File size: 80,279 Bytes
9375c9a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
// Copyright (C) 2015  Davis E. King (davis@dlib.net)
// License: Boost Software License   See LICENSE.txt for the full license.
#undef DLIB_DNn_CORE_ABSTRACT_H_
#ifdef DLIB_DNn_CORE_ABSTRACT_H_

#include "../cuda/tensor_abstract.h"
#include <memory>
#include <type_traits>
#include <tuple>
#include <vector>
#include "../rand.h"


namespace dlib
{

// ----------------------------------------------------------------------------------------

    template <
        typename... T 
        >
    auto tuple_tail(
        const std::tuple<T...>& item 
    );
    /*!
        ensures
            - returns a tuple that contains everything in item except for tuple_head(item).
              The items will be in the same order as they are in item, just without
              tuple_head(item).
            - This function will correctly handle nested tuples.
    !*/

    template <typename... T>
    auto tuple_head (
        const std::tuple<T...>& item
    ); 
    /*!
        ensures
            - returns a copy of the first thing in the tuple that isn't a std::tuple.
              Essentially, this function calls std::get<0>() recursively on item until
              a non-std::tuple object is found.
    !*/

// ----------------------------------------------------------------------------------------

    template <typename T>
    double get_learning_rate_multiplier(
        const T& obj
    ); 
    /*!
        ensures
            - if (obj has a get_learning_rate_multiplier() member function) then
                - returns obj.get_learning_rate_multiplier()
            - else
                - returns 1
    !*/

    template <typename T>
    void set_learning_rate_multiplier(
        T& obj,
        double learning_rate_multiplier
    );
    /*!
        requires
            - learning_rate_multiplier >= 0
        ensures
            - if (obj has a set_learning_rate_multiplier() member function) then
                - calls obj.set_learning_rate_multiplier(learning_rate_multiplier)
            - else
                - does nothing
    !*/

// ----------------------------------------------------------------------------------------

    template <typename T>
    double get_bias_learning_rate_multiplier(
        const T& obj
    );
    /*!
        ensures
            - if (obj has a get_bias_learning_rate_multiplier() member function) then
                - returns obj.get_bias_learning_rate_multiplier()
            - else
                - returns 1
    !*/

    template <typename T>
    void set_bias_learning_rate_multiplier(
        T& obj,
        double bias_learning_rate_multiplier
    );
    /*!
        requires
            - bias_learning_rate_multiplier >= 0
        ensures
            - if (obj has a set_bias_learning_rate_multiplier() member function) then
                - calls obj.set_bias_learning_rate_multiplier(bias_learning_rate_multiplier)
            - else
                - does nothing
    !*/

// ----------------------------------------------------------------------------------------

    template <typename T>
    double get_weight_decay_multiplier(
        const T& obj
    );
    /*!
        ensures
            - if (obj has a get_weight_decay_multiplier() member function) then
                - returns obj.get_weight_decay_multiplier()
            - else
                - returns 1
    !*/

    template <typename T>
    void set_weight_decay_multiplier(
        T& obj,
        double weight_decay_multiplier
    );
    /*!
        requires
            - weight_decay_multiplier >= 0
        ensures
            - if (obj has a set_weight_decay_multiplier() member function) then
                - calls obj.set_weight_decay_multiplier(weight_decay_multiplier)
            - else
                - does nothing
    !*/

// ----------------------------------------------------------------------------------------

    template <typename T>
    double get_bias_weight_decay_multiplier(
        const T& obj
    ); 
    /*!
        ensures
            - if (obj has a get_bias_weight_decay_multiplier() member function) then
                - returns obj.get_bias_weight_decay_multiplier()
            - else
                - returns 1
    !*/

    template <typename T>
    void set_bias_weight_decay_multiplier(
        T& obj,
        double bias_weight_decay_multiplier
    );
    /*!
        requires:
            - bias_weight_decay_multiplier >= 0
        ensures
            - if (obj has a set_bias_weight_decay_multiplier() member function) then
                - calls obj.set_bias_weight_decay_multiplier(bias_weight_decay_multiplier)
            - else
                - does nothing
    !*/

// ----------------------------------------------------------------------------------------

    template <typename T>
    void disable_bias(
        T& obj
    );
    /*!
        ensures
            - if (obj has a disable_bias() member function) then
                - calls obj.disable_bias()
            - else
                - does nothing
    !*/

// ----------------------------------------------------------------------------------------

    bool dnn_prefer_fastest_algorithms(
    );
    /*!
        ensures
            - If dlib should prefer to use fast algorithms rather than ones that use less
              RAM then this function returns true and false otherwise.
            - On program startup this function will default to true.
    !*/

    void set_dnn_prefer_fastest_algorithms(
    );
    /*!
        ensures
            - #dnn_prefer_fastest_algorithms() == true
    !*/

    void set_dnn_prefer_smallest_algorithms(
    );
    /*!
        ensures
            - #dnn_prefer_fastest_algorithms() == false 
    !*/

// ----------------------------------------------------------------------------------------

    template <
        typename T
        >
    class sstack
    {
        /*!
            WHAT THIS OBJECT REPRESENTS
                This is a basic stack of T objects.  It contains no data itself but simply
                points to a memory range of T object and allows you to access that block of
                T objects as a stack.
        !*/

    public:
        typedef T value_type;

        sstack() = delete;

        sstack (
            T* data,
            size_t s
        );
        /*!
            ensures
                - #size() == s
                - #top() == *data
                - #pop(i).top() == data[i]
        !*/

        const T& top(
        ) const;
        /*!
            requires
                - size() != 0
            ensures
                - returns the top element of the stack.
        !*/

        T& top(
        );
        /*!
            requires
                - size() != 0
            ensures
                - returns the top element of the stack.  
        !*/

        size_t size(
        ) const;
        /*!
            ensures
                - returns the number of elements in this stack.  
        !*/

        sstack pop(
            size_t num = 1
        ); 
        /*!
            requires
                - num <= size()
            ensures
                - returns a reference to the sub-stack S such that:
                    - S.size() == size()-num.
                    - S.top() is num elements down the stack. 
        !*/
    };

    template <
        typename T
        >
    sstack<T> make_sstack(
        std::vector<T>& item
    ) { return sstack<T>(item.data(), item.size()); }
    /*!
        ensures
            - returns a sstack that sits on top of the given std::vector.
    !*/

// ----------------------------------------------------------------------------------------

    template <
        typename LAYER_DETAILS, 
        typename SUBNET
        >
    class add_layer
    {
        /*!
            REQUIREMENTS ON LAYER_DETAILS
                - Must be a type that implements the EXAMPLE_COMPUTATIONAL_LAYER_ interface
                  defined in layers_abstract.h

            REQUIREMENTS ON SUBNET
                - One of the following must be true:
                    - SUBNET implements the EXAMPLE_INPUT_LAYER interface defined in
                      input_abstract.h.
                    - SUBNET is an add_layer object.
                    - SUBNET is an add_tag_layer object.
                    - SUBNET is an add_skip_layer object.
                    - SUBNET is a repeat object.

            WHAT THIS OBJECT REPRESENTS
                This object represents a deep neural network.  In particular, it is a tool
                for adding another layer on top of the neural network of type SUBNET, which
                is specified as a template argument.  The specific layer added is defined
                by the LAYER_DETAILS details template argument.
        !*/

    public:
        typedef LAYER_DETAILS layer_details_type;
        typedef SUBNET subnet_type;
        typedef typename subnet_type::input_type input_type;
        // num_computational_layers will always give the number of layers in the network
        // that transform tensors (i.e. layers defined by something that implements the
        // EXAMPLE_COMPUTATIONAL_LAYER_ interface).  This is all the layers except for
        // loss, tag, and skip layers.
        const static size_t num_computational_layers = subnet_type::num_computational_layers + 1;
        // num_layers counts all the layers in the network regardless of their type.  
        const static size_t num_layers = subnet_type::num_layers + 1;

        add_layer(
        );
        /*!
            ensures
                - default constructs all the layers in this network.
                - #sample_expansion_factor() == 0
        !*/

        add_layer(const add_layer&) = default;
        add_layer(add_layer&&) = default;
        add_layer& operator=(add_layer&&) = default;
        add_layer& operator=(const add_layer&) = default;
        /*!
            ensures
                - this object is copyable and movable.
        !*/

        template <typename T, typename U>
        add_layer(
            const add_layer<T,U>& item
        );
        /*!
            ensures
                - This constructor allows you to copy neural network objects from one to
                  another as long as their corresponding layers can be constructed from
                  each other.
                - #layer_details() == layer_details_type(item.layer_details())
                - #subnet()        == subnet_type(item.subnet())
                - #sample_expansion_factor() == item.sample_expansion_factor()
        !*/

        template <typename ...T, typename LD, typename ...U>
        add_layer(
            const std::tuple<LD,U...>& layer_det, 
            T&& ...args
        );
        /*!
            ensures
                - #layer_details() == layer_details_type(tuple_head(layer_det))
                - #subnet()        == subnet_type(tuple_tail(layer_det),args)
                - #sample_expansion_factor() == 0 
        !*/

        template <typename ...T>
        add_layer(
            const layer_details_type& layer_det, 
            T&& ...args
        );
        /*!
            ensures
                - #layer_details() == layer_details_type(layer_det)
                - #subnet()        == subnet_type(args)
                - #sample_expansion_factor() == 0 
        !*/

        template <typename ...T>
        add_layer(
            T&& ...args
        );
        /*!
            ensures
                - This version of the constructor is only called if layer_details_type
                  can't be constructed from the first thing in args.  In this case, the
                  args are simply passed on to the sub layers in their entirety.
                - #layer_details() == layer_details_type()
                - #subnet()        == subnet_type(args)
                - #sample_expansion_factor() == 0 
        !*/

        template <typename ...T>
        add_layer(
            layer_details_type&& layer_det, 
            T&& ...args
        );
        /*!
            ensures
                - #layer_details() == layer_det
                - #subnet()        == subnet_type(args)
                - #sample_expansion_factor() == 0 
        !*/

        template <typename forward_iterator>
        void to_tensor (
            forward_iterator ibegin,
            forward_iterator iend,
            resizable_tensor& data
        ) const;
        /*!
            requires
                - [ibegin, iend) is an iterator range over input_type objects.
                - std::distance(ibegin,iend) > 0
            ensures
                - Converts the iterator range into a tensor and stores it into #data.
                - #data.num_samples()%distance(ibegin,iend) == 0. 
                - #sample_expansion_factor() == #data.num_samples()/distance(ibegin,iend).
                - #sample_expansion_factor() > 0
                - The data in the ith sample of #data corresponds to the input_type object
                  *(ibegin+i/#sample_expansion_factor()).
                - Invokes data.async_copy_to_device() so that the data begins transferring
                  to the GPU device, if present.
                - This function is implemented by calling the to_tensor() routine defined
                  at the input layer of this network.  
        !*/

        unsigned int sample_expansion_factor (
        ) const;
        /*!
            ensures
                - When to_tensor() is invoked on this network's input layer it converts N
                  input objects into M samples, all stored inside a resizable_tensor.  It
                  is always the case that M is some integer multiple of N.
                  sample_expansion_factor() returns the value of this multiplier.  To be
                  very specific, it is always true that M==I*N where I is some integer.
                  This integer I is what is returned by sample_expansion_factor().
        !*/

        const subnet_type& subnet(
        ) const; 
        /*!
            ensures
                - returns the immediate subnetwork of *this network.  
        !*/

        subnet_type& subnet(
        );
        /*!
            ensures
                - returns the immediate subnetwork of *this network.  
        !*/

        const layer_details_type& layer_details(
        ) const; 
        /*!
            ensures
                - returns the layer_details_type instance that defines the behavior of the
                  layer at the top of this network.  I.e. returns the layer details that
                  defines the behavior of the layer nearest to the network output rather
                  than the input layer.
        !*/

        layer_details_type& layer_details(
        );
        /*!
            ensures
                - returns the layer_details_type instance that defines the behavior of the
                  layer at the top of this network.  I.e. returns the layer details that
                  defines the behavior of the layer nearest to the network output rather
                  than the input layer.
        !*/

        template <typename forward_iterator>
        const tensor& operator() (
            forward_iterator ibegin,
            forward_iterator iend
        );
        /*!
            requires
                - [ibegin, iend) is an iterator range over input_type objects.
                - std::distance(ibegin,iend) > 0
            ensures
                - runs [ibegin,iend) through the network and returns the results.
                  In particular, this function performs:
                    to_tensor(ibegin,iend,temp_tensor);
                    return forward(temp_tensor);
                - The return value from this function is also available in #get_output().
                  i.e. this function returns #get_output().
                - have_same_dimensions(#get_gradient_input(), #get_output()) == true.
                - All elements of #get_gradient_input() are set to 0. 
                  i.e. calling this function clears out #get_gradient_input() and ensures
                  it has the same dimensions as the most recent output.
        !*/

        const tensor& operator() (
            const input_type& x
        );
        /*!
            ensures
                - runs a single x through the network and returns the output.
                  I.e. returns (*this)(&x, &x+1);
        !*/

        const tensor& forward(
            const tensor& x
        );
        /*!
            requires
                - sample_expansion_factor() != 0
                  (i.e. to_tensor() must have been called to set sample_expansion_factor()
                  to something non-zero.)
                - x.num_samples()%sample_expansion_factor() == 0
                - x.num_samples() > 0
            ensures
                - Runs x through the network and returns the results.  In particular, this
                  function performs the equivalent of:
                    subnet().forward(x);
                    if (this is the first time forward() has been called) then
                        layer_details().setup(subnet());
                    layer_details().forward(subnet(), get_output());
                - The return value from this function is also available in #get_output().
                  i.e. this function returns #get_output().
                - have_same_dimensions(#get_gradient_input(), #get_output()) == true
                - All elements of #get_gradient_input() are set to 0. 
                  i.e. calling this function clears out #get_gradient_input() and ensures
                  it has the same dimensions as the most recent output.
        !*/

        const tensor& get_output(
        ) const;
        /*!
            ensures
                - returns the output for the last tensor that was run through the network.
                  If nothing has been run through the network yet then returns an empty
                  tensor. 
        !*/

        tensor& get_gradient_input(
        );
        /*!
            ensures
                - returns the error gradient for this network.  That is, this is the error
                  gradient that this network will use to compute parameter gradients when
                  back_propagate_error() is called.  Therefore, when performing back
                  propagation, layers that sit on top of this network layer write their
                  back-propagated error gradients into get_gradient_input().  Or to put it
                  another way, during back-propagation, layers take the contents of their
                  get_gradient_input() and back-propagate it through themselves and store
                  the result into their subnetwork's get_gradient_input().

                  This means you should consider get_gradient_input() as an input to the
                  back_propagate_error() method.  
        !*/

        const tensor& get_final_data_gradient(
        ) const;
        /*!
            ensures
                - if back_propagate_error() has been called to back-propagate a gradient
                  through this network then you can call get_final_data_gradient() to
                  obtain the last data gradient computed.  That is, this function returns
                  the gradient of the network with respect to its inputs.
                - Note that there is only one "final data gradient" for an entire network,
                  not one per layer, since there is only one input to the entire network.
        !*/

        const tensor& get_parameter_gradient(
        ) const; 
        /*!
            ensures
                - if back_propagate_error() has been called then you can call
                  get_parameter_gradient() to find the gradient of this layer's parameters.
                  When we update the parameters by calling update_parameters(), it will use
                  the gradient in get_parameter_gradient() to perform the update.
                  Therefore, you should consider get_parameter_gradient() as an input to
                  update_parameters().
        !*/

        tensor& get_parameter_gradient (
        ); 
        /*!
            ensures
                - returns a non-const reference to the tensor returned by the above
                  get_parameter_gradient() method.  You could use this method to modify the
                  parameter gradient in some way before invoking update_parameters().
        !*/

        void back_propagate_error(
            const tensor& x
        );
        /*!
            requires
                - forward(x) was called to forward propagate x though the network.
                  Moreover, this was the most recent call to forward() and x has not been
                  subsequently modified in any way.
                - get_gradient_input() has been set equal to the gradient of this network's
                  output with respect to some loss function.
            ensures
                - Back propagates the error gradient, get_gradient_input(), through this
                  network and computes parameter and data gradients, via backpropagation.
                  Specifically, this function populates get_final_data_gradient() and also,
                  for each layer, the tensor returned by get_parameter_gradient().
                - All elements of #get_gradient_input() are set to 0. 
                - have_same_dimensions(#get_final_data_gradient(), x) == true.
                - have_same_dimensions(#get_parameter_gradient(), layer_details().get_layer_params()) == true.
                - #get_final_data_gradient() contains the gradient of the network with
                  respect to x.
        !*/

        void back_propagate_error(
            const tensor& x, 
            const tensor& gradient_input
        );
        /*!
            requires
                - forward(x) was called to forward propagate x though the network.
                  Moreover, this was the most recent call to forward() and x has not been
                  subsequently modified in any way.
                - have_same_dimensions(gradient_input, get_output()) == true
            ensures
                - This function is identical to the version of back_propagate_error()
                  defined immediately above except that it back-propagates gradient_input
                  through the network instead of get_gradient_input().  Therefore, this
                  version of back_propagate_error() is equivalent to performing:
                    get_gradient_input() = gradient_input;
                    back_propagate_error(x);
                  Except that calling back_propagate_error(x,gradient_input) avoids the
                  copy and is therefore slightly more efficient.
                - All elements of #get_gradient_input() are set to 0. 
                - have_same_dimensions(#get_final_data_gradient(), x) == true.
                - have_same_dimensions(#get_parameter_gradient(), layer_details().get_layer_params()) == true.
                - #get_final_data_gradient() contains the gradient of the network with
                  respect to x.
        !*/

        template <typename solver_type>
        void update_parameters(
            sstack<solver_type> solvers, 
            double learning_rate
        );
        /*!
            requires
                - solver_type is an implementation of the EXAMPLE_SOLVER interface defined
                  in solvers_abstract.h
                - back_propagate_error() has been called.
                - The given solvers have only ever been used with this network.  That is,
                  if you want to call update_parameters() on some other neural network
                  object then you must NOT reuse the same solvers object.
                - solvers.size() >= num_computational_layers
                - 0 < learning_rate <= 1
            ensures
                - Updates all the parameters in the network.  In particular, we pass each
                  layer's parameter gradient (i.e. the tensor returned by the layer's
                  get_parameter_gradient() member) through that layer's corresponding
                  solver object.  This produces a parameter delta vector which we add to
                  the layer's parameters.
                - The solvers use the given learning rate.
        !*/

        template <typename solver_type>
        void update_parameters(std::vector<solver_type>& solvers, double learning_rate)
        { update_parameters(make_sstack(solvers), learning_rate); }
        /*!
            Convenience method for calling update_parameters()
        !*/

        void clean(
        );
        /*!
            ensures
                - Causes the network to forget about everything but its parameters.  
                  That is, for each layer we will have:
                    - get_output().num_samples() == 0
                    - get_gradient_input().num_samples() == 0
                  However, running new input data though this network will still produce
                  the same output it would have produced regardless of any calls to
                  clean().  The purpose of clean() is to compact the network object prior
                  to saving it to disk so that it takes up less space and the IO is
                  quicker.
                - This also calls the .clean() method on any layer details objects that 
                  define a .clean() method.
        !*/

    };

    template <typename T, typename U> 
    std::ostream& operator<<(std::ostream& out, const add_layer<T,U>& item);
    /*!
        prints the network architecture to the given output stream.
    !*/

    template <typename T, typename U> 
    void serialize(const add_layer<T,U>& item, std::ostream& out);
    template <typename T, typename U> 
    void deserialize(add_layer<T,U>& item, std::istream& in);
    /*!
        provides serialization support  
    !*/

// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------

    class no_label_type;

    template <
        typename LOSS_DETAILS, 
        typename SUBNET
        >
    class add_loss_layer
    {
        /*!
            REQUIREMENTS ON LOSS_DETAILS 
                - Must be a type that implements the EXAMPLE_LOSS_LAYER_ interface defined
                  in loss_abstract.h

            REQUIREMENTS ON SUBNET
                - One of the following must be true:
                    - SUBNET is an add_layer object.
                    - SUBNET is an add_tag_layer object.
                    - SUBNET is an add_skip_layer object.
                    - SUBNET is a repeat object.

            WHAT THIS OBJECT REPRESENTS
                This object represents a deep neural network.  In particular, it is a tool
                for adding a loss layer on top of the neural network of type SUBNET, which
                is specified as a template argument.  The specific layer added is defined
                by the LOSS_DETAILS details template argument.  Importantly, a loss layer
                is the last layer in a deep neural network.  So once it is added you can't
                add any other layers of any type.
        !*/

    public:
        typedef LOSS_DETAILS loss_details_type;
        typedef SUBNET subnet_type;
        typedef typename subnet_type::input_type input_type;
        const static size_t num_computational_layers = subnet_type::num_computational_layers;
        const static size_t num_layers = subnet_type::num_layers + 1;
        // If LOSS_DETAILS is an unsupervised loss then training_label_type==no_label_type.
        // Otherwise it is defined as follows:
        typedef typename LOSS_DETAILS::training_label_type training_label_type;
        // Similarly, if LOSS_DETAILS doesn't provide any output conversion then
        // output_label_type==no_label_type.
        typedef typename LOSS_DETAILS::output_label_type output_label_type;



        add_loss_layer() = default;
        /*!
            ensures
                - default constructs all the layers in this network.
        !*/

        add_loss_layer(const add_loss_layer&) = default;
        add_loss_layer(add_loss_layer&&) = default;
        add_loss_layer& operator=(add_loss_layer&&) = default;
        add_loss_layer& operator=(const add_loss_layer&) = default;
        /*!
            ensures
                - this object is copyable and movable.
        !*/

        template <typename T, typename U>
        add_loss_layer(
            const add_loss_layer<T,U>& item
        );
        /*!
            ensures
                - This constructor allows you to copy neural network objects from one to
                  another as long as their corresponding layers can be constructed from
                  each other.
                - #loss_details() == loss_details_type(item.loss_details())
                - #subnet()       == subnet_type(item.subnet())
        !*/

        template <typename ...T>
        add_loss_layer(
            const LOSS_DETAILS& layer_det, 
            T&& ...args
        ); 
        /*!
            ensures
                - #loss_details() == loss_details_type(layer_det)
                - #subnet()       == subnet_type(args)
        !*/

        template <typename ...T>
        add_loss_layer(
            LOSS_DETAILS&& layer_det, 
            T&& ...args
        );
        /*!
            ensures
                - #loss_details() == loss_details_type(layer_det)
                - #subnet()       == subnet_type(args)
        !*/

        template <typename ...T>
        add_loss_layer(
            T&& ...args
        ); 
        /*!
            ensures
                - This version of the constructor is only called if loss_details_type can't
                  be constructed from the first thing in args.  In this case, the args are
                  simply passed on to the sub layers in their entirety.
                - #loss_details() == loss_details_type()
                - #subnet()       == subnet_type(args)
        !*/

        const subnet_type& subnet(
        ) const; 
        /*!
            ensures
                - returns the immediate subnetwork of *this network.  
        !*/

        subnet_type& subnet(
        ); 
        /*!
            ensures
                - returns the immediate subnetwork of *this network.  
        !*/

        const loss_details_type& loss_details(
        ) const; 
        /*!
            ensures
                - returns the loss_details_type instance that defines the behavior of the
                  loss layer used by this network.
        !*/

        loss_details_type& loss_details(
        ); 
        /*!
            ensures
                - returns the loss_details_type instance that defines the behavior of the
                  loss layer used by this network.
        !*/

        template <typename forward_iterator>
        void to_tensor (
            forward_iterator ibegin,
            forward_iterator iend,
            resizable_tensor& data
        ) const;
        /*!
            requires
                - [ibegin, iend) is an iterator range over input_type objects.
                - std::distance(ibegin,iend) > 0
            ensures
                - Converts the iterator range into a tensor and stores it into #data.
                - #data.num_samples()%distance(ibegin,iend) == 0. 
                - #sample_expansion_factor() == #data.num_samples()/distance(ibegin,iend).
                - #sample_expansion_factor() > 0
                - The data in the ith sample of #data corresponds to the input_type object
                  *(ibegin+i/sample_expansion_factor()).
                - Invokes data.async_copy_to_device() so that the data begins transferring
                  to the GPU device, if present.
                - This function is implemented by calling the to_tensor() routine defined
                  at the input layer of this network.  
        !*/

        unsigned int sample_expansion_factor (
        ) const;
        /*!
            ensures
                - When to_tensor() is invoked on this network's input layer it converts N
                  input objects into M samples, all stored inside a resizable_tensor.  It
                  is always the case that M is some integer multiple of N.
                  sample_expansion_factor() returns the value of this multiplier.  To be
                  very specific, it is always true that M==I*N where I is some integer.
                  This integer I is what is returned by sample_expansion_factor().
        !*/

    // -------------

        const tensor& forward(const tensor& x
        ); 
        /*!
            requires
                - sample_expansion_factor() != 0
                  (i.e. to_tensor() must have been called to set sample_expansion_factor()
                  to something non-zero.)
                - x.num_samples()%sample_expansion_factor() == 0
                - x.num_samples() > 0
            ensures
                - Runs x through the network and returns the results as a tensor.  In particular,
                  this function just performs:
                    return subnet().forward(x);
                  So if you want to get the outputs as an output_label_type then call one of the
                  methods below instead, like operator().
                - The return value from this function is also available in #subnet().get_output().
                  i.e. this function returns #subnet().get_output().
                - have_same_dimensions(#subnet().get_gradient_input(), #subnet().get_output()) == true
                - All elements of #subnet().get_gradient_input() are set to 0. 
                  i.e. calling this function clears out #subnet().get_gradient_input() and ensures
                  it has the same dimensions as the most recent output.
        !*/

        template <typename output_iterator>
        void operator() (
            const tensor& x, 
            output_iterator obegin
        );
        /*!
            requires
                - sample_expansion_factor() != 0
                  (i.e. to_tensor() must have been called to set sample_expansion_factor()
                  to something non-zero.)
                - x.num_samples()%sample_expansion_factor() == 0
                - x.num_samples() > 0
                - obegin == iterator pointing to the start of a range of
                  x.num_samples()/sample_expansion_factor() output_label_type elements.
            ensures
                - runs x through the network and writes the output to the range at obegin.
                - loss_details().to_label() is used to write the network output into
                  obegin.
        !*/

        template <typename forward_iterator, typename label_iterator>
        void operator() (
            forward_iterator ibegin,
            forward_iterator iend,
            label_iterator obegin
        );
        /*!
            requires
                - [ibegin, iend) is an iterator range over input_type objects.
                - std::distance(ibegin,iend) > 0
                - obegin == iterator pointing to the start of a range of
                  std::distance(ibegin,iend) output_label_type elements.
            ensures
                - runs [ibegin,iend) through the network and writes the output to the range
                  at obegin.
                - loss_details().to_label() is used to write the network output into
                  obegin.
        !*/

    // -------------

        const output_label_type& operator() (
            const input_type& x
        );
        /*!
            ensures
                - runs a single object, x, through the network and returns the output.
                - loss_details().to_label() is used to convert the network output into a
                  output_label_type.
        !*/

        template <typename iterable_type>
        std::vector<output_label_type> operator() (
            const iterable_type& data,
            size_t batch_size = 128
        );
        /*!
            requires
                - batch_size > 0
                - data must have a .begin() and .end() that supply iterators over a
                  sequence of input_type elements.  E.g. data could have a type of
                  std::vector<input_type>
            ensures
                - runs all the objects in data through the network and returns their
                  predicted labels.  This means this function returns a vector V such that:
                    - V.size() == data.size()
                    - for all valid i: V[i] == the predicted label of data[i].
                - Elements of data are run through the network in batches of batch_size
                  items.  Using a batch_size > 1 can be faster because it better exploits
                  the available hardware parallelism.
                - loss_details().to_label() is used to convert the network output into a
                  output_label_type.
        !*/

        template <typename ...T>
        const output_label_type& process (
            const input_type& x, 
            T&& ...args
        );
        /*!
            ensures
                - This function is just like (*this)(x), i.e. it runs a single object, x,
                  through the network and returns the output.  But we additionally pass the 
                  given args to loss_details().to_label() as the 4th argument (or more,
                  depending on how many things are in args) when converting the network
                  output to an output_label_type.  This is useful, for instance, with loss
                  layers like loss_mmod_ which has an optional adjust_threshold argument to
                  to_label() that adjusts the detection threshold.  Therefore, for such
                  networks you could call them like: net.process(some_image, -0.5), and -0.5
                  would be passed so the adjust_threshold argument of to_tensor().
        !*/

        template <typename iterable_type, typename ...T>
        std::vector<output_label_type> process_batch (
            const iterable_type& data, 
            size_t batch_size, 
            T&& ...args
        );
        /*!
            requires
                - batch_size > 0
                - data must have a .begin() and .end() that supply iterators over a
                  sequence of input_type elements.  E.g. data could have a type of
                  std::vector<input_type>
            ensures
                - This function is just like (*this)(data,batch_size), i.e. it runs a
                  bunch of objects through the network and returns the outputs.  But we
                  additionally pass the given args to loss_details().to_label() as the 4th
                  argument (or more, depending on how many things are in args) when
                  converting the network output to output_label_types.  This is useful,
                  for instance, with loss layers like loss_mmod_ which has an optional
                  adjust_threshold argument to to_label() that adjusts the detection
                  threshold.  Therefore, for such networks you could call them like:
                  net.process_batch(std::vector<image_type>({some_image, another_image}), 128, -0.5), 
                  and -0.5 would be passed so the adjust_threshold argument of to_tensor().
        !*/

    // -------------

        template <typename label_iterator>
        double compute_loss (
            const tensor& x,
            label_iterator lbegin 
        );
        /*!
            requires
                - sample_expansion_factor() != 0
                  (i.e. to_tensor() must have been called to set sample_expansion_factor()
                  to something non-zero.)
                - x.num_samples()%sample_expansion_factor() == 0
                - x.num_samples() > 0
                - lbegin == iterator pointing to the start of a range of
                  x.num_samples()/sample_expansion_factor() training_label_type elements.
            ensures
                - runs x through the network, compares the output to the expected output
                  pointed to by lbegin, and returns the resulting loss. 
                - for all valid k:
                    - the expected label of the kth sample in x is *(lbegin+k/sample_expansion_factor()).
                - This function does not update the network parameters.
                - For sub-layers that are immediate inputs into the loss layer, we also populate the
                  sub-layer's get_gradient_input() tensor with the gradient of the loss with respect
                  to the sub-layer's output.
        !*/

        template <typename forward_iterator, typename label_iterator>
        double compute_loss (
            forward_iterator ibegin,
            forward_iterator iend,
            label_iterator lbegin 
        );
        /*!
            requires
                - [ibegin, iend) is an iterator range over input_type objects.
                - std::distance(ibegin,iend) > 0
                - lbegin == iterator pointing to the start of a range of
                  std::distance(ibegin,iend) training_label_type elements.
            ensures
                - runs [ibegin,iend) through the network, compares the output to the
                  expected output pointed to by lbegin, and returns the resulting loss. 
                - for all valid k:
                    - the expected label of *(ibegin+k) is *(lbegin+k).
                - This function does not update the network parameters.
                - For sub-layers that are immediate inputs into the loss layer, we also populate the
                  sub-layer's get_gradient_input() tensor with the gradient of the loss with respect
                  to the sub-layer's output.
        !*/

    // -------------

        double compute_loss (
            const tensor& x
        );
        /*!
            requires
                - LOSS_DETAILS is an unsupervised loss.  i.e. training_label_type==no_label_type.
                - sample_expansion_factor() != 0
                  (i.e. to_tensor() must have been called to set sample_expansion_factor()
                  to something non-zero.)
                - x.num_samples()%sample_expansion_factor() == 0
                - x.num_samples() > 0
            ensures
                - runs x through the network and returns the resulting loss. 
                - This function does not update the network parameters.
                - For sub-layers that are immediate inputs into the loss layer, we also populate the
                  sub-layer's get_gradient_input() tensor with the gradient of the loss with respect
                  to the sub-layer's output.
        !*/

        template <typename forward_iterator>
        double compute_loss (
            forward_iterator ibegin,
            forward_iterator iend,
        );
        /*!
            requires
                - LOSS_DETAILS is an unsupervised loss.  i.e. training_label_type==no_label_type.
                - [ibegin, iend) is an iterator range over input_type objects.
                - std::distance(ibegin,iend) > 0
            ensures
                - runs [ibegin,iend) through the network and returns the resulting loss. 
                - This function does not update the network parameters.
                - For sub-layers that are immediate inputs into the loss layer, we also populate the
                  sub-layer's get_gradient_input() tensor with the gradient of the loss with respect
                  to the sub-layer's output.
        !*/

    // -------------

        template <typename label_iterator>
        double compute_parameter_gradients (
            const tensor& x,
            label_iterator lbegin
        );
        /*!
            requires
                - sample_expansion_factor() != 0
                  (i.e. to_tensor() must have been called to set sample_expansion_factor()
                  to something non-zero.)
                - x.num_samples()%sample_expansion_factor() == 0
                - x.num_samples() > 0
                - lbegin == iterator pointing to the start of a range of
                  x.num_samples()/sample_expansion_factor() training_label_type elements.
            ensures
                - runs x through the network, compares the output to the expected output
                  pointed to by lbegin, and computes parameter and data gradients with
                  respect to the loss, via backpropagation.  Specifically, this function
                  updates get_final_data_gradient() and also, for each layer, the tensor
                  returned by get_parameter_gradient().
                - for all valid k:
                    - the expected label of the kth sample in x is *(lbegin+k/sample_expansion_factor()).
                - returns compute_loss(x,lbegin)
        !*/

        template <typename forward_iterator, typename label_iterator>
        double compute_parameter_gradients (
            forward_iterator ibegin,
            forward_iterator iend,
            label_iterator lbegin
        );
        /*!
            requires
                - [ibegin, iend) is an iterator range over input_type objects.
                - std::distance(ibegin,iend) > 0
                - lbegin == iterator pointing to the start of a range of
                  std::distance(ibegin,iend) training_label_type elements.
            ensures
                - runs [ibegin,iend) through the network, compares the output to the
                  expected output pointed to by lbegin, and computes parameter and data
                  gradients with respect to the loss, via backpropagation.  Specifically,
                  this function updates get_final_data_gradient() and also, for each layer,
                  the tensor returned by get_parameter_gradient().
                - for all valid k:
                    - the expected label of *(ibegin+k) is *(lbegin+k).
                - returns compute_loss(ibegin,iend,lbegin)
        !*/

        double compute_parameter_gradients (
            const tensor& x
        );
        /*!
            requires
                - LOSS_DETAILS is an unsupervised loss.  i.e. training_label_type==no_label_type.
                - sample_expansion_factor() != 0
                  (i.e. to_tensor() must have been called to set sample_expansion_factor()
                  to something non-zero.)
                - x.num_samples()%sample_expansion_factor() == 0
                - x.num_samples() > 0
            ensures
                - runs x through the network and computes parameter and data gradients with
                  respect to the loss, via backpropagation.  Specifically, this function
                  updates get_final_data_gradient() and also, for each layer, the tensor
                  returned by get_parameter_gradient().
                - returns compute_loss(x)
        !*/

        template <typename forward_iterator>
        double compute_parameter_gradients (
            forward_iterator ibegin,
            forward_iterator iend
        );
        /*!
            requires
                - LOSS_DETAILS is an unsupervised loss.  i.e. training_label_type==no_label_type.
                - [ibegin, iend) is an iterator range over input_type objects.
                - std::distance(ibegin,iend) > 0
            ensures
                - runs [ibegin,iend) through the network and computes parameter and data
                  gradients with respect to the loss, via backpropagation.  Specifically,
                  this function updates get_final_data_gradient() and also, for each layer,
                  the tensor returned by get_parameter_gradient().
                - returns compute_loss(ibegin,iend)
        !*/

        template <typename solver_type>
        void update_parameters (
            sstack<solver_type> solvers,
            double learning_rate
        );
        /*!
            requires
                - solver_type is an implementation of the EXAMPLE_SOLVER interface defined
                  in solvers_abstract.h
                - compute_parameter_gradients() has been called.
                - The given solvers have only ever been used with this network.  That
                  is, if you want to call update_parameters() on some other neural network
                  object then you must NOT reuse the same solvers object.
                - solvers.size() >= num_computational_layers
                - 0 < learning_rate <= 1
            ensures
                - Updates all the parameters in the network.  In particular, we pass each
                  layer's parameter gradient (i.e. the tensor returned by the layer's
                  get_parameter_gradient() member) through that layer's corresponding
                  solver object.  This produces a parameter delta vector which we add to
                  the layer's parameters.
                - The solvers use the given learning rate.
        !*/

        template <typename solver_type>
        void update_parameters(std::vector<solver_type>& solvers, double learning_rate
        ) { update_parameters(make_sstack(solvers), learning_rate); }
        /*!
            Convenience method for calling update_parameters()
        !*/

        void back_propagate_error(
            const tensor& x
        );
        /*!
            requires
                - forward(x) was called to forward propagate x though the network.
                  Moreover, this was the most recent call to forward() and x has not been
                  subsequently modified in any way.
                - subnet().get_gradient_input() has been set equal to the gradient of this network's
                  output with respect to the loss function (generally this will be done by calling
                  compute_loss()).
            ensures
                - Back propagates the error gradient, subnet().get_gradient_input(), through this
                  network and computes parameter and data gradients, via backpropagation.
                  Specifically, this function populates get_final_data_gradient() and also,
                  for each layer, the tensor returned by get_parameter_gradient().
                - All elements of #subnet().get_gradient_input() are set to 0. 
                - have_same_dimensions(#get_final_data_gradient(), x) == true.
                - #get_final_data_gradient() contains the gradient of the network with
                  respect to x.
        !*/

        void back_propagate_error(
            const tensor& x, 
            const tensor& gradient_input
        );
        /*!
            requires
                - forward(x) was called to forward propagate x though the network.
                  Moreover, this was the most recent call to forward() and x has not been
                  subsequently modified in any way.
                - have_same_dimensions(gradient_input, subnet().get_output()) == true
            ensures
                - This function is identical to the version of back_propagate_error()
                  defined immediately above except that it back-propagates gradient_input
                  through the network instead of subnet().get_gradient_input().  Therefore, this
                  version of back_propagate_error() is equivalent to performing:
                    subnet().get_gradient_input() = gradient_input;
                    back_propagate_error(x);
                  Except that calling back_propagate_error(x,gradient_input) avoids the
                  copy and is therefore slightly more efficient.
                - All elements of #subnet.get_gradient_input() are set to 0. 
                - have_same_dimensions(#get_final_data_gradient(), x) == true.
                - #get_final_data_gradient() contains the gradient of the network with
                  respect to x.
        !*/

        const tensor& get_final_data_gradient(
        ) const; 
        /*!
            ensures
                - if back_propagate_error() has been called to back-propagate a gradient
                  through this network then you can call get_final_data_gradient() to
                  obtain the last data gradient computed.  That is, this function returns
                  the gradient of the network with respect to its inputs.
                - Note that there is only one "final data gradient" for an entire network,
                  not one per layer, since there is only one input to the entire network.
        !*/


    // -------------

        void clean (
        );
        /*!
            ensures
                - Causes the network to forget about everything but its parameters.  
                - invokes subnet().clean()
        !*/
    };

    template <typename T, typename U> 
    std::ostream& operator<<(std::ostream& out, const add_loss_layer<T,U>& item);
    /*!
        prints the network architecture to the given output stream.
    !*/

    template <typename T, typename U> 
    void serialize(const add_loss_layer<T,U>& item, std::ostream& out);
    template <typename T, typename U> 
    void deserialize(add_loss_layer<T,U>& item, std::istream& in);
    /*!
        provides serialization support  
    !*/

// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------

    template <typename ...T>
    decorator_repeat_group<T...> repeat_group (
        T&& ...args
    );
    /*!
        ensures
            - Decorates a group of variables.  This is essentially like std::make_tuple()
              except it's only purpose is to group variables together so they can be passed
              to the repeat object's constructor.
    !*/

    template <
        size_t num,
        template<typename> class REPEATED_LAYER, 
        typename SUBNET
        >
    class repeat 
    {
        /*!
            REQUIREMENTS ON num
                - num > 0

            REQUIREMENTS ON REPEATED_LAYER
                - REPEATED_LAYER must be a template that stacks more layers onto a deep neural
                  network.  For example, if net_type were a network without a loss layer,
                  then it should be legal to create a deeper network with a type of
                  REPEATED_LAYER<net_type>.

            REQUIREMENTS ON SUBNET
                - One of the following must be true:
                    - SUBNET is an add_layer object.
                    - SUBNET is an add_tag_layer object.
                    - SUBNET is an add_skip_layer object.
                    - SUBNET is a repeat object.

            WHAT THIS OBJECT REPRESENTS
                This object adds more layers to a deep neural network.  In particular, it
                adds REPEATED_LAYER on top of SUBNET num times.  So for example, if num were 2 then
                repeat<2,REPEATED_LAYER,SUBNET> would create a network equivalent to REPEATED_LAYER<REPEATED_LAYER<SUBNET>>.

                Also, this object provides an interface identical to the one defined by the
                add_layer object except that we add the num_repetitions() and
                get_repeated_layer() methods.  These additions are shown below along with
                some additional explanatory comments.
        !*/

    public:

        typedef SUBNET subnet_type;
        typedef typename SUBNET::input_type input_type;
        const static size_t num_computational_layers = (REPEATED_LAYER<SUBNET>::num_computational_layers-SUBNET::num_computational_layers)*num + SUBNET::num_computational_layers;
        const static size_t num_layers = (REPEATED_LAYER<SUBNET>::num_layers-SUBNET::num_layers)*num + SUBNET::num_layers;
        typedef REPEATED_LAYER<an_unspecified_input_type> repeated_layer_type;

        template <typename T, typename ...U>
        repeat(
            T arg1,
            U ...args2
        );
        /*!
            ensures
                - arg1 is used to initialize the num_repetitions() copies of REPEATED_LAYER inside
                  this object.  That is, all the REPEATED_LAYER elements are initialized identically
                  by being given copies of arg1.
                - The rest of the arguments to the constructor, i.e. args2, are passed to
                  SUBNET's constructor.  
        !*/

        template <typename ...T, typename ...U>
        repeat(
            decorator_repeat_group<T...>&& arg1,
            U ...args2
        );
        /*!
            ensures
                - arg1 is used to initialize the num_repetitions() copies of REPEATED_LAYER inside
                  this object.  That is, all the REPEATED_LAYER elements are initialized identically
                  by being given copies of an undecorated arg1.
                - The rest of the arguments to the constructor, i.e. args2, are passed to
                  SUBNET's constructor.  
        !*/

        size_t num_repetitions (
        ) const; 
        /*!
            ensures
                - returns num (i.e. the number of times REPEATED_LAYER was stacked on top of SUBNET)
        !*/

        const repeated_layer_type& get_repeated_layer (
            size_t i 
        ) const;
        /*!
            requires
                - i < num_repetitions()
            ensures
                - returns a reference to the i-th instance of REPEATED_LAYER.  For example,
                  get_repeated_layer(0) returns the instance of REPEATED_LAYER that is on the top of
                  the network while get_repeated_layer(num_repetitions()-1) returns the
                  instance of REPEATED_LAYER that is stacked immediately on top of SUBNET.
        !*/

        repeated_layer_type& get_repeated_layer (
            size_t i 
        );
        /*!
            requires
                - i < num_repetitions()
            ensures
                - returns a reference to the i-th instance of REPEATED_LAYER.  For example,
                  get_repeated_layer(0) returns the instance of REPEATED_LAYER that is on the top of
                  the network while get_repeated_layer(num_repetitions()-1) returns the
                  instance of REPEATED_LAYER that is stacked immediately on top of SUBNET.
        !*/

        const subnet_type& subnet(
        ) const; 
        /*!
            ensures
                - returns the SUBNET base network that repeat sits on top of.  If you want
                  to access the REPEATED_LAYER components then you must use get_repeated_layer(). 
        !*/

        subnet_type& subnet(
        ); 
        /*!
            ensures
                - returns the SUBNET base network that repeat sits on top of.  If you want
                  to access the REPEATED_LAYER components then you must use get_repeated_layer(). 
        !*/
    };

    template < size_t num, template<typename> class T, typename U >
    std::ostream& operator<<(std::ostream& out, const repeat<num,T,U>& item);
    /*!
        prints the network architecture to the given output stream.
    !*/

    template < size_t num, template<typename> class T, typename U >
    void serialize(const repeat<num,T,U>& item, std::ostream& out);
    template < size_t num, template<typename> class T, typename U >
    void deserialize(repeat<num,T,U>& item, std::istream& in);
    /*!
        provides serialization support  
    !*/

// ----------------------------------------------------------------------------------------

    template <
        unsigned long ID, 
        typename SUBNET
        >
    class add_tag_layer
    {
        /*!
            REQUIREMENTS ON SUBNET
                - One of the following must be true:
                    - SUBNET implements the EXAMPLE_INPUT_LAYER interface defined in
                      input_abstract.h.
                    - SUBNET is an add_layer object.
                    - SUBNET is an add_tag_layer object.
                    - SUBNET is an add_skip_layer object.
                    - SUBNET is a repeat object.

            WHAT THIS OBJECT REPRESENTS
                This object adds a new layer to a deep neural network.  However, this layer
                simply performs the identity transform.  This means it is a no-op and its
                presence does not change the behavior of the network.  It exists solely to
                be used by add_skip_layer to reference a particular part of a network.

                Also, this object provides an interface identical to the one defined by the
                add_layer object.
        !*/
    };

    template <unsigned long ID, typename U> 
    std::ostream& operator<<(std::ostream& out, const add_tag_layer<ID,U>& item);
    /*!
        prints the network architecture to the given output stream.
    !*/

    template <unsigned long ID, typename U> 
    void serialize(const add_tag_layer<ID,U>& item, std::ostream& out);
    template <unsigned long ID, typename U> 
    void deserialize(add_tag_layer<ID,U>& item, std::istream& in);
    /*!
        provides serialization support  
    !*/

    template <typename SUBNET> using tag1  = add_tag_layer< 1, SUBNET>;
    template <typename SUBNET> using tag2  = add_tag_layer< 2, SUBNET>;
    template <typename SUBNET> using tag3  = add_tag_layer< 3, SUBNET>;
    template <typename SUBNET> using tag4  = add_tag_layer< 4, SUBNET>;
    template <typename SUBNET> using tag5  = add_tag_layer< 5, SUBNET>;
    template <typename SUBNET> using tag6  = add_tag_layer< 6, SUBNET>;
    template <typename SUBNET> using tag7  = add_tag_layer< 7, SUBNET>;
    template <typename SUBNET> using tag8  = add_tag_layer< 8, SUBNET>;
    template <typename SUBNET> using tag9  = add_tag_layer< 9, SUBNET>;
    template <typename SUBNET> using tag10 = add_tag_layer<10, SUBNET>;

    template <template<typename SUBNET> class tag>
    struct tag_id
    {
        /*!
            REQUIREMENTS ON tag
                Tag should be an add_tag_layer template such as tag1, tag2, etc.

            WHAT THIS OBJECT REPRESENTS
                This is a tool for finding the numeric ID of a tag layer.  For example,
                tag_id<tag3>::id == 3.
        !*/

        const static unsigned long id;
    };

// ----------------------------------------------------------------------------------------

    template <
        template<typename> class TAG_TYPE, 
        typename SUBNET
        >
    class add_skip_layer
    {
        /*!
            REQUIREMENTS ON SUBNET
                - One of the following must be true:
                    - SUBNET is an add_layer object.
                    - SUBNET is an add_tag_layer object.
                    - SUBNET is an add_skip_layer object.
                    - SUBNET is a repeat object.

            WHAT THIS OBJECT REPRESENTS
                This object adds a new layer to a deep neural network which draws its
                inputs from layer<TAG_TYPE>(subnet()) and performs the identity transform.

                Also, this object provides an interface identical to the one defined by the
                add_layer object.
        !*/
    };

    template <template<typename> class T, typename U>
    std::ostream& operator<<(std::ostream& out, const add_skip_layer<T,U>& item);
    /*!
        prints the network architecture to the given output stream.
    !*/

    template <template<typename> class T, typename U>
    void serialize(const add_skip_layer<T,U>& item, std::ostream& out);
    template <template<typename> class T, typename U>
    void deserialize(add_skip_layer<T,U>& item, std::istream& in);
    /*!
        provides serialization support  
    !*/

    template <typename SUBNET> using skip1  = add_skip_layer< tag1, SUBNET>;
    template <typename SUBNET> using skip2  = add_skip_layer< tag2, SUBNET>;
    template <typename SUBNET> using skip3  = add_skip_layer< tag3, SUBNET>;
    template <typename SUBNET> using skip4  = add_skip_layer< tag4, SUBNET>;
    template <typename SUBNET> using skip5  = add_skip_layer< tag5, SUBNET>;
    template <typename SUBNET> using skip6  = add_skip_layer< tag6, SUBNET>;
    template <typename SUBNET> using skip7  = add_skip_layer< tag7, SUBNET>;
    template <typename SUBNET> using skip8  = add_skip_layer< tag8, SUBNET>;
    template <typename SUBNET> using skip9  = add_skip_layer< tag9, SUBNET>;
    template <typename SUBNET> using skip10 = add_skip_layer<tag10, SUBNET>;

// ----------------------------------------------------------------------------------------

    template <
        unsigned int i, 
        typename net_type
        >
    auto& layer (
        net_type& n
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
            - i < net_type::num_layers
        ensures
            - This function allows you to access any layer in a network by its layer index
              i.  Therefore, it will walk i steps down the network and return the layer
              object there.  Since networks can be big, the best way to find layer index
              numbers is to print a network to the screen since the print out will include
              indexes for each layer.
            - In general, this function chains together i calls to n.subnet() and returns
              the result.  So for example:
                - if (i == 0)
                    - returns n
                - else if (i == 1)
                    - returns n.subnet()
                - else if (i == 2)
                    - returns n.subnet().subnet()
                - else if (i == 3)
                    - returns n.subnet().subnet().subnet()
                - else
                    - etc.
              Except that when it hits a repeat layer it recurses into the repeated layers
              contained inside.  That is, if the layer index indicates a layer in a repeat
              object this function will make the appropriate call to get_repeated_layer()
              and do the right thing.
    !*/

    template <
        template<typename> class Match, 
        typename net_type 
        >
    auto& layer (
        net_type& n
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
        ensures
            - returns the first layer in n that is of type Match.  E.g. if net_type is
              fc<relu<fc<input<sample_type>>>> then calling layer<relu>(n) would return
              layer<1>(n), that is, a reference to the relu layer.
    !*/

    template <
        template<typename> class Match, 
        unsigned int i, 
        typename net_type
        >
    auto& layer (
        net_type& n
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
        ensures
            - returns layer<i>(layer<Match>(n))
    !*/

// ----------------------------------------------------------------------------------------

    template <typename net_type>
    auto& input_layer (
        net_type& net
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
        ensures
            - returns the input later of the given network object.  Specifically, this
              function is equivalent to calling:
                layer<net_type::num_layers-1>(net);
              That is, you get the input layer details object for the network.
    !*/

// ----------------------------------------------------------------------------------------

    template <
        typename net_type,
        typename visitor
        >
    void visit_layer_parameters(
        net_type& net,
        visitor v
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
            - v is a function object with a signature equivalent to: 
                v(size_t idx, tensor& t)
              or:
                v(tensor& t)
        ensures
            - Loops over all the computational layers (i.e. layers with parameters, as
              opposed to loss, tag, or input layers) in net and passes their parameters to
              v().  To be specific, this function essentially performs the following:

                size_t computational_layer_idx = 0;
                for (size_t i = 0; i < net_type::num_layers; ++i)
                {
                    if (layer<i>(net) is a computational layer)
                    {
                        v(computational_layer_idx, layer<i>(net).layer_details().get_layer_params());
                        ++computational_layer_idx;
                    }
                }
            - When v() is called, the first argument is always < net_type::num_computational_layers.
    !*/

// ----------------------------------------------------------------------------------------

    template <
        typename net_type,
        typename visitor
        >
    void visit_layer_parameter_gradients(
        net_type& net,
        visitor v
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
            - v is a function object with a signature equivalent to: 
                v(size_t idx, tensor& t)
              or:
                v(tensor& t)
        ensures
            - Loops over all the computational layers (i.e. layers with parameters, as
              opposed to loss, tag, or input layers) in net and passes their parameter
              gradients to v().  To be specific, this function essentially performs the
              following:

                size_t computational_layer_idx = 0;
                for (size_t i = 0; i < net_type::num_layers; ++i)
                {
                    if (layer<i>(net) is a computational layer)
                    {
                        v(computational_layer_idx, layer<i>(net).get_parameter_gradient());
                        ++computational_layer_idx;
                    }
                }
            - When v() is called, the first argument is always < net_type::num_computational_layers.
    !*/

// ----------------------------------------------------------------------------------------

    template <
        typename net_type,
        typename visitor
        >
    void visit_layers(
        net_type& net,
        visitor v
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
            - v is a function object with a signature equivalent to: 
                v(size_t idx, any_net_type& t)
              or:
                v(any_net_type& t)
              That is, it takes an optional size_t and then any of the network types such as
              add_layer, add_loss_layer, etc.
        ensures
            - Loops over all the layers in net and calls v() on them.  To be specific, this
              function essentially performs the following:

                for (size_t i = 0; i < net_type::num_layers; ++i)
                    v(i, layer<i>(net));
    !*/

    template <
        typename net_type,
        typename visitor
        >
    void visit_layers_backwards(
        net_type& net,
        visitor v
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
            - v is a function object with a signature equivalent to: 
                v(size_t idx, any_net_type& t)
              or:
                v(any_net_type& t)
              That is, it takes an optional size_t and then any of the network types such as
              add_layer, add_loss_layer, etc.
        ensures
            - Loops over all the layers in net and calls v() on them.  The loop happens in
              the reverse order of visit_layers().  To be specific, this function
              essentially performs the following:

                for (size_t i = net_type::num_layers; i != 0; --i)
                    v(i-1, layer<i-1>(net));
    !*/

// ----------------------------------------------------------------------------------------

    template <
        typename net_type,
        typename visitor
        >
    void visit_computational_layers(
        net_type& net,
        visitor v
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
            - v is a function object with a signature equivalent to: 
                v(size_t idx, any_computational_layer& t)
              or:
                v(any_computational_layer& t)
              That is, it takes an optional size_t and then any of the computational layers.  E.g.
              one of the layer types defined in dlib/dnn/layers_abstract.h like fc_ or conv_.
        ensures
            - Loops over all the computational layers in net and calls v() on them.  To be specific, this
              function essentially performs the following:

                for (size_t i = 0; i < net_type::num_layers; ++i)
                    if (layer<i>(net) is an add_layer type, i.e. it adds a computational layer)
                        v(i, layer<i>(net).layer_details());
    !*/

    template <
        size_t begin,
        size_t end,
        typename net_type,
        typename visitor
        >
    void visit_computational_layers_range(
        net_type& net,
        visitor v
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
            - v is a function object with a signature equivalent to: 
                v(size_t idx, any_computational_layer& t)
              or:
                v(any_computational_layer& t)
              That is, it takes an optional size_t and then any of the computational layers.  E.g.
              one of the layer types defined in dlib/dnn/layers_abstract.h like fc_ or conv_.
        ensures
            - Loops over all the computational layers in the range [begin,end) in net and calls v()
              on them.  To be specific, this function essentially performs the following:

                for (size_t i = begin; i < end; ++i)
                    if (layer<i>(net) is an add_layer type, i.e. it adds a computational layer)
                        v(i, layer<i>(net).layer_details());
    !*/

// ----------------------------------------------------------------------------------------

    template <
        size_t begin,
        size_t end,
        typename net_type,
        typename visitor
        >
    void visit_layers_range(
        net_type& net,
        visitor v
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
            - v is a function object with a signature equivalent to: 
                v(size_t idx, any_net_type& t)
              or:
                v(any_net_type& t)
              That is, it takes an optional size_t and then any of the network types such as
              add_layer, add_loss_layer, etc.
            - begin <= end <= net_type::num_layers
        ensures
            - Loops over the layers in the range [begin,end) in net and calls v() on them.
              To be specific, this function essentially performs the following:

                for (size_t i = begin; i < end; ++i)
                    v(i, layer<i>(net));
    !*/

    template <
        size_t begin,
        size_t end,
        typename net_type,
        typename visitor
        >
    void visit_layers_backwards_range(
        net_type& net,
        visitor v
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
            - v is a function object with a signature equivalent to: 
                v(size_t idx, any_net_type& t)
              or:
                v(any_net_type& t)
              That is, it takes an optional size_t and then any of the network types such as
              add_layer, add_loss_layer, etc.
            - begin <= end <= net_type::num_layers
        ensures
            - Loops over the layers in the range [begin,end) in net and calls v() on them.
              The loop happens in the reverse order of visit_layers_range().  To be specific,
              this function essentially performs the following:

                for (size_t i = end; i != begin; --i)
                    v(i-1, layer<i-1>(net));
    !*/

// ----------------------------------------------------------------------------------------

    template <
        unsigned long tag_id,
        typename net_type,
        typename visitor
        >
    void visit_layers_until_tag(
        net_type& net,
        visitor v
    );
    /*!
        requires
            - net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
              add_tag_layer.
            - v is a function object with a signature equivalent to: 
                v(any_net_type& t)
              That is, it must take any of the network types such as add_layer,
              add_loss_layer, etc.
        ensures
            - Loops over all the layers in net beginning with layer<0>(net) and going until
              a tag layer with an ID of tag_id is encountered.  To be specific, this
              function essentially performs the following:

                size_t i = 0;
                while(layer<i>(net) isn't an add_tag_layer with ID == tag_id) {
                    v(layer<i>(net));
                    ++i;
                }
                v(layer<i>(net));  // also visits the tag layer itself at the very end.
    !*/

// ----------------------------------------------------------------------------------------

    struct layer_test_results
    {
        std::string log;
        bool was_good;

        operator bool() const { return was_good; }
    };

    inline std::ostream& operator<< (std::ostream& out, const layer_test_results& item)
    {
        out << item.log;
        return out;
    }

    template <
        typename layer_details_type
        >
    layer_test_results test_layer (
        layer_details_type l
    );
    /*!
        ensures
            - Checks if l correctly implements the EXAMPLE_COMPUTATIONAL_LAYER_ interface
              defined in layers_abstract.h.  Importantly, it computes numerical approximations 
              to the gradients and compares them to the outputs of the layer.  
            - The results of the testing are returned.  In particular, if the returned object
              is RESULT then we will have:
                - RESULT.was_good == false if and only if the layer failed the testing.
                - RESULT.log == a string describing why the testing failed if was_good==false.
            - Note that this function is only capable of checking layers that take
              arbitrary subnetworks as input.  So if you have designed a layer that expects
              only a certain restricted type of subnetwork then you might get a compile or
              runtime error when you call this function.
    !*/

// ----------------------------------------------------------------------------------------

}

#endif // DLIB_DNn_CORE_ABSTRACT_H_