2D convolution operator (NHWC layout) backed by Eigen on CPU and cuDNN on GPU. More...

#include <operators.h>

Inheritance diagram for opennn::ConvolutionOp:

Public Member Functions
void	set (Index input_h, Index input_w, Index kernels_n, Index kernel_h, Index kernel_w, Index kernel_c, Index row_stride, Index column_stride, Index padding_h, Index padding_w, Type compute_dtype)
	Configures the convolution geometry and compute precision.

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	destroy_cuda () override
	Releases CUDA resources owned by the operator; called from destructors.

	~ConvolutionOp () override

	ConvolutionOp ()=default

	ConvolutionOp (const ConvolutionOp &)=delete

ConvolutionOp &	operator= (const ConvolutionOp &)=delete

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.

void	apply_cpu (const TensorView &input, TensorView &output)
	CPU forward path; im2col + GEMM.

void	apply_gpu (const TensorView &input, TensorView &output, cudnnActivationDescriptor_t fused_activation=nullptr)
	GPU forward path; runs cuDNN convolution with an optional fused activation.

void	apply_delta (const TensorView &input, const TensorView &output_delta, TensorView &input_delta) const
	Computes input_delta from output_delta and updates weight/bias gradients.

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.

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	input_height = 0

Index	input_width = 0

Index	kernels_number = 0

Index	kernel_height = 0

Index	kernel_width = 0

Index	kernel_channels = 0

Index	padding_height = 0

Index	padding_width = 0

Type	compute_dtype = Type::FP32

TensorView	weights

TensorView	bias

TensorView	weight_gradient

TensorView	bias_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

2D convolution operator (NHWC layout) backed by Eigen on CPU and cuDNN on GPU.

Constructor & Destructor Documentation

◆ ~ConvolutionOp()

opennn::ConvolutionOp::~ConvolutionOp ( )

inlineoverride

◆ ConvolutionOp() [1/2]

opennn::ConvolutionOp::ConvolutionOp ( )

default

◆ ConvolutionOp() [2/2]

opennn::ConvolutionOp::ConvolutionOp ( const ConvolutionOp & )

delete

Member Function Documentation

◆ apply_cpu()

void opennn::ConvolutionOp::apply_cpu	(	const TensorView &	input,
		TensorView &	output )

CPU forward path; im2col + GEMM.

◆ apply_delta()

void opennn::ConvolutionOp::apply_delta	(	const TensorView &	input,
		const TensorView &	output_delta,
		TensorView &	input_delta ) const

Computes input_delta from output_delta and updates weight/bias gradients.

◆ apply_gpu()

void opennn::ConvolutionOp::apply_gpu	(	const TensorView &	input,
		TensorView &	output,
		cudnnActivationDescriptor_t	fused_activation = nullptr )

GPU forward path; runs cuDNN convolution with an optional fused activation.

◆ back_propagate()

void opennn::ConvolutionOp::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.

◆ destroy_cuda()

void opennn::ConvolutionOp::destroy_cuda ( )

overridevirtual

Releases CUDA resources owned by the operator; called from destructors.

Reimplemented from opennn::Operator.

◆ forward_propagate()

void opennn::ConvolutionOp::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.

◆ link_gradients()

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

overridevirtual

Binds gradient views provided by the hosting layer.

Reimplemented from opennn::Operator.

◆ link_parameters()

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

overridevirtual

Binds parameter views provided by the hosting layer.

Reimplemented from opennn::Operator.

◆ operator=()

ConvolutionOp & opennn::ConvolutionOp::operator= ( const ConvolutionOp & )

delete

◆ parameter_specs()

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

overridevirtual

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

Reimplemented from opennn::Operator.

◆ set()

void opennn::ConvolutionOp::set	(	Index	input_h,
		Index	input_w,
		Index	kernels_n,
		Index	kernel_h,
		Index	kernel_w,
		Index	kernel_c,
		Index	row_stride,
		Index	column_stride,
		Index	padding_h,
		Index	padding_w,
		Type	compute_dtype )

Configures the convolution geometry and compute precision.

Parameters

input_h	Input height in pixels.
input_w	Input width in pixels.
kernels_n	Number of output channels (kernels).
kernel_h	Kernel height.
kernel_w	Kernel width.
kernel_c	Kernel channel count (matches input channels).
row_stride	Vertical stride.
column_stride	Horizontal stride.
padding_h	Vertical zero-padding on each side.
padding_w	Horizontal zero-padding on each side.
compute_dtype	Dtype used for the matmul computation.

◆ set_parameters_glorot()

void opennn::ConvolutionOp::set_parameters_glorot ( )

overridevirtual

Initializes parameters using Glorot (Xavier) initialization.

Reimplemented from opennn::Operator.

◆ set_parameters_random()

void opennn::ConvolutionOp::set_parameters_random ( )

overridevirtual

Initializes parameters with random values.

Reimplemented from opennn::Operator.

Member Data Documentation

◆ bias

TensorView opennn::ConvolutionOp::bias

◆ bias_gradient

TensorView opennn::ConvolutionOp::bias_gradient

◆ compute_dtype

Type opennn::ConvolutionOp::compute_dtype = Type::FP32

◆ input_height

Index opennn::ConvolutionOp::input_height = 0

◆ input_width

Index opennn::ConvolutionOp::input_width = 0

◆ kernel_channels

Index opennn::ConvolutionOp::kernel_channels = 0

◆ kernel_height

Index opennn::ConvolutionOp::kernel_height = 0

◆ kernel_width

Index opennn::ConvolutionOp::kernel_width = 0

◆ kernels_number

Index opennn::ConvolutionOp::kernels_number = 0

◆ padding_height

Index opennn::ConvolutionOp::padding_height = 0

◆ padding_width

Index opennn::ConvolutionOp::padding_width = 0

◆ weight_gradient

TensorView opennn::ConvolutionOp::weight_gradient

◆ weights

TensorView opennn::ConvolutionOp::weights

Public Member Functions

Public Attributes

Detailed Description

Constructor & Destructor Documentation

◆ ~ConvolutionOp()

◆ ConvolutionOp() [1/2]

◆ ConvolutionOp() [2/2]

Member Function Documentation

◆ apply_cpu()

◆ apply_delta()

◆ apply_gpu()

◆ back_propagate()

◆ destroy_cuda()

◆ forward_propagate()

◆ link_gradients()

◆ link_parameters()

◆ operator=()

◆ parameter_specs()

◆ set()

◆ set_parameters_glorot()

◆ set_parameters_random()

Member Data Documentation

◆ bias

◆ bias_gradient

◆ compute_dtype

◆ input_height

◆ input_width

◆ kernel_channels

◆ kernel_height

◆ kernel_width

◆ kernels_number

◆ padding_height

◆ padding_width

◆ weight_gradient

◆ weights