unet_2d_blocks.py 117 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
    transformer_layers_per_block=1,
42
    num_attention_heads=None,
43
    resnet_groups=None,
44
    cross_attention_dim=None,
Patrick von Platen's avatar
Patrick von Platen committed
45
    downsample_padding=None,
46
    dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
47
    use_linear_projection=False,
48
    only_cross_attention=False,
49
    upcast_attention=False,
Will Berman's avatar
Will Berman committed
50
    resnet_time_scale_shift="default",
51
52
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
53
    cross_attention_norm=None,
54
    attention_head_dim=None,
55
    downsample_type=None,
56
):
57
58
59
60
61
62
63
    # 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
64
65
66
    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
    if down_block_type == "DownBlock2D":
        return DownBlock2D(
67
68
69
70
71
72
73
            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,
74
            resnet_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
75
            downsample_padding=downsample_padding,
Will Berman's avatar
Will Berman committed
76
77
78
79
80
81
82
83
84
85
86
87
88
            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,
89
90
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
91
        )
Patrick von Platen's avatar
Patrick von Platen committed
92
    elif down_block_type == "AttnDownBlock2D":
93
94
95
96
        if add_downsample is False:
            downsample_type = None
        else:
            downsample_type = downsample_type or "conv"  # default to 'conv'
Patrick von Platen's avatar
Patrick von Platen committed
97
        return AttnDownBlock2D(
98
99
100
101
102
103
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
104
            resnet_groups=resnet_groups,
105
            downsample_padding=downsample_padding,
106
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
107
            resnet_time_scale_shift=resnet_time_scale_shift,
108
            downsample_type=downsample_type,
109
        )
Patrick von Platen's avatar
Patrick von Platen committed
110
    elif down_block_type == "CrossAttnDownBlock2D":
111
        if cross_attention_dim is None:
112
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
Patrick von Platen's avatar
Patrick von Platen committed
113
        return CrossAttnDownBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
114
            num_layers=num_layers,
115
            transformer_layers_per_block=transformer_layers_per_block,
Patrick von Platen's avatar
Patrick von Platen committed
116
117
118
119
120
121
            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,
122
            resnet_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
123
            downsample_padding=downsample_padding,
124
            cross_attention_dim=cross_attention_dim,
125
            num_attention_heads=num_attention_heads,
126
            dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
127
            use_linear_projection=use_linear_projection,
128
            only_cross_attention=only_cross_attention,
129
            upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
            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,
145
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
146
            resnet_time_scale_shift=resnet_time_scale_shift,
147
148
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
149
            only_cross_attention=only_cross_attention,
150
            cross_attention_norm=cross_attention_norm,
Patrick von Platen's avatar
Patrick von Platen committed
151
        )
Patrick von Platen's avatar
Patrick von Platen committed
152
153
    elif down_block_type == "SkipDownBlock2D":
        return SkipDownBlock2D(
154
155
156
157
158
159
160
161
            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
162
            resnet_time_scale_shift=resnet_time_scale_shift,
163
        )
Patrick von Platen's avatar
Patrick von Platen committed
164
165
    elif down_block_type == "AttnSkipDownBlock2D":
        return AttnSkipDownBlock2D(
166
167
168
169
170
171
172
            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,
173
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
174
            resnet_time_scale_shift=resnet_time_scale_shift,
175
        )
176
177
178
179
180
181
182
183
    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,
184
            resnet_groups=resnet_groups,
185
            downsample_padding=downsample_padding,
Will Berman's avatar
Will Berman committed
186
            resnet_time_scale_shift=resnet_time_scale_shift,
187
        )
Will Berman's avatar
Will Berman committed
188
189
190
191
192
193
194
195
196
197
    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,
198
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
199
            resnet_time_scale_shift=resnet_time_scale_shift,
Will Berman's avatar
Will Berman committed
200
        )
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
    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,
221
            attention_head_dim=attention_head_dim,
222
223
            add_self_attention=True if not add_downsample else False,
        )
Will Berman's avatar
Will Berman committed
224
    raise ValueError(f"{down_block_type} does not exist.")
225
226
227
228
229
230


def get_up_block(
    up_block_type,
    num_layers,
    in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
231
232
    out_channels,
    prev_output_channel,
233
234
235
236
    temb_channels,
    add_upsample,
    resnet_eps,
    resnet_act_fn,
237
    transformer_layers_per_block=1,
238
    num_attention_heads=None,
239
    resnet_groups=None,
240
    cross_attention_dim=None,
241
    dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
242
    use_linear_projection=False,
243
    only_cross_attention=False,
244
    upcast_attention=False,
Will Berman's avatar
Will Berman committed
245
    resnet_time_scale_shift="default",
246
247
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
248
    cross_attention_norm=None,
249
    attention_head_dim=None,
250
    upsample_type=None,
251
):
252
253
254
255
256
257
258
    # 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
259
260
261
    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
    if up_block_type == "UpBlock2D":
        return UpBlock2D(
262
263
            num_layers=num_layers,
            in_channels=in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
264
265
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
266
267
268
269
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
270
            resnet_groups=resnet_groups,
Will Berman's avatar
Will Berman committed
271
272
273
274
275
276
277
278
279
280
281
282
283
284
            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,
285
286
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
287
        )
Patrick von Platen's avatar
Patrick von Platen committed
288
    elif up_block_type == "CrossAttnUpBlock2D":
289
290
        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
291
        return CrossAttnUpBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
292
            num_layers=num_layers,
293
            transformer_layers_per_block=transformer_layers_per_block,
Patrick von Platen's avatar
Patrick von Platen committed
294
295
296
297
298
299
300
            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,
301
            resnet_groups=resnet_groups,
302
            cross_attention_dim=cross_attention_dim,
303
            num_attention_heads=num_attention_heads,
304
            dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
305
            use_linear_projection=use_linear_projection,
306
            only_cross_attention=only_cross_attention,
307
            upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
            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,
324
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
325
            resnet_time_scale_shift=resnet_time_scale_shift,
326
327
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
328
            only_cross_attention=only_cross_attention,
329
            cross_attention_norm=cross_attention_norm,
Patrick von Platen's avatar
Patrick von Platen committed
330
        )
Patrick von Platen's avatar
Patrick von Platen committed
331
    elif up_block_type == "AttnUpBlock2D":
332
333
334
335
336
        if add_upsample is False:
            upsample_type = None
        else:
            upsample_type = upsample_type or "conv"  # default to 'conv'

Patrick von Platen's avatar
Patrick von Platen committed
337
        return AttnUpBlock2D(
338
339
            num_layers=num_layers,
            in_channels=in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
340
341
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
342
343
344
            temb_channels=temb_channels,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
345
            resnet_groups=resnet_groups,
346
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
347
            resnet_time_scale_shift=resnet_time_scale_shift,
348
            upsample_type=upsample_type,
349
        )
Patrick von Platen's avatar
Patrick von Platen committed
350
351
    elif up_block_type == "SkipUpBlock2D":
        return SkipUpBlock2D(
352
353
354
355
356
357
358
359
            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
360
            resnet_time_scale_shift=resnet_time_scale_shift,
361
        )
Patrick von Platen's avatar
Patrick von Platen committed
362
363
    elif up_block_type == "AttnSkipUpBlock2D":
        return AttnSkipUpBlock2D(
364
365
366
367
368
369
370
371
            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,
372
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
373
            resnet_time_scale_shift=resnet_time_scale_shift,
374
        )
375
376
377
378
379
380
381
382
    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,
383
            resnet_groups=resnet_groups,
Will Berman's avatar
Will Berman committed
384
            resnet_time_scale_shift=resnet_time_scale_shift,
YiYi Xu's avatar
YiYi Xu committed
385
            temb_channels=temb_channels,
386
        )
Will Berman's avatar
Will Berman committed
387
388
389
390
391
392
393
394
395
    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,
396
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
397
            resnet_time_scale_shift=resnet_time_scale_shift,
YiYi Xu's avatar
YiYi Xu committed
398
            temb_channels=temb_channels,
Will Berman's avatar
Will Berman committed
399
        )
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
    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,
420
            attention_head_dim=attention_head_dim,
421
422
        )

423
    raise ValueError(f"{up_block_type} does not exist.")
424
425


Patrick von Platen's avatar
Patrick von Platen committed
426
427
428
429
430
class UNetMidBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
431
        dropout: float = 0.0,
432
        num_layers: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
433
        resnet_eps: float = 1e-6,
YiYi Xu's avatar
YiYi Xu committed
434
        resnet_time_scale_shift: str = "default",  # default, spatial
Patrick von Platen's avatar
Patrick von Platen committed
435
436
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
437
        resnet_pre_norm: bool = True,
Will Berman's avatar
Will Berman committed
438
        add_attention: bool = True,
439
        attention_head_dim=1,
Patrick von Platen's avatar
Patrick von Platen committed
440
441
442
        output_scale_factor=1.0,
    ):
        super().__init__()
443
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
Will Berman's avatar
Will Berman committed
444
        self.add_attention = add_attention
Patrick von Platen's avatar
Patrick von Platen committed
445

446
447
        # there is always at least one resnet
        resnets = [
448
            ResnetBlock2D(
449
450
451
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
452
                eps=resnet_eps,
453
454
455
456
457
458
                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
459
            )
460
461
        ]
        attentions = []
Patrick von Platen's avatar
Patrick von Platen committed
462

463
464
465
466
467
468
        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

469
        for _ in range(num_layers):
Will Berman's avatar
Will Berman committed
470
471
            if self.add_attention:
                attentions.append(
472
                    Attention(
Will Berman's avatar
Will Berman committed
473
                        in_channels,
474
475
                        heads=in_channels // attention_head_dim,
                        dim_head=attention_head_dim,
Will Berman's avatar
Will Berman committed
476
477
                        rescale_output_factor=output_scale_factor,
                        eps=resnet_eps,
YiYi Xu's avatar
YiYi Xu committed
478
479
                        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,
480
481
482
483
                        residual_connection=True,
                        bias=True,
                        upcast_softmax=True,
                        _from_deprecated_attn_block=True,
Will Berman's avatar
Will Berman committed
484
                    )
485
                )
Will Berman's avatar
Will Berman committed
486
487
488
            else:
                attentions.append(None)

489
            resnets.append(
490
                ResnetBlock2D(
491
492
493
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
494
                    eps=resnet_eps,
495
496
497
498
499
500
501
                    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
502
503
            )

504
505
506
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

Will Berman's avatar
Will Berman committed
507
    def forward(self, hidden_states, temb=None):
Patrick von Platen's avatar
Patrick von Platen committed
508
        hidden_states = self.resnets[0](hidden_states, temb)
509
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
Will Berman's avatar
Will Berman committed
510
            if attn is not None:
YiYi Xu's avatar
YiYi Xu committed
511
                hidden_states = attn(hidden_states, temb=temb)
Patrick von Platen's avatar
Patrick von Platen committed
512
            hidden_states = resnet(hidden_states, temb)
Patrick von Platen's avatar
Patrick von Platen committed
513

514
        return hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
515

516

Patrick von Platen's avatar
Patrick von Platen committed
517
518
519
520
521
522
523
class UNetMidBlock2DCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
524
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
525
526
527
528
529
        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,
530
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
531
532
        output_scale_factor=1.0,
        cross_attention_dim=1280,
533
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
534
        use_linear_projection=False,
535
        upcast_attention=False,
Patrick von Platen's avatar
Patrick von Platen committed
536
537
538
    ):
        super().__init__()

539
        self.has_cross_attention = True
540
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
541
542
543
544
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

        # there is always at least one resnet
        resnets = [
545
            ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
                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):
561
562
563
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
564
565
                        num_attention_heads,
                        in_channels // num_attention_heads,
566
                        in_channels=in_channels,
567
                        num_layers=transformer_layers_per_block,
568
569
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
570
                        use_linear_projection=use_linear_projection,
571
                        upcast_attention=upcast_attention,
572
573
574
575
576
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
577
578
                        num_attention_heads,
                        in_channels // num_attention_heads,
579
580
581
582
583
                        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
584
585
                )
            resnets.append(
586
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
                    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)

603
    def forward(
604
605
606
607
608
609
610
611
        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
612
613
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
614
615
616
617
            hidden_states = attn(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                cross_attention_kwargs=cross_attention_kwargs,
618
619
                attention_mask=attention_mask,
                encoder_attention_mask=encoder_attention_mask,
620
621
                return_dict=False,
            )[0]
Will Berman's avatar
Will Berman committed
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
            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,
639
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
640
641
        output_scale_factor=1.0,
        cross_attention_dim=1280,
642
        skip_time_act=False,
643
        only_cross_attention=False,
644
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
645
646
647
648
649
    ):
        super().__init__()

        self.has_cross_attention = True

650
        self.attention_head_dim = attention_head_dim
Will Berman's avatar
Will Berman committed
651
652
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

653
        self.num_heads = in_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
654
655
656
657
658
659
660
661
662
663
664
665
666
667

        # 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,
668
                skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
669
670
671
672
673
            )
        ]
        attentions = []

        for _ in range(num_layers):
674
675
676
677
            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

Will Berman's avatar
Will Berman committed
678
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
679
                Attention(
Will Berman's avatar
Will Berman committed
680
681
682
                    query_dim=in_channels,
                    cross_attention_dim=in_channels,
                    heads=self.num_heads,
683
                    dim_head=self.attention_head_dim,
Will Berman's avatar
Will Berman committed
684
685
686
687
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
688
                    only_cross_attention=only_cross_attention,
689
                    cross_attention_norm=cross_attention_norm,
690
                    processor=processor,
Will Berman's avatar
Will Berman committed
691
692
693
694
695
696
697
698
699
700
701
702
703
704
                )
            )
            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,
705
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
706
707
708
709
710
711
                )
            )

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

712
    def forward(
713
714
715
716
717
718
719
        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,
720
721
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
722
723
724
725
726
727
728
729
730
731
732
733

        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
734
735
736
737
738
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            # attn
            hidden_states = attn(
                hidden_states,
739
                encoder_hidden_states=encoder_hidden_states,
740
                attention_mask=mask,
741
                **cross_attention_kwargs,
Will Berman's avatar
Will Berman committed
742
743
744
            )

            # resnet
Patrick von Platen's avatar
Patrick von Platen committed
745
746
747
748
749
            hidden_states = resnet(hidden_states, temb)

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
750
class AttnDownBlock2D(nn.Module):
751
752
753
754
755
756
757
758
759
760
761
762
    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,
763
        attention_head_dim=1,
764
        output_scale_factor=1.0,
765
        downsample_padding=1,
766
        downsample_type="conv",
767
768
769
770
    ):
        super().__init__()
        resnets = []
        attentions = []
771
        self.downsample_type = downsample_type
772

773
774
775
776
777
778
        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

779
780
781
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
782
                ResnetBlock2D(
783
784
785
786
787
788
789
790
791
792
793
794
795
                    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(
796
                Attention(
797
                    out_channels,
798
799
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
800
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
801
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
802
                    norm_num_groups=resnet_groups,
803
804
805
806
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
807
808
809
810
811
812
                )
            )

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

813
        if downsample_type == "conv":
814
            self.downsamplers = nn.ModuleList(
815
816
                [
                    Downsample2D(
817
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
818
819
                    )
                ]
820
            )
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
        elif downsample_type == "resnet":
            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,
                        down=True,
                    )
                ]
            )
839
840
841
        else:
            self.downsamplers = None

842
    def forward(self, hidden_states, temb=None, upsample_size=None):
843
844
845
846
847
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb)
            hidden_states = attn(hidden_states)
848
            output_states = output_states + (hidden_states,)
849
850
851

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
852
853
854
855
                if self.downsample_type == "resnet":
                    hidden_states = downsampler(hidden_states, temb=temb)
                else:
                    hidden_states = downsampler(hidden_states)
856
857
858
859
860
861

            output_states += (hidden_states,)

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
862
class CrossAttnDownBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
863
864
865
866
867
868
869
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
870
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
871
872
873
874
875
        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,
876
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
877
878
879
880
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        downsample_padding=1,
        add_downsample=True,
881
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
882
        use_linear_projection=False,
883
        only_cross_attention=False,
884
        upcast_attention=False,
Patrick von Platen's avatar
Patrick von Platen committed
885
886
887
888
889
    ):
        super().__init__()
        resnets = []
        attentions = []

890
        self.has_cross_attention = True
891
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
892
893
894
895

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
896
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
897
898
899
900
901
902
903
904
905
906
907
908
                    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,
                )
            )
909
910
911
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
912
913
                        num_attention_heads,
                        out_channels // num_attention_heads,
914
                        in_channels=out_channels,
915
                        num_layers=transformer_layers_per_block,
916
917
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
918
                        use_linear_projection=use_linear_projection,
919
                        only_cross_attention=only_cross_attention,
920
                        upcast_attention=upcast_attention,
921
922
923
924
925
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
926
927
                        num_attention_heads,
                        out_channels // num_attention_heads,
928
929
930
931
932
                        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
933
934
935
936
937
938
939
940
                )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
941
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
Patrick von Platen's avatar
Patrick von Platen committed
942
943
944
945
946
947
                    )
                ]
            )
        else:
            self.downsamplers = None

948
949
        self.gradient_checkpointing = False

950
    def forward(
951
952
953
954
955
956
957
        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,
Will Berman's avatar
Will Berman committed
958
        additional_residuals=None,
959
    ):
Patrick von Platen's avatar
Patrick von Platen committed
960
961
        output_states = ()

Will Berman's avatar
Will Berman committed
962
963
964
        blocks = list(zip(self.resnets, self.attentions))

        for i, (resnet, attn) in enumerate(blocks):
965
966
            if self.training and self.gradient_checkpointing:

Will Berman's avatar
Will Berman committed
967
                def create_custom_forward(module, return_dict=None):
968
                    def custom_forward(*inputs):
Will Berman's avatar
Will Berman committed
969
970
971
972
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)
973
974
975

                    return custom_forward

976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
                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]
994
995
            else:
                hidden_states = resnet(hidden_states, temb)
996
997
998
999
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
1000
1001
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
1002
1003
                    return_dict=False,
                )[0]
1004

Will Berman's avatar
Will Berman committed
1005
1006
1007
1008
            # apply additional residuals to the output of the last pair of resnet and attention blocks
            if i == len(blocks) - 1 and additional_residuals is not None:
                hidden_states = hidden_states + additional_residuals

1009
            output_states = output_states + (hidden_states,)
Patrick von Platen's avatar
Patrick von Platen committed
1010
1011
1012
1013
1014

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

1015
            output_states = output_states + (hidden_states,)
Patrick von Platen's avatar
Patrick von Platen committed
1016
1017
1018
1019

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
1020
class DownBlock2D(nn.Module):
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
    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
1035
        downsample_padding=1,
1036
1037
1038
1039
1040
1041
1042
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1043
                ResnetBlock2D(
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
                    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
1061
1062
                [
                    Downsample2D(
1063
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
Patrick von Platen's avatar
Patrick von Platen committed
1064
1065
                    )
                ]
1066
1067
1068
1069
            )
        else:
            self.downsamplers = None

1070
1071
        self.gradient_checkpointing = False

1072
1073
1074
1075
    def forward(self, hidden_states, temb=None):
        output_states = ()

        for resnet in self.resnets:
1076
1077
1078
1079
1080
1081
1082
1083
            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

1084
1085
1086
1087
1088
1089
1090
1091
                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
                    )
1092
1093
1094
            else:
                hidden_states = resnet(hidden_states, temb)

1095
            output_states = output_states + (hidden_states,)
1096
1097
1098
1099
1100

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

1101
            output_states = output_states + (hidden_states,)
1102
1103
1104
1105

        return hidden_states, output_states


1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
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(
1128
                ResnetBlock2D(
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
                    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(
1148
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
                    )
                ]
            )
        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


1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
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,
1178
        attention_head_dim=1,
1179
1180
1181
1182
1183
1184
1185
1186
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []
        attentions = []

1187
1188
1189
1190
1191
1192
        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

1193
1194
1195
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1196
                ResnetBlock2D(
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
                    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(
1210
                Attention(
1211
                    out_channels,
1212
1213
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1214
1215
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
1216
                    norm_num_groups=resnet_groups,
1217
1218
1219
1220
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
                )
            )

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

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
1231
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
                    )
                ]
            )
        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
1250
class AttnSkipDownBlock2D(nn.Module):
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
    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,
1262
        attention_head_dim=1,
1263
1264
1265
1266
1267
1268
1269
        output_scale_factor=np.sqrt(2.0),
        add_downsample=True,
    ):
        super().__init__()
        self.attentions = nn.ModuleList([])
        self.resnets = nn.ModuleList([])

1270
1271
1272
1273
1274
1275
        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

1276
1277
1278
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            self.resnets.append(
1279
                ResnetBlock2D(
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
                    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(
1294
                Attention(
1295
                    out_channels,
1296
1297
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1298
1299
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
1300
1301
1302
1303
1304
                    norm_num_groups=32,
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1305
1306
1307
1308
                )
            )

        if add_downsample:
1309
            self.resnet_down = ResnetBlock2D(
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
                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,
1320
                use_in_shortcut=True,
1321
1322
1323
                down=True,
                kernel="fir",
            )
1324
            self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
            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
1351
class SkipDownBlock2D(nn.Module):
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
    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(
1373
                ResnetBlock2D(
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
                    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:
1389
            self.resnet_down = ResnetBlock2D(
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
                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,
1400
                use_in_shortcut=True,
1401
1402
1403
                down=True,
                kernel="fir",
            )
1404
            self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
            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
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
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,
1445
        skip_time_act=False,
Will Berman's avatar
Will Berman committed
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
    ):
        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,
1464
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
                )
            )

        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,
1484
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
                        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

1506
1507
1508
1509
1510
1511
1512
1513
                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
1514
1515
1516
            else:
                hidden_states = resnet(hidden_states, temb)

1517
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1518
1519
1520
1521
1522

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

1523
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540

        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,
1541
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
1542
1543
1544
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_downsample=True,
1545
        skip_time_act=False,
1546
        only_cross_attention=False,
1547
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
1548
1549
1550
1551
1552
1553
1554
1555
    ):
        super().__init__()

        self.has_cross_attention = True

        resnets = []
        attentions = []

1556
1557
        self.attention_head_dim = attention_head_dim
        self.num_heads = out_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572

        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,
1573
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1574
1575
                )
            )
1576
1577
1578
1579
1580

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

Will Berman's avatar
Will Berman committed
1581
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
1582
                Attention(
Will Berman's avatar
Will Berman committed
1583
1584
1585
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
1586
                    dim_head=attention_head_dim,
Will Berman's avatar
Will Berman committed
1587
1588
1589
1590
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
1591
                    only_cross_attention=only_cross_attention,
1592
                    cross_attention_norm=cross_attention_norm,
1593
                    processor=processor,
Will Berman's avatar
Will Berman committed
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
                )
            )
        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,
1613
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1614
1615
1616
1617
1618
1619
1620
1621
1622
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

1623
    def forward(
1624
1625
1626
1627
1628
1629
1630
        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,
1631
    ):
Will Berman's avatar
Will Berman committed
1632
        output_states = ()
1633
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
Will Berman's avatar
Will Berman committed
1634

1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
        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
1646
        for resnet, attn in zip(self.resnets, self.attentions):
1647
            if self.training and self.gradient_checkpointing:
Will Berman's avatar
Will Berman committed
1648

1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
                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,
1663
                    mask,
1664
1665
1666
1667
1668
1669
1670
1671
                    cross_attention_kwargs,
                )[0]
            else:
                hidden_states = resnet(hidden_states, temb)

                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
1672
                    attention_mask=mask,
1673
1674
                    **cross_attention_kwargs,
                )
Will Berman's avatar
Will Berman committed
1675

1676
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1677
1678
1679
1680
1681

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

1682
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1683
1684
1685
1686

        return hidden_states, output_states


1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
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

1745
1746
1747
1748
1749
1750
1751
1752
                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
                    )
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
            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,
1776
        attention_head_dim: int = 64,
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
        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,
1809
1810
                    out_channels // attention_head_dim,
                    attention_head_dim,
1811
1812
1813
1814
                    cross_attention_dim=cross_attention_dim,
                    temb_channels=temb_channels,
                    attention_bias=True,
                    add_self_attention=add_self_attention,
1815
                    cross_attention_norm="layer_norm",
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
                    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(
1831
1832
1833
1834
1835
1836
1837
        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,
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
    ):
        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

1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
                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,
                )
1870
1871
1872
1873
1874
1875
1876
1877
            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,
1878
                    encoder_attention_mask=encoder_attention_mask,
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
                )

            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
1893
class AttnUpBlock2D(nn.Module):
1894
1895
1896
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
1897
1898
        prev_output_channel: int,
        out_channels: int,
1899
1900
1901
1902
1903
1904
1905
1906
        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,
1907
        attention_head_dim=1,
1908
        output_scale_factor=1.0,
1909
        upsample_type="conv",
1910
1911
1912
1913
1914
    ):
        super().__init__()
        resnets = []
        attentions = []

1915
1916
        self.upsample_type = upsample_type

1917
1918
1919
1920
1921
1922
        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

1923
        for i in range(num_layers):
Patrick von Platen's avatar
Patrick von Platen committed
1924
1925
1926
            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

1927
            resnets.append(
1928
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
1929
1930
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
                    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(
1942
                Attention(
Patrick von Platen's avatar
Patrick von Platen committed
1943
                    out_channels,
1944
1945
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1946
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
1947
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
1948
                    norm_num_groups=resnet_groups,
1949
1950
1951
1952
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1953
1954
1955
1956
1957
1958
                )
            )

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

1959
        if upsample_type == "conv":
Patrick von Platen's avatar
Patrick von Platen committed
1960
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
        elif upsample_type == "resnet":
            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,
                        up=True,
                    )
                ]
            )
1979
1980
1981
        else:
            self.upsamplers = None

1982
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
        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:
1994
1995
1996
1997
                if self.upsample_type == "resnet":
                    hidden_states = upsampler(hidden_states, temb=temb)
                else:
                    hidden_states = upsampler(hidden_states)
1998
1999
2000
2001

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2002
class CrossAttnUpBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
2003
2004
2005
2006
2007
2008
2009
2010
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
2011
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
2012
2013
2014
2015
2016
        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,
2017
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
2018
2019
2020
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
2021
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
2022
        use_linear_projection=False,
2023
        only_cross_attention=False,
2024
        upcast_attention=False,
Patrick von Platen's avatar
Patrick von Platen committed
2025
2026
2027
2028
2029
    ):
        super().__init__()
        resnets = []
        attentions = []

2030
        self.has_cross_attention = True
2031
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
2032
2033
2034
2035
2036
2037

        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(
2038
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
                    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,
                )
            )
2051
2052
2053
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
2054
2055
                        num_attention_heads,
                        out_channels // num_attention_heads,
2056
                        in_channels=out_channels,
2057
                        num_layers=transformer_layers_per_block,
2058
2059
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
2060
                        use_linear_projection=use_linear_projection,
2061
                        only_cross_attention=only_cross_attention,
2062
                        upcast_attention=upcast_attention,
2063
2064
2065
2066
2067
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
2068
2069
                        num_attention_heads,
                        out_channels // num_attention_heads,
2070
2071
2072
2073
2074
                        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
2075
2076
2077
2078
2079
2080
2081
2082
2083
                )
        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

2084
2085
2086
2087
        self.gradient_checkpointing = False

    def forward(
        self,
2088
2089
2090
2091
2092
2093
2094
2095
        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,
2096
    ):
Patrick von Platen's avatar
Patrick von Platen committed
2097
2098
2099
2100
2101
2102
        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)

2103
2104
            if self.training and self.gradient_checkpointing:

Will Berman's avatar
Will Berman committed
2105
                def create_custom_forward(module, return_dict=None):
2106
                    def custom_forward(*inputs):
Will Berman's avatar
Will Berman committed
2107
2108
2109
2110
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)
2111
2112
2113

                    return custom_forward

2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
                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]
2132
2133
            else:
                hidden_states = resnet(hidden_states, temb)
2134
2135
2136
2137
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
2138
2139
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
2140
2141
                    return_dict=False,
                )[0]
Patrick von Platen's avatar
Patrick von Platen committed
2142
2143
2144

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2145
                hidden_states = upsampler(hidden_states, upsample_size)
Patrick von Platen's avatar
Patrick von Platen committed
2146
2147
2148
2149

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2150
class UpBlock2D(nn.Module):
2151
2152
2153
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
2154
2155
        prev_output_channel: int,
        out_channels: int,
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
        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
2171
2172
2173
            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

2174
            resnets.append(
2175
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2176
2177
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
                    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
2192
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
2193
2194
2195
        else:
            self.upsamplers = None

2196
2197
        self.gradient_checkpointing = False

2198
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
2199
2200
2201
2202
2203
2204
        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)

2205
2206
2207
2208
2209
2210
2211
2212
            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

2213
2214
2215
2216
2217
2218
2219
2220
                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
                    )
2221
2222
            else:
                hidden_states = resnet(hidden_states, temb)
2223
2224
2225

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2226
                hidden_states = upsampler(hidden_states, upsample_size)
2227
2228

        return hidden_states
2229
2230


2231
2232
2233
2234
2235
2236
2237
2238
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
2239
        resnet_time_scale_shift: str = "default",  # default, spatial
2240
2241
2242
2243
2244
        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
2245
        temb_channels=None,
2246
2247
2248
2249
2250
2251
2252
2253
    ):
        super().__init__()
        resnets = []

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

            resnets.append(
2254
                ResnetBlock2D(
2255
2256
                    in_channels=input_channels,
                    out_channels=out_channels,
YiYi Xu's avatar
YiYi Xu committed
2257
                    temb_channels=temb_channels,
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
                    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
2275
    def forward(self, hidden_states, temb=None):
2276
        for resnet in self.resnets:
YiYi Xu's avatar
YiYi Xu committed
2277
            hidden_states = resnet(hidden_states, temb=temb)
2278
2279
2280
2281
2282
2283
2284
2285

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

        return hidden_states


2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
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,
2298
        attention_head_dim=1,
2299
2300
        output_scale_factor=1.0,
        add_upsample=True,
YiYi Xu's avatar
YiYi Xu committed
2301
        temb_channels=None,
2302
2303
2304
2305
2306
    ):
        super().__init__()
        resnets = []
        attentions = []

2307
2308
2309
2310
2311
2312
        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

2313
2314
2315
2316
        for i in range(num_layers):
            input_channels = in_channels if i == 0 else out_channels

            resnets.append(
2317
                ResnetBlock2D(
2318
2319
                    in_channels=input_channels,
                    out_channels=out_channels,
YiYi Xu's avatar
YiYi Xu committed
2320
                    temb_channels=temb_channels,
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
                    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(
2331
                Attention(
2332
                    out_channels,
2333
2334
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
2335
2336
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
2337
                    norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None,
YiYi Xu's avatar
YiYi Xu committed
2338
                    spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
2339
2340
2341
2342
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
                )
            )

        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
2354
    def forward(self, hidden_states, temb=None):
2355
        for resnet, attn in zip(self.resnets, self.attentions):
YiYi Xu's avatar
YiYi Xu committed
2356
2357
            hidden_states = resnet(hidden_states, temb=temb)
            hidden_states = attn(hidden_states, temb=temb)
2358
2359
2360
2361
2362
2363
2364
2365

        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
2366
class AttnSkipUpBlock2D(nn.Module):
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
    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,
2379
        attention_head_dim=1,
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
        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(
2392
                ResnetBlock2D(
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
                    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,
                )
            )

2407
2408
2409
2410
2411
2412
        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

2413
        self.attentions.append(
2414
            Attention(
2415
                out_channels,
2416
2417
                heads=out_channels // attention_head_dim,
                dim_head=attention_head_dim,
2418
2419
                rescale_output_factor=output_scale_factor,
                eps=resnet_eps,
2420
2421
2422
2423
2424
                norm_num_groups=32,
                residual_connection=True,
                bias=True,
                upcast_softmax=True,
                _from_deprecated_attn_block=True,
2425
2426
2427
2428
2429
            )
        )

        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
        if add_upsample:
2430
            self.resnet_up = ResnetBlock2D(
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
                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,
2442
                use_in_shortcut=True,
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
                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
2485
class SkipUpBlock2D(nn.Module):
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
    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(
2510
                ResnetBlock2D(
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
                    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:
2527
            self.resnet_up = ResnetBlock2D(
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
                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,
2539
                use_in_shortcut=True,
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
                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
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595


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,
2596
        skip_time_act=False,
Will Berman's avatar
Will Berman committed
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
    ):
        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,
2617
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
                )
            )

        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,
2637
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
                        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

2662
2663
2664
2665
2666
2667
2668
2669
                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
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
            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,
2694
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
2695
2696
2697
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
2698
        skip_time_act=False,
2699
        only_cross_attention=False,
2700
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
2701
2702
2703
2704
2705
2706
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
2707
        self.attention_head_dim = attention_head_dim
Will Berman's avatar
Will Berman committed
2708

2709
        self.num_heads = out_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726

        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,
2727
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2728
2729
                )
            )
2730
2731
2732
2733
2734

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

Will Berman's avatar
Will Berman committed
2735
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
2736
                Attention(
Will Berman's avatar
Will Berman committed
2737
2738
2739
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
2740
                    dim_head=self.attention_head_dim,
Will Berman's avatar
Will Berman committed
2741
2742
2743
2744
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
2745
                    only_cross_attention=only_cross_attention,
2746
                    cross_attention_norm=cross_attention_norm,
2747
                    processor=processor,
Will Berman's avatar
Will Berman committed
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
                )
            )
        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,
2767
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
2779
2780
2781
2782
2783
2784
2785
2786
        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
2787
    ):
2788
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800

        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
2801
2802
2803
2804
2805
2806
2807
        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)

2808
            if self.training and self.gradient_checkpointing:
Will Berman's avatar
Will Berman committed
2809

2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
                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,
2824
                    mask,
2825
2826
2827
2828
2829
2830
2831
2832
                    cross_attention_kwargs,
                )[0]
            else:
                hidden_states = resnet(hidden_states, temb)

                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
2833
                    attention_mask=mask,
2834
2835
                    **cross_attention_kwargs,
                )
Will Berman's avatar
Will Berman committed
2836
2837
2838
2839
2840
2841

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

        return hidden_states
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
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


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

2906
2907
2908
2909
2910
2911
2912
2913
                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
                    )
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
            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,
2935
        attention_head_dim=1,  # attention dim_head
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
        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
2949
        self.attention_head_dim = attention_head_dim
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984

        # 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,
2985
                    k_out_channels // attention_head_dim
2986
                    if (i == num_layers - 1)
2987
2988
                    else out_channels // attention_head_dim,
                    attention_head_dim,
2989
2990
2991
2992
                    cross_attention_dim=cross_attention_dim,
                    temb_channels=temb_channels,
                    attention_bias=True,
                    add_self_attention=add_self_attention,
2993
                    cross_attention_norm="layer_norm",
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
                    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,
3010
3011
3012
3013
3014
3015
3016
3017
        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,
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
    ):
        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

3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
                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]
3052
3053
3054
3055
3056
3057
3058
3059
            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,
3060
                    encoder_attention_mask=encoder_attention_mask,
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
                )

        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,
3099
        cross_attention_norm: Optional[str] = None,
3100
3101
3102
3103
3104
3105
3106
3107
        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
3108
            self.attn1 = Attention(
3109
3110
3111
3112
3113
3114
                query_dim=dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                cross_attention_dim=None,
3115
                cross_attention_norm=None,
3116
3117
3118
3119
            )

        # 2. Cross-Attn
        self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
Patrick von Platen's avatar
Patrick von Platen committed
3120
        self.attn2 = Attention(
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
            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,
3139
3140
3141
3142
3143
3144
3145
3146
        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,
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
    ):
        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,
3160
                attention_mask=attention_mask,
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
                **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,
3175
            attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask,
3176
3177
3178
3179
3180
3181
3182
            **cross_attention_kwargs,
        )
        attn_output = self._to_4d(attn_output, height, weight)

        hidden_states = attn_output + hidden_states

        return hidden_states