GCC Code Coverage Report

Line	Branch	Exec	Source
1			module athena__conv2d_layer
2			!! Module containing implementation of a 2D convolutional layer
3			!!
4			!! This module implements 2D convolution for processing image-like data.
5			!! Applies learnable filters to extract spatial features.
6			!!
7			!! Mathematical operation:
8			!! \[ y_{i,j,k} =
9			!! \sigma\left(
10			!! \sum_{c=1}^{C_{in}}
11			!! \sum_{m=0}^{K_h-1}
12			!! \sum_{n=0}^{K_w-1}
13			!! x_{i+m,j+n,c} \cdot w_{m,n,c,k} + b_k
14			!! \right)
15			!! \]
16			!!
17			!! where:
18			!! - \((i,j)\) are spatial coordinates in the output
19			!! - \(k\) is the output channel (filter) index
20			!! - \((m,n)\) are kernel offsets
21			!! - \(c\) is the input channel index
22			!! - \(K_h, K_w\) are kernel dimensions
23			!! - \(\sigma\) is the activation function
24			!!
25			!! Shape: \((W, H, C_{in}) \rightarrow (W', H', C_{out})\)
26			use coreutils, only: real32, stop_program
27			use athena__base_layer, only: conv_layer_type, base_layer_type
28			use athena__pad2d_layer, only: pad2d_layer_type
29			use athena__misc_types, only: base_init_type, &
30			onnx_node_type, onnx_initialiser_type, onnx_tensor_type
31			use athena__misc_types, only: base_actv_type, base_init_type
32			use diffstruc, only: array_type
33			use athena__diffstruc_extd, only: conv2d, add_bias
34			use athena__initialiser_data, only: data_init_type
35			implicit none
36
37
38			private
39
40			public :: conv2d_layer_type
41			public :: read_conv2d_layer
42
43
44			type, extends(conv_layer_type) :: conv2d_layer_type
45			!! Type for 2D convolutional layer with overloaded procedures
46			type(array_type), dimension(2) :: z
47			!! Temporary arrays for forward propagation
48			contains
49			procedure, pass(this) :: set_hyperparams => set_hyperparams_conv2d
50			!! Set hyperparameters for 2D convolutional layer
51			procedure, pass(this) :: read => read_conv2d
52			!! Read 2D convolutional layer from file
53			procedure, pass(this) :: build_from_onnx => build_from_onnx_conv2d
54			!! Build 2D convolutional layer from ONNX node and initialiser
55
56			procedure, pass(this) :: forward => forward_conv2d
57			!! Forward propagation derived type handler
58
59			final :: finalise_conv2d
60			!! Finalise 2D convolutional layer
61			end type conv2d_layer_type
62
63			interface conv2d_layer_type
64			!! Interface for setting up the 2D convolutional layer
65			module function layer_setup( &
66			input_shape, &
67			num_filters, kernel_size, stride, dilation, padding, &
68			use_bias, &
69			activation, &
70			kernel_initialiser, bias_initialiser, &
71			verbose ) result(layer)
72			!! Set up the 2D convolutional layer
73			integer, dimension(:), optional, intent(in) :: input_shape
74			!! Input shape
75			integer, optional, intent(in) :: num_filters
76			!! Number of filters
77			integer, dimension(..), optional, intent(in) :: kernel_size
78			!! Kernel size
79			integer, dimension(..), optional, intent(in) :: stride
80			!! Stride
81			integer, dimension(..), optional, intent(in) :: dilation
82			!! Dilation
83			logical, optional, intent(in) :: use_bias
84			!! Use bias
85			class(*), optional, intent(in) :: activation, &
86			kernel_initialiser, bias_initialiser
87			!! Activation function, kernel initialiser, bias initialiser
88			character(*), optional, intent(in) :: padding
89			!! Padding method
90			integer, optional, intent(in) :: verbose
91			!! Verbosity level
92			type(conv2d_layer_type) :: layer
93			!! Instance of the 2D convolutional layer
94			end function layer_setup
95			end interface conv2d_layer_type
96
97
98
99			contains
100
101			!###############################################################################
102			!!! finalise layer
103			!###############################################################################
104		−	subroutine finalise_conv2d(this)
105			!! Finalise 2D convolutional layer
106			implicit none
107
108			! Arguments
109			type(conv2d_layer_type), intent(inout) :: this
110			!! Instance of the 2D convolutional layer
111
112		−	if(allocated(this%dil)) deallocate(this%dil)
113		−	if(allocated(this%knl)) deallocate(this%knl)
114		−	if(allocated(this%stp)) deallocate(this%stp)
115
116		−	if(allocated(this%input_shape)) deallocate(this%input_shape)
117		−	if(allocated(this%output)) deallocate(this%output)
118		−	if(allocated(this%pad_layer)) deallocate(this%pad_layer)
119
120		−	end subroutine finalise_conv2d
121			!###############################################################################
122
123
124			!##############################################################################!
125			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
126			!##############################################################################!
127
128
129			!###############################################################################
130		−	module function layer_setup( &
131		−	input_shape, &
132			num_filters, kernel_size, stride, dilation, padding, &
133			use_bias, &
134			activation, &
135			kernel_initialiser, bias_initialiser, &
136		−	verbose ) result(layer)
137			!! Set up the 2D convolutional layer
138			use athena__activation, only: activation_setup
139			use athena__initialiser, only: initialiser_setup
140			implicit none
141
142			! Arguments
143			integer, dimension(:), optional, intent(in) :: input_shape
144			!! Input shape
145			integer, optional, intent(in) :: num_filters
146			!! Number of filters
147			integer, dimension(..), optional, intent(in) :: kernel_size
148			!! Kernel size
149			integer, dimension(..), optional, intent(in) :: stride
150			!! Stride
151			integer, dimension(..), optional, intent(in) :: dilation
152			!! Dilation
153			logical, optional, intent(in) :: use_bias
154			!! Use bias
155			class(*), optional, intent(in) :: activation
156			!! Activation function
157			class(*), optional, intent(in) :: kernel_initialiser, bias_initialiser
158			!! Activation function, kernel initialiser, and bias initialiser
159			character(*), optional, intent(in) :: padding
160			!! Padding method
161			integer, optional, intent(in) :: verbose
162			!! Verbosity level
163
164			type(conv2d_layer_type) :: layer
165			!! Instance of the 2D convolutional layer
166
167			! Local variables
168			integer :: verbose_ = 0
169			!! Verbosity level
170			integer :: num_filters_
171			!! Number of filters
172			logical :: use_bias_ = .true.
173			!! Use bias
174			character(len=20) :: padding_
175			!! Padding
176			integer, dimension(2) :: kernel_size_, stride_, dilation_
177			!! Kernel size and stride
178		−	class(base_actv_type), allocatable :: activation_
179			!! Activation function
180		−	class(base_init_type), allocatable :: kernel_initialiser_, bias_initialiser_
181			!! Kernel and bias initialisers
182
183		−	if(present(verbose)) verbose_ = verbose
184
185
186			!---------------------------------------------------------------------------
187			! Set use_bias
188			!---------------------------------------------------------------------------
189		−	if(present(use_bias)) use_bias_ = use_bias
190
191
192			!---------------------------------------------------------------------------
193			! Set activation functions based on input name
194			!---------------------------------------------------------------------------
195		−	if(present(activation))then
196		−	activation_ = activation_setup(activation)
197			else
198		−	activation_ = activation_setup("none")
199			end if
200
201
202			!---------------------------------------------------------------------------
203			! Define weights (kernels) and biases initialisers
204			!---------------------------------------------------------------------------
205		−	if(present(kernel_initialiser))then
206		−	kernel_initialiser_ = initialiser_setup(kernel_initialiser)
207			end if
208		−	if(present(bias_initialiser))then
209		−	bias_initialiser_ = initialiser_setup(bias_initialiser)
210			end if
211
212
213			!---------------------------------------------------------------------------
214			! Set up number of filters
215			!---------------------------------------------------------------------------
216		−	if(present(num_filters))then
217		−	num_filters_ = num_filters
218			else
219		−	num_filters_ = 32
220			end if
221
222
223			!---------------------------------------------------------------------------
224			! Set up kernel size
225			!---------------------------------------------------------------------------
226		−	if(present(kernel_size))then
227		−	select rank(kernel_size)
228			rank(0)
229		−	kernel_size_ = kernel_size
230			rank(1)
231		−	kernel_size_(1) = kernel_size(1)
232		−	if(size(kernel_size,dim=1).eq.1)then
233		−	kernel_size_(2) = kernel_size(1)
234		−	elseif(size(kernel_size,dim=1).eq.2)then
235		−	kernel_size_(2) = kernel_size(2)
236			end if
237			end select
238			else
239		−	kernel_size_ = 3
240			end if
241
242
243			!---------------------------------------------------------------------------
244			! Set up padding name
245			!---------------------------------------------------------------------------
246		−	if(present(padding))then
247		−	padding_ = padding
248			else
249		−	padding_ = "valid"
250			end if
251
252
253			!---------------------------------------------------------------------------
254			! Set up stride
255			!---------------------------------------------------------------------------
256		−	if(present(stride))then
257		−	select rank(stride)
258			rank(0)
259		−	stride_ = stride
260			rank(1)
261		−	stride_(1) = stride(1)
262		−	if(size(stride,dim=1).eq.1)then
263		−	stride_(2) = stride(1)
264		−	elseif(size(stride,dim=1).eq.2)then
265		−	stride_(2) = stride(2)
266			end if
267			end select
268			else
269		−	stride_ = 1
270			end if
271
272
273			!---------------------------------------------------------------------------
274			! Set up dilation
275			!---------------------------------------------------------------------------
276		−	if(present(dilation))then
277		−	select rank(dilation)
278			rank(0)
279		−	dilation_ = dilation
280			rank(1)
281		−	dilation_(1) = dilation(1)
282		−	if(size(dilation,dim=1).eq.1)then
283		−	dilation_(2) = dilation(1)
284		−	elseif(size(dilation,dim=1).eq.2)then
285		−	dilation_(2) = dilation(2)
286			end if
287			end select
288			else
289		−	dilation_ = 1
290			end if
291
292
293			!---------------------------------------------------------------------------
294			! Set hyperparameters
295			!---------------------------------------------------------------------------
296			call layer%set_hyperparams( &
297			num_filters = num_filters_, &
298			kernel_size = kernel_size_, stride = stride_, dilation = dilation_, &
299			padding = padding_, &
300			use_bias = use_bias_, &
301			activation = activation_, &
302			kernel_initialiser = kernel_initialiser_, &
303			bias_initialiser = bias_initialiser_, &
304			verbose = verbose_ &
305		−	)
306
307
308			!---------------------------------------------------------------------------
309			! Initialise layer shape
310			!---------------------------------------------------------------------------
311		−	if(present(input_shape)) call layer%init(input_shape=input_shape)
312
313		−	end function layer_setup
314			!###############################################################################
315
316
317			!###############################################################################
318		−	subroutine set_hyperparams_conv2d( &
319			this, &
320			num_filters, &
321			kernel_size, stride, dilation, &
322			padding, &
323			use_bias, &
324			activation, &
325			kernel_initialiser, bias_initialiser, &
326			verbose &
327			)
328			!! Set hyperparameters for 2D convolutional layer
329			use athena__activation, only: activation_setup
330			use athena__initialiser, only: get_default_initialiser, initialiser_setup
331			use coreutils, only: to_lower
332			implicit none
333
334			! Arguments
335			class(conv2d_layer_type), intent(inout) :: this
336			!! Instance of the 2D convolutional layer
337			integer, intent(in) :: num_filters
338			!! Number of filters
339			integer, dimension(2), intent(in) :: kernel_size, stride, dilation
340			!! Kernel size, stride, dilation
341			character(*), intent(in) :: padding
342			!! Padding
343			logical, intent(in) :: use_bias
344			!! Use bias
345			class(base_actv_type), allocatable, intent(in) :: activation
346			!! Activation function
347			class(base_init_type), allocatable, intent(in) :: &
348			kernel_initialiser, bias_initialiser
349			!! Kernel and bias initialisers
350			integer, optional, intent(in) :: verbose
351			!! Verbosity level
352
353			! Local variables
354			character(len=20) :: padding_
355			character(len=256) :: buffer
356
357		−	this%name = "conv2d"
358		−	this%type = "conv"
359		−	this%input_rank = 3
360		−	this%output_rank = 3
361		−	this%use_bias = use_bias
362		−	if(allocated(this%dil)) deallocate(this%dil)
363		−	if(allocated(this%knl)) deallocate(this%knl)
364		−	if(allocated(this%stp)) deallocate(this%stp)
365			allocate( &
366		−	this%dil(this%input_rank-1), &
367		−	this%knl(this%input_rank-1), &
368		−	this%stp(this%input_rank-1) &
369		−	)
370		−	this%dil = dilation
371		−	this%knl = kernel_size
372		−	this%stp = stride
373		−	this%num_filters = num_filters
374		−	padding_ = trim(adjustl(padding))
375
376		−	select case(trim(adjustl(to_lower(padding_))))
377			case("valid", "none", "")
378			case default
379			this%pad_layer = pad2d_layer_type( &
380		−	padding = [ (this%knl-1)/2 ], &
381			method = padding_ &
382		−	)
383			end select
384		−	if(allocated(this%activation)) deallocate(this%activation)
385		−	if(.not.allocated(activation))then
386		−	this%activation = activation_setup("none")
387			else
388		−	allocate(this%activation, source=activation)
389			end if
390		−	if(allocated(this%kernel_init)) deallocate(this%kernel_init)
391		−	if(.not.allocated(kernel_initialiser))then
392		−	buffer = get_default_initialiser(this%activation%name)
393		−	this%kernel_init = initialiser_setup(buffer)
394			else
395		−	allocate(this%kernel_init, source=kernel_initialiser)
396			end if
397		−	if(allocated(this%bias_init)) deallocate(this%bias_init)
398		−	if(.not.allocated(bias_initialiser))then
399			buffer = get_default_initialiser( &
400			this%activation%name, &
401			is_bias=.true. &
402		−	)
403		−	this%bias_init = initialiser_setup(buffer)
404			else
405		−	allocate(this%bias_init, source=bias_initialiser)
406			end if
407		−	if(present(verbose))then
408		−	if(abs(verbose).gt.0)then
409			write(*,'("CONV2D activation function: ",A)') &
410		−	trim(this%activation%name)
411			write(*,'("CONV2D kernel initialiser: ",A)') &
412		−	trim(this%kernel_init%name)
413			write(*,'("CONV2D bias initialiser: ",A)') &
414		−	trim(this%bias_init%name)
415			end if
416			end if
417
418		−	end subroutine set_hyperparams_conv2d
419			!###############################################################################
420
421
422			!##############################################################################!
423			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
424			!##############################################################################!
425
426
427			!###############################################################################
428		−	subroutine read_conv2d(this, unit, verbose)
429			!! Read 2D convolutional layer from file
430			use athena__tools_infile, only: assign_val, assign_vec, move
431			use coreutils, only: to_lower, to_upper, icount
432			use athena__activation, only: read_activation
433			use athena__initialiser, only: initialiser_setup
434			implicit none
435
436			! Arguments
437			class(conv2d_layer_type), intent(inout) :: this
438			!! Instance of the 2D convolutional layer
439			integer, intent(in) :: unit
440			!! Unit number
441			integer, optional, intent(in) :: verbose
442			!! Verbosity level
443
444			! Local variables
445			integer :: stat
446			!! Status of read
447			integer :: verbose_ = 0
448			!! Verbosity level
449			integer :: j, k, l, c, itmp1, iline, num_params
450			!! Loop variables and temporary integer
451			integer :: num_filters
452			!! Number of filters
453			logical :: use_bias = .true.
454			!! Whether to use bias
455			character(14) :: kernel_initialiser_name='', bias_initialiser_name=''
456			!! Kernel and bias initialisers
457			character(20) :: padding='', activation_name=''
458			!! Padding and activation function
459		−	class(base_actv_type), allocatable :: activation
460			!! Activation function
461		−	class(base_init_type), allocatable :: kernel_initialiser, bias_initialiser
462			!! Initialisers
463			character(256) :: buffer, tag, err_msg
464			!! Buffer, tag, and error message
465			integer, dimension(2) :: kernel_size, stride, dilation
466			!! Kernel size and stride
467			integer, dimension(3) :: input_shape
468			!! Input shape
469		−	real(real32), allocatable, dimension(:) :: data_list
470			!! Data list
471			integer :: param_line, final_line
472			!! Parameter line number
473
474
475			! Initialise optional arguments
476			!---------------------------------------------------------------------------
477		−	if(present(verbose)) verbose_ = verbose
478
479
480			! Loop over tags in layer card
481			!---------------------------------------------------------------------------
482		−	iline = 0
483		−	param_line = 0
484		−	final_line = 0
485		−	tag_loop: do
486
487			! Check for end of file
488			!------------------------------------------------------------------------
489		−	read(unit,'(A)',iostat=stat) buffer
490		−	if(stat.ne.0)then
491			write(err_msg,'("file encountered error (EoF?) before END ",A)') &
492		−	to_upper(this%name)
493		−	call stop_program(err_msg)
494		−	return
495			end if
496		−	if(trim(adjustl(buffer)).eq."") cycle tag_loop
497
498			! Check for end of layer card
499			!------------------------------------------------------------------------
500		−	if(trim(adjustl(buffer)).eq."END "//to_upper(trim(this%name)))then
501		−	final_line = iline
502		−	backspace(unit)
503		−	exit tag_loop
504			end if
505		−	iline = iline + 1
506
507		−	tag=trim(adjustl(buffer))
508		−	if(scan(buffer,"=").ne.0) tag=trim(tag(:scan(tag,"=")-1))
509
510			! Read parameters from save file
511			!------------------------------------------------------------------------
512		−	select case(trim(tag))
513			case("INPUT_SHAPE")
514		−	call assign_vec(buffer, input_shape, itmp1)
515			case("NUM_FILTERS")
516		−	call assign_val(buffer, num_filters, itmp1)
517			case("KERNEL_SIZE")
518		−	call assign_vec(buffer, kernel_size, itmp1)
519			case("STRIDE")
520		−	call assign_vec(buffer, stride, itmp1)
521			case("DILATION")
522		−	call assign_vec(buffer, dilation, itmp1)
523			case("USE_BIAS")
524		−	call assign_val(buffer, use_bias, itmp1)
525			case("PADDING")
526		−	call assign_val(buffer, padding, itmp1)
527		−	padding = to_lower(padding)
528			case("ACTIVATION")
529		−	iline = iline - 1
530		−	backspace(unit)
531		−	activation = read_activation(unit, iline)
532			case("KERNEL_INITIALISER", "KERNEL_INIT", "KERNEL_INITIALIZER")
533		−	call assign_val(buffer, kernel_initialiser_name, itmp1)
534			case("BIAS_INITIALISER", "BIAS_INIT", "BIAS_INITIALIZER")
535		−	call assign_val(buffer, bias_initialiser_name, itmp1)
536			case("WEIGHTS")
537		−	kernel_initialiser_name = 'zeros'
538		−	bias_initialiser_name = 'zeros'
539		−	param_line = iline
540			case default
541			! Don't look for "e" due to scientific notation of numbers
542			! ... i.e. exponent (E+00)
543		−	if(scan(to_lower(trim(adjustl(buffer))),&
544			'abcdfghijklmnopqrstuvwxyz').eq.0)then
545		−	cycle tag_loop
546		−	elseif(tag(:3).eq.'END')then
547		−	cycle tag_loop
548			end if
549			write(err_msg,'("Unrecognised line in input file: ",A)') &
550		−	trim(adjustl(buffer))
551		−	call stop_program(err_msg)
552		−	return
553			end select
554			end do tag_loop
555		−	kernel_initialiser = initialiser_setup(kernel_initialiser_name)
556		−	bias_initialiser = initialiser_setup(bias_initialiser_name)
557
558
559			! Set hyperparameters and initialise layer
560			!---------------------------------------------------------------------------
561			call this%set_hyperparams( &
562			num_filters = num_filters, &
563			kernel_size = kernel_size, stride = stride, dilation = dilation, &
564			padding = padding, &
565			use_bias = use_bias, &
566			activation = activation, &
567			kernel_initialiser = kernel_initialiser, &
568			bias_initialiser = bias_initialiser, &
569			verbose = verbose_ &
570		−	)
571		−	call this%init(input_shape = input_shape)
572
573
574			! Check if WEIGHTS card was found
575			!---------------------------------------------------------------------------
576		−	if(param_line.eq.0)then
577		−	write(0,*) "WARNING: WEIGHTS card in "//to_upper(trim(this%name))//" not found"
578			else
579		−	call move(unit, param_line - iline, iostat=stat)
580		−	num_params = product(this%knl) * input_shape(3) * num_filters
581		−	allocate(data_list(num_params), source=0._real32)
582		−	c = 1
583		−	k = 1
584		−	data_concat_loop: do while(c.le.num_params)
585		−	read(unit,'(A)',iostat=stat) buffer
586		−	if(stat.ne.0) exit data_concat_loop
587		−	k = icount(buffer)
588		−	read(buffer,*,iostat=stat) (data_list(j),j=c,c+k-1)
589		−	c = c + k
590			end do data_concat_loop
591		−	this%params(1)%val(:,1) = data_list
592		−	deallocate(data_list)
593		−	allocate(data_list(num_filters), source=0._real32)
594		−	c = 1
595		−	k = 1
596		−	data_concat_loop2: do while(c.le.num_filters)
597		−	read(unit,'(A)',iostat=stat) buffer
598		−	if(stat.ne.0) exit data_concat_loop2
599		−	k = icount(buffer)
600		−	read(buffer,*,iostat=stat) (data_list(j),j=c,c+k-1)
601		−	c = c + k
602			end do data_concat_loop2
603		−	this%params(2)%val(:,1) = data_list
604		−	deallocate(data_list)
605
606			! Check for end of weights card
607			!------------------------------------------------------------------------
608		−	read(unit,'(A)') buffer
609		−	if(trim(adjustl(buffer)).ne."END WEIGHTS")then
610		−	write(0,*) trim(adjustl(buffer))
611		−	call stop_program("END WEIGHTS not where expected")
612		−	return
613			end if
614			end if
615
616
617			! Check for end of layer card
618			!---------------------------------------------------------------------------
619		−	call move(unit, final_line - iline, iostat=stat)
620		−	read(unit,'(A)') buffer
621		−	if(trim(adjustl(buffer)).ne."END "//to_upper(trim(this%name)))then
622		−	write(0,*) trim(adjustl(buffer))
623		−	write(err_msg,'("END ",A," not where expected")') to_upper(this%name)
624		−	call stop_program(err_msg)
625		−	return
626			end if
627
628		−	end subroutine read_conv2d
629			!###############################################################################
630
631
632			!###############################################################################
633		−	function read_conv2d_layer(unit, verbose) result(layer)
634			!! Read 2D convolutional layer from file and return layer
635			implicit none
636
637			! Arguments
638			integer, intent(in) :: unit
639			!! Unit number
640			integer, optional, intent(in) :: verbose
641			!! Verbosity level
642			class(base_layer_type), allocatable :: layer
643			!! Instance of the 2D convolutional layer
644
645			! Local variables
646			integer :: verbose_ = 0
647			!! Verbosity level
648
649		−	if(present(verbose)) verbose_ = verbose
650		−	allocate(layer, source=conv2d_layer_type())
651		−	call layer%read(unit, verbose=verbose_)
652
653		−	end function read_conv2d_layer
654			!###############################################################################
655
656
657			!###############################################################################
658		−	subroutine build_from_onnx_conv2d( &
659		−	this, node, initialisers, value_info, verbose &
660			)
661			!! Read ONNX attributes for 2D convolutional layer
662			use athena__activation, only: activation_setup
663			implicit none
664
665			! Arguments
666			class(conv2d_layer_type), intent(inout) :: this
667			!! Instance of the 2D convolutional layer
668			type(onnx_node_type), intent(in) :: node
669			!! ONNX node information
670			type(onnx_initialiser_type), dimension(:), intent(in) :: initialisers
671			!! ONNX initialiser information
672			type(onnx_tensor_type), dimension(:), intent(in) :: value_info
673			!! ONNX value info information
674			integer, intent(in) :: verbose
675			!! Verbosity level
676
677			! Local variables
678			integer :: i, weight_idx, bias_idx
679			!! Loop index and temporary integer
680			integer :: num_filters
681			!! Number of filters
682			logical :: use_bias = .true.
683			!! Whether to use bias
684			integer, dimension(2) :: padding, stride, kernel_size, dilation
685			!! Padding, stride, kernel size, and dilation
686		−	integer, dimension(:), allocatable :: dims
687			!! Dimensions
688			character(256) :: val
689			!! Attribute value
690		−	class(base_actv_type), allocatable :: activation
691			!! Activation function
692		−	class(base_init_type), allocatable :: kernel_initialiser, bias_initialiser
693
694		−	do i = 1, size(node%attributes)
695		−	val = node%attributes(i)%val
696		−	select case(trim(adjustl(node%attributes(i)%name)))
697			case("pads")
698		−	read(val,*) padding
699			case("strides")
700		−	read(val,*) stride
701			case("kernel_shape")
702		−	read(val,*) kernel_size
703			case("dilations")
704		−	read(val,*) dilation
705			case default
706			! Do nothing
707		−	write(0,*) "WARNING: Unrecognised attribute in ONNX CONV2D layer: ", &
708		−	trim(adjustl(node%attributes(i)%name))
709			end select
710			end do
711
712		−	weight_idx = -1
713		−	bias_idx = -1
714		−	allocate(dims(0))
715		−	if(size(initialisers).lt.1)then
716		−	call stop_program("ONNX CONV2D layer requires at least 1 initialiser")
717		−	return
718			else
719			! check which initialiser has weights and which has biases,
720			! look for dimensions
721		−	do i = 1, size(initialisers)
722		−	if(allocated(initialisers(i)%dims))then
723		−	dims = [ dims, product(initialisers(i)%dims) ]
724			end if
725			end do
726			end if
727			! if both weight and bias have dimension 1, check which is larger and that
728			! the division of it by the kernel size is equal to the length of the other
729		−	select case(size(dims))
730			case(1)
731			! check if the division of dims by the kernel size is integer
732		−	if(mod(dims(1), product(kernel_size)).eq.0)then
733		−	weight_idx = 1
734			else
735			call stop_program("ONNX CONV2D layer initialiser dimensions do not &
736		−	&match kernel size")
737		−	return
738			end if
739		−	use_bias = .false.
740			case(2)
741			! check which is weight and which is bias
742		−	if(mod(dims(1), product(kernel_size)).eq.0 .and. &
743		−	dims(1)/product(kernel_size).eq.dims(2))then
744		−	weight_idx = 1
745		−	bias_idx = 2
746		−	elseif(mod(dims(2), product(kernel_size)).eq.0 .and. &
747		−	dims(2)/product(kernel_size).eq.dims(1))then
748		−	weight_idx = 2
749		−	bias_idx = 1
750			else
751			call stop_program("ONNX CONV2D layer initialiser dimensions do not &
752		−	&match kernel size")
753		−	return
754			end if
755			case default
756		−	call stop_program("ONNX CONV2D layer number of initialisers not supported")
757		−	return
758			end select
759		−	num_filters = dims(weight_idx) / product(kernel_size)
760		−	if(num_filters .ne. value_info(1)%dims(2))then
761			call stop_program("ONNX CONV2D layer number of filters does not match &
762		−	&value info")
763		−	return
764			end if
765
766		−	kernel_initialiser = data_init_type( data = initialisers(weight_idx)%data )
767		−	if(use_bias)then
768		−	bias_initialiser = data_init_type( data = initialisers(bias_idx)%data )
769			end if
770
771		−	activation = activation_setup("none")
772			call this%set_hyperparams( &
773			num_filters = num_filters, &
774			kernel_size = kernel_size, stride = stride, &
775			dilation = dilation, &
776			padding = "valid", &
777			use_bias = .true., &
778			activation = activation, &
779			verbose = verbose, &
780			kernel_initialiser = kernel_initialiser, &
781			bias_initialiser = bias_initialiser &
782		−	)
783
784		−	end subroutine build_from_onnx_conv2d
785			!###############################################################################
786
787
788			!##############################################################################!
789			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
790			!##############################################################################!
791
792
793			!###############################################################################
794		−	subroutine forward_conv2d(this, input)
795			!! Forward propagation
796			implicit none
797
798			! Arguments
799			class(conv2d_layer_type), intent(inout) :: this
800			!! Instance of the 2D convolutional layer
801			class(array_type), dimension(:,:), intent(in) :: input
802			!! Input values
803
804			! Local variables
805			type(array_type), pointer :: ptr
806			!! Pointer array
807
808
809			! Generate outputs from weights, biases, and inputs
810			!---------------------------------------------------------------------------
811		−	select case(allocated(this%pad_layer))
812			case(.true.)
813		−	call this%pad_layer%forward(input)
814		−	ptr => conv2d(this%pad_layer%output(1,1), this%params(1), &
815			this%stp, this%dil &
816		−	)
817			case default
818		−	ptr => conv2d(input(1,1), this%params(1), this%stp, this%dil)
819			end select
820		−	ptr => add_bias(ptr, this%params(2), dim=3, dim_act_on_shape=.true.)
821
822			! Apply activation function to activation
823			!---------------------------------------------------------------------------
824		−	call this%output(1,1)%zero_grad()
825		−	if(trim(this%activation%name) .eq. "none") then
826		−	call this%output(1,1)%assign_and_deallocate_source(ptr)
827			else
828		−	ptr => this%activation%apply(ptr)
829		−	call this%output(1,1)%assign_and_deallocate_source(ptr)
830			end if
831		−	this%output(1,1)%is_temporary = .false.
832
833		−	end subroutine forward_conv2d
834			!###############################################################################
835
836		−	end module athena__conv2d_layer
837