File size: 72,328 Bytes
d3b8c8f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
28c5370
 
 
d3b8c8f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3

# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto.  Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.

# E-RADIO (FasterViTv2) model from
# Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. "AM-RADIO: Agglomerative Model--Reduce All Domains Into One." arXiv preprint arXiv:2312.06709 (2023).

# based on FasterViT, Swin Transformer, YOLOv8

# FasterViT:
# Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M. Alvarez, Jan Kautz, and Pavlo Molchanov. "FasterViT: Fast Vision Transformers with Hierarchical Attention." arXiv preprint arXiv:2306.06189 (2023).

import timm
import torch
import torch.nn as nn
from timm.models.registry import register_model

from timm.models.layers import trunc_normal_, DropPath, LayerNorm2d
import numpy as np
import torch.nn.functional as F
import math
import warnings

#######################
## Codebase from YOLOv8
## BEGINNING
#######################

class C2f(nn.Module):
    """Faster Implementation of CSP Bottleneck with 2 convolutions."""
    """From YOLOv8 codebase"""
    def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, drop_path=None):  # ch_in, ch_out, number, shortcut, groups, expansion
        super().__init__()
        if drop_path is None:
            drop_path = [0.0] * n

        self.c = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, 2 * self.c, 1, 1)
        self.cv2 = Conv((2 + n) * self.c, c2, 1)  # optional act=FReLU(c2)
        self.m = nn.ModuleList(Bottleneck(self.c, self.c, shortcut, g, k=((3, 3), (3, 3)), e=1.0, drop_path=drop_path[i]) for i in range(n))

    def forward(self, x):
        """Forward pass through C2f layer."""
        y = list(self.cv1(x).chunk(2, 1))
        y.extend(m(y[-1]) for m in self.m)
        return self.cv2(torch.cat(y, 1))

    def forward_split(self, x):
        """Forward pass using split() instead of chunk()."""
        y = list(self.cv1(x).split((self.c, self.c), 1))
        y.extend(m(y[-1]) for m in self.m)
        return self.cv2(torch.cat(y, 1))

class Bottleneck(nn.Module):
    """Standard bottleneck."""

    def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5, drop_path=0.0):  # ch_in, ch_out, shortcut, groups, kernels, expand
        super().__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, k[0], 1)
        self.cv2 = Conv(c_, c2, k[1], 1, g=g)
        self.add = shortcut and c1 == c2
        self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()

    def forward(self, x):
        """'forward()' applies the YOLOv5 FPN to input data."""
        return x + self.drop_path1(self.cv2(self.cv1(x))) if self.add else self.cv2(self.cv1(x))


class Conv(nn.Module):
    """Modified to support layer fusion"""
    default_act = nn.SiLU()  # default activation

    def __init__(self, a, b, kernel_size=1, stride=1, padding=None, g=1, dilation=1, bn_weight_init=1, bias=False, act=True):
        super().__init__()

        self.conv = torch.nn.Conv2d(a, b, kernel_size, stride, autopad(kernel_size, padding, dilation), dilation, g, bias=False)
        if 1:
            self.bn = torch.nn.BatchNorm2d(b)
            torch.nn.init.constant_(self.bn.weight, bn_weight_init)
            torch.nn.init.constant_(self.bn.bias, 0)
        self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()


    def forward(self,x):
        x = self.conv(x)
        x = self.bn(x)
        x = self.act(x)
        return x

    @torch.no_grad()
    def switch_to_deploy(self):
        # return 1
        if not isinstance(self.bn, nn.Identity):
            c, bn = self.conv, self.bn
            w = bn.weight / (bn.running_var + bn.eps) ** 0.5
            w = c.weight * w[:, None, None, None]
            b = bn.bias - bn.running_mean * bn.weight / \
                (bn.running_var + bn.eps)**0.5

            self.conv.weight.data.copy_(w)
            self.conv.bias = nn.Parameter(b)

            self.bn = nn.Identity()

def autopad(k, p=None, d=1):  # kernel, padding, dilation
    """Pad to 'same' shape outputs."""
    if d > 1:
        k = d * (k - 1) + 1 if isinstance(k, int) else [d * (x - 1) + 1 for x in k]  # actual kernel-size
    if p is None:
        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
    return p


#######################
## Codebase from YOLOv8
## END
#######################

def pixel_unshuffle(data, factor=2):
    # performs nn.PixelShuffle(factor) in reverse, torch has some bug for ONNX and TRT, so doing it manually
    B, C, H, W = data.shape
    return data.view(B, C, factor, H//factor, factor, W//factor).permute(0,1,2,4,3,5).reshape(B, -1, H//factor, W//factor)

class SwiGLU(nn.Module):
    # should be more advanced, but doesnt improve results so far
    def forward(self, x):
        x, gate = x.chunk(2, dim=-1)
        return F.silu(gate) * x


def window_partition(x, window_size):
    """
    Function for partitioning image into windows and later do windowed attention
    Args:
        x: (B, C, H, W)
        window_size: window size
    Returns:
        windows - local window features (num_windows*B, window_size*window_size, C)
        (Hp, Wp) -  the size of the padded image
    """
    B, C, H, W = x.shape

    if window_size == 0 or (window_size==H and window_size==W):
        windows = x.flatten(2).transpose(1, 2)
        Hp, Wp = H, W
    else:
        pad_h = (window_size - H % window_size) % window_size
        pad_w = (window_size - W % window_size) % window_size
        if pad_h > 0 or pad_w > 0:
            x = F.pad(x, (0, pad_w, 0, pad_h), mode="reflect")
        Hp, Wp = H + pad_h, W + pad_w

        x = x.view(B, C, Hp // window_size, window_size, Wp // window_size, window_size)
        windows = x.permute(0, 2, 4, 3, 5, 1).reshape(-1, window_size*window_size, C)

    return windows, (Hp, Wp)

class Conv2d_BN(nn.Module):
    '''
    Conv2d + BN layer with folding capability to speed up inference
    Can be merged with Conv() function with additional arguments
    '''
    def __init__(self, a, b, kernel_size=1, stride=1, padding=0, dilation=1, groups=1, bn_weight_init=1, bias=False):
        super().__init__()
        self.conv = torch.nn.Conv2d(a, b, kernel_size, stride, padding, dilation, groups, bias=False)
        if 1:
            self.bn = torch.nn.BatchNorm2d(b)
            torch.nn.init.constant_(self.bn.weight, bn_weight_init)
            torch.nn.init.constant_(self.bn.bias, 0)

    def forward(self,x):
        x = self.conv(x)
        x = self.bn(x)
        return x

    @torch.no_grad()
    def switch_to_deploy(self):
        if not isinstance(self.bn, nn.Identity):
            c, bn = self.conv, self.bn
            w = bn.weight / (bn.running_var + bn.eps) ** 0.5
            w = c.weight * w[:, None, None, None]
            b = bn.bias - bn.running_mean * bn.weight / \
                (bn.running_var + bn.eps)**0.5
            self.conv.weight.data.copy_(w)
            self.conv.bias = nn.Parameter(b)
            self.bn = nn.Identity()



def window_reverse(windows, window_size, H, W, pad_hw):
    """
    Windows to the full feature map
    Args:
        windows: local window features (num_windows*B, window_size, window_size, C)
        window_size: Window size
        H: Height of image
        W: Width of image
        pad_w - a tuple of image passing used in windowing step
    Returns:
        x: (B, C, H, W)

    """
    # print(f"window_reverse, windows.shape {windows.shape}")
    Hp, Wp = pad_hw
    if window_size == 0 or (window_size==H and window_size==W):
        B = int(windows.shape[0] / (Hp * Wp / window_size / window_size))
        x = windows.transpose(1, 2).view(B, -1, H, W)
    else:
        B = int(windows.shape[0] / (Hp * Wp / window_size / window_size))
        x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
        x = x.permute(0, 5, 1, 3, 2, 4).reshape(B,windows.shape[2], Hp, Wp)

        if Hp > H or Wp > W:
            x = x[:, :, :H, :W, ].contiguous()

    return x



class PosEmbMLPSwinv2D(nn.Module):
    """
    2D positional embedding from Swin Transformer v2
    Added functionality to store the positional embedding in the model and not recompute it every time
    """
    def __init__(
        self, window_size, pretrained_window_size, num_heads, seq_length, no_log=False, cpb_mlp_hidden=512,
    ):
        super().__init__()
        self.window_size = window_size
        self.num_heads = num_heads
        # mlp to generate continuous relative position bias
        self.cpb_mlp = nn.Sequential(
            nn.Linear(2, cpb_mlp_hidden, bias=True),
            nn.ReLU(inplace=True),
            nn.Linear(cpb_mlp_hidden, num_heads, bias=False),
        )

        self.grid_exists = False
        self.seq_length = seq_length
        self.deploy = False
        self.num_heads = num_heads
        self.no_log = no_log
        self.pretrained_window_size = pretrained_window_size
        self.relative_bias_window_size = window_size

        relative_coords_table, relative_position_index, relative_bias = self.relative_bias_initialization(window_size, num_heads,
                                                                                                     pretrained_window_size, seq_length,
                                                                                                     no_log)

        self.register_buffer("relative_coords_table", relative_coords_table)
        self.register_buffer("relative_position_index", relative_position_index)
        self.register_buffer("relative_bias", relative_bias)  # for EMA

    def relative_bias_initialization(self, window_size, num_heads, pretrained_window_size, seq_length, no_log):
        # as in separate function to support window size chage after model weights loading
        relative_coords_h = torch.arange(
            -(window_size[0] - 1), window_size[0], dtype=torch.float32
        )
        relative_coords_w = torch.arange(
            -(window_size[1] - 1), window_size[1], dtype=torch.float32
        )
        relative_coords_table = (
            torch.stack(torch.meshgrid([relative_coords_h, relative_coords_w]))
            .permute(1, 2, 0)
            .contiguous()
            .unsqueeze(0)
        )  # 1, 2*Wh-1, 2*Ww-1, 2
        if pretrained_window_size[0] > 0:
            relative_coords_table[:, :, :, 0] /= pretrained_window_size[0] - 1
            relative_coords_table[:, :, :, 1] /= pretrained_window_size[1] - 1
        else:
            relative_coords_table[:, :, :, 0] /= self.window_size[0] - 1
            relative_coords_table[:, :, :, 1] /= self.window_size[1] - 1

        if not no_log:
            relative_coords_table *= 8  # normalize to -8, 8
            relative_coords_table = (
                torch.sign(relative_coords_table)
                * torch.log2(torch.abs(relative_coords_table) + 1.0)
                / np.log2(8)
            )

        # get pair-wise relative position index for each token inside the window
        coords_h = torch.arange(self.window_size[0])
        coords_w = torch.arange(self.window_size[1])
        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
        relative_coords = (
            coords_flatten[:, :, None] - coords_flatten[:, None, :]
        )  # 2, Wh*Ww, Wh*Ww
        relative_coords = relative_coords.permute(
            1, 2, 0
        ).contiguous()  # Wh*Ww, Wh*Ww, 2
        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww

        relative_bias = torch.zeros(1, num_heads, seq_length, seq_length)

        self.relative_bias_window_size = window_size

        return relative_coords_table, relative_position_index, relative_bias


    def switch_to_deploy(self):
        self.deploy = True
        self.grid_exists = True

    def forward(self, input_tensor):
        # for efficiency, we want this forward to be folded into a single operation (sum)
        # if resolution stays the same, then we dont need to recompute MLP layers

        if not self.deploy or self.training:
            self.grid_exists = False

        #compare if all elements in self.window_size list match those in self.relative_bias_window_size
        if not all([self.window_size[i] == self.relative_bias_window_size[i] for i in range(len(self.window_size))]):
            relative_coords_table, relative_position_index, relative_bias = self.relative_bias_initialization(self.window_size, self.num_heads,
                                                                                                        self.pretrained_window_size, self.seq_length,
                                                                                                        self.no_log)

            self.relative_coords_table = relative_coords_table.to(self.relative_coords_table.device)
            self.relative_position_index = relative_position_index.to(self.relative_position_index.device)
            self.relative_bias = relative_bias.to(self.relative_bias.device)

        if self.deploy and self.grid_exists:
            input_tensor = input_tensor + self.relative_bias
            return input_tensor

        if 1:
            self.grid_exists = True

            relative_position_bias_table = self.cpb_mlp(
                self.relative_coords_table
            ).view(-1, self.num_heads)
            relative_position_bias = relative_position_bias_table[
                self.relative_position_index.view(-1)
            ].view(
                self.window_size[0] * self.window_size[1],
                self.window_size[0] * self.window_size[1],
                -1,
            )  # Wh*Ww,Wh*Ww,nH

            relative_position_bias = relative_position_bias.permute(
                2, 0, 1
            ).contiguous()  # nH, Wh*Ww, Wh*Ww
            relative_position_bias = 16 * torch.sigmoid(relative_position_bias)

            self.relative_bias = relative_position_bias.unsqueeze(0)

        input_tensor = input_tensor + self.relative_bias
        return input_tensor


class GRAAttentionBlock(nn.Module):
    def __init__(self, window_size, dim_in, dim_out,
                 num_heads, drop_path=0., qk_scale=None, qkv_bias=False,
                 norm_layer=nn.LayerNorm, layer_scale=None,
                  use_swiglu=True,
                  subsample_ratio=1, dim_ratio=1, conv_base=False,
                  do_windowing=True, multi_query=False, use_shift=0,
                  cpb_mlp_hidden=512, conv_groups_ratio=0):
        '''
        Global Resolution Attention Block , see README for details
        Attention with subsampling to get a bigger receptive field for attention
        conv_base - use conv2d instead of avgpool2d for downsample / upsample


        '''
        super().__init__()

        self.shift_size=window_size//2 if use_shift else 0

        self.do_windowing = do_windowing
        self.subsample_ratio = subsample_ratio



        if do_windowing:
            if conv_base:
                    self.downsample_op = nn.Conv2d(dim_in, dim_out, kernel_size=subsample_ratio, stride=subsample_ratio) if subsample_ratio > 1 else nn.Identity()


                    self.downsample_mixer = nn.Identity()
                    self.upsample_mixer = nn.Identity()
                    self.upsample_op = nn.ConvTranspose2d(dim_in, dim_out, kernel_size=subsample_ratio, stride=subsample_ratio) if subsample_ratio > 1 else nn.Identity()
            else:
                self.downsample_op = nn.AvgPool2d(kernel_size=subsample_ratio, stride=subsample_ratio) if subsample_ratio > 1 else nn.Identity()
                self.downsample_mixer = Conv2d_BN(dim_in, dim_out, kernel_size=1, stride=1) if subsample_ratio > 1 else nn.Identity()
                self.upsample_mixer = nn.Upsample(scale_factor=subsample_ratio, mode='nearest') if subsample_ratio > 1 else nn.Identity()
                self.upsample_op = Conv2d_BN(dim_in, dim_out, kernel_size=1, stride=1, padding=0, bias=False) if subsample_ratio > 1 else nn.Identity()


        # in case there is no downsampling conv we want to have it separately
        # will help with information propagation between windows
        if subsample_ratio == 1:
            # conv_groups_ratio=0
            self.pre_conv = Conv2d_BN(dim_in, dim_in, kernel_size=3, stride=1, padding=1, groups=max(1,int(conv_groups_ratio*dim_in)), bias=False)
            # self.pre_conv = nn.Conv2d(dim_in, dim_in, kernel_size=3, stride=1, padding=1, groups=max(1,int(conv_groups_ratio*dim_in)), bias=False)
            # self.pre_conv_act = nn.ReLU6()
            #for simplicity:
            self.pre_conv_act = nn.Identity()
            if conv_groups_ratio == -1:
                self.pre_conv = nn.Identity()
                self.pre_conv_act = nn.Identity()

        self.window_size = window_size

        self.norm1 = norm_layer(dim_in)

        self.attn = WindowAttention(
            dim_in,
            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
            resolution=window_size,
            seq_length=window_size**2, dim_out=dim_in, multi_query=multi_query,
            shift_size=self.shift_size, cpb_mlp_hidden=cpb_mlp_hidden)

        self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()

        use_layer_scale = layer_scale is not None and type(layer_scale) in [int, float]
        self.gamma1 = nn.Parameter(layer_scale * torch.ones(dim_in))  if use_layer_scale else 1

        ### mlp layer
        mlp_ratio = 4
        self.norm2 = norm_layer(dim_in)
        mlp_hidden_dim = int(dim_in * mlp_ratio)

        activation = nn.GELU if not use_swiglu else SwiGLU
        mlp_hidden_dim = int((4 * dim_in * 1 / 2) / 64) * 64 if use_swiglu else mlp_hidden_dim

        self.mlp = Mlp(in_features=dim_in, hidden_features=mlp_hidden_dim, act_layer=activation, use_swiglu=use_swiglu)

        self.gamma2 = nn.Parameter(layer_scale * torch.ones(dim_in)) if layer_scale else 1
        self.drop_path2=DropPath(drop_path) if drop_path > 0. else nn.Identity()


    def forward(self, x):
        skip_connection = x
        attn_mask = None

        # in case there is no downsampling conv we want to have it separately
        # will help with information propagation
        if self.subsample_ratio == 1:
            x = self.pre_conv_act(self.pre_conv(x)) + skip_connection

        if self.do_windowing:
            # performing windowing if required
            x = self.downsample_op(x)
            x = self.downsample_mixer(x)

            if self.window_size>0:
                H, W = x.shape[2], x.shape[3]

            if self.shift_size > 0 and H>self.window_size and W>self.window_size:
                # @swin like cyclic shift, doesnt show better performance
                x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(2, 3))

            x, pad_hw = window_partition(x, self.window_size)

            if self.shift_size > 0 and H>self.window_size and W>self.window_size:
                # set atten matrix to have -100 and the top right square
                # attn[:, :, :-self.shift_size, -self.shift_size:] = -100.0
                # calculate attention mask for SW-MSA
                # not used in final version, can be useful for some cases especially for high res
                H, W = pad_hw
                img_mask = torch.zeros((1, H, W, 1), device=x.device)  # 1 H W 1
                h_slices = (slice(0, -self.window_size),
                            slice(-self.window_size, -self.shift_size),
                            slice(-self.shift_size, None))
                w_slices = (slice(0, -self.window_size),
                            slice(-self.window_size, -self.shift_size),
                            slice(-self.shift_size, None))
                cnt = 0
                for h in h_slices:
                    for w in w_slices:
                        img_mask[:, h, w, :] = cnt
                        cnt += 1
                img_mask = img_mask.transpose(1,2).transpose(1,3)
                mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1

                mask_windows = mask_windows[0].view(-1, self.window_size * self.window_size)
                attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
                attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))

        # window attention
        x = x + self.drop_path1(self.gamma1*self.attn(self.norm1(x), attn_mask=attn_mask)) # or pass H,W
        # mlp layer
        x = x + self.drop_path2(self.gamma2*self.mlp(self.norm2(x)))

        if self.do_windowing:
            if self.window_size > 0:
                x = window_reverse(x, self.window_size, H, W, pad_hw)

            # reverse cyclic shift
            if self.shift_size > 0 and H>self.window_size and W>self.window_size:
                # @swin like cyclic shift, not tested
                x = torch.roll(x, shifts=(self.shift_size, self.shift_size), dims=(2, 3))

            x = self.upsample_mixer(x)
            x = self.upsample_op(x)


            if x.shape[2] != skip_connection.shape[2] or x.shape[3] != skip_connection.shape[3]:
                x = torch.nn.functional.pad(x, ( 0, -x.shape[3] + skip_connection.shape[3], 0, -x.shape[2] + skip_connection.shape[2]), mode="reflect")
        # need to add skip connection because downsampling and upsampling will break residual connection
        # 0.5 is needed to make sure that the skip connection is not too strong
        # in case of no downsample / upsample we can show that 0.5 compensates for the residual connection
        x = 0.5 * x + 0.5 * skip_connection
        return x




class MultiResolutionAttention(nn.Module):
    """
    MultiResolutionAttention (MRA) module
    The idea is to use multiple attention blocks with different resolution
    Feature maps are downsampled / upsampled for each attention block on different blocks
    Every attention block supports windowing
    """

    def __init__(self, window_size, sr_ratio,
                 dim, dim_ratio, num_heads,
                 do_windowing=True,
                 layer_scale=1e-5, norm_layer=nn.LayerNorm,
                 drop_path = 0, qkv_bias=False, qk_scale=1.0,
                 use_swiglu=True, multi_query=False, conv_base=False,
                 use_shift=0, cpb_mlp_hidden=512, conv_groups_ratio=0) -> None:
        """
        Args:
            input_resolution: input image resolution
            window_size: window size
            compression_ratio: compression ratio
            max_depth: maximum depth of the GRA module
            use_shift: do window shifting
        """
        super().__init__()

        depth = len(sr_ratio)

        self.attention_blocks = nn.ModuleList()


        for i in range(depth):
            subsample_ratio = sr_ratio[i]
            if len(window_size) > i:
                window_size_local = window_size[i]
            else:
                window_size_local = window_size[0]

            self.attention_blocks.append(GRAAttentionBlock(window_size=window_size_local,
                                            dim_in=dim, dim_out=dim, num_heads=num_heads,
                                            qkv_bias=qkv_bias, qk_scale=qk_scale, norm_layer=norm_layer,
                                            layer_scale=layer_scale, drop_path=drop_path,
                                            use_swiglu=use_swiglu, subsample_ratio=subsample_ratio, dim_ratio=dim_ratio,
                                            do_windowing=do_windowing, multi_query=multi_query, conv_base=conv_base,
                                            use_shift=use_shift, cpb_mlp_hidden=cpb_mlp_hidden, conv_groups_ratio=conv_groups_ratio),
                                        )

    def forward(self, x):

        for attention_block in self.attention_blocks:
            x = attention_block(x)

        return x



class Mlp(nn.Module):
    """
    Multi-Layer Perceptron (MLP) block
    """

    def __init__(self,
                 in_features,
                 hidden_features=None,
                 out_features=None,
                 act_layer=nn.GELU,
                 use_swiglu=True,
                 drop=0.):
        """
        Args:
            in_features: input features dimension.
            hidden_features: hidden features dimension.
            out_features: output features dimension.
            act_layer: activation function.
            drop: dropout rate.
        """

        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features * (2 if use_swiglu else 1), bias=False)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features, bias=False)

    def forward(self, x):
        x_size = x.size()
        x = x.view(-1, x_size[-1])
        x = self.fc1(x)
        x = self.act(x)
        x = self.fc2(x)
        x = x.view(x_size)
        return x

class Downsample(nn.Module):
    """
    Down-sampling block
    Pixel Unshuffle is used for down-sampling, works great accuracy - wise but takes 10% more TRT time
    """

    def __init__(self,
                 dim,
                 shuffle = False,
                 ):
        """
        Args:
            dim: feature size dimension.
            shuffle: idea with
            keep_dim: bool argument for maintaining the resolution.
        """

        super().__init__()
        dim_out = 2 * dim

        if shuffle:
            self.norm = lambda x: pixel_unshuffle(x, factor=2)
            self.reduction = Conv2d_BN(dim*4, dim_out, 1, 1, 0, bias=False)
            # pixel unshuffleging works well but doesnt provide any speedup
        else:
            # removed layer norm for better, in this formulation we are getting 10% better speed
            # LayerNorm for high resolution inputs will be a pain as it pools over the entire spatial dimension
            # therefore we remove it compared to the original implementation in FasterViTv1
            self.norm = nn.Identity()
            self.reduction = Conv2d_BN(dim, dim_out, 3, 2, 1, bias=False)


    def forward(self, x):
        x = self.norm(x)
        x = self.reduction(x)
        return x


class PatchEmbed(nn.Module):
    """
    Patch embedding block
    Used to convert image into an initial set of feature maps with lower resolution
    """

    def __init__(self, in_chans=3, in_dim=64, dim=96, shuffle_down=False):
        """
        Args:
            in_chans: number of input channels.
            in_dim: intermediate feature size dimension to speed up stem.
            dim: final stem channel number
            shuffle_down: use PixelUnshuffle for down-sampling, effectively increases the receptive field
        """

        super().__init__()
        # shuffle_down = False
        if not shuffle_down:
            self.proj = nn.Identity()
            self.conv_down = nn.Sequential(
                Conv2d_BN(in_chans, in_dim, 3, 2, 1, bias=False),
                nn.ReLU(),
                Conv2d_BN(in_dim, dim, 3, 2, 1, bias=False),
                nn.ReLU()
                )
        else:
            self.proj = lambda x: pixel_unshuffle(x, factor=4)
            self.conv_down = nn.Sequential(Conv2d_BN(in_chans*16, dim, 3, 1, 1),
                                           nn.ReLU(),
                                           )

    def forward(self, x):
        x = self.proj(x)
        x = self.conv_down(x)
        return x



class ConvBlock(nn.Module):
    """
    Convolutional block, used in first couple of stages
    Experimented with plan resnet-18 like modules, they are the best in terms of throughput
    Finally, YOLOv8 idea seem to work fine (resnet-18 like block with squeezed feature dimension, and feature concatendation at the end)
    """
    def __init__(self, dim,
                 drop_path=0.,
                 layer_scale=None,
                 kernel_size=3,
                 ):
        super().__init__()

        self.conv1 = Conv2d_BN(dim, dim, kernel_size=kernel_size, stride=1, padding=1)
        self.act1 = nn.GELU()

        self.conv2 = Conv2d_BN(dim, dim, kernel_size=kernel_size, stride=1, padding=1)

        self.layer_scale = layer_scale
        if layer_scale is not None and type(layer_scale) in [int, float]:
            self.gamma = nn.Parameter(layer_scale * torch.ones(dim))
            self.layer_scale = True
        else:
            self.layer_scale = False
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()

    def forward(self, x):
        input = x

        x = self.conv1(x)
        x = self.act1(x)
        x = self.conv2(x)

        if self.layer_scale:
            x = x * self.gamma.view(1, -1, 1, 1)
        x = input + self.drop_path(x)
        return x


class WindowAttention(nn.Module):
    # Windowed Attention from SwinV2
    # use a MLP trick to deal with various input image resolutions, then fold it to improve speed

    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, resolution=0,
                 seq_length=0, dim_out=None, multi_query=False, shift_size=0, cpb_mlp_hidden=512):
        # taken from EdgeViT and tweaked with attention bias.
        super().__init__()
        if not dim_out: dim_out = dim
        self.shift_size = shift_size
        self.multi_query = multi_query
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.head_dim = dim // num_heads

        self.dim_internal = dim

        self.scale = qk_scale or head_dim ** -0.5
        if not multi_query:
            self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        else:
            self.qkv = nn.Linear(dim, dim + 2*self.head_dim, bias=qkv_bias)

        self.proj = nn.Linear(dim, dim_out, bias=False)
        # attention positional bias
        self.pos_emb_funct = PosEmbMLPSwinv2D(window_size=[resolution, resolution],
                                              pretrained_window_size=[resolution, resolution],
                                              num_heads=num_heads,
                                              seq_length=seq_length,
                                              cpb_mlp_hidden=cpb_mlp_hidden)

        self.resolution = resolution

    def forward(self, x, attn_mask = None):
        B, N, C = x.shape

        if not self.multi_query:
            qkv = self.qkv(x).reshape(B, -1, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
            q, k, v = qkv[0], qkv[1], qkv[2]
        else:
            qkv = self.qkv(x)
            (q, k, v) = qkv.split([self.dim_internal, self.head_dim, self.head_dim], dim=2)

            q = q.reshape(B, -1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
            k = k.reshape(B, -1, 1, C // self.num_heads).permute(0, 2, 1, 3)
            v = v.reshape(B, -1, 1, C // self.num_heads).permute(0, 2, 1, 3)

        attn = (q @ k.transpose(-2, -1)) * self.scale

        attn = self.pos_emb_funct(attn)

        #add window shift
        if attn_mask is not None:
            nW = attn_mask.shape[0]
            attn = attn.view(B // nW, nW, self.num_heads, N, N) + attn_mask.unsqueeze(1).unsqueeze(0)
            attn = attn.view(-1, self.num_heads, N, N)

        attn = attn.softmax(dim=-1)
        x = (attn @ v).transpose(1, 2).reshape(B, -1, C)
        x = self.proj(x)
        return x



class FasterViTLayer(nn.Module):
    """
    fastervitlayer
    """

    def __init__(self,
                 dim,
                 depth,
                 num_heads,
                 window_size,
                 conv=False,
                 downsample=True,
                 mlp_ratio=4.,
                 qkv_bias=False,
                 qk_scale=None,
                 norm_layer=nn.LayerNorm,
                 drop_path=0.,
                 layer_scale=None,
                 layer_scale_conv=None,
                 sr_dim_ratio=1,
                 sr_ratio=1,
                 multi_query=False,
                 use_swiglu=True,
                 yolo_arch=False,
                 downsample_shuffle=False,
                 conv_base=False,
                 use_shift=False,
                 cpb_mlp_hidden=512,
                 conv_groups_ratio=0,
                 verbose: bool = True,

    ):
        """
        Args:
            dim: feature size dimension.
            depth: number of layers in each stage.
            input_resolution: input image resolution.
            window_size: window size in each stage.
            downsample: bool argument for down-sampling.
            mlp_ratio: MLP ratio.
            num_heads: number of heads in each stage.
            qkv_bias: bool argument for query, key, value learnable bias.
            qk_scale: bool argument to scaling query, key.
            drop: dropout rate.
            attn_drop: attention dropout rate.
            drop_path: drop path rate.
            norm_layer: normalization layer.
            layer_scale: layer scaling coefficient.
            use_shift: SWIN like window shifting for half the window size for every alternating layer (considering multi-resolution)
            conv_groups_ratio: group ratio for conv when no subsampling in multi-res attention
        """

        super().__init__()
        self.conv = conv
        self.yolo_arch=False
        self.verbose = verbose
        if conv:
            if not yolo_arch:
                self.blocks = nn.ModuleList([
                    ConvBlock(dim=dim,
                            drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
                            layer_scale=layer_scale_conv)
                    for i in range(depth)])
                self.blocks = nn.Sequential(*self.blocks)
            else:
                self.blocks = C2f(dim,dim,n=depth,shortcut=True,e=0.5)
                self.yolo_arch=True
        else:
            if not isinstance(window_size, list): window_size = [window_size]
            self.window_size = window_size[0]
            self.do_single_windowing = True
            if not isinstance(sr_ratio, list): sr_ratio = [sr_ratio]
            self.sr_ratio = sr_ratio
            if any([sr!=1 for sr in sr_ratio]) or len(set(window_size))>1:
                self.do_single_windowing = False
                do_windowing = True
            else:
                self.do_single_windowing = True
                do_windowing = False

            #for v2_2
            if conv_groups_ratio != -1:
                self.do_single_windowing = False
                do_windowing = True

            self.blocks = nn.ModuleList()
            for i in range(depth):
                self.blocks.append(
                    MultiResolutionAttention(window_size=window_size,
                                             sr_ratio=sr_ratio,
                                             dim=dim,
                                             dim_ratio = sr_dim_ratio,
                                             num_heads=num_heads,
                                             norm_layer=norm_layer,
                                             drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
                                             layer_scale=layer_scale,
                                             qkv_bias=qkv_bias,
                                             qk_scale=qk_scale,
                                             use_swiglu=use_swiglu,
                                             do_windowing=do_windowing,
                                             multi_query=multi_query,
                                             conv_base=conv_base,
                                             cpb_mlp_hidden=cpb_mlp_hidden,
                                             use_shift =0 if ((not use_shift) or ((i) % 2 == 0)) else True    ,
                                             conv_groups_ratio=conv_groups_ratio,
                    ))
            self.blocks = nn.Sequential(*self.blocks)

        self.transformer = not conv
        self.downsample = None if not downsample else Downsample(dim=dim, shuffle=downsample_shuffle)


    def forward(self, x):
        B, C, H, W = x.shape

        # do padding for transforemr
        interpolate = True
        if self.transformer and interpolate:
            # Windowed Attention will split feature map into windows with the size of window_size x window_size
            # if the resolution is not divisible by window_size, we need to interpolate the feature map
            # can be done via padding, but doing so after training hurts the model performance.
            # interpolation affects the performance as well, but not as much as padding
            if isinstance(self.window_size, list) or isinstance(self.window_size, tuple):
                current_max_window_size = max(self.window_size)
            else:
                current_max_window_size = self.window_size

            max_window_size = max([res_upsample*current_max_window_size for res_upsample in self.sr_ratio])
            if H % max_window_size != 0 or W % max_window_size != 0:
                new_h = int(np.ceil(H/max_window_size)*max_window_size)
                new_w = int(np.ceil(W/max_window_size)*max_window_size)
                x = F.interpolate(x, size=(new_h, new_w), mode='nearest')
                if self.verbose:
                    warnings.warn(f"Choosen window size is not optimal for given resolution. Interpolation of features maps will be done and it can affect the performance. Max window size is {max_window_size}, feature map size is {H}x{W}, interpolated feature map size is {new_h}x{new_w}.")


        if self.transformer and self.do_single_windowing:
            H, W = x.shape[2], x.shape[3]
            x, pad_hw = window_partition(x, self.window_size)

        #run main blocks
        x = self.blocks(x)

        if self.transformer and self.do_single_windowing:
            x = window_reverse(x, self.window_size, H, W, pad_hw)

        if self.transformer and interpolate:
            #lets keep original resolution, might be not ideal, but for the upsampling tower we need to keep the expected resolution.
            x = F.interpolate(x, size=(H, W), mode='nearest')

        if self.downsample is None:
            return x, x

        return self.downsample(x), x  # changing to output pre downsampled features


class InterpolateLayer(nn.Module):
    def __init__(self, size=None, scale_factor=None, mode='nearest'):
        super(InterpolateLayer, self).__init__()
        self.size = size
        self.scale_factor = scale_factor
        self.mode = mode

    def forward(self, x):
        return F.interpolate(x, size=self.size, scale_factor=self.scale_factor, mode=self.mode)


class HiResNeck(nn.Module):
    """
    The block is used to output dense features from all stages
    Otherwise, by default, only the last stage features are returned with FasterViTv2
    """
    def __init__(self, dim, depths, neck_start_stage, full_features_head_dim, downsample_enabled):

        '''
        Hi Resolution neck to support output of high res features that are useful for dense tasks.
        depths - total number of layers in the base model
        neck_start_stage - when to start the neck, 0 - start from the first stage, 1 - start from the second stage etc.
                            earlier layers result in higher resolution features at the cost of compute
        full_features_head_dim - number of channels in the dense features head
        '''
        super().__init__()
        # create feature projection layers for segmentation output
        self.neck_features_proj = nn.ModuleList()
        self.neck_start_stage = neck_start_stage
        upsample_ratio = 1
        for i in range(len(depths)):
            level_n_features_output = int(dim * 2 ** i)

            if self.neck_start_stage > i: continue

            if (upsample_ratio > 1) or full_features_head_dim!=level_n_features_output:
                feature_projection = nn.Sequential()
                if False:
                    feature_projection.add_module("norm",nn.BatchNorm2d(level_n_features_output)) #fast, but worse
                    feature_projection.add_module("dconv", nn.ConvTranspose2d(level_n_features_output,
                                                                            full_features_head_dim, kernel_size=upsample_ratio, stride=upsample_ratio))
                else:
                    # B, in_channels, H, W -> B, in_channels, H*upsample_ratio, W*upsample_ratio
                    # print("upsample ratio", upsample_ratio, level_n_features_output, level_n_features_output)
                    feature_projection.add_module("upsample", InterpolateLayer(scale_factor=upsample_ratio, mode='nearest'))
                    feature_projection.add_module("conv1", nn.Conv2d(level_n_features_output, level_n_features_output, kernel_size=3, stride=1, padding=1, groups=level_n_features_output))
                    feature_projection.add_module("norm",nn.BatchNorm2d(level_n_features_output))
                    # B, in_channels, H*upsample_ratio, W*upsample_ratio -> B, full_features_head_dim, H*upsample_ratio, W*upsample_ratio
                    feature_projection.add_module("conv2", nn.Conv2d(level_n_features_output, full_features_head_dim, kernel_size=1, stride=1, padding=0))
            else:
                feature_projection = nn.Sequential()

            self.neck_features_proj.append(feature_projection)

            if i>0 and downsample_enabled[i]:
                upsample_ratio *= 2

    def forward(self, x, il_level=-1, full_features=None):
        if self.neck_start_stage > il_level:
            return full_features

        if full_features is None:
            full_features = self.neck_features_proj[il_level - self.neck_start_stage](x)
        else:
            #upsample torch tensor x to match full_features size, and add to full_features
            feature_projection = self.neck_features_proj[il_level - self.neck_start_stage](x)
            if feature_projection.shape[2] != full_features.shape[2] or feature_projection.shape[3] != full_features.shape[3]:
                feature_projection = torch.nn.functional.pad(feature_projection, ( 0, -feature_projection.shape[3] + full_features.shape[3], 0, -feature_projection.shape[2] + full_features.shape[2]))
            full_features = full_features + feature_projection
        return full_features

class FasterViT(nn.Module):
    """
    FasterViT
    """

    def __init__(self,
                 dim,
                 in_dim,
                 depths,
                 window_size,
                 mlp_ratio,
                 num_heads,
                 drop_path_rate=0.2,
                 in_chans=3,
                 num_classes=1000,
                 qkv_bias=False,
                 qk_scale=None,
                 layer_scale=None,
                 layer_scale_conv=None,
                 layer_norm_last=False,
                 sr_ratio = [1, 1, 1, 1],
                 max_depth = -1,
                 conv_base=False,
                 use_swiglu=False,
                 multi_query=False,
                 norm_layer=nn.LayerNorm,
                 drop_uniform=False,
                 yolo_arch=False,
                 shuffle_down=False,
                 downsample_shuffle=False,
                 return_full_features=False,
                 full_features_head_dim=128,
                 neck_start_stage=1,
                 use_neck=False,
                 use_shift=False,
                 cpb_mlp_hidden=512,
                 conv_groups_ratio=0,
                 verbose: bool = False,
                 **kwargs):
        """
        Args:
            dim: feature size dimension.
            depths: number of layers in each stage.
            window_size: window size in each stage.
            mlp_ratio: MLP ratio.
            num_heads: number of heads in each stage.
            drop_path_rate: drop path rate.
            in_chans: number of input channels.
            num_classes: number of classes.
            qkv_bias: bool argument for query, key, value learnable bias.
            qk_scale: bool argument to scaling query, key.
            drop_rate: dropout rate.
            attn_drop_rate: attention dropout rate.
            norm_layer: normalization layer.
            layer_scale: layer scaling coefficient.
            return_full_features: output dense features as well as logits
            full_features_head_dim: number of channels in the dense features head
            neck_start_stage: a stage id to start full feature neck. Model has 4 stages, indix starts with 0
                                for 224 resolution, the output of the stage before downsample:
                                stage 0: 56x56, stage 1: 28x28, stage 2: 14x14, stage 3: 7x7
            use_neck: even for summarization embedding use neck
            use_shift: SWIN like window shifting but without masking attention
            conv_groups_ratio: will be used for conv blocks where there is no multires attention,
                                if 0 then normal conv,
                                if 1 then channels are independent,
                                if -1 then no conv at all

        """
        super().__init__()

        num_features = int(dim * 2 ** (len(depths) - 1))
        self.num_classes = num_classes
        self.patch_embed = PatchEmbed(in_chans=in_chans, in_dim=in_dim, dim=dim, shuffle_down=shuffle_down)
        # set return_full_features true if we want to return full features from all stages
        self.return_full_features = return_full_features
        self.use_neck = use_neck

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
        if drop_uniform:
            dpr = [drop_path_rate for x in range(sum(depths))]

        if not isinstance(max_depth, list): max_depth = [max_depth] * len(depths)

        self.levels = nn.ModuleList()
        for i in range(len(depths)):
            conv = True if (i == 0 or i == 1) else False

            level = FasterViTLayer(dim=int(dim * 2 ** i),
                                   depth=depths[i],
                                   num_heads=num_heads[i],
                                   window_size=window_size[i],
                                   mlp_ratio=mlp_ratio,
                                   qkv_bias=qkv_bias,
                                   qk_scale=qk_scale,
                                   conv=conv,
                                   drop_path=dpr[sum(depths[:i]):sum(depths[:i + 1])],
                                   downsample=(i < len(depths) - 1),
                                   layer_scale=layer_scale,
                                   layer_scale_conv=layer_scale_conv,
                                   sr_ratio=sr_ratio[i],
                                   use_swiglu=use_swiglu,
                                   multi_query=multi_query,
                                   norm_layer=norm_layer,
                                   yolo_arch=yolo_arch,
                                   downsample_shuffle=downsample_shuffle,
                                   conv_base=conv_base,
                                   cpb_mlp_hidden=cpb_mlp_hidden,
                                   use_shift=use_shift,
                                   conv_groups_ratio=conv_groups_ratio,
                                   verbose=verbose)

            self.levels.append(level)

        if self.return_full_features or self.use_neck:
            #num_heads
            downsample_enabled = [self.levels[i-1].downsample is not None for i in range(len(self.levels))]
            self.high_res_neck = HiResNeck(dim, depths, neck_start_stage, full_features_head_dim, downsample_enabled)

        self.switched_to_deploy = False

        self.norm = LayerNorm2d(num_features) if layer_norm_last else nn.BatchNorm2d(num_features)
        self.avgpool = nn.AdaptiveAvgPool2d(1)
        self.head = nn.Linear(num_features, num_classes) if num_classes > 0 else nn.Identity()
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, LayerNorm2d):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.BatchNorm2d):
            nn.init.ones_(m.weight)
            nn.init.zeros_(m.bias)

    @torch.jit.ignore
    def no_weight_decay_keywords(self):
        return {'rpb'}

    def forward_features(self, x):
        _, _, H, W = x.shape
        if H % 32 != 0 or W % 32 != 0:
            raise ValueError(f"E-RADIO requires input dimensions to be divisible by 32 but got H x W: {H} x {W}")
        x = self.patch_embed(x)
        full_features = None
        for il, level in enumerate(self.levels):
            x, pre_downsample_x = level(x)

            if self.return_full_features or self.use_neck:
                full_features = self.high_res_neck(pre_downsample_x, il, full_features)

        # x = self.norm(full_features if (self.return_full_features or self.use_neck) else x)
        x = self.norm(x) # new version for

        if not self.return_full_features:
            return x, None

        return x, full_features

    def forward(self, x):
        x, full_features = self.forward_features(x)

        x = self.avgpool(x)
        x = torch.flatten(x, 1)

        x = self.head(x)
        if full_features is not None:
            return x, full_features
        return x

    def switch_to_deploy(self):
        '''
        A method to perform model self-compression
        merges BN into conv layers
        converts MLP relative positional bias into precomputed buffers
        '''
        if not self.switched_to_deploy:
            for level in [self.patch_embed, self.levels, self.head]:
                for module in level.modules():
                    if hasattr(module, 'switch_to_deploy'):
                        module.switch_to_deploy()
        self.switched_to_deploy = True


    def change_window_size(self, new_window_size):
        """
        FasterViT employs windowed attention, which may be sensitive to the choice of this parameter,
        especially in cases of uneven partitioning of the feature maps.
        FasterViT allows for the adjustment of the window size after training,
        making it adaptable to different input image resolutions.
        The recommended values for window size based on input resolution are as follows:

        Input Resolution | Window Size
        224 | 7
        256 | 8
        386 | 12
        512 | 16
        Ideally, the window size should be a factor of the input resolution. In the third stage, we divide the resolution by 16, so the window size should be
        img_res/16/2
        for the third stage and img_res/32 for the last stage. While this can be applied in a brute-force manner, a better way is to do model.change_window_size.
        Manual way to change resolution -> model.change_window_size(resolution)
        """
        window_size = new_window_size
        print(f"Setting window size to {window_size}")
        for module in self.modules():
            if hasattr(module, "window_size"):
                # check if tuple or a number
                if isinstance(module.window_size, tuple):
                    if module.window_size[0] != window_size:
                        module.window_size = (window_size, window_size)
                elif isinstance(module.window_size, list):
                    if module.window_size[0] != window_size:
                        module.window_size = [window_size, window_size]
                else:
                    module.window_size = window_size


    def set_optimal_window_size(self, image_dim, max_window_size = 16):
        """
        Using hand picked window size for various resolutions.

        FasterViT employs windowed attention, which may be sensitive to the choice of this parameter,
        especially in cases of uneven partitioning of the feature maps.
        FasterViT allows for the adjustment of the window size after training,
        making it adaptable to different input image resolutions.
        The recommended values for window size based on input resolution are as follows:

        Input Resolution | Window Size
        224 | 7
        256 | 8
        386 | 12
        512 | 16
        Ideally, the window size should be a factor of the input resolution. In the third stage, we divide the resolution by 16, so the window size should be
        img_res/16/2
        for the third stage and img_res/32 for the last stage. While this can be applied in a brute-force manner, a better way is to do model.change_window_size.
        Manual way to change resolution -> model.change_window_size(resolution)

        """
        # import math

        def divisorGenerator(n):
            large_divisors = []
            for i in range(1, int(math.sqrt(n) + 1)):
                if n % i == 0:
                    yield i
                    if i*i != n:
                        large_divisors.append(n / i)
            for divisor in reversed(large_divisors):
                yield divisor

        if isinstance(image_dim, list) or isinstance(image_dim, tuple):
            image_dim = min(image_dim)

        # we do windowed attention in the 3rd stage for the first time, therefore //16,
        # we do subsampled attention with downsample by 2 so need to get //32 actually
        # ideally we should rewrite this to be dependent on the structure of the model like what if subsampled is removed etc
        all_divisors = np.array(list(divisorGenerator(image_dim//32)))
        new_window_size = int(min(all_divisors[all_divisors <= max_window_size][-1], max_window_size))

        # for image_dim in [128, 224, 256, 384, 512, 768, 1024]:
        #     all_divisors = np.array(list(divisorGenerator(image_dim//32)))
        #     new_window_size = int(min(all_divisors[all_divisors <= max_window_size][-1], max_window_size))
        #     print(f"Setting window size to {new_window_size} for image resolution {image_dim}")

        self.change_window_size(new_window_size = new_window_size)

# 83.44200001953125
@register_model
def fastervit2_small(pretrained=False, **kwargs): #,
    model = FasterViT(depths=[3, 3, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=96,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.2,
                     sr_ratio=[1, 1, [1, 2], 1],
                     use_swiglu=False,
                     downsample_shuffle=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

# 82.61
@register_model
def fastervit2_tiny(pretrained=False, **kwargs): #,
    model = FasterViT(depths=[1, 3, 4, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=80,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.2,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     downsample_shuffle=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

#'top1', 84.31800001220704
@register_model
def fastervit2_base(pretrained=False, **kwargs):
    model = FasterViT(depths=[3, 3, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=128,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.2,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     conv_base=True,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

#84.39999999267579
@register_model
def fastervit2_base_v1(pretrained=False, **kwargs):
    model = FasterViT(depths=[4, 4, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=128,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.2,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     conv_base=True,
                     downsample_shuffle=False,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

@register_model
def fastervit2_base_fullres1(pretrained=False, **kwargs):
    model = FasterViT(depths=[3, 3, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=128,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.2,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     conv_base=True,
                     use_neck=True,
                     full_features_head_dim=1024,
                     neck_start_stage=2,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

@register_model
def fastervit2_base_fullres2(pretrained=False, **kwargs):
    model = FasterViT(depths=[3, 3, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=128,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.2,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     conv_base=True,
                     use_neck=True,
                     full_features_head_dim=512,
                     neck_start_stage=1,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

@register_model
def fastervit2_base_fullres3(pretrained=False, **kwargs):
    model = FasterViT(depths=[3, 3, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=128,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.2,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     conv_base=True,
                     use_neck=True,
                     full_features_head_dim=256,
                     neck_start_stage=1,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

@register_model
def fastervit2_base_fullres4(pretrained=False, **kwargs):
    model = FasterViT(depths=[3, 3, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=128,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.2,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     conv_base=True,
                     use_neck=True,
                     full_features_head_dim=256,
                     neck_start_stage=2,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

@register_model
def fastervit2_base_fullres5(pretrained=False, **kwargs):
    model = FasterViT(depths=[3, 3, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=128,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.2,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     conv_base=True,
                     use_neck=True,
                     full_features_head_dim=512,
                     neck_start_stage=2,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

#84.87
@register_model
def fastervit2_large(pretrained=False, **kwargs):
    model = FasterViT(depths=[3, 3, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=128+64,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.3,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     yolo_arch=False,
                     shuffle_down=False,
                     cpb_mlp_hidden=64,
                     conv_base=True,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

@register_model
def fastervit2_large_fullres(pretrained=False, **kwargs):
    model = FasterViT(
        depths=[3, 3, 5, 5],
        num_heads=[2, 4, 8, 16],
        window_size=[None, None, [7, 7], 7],
        dim=192,
        in_dim=64,
        mlp_ratio=4,
        drop_path_rate=0.0,
        sr_ratio=[1, 1, [2, 1], 1],
        use_swiglu=False,
        yolo_arch=True,
        shuffle_down=False,
        conv_base=True,
        use_neck=True,
        full_features_head_dim=1536,
        neck_start_stage=2,
        **kwargs,
    )
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model


@register_model
def fastervit2_large_fullres_ws8(pretrained=False, **kwargs):
    model = FasterViT(
        depths=[3, 3, 5, 5],
        num_heads=[2, 4, 8, 16],
        window_size=[None, None, [8, 8], 8],
        dim=192,
        in_dim=64,
        mlp_ratio=4,
        drop_path_rate=0.0,
        sr_ratio=[1, 1, [2, 1], 1],
        use_swiglu=False,
        yolo_arch=True,
        shuffle_down=False,
        conv_base=True,
        use_neck=True,
        full_features_head_dim=1536,
        neck_start_stage=2,
        **kwargs,
    )
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model


@register_model
def fastervit2_large_fullres_ws16(pretrained=False, **kwargs):
    model = FasterViT(
        depths=[3, 3, 5, 5],
        num_heads=[2, 4, 8, 16],
        window_size=[None, None, [16, 16], 16],
        dim=192,
        in_dim=64,
        mlp_ratio=4,
        drop_path_rate=0.0,
        sr_ratio=[1, 1, [2, 1], 1],
        use_swiglu=False,
        yolo_arch=True,
        shuffle_down=False,
        conv_base=True,
        use_neck=True,
        full_features_head_dim=1536,
        neck_start_stage=2,
        **kwargs,
    )
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model


@register_model
def fastervit2_large_fullres_ws32(pretrained=False, **kwargs):
    model = FasterViT(
        depths=[3, 3, 5, 5],
        num_heads=[2, 4, 8, 16],
        window_size=[None, None, [32, 32], 32],
        dim=192,
        in_dim=64,
        mlp_ratio=4,
        drop_path_rate=0.0,
        sr_ratio=[1, 1, [2, 1], 1],
        use_swiglu=False,
        yolo_arch=True,
        shuffle_down=False,
        conv_base=True,
        use_neck=True,
        full_features_head_dim=1536,
        neck_start_stage=2,
        **kwargs,
    )
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

#85.23% top1
@register_model
def fastervit2_xlarge(pretrained=False, **kwargs):
    model = FasterViT(depths=[3, 3, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=128+128+64,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.4,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     yolo_arch=False,
                     shuffle_down=False,
                     cpb_mlp_hidden=64,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

@register_model
def fastervit2_huge(pretrained=False, **kwargs):
    model = FasterViT(depths=[3, 3, 5, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=128+128+128+64,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.2,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model


# 81.61
@register_model
def fastervit2_xtiny(pretrained=False, **kwargs): #,
    model = FasterViT(depths=[1, 3, 4, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=64,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.1,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     downsample_shuffle=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     cpb_mlp_hidden=64,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model


# 80.19
@register_model
def fastervit2_xxtiny(pretrained=False, **kwargs): #,
    model = FasterViT(depths=[1, 3, 4, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=48,
                     in_dim=64,
                     mlp_ratio=4,
                     drop_path_rate=0.05,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     downsample_shuffle=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     cpb_mlp_hidden=64,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

@register_model
# 77.0
def fastervit2_xxxtiny(pretrained=False, **kwargs): #,
    model = FasterViT(depths=[1, 3, 4, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=32,
                     in_dim=32,
                     mlp_ratio=4,
                     drop_path_rate=0.0,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     downsample_shuffle=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     cpb_mlp_hidden=64,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model


@register_model
def fastervit2_xxxtiny_fullres(pretrained=False, **kwargs):
    model = FasterViT(depths=[1, 3, 4, 5],
                     num_heads=[2, 4, 8, 16],
                     window_size=[8, 8, [7, 7], 7],
                     dim=32,
                     in_dim=32,
                     mlp_ratio=4,
                     drop_path_rate=0.0,
                     sr_ratio=[1, 1, [2, 1], 1],
                     use_swiglu=False,
                     downsample_shuffle=False,
                     yolo_arch=True,
                     shuffle_down=False,
                     cpb_mlp_hidden=64,
                     use_neck=True,
                     full_features_head_dim=128,
                     neck_start_stage=1,
                     conv_groups_ratio = 1,
                     **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

@register_model
def eradio_xxxtiny(pretrained=False, **kwargs):  # ,
    model = FasterViT(
        depths=[1, 3, 4, 5],
        num_heads=[2, 4, 8, 16],
        window_size=[None, None, [16, 16], 16],
        dim=32,
        in_dim=32,
        mlp_ratio=4,
        drop_path_rate=0.0,
        sr_ratio=[1, 1, [2, 1], 1],
        use_swiglu=False,
        yolo_arch=True,
        shuffle_down=False,
        conv_base=True,
        use_neck=True,
        full_features_head_dim=256,
        neck_start_stage=2,
        **kwargs,
    )
    if pretrained:
        model.load_state_dict(torch.load(pretrained))
    return model

@register_model
def eradio_xxxtiny_8x_ws12(pretrained=False, **kwargs):
    model = FasterViT(depths=[1, 3, 4, 5],
        num_heads=[2, 4, 8, 16],
        window_size=[None, None, [12, 12], 12],
        dim=32,
        in_dim=32,
        mlp_ratio=4,
        drop_path_rate=0.0,
        sr_ratio=[1, 1, [2, 1], 1],
        use_swiglu=False,
        downsample_shuffle=False,
        yolo_arch=True,
        shuffle_down=False,
        cpb_mlp_hidden=64,
        use_neck=True,
        full_features_head_dim=256,
        neck_start_stage=2,
        conv_groups_ratio = 1,
        **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model


@register_model
def eradio_xxxtiny_8x_ws16(pretrained=False, **kwargs):
    model = FasterViT(depths=[1, 3, 4, 5],
        num_heads=[2, 4, 8, 16],
        window_size=[None, None, [16, 16], 16],
        dim=32,
        in_dim=32,
        mlp_ratio=4,
        drop_path_rate=0.0,
        sr_ratio=[1, 1, [2, 1], 1],
        use_swiglu=False,
        downsample_shuffle=False,
        yolo_arch=True,
        shuffle_down=False,
        cpb_mlp_hidden=64,
        use_neck=True,
        full_features_head_dim=256,
        neck_start_stage=1,
        conv_groups_ratio = 1,
        **kwargs)
    if pretrained:
        model.load_state_dict(torch.load(pretrained)["state_dict"])
    return model

@register_model
def eradio(pretrained=False, **kwargs):
    return fastervit2_large_fullres_ws16(pretrained=pretrained, **kwargs)