unet_2d_blocks.py 114 KB
Newer Older
Patrick von Platen's avatar
Patrick von Platen committed
1
# Copyright 2023 The HuggingFace Team. All rights reserved.
Patrick von Platen's avatar
Patrick von Platen committed
2
3
4
5
6
7
8
9
10
11
12
13
#
# 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
from typing import Any, Dict, Optional, Tuple
15

16
import numpy as np
17
import torch
18
import torch.nn.functional as F
Patrick von Platen's avatar
Patrick von Platen committed
19
20
from torch import nn

21
from ..utils import is_torch_version, logging
22
from .attention import AdaGroupNorm
23
from .attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
24
from .dual_transformer_2d import DualTransformer2DModel
25
from .resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D
26
from .transformer_2d import Transformer2DModel
Patrick von Platen's avatar
Patrick von Platen committed
27
28


29
30
31
logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


32
33
34
35
36
37
38
39
40
def get_down_block(
    down_block_type,
    num_layers,
    in_channels,
    out_channels,
    temb_channels,
    add_downsample,
    resnet_eps,
    resnet_act_fn,
41
    num_attention_heads=None,
42
    resnet_groups=None,
43
    cross_attention_dim=None,
Patrick von Platen's avatar
Patrick von Platen committed
44
    downsample_padding=None,
45
    dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
46
    use_linear_projection=False,
47
    only_cross_attention=False,
48
    upcast_attention=False,
Will Berman's avatar
Will Berman committed
49
    resnet_time_scale_shift="default",
50
51
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
52
    cross_attention_norm=None,
53
    attention_head_dim=None,
54
):
55
56
57
58
59
60
61
    # If attn head dim is not defined, we default it to the number of heads
    if attention_head_dim is None:
        logger.warn(
            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
        )
        attention_head_dim = num_attention_heads

Patrick von Platen's avatar
Patrick von Platen committed
62
63
64
    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
    if down_block_type == "DownBlock2D":
        return DownBlock2D(
65
66
67
68
69
70
71
            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,
72
            resnet_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
73
            downsample_padding=downsample_padding,
Will Berman's avatar
Will Berman committed
74
75
76
77
78
79
80
81
82
83
84
85
86
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "ResnetDownsampleBlock2D":
        return ResnetDownsampleBlock2D(
            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,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
87
88
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
89
        )
Patrick von Platen's avatar
Patrick von Platen committed
90
91
    elif down_block_type == "AttnDownBlock2D":
        return AttnDownBlock2D(
92
93
94
95
96
97
98
            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,
99
            resnet_groups=resnet_groups,
100
            downsample_padding=downsample_padding,
101
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
102
            resnet_time_scale_shift=resnet_time_scale_shift,
103
        )
Patrick von Platen's avatar
Patrick von Platen committed
104
    elif down_block_type == "CrossAttnDownBlock2D":
105
        if cross_attention_dim is None:
106
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
Patrick von Platen's avatar
Patrick von Platen committed
107
        return CrossAttnDownBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
108
109
110
111
112
113
114
            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,
115
            resnet_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
116
            downsample_padding=downsample_padding,
117
            cross_attention_dim=cross_attention_dim,
118
            num_attention_heads=num_attention_heads,
119
            dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
120
            use_linear_projection=use_linear_projection,
121
            only_cross_attention=only_cross_attention,
122
            upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "SimpleCrossAttnDownBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D")
        return SimpleCrossAttnDownBlock2D(
            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,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
138
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
139
            resnet_time_scale_shift=resnet_time_scale_shift,
140
141
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
142
            only_cross_attention=only_cross_attention,
143
            cross_attention_norm=cross_attention_norm,
Patrick von Platen's avatar
Patrick von Platen committed
144
        )
Patrick von Platen's avatar
Patrick von Platen committed
145
146
    elif down_block_type == "SkipDownBlock2D":
        return SkipDownBlock2D(
147
148
149
150
151
152
153
154
            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,
Will Berman's avatar
Will Berman committed
155
            resnet_time_scale_shift=resnet_time_scale_shift,
156
        )
Patrick von Platen's avatar
Patrick von Platen committed
157
158
    elif down_block_type == "AttnSkipDownBlock2D":
        return AttnSkipDownBlock2D(
159
160
161
162
163
164
165
            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,
166
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
167
            resnet_time_scale_shift=resnet_time_scale_shift,
168
        )
169
170
171
172
173
174
175
176
    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,
177
            resnet_groups=resnet_groups,
178
            downsample_padding=downsample_padding,
Will Berman's avatar
Will Berman committed
179
            resnet_time_scale_shift=resnet_time_scale_shift,
180
        )
Will Berman's avatar
Will Berman committed
181
182
183
184
185
186
187
188
189
190
    elif down_block_type == "AttnDownEncoderBlock2D":
        return AttnDownEncoderBlock2D(
            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,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
191
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
192
            resnet_time_scale_shift=resnet_time_scale_shift,
Will Berman's avatar
Will Berman committed
193
        )
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
    elif down_block_type == "KDownBlock2D":
        return KDownBlock2D(
            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,
        )
    elif down_block_type == "KCrossAttnDownBlock2D":
        return KCrossAttnDownBlock2D(
            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,
            cross_attention_dim=cross_attention_dim,
214
            attention_head_dim=attention_head_dim,
215
216
            add_self_attention=True if not add_downsample else False,
        )
Will Berman's avatar
Will Berman committed
217
    raise ValueError(f"{down_block_type} does not exist.")
218
219
220
221
222
223


def get_up_block(
    up_block_type,
    num_layers,
    in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
224
225
    out_channels,
    prev_output_channel,
226
227
228
229
    temb_channels,
    add_upsample,
    resnet_eps,
    resnet_act_fn,
230
    num_attention_heads=None,
231
    resnet_groups=None,
232
    cross_attention_dim=None,
233
    dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
234
    use_linear_projection=False,
235
    only_cross_attention=False,
236
    upcast_attention=False,
Will Berman's avatar
Will Berman committed
237
    resnet_time_scale_shift="default",
238
239
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
240
    cross_attention_norm=None,
241
    attention_head_dim=None,
242
):
243
244
245
246
247
248
249
    # If attn head dim is not defined, we default it to the number of heads
    if attention_head_dim is None:
        logger.warn(
            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
        )
        attention_head_dim = num_attention_heads

Patrick von Platen's avatar
Patrick von Platen committed
250
251
252
    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
    if up_block_type == "UpBlock2D":
        return UpBlock2D(
253
254
            num_layers=num_layers,
            in_channels=in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
255
256
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
257
258
259
260
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
261
            resnet_groups=resnet_groups,
Will Berman's avatar
Will Berman committed
262
263
264
265
266
267
268
269
270
271
272
273
274
275
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "ResnetUpsampleBlock2D":
        return ResnetUpsampleBlock2D(
            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,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
276
277
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
278
        )
Patrick von Platen's avatar
Patrick von Platen committed
279
    elif up_block_type == "CrossAttnUpBlock2D":
280
281
        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
282
        return CrossAttnUpBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
283
284
285
286
287
288
289
290
            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,
291
            resnet_groups=resnet_groups,
292
            cross_attention_dim=cross_attention_dim,
293
            num_attention_heads=num_attention_heads,
294
            dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
295
            use_linear_projection=use_linear_projection,
296
            only_cross_attention=only_cross_attention,
297
            upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "SimpleCrossAttnUpBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D")
        return SimpleCrossAttnUpBlock2D(
            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,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
314
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
315
            resnet_time_scale_shift=resnet_time_scale_shift,
316
317
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
318
            only_cross_attention=only_cross_attention,
319
            cross_attention_norm=cross_attention_norm,
Patrick von Platen's avatar
Patrick von Platen committed
320
        )
Patrick von Platen's avatar
Patrick von Platen committed
321
322
    elif up_block_type == "AttnUpBlock2D":
        return AttnUpBlock2D(
323
324
            num_layers=num_layers,
            in_channels=in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
325
326
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
327
328
329
330
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
331
            resnet_groups=resnet_groups,
332
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
333
            resnet_time_scale_shift=resnet_time_scale_shift,
334
        )
Patrick von Platen's avatar
Patrick von Platen committed
335
336
    elif up_block_type == "SkipUpBlock2D":
        return SkipUpBlock2D(
337
338
339
340
341
342
343
344
            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,
Will Berman's avatar
Will Berman committed
345
            resnet_time_scale_shift=resnet_time_scale_shift,
346
        )
Patrick von Platen's avatar
Patrick von Platen committed
347
348
    elif up_block_type == "AttnSkipUpBlock2D":
        return AttnSkipUpBlock2D(
349
350
351
352
353
354
355
356
            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,
357
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
358
            resnet_time_scale_shift=resnet_time_scale_shift,
359
        )
360
361
362
363
364
365
366
367
    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,
368
            resnet_groups=resnet_groups,
Will Berman's avatar
Will Berman committed
369
            resnet_time_scale_shift=resnet_time_scale_shift,
YiYi Xu's avatar
YiYi Xu committed
370
            temb_channels=temb_channels,
371
        )
Will Berman's avatar
Will Berman committed
372
373
374
375
376
377
378
379
380
    elif up_block_type == "AttnUpDecoderBlock2D":
        return AttnUpDecoderBlock2D(
            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,
            resnet_groups=resnet_groups,
381
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
382
            resnet_time_scale_shift=resnet_time_scale_shift,
YiYi Xu's avatar
YiYi Xu committed
383
            temb_channels=temb_channels,
Will Berman's avatar
Will Berman committed
384
        )
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
    elif up_block_type == "KUpBlock2D":
        return KUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
        )
    elif up_block_type == "KCrossAttnUpBlock2D":
        return KCrossAttnUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            cross_attention_dim=cross_attention_dim,
405
            attention_head_dim=attention_head_dim,
406
407
        )

408
    raise ValueError(f"{up_block_type} does not exist.")
409
410


Patrick von Platen's avatar
Patrick von Platen committed
411
412
413
414
415
class UNetMidBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
416
        dropout: float = 0.0,
417
        num_layers: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
418
        resnet_eps: float = 1e-6,
YiYi Xu's avatar
YiYi Xu committed
419
        resnet_time_scale_shift: str = "default",  # default, spatial
Patrick von Platen's avatar
Patrick von Platen committed
420
421
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
422
        resnet_pre_norm: bool = True,
Will Berman's avatar
Will Berman committed
423
        add_attention: bool = True,
424
        attention_head_dim=1,
Patrick von Platen's avatar
Patrick von Platen committed
425
426
427
        output_scale_factor=1.0,
    ):
        super().__init__()
428
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
Will Berman's avatar
Will Berman committed
429
        self.add_attention = add_attention
Patrick von Platen's avatar
Patrick von Platen committed
430

431
432
        # there is always at least one resnet
        resnets = [
433
            ResnetBlock2D(
434
435
436
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
437
                eps=resnet_eps,
438
439
440
441
442
443
                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
444
            )
445
446
        ]
        attentions = []
Patrick von Platen's avatar
Patrick von Platen committed
447

448
449
450
451
452
453
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
            )
            attention_head_dim = in_channels

454
        for _ in range(num_layers):
Will Berman's avatar
Will Berman committed
455
456
            if self.add_attention:
                attentions.append(
457
                    Attention(
Will Berman's avatar
Will Berman committed
458
                        in_channels,
459
460
                        heads=in_channels // attention_head_dim,
                        dim_head=attention_head_dim,
Will Berman's avatar
Will Berman committed
461
462
                        rescale_output_factor=output_scale_factor,
                        eps=resnet_eps,
YiYi Xu's avatar
YiYi Xu committed
463
464
                        norm_num_groups=resnet_groups if resnet_time_scale_shift == "default" else None,
                        spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
465
466
467
468
                        residual_connection=True,
                        bias=True,
                        upcast_softmax=True,
                        _from_deprecated_attn_block=True,
Will Berman's avatar
Will Berman committed
469
                    )
470
                )
Will Berman's avatar
Will Berman committed
471
472
473
            else:
                attentions.append(None)

474
            resnets.append(
475
                ResnetBlock2D(
476
477
478
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
479
                    eps=resnet_eps,
480
481
482
483
484
485
486
                    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
487
488
            )

489
490
491
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

Will Berman's avatar
Will Berman committed
492
    def forward(self, hidden_states, temb=None):
Patrick von Platen's avatar
Patrick von Platen committed
493
        hidden_states = self.resnets[0](hidden_states, temb)
494
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
Will Berman's avatar
Will Berman committed
495
            if attn is not None:
YiYi Xu's avatar
YiYi Xu committed
496
                hidden_states = attn(hidden_states, temb=temb)
Patrick von Platen's avatar
Patrick von Platen committed
497
            hidden_states = resnet(hidden_states, temb)
Patrick von Platen's avatar
Patrick von Platen committed
498

499
        return hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
500

501

Patrick von Platen's avatar
Patrick von Platen committed
502
503
504
505
506
507
508
509
510
511
512
513
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,
514
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
515
516
        output_scale_factor=1.0,
        cross_attention_dim=1280,
517
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
518
        use_linear_projection=False,
519
        upcast_attention=False,
Patrick von Platen's avatar
Patrick von Platen committed
520
521
522
    ):
        super().__init__()

523
        self.has_cross_attention = True
524
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
525
526
527
528
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

        # there is always at least one resnet
        resnets = [
529
            ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
                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):
545
546
547
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
548
549
                        num_attention_heads,
                        in_channels // num_attention_heads,
550
551
552
553
                        in_channels=in_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
554
                        use_linear_projection=use_linear_projection,
555
                        upcast_attention=upcast_attention,
556
557
558
559
560
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
561
562
                        num_attention_heads,
                        in_channels // num_attention_heads,
563
564
565
566
567
                        in_channels=in_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
Patrick von Platen's avatar
Patrick von Platen committed
568
569
                )
            resnets.append(
570
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
                    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)

587
    def forward(
588
589
590
591
592
593
594
595
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
    ) -> torch.FloatTensor:
Patrick von Platen's avatar
Patrick von Platen committed
596
597
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
598
599
600
601
            hidden_states = attn(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                cross_attention_kwargs=cross_attention_kwargs,
602
603
                attention_mask=attention_mask,
                encoder_attention_mask=encoder_attention_mask,
604
605
                return_dict=False,
            )[0]
Will Berman's avatar
Will Berman committed
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
            hidden_states = resnet(hidden_states, temb)

        return hidden_states


class UNetMidBlock2DSimpleCrossAttn(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,
623
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
624
625
        output_scale_factor=1.0,
        cross_attention_dim=1280,
626
        skip_time_act=False,
627
        only_cross_attention=False,
628
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
629
630
631
632
633
    ):
        super().__init__()

        self.has_cross_attention = True

634
        self.attention_head_dim = attention_head_dim
Will Berman's avatar
Will Berman committed
635
636
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

637
        self.num_heads = in_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
638
639
640
641
642
643
644
645
646
647
648
649
650
651

        # there is always at least one resnet
        resnets = [
            ResnetBlock2D(
                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,
652
                skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
653
654
655
656
657
            )
        ]
        attentions = []

        for _ in range(num_layers):
658
659
660
661
            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

Will Berman's avatar
Will Berman committed
662
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
663
                Attention(
Will Berman's avatar
Will Berman committed
664
665
666
                    query_dim=in_channels,
                    cross_attention_dim=in_channels,
                    heads=self.num_heads,
667
                    dim_head=self.attention_head_dim,
Will Berman's avatar
Will Berman committed
668
669
670
671
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
672
                    only_cross_attention=only_cross_attention,
673
                    cross_attention_norm=cross_attention_norm,
674
                    processor=processor,
Will Berman's avatar
Will Berman committed
675
676
677
678
679
680
681
682
683
684
685
686
687
688
                )
            )
            resnets.append(
                ResnetBlock2D(
                    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,
689
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
690
691
692
693
694
695
                )
            )

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

696
    def forward(
697
698
699
700
701
702
703
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
704
705
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
706
707
708
709
710
711
712
713
714
715
716
717

        if attention_mask is None:
            # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
            mask = None if encoder_hidden_states is None else encoder_attention_mask
        else:
            # when attention_mask is defined: we don't even check for encoder_attention_mask.
            # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
            # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
            #       then we can simplify this whole if/else block to:
            #         mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
            mask = attention_mask

Will Berman's avatar
Will Berman committed
718
719
720
721
722
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            # attn
            hidden_states = attn(
                hidden_states,
723
                encoder_hidden_states=encoder_hidden_states,
724
                attention_mask=mask,
725
                **cross_attention_kwargs,
Will Berman's avatar
Will Berman committed
726
727
728
            )

            # resnet
Patrick von Platen's avatar
Patrick von Platen committed
729
730
731
732
733
            hidden_states = resnet(hidden_states, temb)

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
734
class AttnDownBlock2D(nn.Module):
735
736
737
738
739
740
741
742
743
744
745
746
    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,
747
        attention_head_dim=1,
748
        output_scale_factor=1.0,
749
        downsample_padding=1,
750
751
752
753
754
755
        add_downsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

756
757
758
759
760
761
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

762
763
764
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
765
                ResnetBlock2D(
766
767
768
769
770
771
772
773
774
775
776
777
778
                    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(
779
                Attention(
780
                    out_channels,
781
782
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
783
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
784
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
785
                    norm_num_groups=resnet_groups,
786
787
788
789
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
790
791
792
793
794
795
796
797
                )
            )

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

        if add_downsample:
            self.downsamplers = nn.ModuleList(
798
799
                [
                    Downsample2D(
800
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
801
802
                    )
                ]
803
804
805
806
            )
        else:
            self.downsamplers = None

807
    def forward(self, hidden_states, temb=None, upsample_size=None):
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
        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
824
class CrossAttnDownBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
825
826
827
828
829
830
831
832
833
834
835
836
    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,
837
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
838
839
840
841
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        downsample_padding=1,
        add_downsample=True,
842
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
843
        use_linear_projection=False,
844
        only_cross_attention=False,
845
        upcast_attention=False,
Patrick von Platen's avatar
Patrick von Platen committed
846
847
848
849
850
    ):
        super().__init__()
        resnets = []
        attentions = []

851
        self.has_cross_attention = True
852
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
853
854
855
856

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
857
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
858
859
860
861
862
863
864
865
866
867
868
869
                    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,
                )
            )
870
871
872
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
873
874
                        num_attention_heads,
                        out_channels // num_attention_heads,
875
876
877
878
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
879
                        use_linear_projection=use_linear_projection,
880
                        only_cross_attention=only_cross_attention,
881
                        upcast_attention=upcast_attention,
882
883
884
885
886
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
887
888
                        num_attention_heads,
                        out_channels // num_attention_heads,
889
890
891
892
893
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
Patrick von Platen's avatar
Patrick von Platen committed
894
895
896
897
898
899
900
901
                )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
902
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
Patrick von Platen's avatar
Patrick von Platen committed
903
904
905
906
907
908
                    )
                ]
            )
        else:
            self.downsamplers = None

909
910
        self.gradient_checkpointing = False

911
    def forward(
912
913
914
915
916
917
918
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
919
    ):
Patrick von Platen's avatar
Patrick von Platen committed
920
921
922
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
923
924
            if self.training and self.gradient_checkpointing:

Will Berman's avatar
Will Berman committed
925
                def create_custom_forward(module, return_dict=None):
926
                    def custom_forward(*inputs):
Will Berman's avatar
Will Berman committed
927
928
929
930
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)
931
932
933

                    return custom_forward

934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    None,  # timestep
                    None,  # class_labels
                    cross_attention_kwargs,
                    attention_mask,
                    encoder_attention_mask,
                    **ckpt_kwargs,
                )[0]
952
953
            else:
                hidden_states = resnet(hidden_states, temb)
954
955
956
957
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
958
959
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
960
961
                    return_dict=False,
                )[0]
962

963
            output_states = output_states + (hidden_states,)
Patrick von Platen's avatar
Patrick von Platen committed
964
965
966
967
968

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

969
            output_states = output_states + (hidden_states,)
Patrick von Platen's avatar
Patrick von Platen committed
970
971
972
973

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
974
class DownBlock2D(nn.Module):
975
976
977
978
979
980
981
982
983
984
985
986
987
988
    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
989
        downsample_padding=1,
990
991
992
993
994
995
996
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
997
                ResnetBlock2D(
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
                    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
1015
1016
                [
                    Downsample2D(
1017
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
Patrick von Platen's avatar
Patrick von Platen committed
1018
1019
                    )
                ]
1020
1021
1022
1023
            )
        else:
            self.downsamplers = None

1024
1025
        self.gradient_checkpointing = False

1026
1027
1028
1029
    def forward(self, hidden_states, temb=None):
        output_states = ()

        for resnet in self.resnets:
1030
1031
1032
1033
1034
1035
1036
1037
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

1038
1039
1040
1041
1042
1043
1044
1045
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
1046
1047
1048
            else:
                hidden_states = resnet(hidden_states, temb)

1049
            output_states = output_states + (hidden_states,)
1050
1051
1052
1053
1054

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

1055
            output_states = output_states + (hidden_states,)
1056
1057
1058
1059

        return hidden_states, output_states


1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
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(
1082
                ResnetBlock2D(
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
                    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(
1102
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
                    )
                ]
            )
        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


1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
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,
1132
        attention_head_dim=1,
1133
1134
1135
1136
1137
1138
1139
1140
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []
        attentions = []

1141
1142
1143
1144
1145
1146
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

1147
1148
1149
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1150
                ResnetBlock2D(
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
                    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(
1164
                Attention(
1165
                    out_channels,
1166
1167
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1168
1169
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
1170
                    norm_num_groups=resnet_groups,
1171
1172
1173
1174
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
                )
            )

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

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
1185
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
                    )
                ]
            )
        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
1204
class AttnSkipDownBlock2D(nn.Module):
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
    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,
1216
        attention_head_dim=1,
1217
1218
1219
1220
1221
1222
1223
        output_scale_factor=np.sqrt(2.0),
        add_downsample=True,
    ):
        super().__init__()
        self.attentions = nn.ModuleList([])
        self.resnets = nn.ModuleList([])

1224
1225
1226
1227
1228
1229
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

1230
1231
1232
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            self.resnets.append(
1233
                ResnetBlock2D(
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
                    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(
1248
                Attention(
1249
                    out_channels,
1250
1251
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1252
1253
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
1254
1255
1256
1257
1258
                    norm_num_groups=32,
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1259
1260
1261
1262
                )
            )

        if add_downsample:
1263
            self.resnet_down = ResnetBlock2D(
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
                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,
1274
                use_in_shortcut=True,
1275
1276
1277
                down=True,
                kernel="fir",
            )
1278
            self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
            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
1305
class SkipDownBlock2D(nn.Module):
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
    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(
1327
                ResnetBlock2D(
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
                    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:
1343
            self.resnet_down = ResnetBlock2D(
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
                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,
1354
                use_in_shortcut=True,
1355
1356
1357
                down=True,
                kernel="fir",
            )
1358
            self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
            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


Will Berman's avatar
Will Berman committed
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
class ResnetDownsampleBlock2D(nn.Module):
    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,
1399
        skip_time_act=False,
Will Berman's avatar
Will Berman committed
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
                ResnetBlock2D(
                    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,
1418
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_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,
1438
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    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

1460
1461
1462
1463
1464
1465
1466
1467
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
Will Berman's avatar
Will Berman committed
1468
1469
1470
            else:
                hidden_states = resnet(hidden_states, temb)

1471
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1472
1473
1474
1475
1476

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

1477
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494

        return hidden_states, output_states


class SimpleCrossAttnDownBlock2D(nn.Module):
    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,
1495
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
1496
1497
1498
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_downsample=True,
1499
        skip_time_act=False,
1500
        only_cross_attention=False,
1501
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
1502
1503
1504
1505
1506
1507
1508
1509
    ):
        super().__init__()

        self.has_cross_attention = True

        resnets = []
        attentions = []

1510
1511
        self.attention_head_dim = attention_head_dim
        self.num_heads = out_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
                ResnetBlock2D(
                    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,
1527
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1528
1529
                )
            )
1530
1531
1532
1533
1534

            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

Will Berman's avatar
Will Berman committed
1535
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
1536
                Attention(
Will Berman's avatar
Will Berman committed
1537
1538
1539
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
1540
                    dim_head=attention_head_dim,
Will Berman's avatar
Will Berman committed
1541
1542
1543
1544
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
1545
                    only_cross_attention=only_cross_attention,
1546
                    cross_attention_norm=cross_attention_norm,
1547
                    processor=processor,
Will Berman's avatar
Will Berman committed
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
                )
            )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_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,
1567
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1568
1569
1570
1571
1572
1573
1574
1575
1576
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

1577
    def forward(
1578
1579
1580
1581
1582
1583
1584
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
1585
    ):
Will Berman's avatar
Will Berman committed
1586
        output_states = ()
1587
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
Will Berman's avatar
Will Berman committed
1588

1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
        if attention_mask is None:
            # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
            mask = None if encoder_hidden_states is None else encoder_attention_mask
        else:
            # when attention_mask is defined: we don't even check for encoder_attention_mask.
            # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
            # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
            #       then we can simplify this whole if/else block to:
            #         mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
            mask = attention_mask

Will Berman's avatar
Will Berman committed
1600
        for resnet, attn in zip(self.resnets, self.attentions):
1601
            if self.training and self.gradient_checkpointing:
Will Berman's avatar
Will Berman committed
1602

1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            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, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
1617
                    mask,
1618
1619
1620
1621
1622
1623
1624
1625
                    cross_attention_kwargs,
                )[0]
            else:
                hidden_states = resnet(hidden_states, temb)

                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
1626
                    attention_mask=mask,
1627
1628
                    **cross_attention_kwargs,
                )
Will Berman's avatar
Will Berman committed
1629

1630
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1631
1632
1633
1634
1635

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

1636
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1637
1638
1639
1640

        return hidden_states, output_states


1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
class KDownBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 4,
        resnet_eps: float = 1e-5,
        resnet_act_fn: str = "gelu",
        resnet_group_size: int = 32,
        add_downsample=False,
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            groups = in_channels // resnet_group_size
            groups_out = out_channels // resnet_group_size

            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    dropout=dropout,
                    temb_channels=temb_channels,
                    groups=groups,
                    groups_out=groups_out,
                    eps=resnet_eps,
                    non_linearity=resnet_act_fn,
                    time_embedding_norm="ada_group",
                    conv_shortcut_bias=False,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            # YiYi's comments- might be able to use FirDownsample2D, look into details later
            self.downsamplers = nn.ModuleList([KDownsample2D()])
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    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

1699
1700
1701
1702
1703
1704
1705
1706
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
            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)

        return hidden_states, output_states


class KCrossAttnDownBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        cross_attention_dim: int,
        dropout: float = 0.0,
        num_layers: int = 4,
        resnet_group_size: int = 32,
        add_downsample=True,
1730
        attention_head_dim: int = 64,
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
        add_self_attention: bool = False,
        resnet_eps: float = 1e-5,
        resnet_act_fn: str = "gelu",
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            groups = in_channels // resnet_group_size
            groups_out = out_channels // resnet_group_size

            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    dropout=dropout,
                    temb_channels=temb_channels,
                    groups=groups,
                    groups_out=groups_out,
                    eps=resnet_eps,
                    non_linearity=resnet_act_fn,
                    time_embedding_norm="ada_group",
                    conv_shortcut_bias=False,
                )
            )
            attentions.append(
                KAttentionBlock(
                    out_channels,
1763
1764
                    out_channels // attention_head_dim,
                    attention_head_dim,
1765
1766
1767
1768
                    cross_attention_dim=cross_attention_dim,
                    temb_channels=temb_channels,
                    attention_bias=True,
                    add_self_attention=add_self_attention,
1769
                    cross_attention_norm="layer_norm",
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
                    group_size=resnet_group_size,
                )
            )

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

        if add_downsample:
            self.downsamplers = nn.ModuleList([KDownsample2D()])
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    def forward(
1785
1786
1787
1788
1789
1790
1791
        self,
        hidden_states: torch.FloatTensor,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
    ):
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    temb,
                    attention_mask,
                    cross_attention_kwargs,
                    encoder_attention_mask,
                    **ckpt_kwargs,
                )
1824
1825
1826
1827
1828
1829
1830
1831
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
1832
                    encoder_attention_mask=encoder_attention_mask,
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
                )

            if self.downsamplers is None:
                output_states += (None,)
            else:
                output_states += (hidden_states,)

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

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
1847
class AttnUpBlock2D(nn.Module):
1848
1849
1850
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
1851
1852
        prev_output_channel: int,
        out_channels: int,
1853
1854
1855
1856
1857
1858
1859
1860
        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,
1861
        attention_head_dim=1,
1862
1863
1864
1865
1866
1867
1868
        output_scale_factor=1.0,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []
        attentions = []

1869
1870
1871
1872
1873
1874
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

1875
        for i in range(num_layers):
Patrick von Platen's avatar
Patrick von Platen committed
1876
1877
1878
            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

1879
            resnets.append(
1880
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
1881
1882
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
                    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(
1894
                Attention(
Patrick von Platen's avatar
Patrick von Platen committed
1895
                    out_channels,
1896
1897
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1898
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
1899
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
1900
                    norm_num_groups=resnet_groups,
1901
1902
1903
1904
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1905
1906
1907
1908
1909
1910
1911
                )
            )

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

        if add_upsample:
Patrick von Platen's avatar
Patrick von Platen committed
1912
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
1913
1914
1915
        else:
            self.upsamplers = None

1916
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
        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
1933
class CrossAttnUpBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
    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,
1947
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
1948
1949
1950
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
1951
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
1952
        use_linear_projection=False,
1953
        only_cross_attention=False,
1954
        upcast_attention=False,
Patrick von Platen's avatar
Patrick von Platen committed
1955
1956
1957
1958
1959
    ):
        super().__init__()
        resnets = []
        attentions = []

1960
        self.has_cross_attention = True
1961
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
1962
1963
1964
1965
1966
1967

        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(
1968
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
                    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,
                )
            )
1981
1982
1983
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
1984
1985
                        num_attention_heads,
                        out_channels // num_attention_heads,
1986
1987
1988
1989
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
1990
                        use_linear_projection=use_linear_projection,
1991
                        only_cross_attention=only_cross_attention,
1992
                        upcast_attention=upcast_attention,
1993
1994
1995
1996
1997
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
1998
1999
                        num_attention_heads,
                        out_channels // num_attention_heads,
2000
2001
2002
2003
2004
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
Patrick von Platen's avatar
Patrick von Platen committed
2005
2006
2007
2008
2009
2010
2011
2012
2013
                )
        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

2014
2015
2016
2017
        self.gradient_checkpointing = False

    def forward(
        self,
2018
2019
2020
2021
2022
2023
2024
2025
        hidden_states: torch.FloatTensor,
        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        upsample_size: Optional[int] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
2026
    ):
Patrick von Platen's avatar
Patrick von Platen committed
2027
2028
2029
2030
2031
2032
        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)

2033
2034
            if self.training and self.gradient_checkpointing:

Will Berman's avatar
Will Berman committed
2035
                def create_custom_forward(module, return_dict=None):
2036
                    def custom_forward(*inputs):
Will Berman's avatar
Will Berman committed
2037
2038
2039
2040
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)
2041
2042
2043

                    return custom_forward

2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    None,  # timestep
                    None,  # class_labels
                    cross_attention_kwargs,
                    attention_mask,
                    encoder_attention_mask,
                    **ckpt_kwargs,
                )[0]
2062
2063
            else:
                hidden_states = resnet(hidden_states, temb)
2064
2065
2066
2067
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
2068
2069
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
2070
2071
                    return_dict=False,
                )[0]
Patrick von Platen's avatar
Patrick von Platen committed
2072
2073
2074

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2075
                hidden_states = upsampler(hidden_states, upsample_size)
Patrick von Platen's avatar
Patrick von Platen committed
2076
2077
2078
2079

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2080
class UpBlock2D(nn.Module):
2081
2082
2083
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
2084
2085
        prev_output_channel: int,
        out_channels: int,
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
        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
2101
2102
2103
            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

2104
            resnets.append(
2105
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2106
2107
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
                    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
2122
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
2123
2124
2125
        else:
            self.upsamplers = None

2126
2127
        self.gradient_checkpointing = False

2128
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
2129
2130
2131
2132
2133
2134
        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)

2135
2136
2137
2138
2139
2140
2141
2142
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

2143
2144
2145
2146
2147
2148
2149
2150
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
2151
2152
            else:
                hidden_states = resnet(hidden_states, temb)
2153
2154
2155

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2156
                hidden_states = upsampler(hidden_states, upsample_size)
2157
2158

        return hidden_states
2159
2160


2161
2162
2163
2164
2165
2166
2167
2168
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,
YiYi Xu's avatar
YiYi Xu committed
2169
        resnet_time_scale_shift: str = "default",  # default, spatial
2170
2171
2172
2173
2174
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_upsample=True,
YiYi Xu's avatar
YiYi Xu committed
2175
        temb_channels=None,
2176
2177
2178
2179
2180
2181
2182
2183
    ):
        super().__init__()
        resnets = []

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

            resnets.append(
2184
                ResnetBlock2D(
2185
2186
                    in_channels=input_channels,
                    out_channels=out_channels,
YiYi Xu's avatar
YiYi Xu committed
2187
                    temb_channels=temb_channels,
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
                    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

YiYi Xu's avatar
YiYi Xu committed
2205
    def forward(self, hidden_states, temb=None):
2206
        for resnet in self.resnets:
YiYi Xu's avatar
YiYi Xu committed
2207
            hidden_states = resnet(hidden_states, temb=temb)
2208
2209
2210
2211
2212
2213
2214
2215

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

        return hidden_states


2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
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,
2228
        attention_head_dim=1,
2229
2230
        output_scale_factor=1.0,
        add_upsample=True,
YiYi Xu's avatar
YiYi Xu committed
2231
        temb_channels=None,
2232
2233
2234
2235
2236
    ):
        super().__init__()
        resnets = []
        attentions = []

2237
2238
2239
2240
2241
2242
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

2243
2244
2245
2246
        for i in range(num_layers):
            input_channels = in_channels if i == 0 else out_channels

            resnets.append(
2247
                ResnetBlock2D(
2248
2249
                    in_channels=input_channels,
                    out_channels=out_channels,
YiYi Xu's avatar
YiYi Xu committed
2250
                    temb_channels=temb_channels,
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
                    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(
2261
                Attention(
2262
                    out_channels,
2263
2264
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
2265
2266
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
2267
                    norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None,
YiYi Xu's avatar
YiYi Xu committed
2268
                    spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
2269
2270
2271
2272
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
                )
            )

        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

YiYi Xu's avatar
YiYi Xu committed
2284
    def forward(self, hidden_states, temb=None):
2285
        for resnet, attn in zip(self.resnets, self.attentions):
YiYi Xu's avatar
YiYi Xu committed
2286
2287
            hidden_states = resnet(hidden_states, temb=temb)
            hidden_states = attn(hidden_states, temb=temb)
2288
2289
2290
2291
2292
2293
2294
2295

        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
2296
class AttnSkipUpBlock2D(nn.Module):
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
    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,
2309
        attention_head_dim=1,
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
        output_scale_factor=np.sqrt(2.0),
        add_upsample=True,
    ):
        super().__init__()
        self.attentions = nn.ModuleList([])
        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(
2322
                ResnetBlock2D(
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
                    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,
                )
            )

2337
2338
2339
2340
2341
2342
        if attention_head_dim is None:
            logger.warn(
                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
            )
            attention_head_dim = out_channels

2343
        self.attentions.append(
2344
            Attention(
2345
                out_channels,
2346
2347
                heads=out_channels // attention_head_dim,
                dim_head=attention_head_dim,
2348
2349
                rescale_output_factor=output_scale_factor,
                eps=resnet_eps,
2350
2351
2352
2353
2354
                norm_num_groups=32,
                residual_connection=True,
                bias=True,
                upcast_softmax=True,
                _from_deprecated_attn_block=True,
2355
2356
2357
2358
2359
            )
        )

        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
        if add_upsample:
2360
            self.resnet_up = ResnetBlock2D(
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
                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,
2372
                use_in_shortcut=True,
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
                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
2415
class SkipUpBlock2D(nn.Module):
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
    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(
2440
                ResnetBlock2D(
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
                    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:
2457
            self.resnet_up = ResnetBlock2D(
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
                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,
2469
                use_in_shortcut=True,
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
                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
Will Berman's avatar
Will Berman committed
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525


class ResnetUpsampleBlock2D(nn.Module):
    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_groups: int = 32,
        resnet_pre_norm: bool = True,
        output_scale_factor=1.0,
        add_upsample=True,
2526
        skip_time_act=False,
Will Berman's avatar
Will Berman committed
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
    ):
        super().__init__()
        resnets = []

        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(
                ResnetBlock2D(
                    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,
2547
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_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,
2567
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=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)

            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs)

                    return custom_forward

2592
2593
2594
2595
2596
2597
2598
2599
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
Will Berman's avatar
Will Berman committed
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
            else:
                hidden_states = resnet(hidden_states, temb)

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

        return hidden_states


class SimpleCrossAttnUpBlock2D(nn.Module):
    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,
2624
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
2625
2626
2627
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
2628
        skip_time_act=False,
2629
        only_cross_attention=False,
2630
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
2631
2632
2633
2634
2635
2636
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
2637
        self.attention_head_dim = attention_head_dim
Will Berman's avatar
Will Berman committed
2638

2639
        self.num_heads = out_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656

        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(
                ResnetBlock2D(
                    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,
2657
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2658
2659
                )
            )
2660
2661
2662
2663
2664

            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

Will Berman's avatar
Will Berman committed
2665
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
2666
                Attention(
Will Berman's avatar
Will Berman committed
2667
2668
2669
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
2670
                    dim_head=self.attention_head_dim,
Will Berman's avatar
Will Berman committed
2671
2672
2673
2674
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
2675
                    only_cross_attention=only_cross_attention,
2676
                    cross_attention_norm=cross_attention_norm,
2677
                    processor=processor,
Will Berman's avatar
Will Berman committed
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
                )
            )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList(
                [
                    ResnetBlock2D(
                        in_channels=out_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,
2697
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
2709
2710
2711
2712
2713
2714
2715
2716
        hidden_states: torch.FloatTensor,
        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        upsample_size: Optional[int] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
Will Berman's avatar
Will Berman committed
2717
    ):
2718
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730

        if attention_mask is None:
            # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
            mask = None if encoder_hidden_states is None else encoder_attention_mask
        else:
            # when attention_mask is defined: we don't even check for encoder_attention_mask.
            # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
            # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
            #       then we can simplify this whole if/else block to:
            #         mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
            mask = attention_mask

Will Berman's avatar
Will Berman committed
2731
2732
2733
2734
2735
2736
2737
        for resnet, attn in zip(self.resnets, self.attentions):
            # resnet
            # 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)

2738
            if self.training and self.gradient_checkpointing:
Will Berman's avatar
Will Berman committed
2739

2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            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, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
2754
                    mask,
2755
2756
2757
2758
2759
2760
2761
2762
                    cross_attention_kwargs,
                )[0]
            else:
                hidden_states = resnet(hidden_states, temb)

                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
2763
                    attention_mask=mask,
2764
2765
                    **cross_attention_kwargs,
                )
Will Berman's avatar
Will Berman committed
2766
2767
2768
2769
2770
2771

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

        return hidden_states
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835


class KUpBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 5,
        resnet_eps: float = 1e-5,
        resnet_act_fn: str = "gelu",
        resnet_group_size: Optional[int] = 32,
        add_upsample=True,
    ):
        super().__init__()
        resnets = []
        k_in_channels = 2 * out_channels
        k_out_channels = in_channels
        num_layers = num_layers - 1

        for i in range(num_layers):
            in_channels = k_in_channels if i == 0 else out_channels
            groups = in_channels // resnet_group_size
            groups_out = out_channels // resnet_group_size

            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=k_out_channels if (i == num_layers - 1) else out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=groups,
                    groups_out=groups_out,
                    dropout=dropout,
                    non_linearity=resnet_act_fn,
                    time_embedding_norm="ada_group",
                    conv_shortcut_bias=False,
                )
            )

        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([KUpsample2D()])
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
        res_hidden_states_tuple = res_hidden_states_tuple[-1]
        if res_hidden_states_tuple is not None:
            hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)

        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

2836
2837
2838
2839
2840
2841
2842
2843
                if is_torch_version(">=", "1.11.0"):
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
                    )
                else:
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(resnet), hidden_states, temb
                    )
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
            else:
                hidden_states = resnet(hidden_states, temb)

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

        return hidden_states


class KCrossAttnUpBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 4,
        resnet_eps: float = 1e-5,
        resnet_act_fn: str = "gelu",
        resnet_group_size: int = 32,
2865
        attention_head_dim=1,  # attention dim_head
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
        cross_attention_dim: int = 768,
        add_upsample: bool = True,
        upcast_attention: bool = False,
    ):
        super().__init__()
        resnets = []
        attentions = []

        is_first_block = in_channels == out_channels == temb_channels
        is_middle_block = in_channels != out_channels
        add_self_attention = True if is_first_block else False

        self.has_cross_attention = True
2879
        self.attention_head_dim = attention_head_dim
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914

        # in_channels, and out_channels for the block (k-unet)
        k_in_channels = out_channels if is_first_block else 2 * out_channels
        k_out_channels = in_channels

        num_layers = num_layers - 1

        for i in range(num_layers):
            in_channels = k_in_channels if i == 0 else out_channels
            groups = in_channels // resnet_group_size
            groups_out = out_channels // resnet_group_size

            if is_middle_block and (i == num_layers - 1):
                conv_2d_out_channels = k_out_channels
            else:
                conv_2d_out_channels = None

            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    conv_2d_out_channels=conv_2d_out_channels,
                    temb_channels=temb_channels,
                    eps=resnet_eps,
                    groups=groups,
                    groups_out=groups_out,
                    dropout=dropout,
                    non_linearity=resnet_act_fn,
                    time_embedding_norm="ada_group",
                    conv_shortcut_bias=False,
                )
            )
            attentions.append(
                KAttentionBlock(
                    k_out_channels if (i == num_layers - 1) else out_channels,
2915
                    k_out_channels // attention_head_dim
2916
                    if (i == num_layers - 1)
2917
2918
                    else out_channels // attention_head_dim,
                    attention_head_dim,
2919
2920
2921
2922
                    cross_attention_dim=cross_attention_dim,
                    temb_channels=temb_channels,
                    attention_bias=True,
                    add_self_attention=add_self_attention,
2923
                    cross_attention_norm="layer_norm",
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
                    upcast_attention=upcast_attention,
                )
            )

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

        if add_upsample:
            self.upsamplers = nn.ModuleList([KUpsample2D()])
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
2940
2941
2942
2943
2944
2945
2946
2947
        hidden_states: torch.FloatTensor,
        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        upsample_size: Optional[int] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
    ):
        res_hidden_states_tuple = res_hidden_states_tuple[-1]
        if res_hidden_states_tuple is not None:
            hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)

        for resnet, attn in zip(self.resnets, self.attentions):
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    temb,
                    attention_mask,
                    cross_attention_kwargs,
                    encoder_attention_mask,
                    **ckpt_kwargs,
                )[0]
2982
2983
2984
2985
2986
2987
2988
2989
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
2990
                    encoder_attention_mask=encoder_attention_mask,
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
                )

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

        return hidden_states


# can potentially later be renamed to `No-feed-forward` attention
class KAttentionBlock(nn.Module):
    r"""
    A basic Transformer block.

    Parameters:
        dim (`int`): The number of channels in the input and output.
        num_attention_heads (`int`): The number of heads to use for multi-head attention.
        attention_head_dim (`int`): The number of channels in each head.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
        num_embeds_ada_norm (:
            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
        attention_bias (:
            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
    """

    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        dropout: float = 0.0,
        cross_attention_dim: Optional[int] = None,
        attention_bias: bool = False,
        upcast_attention: bool = False,
        temb_channels: int = 768,  # for ada_group_norm
        add_self_attention: bool = False,
3029
        cross_attention_norm: Optional[str] = None,
3030
3031
3032
3033
3034
3035
3036
3037
        group_size: int = 32,
    ):
        super().__init__()
        self.add_self_attention = add_self_attention

        # 1. Self-Attn
        if add_self_attention:
            self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
Patrick von Platen's avatar
Patrick von Platen committed
3038
            self.attn1 = Attention(
3039
3040
3041
3042
3043
3044
                query_dim=dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                cross_attention_dim=None,
3045
                cross_attention_norm=None,
3046
3047
3048
3049
            )

        # 2. Cross-Attn
        self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
Patrick von Platen's avatar
Patrick von Platen committed
3050
        self.attn2 = Attention(
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
            query_dim=dim,
            cross_attention_dim=cross_attention_dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
            upcast_attention=upcast_attention,
            cross_attention_norm=cross_attention_norm,
        )

    def _to_3d(self, hidden_states, height, weight):
        return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1)

    def _to_4d(self, hidden_states, height, weight):
        return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight)

    def forward(
        self,
3069
3070
3071
3072
3073
3074
3075
3076
        hidden_states: torch.FloatTensor,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        # TODO: mark emb as non-optional (self.norm2 requires it).
        #       requires assessing impact of change to positional param interface.
        emb: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}

        # 1. Self-Attention
        if self.add_self_attention:
            norm_hidden_states = self.norm1(hidden_states, emb)

            height, weight = norm_hidden_states.shape[2:]
            norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)

            attn_output = self.attn1(
                norm_hidden_states,
                encoder_hidden_states=None,
3090
                attention_mask=attention_mask,
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
                **cross_attention_kwargs,
            )
            attn_output = self._to_4d(attn_output, height, weight)

            hidden_states = attn_output + hidden_states

        # 2. Cross-Attention/None
        norm_hidden_states = self.norm2(hidden_states, emb)

        height, weight = norm_hidden_states.shape[2:]
        norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)
        attn_output = self.attn2(
            norm_hidden_states,
            encoder_hidden_states=encoder_hidden_states,
3105
            attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask,
3106
3107
3108
3109
3110
3111
3112
            **cross_attention_kwargs,
        )
        attn_output = self._to_4d(attn_output, height, weight)

        hidden_states = attn_output + hidden_states

        return hidden_states