File size: 70,627 Bytes
0b06187 | 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 | import math
import copy
import torch
from torch import nn, einsum
import torch.nn.functional as F
from functools import partial
from pathlib import Path
from torch.optim import Adam
from torch.cuda.amp import autocast, GradScaler
from tqdm import tqdm
from einops import rearrange
from einops_exts import check_shape, rearrange_many
from rotary_embedding_torch import RotaryEmbedding
from ddpm.text import tokenize, bert_embed, BERT_MODEL_DIM
from torch.utils.data import DataLoader
# from vq_gan_3d.model.vqgan import VQGAN
from collections import defaultdict
# 检查x是否不为None
def exists(x):
return x is not None
# 无操作函数,用作占位符
def noop(*args, **kwargs):
pass
# 判断数字n是否为奇数
def is_odd(n):
return (n % 2) == 1
# 如果val存在,则返回val;否则返回d()(如果d是可调用的)或d
def default(val, d):
if exists(val):
return val
return d() if callable(d) else d
# 创建一个无限生成器,循环遍历数据加载器dl
def cycle(dl):
while True:
for data in dl:
yield data
# 将num分成大小为divisor的组,处理余数
def num_to_groups(num, divisor):
groups = num // divisor
remainder = num % divisor
arr = [divisor] * groups
if remainder > 0:
arr.append(remainder)
return arr
# 根据概率prob生成一个掩码张量
def prob_mask_like(shape, prob, device):
if prob == 1:
return torch.ones(shape, device=device, dtype=torch.bool)
elif prob == 0:
return torch.zeros(shape, device=device, dtype=torch.bool)
else:
return torch.zeros(shape, device=device).float().uniform_(0, 1) < prob
# 检查x是否为字符串列表或元组
def is_list_str(x):
if not isinstance(x, (list, tuple)):
return False
return all([type(el) == str for el in x])
class RelativePositionBias(nn.Module):
def __init__(
self,
heads=8,
num_buckets=32,
max_distance=128
):
super().__init__()
self.num_buckets = num_buckets
self.max_distance = max_distance
self.relative_attention_bias = nn.Embedding(num_buckets, heads)
@staticmethod
def _relative_position_bucket(relative_position, num_buckets=32, max_distance=128):
ret = 0
n = -relative_position
num_buckets //= 2
ret += (n < 0).long() * num_buckets
n = torch.abs(n)
max_exact = num_buckets // 2
is_small = n < max_exact
val_if_large = max_exact + (
torch.log(n.float() / max_exact) / math.log(max_distance /
max_exact) * (num_buckets - max_exact)
).long()
val_if_large = torch.min(
val_if_large, torch.full_like(val_if_large, num_buckets - 1))
ret += torch.where(is_small, n, val_if_large)
return ret
def forward(self, n, device):
q_pos = torch.arange(n, dtype=torch.long, device=device)
k_pos = torch.arange(n, dtype=torch.long, device=device)
rel_pos = rearrange(k_pos, 'j -> 1 j') - rearrange(q_pos, 'i -> i 1')
rp_bucket = self._relative_position_bucket(
rel_pos, num_buckets=self.num_buckets, max_distance=self.max_distance)
values = self.relative_attention_bias(rp_bucket)
return rearrange(values, 'i j h -> h i j')
class EMA():
def __init__(self, beta):
super().__init__()
self.beta = beta
def update_model_average(self, ma_model, current_model):
for current_params, ma_params in zip(current_model.parameters(), ma_model.parameters()):
old_weight, up_weight = ma_params.data, current_params.data
ma_params.data = self.update_average(old_weight, up_weight)
def update_average(self, old, new):
if old is None:
return new
return old * self.beta + (1 - self.beta) * new
class Residual(nn.Module):
def __init__(self, fn):
super().__init__()
self.fn = fn
def forward(self, x, *args, **kwargs):
return self.fn(x, *args, **kwargs) + x
class SinusoidalPosEmb(nn.Module):
def __init__(self, dim):
super().__init__()
self.dim = dim
def forward(self, x):
device = x.device
half_dim = self.dim // 2
emb = math.log(10000) / (half_dim - 1)
emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
emb = x[:, None] * emb[None, :]
emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
return emb
def Upsample(dim):
return nn.ConvTranspose3d(dim, dim, (1, 4, 4), (1, 2, 2), (0, 1, 1))
def Downsample(dim):
return nn.Conv3d(dim, dim, (1, 4, 4), (1, 2, 2), (0, 1, 1))
class LayerNorm(nn.Module):
def __init__(self, dim, eps=1e-5):
super().__init__()
self.eps = eps
self.gamma = nn.Parameter(torch.ones(1, dim, 1, 1, 1))
def forward(self, x):
var = torch.var(x, dim=1, unbiased=False, keepdim=True)
mean = torch.mean(x, dim=1, keepdim=True)
return (x - mean) / (var + self.eps).sqrt() * self.gamma
class PreNorm(nn.Module):
def __init__(self, dim, fn):
super().__init__()
self.fn = fn
self.norm = LayerNorm(dim)
def forward(self, x, **kwargs):
x = self.norm(x)
return self.fn(x, **kwargs)
class Block(nn.Module):
def __init__(self, dim, dim_out, groups=8):
super().__init__()
self.proj = nn.Conv3d(dim, dim_out, (1, 3, 3), padding=(0, 1, 1))
self.norm = nn.GroupNorm(groups, dim_out)
self.act = nn.SiLU()
def forward(self, x, scale_shift=None):
x = self.proj(x)
x = self.norm(x)
if exists(scale_shift):
scale, shift = scale_shift
x = x * (scale + 1) + shift
return self.act(x)
class ResnetBlock(nn.Module):
def __init__(self, dim, dim_out, *, time_emb_dim=None, groups=8):
super().__init__()
self.mlp = nn.Sequential(
nn.SiLU(),
nn.Linear(time_emb_dim, dim_out * 2)
) if exists(time_emb_dim) else None
self.block1 = Block(dim, dim_out, groups=groups)
self.block2 = Block(dim_out, dim_out, groups=groups)
self.res_conv = nn.Conv3d(
dim, dim_out, 1) if dim != dim_out else nn.Identity()
def forward(self, x, time_emb=None):
scale_shift = None
if exists(self.mlp):
assert exists(time_emb), 'time emb must be passed in'
time_emb = self.mlp(time_emb)
time_emb = rearrange(time_emb, 'b c -> b c 1 1 1')
scale_shift = time_emb.chunk(2, dim=1)
h = self.block1(x, scale_shift=scale_shift)
h = self.block2(h)
return h + self.res_conv(x)
class SpatialLinearAttention(nn.Module):
def __init__(self, dim, heads=4, dim_head=32):
super().__init__()
self.scale = dim_head ** -0.5
self.heads = heads
hidden_dim = dim_head * heads
self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
self.to_out = nn.Conv2d(hidden_dim, dim, 1)
def forward(self, x):
b, c, f, h, w = x.shape
x = rearrange(x, 'b c f h w -> (b f) c h w')
qkv = self.to_qkv(x).chunk(3, dim=1)
q, k, v = rearrange_many(
qkv, 'b (h c) x y -> b h c (x y)', h=self.heads)
q = q.softmax(dim=-2)
k = k.softmax(dim=-1)
q = q * self.scale
context = torch.einsum('b h d n, b h e n -> b h d e', k, v)
out = torch.einsum('b h d e, b h d n -> b h e n', context, q)
out = rearrange(out, 'b h c (x y) -> b (h c) x y',
h=self.heads, x=h, y=w)
out = self.to_out(out)
return rearrange(out, '(b f) c h w -> b c f h w', b=b)
class EinopsToAndFrom(nn.Module):
def __init__(self, from_einops, to_einops, fn):
super().__init__()
self.from_einops = from_einops
self.to_einops = to_einops
self.fn = fn
def forward(self, x, **kwargs):
shape = x.shape
reconstitute_kwargs = dict(
tuple(zip(self.from_einops.split(' '), shape)))
x = rearrange(x, f'{self.from_einops} -> {self.to_einops}')
x = self.fn(x, **kwargs)
x = rearrange(
x, f'{self.to_einops} -> {self.from_einops}', **reconstitute_kwargs)
return x
class Attention(nn.Module):
def __init__(
self,
dim,
heads=4,
dim_head=32,
rotary_emb=None
):
super().__init__()
self.scale = dim_head ** -0.5
self.heads = heads
hidden_dim = dim_head * heads
self.rotary_emb = rotary_emb
self.to_qkv = nn.Linear(dim, hidden_dim * 3, bias=False)
self.to_out = nn.Linear(hidden_dim, dim, bias=False)
def forward(
self,
x,
pos_bias=None,
focus_present_mask=None
):
n, device = x.shape[-2], x.device
qkv = self.to_qkv(x).chunk(3, dim=-1)
if exists(focus_present_mask) and focus_present_mask.all():
values = qkv[-1]
return self.to_out(values)
q, k, v = rearrange_many(qkv, '... n (h d) -> ... h n d', h=self.heads)
q = q * self.scale
if exists(self.rotary_emb):
q = self.rotary_emb.rotate_queries_or_keys(q)
k = self.rotary_emb.rotate_queries_or_keys(k)
sim = einsum('... h i d, ... h j d -> ... h i j', q, k)
if exists(pos_bias):
sim = sim + pos_bias
if exists(focus_present_mask) and not (~focus_present_mask).all():
attend_all_mask = torch.ones(
(n, n), device=device, dtype=torch.bool)
attend_self_mask = torch.eye(n, device=device, dtype=torch.bool)
mask = torch.where(
rearrange(focus_present_mask, 'b -> b 1 1 1 1'),
rearrange(attend_self_mask, 'i j -> 1 1 1 i j'),
rearrange(attend_all_mask, 'i j -> 1 1 1 i j'),
)
sim = sim.masked_fill(~mask, -torch.finfo(sim.dtype).max)
sim = sim - sim.amax(dim=-1, keepdim=True).detach()
attn = sim.softmax(dim=-1)
out = einsum('... h i j, ... h j d -> ... h i d', attn, v)
out = rearrange(out, '... h n d -> ... n (h d)')
return self.to_out(out)
class Unet3D(nn.Module):
def __init__(
self,
dim,
cond_dim=None,
out_dim=None,
dim_mults=(1, 2, 4, 8),
channels=3,
attn_heads=8,
attn_dim_head=32,
use_bert_text_cond=False,
init_dim=None,
init_kernel_size=7,
use_sparse_linear_attn=True,
resnet_groups=8
):
super().__init__()
self.channels = channels
rotary_emb = RotaryEmbedding(min(32, attn_dim_head))
def temporal_attn(dim): return EinopsToAndFrom('b c f h w', 'b (h w) f c', Attention(
dim, heads=attn_heads, dim_head=attn_dim_head, rotary_emb=rotary_emb))
self.time_rel_pos_bias = RelativePositionBias(
heads=attn_heads, max_distance=32)
init_dim = default(init_dim, dim)
assert is_odd(init_kernel_size)
init_padding = init_kernel_size // 2
self.init_conv = nn.Conv3d(channels, init_dim, (1, init_kernel_size,
init_kernel_size), padding=(0, init_padding, init_padding))
self.init_temporal_attn = Residual(
PreNorm(init_dim, temporal_attn(init_dim)))
dims = [init_dim, *map(lambda m: dim * m, dim_mults)]
in_out = list(zip(dims[:-1], dims[1:]))
time_dim = dim * 4
self.time_mlp = nn.Sequential(
SinusoidalPosEmb(dim),
nn.Linear(dim, time_dim),
nn.GELU(),
nn.Linear(time_dim, time_dim)
)
self.has_cond = exists(cond_dim) or use_bert_text_cond
cond_dim = BERT_MODEL_DIM if use_bert_text_cond else cond_dim
self.null_cond_emb = nn.Parameter(
torch.randn(1, cond_dim)) if self.has_cond else None
cond_dim = time_dim + int(cond_dim or 0)
self.downs = nn.ModuleList([])
self.ups = nn.ModuleList([])
num_resolutions = len(in_out)
block_klass = partial(ResnetBlock, groups=resnet_groups)
block_klass_cond = partial(block_klass, time_emb_dim=cond_dim)
for ind, (dim_in, dim_out) in enumerate(in_out):
is_last = ind >= (num_resolutions - 1)
self.downs.append(nn.ModuleList([
block_klass_cond(dim_in, dim_out),
block_klass_cond(dim_out, dim_out),
Residual(PreNorm(dim_out, SpatialLinearAttention(
dim_out, heads=attn_heads))) if use_sparse_linear_attn else nn.Identity(),
Residual(PreNorm(dim_out, temporal_attn(dim_out))),
Downsample(dim_out) if not is_last else nn.Identity()
]))
mid_dim = dims[-1]
self.mid_block1 = block_klass_cond(mid_dim, mid_dim)
spatial_attn = EinopsToAndFrom(
'b c f h w', 'b f (h w) c', Attention(mid_dim, heads=attn_heads))
self.mid_spatial_attn = Residual(PreNorm(mid_dim, spatial_attn))
self.mid_temporal_attn = Residual(
PreNorm(mid_dim, temporal_attn(mid_dim)))
self.mid_block2 = block_klass_cond(mid_dim, mid_dim)
for ind, (dim_in, dim_out) in enumerate(reversed(in_out)):
is_last = ind >= (num_resolutions - 1)
self.ups.append(nn.ModuleList([
block_klass_cond(dim_out * 2, dim_in),
block_klass_cond(dim_in, dim_in),
Residual(PreNorm(dim_in, SpatialLinearAttention(
dim_in, heads=attn_heads))) if use_sparse_linear_attn else nn.Identity(),
Residual(PreNorm(dim_in, temporal_attn(dim_in))),
Upsample(dim_in) if not is_last else nn.Identity()
]))
out_dim = default(out_dim, channels)
self.final_conv = nn.Sequential(
block_klass(dim * 2, dim),
nn.Conv3d(dim, out_dim, 1)
)
def forward_with_cond_scale(
self,
*args,
cond_scale=2.,
**kwargs
):
logits = self.forward(*args, null_cond_prob=0., **kwargs)
if cond_scale == 1 or not self.has_cond:
return logits
null_logits = self.forward(*args, null_cond_prob=1., **kwargs)
return null_logits + (logits - null_logits) * cond_scale
def forward(
self,
x,
time,
cond=None,
null_cond_prob=0.,
focus_present_mask=None,
prob_focus_present=0.
):
if cond is None:
cond = torch.zeros((1, 16))
cond[0, -1] = 1.0
assert not (self.has_cond and not exists(cond)
), 'cond must be passed in if cond_dim specified'
batch, device = x.shape[0], x.device
focus_present_mask = default(focus_present_mask, lambda: prob_mask_like(
(batch,), prob_focus_present, device=device))
time_rel_pos_bias = self.time_rel_pos_bias(x.shape[2], device=x.device)
x = self.init_conv(x)
r = x.clone()
x = self.init_temporal_attn(x, pos_bias=time_rel_pos_bias)
t = self.time_mlp(time) if exists(self.time_mlp) else None
if self.has_cond:
batch, device = x.shape[0], x.device
mask = prob_mask_like((batch,), null_cond_prob, device=device)
cond = cond.to(device)
cond = torch.where(rearrange(mask, 'b -> b 1'),
self.null_cond_emb, cond)
t = torch.cat((t, cond), dim=-1)
h = []
for block1, block2, spatial_attn, temporal_attn, downsample in self.downs:
x = block1(x, t)
x = block2(x, t)
x = spatial_attn(x)
x = temporal_attn(x, pos_bias=time_rel_pos_bias,
focus_present_mask=focus_present_mask)
h.append(x)
x = downsample(x)
x = self.mid_block1(x, t)
x = self.mid_spatial_attn(x)
x = self.mid_temporal_attn(
x, pos_bias=time_rel_pos_bias, focus_present_mask=focus_present_mask)
x = self.mid_block2(x, t)
for block1, block2, spatial_attn, temporal_attn, upsample in self.ups:
x = torch.cat((x, h.pop()), dim=1)
x = block1(x, t)
x = block2(x, t)
x = spatial_attn(x)
x = temporal_attn(x, pos_bias=time_rel_pos_bias,
focus_present_mask=focus_present_mask)
x = upsample(x)
x = torch.cat((x, r), dim=1)
return self.final_conv(x)
def extract(a, t, x_shape):
b, *_ = t.shape
out = a.gather(-1, t)
return out.reshape(b, *((1,) * (len(x_shape) - 1)))
def cosine_beta_schedule(timesteps, s=0.008):
steps = timesteps + 1
x = torch.linspace(0, timesteps, steps, dtype=torch.float64)
alphas_cumprod = torch.cos(
((x / timesteps) + s) / (1 + s) * torch.pi * 0.5) ** 2
alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
return torch.clip(betas, 0, 0.9999)
class GaussianDiffusion_Nolatent(nn.Module):
def __init__(
self,
denoise_fn,
*,
image_size,
num_frames,
text_use_bert_cls=False,
channels=2,
timesteps=1000,
loss_type='l1',
use_dynamic_thres=False,
dynamic_thres_percentile=0.9,
device=None,
use_guide=True,
# vqgan_ckpt=None,
):
super().__init__()
self.channels = channels
self.image_size = image_size
self.num_frames = num_frames
self.denoise_fn = denoise_fn
# if vqgan_ckpt:
# self.vqgan = VQGAN.load_from_checkpoint(vqgan_ckpt).cuda()
# self.vqgan.eval()
# else:
# self.vqgan = None
self.device=device
betas = cosine_beta_schedule(timesteps)
alphas = 1. - betas
alphas_cumprod = torch.cumprod(alphas, axis=0)
alphas_cumprod_prev = F.pad(alphas_cumprod[:-1], (1, 0), value=1.)
timesteps, = betas.shape
self.num_timesteps = int(timesteps)
print("timesteps : ", timesteps)
self.loss_type = loss_type
self.use_guide = use_guide
def register_buffer(name, val): return self.register_buffer(
name, val.to(torch.float32))
register_buffer('betas', betas)
register_buffer('alphas_cumprod', alphas_cumprod)
register_buffer('alphas_cumprod_prev', alphas_cumprod_prev)
register_buffer('sqrt_alphas_cumprod', torch.sqrt(alphas_cumprod))
register_buffer('sqrt_one_minus_alphas_cumprod',
torch.sqrt(1. - alphas_cumprod))
register_buffer('log_one_minus_alphas_cumprod',
torch.log(1. - alphas_cumprod))
register_buffer('sqrt_recip_alphas_cumprod',
torch.sqrt(1. / alphas_cumprod))
register_buffer('sqrt_recipm1_alphas_cumprod',
torch.sqrt(1. / alphas_cumprod - 1))
posterior_variance = betas * \
(1. - alphas_cumprod_prev) / (1. - alphas_cumprod)
register_buffer('posterior_variance', posterior_variance)
register_buffer('posterior_log_variance_clipped',
torch.log(posterior_variance.clamp(min=1e-20)))
register_buffer('posterior_mean_coef1', betas *
torch.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))
register_buffer('posterior_mean_coef2', (1. - alphas_cumprod_prev)
* torch.sqrt(alphas) / (1. - alphas_cumprod))
self.text_use_bert_cls = text_use_bert_cls
self.use_dynamic_thres = use_dynamic_thres
self.dynamic_thres_percentile = dynamic_thres_percentile
# 计算扩散过程中的均值和方差,用于定义 q(x_t|x_0) 分布
# x_start:原始数据 x_0
def q_mean_variance(self, x_start, t):
mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
variance = extract(1. - self.alphas_cumprod, t, x_start.shape)
log_variance = extract(
self.log_one_minus_alphas_cumprod, t, x_start.shape)
return mean, variance, log_variance
# 从带噪声的样本 x_t 和噪声 epsilon 预测 原始数据 x_0
# 输出:预测的原始数据 x_0
def predict_start_from_noise(self, x_t, t, noise):
return (
extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
)
# 计算后验分布 q(x_{t-1}|x_t, x_0) 的均值和方差
def q_posterior(self, x_start, x_t, t):
posterior_mean = (
extract(self.posterior_mean_coef1, t, x_t.shape) * x_start +
extract(self.posterior_mean_coef2, t, x_t.shape) * x_t
)
posterior_variance = extract(self.posterior_variance, t, x_t.shape)
posterior_log_variance_clipped = extract(
self.posterior_log_variance_clipped, t, x_t.shape)
return posterior_mean, posterior_variance, posterior_log_variance_clipped
# 根据去噪函数的预测,计算后验分布 p(x_{t-1}|x_t)
def p_mean_variance(self, x, t, clip_denoised: bool, cond=None, cond_scale=1.):
if isinstance(self.denoise_fn, torch.nn.DataParallel):
noise = self.denoise_fn.module.forward_with_cond_scale(x, t, cond=cond, cond_scale=cond_scale)
else:
noise = self.denoise_fn.forward_with_cond_scale(x, t, cond=cond, cond_scale=cond_scale)
x_recon = self.predict_start_from_noise(
x, t=t, noise=noise)
if clip_denoised:
s = 1.
if self.use_dynamic_thres:
s = torch.quantile(
rearrange(x_recon, 'b ... -> b (...)').abs(),
self.dynamic_thres_percentile,
dim=-1
)
s.clamp_(min=1.)
s = s.view(-1, *((1,) * (x_recon.ndim - 1)))
x_recon = x_recon.clamp(-s, s) / s
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(
x_start=x_recon, x_t=x, t=t)
return model_mean, posterior_variance, posterior_log_variance
# 从后验分布 p(x_{t-1}|x_t) 采样 x_{t-1}
#@torch.inference_mode()
def p_sample_v2(self, x, t, cond=None, cond_scale=1., clip_denoised=True):
b, *_ = x.shape
model_mean, _, model_log_variance = self.p_mean_variance(
x=x, t=t, clip_denoised=clip_denoised, cond=cond, cond_scale=cond_scale)
noise = torch.randn_like(x)
nonzero_mask = (1 - (t == 0).float()).reshape(b,
*((1,) * (len(x.shape) - 1)))
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
@torch.inference_mode()
def p_sample(self, x, t, cond=None, cond_scale=1., clip_denoised=True):
b, *_ = x.shape
model_mean, _, model_log_variance = self.p_mean_variance(
x=x, t=t, clip_denoised=clip_denoised, cond=cond, cond_scale=cond_scale)
noise = torch.randn_like(x)
nonzero_mask = (1 - (t == 0).float()).reshape(b,
*((1,) * (len(x.shape) - 1)))
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
# @torch.inference_mode()
def p_sample_loop_v2(self, shape_image, shape_mask, cond=None, cond_scale=1., device=None, image=None):
b = shape_image[0]
img = torch.randn(shape_image, device=device)
mask = torch.randn(shape_mask, device=device)
input = torch.cat((img, mask), dim=1)
real_img = image
input_guided = input
N = 2
R = 3
B = 1
recurrent = [0] * self.num_timesteps
for i in range(self.num_timesteps):
if i % R == 0:
recurrent[i] = R
i = self.num_timesteps - 1
while i >= 0:
# print(i)
if self.use_guide is not None and i < 250:
# if self.use_guide is not None:
input_with_grad = input_guided.clone().detach().requires_grad_(True)
loss = 0
t = torch.full((b,), i, dtype=torch.long, device=device)
real_noisy_image = self.q_sample(x_start=real_img, t=t)
for _ in range(N):
input_sampled = input_with_grad
input_sampled = self.p_sample_v2(input_sampled, torch.full(
(b,), i, device=device, dtype=torch.long), cond=cond, cond_scale=cond_scale)
# print(torch.split(input_sampled, 1, dim=1).shape)
loss+=F.mse_loss(torch.split(input_sampled, 1, dim=1)[0], real_noisy_image)
loss /= N
loss.backward()
update = torch.clamp(input_with_grad.grad * 10000.0, -1.5, 1.5)
# input_guided= input_guided- update
# print(update[:, 0, :, :, :])
input_guided[:, 0, :, :, :] = input_guided[:, 0, :, :, :] - update[:, 0, :, :, :]
# input_guided[:, 1:6, :, :, :] = input_guided[:, 1:6, :, :, :] - 0.01*update[:, 0, :, :, :]
input_with_grad.grad.zero_()
# else:
# input_guided = input
with torch.no_grad():
# input = self.p_sample_v2(input, torch.full(
# (b,), i, device=device, dtype=torch.long), cond=cond, cond_scale=cond_scale)
input_guided = self.p_sample_v2(input_guided, torch.full(
(b,), i, device=device, dtype=torch.long), cond=cond, cond_scale=cond_scale)
# if i % R == 0 and recurrent[i]!=0 :
# recurrent[i] -= 1
# with torch.no_grad():
# for _ in range(B):
# input_guided=self.q_sample_one_step(input_guided, torch.full((b,), i, device=device, dtype=torch.long))
# i += 1
i -= 1
return input_guided
@torch.inference_mode()
def p_sample_loop(self, shape_image, shape_mask, cond=None, cond_scale=1., device=None, image=None):
b = shape_image[0]
img = torch.randn(shape_image, device=device)
mask = torch.randn(shape_mask, device=device)
input = torch.cat((img, mask), dim=1)
real_img = image
R = 2
recurrent = [0] * self.num_timesteps
for i in range(self.num_timesteps):
if i % R == 0:
recurrent[i] = R
i = self.num_timesteps - 1
while i >= 0:
# print(i)
# if self.use_guide is not None and i < 250:
# if self.use_guide is not None and i > 200:
if self.use_guide is not None:
t = torch.full((b,), i, dtype=torch.long, device=device)
real_noisy_image = self.q_sample(x_start=real_img, t=t)
input[:, 0, :, :, :] = real_noisy_image[:, 0, :, :, :].clone()
# with torch.no_grad():
input = self.p_sample(input, torch.full(
(b,), i, device=device, dtype=torch.long), cond=cond, cond_scale=cond_scale)
# if i % 5 == 0 and recurrent[i]!=0 :
# recurrent[i] -= 1
# with torch.no_grad():
# for _ in range(5):
# input=self.q_sample_one_step(input, torch.full((b,), i, device=device, dtype=torch.long))
# i += 1
i -= 1
return input
@torch.inference_mode()
def p_sample_loop_v4(self, shape_image, shape_mask, cond=None, cond_scale=1., device=None, image=None):
b = shape_image[0]
img = torch.randn(shape_image, device=device)
mask = torch.randn(shape_mask, device=device)
input = torch.cat((img, mask), dim=1)
real_img = image
R = 2
recurrent = [0] * self.num_timesteps
for i in range(self.num_timesteps):
if i % R == 0:
recurrent[i] = R
i = self.num_timesteps - 1
while i >= 0:
print(i)
# if self.use_guide is not None and i < 250:
if self.use_guide is not None and i > 100:
# if self.use_guide is not None:
t = torch.full((b,), i, dtype=torch.long, device=device)
real_noisy_image = self.q_sample(x_start=real_img, t=t)
input[:, 0, :, :, :] = real_noisy_image[:, 0, :, :, :].clone()
# with torch.no_grad():
input = self.p_sample(input, torch.full(
(b,), i, device=device, dtype=torch.long), cond=cond, cond_scale=cond_scale)
else:
input = self.p_sample(input, torch.full(
(b,), i, device=device, dtype=torch.long), cond=cond, cond_scale=cond_scale)
# if i % 5 == 0 and recurrent[i]!=0 :
# recurrent[i] -= 1
# with torch.no_grad():
# for _ in range(5):
# input=self.q_sample_one_step(input, torch.full((b,), i, device=device, dtype=torch.long))
# i += 1
i -= 1
return input
@torch.inference_mode()
def p_sample_loop_v3(self, shape_image, shape_mask, cond=None, cond_scale=1., device=None, image=None):
b = shape_image[0]
img = torch.randn(shape_image, device=device)
mask = torch.randn(shape_mask, device=device)
input = torch.cat((img, mask), dim=1)
i = self.num_timesteps - 1
while i >= 0:
input = self.p_sample(input, torch.full(
(b,), i, device=device, dtype=torch.long), cond=cond, cond_scale=cond_scale)
i -= 1
return input
@torch.inference_mode()
def p_sample_loop_v3_image_only(self, shape_image, shape_mask, cond=None, cond_scale=1., device=None, image=None):
b = shape_image[0]
img = torch.randn(shape_image, device=device)
input = img
i = self.num_timesteps - 1
while i >= 0:
input = self.p_sample(input, torch.full(
(b,), i, device=device, dtype=torch.long), cond=cond, cond_scale=cond_scale)
i -= 1
return input
# for i in reversed(range(0, self.num_timesteps)):
# t = torch.full((b,), i, dtype=torch.long, device=device)
# real_noisy_image = self.q_sample(x_start=real_img, t=t)
# input[:, 0, :, :, :] = real_noisy_image[:, 0, :, :, :].clone()
# with torch.no_grad():
# input = self.p_sample_v2(input, torch.full(
# (b,), i, device=device, dtype=torch.long), cond=cond, cond_scale=cond_scale)
# return input
# @torch.inference_mode()
# def p_sample_loop(self, shape, cond=None, cond_scale=1., device=None):
# b = shape[0]
# img = torch.randn(shape, device=device)
# mask = torch.randn(shape, device=device)
# input = torch.cat((img, mask), dim=1)
# for i in tqdm(reversed(range(0, self.num_timesteps)), desc='sampling loop time step', total=self.num_timesteps):
# input = self.p_sample(input, torch.full(
# (b,), i, device=device, dtype=torch.long), cond=cond, cond_scale=cond_scale)
# return input
# 向原始数据 x_0 添加噪声,生成 x_t
def q_sample(self, x_start, t, noise=None):
noise = default(noise, lambda: torch.randn_like(x_start))
return (
extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
extract(self.sqrt_one_minus_alphas_cumprod,
t, x_start.shape) * noise
)
# 对当前样本 x_{t-1} 添加一步噪声,生成 x_t
# t 是目标时间步
def q_sample_one_step(self, x_prev, t):
beta_t = extract(self.betas, t, x_prev.shape)
alpha_t = 1.0 - beta_t
sqrt_alpha_t = torch.sqrt(alpha_t)
sqrt_beta_t = torch.sqrt(beta_t)
noise = torch.randn_like(x_prev)
x_t = sqrt_alpha_t * x_prev + sqrt_beta_t * noise
return x_t
def p_losses(self, x_start, t, mask_start, cond=None, noise_x=None, noise_m=None, **kwargs):
device = x_start.device
x_start = x_start.to(device=device, dtype=torch.float32)
mask_start = mask_start.to(device=device, dtype=torch.float32)
noise_x = default(noise_x, lambda: torch.randn_like(x_start))
noise_m = default(noise_m, lambda: torch.randn_like(mask_start))
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise_x)
m_noisy = self.q_sample(x_start=mask_start, t=t, noise=noise_m)
input = torch.cat((x_noisy, m_noisy), dim=1)
if is_list_str(cond):
cond = bert_embed(
tokenize(cond), return_cls_repr=self.text_use_bert_cls)
cond = cond.to(device)
recon = self.denoise_fn(**dict(x=input, time=t, cond=cond, **kwargs))
# print(recon.size())
x_recon = recon[:,0,:,:,:]
x_recon = x_recon.unsqueeze(1)
#x_recon = x_recon.squeeze(1)
m_recon = recon[:,1:(recon.size()[1]),:,:,:]
#x_recon, m_recon = torch.split(recon, 1, dim=1)
m_recon = m_recon.squeeze(1)
# noise_x = noise_x.squeeze(1)
noise_m = noise_m.squeeze(1)
# print(noise_x.size())
# print(x_recon.size())
# print(noise_m.size())
# print(m_recon.size())
# print(m_recon.shape)
# print(noise_m.shape)
if self.loss_type == 'l1':
loss = F.l1_loss(noise_x, x_recon) + F.l1_loss(noise_m, m_recon)
elif self.loss_type == 'l2':
loss = F.mse_loss(noise_x, x_recon) + F.mse_loss(noise_m, m_recon)
else:
raise NotImplementedError()
return loss
def p_losses_image_only(self, x_start, t, cond=None, noise_x=None,**kwargs):
device = x_start.device
x_start = x_start.to(device=device, dtype=torch.float32)
noise_x = default(noise_x, lambda: torch.randn_like(x_start))
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise_x)
input = x_noisy
if is_list_str(cond):
cond = bert_embed(
tokenize(cond), return_cls_repr=self.text_use_bert_cls)
cond = cond.to(device)
recon = self.denoise_fn(**dict(x=input, time=t, cond=cond, **kwargs))
x_recon = recon
# x_recon = recon[:,0,:,:,:]
#x_recon = x_recon.squeeze(1)
#x_recon, m_recon = torch.split(recon, 1, dim=1)
# noise_x = noise_x.squeeze(1)
# print(m_recon.shape)
# print(noise_m.shape)
if self.loss_type == 'l1':
loss = F.l1_loss(noise_x, x_recon)
elif self.loss_type == 'l2':
loss = F.mse_loss(noise_x, x_recon)
else:
raise NotImplementedError()
return loss
def forward(self, x, mask, *args, **kwargs):
b, device, img_size, = x.shape[0], x.device, self.image_size
# check_shape(x, 'b c f h w', c=self.channels,
# f=self.num_frames, h=img_size, w=img_size)
t = torch.randint(0, self.num_timesteps, (b,), device=device).long().to(self.device)
return self.p_losses(**dict(x_start=x, t=t, mask_start=mask, *args, **kwargs))
# def forward(self, x, mask, *args, **kwargs):
# b, device, img_size, = x.shape[0], x.device, self.image_size
# # check_shape(x, 'b c f h w', c=self.channels,
# # f=self.num_frames, h=img_size, w=img_size)
# t = torch.randint(0, self.num_timesteps, (b,), device=device).long().to(self.device)
# return self.p_losses_image_only(**dict(x_start=x, t=t, *args, **kwargs))
# @torch.inference_mode()
def p_sample_loop_guidance(self, shape_image, shape_mask, cond=None, cond_scale=1., device=None, image=None, degrade_mask=None, delta=1.5):
if degrade_mask is None: # default is all 1
degrade_mask = torch.ones(shape_image, device=device)
degrade_mask = degrade_mask.to(device=device, dtype=torch.float32)
device = self.betas.device
b = shape_image[0]
init_noise = torch.randn_like(image)
step_noise_list = []
for step in range(self.num_timesteps):
t = torch.full((image.shape[0],), step, device=device, dtype=torch.long)
step_noise = self.q_sample(image, t, noise=init_noise)
step_noise_list.append(step_noise)
img_noisy = torch.stack(step_noise_list) # [T, B, C, D, H, W]
img = torch.randn(shape_image, device=device)
mask = torch.randn(shape_mask, device=device)
pair = torch.cat((img, mask), dim=1)
TCOUNT = 2 # 2
RSTEP = 1 # 10
GRAD_STEP = 3 # 7
print(f'TCOUNT: {TCOUNT}, RSTEP: {RSTEP}, GRAD_STEP: {GRAD_STEP}')
recurrent = [0] * self.num_timesteps
for i in range(self.num_timesteps):
if i % RSTEP == 0:
recurrent[i] = TCOUNT
i = self.num_timesteps - 1
print('degrade_mask_sum_check:', degrade_mask.sum().item())
while i >= 0:
# real_noisy_image = self.q_sample(x_start=image, t=t)
pair[:, :1] = img_noisy[i] * degrade_mask + pair[:, :1] * (1 - degrade_mask)
# pair[:, :1] = img_noisy[i]
pair = pair.clone().detach()
for g in range(GRAD_STEP):
# break
if i > 150:
break
with torch.enable_grad():
pair_with_grad = pair.detach().clone().requires_grad_(True)
t = torch.full((b,), i, device=device, dtype=torch.long)
self.loss_type = 'l1'
loss = self.p_losses_guidance(pair_with_grad, t, init_noise, degrade_mask=degrade_mask)
print('t:', i, 'g', g, 'loss:', loss.item())
grad = torch.autograd.grad(loss, pair_with_grad)[0]
grad[:, :1] = grad[:, :1] * (1 - degrade_mask)
grad_norm = grad.flatten(start_dim=1).norm( dim=1, keepdim=True).unsqueeze(-1).unsqueeze(-1).unsqueeze(-1)
grad = grad / (grad_norm + 1e-8)
pair = pair - grad * delta
t = torch.full((b,), i, device=device, dtype=torch.long)
pair = self.p_sample(pair, t, cond=cond, cond_scale=cond_scale).clone()
# if recurrent[i] > 0 and i % RSTEP == 0:
# recurrent[i] -= 1
# for _ in range(RSTEP):
# with torch.no_grad():
# t = torch.full((b,), i, device=device, dtype=torch.long)
# pair = self.q_sample_one_step(pair, t)
# i += 1
# if i >= self.num_timesteps:
# break
i -= 1
return pair
def p_losses_guidance(self, pair_noisy, t, noise, cond=None, degrade_mask=None, **kwargs):
device = pair_noisy.device
if is_list_str(cond):
cond = bert_embed(
tokenize(cond), return_cls_repr=self.text_use_bert_cls)
cond = cond.to(device)
x_recon = self.denoise_fn(pair_noisy, t, cond=cond, **kwargs)[:, :1]
if degrade_mask is not None:
x_recon = x_recon * degrade_mask
noise = noise * degrade_mask
if self.loss_type == 'l1':
loss = F.l1_loss(noise, x_recon, reduction='sum')
elif self.loss_type == 'l2':
loss = F.mse_loss(noise, x_recon, reduction='sum')
else:
raise NotImplementedError()
if degrade_mask is not None:
loss = loss / degrade_mask.sum()
else:
loss = loss / noise.numel()
return loss
def p_sample_loop_universal_guidance(
self,
shape_image,
shape_mask,
cond=None,
cond_scale=1.,
device=None,
image=None,
degrade_mask=None,
num_guidance_steps=3,
guidance_scale=1.5,
guidance_start_t=-1,
recurrent_steps=1,
loss_type='l1',
use_ddim=True,
ddim_steps=50,
ddim_eta=0.0,
guidance_strategy='equal_distance',
proc=True,
):
"""
Universal guidance for continuous diffusion model supporting both DDPM and DDIM.
This function performs gradient-based guidance to match the degraded region of the
predicted clean image with the real clean image.
Args:
shape_image: Shape of image to generate
shape_mask: Shape of mask to generate
cond: Optional conditioning
cond_scale: Conditioning scale
device: Device to use
image: Real clean image (ground truth for degraded region)
degrade_mask: Binary mask indicating degraded region (1=degraded, 0=clean)
num_guidance_steps: Number of gradient descent iterations per guided timestep
guidance_scale: Step size for gradient descent (delta)
guidance_start_t: Number of timesteps/steps to apply guidance
recurrent_steps: Number of denoise-renoise cycles per timestep (DDPM and DDIM)
loss_type: 'l1' or 'l2' for guidance loss
use_ddim: Whether to use DDIM sampling instead of DDPM
ddim_steps: Number of steps for DDIM sampling
ddim_eta: Stochasticity parameter for DDIM (0=deterministic, 1=DDPM-like)
guidance_strategy: 'last_n' or 'equal_distance' - how to distribute guidance steps
proc: Whether to show progress bar
Returns:
Generated pair (image + mask concatenated)
"""
if degrade_mask is None:
degrade_mask = torch.ones(shape_image, device=device)
degrade_mask = degrade_mask.to(device=device, dtype=torch.float32)
device = self.betas.device
b = shape_image[0]
# Pre-compute noisy versions of input image at all timesteps (for replacement strategy)
init_noise = torch.randn_like(image)
step_noise_list = []
for step in range(self.num_timesteps):
t = torch.full((image.shape[0],), step, device=device, dtype=torch.long)
step_noise = self.q_sample(image, t, noise=init_noise)
step_noise_list.append(step_noise)
img_noisy = torch.stack(step_noise_list) # [T, B, C, D, H, W]
# Initialize random samples
img = torch.randn(shape_image, device=device)
mask = torch.randn(shape_mask, device=device)
pair = torch.cat((img, mask), dim=1)
if use_ddim:
# DDIM sampling with universal guidance
return self._ddim_sample_universal_guidance(
pair=pair,
img_noisy=img_noisy,
real_image=image,
degrade_mask=degrade_mask,
cond=cond,
cond_scale=cond_scale,
num_guidance_steps=num_guidance_steps,
guidance_scale=guidance_scale,
guidance_start_t=guidance_start_t,
recurrent_steps=recurrent_steps,
loss_type=loss_type,
ddim_steps=ddim_steps,
ddim_eta=ddim_eta,
guidance_strategy=guidance_strategy,
proc=proc,
)
else:
# DDPM sampling with universal guidance
return self._ddpm_sample_universal_guidance(
pair=pair,
img_noisy=img_noisy,
real_image=image,
degrade_mask=degrade_mask,
cond=cond,
cond_scale=cond_scale,
num_guidance_steps=num_guidance_steps,
guidance_scale=guidance_scale,
guidance_start_t=guidance_start_t,
recurrent_steps=recurrent_steps,
loss_type=loss_type,
guidance_strategy=guidance_strategy,
proc=proc,
)
def _ddpm_sample_universal_guidance(
self,
pair,
img_noisy,
real_image,
degrade_mask,
cond,
cond_scale,
num_guidance_steps,
guidance_scale,
guidance_start_t,
recurrent_steps,
loss_type,
guidance_strategy,
proc,
):
"""DDPM sampling with universal guidance."""
device = pair.device
b = pair.shape[0]
# Build guidance schedule
guidance_schedule = self._build_guidance_schedule(
total_steps=self.num_timesteps,
num_guidance=guidance_start_t,
strategy=guidance_strategy,
)
print(f'DDPM Universal Guidance Config:')
print(f' num_guidance_steps: {num_guidance_steps}')
print(f' recurrent_steps: {recurrent_steps}')
print(f' guidance_scale: {guidance_scale}')
print(f' guidance_start_t: {guidance_start_t}')
print(f' guidance_strategy: {guidance_strategy}')
print(f' loss_type: {loss_type}')
print(f' degrade_mask sum: {degrade_mask.sum().item()}')
print(f' guidance at {len(guidance_schedule)} timesteps: {sorted(list(guidance_schedule))[:10]}{"..." if len(guidance_schedule) > 10 else ""}')
iterator = range(self.num_timesteps - 1, -1, -1)
if proc:
from tqdm import tqdm
iterator = tqdm(iterator, desc='DDPM + Universal Guidance', leave=False)
for i in iterator:
t = torch.full((b,), i, device=device, dtype=torch.long)
# Replace degraded region with noisy ground truth
pair[:, :1] = img_noisy[i] * degrade_mask + pair[:, :1] * (1 - degrade_mask)
# Apply guidance if this timestep is in the schedule
if i in guidance_schedule:
pair = pair.clone().detach()
for recurrent_idx in range(recurrent_steps):
# Gradient descent iterations
for g in range(num_guidance_steps):
with torch.enable_grad():
pair_with_grad = pair.detach().clone().requires_grad_(True)
# Compute guidance loss
loss = self._compute_guidance_loss(
pair_with_grad,
t,
real_image,
degrade_mask,
loss_type,
cond,
)
if (i % 10 == 0 or i < 3) and g == 0 and recurrent_idx == 0:
print(f' t={i}: loss={loss.item():.6f}')
# Compute gradient
grad = torch.autograd.grad(loss, pair_with_grad)[0]
# Only apply gradient to non-degraded region of image channel
grad[:, :1] = grad[:, :1] * (1 - degrade_mask)
# Normalize gradient
grad_norm = grad.flatten(start_dim=1).norm(dim=1, keepdim=True)
grad_norm = grad_norm.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1)
grad = grad / (grad_norm + 1e-8)
# Update pair
pair = pair - grad * guidance_scale
# Recurrent refinement: denoise one step and add noise back
# (Only if not the last recurrent step)
if recurrent_idx < recurrent_steps - 1 and i > 0:
with torch.no_grad():
# Denoise one step
pair = self.p_sample(pair, t, cond=cond, cond_scale=cond_scale)
# Add noise back: x_{t-1} → x_t
pair = self.q_sample_one_step(pair, t)
# Standard denoising step
with torch.no_grad():
pair = self.p_sample(pair, t, cond=cond, cond_scale=cond_scale).clone()
return pair
def _ddim_sample_universal_guidance(
self,
pair,
img_noisy,
real_image,
degrade_mask,
cond,
cond_scale,
num_guidance_steps,
guidance_scale,
guidance_start_t,
recurrent_steps,
loss_type,
ddim_steps,
ddim_eta,
guidance_strategy,
proc,
):
"""DDIM sampling with universal guidance."""
device = pair.device
b = pair.shape[0]
# Build DDIM timestep schedule
step = self.num_timesteps // ddim_steps
timesteps = torch.arange(0, self.num_timesteps, step, device=device).long()
timesteps = torch.flip(timesteps, dims=[0]) # Reverse for denoising
# Build guidance schedule based on DDIM steps
guidance_schedule = self._build_guidance_schedule(
total_steps=len(timesteps),
num_guidance=guidance_start_t,
strategy=guidance_strategy,
)
print(f'DDIM Universal Guidance Config:')
print(f' ddim_steps: {ddim_steps}')
print(f' num_guidance_steps: {num_guidance_steps}')
print(f' recurrent_steps: {recurrent_steps}')
print(f' guidance_scale: {guidance_scale}')
print(f' guidance_start_t: {guidance_start_t}')
print(f' guidance_strategy: {guidance_strategy}')
print(f' ddim_eta: {ddim_eta}')
print(f' loss_type: {loss_type}')
print(f' degrade_mask sum: {degrade_mask.sum().item()}')
print(f' guidance at {len(guidance_schedule)} DDIM steps (indices): {sorted(list(guidance_schedule))[:10]}{"..." if len(guidance_schedule) > 10 else ""}')
iterator = enumerate(timesteps.tolist())
if proc:
from tqdm import tqdm
iterator = enumerate(tqdm(timesteps.tolist(), desc='DDIM + Universal Guidance', leave=False))
for idx, t_val in iterator:
t = torch.full((b,), t_val, device=device, dtype=torch.long)
# Determine next timestep
if idx + 1 < len(timesteps):
t_next = timesteps[idx + 1]
else:
t_next = torch.tensor(-1, device=device)
# Replace degraded region with noisy ground truth before guidance
pair[:, :1] = img_noisy[t_val] * degrade_mask + pair[:, :1] * (1 - degrade_mask)
# Apply guidance if this DDIM step index is in the schedule
if idx in guidance_schedule:
pair = pair.clone().detach()
for recurrent_idx in range(recurrent_steps):
# Gradient descent iterations
for g in range(num_guidance_steps):
with torch.enable_grad():
pair_with_grad = pair.detach().clone().requires_grad_(True)
# Compute guidance loss
loss = self._compute_guidance_loss(
pair_with_grad,
t,
real_image,
degrade_mask,
loss_type,
cond,
)
if (idx % 5 == 0 or idx < 3) and g == 0 and recurrent_idx == 0:
print(f' DDIM step {idx} (t={t_val}): loss={loss.item():.6f}')
# Compute gradient
grad = torch.autograd.grad(loss, pair_with_grad)[0]
# Only apply gradient to non-degraded region of image channel
grad[:, :1] = grad[:, :1] * (1 - degrade_mask)
# Normalize gradient
grad_norm = grad.flatten(start_dim=1).norm(dim=1, keepdim=True)
grad_norm = grad_norm.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1)
grad = grad / (grad_norm + 1e-8)
# Update pair
pair = pair - grad * guidance_scale
# Recurrent refinement: DDIM step and add noise back
# (Only if not the last recurrent step and not the final timestep)
if recurrent_idx < recurrent_steps - 1 and t_next >= 0:
with torch.no_grad():
# Apply one DDIM denoising step
pair_denoised = self._ddim_step(
pair,
t,
t_next,
cond=cond,
cond_scale=cond_scale,
eta=ddim_eta,
)
# Re-noise back to current timestep t using DDIM forward process
# q(x_t | x_{t-1}) for DDIM: deterministically add noise back
alpha_t = extract(self.alphas_cumprod, t, pair.shape)
alpha_t_next = extract(self.alphas_cumprod, t_next.expand(pair.shape[0]), pair.shape)
# Predict x0 from denoised sample at t_next
noise_pred = self.denoise_fn(pair_denoised, t_next.expand(b), cond=cond)[:, :pair.shape[1]]
x0_from_denoised = (pair_denoised - torch.sqrt(1 - alpha_t_next) * noise_pred) / torch.sqrt(alpha_t_next)
# Re-noise to timestep t: x_t = sqrt(alpha_t) * x0 + sqrt(1-alpha_t) * noise
noise = torch.randn_like(pair)
pair = torch.sqrt(alpha_t) * x0_from_denoised + torch.sqrt(1 - alpha_t) * noise
# Replace degraded region with noisy ground truth after guidance
pair[:, :1] = img_noisy[t_val] * degrade_mask + pair[:, :1] * (1 - degrade_mask)
# DDIM denoising step
with torch.no_grad():
pair = self._ddim_step(
pair,
t,
t_next,
cond=cond,
cond_scale=cond_scale,
eta=ddim_eta,
)
return pair
def _compute_guidance_loss(
self,
pair_noisy,
t,
real_image,
degrade_mask,
loss_type,
cond,
):
"""
Compute guidance loss for universal guidance.
The loss is the L1/L2 distance between:
- degrade_mask * predicted clean image
- degrade_mask * real clean image
"""
device = pair_noisy.device
# Predict noise from the noisy pair
noise_pred = self.denoise_fn(pair_noisy, t, cond=cond)[:, :1]
# Predict clean image (x_0) from noise prediction
# x_0 = (x_t - sqrt(1-alpha_t) * noise) / sqrt(alpha_t)
x0_pred = self.predict_start_from_noise(pair_noisy[:, :1], t, noise_pred)
# Apply mask to both predicted and real clean images
if degrade_mask is not None:
x0_pred_masked = x0_pred * degrade_mask
real_image_masked = real_image * degrade_mask
else:
x0_pred_masked = x0_pred
real_image_masked = real_image
# Compute loss between masked regions
if loss_type == 'l1':
loss = F.l1_loss(x0_pred_masked, real_image_masked, reduction='sum')
elif loss_type == 'l2':
loss = F.mse_loss(x0_pred_masked, real_image_masked, reduction='sum')
else:
raise NotImplementedError(f'Unknown loss type: {loss_type}')
# Normalize by mask size
if degrade_mask is not None:
loss = loss / degrade_mask.sum().clamp(min=1.0)
else:
loss = loss / real_image_masked.numel()
return loss
def _ddim_step(self, x, t, t_next, cond=None, cond_scale=1., eta=0.0):
"""
Single DDIM denoising step.
Args:
x: Current noisy sample
t: Current timestep tensor
t_next: Next timestep (scalar tensor or -1 for final step)
cond: Optional conditioning
cond_scale: Conditioning scale
eta: Stochasticity parameter (0=deterministic, 1=DDPM-like)
"""
# Predict noise
noise_pred = self.denoise_fn(x, t, cond=cond)
# Apply classifier-free guidance if cond_scale != 1
if cond is not None and cond_scale != 1.:
noise_pred_uncond = self.denoise_fn(x, t, cond=None)
noise_pred = noise_pred_uncond + cond_scale * (noise_pred - noise_pred_uncond)
# Extract coefficients
alpha_t = extract(self.alphas_cumprod, t, x.shape)
if t_next >= 0:
alpha_t_next = extract(self.alphas_cumprod, t_next.expand(x.shape[0]), x.shape)
else:
alpha_t_next = torch.ones_like(alpha_t)
# Predict x0
pred_x0 = (x - torch.sqrt(1 - alpha_t) * noise_pred) / torch.sqrt(alpha_t)
# Compute direction pointing to x_t
sigma_t = eta * torch.sqrt((1 - alpha_t_next) / (1 - alpha_t) * (1 - alpha_t / alpha_t_next))
# Compute x_{t-1}
dir_xt = torch.sqrt(1 - alpha_t_next - sigma_t ** 2) * noise_pred
if t_next >= 0:
noise = torch.randn_like(x)
x_next = torch.sqrt(alpha_t_next) * pred_x0 + dir_xt + sigma_t * noise
else:
x_next = torch.sqrt(alpha_t_next) * pred_x0 + dir_xt
return x_next
def _build_guidance_schedule(self, total_steps, num_guidance, strategy='last_n'):
"""
Build a set of step indices where guidance should be applied.
Args:
total_steps: Total number of denoising steps
num_guidance: Number of steps to apply guidance
strategy: 'last_n' or 'equal_distance'
- 'last_n': Apply guidance at the last N steps (smallest timesteps)
- 'equal_distance': Distribute N guidance steps evenly across all steps
Returns:
Set of step indices where guidance should be applied
"""
num_guidance = min(num_guidance, total_steps)
if strategy == 'last_n':
# For DDPM iterating t from (total_steps-1) down to 0:
# Last N steps means the smallest timestep values: {0, 1, ..., N-1}
return set(range(num_guidance))
elif strategy == 'equal_distance':
# Distribute guidance steps evenly across the timeline
if num_guidance == 0:
return set()
if num_guidance >= total_steps:
return set(range(total_steps))
# Calculate spacing
spacing = total_steps / num_guidance
indices = []
for i in range(num_guidance):
idx = int(i * spacing)
indices.append(idx)
return set(indices)
else:
raise ValueError(f"Unknown guidance strategy: {strategy}. Use 'last_n' or 'equal_distance'")
def p_sample_loop_gen(self, shape_image, shape_mask, cond=None, cond_scale=1., device=None, mask=None, degrade_mask=None):
if degrade_mask is None: # default is all 1
degrade_mask = torch.ones(shape_image, device=device)
degrade_mask = degrade_mask.to(device=device, dtype=torch.float32)
device = self.betas.device
b = shape_image[0]
init_noise = torch.randn_like(mask)
step_noise_list = []
for step in range(self.num_timesteps):
t = torch.full((mask.shape[0],), step, device=device, dtype=torch.long)
step_noise = self.q_sample(mask, t, noise=init_noise)
step_noise_list.append(step_noise)
mask_noisy = torch.stack(step_noise_list) # [T, B, C, D, H, W]
img = torch.randn(shape_image, device=device)
mask = torch.randn(shape_mask, device=device)
pair = torch.cat((img, mask), dim=1)
# print(shape_mask)
TCOUNT = 2 # 2
RSTEP = 1 # 10
GRAD_STEP = 0 # 7
print(f'TCOUNT: {TCOUNT}, RSTEP: {RSTEP}, GRAD_STEP: {GRAD_STEP}')
recurrent = [0] * self.num_timesteps
for i in range(self.num_timesteps):
if i % RSTEP == 0:
recurrent[i] = TCOUNT
i = self.num_timesteps - 1
while i >= 0:
# real_noisy_mask = self.q_sample(x_start=mask, t=t)
pair[:, 1:] = mask_noisy[i] * degrade_mask + pair[:, 1:] * (1 - degrade_mask)
pair = pair.clone().detach()
for g in range(GRAD_STEP):
if i > 10:
break
with torch.enable_grad():
pair_with_grad = pair.detach().clone().requires_grad_(True)
t = torch.full((b,), i, device=device, dtype=torch.long)
self.loss_type = 'l1'
loss = self.p_losses_guidance_gen(pair_with_grad, t, init_noise, degrade_mask=degrade_mask)
print('t:', i, 'g', g, 'loss:', loss.item())
grad = torch.autograd.grad(loss, pair_with_grad)[0][:, :1]
grad_norm = grad.flatten(start_dim=1).norm( dim=1, keepdim=True).unsqueeze(-1).unsqueeze(-1).unsqueeze(-1)
grad = grad / (grad_norm + 1e-8)
pair[:, :1] = pair[:, :1] - grad * 0.5
t = torch.full((b,), i, device=device, dtype=torch.long)
pair = self.p_sample(pair, t, cond=cond, cond_scale=cond_scale).clone()
# if recurrent[i] > 0 and i % RSTEP == 0:
# recurrent[i] -= 1
# for _ in range(RSTEP):
# with torch.no_grad():
# t = torch.full((b,), i, device=device, dtype=torch.long)
# pair = self.q_sample_one_step(pair, t)
# i += 1
# if i >= self.num_timesteps:
# break
i -= 1
return pair
def p_losses_guidance_gen(self, pair_noisy, t, noise, cond=None, degrade_mask=None, **kwargs):
device = pair_noisy.device
if is_list_str(cond):
cond = bert_embed(
tokenize(cond), return_cls_repr=self.text_use_bert_cls)
cond = cond.to(device)
x_recon = self.denoise_fn(pair_noisy, t, cond=cond, **kwargs)[:, 1:]
if self.loss_type == 'l1':
loss = F.l1_loss(noise, x_recon, reduction='sum')
elif self.loss_type == 'l2':
loss = F.mse_loss(noise, x_recon, reduction='sum')
else:
raise NotImplementedError()
if degrade_mask is not None:
loss = loss / degrade_mask.sum()
else:
loss = loss / noise.numel()
return loss
class Trainer(object):
def __init__(
self,
diffusion_model,
cfg,
dataset=None,
*,
ema_decay=0.995,
train_batch_size=32,
train_lr=1e-4,
train_num_steps=100000,
gradient_accumulate_every=2,
amp=False,
step_start_ema=2000,
update_ema_every=10,
save_and_sample_every=1000,
results_folder='./results',
max_grad_norm=None,
num_workers=4,
device=None,
):
super().__init__()
self.model = diffusion_model
self.ema = EMA(ema_decay)
self.ema_model = copy.deepcopy(self.model)
self.update_ema_every = update_ema_every
self.step_start_ema = step_start_ema
self.save_and_sample_every = save_and_sample_every
self.batch_size = train_batch_size
self.image_size = diffusion_model.image_size
self.gradient_accumulate_every = gradient_accumulate_every
self.train_num_steps = train_num_steps
self.device = device
self.cfg = cfg
self.ds = dataset
dl = DataLoader(self.ds, batch_size=train_batch_size,
shuffle=True, pin_memory=True, num_workers=num_workers)
self.len_dataloader = len(dl)
print("len_dl ", len(dl))
self.dl = cycle(dl)
print(f'found {len(self.ds)} videos as gif files')
assert len(
self.ds) > 0, 'need to have at least 1 video to start training (although 1 is not great, try 100k)'
self.opt = Adam(diffusion_model.parameters(), lr=train_lr)
self.step = 0
self.amp = amp
self.scaler = GradScaler(enabled=amp)
self.max_grad_norm = max_grad_norm
self.results_folder = Path(results_folder)
self.results_folder.mkdir(exist_ok=True, parents=True)
self.reset_parameters()
def reset_parameters(self):
self.ema_model.load_state_dict(self.model.state_dict())
def step_ema(self):
if self.step < self.step_start_ema:
self.reset_parameters()
return
self.ema.update_model_average(self.ema_model, self.model)
def save(self, milestone):
data = {
'step': self.step,
'model': self.model.state_dict(),
'ema': self.ema_model.state_dict(),
'scaler': self.scaler.state_dict(),
'optimizer': self.opt.state_dict() # 保存优化器状态
}
torch.save(data, str(self.results_folder / f'model-{milestone}.pt'))
def load(self, milestone, map_location=None, **kwargs):
if milestone == -1:
all_milestones = [int(p.stem.split('-')[-1])
for p in Path(self.results_folder).glob('**/*.pt')]
assert len(
all_milestones) > 0, 'need to have at least one milestone to load from latest checkpoint (milestone == -1)'
milestone = max(all_milestones)
if map_location:
data = torch.load(milestone, map_location=map_location)
else:
import os
data = torch.load(os.path.join(self.results_folder, f'model-{milestone}.pt'))
self.step = data['step']
self.model.load_state_dict(data['model'], **kwargs)
self.ema_model.load_state_dict(data['ema'], **kwargs)
self.scaler.load_state_dict(data['scaler'])
self.opt.load_state_dict(data['optimizer'])
def train(
self,
prob_focus_present=0.,
focus_present_mask=None,
log_fn=noop
):
assert callable(log_fn)
while self.step < self.train_num_steps:
for i in range(self.gradient_accumulate_every):
data_frame = next(self.dl)
data = data_frame['img'].to(self.device)
mask_sdf = data_frame['mask_sdf'].to(self.device)
# print("Mask Sum: ", mask.sum())
# print(data_frame['name'])
with autocast(enabled=self.amp):
loss = self.model(**dict(
x=data,
mask=mask_sdf,
prob_focus_present=prob_focus_present,
focus_present_mask=focus_present_mask)
)
self.scaler.scale(
loss / self.gradient_accumulate_every).backward()
print(f'{self.step}: {loss.item()}')
log = {'loss': loss.item()}
if exists(self.max_grad_norm):
self.scaler.unscale_(self.opt)
nn.utils.clip_grad_norm_(
self.model.parameters(), self.max_grad_norm)
self.scaler.step(self.opt)
self.scaler.update()
self.opt.zero_grad()
if self.step % self.update_ema_every == 0:
self.step_ema()
if self.step != 0 and self.step % self.save_and_sample_every == 0:
self.ema_model.eval()
with torch.no_grad():
milestone = self.step // self.save_and_sample_every
self.save(milestone)
log_fn(log)
self.step += 1
print('training completed')
# _extract_into_tensor 函数的作用是从一个给定的数组(arr)中提取与时间步(timesteps)相对应的值,
# 并将这些值广播成指定的形状(broadcast_shape)。
# 该函数主要用于处理扩散模型中的参数提取和广播操作,确保不同时间步的参数能够与图像数据进行匹配
def _extract_into_tensor(arr, timesteps, broadcast_shape):
res = arr[timesteps].float()
while len(res.shape) < len(broadcast_shape):
res = res[..., None]
return res.expand(broadcast_shape)
|