Spaces:
Runtime error
Runtime error
File size: 62,890 Bytes
a220803 |
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 |
import torch
import os
import json
from dataclasses import dataclass
from einops import rearrange, repeat
from typing import Any, Dict, Optional, Tuple
from diffusers.models import Transformer2DModel
from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate
from diffusers.models.embeddings import get_1d_sincos_pos_embed_from_grid, ImagePositionalEmbeddings, CaptionProjection
# from diffusers.models.embeddings import PatchEmbed, CombinedTimestepSizeEmbeddings
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
import torch
import torch.nn.functional as F
from torch import nn
from opensora.models.diffusion.utils.pos_embed import get_1d_sincos_pos_embed
from .modules import PatchEmbed, BasicTransformerBlock, BasicTransformerBlock_, AdaLayerNormSingle, Transformer3DModelOutput
class Latte(ModelMixin, ConfigMixin):
_supports_gradient_checkpointing = True
"""
A 2D Transformer model for image-like data.
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*):
The number of channels in the input and output (specify if the input is **continuous**).
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.
cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
This is fixed during training since it is used to learn a number of position embeddings.
num_vector_embeds (`int`, *optional*):
The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
Includes the class for the masked latent pixel.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
num_embeds_ada_norm ( `int`, *optional*):
The number of diffusion steps used during training. Pass if at least one of the norm_layers is
`AdaLayerNorm`. This is fixed during training since 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 steps than `num_embeds_ada_norm`.
attention_bias (`bool`, *optional*):
Configure if the `TransformerBlocks` attention should contain a bias parameter.
"""
@register_to_config
def __init__(
self,
num_attention_heads: int = 16,
patch_size_t: int = 1,
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,
double_self_attention: bool = False,
upcast_attention: bool = False,
norm_type: str = "layer_norm",
norm_elementwise_affine: bool = True,
norm_eps: float = 1e-5,
attention_type: str = "default",
caption_channels: int = None,
video_length: int = 16,
attention_mode: str = 'flash'
):
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
self.video_length = video_length
conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
# 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
# Define whether input is continuous or discrete depending on configuration
self.is_input_continuous = (in_channels is not None) and (patch_size is None)
self.is_input_vectorized = num_vector_embeds is not None
self.is_input_patches = in_channels is not None and patch_size is not None
if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
deprecation_message = (
f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
" incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
" Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
" results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
" would be very nice if you could open a Pull request for the `transformer/config.json` file"
)
deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
norm_type = "ada_norm"
if self.is_input_continuous and self.is_input_vectorized:
raise ValueError(
f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
" sure that either `in_channels` or `num_vector_embeds` is None."
)
elif self.is_input_vectorized and self.is_input_patches:
raise ValueError(
f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
" sure that either `num_vector_embeds` or `num_patches` is None."
)
elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
raise ValueError(
f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
)
# 2. Define input layers
if self.is_input_continuous:
self.in_channels = in_channels
self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
if use_linear_projection:
self.proj_in = linear_cls(in_channels, inner_dim)
else:
self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
elif self.is_input_vectorized:
assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
self.height = sample_size[0]
self.width = sample_size[1]
self.num_vector_embeds = num_vector_embeds
self.num_latent_pixels = self.height * self.width
self.latent_image_embedding = ImagePositionalEmbeddings(
num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
)
elif self.is_input_patches:
assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
self.height = sample_size[0]
self.width = sample_size[1]
self.patch_size = patch_size
interpolation_scale = self.config.sample_size[0] // 64 # => 64 (= 512 pixart) has interpolation scale 1
interpolation_scale = max(interpolation_scale, 1)
self.pos_embed = PatchEmbed(
height=sample_size[0],
width=sample_size[1],
patch_size=patch_size,
in_channels=in_channels,
embed_dim=inner_dim,
interpolation_scale=interpolation_scale,
)
# 3. Define transformers blocks
self.transformer_blocks = nn.ModuleList(
[
BasicTransformerBlock_(
inner_dim,
num_attention_heads,
attention_head_dim,
dropout=dropout,
cross_attention_dim=None, ############## unconditon do not need cross attn
activation_fn=activation_fn,
num_embeds_ada_norm=num_embeds_ada_norm,
attention_bias=attention_bias,
only_cross_attention=only_cross_attention,
double_self_attention=False,
upcast_attention=upcast_attention,
norm_type=norm_type,
norm_elementwise_affine=norm_elementwise_affine,
norm_eps=norm_eps,
attention_type=attention_type,
attention_mode=attention_mode,
)
for d in range(num_layers)
]
)
# Define temporal transformers blocks
self.temporal_transformer_blocks = nn.ModuleList(
[
BasicTransformerBlock_( # one attention
inner_dim,
num_attention_heads, # num_attention_heads
attention_head_dim, # attention_head_dim 72
dropout=dropout,
cross_attention_dim=None,
activation_fn=activation_fn,
num_embeds_ada_norm=num_embeds_ada_norm,
attention_bias=attention_bias,
only_cross_attention=only_cross_attention,
double_self_attention=False,
upcast_attention=upcast_attention,
norm_type=norm_type,
norm_elementwise_affine=norm_elementwise_affine,
norm_eps=norm_eps,
attention_type=attention_type,
attention_mode=attention_mode,
)
for d in range(num_layers)
]
)
# 4. Define output layers
self.out_channels = in_channels if out_channels is None else out_channels
if self.is_input_continuous:
# TODO: should use out_channels for continuous projections
if use_linear_projection:
self.proj_out = linear_cls(inner_dim, in_channels)
else:
self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
elif self.is_input_vectorized:
self.norm_out = nn.LayerNorm(inner_dim)
self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
elif self.is_input_patches and norm_type != "ada_norm_single":
self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
elif self.is_input_patches and norm_type == "ada_norm_single":
self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim ** 0.5)
self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
# 5. PixArt-Alpha blocks.
self.adaln_single = None
self.use_additional_conditions = False
if norm_type == "ada_norm_single":
# self.use_additional_conditions = self.config.sample_size[0] == 128 # False, 128 -> 1024
# TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
# additional conditions until we find better name
self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
self.caption_projection = None
if caption_channels is not None:
self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)
self.gradient_checkpointing = False
# define temporal positional embedding
# temp_pos_embed = self.get_1d_sincos_temp_embed(inner_dim, video_length) # 1152 hidden size
# self.register_buffer("temp_pos_embed", torch.from_numpy(temp_pos_embed).float().unsqueeze(0), persistent=False)
interpolation_scale = self.config.video_length // 5 # => 5 (= 5 our causalvideovae) has interpolation scale 1
interpolation_scale = max(interpolation_scale, 1)
temp_pos_embed = get_1d_sincos_pos_embed(inner_dim, video_length, interpolation_scale=interpolation_scale) # 1152 hidden size
self.register_buffer("temp_pos_embed", torch.from_numpy(temp_pos_embed).float().unsqueeze(0), persistent=False)
def _set_gradient_checkpointing(self, module, value=False):
self.gradient_checkpointing = value
def forward(
self,
hidden_states: torch.Tensor,
timestep: Optional[torch.LongTensor] = None,
encoder_hidden_states: Optional[torch.Tensor] = None,
added_cond_kwargs: Dict[str, torch.Tensor] = None,
class_labels: Optional[torch.LongTensor] = None,
cross_attention_kwargs: Dict[str, Any] = None,
attention_mask: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
use_image_num: int = 0,
enable_temporal_attentions: bool = True,
return_dict: bool = True,
):
"""
The [`Transformer2DModel`] forward method.
Args:
hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, frame, channel, height, width)` if continuous):
Input `hidden_states`.
encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
self-attention.
timestep ( `torch.LongTensor`, *optional*):
Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
`AdaLayerZeroNorm`.
cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
`self.processor` in
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
attention_mask ( `torch.Tensor`, *optional*):
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
negative values to the attention scores corresponding to "discard" tokens.
encoder_attention_mask ( `torch.Tensor`, *optional*):
Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
* Mask `(batch, sequence_length)` True = keep, False = discard.
* Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
above. This bias will be added to the cross-attention scores.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
tuple.
Returns:
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
`tuple` where the first element is the sample tensor.
"""
input_batch_size, c, frame, h, w = hidden_states.shape
frame = frame - use_image_num
hidden_states = rearrange(hidden_states, 'b c f h w -> (b f) c h w').contiguous()
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
# we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
# we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
# expects mask of shape:
# [batch, key_tokens]
# adds singleton query_tokens dimension:
# [batch, 1, key_tokens]
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
if attention_mask is not None and attention_mask.ndim == 2:
# assume that mask is expressed as:
# (1 = keep, 0 = discard)
# convert mask into a bias that can be added to attention scores:
# (keep = +0, discard = -10000.0)
attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
attention_mask = attention_mask.unsqueeze(1)
# convert encoder_attention_mask to a bias the same way we do for attention_mask
# if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: # ndim == 2 means no image joint
# encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
# encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
# encoder_attention_mask = repeat(encoder_attention_mask, 'b 1 l -> (b f) 1 l', f=frame).contiguous()
# elif encoder_attention_mask is not None and encoder_attention_mask.ndim == 3: # ndim == 3 means image joint
# encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
# encoder_attention_mask_video = encoder_attention_mask[:, :1, ...]
# encoder_attention_mask_video = repeat(encoder_attention_mask_video, 'b 1 l -> b (1 f) l',
# f=frame).contiguous()
# encoder_attention_mask_image = encoder_attention_mask[:, 1:, ...]
# encoder_attention_mask = torch.cat([encoder_attention_mask_video, encoder_attention_mask_image], dim=1)
# encoder_attention_mask = rearrange(encoder_attention_mask, 'b n l -> (b n) l').contiguous().unsqueeze(1)
# Retrieve lora scale.
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
# 1. Input
if self.is_input_patches: # here
height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
num_patches = height * width
hidden_states = self.pos_embed(hidden_states.to(self.dtype)) # alrady add positional embeddings
if self.adaln_single is not None:
if self.use_additional_conditions and added_cond_kwargs is None:
raise ValueError(
"`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
)
# batch_size = hidden_states.shape[0]
batch_size = input_batch_size
timestep, embedded_timestep = self.adaln_single(
timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
)
# 2. Blocks
# if self.caption_projection is not None:
# batch_size = hidden_states.shape[0]
# encoder_hidden_states = self.caption_projection(encoder_hidden_states) # 3 120 1152
#
# if use_image_num != 0 and self.training:
# encoder_hidden_states_video = encoder_hidden_states[:, :1, ...]
# encoder_hidden_states_video = repeat(encoder_hidden_states_video, 'b 1 t d -> b (1 f) t d',
# f=frame).contiguous()
# encoder_hidden_states_image = encoder_hidden_states[:, 1:, ...]
# encoder_hidden_states = torch.cat([encoder_hidden_states_video, encoder_hidden_states_image], dim=1)
# encoder_hidden_states_spatial = rearrange(encoder_hidden_states, 'b f t d -> (b f) t d').contiguous()
# else:
# encoder_hidden_states_spatial = repeat(encoder_hidden_states, 'b t d -> (b f) t d',
# f=frame).contiguous()
# prepare timesteps for spatial and temporal block
timestep_spatial = repeat(timestep, 'b d -> (b f) d', f=frame + use_image_num).contiguous()
timestep_temp = repeat(timestep, 'b d -> (b p) d', p=num_patches).contiguous()
for i, (spatial_block, temp_block) in enumerate(zip(self.transformer_blocks, self.temporal_transformer_blocks)):
if self.training and self.gradient_checkpointing:
hidden_states = torch.utils.checkpoint.checkpoint(
spatial_block,
hidden_states,
attention_mask,
None, # encoder_hidden_states_spatial
None, # encoder_attention_mask
timestep_spatial,
cross_attention_kwargs,
class_labels,
use_reentrant=False,
)
if enable_temporal_attentions:
hidden_states = rearrange(hidden_states, '(b f) t d -> (b t) f d', b=input_batch_size).contiguous()
if use_image_num != 0: # image-video joitn training
hidden_states_video = hidden_states[:, :frame, ...]
hidden_states_image = hidden_states[:, frame:, ...]
if i == 0:
hidden_states_video = hidden_states_video + self.temp_pos_embed
hidden_states_video = torch.utils.checkpoint.checkpoint(
temp_block,
hidden_states_video,
None, # attention_mask
None, # encoder_hidden_states
None, # encoder_attention_mask
timestep_temp,
cross_attention_kwargs,
class_labels,
use_reentrant=False,
)
hidden_states = torch.cat([hidden_states_video, hidden_states_image], dim=1)
hidden_states = rearrange(hidden_states, '(b t) f d -> (b f) t d',
b=input_batch_size).contiguous()
else:
if i == 0:
hidden_states = hidden_states + self.temp_pos_embed
hidden_states = torch.utils.checkpoint.checkpoint(
temp_block,
hidden_states,
None, # attention_mask
None, # encoder_hidden_states
None, # encoder_attention_mask
timestep_temp,
cross_attention_kwargs,
class_labels,
use_reentrant=False,
)
hidden_states = rearrange(hidden_states, '(b t) f d -> (b f) t d',
b=input_batch_size).contiguous()
else:
hidden_states = spatial_block(
hidden_states,
attention_mask,
None, # encoder_hidden_states_spatial
None, # encoder_attention_mask
timestep_spatial,
cross_attention_kwargs,
class_labels,
)
if enable_temporal_attentions:
hidden_states = rearrange(hidden_states, '(b f) t d -> (b t) f d', b=input_batch_size).contiguous()
if use_image_num != 0 and self.training:
hidden_states_video = hidden_states[:, :frame, ...]
hidden_states_image = hidden_states[:, frame:, ...]
hidden_states_video = temp_block(
hidden_states_video,
None, # attention_mask
None, # encoder_hidden_states
None, # encoder_attention_mask
timestep_temp,
cross_attention_kwargs,
class_labels,
)
hidden_states = torch.cat([hidden_states_video, hidden_states_image], dim=1)
hidden_states = rearrange(hidden_states, '(b t) f d -> (b f) t d',
b=input_batch_size).contiguous()
else:
if i == 0:
hidden_states = hidden_states + self.temp_pos_embed
hidden_states = temp_block(
hidden_states,
None, # attention_mask
None, # encoder_hidden_states
None, # encoder_attention_mask
timestep_temp,
cross_attention_kwargs,
class_labels,
)
hidden_states = rearrange(hidden_states, '(b t) f d -> (b f) t d',
b=input_batch_size).contiguous()
if self.is_input_patches:
if self.config.norm_type != "ada_norm_single":
conditioning = self.transformer_blocks[0].norm1.emb(
timestep, class_labels, hidden_dtype=hidden_states.dtype
)
shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
hidden_states = self.proj_out_2(hidden_states)
elif self.config.norm_type == "ada_norm_single":
embedded_timestep = repeat(embedded_timestep, 'b d -> (b f) d', f=frame + use_image_num).contiguous()
shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
hidden_states = self.norm_out(hidden_states)
# Modulation
hidden_states = hidden_states * (1 + scale) + shift
hidden_states = self.proj_out(hidden_states)
# unpatchify
if self.adaln_single is None:
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)
)
output = rearrange(output, '(b f) c h w -> b c f h w', b=input_batch_size).contiguous()
if not return_dict:
return (output,)
return Transformer3DModelOutput(sample=output)
# def get_1d_sincos_temp_embed(self, embed_dim, length):
# pos = torch.arange(0, length).unsqueeze(1)
# return get_1d_sincos_pos_embed_from_grid(embed_dim, pos)
@classmethod
def from_pretrained_2d(cls, pretrained_model_path, subfolder=None, **kwargs):
if subfolder is not None:
pretrained_model_path = os.path.join(pretrained_model_path, subfolder)
config_file = os.path.join(pretrained_model_path, 'config.json')
if not os.path.isfile(config_file):
raise RuntimeError(f"{config_file} does not exist")
with open(config_file, "r") as f:
config = json.load(f)
model = cls.from_config(config, **kwargs)
# model_files = [
# os.path.join(pretrained_model_path, 'diffusion_pytorch_model.bin'),
# os.path.join(pretrained_model_path, 'diffusion_pytorch_model.safetensors')
# ]
# model_file = None
# for fp in model_files:
# if os.path.exists(fp):
# model_file = fp
# if not model_file:
# raise RuntimeError(f"{model_file} does not exist")
# if model_file.split(".")[-1] == "safetensors":
# from safetensors import safe_open
# state_dict = {}
# with safe_open(model_file, framework="pt", device="cpu") as f:
# for key in f.keys():
# state_dict[key] = f.get_tensor(key)
# else:
# state_dict = torch.load(model_file, map_location="cpu")
# for k, v in model.state_dict().items():
# if 'temporal_transformer_blocks' in k:
# state_dict.update({k: v})
# model.load_state_dict(state_dict)
return model
def forward_with_cfg(self, x, timestep, class_labels=None, cfg_scale=7.0, attention_mask=None):
"""
Forward pass of Latte, but also batches the unconditional forward pass for classifier-free guidance.
"""
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
half = x[: len(x) // 2]
combined = torch.cat([half, half], dim=0)
model_out = self.forward(combined, timestep, class_labels=class_labels, attention_mask=attention_mask)
# For exact reproducibility reasons, we apply classifier-free guidance on only
# three channels by default. The standard approach to cfg applies it to all channels.
# This can be done by uncommenting the following line and commenting-out the line following that.
eps, rest = model_out[:, :, :self.in_channels], model_out[:, :, self.in_channels:]
cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
eps = torch.cat([half_eps, half_eps], dim=0)
return torch.cat([eps, rest], dim=2)
class LatteT2V(ModelMixin, ConfigMixin):
_supports_gradient_checkpointing = True
"""
A 2D Transformer model for image-like data.
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*):
The number of channels in the input and output (specify if the input is **continuous**).
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.
cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
This is fixed during training since it is used to learn a number of position embeddings.
num_vector_embeds (`int`, *optional*):
The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
Includes the class for the masked latent pixel.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
num_embeds_ada_norm ( `int`, *optional*):
The number of diffusion steps used during training. Pass if at least one of the norm_layers is
`AdaLayerNorm`. This is fixed during training since 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 steps than `num_embeds_ada_norm`.
attention_bias (`bool`, *optional*):
Configure if the `TransformerBlocks` attention should contain a bias parameter.
"""
@register_to_config
def __init__(
self,
num_attention_heads: int = 16,
patch_size_t: int = 1,
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,
double_self_attention: bool = False,
upcast_attention: bool = False,
norm_type: str = "layer_norm",
norm_elementwise_affine: bool = True,
norm_eps: float = 1e-5,
attention_type: str = "default",
caption_channels: int = None,
video_length: int = 16,
attention_mode: str = 'flash'
):
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
self.video_length = video_length
conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
# 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
# Define whether input is continuous or discrete depending on configuration
self.is_input_continuous = (in_channels is not None) and (patch_size is None)
self.is_input_vectorized = num_vector_embeds is not None
self.is_input_patches = in_channels is not None and patch_size is not None
if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
deprecation_message = (
f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
" incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
" Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
" results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
" would be very nice if you could open a Pull request for the `transformer/config.json` file"
)
deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
norm_type = "ada_norm"
if self.is_input_continuous and self.is_input_vectorized:
raise ValueError(
f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
" sure that either `in_channels` or `num_vector_embeds` is None."
)
elif self.is_input_vectorized and self.is_input_patches:
raise ValueError(
f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
" sure that either `num_vector_embeds` or `num_patches` is None."
)
elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
raise ValueError(
f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
)
# 2. Define input layers
if self.is_input_continuous:
self.in_channels = in_channels
self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
if use_linear_projection:
self.proj_in = linear_cls(in_channels, inner_dim)
else:
self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
elif self.is_input_vectorized:
assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
self.height = sample_size[0]
self.width = sample_size[1]
self.num_vector_embeds = num_vector_embeds
self.num_latent_pixels = self.height * self.width
self.latent_image_embedding = ImagePositionalEmbeddings(
num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
)
elif self.is_input_patches:
assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
self.height = sample_size[0]
self.width = sample_size[1]
self.patch_size = patch_size
interpolation_scale = self.config.sample_size[0] // 64 # => 64 (= 512 pixart) has interpolation scale 1
interpolation_scale = max(interpolation_scale, 1)
self.pos_embed = PatchEmbed(
height=sample_size[0],
width=sample_size[1],
patch_size=patch_size,
in_channels=in_channels,
embed_dim=inner_dim,
interpolation_scale=interpolation_scale,
)
# 3. Define transformers blocks, spatial attention
self.transformer_blocks = nn.ModuleList(
[
BasicTransformerBlock(
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,
double_self_attention=double_self_attention,
upcast_attention=upcast_attention,
norm_type=norm_type,
norm_elementwise_affine=norm_elementwise_affine,
norm_eps=norm_eps,
attention_type=attention_type,
attention_mode=attention_mode
)
for d in range(num_layers)
]
)
# Define temporal transformers blocks
self.temporal_transformer_blocks = nn.ModuleList(
[
BasicTransformerBlock_( # one attention
inner_dim,
num_attention_heads, # num_attention_heads
attention_head_dim, # attention_head_dim 72
dropout=dropout,
cross_attention_dim=None,
activation_fn=activation_fn,
num_embeds_ada_norm=num_embeds_ada_norm,
attention_bias=attention_bias,
only_cross_attention=only_cross_attention,
double_self_attention=False,
upcast_attention=upcast_attention,
norm_type=norm_type,
norm_elementwise_affine=norm_elementwise_affine,
norm_eps=norm_eps,
attention_type=attention_type,
attention_mode=attention_mode
)
for d in range(num_layers)
]
)
# 4. Define output layers
self.out_channels = in_channels if out_channels is None else out_channels
if self.is_input_continuous:
# TODO: should use out_channels for continuous projections
if use_linear_projection:
self.proj_out = linear_cls(inner_dim, in_channels)
else:
self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
elif self.is_input_vectorized:
self.norm_out = nn.LayerNorm(inner_dim)
self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
elif self.is_input_patches and norm_type != "ada_norm_single":
self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
elif self.is_input_patches and norm_type == "ada_norm_single":
self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim ** 0.5)
self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
# 5. PixArt-Alpha blocks.
self.adaln_single = None
self.use_additional_conditions = False
if norm_type == "ada_norm_single":
# self.use_additional_conditions = self.config.sample_size[0] == 128 # False, 128 -> 1024
# TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
# additional conditions until we find better name
self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
self.caption_projection = None
if caption_channels is not None:
self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)
self.gradient_checkpointing = False
# define temporal positional embedding
# temp_pos_embed = self.get_1d_sincos_temp_embed(inner_dim, video_length) # 1152 hidden size
interpolation_scale = self.config.video_length // 5 # => 5 (= 5 our causalvideovae) has interpolation scale 1
interpolation_scale = max(interpolation_scale, 1)
temp_pos_embed = get_1d_sincos_pos_embed(inner_dim, video_length, interpolation_scale=interpolation_scale) # 1152 hidden size
self.register_buffer("temp_pos_embed", torch.from_numpy(temp_pos_embed).float().unsqueeze(0), persistent=False)
def _set_gradient_checkpointing(self, module, value=False):
self.gradient_checkpointing = value
def forward(
self,
hidden_states: torch.Tensor,
timestep: Optional[torch.LongTensor] = None,
encoder_hidden_states: Optional[torch.Tensor] = None,
added_cond_kwargs: Dict[str, torch.Tensor] = None,
class_labels: Optional[torch.LongTensor] = None,
cross_attention_kwargs: Dict[str, Any] = None,
attention_mask: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
use_image_num: int = 0,
enable_temporal_attentions: bool = True,
return_dict: bool = True,
):
"""
The [`Transformer2DModel`] forward method.
Args:
hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, frame, channel, height, width)` if continuous):
Input `hidden_states`.
encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
self-attention.
timestep ( `torch.LongTensor`, *optional*):
Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
`AdaLayerZeroNorm`.
cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
`self.processor` in
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
attention_mask ( `torch.Tensor`, *optional*):
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
negative values to the attention scores corresponding to "discard" tokens.
encoder_attention_mask ( `torch.Tensor`, *optional*):
Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
* Mask `(batch, sequence_length)` True = keep, False = discard.
* Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
above. This bias will be added to the cross-attention scores.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
tuple.
Returns:
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
`tuple` where the first element is the sample tensor.
"""
input_batch_size, c, frame, h, w = hidden_states.shape
# print(hidden_states.shape, input_batch_size, c, frame, h, w, use_image_num)
# print(timestep)
# print(encoder_hidden_states.shape)
# print(encoder_attention_mask.shape)
frame = frame - use_image_num # 20-4=16
hidden_states = rearrange(hidden_states, 'b c f h w -> (b f) c h w').contiguous()
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
# we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
# we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
# expects mask of shape:
# [batch, key_tokens]
# adds singleton query_tokens dimension:
# [batch, 1, key_tokens]
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
if attention_mask is not None and attention_mask.ndim == 2:
# assume that mask is expressed as:
# (1 = keep, 0 = discard)
# convert mask into a bias that can be added to attention scores:
# (keep = +0, discard = -10000.0)
attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
attention_mask = attention_mask.unsqueeze(1)
attention_mask = attention_mask.to(self.dtype)
# 1 + 4, 1 -> video condition, 4 -> image condition
# convert encoder_attention_mask to a bias the same way we do for attention_mask
if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: # ndim == 2 means no image joint
encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
encoder_attention_mask = repeat(encoder_attention_mask, 'b 1 l -> (b f) 1 l', f=frame).contiguous()
encoder_attention_mask = encoder_attention_mask.to(self.dtype)
elif encoder_attention_mask is not None and encoder_attention_mask.ndim == 3: # ndim == 3 means image joint
encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
encoder_attention_mask_video = encoder_attention_mask[:, :1, ...]
encoder_attention_mask_video = repeat(encoder_attention_mask_video, 'b 1 l -> b (1 f) l',
f=frame).contiguous()
encoder_attention_mask_image = encoder_attention_mask[:, 1:, ...]
encoder_attention_mask = torch.cat([encoder_attention_mask_video, encoder_attention_mask_image], dim=1)
encoder_attention_mask = rearrange(encoder_attention_mask, 'b n l -> (b n) l').contiguous().unsqueeze(1)
encoder_attention_mask = encoder_attention_mask.to(self.dtype)
# Retrieve lora scale.
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
# 1. Input
if self.is_input_patches: # here
height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
num_patches = height * width
hidden_states = self.pos_embed(hidden_states.to(self.dtype)) # alrady add positional embeddings
if self.adaln_single is not None:
if self.use_additional_conditions and added_cond_kwargs is None:
raise ValueError(
"`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
)
# batch_size = hidden_states.shape[0]
batch_size = input_batch_size
timestep, embedded_timestep = self.adaln_single(
timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
)
# 2. Blocks
if self.caption_projection is not None:
batch_size = hidden_states.shape[0]
encoder_hidden_states = self.caption_projection(encoder_hidden_states.to(self.dtype)) # 3 120 1152
if use_image_num != 0 and self.training:
encoder_hidden_states_video = encoder_hidden_states[:, :1, ...]
encoder_hidden_states_video = repeat(encoder_hidden_states_video, 'b 1 t d -> b (1 f) t d', f=frame).contiguous()
encoder_hidden_states_image = encoder_hidden_states[:, 1:, ...]
encoder_hidden_states = torch.cat([encoder_hidden_states_video, encoder_hidden_states_image], dim=1)
encoder_hidden_states_spatial = rearrange(encoder_hidden_states, 'b f t d -> (b f) t d').contiguous()
else:
encoder_hidden_states_spatial = repeat(encoder_hidden_states, 'b t d -> (b f) t d', f=frame).contiguous()
# prepare timesteps for spatial and temporal block
timestep_spatial = repeat(timestep, 'b d -> (b f) d', f=frame + use_image_num).contiguous()
timestep_temp = repeat(timestep, 'b d -> (b p) d', p=num_patches).contiguous()
for i, (spatial_block, temp_block) in enumerate(zip(self.transformer_blocks, self.temporal_transformer_blocks)):
if self.training and self.gradient_checkpointing:
hidden_states = torch.utils.checkpoint.checkpoint(
spatial_block,
hidden_states,
attention_mask,
encoder_hidden_states_spatial,
encoder_attention_mask,
timestep_spatial,
cross_attention_kwargs,
class_labels,
use_reentrant=False,
)
if enable_temporal_attentions:
hidden_states = rearrange(hidden_states, '(b f) t d -> (b t) f d', b=input_batch_size).contiguous()
if use_image_num != 0: # image-video joitn training
hidden_states_video = hidden_states[:, :frame, ...]
hidden_states_image = hidden_states[:, frame:, ...]
if i == 0:
hidden_states_video = hidden_states_video + self.temp_pos_embed
hidden_states_video = torch.utils.checkpoint.checkpoint(
temp_block,
hidden_states_video,
None, # attention_mask
None, # encoder_hidden_states
None, # encoder_attention_mask
timestep_temp,
cross_attention_kwargs,
class_labels,
use_reentrant=False,
)
hidden_states = torch.cat([hidden_states_video, hidden_states_image], dim=1)
hidden_states = rearrange(hidden_states, '(b t) f d -> (b f) t d',
b=input_batch_size).contiguous()
else:
if i == 0:
hidden_states = hidden_states + self.temp_pos_embed
hidden_states = torch.utils.checkpoint.checkpoint(
temp_block,
hidden_states,
None, # attention_mask
None, # encoder_hidden_states
None, # encoder_attention_mask
timestep_temp,
cross_attention_kwargs,
class_labels,
use_reentrant=False,
)
hidden_states = rearrange(hidden_states, '(b t) f d -> (b f) t d',
b=input_batch_size).contiguous()
else:
hidden_states = spatial_block(
hidden_states,
attention_mask,
encoder_hidden_states_spatial,
encoder_attention_mask,
timestep_spatial,
cross_attention_kwargs,
class_labels,
)
if enable_temporal_attentions:
# b c f h w, f = 16 + 4
hidden_states = rearrange(hidden_states, '(b f) t d -> (b t) f d', b=input_batch_size).contiguous()
if use_image_num != 0 and self.training:
hidden_states_video = hidden_states[:, :frame, ...]
hidden_states_image = hidden_states[:, frame:, ...]
hidden_states_video = temp_block(
hidden_states_video,
None, # attention_mask
None, # encoder_hidden_states
None, # encoder_attention_mask
timestep_temp,
cross_attention_kwargs,
class_labels,
)
hidden_states = torch.cat([hidden_states_video, hidden_states_image], dim=1)
hidden_states = rearrange(hidden_states, '(b t) f d -> (b f) t d',
b=input_batch_size).contiguous()
else:
if i == 0:
hidden_states = hidden_states + self.temp_pos_embed
hidden_states = temp_block(
hidden_states,
None, # attention_mask
None, # encoder_hidden_states
None, # encoder_attention_mask
timestep_temp,
cross_attention_kwargs,
class_labels,
)
hidden_states = rearrange(hidden_states, '(b t) f d -> (b f) t d',
b=input_batch_size).contiguous()
if self.is_input_patches:
if self.config.norm_type != "ada_norm_single":
conditioning = self.transformer_blocks[0].norm1.emb(
timestep, class_labels, hidden_dtype=hidden_states.dtype
)
shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
hidden_states = self.proj_out_2(hidden_states)
elif self.config.norm_type == "ada_norm_single":
embedded_timestep = repeat(embedded_timestep, 'b d -> (b f) d', f=frame + use_image_num).contiguous()
shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
hidden_states = self.norm_out(hidden_states)
# Modulation
hidden_states = hidden_states * (1 + scale) + shift
hidden_states = self.proj_out(hidden_states)
# unpatchify
if self.adaln_single is None:
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)
)
output = rearrange(output, '(b f) c h w -> b c f h w', b=input_batch_size).contiguous()
if not return_dict:
return (output,)
return Transformer3DModelOutput(sample=output)
# def get_1d_sincos_temp_embed(self, embed_dim, length):
# pos = torch.arange(0, length).unsqueeze(1)
# return get_1d_sincos_pos_embed_from_grid(embed_dim, pos)
@classmethod
def from_pretrained_2d(cls, pretrained_model_path, subfolder=None, **kwargs):
if subfolder is not None:
pretrained_model_path = os.path.join(pretrained_model_path, subfolder)
config_file = os.path.join(pretrained_model_path, 'config.json')
if not os.path.isfile(config_file):
raise RuntimeError(f"{config_file} does not exist")
with open(config_file, "r") as f:
config = json.load(f)
model = cls.from_config(config, **kwargs)
# model_files = [
# os.path.join(pretrained_model_path, 'diffusion_pytorch_model.bin'),
# os.path.join(pretrained_model_path, 'diffusion_pytorch_model.safetensors')
# ]
# model_file = None
# for fp in model_files:
# if os.path.exists(fp):
# model_file = fp
# if not model_file:
# raise RuntimeError(f"{model_file} does not exist")
# if model_file.split(".")[-1] == "safetensors":
# from safetensors import safe_open
# state_dict = {}
# with safe_open(model_file, framework="pt", device="cpu") as f:
# for key in f.keys():
# state_dict[key] = f.get_tensor(key)
# else:
# state_dict = torch.load(model_file, map_location="cpu")
# for k, v in model.state_dict().items():
# if 'temporal_transformer_blocks' in k:
# state_dict.update({k: v})
# model.load_state_dict(state_dict)
return model
# depth = num_layers * 2
def Latte_XL_122(**kwargs):
return Latte(num_layers=28, attention_head_dim=72, num_attention_heads=16, patch_size_t=1, patch_size=2, norm_type="ada_norm_single", **kwargs)
def LatteClass_XL_122(**kwargs):
return Latte(num_layers=28, attention_head_dim=72, num_attention_heads=16, patch_size_t=1, patch_size=2, norm_type="ada_norm_zero", **kwargs)
def LatteT2V_XL_122(**kwargs):
return LatteT2V(num_layers=28, attention_head_dim=72, num_attention_heads=16, patch_size_t=1, patch_size=2,
norm_type="ada_norm_single", caption_channels=4096, cross_attention_dim=1152, **kwargs)
Latte_models = {
"Latte-XL/122": Latte_XL_122,
"LatteClass-XL/122": LatteClass_XL_122,
"LatteT2V-XL/122": LatteT2V_XL_122,
}
if __name__ == '__main__':
a = json.load(open('./config.json', 'r'))
model = LatteT2V.from_config(a)
ckpt = torch.load(r"E:\下载\t2v.pt", map_location='cpu')['model']
model.load_state_dict(ckpt)
print(model) |