"docs/vscode:/vscode.git/clone" did not exist on "f9d5a9324d77169d486a60f3b4b267c74149b982"
unet_2d_blocks.py 54.7 KB
Newer Older
Patrick von Platen's avatar
Patrick von Platen committed
1
2
3
4
5
6
7
8
9
10
11
12
13
# Copyright 2022 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.
14
import numpy as np
15
import torch
Patrick von Platen's avatar
Patrick von Platen committed
16
17
from torch import nn

18
19
from .attention import AttentionBlock, SpatialTransformer
from .resnet import Downsample2D, FirDownsample2D, FirUpsample2D, ResnetBlock2D, Upsample2D
Patrick von Platen's avatar
Patrick von Platen committed
20
21


22
23
24
25
26
27
28
29
30
31
def get_down_block(
    down_block_type,
    num_layers,
    in_channels,
    out_channels,
    temb_channels,
    add_downsample,
    resnet_eps,
    resnet_act_fn,
    attn_num_head_channels,
32
    resnet_groups=None,
33
    cross_attention_dim=None,
Patrick von Platen's avatar
Patrick von Platen committed
34
    downsample_padding=None,
35
):
Patrick von Platen's avatar
Patrick von Platen committed
36
37
38
    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
    if down_block_type == "DownBlock2D":
        return DownBlock2D(
39
40
41
42
43
44
45
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
46
            resnet_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
47
            downsample_padding=downsample_padding,
48
        )
Patrick von Platen's avatar
Patrick von Platen committed
49
50
    elif down_block_type == "AttnDownBlock2D":
        return AttnDownBlock2D(
51
52
53
54
55
56
57
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
58
            resnet_groups=resnet_groups,
59
            downsample_padding=downsample_padding,
60
61
            attn_num_head_channels=attn_num_head_channels,
        )
Patrick von Platen's avatar
Patrick von Platen committed
62
    elif down_block_type == "CrossAttnDownBlock2D":
63
        if cross_attention_dim is None:
64
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
Patrick von Platen's avatar
Patrick von Platen committed
65
        return CrossAttnDownBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
66
67
68
69
70
71
72
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
73
            resnet_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
74
            downsample_padding=downsample_padding,
75
            cross_attention_dim=cross_attention_dim,
Patrick von Platen's avatar
Patrick von Platen committed
76
77
            attn_num_head_channels=attn_num_head_channels,
        )
Patrick von Platen's avatar
Patrick von Platen committed
78
79
    elif down_block_type == "SkipDownBlock2D":
        return SkipDownBlock2D(
80
81
82
83
84
85
86
87
88
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            downsample_padding=downsample_padding,
        )
Patrick von Platen's avatar
Patrick von Platen committed
89
90
    elif down_block_type == "AttnSkipDownBlock2D":
        return AttnSkipDownBlock2D(
91
92
93
94
95
96
97
98
99
100
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            downsample_padding=downsample_padding,
            attn_num_head_channels=attn_num_head_channels,
        )
101
102
103
104
105
106
107
108
    elif down_block_type == "DownEncoderBlock2D":
        return DownEncoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
109
            resnet_groups=resnet_groups,
110
111
            downsample_padding=downsample_padding,
        )
112
113
114
115
116
117


def get_up_block(
    up_block_type,
    num_layers,
    in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
118
119
    out_channels,
    prev_output_channel,
120
121
122
123
124
    temb_channels,
    add_upsample,
    resnet_eps,
    resnet_act_fn,
    attn_num_head_channels,
125
    resnet_groups=None,
126
    cross_attention_dim=None,
127
):
Patrick von Platen's avatar
Patrick von Platen committed
128
129
130
    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
    if up_block_type == "UpBlock2D":
        return UpBlock2D(
131
132
            num_layers=num_layers,
            in_channels=in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
133
134
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
135
136
137
138
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
139
            resnet_groups=resnet_groups,
140
        )
Patrick von Platen's avatar
Patrick von Platen committed
141
    elif up_block_type == "CrossAttnUpBlock2D":
142
143
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
Patrick von Platen's avatar
Patrick von Platen committed
144
        return CrossAttnUpBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
145
146
147
148
149
150
151
152
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
153
            resnet_groups=resnet_groups,
154
            cross_attention_dim=cross_attention_dim,
Patrick von Platen's avatar
Patrick von Platen committed
155
156
            attn_num_head_channels=attn_num_head_channels,
        )
Patrick von Platen's avatar
Patrick von Platen committed
157
158
    elif up_block_type == "AttnUpBlock2D":
        return AttnUpBlock2D(
159
160
            num_layers=num_layers,
            in_channels=in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
161
162
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
163
164
165
166
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
167
            resnet_groups=resnet_groups,
168
169
            attn_num_head_channels=attn_num_head_channels,
        )
Patrick von Platen's avatar
Patrick von Platen committed
170
171
    elif up_block_type == "SkipUpBlock2D":
        return SkipUpBlock2D(
172
173
174
175
176
177
178
179
180
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
        )
Patrick von Platen's avatar
Patrick von Platen committed
181
182
    elif up_block_type == "AttnSkipUpBlock2D":
        return AttnSkipUpBlock2D(
183
184
185
186
187
188
189
190
191
192
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            attn_num_head_channels=attn_num_head_channels,
        )
193
194
195
196
197
198
199
200
    elif up_block_type == "UpDecoderBlock2D":
        return UpDecoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
201
            resnet_groups=resnet_groups,
202
        )
203
    raise ValueError(f"{up_block_type} does not exist.")
204
205


Patrick von Platen's avatar
Patrick von Platen committed
206
207
208
209
210
class UNetMidBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
211
        dropout: float = 0.0,
212
        num_layers: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
213
214
215
216
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
217
        resnet_pre_norm: bool = True,
218
        attn_num_head_channels=1,
Patrick von Platen's avatar
Patrick von Platen committed
219
        attention_type="default",
Patrick von Platen's avatar
Patrick von Platen committed
220
        output_scale_factor=1.0,
221
        **kwargs,
Patrick von Platen's avatar
Patrick von Platen committed
222
223
224
    ):
        super().__init__()

Patrick von Platen's avatar
Patrick von Platen committed
225
        self.attention_type = attention_type
226
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
Patrick von Platen's avatar
Patrick von Platen committed
227

228
229
        # there is always at least one resnet
        resnets = [
230
            ResnetBlock2D(
231
232
233
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
234
                eps=resnet_eps,
235
236
237
238
239
240
                groups=resnet_groups,
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
Patrick von Platen's avatar
Patrick von Platen committed
241
            )
242
243
        ]
        attentions = []
Patrick von Platen's avatar
Patrick von Platen committed
244

245
246
        for _ in range(num_layers):
            attentions.append(
247
                AttentionBlock(
248
249
250
                    in_channels,
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
251
                    eps=resnet_eps,
252
                    num_groups=resnet_groups,
253
                )
254
            )
255
            resnets.append(
256
                ResnetBlock2D(
257
258
259
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
260
                    eps=resnet_eps,
261
262
263
264
265
266
267
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
Patrick von Platen's avatar
Patrick von Platen committed
268
269
            )

270
271
272
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

Patrick von Platen's avatar
Patrick von Platen committed
273
274
    def forward(self, hidden_states, temb=None, encoder_states=None):
        hidden_states = self.resnets[0](hidden_states, temb)
275
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
Patrick von Platen's avatar
Patrick von Platen committed
276
277
            if self.attention_type == "default":
                hidden_states = attn(hidden_states)
278
            else:
Patrick von Platen's avatar
Patrick von Platen committed
279
280
                hidden_states = attn(hidden_states, encoder_states)
            hidden_states = resnet(hidden_states, temb)
Patrick von Platen's avatar
Patrick von Platen committed
281

282
        return hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
283

284

Patrick von Platen's avatar
Patrick von Platen committed
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
class UNetMidBlock2DCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        attn_num_head_channels=1,
        attention_type="default",
        output_scale_factor=1.0,
        cross_attention_dim=1280,
        **kwargs,
    ):
        super().__init__()

        self.attention_type = attention_type
306
        self.attn_num_head_channels = attn_num_head_channels
Patrick von Platen's avatar
Patrick von Platen committed
307
308
309
310
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

        # there is always at least one resnet
        resnets = [
311
            ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=resnet_groups,
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
            )
        ]
        attentions = []

        for _ in range(num_layers):
            attentions.append(
                SpatialTransformer(
                    in_channels,
                    attn_num_head_channels,
                    in_channels // attn_num_head_channels,
                    depth=1,
                    context_dim=cross_attention_dim,
334
                    num_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
335
336
337
                )
            )
            resnets.append(
338
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
    def set_attention_slice(self, slice_size):
        if slice_size is not None and self.attn_num_head_channels % slice_size != 0:
            raise ValueError(
                f"Make sure slice_size {slice_size} is a divisor of "
                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
            )
        if slice_size is not None and slice_size > self.attn_num_head_channels:
            raise ValueError(
                f"Chunk_size {slice_size} has to be smaller or equal to "
                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
            )

        for attn in self.attentions:
            attn._set_attention_slice(slice_size)

370
371
372
373
    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
        for attn in self.attentions:
            attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)

Patrick von Platen's avatar
Patrick von Platen committed
374
375
376
377
378
379
380
381
382
    def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            hidden_states = attn(hidden_states, encoder_hidden_states)
            hidden_states = resnet(hidden_states, temb)

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
383
class AttnDownBlock2D(nn.Module):
384
385
386
387
388
389
390
391
392
393
394
395
396
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        attn_num_head_channels=1,
Patrick von Platen's avatar
Patrick von Platen committed
397
        attention_type="default",
398
        output_scale_factor=1.0,
399
        downsample_padding=1,
400
401
402
403
404
405
        add_downsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

Patrick von Platen's avatar
Patrick von Platen committed
406
407
        self.attention_type = attention_type

408
409
410
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
411
                ResnetBlock2D(
412
413
414
415
416
417
418
419
420
421
422
423
424
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            attentions.append(
425
                AttentionBlock(
426
427
428
                    out_channels,
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
429
                    eps=resnet_eps,
430
                    num_groups=resnet_groups,
431
432
433
434
435
436
437
438
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
439
440
441
442
443
                [
                    Downsample2D(
                        in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
                    )
                ]
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
            )
        else:
            self.downsamplers = None

    def forward(self, hidden_states, temb=None):
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb)
            hidden_states = attn(hidden_states)
            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


Patrick von Platen's avatar
Patrick von Platen committed
465
class CrossAttnDownBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        attn_num_head_channels=1,
        cross_attention_dim=1280,
        attention_type="default",
        output_scale_factor=1.0,
        downsample_padding=1,
        add_downsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.attention_type = attention_type
490
        self.attn_num_head_channels = attn_num_head_channels
Patrick von Platen's avatar
Patrick von Platen committed
491
492
493
494

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
495
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            attentions.append(
                SpatialTransformer(
                    out_channels,
                    attn_num_head_channels,
                    out_channels // attn_num_head_channels,
                    depth=1,
                    context_dim=cross_attention_dim,
515
                    num_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
                )
            )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
                        in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
                    )
                ]
            )
        else:
            self.downsamplers = None

532
533
        self.gradient_checkpointing = False

534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
    def set_attention_slice(self, slice_size):
        if slice_size is not None and self.attn_num_head_channels % slice_size != 0:
            raise ValueError(
                f"Make sure slice_size {slice_size} is a divisor of "
                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
            )
        if slice_size is not None and slice_size > self.attn_num_head_channels:
            raise ValueError(
                f"Chunk_size {slice_size} has to be smaller or equal to "
                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
            )

        for attn in self.attentions:
            attn._set_attention_slice(slice_size)

549
550
551
552
    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
        for attn in self.attentions:
            attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)

Patrick von Platen's avatar
Patrick von Platen committed
553
554
555
556
    def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
            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)
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn), hidden_states, encoder_hidden_states
                )
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(hidden_states, context=encoder_hidden_states)

Patrick von Platen's avatar
Patrick von Platen committed
573
574
575
576
577
578
579
580
581
582
583
            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


Patrick von Platen's avatar
Patrick von Platen committed
584
class DownBlock2D(nn.Module):
585
586
587
588
589
590
591
592
593
594
595
596
597
598
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_downsample=True,
Patrick von Platen's avatar
Patrick von Platen committed
599
        downsample_padding=1,
600
601
602
603
604
605
606
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
607
                ResnetBlock2D(
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
Patrick von Platen's avatar
Patrick von Platen committed
625
626
627
628
629
                [
                    Downsample2D(
                        in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
                    )
                ]
630
631
632
633
            )
        else:
            self.downsamplers = None

634
635
        self.gradient_checkpointing = False

636
637
638
639
    def forward(self, hidden_states, temb=None):
        output_states = ()

        for resnet in self.resnets:
640
641
642
643
644
645
646
647
648
649
650
651
            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)

652
653
654
655
656
657
658
659
660
661
662
            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


663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
class DownEncoderBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
685
                ResnetBlock2D(
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
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=None,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
                        in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
                    )
                ]
            )
        else:
            self.downsamplers = None

    def forward(self, hidden_states):
        for resnet in self.resnets:
            hidden_states = resnet(hidden_states, temb=None)

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
                hidden_states = downsampler(hidden_states)

        return hidden_states


723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
class AttnDownEncoderBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        attn_num_head_channels=1,
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []
        attentions = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
747
                ResnetBlock2D(
748
749
750
751
752
753
754
755
756
757
758
759
760
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=None,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            attentions.append(
761
                AttentionBlock(
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
                    out_channels,
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
                    num_groups=resnet_groups,
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
                        in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
                    )
                ]
            )
        else:
            self.downsamplers = None

    def forward(self, hidden_states):
        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb=None)
            hidden_states = attn(hidden_states)

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
                hidden_states = downsampler(hidden_states)

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
796
class AttnSkipDownBlock2D(nn.Module):
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
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_pre_norm: bool = True,
        attn_num_head_channels=1,
        attention_type="default",
        output_scale_factor=np.sqrt(2.0),
        downsample_padding=1,
        add_downsample=True,
    ):
        super().__init__()
        self.attentions = nn.ModuleList([])
        self.resnets = nn.ModuleList([])

        self.attention_type = attention_type

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            self.resnets.append(
823
                ResnetBlock2D(
824
825
826
827
828
829
830
831
832
833
834
835
836
837
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min(in_channels // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            self.attentions.append(
838
                AttentionBlock(
839
840
841
842
843
844
845
846
                    out_channels,
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
                )
            )

        if add_downsample:
847
            self.resnet_down = ResnetBlock2D(
848
849
850
851
852
853
854
855
856
857
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
858
                use_in_shortcut=True,
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
                down=True,
                kernel="fir",
            )
            self.downsamplers = nn.ModuleList([FirDownsample2D(in_channels, out_channels=out_channels)])
            self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
        else:
            self.resnet_down = None
            self.downsamplers = None
            self.skip_conv = None

    def forward(self, hidden_states, temb=None, skip_sample=None):
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb)
            hidden_states = attn(hidden_states)
            output_states += (hidden_states,)

        if self.downsamplers is not None:
            hidden_states = self.resnet_down(hidden_states, temb)
            for downsampler in self.downsamplers:
                skip_sample = downsampler(skip_sample)

            hidden_states = self.skip_conv(skip_sample) + hidden_states

            output_states += (hidden_states,)

        return hidden_states, output_states, skip_sample


Patrick von Platen's avatar
Patrick von Platen committed
889
class SkipDownBlock2D(nn.Module):
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_pre_norm: bool = True,
        output_scale_factor=np.sqrt(2.0),
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        self.resnets = nn.ModuleList([])

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            self.resnets.append(
911
                ResnetBlock2D(
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
                    in_channels=in_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min(in_channels // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        if add_downsample:
927
            self.resnet_down = ResnetBlock2D(
928
929
930
931
932
933
934
935
936
937
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
938
                use_in_shortcut=True,
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
                down=True,
                kernel="fir",
            )
            self.downsamplers = nn.ModuleList([FirDownsample2D(in_channels, out_channels=out_channels)])
            self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
        else:
            self.resnet_down = None
            self.downsamplers = None
            self.skip_conv = None

    def forward(self, hidden_states, temb=None, skip_sample=None):
        output_states = ()

        for resnet in self.resnets:
            hidden_states = resnet(hidden_states, temb)
            output_states += (hidden_states,)

        if self.downsamplers is not None:
            hidden_states = self.resnet_down(hidden_states, temb)
            for downsampler in self.downsamplers:
                skip_sample = downsampler(skip_sample)

            hidden_states = self.skip_conv(skip_sample) + hidden_states

            output_states += (hidden_states,)

        return hidden_states, output_states, skip_sample


Patrick von Platen's avatar
Patrick von Platen committed
968
class AttnUpBlock2D(nn.Module):
969
970
971
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
972
973
        prev_output_channel: int,
        out_channels: int,
974
975
976
977
978
979
980
981
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
Patrick von Platen's avatar
Patrick von Platen committed
982
        attention_type="default",
983
984
985
986
987
988
989
990
        attn_num_head_channels=1,
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

Patrick von Platen's avatar
Patrick von Platen committed
991
992
        self.attention_type = attention_type

993
        for i in range(num_layers):
Patrick von Platen's avatar
Patrick von Platen committed
994
995
996
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

997
            resnets.append(
998
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
999
1000
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            attentions.append(
1012
                AttentionBlock(
Patrick von Platen's avatar
Patrick von Platen committed
1013
                    out_channels,
1014
1015
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
1016
                    eps=resnet_eps,
1017
                    num_groups=resnet_groups,
1018
1019
1020
1021
1022
1023
1024
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
Patrick von Platen's avatar
Patrick von Platen committed
1025
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
        else:
            self.upsamplers = None

    def forward(self, hidden_states, res_hidden_states_tuple, temb=None):
        for resnet, attn in zip(self.resnets, self.attentions):
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

            hidden_states = resnet(hidden_states, temb)
            hidden_states = attn(hidden_states)

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
                hidden_states = upsampler(hidden_states)

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
1046
class CrossAttnUpBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        attn_num_head_channels=1,
        cross_attention_dim=1280,
        attention_type="default",
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.attention_type = attention_type
1071
        self.attn_num_head_channels = attn_num_head_channels
Patrick von Platen's avatar
Patrick von Platen committed
1072
1073
1074
1075
1076
1077

        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            resnets.append(
1078
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            attentions.append(
                SpatialTransformer(
                    out_channels,
                    attn_num_head_channels,
                    out_channels // attn_num_head_channels,
                    depth=1,
                    context_dim=cross_attention_dim,
1098
                    num_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
                )
            )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

1109
1110
        self.gradient_checkpointing = False

1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
    def set_attention_slice(self, slice_size):
        if slice_size is not None and self.attn_num_head_channels % slice_size != 0:
            raise ValueError(
                f"Make sure slice_size {slice_size} is a divisor of "
                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
            )
        if slice_size is not None and slice_size > self.attn_num_head_channels:
            raise ValueError(
                f"Chunk_size {slice_size} has to be smaller or equal to "
                f"the number of heads used in cross_attention {self.attn_num_head_channels}"
            )

        for attn in self.attentions:
            attn._set_attention_slice(slice_size)

1126
1127
        self.gradient_checkpointing = False

1128
1129
1130
1131
    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
        for attn in self.attentions:
            attn._set_use_memory_efficient_attention_xformers(use_memory_efficient_attention_xformers)

1132
1133
1134
1135
1136
1137
    def forward(
        self,
        hidden_states,
        res_hidden_states_tuple,
        temb=None,
        encoder_hidden_states=None,
1138
        upsample_size=None,
1139
    ):
Patrick von Platen's avatar
Patrick von Platen committed
1140
1141
1142
1143
1144
1145
        for resnet, attn in zip(self.resnets, self.attentions):
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
            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)
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn), hidden_states, encoder_hidden_states
                )
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(hidden_states, context=encoder_hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
1161
1162
1163

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
1164
                hidden_states = upsampler(hidden_states, upsample_size)
Patrick von Platen's avatar
Patrick von Platen committed
1165
1166
1167
1168

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
1169
class UpBlock2D(nn.Module):
1170
1171
1172
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
1173
1174
        prev_output_channel: int,
        out_channels: int,
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
Patrick von Platen's avatar
Patrick von Platen committed
1190
1191
1192
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

1193
            resnets.append(
1194
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
1195
1196
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
Patrick von Platen's avatar
Patrick von Platen committed
1211
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
1212
1213
1214
        else:
            self.upsamplers = None

1215
1216
        self.gradient_checkpointing = False

1217
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
1218
1219
1220
1221
1222
1223
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
            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)
1235
1236
1237

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
1238
                hidden_states = upsampler(hidden_states, upsample_size)
1239
1240

        return hidden_states
1241
1242


1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
class UpDecoderBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            input_channels = in_channels if i == 0 else out_channels

            resnets.append(
1265
                ResnetBlock2D(
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
                    in_channels=input_channels,
                    out_channels=out_channels,
                    temb_channels=None,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

    def forward(self, hidden_states):
        for resnet in self.resnets:
            hidden_states = resnet(hidden_states, temb=None)

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
                hidden_states = upsampler(hidden_states)

        return hidden_states


1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
class AttnUpDecoderBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        attn_num_head_channels=1,
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

        for i in range(num_layers):
            input_channels = in_channels if i == 0 else out_channels

            resnets.append(
1321
                ResnetBlock2D(
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
                    in_channels=input_channels,
                    out_channels=out_channels,
                    temb_channels=None,
                    eps=resnet_eps,
                    groups=resnet_groups,
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )
            attentions.append(
1335
                AttentionBlock(
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
                    out_channels,
                    num_head_channels=attn_num_head_channels,
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
                    num_groups=resnet_groups,
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

    def forward(self, hidden_states):
        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb=None)
            hidden_states = attn(hidden_states)

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
                hidden_states = upsampler(hidden_states)

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
1364
class AttnSkipUpBlock2D(nn.Module):
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_pre_norm: bool = True,
        attn_num_head_channels=1,
        attention_type="default",
        output_scale_factor=np.sqrt(2.0),
        upsample_padding=1,
        add_upsample=True,
    ):
        super().__init__()
        self.attentions = nn.ModuleList([])
        self.resnets = nn.ModuleList([])

        self.attention_type = attention_type

        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            self.resnets.append(
1394
                ResnetBlock2D(
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min(resnet_in_channels + res_skip_channels // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.attentions.append(
1410
            AttentionBlock(
1411
1412
1413
1414
1415
1416
1417
1418
1419
                out_channels,
                num_head_channels=attn_num_head_channels,
                rescale_output_factor=output_scale_factor,
                eps=resnet_eps,
            )
        )

        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
        if add_upsample:
1420
            self.resnet_up = ResnetBlock2D(
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                groups_out=min(out_channels // 4, 32),
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
1432
                use_in_shortcut=True,
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
                up=True,
                kernel="fir",
            )
            self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
            self.skip_norm = torch.nn.GroupNorm(
                num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
            )
            self.act = nn.SiLU()
        else:
            self.resnet_up = None
            self.skip_conv = None
            self.skip_norm = None
            self.act = None

    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None):
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

            hidden_states = resnet(hidden_states, temb)

        hidden_states = self.attentions[0](hidden_states)

        if skip_sample is not None:
            skip_sample = self.upsampler(skip_sample)
        else:
            skip_sample = 0

        if self.resnet_up is not None:
            skip_sample_states = self.skip_norm(hidden_states)
            skip_sample_states = self.act(skip_sample_states)
            skip_sample_states = self.skip_conv(skip_sample_states)

            skip_sample = skip_sample + skip_sample_states

            hidden_states = self.resnet_up(hidden_states, temb)

        return hidden_states, skip_sample


Patrick von Platen's avatar
Patrick von Platen committed
1475
class SkipUpBlock2D(nn.Module):
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
    def __init__(
        self,
        in_channels: int,
        prev_output_channel: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_pre_norm: bool = True,
        output_scale_factor=np.sqrt(2.0),
        add_upsample=True,
        upsample_padding=1,
    ):
        super().__init__()
        self.resnets = nn.ModuleList([])

        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            self.resnets.append(
1500
                ResnetBlock2D(
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=min((resnet_in_channels + res_skip_channels) // 4, 32),
                    groups_out=min(out_channels // 4, 32),
                    dropout=dropout,
                    time_embedding_norm=resnet_time_scale_shift,
                    non_linearity=resnet_act_fn,
                    output_scale_factor=output_scale_factor,
                    pre_norm=resnet_pre_norm,
                )
            )

        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
        if add_upsample:
1517
            self.resnet_up = ResnetBlock2D(
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
                in_channels=out_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=min(out_channels // 4, 32),
                groups_out=min(out_channels // 4, 32),
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
1529
                use_in_shortcut=True,
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
                up=True,
                kernel="fir",
            )
            self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
            self.skip_norm = torch.nn.GroupNorm(
                num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
            )
            self.act = nn.SiLU()
        else:
            self.resnet_up = None
            self.skip_conv = None
            self.skip_norm = None
            self.act = None

    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, skip_sample=None):
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

            hidden_states = resnet(hidden_states, temb)

        if skip_sample is not None:
            skip_sample = self.upsampler(skip_sample)
        else:
            skip_sample = 0

        if self.resnet_up is not None:
            skip_sample_states = self.skip_norm(hidden_states)
            skip_sample_states = self.act(skip_sample_states)
            skip_sample_states = self.skip_conv(skip_sample_states)

            skip_sample = skip_sample + skip_sample_states

            hidden_states = self.resnet_up(hidden_states, temb)

        return hidden_states, skip_sample