dcnnsol::nonstrided_conv_layer_base< SPI, SPO, KSP, PP > Struct Template Reference

Utility base class for the convolutional layer. More...

#include <dcnnelements.hpp>

Inheritance diagram for dcnnsol::nonstrided_conv_layer_base< SPI, SPO, KSP, PP >:

Public Types
using	CSPI = typename SPI::channel_policy
using	CSPO = typename SPO::channel_policy
Data classes
using	input_data = dcnnsol::image_data<SPI, PP>
	Input activations.
using	weights = dcnnsol::conv_weights<KSP, CSPI, CSPO, PP>
	Model weights.
using	output_data = dcnnsol::image_data<SPO, PP>
	Output activations.
Dimension tags
using	height_in_tag = typename SPI::height_tag
	Input image height.
using	width_in_tag = typename SPI::width_tag
	Input image width.
using	channel_in_tag = typename SPI::channel_tag
	Input image channel.
using	kernel_height_tag = typename KSP::kernel_height_tag
	Kernel height.
using	kernel_width_tag = typename KSP::kernel_width_tag
	Kernel width.
using	height_out_tag = typename SPO::height_tag
	Output image height.
using	width_out_tag = typename SPO::width_tag
	Output image width.
using	channel_out_tag = typename SPO::channel_tag
	Output image channel.

Static Public Member Functions
Check functions
static tagged::range_class< batch_tag >	forward_check (const input_data &ind, const weights &wtd, output_data &outd)
	Check argument compatibility for the forward function.

Static Public Attributes
static constexpr std::size_t	fanin = khr.size() * kwr.size() * cir.size()
	Number of inputs for each output element.
Dimension ranges
Each dimension range allows enumeration of all indexes in that dimension, using the for(:) loop
static constexpr auto	hir = SPI::hr
	Input image height range.
static constexpr auto	wir = SPI::wr
	Input image width range.
static constexpr auto	cir = SPI::cr
	Input image channel range.
static constexpr auto	khr = KSP::khr
	Kernel height range.
static constexpr auto	kwr = KSP::kwr
	Kernel width range.
static constexpr auto	hor = SPO::hr
	Output image height range.
static constexpr auto	wor = SPO::wr
	Output image width range.
static constexpr auto	cor = SPO::cr
	Output image channel range.
Input-kernel-output index convertors
Each quadruplet of functors serves for a different order of loop nesting, i.e. loops like for ( auto top_index_value : top_index_range ) { auto mapping_functor = mf(top_index_value) for ( auto mid_index_value : rf(top_index_value) ) { auto bottom_index_value = mapping_functor(mid_index_value); /*...*/ } } Each -rf functor computes a mid-index range from a top-index value. The purpose is to handle clipping of valid index ranges at the borders of the images because output images have the same dimensions as the input images despite being generated by convolution from neighboring values. Thus, the available kernel-index range must be clipped for an output or input index at a border and vice versa. Each -mf functor produces another functor from a top-index value that converts a mid-index value to a bottom-index value. The produced functor handles the required shifts (and sign changes) that result from the top-index value. In addition, it also handles the fact that kernels are indexed by [0,3) instead of [-1,1]. h- and w- functors work along image heights and widths, respectively.
static constexpr conv_ok_range_functor< tagged::selector< kernel_height_tag >, tagged::selector< height_out_tag >, hir.size(), khr.size(), 1 >	hokrf {}
	output-height index to kernel-height range
static constexpr conv_oki_map_functor< tagged::selector< height_in_tag >, tagged::selector< kernel_height_tag >, tagged::selector< height_out_tag >, khr.size(), 1 >	hokimf {}
	output-height index to kernel-height index to input-height index
static constexpr conv_ok_range_functor< tagged::selector< kernel_width_tag >, tagged::selector< width_out_tag >, wir.size(), kwr.size(), 1 >	wokrf {}
	output-width index to kernel-width range
static constexpr conv_oki_map_functor< tagged::selector< width_in_tag >, tagged::selector< kernel_width_tag >, tagged::selector< width_out_tag >, kwr.size(), 1 >	wokimf {}
	output-width index to kernel-width index to input-width index
static constexpr conv_ik_range_functor< tagged::selector< height_in_tag >, tagged::selector< kernel_height_tag >, hir.size(), khr.size(), 1 >	hikrf {}
	input-height index to kernel-height range
static constexpr conv_iko_map_functor< tagged::selector< height_in_tag >, tagged::selector< kernel_height_tag >, tagged::selector< height_out_tag >, khr.size(), 1 >	hikomf {}
	input-height index to kernel-height index to output-height index
static constexpr conv_ik_range_functor< tagged::selector< width_in_tag >, tagged::selector< kernel_width_tag >, wir.size(), kwr.size(), 1 >	wikrf {}
	input-width index to kernel-width range
static constexpr conv_iko_map_functor< tagged::selector< width_in_tag >, tagged::selector< kernel_width_tag >, tagged::selector< width_out_tag >, kwr.size(), 1 >	wikomf {}
	input-width index to kernel-width index to output-width index
static constexpr conv_ko_range_functor< tagged::selector< kernel_height_tag >, tagged::selector< height_out_tag >, hir.size(), khr.size(), 1 >	hkorf {}
	kernel-height index to output-height range
static constexpr conv_koi_map_functor< tagged::selector< kernel_height_tag >, tagged::selector< height_out_tag >, tagged::selector< height_in_tag >, khr.size(), 1 >	hkoimf {}
	kernel-height index to output-height index to input-height index
static constexpr conv_ko_range_functor< tagged::selector< kernel_width_tag >, tagged::selector< width_out_tag >, wir.size(), kwr.size(), 1 >	wkorf {}
	kernel-width index to output-width range
static constexpr conv_koi_map_functor< tagged::selector< kernel_width_tag >, tagged::selector< width_out_tag >, tagged::selector< width_in_tag >, kwr.size(), 1 >	wkoimf {}
	kernel-width index to output-width index to input-width index

Detailed Description

template<typename SPI, typename SPO, typename KSP, is_policy PP>
struct dcnnsol::nonstrided_conv_layer_base< SPI, SPO, KSP, PP >

Utility base class for the convolutional layer.

Template Parameters

SPI	Input size policy
SPO	Output size policy
KSP	Kernel size policy
PP	Permutation policy

Member Function Documentation

forward_check()

template<typename SPI, typename SPO, typename KSP, is_policy PP>

static tagged::range_class< batch_tag > dcnnasgn::nonstrided_conv_layer_base< SPI, SPO, KSP, PP >::forward_check ( const input_data & ind, const weights & wtd, output_data & outd )

inlinestatic

Check argument compatibility for the forward function.

Parameters

ind	Input activations
wtd	Weights
outd	Output activations

Returns

The minibatch range

---

The documentation for this struct was generated from the following file:

• asgn/dcnnelements.hpp