forked from kohya-ss/sd-scripts
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathoriginal_unet.py
1919 lines (1518 loc) · 68.7 KB
/
original_unet.py
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
863
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
# Diffusers 0.10.2からStable Diffusionに必要な部分だけを持ってくる
# 条件分岐等で不要な部分は削除している
# コードの多くはDiffusersからコピーしている
# 制約として、モデルのstate_dictがDiffusers 0.10.2のものと同じ形式である必要がある
# Copy from Diffusers 0.10.2 for Stable Diffusion. Most of the code is copied from Diffusers.
# Unnecessary parts are deleted by condition branching.
# As a constraint, the state_dict of the model must be in the same format as that of Diffusers 0.10.2
"""
v1.5とv2.1の相違点は
- attention_head_dimがintかlist[int]か
- cross_attention_dimが768か1024か
- use_linear_projection: trueがない（=False, 1.5）かあるか
- upcast_attentionがFalse(1.5)かTrue(2.1)か
- （以下は多分無視していい）
- sample_sizeが64か96か
- dual_cross_attentionがあるかないか
- num_class_embedsがあるかないか
- only_cross_attentionがあるかないか
v1.5
{
"_class_name": "UNet2DConditionModel",
"_diffusers_version": "0.6.0",
"act_fn": "silu",
"attention_head_dim": 8,
"block_out_channels": [
320,
640,
1280,
1280
],
"center_input_sample": false,
"cross_attention_dim": 768,
"down_block_types": [
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"DownBlock2D"
],
"downsample_padding": 1,
"flip_sin_to_cos": true,
"freq_shift": 0,
"in_channels": 4,
"layers_per_block": 2,
"mid_block_scale_factor": 1,
"norm_eps": 1e-05,
"norm_num_groups": 32,
"out_channels": 4,
"sample_size": 64,
"up_block_types": [
"UpBlock2D",
"CrossAttnUpBlock2D",
"CrossAttnUpBlock2D",
"CrossAttnUpBlock2D"
]
}
v2.1
{
"_class_name": "UNet2DConditionModel",
"_diffusers_version": "0.10.0.dev0",
"act_fn": "silu",
"attention_head_dim": [
5,
10,
20,
20
],
"block_out_channels": [
320,
640,
1280,
1280
],
"center_input_sample": false,
"cross_attention_dim": 1024,
"down_block_types": [
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"DownBlock2D"
],
"downsample_padding": 1,
"dual_cross_attention": false,
"flip_sin_to_cos": true,
"freq_shift": 0,
"in_channels": 4,
"layers_per_block": 2,
"mid_block_scale_factor": 1,
"norm_eps": 1e-05,
"norm_num_groups": 32,
"num_class_embeds": null,
"only_cross_attention": false,
"out_channels": 4,
"sample_size": 96,
"up_block_types": [
"UpBlock2D",
"CrossAttnUpBlock2D",
"CrossAttnUpBlock2D",
"CrossAttnUpBlock2D"
],
"use_linear_projection": true,
"upcast_attention": true
}
"""
import math
from types import SimpleNamespace
from typing import Dict, Optional, Tuple, Union
import torch
from torch import nn
from torch.nn import functional as F
from einops import rearrange
from library.utils import setup_logging
setup_logging()
import logging
logger = logging.getLogger(__name__)
BLOCK_OUT_CHANNELS: Tuple[int] = (320, 640, 1280, 1280)
TIMESTEP_INPUT_DIM = BLOCK_OUT_CHANNELS[0]
TIME_EMBED_DIM = BLOCK_OUT_CHANNELS[0] * 4
IN_CHANNELS: int = 4
OUT_CHANNELS: int = 4
LAYERS_PER_BLOCK: int = 2
LAYERS_PER_BLOCK_UP: int = LAYERS_PER_BLOCK + 1
TIME_EMBED_FLIP_SIN_TO_COS: bool = True
TIME_EMBED_FREQ_SHIFT: int = 0
NORM_GROUPS: int = 32
NORM_EPS: float = 1e-5
TRANSFORMER_NORM_NUM_GROUPS = 32
DOWN_BLOCK_TYPES = ["CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D"]
UP_BLOCK_TYPES = ["UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"]
# region memory efficient attention
# FlashAttentionを使うCrossAttention
# based on https://github.com/lucidrains/memory-efficient-attention-pytorch/blob/main/memory_efficient_attention_pytorch/flash_attention.py
# LICENSE MIT https://github.com/lucidrains/memory-efficient-attention-pytorch/blob/main/LICENSE
# constants
EPSILON = 1e-6
# helper functions
def exists(val):
return val is not None
def default(val, d):
return val if exists(val) else d
# flash attention forwards and backwards
# https://arxiv.org/abs/2205.14135
class FlashAttentionFunction(torch.autograd.Function):
@staticmethod
@torch.no_grad()
def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size):
"""Algorithm 2 in the paper"""
device = q.device
dtype = q.dtype
max_neg_value = -torch.finfo(q.dtype).max
qk_len_diff = max(k.shape[-2] - q.shape[-2], 0)
o = torch.zeros_like(q)
all_row_sums = torch.zeros((*q.shape[:-1], 1), dtype=dtype, device=device)
all_row_maxes = torch.full((*q.shape[:-1], 1), max_neg_value, dtype=dtype, device=device)
scale = q.shape[-1] ** -0.5
if not exists(mask):
mask = (None,) * math.ceil(q.shape[-2] / q_bucket_size)
else:
mask = rearrange(mask, "b n -> b 1 1 n")
mask = mask.split(q_bucket_size, dim=-1)
row_splits = zip(
q.split(q_bucket_size, dim=-2),
o.split(q_bucket_size, dim=-2),
mask,
all_row_sums.split(q_bucket_size, dim=-2),
all_row_maxes.split(q_bucket_size, dim=-2),
)
for ind, (qc, oc, row_mask, row_sums, row_maxes) in enumerate(row_splits):
q_start_index = ind * q_bucket_size - qk_len_diff
col_splits = zip(
k.split(k_bucket_size, dim=-2),
v.split(k_bucket_size, dim=-2),
)
for k_ind, (kc, vc) in enumerate(col_splits):
k_start_index = k_ind * k_bucket_size
attn_weights = torch.einsum("... i d, ... j d -> ... i j", qc, kc) * scale
if exists(row_mask):
attn_weights.masked_fill_(~row_mask, max_neg_value)
if causal and q_start_index < (k_start_index + k_bucket_size - 1):
causal_mask = torch.ones((qc.shape[-2], kc.shape[-2]), dtype=torch.bool, device=device).triu(
q_start_index - k_start_index + 1
)
attn_weights.masked_fill_(causal_mask, max_neg_value)
block_row_maxes = attn_weights.amax(dim=-1, keepdims=True)
attn_weights -= block_row_maxes
exp_weights = torch.exp(attn_weights)
if exists(row_mask):
exp_weights.masked_fill_(~row_mask, 0.0)
block_row_sums = exp_weights.sum(dim=-1, keepdims=True).clamp(min=EPSILON)
new_row_maxes = torch.maximum(block_row_maxes, row_maxes)
exp_values = torch.einsum("... i j, ... j d -> ... i d", exp_weights, vc)
exp_row_max_diff = torch.exp(row_maxes - new_row_maxes)
exp_block_row_max_diff = torch.exp(block_row_maxes - new_row_maxes)
new_row_sums = exp_row_max_diff * row_sums + exp_block_row_max_diff * block_row_sums
oc.mul_((row_sums / new_row_sums) * exp_row_max_diff).add_((exp_block_row_max_diff / new_row_sums) * exp_values)
row_maxes.copy_(new_row_maxes)
row_sums.copy_(new_row_sums)
ctx.args = (causal, scale, mask, q_bucket_size, k_bucket_size)
ctx.save_for_backward(q, k, v, o, all_row_sums, all_row_maxes)
return o
@staticmethod
@torch.no_grad()
def backward(ctx, do):
"""Algorithm 4 in the paper"""
causal, scale, mask, q_bucket_size, k_bucket_size = ctx.args
q, k, v, o, l, m = ctx.saved_tensors
device = q.device
max_neg_value = -torch.finfo(q.dtype).max
qk_len_diff = max(k.shape[-2] - q.shape[-2], 0)
dq = torch.zeros_like(q)
dk = torch.zeros_like(k)
dv = torch.zeros_like(v)
row_splits = zip(
q.split(q_bucket_size, dim=-2),
o.split(q_bucket_size, dim=-2),
do.split(q_bucket_size, dim=-2),
mask,
l.split(q_bucket_size, dim=-2),
m.split(q_bucket_size, dim=-2),
dq.split(q_bucket_size, dim=-2),
)
for ind, (qc, oc, doc, row_mask, lc, mc, dqc) in enumerate(row_splits):
q_start_index = ind * q_bucket_size - qk_len_diff
col_splits = zip(
k.split(k_bucket_size, dim=-2),
v.split(k_bucket_size, dim=-2),
dk.split(k_bucket_size, dim=-2),
dv.split(k_bucket_size, dim=-2),
)
for k_ind, (kc, vc, dkc, dvc) in enumerate(col_splits):
k_start_index = k_ind * k_bucket_size
attn_weights = torch.einsum("... i d, ... j d -> ... i j", qc, kc) * scale
if causal and q_start_index < (k_start_index + k_bucket_size - 1):
causal_mask = torch.ones((qc.shape[-2], kc.shape[-2]), dtype=torch.bool, device=device).triu(
q_start_index - k_start_index + 1
)
attn_weights.masked_fill_(causal_mask, max_neg_value)
exp_attn_weights = torch.exp(attn_weights - mc)
if exists(row_mask):
exp_attn_weights.masked_fill_(~row_mask, 0.0)
p = exp_attn_weights / lc
dv_chunk = torch.einsum("... i j, ... i d -> ... j d", p, doc)
dp = torch.einsum("... i d, ... j d -> ... i j", doc, vc)
D = (doc * oc).sum(dim=-1, keepdims=True)
ds = p * scale * (dp - D)
dq_chunk = torch.einsum("... i j, ... j d -> ... i d", ds, kc)
dk_chunk = torch.einsum("... i j, ... i d -> ... j d", ds, qc)
dqc.add_(dq_chunk)
dkc.add_(dk_chunk)
dvc.add_(dv_chunk)
return dq, dk, dv, None, None, None, None
# endregion
def get_parameter_dtype(parameter: torch.nn.Module):
return next(parameter.parameters()).dtype
def get_parameter_device(parameter: torch.nn.Module):
return next(parameter.parameters()).device
def get_timestep_embedding(
timesteps: torch.Tensor,
embedding_dim: int,
flip_sin_to_cos: bool = False,
downscale_freq_shift: float = 1,
scale: float = 1,
max_period: int = 10000,
):
"""
This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.
:param timesteps: a 1-D Tensor of N indices, one per batch element.
These may be fractional.
:param embedding_dim: the dimension of the output. :param max_period: controls the minimum frequency of the
embeddings. :return: an [N x dim] Tensor of positional embeddings.
"""
assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
half_dim = embedding_dim // 2
exponent = -math.log(max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device)
exponent = exponent / (half_dim - downscale_freq_shift)
emb = torch.exp(exponent)
emb = timesteps[:, None].float() * emb[None, :]
# scale embeddings
emb = scale * emb
# concat sine and cosine embeddings
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
# flip sine and cosine embeddings
if flip_sin_to_cos:
emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
# zero pad
if embedding_dim % 2 == 1:
emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
return emb
# Deep Shrink: We do not common this function, because minimize dependencies.
def resize_like(x, target, mode="bicubic", align_corners=False):
org_dtype = x.dtype
if org_dtype == torch.bfloat16:
x = x.to(torch.float32)
if x.shape[-2:] != target.shape[-2:]:
if mode == "nearest":
x = F.interpolate(x, size=target.shape[-2:], mode=mode)
else:
x = F.interpolate(x, size=target.shape[-2:], mode=mode, align_corners=align_corners)
if org_dtype == torch.bfloat16:
x = x.to(org_dtype)
return x
class SampleOutput:
def __init__(self, sample):
self.sample = sample
class TimestepEmbedding(nn.Module):
def __init__(self, in_channels: int, time_embed_dim: int, act_fn: str = "silu", out_dim: int = None):
super().__init__()
self.linear_1 = nn.Linear(in_channels, time_embed_dim)
self.act = None
if act_fn == "silu":
self.act = nn.SiLU()
elif act_fn == "mish":
self.act = nn.Mish()
if out_dim is not None:
time_embed_dim_out = out_dim
else:
time_embed_dim_out = time_embed_dim
self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim_out)
def forward(self, sample):
sample = self.linear_1(sample)
if self.act is not None:
sample = self.act(sample)
sample = self.linear_2(sample)
return sample
class Timesteps(nn.Module):
def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float):
super().__init__()
self.num_channels = num_channels
self.flip_sin_to_cos = flip_sin_to_cos
self.downscale_freq_shift = downscale_freq_shift
def forward(self, timesteps):
t_emb = get_timestep_embedding(
timesteps,
self.num_channels,
flip_sin_to_cos=self.flip_sin_to_cos,
downscale_freq_shift=self.downscale_freq_shift,
)
return t_emb
class ResnetBlock2D(nn.Module):
def __init__(
self,
in_channels,
out_channels,
):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.norm1 = torch.nn.GroupNorm(num_groups=NORM_GROUPS, num_channels=in_channels, eps=NORM_EPS, affine=True)
self.conv1 = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
self.time_emb_proj = torch.nn.Linear(TIME_EMBED_DIM, out_channels)
self.norm2 = torch.nn.GroupNorm(num_groups=NORM_GROUPS, num_channels=out_channels, eps=NORM_EPS, affine=True)
self.conv2 = torch.nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
# if non_linearity == "swish":
self.nonlinearity = lambda x: F.silu(x)
self.use_in_shortcut = self.in_channels != self.out_channels
self.conv_shortcut = None
if self.use_in_shortcut:
self.conv_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
def forward(self, input_tensor, temb):
hidden_states = input_tensor
hidden_states = self.norm1(hidden_states)
hidden_states = self.nonlinearity(hidden_states)
hidden_states = self.conv1(hidden_states)
temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None]
hidden_states = hidden_states + temb
hidden_states = self.norm2(hidden_states)
hidden_states = self.nonlinearity(hidden_states)
hidden_states = self.conv2(hidden_states)
if self.conv_shortcut is not None:
input_tensor = self.conv_shortcut(input_tensor)
output_tensor = input_tensor + hidden_states
return output_tensor
class DownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
add_downsample=True,
):
super().__init__()
self.has_cross_attention = False
resnets = []
for i in range(LAYERS_PER_BLOCK):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
)
)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = [Downsample2D(out_channels, out_channels=out_channels)]
else:
self.downsamplers = None
self.gradient_checkpointing = False
def set_use_memory_efficient_attention(self, xformers, mem_eff):
pass
def set_use_sdpa(self, sdpa):
pass
def forward(self, hidden_states, temb=None):
output_states = ()
for resnet in self.resnets:
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
else:
hidden_states = resnet(hidden_states, temb)
output_states += (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states)
output_states += (hidden_states,)
return hidden_states, output_states
class Downsample2D(nn.Module):
def __init__(self, channels, out_channels):
super().__init__()
self.channels = channels
self.out_channels = out_channels
self.conv = nn.Conv2d(self.channels, self.out_channels, 3, stride=2, padding=1)
def forward(self, hidden_states):
assert hidden_states.shape[1] == self.channels
hidden_states = self.conv(hidden_states)
return hidden_states
class CrossAttention(nn.Module):
def __init__(
self,
query_dim: int,
cross_attention_dim: Optional[int] = None,
heads: int = 8,
dim_head: int = 64,
upcast_attention: bool = False,
):
super().__init__()
inner_dim = dim_head * heads
cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
self.upcast_attention = upcast_attention
self.scale = dim_head**-0.5
self.heads = heads
self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=False)
self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=False)
self.to_out = nn.ModuleList([])
self.to_out.append(nn.Linear(inner_dim, query_dim))
# no dropout here
self.use_memory_efficient_attention_xformers = False
self.use_memory_efficient_attention_mem_eff = False
self.use_sdpa = False
# Attention processor
self.processor = None
def set_use_memory_efficient_attention(self, xformers, mem_eff):
self.use_memory_efficient_attention_xformers = xformers
self.use_memory_efficient_attention_mem_eff = mem_eff
def set_use_sdpa(self, sdpa):
self.use_sdpa = sdpa
def reshape_heads_to_batch_dim(self, tensor):
batch_size, seq_len, dim = tensor.shape
head_size = self.heads
tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
return tensor
def reshape_batch_dim_to_heads(self, tensor):
batch_size, seq_len, dim = tensor.shape
head_size = self.heads
tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
return tensor
def set_processor(self):
return self.processor
def get_processor(self):
return self.processor
def forward(self, hidden_states, context=None, mask=None, **kwargs):
if self.processor is not None:
(
hidden_states,
encoder_hidden_states,
attention_mask,
) = translate_attention_names_from_diffusers(
hidden_states=hidden_states, context=context, mask=mask, **kwargs
)
return self.processor(
attn=self,
hidden_states=hidden_states,
encoder_hidden_states=context,
attention_mask=mask,
**kwargs
)
if self.use_memory_efficient_attention_xformers:
return self.forward_memory_efficient_xformers(hidden_states, context, mask)
if self.use_memory_efficient_attention_mem_eff:
return self.forward_memory_efficient_mem_eff(hidden_states, context, mask)
if self.use_sdpa:
return self.forward_sdpa(hidden_states, context, mask)
query = self.to_q(hidden_states)
context = context if context is not None else hidden_states
key = self.to_k(context)
value = self.to_v(context)
query = self.reshape_heads_to_batch_dim(query)
key = self.reshape_heads_to_batch_dim(key)
value = self.reshape_heads_to_batch_dim(value)
hidden_states = self._attention(query, key, value)
# linear proj
hidden_states = self.to_out[0](hidden_states)
# hidden_states = self.to_out[1](hidden_states) # no dropout
return hidden_states
def _attention(self, query, key, value):
if self.upcast_attention:
query = query.float()
key = key.float()
attention_scores = torch.baddbmm(
torch.empty(query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device),
query,
key.transpose(-1, -2),
beta=0,
alpha=self.scale,
)
attention_probs = attention_scores.softmax(dim=-1)
# cast back to the original dtype
attention_probs = attention_probs.to(value.dtype)
# compute attention output
hidden_states = torch.bmm(attention_probs, value)
# reshape hidden_states
hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
return hidden_states
# TODO support Hypernetworks
def forward_memory_efficient_xformers(self, x, context=None, mask=None):
import xformers.ops
h = self.heads
q_in = self.to_q(x)
context = context if context is not None else x
context = context.to(x.dtype)
k_in = self.to_k(context)
v_in = self.to_v(context)
q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b n h d", h=h), (q_in, k_in, v_in))
del q_in, k_in, v_in
q = q.contiguous()
k = k.contiguous()
v = v.contiguous()
out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None) # 最適なのを選んでくれる
out = rearrange(out, "b n h d -> b n (h d)", h=h)
out = self.to_out[0](out)
return out
def forward_memory_efficient_mem_eff(self, x, context=None, mask=None):
flash_func = FlashAttentionFunction
q_bucket_size = 512
k_bucket_size = 1024
h = self.heads
q = self.to_q(x)
context = context if context is not None else x
context = context.to(x.dtype)
k = self.to_k(context)
v = self.to_v(context)
del context, x
q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v))
out = flash_func.apply(q, k, v, mask, False, q_bucket_size, k_bucket_size)
out = rearrange(out, "b h n d -> b n (h d)")
out = self.to_out[0](out)
return out
def forward_sdpa(self, x, context=None, mask=None):
h = self.heads
q_in = self.to_q(x)
context = context if context is not None else x
context = context.to(x.dtype)
k_in = self.to_k(context)
v_in = self.to_v(context)
q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q_in, k_in, v_in))
del q_in, k_in, v_in
out = F.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
out = rearrange(out, "b h n d -> b n (h d)", h=h)
out = self.to_out[0](out)
return out
def translate_attention_names_from_diffusers(
hidden_states: torch.FloatTensor,
context: Optional[torch.FloatTensor] = None,
mask: Optional[torch.FloatTensor] = None,
# HF naming
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None
):
# translate from hugging face diffusers
context = context if context is not None else encoder_hidden_states
# translate from hugging face diffusers
mask = mask if mask is not None else attention_mask
return hidden_states, context, mask
# feedforward
class GEGLU(nn.Module):
r"""
A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202.
Parameters:
dim_in (`int`): The number of channels in the input.
dim_out (`int`): The number of channels in the output.
"""
def __init__(self, dim_in: int, dim_out: int):
super().__init__()
self.proj = nn.Linear(dim_in, dim_out * 2)
def gelu(self, gate):
if gate.device.type != "mps":
return F.gelu(gate)
# mps: gelu is not implemented for float16
return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
def forward(self, hidden_states):
hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1)
return hidden_states * self.gelu(gate)
class FeedForward(nn.Module):
def __init__(
self,
dim: int,
):
super().__init__()
inner_dim = int(dim * 4) # mult is always 4
self.net = nn.ModuleList([])
# project in
self.net.append(GEGLU(dim, inner_dim))
# project dropout
self.net.append(nn.Identity()) # nn.Dropout(0)) # dummy for dropout with 0
# project out
self.net.append(nn.Linear(inner_dim, dim))
def forward(self, hidden_states):
for module in self.net:
hidden_states = module(hidden_states)
return hidden_states
class BasicTransformerBlock(nn.Module):
def __init__(
self, dim: int, num_attention_heads: int, attention_head_dim: int, cross_attention_dim: int, upcast_attention: bool = False
):
super().__init__()
# 1. Self-Attn
self.attn1 = CrossAttention(
query_dim=dim,
cross_attention_dim=None,
heads=num_attention_heads,
dim_head=attention_head_dim,
upcast_attention=upcast_attention,
)
self.ff = FeedForward(dim)
# 2. Cross-Attn
self.attn2 = CrossAttention(
query_dim=dim,
cross_attention_dim=cross_attention_dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
upcast_attention=upcast_attention,
)
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
# 3. Feed-forward
self.norm3 = nn.LayerNorm(dim)
def set_use_memory_efficient_attention(self, xformers: bool, mem_eff: bool):
self.attn1.set_use_memory_efficient_attention(xformers, mem_eff)
self.attn2.set_use_memory_efficient_attention(xformers, mem_eff)
def set_use_sdpa(self, sdpa: bool):
self.attn1.set_use_sdpa(sdpa)
self.attn2.set_use_sdpa(sdpa)
def forward(self, hidden_states, context=None, timestep=None):
# 1. Self-Attention
norm_hidden_states = self.norm1(hidden_states)
hidden_states = self.attn1(norm_hidden_states) + hidden_states
# 2. Cross-Attention
norm_hidden_states = self.norm2(hidden_states)
hidden_states = self.attn2(norm_hidden_states, context=context) + hidden_states
# 3. Feed-forward
hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
return hidden_states
class Transformer2DModel(nn.Module):
def __init__(
self,
num_attention_heads: int = 16,
attention_head_dim: int = 88,
in_channels: Optional[int] = None,
cross_attention_dim: Optional[int] = None,
use_linear_projection: bool = False,
upcast_attention: bool = False,
):
super().__init__()
self.in_channels = in_channels
self.num_attention_heads = num_attention_heads
self.attention_head_dim = attention_head_dim
inner_dim = num_attention_heads * attention_head_dim
self.use_linear_projection = use_linear_projection
self.norm = torch.nn.GroupNorm(num_groups=TRANSFORMER_NORM_NUM_GROUPS, num_channels=in_channels, eps=1e-6, affine=True)
if use_linear_projection:
self.proj_in = nn.Linear(in_channels, inner_dim)
else:
self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
self.transformer_blocks = nn.ModuleList(
[
BasicTransformerBlock(
inner_dim,
num_attention_heads,
attention_head_dim,
cross_attention_dim=cross_attention_dim,
upcast_attention=upcast_attention,
)
]
)
if use_linear_projection:
self.proj_out = nn.Linear(in_channels, inner_dim)
else:
self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
def set_use_memory_efficient_attention(self, xformers, mem_eff):
for transformer in self.transformer_blocks:
transformer.set_use_memory_efficient_attention(xformers, mem_eff)
def set_use_sdpa(self, sdpa):
for transformer in self.transformer_blocks:
transformer.set_use_sdpa(sdpa)
def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True):
# 1. Input
batch, _, height, weight = hidden_states.shape
residual = hidden_states
hidden_states = self.norm(hidden_states)
if not self.use_linear_projection:
hidden_states = self.proj_in(hidden_states)
inner_dim = hidden_states.shape[1]
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
else:
inner_dim = hidden_states.shape[1]
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
hidden_states = self.proj_in(hidden_states)
# 2. Blocks
for block in self.transformer_blocks:
hidden_states = block(hidden_states, context=encoder_hidden_states, timestep=timestep)
# 3. Output
if not self.use_linear_projection:
hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
hidden_states = self.proj_out(hidden_states)
else:
hidden_states = self.proj_out(hidden_states)
hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
output = hidden_states + residual
if not return_dict:
return (output,)
return SampleOutput(sample=output)
class CrossAttnDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
add_downsample=True,
cross_attention_dim=1280,
attn_num_head_channels=1,
use_linear_projection=False,
upcast_attention=False,
):
super().__init__()
self.has_cross_attention = True
resnets = []
attentions = []
self.attn_num_head_channels = attn_num_head_channels
for i in range(LAYERS_PER_BLOCK):
in_channels = in_channels if i == 0 else out_channels
resnets.append(ResnetBlock2D(in_channels=in_channels, out_channels=out_channels))
attentions.append(
Transformer2DModel(
attn_num_head_channels,
out_channels // attn_num_head_channels,
in_channels=out_channels,
cross_attention_dim=cross_attention_dim,
use_linear_projection=use_linear_projection,
upcast_attention=upcast_attention,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList([Downsample2D(out_channels, out_channels)])
else:
self.downsamplers = None
self.gradient_checkpointing = False
def set_use_memory_efficient_attention(self, xformers, mem_eff):
for attn in self.attentions:
attn.set_use_memory_efficient_attention(xformers, mem_eff)
def set_use_sdpa(self, sdpa):
for attn in self.attentions: