer::back_propagate	(	ForwardPropagation &	,
		BackPropagation &	,
		size_t	) const

inlinevirtualnoexcept

Backward pass: propagates gradients through this layer.

Receives the forward intermediates, the BackPropagation buffer in which to accumulate gradients, and this layer's index inside the network. Throws by default; concrete trainable layers must override.

Reimplemented in opennn::Addition, opennn::Convolutional, opennn::ConvolutionalRelu, opennn::Dense, opennn::DenseRelu, opennn::Embedding, opennn::Flatten, opennn::MultiHeadAttention, opennn::Normalization3d, opennn::Pooling3d, opennn::Pooling, and opennn::Recurrent.

◆ distribute_to_operators()

void opennn::Layer::distribute_to_operators	(	vector< TensorView > &	views,
		void(Operator::*	link )(const vector< TensorView > &),
		vector< pair< Shape, Type > >(Operator::*	specs )() const )

protected

Generic helper used by the redistribute_*_to_operators() routines.

Parameters

views	Source TensorViews to slice.
link	Operator member function that receives each sub-view list.
specs	Operator member function returning the per-Operator spec list.

◆ forward_propagate()

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

inlinevirtualnoexcept

Forward pass: reads inputs from fp and writes outputs into fp.

Parameters

fp	ForwardPropagation buffer slice owned by the caller.
layer	Index of this layer inside the network.
is_training	True during training (enables dropout, BN updates, etc.).

The default implementation runs every Operator returned by get_operators() in order. Leaf layers override this directly.

Reimplemented in opennn::ConvolutionalRelu, opennn::MultiHeadAttention, opennn::Recurrent, opennn::Scaling, and opennn::Unscaling.

◆ from_JSON()

virtual void opennn::Layer::from_JSON ( const JsonDocument & document )

virtual

Loads the layer configuration (hyperparameters) from JSON.

Parameters

document Parsed JSON document for this layer.

◆ get_backward_dtypes()

vector< Type > opennn::Layer::get_backward_dtypes ( Index b ) const

inline

Dtype-only view of get_backward_specs() for batch size b.

◆ get_backward_shapes()

vector< Shape > opennn::Layer::get_backward_shapes ( Index b ) const

inline

Shape-only view of get_backward_specs() for batch size b.

◆ get_backward_specs()

virtual vector< pair< Shape, Type > > opennn::Layer::get_backward_specs ( Index batch_size ) const

inlinevirtual

Specifications of the backward intermediate buffers for one batch.

Parameters

batch_size Number of samples in the batch.

Returns: Vector of (Shape, Type) pairs. Empty for non-trainable layers.

Reimplemented in opennn::Addition, opennn::MultiHeadAttention, and opennn::Recurrent.

◆ get_compute_dtype()

Type opennn::Layer::get_compute_dtype ( ) const

inline

Numerical type used for forward/backward computation.

Returns: Compute dtype (FP32 by default; FP16/BF16 in mixed precision).

◆ get_forward_dtypes()

vector< Type > opennn::Layer::get_forward_dtypes ( Index b ) const

inline

Dtype-only view of get_forward_specs() for batch size b.

◆ get_forward_shapes()

vector< Shape > opennn::Layer::get_forward_shapes ( Index b ) const

inline

Shape-only view of get_forward_specs() for batch size b.

◆ get_forward_specs()

virtual vector< pair< Shape, Type > > opennn::Layer::get_forward_specs ( Index batch_size ) const

inlinevirtual

Specifications of the forward intermediate buffers for one batch.

Parameters

batch_size Number of samples in the batch.

Returns: Vector of (Shape, Type) pairs. The last entry is the layer output, wired to downstream layers.

Reimplemented in opennn::Convolutional, opennn::ConvolutionalRelu, opennn::Dense, opennn::DenseRelu, opennn::Embedding, opennn::MultiHeadAttention, opennn::Normalization3d, opennn::Pooling3d, opennn::Pooling, and opennn::Recurrent.

◆ get_input_shape()

virtual Shape opennn::Layer::get_input_shape ( ) const

pure virtual

Per-sample input shape of the layer (excluding the batch dimension).

Returns: Shape required by forward_propagate() for a single sample.

Implemented in opennn::Addition, opennn::Bounding, opennn::Convolutional, opennn::ConvolutionalRelu, opennn::Dense, opennn::DenseRelu, opennn::Embedding, opennn::Flatten, opennn::MultiHeadAttention, opennn::Normalization3d, opennn::Pooling3d, opennn::Pooling, opennn::Recurrent, opennn::Scaling, and opennn::Unscaling.

◆ get_inputs_number()

Index opennn::Layer::get_inputs_number ( ) const

inline

Total number of scalar inputs per sample (product of input dims).

Returns: Flat input size.

◆ get_is_trainable()

bool opennn::Layer::get_is_trainable ( ) const

inline

Whether this layer has trainable parameters.

Returns: True if gradients should flow through this layer during training.

◆ get_label()

const string & opennn::Layer::get_label ( ) const

inline

Returns the user-assigned label of this layer.

Returns: Reference to the label string (default "my_layer").

◆ get_name()

const string & opennn::Layer::get_name ( ) const

inline

Returns the canonical type name of this layer.

Returns: Reference to the name string (e.g. "Dense", "Convolutional").

◆ get_operators()

virtual vector< Operator * > opennn::Layer::get_operators ( )

inlinevirtual

Returns the operators that compose this layer, if any.

Returns: Vector of non-owning Operator pointers; empty for leaf layers.

Composite layers (e.g. those mixing convolution and activation) override this so that the base forward_propagate() can iterate through them.

Reimplemented in opennn::Addition, opennn::Bounding, opennn::Convolutional, opennn::ConvolutionalRelu, opennn::Dense, opennn::DenseRelu, opennn::Embedding, opennn::Flatten, opennn::MultiHeadAttention, opennn::Normalization3d, opennn::Pooling3d, opennn::Pooling, opennn::Scaling, and opennn::Unscaling.

◆ get_output_activation()

virtual Activation::Function opennn::Layer::get_output_activation ( ) const

inlinevirtual

Activation function fused at the end of this layer, if any.

Returns: Activation::Function::Identity by default; layers like DenseRelu / ConvolutionalRelu override this.

Reimplemented in opennn::Convolutional, opennn::ConvolutionalRelu, opennn::Dense, and opennn::DenseRelu.

◆ get_output_shape()

virtual Shape opennn::Layer::get_output_shape ( ) const

pure virtual

Per-sample output shape of the layer (excluding the batch dimension).

Returns: Shape produced by forward_propagate() for a single sample.

Implemented in opennn::Addition, opennn::Bounding, opennn::Convolutional, opennn::ConvolutionalRelu, opennn::Dense, opennn::DenseRelu, opennn::Embedding, opennn::Flatten, opennn::MultiHeadAttention, opennn::Normalization3d, opennn::Pooling3d, opennn::Pooling, opennn::Recurrent, opennn::Scaling, and opennn::Unscaling.

◆ get_outputs_number()

Index opennn::Layer::get_outputs_number ( ) const

inline

Total number of scalar outputs per sample (product of output dims).

Returns: Flat output size.

◆ get_parameter_dtypes()

vector< Type > opennn::Layer::get_parameter_dtypes ( ) const

inline

Dtype-only view of get_parameter_specs().

◆ get_parameter_shapes()

vector< Shape > opennn::Layer::get_parameter_shapes ( ) const

inline

Shape-only view of get_parameter_specs().

◆ get_parameter_specs()

virtual vector< pair< Shape, Type > > opennn::Layer::get_parameter_specs ( ) const

virtual

Specifications of the trainable parameter tensors owned by this layer.

Returns: Vector of (Shape, Type) pairs, one per parameter tensor.

Reimplemented in opennn::Recurrent.

◆ get_parameter_views() [1/2]

vector< TensorView > & opennn::Layer::get_parameter_views ( )

inline

Mutable access to this layer's parameter TensorViews.

◆ get_parameter_views() [2/2]

const vector< TensorView > & opennn::Layer::get_parameter_views ( ) const

inline

Read-only access to this layer's parameter TensorViews.

◆ get_parameters_number()

Index opennn::Layer::get_parameters_number ( ) const

Total number of trainable parameters in this layer.

Returns: Sum of the sizes of all parameter TensorViews.

◆ get_state_shapes()

vector< Shape > opennn::Layer::get_state_shapes ( ) const

inline

Shape-only view of get_state_specs().

◆ get_state_specs()

virtual vector< pair< Shape, Type > > opennn::Layer::get_state_specs ( ) const

virtual

Specifications of the persistent state tensors of this layer.

Returns: Vector of (Shape, Type) pairs, one per state tensor (e.g. running statistics for batch normalization).

◆ get_state_views() [1/2]

vector< TensorView > & opennn::Layer::get_state_views ( )

inline

Mutable access to this layer's state TensorViews.

◆ get_state_views() [2/2]

const vector< TensorView > & opennn::Layer::get_state_views ( ) const

inline

Read-only access to this layer's state TensorViews.

◆ get_type()

LayerType opennn::Layer::get_type ( ) const

inline

Returns the LayerType enumerator for this layer.

Returns: Layer type tag set by the concrete subclass.

◆ link_parameters()

virtual float * opennn::Layer::link_parameters ( float * pointer )

virtual

Wires this layer's parameter TensorViews onto an external buffer.

Parameters

pointer Start of the parameter region inside the network's parameter arena.

Returns: Pointer advanced past the bytes consumed by this layer.

◆ link_states()

virtual float * opennn::Layer::link_states ( float * pointer )

virtual

Wires this layer's state TensorViews onto an external buffer.

Parameters

pointer Start of the state region inside the network's state arena.

Returns: Pointer advanced past the bytes consumed by this layer.

◆ link_views()

float * opennn::Layer::link_views	(	float *	pointer,
		const vector< Shape > &	shapes,
		vector< TensorView > &	views,
		const char *	tag ) const

protected

Builds views over a contiguous float buffer using shapes.

Parameters

pointer	Start of the buffer to slice.
shapes	Per-view shapes.
views	Output vector of TensorViews, populated in order.
tag	Diagnostic label used in error messages.

Returns: Pointer advanced past the bytes covered by shapes.

◆ load_state_from_JSON()

virtual void opennn::Layer::load_state_from_JSON ( const JsonDocument & document )

virtual

Loads parameter and state tensors from a JSON document.

Parameters

document Parsed JSON document containing tensor data.

◆ on_compute_dtype_changed()

virtual void opennn::Layer::on_compute_dtype_changed ( )

inlinevirtual

Hook invoked after set_compute_dtype() mutates the dtype.

Subclasses override this to reconfigure cached dtype-dependent resources (CUDA descriptors, cached scratch buffers, etc.).

Reimplemented in opennn::Convolutional, opennn::ConvolutionalRelu, opennn::Dense, opennn::DenseRelu, and opennn::MultiHeadAttention.

◆ print()

virtual void opennn::Layer::print ( ) const

inlinevirtual

Prints a human-readable summary of the layer to stdout.

Reimplemented in opennn::Unscaling.

◆ read_JSON_body()

virtual void opennn::Layer::read_JSON_body ( const Json * )

inlinevirtual

Subclass hook for parsing the body of from_JSON().

Receives the raw JSON node for this layer. Overridden by concrete layers to read their specific fields after the base class has read the common ones.

Reimplemented in opennn::Bounding, opennn::Convolutional, opennn::ConvolutionalRelu, opennn::Dense, opennn::Embedding, opennn::MultiHeadAttention, opennn::Normalization3d, opennn::Pooling3d, opennn::Pooling, opennn::Recurrent, opennn::Scaling, and opennn::Unscaling.

◆ redistribute_parameter_gradients_to_operators()

void opennn::Layer::redistribute_parameter_gradients_to_operators ( vector< TensorView > & gradient_views )

inline

Forwards externally provided gradient views down to each Operator.

Parameters

gradient_views Per-parameter gradient TensorViews aligned with this layer's parameter specs.

◆ redistribute_parameters_to_operators()

void opennn::Layer::redistribute_parameters_to_operators ( )

inline

Forwards the current parameter views down to each composing Operator.

Composite layers slice their shared parameter buffer and hand sub-views to their Operators so the operators can run independently.

◆ redistribute_states_to_operators()

void opennn::Layer::redistribute_states_to_operators ( )

inline

Forwards the current state views down to each composing Operator.

◆ set_compute_dtype()

void opennn::Layer::set_compute_dtype ( Type new_compute_dtype )

inline

Sets the compute dtype and triggers on_compute_dtype_changed().

Parameters

new_compute_dtype Desired compute dtype.

◆ set_input_shape()

virtual void opennn::Layer::set_input_shape ( const Shape & )

virtual

Sets the per-sample input shape of this layer.

The argument is the new input shape, excluding the batch dimension. Subclasses override this to also resize the output shape and any shape-dependent parameters.

Reimplemented in opennn::Addition, opennn::Bounding, opennn::Convolutional, opennn::ConvolutionalRelu, opennn::Dense, opennn::DenseRelu, opennn::Flatten, opennn::MultiHeadAttention, opennn::Normalization3d, opennn::Pooling3d, opennn::Pooling, opennn::Recurrent, opennn::Scaling, and opennn::Unscaling.

◆ set_label()

void opennn::Layer::set_label ( string new_label )

inline

Sets the human-readable label of this layer.

Parameters

new_label Label moved into the layer.

◆ set_output_shape()

virtual void opennn::Layer::set_output_shape ( const Shape & )

virtual

Sets the per-sample output shape of this layer.

The argument is the new output shape, excluding the batch dimension. Subclasses override this to also resize parameters whose dimensions depend on the output shape.

Reimplemented in opennn::Bounding, opennn::Dense, opennn::DenseRelu, opennn::Recurrent, opennn::Scaling, and opennn::Unscaling.

◆ to_JSON()

virtual void opennn::Layer::to_JSON ( JsonWriter & writer ) const

virtual

Writes the layer configuration to JSON.

Parameters

writer Streaming JSON writer.

◆ write_JSON_body()

virtual void opennn::Layer::write_JSON_body ( JsonWriter & ) const

inlinevirtual

Subclass hook for emitting the body of to_JSON().

Receives the streaming JSON writer; overridden by concrete layers to emit their subclass-specific fields.

Reimplemented in opennn::Bounding, opennn::Convolutional, opennn::ConvolutionalRelu, opennn::Embedding, opennn::MultiHeadAttention, opennn::Pooling3d, opennn::Pooling, opennn::Recurrent, opennn::Scaling, and opennn::Unscaling.

Member Data Documentation

◆ compute_dtype

Type opennn::Layer::compute_dtype = Type::FP32

protected

Numerical type used for forward and backward computation.

◆ is_first_layer

bool opennn::Layer::is_first_layer = false

protected

True if this layer is the network's input layer.

◆ is_trainable

bool opennn::Layer::is_trainable = true

protected

True if the layer has parameters that participate in training.

◆ label

string opennn::Layer::label = "my_layer"

protected

User-visible label for this layer instance (default "my_layer").

◆ layer_type

LayerType opennn::Layer::layer_type = LayerType::Dense

protected

Layer type tag set by the subclass.

◆ layers

vector<unique_ptr<Layer> > opennn::Layer::layers

protected

Sub-layers, when this layer is itself a composite.

◆ name

string opennn::Layer::name = "layer"

protected

Canonical type name set by the subclass (e.g. "dense").

◆ parameters

vector<TensorView> opennn::Layer::parameters

protected

Parameter TensorViews bound to the network's parameter arena.

◆ states

vector<TensorView> opennn::Layer::states

protected

State TensorViews bound to the network's state arena.

Public Member Functions

Protected Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ ~Layer()

◆ Layer()

Member Function Documentation

◆ back_propagate()

◆ distribute_to_operators()

◆ forward_propagate()

◆ from_JSON()

◆ get_backward_dtypes()

◆ get_backward_shapes()

◆ get_backward_specs()

◆ get_compute_dtype()

◆ get_forward_dtypes()

◆ get_forward_shapes()

◆ get_forward_specs()

◆ get_input_shape()

◆ get_inputs_number()

◆ get_is_trainable()

◆ get_label()

◆ get_name()

◆ get_operators()

◆ get_output_activation()

◆ get_output_shape()

◆ get_outputs_number()

◆ get_parameter_dtypes()

◆ get_parameter_shapes()

◆ get_parameter_specs()

◆ get_parameter_views() [1/2]

◆ get_parameter_views() [2/2]

◆ get_parameters_number()

◆ get_state_shapes()

◆ get_state_specs()

◆ get_state_views() [1/2]

◆ get_state_views() [2/2]

◆ get_type()

◆ link_parameters()

◆ link_states()

◆ link_views()

◆ load_state_from_JSON()

◆ on_compute_dtype_changed()

◆ print()

◆ read_JSON_body()

◆ redistribute_parameter_gradients_to_operators()

◆ redistribute_parameters_to_operators()

◆ redistribute_states_to_operators()

◆ set_compute_dtype()

◆ set_input_shape()

◆ set_label()

◆ set_output_shape()

◆ to_JSON()

◆ write_JSON_body()

Member Data Documentation

◆ compute_dtype

◆ is_first_layer

◆ is_trainable

◆ label

◆ layer_type

◆ layers

◆ name

◆ parameters

◆ states