Spaces:
Paused
Paused
File size: 39,484 Bytes
4f2a492 |
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 |
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Any, Dict, Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.utils.checkpoint
from ..configuration_utils import ConfigMixin, register_to_config
from ..loaders import UNet2DConditionLoadersMixin
from ..utils import logging
from .attention_processor import (
ADDED_KV_ATTENTION_PROCESSORS,
CROSS_ATTENTION_PROCESSORS,
AttentionProcessor,
AttnAddedKVProcessor,
AttnProcessor,
)
from .embeddings import TimestepEmbedding, Timesteps
from .modeling_utils import ModelMixin
from .transformer_temporal import TransformerTemporalModel
from .unet_2d_blocks import UNetMidBlock2DCrossAttn
from .unet_2d_condition import UNet2DConditionModel
from .unet_3d_blocks import (
CrossAttnDownBlockMotion,
CrossAttnUpBlockMotion,
DownBlockMotion,
UNetMidBlockCrossAttnMotion,
UpBlockMotion,
get_down_block,
get_up_block,
)
from .unet_3d_condition import UNet3DConditionOutput
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class MotionModules(nn.Module):
def __init__(
self,
in_channels: int,
layers_per_block: int = 2,
num_attention_heads: int = 8,
attention_bias: bool = False,
cross_attention_dim: Optional[int] = None,
activation_fn: str = "geglu",
norm_num_groups: int = 32,
max_seq_length: int = 32,
):
super().__init__()
self.motion_modules = nn.ModuleList([])
for i in range(layers_per_block):
self.motion_modules.append(
TransformerTemporalModel(
in_channels=in_channels,
norm_num_groups=norm_num_groups,
cross_attention_dim=cross_attention_dim,
activation_fn=activation_fn,
attention_bias=attention_bias,
num_attention_heads=num_attention_heads,
attention_head_dim=in_channels // num_attention_heads,
positional_embeddings="sinusoidal",
num_positional_embeddings=max_seq_length,
)
)
class MotionAdapter(ModelMixin, ConfigMixin):
@register_to_config
def __init__(
self,
block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
motion_layers_per_block: int = 2,
motion_mid_block_layers_per_block: int = 1,
motion_num_attention_heads: int = 8,
motion_norm_num_groups: int = 32,
motion_max_seq_length: int = 32,
use_motion_mid_block: bool = True,
):
"""Container to store AnimateDiff Motion Modules
Args:
block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
The tuple of output channels for each UNet block.
motion_layers_per_block (`int`, *optional*, defaults to 2):
The number of motion layers per UNet block.
motion_mid_block_layers_per_block (`int`, *optional*, defaults to 1):
The number of motion layers in the middle UNet block.
motion_num_attention_heads (`int`, *optional*, defaults to 8):
The number of heads to use in each attention layer of the motion module.
motion_norm_num_groups (`int`, *optional*, defaults to 32):
The number of groups to use in each group normalization layer of the motion module.
motion_max_seq_length (`int`, *optional*, defaults to 32):
The maximum sequence length to use in the motion module.
use_motion_mid_block (`bool`, *optional*, defaults to True):
Whether to use a motion module in the middle of the UNet.
"""
super().__init__()
down_blocks = []
up_blocks = []
for i, channel in enumerate(block_out_channels):
output_channel = block_out_channels[i]
down_blocks.append(
MotionModules(
in_channels=output_channel,
norm_num_groups=motion_norm_num_groups,
cross_attention_dim=None,
activation_fn="geglu",
attention_bias=False,
num_attention_heads=motion_num_attention_heads,
max_seq_length=motion_max_seq_length,
layers_per_block=motion_layers_per_block,
)
)
if use_motion_mid_block:
self.mid_block = MotionModules(
in_channels=block_out_channels[-1],
norm_num_groups=motion_norm_num_groups,
cross_attention_dim=None,
activation_fn="geglu",
attention_bias=False,
num_attention_heads=motion_num_attention_heads,
layers_per_block=motion_mid_block_layers_per_block,
max_seq_length=motion_max_seq_length,
)
else:
self.mid_block = None
reversed_block_out_channels = list(reversed(block_out_channels))
output_channel = reversed_block_out_channels[0]
for i, channel in enumerate(reversed_block_out_channels):
output_channel = reversed_block_out_channels[i]
up_blocks.append(
MotionModules(
in_channels=output_channel,
norm_num_groups=motion_norm_num_groups,
cross_attention_dim=None,
activation_fn="geglu",
attention_bias=False,
num_attention_heads=motion_num_attention_heads,
max_seq_length=motion_max_seq_length,
layers_per_block=motion_layers_per_block + 1,
)
)
self.down_blocks = nn.ModuleList(down_blocks)
self.up_blocks = nn.ModuleList(up_blocks)
def forward(self, sample):
pass
class UNetMotionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
r"""
A modified conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a
sample shaped output.
This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
for all models (such as downloading or saving).
"""
_supports_gradient_checkpointing = True
@register_to_config
def __init__(
self,
sample_size: Optional[int] = None,
in_channels: int = 4,
out_channels: int = 4,
down_block_types: Tuple[str, ...] = (
"CrossAttnDownBlockMotion",
"CrossAttnDownBlockMotion",
"CrossAttnDownBlockMotion",
"DownBlockMotion",
),
up_block_types: Tuple[str, ...] = (
"UpBlockMotion",
"CrossAttnUpBlockMotion",
"CrossAttnUpBlockMotion",
"CrossAttnUpBlockMotion",
),
block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
layers_per_block: int = 2,
downsample_padding: int = 1,
mid_block_scale_factor: float = 1,
act_fn: str = "silu",
norm_num_groups: int = 32,
norm_eps: float = 1e-5,
cross_attention_dim: int = 1280,
use_linear_projection: bool = False,
num_attention_heads: Union[int, Tuple[int, ...]] = 8,
motion_max_seq_length: int = 32,
motion_num_attention_heads: int = 8,
use_motion_mid_block: int = True,
encoder_hid_dim: Optional[int] = None,
encoder_hid_dim_type: Optional[str] = None,
):
super().__init__()
self.sample_size = sample_size
# Check inputs
if len(down_block_types) != len(up_block_types):
raise ValueError(
f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
)
if len(block_out_channels) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
)
if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
)
# input
conv_in_kernel = 3
conv_out_kernel = 3
conv_in_padding = (conv_in_kernel - 1) // 2
self.conv_in = nn.Conv2d(
in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
)
# time
time_embed_dim = block_out_channels[0] * 4
self.time_proj = Timesteps(block_out_channels[0], True, 0)
timestep_input_dim = block_out_channels[0]
self.time_embedding = TimestepEmbedding(
timestep_input_dim,
time_embed_dim,
act_fn=act_fn,
)
if encoder_hid_dim_type is None:
self.encoder_hid_proj = None
# class embedding
self.down_blocks = nn.ModuleList([])
self.up_blocks = nn.ModuleList([])
if isinstance(num_attention_heads, int):
num_attention_heads = (num_attention_heads,) * len(down_block_types)
# down
output_channel = block_out_channels[0]
for i, down_block_type in enumerate(down_block_types):
input_channel = output_channel
output_channel = block_out_channels[i]
is_final_block = i == len(block_out_channels) - 1
down_block = get_down_block(
down_block_type,
num_layers=layers_per_block,
in_channels=input_channel,
out_channels=output_channel,
temb_channels=time_embed_dim,
add_downsample=not is_final_block,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
cross_attention_dim=cross_attention_dim,
num_attention_heads=num_attention_heads[i],
downsample_padding=downsample_padding,
use_linear_projection=use_linear_projection,
dual_cross_attention=False,
temporal_num_attention_heads=motion_num_attention_heads,
temporal_max_seq_length=motion_max_seq_length,
)
self.down_blocks.append(down_block)
# mid
if use_motion_mid_block:
self.mid_block = UNetMidBlockCrossAttnMotion(
in_channels=block_out_channels[-1],
temb_channels=time_embed_dim,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
output_scale_factor=mid_block_scale_factor,
cross_attention_dim=cross_attention_dim,
num_attention_heads=num_attention_heads[-1],
resnet_groups=norm_num_groups,
dual_cross_attention=False,
temporal_num_attention_heads=motion_num_attention_heads,
temporal_max_seq_length=motion_max_seq_length,
)
else:
self.mid_block = UNetMidBlock2DCrossAttn(
in_channels=block_out_channels[-1],
temb_channels=time_embed_dim,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
output_scale_factor=mid_block_scale_factor,
cross_attention_dim=cross_attention_dim,
num_attention_heads=num_attention_heads[-1],
resnet_groups=norm_num_groups,
dual_cross_attention=False,
)
# count how many layers upsample the images
self.num_upsamplers = 0
# up
reversed_block_out_channels = list(reversed(block_out_channels))
reversed_num_attention_heads = list(reversed(num_attention_heads))
output_channel = reversed_block_out_channels[0]
for i, up_block_type in enumerate(up_block_types):
is_final_block = i == len(block_out_channels) - 1
prev_output_channel = output_channel
output_channel = reversed_block_out_channels[i]
input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
# add upsample block for all BUT final layer
if not is_final_block:
add_upsample = True
self.num_upsamplers += 1
else:
add_upsample = False
up_block = get_up_block(
up_block_type,
num_layers=layers_per_block + 1,
in_channels=input_channel,
out_channels=output_channel,
prev_output_channel=prev_output_channel,
temb_channels=time_embed_dim,
add_upsample=add_upsample,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
cross_attention_dim=cross_attention_dim,
num_attention_heads=reversed_num_attention_heads[i],
dual_cross_attention=False,
resolution_idx=i,
use_linear_projection=use_linear_projection,
temporal_num_attention_heads=motion_num_attention_heads,
temporal_max_seq_length=motion_max_seq_length,
)
self.up_blocks.append(up_block)
prev_output_channel = output_channel
# out
if norm_num_groups is not None:
self.conv_norm_out = nn.GroupNorm(
num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
)
self.conv_act = nn.SiLU()
else:
self.conv_norm_out = None
self.conv_act = None
conv_out_padding = (conv_out_kernel - 1) // 2
self.conv_out = nn.Conv2d(
block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
)
@classmethod
def from_unet2d(
cls,
unet: UNet2DConditionModel,
motion_adapter: Optional[MotionAdapter] = None,
load_weights: bool = True,
):
has_motion_adapter = motion_adapter is not None
# based on https://github.com/guoyww/AnimateDiff/blob/895f3220c06318ea0760131ec70408b466c49333/animatediff/models/unet.py#L459
config = unet.config
config["_class_name"] = cls.__name__
down_blocks = []
for down_blocks_type in config["down_block_types"]:
if "CrossAttn" in down_blocks_type:
down_blocks.append("CrossAttnDownBlockMotion")
else:
down_blocks.append("DownBlockMotion")
config["down_block_types"] = down_blocks
up_blocks = []
for down_blocks_type in config["up_block_types"]:
if "CrossAttn" in down_blocks_type:
up_blocks.append("CrossAttnUpBlockMotion")
else:
up_blocks.append("UpBlockMotion")
config["up_block_types"] = up_blocks
if has_motion_adapter:
config["motion_num_attention_heads"] = motion_adapter.config["motion_num_attention_heads"]
config["motion_max_seq_length"] = motion_adapter.config["motion_max_seq_length"]
config["use_motion_mid_block"] = motion_adapter.config["use_motion_mid_block"]
# Need this for backwards compatibility with UNet2DConditionModel checkpoints
if not config.get("num_attention_heads"):
config["num_attention_heads"] = config["attention_head_dim"]
model = cls.from_config(config)
if not load_weights:
return model
model.conv_in.load_state_dict(unet.conv_in.state_dict())
model.time_proj.load_state_dict(unet.time_proj.state_dict())
model.time_embedding.load_state_dict(unet.time_embedding.state_dict())
for i, down_block in enumerate(unet.down_blocks):
model.down_blocks[i].resnets.load_state_dict(down_block.resnets.state_dict())
if hasattr(model.down_blocks[i], "attentions"):
model.down_blocks[i].attentions.load_state_dict(down_block.attentions.state_dict())
if model.down_blocks[i].downsamplers:
model.down_blocks[i].downsamplers.load_state_dict(down_block.downsamplers.state_dict())
for i, up_block in enumerate(unet.up_blocks):
model.up_blocks[i].resnets.load_state_dict(up_block.resnets.state_dict())
if hasattr(model.up_blocks[i], "attentions"):
model.up_blocks[i].attentions.load_state_dict(up_block.attentions.state_dict())
if model.up_blocks[i].upsamplers:
model.up_blocks[i].upsamplers.load_state_dict(up_block.upsamplers.state_dict())
model.mid_block.resnets.load_state_dict(unet.mid_block.resnets.state_dict())
model.mid_block.attentions.load_state_dict(unet.mid_block.attentions.state_dict())
if unet.conv_norm_out is not None:
model.conv_norm_out.load_state_dict(unet.conv_norm_out.state_dict())
if unet.conv_act is not None:
model.conv_act.load_state_dict(unet.conv_act.state_dict())
model.conv_out.load_state_dict(unet.conv_out.state_dict())
if has_motion_adapter:
model.load_motion_modules(motion_adapter)
# ensure that the Motion UNet is the same dtype as the UNet2DConditionModel
model.to(unet.dtype)
return model
def freeze_unet2d_params(self) -> None:
"""Freeze the weights of just the UNet2DConditionModel, and leave the motion modules
unfrozen for fine tuning.
"""
# Freeze everything
for param in self.parameters():
param.requires_grad = False
# Unfreeze Motion Modules
for down_block in self.down_blocks:
motion_modules = down_block.motion_modules
for param in motion_modules.parameters():
param.requires_grad = True
for up_block in self.up_blocks:
motion_modules = up_block.motion_modules
for param in motion_modules.parameters():
param.requires_grad = True
if hasattr(self.mid_block, "motion_modules"):
motion_modules = self.mid_block.motion_modules
for param in motion_modules.parameters():
param.requires_grad = True
def load_motion_modules(self, motion_adapter: Optional[MotionAdapter]) -> None:
for i, down_block in enumerate(motion_adapter.down_blocks):
self.down_blocks[i].motion_modules.load_state_dict(down_block.motion_modules.state_dict())
for i, up_block in enumerate(motion_adapter.up_blocks):
self.up_blocks[i].motion_modules.load_state_dict(up_block.motion_modules.state_dict())
# to support older motion modules that don't have a mid_block
if hasattr(self.mid_block, "motion_modules"):
self.mid_block.motion_modules.load_state_dict(motion_adapter.mid_block.motion_modules.state_dict())
def save_motion_modules(
self,
save_directory: str,
is_main_process: bool = True,
safe_serialization: bool = True,
variant: Optional[str] = None,
push_to_hub: bool = False,
**kwargs,
) -> None:
state_dict = self.state_dict()
# Extract all motion modules
motion_state_dict = {}
for k, v in state_dict.items():
if "motion_modules" in k:
motion_state_dict[k] = v
adapter = MotionAdapter(
block_out_channels=self.config["block_out_channels"],
motion_layers_per_block=self.config["layers_per_block"],
motion_norm_num_groups=self.config["norm_num_groups"],
motion_num_attention_heads=self.config["motion_num_attention_heads"],
motion_max_seq_length=self.config["motion_max_seq_length"],
use_motion_mid_block=self.config["use_motion_mid_block"],
)
adapter.load_state_dict(motion_state_dict)
adapter.save_pretrained(
save_directory=save_directory,
is_main_process=is_main_process,
safe_serialization=safe_serialization,
variant=variant,
push_to_hub=push_to_hub,
**kwargs,
)
@property
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
def attn_processors(self) -> Dict[str, AttentionProcessor]:
r"""
Returns:
`dict` of attention processors: A dictionary containing all attention processors used in the model with
indexed by its weight name.
"""
# set recursively
processors = {}
def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
if hasattr(module, "get_processor"):
processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
for sub_name, child in module.named_children():
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
return processors
for name, module in self.named_children():
fn_recursive_add_processors(name, module, processors)
return processors
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
r"""
Sets the attention processor to use to compute attention.
Parameters:
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
The instantiated processor class or a dictionary of processor classes that will be set as the processor
for **all** `Attention` layers.
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
processor. This is strongly recommended when setting trainable attention processors.
"""
count = len(self.attn_processors.keys())
if isinstance(processor, dict) and len(processor) != count:
raise ValueError(
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
)
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
if hasattr(module, "set_processor"):
if not isinstance(processor, dict):
module.set_processor(processor)
else:
module.set_processor(processor.pop(f"{name}.processor"))
for sub_name, child in module.named_children():
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
for name, module in self.named_children():
fn_recursive_attn_processor(name, module, processor)
# Copied from diffusers.models.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
"""
Sets the attention processor to use [feed forward
chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
Parameters:
chunk_size (`int`, *optional*):
The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
over each tensor of dim=`dim`.
dim (`int`, *optional*, defaults to `0`):
The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
or dim=1 (sequence length).
"""
if dim not in [0, 1]:
raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
# By default chunk size is 1
chunk_size = chunk_size or 1
def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
if hasattr(module, "set_chunk_feed_forward"):
module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
for child in module.children():
fn_recursive_feed_forward(child, chunk_size, dim)
for module in self.children():
fn_recursive_feed_forward(module, chunk_size, dim)
# Copied from diffusers.models.unet_3d_condition.UNet3DConditionModel.disable_forward_chunking
def disable_forward_chunking(self) -> None:
def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
if hasattr(module, "set_chunk_feed_forward"):
module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
for child in module.children():
fn_recursive_feed_forward(child, chunk_size, dim)
for module in self.children():
fn_recursive_feed_forward(module, None, 0)
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
def set_default_attn_processor(self) -> None:
"""
Disables custom attention processors and sets the default attention implementation.
"""
if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnAddedKVProcessor()
elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnProcessor()
else:
raise ValueError(
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
)
self.set_attn_processor(processor)
def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
if isinstance(module, (CrossAttnDownBlockMotion, DownBlockMotion, CrossAttnUpBlockMotion, UpBlockMotion)):
module.gradient_checkpointing = value
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.enable_freeu
def enable_freeu(self, s1: float, s2: float, b1: float, b2: float) -> None:
r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
The suffixes after the scaling factors represent the stage blocks where they are being applied.
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
Args:
s1 (`float`):
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
mitigate the "oversmoothing effect" in the enhanced denoising process.
s2 (`float`):
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
mitigate the "oversmoothing effect" in the enhanced denoising process.
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
"""
for i, upsample_block in enumerate(self.up_blocks):
setattr(upsample_block, "s1", s1)
setattr(upsample_block, "s2", s2)
setattr(upsample_block, "b1", b1)
setattr(upsample_block, "b2", b2)
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.disable_freeu
def disable_freeu(self) -> None:
"""Disables the FreeU mechanism."""
freeu_keys = {"s1", "s2", "b1", "b2"}
for i, upsample_block in enumerate(self.up_blocks):
for k in freeu_keys:
if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
setattr(upsample_block, k, None)
def forward(
self,
sample: torch.FloatTensor,
timestep: Union[torch.Tensor, float, int],
encoder_hidden_states: torch.Tensor,
timestep_cond: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
mid_block_additional_residual: Optional[torch.Tensor] = None,
return_dict: bool = True,
) -> Union[UNet3DConditionOutput, Tuple[torch.Tensor]]:
r"""
The [`UNetMotionModel`] forward method.
Args:
sample (`torch.FloatTensor`):
The noisy input tensor with the following shape `(batch, num_frames, channel, height, width`.
timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
encoder_hidden_states (`torch.FloatTensor`):
The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
through the `self.time_embedding` layer to obtain the timestep embeddings.
attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
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.
cross_attention_kwargs (`dict`, *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).
down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
A tuple of tensors that if specified are added to the residuals of down unet blocks.
mid_block_additional_residual: (`torch.Tensor`, *optional*):
A tensor that if specified is added to the residual of the middle unet block.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~models.unet_3d_condition.UNet3DConditionOutput`] instead of a plain
tuple.
Returns:
[`~models.unet_3d_condition.UNet3DConditionOutput`] or `tuple`:
If `return_dict` is True, an [`~models.unet_3d_condition.UNet3DConditionOutput`] is returned, otherwise
a `tuple` is returned where the first element is the sample tensor.
"""
# By default samples have to be AT least a multiple of the overall upsampling factor.
# The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
# However, the upsampling interpolation output size can be forced to fit any upsampling size
# on the fly if necessary.
default_overall_up_factor = 2**self.num_upsamplers
# upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
forward_upsample_size = False
upsample_size = None
if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
logger.info("Forward upsample size to force interpolation output size.")
forward_upsample_size = True
# prepare attention_mask
if attention_mask is not None:
attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
attention_mask = attention_mask.unsqueeze(1)
# 1. time
timesteps = timestep
if not torch.is_tensor(timesteps):
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
# This would be a good case for the `match` statement (Python 3.10+)
is_mps = sample.device.type == "mps"
if isinstance(timestep, float):
dtype = torch.float32 if is_mps else torch.float64
else:
dtype = torch.int32 if is_mps else torch.int64
timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
elif len(timesteps.shape) == 0:
timesteps = timesteps[None].to(sample.device)
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
num_frames = sample.shape[2]
timesteps = timesteps.expand(sample.shape[0])
t_emb = self.time_proj(timesteps)
# timesteps does not contain any weights and will always return f32 tensors
# but time_embedding might actually be running in fp16. so we need to cast here.
# there might be better ways to encapsulate this.
t_emb = t_emb.to(dtype=self.dtype)
emb = self.time_embedding(t_emb, timestep_cond)
emb = emb.repeat_interleave(repeats=num_frames, dim=0)
if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
if "image_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
)
image_embeds = added_cond_kwargs.get("image_embeds")
image_embeds = self.encoder_hid_proj(image_embeds).to(encoder_hidden_states.dtype)
encoder_hidden_states = torch.cat([encoder_hidden_states, image_embeds], dim=1)
encoder_hidden_states = encoder_hidden_states.repeat_interleave(repeats=num_frames, dim=0)
# 2. pre-process
sample = sample.permute(0, 2, 1, 3, 4).reshape((sample.shape[0] * num_frames, -1) + sample.shape[3:])
sample = self.conv_in(sample)
# 3. down
down_block_res_samples = (sample,)
for downsample_block in self.down_blocks:
if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
sample, res_samples = downsample_block(
hidden_states=sample,
temb=emb,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
num_frames=num_frames,
cross_attention_kwargs=cross_attention_kwargs,
)
else:
sample, res_samples = downsample_block(hidden_states=sample, temb=emb, num_frames=num_frames)
down_block_res_samples += res_samples
if down_block_additional_residuals is not None:
new_down_block_res_samples = ()
for down_block_res_sample, down_block_additional_residual in zip(
down_block_res_samples, down_block_additional_residuals
):
down_block_res_sample = down_block_res_sample + down_block_additional_residual
new_down_block_res_samples += (down_block_res_sample,)
down_block_res_samples = new_down_block_res_samples
# 4. mid
if self.mid_block is not None:
# To support older versions of motion modules that don't have a mid_block
if hasattr(self.mid_block, "motion_modules"):
sample = self.mid_block(
sample,
emb,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
num_frames=num_frames,
cross_attention_kwargs=cross_attention_kwargs,
)
else:
sample = self.mid_block(
sample,
emb,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
)
if mid_block_additional_residual is not None:
sample = sample + mid_block_additional_residual
# 5. up
for i, upsample_block in enumerate(self.up_blocks):
is_final_block = i == len(self.up_blocks) - 1
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
# if we have not reached the final block and need to forward the
# upsample size, we do it here
if not is_final_block and forward_upsample_size:
upsample_size = down_block_res_samples[-1].shape[2:]
if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
sample = upsample_block(
hidden_states=sample,
temb=emb,
res_hidden_states_tuple=res_samples,
encoder_hidden_states=encoder_hidden_states,
upsample_size=upsample_size,
attention_mask=attention_mask,
num_frames=num_frames,
cross_attention_kwargs=cross_attention_kwargs,
)
else:
sample = upsample_block(
hidden_states=sample,
temb=emb,
res_hidden_states_tuple=res_samples,
upsample_size=upsample_size,
num_frames=num_frames,
)
# 6. post-process
if self.conv_norm_out:
sample = self.conv_norm_out(sample)
sample = self.conv_act(sample)
sample = self.conv_out(sample)
# reshape to (batch, channel, framerate, width, height)
sample = sample[None, :].reshape((-1, num_frames) + sample.shape[1:]).permute(0, 2, 1, 3, 4)
if not return_dict:
return (sample,)
return UNet3DConditionOutput(sample=sample)
|