File size: 43,884 Bytes
bc3092e |
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 |
import torch
import torch as th
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from contextlib import redirect_stdout
from einops import rearrange
from PrimeDepth_inference.ldm.modules.diffusionmodules.util import (
conv_nd,
linear,
zero_module,
timestep_embedding,
checkpoint
)
from PrimeDepth_inference.ldm.modules.attention import SpatialTransformer
from PrimeDepth_inference.ldm.modules.diffusionmodules.openaimodel import (
UNetModel,
TimestepEmbedSequential,
ResBlock as ResBlock_orig,
Downsample,
Upsample,
AttentionBlock,
TimestepBlock
)
from PrimeDepth_inference.ldm.modules.diffusionmodules.model import Decoder
from PrimeDepth_inference.ldm.util import exists
class PrimeDepthLabeller(nn.Module):
def __init__(
self,
image_size,
in_channels,
model_channels,
out_channels,
num_res_blocks,
attention_resolutions,
label_decoder_config,
dropout=0,
channel_mult=(1, 2, 4, 8),
conv_resample=True,
dims=2,
use_checkpoint=False,
use_fp16=False,
num_heads=-1,
num_head_channels=-1,
num_heads_upsample=-1,
use_scale_shift_norm=False,
resblock_updown=False,
use_new_attention_order=False,
use_spatial_transformer=False, # custom transformer support
transformer_depth=1, # custom transformer support
context_dim=None, # custom transformer support
n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
legacy=False,
use_linear_in_transformer=False,
infusion2refiner='cat', # how to infuse intermediate information into the refiner? {'add', 'cat', None}
refiner_model_ratio=1.0, # ratio of the refiner size compared to the base model. [0, 1]
label_mode={
'depth': 3,
'segmentation': 20
}, # label mode and corresponding number of output channels
use_self_attn_maps=True, # use self attention maps from the base model
use_cross_attn_maps=True, # use cross attention maps from the base model
use_feature_maps=True, # use intermediate features from the base model
n_ca_maps=77, # number of cross-attention maps
n_sa_maps=64, # number of self-attention maps
channels2predictor=512,
zero_convs=True, # use zero-convolutions for the infusion of features from base to refiner
scale_input_features_with_model=False, # scale input features from base to refiner with refiner model size
):
assert infusion2refiner in ('cat', 'add', 'att', None), f'infusion2refiner needs to be cat, add, att or None, but not {infusion2refiner}'
super().__init__()
self.infusion2refiner = infusion2refiner
self.in_ch_factor = 1 if infusion2refiner in ('add', 'att') else 2
self.refiner_model_ratio = refiner_model_ratio
self.out_channels = out_channels
self.dims = 2
self.model_channels = model_channels
self.label_mode = label_mode
self.use_self_attn_maps = use_self_attn_maps
self.use_cross_attn_maps = use_cross_attn_maps
self.use_feature_maps = use_feature_maps
self.n_ca_maps = n_ca_maps
self.n_sa_maps = n_sa_maps
self.channels2predictor = channels2predictor
self.zero_convs = zero_convs
with redirect_stdout(None):
################# dummy base model to gather information #################
base_model = UNetModel(
image_size=image_size, in_channels=in_channels, model_channels=model_channels,
out_channels=out_channels, num_res_blocks=num_res_blocks,
attention_resolutions=attention_resolutions, dropout=dropout, channel_mult=channel_mult,
conv_resample=conv_resample, dims=dims, use_checkpoint=use_checkpoint,
use_fp16=use_fp16, num_heads=num_heads, num_head_channels=num_head_channels,
num_heads_upsample=num_heads_upsample, use_scale_shift_norm=use_scale_shift_norm,
resblock_updown=resblock_updown, use_new_attention_order=use_new_attention_order,
use_spatial_transformer=use_spatial_transformer, transformer_depth=transformer_depth,
context_dim=context_dim, n_embed=n_embed, legacy=legacy,
use_linear_in_transformer=use_linear_in_transformer,
)
#####################################################################
### Gathering information about Attention Block Existince ###
#####################################################################
blockwise_attn_dec = [torch.tensor(False)]
blockwise_attn_enc = []
heads_enc = []
heads_dec = [0]
for module in base_model.input_blocks:
has_att = False
heads = 0
for mod in list(module.children()):
if isinstance(mod, SpatialTransformer):
has_att = True
heads = mod.transformer_blocks[0].attn1.heads
heads_enc.append(heads)
blockwise_attn_enc.append(has_att)
blockwise_middle = False
for mod in list(base_model.middle_block.children()):
heads = 0
if isinstance(mod, SpatialTransformer):
blockwise_middle = True
heads_middle = mod.transformer_blocks[0].attn1.heads
for module in base_model.output_blocks:
has_att = False
heads = 0
for mod in list(module.children()):
if isinstance(mod, SpatialTransformer):
has_att = True
heads = mod.transformer_blocks[0].attn1.heads
heads_dec.append(heads)
blockwise_attn_dec.append(has_att)
additional_attention_blocks = torch.tensor(blockwise_attn_enc[::-1]) + torch.tensor(blockwise_attn_dec[:-1])
additional_attention_blocks[0] = torch.tensor([blockwise_middle]) or additional_attention_blocks[0]
additional_attention_channels = []
ch_per_head = n_ca_maps * use_cross_attn_maps + n_sa_maps * use_self_attn_maps
for h_enc, h_dec in zip(heads_enc[::-1], heads_dec[:-1]):
additional_attention_channels.append(max(h_enc, h_dec) * ch_per_head)
additional_attention_channels[0] = max(heads_middle * ch_per_head, additional_attention_channels[0])
self.refiner = Refiner(
image_size=image_size, in_channels=in_channels, model_channels=model_channels,
out_channels=channels2predictor, num_res_blocks=num_res_blocks,
attention_resolutions=attention_resolutions, dropout=dropout, channel_mult=channel_mult,
conv_resample=conv_resample, dims=dims, use_checkpoint=use_checkpoint,
use_fp16=use_fp16, num_heads=num_heads, num_head_channels=num_head_channels,
num_heads_upsample=num_heads_upsample, use_scale_shift_norm=use_scale_shift_norm,
resblock_updown=resblock_updown, use_new_attention_order=use_new_attention_order,
use_spatial_transformer=use_spatial_transformer, transformer_depth=transformer_depth,
context_dim=context_dim, n_embed=n_embed, legacy=legacy,
use_linear_in_transformer=use_linear_in_transformer,
infusion2refiner=infusion2refiner, refiner_model_ratio=refiner_model_ratio,
additional_attention_channels=additional_attention_channels,
additional_attention_blocks=additional_attention_blocks,
scale_input_features_with_model=scale_input_features_with_model,
use_feature_maps=use_feature_maps,
) # initialise pretrained model
# self.enc_zero_convs_out = nn.ModuleList([])
self.enc_zero_convs_in = nn.ModuleList([])
# self.middle_block_out = None
self.middle_block_in = None
# self.dec_zero_convs_out = nn.ModuleList([])
self.dec_zero_convs_in = nn.ModuleList([])
ch_inout_refiner = {'enc': [], 'mid': [], 'dec': []}
ch_inout_base = {'enc': [], 'mid': [], 'dec': []}
################# Gather Channel Sizes #################
for module in base_model.input_blocks:
if isinstance(module[0], nn.Conv2d):
ch_inout_base['enc'].append((module[0].in_channels, module[0].out_channels))
elif isinstance(module[0], (ResBlock, ResBlock_orig)):
ch_inout_base['enc'].append((module[0].channels, module[0].out_channels))
elif isinstance(module[0], Downsample):
ch_inout_base['enc'].append((module[0].channels, module[-1].out_channels))
ch_inout_base['mid'].append((base_model.middle_block[0].channels, base_model.middle_block[-1].out_channels))
for module in self.refiner.output_blocks:
if isinstance(module[0], nn.Conv2d):
ch_inout_refiner['dec'].append((module[0].in_channels, module[0].out_channels))
elif isinstance(module[0], (ResBlock, ResBlock_orig)):
ch_inout_refiner['dec'].append((module[0].channels, module[0].out_channels))
elif isinstance(module[-1], Upsample):
ch_inout_refiner['dec'].append((module[0].channels, module[-1].out_channels))
for module in base_model.output_blocks:
if isinstance(module[0], nn.Conv2d):
ch_inout_base['dec'].append((module[0].in_channels, module[0].out_channels))
elif isinstance(module[0], (ResBlock, ResBlock_orig)):
ch_inout_base['dec'].append((module[0].channels, module[0].out_channels))
elif isinstance(module[-1], Upsample):
ch_inout_base['dec'].append((module[0].channels, module[-1].out_channels))
self.ch_inout_refiner = ch_inout_refiner
self.ch_inout_base = ch_inout_base
########################################
### Build Zero-Convolutions ###
########################################
# infusion2refiner
self.convs_features = nn.ModuleList()
self.convs_attentions = nn.ModuleList()
self.convs_merge = nn.ModuleList()
############## kill ################
self.kill_blockwise_attn_enc = blockwise_attn_enc[::-1]
self.kill_blockwise_attn_dec = blockwise_attn_dec
####################################
for io_ch_base_enc, io_ch_base_dec, ch_features, ch_attn, attn_heads_enc, attn_heads_dec in zip(
ch_inout_base['enc'][::-1],
ch_inout_base['dec'],
self.refiner.ch_feature_stream,
self.refiner.ch_attn_stream,
heads_enc[::-1], # blockwise_attn_enc[::-1],
heads_dec, # blockwise_attn_dec
):
if self.use_feature_maps:
self.convs_features.append(self.make_zero_conv(
# in_channels=ch_base_enc[-1] + ch_base_dec[-1], out_channels=ch_features))
in_channels=io_ch_base_dec[0], out_channels=ch_features, zero=zero_convs))
else:
self.convs_features.append(None)
if self.use_self_attn_maps or self.use_cross_attn_maps:
ch_attn_proj = ch_attn if ch_attn > 0 else ch_features
ch_attn_in = (n_ca_maps * self.use_cross_attn_maps + n_sa_maps * self.use_self_attn_maps)
# if attn_enc or attn_dec or blockwise_middle:
if attn_heads_enc + attn_heads_dec + heads_middle:
self.convs_attentions.append(self.make_zero_conv(
in_channels=ch_attn_in * (heads_middle + attn_heads_enc + attn_heads_dec),
out_channels=ch_attn_proj, zero=zero_convs))
blockwise_middle = False
heads_middle = 0
else:
self.convs_attentions.append(None)
ch_attn_proj = 0
if ch_features + ch_attn_proj != 0:
self.convs_merge.append(self.make_zero_conv(
in_channels=ch_features + ch_attn_proj, out_channels=ch_features + ch_attn, zero=zero_convs))
else:
self.convs_merge.append(None)
else:
if ch_features != 0:
self.convs_merge.append(self.make_zero_conv(
in_channels=ch_features, out_channels=ch_features + ch_attn, zero=zero_convs))
else:
self.convs_merge.append(None)
label_channels = 0
for key in label_mode:
label_channels += label_mode[key]
self.label_channels = label_channels
self.label_predictor = LabelDecoder(
input_channels=channels2predictor,
channels_out=label_mode['segmentation'],
label_decoder_config=label_decoder_config,
)
self.depth_predictor = LabelDecoder(
input_channels=channels2predictor,
channels_out=label_mode['depth'],
label_decoder_config=label_decoder_config,
)
def make_zero_conv(self, in_channels, out_channels=None, zero=True):
in_channels = in_channels
out_channels = out_channels or in_channels
if zero:
return zero_module(conv_nd(self.dims, in_channels, out_channels, 1, padding=0))
return conv_nd(self.dims, in_channels, out_channels, 1, padding=0)
def infuse(self, stream, infusion, mlp, variant, emb, scale=1.0):
if variant == 'add':
stream = stream + mlp(infusion, emb) * scale
elif variant == 'cat':
stream = torch.cat([stream, mlp(infusion, emb) * scale], dim=1)
elif variant == 'att':
stream = mlp(stream, infusion)
return stream
def prepare_attn_map_features(self, attn_map, attn_version, x_shape, n_slices=None, patches=None):
'''
Preprocesses the attention map features based on the specified attention version.
Parameters:
attn_map (ndarray): The attention map tensor of shape (bs, heads, pixels, tokens).
attn_version (str): The version of attention to be applied. Possible values are 'self' or 'cross'.
n_slices (int, optional): The number of slices. Default is None.
patches (int, optional): The number of patches per row and column. Default is None.
Returns:
ndarray: The preprocessed attention map tensor.
'''
assert n_slices is not None or patches is not None, '!Either <n_slices> or <patches> has to be specified!'
bs, heads, pixels, tokens = attn_map.shape
b, c, h, w = x_shape
# h = w = int(np.sqrt(pixels))
if attn_version == 'self':
# normalise first, patch after
if patches is not None:
h_p = max(1, h // patches)
w_p = max(1, w // patches)
h_mod = 8 - (h % 8)
w_mod = 8 - (w % 8)
attn_map = rearrange(attn_map, 'b head p (h w) -> (b head) p h w', h=h, w=w)
if h_mod != 0 or w_mod != 0:
attn_map = F.interpolate(attn_map, size=(h + h_mod, w + w_mod), mode="bilinear")
attn_map = nn.AvgPool2d(kernel_size=(h_p, w_p), stride=(h_p, w_p))(attn_map)
if h_mod != 0 or w_mod != 0:
attn_map = F.interpolate(attn_map, size=(patches, patches), mode="bilinear")
# normalisation of self-attention maps across head dimensions
attn_map = attn_map / attn_map.max(dim=-3, keepdim=True)[0].max(dim=-2, keepdim=True)[0].max(dim=-1, keepdim=True)[0]
attn_map = rearrange(attn_map, '(b head) (h w) h_p w_p -> b (head h_p w_p) h w', h=h, w=w, b=bs, h_p=patches)
else:
attn_map = rearrange(attn_map, 'b head (n slice) p -> b head n slice p', n=n_slices).mean(dim=3)
attn_map = rearrange(attn_map, 'b head n (h w) -> b (head n) h w', h=h, w=w)
elif attn_version == 'cross':
# normalisation of cross-attention maps
# normalisation not necessary because maximum values for each head are 1 by construction
attn_map = rearrange(attn_map, 'b head p (n slice) -> b head n slice p', n=n_slices).mean(dim=3)
attn_map = rearrange(attn_map, 'b head n (h w) -> b (head n) h w', h=h, w=w)
return attn_map
def get_used_attn_maps(self, attn_maps):
if self.use_self_attn_maps and self.use_cross_attn_maps:
return attn_maps
elif self.use_self_attn_maps:
return [attn_maps[0]]
elif self.use_cross_attn_maps:
return [attn_maps[1]]
return None
def forward(self, x, timesteps, context, base_model, return_labels=False, return_attn=False, exclude_depth=False, exclude_segmentation=False, **kwargs):
if not return_labels:
return base_model(x=x, timesteps=timesteps, context=context, **kwargs)
#########################
### PREPARATION ###
#########################
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
emb = base_model.time_embed(t_emb)
h_base = x.type(base_model.dtype)
hs_base = []
attn_maps_base_enc = []
it_feature_convs = iter(self.convs_features)
if self.use_cross_attn_maps or self.use_self_attn_maps:
it_attn_convs = iter(self.convs_attentions)
it_merge_convs = iter(self.convs_merge)
used_attn_maps_base = {'self': {}, 'cross': {}}
used_attn_maps_refiner = {'self': {}, 'cross': {}}
#########################
### INPUT BLOCK ###
#########################
for i, module_base in enumerate(base_model.input_blocks):
shape = h_base.shape
h_base, attn_maps_base = module_base(h_base, emb, context, return_attn_maps=True)
hs_base.append(h_base)
if attn_maps_base is not None:
attn_maps_base = [
self.prepare_attn_map_features(attn_maps_base[0], 'self', patches=int(np.sqrt(self.n_sa_maps)), x_shape=shape),
self.prepare_attn_map_features(attn_maps_base[1], 'cross', n_slices=self.n_ca_maps, x_shape=h_base.shape)
]
if return_attn:
used_attn_maps_base['self'][f'input_block-{i}'] = attn_maps_base[0]
used_attn_maps_base['cross'][f'input_block-{i}'] = attn_maps_base[1]
attn_maps_base_enc.append(attn_maps_base)
h_base, attn_maps_base = base_model.middle_block(h_base, emb, context, return_attn_maps=True)
attn_maps_base = [
self.prepare_attn_map_features(attn_maps_base[0], 'self', patches=int(np.sqrt(self.n_sa_maps)), x_shape=shape),
self.prepare_attn_map_features(attn_maps_base[1], 'cross', n_slices=self.n_ca_maps, x_shape=h_base.shape)
]
h_refiner = None
if return_attn:
used_attn_maps_base['self']['mid_block'] = attn_maps_base[0]
used_attn_maps_base['cross']['mid_block'] = attn_maps_base[1]
#######################################
### OUTPUT BLOCK WITH REFINER ###
#######################################
for i, (module_base, module_refiner) in enumerate(zip(
base_model.output_blocks,
self.refiner.output_blocks
)):
attn_maps_enc = attn_maps_base_enc.pop()
# prepare attention maps
if self.use_self_attn_maps or self.use_cross_attn_maps:
if attn_maps_base is not None:
attn_maps_all = torch.cat(self.get_used_attn_maps(attn_maps_base), dim=1)
if attn_maps_enc is not None:
attn_maps_enc = torch.cat(self.get_used_attn_maps(attn_maps_enc), dim=1)
if attn_maps_all.shape[-2:] != attn_maps_enc.shape[-2:]:
attn_maps_all = F.interpolate(attn_maps_all, size=attn_maps_enc.shape[-2:], mode="bilinear")
attn_maps_all = torch.cat([
attn_maps_all,
attn_maps_enc
], dim=1)
elif attn_maps_enc is not None:
attn_maps_all = torch.cat(self.get_used_attn_maps(attn_maps_enc), dim=1)
h_base = th.cat([h_base, hs_base.pop()], dim=1)
if len(hs_base) != 0:
output_size = hs_base[-1].shape
else:
output_size = None
# process features and attention maps for Refiner
# get zero-convolutions
feature_conv = next(it_feature_convs)
if self.use_cross_attn_maps or self.use_self_attn_maps:
attn_conv = next(it_attn_convs)
else:
attn_conv = None
merge_conv = next(it_merge_convs)
# process inputs from base model
merge_input = []
if feature_conv is not None:
features_base = feature_conv(h_base)
merge_input.append(features_base)
if attn_conv is not None:
features_attn = attn_conv(attn_maps_all)
merge_input.append(features_attn)
if merge_conv is not None:
features_merged = merge_conv(
torch.cat(merge_input, dim=1))
else:
features_merged = None
# process with Refiner
if h_refiner is None:
# there should always be either attention features of feature maps
if return_attn:
h_refiner, attn_maps_refiner = module_refiner(features_merged, emb, context, return_attn_maps=True, output_size=output_size)
else:
h_refiner = module_refiner(features_merged, emb, context, return_attn_maps=False, output_size=output_size)
else:
if features_merged is not None:
merged_input = torch.cat([h_refiner, features_merged], dim=1)
else:
merged_input = h_refiner
if return_attn:
h_refiner, attn_maps_refiner = module_refiner(merged_input, emb, context, return_attn_maps=True, output_size=output_size)
else:
h_refiner = module_refiner(merged_input, emb, context, return_attn_maps=False, output_size=output_size)
##### compute additional features for refiner #####
shape = h_base.shape
h_base, attn_maps_base = module_base(h_base, emb, context, return_attn_maps=True, output_size=output_size)
if attn_maps_base is not None:
attn_maps_base = [
self.prepare_attn_map_features(attn_maps_base[0], 'self', patches=int(np.sqrt(self.n_sa_maps)), x_shape=shape),
self.prepare_attn_map_features(attn_maps_base[1], 'cross', n_slices=self.n_ca_maps, x_shape=shape)
]
if return_attn:
attn_maps_refiner = [
self.prepare_attn_map_features(attn_maps_refiner[0], 'self', patches=int(np.sqrt(self.n_sa_maps)), x_shape=shape),
self.prepare_attn_map_features(attn_maps_refiner[1], 'cross', n_slices=self.n_ca_maps, x_shape=shape)
]
used_attn_maps_base['self'][f'out_block-{i}'] = attn_maps_base[0]
used_attn_maps_base['cross'][f'out_block-{i}'] = attn_maps_base[1]
used_attn_maps_refiner['self'][f'out_block-{i}'] = attn_maps_refiner[0]
used_attn_maps_refiner['cross'][f'out_block-{i}'] = attn_maps_refiner[1]
#########################################################################
### LABEL CALCULATION WITH PREDICTION HEADS FROM REFINER OUTPUT ###
#########################################################################
if return_attn:
return used_attn_maps_base, used_attn_maps_refiner
labels = {}
h_refiner = self.refiner.out(h_refiner)
if not exclude_depth:
labels['depth'] = self.depth_predictor(h_refiner)
if not exclude_segmentation:
labels['segmentation'] = self.label_predictor(h_refiner)
return base_model.out(h_base), labels
class Refiner(nn.Module):
"""
The full UNet model with attention and timestep embedding.
:param in_channels: channels in the input Tensor.
:param model_channels: base channel count for the model.
:param out_channels: channels in the output Tensor.
:param num_res_blocks: number of residual blocks per downsample.
:param attention_resolutions: a collection of downsample rates at which
attention will take place. May be a set, list, or tuple.
For example, if this contains 4, then at 4x downsampling, attention
will be used.
:param dropout: the dropout probability.
:param channel_mult: channel multiplier for each level of the UNet.
:param conv_resample: if True, use learned convolutions for upsampling and
downsampling.
:param dims: determines if the signal is 1D, 2D, or 3D.
:param use_checkpoint: use gradient checkpointing to reduce memory usage.
:param num_heads: the number of attention heads in each attention layer.
:param num_heads_channels: if specified, ignore num_heads and instead use
a fixed channel width per attention head.
:param num_heads_upsample: works with num_heads to set a different number
of heads for upsampling. Deprecated.
:param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
:param resblock_updown: use residual blocks for up/downsampling.
:param use_new_attention_order: use a different attention pattern for potentially
increased efficiency.
"""
def __init__(
self,
image_size,
in_channels,
model_channels,
out_channels,
num_res_blocks,
attention_resolutions,
dropout=0,
channel_mult=(1, 2, 4, 8),
conv_resample=True,
dims=2,
use_checkpoint=False,
use_fp16=False,
num_heads=-1,
num_head_channels=-1,
num_heads_upsample=-1,
use_scale_shift_norm=False,
resblock_updown=False,
use_new_attention_order=False,
use_spatial_transformer=False, # custom transformer support
transformer_depth=1, # custom transformer support
context_dim=None, # custom transformer support
n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
legacy=True,
disable_self_attentions=None,
num_attention_blocks=None,
use_linear_in_transformer=False,
infusion2refiner='cat', # how to infuse intermediate information into the refiner? {'add', 'cat', None}
refiner_model_ratio=1.0,
additional_attention_channels=1128, # additional features from attention maps
additional_attention_blocks=None, # boolean list stating existance of attention layers in blocks
scale_input_features_with_model=False, # scale input features with ratio for model channels
use_feature_maps=True, # use intermediate features from the base model
):
super().__init__()
if use_spatial_transformer:
assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
if context_dim is not None:
assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
from omegaconf.listconfig import ListConfig
if type(context_dim) == ListConfig:
context_dim = list(context_dim)
self.infusion2refiner = infusion2refiner
if scale_input_features_with_model:
infusion_factor = 1
else:
infusion_factor = 1 / refiner_model_ratio
cat_infusion = 1 if infusion2refiner == 'cat' else 0
if num_heads_upsample == -1:
num_heads_upsample = num_heads
if num_heads == -1:
assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
if num_head_channels == -1:
assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
self.image_size = image_size
self.in_channels = in_channels
self.out_channels = out_channels
if isinstance(num_res_blocks, int):
self.num_res_blocks = len(channel_mult) * [num_res_blocks]
else:
if len(num_res_blocks) != len(channel_mult):
raise ValueError("provide num_res_blocks either as an int (globally constant) or "
"as a list/tuple (per-level) with the same length as channel_mult")
self.num_res_blocks = num_res_blocks
if disable_self_attentions is not None:
# should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
assert len(disable_self_attentions) == len(channel_mult)
if num_attention_blocks is not None:
assert len(num_attention_blocks) == len(self.num_res_blocks)
assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
f"attention will still not be set.")
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mult
self.conv_resample = conv_resample
self.use_checkpoint = use_checkpoint
self.dtype = th.float16 if use_fp16 else th.float32
self.num_heads = num_heads
self.num_head_channels = num_head_channels
self.num_heads_upsample = num_heads_upsample
self.predict_codebook_ids = n_embed is not None
time_embed_dim = model_channels * 4
additional_attention_channels = [int(attn_ch * refiner_model_ratio) for attn_ch in additional_attention_channels]
self.additional_attention_channels = additional_attention_channels
self.scale_input_features_with_model = scale_input_features_with_model
self.use_feature_maps = use_feature_maps
model_channels = max(1, int(model_channels * refiner_model_ratio))
self.model_channels = model_channels
self.refiner_model_ratio = refiner_model_ratio
self._feature_size = model_channels
input_block_chans = [model_channels]
ch = model_channels
ds = 1
############# START - getting rid of the encoder initialisation #############
for level, mult in enumerate(channel_mult):
for nr in range(self.num_res_blocks[level]):
ch = mult * model_channels
input_block_chans.append(ch)
if level != len(channel_mult) - 1:
input_block_chans.append(ch)
ds *= 2
if num_head_channels == -1:
dim_head = max(num_heads, ch // num_heads)
else:
# custom code for smaller models - start
num_head_channels = find_denominator(ch, min(ch, self.num_head_channels))
# custom code for smaller models - end
num_heads = ch // num_head_channels
dim_head = num_head_channels
if legacy:
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
############# END - getting rid of the encoder initialisation #############
self.ch_label_stream = []
self.ch_feature_stream = []
self.ch_attn_stream = []
self.output_blocks = nn.ModuleList([])
attentions_exist = iter(additional_attention_blocks)
additional_attn_ch = iter(additional_attention_channels)
out_ch = 0
for level, mult in list(enumerate(channel_mult))[::-1]:
for i in range(self.num_res_blocks[level] + 1):
attn_flag = next(attentions_exist)
additional_attention_features = next(additional_attn_ch)
ich = input_block_chans.pop()
if level == len(channel_mult) - 1 and i == 0:
self.ch_feature_stream.append(int(ch * cat_infusion * infusion_factor) * 2 if self.use_feature_maps else 0)
self.ch_label_stream.append(0)
else:
self.ch_feature_stream.append(int(ch + (ich + ch) * cat_infusion * infusion_factor) - ch if self.use_feature_maps else 0)
self.ch_label_stream.append(ich)
self.ch_attn_stream.append(additional_attention_features * attn_flag)
layers = [
ResBlock(
# int(ch * cat_infusion * infusion_factor) * 2 + additional_attention_features * attn_flag if level == len(channel_mult) - 1 and i == 0 else int(
# ch + (ich + ch) * cat_infusion * infusion_factor) + additional_attention_features * attn_flag,
int(ch * cat_infusion * infusion_factor) * 2 * self.use_feature_maps + additional_attention_features * attn_flag if level == len(channel_mult) - 1 and i == 0 else int(
(ich + ch) * cat_infusion * infusion_factor) * self.use_feature_maps + ch + additional_attention_features * attn_flag,
# ch + (ich + ch) * cat_infusion * infusion_factor) + additional_attention_features * attn_flag,
time_embed_dim,
dropout,
out_channels=model_channels * mult,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
)
]
ch = model_channels * mult
if ds in attention_resolutions:
if num_head_channels == -1:
dim_head = max(num_heads, ch // num_heads)
else:
# custom code for smaller models - start
num_head_channels = find_denominator(ch, min(ch, self.num_head_channels))
# custom code for smaller models - end
num_heads = ch // num_head_channels
dim_head = num_head_channels
if legacy:
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
if exists(disable_self_attentions):
disabled_sa = disable_self_attentions[level]
else:
disabled_sa = False
if not exists(num_attention_blocks) or i < num_attention_blocks[level]:
layers.append(
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads_upsample,
num_head_channels=dim_head,
use_new_attention_order=use_new_attention_order,
) if not use_spatial_transformer else SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
use_checkpoint=use_checkpoint
)
)
if level and i == self.num_res_blocks[level]:
out_ch = ch
layers.append(
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=out_ch,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
up=True,
)
if resblock_updown
else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
)
ds //= 2
self.output_blocks.append(TimestepEmbedSequential(*layers))
self._feature_size += ch
self.out = nn.Sequential(
normalization(ch),
nn.SiLU(),
zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
)
class LabelDecoder(Decoder):
def __init__(self, input_channels, channels_out, label_decoder_config):
super().__init__(**label_decoder_config)
self.channels_out = channels_out
block_in = self.conv_out.in_channels
self.conv_in = nn.Conv2d(input_channels, 512, 3, 1, 1)
self.conv_out = nn.Conv2d(block_in, channels_out, 3, 1, 1)
def find_denominator(number, start):
if start >= number:
return number
while (start != 0):
residual = number % start
if residual == 0:
return start
start -= 1
def normalization(channels):
"""
Make a standard normalization layer.
:param channels: number of input channels.
:return: an nn.Module for normalization.
"""
if find_denominator(channels, 32) < 32:
print(f'[USING GROUPNORM OVER LESS CHANNELS ({find_denominator(channels, 32)}) FOR {channels} CHANNELS]')
return GroupNorm_leq32(find_denominator(channels, 32), channels)
class GroupNorm_leq32(nn.GroupNorm):
def forward(self, x):
return super().forward(x.float()).type(x.dtype)
class ResBlock(TimestepBlock):
"""
A residual block that can optionally change the number of channels.
:param channels: the number of input channels.
:param emb_channels: the number of timestep embedding channels.
:param dropout: the rate of dropout.
:param out_channels: if specified, the number of out channels.
:param use_conv: if True and out_channels is specified, use a spatial
convolution instead of a smaller 1x1 convolution to change the
channels in the skip connection.
:param dims: determines if the signal is 1D, 2D, or 3D.
:param use_checkpoint: if True, use gradient checkpointing on this module.
:param up: if True, use this block for upsampling.
:param down: if True, use this block for downsampling.
"""
def __init__(
self,
channels,
emb_channels,
dropout,
out_channels=None,
use_conv=False,
use_scale_shift_norm=False,
dims=2,
use_checkpoint=False,
up=False,
down=False,
):
super().__init__()
self.channels = channels
self.emb_channels = emb_channels
self.dropout = dropout
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.use_checkpoint = use_checkpoint
self.use_scale_shift_norm = use_scale_shift_norm
self.in_layers = nn.Sequential(
normalization(channels),
nn.SiLU(),
conv_nd(dims, channels, self.out_channels, 3, padding=1),
)
self.updown = up or down
if up:
self.h_upd = Upsample(channels, False, dims)
self.x_upd = Upsample(channels, False, dims)
elif down:
self.h_upd = Downsample(channels, False, dims)
self.x_upd = Downsample(channels, False, dims)
else:
self.h_upd = self.x_upd = nn.Identity()
self.emb_layers = nn.Sequential(
nn.SiLU(),
linear(
emb_channels,
2 * self.out_channels if use_scale_shift_norm else self.out_channels,
),
)
self.out_layers = nn.Sequential(
normalization(self.out_channels),
nn.SiLU(),
nn.Dropout(p=dropout),
zero_module(
conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
),
)
if self.out_channels == channels:
self.skip_connection = nn.Identity()
elif use_conv:
self.skip_connection = conv_nd(
dims, channels, self.out_channels, 3, padding=1
)
else:
self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
def forward(self, x, emb):
"""
Apply the block to a Tensor, conditioned on a timestep embedding.
:param x: an [N x C x ...] Tensor of features.
:param emb: an [N x emb_channels] Tensor of timestep embeddings.
:return: an [N x C x ...] Tensor of outputs.
"""
return checkpoint(
self._forward, (x, emb), self.parameters(), self.use_checkpoint
)
def _forward(self, x, emb):
if self.updown:
in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
h = in_rest(x)
h = self.h_upd(h)
x = self.x_upd(x)
h = in_conv(h)
else:
h = self.in_layers(x)
emb_out = self.emb_layers(emb).type(h.dtype)
while len(emb_out.shape) < len(h.shape):
emb_out = emb_out[..., None]
if self.use_scale_shift_norm:
out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
scale, shift = th.chunk(emb_out, 2, dim=1)
h = out_norm(h) * (1 + scale) + shift
h = out_rest(h)
else:
h = h + emb_out
h = self.out_layers(h)
return self.skip_connection(x) + h
|