GCC Code Coverage Report

Line	Branch	Exec	Source
1		✗	submodule(athena__network) athena__network_submodule
2			!! Submodule containing implementations for the network module
3			#ifdef _OPENMP
4			use omp_lib
5			#endif
6			use coreutils, only: stop_program, print_warning, to_lower
7			use athena__misc_ml, only: shuffle
8
9			use athena__accuracy, only: categorical_score, mae_score, mse_score, r2_score
10			use athena__base_layer, only: learnable_layer_type, merge_layer_type
11			#if defined(GFORTRAN)
12			use athena__container_layer, only: container_reduction
13			#endif
14
15			use athena__container_layer, only: &
16			list_of_layer_types, allocate_list_of_layer_types, &
17			list_of_onnx_layer_creators, allocate_list_of_onnx_layer_creators
18
19			! Layer types
20			use athena__flatten_layer, only: flatten_layer_type
21			use athena__add_layer, only: add_layer_type
22			use athena__concat_layer, only: concat_layer_type
23			use athena__input_layer, only: input_layer_type
24			use athena__msgpass_layer, only: msgpass_layer_type
25			use athena__recurrent_layer, only: recurrent_layer_type
26
27			! #ifdef _OPENMP
28			! !$omp declare reduction( &
29			! !$omp& network_reduction : network_type:omp_out%network_reduction(omp_in)) &
30			! !$omp& initialiser(omp_priv = omp_orig)
31			! #endif
32
33			contains
34
35			!###############################################################################
36		−	module subroutine network_reduction(this, source)
37			!! Procedure to add two networks together
38			implicit none
39
40			! Arguments
41			class(network_type), intent(inout) :: this
42			!! Instance of network
43			type(network_type), intent(in) :: source
44			!! Instance of network to be added to this
45
46			! Local variables
47			integer :: i
48			!! Loop index
49
50		−	this%metrics(1)%val = this%metrics(1)%val + source%metrics(1)%val
51		−	this%metrics(2)%val = this%metrics(2)%val + source%metrics(2)%val
52		−	do i=1,size(this%model)
53		−	select type(layer_this => this%model(i)%layer)
54			class is(learnable_layer_type)
55		−	select type(layer_source => source%model(i)%layer)
56			class is(learnable_layer_type)
57		−	call layer_this%reduce(layer_source)
58			end select
59			end select
60			end do
61
62		−	end subroutine network_reduction
63			!###############################################################################
64
65
66			!###############################################################################
67		−	module subroutine network_copy(this, source)
68			!! Procedure to copy a network
69			implicit none
70
71			! Arguments
72			class(network_type), intent(inout) :: this
73			!! Instance of network
74			type(network_type), intent(in), target :: source
75			!! Instance of network to be copied
76
77			! Local variables
78			integer :: i
79			!! Loop index
80
81
82		−	this%metrics = source%metrics
83		−	this%model = source%model
84		−	this%num_layers = source%num_layers
85		−	this%batch_size = source%batch_size
86		−	this%num_params = source%num_params
87		−	this%num_outputs = source%num_outputs
88		−	this%optimiser = source%optimiser
89		−	this%vertex_order = source%vertex_order
90		−	this%root_vertices = source%root_vertices
91		−	this%leaf_vertices = source%leaf_vertices
92		−	this%loss = source%loss
93		−	this%get_accuracy => source%get_accuracy
94		−	this%auto_graph = source%auto_graph
95
96		−	end subroutine network_copy
97			!###############################################################################
98
99
100			!##############################################################################!
101			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
102			!##############################################################################!
103
104
105			!###############################################################################
106		−	module subroutine build_vertex_order(this)
107			!! Generate the order of the layers in the network
108			!!
109			!! This module contains the subroutine to generate the order of the layers
110			!! in the network. The order is generated by depth first search (DFS) on the
111			!! graph of the network.
112			implicit none
113
114			! Arguments
115			class(network_type), intent(inout) :: this
116			!! Instance of network
117
118			! Local variables
119			integer :: i, order_index
120			!! Loop index
121		−	logical, dimension(this%auto_graph%num_vertices) :: visited
122			!! Array to store whether a vertex has been
123
124		−	visited = .false.
125		−	if(allocated(this%vertex_order)) deallocate(this%vertex_order)
126		−	allocate(this%vertex_order(this%auto_graph%num_vertices), source=0)
127
128		−	order_index = 0
129		−	do i = this%auto_graph%num_vertices, 1, -1
130		−	if(.not.visited(i)) call this%dfs( &
131			i, visited, this%vertex_order, order_index &
132		−	)
133			end do
134
135		−	end subroutine build_vertex_order
136			!###############################################################################
137
138
139			!###############################################################################
140		−	recursive module subroutine dfs( &
141		−	this, vertex_index, visited, order, order_index &
142			)
143			!! Depth first search algorithm
144			implicit none
145
146			! Arguments
147			class(network_type), intent(in) :: this
148			!! Instance of network
149			integer, intent(in) :: vertex_index
150			!! Index of the vertex to start the search from
151			logical, dimension(this%auto_graph%num_vertices), intent(inout) :: visited
152			!! Array to store whether a vertex has been visited
153			integer, dimension(this%auto_graph%num_vertices), intent(inout) :: order
154			!! Array to store the order of the vertices
155			integer, intent(inout) :: order_index
156			!! Index of the current vertex in the order array
157
158			! Local variables
159			integer :: i
160			!! Loop index
161
162		−	visited(vertex_index) = .true.
163		−	do i = 1, this%auto_graph%num_vertices, 1
164		−	if(this%auto_graph%adjacency(i,vertex_index).ne.0)then
165		−	if(.not.visited(i)) call this%dfs(i, visited, order, order_index)
166			end if
167			end do
168		−	order_index = order_index + 1
169		−	order(order_index) = vertex_index
170
171		−	end subroutine dfs
172			!###############################################################################
173
174
175			!###############################################################################
176		−	module subroutine build_root_vertices(this)
177			!! Calculate the root vertices of the network
178			implicit none
179
180			! Arguments
181			class(network_type), intent(inout) :: this
182			!! Instance of network
183
184			! Local variables
185			integer :: i
186			!! Loop index
187
188		−	if(allocated(this%root_vertices)) deallocate(this%root_vertices)
189		−	allocate(this%root_vertices(0))
190			! from = 1
191		−	do i = 1, this%auto_graph%num_vertices
192		−	if(all(this%auto_graph%adjacency(:,i).eq.0))then
193		−	this%root_vertices = [this%root_vertices, i]
194			! to = from + this%model(i)layer%num_input_data - 1
195			! this%root_bounds = [ this%root_bounds, reshape([from,to], [2,1]) ]
196			! from = to + 1
197			end if
198			end do
199		−	end subroutine build_root_vertices
200			!###############################################################################
201
202
203			!###############################################################################
204		−	module subroutine build_leaf_vertices(this)
205			!! Calculate the output vertices of the network
206			implicit none
207
208			! Arguments
209			class(network_type), intent(inout) :: this
210			!! Instance of network
211
212			! Local variables
213			integer :: i
214			!! Loop index
215
216		−	if(allocated(this%leaf_vertices)) deallocate(this%leaf_vertices)
217		−	allocate(this%leaf_vertices(0))
218		−	do i = 1, this%auto_graph%num_vertices
219		−	if(all(this%auto_graph%adjacency(i,:).eq.0))then
220		−	this%leaf_vertices = [this%leaf_vertices, i]
221			end if
222			end do
223		−	end subroutine build_leaf_vertices
224			!###############################################################################
225
226
227
228
229
230			!##############################################################################!
231			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
232			!##############################################################################!
233
234
235			!###############################################################################
236		−	module subroutine print(this, file)
237			!! Print the network to a file
238			use coreutils, only: to_upper
239			use athena__io_utils, only: athena__version__
240			implicit none
241
242			! Arguments
243			class(network_type), intent(in) :: this
244			!! Instance of network
245			character(*), intent(in) :: file
246			!! File to print the network to
247
248			! Local variables
249			integer :: l, v, e, vertex_index, unit
250			!! Loop index
251			integer :: operator_in, operator_out
252			!! Operators for the layer
253			character(3) :: operator_str
254			!! String to store the operator
255			character(256) :: suffix, fmt
256			!! Suffix for the layer
257		−	integer, dimension(:), allocatable :: input_list, output_list
258
259		−	open(newunit=unit,file=file,status='replace')
260
261		−	write(unit,'("NETWORK_SETTINGS")')
262		−	write(unit,'(3X,"ATHENA_VERSION = ",A)') trim(adjustl(athena__version__))
263		−	if(allocated(this%name)) write(unit,'(3X,"NAME = ",A)') trim(adjustl(this%name))
264		−	write(unit,'(3X,"EPOCH = ",I0)') this%epoch
265		−	write(unit,'(3X,"BATCH_SIZE = ",I0)') this%batch_size
266		−	write(unit,'(3X,"ACCURACY = ",F0.9)') this%accuracy_val
267		−	write(unit,'(3X,"LOSS = ",F0.9)') this%loss_val
268		−	if(allocated(this%accuracy_method))then
269		−	write(unit,'(3X,"ACCURACY_METHOD = ",A)') trim(adjustl(this%accuracy_method))
270			end if
271		−	if(allocated(this%loss_method))then
272		−	write(unit,'(3X,"LOSS_METHOD = ",A)') trim(adjustl(this%loss_method))
273			end if
274		−	if(allocated(this%optimiser))then
275		−	write(unit,'(3X,"OPTIMISER: ",A)') trim(adjustl(this%optimiser%name))
276		−	call this%optimiser%print_to_unit(unit=unit)
277		−	write(unit,'(3X,"END OPTIMISER")')
278			end if
279		−	write(unit,'("END NETWORK_SETTINGS")')
280
281		−	do v = 1, size(this%vertex_order,dim=1), 1
282		−	l = this%vertex_order(v)
283		−	operator_in = -1
284		−	operator_out = -1
285		−	allocate(input_list(0), output_list(0))
286		−	do e = 1, this%auto_graph%num_edges
287		−	if(-this%auto_graph%edge(e)%index(2).eq.l)then
288		−	if(operator_in.gt.0.and.this%auto_graph%edge(e)%id.ne.operator_in)then
289		−	write(0,*) "WARNING: multiple operators for layer ", l
290		−	write(0,*) " using operator ", this%auto_graph%edge(e)%id
291			end if
292		−	operator_in = this%auto_graph%edge(e)%id
293			vertex_index = &
294		−	findloc( this%vertex_order, this%auto_graph%edge(e)%index(1), 1 )
295		−	input_list = [ input_list, vertex_index ]
296			end if
297		−	if(this%auto_graph%edge(e)%index(1).eq.l)then
298		−	if(operator_out.gt.0.and.this%auto_graph%edge(e)%id.ne.operator_out)then
299		−	write(0,*) "WARNING: multiple operators for layer ", l
300		−	write(0,*) " using operator ", this%auto_graph%edge(e)%id
301			end if
302		−	operator_in = this%auto_graph%edge(e)%id
303			vertex_index = &
304		−	findloc( this%vertex_order, this%auto_graph%edge(e)%index(2), 1 )
305		−	output_list = [ output_list, vertex_index ]
306			end if
307			end do
308
309		−	suffix = ""
310		−	select case(operator_in)
311			case(1)
312		−	operator_str = " ||"
313			case(2)
314		−	operator_str = " +"
315			case(3)
316		−	operator_str = " *"
317			end select
318			! get size of input_list and make the formatted string
319		−	if(size(input_list).eq.0)then
320		−	write(suffix,'(A," []")') trim(operator_str)
321			else
322		−	write(fmt,'("(A,A,"" ["",",I0,"(1X,I0),"" ]"")")') size(input_list)
323		−	write(suffix,fmt) trim(suffix), operator_str, input_list
324			end if
325			! select case(operator_out)
326			! case(1)
327			! operator_str = " ||"
328			! case(2)
329			! operator_str = " +"
330			! case(3)
331			! operator_str = " *"
332			! end select
333			! if(size(output_list).gt.0)then
334			! write(fmt,'("(A,A,"" ["",",I0,"(1X,I0),"" ]"")")') size(output_list)
335			! write(suffix,fmt) trim(suffix), operator_str, output_list
336			! end if
337
338		−	write(unit,'(A,A)') to_upper(trim(this%model(l)%layer%name)), trim(suffix)
339		−	call this%model(l)%layer%print(unit=unit, print_header_footer=.false.)
340
341		−	write(unit,'("END ",A)') to_upper(trim(this%model(l)%layer%name))
342		−	deallocate(input_list, output_list)
343			end do
344		−	close(unit)
345
346		−	end subroutine print
347			!###############################################################################
348
349
350			!###############################################################################
351		−	module subroutine read(this, file)
352			!! Read the network from a file
353			use coreutils, only: icount
354			implicit none
355
356			! Arguments
357			class(network_type), intent(inout) :: this
358			!! Instance of network
359			character(*), intent(in) :: file
360			!! File to read the network from
361
362			! Local variables
363			integer :: i, unit, stat, itmp1
364			!! Loop index
365		−	integer, dimension(:), allocatable :: input_list, output_list
366			!!! List of input and output layers
367			character(256) :: buffer, err_msg, input_str, output_str
368			!! Buffer for reading lines from file
369			character(20) :: name
370			!! Name of the layer
371			character(2) :: operator_in, operator_out
372			!! Operator for the layer
373			integer :: layer_index
374			!! Index of the layer in the list of layer types
375
376
377		−	if(.not.allocated(list_of_layer_types))then
378		−	call allocate_list_of_layer_types()
379			end if
380
381		−	open(newunit=unit,file=file,action='read')
382		−	i = 0
383		−	card_loop: do
384		−	i = i + 1
385		−	read(unit,'(A)',iostat=stat) buffer
386		−	if(stat.lt.0)then
387		−	exit card_loop
388		−	elseif(stat.gt.0)then
389		−	call stop_program("error encountered in network read")
390		−	return
391			end if
392		−	if(trim(adjustl(buffer)).eq."") cycle card_loop
393
394			!! check if a tag line
395		−	if(scan(buffer,'=').ne.0)then
396		−	write(0,*) "WARNING: unexpected line in read file"
397		−	write(0,*) trim(buffer)
398		−	write(0,*) " skipping..."
399		−	cycle card_loop
400			end if
401
402			!! check for card
403		−	name = trim(adjustl(buffer(1:scan(buffer,' ')-1)))
404		−	if(name.eq."NETWORK_SETTINGS")then
405		−	call this%read_network_settings(unit)
406		−	cycle card_loop
407			end if
408		−	buffer = trim(adjustl(buffer(scan(buffer,' ')+1:)))
409		−	operator_in = trim(adjustl(buffer(1:scan(buffer,' ')-1)))
410		−	buffer = trim(adjustl(buffer(scan(buffer,' ')+1:)))
411		−	input_str = trim(adjustl(buffer(1:scan(buffer,']'))))
412		−	if(scan(input_str,'[').ne.0)then
413			input_str = &
414		−	trim(adjustl(input_str(scan(input_str,'[')+1:scan(input_str,']')-1)))
415		−	itmp1 = icount(input_str)
416		−	allocate(input_list(itmp1))
417		−	read(input_str,*) input_list
418			else
419		−	allocate(input_list, source = [-1])
420			end if
421		−	buffer = trim(adjustl(buffer(scan(buffer,']')+1:)))
422		−	operator_out = trim(adjustl(buffer(1:scan(buffer,' ')-1)))
423		−	buffer = trim(adjustl(buffer(scan(buffer,' ')+1:)))
424		−	output_str = trim(adjustl(buffer(1:scan(buffer,']'))))
425		−	if(scan(output_str,'[').ne.0)then
426			output_str = &
427		−	trim(adjustl(output_str(scan(output_str,'[')+1:scan(output_str,']')-1)))
428		−	itmp1 = icount(output_str)
429		−	allocate(output_list(itmp1))
430		−	read(output_str,*) output_list
431			else
432		−	allocate(output_list(0))
433			end if
434		−	name = trim(adjustl(to_lower(name)))
435			layer_index = &
436			findloc( &
437		−	[ list_of_layer_types(:)%name ], &
438			name, &
439			dim = 1 &
440		−	)
441		−	if(layer_index.eq.0)then
442		−	write(err_msg,'("unrecognised card ''",A)') trim(adjustl(buffer))
443		−	call stop_program(err_msg)
444		−	return
445			end if
446			call this%add( &
447			list_of_layer_types(layer_index)%read_ptr(unit), &
448			input_list = input_list, &
449			operator = operator_in &
450		−	)
451		−	if(allocated(input_list)) deallocate(input_list)
452		−	if(allocated(output_list)) deallocate(output_list)
453			end do card_loop
454		−	close(unit)
455
456		−	end subroutine read
457			!###############################################################################
458
459
460			!###############################################################################
461		−	module subroutine read_network_settings(this, unit)
462			!! Read the network settings from a file
463			use athena__tools_infile, only: assign_val, assign_vec
464			use coreutils, only: to_lower, to_upper, icount
465			implicit none
466
467			! Arguments
468			class(network_type), intent(inout) :: this
469			!! Instance of network
470			integer, intent(in) :: unit
471			!! File unit
472
473			! Local variables
474			integer :: stat
475			!! File status
476			integer :: itmp1
477			!! Temporary integer
478			character(20) :: accuracy_method, loss_method
479			!! Methods for accuracy and loss
480			character(256) :: buffer, tag, err_msg, name_
481			!! Buffer for reading lines, tag for identifying lines, error message
482
483
484			! Loop over tags in layer card
485			!---------------------------------------------------------------------------
486		−	accuracy_method = ""
487		−	loss_method = ""
488		−	tag_loop: do
489
490			! Check for end of file
491			!------------------------------------------------------------------------
492		−	read(unit,'(A)',iostat=stat) buffer
493		−	if(stat.ne.0)then
494			write(err_msg,'("file encountered error (EoF?) before END ",A)') &
495		−	to_upper(this%name)
496		−	call stop_program(err_msg)
497		−	return
498			end if
499		−	if(trim(adjustl(buffer)).eq."") cycle tag_loop
500
501			! Check for end of layer card
502			!------------------------------------------------------------------------
503		−	if(trim(adjustl(buffer)).eq."END NETWORK_SETTINGS")then
504		−	backspace(unit)
505		−	exit tag_loop
506			end if
507
508		−	tag=trim(adjustl(buffer))
509		−	if(scan(buffer,"=").ne.0) tag=trim(tag(:scan(tag,"=")-1))
510		−	if(scan(buffer,":").ne.0) tag=trim(tag(:scan(tag,":")-1))
511
512			! Read parameters from save file
513			!------------------------------------------------------------------------
514		−	select case(trim(tag))
515			case("ATHENA_VERSION")
516			! Ignore this tag, it is only for information
517			case("NAME")
518		−	call assign_val(buffer, name_, itmp1)
519		−	if(len(trim(adjustl(name_))) .gt. 0)then
520		−	this%name = trim(adjustl(name_))
521			end if
522			case("EPOCH")
523		−	call assign_val(buffer, this%epoch, itmp1)
524			case("BATCH_SIZE")
525		−	call assign_val(buffer, this%batch_size, itmp1)
526			case("ACCURACY")
527		−	call assign_val(buffer, this%accuracy_val, itmp1)
528			case("LOSS")
529		−	call assign_val(buffer, this%loss_val, itmp1)
530			case("ACCURACY_METHOD")
531		−	call assign_val(buffer, accuracy_method, itmp1)
532		−	call this%set_accuracy(accuracy_method)
533			case("LOSS_METHOD")
534		−	call assign_val(buffer, loss_method, itmp1)
535		−	call this%set_loss(loss_method)
536			case("OPTIMISER")
537		−	backspace(unit)
538		−	call this%read_optimiser_settings(unit)
539			case default
540			! Don't look for "e" due to scientific notation of numbers
541			! ... i.e. exponent (E+00)
542		−	if(scan(to_lower(trim(adjustl(buffer))),&
543			'abcdfghijklmnopqrstuvwxyz').eq.0)then
544		−	cycle tag_loop
545		−	elseif(tag(:3).eq.'END')then
546		−	cycle tag_loop
547			end if
548			write(err_msg,'("Unrecognised line in input file: ",A)') &
549		−	trim(adjustl(buffer))
550		−	call stop_program(err_msg)
551		−	return
552			end select
553			end do tag_loop
554
555
556			! Check for end of layer card
557			!---------------------------------------------------------------------------
558		−	read(unit,'(A)') buffer
559		−	if(trim(adjustl(buffer)).ne."END NETWORK_SETTINGS")then
560		−	write(0,*) trim(adjustl(buffer))
561		−	write(err_msg,'("END NETWORK_SETTINGS not where expected")')
562		−	call stop_program(err_msg)
563		−	return
564			end if
565
566			end subroutine read_network_settings
567			!-------------------------------------------------------------------------------
568		−	module subroutine read_optimiser_settings(this, unit)
569			!! Read the optimiser settings from a file
570			use coreutils, only: to_lower, to_upper, icount
571			use athena__optimiser, only: &
572			sgd_optimiser_type, adam_optimiser_type, rmsprop_optimiser_type, &
573			adagrad_optimiser_type, base_optimiser_type
574			implicit none
575
576			! Arguments
577			class(network_type), intent(inout) :: this
578			!! Instance of network
579			integer, intent(in) :: unit
580			!! File unit
581
582			! Local variables
583			integer :: stat
584			!! File status
585			character(20) :: optimiser_name
586			!! Name of the optimiser
587			character(256) :: buffer, err_msg, tmp
588			!! Buffer for reading lines, error message
589
590			! Read until end of optimiser settings
591		−	read(unit,'(A)',iostat=stat) buffer
592		−	if(stat.ne.0)then
593			write(err_msg,'("file encountered error (EoF?) before END ",A)') &
594		−	to_upper(this%name)
595		−	call stop_program(err_msg)
596		−	return
597			end if
598		−	read(buffer,*) tmp, optimiser_name
599
600		−	select case(trim(adjustl(to_lower(optimiser_name))))
601			case("sgd")
602		−	this%optimiser = sgd_optimiser_type()
603			case("adam")
604		−	this%optimiser = adam_optimiser_type()
605			case("rmsprop")
606		−	this%optimiser = rmsprop_optimiser_type()
607			case("adagrad")
608		−	this%optimiser = adagrad_optimiser_type()
609			case("","base")
610		−	this%optimiser = base_optimiser_type()
611			case default
612		−	write(err_msg,'("Unrecognised optimiser: ",A)') trim(adjustl(optimiser_name))
613		−	call stop_program(err_msg)
614		−	return
615			end select
616		−	call this%optimiser%read(unit)
617
618			end subroutine read_optimiser_settings
619			!###############################################################################
620
621
622			!###############################################################################
623		−	module subroutine build_from_onnx( &
624		−	this, nodes, initialisers, inputs, value_info, verbose &
625			)
626			!! Build network from ONNX nodes and initialisers
627			use coreutils, only: to_lower
628			implicit none
629
630			! Arguments
631			class(network_type), intent(inout) :: this
632			!! Instance of network
633			type(onnx_node_type), dimension(:), intent(in) :: nodes
634			!! Array of ONNX nodes
635			type(onnx_initialiser_type), dimension(:), intent(in) :: initialisers
636			!! Array of ONNX initialisers
637			type(onnx_tensor_type), dimension(:), intent(in) :: inputs
638			!! Array of ONNX inputs
639			type(onnx_tensor_type), dimension(:), intent(in) :: value_info
640			!! Array of ONNX value infos
641			integer, optional, intent(in) :: verbose
642			!! Verbosity level
643
644			! Local variables
645			integer :: i, j, k, j_out, layer_index
646			!! Loop indices
647			integer :: verbose_ = 0
648			!! Verbosity level
649			character(20) :: op_type
650			!! Lowercase op_type
651			character(64) :: tmp_name
652			!! Temporary name for matching
653			character(256) :: err_msg
654			!! Error message
655		−	integer, dimension(:), allocatable :: input_shape
656			!! Shape of input layer
657		−	integer, dimension(:), allocatable :: input_list
658			!! List of input layers
659		−	type(onnx_initialiser_type), dimension(:), allocatable :: init_list
660			!! List of initialisers for a specific node
661		−	type(onnx_tensor_type), dimension(:), allocatable :: value_info_list
662			!! List of value info tensors
663
664		−	verbose_ = 0
665		−	if(present(verbose)) verbose_ = verbose
666
667
668		−	if(.not.allocated(list_of_onnx_layer_creators))then
669		−	call allocate_list_of_onnx_layer_creators()
670			end if
671
672		−	do i = 1, size(inputs)
673		−	write(*,*) "Processing ONNX input: ", trim(inputs(i)%name)
674		−	input_shape = inputs(i)%dims(size(inputs(i)%dims):2:-1)
675
676			call this%add( &
677			input_layer_type(input_shape, index=i) &
678		−	)
679			end do
680
681			! Loop through nodes and create layers
682		−	do i = 1, size(nodes)
683		−	if(verbose_.gt.0) write(*,*) "Processing ONNX node: ", trim(nodes(i)%name), &
684		−	" (", trim(nodes(i)%op_type), ")"
685		−	op_type = trim(adjustl(nodes(i)%op_type))
686
687			layer_index = &
688			findloc( &
689		−	[ list_of_onnx_layer_creators(:)%op_type ], &
690			op_type, &
691			dim = 1 &
692		−	)
693		−	if(layer_index.eq.0)then
694		−	write(err_msg,'("unrecognised op_type ''",A)') trim(adjustl(nodes(i)%op_type))
695		−	call stop_program(err_msg)
696		−	return
697			end if
698
699			! find all input layers and initialisers for this node
700			! ... i.e. check over inputs for name matches
701		−	j_out = 0
702		−	allocate(init_list(0))
703		−	allocate(input_list(0))
704		−	allocate(value_info_list(0))
705		−	do j = 1, size(nodes(i)%inputs)
706		−	do k = 1, size(initialisers)
707		−	if(trim(nodes(i)%inputs(j)) .eq. trim(initialisers(k)%name))then
708		−	init_list = [ init_list, initialisers(k) ]
709			end if
710			end do
711		−	do k = 1, size(inputs)
712		−	if(trim(nodes(i)%inputs(j)) .eq. trim(inputs(k)%name))then
713		−	input_list = [ input_list, k ]
714			end if
715			end do
716		−	tmp_name = trim(nodes(i)%inputs(j))
717		−	if(index(tmp_name, "_output").gt.0) &
718		−	tmp_name = trim(tmp_name(:index(tmp_name, "_output")-1))
719		−	do k = 1, size(nodes)
720		−	if(trim(tmp_name) .eq. trim(nodes(k)%name))then
721		−	input_list = [ input_list, k + size(inputs) ]
722			end if
723			end do
724			end do
725		−	do j = 1, size(nodes(i)%outputs)
726		−	do k = 1, size(value_info)
727		−	if(trim(nodes(i)%outputs(j)) .eq. trim(value_info(k)%name))then
728		−	value_info_list = [ value_info_list, value_info(k) ]
729			end if
730			end do
731			end do
732		−	if(size(init_list)+size(input_list).ne.size(nodes(i)%inputs))then
733		−	if(verbose_.gt.0)then
734		−	write(0,*) "WARNING: not all inputs found for node ", &
735		−	trim(nodes(i)%name)
736			end if
737			end if
738
739			! assume default operator
740
741			call this%add( &
742			list_of_onnx_layer_creators(layer_index)%create_ptr( &
743		−	nodes(i), init_list, value_info_list &
744			), &
745			input_list = input_list &
746			! operator = operator_in &
747		−	)
748		−	deallocate(input_list)
749		−	deallocate(init_list)
750		−	deallocate(value_info_list)
751			end do
752
753		−	if(verbose_.gt.0) write(*,*) "ONNX model built with ", this%num_layers, " layers."
754
755		−	end subroutine build_from_onnx
756			!###############################################################################
757
758
759			!##############################################################################!
760			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
761			!##############################################################################!
762
763
764			!###############################################################################
765		−	module subroutine add(this, layer, input_list, output_list, operator)
766			!! Add a layer to the network
767			implicit none
768
769			! Arguments
770			class(network_type), intent(inout) :: this
771			!! Instance of network
772			class(base_layer_type), intent(in) :: layer
773			!! Layer to add to the network
774			integer, dimension(:), optional, intent(in) :: input_list
775			!! List of input layers
776			integer, dimension(:), optional, intent(in) :: output_list
777			!! List of output layers
778			class(*), optional, intent(in) :: operator
779			!! Operator to use to connect the layers
780
781			! Local variables
782			integer :: i, vertex_index
783			!! Loop index
784			integer :: operator_
785			!! Operator to use to connect the layers
786			character(256) :: err_msg
787			!! Error message
788			integer, dimension(2) :: vertex_indices
789			!! Indices of the vertices to connect
790		−	type(container_layer_type), allocatable, dimension(:) :: model
791			!! Model to add the layer to
792
793
794
795		−	if(.not.allocated(this%model))then
796		−	this%model = [container_layer_type()]
797		−	this%num_layers = 1
798			else
799		−	allocate(model(size(this%model,dim=1)+1))
800		−	do i = 1, size(this%model,dim=1)
801		−	allocate(model(i)%layer, source=this%model(i)%layer)
802			end do
803		−	call move_alloc(model, this%model)
804		−	this%num_layers = this%num_layers + 1
805			end if
806		−	allocate(this%model(size(this%model,dim=1))%layer, source=layer)
807		−	this%model(size(this%model,dim=1))%layer%id = this%num_layers
808
809
810		−	operator_ = 1
811		−	if(present(operator))then
812			select type(operator)
813			type is(integer)
814		−	operator_ = operator
815			type is(character(*))
816		−	select case(trim(to_lower(operator)))
817			case("||", "concat", "concatenate", "append")
818		−	operator_ = 1
819			case("+", "add")
820		−	operator_ = 2
821			case("*", "x", "mul", "multiply")
822		−	operator_ = 3
823			end select
824			end select
825			end if
826		−	if(operator_.gt.2.or.operator_.lt.1)then
827		−	call stop_program("invalid operator")
828		−	return
829			end if
830
831			! edge_index(1) = index of the previous layer
832			! abs(edge_index(2)) = index of the current layer
833			! the -ve sign of edge_index(2) indicates that the edge goes from the
834			! previous layer to the current layer
835			! i.e. forward pass flows from positive to negative
836			! adjacency(i,:) is all of the layers that i feeds forward to
837			! adjacency(:,i) is all of the layers that feed forward to i
838			! (i.e. the backward pass)
839		−	this%auto_graph%directed = .true.
840			call this%auto_graph%add_vertex( &
841			feature=[1._real32], id=this%num_layers, update_adjacency=.true. &
842		−	)
843		−	if(present(input_list))then
844		−	do i = 1, size(input_list), 1
845		−	if(input_list(i).eq.0)then
846		−	vertex_index = 0
847			elseif( &
848		−	input_list(i).le.-this%auto_graph%num_vertices .or. &
849			input_list(i).gt.this%auto_graph%num_vertices &
850			)then
851			write(err_msg, &
852			'("input vertex index ",I0," out of range (",I0,":",I0,")")' &
853		−	) &
854		−	input_list(i), &
855		−	-this%auto_graph%num_vertices +1, &
856		−	this%auto_graph%num_vertices
857		−	call stop_program(err_msg)
858		−	return
859		−	elseif(input_list(i).lt.0)then
860		−	vertex_index = this%auto_graph%num_vertices + input_list(i)
861			else
862			vertex_index = findloc( &
863		−	[this%auto_graph%vertex(:)%id], &
864		−	input_list(i), 1 &
865		−	)
866			end if
867		−	vertex_indices = [ vertex_index, -this%auto_graph%num_vertices ]
868			call this%auto_graph%add_edge( &
869			index = vertex_indices, &
870			feature = [ 1._real32 ], &
871			id = operator_, &
872			update_adjacency = .true. &
873		−	)
874			end do
875		−	elseif(trim(layer%type).ne."inpt".and.this%auto_graph%num_vertices.gt.1)then
876			vertex_indices = [ &
877			this%auto_graph%num_vertices - 1, &
878			-this%auto_graph%num_vertices &
879		−	]
880			call this%auto_graph%add_edge( &
881			index = vertex_indices, &
882			feature = [ 1._real32 ], &
883			id = operator_, &
884			update_adjacency = .true. &
885		−	)
886			end if
887
888		−	if(present(output_list))then
889		−	do i = 1, size(output_list), 1
890			vertex_index = findloc( &
891		−	[this%auto_graph%vertex(:)%id], &
892		−	output_list(i), 1 &
893		−	)
894		−	vertex_indices = [ this%auto_graph%num_vertices, -vertex_index ]
895			call this%auto_graph%add_edge( &
896			index = vertex_indices, &
897			feature = [ 1._real32 ], &
898			id = operator_, &
899			update_adjacency = .true. &
900		−	)
901			end do
902			end if
903
904		−	end subroutine add
905			!###############################################################################
906
907
908			!###############################################################################
909		−	module function network_setup( &
910		−	layers, optimiser, loss_method, accuracy_method, &
911			metrics, batch_size &
912		−	) result(network)
913			!! Setup the network
914			implicit none
915
916			! Arguments
917			type(container_layer_type), dimension(:), intent(in) :: layers
918			!! Layers to add to the network
919			class(base_optimiser_type), optional, intent(in) :: optimiser
920			!! Optimiser to use for training
921			class(*), optional, intent(in) :: loss_method
922			!! Loss method
923			character(*), optional, intent(in) :: accuracy_method
924			!! Accuracy method
925			class(*), dimension(..), optional, intent(in) :: metrics
926			!! Metrics
927			integer, optional, intent(in) :: batch_size
928			!! Batch size
929
930			type(network_type) :: network
931			!! Network to setup
932
933			! Local variables
934			integer :: l
935			!! Loop index
936
937
938			!---------------------------------------------------------------------------
939			! Handle optional arguments
940			!---------------------------------------------------------------------------
941		−	if(present(loss_method)) call network%set_loss(loss_method)
942		−	if(present(accuracy_method)) call network%set_accuracy(accuracy_method)
943		−	if(present(metrics)) call network%set_metrics(metrics)
944		−	if(present(batch_size)) network%batch_size = batch_size
945		−	network%auto_graph%directed = .true.
946
947
948			!---------------------------------------------------------------------------
949			! Add layers to network
950			!---------------------------------------------------------------------------
951		−	do l = 1, size(layers)
952		−	call network%add(layers(l)%layer)
953			end do
954
955
956			!---------------------------------------------------------------------------
957			! Compile network if optimiser present
958			!---------------------------------------------------------------------------
959		−	if(present(optimiser)) call network%compile(optimiser)
960
961		−	end function network_setup
962			!###############################################################################
963
964
965			!###############################################################################
966		−	module subroutine set_metrics(this, metrics)
967			!! Set the metrics for the network
968			use coreutils, only: to_lower
969			implicit none
970
971			! Arguments
972			class(network_type), intent(inout) :: this
973			!! Instance of network
974			class(*), dimension(..), intent(in) :: metrics
975			!! Metrics
976
977			! Local variables
978			integer :: i
979			!! Loop index
980
981
982		−	this%metrics%active = .false.
983		−	this%metrics(1)%key = "loss"
984		−	this%metrics(2)%key = "accuracy"
985		−	this%metrics%threshold = 1.E-1_real32
986			select rank(metrics)
987			#if defined(GFORTRAN)
988			rank(0)
989			select type(metrics)
990			type is(character(*))
991			! ERROR: ifort cannot identify that the rank of metrics has been ...
992			! ... identified as scalar here
993			where(to_lower(trim(metrics)).eq.this%metrics%key)
994			this%metrics%active = .true.
995			end where
996			end select
997			#endif
998			rank(1)
999		−	select type(metrics)
1000			type is(character(*))
1001		−	do i=1,size(metrics,1)
1002		−	where(to_lower(trim(metrics(i))).eq.this%metrics%key)
1003			this%metrics%active = .true.
1004			end where
1005			end do
1006			type is(metric_dict_type)
1007		−	if(size(metrics,1).eq.2)then
1008		−	this%metrics(:2) = metrics(:2)
1009			else
1010		−	call stop_program("invalid length array for metric_dict_type")
1011		−	return
1012			end if
1013			end select
1014			rank default
1015		−	call stop_program("provided metrics rank in compile invalid")
1016		−	return
1017			end select
1018
1019			end subroutine set_metrics
1020			!###############################################################################
1021
1022
1023			!###############################################################################
1024		−	module subroutine set_loss(this, loss_method, verbose)
1025			!! Set the loss method for the network
1026			use coreutils, only: to_lower
1027			use athena__loss, only: &
1028			bce_loss_type, &
1029			cce_loss_type, &
1030			mae_loss_type, &
1031			mse_loss_type, &
1032			nll_loss_type, &
1033			huber_loss_type
1034			implicit none
1035
1036			! Arguments
1037			class(network_type), intent(inout) :: this
1038			!! Instance of network
1039			class(*), intent(in) :: loss_method
1040			!! Loss method
1041			integer, optional, intent(in) :: verbose
1042			!! Verbosity level
1043
1044			! Local variables
1045			integer :: verbose_
1046			!! Verbosity level
1047		−	character(len=:), allocatable :: loss_method_
1048			!! Loss method
1049			character(256) :: err_msg
1050			!! Error message
1051
1052
1053		−	if(present(verbose))then
1054		−	verbose_ = verbose
1055			else
1056		−	verbose_ = 0
1057			end if
1058
1059			!---------------------------------------------------------------------------
1060			! Handle analogous definitions
1061			!---------------------------------------------------------------------------
1062
1063			!---------------------------------------------------------------------------
1064			! Set loss method
1065			!---------------------------------------------------------------------------
1066			select type(loss_method)
1067			class is(base_loss_type)
1068		−	this%loss = loss_method
1069		−	if(verbose_.gt.0) write(*,*) "Loss method: ", trim(loss_method%name)
1070		−	loss_method_ = trim(loss_method%name)
1071			type is(character(*))
1072		−	loss_method_ = to_lower(loss_method)
1073		−	select case(loss_method)
1074			case("binary_crossentropy")
1075		−	loss_method_ = "bce"
1076			case("categorical_crossentropy")
1077		−	loss_method_ = "cce"
1078			case("mean_absolute_error")
1079		−	loss_method_ = "mae"
1080			case("mean_squared_error")
1081		−	loss_method_ = "mse"
1082			case("negative_log_likelihood")
1083		−	loss_method_ = "nll"
1084			case("huber")
1085		−	loss_method_ = "hub"
1086			end select
1087		−	select case(loss_method_)
1088			case("bce")
1089		−	this%loss = bce_loss_type()
1090		−	if(verbose_.gt.0) write(*,*) "Loss method: Binary Cross Entropy"
1091			case("cce")
1092		−	this%loss = cce_loss_type()
1093		−	if(verbose_.gt.0) write(*,*) "Loss method: Categorical Cross Entropy"
1094			case("mae")
1095		−	this%loss = mae_loss_type()
1096		−	if(verbose_.gt.0) write(*,*) "Loss method: Mean Absolute Error"
1097			case("mse")
1098		−	this%loss = mse_loss_type()
1099		−	if(verbose_.gt.0) write(*,*) "Loss method: Mean Squared Error"
1100			case("nll")
1101		−	this%loss = nll_loss_type()
1102		−	if(verbose_.gt.0) write(*,*) "Loss method: Negative Log Likelihood"
1103			case("hub")
1104		−	this%loss = huber_loss_type()
1105		−	if(verbose_.gt.0) write(*,*) "Loss method: Huber"
1106			case default
1107			write(err_msg,'(A)') &
1108			"No loss method provided" // &
1109			achar(13) // achar(10) // &
1110		−	"Failed loss method: "//trim(loss_method_)
1111		−	call stop_program(trim(err_msg))
1112		−	return
1113			end select
1114			end select
1115		−	this%loss_method = loss_method_
1116
1117		−	end subroutine set_loss
1118			!###############################################################################
1119
1120
1121			!###############################################################################
1122		−	module subroutine set_accuracy(this, accuracy_method, verbose)
1123			!! Set the accuracy method for the network
1124			use coreutils, only: to_lower
1125			use athena__accuracy, only: &
1126			categorical_score, &
1127			mae_score, &
1128			mse_score, &
1129			rmse_score, &
1130			r2_score
1131			implicit none
1132
1133			! Arguments
1134			class(network_type), intent(inout) :: this
1135			!! Instance of network
1136			character(*), intent(in) :: accuracy_method
1137			!! Accuracy method
1138			integer, optional, intent(in) :: verbose
1139			!! Verbosity level
1140
1141			! Local variables
1142			integer :: verbose_
1143			!! Verbosity level
1144		−	character(len=:), allocatable :: accuracy_method_
1145			!! Accuracy method
1146			character(256) :: err_msg
1147			!! Error message
1148
1149
1150		−	if(present(verbose))then
1151		−	verbose_ = verbose
1152			else
1153		−	verbose_ = 0
1154			end if
1155
1156			!---------------------------------------------------------------------------
1157			! Handle analogous definitions
1158			!---------------------------------------------------------------------------
1159		−	accuracy_method_ = to_lower(accuracy_method)
1160		−	select case(accuracy_method)
1161			case("categorical")
1162		−	accuracy_method_ = "cat"
1163			case("mean_absolute_error")
1164		−	accuracy_method_ = "mae"
1165			case("mean_squared_error")
1166		−	accuracy_method_ = "mse"
1167			case("root_mean_squared_error")
1168		−	accuracy_method_ = "rmse"
1169			case("r2", "r^2", "r squared")
1170		−	accuracy_method_ = "r2"
1171			end select
1172
1173			!---------------------------------------------------------------------------
1174			! Set accuracy method
1175			!---------------------------------------------------------------------------
1176		−	select case(accuracy_method_)
1177			case("cat")
1178		−	this%get_accuracy => categorical_score
1179		−	if(verbose_.gt.0) write(*,*) "Accuracy method: Categorical "
1180			case("mae")
1181		−	this%get_accuracy => mae_score
1182		−	if(verbose_.gt.0) write(*,*) "Accuracy method: Mean Absolute Error"
1183			case("mse")
1184		−	this%get_accuracy => mse_score
1185		−	if(verbose_.gt.0) write(*,*) "Accuracy method: Mean Squared Error"
1186			case("rmse")
1187		−	this%get_accuracy => rmse_score
1188		−	if(verbose_.gt.0) write(*,*) "Accuracy method: Root Mean Squared Error"
1189			case("r2")
1190		−	this%get_accuracy => r2_score
1191		−	if(verbose_.gt.0) write(*,*) "Accuracy method: R^2"
1192			case default
1193			write(err_msg,'(A)') &
1194			"No accuracy method provided" // &
1195			achar(13) // achar(10) // &
1196		−	"Failed accuracy method: "//trim(accuracy_method_)
1197		−	call stop_program(trim(err_msg))
1198		−	return
1199			end select
1200		−	this%accuracy_method = accuracy_method_
1201
1202		−	end subroutine set_accuracy
1203			!###############################################################################
1204
1205
1206			!###############################################################################
1207		−	module subroutine reset(this)
1208			!! Reset the network
1209			implicit none
1210
1211			! Arguments
1212			class(network_type), intent(inout) :: this
1213			!! Instance of network
1214
1215		−	this%epoch = 0
1216		−	this%accuracy_val = 0._real32
1217		−	this%loss_val = huge(1._real32)
1218		−	this%batch_size = 0
1219		−	this%num_layers = 0
1220		−	this%num_outputs = 0
1221		−	if(allocated(this%optimiser)) deallocate(this%optimiser)
1222		−	call this%set_metrics(["loss"])
1223		−	if(allocated(this%model)) deallocate(this%model)
1224		−	if(allocated(this%loss)) deallocate(this%loss)
1225		−	this%get_accuracy => null()
1226
1227		−	if(allocated(this%vertex_order)) deallocate(this%vertex_order)
1228		−	if(allocated(this%leaf_vertices)) deallocate(this%leaf_vertices)
1229		−	if(allocated(this%root_vertices)) deallocate(this%root_vertices)
1230		−	this%auto_graph = graph_type(directed=.true.)
1231
1232		−	end subroutine reset
1233			!###############################################################################
1234
1235
1236			!###############################################################################
1237		−	module subroutine compile( &
1238			this, optimiser, loss_method, accuracy_method, &
1239			metrics, batch_size, verbose &
1240			)
1241			!! Compile the network
1242			implicit none
1243
1244			! Arguments
1245			class(network_type), intent(inout) :: this
1246			!! Instance of network
1247			class(base_optimiser_type), optional, intent(in) :: optimiser
1248			!! Optimiser to use for training
1249			class(*), optional, intent(in) :: loss_method
1250			!! Loss method
1251			character(*), optional, intent(in) :: accuracy_method
1252			!! Accuracy method
1253			class(*), dimension(..), optional, intent(in) :: metrics
1254			!! Metrics
1255			integer, optional, intent(in) :: batch_size
1256			!! Batch size
1257			integer, optional, intent(in) :: verbose
1258			!! Verbosity level
1259
1260			! Local variables
1261			integer :: i, j, k, child_id, parent_id, layer_id, num_inputs, input_rank
1262			!! Loop index
1263			integer :: parent_vertex, vertex_idx
1264			!! Vertex indices
1265			integer :: layer_rank, parent_rank, operator
1266			!! Ranks of layers
1267			integer :: verbose_ = 0
1268			!! Verbosity level
1269			logical :: use_graph_input = .false.
1270			!! Boolean whether to use graph input
1271			logical :: l_flatten_child, l_set_input_shape
1272			!! Booleans whether to flatten child or set input shape
1273		−	integer, dimension(:), allocatable :: input_shape, &
1274		−	child_vertices, parent_vertices, output_ranks, parent_ids
1275			!! Shapes of the input and output of the layers
1276		−	integer, dimension(:,:), allocatable :: merge_shape
1277			!! Shapes of the inputs to merge layers
1278			class(base_layer_type), allocatable :: &
1279		−	t_input_layer, t_flatten_layer, t_merge_layer
1280			!! Temporary input, flatten, and merge layers
1281
1282
1283			!---------------------------------------------------------------------------
1284			! Initialise optional arguments
1285			!---------------------------------------------------------------------------
1286		−	if(present(verbose)) verbose_ = verbose
1287
1288
1289			!---------------------------------------------------------------------------
1290			! Initialise metrics
1291			!---------------------------------------------------------------------------
1292		−	if(present(metrics)) call this%set_metrics(metrics)
1293
1294
1295			!---------------------------------------------------------------------------
1296			! Initialise loss and accuracy methods
1297			!---------------------------------------------------------------------------
1298		−	if(present(loss_method)) call this%set_loss(loss_method, verbose_)
1299		−	if(present(accuracy_method)) &
1300		−	call this%set_accuracy(accuracy_method, verbose_)
1301
1302
1303			!---------------------------------------------------------------------------
1304			! Check for input layers at root vertices
1305			!---------------------------------------------------------------------------
1306		−	this%auto_graph%directed = .true.
1307		−	call this%build_root_vertices()
1308		−	do i = 1, size(this%root_vertices)
1309		−	layer_id = this%auto_graph%vertex(this%root_vertices(i))%id
1310		−	if(.not.allocated(this%model(layer_id)%layer%input_shape))then
1311		−	call stop_program("input_shape of first layer not defined")
1312		−	return
1313			end if
1314		−	use_graph_input = .false.
1315		−	select type( root => this%model(layer_id)%layer)
1316			class is(input_layer_type)
1317		−	cycle
1318			class is(learnable_layer_type)
1319		−	input_shape = root%input_shape
1320		−	use_graph_input = root%use_graph_input
1321			class default
1322		−	input_shape = root%input_shape
1323			end select
1324			t_input_layer = input_layer_type(&
1325			input_shape = input_shape, &
1326			index = i, &
1327			use_graph_input = use_graph_input, &
1328			verbose=verbose_ &
1329		−	)
1330			call this%add( &
1331		−	t_input_layer, output_list = [ this%model(layer_id)%layer%id ] &
1332		−	)
1333			! NEED TO CALL layer%init?
1334		−	deallocate(input_shape)
1335		−	deallocate(t_input_layer)
1336		−	this%root_vertices(i) = this%num_layers
1337		−	if(i.eq.1)then
1338		−	do j = 1, this%auto_graph%num_edges
1339		−	if(this%auto_graph%edge(j)%index(1).eq.0) &
1340		−	this%auto_graph%edge(j)%index(1) = this%num_layers
1341			end do
1342			end if
1343			end do
1344		−	call this%auto_graph%generate_adjacency()
1345
1346
1347			!---------------------------------------------------------------------------
1348			! Identify whether input is graph type
1349			!---------------------------------------------------------------------------
1350		−	if( &
1351			this%model( &
1352		−	this%auto_graph%vertex(this%root_vertices(1))%id &
1353			)%layer%use_graph_input &
1354			)then
1355		−	this%use_graph_input = .true.
1356			else
1357		−	this%use_graph_input = .false.
1358			end if
1359
1360
1361			!---------------------------------------------------------------------------
1362			! Check for zero input rank layers
1363			!---------------------------------------------------------------------------
1364		−	do i = 1, size(this%auto_graph%vertex, dim = 1)
1365		−	layer_id = this%auto_graph%vertex(i)%id
1366		−	if(this%model(layer_id)%layer%input_rank.eq.0)then
1367			parent_ids = pack( &
1368		−	[ ( &
1369			this%auto_graph%vertex(j)%id, &
1370		−	j = 1, size(this%auto_graph%adjacency(:,i)) &
1371			) ], &
1372		−	this%auto_graph%adjacency(:,i) .ne. 0 &
1373		−	)
1374		−	if(size(parent_ids).eq.0) cycle
1375		−	output_ranks = [ ( this%model(parent_ids(j))%layer%output_rank, &
1376		−	j=1,size(parent_ids) ) ]
1377		−	if(any(output_ranks.ne.output_ranks(1)))then
1378		−	write(0,*) output_ranks
1379			call stop_program( &
1380		−	"input rank of layer "//trim(this%model(layer_id)%layer%name) // &
1381			" is zero, but multiple parents with different output ranks" &
1382		−	)
1383		−	return
1384			end if
1385		−	input_rank = this%model(parent_ids(1))%layer%output_rank
1386		−	call this%model(layer_id)%layer%set_rank( &
1387			input_rank = input_rank, &
1388			output_rank = input_rank &
1389		−	)
1390			end if
1391			end do
1392
1393
1394			!---------------------------------------------------------------------------
1395			! Check for required flatten layers
1396			!---------------------------------------------------------------------------
1397		−	i = 0
1398		−	flatten_loop: do
1399		−	i = i + 1
1400		−	if(i.gt.this%auto_graph%num_vertices) exit flatten_loop
1401		−	layer_id = this%auto_graph%vertex(i)%id
1402
1403			! get all child vertices
1404			child_vertices = pack( &
1405		−	[(j, j=1,size(this%auto_graph%adjacency(i,:)))], &
1406		−	this%auto_graph%adjacency(i,:) .ne. 0 &
1407		−	)
1408		−	child_loop: do j = 1, size(child_vertices)
1409			! Get layer ID (needed for add() function's output_list parameter)
1410		−	child_id = this%auto_graph%vertex(child_vertices(j))%id
1411		−	if(trim(this%model(layer_id)%layer%type).eq."flat") cycle child_loop
1412		−	if( this%model(layer_id)%layer%output_rank .eq. &
1413		−	this%model(child_id)%layer%input_rank ) cycle child_loop
1414		−	if(this%model(layer_id)%layer%output_rank.eq.0) cycle child_loop
1415
1416			! get all parent vertices of the child vertex
1417			parent_vertices = pack( &
1418		−	[(k, k=1,size(this%auto_graph%adjacency(:,child_vertices(j))))], &
1419		−	this%auto_graph%adjacency(:,child_vertices(j)) .ne. 0 &
1420		−	)
1421		−	l_flatten_child = .true.
1422		−	do k = 1, size(parent_vertices)
1423		−	parent_id = this%auto_graph%vertex(parent_vertices(k))%id
1424			!check if ranks match, rather than input and output shapes
1425		−	if( this%model(layer_id)%layer%output_rank .ne. &
1426			this%model(parent_id)%layer%input_rank &
1427		−	) l_flatten_child = .false.
1428			end do
1429			t_flatten_layer = flatten_layer_type( &
1430		−	input_rank = this%model(layer_id)%layer%output_rank &
1431		−	)
1432
1433		−	if(l_flatten_child)then
1434			! add flatten layer in the place of the child layer
1435		−	operator = this%auto_graph%edge( &
1436		−	this%auto_graph%adjacency(parent_vertices(1),child_vertices(j)) &
1437		−	)%id
1438			call this%auto_graph%remove_edges( &
1439			indices = [ &
1440		−	this%auto_graph%adjacency( &
1441		−	parent_vertices(:),child_vertices(j) &
1442			) &
1443			] &
1444		−	)
1445			call this%add( &
1446			t_flatten_layer, &
1447		−	input_list=[parent_vertices(:)], output_list=[child_id], &
1448			operator=operator &
1449		−	)
1450			else
1451			! add flatten layer between the current layer and the child layer
1452			call this%auto_graph%remove_edges( &
1453		−	indices = [this%auto_graph%adjacency(i,child_vertices(j))] &
1454		−	)
1455			call this%add( &
1456			t_flatten_layer, input_list = [i], output_list = [child_id], &
1457			operator=operator &
1458		−	)
1459			end if
1460		−	deallocate(t_flatten_layer)
1461		−	deallocate(child_vertices)
1462		−	cycle flatten_loop
1463			end do child_loop
1464		−	deallocate(child_vertices)
1465			end do flatten_loop
1466		−	call this%build_vertex_order()
1467
1468
1469			!---------------------------------------------------------------------------
1470			! Check for required merge layers
1471			!---------------------------------------------------------------------------
1472		−	i = 0
1473		−	merge_loop: do
1474		−	i = i + 1
1475		−	if(i.gt.this%auto_graph%num_vertices) exit merge_loop
1476		−	layer_id = this%auto_graph%vertex(i)%id
1477		−	if(this%model(layer_id)%layer%type.eq."merg") cycle merge_loop
1478
1479			! get all child vertices
1480			parent_vertices = pack( &
1481		−	[(j, j=1,size(this%auto_graph%adjacency(:,i)))], &
1482		−	this%auto_graph%adjacency(:,i) .ne. 0 &
1483		−	)
1484		−	if(size(parent_vertices).le.1) cycle merge_loop
1485
1486			! get edge id for merge layer
1487		−	operator = this%auto_graph%edge( &
1488		−	this%auto_graph%adjacency(parent_vertices(1),i) &
1489		−	)%id
1490
1491			! remove edges from parents to this layer
1492		−	do j = 1, size(parent_vertices)
1493			call this%auto_graph%remove_edges( &
1494		−	indices = [this%auto_graph%adjacency(parent_vertices(j),i)] &
1495		−	)
1496			end do
1497			parent_ids = &
1498		−	[ ( &
1499		−	this%auto_graph%vertex(parent_vertices(k))%id, &
1500			k = 1, size(parent_vertices) &
1501		−	) ]
1502		−	select case(operator)
1503			case(1) ! concatenate
1504			t_merge_layer = concat_layer_type( &
1505			input_layer_ids = parent_ids, &
1506		−	input_rank = this%model(layer_id)%layer%input_rank &
1507		−	)
1508			case(2) ! add
1509			t_merge_layer = add_layer_type( &
1510			input_layer_ids = parent_ids, &
1511		−	input_rank = this%model(layer_id)%layer%input_rank &
1512		−	)
1513			! case(3) ! multiply
1514			! t_merge_layer = multiply_layer_type( &
1515			! input_layer_ids = parent_vertices &
1516			! )
1517			case default
1518		−	write(0,*) "invalid merge operator: ", operator
1519		−	call stop_program("invalid merge operator")
1520		−	return
1521			end select
1522		−	t_merge_layer%use_graph_input = this%model(layer_id)%layer%use_graph_input
1523		−	t_merge_layer%use_graph_output = t_merge_layer%use_graph_input
1524			call this%add( &
1525			t_merge_layer, &
1526			input_list = parent_ids, &
1527			output_list = [layer_id] &
1528		−	)
1529		−	deallocate(t_merge_layer)
1530			end do merge_loop
1531		−	call this%build_vertex_order()
1532
1533
1534			! Update number of layers
1535			!---------------------------------------------------------------------------
1536		−	this%num_layers = size(this%model,dim=1)
1537
1538
1539
1540			!---------------------------------------------------------------------------
1541			! Initialise layers
1542			!---------------------------------------------------------------------------
1543		−	do i = 1, size(this%vertex_order, dim = 1)
1544		−	vertex_idx = this%vertex_order(i)
1545		−	layer_id = this%auto_graph%vertex(vertex_idx)%id
1546		−	if(allocated(this%model(layer_id)%layer%input_shape))then
1547		−	l_set_input_shape = .false.
1548			else
1549		−	l_set_input_shape = .true.
1550			end if
1551		−	if(l_set_input_shape)then
1552		−	layer_rank = this%model(layer_id)%layer%input_rank
1553		−	parent_rank = 0
1554
1555		−	select type( layer => this%model(layer_id)%layer )
1556			class is(merge_layer_type)
1557			! loop over all parent layers
1558			allocate( &
1559		−	merge_shape( &
1560		−	this%model(layer_id)%layer%input_rank, &
1561			size(layer%input_layer_ids) &
1562			) &
1563		−	)
1564		−	do k = 1, size(layer%input_layer_ids)
1565		−	merge_shape(:,k) = &
1566		−	this%model(layer%input_layer_ids(k))%layer%output_shape
1567			end do
1568		−	input_shape = layer%calc_input_shape(merge_shape)
1569		−	deallocate(merge_shape)
1570			class default
1571
1572			allocate( &
1573		−	input_shape(this%model(layer_id)%layer%input_rank), &
1574			source = 0 &
1575		−	)
1576		−	do j = 1, this%auto_graph%num_vertices
1577		−	if(this%auto_graph%adjacency(j,vertex_idx).eq.0) cycle
1578		−	parent_id = this%auto_graph%vertex(j)%id
1579		−	parent_rank = this%model(parent_id)%layer%output_rank
1580
1581		−	if(layer_rank .eq. parent_rank)then
1582		−	input_shape(:) = input_shape(:) + &
1583		−	this%model(parent_id)%layer%output_shape
1584		−	elseif(layer_rank .eq. 1)then
1585		−	input_shape(1) = input_shape(1) + &
1586		−	product( this%model(parent_id)%layer%output_shape )
1587			end if
1588			end do
1589			end select
1590		−	call this%model(layer_id)%layer%init( &
1591			input_shape = input_shape, &
1592			verbose = verbose_ &
1593		−	)
1594		−	deallocate(input_shape)
1595			end if
1596		−	if(verbose_.gt.0)then
1597		−	write(*,*) "layer: ", layer_id, this%model(layer_id)%layer%type
1598		−	write(*,*) this%model(layer_id)%layer%input_shape
1599		−	write(*,*) this%model(layer_id)%layer%output_shape
1600			end if
1601			end do
1602
1603
1604			! Set number of outputs
1605			!---------------------------------------------------------------------------
1606		−	this%num_outputs = 0
1607		−	call this%build_leaf_vertices()
1608		−	do i = 1, size(this%leaf_vertices,1)
1609			this%num_outputs = this%num_outputs + &
1610			product( &
1611		−	this%model( &
1612		−	this%auto_graph%vertex(this%leaf_vertices(i))%id &
1613			)%layer%output_shape &
1614		−	)
1615			end do
1616		−	if( &
1617			this%model( &
1618		−	this%auto_graph%vertex(this%leaf_vertices(1))%id &
1619			)%layer%use_graph_output &
1620			)then
1621		−	this%use_graph_output = .true.
1622			else
1623		−	this%use_graph_output = .false.
1624			end if
1625
1626
1627			!---------------------------------------------------------------------------
1628			! Confirm input_shape of each layer matches data going into it
1629			!---------------------------------------------------------------------------
1630		−	do i = 1, size(this%vertex_order, dim = 1)
1631		−	vertex_idx = this%vertex_order(i)
1632		−	layer_id = this%auto_graph%vertex(vertex_idx)%id
1633		−	if(this%model(layer_id)%layer%type.eq."inpt") cycle
1634
1635			! Get all parent vertices that feed into this layer
1636			parent_vertices = pack( &
1637		−	[(j, j=1,size(this%auto_graph%adjacency(:,vertex_idx)))], &
1638		−	this%auto_graph%adjacency(:,vertex_idx) .ne. 0 &
1639		−	)
1640		−	if(size(parent_vertices).eq.0) cycle
1641		−	select type( layer => this%model(layer_id)%layer )
1642			class is(merge_layer_type)
1643		−	operator = layer%merge_mode
1644			class default
1645		−	if(size(parent_vertices).gt.1)then
1646			call stop_program( &
1647			"layer "//trim(layer%name)// &
1648			" is not a merge layer but has multiple inputs" &
1649		−	)
1650		−	return
1651			end if
1652			end select
1653
1654			! Calculate expected input size from parent layers
1655		−	num_inputs = 0
1656		−	do j = 1, size(parent_vertices)
1657		−	parent_vertex = parent_vertices(j)
1658
1659		−	select case(operator)
1660			case(1) ! pointwise - all inputs should have same size
1661		−	if(num_inputs.eq.0)then
1662		−	if(this%model(layer_id)%layer%use_graph_input)then
1663		−	num_inputs = this%model(parent_vertex)%layer%output_shape(1)
1664			else
1665		−	num_inputs = product(this%model(parent_vertex)%layer%output_shape)
1666			end if
1667			end if
1668			case(2) ! concatenate
1669		−	if(this%model(layer_id)%layer%use_graph_input)then
1670			num_inputs = num_inputs + &
1671		−	this%model(parent_vertex)%layer%output_shape(1)
1672			else
1673			num_inputs = num_inputs + &
1674		−	product(this%model(parent_vertex)%layer%output_shape)
1675			end if
1676			end select
1677			end do
1678
1679			! Verify calculated input size matches layer's expected input size
1680		−	if(this%model(layer_id)%layer%use_graph_input)then
1681		−	if(num_inputs.ne.this%model(layer_id)%layer%input_shape(1) .and. &
1682			num_inputs.ne.0)then
1683		−	write(*,*) "Expected:", num_inputs, "Got:", &
1684		−	this%model(layer_id)%layer%input_shape(1)
1685			call stop_program( &
1686			"input_shape of layer "//&
1687		−	trim(this%model(layer_id)%layer%name)// &
1688			" does not match data going into it" &
1689		−	)
1690			end if
1691			else
1692		−	if(num_inputs.ne.product(this%model(layer_id)%layer%input_shape) .and. &
1693			num_inputs.ne.0)then
1694		−	write(*,*) "Expected:", num_inputs, "Got:", &
1695		−	product(this%model(layer_id)%layer%input_shape)
1696			call stop_program( &
1697			"input_shape of layer "//&
1698		−	trim(this%model(layer_id)%layer%name)// &
1699			" does not match data going into it" &
1700		−	)
1701			end if
1702			end if
1703
1704			end do
1705
1706			!---------------------------------------------------------------------------
1707			! Initialise optimiser
1708			!---------------------------------------------------------------------------
1709		−	this%num_params = this%get_num_params()
1710		−	if(present(optimiser))then
1711		−	this%optimiser = optimiser
1712			end if
1713		−	if(.not.allocated(this%optimiser))then
1714		−	call stop_program("No optimiser is defined for the network")
1715		−	return
1716			else
1717		−	call this%optimiser%init(num_params=this%num_params)
1718			end if
1719
1720
1721			!---------------------------------------------------------------------------
1722			! Set batch size, if provided
1723			!---------------------------------------------------------------------------
1724		−	if(present(batch_size)) this%batch_size = batch_size
1725
1726		−	end subroutine compile
1727			!###############################################################################
1728
1729
1730			!###############################################################################
1731		−	module subroutine set_batch_size(this, batch_size)
1732			!! Set the batch size for the network
1733			implicit none
1734
1735			! Arguments
1736			class(network_type), intent(inout) :: this
1737			!! Instance of network
1738			integer, intent(in) :: batch_size
1739			!! Batch size
1740
1741			! Local variables
1742			integer :: l
1743			!! Loop index
1744
1745
1746		−	this%batch_size = batch_size
1747
1748		−	end subroutine set_batch_size
1749			!###############################################################################
1750
1751
1752			!###############################################################################
1753		−	module subroutine reset_state(this)
1754			!! Reset the hidden state of all layers in the network
1755			implicit none
1756
1757			! Arguments
1758			class(network_type), intent(inout) :: this
1759			!! Instance of network
1760
1761			! Local variables
1762			integer :: l
1763			!! Loop index
1764
1765		−	do l = 1, size(this%model, dim = 1)
1766		−	select type( layer => this%model(l)%layer )
1767			class is(recurrent_layer_type)
1768		−	call layer%reset_state()
1769			end select
1770			end do
1771
1772		−	end subroutine reset_state
1773			!###############################################################################
1774
1775
1776			!###############################################################################
1777		−	module function layer_from_id(this, id) result(layer)
1778			!! Get layer from its ID
1779			implicit none
1780
1781			! Arguments
1782			class(network_type), intent(in), target :: this
1783			!! Instance of network
1784			integer, intent(in) :: id
1785			!! Layer ID
1786
1787			class(base_layer_type), pointer :: layer
1788			!! Layer
1789
1790			! Local variables
1791			integer :: i, itmp1
1792			!! Loop index
1793
1794		−	itmp1 = 0
1795		−	do i = 1, size(this%model, dim = 1)
1796		−	if(this%model(i)%layer%id.eq.id)then
1797		−	if(itmp1.ne.0)then
1798		−	call stop_program("multiple layers with same ID found")
1799		−	return
1800			end if
1801		−	layer => this%model(i)%layer
1802		−	itmp1 = itmp1 + 1
1803			end if
1804			end do
1805
1806		−	end function layer_from_id
1807			!###############################################################################
1808
1809
1810			!##############################################################################!
1811			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
1812			!##############################################################################!
1813
1814
1815			!###############################################################################
1816		−	module function get_sample_ptr( &
1817			input, start_index, end_index, batch_size &
1818			) result(sample_ptr)
1819			!! Get samples of batch size from a real array
1820			implicit none
1821
1822			! Arguments
1823			integer, intent(in) :: start_index, end_index
1824			!! Start and end indices
1825			integer, intent(in) :: batch_size
1826			!! Batch size
1827			real(real32), dimension(..), intent(in), target :: input
1828			!! Input array
1829
1830			real(real32), pointer :: sample_ptr(:,:)
1831			!! Pointer to sample
1832
1833
1834			select rank(input)
1835			rank(2)
1836		−	sample_ptr(1:size(input(:,1)),1:end_index-start_index+1) => &
1837		−	input(:,start_index:end_index)
1838			rank(3)
1839		−	sample_ptr(1:size(input(:,:,1)),1:end_index-start_index+1) => &
1840		−	input(:,:,start_index:end_index)
1841			rank(4)
1842		−	sample_ptr(1:size(input(:,:,:,1)),1:end_index-start_index+1) => &
1843		−	input(:,:,:,start_index:end_index)
1844			rank(5)
1845		−	sample_ptr(1:size(input(:,:,:,:,1)),1:end_index-start_index+1) => &
1846		−	input(:,:,:,:,start_index:end_index)
1847			rank(6)
1848		−	sample_ptr(1:size(input(:,:,:,:,:,1)),1:end_index-start_index+1) => &
1849		−	input(:,:,:,:,:,start_index:end_index)
1850			rank default
1851		−	sample_ptr => null()
1852			end select
1853
1854		−	end function get_sample_ptr
1855			!-------------------------------------------------------------------------------
1856		−	module function get_sample_array( &
1857		−	input, start_index, end_index, batch_size, as_graph &
1858			) result(sample)
1859			!! Get samples of batch size from a derived type array
1860			implicit none
1861
1862			! Arguments
1863			integer, intent(in) :: start_index, end_index
1864			!! Start and end indices
1865			integer, intent(in) :: batch_size
1866			!! Batch size
1867			class(array_type), dimension(:,:), intent(in) :: input
1868			!! Input array
1869			logical, intent(in) :: as_graph
1870			!! Boolean whether to treat the input as a graph
1871
1872			type(array_type), dimension(:,:), allocatable :: sample
1873			!! Sample array
1874
1875			! Local variables
1876			integer :: i, j
1877			!! Loop index
1878
1879		−	if(as_graph)then
1880		−	allocate(sample(size(input,1), batch_size))
1881		−	do i = 1, size(input,1)
1882		−	do j = start_index, end_index, 1
1883		−	sample(i, j - start_index + 1)%val = input(i, j)%val
1884			end do
1885			end do
1886			else
1887		−	allocate(sample(size(input,1), size(input,2)))
1888		−	do i = 1, size(input,1)
1889		−	do j = 1, size(input,2)
1890		−	call sample(i,j)%allocate(array_shape=[input(i,j)%shape, &
1891		−	end_index - start_index + 1])
1892		−	sample(i,j)%val = get_sample_ptr( &
1893		−	input(i,j)%val, start_index, end_index, batch_size &
1894		−	)
1895			end do
1896			end do
1897			end if
1898
1899		−	end function get_sample_array
1900			!-------------------------------------------------------------------------------
1901		−	module function get_sample_graph1d( &
1902		−	input, start_index, end_index, batch_size &
1903		−	) result(sample)
1904			!! Get samples of batch size from a graph
1905			implicit none
1906
1907			! Arguments
1908			integer, intent(in) :: start_index, end_index
1909			!! Start and end indices
1910			integer, intent(in) :: batch_size
1911			!! Batch size
1912			class(graph_type), dimension(:), intent(in) :: input
1913			!! Input array
1914
1915			type(graph_type), dimension(1, batch_size) :: sample
1916			!! Sample array
1917
1918		−	sample(1,1:batch_size) = input(start_index:end_index)
1919
1920		−	end function get_sample_graph1d
1921			!-------------------------------------------------------------------------------
1922		−	module function get_sample_graph2d( &
1923		−	input, start_index, end_index, batch_size &
1924		−	) result(sample)
1925			!! Get samples of batch size from a graph
1926			implicit none
1927
1928			! Arguments
1929			integer, intent(in) :: start_index, end_index
1930			!! Start and end indices
1931			integer, intent(in) :: batch_size
1932			!! Batch size
1933			class(graph_type), dimension(:,:), intent(in) :: input
1934			!! Input array
1935
1936		−	type(graph_type), dimension(size(input,1), batch_size) :: sample
1937			!! Sample array
1938
1939		−	sample(1:size(input,1),1:batch_size) = input(:,start_index:end_index)
1940
1941		−	end function get_sample_graph2d
1942			!###############################################################################
1943
1944
1945			!###############################################################################
1946		−	pure module function get_num_params(this) result(num_params)
1947			!! Get the number of learnable parameters in the network
1948			implicit none
1949
1950			! Arguments
1951			class(network_type), intent(in) :: this
1952			!! Instance of network
1953			integer :: num_params
1954			!! Number of parameters
1955
1956			! Local variables
1957			integer :: l, i
1958			!! Loop index
1959
1960		−	num_params = 0
1961		−	do l = 1, this%num_layers
1962		−	select type(current => this%model(l)%layer)
1963			class is(learnable_layer_type)
1964		−	do i = 1, size(current%params)
1965		−	num_params = num_params + size(current%params(i)%val, 1)
1966			end do
1967			end select
1968			end do
1969
1970		−	end function get_num_params
1971			!###############################################################################
1972
1973
1974			!###############################################################################
1975		−	pure module function get_params(this) result(params)
1976			!! Get learnable parameters
1977			implicit none
1978
1979			! Arguments
1980			class(network_type), intent(in) :: this
1981			!! Instance of network
1982			real(real32), dimension(this%num_params) :: params
1983			!! Parameters
1984
1985			! Local variables
1986			integer :: l, i, start_idx, end_idx
1987			!! Loop index
1988
1989		−	start_idx = 0
1990		−	end_idx = 0
1991		−	do l = 1, this%num_layers
1992		−	select type(current => this%model(l)%layer)
1993			class is(learnable_layer_type)
1994		−	do i = 1, size(current%params)
1995		−	start_idx = end_idx + 1
1996		−	end_idx = end_idx + size(current%params(i)%val, 1)
1997		−	params(start_idx:end_idx) = current%params(i)%val(:,1)
1998			end do
1999			end select
2000			end do
2001
2002		−	end function get_params
2003			!###############################################################################
2004
2005
2006			!###############################################################################
2007		−	module subroutine set_params(this, params)
2008			!! Set learnable parameters
2009			implicit none
2010
2011			! Arguments
2012			class(network_type), intent(inout) :: this
2013			!! Instance of network
2014			real(real32), dimension(this%num_params), intent(in) :: params
2015			!! Parameters
2016
2017			! Local variables
2018			integer :: l, i, start_idx, end_idx
2019			!! Loop index
2020
2021		−	start_idx = 0
2022		−	end_idx = 0
2023		−	do l = 1, this%num_layers
2024		−	select type(current => this%model(l)%layer)
2025			class is(learnable_layer_type)
2026		−	do i = 1, size(current%params)
2027		−	start_idx = end_idx + 1
2028		−	end_idx = end_idx + size(current%params(i)%val, 1)
2029		−	current%params(i)%val(:,1) = params(start_idx:end_idx)
2030			end do
2031			! call current%set_params(params(start_idx:end_idx))
2032			end select
2033			end do
2034
2035		−	end subroutine set_params
2036			!###############################################################################
2037
2038
2039			!###############################################################################
2040		−	pure module function get_gradients(this) result(gradients)
2041			!! Get gradients
2042			implicit none
2043
2044			! Arguments
2045			class(network_type), intent(in) :: this
2046			!! Instance of network
2047			real(real32), dimension(this%num_params) :: gradients
2048			!! Gradients
2049
2050			! Local variables
2051			integer :: l, i, start_idx, end_idx
2052			!! Loop index
2053
2054		−	start_idx = 0
2055		−	end_idx = 0
2056		−	do l = 1, this%num_layers
2057		−	select type(current => this%model(l)%layer)
2058			class is(learnable_layer_type)
2059		−	do i = 1, size(current%params)
2060		−	if(associated(current%params(i)%grad))then
2061		−	start_idx = end_idx + 1
2062		−	end_idx = end_idx + size(current%params(i)%val, 1)
2063		−	gradients(start_idx:end_idx) = [ &
2064		−	sum(current%params(i)%grad%val, dim=2) / &
2065		−	real(size(current%params(i)%grad%val, dim=2), real32) &
2066		−	]
2067			end if
2068			end do
2069			end select
2070			end do
2071		−	call this%optimiser%clip_dict%apply(size(gradients),gradients)
2072
2073		−	end function get_gradients
2074			!###############################################################################
2075
2076
2077			!###############################################################################
2078		−	module subroutine set_gradients(this, gradients)
2079			!! Set gradients
2080			implicit none
2081
2082			! Arguments
2083			class(network_type), intent(inout) :: this
2084			!! Instance of network
2085			real(real32), dimension(..), intent(in) :: gradients
2086			!! Gradients
2087
2088			! Local variables
2089			integer :: l, start_idx, end_idx
2090			!! Loop index
2091
2092		−	start_idx = 0
2093		−	end_idx = 0
2094		−	do l = 1, this%num_layers
2095		−	select type(current => this%model(l)%layer)
2096			class is(learnable_layer_type)
2097		−	start_idx = end_idx + 1
2098		−	end_idx = end_idx + current%num_params
2099		−	select rank(gradients)
2100			rank(0)
2101		−	call current%set_gradients(gradients)
2102			rank(1)
2103		−	call current%set_gradients(gradients(start_idx:end_idx))
2104			end select
2105			end select
2106			end do
2107
2108		−	end subroutine set_gradients
2109			!###############################################################################
2110
2111
2112			!###############################################################################
2113		−	module subroutine reset_gradients(this)
2114			!! Reset gradients
2115			implicit none
2116
2117			! Arguments
2118			class(network_type), intent(inout) :: this
2119			!! Instance of network
2120
2121			! Local variables
2122			integer :: l, i
2123			!! Loop index
2124
2125		−	do l = 1, this%num_layers
2126		−	select type(current => this%model(l)%layer)
2127			class is(learnable_layer_type)
2128		−	do i = 1, size(current%params)
2129		−	call current%params(i)%zero_grad()
2130			end do
2131			end select
2132			end do
2133
2134		−	end subroutine reset_gradients
2135			!###############################################################################
2136
2137
2138			!###############################################################################
2139		−	module function get_output_shape(this) result(output_shape)
2140			!! Get the output of the network
2141			implicit none
2142
2143			! Arguments
2144			class(network_type), intent(in) :: this
2145			!! Instance of network
2146			integer, dimension(2) :: output_shape
2147			!! Output shape
2148
2149			! Local variables
2150			integer :: i, layer_idx
2151			!! Loop indices
2152
2153
2154			! array data: [ layer idx, empty ]
2155			! graph data: [ vertex/edge idx, sample idx]
2156
2157		−	if(this%use_graph_output)then
2158		−	output_shape = [2, this%batch_size]
2159		−	do i = 1, size(this%leaf_vertices,1), 1
2160		−	layer_idx = this%auto_graph%vertex(this%leaf_vertices(i))%id
2161		−	if(size(this%model(layer_idx)%layer%output,2).ne.this%batch_size)then
2162			call stop_program( &
2163			"Inconsistent batch size in output layers" &
2164		−	)
2165		−	return
2166			end if
2167			output_shape(1) = output_shape(1) + &
2168		−	size( this%model(layer_idx)%layer%output, 1 )
2169			end do
2170			else
2171		−	output_shape = [0, 1]
2172		−	do i = 1, size(this%leaf_vertices,1)
2173		−	layer_idx = this%auto_graph%vertex(this%leaf_vertices(i))%id
2174		−	if(size(this%model(layer_idx)%layer%output,2).ne.1)then
2175			call stop_program( &
2176			"Inconsistent size of dimension 2 in output layers" &
2177		−	)
2178		−	return
2179			end if
2180			output_shape(1) = &
2181		−	output_shape(1) + size( this%model(layer_idx)%layer%output, 1 )
2182			end do
2183			end if
2184
2185		−	end function get_output_shape
2186			!-------------------------------------------------------------------------------
2187		−	module function get_output(this) result(output)
2188			!! Get the output of the network
2189			implicit none
2190
2191			! Arguments
2192			class(network_type), intent(in) :: this
2193			!! Instance of network
2194			type(array_type), dimension(:,:), allocatable :: output
2195			!! Output
2196
2197			! Local variables
2198			integer :: i, start_idx, end_idx, layer_idx, output_id
2199			!! Loop indices
2200			integer, dimension(2) :: output_shape
2201			!! Output shape
2202		−	integer, dimension(this%num_outputs) :: output_ids
2203			!! Output IDs
2204
2205
2206			! array data: [ layer idx, empty ]
2207			! graph data: [ vertex/edge idx, sample idx]
2208
2209		−	if(this%use_graph_output)then
2210		−	output_shape = [2, this%batch_size]
2211		−	do i = 1, size(this%leaf_vertices,1), 1
2212		−	layer_idx = this%auto_graph%vertex(this%leaf_vertices(i))%id
2213		−	if(size(this%model(layer_idx)%layer%output,2).ne.this%batch_size)then
2214			call stop_program( &
2215			"Inconsistent batch size in output layers" &
2216		−	)
2217		−	return
2218			end if
2219		−	output_id = this%model(layer_idx)%layer%id
2220		−	output_ids(output_id) = size( this%model(layer_idx)%layer%output, 1 )
2221		−	output_shape(1) = output_shape(1) + output_ids(output_id)
2222			end do
2223		−	allocate(output(output_shape(1), output_shape(2)))
2224		−	do i = 1, size(this%leaf_vertices,1)
2225		−	layer_idx = this%auto_graph%vertex(this%leaf_vertices(i))%id
2226		−	output_id = sum(output_ids(1:this%model(layer_idx)%layer%id-1)) + 1
2227		−	output(output_id,:) = this%model(layer_idx)%layer%output(1,:)
2228		−	if(output_ids(this%model(layer_idx)%layer%id).gt.1)then
2229		−	output(output_id+1,:) = this%model(layer_idx)%layer%output(2,:)
2230			end if
2231			end do
2232			else
2233		−	output_shape = [0, 1]
2234		−	do i = 1, size(this%leaf_vertices,1)
2235		−	layer_idx = this%auto_graph%vertex(this%leaf_vertices(i))%id
2236		−	if(size(this%model(layer_idx)%layer%output,2).ne.1)then
2237			call stop_program( &
2238			"Inconsistent size of dimension 2 in output layers" &
2239		−	)
2240		−	return
2241			end if
2242			output_shape(1) = &
2243		−	output_shape(1) + size( this%model(layer_idx)%layer%output, 1 )
2244		−	output_id = this%model(layer_idx)%layer%id
2245		−	output_ids(output_id) = size( this%model(layer_idx)%layer%output, 1 )
2246			end do
2247		−	allocate(output(output_shape(1), output_shape(2)))
2248		−	start_idx = 1
2249		−	end_idx = 0
2250		−	do i = 1, size(this%leaf_vertices,1)
2251		−	layer_idx = this%auto_graph%vertex(this%leaf_vertices(i))%id
2252		−	output_id = this%model(layer_idx)%layer%id
2253		−	end_idx = end_idx + output_ids(output_id)
2254		−	output(start_idx:end_idx,1) = this%model(layer_idx)%layer%output(:,1)
2255		−	start_idx = end_idx + 1
2256			end do
2257			end if
2258
2259		−	end function get_output
2260			!-------------------------------------------------------------------------------
2261		−	module subroutine extract_output_real(this, output)
2262			!! Get the output of the network as real array
2263			implicit none
2264
2265			! Arguments
2266			class(network_type), intent(in) :: this
2267			! Instance of network
2268			real(real32), dimension(..), allocatable, intent(out) :: output
2269			!! Output
2270
2271			! Local variables
2272			integer :: layer_id
2273			!! Layer ID
2274			character(len=10) :: rank_str
2275			!! String for rank
2276
2277			! check if number of leaf vertices is 1
2278		−	if(size(this%leaf_vertices,1).gt.1)then
2279			call print_warning("Output extraction to real array only works for single &
2280		−	&output networks")
2281		−	return
2282			end if
2283
2284			! Get output from the first (and only) leaf vertex
2285		−	layer_id = this%auto_graph%vertex(this%leaf_vertices(1))%id
2286		−	call this%model(layer_id)%layer%output(1,1)%extract(output)
2287
2288			end subroutine extract_output_real
2289			!###############################################################################
2290
2291
2292			!##############################################################################!
2293			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
2294			!##############################################################################!
2295
2296
2297			!###############################################################################
2298		−	module function accuracy_eval(this, output, start_index, end_index) &
2299			result(accuracy)
2300			!! Get the loss for the output
2301			implicit none
2302
2303			! Arguments
2304			class(network_type), intent(in) :: this
2305			!! Instance of network
2306			class(*), dimension(:,:), intent(in) :: output
2307			!! Output
2308			integer, intent(in) :: start_index, end_index
2309			!! Start and end batch indices
2310
2311			real(real32) :: accuracy
2312			!! Loss value
2313
2314			! Local variables
2315			integer :: s, s_idx
2316			!! Loop index
2317
2318		−	accuracy = 0._real32
2319		−	select type(output)
2320			type is(graph_type)
2321		−	do s = start_index, end_index, 1
2322		−	s_idx = s - start_index + 1
2323			accuracy = accuracy + sum( this%get_accuracy( &
2324		−	this%model(this%leaf_vertices(1))%layer%output(1,s_idx)%val, &
2325		−	output(1,s)%vertex_features &
2326		−	) ) / output(1,s)%num_vertices
2327		−	if( &
2328		−	this%model(this%leaf_vertices(1))%layer%output_shape(2).gt.0 &
2329		−	)then
2330			accuracy = accuracy + sum( this%get_accuracy( &
2331		−	this%model(this%leaf_vertices(1))%layer%output(2,s_idx)%val, &
2332		−	output(1,s)%edge_features &
2333		−	) ) / output(1,s)%num_edges
2334			end if
2335			end do
2336			type is(real(real32))
2337			accuracy = sum( &
2338			this%get_accuracy( &
2339		−	this%model(this%leaf_vertices(1))%layer%output(1,1)%val, &
2340		−	output(:,start_index:end_index:1) &
2341		−	))
2342			type is(integer)
2343			accuracy = sum( &
2344			this%get_accuracy( &
2345		−	this%model(this%leaf_vertices(1))%layer%output(1,1)%val, &
2346		−	real(output(:,start_index:end_index:1),real32) &
2347		−	))
2348			class is(array_type)
2349			accuracy = sum( &
2350			this%get_accuracy( &
2351		−	this%model(this%leaf_vertices(1))%layer%output(1,1)%val, &
2352		−	output(1,1)%val(:,start_index:end_index:1) &
2353		−	))
2354			end select
2355
2356		−	end function accuracy_eval
2357			!###############################################################################
2358
2359
2360			!###############################################################################
2361		−	module function loss_eval(this, start_index, end_index) result(loss)
2362			!! Get the loss for the output
2363			implicit none
2364
2365			! Arguments
2366			class(network_type), intent(inout), target :: this
2367			!! Instance of network
2368			integer, intent(in) :: start_index, end_index
2369			!! Start and end batch indices
2370
2371			type(array_type), pointer :: loss
2372			!! Loss value
2373
2374			! Local variables
2375			integer :: i, s
2376			!! Loop index
2377		−	type(array_type), pointer :: expected(:,:), predicted(:,:)
2378
2379
2380		−	if(this%use_graph_output)then
2381			expected(1:2, 1: end_index - start_index + 1) => &
2382		−	this%expected_array( :, start_index:end_index )
2383			else
2384		−	allocate(expected(size(this%expected_array,1), size(this%expected_array,2)))
2385		−	do s = 1, size(this%expected_array,2)
2386		−	do i = 1, size(this%expected_array,1)
2387		−	call expected(i,s)%allocate( &
2388		−	array_shape = [ &
2389			this%expected_array(i,s)%shape, &
2390		−	size(this%expected_array(i,s)%val,2) &
2391			] &
2392		−	)
2393		−	expected(i,s)%val = this%expected_array(i,s)%val(:, &
2394		−	start_index:end_index:1)
2395			end do
2396			end do
2397			end if
2398
2399		−	predicted => this%model(this%leaf_vertices(1))%layer%output
2400			loss => this%loss%compute( &
2401			predicted, &
2402			expected &
2403		−	)
2404
2405		−	end function loss_eval
2406			!###############################################################################
2407
2408
2409			!##############################################################################!
2410			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
2411			!##############################################################################!
2412
2413
2414			!###############################################################################
2415		−	module subroutine forward_generic2d(this, input)
2416			!! Forward pass for array derived type input
2417			implicit none
2418
2419			! Arguments
2420			class(network_type), intent(inout), target :: this
2421			!! Instance of network
2422			class(*), dimension(:,:), intent(in) :: input
2423			!! Input
2424
2425			! Local variables
2426			integer :: l, i, j, vertex_idx, layer_id, parent_id
2427			!! Loop index and vertex index
2428			integer :: input_idx
2429			!! Index of input layer
2430			integer :: num_input_layers
2431			!! Number of input layers
2432		−	type(array_type), pointer :: input_ptr(:,:) => null()
2433		−	type(array_ptr_type), dimension(:), allocatable :: input_list
2434
2435
2436		−	select type(input)
2437			type is(graph_type)
2438		−	do j = 1, this%batch_size
2439		−	if(any(input(1,j)%adj_ja(1,:).gt.input(1,j)%num_vertices))then
2440			call stop_program( &
2441			"input graph has more vertices than expected" &
2442		−	)
2443			end if
2444			end do
2445			end select
2446			! Forward pass
2447			!---------------------------------------------------------------------------
2448		−	do l = 1, size(this%vertex_order,1)
2449		−	vertex_idx = this%vertex_order(l)
2450		−	layer_id = this%auto_graph%vertex(vertex_idx)%id
2451		−	num_input_layers = count(this%auto_graph%adjacency(:,vertex_idx).gt.0)
2452		−	if(num_input_layers.eq.0)then
2453		−	select type(layer => this%model(layer_id)%layer)
2454			class is(input_layer_type)
2455		−	select type(input)
2456			type is(graph_type)
2457		−	call layer%set_input_graph( [ input(layer%index, :) ] )
2458		−	cycle
2459			class is(array_type)
2460		−	call layer%forward(input(layer%index:layer%index,:))
2461		−	do concurrent(i=1:size(layer%output,1), j=1:size(layer%output,2))
2462		−	call layer%output(i,j)%set_requires_grad(.false.)
2463			end do
2464		−	cycle
2465			type is(real(real32))
2466		−	allocate(input_ptr(1,1))
2467		−	call input_ptr(1,1)%allocate(shape(input))
2468		−	call input_ptr(1,1)%set(input)
2469		−	call layer%forward(input_ptr)
2470		−	call layer%output(1,1)%set_requires_grad(.false.)
2471		−	deallocate(input_ptr)
2472		−	input_ptr => null()
2473		−	cycle
2474			class default
2475			call stop_program( &
2476			"input type for layer "// &
2477			trim(layer%name) // &
2478			" is not supported" &
2479		−	)
2480			end select
2481			class default
2482		−	return
2483			end select
2484		−	elseif(num_input_layers.eq.1)then
2485		−	j = maxloc(this%auto_graph%adjacency(:,vertex_idx),dim=1)
2486		−	input_idx = findloc(this%root_vertices, j, dim=1)
2487		−	parent_id = this%auto_graph%vertex(j)%id
2488		−	input_ptr => this%model(parent_id)%layer%output
2489		−	select type(input)
2490			type is(graph_type)
2491		−	call this%model(layer_id)%layer%set_graph( [ input(1,:) ] )
2492			end select
2493			else
2494		−	allocate(input_list(num_input_layers))
2495		−	i = 0
2496		−	do j = 1, size(this%vertex_order,1)
2497		−	if(this%auto_graph%adjacency(j,vertex_idx).gt.0)then
2498		−	i = i + 1
2499		−	parent_id = this%auto_graph%vertex(j)%id
2500		−	input_list(i)%array => this%model(parent_id)%layer%output
2501			end if
2502			end do
2503			end if
2504
2505		−	select type(layer => this%model(layer_id)%layer)
2506			class is(merge_layer_type)
2507		−	call layer%combine(input_list)
2508		−	deallocate(input_list)
2509			class default
2510		−	call layer%forward(input_ptr)
2511		−	input_ptr => null()
2512			end select
2513
2514			end do
2515
2516		−	end subroutine forward_generic2d
2517			!-------------------------------------------------------------------------------
2518		−	module function forward_eval(this, input) result(output)
2519			!! Forward pass for evaluation
2520			implicit none
2521
2522			! Arguments
2523			class(network_type), intent(inout), target :: this
2524			!! Instance of network
2525			class(*), dimension(:,:), intent(in) :: input
2526			!! Input
2527
2528			type(array_type), pointer :: output(:,:)
2529			!! Output
2530
2531		−	call this%forward(input)
2532		−	output => this%model(this%leaf_vertices(1))%layer%output
2533
2534		−	end function forward_eval
2535			!-------------------------------------------------------------------------------
2536		−	module function forward_eval_multi(this, input) result(output)
2537			!! Forward pass for evaluation
2538			implicit none
2539
2540			! Arguments
2541			class(network_type), intent(inout), target :: this
2542			!! Instance of network
2543			class(*), dimension(:,:), intent(in) :: input
2544			!! Input
2545
2546			type(array_ptr_type), pointer :: output(:)
2547			!! Output
2548
2549			! Local variables
2550			integer :: l
2551			!! Loop index
2552
2553		−	call this%forward(input)
2554		−	allocate(output(size(this%leaf_vertices,1)))
2555		−	do l = 1, size(this%leaf_vertices,1)
2556		−	output(l)%array => this%model(this%leaf_vertices(l))%layer%output
2557			end do
2558
2559		−	end function forward_eval_multi
2560			!###############################################################################
2561
2562
2563			!###############################################################################
2564		−	module subroutine update(this)
2565			!! Update the network
2566			implicit none
2567
2568			! Arguments
2569			class(network_type), intent(inout) :: this
2570			!! Instance of network
2571		−	real(real32), dimension(this%num_params) :: params, gradients
2572			!! Parameters and gradients
2573
2574			! Local variables
2575			integer :: l, i, start_idx, end_idx
2576			!! Loop index
2577
2578
2579			!---------------------------------------------------------------------------
2580			! Increment optimiser iteration counter
2581			!---------------------------------------------------------------------------
2582		−	if(this%optimiser%lr_decay%iterate_per_epoch)then
2583		−	if(this%epoch.gt.this%optimiser%epoch)then
2584		−	this%optimiser%epoch = this%epoch
2585		−	this%optimiser%iter = this%optimiser%iter + 1
2586			end if
2587			else
2588		−	this%optimiser%iter = this%optimiser%iter + 1
2589			end if
2590
2591
2592			!---------------------------------------------------------------------------
2593			! Get learnable parameters and gradients
2594			!---------------------------------------------------------------------------
2595		−	start_idx = 0
2596		−	end_idx = 0
2597		−	do l = 1, this%num_layers
2598		−	select type(current => this%model(l)%layer)
2599			class is(learnable_layer_type)
2600		−	do i = 1, size(current%params)
2601		−	start_idx = end_idx + 1
2602		−	end_idx = end_idx + size(current%params(i)%val, 1)
2603		−	params(start_idx:end_idx) = current%params(i)%val(:,1)
2604		−	if(.not.associated(current%params(i)%grad))then
2605			call stop_program( &
2606			"Gradient not allocated for parameters in layer "// &
2607			trim(current%name) // &
2608			"." &
2609		−	)
2610			end if
2611		−	select case(size(current%params(i)%grad%val,2))
2612			case(1)
2613		−	gradients(start_idx:end_idx) = current%params(i)%grad%val(:,1)
2614			case default
2615		−	gradients(start_idx:end_idx) = [ &
2616		−	sum(current%params(i)%grad%val, dim=2) / &
2617		−	real(size(current%params(i)%grad%val, dim=2), real32) &
2618		−	]
2619			end select
2620			end do
2621			end select
2622			end do
2623			! have an if statement of whether to apply clipping to to gradients of
2624			! each layer individually or collectively to the all gradients at once
2625		−	call this%optimiser%clip_dict%apply(size(gradients),gradients)
2626
2627			!---------------------------------------------------------------------------
2628			! Update layers of learnable layer types
2629			!---------------------------------------------------------------------------
2630		−	call this%optimiser%minimise(params, gradients)
2631		−	call this%set_params(params)
2632		−	call this%reset_gradients()
2633
2634		−	end subroutine update
2635			!###############################################################################
2636
2637
2638			!##############################################################################!
2639			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
2640			!##############################################################################!
2641
2642
2643			!###############################################################################
2644		−	module subroutine nullify_graph(this)
2645			!! Nullify the input graph
2646			implicit none
2647
2648			! Arguments
2649			class(network_type), intent(inout) :: this
2650			!! Instance of network
2651
2652			! Local variables
2653			integer :: l
2654
2655		−	do l = 1, this%num_layers
2656		−	call this%model(l)%layer%nullify_graph()
2657			end do
2658
2659		−	end subroutine nullify_graph
2660			!###############################################################################
2661
2662
2663			!##############################################################################!
2664			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
2665			!##############################################################################!
2666
2667
2668			!###############################################################################
2669		−	module function save_input_to_network( this, input ) result(num_samples)
2670			!! Save input to network
2671			implicit none
2672
2673			! Arguments
2674			class(network_type), intent(inout) :: this
2675			!! Instance of network
2676			class(*), dimension(..), intent(in) :: input
2677			!! Input
2678
2679			integer :: num_samples
2680			!! Number of samples
2681
2682			! Local variables
2683			integer :: i, j, l, ip, input_rank, num_inputs
2684			!! Loop index
2685			integer :: num_input_layers
2686			!! Number of input layers
2687			logical :: l_valid_rank_type
2688			!! Boolean whether rank type is valid
2689			character(256) :: err_msg
2690			!! Error message
2691
2692		−	num_samples = get_num_samples(this, input)
2693		−	num_input_layers = size(this%root_vertices, 1)
2694		−	if(allocated(this%input_array))then
2695		−	do i = 1, size(this%input_array, 1)
2696		−	do j = 1, size(this%input_array, 2)
2697		−	call this%input_array(i,j)%deallocate()
2698			end do
2699			end do
2700		−	deallocate(this%input_array)
2701			end if
2702		−	if(allocated(this%input_graph)) deallocate(this%input_graph)
2703
2704			! Determine the rank of the input
2705			!---------------------------------------------------------------------------
2706			select rank(input)
2707			rank(0)
2708			rank(1)
2709			rank(2)
2710		−	select type(input)
2711			class is(array_type)
2712		−	num_inputs = size(input(1,1)%val, 1)
2713		−	allocate(this%input_array(size(input,1), size(input,2)))
2714		−	do i = 1, size(input,1)
2715		−	do j = 1, size(input,2)
2716		−	call this%input_array(i,j)%assign_shallow(input(i,j))
2717			end do
2718			end do
2719		−	return
2720			class default
2721		−	input_rank = rank(input)
2722		−	num_inputs = size(input) / num_samples
2723		−	allocate(this%input_array(1,1))
2724		−	call this%input_array(1,1)%allocate(array_shape=[num_inputs, num_samples])
2725			end select
2726			rank default
2727		−	input_rank = rank(input)
2728		−	num_inputs = size(input) / num_samples
2729		−	allocate(this%input_array(1,1))
2730		−	call this%input_array(1,1)%allocate(array_shape=shape(input))
2731			end select
2732		−	l_valid_rank_type = .false.
2733
2734
2735			! Process input based on its rank
2736			!---------------------------------------------------------------------------
2737			rank_select: select rank(input)
2738			rank(0)
2739			select type(input)
2740		−	type is(real); exit rank_select
2741		−	class default; l_valid_rank_type = .true.
2742			end select
2743		−	if(num_input_layers.ne.1)then
2744			call stop_program( &
2745			"number of input arrays does not match expected number of &
2746			&input layers" &
2747		−	)
2748		−	return
2749			end if
2750		−	select type(input)
2751			class is(array_type)
2752		−	allocate(this%input_array(1,1))
2753		−	call handle_array_type(input, this%input_array(1,1), num_samples)
2754			type is(array_ptr_type)
2755		−	allocate(this%input_array(size(input%array,1), size(input%array,2)))
2756		−	do i = 1, size(input%array,1)
2757		−	do j = 1, size(input%array,2)
2758			call handle_array_type( &
2759		−	input%array(i,j), this%input_array(i,j), num_samples &
2760		−	)
2761			end do
2762			end do
2763			end select
2764			rank(1)
2765			select type(input)
2766			type is(real(real32))
2767		−	exit rank_select
2768			type is(graph_type)
2769		−	allocate(this%input_graph(num_input_layers, num_samples))
2770		−	this%input_graph(1,:) = input(:)
2771		−	return
2772			class default
2773		−	l_valid_rank_type = .true.
2774			end select
2775		−	if(size(input,1).ne.num_input_layers)then
2776			call stop_program( &
2777			"number of input arrays does not match expected number of &
2778			&input layers" &
2779		−	)
2780		−	return
2781			end if
2782		−	select type(input)
2783			class is(array_type)
2784		−	allocate(this%input_array(1,size(input,1)))
2785		−	do l = 1, size(input,1)
2786		−	call handle_array_type(input(l), this%input_array(1,l), num_samples)
2787			end do
2788			type is(array_ptr_type)
2789		−	call stop_program("Use of array_ptr_type with rank 1 input not yet supported")
2790		−	return
2791			! ip = 0
2792			! do l = 1, size(input,1)
2793			! do i = 1, size(input%array,1)
2794			! ip = ip + 1
2795			! do j = 1, size(input%array,2)
2796			! call handle_array_type( &
2797			! input(l)%array(i,j), this%input_array(ip,j), num_samples &
2798			! )
2799			! end do
2800			! end do
2801			! end do
2802			end select
2803			rank(2)
2804		−	select type(input)
2805			type is(real(real32))
2806		−	this%input_array(1,1)%val = reshape(input, [num_inputs, num_samples])
2807		−	l_valid_rank_type = .true.
2808			type is(graph_type)
2809		−	num_samples = size(input, dim=2)
2810		−	allocate(this%input_graph(num_input_layers, num_samples))
2811		−	this%input_graph(:,:) = input(:,:)
2812		−	return
2813			type is(array_type)
2814		−	call stop_program("SHOULD NOT GET HERE")
2815		−	this%input_array = input
2816		−	l_valid_rank_type = .true.
2817			end select
2818			rank(3)
2819		−	select type(input)
2820			type is(real(real32))
2821		−	call this%input_array(1,1)%set(input)
2822		−	l_valid_rank_type = .true.
2823			end select
2824			rank(4)
2825		−	select type(input)
2826			type is(real(real32))
2827		−	call this%input_array(1,1)%set(input)
2828		−	l_valid_rank_type = .true.
2829			end select
2830			rank(5)
2831		−	select type(input)
2832			type is(real(real32))
2833		−	call this%input_array(1,1)%set(input)
2834		−	l_valid_rank_type = .true.
2835			end select
2836			end select rank_select
2837
2838		−	if(.not.l_valid_rank_type)then
2839		✗	write(err_msg,'("Unknown input type for rank ",I0)') input_rank
2840		✗	call stop_program(err_msg)
2841		✗	return
2842			end if
2843
2844			contains
2845
2846		6	function get_num_samples(network, input) result(num_samples)
2847			implicit none
2848			!! Get the number of samples in the input
2849
2850			! Arguments
2851			type(network_type), intent(in) :: network
2852			!! Instance of network
2853			class(*), dimension(..), intent(in) :: input
2854			!! Input
2855			integer :: num_samples
2856			!! Number of samples
2857
2858			! Local variables
2859			integer :: layer_id
2860			!! Layer ID
2861			logical :: use_graph_input
2862			!! Whether to use graph input
2863
2864		6	num_samples = 0
2865	4/8 ✗ Branch 0 (2→3) not taken. ✓ Branch 1 (2→4) taken 6 times. ✗ Branch 2 (4→5) not taken. ✓ Branch 3 (4→6) taken 6 times. ✗ Branch 4 (6→7) not taken. ✓ Branch 5 (6→8) taken 6 times. ✗ Branch 6 (8→9) not taken. ✓ Branch 7 (8→10) taken 6 times.	6	layer_id = network%auto_graph%vertex(network%root_vertices(1))%id
2866	2/4 ✗ Branch 0 (10→11) not taken. ✓ Branch 1 (10→12) taken 6 times. ✗ Branch 2 (12→13) not taken. ✓ Branch 3 (12→14) taken 6 times.	6	use_graph_input = network%model(layer_id)%layer%use_graph_input
2867			select rank(input)
2868			rank(0)
2869		✗	select type(input)
2870			class is(array_type)
2871		✗	num_samples = size(input%val, 2)
2872			class is(array_ptr_type)
2873		✗	num_samples = size(input%array(1,1)%val, 2)
2874			class default
2875		✗	call stop_program("Unknown input type in get_num_samples for rank 0")
2876		✗	return
2877			end select
2878			rank(1)
2879		✗	select type(input)
2880			class is(array_type)
2881		✗	if(use_graph_input)then
2882		✗	num_samples = size(input)
2883			else
2884		✗	num_samples = size(input(1)%val, 2)
2885			end if
2886			class is(array_ptr_type)
2887		✗	if(use_graph_input)then
2888		✗	num_samples = size(input(1)%array, 2)
2889			else
2890		✗	num_samples = size(input(1)%array(1,1)%val, 2)
2891			end if
2892			class is(graph_type)
2893		✗	num_samples = size(input, dim=1)
2894			type is(real)
2895		✗	num_samples = size(input, rank(input))
2896			class default
2897		✗	call stop_program("Unknown input type in get_num_samples for rank 1")
2898		✗	return
2899			end select
2900			rank(2)
2901	1/2 ✓ Branch 0 (111→112) taken 6 times. ✗ Branch 1 (111→114) not taken.	6	select type(input)
2902			class is(array_type)
2903	1/2 ✗ Branch 0 (92→93) not taken. ✓ Branch 1 (92→94) taken 1 times.	2	if(use_graph_input)then
2904		✗	num_samples = size(input, 2)
2905			else
2906	4/8 ✗ Branch 0 (94→95) not taken. ✓ Branch 1 (94→96) taken 1 times. ✗ Branch 2 (96→97) not taken. ✓ Branch 3 (96→98) taken 1 times. ✗ Branch 4 (98→99) not taken. ✓ Branch 5 (98→100) taken 1 times. ✗ Branch 6 (100→101) not taken. ✓ Branch 7 (100→102) taken 1 times.	1	num_samples = size(input(1,1)%val, 2)
2907			end if
2908			class is(graph_type)
2909		4	num_samples = size(input, dim=2)
2910			type is(real)
2911		1	num_samples = size(input, rank(input))
2912			class default
2913		✗	call stop_program("Unknown input type in get_num_samples for rank 2")
2914		✗	return
2915			end select
2916			rank(3)
2917		✗	select type(input)
2918			type is(real)
2919		✗	num_samples = size(input, rank(input))
2920			class default
2921		✗	call stop_program("Unknown input type in get_num_samples for rank 3")
2922		✗	return
2923			end select
2924			rank(4)
2925		✗	select type(input)
2926			type is(real)
2927		✗	num_samples = size(input, rank(input))
2928			class default
2929		✗	call stop_program("Unknown input type in get_num_samples for rank 4")
2930		✗	return
2931			end select
2932			rank(5)
2933		✗	select type(input)
2934			type is(real)
2935		✗	num_samples = size(input, rank(input))
2936			class default
2937		✗	call stop_program("Unknown input type in get_num_samples for rank 5")
2938		✗	return
2939			end select
2940			rank default
2941		✗	call stop_program("Unknown input rank in get_num_samples")
2942		✗	return
2943			end select
2944
2945		6	end function get_num_samples
2946
2947
2948		✗	subroutine handle_array_type(input, output, num_samples)
2949			!! Handle array type input
2950
2951			! Arguments
2952			class(array_type), intent(in) :: input
2953			!! Input
2954			type(array_type), intent(out) :: output
2955			!! Output
2956			integer, intent(in) :: num_samples
2957			!! Number of samples
2958
2959		✗	if(size(input%val,2).ne.num_samples)then
2960		✗	call stop_program("number of samples in input arrays do not match")
2961		✗	return
2962			end if
2963			call output%allocate( array_shape = &
2964		✗	[ product(input%shape(1:input%rank)), num_samples ] &
2965		✗	)
2966		✗	output%val = input%val
2967			end subroutine handle_array_type
2968
2969			end function save_input_to_network
2970			!-------------------------------------------------------------------------------
2971		−	module subroutine save_output_to_network( this, output )
2972			!! Save output to network
2973			implicit none
2974
2975			! Arguments
2976			class(network_type), intent(inout) :: this
2977			!! Instance of network
2978			class(*), dimension(:,:), intent(in) :: output
2979			!! Output
2980
2981			! Local variables
2982			integer :: i, j, s
2983			!! Loop indices
2984
2985		−	if(allocated(this%expected_array))then
2986		−	do i = 1, size(this%expected_array, 1)
2987		−	do j = 1, size(this%expected_array, 2)
2988		−	call this%expected_array(i,j)%deallocate()
2989			end do
2990			end do
2991		−	deallocate(this%expected_array)
2992			end if
2993
2994		−	select type(output)
2995			type is(graph_type)
2996		−	allocate(this%expected_array(2,size(output,2)))
2997		−	do s = 1, size(output,2)
2998		−	if(this%expected_array(1,s)%allocated) &
2999		−	call this%expected_array(1,s)%deallocate()
3000		−	if(this%expected_array(2,s)%allocated) &
3001		−	call this%expected_array(2,s)%deallocate()
3002		−	call this%expected_array(1,s)%allocate( &
3003		−	array_shape = [ &
3004		−	output(1,s)%num_vertex_features, output(1,s)%num_vertices &
3005			] &
3006		−	)
3007		−	call this%expected_array(1,s)%zero_grad()
3008		−	call this%expected_array(1,s)%set_requires_grad(.false.)
3009		−	call this%expected_array(1,s)%set( output(1,s)%vertex_features )
3010		−	this%expected_array(1,s)%is_temporary = .false.
3011		−	if(output(1,s)%num_edge_features.le.0) cycle
3012		−	call this%expected_array(2,s)%allocate( &
3013		−	array_shape = [ &
3014		−	output(1,s)%num_edge_features, output(1,s)%num_edges &
3015			] &
3016		−	)
3017		−	call this%expected_array(2,s)%set_requires_grad(.false.)
3018		−	call this%expected_array(2,s)%set( output(1,s)%edge_features )
3019		−	this%expected_array(2,s)%is_temporary = .false.
3020			end do
3021			class is(array_type)
3022		−	allocate(this%expected_array(size(output,1),size(output,2)))
3023		−	do s = 1, size(output,2)
3024		−	do i = 1, size(output,1)
3025		−	if(this%expected_array(i,s)%allocated) &
3026		−	call this%expected_array(i,s)%deallocate()
3027		−	call this%expected_array(i,s)%allocate( &
3028		−	array_shape = [ &
3029		−	output(i,s)%shape, size(output(i,s)%val,2) &
3030			] &
3031		−	)
3032		−	call this%expected_array(i,s)%set_requires_grad(.false.)
3033		−	call this%expected_array(i,s)%set( output(i,s)%val )
3034		−	this%expected_array(i,s)%is_temporary = .false.
3035			end do
3036			end do
3037			type is(real)
3038		−	allocate(this%expected_array(1,1))
3039		−	if(this%expected_array(1,1)%allocated) &
3040		−	call this%expected_array(1,1)%deallocate()
3041		−	call this%expected_array(1,1)%allocate( &
3042			array_shape = [ size(output,1), size(output,2) ] &
3043		−	)
3044		−	call this%expected_array(1,1)%set_requires_grad(.false.)
3045		−	call this%expected_array(1,1)%set( output )
3046		−	this%expected_array(1,1)%is_temporary = .false.
3047			type is(integer)
3048		−	allocate(this%expected_array(1,1))
3049		−	if(this%expected_array(1,1)%allocated) &
3050		−	call this%expected_array(1,1)%deallocate()
3051		−	call this%expected_array(1,1)%allocate( &
3052			array_shape = [ size(output,1), size(output,2) ] &
3053		−	)
3054		−	call this%expected_array(1,1)%set_requires_grad(.false.)
3055		−	this%expected_array(1,1)%val = real(output, real32)
3056		−	this%expected_array(1,1)%is_temporary = .false.
3057			class default
3058		−	call stop_program("output type not supported in training")
3059			end select
3060
3061		−	end subroutine save_output_to_network
3062			!###############################################################################
3063
3064
3065			!##############################################################################!
3066			! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * !
3067			!##############################################################################!
3068
3069
3070			!###############################################################################
3071		−	module subroutine train( &
3072		−	this, input, output, num_epochs, batch_size, &
3073			plateau_threshold, shuffle_batches, batch_print_step, verbose &
3074			)
3075			!! Train the network
3076			!!
3077			!! This function trains the network on the input data for a number of
3078			!! epochs. The input data is split into batches of size batch_size and
3079			!! the network is trained on each batch. The network is trained using
3080			!! the optimiser specified in the network object.
3081			use athena__tools_infile, only: stop_check
3082			implicit none
3083
3084			! Arguments
3085			class(network_type), intent(inout) :: this
3086			!! Instance of network
3087			class(*), dimension(..), intent(in) :: input
3088			!! Input data
3089			class(*), dimension(:,:), intent(in) :: output
3090			!! Output data
3091			integer, intent(in) :: num_epochs
3092			!! Number of epochs
3093			integer, optional, intent(in) :: batch_size
3094			!! Batch size
3095			real(real32), optional, intent(in) :: plateau_threshold
3096			!! Plateau threshold
3097			logical, optional, intent(in) :: shuffle_batches
3098			!! Shuffle batches
3099			integer, optional, intent(in) :: batch_print_step
3100			!! Batch print step
3101			integer, optional, intent(in) :: verbose
3102			!! Verbosity level
3103
3104			! Training parameters
3105			real(real32) :: batch_loss, batch_accuracy, avg_loss, avg_accuracy
3106			!! Loss and accuracy
3107
3108			! learning parameters
3109			integer :: l, num_samples
3110			!! Loop index
3111			integer :: num_batches
3112			!! Number of batches
3113			integer :: converged
3114			!! Convergence flag
3115			integer :: window_width
3116			!! Length of convergence check window
3117			integer :: verbose_
3118			!! Verbosity level
3119			integer :: batch_print_step_
3120			!! Batch print step
3121			real(real32) :: plateau_threshold_
3122			!! Plateau threshold
3123			logical :: shuffle_batches_
3124			!! Shuffle batches
3125
3126			! Training loop variables
3127			integer :: epoch, batch, start_index, end_index
3128			!! Loop index
3129		−	integer, allocatable, dimension(:) :: batch_order
3130			!! Batch order
3131
3132			integer :: i, j, s, time, time_old, clock_rate
3133			!! Loop index
3134
3135		−	class(*), allocatable :: data_poly(:,:)
3136			type(array_type), pointer :: loss => null()
3137
3138			#ifdef _OPENMP
3139			type(network_type) :: this_copy
3140			!! Copy of network
3141			#endif
3142			! integer :: timer_start = 0, timer_stop = 0, timer_sum = 0, timer_tot = 0
3143			! integer :: forward_timer = 0, backward_timer = 0, update_timer = 0
3144
3145
3146			!---------------------------------------------------------------------------
3147			! Check loss and accuracy methods are set
3148			!---------------------------------------------------------------------------
3149		−	if(.not.allocated(this%loss))then
3150		−	call stop_program("loss method not set")
3151		−	return
3152			end if
3153		−	if(.not.associated(this%get_accuracy))then
3154		−	call stop_program("accuracy method not set")
3155		−	return
3156			end if
3157
3158
3159			!---------------------------------------------------------------------------
3160			! Initialise optional arguments
3161			!---------------------------------------------------------------------------
3162		−	verbose_ = 0
3163		−	batch_print_step_ = 20
3164		−	plateau_threshold_ = 0._real32
3165		−	shuffle_batches_ = .true.
3166		−	if(present(plateau_threshold)) plateau_threshold_ = plateau_threshold
3167		−	if(present(shuffle_batches)) shuffle_batches_ = shuffle_batches
3168		−	if(present(batch_print_step)) batch_print_step_ = batch_print_step
3169		−	if(present(verbose)) verbose_ = verbose
3170		−	if(present(batch_size)) this%batch_size = batch_size
3171
3172
3173			!---------------------------------------------------------------------------
3174			! Initialise monitoring variables
3175			!---------------------------------------------------------------------------
3176		−	window_width = max(ceiling(500._real32/this%batch_size),1)
3177		−	do i = 1, size(this%metrics,dim=1)
3178		−	this%metrics(i)%window_width = window_width
3179			end do
3180
3181
3182			!---------------------------------------------------------------------------
3183			! Save input and output to network
3184			!---------------------------------------------------------------------------
3185		−	num_samples = this%save_input( input )
3186		−	call this%save_output( output )
3187		−	if(size(output,2).ne.num_samples.and.this%use_graph_output)then
3188		−	call stop_program("number of samples in input and output do not match")
3189		−	return
3190			end if
3191
3192
3193			!---------------------------------------------------------------------------
3194			! If parallel, initialise slices
3195			!---------------------------------------------------------------------------
3196		−	select type(output)
3197			type is(graph_type)
3198		−	num_batches = size(output,dim=2) / this%batch_size
3199			class is(array_type)
3200		−	if(this%use_graph_output)then
3201		−	num_batches = size(output,dim=2) / this%batch_size
3202			else
3203		−	num_batches = size(output(1,1)%val,dim=2) / this%batch_size
3204			end if
3205			class default
3206		−	num_batches = size(output,dim=2) / this%batch_size
3207			end select
3208		−	allocate(batch_order(num_batches))
3209		−	do batch = 1, num_batches
3210		−	batch_order(batch) = batch
3211			end do
3212
3213
3214			!---------------------------------------------------------------------------
3215			! Set/reset batch size for training
3216			!---------------------------------------------------------------------------
3217		−	call this%set_batch_size(this%batch_size)
3218
3219
3220			!---------------------------------------------------------------------------
3221			! Turn off inference booleans
3222			!---------------------------------------------------------------------------
3223		−	do l = 1, this%num_layers
3224		−	this%model(l)%layer%inference = .false.
3225			end do
3226
3227
3228			!---------------------------------------------------------------------------
3229			! Query system clock
3230			!---------------------------------------------------------------------------
3231		−	call system_clock(time, count_rate = clock_rate)
3232
3233
3234		−	epoch_loop: do epoch = 1, num_epochs
3235		−	this%epoch = epoch
3236			!------------------------------------------------------------------------
3237			! Shuffle batch order at the start of each epoch
3238			!------------------------------------------------------------------------
3239		−	if(shuffle_batches_)then
3240		−	call shuffle(batch_order)
3241			end if
3242
3243		−	avg_loss = 0._real32
3244		−	avg_accuracy = 0._real32
3245
3246			!------------------------------------------------------------------------
3247			! Batch loop
3248			! ... split data up into minibatches for training
3249			!------------------------------------------------------------------------
3250		−	batch_loop: do batch = 1, num_batches
3251
3252
3253			! Set batch start and end index
3254			!---------------------------------------------------------------------
3255		−	start_index = (batch_order(batch) - 1) * this%batch_size + 1
3256		−	end_index = batch_order(batch) * this%batch_size
3257
3258
3259			! Forward pass
3260			!---------------------------------------------------------------------
3261			! call system_clock(timer_start)
3262		−	select case(this%use_graph_input)
3263			case(.true.)
3264		−	data_poly = get_sample( &
3265			this%input_graph, start_index, end_index, this%batch_size &
3266		−	)
3267			case default
3268			data_poly = get_sample( &
3269			this%input_array, start_index, end_index, this%batch_size, &
3270			as_graph = .false. &
3271		−	)
3272			end select
3273		−	call this%forward(data_poly)
3274		−	deallocate(data_poly)
3275			! call system_clock(timer_stop)
3276			! forward_timer = forward_timer + timer_stop - timer_start
3277
3278
3279			! Backward pass
3280			!---------------------------------------------------------------------
3281			! call system_clock(timer_start)
3282		−	loss => this%loss_eval(start_index, end_index)
3283		−	loss%is_temporary = .false.
3284		−	call loss%grad_reverse(reset_graph=.true.)
3285			! call system_clock(timer_stop)
3286			! backward_timer = backward_timer + timer_stop - timer_start
3287
3288
3289			! Compute loss and accuracy (for monitoring)
3290			!---------------------------------------------------------------------
3291		−	batch_loss = sum(loss%val)
3292		−	batch_accuracy = this%accuracy_eval(output, start_index, end_index)
3293
3294
3295			! Average metric over batch size and store
3296			! Check metric convergence
3297			!---------------------------------------------------------------------
3298		−	avg_loss = avg_loss + batch_loss
3299		−	avg_accuracy = avg_accuracy + batch_accuracy
3300		−	call this%metrics(1)%append(batch_loss)
3301		−	call this%metrics(2)%append(batch_accuracy / this%batch_size)
3302		−	do i = 1, size(this%metrics,dim=1)
3303		−	call this%metrics(i)%check(plateau_threshold_, converged)
3304		−	if(converged.ne.0)then
3305		−	exit epoch_loop
3306			end if
3307			end do
3308
3309
3310			! Update weights and biases using optimisation algorithm
3311			!---------------------------------------------------------------------
3312			! call system_clock(timer_start)
3313		−	call this%update()
3314			! call system_clock(timer_stop)
3315			! update_timer = update_timer + timer_stop - timer_start
3316		−	call loss%nullify_graph()
3317		−	deallocate(loss)
3318		−	nullify(loss)
3319
3320
3321			! Print batch results
3322			!---------------------------------------------------------------------
3323		−	if(abs(verbose_).gt.0.and.&
3324			(batch.eq.1.or.abs(mod(batch,batch_print_step_)).lt.1.E-6))then
3325			write(6,'("epoch=",I0,", batch=",I0,&
3326			&", learning_rate=",F0.3,", loss=",F0.3,", accuracy=",F0.3)' &
3327			) &
3328		−	this%epoch, batch, &
3329			this%optimiser%lr_decay%get_lr( &
3330			this%optimiser%learning_rate, this%optimiser%iter &
3331		−	), &
3332		−	avg_loss/batch, &
3333		−	avg_accuracy/(batch*this%batch_size)
3334			end if
3335
3336
3337			! Time check
3338			!---------------------------------------------------------------------
3339		−	if(verbose_.eq.-2)then
3340		−	time_old = time
3341		−	call system_clock(time)
3342			write(*,'("time check: ",F5.3," seconds")') &
3343		−	real(time-time_old)/clock_rate
3344		−	time_old = time
3345			end if
3346
3347
3348			! Check for user-name stop file
3349			!---------------------------------------------------------------------
3350		−	if(stop_check())then
3351		−	write(0,*) "STOPCAR ENCOUNTERED"
3352		−	write(0,*) "Exiting training loop..."
3353		−	exit epoch_loop
3354			end if
3355
3356			end do batch_loop
3357
3358
3359			! Print epoch summary results
3360			!------------------------------------------------------------------------
3361		−	if(verbose_.eq.0)then
3362			write(6,'("epoch=",I0,&
3363			&", learning_rate=",F0.3,", val_loss=",F0.3,&
3364			&", val_accuracy=",F0.3)' &
3365			) &
3366		−	this%epoch, &
3367			this%optimiser%lr_decay%get_lr( &
3368			this%optimiser%learning_rate, this%optimiser%iter &
3369		−	), &
3370		−	this%metrics(1)%val, this%metrics(2)%val
3371			end if
3372
3373
3374			end do epoch_loop
3375
3376			! write(*,*) "forward timer: ", real(forward_timer)/clock_rate
3377			! write(*,*) "backward timer: ", real(backward_timer)/clock_rate
3378			! write(*,*) "update timer: ", real(update_timer)/clock_rate
3379
3380		−	end subroutine train
3381			!###############################################################################
3382
3383
3384			!###############################################################################
3385		−	module subroutine test( &
3386		−	this, input, output, verbose &
3387			)
3388			!! Test the network
3389			implicit none
3390
3391			! Arguments
3392			class(network_type), intent(inout) :: this
3393			!! Instance of network
3394			class(*), dimension(..), intent(in) :: input
3395			!! Input data
3396			class(*), dimension(:,:), intent(in) :: output
3397			!! Output data
3398			integer, optional, intent(in) :: verbose
3399			!! Verbosity level
3400
3401			! Local variables
3402			integer :: l, sample, num_samples
3403			!! Loop index
3404			integer :: verbose_
3405			!! Verbosity level
3406			real(real32) :: acc_val, loss_val
3407			!! Loss and accuracy
3408		−	class(*), allocatable, dimension(:,:) :: data_poly
3409			!! Polymorphic data array
3410			type(array_type), pointer :: loss => null()
3411			!! Loss
3412
3413
3414			!---------------------------------------------------------------------------
3415			! Initialise optional arguments
3416			!---------------------------------------------------------------------------
3417		−	if(present(verbose))then
3418		−	verbose_ = verbose
3419			else
3420		−	verbose_ = 0
3421			end if
3422
3423		−	do l = 1, size(this%metrics,dim=1)
3424		−	this%metrics(l)%val = 0._real32
3425			end do
3426		−	loss_val = 0._real32
3427		−	acc_val = 0._real32
3428
3429
3430		−	num_samples = this%save_input( input )
3431
3432
3433			!---------------------------------------------------------------------------
3434			! Reset batch size for testing
3435			!---------------------------------------------------------------------------
3436		−	call this%set_batch_size(1)
3437
3438
3439			!---------------------------------------------------------------------------
3440			! Turn on inference booleans
3441			!---------------------------------------------------------------------------
3442		−	do l = 1, this%num_layers
3443		−	this%model(l)%layer%inference = .true.
3444			end do
3445
3446
3447			!---------------------------------------------------------------------------
3448			! Testing loop
3449			!---------------------------------------------------------------------------
3450		−	test_loop1: do sample = 1, num_samples
3451
3452			! Forward pass
3453			!------------------------------------------------------------------------
3454		−	select case(this%use_graph_input)
3455			case(.true.)
3456		−	data_poly = get_sample( &
3457			this%input_graph, sample, sample, 1 &
3458		−	)
3459			case default
3460			data_poly = get_sample_array( &
3461			this%input_array, sample, sample, 1, &
3462			as_graph = .false. &
3463		−	)
3464			end select
3465		−	call this%forward(data_poly)
3466		−	deallocate(data_poly)
3467
3468
3469			! Compute loss and accuracy (for monitoring)
3470			!------------------------------------------------------------------------
3471		−	loss => this%loss_eval(sample, sample)
3472		−	loss_val = sum(loss%val)
3473		−	call loss%nullify_graph()
3474		−	deallocate(loss)
3475		−	nullify(loss)
3476		−	acc_val = this%accuracy_eval(output, sample, sample)
3477
3478		−	this%metrics(2)%val = this%metrics(2)%val + acc_val
3479		−	this%metrics(1)%val = this%metrics(1)%val + loss_val
3480
3481			end do test_loop1
3482
3483
3484			! Normalise metrics by number of samples
3485			!---------------------------------------------------------------------------
3486		−	this%accuracy_val = this%metrics(2)%val / real(num_samples, real32)
3487		−	this%loss_val = this%metrics(1)%val / real(num_samples, real32)
3488
3489		−	end subroutine test
3490			!###############################################################################
3491
3492
3493			!###############################################################################
3494		−	module function predict_real( &
3495			this, input, verbose &
3496			) result(output)
3497			!! Predict the output for a 1D input
3498			implicit none
3499
3500			! Arguments
3501			class(network_type), intent(inout) :: this
3502			!! Instance of network
3503			real(real32), dimension(..), intent(in) :: input
3504			!! Input
3505			integer, optional, intent(in) :: verbose
3506			!! Verbosity level
3507
3508			! Local variables
3509			integer :: l
3510			!! Loop index
3511			real(real32), dimension(:,:), allocatable :: output
3512			!! Output
3513			integer :: verbose_, batch_size
3514			!! Verbosity level
3515
3516
3517			!---------------------------------------------------------------------------
3518			! Initialise optional arguments
3519			!---------------------------------------------------------------------------
3520		−	if(present(verbose))then
3521		−	verbose_ = verbose
3522			else
3523		−	verbose_ = 0
3524			end if
3525
3526			select rank(input)
3527			rank(2)
3528		−	batch_size = size(input,dim=2)
3529			rank(3)
3530		−	batch_size = size(input,dim=3)
3531			rank(4)
3532		−	batch_size = size(input,dim=4)
3533			rank(5)
3534		−	batch_size = size(input,dim=5)
3535			rank(6)
3536		−	batch_size = size(input,dim=6)
3537			rank default
3538		−	batch_size = size(input,dim=rank(input))
3539			end select
3540
3541
3542			!---------------------------------------------------------------------------
3543			! Reset batch size for testing
3544			!---------------------------------------------------------------------------
3545		−	call this%set_batch_size(batch_size)
3546
3547
3548			!---------------------------------------------------------------------------
3549			! Turn on inference booleans
3550			!---------------------------------------------------------------------------
3551		−	do l = 1, this%num_layers
3552		−	this%model(l)%layer%inference = .true.
3553			end do
3554
3555
3556			!---------------------------------------------------------------------------
3557			! Predict
3558			!---------------------------------------------------------------------------
3559		−	call this%forward(get_sample(input, 1, batch_size, batch_size))
3560
3561		−	output = this%model(this%leaf_vertices(1))%layer%output(1,1)%val
3562
3563		−	end function predict_real
3564			!###############################################################################
3565
3566
3567			!###############################################################################
3568		−	module function predict_graph1d( this, input, verbose ) result(output)
3569			!! Predict the output for a graph input
3570			implicit none
3571
3572			! Arguments
3573			class(network_type), intent(inout) :: this
3574			!! Instance of network
3575			type(graph_type), dimension(:), intent(in) :: input
3576			!! Input graph
3577			integer, optional, intent(in) :: verbose
3578			!! Verbosity level
3579
3580			! Local variables
3581			integer :: l, s
3582			!! Loop index
3583		−	type(graph_type), dimension(size(this%leaf_vertices),size(input)) :: output
3584			!! Output graph
3585			integer :: verbose_ = 0, batch_size
3586			!! Verbosity level
3587
3588
3589			!---------------------------------------------------------------------------
3590			! Initialise optional arguments
3591			!---------------------------------------------------------------------------
3592		−	if(present(verbose)) verbose_ = verbose
3593
3594			!---------------------------------------------------------------------------
3595			! Reset batch size for testing
3596			!---------------------------------------------------------------------------
3597		−	batch_size = size(input)
3598		−	call this%set_batch_size(batch_size)
3599
3600
3601			!---------------------------------------------------------------------------
3602			! Turn on inference booleans
3603			!---------------------------------------------------------------------------
3604		−	do l = 1, this%num_layers
3605		−	this%model(l)%layer%inference = .true.
3606			end do
3607
3608
3609			!---------------------------------------------------------------------------
3610			! Predict
3611			!---------------------------------------------------------------------------
3612		−	call this%forward(get_sample(input, 1, batch_size, batch_size))
3613
3614		−	do l = 1, size(this%leaf_vertices)
3615		−	do s = 1, batch_size
3616		−	output(l,s)%num_vertices = input(s)%num_vertices
3617		−	output(l,s)%num_edges = input(s)%num_edges
3618		−	output(l,s)%num_vertex_features = this%model( &
3619		−	this%leaf_vertices(l) &
3620		−	)%layer%output_shape(1)
3621		−	output(l,s)%num_edge_features = this%model( &
3622		−	this%leaf_vertices(l) &
3623		−	)%layer%output_shape(2)
3624		−	output(l,s)%vertex_features = this%model( &
3625		−	this%leaf_vertices(l) &
3626		−	)%layer%output(1,s)%val
3627		−	if(size(this%model(this%leaf_vertices(l))%layer%output,1).eq.1)then
3628		−	output(l,s)%edge_features = input(s)%edge_features
3629			else
3630		−	output(l,s)%edge_features = this%model( &
3631		−	this%leaf_vertices(l) &
3632		−	)%layer%output(2,s)%val
3633			end if
3634			end do
3635			end do
3636
3637		−	end function predict_graph1d
3638			!-------------------------------------------------------------------------------
3639		−	module function predict_graph2d( this, input, verbose ) result(output)
3640			!! Predict the output for a graph input
3641			implicit none
3642
3643			! Arguments
3644			class(network_type), intent(inout) :: this
3645			!! Instance of network
3646			type(graph_type), dimension(:,:), intent(in) :: input
3647			!! Input graph
3648			integer, optional, intent(in) :: verbose
3649			!! Verbosity level
3650
3651			! Local variables
3652			integer :: l, s
3653			!! Loop index
3654		−	type(graph_type), dimension(size(this%leaf_vertices),size(input,dim=2)) :: &
3655			output
3656			!! Output graph
3657			integer :: verbose_ = 0, batch_size
3658			!! Verbosity level
3659
3660
3661			!---------------------------------------------------------------------------
3662			! Initialise optional arguments
3663			!---------------------------------------------------------------------------
3664		−	if(present(verbose)) verbose_ = verbose
3665
3666			!---------------------------------------------------------------------------
3667			! Reset batch size for testing
3668			!---------------------------------------------------------------------------
3669		−	batch_size = size(input, 2)
3670		−	call this%set_batch_size(batch_size)
3671
3672
3673			!---------------------------------------------------------------------------
3674			! Turn on inference booleans
3675			!---------------------------------------------------------------------------
3676		−	do l = 1, this%num_layers
3677		−	this%model(l)%layer%inference = .true.
3678			end do
3679
3680
3681			!---------------------------------------------------------------------------
3682			! Predict
3683			!---------------------------------------------------------------------------
3684		−	call this%forward(get_sample(input, 1, batch_size, batch_size))
3685
3686		−	do l = 1, size(this%leaf_vertices)
3687		−	do s = 1, batch_size
3688		−	output(l,s)%num_vertices = input(1,s)%num_vertices
3689		−	output(l,s)%num_edges = input(1,s)%num_edges
3690		−	output(l,s)%num_vertex_features = this%model( &
3691		−	this%leaf_vertices(l) &
3692		−	)%layer%output_shape(1)
3693		−	output(l,s)%num_edge_features = this%model( &
3694		−	this%leaf_vertices(l) &
3695		−	)%layer%output_shape(2)
3696		−	output(l,s)%vertex_features = this%model( &
3697		−	this%leaf_vertices(l) &
3698		−	)%layer%output(1,s)%val
3699		−	if(size(this%model(this%leaf_vertices(l))%layer%output,1).eq.1)then
3700		−	output(l,s)%edge_features = input(1,s)%edge_features
3701			else
3702		−	output(l,s)%edge_features = this%model( &
3703		−	this%leaf_vertices(l) &
3704		−	)%layer%output(2,s)%val
3705			end if
3706			end do
3707			end do
3708
3709		−	end function predict_graph2d
3710			!###############################################################################
3711
3712
3713			!###############################################################################
3714		−	module function predict_array_from_real( this, input, output_as_array, verbose ) &
3715			result(output)
3716			!! Predict the output for a generic input
3717			implicit none
3718
3719			! Arguments
3720			class(network_type), intent(inout) :: this
3721			!! Instance of network
3722			class(*), dimension(..), intent(in) :: input
3723			!! Input graph
3724			logical, intent(in) :: output_as_array
3725			!! Whether to output as array
3726			integer, intent(in), optional :: verbose
3727			!! Verbosity level
3728
3729			type(array_type), dimension(:,:), allocatable :: output
3730			!! Predicted output
3731
3732			! Local variables
3733			integer :: l, s, i
3734			!! Loop index
3735			integer :: num_samples
3736			!! Number of samples
3737			integer :: verbose_
3738			!! Verbosity level
3739		−	logical, dimension(:), allocatable :: inference_store
3740
3741
3742			!---------------------------------------------------------------------------
3743			! Initialise optional arguments
3744			!---------------------------------------------------------------------------
3745		−	if(present(verbose))then
3746		−	verbose_ = verbose
3747			else
3748		−	verbose_ = 0
3749			end if
3750		−	if(.not.output_as_array)then
3751		−	call stop_program("predict_array_from_real: output_as_array must be true")
3752		−	return
3753			end if
3754
3755
3756			!---------------------------------------------------------------------------
3757			! Set number of samples for predicting
3758			!---------------------------------------------------------------------------
3759		−	num_samples = this%save_input( input )
3760			! call this%set_batch_size(num_samples)
3761
3762
3763			!---------------------------------------------------------------------------
3764			! Turn on inference booleans
3765			!---------------------------------------------------------------------------
3766		−	allocate(inference_store(this%num_layers))
3767		−	do l = 1, this%num_layers
3768		−	inference_store(l) = this%model(l)%layer%inference
3769		−	this%model(l)%layer%inference = .true.
3770			end do
3771
3772			!---------------------------------------------------------------------------
3773			! Forward pass
3774			!---------------------------------------------------------------------------
3775		−	select case(this%use_graph_input)
3776			case(.true.)
3777		−	call this%forward(this%input_graph)
3778			case default
3779		−	call this%forward(this%input_array)
3780			end select
3781
3782
3783			!---------------------------------------------------------------------------
3784			! Allocate output data
3785			!---------------------------------------------------------------------------
3786		−	allocate(output( &
3787		−	size(this%model(this%leaf_vertices(1))%layer%output, 1), &
3788		−	size(this%model(this%leaf_vertices(1))%layer%output, 2) &
3789		−	))
3790		−	do s = 1, size(this%model(this%leaf_vertices(1))%layer%output, 2)
3791		−	do i = 1, size(this%model(this%leaf_vertices(1))%layer%output, 1)
3792		−	output(i,s) = this%model(this%leaf_vertices(1))%layer%output(i,s)
3793			end do
3794			end do
3795
3796			!---------------------------------------------------------------------------
3797			! Reset inference booleans
3798			!---------------------------------------------------------------------------
3799		−	do l = 1, this%num_layers
3800		−	this%model(l)%layer%inference = inference_store(l)
3801			end do
3802
3803		−	end function predict_array_from_real
3804			!###############################################################################
3805
3806
3807			!###############################################################################
3808		−	module function predict_array( this, input, verbose ) &
3809			result(output)
3810			!! Predict the output for a generic input
3811			implicit none
3812
3813			! Arguments
3814			class(network_type), intent(inout) :: this
3815			!! Instance of network
3816			class(array_type), dimension(..), intent(in) :: input
3817			!! Input graph
3818			integer, intent(in), optional :: verbose
3819			!! Verbosity level
3820
3821			type(array_type), dimension(:,:), allocatable :: output
3822			!! Predicted output
3823
3824			! Local variables
3825			integer :: l, s, i, j, layer_id
3826			!! Loop index
3827			integer :: num_samples
3828			!! Number of samples
3829			integer :: verbose_
3830			!! Verbosity level
3831			integer, dimension(2) :: output_shape
3832			!! Output shape
3833		−	logical, dimension(:), allocatable :: inference_store
3834			!! Inference store
3835
3836
3837			!---------------------------------------------------------------------------
3838			! Initialise optional arguments
3839			!---------------------------------------------------------------------------
3840		−	if(present(verbose))then
3841		−	verbose_ = verbose
3842			else
3843		−	verbose_ = 0
3844			end if
3845
3846
3847			!---------------------------------------------------------------------------
3848			! Set number of samples for predicting
3849			!---------------------------------------------------------------------------
3850		−	num_samples = this%save_input( input )
3851			! call this%set_batch_size(num_samples)
3852
3853
3854			!---------------------------------------------------------------------------
3855			! Turn on inference booleans
3856			!---------------------------------------------------------------------------
3857		−	allocate(inference_store(this%num_layers))
3858		−	do l = 1, this%num_layers
3859		−	inference_store(l) = this%model(l)%layer%inference
3860		−	this%model(l)%layer%inference = .true.
3861			end do
3862
3863			!---------------------------------------------------------------------------
3864			! Forward pass
3865			!---------------------------------------------------------------------------
3866		−	select case(this%use_graph_input)
3867			case(.true.)
3868		−	call this%forward(this%input_graph)
3869			case default
3870		−	call this%forward(this%input_array)
3871			end select
3872
3873
3874			!---------------------------------------------------------------------------
3875			! Allocate output data
3876			!---------------------------------------------------------------------------
3877		−	output_shape = this%get_output_shape()
3878		−	allocate(output(output_shape(1), output_shape(2)))
3879		−	do l = 1, size(this%leaf_vertices)
3880		−	layer_id = this%auto_graph%vertex(this%leaf_vertices(l))%id
3881		−	j = 0
3882		−	do i = 1, size(this%model(layer_id)%layer%output, 1)
3883		−	j = j + 1
3884		−	do s = 1, size(this%model(layer_id)%layer%output, 2)
3885		−	output(j,s) = this%model(layer_id)%layer%output(i,s)
3886			end do
3887			end do
3888			end do
3889
3890			!---------------------------------------------------------------------------
3891			! Reset inference booleans
3892			!---------------------------------------------------------------------------
3893		−	do l = 1, this%num_layers
3894		−	this%model(l)%layer%inference = inference_store(l)
3895			end do
3896
3897		−	end function predict_array
3898			!###############################################################################
3899
3900
3901			!###############################################################################
3902		−	module function predict_generic( this, input, verbose, output_as_graph ) &
3903		−	result(output)
3904			!! Predict the output for a generic input
3905			implicit none
3906
3907			! Arguments
3908			class(network_type), intent(inout) :: this
3909			!! Instance of network
3910			class(*), dimension(:,:), intent(in) :: input
3911			!! Input graph
3912			integer, intent(in), optional :: verbose
3913			!! Verbosity level
3914			logical, intent(in), optional :: output_as_graph
3915			!! Boolean whether to output as graph
3916
3917			class(*), dimension(:,:), allocatable :: output
3918			!! Predicted output
3919
3920			! Local variables
3921			integer :: l, s, i, j, layer_id
3922			!! Loop index
3923			integer :: num_samples
3924			!! Number of samples
3925			integer :: verbose_
3926			!! Verbosity level
3927			logical :: output_as_graph_
3928			!! Output as graph boolean
3929			integer, dimension(2) :: output_shape
3930			!! Output shape
3931
3932
3933			!---------------------------------------------------------------------------
3934			! Initialise optional arguments
3935			!---------------------------------------------------------------------------
3936		−	if(present(verbose))then
3937		−	verbose_ = verbose
3938			else
3939		−	verbose_ = 0
3940			end if
3941
3942		−	if(present(output_as_graph))then
3943		−	output_as_graph_ = output_as_graph
3944			else
3945		−	output_as_graph_ = .false.
3946			end if
3947		−	if(output_as_graph_.and..not.this%use_graph_output)then
3948			call stop_program("output_as_graph is true but network does not use &
3949		−	&graph output")
3950			end if
3951
3952
3953			!---------------------------------------------------------------------------
3954			! Set number of samples for predicting
3955			!---------------------------------------------------------------------------
3956		−	num_samples = this%save_input( input )
3957		−	call this%set_batch_size(num_samples)
3958
3959
3960			!---------------------------------------------------------------------------
3961			! Turn on inference booleans
3962			!---------------------------------------------------------------------------
3963		−	do l = 1, this%num_layers
3964		−	this%model(l)%layer%inference = .true.
3965			end do
3966
3967			!---------------------------------------------------------------------------
3968			! Forward pass
3969			!---------------------------------------------------------------------------
3970		−	select case(this%use_graph_input)
3971			case(.true.)
3972		−	call this%forward(this%input_graph)
3973			case default
3974		−	call this%forward(this%input_array)
3975			end select
3976
3977
3978			!---------------------------------------------------------------------------
3979			! Allocate output data
3980			!---------------------------------------------------------------------------
3981		−	output_shape = this%get_output_shape()
3982		−	if(output_as_graph_)then
3983		−	allocate(output(output_shape(1), output_shape(2)), source = graph_type())
3984			select type(output)
3985			type is(graph_type)
3986		−	select type(input)
3987			type is(graph_type)
3988		−	do l = 1, size(this%leaf_vertices)
3989		−	do s = 1, num_samples
3990		−	output(l,s)%num_vertices = input(1,s)%num_vertices
3991		−	output(l,s)%num_edges = input(1,s)%num_edges
3992		−	output(l,s)%num_vertex_features = this%model( &
3993		−	this%leaf_vertices(l) &
3994		−	)%layer%output_shape(1)
3995		−	output(l,s)%num_edge_features = this%model( &
3996		−	this%leaf_vertices(l) &
3997		−	)%layer%output_shape(2)
3998		−	output(l,s)%vertex_features = this%model( &
3999		−	this%leaf_vertices(l) &
4000		−	)%layer%output(1,s)%val
4001		−	if(size(this%model(this%leaf_vertices(l))%layer%output,1).eq.1)then
4002		−	output(l,s)%edge_features = input(1,s)%edge_features
4003			else
4004		−	output(l,s)%edge_features = this%model( &
4005		−	this%leaf_vertices(l) &
4006		−	)%layer%output(2,s)%val
4007			end if
4008			end do
4009			end do
4010			class default
4011		−	call stop_program("input is not of type graph_type")
4012			end select
4013			class default
4014		−	call stop_program("allocation of output as graph_type failed")
4015			end select
4016			else
4017		−	output_shape = this%get_output_shape()
4018		−	allocate(output(output_shape(1), output_shape(2)), source = array_type())
4019			select type(output)
4020			type is(array_type)
4021		−	do l = 1, size(this%leaf_vertices)
4022		−	layer_id = this%auto_graph%vertex(this%leaf_vertices(l))%id
4023		−	j = 0
4024		−	do i = 1, size(this%model(layer_id)%layer%output, 1)
4025		−	j = j + 1
4026		−	do s = 1, size(this%model(layer_id)%layer%output, 2)
4027		−	output(j,s) = this%model(layer_id)%layer%output(i,s)
4028			end do
4029			end do
4030			end do
4031			end select
4032			end if
4033
4034		−	end function predict_generic
4035			!###############################################################################
4036
4037
4038			!###############################################################################
4039		−	module subroutine print_summary(this)
4040			!! Print a summary of the network architecture
4041			implicit none
4042
4043			! Arguments
4044			class(network_type), intent(in) :: this
4045			!! Instance of network
4046
4047			! Local variables
4048			integer :: i, vertex_idx
4049			!! Loop index and vertex index
4050			integer :: total_params
4051			!! Parameter counts
4052			integer :: layer_params
4053			!! Parameters in current layer
4054			character(len=80) :: line
4055			!! Line separator
4056			character(len=40) :: layer_name
4057			!! Layer name
4058			character(len=30) :: output_shape_str
4059			!! Output shape string
4060			character(len=20) :: param_str
4061			!! Parameter count string
4062			character(len=100) :: fmt
4063			!! Format string
4064
4065		−	line = repeat('_', 80)
4066
4067			! Print header
4068		−	write(*,*)
4069		−	write(*,'(A)') line
4070		−	write(*,'(A)') 'Model Summary'
4071		−	write(*,'(A)') line
4072		−	write(*,'(A35, A25, A15)') 'Layer (type)', 'Output Shape', 'Param #'
4073		−	write(*,'(A)') repeat('=', 80)
4074
4075			! Initialise parameter count
4076		−	total_params = 0
4077
4078			! Print each layer
4079		−	do i = 1, this%num_layers
4080		−	vertex_idx = this%vertex_order(i)
4081		−	associate(layer => this%model(vertex_idx)%layer)
4082			! Get layer name
4083		−	if(allocated(layer%name))then
4084			write(layer_name, '(A," (",A,")")') &
4085		−	trim(layer%name), trim(layer%subtype)
4086			else
4087			write(layer_name, '(A,I0," (",A,")")') &
4088		−	'layer_', i, trim(layer%subtype)
4089			end if
4090
4091			! Get output shape string
4092		−	if(allocated(layer%output_shape))then
4093			! write the general format for output shape
4094		−	write(fmt,'("(""(""",A,"I0,"")"")")') &
4095		−	repeat('I0,", "', size(layer%output_shape)-1)
4096		−	write(output_shape_str, fmt) layer%output_shape
4097			else
4098		−	output_shape_str = '(Not set)'
4099			end if
4100
4101			! Get parameter count
4102		−	layer_params = layer%get_num_params()
4103		−	total_params = total_params + layer_params
4104		−	if(layer_params > 0)then
4105		−	write(param_str, '(I0)') layer_params
4106			else
4107		−	param_str = '0'
4108			end if
4109
4110			! Print layer information
4111		−	write(*,'(A35, A25, A15)') adjustl(trim(layer_name)), &
4112		−	adjustl(trim(output_shape_str)), adjustl(trim(param_str))
4113			end associate
4114			end do
4115
4116			! Print footer
4117		−	write(*,'(A)') repeat('=', 80)
4118		−	write(*,'(A,I0)') 'Number of input vertices: ', size(this%root_vertices)
4119		−	write(*,'(A,I0)') 'Number of output vertices: ', size(this%leaf_vertices)
4120		−	write(*,'(A,I0)') 'Total trainable params: ', total_params
4121		−	write(*,'(A)') line
4122		−	write(*,*)
4123
4124		−	end subroutine print_summary
4125			!###############################################################################
4126
4127		−	end submodule athena__network_submodule
4128