File size: 64,131 Bytes
5c6ebf3 95bb532 5c6ebf3 2e19bed 5c6ebf3 2e19bed 5c6ebf3 2e19bed 5c6ebf3 27831c5 5c6ebf3 27831c5 5c6ebf3 27831c5 5c6ebf3 27831c5 5c6ebf3 95bb532 5c6ebf3 95bb532 5c6ebf3 27831c5 5c6ebf3 95bb532 5c6ebf3 27831c5 5c6ebf3 95bb532 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 |
import os
import warnings
from collections import OrderedDict
from copy import deepcopy
import logging
import math
from typing import Sequence
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as cp
import numpy as np
import cv2
from dataclasses import dataclass
from transformers import PreTrainedModel, PretrainedConfig
from transformers import PretrainedConfig
# from .lavis_base_model import BaseEncoder
# from lavis.common.registry import registry
from torch.nn import Module as BaseModule
from torch.nn import ModuleList
from torch.nn import Sequential
from torch.nn import Linear
from torch import Tensor
from itertools import repeat
import collections.abc
from .configuration_solider import SOLIDERConfig, BACKBONE_NAME2WIDTH
def _ntuple(n):
def parse(x):
if isinstance(x, collections.abc.Iterable):
return x
return tuple(repeat(x, n))
return parse
to_2tuple = _ntuple(2)
def trunc_normal_init(
module: nn.Module,
mean: float = 0,
std: float = 1,
a: float = -2,
b: float = 2,
bias: float = 0,
) -> None:
if hasattr(module, "weight") and module.weight is not None:
# trunc_normal_(module.weight, mean, std, a, b) # type: ignore
_no_grad_trunc_normal_(module.weight, mean, std, a, b) # type: ignore
if hasattr(module, "bias") and module.bias is not None:
nn.init.constant_(module.bias, bias) # type: ignore
def _no_grad_trunc_normal_(
tensor: Tensor, mean: float, std: float, a: float, b: float
) -> Tensor:
# Method based on
# https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
# Modified from
# https://github.com/pytorch/pytorch/blob/master/torch/nn/init.py
def norm_cdf(x):
# Computes standard normal cumulative distribution function
return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
if (mean < a - 2 * std) or (mean > b + 2 * std):
warnings.warn(
"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
"The distribution of values may be incorrect.",
stacklevel=2,
)
with torch.no_grad():
# Values are generated by using a truncated uniform distribution and
# then using the inverse CDF for the normal distribution.
# Get upper and lower cdf values
lower = norm_cdf((a - mean) / std)
upper = norm_cdf((b - mean) / std)
# Uniformly fill tensor with values from [lower, upper], then translate
# to [2lower-1, 2upper-1].
tensor.uniform_(2 * lower - 1, 2 * upper - 1)
# Use inverse cdf transform for normal distribution to get truncated
# standard normal
tensor.erfinv_()
# Transform to proper mean, std
tensor.mul_(std * math.sqrt(2.0))
tensor.add_(mean)
# Clamp to ensure it's in the proper range
tensor.clamp_(min=a, max=b)
return tensor
def trunc_normal_(
tensor: Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
) -> Tensor:
r"""Fills the input Tensor with values drawn from a truncated
normal distribution. The values are effectively drawn from the
normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
with values outside :math:`[a, b]` redrawn until they are within
the bounds. The method used for generating the random values works
best when :math:`a \leq \text{mean} \leq b`.
Modified from
https://github.com/pytorch/pytorch/blob/master/torch/nn/init.py
Args:
tensor (``torch.Tensor``): an n-dimensional `torch.Tensor`.
mean (float): the mean of the normal distribution.
std (float): the standard deviation of the normal distribution.
a (float): the minimum cutoff value.
b (float): the maximum cutoff value.
"""
return _no_grad_trunc_normal_(tensor, mean, std, a, b)
def constant_init(module, val, bias=0):
if hasattr(module, "weight") and module.weight is not None:
nn.init.constant_(module.weight, val)
if hasattr(module, "bias") and module.bias is not None:
nn.init.constant_(module.bias, bias)
def build_norm_layer(norm_cfg, embed_dims):
assert norm_cfg["type"] == "LN"
norm_layer = nn.LayerNorm(embed_dims)
return norm_cfg["type"], norm_layer
class GELU(nn.Module):
r"""Applies the Gaussian Error Linear Units function:
.. math::
\text{GELU}(x) = x * \Phi(x)
where :math:`\Phi(x)` is the Cumulative Distribution Function for
Gaussian Distribution.
Shape:
- Input: :math:`(N, *)` where `*` means, any number of additional
dimensions
- Output: :math:`(N, *)`, same shape as the input
.. image:: scripts/activation_images/GELU.png
Examples::
>>> m = nn.GELU()
>>> input = torch.randn(2)
>>> output = m(input)
"""
def forward(self, input):
return F.gelu(input)
def build_activation_layer(act_cfg):
if act_cfg["type"] == "ReLU":
act_layer = nn.ReLU(inplace=act_cfg["inplace"])
elif act_cfg["type"] == "GELU":
act_layer = GELU()
return act_layer
def build_conv_layer(
conv_cfg, in_channels, out_channels, kernel_size, stride, padding, dilation, bias
):
conv_layer = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
bias=bias,
)
return conv_layer
def drop_path(x, drop_prob=0.0, training=False):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of
residual blocks).
We follow the implementation
https://github.com/rwightman/pytorch-image-models/blob/a2727c1bf78ba0d7b5727f5f95e37fb7f8866b1f/timm/models/layers/drop.py # noqa: E501
"""
if drop_prob == 0.0 or not training:
return x
keep_prob = 1 - drop_prob
# handle tensors with different dimensions, not just 4D tensors.
shape = (x.shape[0],) + (1,) * (x.ndim - 1)
random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
output = x.div(keep_prob) * random_tensor.floor()
return output
class DropPath(nn.Module):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of
residual blocks).
We follow the implementation
https://github.com/rwightman/pytorch-image-models/blob/a2727c1bf78ba0d7b5727f5f95e37fb7f8866b1f/timm/models/layers/drop.py # noqa: E501
Args:
drop_prob (float): Probability of the path to be zeroed. Default: 0.1
"""
def __init__(self, drop_prob=0.1):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
def forward(self, x):
return drop_path(x, self.drop_prob, self.training)
def build_dropout(drop_cfg):
drop_layer = DropPath(drop_cfg["drop_prob"])
return drop_layer
class FFN(BaseModule):
def __init__(
self,
embed_dims=256,
feedforward_channels=1024,
num_fcs=2,
act_cfg=dict(type="ReLU", inplace=True),
ffn_drop=0.0,
dropout_layer=None,
add_identity=True,
init_cfg=None,
**kwargs,
):
super(FFN, self).__init__()
assert num_fcs >= 2, "num_fcs should be no less " f"than 2. got {num_fcs}."
self.embed_dims = embed_dims
self.feedforward_channels = feedforward_channels
self.num_fcs = num_fcs
self.act_cfg = act_cfg
self.activate = build_activation_layer(act_cfg)
layers = []
in_channels = embed_dims
for _ in range(num_fcs - 1):
layers.append(
Sequential(
Linear(in_channels, feedforward_channels),
self.activate,
nn.Dropout(ffn_drop),
)
)
in_channels = feedforward_channels
layers.append(Linear(feedforward_channels, embed_dims))
layers.append(nn.Dropout(ffn_drop))
self.layers = Sequential(*layers)
self.dropout_layer = (
build_dropout(dropout_layer) if dropout_layer else torch.nn.Identity()
)
self.add_identity = add_identity
def forward(self, x, identity=None):
"""Forward function for `FFN`.
The function would add x to the output tensor if residue is None.
"""
out = self.layers(x)
if not self.add_identity:
return self.dropout_layer(out)
if identity is None:
identity = x
return identity + self.dropout_layer(out)
def swin_converter(ckpt):
new_ckpt = OrderedDict()
def correct_unfold_reduction_order(x):
out_channel, in_channel = x.shape
x = x.reshape(out_channel, 4, in_channel // 4)
x = x[:, [0, 2, 1, 3], :].transpose(1, 2).reshape(out_channel, in_channel)
return x
def correct_unfold_norm_order(x):
in_channel = x.shape[0]
x = x.reshape(4, in_channel // 4)
x = x[[0, 2, 1, 3], :].transpose(0, 1).reshape(in_channel)
return x
for k, v in ckpt.items():
if k.startswith("head"):
continue
elif k.startswith("layers"):
new_v = v
if "attn." in k:
new_k = k.replace("attn.", "attn.w_msa.")
elif "mlp." in k:
if "mlp.fc1." in k:
new_k = k.replace("mlp.fc1.", "ffn.layers.0.0.")
elif "mlp.fc2." in k:
new_k = k.replace("mlp.fc2.", "ffn.layers.1.")
else:
new_k = k.replace("mlp.", "ffn.")
elif "downsample" in k:
new_k = k
if "reduction." in k:
new_v = correct_unfold_reduction_order(v)
elif "norm." in k:
new_v = correct_unfold_norm_order(v)
else:
new_k = k
new_k = new_k.replace("layers", "stages", 1)
elif k.startswith("patch_embed"):
new_v = v
if "proj" in k:
new_k = k.replace("proj", "projection")
else:
new_k = k
else:
new_v = v
new_k = k
new_ckpt["backbone." + new_k] = new_v
return new_ckpt
class AdaptivePadding(nn.Module):
"""Applies padding to input (if needed) so that input can get fully covered
by filter you specified. It support two modes "same" and "corner". The
"same" mode is same with "SAME" padding mode in TensorFlow, pad zero around
input. The "corner" mode would pad zero to bottom right.
Args:
kernel_size (int | tuple): Size of the kernel:
stride (int | tuple): Stride of the filter. Default: 1:
dilation (int | tuple): Spacing between kernel elements.
Default: 1
padding (str): Support "same" and "corner", "corner" mode
would pad zero to bottom right, and "same" mode would
pad zero around input. Default: "corner".
Example:
>>> kernel_size = 16
>>> stride = 16
>>> dilation = 1
>>> input = torch.rand(1, 1, 15, 17)
>>> adap_pad = AdaptivePadding(
>>> kernel_size=kernel_size,
>>> stride=stride,
>>> dilation=dilation,
>>> padding="corner")
>>> out = adap_pad(input)
>>> assert (out.shape[2], out.shape[3]) == (16, 32)
>>> input = torch.rand(1, 1, 16, 17)
>>> out = adap_pad(input)
>>> assert (out.shape[2], out.shape[3]) == (16, 32)
"""
def __init__(self, kernel_size=1, stride=1, dilation=1, padding="corner"):
super(AdaptivePadding, self).__init__()
assert padding in ("same", "corner")
kernel_size = to_2tuple(kernel_size)
stride = to_2tuple(stride)
padding = to_2tuple(padding)
dilation = to_2tuple(dilation)
self.padding = padding
self.kernel_size = kernel_size
self.stride = stride
self.dilation = dilation
def get_pad_shape(self, input_shape):
input_h, input_w = input_shape
kernel_h, kernel_w = self.kernel_size
stride_h, stride_w = self.stride
output_h = math.ceil(input_h / stride_h)
output_w = math.ceil(input_w / stride_w)
pad_h = max(
(output_h - 1) * stride_h + (kernel_h - 1) * self.dilation[0] + 1 - input_h,
0,
)
pad_w = max(
(output_w - 1) * stride_w + (kernel_w - 1) * self.dilation[1] + 1 - input_w,
0,
)
return pad_h, pad_w
def forward(self, x):
B, C, h, w = x.shape
pad_h, pad_w = self.get_pad_shape((h, w))
if pad_h > 0 or pad_w > 0:
if self.padding == "corner":
return F.pad(x, [0, pad_w, 0, pad_h]).view(
B, C, h + pad_h, w + pad_w
), (
h + pad_h,
w + pad_w,
)
elif self.padding == "same":
return F.pad(
x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2]
).view(B, C, h + pad_h, w + pad_w), (
h + pad_h,
w + pad_w,
)
return x, (h, w)
class PatchEmbed(BaseModule):
"""Image to Patch Embedding.
We use a conv layer to implement PatchEmbed.
Args:
in_channels (int): The num of input channels. Default: 3
embed_dims (int): The dimensions of embedding. Default: 768
conv_type (str): The config dict for embedding
conv layer type selection. Default: "Conv2d.
kernel_size (int): The kernel_size of embedding conv. Default: 16.
stride (int): The slide stride of embedding conv.
Default: None (Would be set as `kernel_size`).
padding (int | tuple | string ): The padding length of
embedding conv. When it is a string, it means the mode
of adaptive padding, support "same" and "corner" now.
Default: "corner".
dilation (int): The dilation rate of embedding conv. Default: 1.
bias (bool): Bias of embed conv. Default: True.
norm_cfg (dict, optional): Config dict for normalization layer.
Default: None.
input_size (int | tuple | None): The size of input, which will be
used to calculate the out size. Only work when `dynamic_size`
is False. Default: None.
init_cfg (`mmcv.ConfigDict`, optional): The Config for initialization.
Default: None.
"""
def __init__(
self,
in_channels=3,
embed_dims=768,
conv_type="Conv2d",
kernel_size=16,
stride=16,
padding="corner",
dilation=1,
bias=True,
norm_cfg=None,
input_size=None,
init_cfg=None,
):
super(PatchEmbed, self).__init__()
self.embed_dims = embed_dims
if stride is None:
stride = kernel_size
kernel_size = to_2tuple(kernel_size)
stride = to_2tuple(stride)
dilation = to_2tuple(dilation)
if isinstance(padding, str):
self.adap_padding = AdaptivePadding(
kernel_size=kernel_size,
stride=stride,
dilation=dilation,
padding=padding,
)
# disable the padding of conv
padding = 0
else:
self.adap_padding = None
padding = to_2tuple(padding)
self.projection = build_conv_layer(
dict(type=conv_type),
in_channels=in_channels,
out_channels=embed_dims,
kernel_size=kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
bias=bias,
)
if norm_cfg is not None:
self.norm = build_norm_layer(norm_cfg, embed_dims)[1]
else:
self.norm = None
if input_size:
input_size = to_2tuple(input_size)
# `init_out_size` would be used outside to
# calculate the num_patches
# when `use_abs_pos_embed` outside
self.init_input_size = input_size
if self.adap_padding:
pad_h, pad_w = self.adap_padding.get_pad_shape(input_size)
input_h, input_w = input_size
input_h = input_h + pad_h
input_w = input_w + pad_w
input_size = (input_h, input_w)
# https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html
h_out = (
input_size[0] + 2 * padding[0] - dilation[0] * (kernel_size[0] - 1) - 1
) // stride[0] + 1
w_out = (
input_size[1] + 2 * padding[1] - dilation[1] * (kernel_size[1] - 1) - 1
) // stride[1] + 1
self.init_out_size = (h_out, w_out)
else:
self.init_input_size = None
self.init_out_size = None
def forward(self, x):
"""
Args:
x (Tensor): Has shape (B, C, H, W). In most case, C is 3.
Returns:
tuple: Contains merged results and its spatial shape.
- x (Tensor): Has shape (B, out_h * out_w, embed_dims)
- out_size (tuple[int]): Spatial shape of x, arrange as
(out_h, out_w).
"""
if self.adap_padding:
x, _ = self.adap_padding(x)
x = self.projection(x)
B, C, out_h, out_w = x.shape
x = x.view(B, C, out_h * out_w).transpose(1, 2)
if self.norm is not None:
x = self.norm(x)
return x, (out_h, out_w)
class PatchMerging(BaseModule):
"""Merge patch feature map.
This layer groups feature map by kernel_size, and applies norm and linear
layers to the grouped feature map. Our implementation uses `nn.Unfold` to
merge patch, which is about 25% faster than original implementation.
Instead, we need to modify pretrained models for compatibility.
Args:
in_channels (int): The num of input channels.
to gets fully covered by filter and stride you specified..
Default: True.
out_channels (int): The num of output channels.
kernel_size (int | tuple, optional): the kernel size in the unfold
layer. Defaults to 2.
stride (int | tuple, optional): the stride of the sliding blocks in the
unfold layer. Default: None. (Would be set as `kernel_size`)
padding (int | tuple | string ): The padding length of
embedding conv. When it is a string, it means the mode
of adaptive padding, support "same" and "corner" now.
Default: "corner".
dilation (int | tuple, optional): dilation parameter in the unfold
layer. Default: 1.
bias (bool, optional): Whether to add bias in linear layer or not.
Defaults: False.
norm_cfg (dict, optional): Config dict for normalization layer.
Default: dict(type='LN').
init_cfg (dict, optional): The extra config for initialization.
Default: None.
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size=2,
stride=None,
padding="corner",
dilation=1,
bias=False,
norm_cfg=dict(type="LN"),
init_cfg=None,
):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
if stride:
stride = stride
else:
stride = kernel_size
kernel_size = to_2tuple(kernel_size)
stride = to_2tuple(stride)
dilation = to_2tuple(dilation)
if isinstance(padding, str):
self.adap_padding = AdaptivePadding(
kernel_size=kernel_size,
stride=stride,
dilation=dilation,
padding=padding,
)
# disable the padding of unfold
padding = 0
else:
self.adap_padding = None
padding = to_2tuple(padding)
self.sampler = nn.Unfold(
kernel_size=kernel_size, dilation=dilation, padding=padding, stride=stride
)
sample_dim = kernel_size[0] * kernel_size[1] * in_channels
if norm_cfg is not None:
self.norm = build_norm_layer(norm_cfg, sample_dim)[1]
else:
self.norm = None
self.reduction = nn.Linear(sample_dim, out_channels, bias=bias)
def forward(self, x, input_size):
"""
Args:
x (Tensor): Has shape (B, H*W, C_in).
input_size (tuple[int]): The spatial shape of x, arrange as (H, W).
Default: None.
Returns:
tuple: Contains merged results and its spatial shape.
- x (Tensor): Has shape (B, Merged_H * Merged_W, C_out)
- out_size (tuple[int]): Spatial shape of x, arrange as
(Merged_H, Merged_W).
"""
B, L, C = x.shape
assert isinstance(input_size, Sequence), (
f"Expect " f"input_size is " f"`Sequence` " f"but get {input_size}"
)
H, W = input_size
assert L == H * W, "input feature has wrong size"
x = x.view(B, H, W, C).permute([0, 3, 1, 2]) # B, C, H, W
# Use nn.Unfold to merge patch. About 25% faster than original method,
# but need to modify pretrained model for compatibility
if self.adap_padding:
x, (H, W) = self.adap_padding(x)
x = self.sampler(x)
# if kernel_size=2 and stride=2, x should has shape (B, 4*C, H/2*W/2)
out_h = (
H
+ 2 * self.sampler.padding[0]
- self.sampler.dilation[0] * (self.sampler.kernel_size[0] - 1)
- 1
) // self.sampler.stride[0] + 1
out_w = (
W
+ 2 * self.sampler.padding[1]
- self.sampler.dilation[1] * (self.sampler.kernel_size[1] - 1)
- 1
) // self.sampler.stride[1] + 1
x = x.view(B, C * H * W // (out_h * out_w), out_h * out_w)
output_size = (out_h, out_w)
x = x.transpose(1, 2) # B, H/2*W/2, 4*C
x = self.norm(x) if self.norm else x
x = self.reduction(x)
return x, output_size
class WindowMSA(BaseModule):
"""Window based multi-head self-attention (W-MSA) module with relative
position bias.
Args:
embed_dims (int): Number of input channels.
num_heads (int): Number of attention heads.
window_size (tuple[int]): The height and width of the window.
qkv_bias (bool, optional): If True, add a learnable bias to q, k, v.
Default: True.
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Default: None.
attn_drop_rate (float, optional): Dropout ratio of attention weight.
Default: 0.0
proj_drop_rate (float, optional): Dropout ratio of output. Default: 0.
init_cfg (dict | None, optional): The Config for initialization.
Default: None.
"""
def __init__(
self,
embed_dims,
num_heads,
window_size,
qkv_bias=True,
qk_scale=None,
attn_drop_rate=0.0,
proj_drop_rate=0.0,
init_cfg=None,
):
super().__init__()
self.embed_dims = embed_dims
self.window_size = window_size # Wh, Ww
self.num_heads = num_heads
head_embed_dims = embed_dims // num_heads
self.scale = qk_scale or head_embed_dims**-0.5
self.init_cfg = init_cfg
# define a parameter table of relative position bias
self.relative_position_bias_table = nn.Parameter(
torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
) # 2*Wh-1 * 2*Ww-1, nH
# About 2x faster than original impl
Wh, Ww = self.window_size
rel_index_coords = self.double_step_seq(2 * Ww - 1, Wh, 1, Ww)
rel_position_index = rel_index_coords + rel_index_coords.T
rel_position_index = rel_position_index.flip(1).contiguous()
self.register_buffer("relative_position_index", rel_position_index)
self.qkv = nn.Linear(embed_dims, embed_dims * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop_rate)
self.proj = nn.Linear(embed_dims, embed_dims)
self.proj_drop = nn.Dropout(proj_drop_rate)
self.softmax = nn.Softmax(dim=-1)
def init_weights(self):
trunc_normal_(self.relative_position_bias_table, std=0.02)
def forward(self, x, mask, N, C, nW):
"""
Args:
x (tensor): input features with shape of (nW*B, N, C)
mask (tensor | None, Optional): mask with shape of (nW,
Wh*Ww, Wh*Ww), value should be between (-inf, 0].
"""
nWB = x.shape[0]
qkv = (
self.qkv(x)
.reshape(x.shape[0], N, 3, self.num_heads, C // self.num_heads)
.permute(2, 0, 3, 1, 4)
)
# make torchscript happy (cannot use tensor as tuple)
q, k, v = qkv[0], qkv[1], qkv[2]
q = q * self.scale
attn = q @ k.transpose(-2, -1)
relative_position_bias = self.relative_position_bias_table[
self.relative_position_index.view(
(
self.window_size[0]
* self.window_size[1]
* self.window_size[0]
* self.window_size[1],
)
)
].view(
self.window_size[0] * self.window_size[1],
self.window_size[0] * self.window_size[1],
self.num_heads,
) # Wh*Ww,Wh*Ww,nH
relative_position_bias = relative_position_bias.permute(
2, 0, 1
).contiguous() # nH, Wh*Ww, Wh*Ww
attn = attn + relative_position_bias.unsqueeze(0)
if mask is not None:
nW = mask.shape[0]
attn = attn.view(nWB // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
1
).unsqueeze(0)
attn = attn.view(nWB, self.num_heads, N, N)
attn = self.softmax(attn)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(nWB, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
@staticmethod
def double_step_seq(step1, len1, step2, len2):
seq1 = torch.arange(0, step1 * len1, step1)
seq2 = torch.arange(0, step2 * len2, step2)
return (seq1[:, None] + seq2[None, :]).reshape(1, -1)
class ShiftWindowMSA(BaseModule):
"""Shifted Window Multihead Self-Attention Module.
Args:
embed_dims (int): Number of input channels.
num_heads (int): Number of attention heads.
window_size (int): The height and width of the window.
shift_size (int, optional): The shift step of each window towards
right-bottom. If zero, act as regular window-msa. Defaults to 0.
qkv_bias (bool, optional): If True, add a learnable bias to q, k, v.
Default: True
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Defaults: None.
attn_drop_rate (float, optional): Dropout ratio of attention weight.
Defaults: 0.
proj_drop_rate (float, optional): Dropout ratio of output.
Defaults: 0.
dropout_layer (dict, optional): The dropout_layer used before output.
Defaults: dict(type='DropPath', drop_prob=0.).
init_cfg (dict, optional): The extra config for initialization.
Default: None.
"""
def __init__(
self,
embed_dims,
num_heads,
window_size,
shift_size=0,
qkv_bias=True,
qk_scale=None,
attn_drop_rate=0,
proj_drop_rate=0,
dropout_layer=dict(type="DropPath", drop_prob=0.0),
init_cfg=None,
):
super().__init__()
self.window_size = window_size
self.shift_size = shift_size
self.h_slices = (
slice(0, -self.window_size),
slice(-self.window_size, -self.shift_size),
slice(-self.shift_size, None),
)
self.w_slices = (
slice(0, -self.window_size),
slice(-self.window_size, -self.shift_size),
slice(-self.shift_size, None),
)
assert 0 <= self.shift_size < self.window_size
self.w_msa = WindowMSA(
embed_dims=embed_dims,
num_heads=num_heads,
window_size=to_2tuple(window_size),
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop_rate=attn_drop_rate,
proj_drop_rate=proj_drop_rate,
init_cfg=None,
)
self.drop = build_dropout(dropout_layer)
def forward(self, query, hw_shape):
B, L, C = query.shape
H, W = hw_shape
assert L == H * W, "input feature has wrong size"
query = query.view(-1, H, W, C)
# pad feature maps to multiples of window size
pad_r = (self.window_size - W % self.window_size) % self.window_size
pad_b = (self.window_size - H % self.window_size) % self.window_size
query = F.pad(query, (0, 0, 0, pad_r, 0, pad_b))
H_pad = H + pad_b
W_pad = W + pad_r
N = self.window_size**2
nW = H_pad * W_pad // N
# cyclic shift
if self.shift_size > 0:
shifted_query = torch.roll(
query, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
)
# calculate attention mask for SW-MSA
img_mask = torch.zeros((1, H_pad, W_pad, 1), device=query.device)
cnt = 0
for h in self.h_slices:
for w in self.w_slices:
img_mask[:, h, w, :] = cnt
cnt += 1
# nW, window_size, window_size, 1
mask_windows = self.window_partition(img_mask, H_pad, W_pad, 1, nW)
mask_windows = mask_windows.view(nW, N)
attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
attn_mask = attn_mask.masked_fill(
attn_mask != 0, float(-100.0)
).masked_fill(attn_mask == 0, float(0.0))
else:
shifted_query = query
attn_mask = None
# nW*B, window_size, window_size, C
query_windows = self.window_partition(shifted_query, H_pad, W_pad, C, nW)
# nW*B, window_size*window_size, C
query_windows = query_windows.view(-1, N, C)
# W-MSA/SW-MSA (nW*B, window_size*window_size, C)
attn_windows = self.w_msa(query_windows, attn_mask, N, C, nW)
# merge windows
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
# B H' W' C
shifted_x = self.window_reverse(attn_windows, H_pad, W_pad, C, nW)
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(
shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
)
else:
x = shifted_x
if pad_r > 0 or pad_b:
x = x[:, :H, :W, :].contiguous()
x = x.view(-1, H * W, C)
x = self.drop(x)
return x
def window_reverse(self, windows, H, W, C, nW):
"""
Args:
windows: (nW*B, window_size, window_size, C)
H (int): Height of image
W (int): Width of image
Returns:
x: (B, H, W, C)
"""
window_size = self.window_size
x = windows.view(
-1, H // window_size, W // window_size, window_size, window_size, C
)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, H, W, C)
return x
def window_partition(self, x, H, W, C, nW):
"""
Args:
x: (B, H, W, C)
Returns:
windows: (nW*B, window_size, window_size, C)
"""
window_size = self.window_size
x = x.view(
-1,
H // window_size,
window_size,
W // window_size,
window_size,
C,
)
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous()
windows = windows.view(-1, window_size, window_size, C)
return windows
class SwinBlock(BaseModule):
""" "
Args:
embed_dims (int): The feature dimension.
num_heads (int): Parallel attention heads.
feedforward_channels (int): The hidden dimension for FFNs.
window_size (int, optional): The local window scale. Default: 7.
shift (bool, optional): whether to shift window or not. Default False.
qkv_bias (bool, optional): enable bias for qkv if True. Default: True.
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Default: None.
drop_rate (float, optional): Dropout rate. Default: 0.
attn_drop_rate (float, optional): Attention dropout rate. Default: 0.
drop_path_rate (float, optional): Stochastic depth rate. Default: 0.
act_cfg (dict, optional): The config dict of activation function.
Default: dict(type='GELU').
norm_cfg (dict, optional): The config dict of normalization.
Default: dict(type='LN').
with_cp (bool, optional): Use checkpoint or not. Using checkpoint
will save some memory while slowing down the training speed.
Default: False.
init_cfg (dict | list | None, optional): The init config.
Default: None.
"""
def __init__(
self,
embed_dims,
num_heads,
feedforward_channels,
window_size=7,
shift=False,
qkv_bias=True,
qk_scale=None,
drop_rate=0.0,
attn_drop_rate=0.0,
drop_path_rate=0.0,
act_cfg=dict(type="GELU"),
norm_cfg=dict(type="LN"),
with_cp=False,
init_cfg=None,
):
super(SwinBlock, self).__init__()
self.init_cfg = init_cfg
self.with_cp = with_cp
self.norm1 = build_norm_layer(norm_cfg, embed_dims)[1]
self.attn = ShiftWindowMSA(
embed_dims=embed_dims,
num_heads=num_heads,
window_size=window_size,
shift_size=window_size // 2 if shift else 0,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop_rate=attn_drop_rate,
proj_drop_rate=drop_rate,
dropout_layer=dict(type="DropPath", drop_prob=drop_path_rate),
init_cfg=None,
)
self.norm2 = build_norm_layer(norm_cfg, embed_dims)[1]
self.ffn = FFN(
embed_dims=embed_dims,
feedforward_channels=feedforward_channels,
num_fcs=2,
ffn_drop=drop_rate,
dropout_layer=dict(type="DropPath", drop_prob=drop_path_rate),
act_cfg=act_cfg,
add_identity=True,
init_cfg=None,
)
def forward(self, x, hw_shape):
def _inner_forward(x):
identity = x
x = self.norm1(x)
x = self.attn(x, hw_shape)
x = x + identity
identity = x
x = self.norm2(x)
x = self.ffn(x, identity=identity)
return x
if self.with_cp and x.requires_grad:
x = cp.checkpoint(_inner_forward, x)
else:
x = _inner_forward(x)
return x
class SwinBlockSequence(BaseModule):
"""Implements one stage in Swin Transformer.
Args:
embed_dims (int): The feature dimension.
num_heads (int): Parallel attention heads.
feedforward_channels (int): The hidden dimension for FFNs.
depth (int): The number of blocks in this stage.
window_size (int, optional): The local window scale. Default: 7.
qkv_bias (bool, optional): enable bias for qkv if True. Default: True.
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Default: None.
drop_rate (float, optional): Dropout rate. Default: 0.
attn_drop_rate (float, optional): Attention dropout rate. Default: 0.
drop_path_rate (float | list[float], optional): Stochastic depth
rate. Default: 0.
downsample (BaseModule | None, optional): The downsample operation
module. Default: None.
act_cfg (dict, optional): The config dict of activation function.
Default: dict(type='GELU').
norm_cfg (dict, optional): The config dict of normalization.
Default: dict(type='LN').
with_cp (bool, optional): Use checkpoint or not. Using checkpoint
will save some memory while slowing down the training speed.
Default: False.
init_cfg (dict | list | None, optional): The init config.
Default: None.
"""
def __init__(
self,
embed_dims,
num_heads,
feedforward_channels,
depth,
window_size=7,
qkv_bias=True,
qk_scale=None,
drop_rate=0.0,
attn_drop_rate=0.0,
drop_path_rate=0.0,
downsample=None,
act_cfg=dict(type="GELU"),
norm_cfg=dict(type="LN"),
with_cp=False,
init_cfg=None,
):
super().__init__()
if isinstance(drop_path_rate, list):
drop_path_rates = drop_path_rate
assert len(drop_path_rates) == depth
else:
drop_path_rates = [deepcopy(drop_path_rate) for _ in range(depth)]
self.blocks = ModuleList()
for i in range(depth):
block = SwinBlock(
embed_dims=embed_dims,
num_heads=num_heads,
feedforward_channels=feedforward_channels,
window_size=window_size,
shift=False if i % 2 == 0 else True,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
drop_path_rate=drop_path_rates[i],
act_cfg=act_cfg,
norm_cfg=norm_cfg,
with_cp=with_cp,
init_cfg=None,
)
self.blocks.append(block)
self.downsample = downsample
def forward(self, x, hw_shape):
for block in self.blocks:
x = block(x, hw_shape)
if self.downsample:
x_down, down_hw_shape = self.downsample(x, hw_shape)
return x_down, down_hw_shape, x, hw_shape
else:
return x, hw_shape, x, hw_shape
class SwinTransformer(BaseModule):
"""Swin Transformer
A PyTorch implement of : `Swin Transformer:
Hierarchical Vision Transformer using Shifted Windows` -
https://arxiv.org/abs/2103.14030
Inspiration from
https://github.com/microsoft/Swin-Transformer
Args:
pretrain_img_size (int | tuple[int]): The size of input image when
pretrain. Defaults: 224.
in_channels (int): The num of input channels.
Defaults: 3.
embed_dims (int): The feature dimension. Default: 96.
patch_size (int | tuple[int]): Patch size. Default: 4.
window_size (int): Window size. Default: 7.
mlp_ratio (int): Ratio of mlp hidden dim to embedding dim.
Default: 4.
depths (tuple[int]): Depths of each Swin Transformer stage.
Default: (2, 2, 6, 2).
num_heads (tuple[int]): Parallel attention heads of each Swin
Transformer stage. Default: (3, 6, 12, 24).
strides (tuple[int]): The patch merging or patch embedding stride of
each Swin Transformer stage. (In swin, we set kernel size equal to
stride.) Default: (4, 2, 2, 2).
out_indices (tuple[int]): Output from which stages.
Default: (0, 1, 2, 3).
qkv_bias (bool, optional): If True, add a learnable bias to query, key,
value. Default: True
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Default: None.
patch_norm (bool): If add a norm layer for patch embed and patch
merging. Default: True.
drop_rate (float): Dropout rate. Defaults: 0.
attn_drop_rate (float): Attention dropout rate. Default: 0.
drop_path_rate (float): Stochastic depth rate. Defaults: 0.1.
use_abs_pos_embed (bool): If True, add absolute position embedding to
the patch embedding. Defaults: False.
act_cfg (dict): Config dict for activation layer.
Default: dict(type='LN').
norm_cfg (dict): Config dict for normalization layer at
output of backone. Defaults: dict(type='LN').
with_cp (bool, optional): Use checkpoint or not. Using checkpoint
will save some memory while slowing down the training speed.
Default: False.
pretrained (str, optional): model pretrained path. Default: None.
convert_weights (bool): The flag indicates whether the
pre-trained model is from the original repo. We may need
to convert some keys to make it compatible.
Default: False.
frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
-1 means not freezing any parameters.
init_cfg (dict, optional): The Config for initialization.
Defaults to None.
"""
def __init__(
self,
pretrain_img_size=224,
in_channels=3,
embed_dims=96,
patch_size=4,
window_size=7,
mlp_ratio=4,
depths=(2, 2, 6, 2),
num_heads=(3, 6, 12, 24),
strides=(4, 2, 2, 2),
out_indices=(0, 1, 2, 3),
qkv_bias=True,
qk_scale=None,
patch_norm=True,
drop_rate=0.0,
attn_drop_rate=0.0,
drop_path_rate=0.1,
use_abs_pos_embed=False,
act_cfg=dict(type="GELU"),
norm_cfg=dict(type="LN"),
with_cp=False,
pretrained=None,
convert_weights=False,
frozen_stages=-1,
init_cfg=None,
# NOTE: This is my modification based on SOLIDER
semantic_weight=0.5,
freeze_semantic_embedding=False,
):
self.convert_weights = convert_weights
self.frozen_stages = frozen_stages
if isinstance(pretrain_img_size, int):
pretrain_img_size = to_2tuple(pretrain_img_size)
elif isinstance(pretrain_img_size, tuple):
if len(pretrain_img_size) == 1:
pretrain_img_size = to_2tuple(pretrain_img_size[0])
assert len(pretrain_img_size) == 2, (
f"The size of image should have length 1 or 2, "
f"but got {len(pretrain_img_size)}"
)
assert not (
init_cfg and pretrained
), "init_cfg and pretrained cannot be specified at the same time"
if isinstance(pretrained, str):
warnings.warn(
"DeprecationWarning: pretrained is deprecated, "
'please use "init_cfg" instead'
)
self.init_cfg = dict(type="Pretrained", checkpoint=pretrained)
elif pretrained is None:
self.init_cfg = init_cfg
else:
raise TypeError("pretrained must be a str or None")
super(SwinTransformer, self).__init__()
num_layers = len(depths)
self.out_indices = out_indices
self.use_abs_pos_embed = use_abs_pos_embed
assert strides[0] == patch_size, "Use non-overlapping patch embed."
self.patch_embed = PatchEmbed(
in_channels=in_channels,
embed_dims=embed_dims,
conv_type="Conv2d",
kernel_size=patch_size,
stride=strides[0],
norm_cfg=norm_cfg if patch_norm else None,
init_cfg=None,
)
if self.use_abs_pos_embed:
patch_row = pretrain_img_size[0] // patch_size
patch_col = pretrain_img_size[1] // patch_size
num_patches = patch_row * patch_col
self.absolute_pos_embed = nn.Parameter(
torch.zeros((1, num_patches, embed_dims))
)
self.drop_after_pos = nn.Dropout(p=drop_rate)
# set stochastic depth decay rule
total_depth = sum(depths)
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, total_depth)]
self.stages = ModuleList()
in_channels = embed_dims
for i in range(num_layers):
if i < num_layers - 1:
downsample = PatchMerging(
in_channels=in_channels,
out_channels=2 * in_channels,
stride=strides[i + 1],
norm_cfg=norm_cfg if patch_norm else None,
init_cfg=None,
)
else:
downsample = None
stage = SwinBlockSequence(
embed_dims=in_channels,
num_heads=num_heads[i],
feedforward_channels=mlp_ratio * in_channels,
depth=depths[i],
window_size=window_size,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
drop_path_rate=dpr[sum(depths[:i]) : sum(depths[: i + 1])],
downsample=downsample,
act_cfg=act_cfg,
norm_cfg=norm_cfg,
with_cp=with_cp,
init_cfg=None,
)
self.stages.append(stage)
if downsample:
in_channels = downsample.out_channels
self.num_features = [int(embed_dims * 2**i) for i in range(num_layers)]
# Add a norm layer for each output
for i in out_indices:
layer = build_norm_layer(norm_cfg, self.num_features[i])[1]
layer_name = f"norm{i}"
self.add_module(layer_name, layer)
# self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.avgpool = nn.AdaptiveAvgPool1d(1)
# semantic embedding
self.semantic_weight = semantic_weight
self.freeze_semantic_embedding = freeze_semantic_embedding
if self.semantic_weight >= 0:
self.semantic_embed_w = ModuleList()
self.semantic_embed_b = ModuleList()
for i in range(len(depths)):
if i >= len(depths) - 1:
i = len(depths) - 2
semantic_embed_w = nn.Linear(2, self.num_features[i + 1])
semantic_embed_b = nn.Linear(2, self.num_features[i + 1])
# TODO: Test with semantic embed unfreeze
if self.freeze_semantic_embedding:
for param in semantic_embed_w.parameters():
param.requires_grad = False
for param in semantic_embed_b.parameters():
param.requires_grad = False
trunc_normal_init(semantic_embed_w, std=0.02, bias=0.0)
trunc_normal_init(semantic_embed_b, std=0.02, bias=0.0)
self.semantic_embed_w.append(semantic_embed_w)
self.semantic_embed_b.append(semantic_embed_b)
self.softplus = nn.Softplus()
def train(self, mode=True):
"""Convert the model into training mode while keep layers freezed."""
super(SwinTransformer, self).train(mode)
self._freeze_stages()
def _freeze_stages(self):
if self.frozen_stages >= 0:
self.patch_embed.eval()
for param in self.patch_embed.parameters():
param.requires_grad = False
if self.use_abs_pos_embed:
self.absolute_pos_embed.requires_grad = False
self.drop_after_pos.eval()
for i in range(1, self.frozen_stages + 1):
if (i - 1) in self.out_indices:
norm_layer = getattr(self, f"norm{i-1}")
norm_layer.eval()
for param in norm_layer.parameters():
param.requires_grad = False
m = self.stages[i - 1]
m.eval()
for param in m.parameters():
param.requires_grad = False
def init_weights(self, pretrained=None):
logger = logging.getLogger("loading parameters.")
if pretrained is None:
logger.warn(
f"No pre-trained weights for "
f"{self.__class__.__name__}, "
f"training start from scratch"
)
if self.use_abs_pos_embed:
trunc_normal_(self.absolute_pos_embed, std=0.02)
for m in self.modules():
if isinstance(m, nn.Linear):
trunc_normal_init(m, std=0.02, bias=0.0)
elif isinstance(m, nn.LayerNorm):
constant_init(m.bias, 0)
constant_init(m.weight, 1.0)
else:
ckpt = torch.load(pretrained, map_location="cpu")
if "teacher" in ckpt:
ckpt = ckpt["teacher"]
if "state_dict" in ckpt:
_state_dict = ckpt["state_dict"]
elif "model" in ckpt:
_state_dict = ckpt["model"]
else:
_state_dict = ckpt
if self.convert_weights:
# supported loading weight from original repo,
_state_dict = swin_converter(_state_dict)
state_dict = OrderedDict()
for k, v in _state_dict.items():
if k.startswith("backbone."):
state_dict[k[9:]] = v
# strip prefix of state_dict
if list(state_dict.keys())[0].startswith("module."):
state_dict = {k[7:]: v for k, v in state_dict.items()}
# reshape absolute position embedding
if state_dict.get("absolute_pos_embed") is not None:
absolute_pos_embed = state_dict["absolute_pos_embed"]
N1, L, C1 = absolute_pos_embed.size()
N2, C2, H, W = self.absolute_pos_embed.size()
if N1 != N2 or C1 != C2 or L != H * W:
logger.warning("Error in loading absolute_pos_embed, pass")
else:
state_dict["absolute_pos_embed"] = (
absolute_pos_embed.view(N2, H, W, C2)
.permute(0, 3, 1, 2)
.contiguous()
)
# interpolate position bias table if needed
relative_position_bias_table_keys = [
k for k in state_dict.keys() if "relative_position_bias_table" in k
]
for table_key in relative_position_bias_table_keys:
table_pretrained = state_dict[table_key]
table_current = self.state_dict()[table_key]
L1, nH1 = table_pretrained.size()
L2, nH2 = table_current.size()
if nH1 != nH2:
logger.warning(f"Error in loading {table_key}, pass")
elif L1 != L2:
S1 = int(L1**0.5)
S2 = int(L2**0.5)
table_pretrained_resized = F.interpolate(
table_pretrained.permute(1, 0).reshape(1, nH1, S1, S1),
size=(S2, S2),
mode="bicubic",
)
state_dict[table_key] = (
table_pretrained_resized.view(nH2, L2)
.permute(1, 0)
.contiguous()
)
res = self.load_state_dict(state_dict, False)
print("unloaded parameters:", res)
def forward(self, x, semantic_weight=None):
if self.semantic_weight >= 0 and semantic_weight == None:
w = torch.ones(x.shape[0], 1) * self.semantic_weight
w = torch.cat([w, 1 - w], axis=-1)
semantic_weight = w.to(x.device)
x, hw_shape = self.patch_embed(x)
if self.use_abs_pos_embed:
x = x + self.absolute_pos_embed
x = self.drop_after_pos(x)
outs = []
for i, stage in enumerate(self.stages):
x, hw_shape, out, out_hw_shape = stage(x, hw_shape)
if self.semantic_weight >= 0:
sw = self.semantic_embed_w[i](semantic_weight).unsqueeze(1)
sb = self.semantic_embed_b[i](semantic_weight).unsqueeze(1)
x = x * self.softplus(sw) + sb
if i in self.out_indices:
norm_layer = getattr(self, f"norm{i}")
out = norm_layer(out)
# out = (
# out.view(-1, out_hw_shape[0], out_hw_shape[1], self.num_features[i])
# .permute(0, 3, 1, 2)
# .contiguous()
# )
outs.append(out)
x = outs[-1]
x_cls = self.avgpool(x.transpose(1, 2)) # B C 1
x = torch.cat([x_cls.transpose(1, 2), x], dim=1)
return x
def swin_base_patch4_window7_224(
img_size=224, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.0, **kwargs
):
model = SwinTransformer(
pretrain_img_size=img_size,
patch_size=4,
window_size=7,
embed_dims=128,
depths=(2, 2, 18, 2),
num_heads=(4, 8, 16, 32),
drop_path_rate=drop_path_rate,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
**kwargs,
)
return model
def swin_small_patch4_window7_224(
img_size=224, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.0, **kwargs
):
model = SwinTransformer(
pretrain_img_size=img_size,
patch_size=4,
window_size=7,
embed_dims=96,
depths=(2, 2, 18, 2),
num_heads=(3, 6, 12, 24),
drop_path_rate=drop_path_rate,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
**kwargs,
)
return model
def swin_tiny_patch4_window7_224(
img_size=224, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.0, **kwargs
):
model = SwinTransformer(
pretrain_img_size=img_size,
patch_size=4,
window_size=7,
embed_dims=96,
depths=(2, 2, 6, 2),
num_heads=(3, 6, 12, 24),
drop_path_rate=drop_path_rate,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
**kwargs,
)
return model
def build_solider(cfg: dict) -> SwinTransformer:
name = cfg["name"]
img_size = cfg["img_size"]
# drop_path_rate = cfg["drop_path_rate"]\
# TODO: Test with drop_path_rate = 0.0
drop_path_rate = 0.1
# drop_rate = cfg["drop_rate"]
drop_rate = 0.0
# attn_drop_rate = cfg["attn_drop_rate"]
attn_drop_rate = 0.0
pretrained = cfg["pretrained"]
# convert_weights = cfg["convert_weights"]
convert_weights = False
semantic_weight = cfg["semantic_weight"]
if name == "swin_tiny_patch4_window7_224":
model = swin_tiny_patch4_window7_224(
img_size=img_size,
drop_path_rate=drop_path_rate,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
pretrained=pretrained,
convert_weights=convert_weights,
semantic_weight=semantic_weight,
)
elif name == "swin_small_patch4_window7_224":
model = swin_small_patch4_window7_224(
img_size=img_size,
drop_path_rate=drop_path_rate,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
pretrained=pretrained,
convert_weights=convert_weights,
semantic_weight=semantic_weight,
)
elif name == "swin_base_patch4_window7_224":
model = swin_base_patch4_window7_224(
img_size=img_size,
drop_path_rate=drop_path_rate,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
pretrained=pretrained,
convert_weights=convert_weights,
semantic_weight=semantic_weight,
)
else:
raise RuntimeError(f"Not support model name: {name}")
if pretrained != "":
if os.path.exists(pretrained):
model.init_weights(pretrained)
else:
warnings.warn(f"pretrained: {pretrained} not exists")
return model
# BACKBONE_NAME2WIDTH = {
# "swin_tiny_patch4_window7_224": 768,
# "swin_small_patch4_window7_224": 768,
# "swin_base_patch4_window7_224": 1024,
# "solider_tiny": 768,
# "solider_small": 768,
# "solider_base": 1024,
# }
SOLIDER_BASE_MODEL_CONFIG_PARAMETERS = {
"pretrain_img_size": [224, 224],
"in_channels": 3,
"embed_dims": 128,
"patch_size": 4,
"window_size": 7,
"mlp_ratio": 4,
"depths": (2, 2, 18, 2),
"num_heads": (4, 8, 16, 32),
"strides": (4, 2, 2, 2),
"out_indices": (0, 1, 2, 3),
"qkv_bias": True,
"qk_scale": None,
"patch_norm": True,
"drop_rate": 0.0,
"attn_drop_rate": 0.0,
"drop_path_rate": 0.0,
"use_abs_pos_embed": False,
"act_cfg": dict(type="GELU"),
"norm_cfg": dict(type="LN"),
"with_cp": False,
"pretrained": None,
"convert_weights": False,
"frozen_stages": -1,
"init_cfg": None,
"semantic_weight": 0.5,
"name": "solider_base",
}
SOLIDER_SMALL_MODEL_CONFIG_PARAMETERS = {
"pretrain_img_size": [224, 224],
"in_channels": 3,
"embed_dims": 96,
"patch_size": 4,
"window_size": 7,
"mlp_ratio": 4,
"depths": (2, 2, 18, 2),
"num_heads": (3, 6, 12, 24),
"strides": (4, 2, 2, 2),
"out_indices": (0, 1, 2, 3),
"qkv_bias": True,
"qk_scale": None,
"patch_norm": True,
"drop_rate": 0.0,
"attn_drop_rate": 0.0,
"drop_path_rate": 0.0,
"use_abs_pos_embed": False,
"act_cfg": dict(type="GELU"),
"norm_cfg": dict(type="LN"),
"with_cp": False,
"pretrained": None,
"convert_weights": False,
"frozen_stages": -1,
"init_cfg": None,
"semantic_weight": 0.5,
"name": "solider_small",
}
SOLIDER_TINY_MODEL_CONFIG_PARAMETERS = {
"pretrain_img_size": [224, 224],
"in_channels": 3,
"embed_dims": 96,
"patch_size": 4,
"window_size": 7,
"mlp_ratio": 4,
"depths": (2, 2, 6, 2),
"num_heads": (3, 6, 12, 24),
"strides": (4, 2, 2, 2),
"out_indices": (0, 1, 2, 3),
"qkv_bias": True,
"qk_scale": None,
"patch_norm": True,
"drop_rate": 0.0,
"attn_drop_rate": 0.0,
"drop_path_rate": 0.0,
"use_abs_pos_embed": False,
"act_cfg": dict(type="GELU"),
"norm_cfg": dict(type="LN"),
"with_cp": False,
"pretrained": None,
"convert_weights": False,
"frozen_stages": -1,
"init_cfg": None,
"semantic_weight": 0.5,
"name": "solider_tiny",
}
SOLIDER_BASE_CONFIG = SOLIDERConfig(**SOLIDER_BASE_MODEL_CONFIG_PARAMETERS)
SOLIDER_SMALL_CONFIG = SOLIDERConfig(**SOLIDER_SMALL_MODEL_CONFIG_PARAMETERS)
SOLIDER_TINY_CONFIG = SOLIDERConfig(**SOLIDER_TINY_MODEL_CONFIG_PARAMETERS)
def build_solider_vision_encoder(weight_path, name="swin_small_patch4_window7_224"):
vision_width = BACKBONE_NAME2WIDTH[name]
return (
build_solider(
{
"name": name,
"img_size": [384, 128],
"pretrained": weight_path,
"semantic_weight": 0.5,
}
),
vision_width,
)
class SOLIDERModel(PreTrainedModel):
config_class = SOLIDERConfig
base_model_prefix = "solider"
def __init__(self, config: SOLIDERConfig):
super().__init__(config)
self.solider = SwinTransformer(
pretrain_img_size=config.pretrain_img_size,
embed_dims=config.embed_dims,
patch_size=config.patch_size,
window_size=config.window_size,
mlp_ratio=config.mlp_ratio,
depths=config.depths,
num_heads=config.num_heads,
strides=config.strides,
out_indices=config.out_indices,
qkv_bias=config.qkv_bias,
qk_scale=config.qk_scale,
patch_norm=config.patch_norm,
drop_rate=config.drop_rate,
attn_drop_rate=config.attn_drop_rate,
drop_path_rate=config.drop_path_rate,
use_abs_pos_embed=config.use_abs_pos_embed,
act_cfg=config.act_cfg,
norm_cfg=config.norm_cfg,
with_cp=config.with_cp,
pretrained=config.pretrained,
convert_weights=config.convert_weights,
frozen_stages=config.frozen_stages,
init_cfg=config.init_cfg,
semantic_weight=config.semantic_weight,
)
self.solider_name = config.name
self.vision_width = BACKBONE_NAME2WIDTH[self.solider_name]
self.hidden_size = self.vision_width
self.config = config
# self.init_weights()
def forward(self, x):
return self.solider(x, None)
# NOTE: Currently not used!
class SoliderEncoder(SwinTransformer):
options = [
"swin_tiny_patch4_window7_224",
"swin_small_patch4_window7_224",
"swin_base_patch4_window7_224",
]
@classmethod
def from_config(cls, cfg, from_pretrained=None):
name = cfg.get("name", "swin_small_patch4_window7_224")
img_size = cfg.get("img_size", [384, 128])
drop_path_rate = cfg.get("drop_path_rate", 0.1)
drop_rate = cfg.get("drop_rate", 0.0)
attn_drop_rate = cfg.get("attn_drop_rate", 0.0)
pretrained = cfg.get("pretrained", None)
convert_weights = cfg.get("convert_weights", False)
semantic_weight = cfg.get("semantic_weight", 0.2)
if name == "swin_tiny_patch4_window7_224" or name == "tiny":
model = swin_tiny_patch4_window7_224(
img_size=img_size,
drop_path_rate=drop_path_rate,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
pretrained=pretrained,
convert_weights=convert_weights,
semantic_weight=semantic_weight,
)
elif name == "swin_small_patch4_window7_224" or name == "small":
model = swin_small_patch4_window7_224(
img_size=img_size,
drop_path_rate=drop_path_rate,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
pretrained=pretrained,
convert_weights=convert_weights,
semantic_weight=semantic_weight,
)
elif name == "swin_base_patch4_window7_224" or name == "base":
model = swin_base_patch4_window7_224(
img_size=img_size,
drop_path_rate=drop_path_rate,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
pretrained=pretrained,
convert_weights=convert_weights,
semantic_weight=semantic_weight,
)
model.vision_width = BACKBONE_NAME2WIDTH[name]
if from_pretrained is not None:
print("begin load pretrained model solider")
state_dict_vision_encoder = torch.load(from_pretrained, map_location="cpu")
msg = model.load_state_dict(state_dict_vision_encoder)
print(msg)
model.config = cfg
return model
def forward_features(self, x):
return SwinTransformer.forward(self, x, None)
|