Reshape layer that collapses every input axis into a single feature axis. More...

#include <flatten_layer.h>

Inheritance diagram for opennn::Flatten:

Public Member Functions
	Flatten (const Shape &input_shape={})
	Constructs a Flatten layer.

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

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

vector< Operator * >	get_operators () override
	Returns the single Flat operator that implements this layer.

void	set (const Shape &)
	Re-initializes the layer.

void	set_input_shape (const Shape &new_input_shape) override
	Updates the input shape; equivalent to calling set().

void	back_propagate (ForwardPropagation &, BackPropagation &, size_t) const noexcept override
	Backward pass: reshapes the output gradient back to the input shape.

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.

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< 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 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	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.

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

Reshape layer that collapses every input axis into a single feature axis.

Given an input of any rank, the output is a 1D tensor of length equal to the product of the input dimensions. Forward and backward passes are pure reshapes (no element movement); the layer has no trainable parameters.

Typically inserted between convolutional / pooling layers and dense layers in CNN classifiers.

Constructor & Destructor Documentation

◆ Flatten()

opennn::Flatten::Flatten ( const Shape & input_shape = {} )

Constructs a Flatten layer.

Parameters

input_shape Per-sample input shape; the output length is its flat size.

Member Function Documentation

◆ back_propagate()

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

overridevirtualnoexcept

Backward pass: reshapes the output gradient back to the input shape.

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

Reimplemented from opennn::Layer.

◆ get_input_shape()

Shape opennn::Flatten::get_input_shape ( ) const

inlineoverridevirtual

Returns the per-sample input shape.

Implements opennn::Layer.

◆ get_operators()

vector< Operator * > opennn::Flatten::get_operators ( )

inlineoverridevirtual

Returns the single Flat operator that implements this layer.

Reimplemented from opennn::Layer.

◆ get_output_shape()

Shape opennn::Flatten::get_output_shape ( ) const

inlineoverridevirtual

Returns the per-sample output shape.

Returns: 1D shape with length equal to the flat size of the input.

Implements opennn::Layer.

◆ set()

void opennn::Flatten::set ( const Shape & )

Re-initializes the layer.

Receives the new per-sample input shape.

◆ set_input_shape()

void opennn::Flatten::set_input_shape ( const Shape & new_input_shape )

inlineoverridevirtual

Updates the input shape; equivalent to calling set().

Parameters

new_input_shape New per-sample input shape.

Reimplemented from opennn::Layer.

Public Member Functions

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ Flatten()

Member Function Documentation

◆ back_propagate()

◆ get_input_shape()

◆ get_operators()

◆ get_output_shape()

◆ set()

◆ set_input_shape()