File size: 54,309 Bytes
e8aa256 |
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 |
import math
from typing import Optional, Union
import torch
from torch import nn
from ...configuration_utils import ConfigMixin, register_to_config
from ...models import ModelMixin
from ...models.attention import FeedForward
from ...models.attention_processor import Attention
from ...models.embeddings import TimestepEmbedding, Timesteps, get_2d_sincos_pos_embed
from ...models.normalization import AdaLayerNorm
from ...models.transformer_2d import Transformer2DModelOutput
from ...utils import logging
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
def _no_grad_trunc_normal_(tensor, mean, std, a, b):
# Cut & paste from PyTorch official master until it's in a few official releases - RW
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
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):
logger.warning(
"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
"The distribution of values may be incorrect."
)
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
l = norm_cdf((a - mean) / std)
u = norm_cdf((b - mean) / std)
# Uniformly fill tensor with values from [l, u], then translate to
# [2l-1, 2u-1].
tensor.uniform_(2 * l - 1, 2 * u - 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, mean=0.0, std=1.0, a=-2.0, b=2.0):
# type: (torch.Tensor, float, float, float, float) -> torch.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`.
Args:
tensor: an n-dimensional `torch.Tensor`
mean: the mean of the normal distribution
std: the standard deviation of the normal distribution
a: the minimum cutoff value
b: the maximum cutoff value
Examples:
>>> w = torch.empty(3, 5) >>> nn.init.trunc_normal_(w)
"""
return _no_grad_trunc_normal_(tensor, mean, std, a, b)
class PatchEmbed(nn.Module):
"""2D Image to Patch Embedding"""
def __init__(
self,
height=224,
width=224,
patch_size=16,
in_channels=3,
embed_dim=768,
layer_norm=False,
flatten=True,
bias=True,
use_pos_embed=True,
):
super().__init__()
num_patches = (height // patch_size) * (width // patch_size)
self.flatten = flatten
self.layer_norm = layer_norm
self.proj = nn.Conv2d(
in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
)
if layer_norm:
self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6)
else:
self.norm = None
self.use_pos_embed = use_pos_embed
if self.use_pos_embed:
pos_embed = get_2d_sincos_pos_embed(embed_dim, int(num_patches**0.5))
self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
def forward(self, latent):
latent = self.proj(latent)
if self.flatten:
latent = latent.flatten(2).transpose(1, 2) # BCHW -> BNC
if self.layer_norm:
latent = self.norm(latent)
if self.use_pos_embed:
return latent + self.pos_embed
else:
return latent
class SkipBlock(nn.Module):
def __init__(self, dim: int):
super().__init__()
self.skip_linear = nn.Linear(2 * dim, dim)
# Use torch.nn.LayerNorm for now, following the original code
self.norm = nn.LayerNorm(dim)
def forward(self, x, skip):
x = self.skip_linear(torch.cat([x, skip], dim=-1))
x = self.norm(x)
return x
# Modified to support both pre-LayerNorm and post-LayerNorm configurations
# Don't support AdaLayerNormZero for now
# Modified from diffusers.models.attention.BasicTransformerBlock
class UTransformerBlock(nn.Module):
r"""
A modification of BasicTransformerBlock which supports pre-LayerNorm and post-LayerNorm configurations.
Parameters:
dim (`int`): The number of channels in the input and output.
num_attention_heads (`int`): The number of heads to use for multi-head attention.
attention_head_dim (`int`): The number of channels in each head.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
activation_fn (`str`, *optional*, defaults to `"geglu"`):
Activation function to be used in feed-forward.
num_embeds_ada_norm (:obj: `int`, *optional*):
The number of diffusion steps used during training. See `Transformer2DModel`.
attention_bias (:obj: `bool`, *optional*, defaults to `False`):
Configure if the attentions should contain a bias parameter.
only_cross_attention (`bool`, *optional*):
Whether to use only cross-attention layers. In this case two cross attention layers are used.
double_self_attention (`bool`, *optional*):
Whether to use two self-attention layers. In this case no cross attention layers are used.
upcast_attention (`bool`, *optional*):
Whether to upcast the query and key to float32 when performing the attention calculation.
norm_elementwise_affine (`bool`, *optional*):
Whether to use learnable per-element affine parameters during layer normalization.
norm_type (`str`, defaults to `"layer_norm"`):
The layer norm implementation to use.
pre_layer_norm (`bool`, *optional*):
Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
as opposed to after ("post-LayerNorm"). Note that `BasicTransformerBlock` uses pre-LayerNorm, e.g.
`pre_layer_norm = True`.
final_dropout (`bool`, *optional*):
Whether to use a final Dropout layer after the feedforward network.
"""
def __init__(
self,
dim: int,
num_attention_heads: int,
attention_head_dim: int,
dropout=0.0,
cross_attention_dim: Optional[int] = None,
activation_fn: str = "geglu",
num_embeds_ada_norm: Optional[int] = None,
attention_bias: bool = False,
only_cross_attention: bool = False,
double_self_attention: bool = False,
upcast_attention: bool = False,
norm_elementwise_affine: bool = True,
norm_type: str = "layer_norm",
pre_layer_norm: bool = True,
final_dropout: bool = False,
):
super().__init__()
self.only_cross_attention = only_cross_attention
self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
self.pre_layer_norm = pre_layer_norm
if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
raise ValueError(
f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
)
# 1. Self-Attn
self.attn1 = Attention(
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
upcast_attention=upcast_attention,
)
# 2. Cross-Attn
if cross_attention_dim is not None or double_self_attention:
self.attn2 = Attention(
query_dim=dim,
cross_attention_dim=cross_attention_dim if not double_self_attention else None,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
upcast_attention=upcast_attention,
) # is self-attn if encoder_hidden_states is none
else:
self.attn2 = None
if self.use_ada_layer_norm:
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
else:
self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
if cross_attention_dim is not None or double_self_attention:
# We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
# I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
# the second cross attention block.
self.norm2 = (
AdaLayerNorm(dim, num_embeds_ada_norm)
if self.use_ada_layer_norm
else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
)
else:
self.norm2 = None
# 3. Feed-forward
self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
def forward(
self,
hidden_states,
attention_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
timestep=None,
cross_attention_kwargs=None,
class_labels=None,
):
# Pre-LayerNorm
if self.pre_layer_norm:
if self.use_ada_layer_norm:
norm_hidden_states = self.norm1(hidden_states, timestep)
else:
norm_hidden_states = self.norm1(hidden_states)
else:
norm_hidden_states = hidden_states
# 1. Self-Attention
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
attn_output = self.attn1(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
attention_mask=attention_mask,
**cross_attention_kwargs,
)
# Post-LayerNorm
if not self.pre_layer_norm:
if self.use_ada_layer_norm:
attn_output = self.norm1(attn_output, timestep)
else:
attn_output = self.norm1(attn_output)
hidden_states = attn_output + hidden_states
if self.attn2 is not None:
# Pre-LayerNorm
if self.pre_layer_norm:
norm_hidden_states = (
self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
)
else:
norm_hidden_states = hidden_states
# TODO (Birch-San): Here we should prepare the encoder_attention mask correctly
# prepare attention mask here
# 2. Cross-Attention
attn_output = self.attn2(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=encoder_attention_mask,
**cross_attention_kwargs,
)
# Post-LayerNorm
if not self.pre_layer_norm:
attn_output = self.norm2(attn_output, timestep) if self.use_ada_layer_norm else self.norm2(attn_output)
hidden_states = attn_output + hidden_states
# 3. Feed-forward
# Pre-LayerNorm
if self.pre_layer_norm:
norm_hidden_states = self.norm3(hidden_states)
else:
norm_hidden_states = hidden_states
ff_output = self.ff(norm_hidden_states)
# Post-LayerNorm
if not self.pre_layer_norm:
ff_output = self.norm3(ff_output)
hidden_states = ff_output + hidden_states
return hidden_states
# Like UTransformerBlock except with LayerNorms on the residual backbone of the block
# Modified from diffusers.models.attention.BasicTransformerBlock
class UniDiffuserBlock(nn.Module):
r"""
A modification of BasicTransformerBlock which supports pre-LayerNorm and post-LayerNorm configurations and puts the
LayerNorms on the residual backbone of the block. This matches the transformer block in the [original UniDiffuser
implementation](https://github.com/thu-ml/unidiffuser/blob/main/libs/uvit_multi_post_ln_v1.py#L104).
Parameters:
dim (`int`): The number of channels in the input and output.
num_attention_heads (`int`): The number of heads to use for multi-head attention.
attention_head_dim (`int`): The number of channels in each head.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
activation_fn (`str`, *optional*, defaults to `"geglu"`):
Activation function to be used in feed-forward.
num_embeds_ada_norm (:obj: `int`, *optional*):
The number of diffusion steps used during training. See `Transformer2DModel`.
attention_bias (:obj: `bool`, *optional*, defaults to `False`):
Configure if the attentions should contain a bias parameter.
only_cross_attention (`bool`, *optional*):
Whether to use only cross-attention layers. In this case two cross attention layers are used.
double_self_attention (`bool`, *optional*):
Whether to use two self-attention layers. In this case no cross attention layers are used.
upcast_attention (`bool`, *optional*):
Whether to upcast the query and key to float() when performing the attention calculation.
norm_elementwise_affine (`bool`, *optional*):
Whether to use learnable per-element affine parameters during layer normalization.
norm_type (`str`, defaults to `"layer_norm"`):
The layer norm implementation to use.
pre_layer_norm (`bool`, *optional*):
Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm
(`pre_layer_norm = False`).
final_dropout (`bool`, *optional*):
Whether to use a final Dropout layer after the feedforward network.
"""
def __init__(
self,
dim: int,
num_attention_heads: int,
attention_head_dim: int,
dropout=0.0,
cross_attention_dim: Optional[int] = None,
activation_fn: str = "geglu",
num_embeds_ada_norm: Optional[int] = None,
attention_bias: bool = False,
only_cross_attention: bool = False,
double_self_attention: bool = False,
upcast_attention: bool = False,
norm_elementwise_affine: bool = True,
norm_type: str = "layer_norm",
pre_layer_norm: bool = False,
final_dropout: bool = True,
):
super().__init__()
self.only_cross_attention = only_cross_attention
self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
self.pre_layer_norm = pre_layer_norm
if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
raise ValueError(
f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
)
# 1. Self-Attn
self.attn1 = Attention(
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
upcast_attention=upcast_attention,
)
# 2. Cross-Attn
if cross_attention_dim is not None or double_self_attention:
self.attn2 = Attention(
query_dim=dim,
cross_attention_dim=cross_attention_dim if not double_self_attention else None,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
upcast_attention=upcast_attention,
) # is self-attn if encoder_hidden_states is none
else:
self.attn2 = None
if self.use_ada_layer_norm:
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
else:
self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
if cross_attention_dim is not None or double_self_attention:
# We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
# I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
# the second cross attention block.
self.norm2 = (
AdaLayerNorm(dim, num_embeds_ada_norm)
if self.use_ada_layer_norm
else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
)
else:
self.norm2 = None
# 3. Feed-forward
self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
def forward(
self,
hidden_states,
attention_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
timestep=None,
cross_attention_kwargs=None,
class_labels=None,
):
# Following the diffusers transformer block implementation, put the LayerNorm on the
# residual backbone
# Pre-LayerNorm
if self.pre_layer_norm:
if self.use_ada_layer_norm:
hidden_states = self.norm1(hidden_states, timestep)
else:
hidden_states = self.norm1(hidden_states)
# 1. Self-Attention
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
attn_output = self.attn1(
hidden_states,
encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
attention_mask=attention_mask,
**cross_attention_kwargs,
)
hidden_states = attn_output + hidden_states
# Following the diffusers transformer block implementation, put the LayerNorm on the
# residual backbone
# Post-LayerNorm
if not self.pre_layer_norm:
if self.use_ada_layer_norm:
hidden_states = self.norm1(hidden_states, timestep)
else:
hidden_states = self.norm1(hidden_states)
if self.attn2 is not None:
# Pre-LayerNorm
if self.pre_layer_norm:
hidden_states = (
self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
)
# TODO (Birch-San): Here we should prepare the encoder_attention mask correctly
# prepare attention mask here
# 2. Cross-Attention
attn_output = self.attn2(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=encoder_attention_mask,
**cross_attention_kwargs,
)
hidden_states = attn_output + hidden_states
# Post-LayerNorm
if not self.pre_layer_norm:
hidden_states = (
self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
)
# 3. Feed-forward
# Pre-LayerNorm
if self.pre_layer_norm:
hidden_states = self.norm3(hidden_states)
ff_output = self.ff(hidden_states)
hidden_states = ff_output + hidden_states
# Post-LayerNorm
if not self.pre_layer_norm:
hidden_states = self.norm3(hidden_states)
return hidden_states
# Modified from diffusers.models.transformer_2d.Transformer2DModel
# Modify the transformer block structure to be U-Net like following U-ViT
# Only supports patch-style input and torch.nn.LayerNorm currently
# https://github.com/baofff/U-ViT
class UTransformer2DModel(ModelMixin, ConfigMixin):
"""
Transformer model based on the [U-ViT](https://github.com/baofff/U-ViT) architecture for image-like data. Compared
to [`Transformer2DModel`], this model has skip connections between transformer blocks in a "U"-shaped fashion,
similar to a U-Net. Supports only continuous (actual embeddings) inputs, which are embedded via a [`PatchEmbed`]
layer and then reshaped to (b, t, d).
Parameters:
num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
in_channels (`int`, *optional*):
Pass if the input is continuous. The number of channels in the input.
out_channels (`int`, *optional*):
The number of output channels; if `None`, defaults to `in_channels`.
num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
norm_num_groups (`int`, *optional*, defaults to `32`):
The number of groups to use when performing Group Normalization.
cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
attention_bias (`bool`, *optional*):
Configure if the TransformerBlocks' attention should contain a bias parameter.
sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
Note that this is fixed at training time as it is used for learning a number of position embeddings. See
`ImagePositionalEmbeddings`.
num_vector_embeds (`int`, *optional*):
Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
Includes the class for the masked latent pixel.
patch_size (`int`, *optional*, defaults to 2):
The patch size to use in the patch embedding.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
The number of diffusion steps used during training. Note that this is fixed at training time as it is used
to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
up to but not more than steps than `num_embeds_ada_norm`.
use_linear_projection (int, *optional*): TODO: Not used
only_cross_attention (`bool`, *optional*):
Whether to use only cross-attention layers. In this case two cross attention layers are used in each
transformer block.
upcast_attention (`bool`, *optional*):
Whether to upcast the query and key to float() when performing the attention calculation.
norm_type (`str`, *optional*, defaults to `"layer_norm"`):
The Layer Normalization implementation to use. Defaults to `torch.nn.LayerNorm`.
block_type (`str`, *optional*, defaults to `"unidiffuser"`):
The transformer block implementation to use. If `"unidiffuser"`, has the LayerNorms on the residual
backbone of each transformer block; otherwise has them in the attention/feedforward branches (the standard
behavior in `diffusers`.)
pre_layer_norm (`bool`, *optional*):
Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm
(`pre_layer_norm = False`).
norm_elementwise_affine (`bool`, *optional*):
Whether to use learnable per-element affine parameters during layer normalization.
use_patch_pos_embed (`bool`, *optional*):
Whether to use position embeddings inside the patch embedding layer (`PatchEmbed`).
final_dropout (`bool`, *optional*):
Whether to use a final Dropout layer after the feedforward network.
"""
@register_to_config
def __init__(
self,
num_attention_heads: int = 16,
attention_head_dim: int = 88,
in_channels: Optional[int] = None,
out_channels: Optional[int] = None,
num_layers: int = 1,
dropout: float = 0.0,
norm_num_groups: int = 32,
cross_attention_dim: Optional[int] = None,
attention_bias: bool = False,
sample_size: Optional[int] = None,
num_vector_embeds: Optional[int] = None,
patch_size: Optional[int] = 2,
activation_fn: str = "geglu",
num_embeds_ada_norm: Optional[int] = None,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
norm_type: str = "layer_norm",
block_type: str = "unidiffuser",
pre_layer_norm: bool = False,
norm_elementwise_affine: bool = True,
use_patch_pos_embed=False,
ff_final_dropout: bool = False,
):
super().__init__()
self.use_linear_projection = use_linear_projection
self.num_attention_heads = num_attention_heads
self.attention_head_dim = attention_head_dim
inner_dim = num_attention_heads * attention_head_dim
# 1. Input
# Only support patch input of shape (batch_size, num_channels, height, width) for now
assert in_channels is not None and patch_size is not None, "Patch input requires in_channels and patch_size."
assert sample_size is not None, "UTransformer2DModel over patched input must provide sample_size"
# 2. Define input layers
self.height = sample_size
self.width = sample_size
self.patch_size = patch_size
self.pos_embed = PatchEmbed(
height=sample_size,
width=sample_size,
patch_size=patch_size,
in_channels=in_channels,
embed_dim=inner_dim,
use_pos_embed=use_patch_pos_embed,
)
# 3. Define transformers blocks
# Modify this to have in_blocks ("downsample" blocks, even though we don't actually downsample), a mid_block,
# and out_blocks ("upsample" blocks). Like a U-Net, there are skip connections from in_blocks to out_blocks in
# a "U"-shaped fashion (e.g. first in_block to last out_block, etc.).
# Quick hack to make the transformer block type configurable
if block_type == "unidiffuser":
block_cls = UniDiffuserBlock
else:
block_cls = UTransformerBlock
self.transformer_in_blocks = nn.ModuleList(
[
block_cls(
inner_dim,
num_attention_heads,
attention_head_dim,
dropout=dropout,
cross_attention_dim=cross_attention_dim,
activation_fn=activation_fn,
num_embeds_ada_norm=num_embeds_ada_norm,
attention_bias=attention_bias,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
norm_type=norm_type,
pre_layer_norm=pre_layer_norm,
norm_elementwise_affine=norm_elementwise_affine,
final_dropout=ff_final_dropout,
)
for d in range(num_layers // 2)
]
)
self.transformer_mid_block = block_cls(
inner_dim,
num_attention_heads,
attention_head_dim,
dropout=dropout,
cross_attention_dim=cross_attention_dim,
activation_fn=activation_fn,
num_embeds_ada_norm=num_embeds_ada_norm,
attention_bias=attention_bias,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
norm_type=norm_type,
pre_layer_norm=pre_layer_norm,
norm_elementwise_affine=norm_elementwise_affine,
final_dropout=ff_final_dropout,
)
# For each skip connection, we use a SkipBlock (concatenation + Linear + LayerNorm) to process the inputs
# before each transformer out_block.
self.transformer_out_blocks = nn.ModuleList(
[
nn.ModuleDict(
{
"skip": SkipBlock(
inner_dim,
),
"block": block_cls(
inner_dim,
num_attention_heads,
attention_head_dim,
dropout=dropout,
cross_attention_dim=cross_attention_dim,
activation_fn=activation_fn,
num_embeds_ada_norm=num_embeds_ada_norm,
attention_bias=attention_bias,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
norm_type=norm_type,
pre_layer_norm=pre_layer_norm,
norm_elementwise_affine=norm_elementwise_affine,
final_dropout=ff_final_dropout,
),
}
)
for d in range(num_layers // 2)
]
)
# 4. Define output layers
self.out_channels = in_channels if out_channels is None else out_channels
# Following the UniDiffuser U-ViT implementation, we process the transformer output with
# a LayerNorm layer with per-element affine params
self.norm_out = nn.LayerNorm(inner_dim)
def forward(
self,
hidden_states,
encoder_hidden_states=None,
timestep=None,
class_labels=None,
cross_attention_kwargs=None,
return_dict: bool = True,
hidden_states_is_embedding: bool = False,
unpatchify: bool = True,
):
"""
Args:
hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
hidden_states
encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
self-attention.
timestep ( `torch.long`, *optional*):
Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
Optional class labels to be applied as an embedding in AdaLayerZeroNorm. Used to indicate class labels
conditioning.
cross_attention_kwargs (*optional*):
Keyword arguments to supply to the cross attention layers, if used.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
hidden_states_is_embedding (`bool`, *optional*, defaults to `False`):
Whether or not hidden_states is an embedding directly usable by the transformer. In this case we will
ignore input handling (e.g. continuous, vectorized, etc.) and directly feed hidden_states into the
transformer blocks.
unpatchify (`bool`, *optional*, defaults to `True`):
Whether to unpatchify the transformer output.
Returns:
[`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`:
[`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When
returning a tuple, the first element is the sample tensor.
"""
# 0. Check inputs
if not unpatchify and return_dict:
raise ValueError(
f"Cannot both define `unpatchify`: {unpatchify} and `return_dict`: {return_dict} since when"
f" `unpatchify` is {unpatchify} the returned output is of shape (batch_size, seq_len, hidden_dim)"
" rather than (batch_size, num_channels, height, width)."
)
# 1. Input
if not hidden_states_is_embedding:
hidden_states = self.pos_embed(hidden_states)
# 2. Blocks
# In ("downsample") blocks
skips = []
for in_block in self.transformer_in_blocks:
hidden_states = in_block(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
timestep=timestep,
cross_attention_kwargs=cross_attention_kwargs,
class_labels=class_labels,
)
skips.append(hidden_states)
# Mid block
hidden_states = self.transformer_mid_block(hidden_states)
# Out ("upsample") blocks
for out_block in self.transformer_out_blocks:
hidden_states = out_block["skip"](hidden_states, skips.pop())
hidden_states = out_block["block"](
hidden_states,
encoder_hidden_states=encoder_hidden_states,
timestep=timestep,
cross_attention_kwargs=cross_attention_kwargs,
class_labels=class_labels,
)
# 3. Output
# Don't support AdaLayerNorm for now, so no conditioning/scale/shift logic
hidden_states = self.norm_out(hidden_states)
# hidden_states = self.proj_out(hidden_states)
if unpatchify:
# unpatchify
height = width = int(hidden_states.shape[1] ** 0.5)
hidden_states = hidden_states.reshape(
shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
)
hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
output = hidden_states.reshape(
shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
)
else:
output = hidden_states
if not return_dict:
return (output,)
return Transformer2DModelOutput(sample=output)
class UniDiffuserModel(ModelMixin, ConfigMixin):
"""
Transformer model for a image-text [UniDiffuser](https://arxiv.org/pdf/2303.06555.pdf) model. This is a
modification of [`UTransformer2DModel`] with input and output heads for the VAE-embedded latent image, the
CLIP-embedded image, and the CLIP-embedded prompt (see paper for more details).
Parameters:
text_dim (`int`): The hidden dimension of the CLIP text model used to embed images.
clip_img_dim (`int`): The hidden dimension of the CLIP vision model used to embed prompts.
num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
in_channels (`int`, *optional*):
Pass if the input is continuous. The number of channels in the input.
out_channels (`int`, *optional*):
The number of output channels; if `None`, defaults to `in_channels`.
num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
norm_num_groups (`int`, *optional*, defaults to `32`):
The number of groups to use when performing Group Normalization.
cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
attention_bias (`bool`, *optional*):
Configure if the TransformerBlocks' attention should contain a bias parameter.
sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
Note that this is fixed at training time as it is used for learning a number of position embeddings. See
`ImagePositionalEmbeddings`.
num_vector_embeds (`int`, *optional*):
Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
Includes the class for the masked latent pixel.
patch_size (`int`, *optional*, defaults to 2):
The patch size to use in the patch embedding.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
The number of diffusion steps used during training. Note that this is fixed at training time as it is used
to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
up to but not more than steps than `num_embeds_ada_norm`.
use_linear_projection (int, *optional*): TODO: Not used
only_cross_attention (`bool`, *optional*):
Whether to use only cross-attention layers. In this case two cross attention layers are used in each
transformer block.
upcast_attention (`bool`, *optional*):
Whether to upcast the query and key to float32 when performing the attention calculation.
norm_type (`str`, *optional*, defaults to `"layer_norm"`):
The Layer Normalization implementation to use. Defaults to `torch.nn.LayerNorm`.
block_type (`str`, *optional*, defaults to `"unidiffuser"`):
The transformer block implementation to use. If `"unidiffuser"`, has the LayerNorms on the residual
backbone of each transformer block; otherwise has them in the attention/feedforward branches (the standard
behavior in `diffusers`.)
pre_layer_norm (`bool`, *optional*):
Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm
(`pre_layer_norm = False`).
norm_elementwise_affine (`bool`, *optional*):
Whether to use learnable per-element affine parameters during layer normalization.
use_patch_pos_embed (`bool`, *optional*):
Whether to use position embeddings inside the patch embedding layer (`PatchEmbed`).
ff_final_dropout (`bool`, *optional*):
Whether to use a final Dropout layer after the feedforward network.
use_data_type_embedding (`bool`, *optional*):
Whether to use a data type embedding. This is only relevant for UniDiffuser-v1 style models; UniDiffuser-v1
is continue-trained from UniDiffuser-v0 on non-publically-available data and accepts a `data_type`
argument, which can either be `1` to use the weights trained on non-publically-available data or `0`
otherwise. This argument is subsequently embedded by the data type embedding, if used.
"""
@register_to_config
def __init__(
self,
text_dim: int = 768,
clip_img_dim: int = 512,
num_text_tokens: int = 77,
num_attention_heads: int = 16,
attention_head_dim: int = 88,
in_channels: Optional[int] = None,
out_channels: Optional[int] = None,
num_layers: int = 1,
dropout: float = 0.0,
norm_num_groups: int = 32,
cross_attention_dim: Optional[int] = None,
attention_bias: bool = False,
sample_size: Optional[int] = None,
num_vector_embeds: Optional[int] = None,
patch_size: Optional[int] = None,
activation_fn: str = "geglu",
num_embeds_ada_norm: Optional[int] = None,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
norm_type: str = "layer_norm",
block_type: str = "unidiffuser",
pre_layer_norm: bool = False,
use_timestep_embedding=False,
norm_elementwise_affine: bool = True,
use_patch_pos_embed=False,
ff_final_dropout: bool = True,
use_data_type_embedding: bool = False,
):
super().__init__()
# 0. Handle dimensions
self.inner_dim = num_attention_heads * attention_head_dim
assert sample_size is not None, "UniDiffuserModel over patched input must provide sample_size"
self.sample_size = sample_size
self.in_channels = in_channels
self.out_channels = in_channels if out_channels is None else out_channels
self.patch_size = patch_size
# Assume image is square...
self.num_patches = (self.sample_size // patch_size) * (self.sample_size // patch_size)
# 1. Define input layers
# 1.1 Input layers for text and image input
# For now, only support patch input for VAE latent image input
self.vae_img_in = PatchEmbed(
height=sample_size,
width=sample_size,
patch_size=patch_size,
in_channels=in_channels,
embed_dim=self.inner_dim,
use_pos_embed=use_patch_pos_embed,
)
self.clip_img_in = nn.Linear(clip_img_dim, self.inner_dim)
self.text_in = nn.Linear(text_dim, self.inner_dim)
# 1.2. Timestep embeddings for t_img, t_text
self.timestep_img_proj = Timesteps(
self.inner_dim,
flip_sin_to_cos=True,
downscale_freq_shift=0,
)
self.timestep_img_embed = (
TimestepEmbedding(
self.inner_dim,
4 * self.inner_dim,
out_dim=self.inner_dim,
)
if use_timestep_embedding
else nn.Identity()
)
self.timestep_text_proj = Timesteps(
self.inner_dim,
flip_sin_to_cos=True,
downscale_freq_shift=0,
)
self.timestep_text_embed = (
TimestepEmbedding(
self.inner_dim,
4 * self.inner_dim,
out_dim=self.inner_dim,
)
if use_timestep_embedding
else nn.Identity()
)
# 1.3. Positional embedding
self.num_text_tokens = num_text_tokens
self.num_tokens = 1 + 1 + num_text_tokens + 1 + self.num_patches
self.pos_embed = nn.Parameter(torch.zeros(1, self.num_tokens, self.inner_dim))
self.pos_embed_drop = nn.Dropout(p=dropout)
trunc_normal_(self.pos_embed, std=0.02)
# 1.4. Handle data type token embeddings for UniDiffuser-V1, if necessary
self.use_data_type_embedding = use_data_type_embedding
if self.use_data_type_embedding:
self.data_type_token_embedding = nn.Embedding(2, self.inner_dim)
self.data_type_pos_embed_token = nn.Parameter(torch.zeros(1, 1, self.inner_dim))
# 2. Define transformer blocks
self.transformer = UTransformer2DModel(
num_attention_heads=num_attention_heads,
attention_head_dim=attention_head_dim,
in_channels=in_channels,
out_channels=out_channels,
num_layers=num_layers,
dropout=dropout,
norm_num_groups=norm_num_groups,
cross_attention_dim=cross_attention_dim,
attention_bias=attention_bias,
sample_size=sample_size,
num_vector_embeds=num_vector_embeds,
patch_size=patch_size,
activation_fn=activation_fn,
num_embeds_ada_norm=num_embeds_ada_norm,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
norm_type=norm_type,
block_type=block_type,
pre_layer_norm=pre_layer_norm,
norm_elementwise_affine=norm_elementwise_affine,
use_patch_pos_embed=use_patch_pos_embed,
ff_final_dropout=ff_final_dropout,
)
# 3. Define output layers
patch_dim = (patch_size**2) * out_channels
self.vae_img_out = nn.Linear(self.inner_dim, patch_dim)
self.clip_img_out = nn.Linear(self.inner_dim, clip_img_dim)
self.text_out = nn.Linear(self.inner_dim, text_dim)
@torch.jit.ignore
def no_weight_decay(self):
return {"pos_embed"}
def forward(
self,
latent_image_embeds: torch.FloatTensor,
image_embeds: torch.FloatTensor,
prompt_embeds: torch.FloatTensor,
timestep_img: Union[torch.Tensor, float, int],
timestep_text: Union[torch.Tensor, float, int],
data_type: Optional[Union[torch.Tensor, float, int]] = 1,
encoder_hidden_states=None,
cross_attention_kwargs=None,
):
"""
Args:
latent_image_embeds (`torch.FloatTensor` of shape `(batch size, latent channels, height, width)`):
Latent image representation from the VAE encoder.
image_embeds (`torch.FloatTensor` of shape `(batch size, 1, clip_img_dim)`):
CLIP-embedded image representation (unsqueezed in the first dimension).
prompt_embeds (`torch.FloatTensor` of shape `(batch size, seq_len, text_dim)`):
CLIP-embedded text representation.
timestep_img (`torch.long` or `float` or `int`):
Current denoising step for the image.
timestep_text (`torch.long` or `float` or `int`):
Current denoising step for the text.
data_type: (`torch.int` or `float` or `int`, *optional*, defaults to `1`):
Only used in UniDiffuser-v1-style models. Can be either `1`, to use weights trained on nonpublic data,
or `0` otherwise.
encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
self-attention.
cross_attention_kwargs (*optional*):
Keyword arguments to supply to the cross attention layers, if used.
Returns:
`tuple`: Returns relevant parts of the model's noise prediction: the first element of the tuple is tbe VAE
image embedding, the second element is the CLIP image embedding, and the third element is the CLIP text
embedding.
"""
batch_size = latent_image_embeds.shape[0]
# 1. Input
# 1.1. Map inputs to shape (B, N, inner_dim)
vae_hidden_states = self.vae_img_in(latent_image_embeds)
clip_hidden_states = self.clip_img_in(image_embeds)
text_hidden_states = self.text_in(prompt_embeds)
num_text_tokens, num_img_tokens = text_hidden_states.size(1), vae_hidden_states.size(1)
# 1.2. Encode image timesteps to single token (B, 1, inner_dim)
if not torch.is_tensor(timestep_img):
timestep_img = torch.tensor([timestep_img], dtype=torch.long, device=vae_hidden_states.device)
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timestep_img = timestep_img * torch.ones(batch_size, dtype=timestep_img.dtype, device=timestep_img.device)
timestep_img_token = self.timestep_img_proj(timestep_img)
# t_img_token does not contain any weights and will always return f32 tensors
# but time_embedding might be fp16, so we need to cast here.
timestep_img_token = timestep_img_token.to(dtype=self.dtype)
timestep_img_token = self.timestep_img_embed(timestep_img_token)
timestep_img_token = timestep_img_token.unsqueeze(dim=1)
# 1.3. Encode text timesteps to single token (B, 1, inner_dim)
if not torch.is_tensor(timestep_text):
timestep_text = torch.tensor([timestep_text], dtype=torch.long, device=vae_hidden_states.device)
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timestep_text = timestep_text * torch.ones(batch_size, dtype=timestep_text.dtype, device=timestep_text.device)
timestep_text_token = self.timestep_text_proj(timestep_text)
# t_text_token does not contain any weights and will always return f32 tensors
# but time_embedding might be fp16, so we need to cast here.
timestep_text_token = timestep_text_token.to(dtype=self.dtype)
timestep_text_token = self.timestep_text_embed(timestep_text_token)
timestep_text_token = timestep_text_token.unsqueeze(dim=1)
# 1.4. Concatenate all of the embeddings together.
if self.use_data_type_embedding:
assert data_type is not None, "data_type must be supplied if the model uses a data type embedding"
if not torch.is_tensor(data_type):
data_type = torch.tensor([data_type], dtype=torch.int, device=vae_hidden_states.device)
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
data_type = data_type * torch.ones(batch_size, dtype=data_type.dtype, device=data_type.device)
data_type_token = self.data_type_token_embedding(data_type).unsqueeze(dim=1)
hidden_states = torch.cat(
[
timestep_img_token,
timestep_text_token,
data_type_token,
text_hidden_states,
clip_hidden_states,
vae_hidden_states,
],
dim=1,
)
else:
hidden_states = torch.cat(
[timestep_img_token, timestep_text_token, text_hidden_states, clip_hidden_states, vae_hidden_states],
dim=1,
)
# 1.5. Prepare the positional embeddings and add to hidden states
# Note: I think img_vae should always have the proper shape, so there's no need to interpolate
# the position embeddings.
if self.use_data_type_embedding:
pos_embed = torch.cat(
[self.pos_embed[:, : 1 + 1, :], self.data_type_pos_embed_token, self.pos_embed[:, 1 + 1 :, :]], dim=1
)
else:
pos_embed = self.pos_embed
hidden_states = hidden_states + pos_embed
hidden_states = self.pos_embed_drop(hidden_states)
# 2. Blocks
hidden_states = self.transformer(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
timestep=None,
class_labels=None,
cross_attention_kwargs=cross_attention_kwargs,
return_dict=False,
hidden_states_is_embedding=True,
unpatchify=False,
)[0]
# 3. Output
# Split out the predicted noise representation.
if self.use_data_type_embedding:
(
t_img_token_out,
t_text_token_out,
data_type_token_out,
text_out,
img_clip_out,
img_vae_out,
) = hidden_states.split((1, 1, 1, num_text_tokens, 1, num_img_tokens), dim=1)
else:
t_img_token_out, t_text_token_out, text_out, img_clip_out, img_vae_out = hidden_states.split(
(1, 1, num_text_tokens, 1, num_img_tokens), dim=1
)
img_vae_out = self.vae_img_out(img_vae_out)
# unpatchify
height = width = int(img_vae_out.shape[1] ** 0.5)
img_vae_out = img_vae_out.reshape(
shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
)
img_vae_out = torch.einsum("nhwpqc->nchpwq", img_vae_out)
img_vae_out = img_vae_out.reshape(
shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
)
img_clip_out = self.clip_img_out(img_clip_out)
text_out = self.text_out(text_out)
return img_vae_out, img_clip_out, text_out
|