Skip to content

model_trainer

sleap_nn.training.model_trainer

This module is to train a sleap-nn model using Lightning.

Classes:

Name Description
ModelTrainer

Train sleap-nn model using PyTorch Lightning.

ModelTrainer

Train sleap-nn model using PyTorch Lightning.

This class is used to create dataloaders, train a sleap-nn model and save the model checkpoints/ logs with options to logging with wandb and csvlogger.

Parameters:

Name Type Description Default
config

OmegaConf dictionary which has the following: (i) data_config: data loading pre-processing configs. (ii) model_config: backbone and head configs to be passed to Model class. (iii) trainer_config: trainer configs like accelerator, optimiser params, etc.

required
train_labels

List of sio.Labels objects for training dataset.

required
val_labels

List of sio.Labels objects for validation dataset.

required
skeletons

List of sio.Skeleton objects in a single slp file.

required
lightning_model

One of the child classes of sleap_nn.training.lightning_modules.LightningModel.

required
model_type

Type of the model. One of single_instance, centered_instance, centroid, bottomup, multi_class_bottomup, multi_class_topdown.

required
backbone_type

Backbone model. One of unet, convnext and swint.

required
trainer

Instance of the lightning.Trainer initialized with loggers and callbacks.

required

Methods:

Name Description
get_model_trainer_from_config

Create a model trainer instance from config.

setup_config

Compute config parameters.

train

Train the lightning model.

Source code in sleap_nn/training/model_trainer.py
 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
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
@attrs.define
class ModelTrainer:
    """Train sleap-nn model using PyTorch Lightning.

    This class is used to create dataloaders, train a sleap-nn model and save the model checkpoints/ logs with options to logging
    with wandb and csvlogger.

    Args:
        config: OmegaConf dictionary which has the following:
                (i) data_config: data loading pre-processing configs.
                (ii) model_config: backbone and head configs to be passed to `Model` class.
                (iii) trainer_config: trainer configs like accelerator, optimiser params, etc.
        train_labels: List of `sio.Labels` objects for training dataset.
        val_labels: List of `sio.Labels` objects for validation dataset.
        skeletons: List of `sio.Skeleton` objects in a single slp file.
        lightning_model: One of the child classes of `sleap_nn.training.lightning_modules.LightningModel`.
        model_type: Type of the model. One of `single_instance`, `centered_instance`, `centroid`, `bottomup`, `multi_class_bottomup`, `multi_class_topdown`.
        backbone_type: Backbone model. One of `unet`, `convnext` and `swint`.
        trainer: Instance of the `lightning.Trainer` initialized with loggers and callbacks.
    """

    config: DictConfig
    _initial_config: Optional[DictConfig] = None
    train_labels: List[sio.Labels] = attrs.field(factory=list)
    val_labels: List[sio.Labels] = attrs.field(factory=list)
    skeletons: Optional[List[sio.Skeleton]] = None

    lightning_model: Optional[LightningModel] = None
    model_type: Optional[str] = None
    backbone_type: Optional[str] = None

    _profilers: dict = {
        "advanced": AdvancedProfiler(),
        "passthrough": PassThroughProfiler(),
        "pytorch": PyTorchProfiler(),
        "simple": SimpleProfiler(),
    }

    trainer: Optional[L.Trainer] = None

    @classmethod
    def get_model_trainer_from_config(
        cls,
        config: DictConfig,
        train_labels: Optional[List[sio.Labels]] = None,
        val_labels: Optional[List[sio.Labels]] = None,
    ):
        """Create a model trainer instance from config."""
        model_trainer = cls(config=config)
        model_trainer._initialize_from_config(
            train_labels=train_labels, val_labels=val_labels
        )
        return model_trainer

    def _initialize_from_config(
        self,
        train_labels: Optional[List[sio.Labels]] = None,
        val_labels: Optional[List[sio.Labels]] = None,
    ):
        """Bring a bare trainer up to a trainable state.

        The single initializer behind BOTH entry points --
        :meth:`get_model_trainer_from_config` and the bare-constructor fallback in
        :meth:`train`. It used to be duplicated between them, and the copy in
        ``train()`` had already drifted: it omitted ``_set_seed()`` (so
        ``ModelTrainer(config).train()`` ran with unseeded weight init and
        augmentation even with ``trainer_config.seed`` set) and the
        video-existence check (so a missing video surfaced later, as a confusing
        failure deep in data loading).

        Args:
            train_labels: Labels to train on. ``None`` (with ``val_labels`` also
                ``None``) loads them from ``data_config.train_labels_path``.
            val_labels: Labels to validate on, or ``None`` to split from
                ``train_labels`` / load from ``data_config.val_labels_path``.

        Raises:
            FileNotFoundError: If any video referenced by the labels is missing.
        """
        # Normalize the config first: it fills in optional sections (e.g.
        # `preprocessing.tiling`), and the bare constructor does no verification
        # of its own, so without this the first access to a missing key raises.
        self.config = verify_training_cfg(self.config)

        # Derived from the config, but only when not explicitly provided to the
        # constructor -- an explicit override must survive.
        if self.model_type is None:
            self.model_type = get_model_type_from_cfg(self.config)
        if self.backbone_type is None:
            self.backbone_type = get_backbone_type_from_cfg(self.config)

        if self.config.trainer_config.seed is not None:
            self._set_seed()

        if train_labels is None and val_labels is None:
            # read labels from paths provided in the config
            train_labels = [
                sio.load_slp(path) for path in self.config.data_config.train_labels_path
            ]
            val_labels = (
                [sio.load_slp(path) for path in self.config.data_config.val_labels_path]
                if self.config.data_config.val_labels_path is not None
                else None
            )
        self._setup_train_val_labels(labels=train_labels, val_labels=val_labels)

        # Snapshot the pre-`setup_config` config: it is written out as
        # `initial_config.yaml` at the end of training.
        if self._initial_config is None:
            self._initial_config = self.config.copy()
        # update config parameters
        self.setup_config()

        # Check if all videos exist across all labels
        all_videos_exist = all(
            video.exists(check_all=True)
            for labels in [*self.train_labels, *self.val_labels]
            for video in labels.videos
        )

        if not all_videos_exist:
            raise FileNotFoundError(
                "One or more video files do not exist or are not accessible."
            )

    def _set_seed(self):
        """Set seed for the current experiment."""
        seed = self.config.trainer_config.seed

        random.seed(seed)

        # torch
        torch.manual_seed(seed)

        # if cuda is available
        if torch.cuda.is_available():
            torch.cuda.manual_seed(seed)

        # lightning
        L.seed_everything(seed)

        # numpy
        np.random.seed(seed)

    def _get_trainer_devices(self):
        """Get trainer devices."""
        trainer_devices = (
            self.config.trainer_config.trainer_devices
            if self.config.trainer_config.trainer_devices is not None
            else "auto"
        )
        if (
            trainer_devices == "auto"
            and OmegaConf.select(
                self.config, "trainer_config.trainer_device_indices", default=None
            )
            is not None
        ):
            trainer_devices = len(
                OmegaConf.select(
                    self.config,
                    "trainer_config.trainer_device_indices",
                    default=None,
                )
            )
        elif trainer_devices == "auto":
            if torch.cuda.is_available():
                trainer_devices = torch.cuda.device_count()
            elif torch.backends.mps.is_available():
                trainer_devices = 1
            elif torch.xpu.is_available():
                trainer_devices = torch.xpu.device_count()
            else:
                trainer_devices = 1
        return trainer_devices

    # Model types whose training targets are segmentation masks (``lf.masks``)
    # rather than keypoint instances. Mask-only labels — no pose skeleton at all —
    # are a normal input for these, so a frame with user masks and zero instances
    # is a perfectly good training frame.
    _MASK_TARGET_MODEL_TYPES = (
        "bottomup_segmentation",
        "centered_instance_segmentation",
        "semantic_segmentation",
        "embedding",
    )

    def _is_training_frame(self, lf: "sio.LabeledFrame") -> bool:
        """Whether a frame carries a usable training target.

        Normally that means user instances. Two model families read something else:

        - The centroid model can train on frames carrying only user centroid
          annotations (``frame.centroids``) with no pose instance — the
          pure-centroid seeding case (label a body-center per animal before any
          keypoints exist).
        - Mask-target models (see ``_MASK_TARGET_MODEL_TYPES``) train on
          ``lf.masks``. Mask-only datasets carry **zero** instances, so counting
          only user instances rejected them outright: every split came back empty
          and ``_validate_nonempty_labels`` raised "No labeled frames available
          for train" before setup finished. That made the `embedding` model
          untrainable on its primary dataset, and would do the same to
          `semantic_segmentation` on mask-only labels.
        """
        if lf.has_user_instances:
            return True
        if self.model_type == "centroid":
            return any(not c.is_predicted for c in lf.centroids)
        if self.model_type in self._MASK_TARGET_MODEL_TYPES:
            return any(not m.is_predicted for m in (getattr(lf, "masks", None) or []))
        return False

    def _count_labeled_frames(
        self, labels_list: List[sio.Labels], user_only: bool = True
    ) -> int:
        """Count labeled frames, optionally filtering to trainable frames only.

        Args:
            labels_list: List of Labels objects to count frames from.
            user_only: If True, count only frames with a training target
                (user instances, or — for the centroid model — user centroids).

        Returns:
            Total count of labeled frames.
        """
        total = 0
        for label in labels_list:
            if user_only:
                total += sum(1 for lf in label if self._is_training_frame(lf))
            else:
                total += len(label)
        return total

    def _filter_to_user_labeled(self, labels: sio.Labels) -> sio.Labels:
        """Filter a Labels object to only include trainable frames.

        Args:
            labels: Labels object to filter.

        Returns:
            New Labels object containing only frames with a training target
            (user instances, or user centroids for the centroid model).
        """
        # Filter labeled frames to only trainable ones
        user_lfs = [lf for lf in labels if self._is_training_frame(lf)]

        # Set instances to user instances only (empty for centroid-only frames)
        for lf in user_lfs:
            lf.instances = lf.user_instances

        # Create new Labels with filtered frames
        return sio.Labels(
            labeled_frames=user_lfs,
            videos=labels.videos,
            skeletons=labels.skeletons,
            tracks=labels.tracks,
            suggestions=labels.suggestions,
            provenance=labels.provenance,
        )

    def _split_centroid_labels(
        self, label: sio.Labels, val_fraction: float, seed: Optional[int]
    ):
        """Train/val split for the centroid model that keeps centroid-only frames.

        `sio.Labels.make_training_splits` returns only frames with user
        instances, so it would drop pure-centroid frames. This reimplements a
        deterministic fractional split over frames with a centroid training
        target (user instances OR user centroids).
        """
        frames = [lf for lf in label if self._is_training_frame(lf)]
        for lf in frames:
            lf.instances = lf.user_instances

        def mk(selected):
            return sio.Labels(
                labeled_frames=selected,
                videos=label.videos,
                skeletons=label.skeletons,
                tracks=label.tracks,
                suggestions=label.suggestions,
                provenance=label.provenance,
            )

        n = len(frames)
        if n <= 1:
            # Too few to hold out a val frame; use the same frame for both so
            # training still runs (mirrors small-dataset behavior elsewhere).
            return mk(list(frames)), mk(list(frames))
        rng = np.random.default_rng(seed)
        order = rng.permutation(n)
        n_val = max(1, int(round(n * val_fraction)))
        val_sel = [frames[i] for i in order[:n_val]]
        train_sel = [frames[i] for i in order[n_val:]]
        return mk(train_sel), mk(val_sel)

    def _setup_train_val_labels(
        self,
        labels: Optional[List[sio.Labels]] = None,
        val_labels: Optional[List[sio.Labels]] = None,
    ):
        """Create train and val labels objects. (Initialize `self.train_labels` and `self.val_labels`)."""
        logger.info(f"Creating train-val split...")
        total_train_lfs = 0
        total_val_lfs = 0
        self.skeletons = labels[0].skeletons

        # Mask-only labels: derive the centroid annotations a centroid model needs
        # from the segmentation masks. Must run BEFORE splitting and counting —
        # such labels have no trainable frame at all until it has (#586 / #674) —
        # and on the raw lists, so every split inherits the derived centroids.
        self._apply_centroids_from_masks(labels, val_labels)

        # Check if we should count only user-labeled frames
        user_instances_only = OmegaConf.select(
            self.config, "data_config.user_instances_only", default=True
        )

        # check if all `.slp` file shave same skeleton structure (if multiple slp file paths are provided)
        # Mask-only labels (e.g. instance segmentation) may carry no skeleton.
        if self.skeletons:
            skeleton = self.skeletons[0]
            for index, train_label in enumerate(labels):
                if not train_label.skeletons:
                    continue
                skel_temp = train_label.skeletons[0]
                skeletons_equal = skeleton.matches(skel_temp)
                if not skeletons_equal:
                    message = f"The skeletons in the training labels: {index + 1} do not match the skeleton in the first training label file."
                    logger.error(message)
                    raise ValueError(message)

        # Check for same-data mode (train = val, for intentional overfitting)
        use_same = OmegaConf.select(
            self.config, "data_config.use_same_data_for_val", default=False
        )

        # Group-aware split (SPEC §5.3): when `data_config.split` is configured and no
        # explicit val labels are provided, partition the training labels by a group key
        # (frame/video/identity) instead of the frame-level random validation split.
        split_cfg = OmegaConf.select(self.config, "data_config.split", default=None)
        use_group_split = (
            split_cfg is not None
            and not use_same
            and (val_labels is None or not len(val_labels))
        )

        if use_same:
            # Same mode: use identical data for train and val (for overfitting)
            logger.info("Using same data for train and val (overfit mode)")
            self.train_labels = labels
            self.val_labels = labels
            total_train_lfs = self._count_labeled_frames(labels, user_instances_only)
            total_val_lfs = total_train_lfs
        elif use_group_split:
            from sleap_nn.data.splitting import split_labels_list_train_val

            logger.info(
                "Using group-aware train/val split "
                f"(split_by={OmegaConf.select(self.config, 'data_config.split.split_by', default='frame')})"
            )
            self.train_labels, self.val_labels = split_labels_list_train_val(
                labels, split_cfg
            )
            total_train_lfs = self._count_labeled_frames(
                self.train_labels, user_instances_only
            )
            total_val_lfs = self._count_labeled_frames(
                self.val_labels, user_instances_only
            )
        elif val_labels is None or not len(val_labels):
            # if val labels are not provided, split from train
            val_fraction = OmegaConf.select(
                self.config, "data_config.validation_fraction", default=0.1
            )
            seed = self.config.trainer_config.seed

            # Warn if resuming from a checkpoint with a potentially different seed
            resume_ckpt = OmegaConf.select(
                self.config, "trainer_config.resume_ckpt_path", default=None
            )
            if resume_ckpt is not None:
                orig_config_path = Path(resume_ckpt).parent / "training_config.yaml"
                if orig_config_path.exists():
                    try:
                        orig_cfg = OmegaConf.load(orig_config_path.as_posix())
                        orig_seed = OmegaConf.select(
                            orig_cfg, "trainer_config.seed", default=None
                        )
                        if orig_seed != seed:
                            logger.warning(
                                f"Current seed ({seed}) differs from the original "
                                f"training seed ({orig_seed}) in {orig_config_path}. "
                                f"This will produce a different train/val split and "
                                f"may cause data leakage between train and val sets. "
                                f"Set `trainer_config.seed: {orig_seed}` to preserve "
                                f"the original split."
                            )
                    except Exception:
                        pass
                else:
                    logger.warning(
                        f"Resuming from checkpoint but could not find "
                        f"{orig_config_path} to verify the train/val split seed. "
                        f"Ensure `trainer_config.seed` matches the original "
                        f"training run to avoid data leakage."
                    )
            for label in labels:
                if self.model_type == "centroid":
                    # Centroid-aware split keeps pure-centroid frames that
                    # make_training_splits would drop (no user instances).
                    train_split, val_split = self._split_centroid_labels(
                        label, val_fraction, seed
                    )
                else:
                    train_split, val_split = label.make_training_splits(
                        n_train=1 - val_fraction, n_val=val_fraction, seed=seed
                    )
                self.train_labels.append(train_split)
                self.val_labels.append(val_split)
                total_train_lfs += len(train_split)
                total_val_lfs += len(val_split)
        else:
            self.train_labels = labels
            self.val_labels = val_labels
            total_train_lfs = self._count_labeled_frames(labels, user_instances_only)
            total_val_lfs = self._count_labeled_frames(val_labels, user_instances_only)

        logger.info(f"# Train Labeled frames: {total_train_lfs}")
        logger.info(f"# Val Labeled frames: {total_val_lfs}")

        # Fail fast on an empty split instead of letting training run for several
        # minutes of setup only to crash with a cryptic IndexError the first time
        # something indexes into an empty Labels (e.g. `_verify_model_input_channels`
        # accessing `self.train_labels[0][0]`).
        self._validate_nonempty_labels(total_train_lfs, "train")
        self._validate_nonempty_labels(total_val_lfs, "validation")

        # Single-instance models assume exactly one instance per frame; fail fast
        # with a clear error if any frame has more than one.
        if self.model_type == "single_instance":
            self._validate_single_instance_labels(self.train_labels, "train")
            self._validate_single_instance_labels(self.val_labels, "validation")

    def _apply_centroids_from_masks(self, *label_lists):
        """Derive user centroids from segmentation masks when configured.

        No-op unless ``data_config.centroids_from_masks`` names a method. Mutates
        the labels in place, before any split or frame count: a mask-only dataset
        (no poses, no centroid annotations) has zero trainable frames until this
        has run, so running it later fails the empty-split guard.

        Args:
            *label_lists: The raw ``List[sio.Labels]`` inputs (train, val).
                ``None`` entries are skipped.
        """
        method = OmegaConf.select(
            self.config, "data_config.centroids_from_masks", default=None
        )
        if not method:
            return
        from sleap_nn.data.instance_centroids import add_centroids_from_masks

        for label_list in label_lists:
            for labels in label_list or []:
                add_centroids_from_masks(labels, method=method)

    def _validate_nonempty_labels(self, n_labeled_frames: int, split_name: str):
        """Ensure a split has at least one trainable labeled frame.

        Args:
            n_labeled_frames: Count of labeled frames usable as training targets
                for this split (user instances, or user centroids for the
                centroid model).
            split_name: Name of the split (e.g. "train", "validation") for the
                error message.

        Raises:
            ValueError: If `n_labeled_frames` is zero.
        """
        if n_labeled_frames == 0:
            message = (
                f"No labeled frames available for {split_name}: none of the "
                f"labeled frame(s) in the provided {split_name} labels contain "
                "user-labeled data usable by this model. Predicted instances "
                "and suggestion frames are not used as training targets (nor "
                "are standalone centroid annotations, except by centroid "
                "models). Verify that the .slp file(s) passed for training "
                "contain user-labeled instances, and that `validation_fraction` "
                "and `use_same_data_for_val` are set as intended."
            )
            logger.error(message)
            raise ValueError(message)

    def _validate_single_instance_labels(
        self, labels: List[sio.Labels], split_name: str
    ):
        """Ensure no frame has more than one instance for single-instance models.

        Single-instance confidence-map generation flattens the instance dimension
        (see `sleap_nn.data.confidence_maps.generate_confmaps`), so a frame with
        more than one instance would silently merge multiple animals into a single
        instance with `n_instances * n_nodes` "nodes" and corrupt training. This
        raises a clear error instead, naming the offending frame.

        Args:
            labels: List of `sio.Labels` objects to validate.
            split_name: Name of the split (e.g. "train", "validation") for the
                error message.

        Raises:
            ValueError: If any frame contains more than one (non-empty) instance.
        """
        user_instances_only = OmegaConf.select(
            self.config, "data_config.user_instances_only", default=True
        )
        for label in labels:
            for lf in label:
                if user_instances_only and lf.user_instances is not None:
                    instances = lf.user_instances
                else:
                    instances = lf.instances
                instances = [inst for inst in instances if not inst.is_empty]
                if len(instances) > 1:
                    video_idx = label.videos.index(lf.video)
                    raise ValueError(
                        f"Single-instance training requires at most one instance "
                        f"per frame, but the {split_name} frame at (video index "
                        f"{video_idx}, frame_idx {lf.frame_idx}) has "
                        f"{len(instances)} instances. Remove the extra instance(s) "
                        f"from this frame, or train a multi-instance (top-down or "
                        f"bottom-up) model instead."
                    )

    def _resolve_crop_centroid(self):
        """Resolve the centroid definition the crops will be centered on.

        The crop center is what makes a crop size sufficient or not, so sizing
        the crop requires knowing it. Resolution is deliberately lenient: this
        runs before `_setup_head_config`, which is where a bad ``anchor_part``
        is reported, so an unresolvable anchor degrades to its fallback here and
        lets that validation raise the good error a moment later.

        Returns:
            ``(anchor_ind, anchor_part, method, fallback)`` -- the first, third
            and fourth for `generate_centroids`, and the node's name for
            reporting. ``anchor_part`` is ``None`` when no anchor is in force,
            including when a configured one could not be resolved and was
            degraded to its fallback.
        """
        leaf_paths = {
            "centered_instance": "centered_instance.confmaps",
            "multi_class_topdown": "multi_class_topdown.confmaps",
            "centered_instance_segmentation": (
                "centered_instance_segmentation.segmentation"
            ),
            "embedding": "embedding.embedding",
        }
        leaf_path = leaf_paths.get(self.model_type)
        if leaf_path is None:
            return None, None, None, None
        head_cfg = OmegaConf.select(
            self.config, f"model_config.head_configs.{leaf_path}", default=None
        )

        anchor_part = OmegaConf.select(head_cfg, "anchor_part", default=None)
        anchor_ind = None
        if anchor_part is not None:
            # Resolve against the skeleton rather than the head's `part_names`,
            # which `_setup_head_config` has not populated yet.
            for labels in self.train_labels:
                if labels.skeletons:
                    names = labels.skeletons[0].node_names
                    if anchor_part in names:
                        anchor_ind = names.index(anchor_part)
                    break

        method, fallback = degrade_anchor_if_unresolved(
            *centroid_method_from_config(head_cfg), anchor_ind
        )
        # Report the name only if the anchor actually took effect; an
        # unresolvable one has degraded to `fallback` and naming it would
        # describe a center that is not being used.
        return (
            anchor_ind,
            (anchor_part if anchor_ind is not None else None),
            method,
            fallback,
        )

    def _compute_crop_padding(self, train_label, max_hw):
        """Return the augmentation margin to add to a computed crop size.

        Args:
            train_label: The `sio.Labels` to measure instances from.
            max_hw: The resolved ``(max_height, max_width)``, so the bounding
                box is measured in the space the crop is taken in.

        Returns:
            Padding in pixels, from the config when set, else derived from the
            enabled geometric augmentations.
        """
        padding = self.config.data_config.preprocessing.crop_padding
        if padding is not None:
            return padding

        aug_config = self.config.data_config.augmentation_config
        if not (
            self.config.data_config.use_augmentations_train
            and aug_config is not None
            and aug_config.geometric is not None
        ):
            return 0

        geo = aug_config.geometric
        # Check if rotation is enabled (via rotation_p or affine_p)
        rotation_enabled = (geo.rotation_p is not None and geo.rotation_p > 0) or (
            geo.rotation_p is None
            and geo.scale_p is None
            and geo.translate_p is None
            and geo.affine_p > 0
        )
        # Check if scale is enabled (via scale_p or affine_p)
        scale_enabled = (geo.scale_p is not None and geo.scale_p > 0) or (
            geo.rotation_p is None
            and geo.scale_p is None
            and geo.translate_p is None
            and geo.affine_p > 0
        )
        if not (rotation_enabled or scale_enabled):
            return 0

        # First find the actual max bbox size from labels
        bbox_size = find_max_instance_bbox_size(
            train_label,
            max_hw=max_hw,
            user_instances_only=self.config.data_config.user_instances_only,
        )
        bbox_size = max(
            bbox_size,
            self.config.data_config.preprocessing.min_crop_size or 100,
        )
        rotation_max = (
            max(abs(geo.rotation_min), abs(geo.rotation_max))
            if rotation_enabled
            else 0.0
        )
        scale_max = geo.scale_max if scale_enabled else 1.0
        return compute_augmentation_padding(
            bbox_size=bbox_size,
            rotation_max=rotation_max,
            scale_max=scale_max,
        )

    def _setup_preprocessing_config(self):
        """Setup preprocessing config.

        Runs in two passes: ``max_height``/``max_width`` must be resolved across
        *every* labels file before any crop size can be computed, because the
        size matcher rescales each frame to them and the crop is taken in that
        rescaled space. Measuring a crop in native pixels under-sizes it for any
        video that gets scaled up (see #2862).
        """
        max_height = self.config.data_config.preprocessing.max_height
        max_width = self.config.data_config.preprocessing.max_width

        # Pass 1: resolve the size-matcher target.
        if max_height is None or max_width is None:
            max_h, max_w = 0, 0
            for train_label in self.train_labels:
                current_max_h, current_max_w = get_max_height_width(train_label)
                max_h = max(max_h, current_max_h)
                max_w = max(max_w, current_max_w)
            self.config.data_config.preprocessing.max_height = max_h
            self.config.data_config.preprocessing.max_width = max_w

        if self.model_type not in CROPPING_MODEL_TYPES:
            return

        max_hw = (
            self.config.data_config.preprocessing.max_height,
            self.config.data_config.preprocessing.max_width,
        )
        (
            anchor_ind,
            anchor_part,
            centroid_method,
            centroid_fallback,
        ) = self._resolve_crop_centroid()
        user_instances_only = self.config.data_config.user_instances_only
        crop_size = self.config.data_config.preprocessing.crop_size

        # Pass 2: size the crop when it was not given.
        was_auto = crop_size is None
        if was_auto:
            max_crop_size = 0
            for train_label in self.train_labels:
                padding = self._compute_crop_padding(train_label, max_hw)
                crop_sz = find_instance_crop_size(
                    labels=train_label,
                    padding=padding,
                    maximum_stride=self.config.model_config.backbone_config[
                        f"{self.backbone_type}"
                    ]["max_stride"],
                    min_crop_size=self.config.data_config.preprocessing.min_crop_size,
                    max_hw=max_hw,
                    anchor_ind=anchor_ind,
                    centroid_method=centroid_method,
                    centroid_fallback=centroid_fallback,
                    user_instances_only=user_instances_only,
                )
                max_crop_size = max(max_crop_size, crop_sz)
            self.config.data_config.preprocessing.crop_size = max_crop_size
            crop_size = max_crop_size
            self._log_crop_size(
                max_crop_size, anchor_ind, anchor_part, centroid_method, max_hw
            )

        # Check the resolved crop size against every labeled instance, however it
        # was arrived at. A computed size is derived from the TRAIN split, so a
        # larger val instance can still clip -- and the validation images are
        # exactly where a user would notice it (#2862).
        self._warn_if_crop_size_clips(
            crop_size,
            max_hw,
            anchor_ind,
            centroid_method,
            centroid_fallback,
            was_auto=was_auto,
        )

    def _log_crop_size(
        self, crop_size, anchor_ind, anchor_part, centroid_method, max_hw
    ):
        """Report the computed crop size and what it was derived from.

        The value depends on the size-matcher scale, the centroid the crop is
        centered on and the augmentation margin, none of which a user can infer
        from the number alone -- an off-center anchor alone can double it, so a
        crop several times the animal's width is expected rather than a bug.
        Warned rather than logged at info so it does not scroll past the person
        who is trying to choose a crop size (#2862).

        Args:
            crop_size: The computed crop size, in size-matched pixels.
            anchor_ind: Index of the anchor node, or ``None``.
            anchor_part: Name of the anchor node, for reporting, or ``None``.
            centroid_method: The resolved centroid method.
            max_hw: The resolved ``(max_height, max_width)``.
        """
        scale = self.config.data_config.preprocessing.scale or 1.0
        if anchor_ind is not None:
            center = (
                f"anchor node {anchor_part!r}"
                if anchor_part is not None
                else f"anchor node index {anchor_ind}"
            )
        else:
            center = f"{centroid_method or 'center_of_mass'} centroid"
        message = (
            f"Computed crop size: {crop_size}px, sized to reach every labeled "
            f"node from the {center} it is centered on"
        )
        if max_hw[0] is not None:
            message += (
                f", measured after size matching to {max_hw[0]}x{max_hw[1]} (HxW)"
            )
        if scale != 1.0:
            message += (
                f". Input scaling {scale} is applied to the crop, so the network "
                f"input is {int(crop_size * scale)}px"
            )
        logger.warning(message + ".")

    def _warn_if_crop_size_clips(
        self,
        crop_size,
        max_hw,
        anchor_ind,
        centroid_method,
        centroid_fallback,
        was_auto=False,
    ):
        """Warn when the crop size in force clips labeled instances.

        Checks both splits and reports them separately, because which split an
        instance falls in determines what the user should do about it.

        The crop size is computed from the TRAINING split only -- the behavior
        SLEAP has had since 1.x, where `find_instance_crop_size` was likewise
        passed `training_labels`. Keeping the crop a property of the training
        data means a validation instance larger than anything in training is not
        covered, so it is called out here rather than left to be discovered in a
        validation visualization (#2862). An explicit crop size is never
        overridden either -- clipping an extremity may well be a deliberate trade
        against GPU memory -- but it should not be silent, since the clipped
        nodes are dropped from the targets.

        Args:
            crop_size: The crop size in force, in size-matched pixels.
            max_hw: The resolved ``(max_height, max_width)``.
            anchor_ind: Index of the anchor node, or ``None``.
            centroid_method: The resolved centroid method.
            centroid_fallback: The reduce method for a non-visible anchor node.
            was_auto: Whether ``crop_size`` was computed rather than configured,
                which changes what the user can do about it.
        """

        def tally(labels_list):
            """Return ``(n_clipped, n_total, max_required)`` over one split."""
            n_clipped = n_total = 0
            max_required = 0.0
            for labels in labels_list:
                clipped, total, required = count_clipped_instances(
                    labels,
                    crop_size=crop_size,
                    max_hw=max_hw,
                    anchor_ind=anchor_ind,
                    centroid_method=centroid_method,
                    centroid_fallback=centroid_fallback,
                    user_instances_only=self.config.data_config.user_instances_only,
                )
                n_clipped += clipped
                n_total += total
                max_required = max(max_required, required)
            return n_clipped, n_total, max_required

        # Tallied per split, because which split an instance is in determines
        # what the user should do about it.
        train_clipped, train_total, train_required = tally(self.train_labels)
        val_clipped, val_total, val_required = tally(self.val_labels)
        n_clipped = train_clipped + val_clipped
        if n_clipped == 0:
            return

        n_total = train_total + val_total
        stride = self.config.model_config.backbone_config[f"{self.backbone_type}"][
            "max_stride"
        ]
        origin = "Computed" if was_auto else "Configured"
        message = (
            f"{origin} crop size {crop_size}px clips {n_clipped} of {n_total} "
            f"labeled instances ({train_clipped} in train, {val_clipped} in "
            f"validation): crops are centered on the instance centroid, and these "
            f"instances have nodes further from it than {crop_size // 2}px. Those "
            f"nodes are dropped from the targets."
        )

        if train_clipped:
            # Only ever suggest a size that covers TRAIN. Deriving it from val
            # too would fit a hyperparameter to held-out data, which is the very
            # thing the train-only sizing exists to avoid.
            suggested = math.ceil(train_required / float(stride)) * int(stride)
            message += (
                f" Set crop_size to {suggested} to contain every training instance."
            )

        if val_clipped:
            # Deliberately no size to set: the crop is sized from train alone so
            # that it generalizes to unseen data, and a val instance bigger than
            # anything in train is a coverage problem in the labels, not a knob.
            plural = "" if val_clipped == 1 else "s"
            message += (
                f" The crop size is sized from the training split alone, so that it "
                f"generalizes to new data rather than being fitted to this project's "
                f"validation set -- so {val_clipped} validation instance{plural}, "
                f"larger than anything labeled for training, "
                f"{'is' if val_clipped == 1 else 'are'} not covered."
            )

        logger.warning(message)

    def _get_confmap_sigma(self, output_stride: int) -> float:
        """Return the active head's confmap sigma (input px), else ``output_stride``.

        Used to size the tiling overlap so a keypoint's Gaussian blob is not split
        across a seam. Segmentation / non-confmap heads have no sigma; callers only
        rely on this for confmap-based tiled models (Phase A).

        Args:
            output_stride: Head output stride, used as the fallback value.

        Returns:
            The confmap Gaussian sigma in input pixels.
        """
        head_cfg = self.config.model_config.head_configs[self.model_type]
        if head_cfg is not None:
            for head_layer in head_cfg:
                sub = head_cfg[head_layer]
                if sub is not None and "sigma" in sub and sub["sigma"] is not None:
                    return float(sub["sigma"])
        return float(output_stride)

    def _setup_tiling_config(self):
        """Auto-size tiling geometry from labels + backbone, writing back to config.

        No-op unless ``data_config.preprocessing.tiling.enabled``. Explicit
        ``tile_size`` / ``overlap`` values are preserved; only ``None`` values are
        auto-sized. Must run after :func:`check_output_strides` (so ``max_stride`` /
        ``output_stride`` are finalized). :func:`check_tiling` then validates and
        snaps the resulting geometry.
        """
        tiling = self.config.data_config.preprocessing.tiling
        if tiling is None or not tiling.enabled:
            return

        backbone_type = self.backbone_type
        backbone_cfg = self.config.model_config.backbone_config[f"{backbone_type}"]
        max_stride = int(backbone_cfg["max_stride"])
        output_stride = int(backbone_cfg["output_stride"])
        convs_per_block = int(backbone_cfg.get("convs_per_block", 2))
        kernel_size = int(backbone_cfg.get("kernel_size", 3))

        # Object extent + instance count across train labels.
        max_bbox_dim = 0.0
        n_instances = 0
        for train_label in self.train_labels:
            bbox = find_max_instance_bbox_size(train_label)
            if bbox > max_bbox_dim:
                max_bbox_dim = bbox
            n_instances += sum(len(lf.instances) for lf in train_label)

        try:
            backbone_margin = compute_backbone_context_margin(
                backbone_type, max_stride, convs_per_block, kernel_size
            )
        except ValueError:
            # Unsupported backbone under tiling; check_tiling emits the hard error.
            backbone_margin = 0

        # tile_size (preserve explicit).
        if tiling.tile_size is None:
            tile_size = compute_suggested_tile_size(
                max_bbox_dim, max_stride, output_stride, backbone_margin
            )
            self.config.data_config.preprocessing.tiling.tile_size = tile_size
            logger.info(
                f"Auto-sized tiling.tile_size={tile_size} "
                f"(max_bbox_dim={max_bbox_dim:.1f}, backbone_margin={backbone_margin})."
            )
        tile_size = int(self.config.data_config.preprocessing.tiling.tile_size)

        # overlap (preserve explicit; conservative + warn when labels sparse).
        if tiling.overlap is None:
            if n_instances < _SPARSE_LABEL_THRESHOLD:
                overlap = (
                    math.ceil(tiling.min_overlap_fraction * tile_size / output_stride)
                    * output_stride
                )
                logger.warning(
                    f"Only {n_instances} labeled instances "
                    f"(< {_SPARSE_LABEL_THRESHOLD}); the object-size estimate is "
                    f"unreliable. Using a conservative tiling.overlap={overlap} "
                    f"({tiling.min_overlap_fraction:.0%} of tile_size). Set "
                    "data_config.preprocessing.tiling.overlap explicitly to override."
                )
            else:
                sigma = self._get_confmap_sigma(output_stride)
                overlap = compute_suggested_tile_overlap(
                    tile_size,
                    max_bbox_dim,
                    sigma,
                    output_stride,
                    backbone_margin,
                    tiling.min_overlap_fraction,
                )
                logger.info(
                    f"Auto-sized tiling.overlap={overlap} "
                    f"(confmap_sigma={sigma}, backbone_margin={backbone_margin})."
                )
            self.config.data_config.preprocessing.tiling.overlap = overlap

        overlap = int(self.config.data_config.preprocessing.tiling.overlap)

        # samples_per_frame default: a conservative grid-tile count for a
        # representative (max-sized) frame, so train coverage ~= one grid pass
        # per frame. Written back into the live config before dataset creation.
        if tiling.samples_per_frame is None:
            scale = float(self.config.data_config.preprocessing.scale)
            rep_h = self.config.data_config.preprocessing.max_height
            rep_w = self.config.data_config.preprocessing.max_width
            if rep_h is None or rep_w is None:
                # Fall back to the first labeled frame's native size.
                shape = getattr(self.train_labels[0].videos[0], "shape", None)
                if shape is not None and len(shape) >= 3:
                    rep_h, rep_w = int(shape[1]), int(shape[2])
                else:
                    img = self.train_labels[0][0].image
                    rep_h, rep_w = int(img.shape[0]), int(img.shape[1])
            sized_hw = (int(rep_h * scale), int(rep_w * scale))
            n_tiles = len(
                generate_tile_grid(
                    sized_hw,
                    tile_size=tile_size,
                    overlap=overlap,
                    output_stride=output_stride,
                    max_stride=max_stride,
                    min_overlap_fraction=float(tiling.min_overlap_fraction),
                )
            )
            self.config.data_config.preprocessing.tiling.samples_per_frame = max(
                1, n_tiles
            )
            logger.info(
                "Auto-sized tiling.samples_per_frame="
                f"{max(1, n_tiles)} (grid tiles for a "
                f"{sized_hw[0]}x{sized_hw[1]} frame)."
            )

    def _setup_head_config(self):
        """Setup node, edge and class names in head config."""
        # if edges and part names aren't set in head configs, get it from labels object.
        head_config = self.config.model_config.head_configs[self.model_type]
        # Skeleton-less model types carry no skeleton at all: mask-only labels
        # for a centroid model trained off `data_config.centroids_from_masks`,
        # the `embedding` model type, and the segmentation heads. None of their
        # heads declares part_names or edges, so an absent skeleton is correct --
        # but indexing `self.skeletons[0]` unconditionally raised
        # `IndexError: list index out of range` here before any guard below could
        # be reached, making those model types unconfigurable. Heads that DO need
        # a skeleton now say so by name instead.
        skeleton = self.skeletons[0] if self.skeletons else None
        skeleton_node_names = list(skeleton.node_names) if skeleton is not None else []
        for key in head_config:
            if "part_names" in head_config[key].keys():
                if head_config[key]["part_names"] is None:
                    if skeleton is None:
                        message = (
                            f"model_config.head_configs.{self.model_type}.{key}"
                            ".part_names is null and the labels carry no skeleton, "
                            "so the node names cannot be inferred. Provide labels "
                            "with a skeleton, or set part_names explicitly."
                        )
                        logger.error(message)
                        raise ValueError(message)
                    self.config.model_config.head_configs[self.model_type][key][
                        "part_names"
                    ] = skeleton.node_names
                elif list(head_config[key]["part_names"]) != skeleton_node_names:
                    # GT confidence-map generation always produces one channel
                    # per node in the skeleton (custom_datasets.py's
                    # generate_confmaps has no part_names/subset parameter), while
                    # the head's own output channel count is len(part_names). An
                    # explicit part_names that's shorter, longer, or reordered
                    # relative to the skeleton silently mismatches those two
                    # channel counts (a confusing tensor-shape error deep in the
                    # loss) or silently mislabels channels (if the same length
                    # but reordered). Catch it here, fail-fast, before any data
                    # loading/model construction.
                    message = (
                        f"model_config.head_configs.{self.model_type}.{key}"
                        f".part_names must exactly match the skeleton's node "
                        f"names (in order) -- partial/reordered subsets are not "
                        f"supported. Got {list(head_config[key]['part_names'])!r}, "
                        f"skeleton has {skeleton_node_names!r}. Set part_names to "
                        f"null to use the full skeleton automatically."
                    )
                    logger.error(message)
                    raise ValueError(message)

            if (
                "anchor_part" in head_config[key].keys()
                and head_config[key]["anchor_part"] is not None
                and self.model_type != "centroid"
            ):
                # `centroid`'s anchor_part deliberately falls back to None (mean
                # of visible nodes) when absent from the skeleton -- see the
                # comment at custom_datasets.py's centroid branch ("must NOT
                # crash"). Every other head type that consumes anchor_part
                # (centered_instance, multi_class_topdown,
                # centered_instance_segmentation) does `nodes.index(anchor_part)`
                # with no such guard, so a typo'd/nonexistent anchor_part
                # currently passes setup cleanly and only fails deep inside
                # dataset construction -- with an error message that
                # misleadingly blames `part_names`, not the actual offending
                # `anchor_part` field. Catch it here instead.
                if head_config[key]["anchor_part"] not in skeleton_node_names:
                    message = (
                        f"model_config.head_configs.{self.model_type}.{key}"
                        f".anchor_part {head_config[key]['anchor_part']!r} is not "
                        f"a node in the skeleton {skeleton_node_names!r}."
                    )
                    logger.error(message)
                    raise ValueError(message)

            if "edges" in head_config[key].keys():
                if head_config[key]["edges"] is None:
                    if skeleton is None:
                        message = (
                            f"model_config.head_configs.{self.model_type}.{key}"
                            ".edges is null and the labels carry no skeleton, so "
                            "the edges cannot be inferred. Provide labels with a "
                            "skeleton, or set edges explicitly."
                        )
                        logger.error(message)
                        raise ValueError(message)
                    edges = [
                        (x.source.name, x.destination.name) for x in skeleton.edges
                    ]
                    self.config.model_config.head_configs[self.model_type][key][
                        "edges"
                    ] = edges

            if "classes" in head_config[key].keys():
                if head_config[key]["classes"] is None:
                    tracks = []
                    for train_label in self.train_labels:
                        tracks.extend(
                            [x.name for x in train_label.tracks if x is not None]
                        )
                    classes = list(set(tracks))
                    if not len(classes):
                        message = (
                            f"No tracks found. ID models need tracks to be defined."
                        )
                        logger.error(message)
                        raise Exception(message)
                    self.config.model_config.head_configs[self.model_type][key][
                        "classes"
                    ] = classes

            # NOTE: no per-class identity UUID is minted. The simplified sleap-io
            # `Identity` (name + metadata, sleap-io #535) matches by NAME across files
            # and retrains, so the class name IS the canonical cross-file identity key;
            # the old train→inference uuid bridge is obsolete. A `class_output ==
            # "identity"` multi_class model emits `sio.Identity(name=<class name>)` at
            # inference (see `predictor._multiclass_identities`).

    def _setup_ckpt_path(self):
        """Setup checkpoint path."""
        # if run_name is None, assign a new dir name
        ckpt_dir = self.config.trainer_config.ckpt_dir
        if ckpt_dir is None or ckpt_dir == "" or ckpt_dir == "None":
            ckpt_dir = "."
            self.config.trainer_config.ckpt_dir = ckpt_dir
        run_name = self.config.trainer_config.run_name
        run_name_is_empty = run_name is None or run_name == "" or run_name == "None"

        # Validate: multi-GPU + disk cache requires explicit run_name
        if run_name_is_empty:
            is_disk_caching = (
                self.config.data_config.data_pipeline_fw
                == "torch_dataset_cache_img_disk"
            )
            num_devices = self._get_trainer_devices()

            if is_disk_caching and num_devices > 1:
                raise ValueError(
                    f"Multi-GPU training with disk caching requires an explicit `run_name`.\n\n"
                    f"Detected {num_devices} device(s) with "
                    f"`data_pipeline_fw='torch_dataset_cache_img_disk'`.\n"
                    f"Without an explicit run_name, each GPU worker generates a different "
                    f"timestamp-based directory, causing cache synchronization failures.\n\n"
                    f"Please provide a run_name using one of these methods:\n"
                    f"  - CLI: sleap-nn train config.yaml trainer_config.run_name=my_experiment\n"
                    f"  - Config file: Set `trainer_config.run_name: my_experiment`\n"
                    f"  - Python API: train(..., run_name='my_experiment')"
                )

            # Auto-generate timestamp-based run_name (safe for single GPU or non-disk-cache)
            sum_train_lfs = sum([len(train_label) for train_label in self.train_labels])
            sum_val_lfs = sum([len(val_label) for val_label in self.val_labels])
            run_name = (
                datetime.now().strftime("%y%m%d_%H%M%S")
                + f".{self.model_type}.n={sum_train_lfs + sum_val_lfs}"
            )

        # If checkpoint path already exists, add suffix to prevent overwriting
        if (Path(ckpt_dir) / run_name).exists() and (
            Path(ckpt_dir) / run_name / "best.ckpt"
        ).exists():
            logger.info(
                f"Checkpoint path already exists: {Path(ckpt_dir) / run_name}... adding suffix to prevent overwriting."
            )
            for i in count(1):
                new_run_name = f"{run_name}-{i}"
                if not (Path(ckpt_dir) / new_run_name).exists():
                    run_name = new_run_name
                    break

        self.config.trainer_config.run_name = run_name

        # set output dir for cache img
        if self.config.data_config.data_pipeline_fw == "torch_dataset_cache_img_disk":
            if self.config.data_config.cache_img_path is None:
                self.config.data_config.cache_img_path = (
                    Path(self.config.trainer_config.ckpt_dir)
                    / self.config.trainer_config.run_name
                )

    def _disable_pretrained_normalize_for_embedding(self):
        """Turn off the pretrained backbone's input normalization for `embedding`.

        `PretrainedBackbone` applies model-specific (ImageNet) mean/std inside its
        forward, and documents its input contract as ``[0, 1]`` -- which is what
        every other model type feeds it. The `embedding` model type breaks that
        contract on purpose: `EmbeddingLightningModule._build_input` per-crop
        standardizes to ~N(0, 1) before the backbone sees the crop, so applying the
        ImageNet shift on top of it hands the stem mean -1.99 / std 4.43 instead of
        the ~N(0, 1) it was pretrained on.

        This is silent -- it trains, it just trains badly. Measured on the gerbil
        re-ID set (DINOv2-with-registers, 3 epochs, seed 0, paired runs): val rank-1
        **0.363** with normalization left on vs **0.920** with it off.

        Rather than let the default quietly cost 2.5x rank-1, force it off and say
        so. Only the pretrained backbone has this knob, so this runs only on that
        branch of :meth:`_verify_model_input_channels`. Note this is
        ``backbone_config.pretrained.normalize`` (input normalization), NOT
        ``head_configs.embedding.embedding.normalize`` (L2-normalizing the output
        vector), which is unrelated and stays on.
        """
        pretrained_cfg = self.config.model_config.backbone_config.pretrained
        if not OmegaConf.select(pretrained_cfg, "normalize", default=True):
            return
        pretrained_cfg.normalize = False
        logger.info(
            "Disabling `model_config.backbone_config.pretrained.normalize` for the "
            "`embedding` model type: the embedding pipeline already per-crop "
            "standardizes to ~N(0, 1), so the backbone's ImageNet mean/std would "
            "double-normalize the input (measured cost: 2.5x val rank-1). Set it "
            "back to `true` only if you also disable the per-crop standardize."
        )

    def _verify_model_input_channels(self):
        """Verify input channels in model_config based on input image and pretrained model weights."""
        # check in channels, verify with img channels / ensure_rgb/ ensure_grayscale
        if self.train_labels[0] is not None:
            img_channels = self.train_labels[0][0].image.shape[-1]
            if self.config.data_config.preprocessing.ensure_rgb:
                img_channels = 3
            if self.config.data_config.preprocessing.ensure_grayscale:
                img_channels = 1
            model_in_channels = self.config.model_config.backbone_config[
                f"{self.backbone_type}"
            ].in_channels
            if model_in_channels != img_channels:
                target_format = "grayscale" if model_in_channels == 1 else "rgb"
                logger.warning(
                    f"Image has {img_channels} channel(s) but model has "
                    f"{model_in_channels} input channel(s). Images will be "
                    f"converted to {target_format} to fit the model architecture."
                )
                self.config.model_config.backbone_config[
                    f"{self.backbone_type}"
                ].in_channels = img_channels
                logger.info(
                    f"Updating backbone in_channels to {img_channels} based on the input image channels."
                )

        # verify input img channels with pretrained model ckpts (if any)
        if (
            self.backbone_type == "convnext" or self.backbone_type == "swint"
        ) and self.config.model_config.backbone_config[
            f"{self.backbone_type}"
        ].pre_trained_weights is not None:
            current_in_channels = self.config.model_config.backbone_config[
                f"{self.backbone_type}"
            ].in_channels
            if current_in_channels != 3:
                logger.warning(
                    f"Image has {current_in_channels} channel(s) but the "
                    f"pretrained {self.backbone_type} backbone requires 3 "
                    "(ImageNet RGB) input channels. Images will be converted "
                    "to rgb to fit the model architecture."
                )
                self.config.model_config.backbone_config[
                    f"{self.backbone_type}"
                ].in_channels = 3
                # The ImageNet stem needs a 3-channel input, but the `embedding`
                # model type keeps its (default grayscale) DATA channels: a 1-channel
                # crop is repeated to 3 in `Model.forward`. Flipping ensure_rgb here
                # would silently turn an explicitly-grayscale re-ID model into an RGB
                # one. For every other model type, sync the data to RGB as before.
                if self.model_type != "embedding":
                    self.config.data_config.preprocessing.ensure_rgb = True
                    self.config.data_config.preprocessing.ensure_grayscale = False
                logger.info(
                    f"Updating backbone in_channels to 3 based on the pretrained model weights."
                )

        # External pretrained (HuggingFace) backbones have 3-channel stems and
        # expect RGB (grayscale is replicated). Force in_channels=3 + ensure_rgb so
        # the data pipeline feeds 3-channel images and the pretrained stem loads.
        elif self.backbone_type == "pretrained":
            if self.config.model_config.backbone_config.pretrained.in_channels != 3:
                self.config.model_config.backbone_config.pretrained.in_channels = 3
                logger.info("Updating pretrained backbone in_channels to 3 (RGB stem).")
            if self.model_type == "embedding":
                # As in the two branches above, the `embedding` model type keeps its
                # own DATA channels: a 1-channel crop is repeated to 3 in
                # `Model.forward`, so the stem is fed correctly either way, and
                # flipping `ensure_rgb` here would silently turn an
                # explicitly-grayscale re-ID model into an RGB one.
                self._disable_pretrained_normalize_for_embedding()
            else:
                self.config.data_config.preprocessing.ensure_rgb = True
                self.config.data_config.preprocessing.ensure_grayscale = False

        elif (
            self.backbone_type == "unet"
            and self.config.model_config.pretrained_backbone_weights is not None
        ):
            if self.config.model_config.pretrained_backbone_weights.endswith(".ckpt"):
                pretrained_backbone_ckpt = torch.load(
                    self.config.model_config.pretrained_backbone_weights,
                    map_location="cpu",  # this will be loaded on cpu as it's just used to get the input channels
                    weights_only=False,
                )
                input_channels = list(pretrained_backbone_ckpt["state_dict"].values())[
                    0
                ].shape[
                    -3
                ]  # get input channels from first layer
                if (
                    self.config.model_config.backbone_config.unet.in_channels
                    != input_channels
                ):
                    self.config.model_config.backbone_config.unet.in_channels = (
                        input_channels
                    )
                    logger.info(
                        f"Updating backbone in_channels to {input_channels} based on the pretrained model weights."
                    )

                    if input_channels == 1:
                        self.config.data_config.preprocessing.ensure_grayscale = True
                        self.config.data_config.preprocessing.ensure_rgb = False
                        logger.info(
                            f"Updating data preprocessing to ensure_grayscale to True based on the pretrained model weights."
                        )
                    elif input_channels == 3 and self.model_type != "embedding":
                        # Embedding keeps its (default grayscale) DATA channels even with a
                        # 3-channel pretrained stem (gray is repeated in Model.forward).
                        self.config.data_config.preprocessing.ensure_rgb = True
                        self.config.data_config.preprocessing.ensure_grayscale = False
                        logger.info(
                            f"Updating data preprocessing to ensure_rgb to True based on the pretrained model weights."
                        )

            elif self.config.model_config.pretrained_backbone_weights.endswith(".h5"):
                input_channels = get_keras_first_layer_channels(
                    self.config.model_config.pretrained_backbone_weights
                )
                if (
                    self.config.model_config.backbone_config.unet.in_channels
                    != input_channels
                ):
                    self.config.model_config.backbone_config.unet.in_channels = (
                        input_channels
                    )
                    logger.info(
                        f"Updating backbone in_channels to {input_channels} based on the pretrained model weights."
                    )

                    if input_channels == 1:
                        self.config.data_config.preprocessing.ensure_grayscale = True
                        self.config.data_config.preprocessing.ensure_rgb = False
                        logger.info(
                            f"Updating data preprocessing to ensure_grayscale to True based on the pretrained model weights."
                        )
                    elif input_channels == 3 and self.model_type != "embedding":
                        # Embedding keeps grayscale data with a 3-channel pretrained stem.
                        self.config.data_config.preprocessing.ensure_rgb = True
                        self.config.data_config.preprocessing.ensure_grayscale = False
                        logger.info(
                            f"Updating data preprocessing to ensure_rgb to True based on the pretrained model weights."
                        )

    def _verify_accelerator_config(self):
        """Verify the configured `trainer_accelerator` is available on this machine.

        A saved training config may have been created on a different machine (e.g.
        `trainer_accelerator: mps` from a Mac, reloaded on a Linux/CUDA box). Passing an
        unavailable accelerator straight to `lightning.Trainer` raises deep inside
        `train()`, after dataset/ckpt setup has already run. This check runs early
        (from `setup_config()`) and falls back to `"auto"` on a mismatch instead,
        letting the existing device-count resolution logic pick the right backend.
        """
        accelerator = self.config.trainer_config.trainer_accelerator

        if accelerator in ("auto", "cpu"):
            return

        if accelerator in ("gpu", "cuda"):
            available = torch.cuda.is_available()
        elif accelerator == "mps":
            available = (
                hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
            )
        else:
            logger.info(
                f"Configured accelerator '{accelerator}' is not a recognized option; "
                "changing it to 'auto' (Lightning will select the best available device)."
            )
            self.config.trainer_config.trainer_accelerator = "auto"
            return

        if available:
            logger.info(
                f"Configured accelerator '{accelerator}' is available on this machine "
                "and will be used."
            )
        else:
            logger.info(
                f"Configured accelerator '{accelerator}' is not available on this "
                "machine; changing it to 'auto' (Lightning will select the best "
                "available device)."
            )
            self.config.trainer_config.trainer_accelerator = "auto"

    def setup_config(self):
        """Compute config parameters."""
        logger.info("Setting up config...")

        # Normalize empty strings to None for optional wandb fields
        if self.config.trainer_config.wandb.prv_runid == "":
            self.config.trainer_config.wandb.prv_runid = None

        # compute preprocessing parameters from the labels objects and fill in the config
        self._setup_preprocessing_config()

        # save skeleton to config
        skeleton_yaml = yaml.safe_load(SkeletonYAMLEncoder().encode(self.skeletons))
        skeleton_names = skeleton_yaml.keys()
        self.config["data_config"]["skeletons"] = []
        for skeleton_name in skeleton_names:
            skl = skeleton_yaml[skeleton_name]
            skl["name"] = skeleton_name
            self.config["data_config"]["skeletons"].append(skl)

        # setup head config - partnames, edges and class names
        self._setup_head_config()

        # Identity-equality gates (SPEC §4.4): the embedding objective's pos/neg
        # sources silently assert "same / different animal"; validate them against the
        # declared `data_config.identity` semantics (error for global_id without global
        # names; warn for unproofread tracklets / non-deduplicated same-frame negatives).
        if self.model_type == "embedding":
            # Real `sio.Identity` annotations ground `global_id` grouping directly (no
            # `track_names_are_global` promise needed).
            has_identities = any(
                bool(getattr(label, "identities", None) or [])
                for label in self.train_labels
            )
            validate_embedding_identity(
                objective=OmegaConf.select(
                    self.config,
                    "model_config.head_configs.embedding.embedding.objective",
                    default=None,
                ),
                identity=OmegaConf.select(
                    self.config, "data_config.identity", default=None
                ),
                has_identities=has_identities,
            )

        # set max stride for the backbone: convnext and swint
        if self.backbone_type == "convnext":
            self.config.model_config.backbone_config.convnext.max_stride = (
                self.config.model_config.backbone_config.convnext.stem_patch_stride
                * (2**3)
                * 2
            )
        elif self.backbone_type == "swint":
            self.config.model_config.backbone_config.swint.max_stride = (
                self.config.model_config.backbone_config.swint.stem_patch_stride
                * (2**3)
                * 2
            )

        # set output stride for backbone from head config and verify max stride
        self.config = check_output_strides(self.config)

        # fail fast on a contradictory / unknown centroid method (#586)
        self.config = check_centroid_methods(self.config)

        # auto-size + validate tiling geometry (no-op unless tiling.enabled)
        self._setup_tiling_config()
        self.config = check_tiling(self.config)

        # verify the configured accelerator is available on this machine
        self._verify_accelerator_config()

        # if trainer_devices is None, set it to "auto"
        if self.config.trainer_config.trainer_devices is None:
            self.config.trainer_config.trainer_devices = (
                "auto"
                if OmegaConf.select(
                    self.config, "trainer_config.trainer_device_indices", default=None
                )
                is None
                else len(
                    OmegaConf.select(
                        self.config,
                        "trainer_config.trainer_device_indices",
                        default=None,
                    )
                )
            )

        # setup checkpoint path (generates run_name if not specified)
        self._setup_ckpt_path()

        # Default wandb run name to trainer run_name if not specified
        # Note: This must come after _setup_ckpt_path() which generates run_name
        if self.config.trainer_config.wandb.name is None:
            self.config.trainer_config.wandb.name = self.config.trainer_config.run_name

        # verify input_channels in model_config based on input image and pretrained model weights
        self._verify_model_input_channels()

    def _setup_model_ckpt_dir(self):
        """Create the model ckpt folder and save ground truth labels."""
        ckpt_path = (
            Path(self.config.trainer_config.ckpt_dir)
            / self.config.trainer_config.run_name
        ).as_posix()
        logger.info(f"Setting up model ckpt dir: `{ckpt_path}`...")

        # Only rank 0 (or non-distributed) should create directories and save files
        if RANK in [0, -1]:
            if not Path(ckpt_path).exists():
                try:
                    Path(ckpt_path).mkdir(parents=True, exist_ok=True)
                except OSError as e:
                    message = f"Cannot create a new folder in {ckpt_path}.\n {e}"
                    logger.error(message)
                    raise OSError(message)
            # Check if we should filter to user-labeled frames only
            user_instances_only = OmegaConf.select(
                self.config, "data_config.user_instances_only", default=True
            )

            # Save train and val ground truth labels
            for idx, (train, val) in enumerate(zip(self.train_labels, self.val_labels)):
                # Filter to user-labeled frames if needed (for evaluation)
                if user_instances_only:
                    train_filtered = self._filter_to_user_labeled(train)
                    val_filtered = self._filter_to_user_labeled(val)
                else:
                    train_filtered = train
                    val_filtered = val

                train_filtered.save(
                    Path(ckpt_path) / f"labels_gt.train.{idx}.slp",
                    restore_original_videos=False,
                )
                val_filtered.save(
                    Path(ckpt_path) / f"labels_gt.val.{idx}.slp",
                    restore_original_videos=False,
                )

            # Save test ground truth labels if test paths are provided
            test_file_path = OmegaConf.select(
                self.config, "data_config.test_file_path", default=None
            )
            if test_file_path is not None:
                # Normalize to list of strings
                if isinstance(test_file_path, str):
                    test_paths = [test_file_path]
                else:
                    test_paths = list(test_file_path)

                for idx, test_path in enumerate(test_paths):
                    # Only save if it's a .slp file (not a video file)
                    if test_path.endswith(".slp") or test_path.endswith(".pkg.slp"):
                        try:
                            test_labels = sio.load_slp(test_path)
                            if user_instances_only:
                                test_filtered = self._filter_to_user_labeled(
                                    test_labels
                                )
                            else:
                                test_filtered = test_labels
                            test_filtered.save(
                                Path(ckpt_path) / f"labels_gt.test.{idx}.slp",
                                restore_original_videos=False,
                            )
                        except Exception as e:
                            logger.warning(
                                f"Could not save test ground truth for {test_path}: {e}"
                            )

    def _setup_viz_datasets(self):
        """Setup dataloaders."""
        data_viz_config = self.config.copy()
        data_viz_config.data_config.data_pipeline_fw = "torch_dataset"

        return get_train_val_datasets(
            train_labels=self.train_labels,
            val_labels=self.val_labels,
            config=data_viz_config,
            rank=-1,
        )

    def _setup_datasets(self):
        """Setup dataloaders."""
        base_cache_img_path = None
        if self.config.data_config.data_pipeline_fw == "torch_dataset_cache_img_memory":
            # check available memory. If insufficient memory, default to disk caching.
            # Account for DataLoader worker memory overhead
            train_num_workers = self.config.trainer_config.train_data_loader.num_workers
            val_num_workers = self.config.trainer_config.val_data_loader.num_workers
            max_num_workers = max(train_num_workers, val_num_workers)

            mem_available = check_cache_memory(
                self.train_labels,
                self.val_labels,
                memory_buffer=MEMORY_BUFFER,
                num_workers=max_num_workers,
            )
            if not mem_available and self.model_type == "embedding":
                # The embedding (appearance / re-ID) model must NOT use the lossy JPEG
                # disk cache (it silently degrades the model — see the disk-cache gate
                # in get_train_val_datasets). Fall back to an uncached, lossless
                # pipeline instead of disk.
                self.config.data_config.data_pipeline_fw = "torch_dataset"
                base_cache_img_path = None
                logger.info(
                    "Insufficient memory for in-memory caching of the `embedding` "
                    "model; falling back to an uncached pipeline (`torch_dataset`) to "
                    "avoid the lossy JPEG disk cache."
                )
            elif not mem_available:
                # Validate: multi-GPU + auto-generated run_name + fallback to disk cache
                original_run_name = self._initial_config.trainer_config.run_name
                run_name_was_auto = (
                    original_run_name is None
                    or original_run_name == ""
                    or original_run_name == "None"
                )
                if run_name_was_auto and self.trainer.num_devices > 1:
                    raise ValueError(
                        f"Memory caching failed and disk caching fallback requires an "
                        f"explicit `run_name` for multi-GPU training.\n\n"
                        f"Detected {self.trainer.num_devices} device(s) with insufficient "
                        f"memory for in-memory caching.\n"
                        f"Without an explicit run_name, each GPU worker generates a different "
                        f"timestamp-based directory, causing cache synchronization failures.\n\n"
                        f"Please provide a run_name using one of these methods:\n"
                        f"  - CLI: sleap-nn train config.yaml trainer_config.run_name=my_experiment\n"
                        f"  - Config file: Set `trainer_config.run_name: my_experiment`\n"
                        f"  - Python API: train(..., run_name='my_experiment')\n\n"
                        f"Alternatively, use `data_pipeline_fw='torch_dataset'` to disable caching."
                    )

                self.config.data_config.data_pipeline_fw = (
                    "torch_dataset_cache_img_disk"
                )
                base_cache_img_path = (
                    Path(self.config.data_config.cache_img_path)
                    if self.config.data_config.cache_img_path is not None
                    else Path(self.config.trainer_config.ckpt_dir)
                    / self.config.trainer_config.run_name
                )
                logger.info(
                    f"Insufficient memory for in-memory caching. `jpg` files will be created for disk-caching."
                )
            self.config.data_config.cache_img_path = base_cache_img_path

        elif self.config.data_config.data_pipeline_fw == "torch_dataset_cache_img_disk":
            # Get cache img path
            base_cache_img_path = (
                Path(self.config.data_config.cache_img_path)
                if self.config.data_config.cache_img_path is not None
                else Path(self.config.trainer_config.ckpt_dir)
                / self.config.trainer_config.run_name
            )

            if self.config.data_config.cache_img_path is None:
                self.config.data_config.cache_img_path = base_cache_img_path

        return get_train_val_datasets(
            train_labels=self.train_labels,
            val_labels=self.val_labels,
            config=self.config,
            rank=self.trainer.global_rank,
        )

    def _setup_loggers_callbacks(self, viz_train_dataset, viz_val_dataset):
        """Create loggers and callbacks."""
        logger.info("Setting up callbacks and loggers...")
        loggers = []
        callbacks = []
        # Checkpoint/early-stop SELECTION metric. Contrastive (embedding) objectives
        # are NOT well-selected by val/loss, so select on a retrieval metric (SPEC §8).
        if self.model_type == "embedding":
            # Touch the objective/sampler so a missing config fails loud here rather
            # than deep in the dataloader.
            OmegaConf.select(
                self.config,
                "model_config.head_configs.embedding.embedding.objective.sampler.kind",
                default=None,
            )
            emb_select = OmegaConf.select(
                self.config, "trainer_config.eval.select_metric", default="rank1"
            )
            # Per-metric selection mode: retrieval / verification-AUC / kNN-accuracy are
            # higher-better; EER is lower-better. Validate the name so a typo cannot
            # silently monitor a key that is never logged (ModelCheckpoint would then
            # never save "best" and a strict EarlyStopping would raise).
            emb_metric_modes = {
                "rank1": "max",
                "mAP": "max",
                "auc": "max",
                "knn_acc": "max",
                "eer": "min",
            }
            if emb_select not in emb_metric_modes:
                raise ValueError(
                    f"trainer_config.eval.select_metric='{emb_select}' is not a valid "
                    f"embedding selection metric; choose one of "
                    f"{'|'.join(emb_metric_modes)}."
                )
            ckpt_monitor = f"eval/val/{emb_select}"
            ckpt_mode = emb_metric_modes[emb_select]
            # The selection metric only exists on eval epochs, so let the
            # checkpointer consider "best" only on those epochs. Same formula the
            # callback uses ((epoch + 1) % frequency), so they cannot drift, and it
            # keeps a saved "best" tied to the epoch its metric was measured on
            # instead of a later epoch carrying a stale value forward.
            ckpt_every_n_epochs = int(
                OmegaConf.select(
                    self.config, "trainer_config.eval.frequency", default=1
                )
            )
        else:
            ckpt_monitor, ckpt_mode = "val/loss", "min"
            ckpt_every_n_epochs = 1

        if self.config.trainer_config.save_ckpt:
            # checkpoint callback
            checkpoint_callback = ModelCheckpoint(
                save_top_k=self.config.trainer_config.model_ckpt.save_top_k,
                save_last=self.config.trainer_config.model_ckpt.save_last,
                dirpath=(
                    Path(self.config.trainer_config.ckpt_dir)
                    / self.config.trainer_config.run_name
                ).as_posix(),
                filename="best",
                # Config-driven monitor/mode (main #690/#692) wins when the user set
                # it away from the "val/loss" schema default (e.g. a seg quality
                # metric). Left at the default, fall back to the model-type-aware
                # default computed above (embedding -> retrieval metric, else val/loss)
                # since contrastive objectives are not well-selected by val/loss.
                monitor=(
                    self.config.trainer_config.model_ckpt.monitor
                    if self.config.trainer_config.model_ckpt.monitor != "val/loss"
                    else ckpt_monitor
                ),
                mode=(
                    self.config.trainer_config.model_ckpt.mode
                    if self.config.trainer_config.model_ckpt.monitor != "val/loss"
                    else ckpt_mode
                ),
                # Only meaningful (and only non-1) when the monitored metric is the
                # eval-gated embedding one; an explicitly configured monitor keeps
                # the every-epoch default.
                every_n_epochs=(
                    1
                    if self.config.trainer_config.model_ckpt.monitor != "val/loss"
                    else ckpt_every_n_epochs
                ),
            )
            callbacks.append(checkpoint_callback)

            # csv log callback
            csv_log_keys = [
                "epoch",
                "train/loss",
                "val/loss",
                "learning_rate",
                "train/time",
                "val/time",
            ]
            # Negative-frame split metrics (train + val), only for model types
            # that support frame-level negatives and only when the feature is
            # enabled. Gating avoids empty columns in the common no-negatives case.
            use_negative_frames = OmegaConf.select(
                self.config, "data_config.use_negative_frames", default=False
            )
            if use_negative_frames and self.model_type in [
                "single_instance",
                "centroid",
                "bottomup",
                "multi_class_bottomup",
            ]:
                # Aggregate split metrics (all four supported model types).
                csv_log_keys.extend(
                    [
                        "train/n_positive",
                        "train/n_negative",
                        "train/loss_positive",
                        "train/loss_negative",
                        "train/loss_positive_unweighted",
                        "train/loss_negative_unweighted",
                        "val/n_positive",
                        "val/n_negative",
                        "val/loss_positive",
                        "val/loss_negative",
                        "val/loss_positive_unweighted",
                        "val/loss_negative_unweighted",
                    ]
                )
                # Per-head split metrics (two-head models only).
                if self.model_type == "bottomup":
                    csv_log_keys.extend(
                        [
                            "train/confmaps_loss_positive",
                            "train/confmaps_loss_negative",
                            "train/paf_loss_positive",
                            "train/paf_loss_negative",
                            "val/confmaps_loss_positive",
                            "val/confmaps_loss_negative",
                            "val/paf_loss_positive",
                            "val/paf_loss_negative",
                        ]
                    )
                elif self.model_type == "multi_class_bottomup":
                    csv_log_keys.extend(
                        [
                            "train/confmaps_loss_positive",
                            "train/confmaps_loss_negative",
                            "train/classmap_loss_positive",
                            "train/classmap_loss_negative",
                            "val/confmaps_loss_positive",
                            "val/confmaps_loss_negative",
                            "val/classmap_loss_positive",
                            "val/classmap_loss_negative",
                        ]
                    )
            # Add model-specific keys for wandb parity
            if self.model_type in [
                "single_instance",
                "centered_instance",
                "multi_class_topdown",
            ]:
                csv_log_keys.extend(
                    [f"train/confmaps/{name}" for name in self.skeletons[0].node_names]
                )
            if self.model_type == "centroid":
                # Foreground/background confmap MSE split -- see
                # `LightningModel._log_confmap_fg_bg_loss`. Always computed for
                # this model type (not gated on `centroid_focal_loss_alpha`),
                # since it's also the diagnostic that motivates whether a
                # focal-style loss would help.
                csv_log_keys.extend(
                    [
                        "train/confmap_loss_fg",
                        "train/confmap_loss_bg",
                        "train/confmap_fg_frac",
                        "val/confmap_loss_fg",
                        "val/confmap_loss_bg",
                        "val/confmap_fg_frac",
                    ]
                )
            if self.model_type == "bottomup":
                csv_log_keys.extend(
                    [
                        "train/confmaps_loss",
                        "train/paf_loss",
                        "val/confmaps_loss",
                        "val/paf_loss",
                    ]
                )
            if self.model_type == "multi_class_bottomup":
                csv_log_keys.extend(
                    [
                        "train/confmaps_loss",
                        "train/classmap_loss",
                        "train/class_accuracy",
                        "val/confmaps_loss",
                        "val/classmap_loss",
                        "val/class_accuracy",
                    ]
                )
            if self.model_type == "multi_class_topdown":
                csv_log_keys.extend(
                    [
                        "train/confmaps_loss",
                        "train/classvector_loss",
                        "train/class_accuracy",
                        "val/confmaps_loss",
                        "val/classvector_loss",
                        "val/class_accuracy",
                    ]
                )
            if self.model_type == "bottomup_segmentation":
                csv_log_keys.extend(
                    [
                        "train/fg_loss",
                        "train/center_loss",
                        "train/offset_loss",
                        "val/fg_loss",
                        "val/center_loss",
                        "val/offset_loss",
                        "val/fg_iou",
                    ]
                )
            if self.model_type == "centered_instance_segmentation":
                csv_log_keys.extend(
                    [
                        "train/fg_loss",
                        "val/fg_loss",
                        "val/fg_iou",
                    ]
                )
            if self.model_type == "semantic_segmentation":
                csv_log_keys.extend(
                    [
                        "train/fg_loss",
                        "val/fg_loss",
                        "val/fg_iou",
                    ]
                )
            # The embedding model's retrieval callback is NOT gated on eval.enabled
            # (ModelCheckpoint selects on its metric, so it always runs -- see the
            # callback branching below). Its keys therefore must NOT sit inside the
            # eval.enabled gate, or the columns would vanish while the callback kept
            # logging them. `train/pos_per_anchor` is a per-epoch training metric.
            if self.model_type == "embedding":
                csv_log_keys.extend(
                    [
                        "train/pos_per_anchor",
                        "eval/val/rank1",
                        "eval/val/mAP",
                        "eval/val/auc",
                        "eval/val/eer",
                        "eval/val/knn_acc",
                    ]
                )
            # Eval-callback keys (only when trainer_config.eval.enabled, mirroring
            # the callback branching below). These are only computed every
            # eval.frequency epochs; CSVLoggerCallback NaN-resets them at the
            # start of each validation epoch so non-eval epochs show NaN instead
            # of silently repeating the last-computed eval value.
            elif self.config.trainer_config.eval.enabled:
                if self.model_type == "centroid":
                    csv_log_keys.extend(
                        [
                            "eval/val/centroid_dist_avg",
                            "eval/val/centroid_dist_median",
                            "eval/val/centroid_dist_p90",
                            "eval/val/centroid_dist_p95",
                            "eval/val/centroid_dist_max",
                            "eval/val/centroid_precision",
                            "eval/val/centroid_recall",
                            "eval/val/centroid_f1",
                            "eval/val/centroid_n_tp",
                            "eval/val/centroid_n_fp",
                            "eval/val/centroid_n_fn",
                        ]
                    )
                elif self.model_type == "semantic_segmentation":
                    csv_log_keys.extend(
                        [
                            "eval/val/fg_mean_iou",
                            "eval/val/fg_mean_cldice",
                            "eval/val/fg_mean_boundary_iou",
                            "eval/val/fg_frame_recall",
                        ]
                    )
                elif self.model_type in (
                    "bottomup_segmentation",
                    "centered_instance_segmentation",
                ):
                    csv_log_keys.extend(
                        [
                            "eval/val/mask_mean_iou",
                            "eval/val/mask_mean_iou_all_gt",
                            "eval/val/mask_mean_cldice",
                            "eval/val/mask_precision",
                            "eval/val/mask_recall",
                            "eval/val/mask_f1",
                            "eval/val/mask_n_tp",
                            "eval/val/mask_n_fp",
                            "eval/val/mask_n_fn",
                        ]
                    )
                else:
                    csv_log_keys.extend(
                        [
                            "eval/val/mOKS",
                            "eval/val/oks_voc_mAP",
                            "eval/val/oks_voc_mAR",
                            "eval/val/distance/avg",
                            "eval/val/distance/p50",
                            "eval/val/distance/p95",
                            "eval/val/distance/p99",
                            "eval/val/mPCK",
                            "eval/val/PCK_5",
                            "eval/val/PCK_10",
                            "eval/val/visibility_precision",
                            "eval/val/visibility_recall",
                        ]
                    )
            csv_logger = CSVLoggerCallback(
                filepath=Path(self.config.trainer_config.ckpt_dir)
                / self.config.trainer_config.run_name
                / "training_log.csv",
                keys=csv_log_keys,
            )
            callbacks.append(csv_logger)

        if self.config.trainer_config.early_stopping.stop_training_on_plateau:
            # early stopping callback
            es_kwargs = dict(
                monitor=ckpt_monitor,
                mode=ckpt_mode,
                verbose=False,
                min_delta=self.config.trainer_config.early_stopping.min_delta,
                patience=self.config.trainer_config.early_stopping.patience,
            )
            if self.model_type == "embedding":
                # The retrieval selection metric is logged only on eval epochs (so it is
                # absent when eval.frequency > 1) and the verification metrics (auc/eer)
                # can be NaN on a degenerate val set/shard. Don't let an absent/NaN value
                # abort the run — pose/seg models keep the strict defaults.
                es_kwargs["strict"] = False
                es_kwargs["check_finite"] = False
            callbacks.append(EarlyStopping(**es_kwargs))

        if self.config.trainer_config.use_wandb:
            # wandb logger
            wandb_config = self.config.trainer_config.wandb
            if wandb_config.wandb_mode == "offline":
                os.environ["WANDB_MODE"] = "offline"
            else:
                if RANK in [0, -1]:
                    wandb.login(key=self.config.trainer_config.wandb.api_key)
            wandb_logger = WandbLogger(
                entity=wandb_config.entity,
                project=wandb_config.project,
                name=wandb_config.name,
                save_dir=(
                    Path(self.config.trainer_config.ckpt_dir)
                    / self.config.trainer_config.run_name
                ).as_posix(),
                id=self.config.trainer_config.wandb.prv_runid,
                group=self.config.trainer_config.wandb.group,
            )
            loggers.append(wandb_logger)

            # Log message about wandb local logs cleanup
            should_delete_wandb_logs = wandb_config.delete_local_logs is True or (
                wandb_config.delete_local_logs is None
                and wandb_config.wandb_mode != "offline"
            )
            if should_delete_wandb_logs:
                logger.info(
                    "WandB local logs will be deleted after training completes. "
                    "To keep logs, set trainer_config.wandb.delete_local_logs=false"
                )

            # save the configs as yaml in the checkpoint dir
            # Mask API key in both configs to prevent saving to disk
            self.config.trainer_config.wandb.api_key = ""
            if self._initial_config is not None:
                self._initial_config.trainer_config.wandb.api_key = ""

        # zmq callbacks
        if self.config.trainer_config.zmq.controller_port is not None:
            controller_address = "tcp://127.0.0.1:" + str(
                self.config.trainer_config.zmq.controller_port
            )
            callbacks.append(TrainingControllerZMQ(address=controller_address))
        if self.config.trainer_config.zmq.publish_port is not None:
            publish_address = "tcp://127.0.0.1:" + str(
                self.config.trainer_config.zmq.publish_port
            )
            callbacks.append(ProgressReporterZMQ(address=publish_address))

        # viz callbacks - use unified callback for all visualization outputs
        if self.config.trainer_config.visualize_preds_during_training:
            viz_dir = (
                Path(self.config.trainer_config.ckpt_dir)
                / self.config.trainer_config.run_name
                / "viz"
            )
            if not Path(viz_dir).exists():
                if RANK in [0, -1]:
                    Path(viz_dir).mkdir(parents=True, exist_ok=True)

            # Get wandb viz config options
            log_wandb = self.config.trainer_config.use_wandb and OmegaConf.select(
                self.config, "trainer_config.wandb.save_viz_imgs_wandb", default=False
            )
            wandb_modes = []
            if log_wandb:
                if OmegaConf.select(
                    self.config, "trainer_config.wandb.viz_enabled", default=True
                ):
                    wandb_modes.append("direct")
                if OmegaConf.select(
                    self.config, "trainer_config.wandb.viz_boxes", default=False
                ):
                    wandb_modes.append("boxes")
                if OmegaConf.select(
                    self.config, "trainer_config.wandb.viz_masks", default=False
                ):
                    wandb_modes.append("masks")

            # Single unified callback handles all visualization outputs
            callbacks.append(
                UnifiedVizCallback(
                    model_trainer=self,
                    train_dataset=viz_train_dataset,
                    val_dataset=viz_val_dataset,
                    model_type=self.model_type,
                    save_local=self.config.trainer_config.save_ckpt,
                    local_save_dir=viz_dir,
                    log_wandb=log_wandb,
                    wandb_modes=wandb_modes if wandb_modes else ["direct"],
                    wandb_box_size=OmegaConf.select(
                        self.config, "trainer_config.wandb.viz_box_size", default=5.0
                    ),
                    wandb_confmap_threshold=OmegaConf.select(
                        self.config,
                        "trainer_config.wandb.viz_confmap_threshold",
                        default=0.1,
                    ),
                    log_wandb_table=OmegaConf.select(
                        self.config, "trainer_config.wandb.log_viz_table", default=False
                    ),
                    img_format=OmegaConf.select(
                        self.config,
                        "trainer_config.viz_img_format",
                        default="png",
                    ),
                )
            )

        # Add custom progress bar with better metric formatting
        if self.config.trainer_config.enable_progress_bar:
            callbacks.append(SleapProgressBar())

        # Add epoch-end evaluation callback if enabled
        # Embedding models are selected on a retrieval metric, so the retrieval
        # callback is REQUIRED (it logs eval/val/<metric> that ModelCheckpoint reads),
        # not gated on eval.enabled. Embedding models are skeleton-less so they must
        # never reach the keypoint EpochEndEvaluationCallback below.
        if self.model_type == "embedding":
            emb_freq = (
                OmegaConf.select(
                    self.config, "trainer_config.eval.frequency", default=1
                )
                or 1
            )
            emb_select = OmegaConf.select(
                self.config, "trainer_config.eval.select_metric", default="rank1"
            )
            callbacks.append(
                EmbeddingEvaluationCallback(
                    eval_frequency=emb_freq, select_metric=emb_select
                )
            )
        elif self.config.trainer_config.eval.enabled:
            if self.model_type == "centroid":
                # Use centroid-specific evaluation with distance-based metrics
                callbacks.append(
                    CentroidEvaluationCallback(
                        videos=self.val_labels[0].videos,
                        eval_frequency=self.config.trainer_config.eval.frequency,
                        match_threshold=self.config.trainer_config.eval.match_threshold,
                    )
                )
            elif self.model_type in (
                "bottomup_segmentation",
                "centered_instance_segmentation",
                "semantic_segmentation",
            ):
                # Segmentation has no keypoint predictions for OKS/PCK. Instance seg
                # (bottomup / centered-instance) recovers per-instance masks on the
                # shared preprocessed grid and reports instance-level mask-IoU mAP /
                # precision / recall (grouping-sensitive). Semantic (whole-frame
                # foreground) has no instances, so it runs the matching-free
                # foreground IoU / clDice / boundary-IoU variant. Both complement the
                # coarse val/fg_iou logged in validation_step.
                callbacks.append(
                    SegmentationEvaluationCallback(
                        eval_frequency=self.config.trainer_config.eval.frequency,
                        match_threshold=self.config.trainer_config.eval.match_threshold,
                        foreground=(self.model_type == "semantic_segmentation"),
                    )
                )
            else:
                # Use standard OKS/PCK evaluation for pose models
                callbacks.append(
                    EpochEndEvaluationCallback(
                        skeleton=self.skeletons[0],
                        videos=self.val_labels[0].videos,
                        eval_frequency=self.config.trainer_config.eval.frequency,
                        oks_stddev=self.config.trainer_config.eval.oks_stddev,
                        oks_scale=self.config.trainer_config.eval.oks_scale,
                    )
                )

        # Sync the tiling sampler + shared epoch tensor each epoch (tiling only).
        tiling = OmegaConf.select(
            self.config, "data_config.preprocessing.tiling", default=None
        )
        if tiling is not None and tiling.enabled:
            callbacks.append(TilingEpochCallback())

        return loggers, callbacks

    def _delete_cache_imgs(self):
        """Delete cache images in disk."""
        base_cache_img_path = Path(self.config.data_config.cache_img_path)
        train_cache_img_path = Path(base_cache_img_path) / "train_imgs"
        val_cache_img_path = Path(base_cache_img_path) / "val_imgs"

        if (train_cache_img_path).exists():
            logger.info(f"Deleting cache imgs from `{train_cache_img_path}`...")
            shutil.rmtree(
                (train_cache_img_path).as_posix(),
                ignore_errors=True,
            )

        if (val_cache_img_path).exists():
            logger.info(f"Deleting cache imgs from `{val_cache_img_path}`...")
            shutil.rmtree(
                (val_cache_img_path).as_posix(),
                ignore_errors=True,
            )

    def train(self):
        """Train the lightning model."""
        logger.info(f"Setting up for training...")
        start_setup_time = time.time()

        # initialize the labels object and update config.
        if not len(self.train_labels) or not len(self.val_labels):
            # Reached when the bare constructor was used (`ModelTrainer(config=...)`)
            # instead of `get_model_trainer_from_config`. This guard used to pass the
            # CONFIG where `_setup_train_val_labels` expects a `List[sio.Labels]`, so
            # it dereferenced `labels[0].skeletons` and died with
            # `ConfigKeyError: Missing key 0` -- it could only ever raise. Both entry
            # points now run the SAME initializer, so this path cannot drift from the
            # factory again (it had already lost `_set_seed` and the video check).
            self._initialize_from_config()

        # create the ckpt dir.
        self._setup_model_ckpt_dir()

        # create the train and val datasets for visualization.
        viz_train_dataset = None
        viz_val_dataset = None
        if self.config.trainer_config.visualize_preds_during_training:
            logger.info(f"Setting up visualization train and val datasets...")
            viz_train_dataset, viz_val_dataset = self._setup_viz_datasets()

        # setup loggers and callbacks for Trainer.
        logger.info(f"Setting up Trainer...")
        loggers, callbacks = self._setup_loggers_callbacks(
            viz_train_dataset=viz_train_dataset, viz_val_dataset=viz_val_dataset
        )
        # set up the strategy (for multi-gpu training)
        strategy = OmegaConf.select(
            self.config, "trainer_config.trainer_strategy", default="auto"
        )
        # set up profilers
        cfg_profiler = self.config.trainer_config.profiler
        profiler = None
        if cfg_profiler is not None:
            if cfg_profiler in self._profilers:
                profiler = self._profilers[cfg_profiler]
            else:
                message = f"{cfg_profiler} is not a valid option. Please choose one of {list(self._profilers.keys())}"
                logger.error(message)
                raise ValueError(message)

        devices = (
            OmegaConf.select(
                self.config, "trainer_config.trainer_device_indices", default=None
            )
            if OmegaConf.select(
                self.config, "trainer_config.trainer_device_indices", default=None
            )
            is not None
            else self.config.trainer_config.trainer_devices
        )
        logger.info(f"Trainer devices: {devices}")

        # if trainer devices is set to less than the number of available GPUs, use the least used GPUs
        if (
            torch.cuda.is_available()
            and self.config.trainer_config.trainer_accelerator != "cpu"
            and isinstance(self.config.trainer_config.trainer_devices, int)
            and self.config.trainer_config.trainer_devices < torch.cuda.device_count()
            and self.config.trainer_config.trainer_device_indices is None
        ):
            devices = [
                int(x)
                for x in np.argsort(get_gpu_memory())[::-1][
                    : self.config.trainer_config.trainer_devices
                ]
            ]
            # Sort device indices in ascending order for NCCL compatibility.
            # NCCL expects devices in consistent ascending order across ranks
            # to properly set up communication rings. Without sorting, DDP may
            # assign multiple ranks to the same GPU, causing "Duplicate GPU detected" errors.
            devices.sort()
            logger.info(f"Using GPUs with most available memory: {devices}")

        # create lightning.Trainer instance.
        # The embedding loader uses a custom group-aware ``batch_sampler`` that shards
        # itself per rank (seed + rank). Lightning's default sampler replacement cannot
        # inject a DistributedSampler into a custom batch_sampler (it raises), so disable
        # replacement for embedding; every other model type passes an explicit
        # DistributedSampler under DDP, so the default (True) is correct for them.
        use_distributed_sampler = self.model_type != "embedding"

        self.trainer = L.Trainer(
            callbacks=callbacks,
            logger=loggers,
            enable_checkpointing=self.config.trainer_config.save_ckpt,
            devices=devices,
            max_epochs=self.config.trainer_config.max_epochs,
            accelerator=self.config.trainer_config.trainer_accelerator,
            enable_progress_bar=self.config.trainer_config.enable_progress_bar,
            strategy=strategy,
            profiler=profiler,
            log_every_n_steps=1,
            use_distributed_sampler=use_distributed_sampler,
        )

        self.trainer.strategy.barrier()

        # setup datasets
        train_dataset, val_dataset = self._setup_datasets()

        # Barrier after dataset creation to ensure all workers wait for disk caching
        # (rank 0 caches to disk, others must wait before reading cached files)
        self.trainer.strategy.barrier()

        # set-up steps per epoch
        train_steps_per_epoch = self.config.trainer_config.train_steps_per_epoch
        tiling = OmegaConf.select(
            self.config, "data_config.preprocessing.tiling", default=None
        )
        if train_steps_per_epoch is None:
            if (
                tiling is not None
                and tiling.enabled
                and tiling.steps_per_epoch is not None
            ):
                # TRAIN decouple: the tiling knob overrides the tile-count length.
                train_steps_per_epoch = tiling.steps_per_epoch
                logger.info(
                    f"train_steps_per_epoch not set; using tiling.steps_per_epoch={train_steps_per_epoch}"
                )
            else:
                train_steps_per_epoch = get_steps_per_epoch(
                    dataset=train_dataset,
                    batch_size=self.config.trainer_config.train_data_loader.batch_size,
                )
                logger.info(
                    f"train_steps_per_epoch not set; computed {train_steps_per_epoch} from training dataset"
                )
        else:
            logger.info(
                f"Using configured train_steps_per_epoch={train_steps_per_epoch}"
            )
        min_train_steps_per_epoch = self.config.trainer_config.min_train_steps_per_epoch
        if min_train_steps_per_epoch > train_steps_per_epoch:
            logger.info(
                f"train_steps_per_epoch={train_steps_per_epoch} is below "
                f"min_train_steps_per_epoch={min_train_steps_per_epoch}; using the minimum"
            )
            train_steps_per_epoch = min_train_steps_per_epoch
        self.config.trainer_config.train_steps_per_epoch = train_steps_per_epoch
        logger.info(f"Final train_steps_per_epoch={train_steps_per_epoch}")

        # VAL: always full-coverage (every grid tile visited once), NOT decoupled.
        val_steps_per_epoch = get_steps_per_epoch(
            dataset=val_dataset,
            batch_size=self.config.trainer_config.val_data_loader.batch_size,
        )

        logger.info(f"Training on {self.trainer.num_devices} device(s)")
        logger.info(f"Training on {self.trainer.strategy.root_device} accelerator")

        # initialize the lightning model.
        # need to initialize after Trainer is initialized (for trainer accelerator)
        logger.info(f"Setting up lightning module for {self.model_type} model...")
        self.lightning_model = LightningModel.get_lightning_model_from_config(
            config=self.config,
        )
        logger.info(f"Backbone model: {self.lightning_model.model.backbone}")
        logger.info(f"Head model: {self.lightning_model.model.head_layers}")
        total_params = sum(p.numel() for p in self.lightning_model.parameters())
        logger.info(f"Total model parameters: {total_params:,}")
        self.config.model_config.total_params = total_params

        # setup dataloaders
        # need to set up dataloaders after Trainer is initialized (for ddp). DistributedSampler depends on the rank
        logger.info(
            f"Input image shape: {train_dataset[0]['image'].shape if 'image' in train_dataset[0] else train_dataset[0]['instance_image'].shape}"
        )
        train_dataloader, val_dataloader = get_train_val_dataloaders(
            train_dataset=train_dataset,
            val_dataset=val_dataset,
            config=self.config,
            rank=self.trainer.global_rank,
            train_steps_per_epoch=self.config.trainer_config.train_steps_per_epoch,
            val_steps_per_epoch=val_steps_per_epoch,
            trainer_devices=self.trainer.num_devices,
        )

        if self.trainer.global_rank == 0:  # save config only in rank 0 process
            ckpt_path = (
                Path(self.config.trainer_config.ckpt_dir)
                / self.config.trainer_config.run_name
            ).as_posix()

            # Overwrite version with current sleap-nn version
            self._initial_config.sleap_nn_version = sleap_nn.__version__
            self.config.sleap_nn_version = sleap_nn.__version__

            OmegaConf.save(
                self._initial_config,
                (Path(ckpt_path) / "initial_config.yaml").as_posix(),
            )

            if self.config.trainer_config.use_wandb:
                if wandb.run is None:
                    wandb.init(
                        dir=(
                            Path(self.config.trainer_config.ckpt_dir)
                            / self.config.trainer_config.run_name
                        ).as_posix(),
                        project=self.config.trainer_config.wandb.project,
                        entity=self.config.trainer_config.wandb.entity,
                        name=self.config.trainer_config.wandb.name,
                        id=self.config.trainer_config.wandb.prv_runid,
                        group=self.config.trainer_config.wandb.group,
                    )

                # Define custom x-axes for wandb metrics
                # Epoch-level metrics use epoch as x-axis, step-level use default global_step
                wandb.define_metric("epoch")

                # Training metrics (train/ prefix for grouping) - all use epoch x-axis
                wandb.define_metric("train/*", step_metric="epoch")
                wandb.define_metric("train/confmaps/*", step_metric="epoch")

                # Validation metrics (val/ prefix for grouping)
                wandb.define_metric("val/*", step_metric="epoch")

                # Evaluation metrics (eval/ prefix for grouping)
                wandb.define_metric("eval/*", step_metric="epoch")

                # Visualization images (need explicit nested paths)
                wandb.define_metric("viz/*", step_metric="epoch")
                wandb.define_metric("viz/train/*", step_metric="epoch")
                wandb.define_metric("viz/val/*", step_metric="epoch")

                self.config.trainer_config.wandb.current_run_id = wandb.run.id
                wandb.config["run_name"] = self.config.trainer_config.wandb.name
                wandb.config["run_config"] = OmegaConf.to_container(
                    self.config, resolve=True
                )

            OmegaConf.save(
                self.config,
                (
                    Path(self.config.trainer_config.ckpt_dir)
                    / self.config.trainer_config.run_name
                    / "training_config.yaml"
                ).as_posix(),
            )

        self.trainer.strategy.barrier()

        # Flag to track if training was interrupted (not completed normally)
        training_interrupted = False

        try:
            logger.info(
                f"Finished trainer set up. [{time.time() - start_setup_time:.1f}s]"
            )
            logger.info(f"Starting training loop...")
            start_train_time = time.time()
            self.trainer.fit(
                self.lightning_model,
                train_dataloader,
                val_dataloader,
                ckpt_path=self.config.trainer_config.resume_ckpt_path,
            )

        except KeyboardInterrupt:
            logger.info("Stopping training...")
            training_interrupted = True

        finally:
            logger.info(
                f"Finished training loop. [{(time.time() - start_train_time) / 60:.1f} min]"
            )
            # Note: wandb.finish() is called in train.py after post-training evaluation

            # delete image disk caching
            if (
                self.config.data_config.data_pipeline_fw
                == "torch_dataset_cache_img_disk"
                and self.config.data_config.delete_cache_imgs_after_training
            ):
                if self.trainer.global_rank == 0:
                    self._delete_cache_imgs()

            # delete viz folder if requested
            if (
                self.config.trainer_config.visualize_preds_during_training
                and not self.config.trainer_config.keep_viz
            ):
                if self.trainer.global_rank == 0:
                    viz_dir = (
                        Path(self.config.trainer_config.ckpt_dir)
                        / self.config.trainer_config.run_name
                        / "viz"
                    )
                    if viz_dir.exists():
                        logger.info(f"Deleting viz folder at {viz_dir}...")
                        shutil.rmtree(viz_dir, ignore_errors=True)

            # Clean up entire run folder if training was interrupted (KeyboardInterrupt)
            if training_interrupted and self.trainer.global_rank == 0:
                run_dir = (
                    Path(self.config.trainer_config.ckpt_dir)
                    / self.config.trainer_config.run_name
                )
                if run_dir.exists():
                    logger.info(
                        f"Training canceled - cleaning up run folder at {run_dir}..."
                    )
                    shutil.rmtree(run_dir, ignore_errors=True)

get_model_trainer_from_config(config, train_labels=None, val_labels=None) classmethod

Create a model trainer instance from config.

Source code in sleap_nn/training/model_trainer.py
@classmethod
def get_model_trainer_from_config(
    cls,
    config: DictConfig,
    train_labels: Optional[List[sio.Labels]] = None,
    val_labels: Optional[List[sio.Labels]] = None,
):
    """Create a model trainer instance from config."""
    model_trainer = cls(config=config)
    model_trainer._initialize_from_config(
        train_labels=train_labels, val_labels=val_labels
    )
    return model_trainer

setup_config()

Compute config parameters.

Source code in sleap_nn/training/model_trainer.py
def setup_config(self):
    """Compute config parameters."""
    logger.info("Setting up config...")

    # Normalize empty strings to None for optional wandb fields
    if self.config.trainer_config.wandb.prv_runid == "":
        self.config.trainer_config.wandb.prv_runid = None

    # compute preprocessing parameters from the labels objects and fill in the config
    self._setup_preprocessing_config()

    # save skeleton to config
    skeleton_yaml = yaml.safe_load(SkeletonYAMLEncoder().encode(self.skeletons))
    skeleton_names = skeleton_yaml.keys()
    self.config["data_config"]["skeletons"] = []
    for skeleton_name in skeleton_names:
        skl = skeleton_yaml[skeleton_name]
        skl["name"] = skeleton_name
        self.config["data_config"]["skeletons"].append(skl)

    # setup head config - partnames, edges and class names
    self._setup_head_config()

    # Identity-equality gates (SPEC §4.4): the embedding objective's pos/neg
    # sources silently assert "same / different animal"; validate them against the
    # declared `data_config.identity` semantics (error for global_id without global
    # names; warn for unproofread tracklets / non-deduplicated same-frame negatives).
    if self.model_type == "embedding":
        # Real `sio.Identity` annotations ground `global_id` grouping directly (no
        # `track_names_are_global` promise needed).
        has_identities = any(
            bool(getattr(label, "identities", None) or [])
            for label in self.train_labels
        )
        validate_embedding_identity(
            objective=OmegaConf.select(
                self.config,
                "model_config.head_configs.embedding.embedding.objective",
                default=None,
            ),
            identity=OmegaConf.select(
                self.config, "data_config.identity", default=None
            ),
            has_identities=has_identities,
        )

    # set max stride for the backbone: convnext and swint
    if self.backbone_type == "convnext":
        self.config.model_config.backbone_config.convnext.max_stride = (
            self.config.model_config.backbone_config.convnext.stem_patch_stride
            * (2**3)
            * 2
        )
    elif self.backbone_type == "swint":
        self.config.model_config.backbone_config.swint.max_stride = (
            self.config.model_config.backbone_config.swint.stem_patch_stride
            * (2**3)
            * 2
        )

    # set output stride for backbone from head config and verify max stride
    self.config = check_output_strides(self.config)

    # fail fast on a contradictory / unknown centroid method (#586)
    self.config = check_centroid_methods(self.config)

    # auto-size + validate tiling geometry (no-op unless tiling.enabled)
    self._setup_tiling_config()
    self.config = check_tiling(self.config)

    # verify the configured accelerator is available on this machine
    self._verify_accelerator_config()

    # if trainer_devices is None, set it to "auto"
    if self.config.trainer_config.trainer_devices is None:
        self.config.trainer_config.trainer_devices = (
            "auto"
            if OmegaConf.select(
                self.config, "trainer_config.trainer_device_indices", default=None
            )
            is None
            else len(
                OmegaConf.select(
                    self.config,
                    "trainer_config.trainer_device_indices",
                    default=None,
                )
            )
        )

    # setup checkpoint path (generates run_name if not specified)
    self._setup_ckpt_path()

    # Default wandb run name to trainer run_name if not specified
    # Note: This must come after _setup_ckpt_path() which generates run_name
    if self.config.trainer_config.wandb.name is None:
        self.config.trainer_config.wandb.name = self.config.trainer_config.run_name

    # verify input_channels in model_config based on input image and pretrained model weights
    self._verify_model_input_channels()

train()

Train the lightning model.

Source code in sleap_nn/training/model_trainer.py
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
def train(self):
    """Train the lightning model."""
    logger.info(f"Setting up for training...")
    start_setup_time = time.time()

    # initialize the labels object and update config.
    if not len(self.train_labels) or not len(self.val_labels):
        # Reached when the bare constructor was used (`ModelTrainer(config=...)`)
        # instead of `get_model_trainer_from_config`. This guard used to pass the
        # CONFIG where `_setup_train_val_labels` expects a `List[sio.Labels]`, so
        # it dereferenced `labels[0].skeletons` and died with
        # `ConfigKeyError: Missing key 0` -- it could only ever raise. Both entry
        # points now run the SAME initializer, so this path cannot drift from the
        # factory again (it had already lost `_set_seed` and the video check).
        self._initialize_from_config()

    # create the ckpt dir.
    self._setup_model_ckpt_dir()

    # create the train and val datasets for visualization.
    viz_train_dataset = None
    viz_val_dataset = None
    if self.config.trainer_config.visualize_preds_during_training:
        logger.info(f"Setting up visualization train and val datasets...")
        viz_train_dataset, viz_val_dataset = self._setup_viz_datasets()

    # setup loggers and callbacks for Trainer.
    logger.info(f"Setting up Trainer...")
    loggers, callbacks = self._setup_loggers_callbacks(
        viz_train_dataset=viz_train_dataset, viz_val_dataset=viz_val_dataset
    )
    # set up the strategy (for multi-gpu training)
    strategy = OmegaConf.select(
        self.config, "trainer_config.trainer_strategy", default="auto"
    )
    # set up profilers
    cfg_profiler = self.config.trainer_config.profiler
    profiler = None
    if cfg_profiler is not None:
        if cfg_profiler in self._profilers:
            profiler = self._profilers[cfg_profiler]
        else:
            message = f"{cfg_profiler} is not a valid option. Please choose one of {list(self._profilers.keys())}"
            logger.error(message)
            raise ValueError(message)

    devices = (
        OmegaConf.select(
            self.config, "trainer_config.trainer_device_indices", default=None
        )
        if OmegaConf.select(
            self.config, "trainer_config.trainer_device_indices", default=None
        )
        is not None
        else self.config.trainer_config.trainer_devices
    )
    logger.info(f"Trainer devices: {devices}")

    # if trainer devices is set to less than the number of available GPUs, use the least used GPUs
    if (
        torch.cuda.is_available()
        and self.config.trainer_config.trainer_accelerator != "cpu"
        and isinstance(self.config.trainer_config.trainer_devices, int)
        and self.config.trainer_config.trainer_devices < torch.cuda.device_count()
        and self.config.trainer_config.trainer_device_indices is None
    ):
        devices = [
            int(x)
            for x in np.argsort(get_gpu_memory())[::-1][
                : self.config.trainer_config.trainer_devices
            ]
        ]
        # Sort device indices in ascending order for NCCL compatibility.
        # NCCL expects devices in consistent ascending order across ranks
        # to properly set up communication rings. Without sorting, DDP may
        # assign multiple ranks to the same GPU, causing "Duplicate GPU detected" errors.
        devices.sort()
        logger.info(f"Using GPUs with most available memory: {devices}")

    # create lightning.Trainer instance.
    # The embedding loader uses a custom group-aware ``batch_sampler`` that shards
    # itself per rank (seed + rank). Lightning's default sampler replacement cannot
    # inject a DistributedSampler into a custom batch_sampler (it raises), so disable
    # replacement for embedding; every other model type passes an explicit
    # DistributedSampler under DDP, so the default (True) is correct for them.
    use_distributed_sampler = self.model_type != "embedding"

    self.trainer = L.Trainer(
        callbacks=callbacks,
        logger=loggers,
        enable_checkpointing=self.config.trainer_config.save_ckpt,
        devices=devices,
        max_epochs=self.config.trainer_config.max_epochs,
        accelerator=self.config.trainer_config.trainer_accelerator,
        enable_progress_bar=self.config.trainer_config.enable_progress_bar,
        strategy=strategy,
        profiler=profiler,
        log_every_n_steps=1,
        use_distributed_sampler=use_distributed_sampler,
    )

    self.trainer.strategy.barrier()

    # setup datasets
    train_dataset, val_dataset = self._setup_datasets()

    # Barrier after dataset creation to ensure all workers wait for disk caching
    # (rank 0 caches to disk, others must wait before reading cached files)
    self.trainer.strategy.barrier()

    # set-up steps per epoch
    train_steps_per_epoch = self.config.trainer_config.train_steps_per_epoch
    tiling = OmegaConf.select(
        self.config, "data_config.preprocessing.tiling", default=None
    )
    if train_steps_per_epoch is None:
        if (
            tiling is not None
            and tiling.enabled
            and tiling.steps_per_epoch is not None
        ):
            # TRAIN decouple: the tiling knob overrides the tile-count length.
            train_steps_per_epoch = tiling.steps_per_epoch
            logger.info(
                f"train_steps_per_epoch not set; using tiling.steps_per_epoch={train_steps_per_epoch}"
            )
        else:
            train_steps_per_epoch = get_steps_per_epoch(
                dataset=train_dataset,
                batch_size=self.config.trainer_config.train_data_loader.batch_size,
            )
            logger.info(
                f"train_steps_per_epoch not set; computed {train_steps_per_epoch} from training dataset"
            )
    else:
        logger.info(
            f"Using configured train_steps_per_epoch={train_steps_per_epoch}"
        )
    min_train_steps_per_epoch = self.config.trainer_config.min_train_steps_per_epoch
    if min_train_steps_per_epoch > train_steps_per_epoch:
        logger.info(
            f"train_steps_per_epoch={train_steps_per_epoch} is below "
            f"min_train_steps_per_epoch={min_train_steps_per_epoch}; using the minimum"
        )
        train_steps_per_epoch = min_train_steps_per_epoch
    self.config.trainer_config.train_steps_per_epoch = train_steps_per_epoch
    logger.info(f"Final train_steps_per_epoch={train_steps_per_epoch}")

    # VAL: always full-coverage (every grid tile visited once), NOT decoupled.
    val_steps_per_epoch = get_steps_per_epoch(
        dataset=val_dataset,
        batch_size=self.config.trainer_config.val_data_loader.batch_size,
    )

    logger.info(f"Training on {self.trainer.num_devices} device(s)")
    logger.info(f"Training on {self.trainer.strategy.root_device} accelerator")

    # initialize the lightning model.
    # need to initialize after Trainer is initialized (for trainer accelerator)
    logger.info(f"Setting up lightning module for {self.model_type} model...")
    self.lightning_model = LightningModel.get_lightning_model_from_config(
        config=self.config,
    )
    logger.info(f"Backbone model: {self.lightning_model.model.backbone}")
    logger.info(f"Head model: {self.lightning_model.model.head_layers}")
    total_params = sum(p.numel() for p in self.lightning_model.parameters())
    logger.info(f"Total model parameters: {total_params:,}")
    self.config.model_config.total_params = total_params

    # setup dataloaders
    # need to set up dataloaders after Trainer is initialized (for ddp). DistributedSampler depends on the rank
    logger.info(
        f"Input image shape: {train_dataset[0]['image'].shape if 'image' in train_dataset[0] else train_dataset[0]['instance_image'].shape}"
    )
    train_dataloader, val_dataloader = get_train_val_dataloaders(
        train_dataset=train_dataset,
        val_dataset=val_dataset,
        config=self.config,
        rank=self.trainer.global_rank,
        train_steps_per_epoch=self.config.trainer_config.train_steps_per_epoch,
        val_steps_per_epoch=val_steps_per_epoch,
        trainer_devices=self.trainer.num_devices,
    )

    if self.trainer.global_rank == 0:  # save config only in rank 0 process
        ckpt_path = (
            Path(self.config.trainer_config.ckpt_dir)
            / self.config.trainer_config.run_name
        ).as_posix()

        # Overwrite version with current sleap-nn version
        self._initial_config.sleap_nn_version = sleap_nn.__version__
        self.config.sleap_nn_version = sleap_nn.__version__

        OmegaConf.save(
            self._initial_config,
            (Path(ckpt_path) / "initial_config.yaml").as_posix(),
        )

        if self.config.trainer_config.use_wandb:
            if wandb.run is None:
                wandb.init(
                    dir=(
                        Path(self.config.trainer_config.ckpt_dir)
                        / self.config.trainer_config.run_name
                    ).as_posix(),
                    project=self.config.trainer_config.wandb.project,
                    entity=self.config.trainer_config.wandb.entity,
                    name=self.config.trainer_config.wandb.name,
                    id=self.config.trainer_config.wandb.prv_runid,
                    group=self.config.trainer_config.wandb.group,
                )

            # Define custom x-axes for wandb metrics
            # Epoch-level metrics use epoch as x-axis, step-level use default global_step
            wandb.define_metric("epoch")

            # Training metrics (train/ prefix for grouping) - all use epoch x-axis
            wandb.define_metric("train/*", step_metric="epoch")
            wandb.define_metric("train/confmaps/*", step_metric="epoch")

            # Validation metrics (val/ prefix for grouping)
            wandb.define_metric("val/*", step_metric="epoch")

            # Evaluation metrics (eval/ prefix for grouping)
            wandb.define_metric("eval/*", step_metric="epoch")

            # Visualization images (need explicit nested paths)
            wandb.define_metric("viz/*", step_metric="epoch")
            wandb.define_metric("viz/train/*", step_metric="epoch")
            wandb.define_metric("viz/val/*", step_metric="epoch")

            self.config.trainer_config.wandb.current_run_id = wandb.run.id
            wandb.config["run_name"] = self.config.trainer_config.wandb.name
            wandb.config["run_config"] = OmegaConf.to_container(
                self.config, resolve=True
            )

        OmegaConf.save(
            self.config,
            (
                Path(self.config.trainer_config.ckpt_dir)
                / self.config.trainer_config.run_name
                / "training_config.yaml"
            ).as_posix(),
        )

    self.trainer.strategy.barrier()

    # Flag to track if training was interrupted (not completed normally)
    training_interrupted = False

    try:
        logger.info(
            f"Finished trainer set up. [{time.time() - start_setup_time:.1f}s]"
        )
        logger.info(f"Starting training loop...")
        start_train_time = time.time()
        self.trainer.fit(
            self.lightning_model,
            train_dataloader,
            val_dataloader,
            ckpt_path=self.config.trainer_config.resume_ckpt_path,
        )

    except KeyboardInterrupt:
        logger.info("Stopping training...")
        training_interrupted = True

    finally:
        logger.info(
            f"Finished training loop. [{(time.time() - start_train_time) / 60:.1f} min]"
        )
        # Note: wandb.finish() is called in train.py after post-training evaluation

        # delete image disk caching
        if (
            self.config.data_config.data_pipeline_fw
            == "torch_dataset_cache_img_disk"
            and self.config.data_config.delete_cache_imgs_after_training
        ):
            if self.trainer.global_rank == 0:
                self._delete_cache_imgs()

        # delete viz folder if requested
        if (
            self.config.trainer_config.visualize_preds_during_training
            and not self.config.trainer_config.keep_viz
        ):
            if self.trainer.global_rank == 0:
                viz_dir = (
                    Path(self.config.trainer_config.ckpt_dir)
                    / self.config.trainer_config.run_name
                    / "viz"
                )
                if viz_dir.exists():
                    logger.info(f"Deleting viz folder at {viz_dir}...")
                    shutil.rmtree(viz_dir, ignore_errors=True)

        # Clean up entire run folder if training was interrupted (KeyboardInterrupt)
        if training_interrupted and self.trainer.global_rank == 0:
            run_dir = (
                Path(self.config.trainer_config.ckpt_dir)
                / self.config.trainer_config.run_name
            )
            if run_dir.exists():
                logger.info(
                    f"Training canceled - cleaning up run folder at {run_dir}..."
                )
                shutil.rmtree(run_dir, ignore_errors=True)