Plain (Elman-style) recurrent layer over fixed-length sequences. More...

#include <recurrent_layer.h>

Inheritance diagram for opennn::Recurrent:

Public Member Functions
	Recurrent (const Shape &input_shape={0, 0}, const Shape &output_shape={0})
	Constructs a Recurrent layer.

Shape	get_input_shape () const override
	Returns the per-sample input shape (time_steps, input_features).

Shape	get_output_shape () const override
	Returns the per-sample output shape.

const TensorView &	get_biases () const
	Read-only access to the bias TensorView.

const TensorView &	get_input_weights () const
	Read-only access to the input-to-hidden weight TensorView.

const TensorView &	get_recurrent_weights () const
	Read-only access to the hidden-to-hidden weight TensorView.

const string &	get_activation_function () const
	Name of the activation function applied to the hidden state.

vector< pair< Shape, Type > >	get_parameter_specs () const override
	Specifications of the parameter tensors.

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

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

void	set (const Shape &input_shape={}, const Shape &output_shape={})
	Re-initializes the layer.

void	set_input_shape (const Shape &) override
	Updates the input shape and re-shapes weight tensors accordingly.

void	set_output_shape (const Shape &) override
	Updates the output shape and re-shapes weight tensors accordingly.

void	set_activation_function (const string &)
	Sets the activation function applied to the hidden state.

void	forward_propagate (ForwardPropagation &, size_t, bool) noexcept override
	Forward pass: unrolls the recurrence over time_steps steps.

void	back_propagate (ForwardPropagation &, BackPropagation &, size_t) const noexcept override
	Backward pass through time (BPTT) over the unrolled recurrence.

void	read_JSON_body (const Json *) override
	Reads the layer-specific JSON body (shapes, activation name).

void	write_JSON_body (JsonWriter &) const override
	Writes the layer-specific JSON body (shapes, activation name).

Public Member Functions inherited from opennn::Layer
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.

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_state_specs () const
	Specifications of the persistent state tensors of this layer.

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 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	from_JSON (const JsonDocument &document)
	Loads the layer configuration (hyperparameters) 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	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.

Additional Inherited Members
Protected Member Functions inherited from opennn::Layer
	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 inherited from opennn::Layer
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

Plain (Elman-style) recurrent layer over fixed-length sequences.

Inputs are rank-2 tensors (time_steps, input_features). At each step the hidden state is updated as h_t = activation(W_x * x_t + W_h * h_{t-1} + b), with shared weights and biases across all time steps.

The layer owns three parameter tensors (input weights, recurrent weights, biases) and applies a configurable elementwise activation.

Constructor & Destructor Documentation

◆ Recurrent()

opennn::Recurrent::Recurrent	(	const Shape &	input_shape = {0, 0},
		const Shape &	output_shape = {0} )

Constructs a Recurrent layer.

Parameters

input_shape	Per-sample input shape (time_steps, input_features).
output_shape	Per-sample output shape (number of hidden units).

Member Function Documentation

◆ back_propagate()

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

overridevirtualnoexcept

Backward pass through time (BPTT) over the unrolled recurrence.

Receives the forward intermediates, the BackPropagation buffer and this layer's index inside the network.

Reimplemented from opennn::Layer.

◆ forward_propagate()

void opennn::Recurrent::forward_propagate	(	ForwardPropagation &	,
		size_t	,
		bool	)

overridevirtualnoexcept

Forward pass: unrolls the recurrence over time_steps steps.

Receives the ForwardPropagation buffer slice, this layer's index and the training flag.

Reimplemented from opennn::Layer.

◆ get_activation_function()

const string & opennn::Recurrent::get_activation_function ( ) const

inline

Name of the activation function applied to the hidden state.

◆ get_backward_specs()

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

overridevirtual

Specifications of the backward intermediate buffers.

Parameters

batch_size Batch size used for sizing.

Returns: Buffers used for time-unrolled gradient accumulation.

Reimplemented from opennn::Layer.

◆ get_biases()

const TensorView & opennn::Recurrent::get_biases ( ) const

inline

Read-only access to the bias TensorView.

◆ get_forward_specs()

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

overridevirtual

Specifications of the forward intermediate buffers.

Parameters

batch_size Batch size used for sizing.

Returns: Per-step combination and activation buffers plus the output.

Reimplemented from opennn::Layer.

◆ get_input_shape()

Shape opennn::Recurrent::get_input_shape ( ) const

inlineoverridevirtual

Returns the per-sample input shape (time_steps, input_features).

Implements opennn::Layer.

◆ get_input_weights()

const TensorView & opennn::Recurrent::get_input_weights ( ) const

inline

Read-only access to the input-to-hidden weight TensorView.

◆ get_output_shape()

Shape opennn::Recurrent::get_output_shape ( ) const

overridevirtual

Returns the per-sample output shape.

Returns: Shape with the number of hidden units.

Implements opennn::Layer.

◆ get_parameter_specs()

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

overridevirtual

Specifications of the parameter tensors.

Returns: Specs for input weights, recurrent weights and biases.

Reimplemented from opennn::Layer.

◆ get_recurrent_weights()

const TensorView & opennn::Recurrent::get_recurrent_weights ( ) const

inline

Read-only access to the hidden-to-hidden weight TensorView.

◆ read_JSON_body()

void opennn::Recurrent::read_JSON_body ( const Json * )

overridevirtual

Reads the layer-specific JSON body (shapes, activation name).

Reimplemented from opennn::Layer.

◆ set()

void opennn::Recurrent::set	(	const Shape &	input_shape = {},
		const Shape &	output_shape = {} )

Re-initializes the layer.

Parameters

input_shape	Per-sample input shape (time_steps, input_features).
output_shape	Per-sample output shape (hidden units).

◆ set_activation_function()

void opennn::Recurrent::set_activation_function ( const string & )

Sets the activation function applied to the hidden state.

Receives the activation name; see Activation::Function for the supported set.

◆ set_input_shape()

void opennn::Recurrent::set_input_shape ( const Shape & )

overridevirtual

Updates the input shape and re-shapes weight tensors accordingly.

Reimplemented from opennn::Layer.

◆ set_output_shape()

void opennn::Recurrent::set_output_shape ( const Shape & )

overridevirtual

Updates the output shape and re-shapes weight tensors accordingly.

Reimplemented from opennn::Layer.

◆ write_JSON_body()

void opennn::Recurrent::write_JSON_body ( JsonWriter & ) const

overridevirtual

Writes the layer-specific JSON body (shapes, activation name).

Reimplemented from opennn::Layer.

Public Member Functions

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ Recurrent()

Member Function Documentation

◆ back_propagate()

◆ forward_propagate()

◆ get_activation_function()

◆ get_backward_specs()

◆ get_biases()

◆ get_forward_specs()

◆ get_input_shape()

◆ get_input_weights()

◆ get_output_shape()

◆ get_parameter_specs()

◆ get_recurrent_weights()

◆ read_JSON_body()

◆ set()

◆ set_activation_function()

◆ set_input_shape()

◆ set_output_shape()

◆ write_JSON_body()