unet_2d_blocks.py 119 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 .activations import get_activation
23
from .attention import AdaGroupNorm
24
from .attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
25
from .dual_transformer_2d import DualTransformer2DModel
26
from .resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D
27
from .transformer_2d import Transformer2DModel
Patrick von Platen's avatar
Patrick von Platen committed
28
29


30
31
32
logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


33
34
35
36
37
38
39
40
41
def get_down_block(
    down_block_type,
    num_layers,
    in_channels,
    out_channels,
    temb_channels,
    add_downsample,
    resnet_eps,
    resnet_act_fn,
42
    transformer_layers_per_block=1,
43
    num_attention_heads=None,
44
    resnet_groups=None,
45
    cross_attention_dim=None,
Patrick von Platen's avatar
Patrick von Platen committed
46
    downsample_padding=None,
47
    dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
48
    use_linear_projection=False,
49
    only_cross_attention=False,
50
    upcast_attention=False,
Will Berman's avatar
Will Berman committed
51
    resnet_time_scale_shift="default",
52
    attention_type="default",
53
54
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
55
    cross_attention_norm=None,
56
    attention_head_dim=None,
57
    downsample_type=None,
58
):
59
60
61
62
63
64
65
    # 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
66
67
68
    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
    if down_block_type == "DownBlock2D":
        return DownBlock2D(
69
70
71
72
73
74
75
            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,
76
            resnet_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
77
            downsample_padding=downsample_padding,
Will Berman's avatar
Will Berman committed
78
79
80
81
82
83
84
85
86
87
88
89
90
            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,
91
92
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
93
        )
Patrick von Platen's avatar
Patrick von Platen committed
94
    elif down_block_type == "AttnDownBlock2D":
95
96
97
98
        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
99
        return AttnDownBlock2D(
100
101
102
103
104
105
            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,
106
            resnet_groups=resnet_groups,
107
            downsample_padding=downsample_padding,
108
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
109
            resnet_time_scale_shift=resnet_time_scale_shift,
110
            downsample_type=downsample_type,
111
        )
Patrick von Platen's avatar
Patrick von Platen committed
112
    elif down_block_type == "CrossAttnDownBlock2D":
113
        if cross_attention_dim is None:
114
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
Patrick von Platen's avatar
Patrick von Platen committed
115
        return CrossAttnDownBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
116
            num_layers=num_layers,
117
            transformer_layers_per_block=transformer_layers_per_block,
Patrick von Platen's avatar
Patrick von Platen committed
118
119
120
121
122
123
            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,
124
            resnet_groups=resnet_groups,
Patrick von Platen's avatar
Patrick von Platen committed
125
            downsample_padding=downsample_padding,
126
            cross_attention_dim=cross_attention_dim,
127
            num_attention_heads=num_attention_heads,
128
            dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
129
            use_linear_projection=use_linear_projection,
130
            only_cross_attention=only_cross_attention,
131
            upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
132
            resnet_time_scale_shift=resnet_time_scale_shift,
133
            attention_type=attention_type,
Will Berman's avatar
Will Berman committed
134
135
136
137
138
139
140
141
142
143
144
145
146
147
        )
    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,
148
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
149
            resnet_time_scale_shift=resnet_time_scale_shift,
150
151
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
152
            only_cross_attention=only_cross_attention,
153
            cross_attention_norm=cross_attention_norm,
Patrick von Platen's avatar
Patrick von Platen committed
154
        )
Patrick von Platen's avatar
Patrick von Platen committed
155
156
    elif down_block_type == "SkipDownBlock2D":
        return SkipDownBlock2D(
157
158
159
160
161
162
163
164
            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
165
            resnet_time_scale_shift=resnet_time_scale_shift,
166
        )
Patrick von Platen's avatar
Patrick von Platen committed
167
168
    elif down_block_type == "AttnSkipDownBlock2D":
        return AttnSkipDownBlock2D(
169
170
171
172
173
174
175
            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,
176
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
177
            resnet_time_scale_shift=resnet_time_scale_shift,
178
        )
179
180
181
182
183
184
185
186
    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,
187
            resnet_groups=resnet_groups,
188
            downsample_padding=downsample_padding,
Will Berman's avatar
Will Berman committed
189
            resnet_time_scale_shift=resnet_time_scale_shift,
190
        )
Will Berman's avatar
Will Berman committed
191
192
193
194
195
196
197
198
199
200
    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,
201
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
202
            resnet_time_scale_shift=resnet_time_scale_shift,
Will Berman's avatar
Will Berman committed
203
        )
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
    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,
224
            attention_head_dim=attention_head_dim,
225
226
            add_self_attention=True if not add_downsample else False,
        )
Will Berman's avatar
Will Berman committed
227
    raise ValueError(f"{down_block_type} does not exist.")
228
229
230
231
232
233


def get_up_block(
    up_block_type,
    num_layers,
    in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
234
235
    out_channels,
    prev_output_channel,
236
237
238
239
    temb_channels,
    add_upsample,
    resnet_eps,
    resnet_act_fn,
240
    transformer_layers_per_block=1,
241
    num_attention_heads=None,
242
    resnet_groups=None,
243
    cross_attention_dim=None,
244
    dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
245
    use_linear_projection=False,
246
    only_cross_attention=False,
247
    upcast_attention=False,
Will Berman's avatar
Will Berman committed
248
    resnet_time_scale_shift="default",
249
    attention_type="default",
250
251
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
252
    cross_attention_norm=None,
253
    attention_head_dim=None,
254
    upsample_type=None,
255
):
256
257
258
259
260
261
262
    # 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
263
264
265
    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
    if up_block_type == "UpBlock2D":
        return UpBlock2D(
266
267
            num_layers=num_layers,
            in_channels=in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
268
269
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
270
271
272
273
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
274
            resnet_groups=resnet_groups,
Will Berman's avatar
Will Berman committed
275
276
277
278
279
280
281
282
283
284
285
286
287
288
            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,
289
290
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
291
        )
Patrick von Platen's avatar
Patrick von Platen committed
292
    elif up_block_type == "CrossAttnUpBlock2D":
293
294
        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
295
        return CrossAttnUpBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
296
            num_layers=num_layers,
297
            transformer_layers_per_block=transformer_layers_per_block,
Patrick von Platen's avatar
Patrick von Platen committed
298
299
300
301
302
303
304
            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,
305
            resnet_groups=resnet_groups,
306
            cross_attention_dim=cross_attention_dim,
307
            num_attention_heads=num_attention_heads,
308
            dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
309
            use_linear_projection=use_linear_projection,
310
            only_cross_attention=only_cross_attention,
311
            upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
312
            resnet_time_scale_shift=resnet_time_scale_shift,
313
            attention_type=attention_type,
Will Berman's avatar
Will Berman committed
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
        )
    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,
329
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
330
            resnet_time_scale_shift=resnet_time_scale_shift,
331
332
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
333
            only_cross_attention=only_cross_attention,
334
            cross_attention_norm=cross_attention_norm,
Patrick von Platen's avatar
Patrick von Platen committed
335
        )
Patrick von Platen's avatar
Patrick von Platen committed
336
    elif up_block_type == "AttnUpBlock2D":
337
338
339
340
341
        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
342
        return AttnUpBlock2D(
343
344
            num_layers=num_layers,
            in_channels=in_channels,
Patrick von Platen's avatar
Patrick von Platen committed
345
346
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
347
348
349
            temb_channels=temb_channels,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
350
            resnet_groups=resnet_groups,
351
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
352
            resnet_time_scale_shift=resnet_time_scale_shift,
353
            upsample_type=upsample_type,
354
        )
Patrick von Platen's avatar
Patrick von Platen committed
355
356
    elif up_block_type == "SkipUpBlock2D":
        return SkipUpBlock2D(
357
358
359
360
361
362
363
364
            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
365
            resnet_time_scale_shift=resnet_time_scale_shift,
366
        )
Patrick von Platen's avatar
Patrick von Platen committed
367
368
    elif up_block_type == "AttnSkipUpBlock2D":
        return AttnSkipUpBlock2D(
369
370
371
372
373
374
375
376
            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,
377
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
378
            resnet_time_scale_shift=resnet_time_scale_shift,
379
        )
380
381
382
383
384
385
386
387
    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,
388
            resnet_groups=resnet_groups,
Will Berman's avatar
Will Berman committed
389
            resnet_time_scale_shift=resnet_time_scale_shift,
YiYi Xu's avatar
YiYi Xu committed
390
            temb_channels=temb_channels,
391
        )
Will Berman's avatar
Will Berman committed
392
393
394
395
396
397
398
399
400
    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,
401
            attention_head_dim=attention_head_dim,
Will Berman's avatar
Will Berman committed
402
            resnet_time_scale_shift=resnet_time_scale_shift,
YiYi Xu's avatar
YiYi Xu committed
403
            temb_channels=temb_channels,
Will Berman's avatar
Will Berman committed
404
        )
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
    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,
425
            attention_head_dim=attention_head_dim,
426
427
        )

428
    raise ValueError(f"{up_block_type} does not exist.")
429
430


431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
class AutoencoderTinyBlock(nn.Module):
    def __init__(self, in_channels: int, out_channels: int, act_fn: str):
        super().__init__()
        act_fn = get_activation(act_fn)
        self.conv = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
            act_fn,
            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
            act_fn,
            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
        )
        self.skip = (
            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
            if in_channels != out_channels
            else nn.Identity()
        )
        self.fuse = nn.ReLU()

    def forward(self, x):
        return self.fuse(self.conv(x) + self.skip(x))


Patrick von Platen's avatar
Patrick von Platen committed
453
454
455
456
457
class UNetMidBlock2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
458
        dropout: float = 0.0,
459
        num_layers: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
460
        resnet_eps: float = 1e-6,
YiYi Xu's avatar
YiYi Xu committed
461
        resnet_time_scale_shift: str = "default",  # default, spatial
Patrick von Platen's avatar
Patrick von Platen committed
462
463
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
464
        resnet_pre_norm: bool = True,
Will Berman's avatar
Will Berman committed
465
        add_attention: bool = True,
466
        attention_head_dim=1,
Patrick von Platen's avatar
Patrick von Platen committed
467
468
469
        output_scale_factor=1.0,
    ):
        super().__init__()
470
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
Will Berman's avatar
Will Berman committed
471
        self.add_attention = add_attention
Patrick von Platen's avatar
Patrick von Platen committed
472

473
474
        # there is always at least one resnet
        resnets = [
475
            ResnetBlock2D(
476
477
478
                in_channels=in_channels,
                out_channels=in_channels,
                temb_channels=temb_channels,
479
                eps=resnet_eps,
480
481
482
483
484
485
                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
486
            )
487
488
        ]
        attentions = []
Patrick von Platen's avatar
Patrick von Platen committed
489

490
491
492
493
494
495
        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

496
        for _ in range(num_layers):
Will Berman's avatar
Will Berman committed
497
498
            if self.add_attention:
                attentions.append(
499
                    Attention(
Will Berman's avatar
Will Berman committed
500
                        in_channels,
501
502
                        heads=in_channels // attention_head_dim,
                        dim_head=attention_head_dim,
Will Berman's avatar
Will Berman committed
503
504
                        rescale_output_factor=output_scale_factor,
                        eps=resnet_eps,
YiYi Xu's avatar
YiYi Xu committed
505
506
                        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,
507
508
509
510
                        residual_connection=True,
                        bias=True,
                        upcast_softmax=True,
                        _from_deprecated_attn_block=True,
Will Berman's avatar
Will Berman committed
511
                    )
512
                )
Will Berman's avatar
Will Berman committed
513
514
515
            else:
                attentions.append(None)

516
            resnets.append(
517
                ResnetBlock2D(
518
519
520
                    in_channels=in_channels,
                    out_channels=in_channels,
                    temb_channels=temb_channels,
521
                    eps=resnet_eps,
522
523
524
525
526
527
528
                    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
529
530
            )

531
532
533
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

Will Berman's avatar
Will Berman committed
534
    def forward(self, hidden_states, temb=None):
Patrick von Platen's avatar
Patrick von Platen committed
535
        hidden_states = self.resnets[0](hidden_states, temb)
536
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
Will Berman's avatar
Will Berman committed
537
            if attn is not None:
YiYi Xu's avatar
YiYi Xu committed
538
                hidden_states = attn(hidden_states, temb=temb)
Patrick von Platen's avatar
Patrick von Platen committed
539
            hidden_states = resnet(hidden_states, temb)
Patrick von Platen's avatar
Patrick von Platen committed
540

541
        return hidden_states
Patrick von Platen's avatar
Patrick von Platen committed
542

543

Patrick von Platen's avatar
Patrick von Platen committed
544
545
546
547
548
549
550
class UNetMidBlock2DCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
551
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
552
553
554
555
556
        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,
557
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
558
559
        output_scale_factor=1.0,
        cross_attention_dim=1280,
560
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
561
        use_linear_projection=False,
562
        upcast_attention=False,
563
        attention_type="default",
Patrick von Platen's avatar
Patrick von Platen committed
564
565
566
    ):
        super().__init__()

567
        self.has_cross_attention = True
568
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
569
570
571
572
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

        # there is always at least one resnet
        resnets = [
573
            ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
                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):
589
590
591
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
592
593
                        num_attention_heads,
                        in_channels // num_attention_heads,
594
                        in_channels=in_channels,
595
                        num_layers=transformer_layers_per_block,
596
597
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
598
                        use_linear_projection=use_linear_projection,
599
                        upcast_attention=upcast_attention,
600
                        attention_type=attention_type,
601
602
603
604
605
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
606
607
                        num_attention_heads,
                        in_channels // num_attention_heads,
608
609
610
611
612
                        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
613
614
                )
            resnets.append(
615
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
                    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)

632
633
        self.gradient_checkpointing = False

634
    def forward(
635
636
637
638
639
640
641
642
        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
643
644
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
645
646
647
648
649
650
651
652
653
654
655
656
            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

                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
ethansmith2000's avatar
ethansmith2000 committed
657
                hidden_states = attn(
658
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
659
660
661
662
663
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
                )[0]
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
            else:
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
                )[0]
                hidden_states = resnet(hidden_states, temb)
Will Berman's avatar
Will Berman committed
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696

        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,
697
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
698
699
        output_scale_factor=1.0,
        cross_attention_dim=1280,
700
        skip_time_act=False,
701
        only_cross_attention=False,
702
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
703
704
705
706
707
    ):
        super().__init__()

        self.has_cross_attention = True

708
        self.attention_head_dim = attention_head_dim
Will Berman's avatar
Will Berman committed
709
710
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)

711
        self.num_heads = in_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
712
713
714
715
716
717
718
719
720
721
722
723
724
725

        # 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,
726
                skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
727
728
729
730
731
            )
        ]
        attentions = []

        for _ in range(num_layers):
732
733
734
735
            processor = (
                AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
            )

Will Berman's avatar
Will Berman committed
736
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
737
                Attention(
Will Berman's avatar
Will Berman committed
738
739
740
                    query_dim=in_channels,
                    cross_attention_dim=in_channels,
                    heads=self.num_heads,
741
                    dim_head=self.attention_head_dim,
Will Berman's avatar
Will Berman committed
742
743
744
745
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
746
                    only_cross_attention=only_cross_attention,
747
                    cross_attention_norm=cross_attention_norm,
748
                    processor=processor,
Will Berman's avatar
Will Berman committed
749
750
751
752
753
754
755
756
757
758
759
760
761
762
                )
            )
            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,
763
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
764
765
766
767
768
769
                )
            )

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

770
    def forward(
771
772
773
774
775
776
777
        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,
778
779
    ):
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
780
781
782
783
784
785
786
787
788
789
790
791

        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
792
793
794
795
796
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            # attn
            hidden_states = attn(
                hidden_states,
797
                encoder_hidden_states=encoder_hidden_states,
798
                attention_mask=mask,
799
                **cross_attention_kwargs,
Will Berman's avatar
Will Berman committed
800
801
802
            )

            # resnet
Patrick von Platen's avatar
Patrick von Platen committed
803
804
805
806
807
            hidden_states = resnet(hidden_states, temb)

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
808
class AttnDownBlock2D(nn.Module):
809
810
811
812
813
814
815
816
817
818
819
820
    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,
821
        attention_head_dim=1,
822
        output_scale_factor=1.0,
823
        downsample_padding=1,
824
        downsample_type="conv",
825
826
827
828
    ):
        super().__init__()
        resnets = []
        attentions = []
829
        self.downsample_type = downsample_type
830

831
832
833
834
835
836
        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

837
838
839
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
840
                ResnetBlock2D(
841
842
843
844
845
846
847
848
849
850
851
852
853
                    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(
854
                Attention(
855
                    out_channels,
856
857
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
858
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
859
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
860
                    norm_num_groups=resnet_groups,
861
862
863
864
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
865
866
867
868
869
870
                )
            )

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

871
        if downsample_type == "conv":
872
            self.downsamplers = nn.ModuleList(
873
874
                [
                    Downsample2D(
875
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
876
877
                    )
                ]
878
            )
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
        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,
                    )
                ]
            )
897
898
899
        else:
            self.downsamplers = None

900
    def forward(self, hidden_states, temb=None, upsample_size=None):
901
902
903
904
905
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb)
            hidden_states = attn(hidden_states)
906
            output_states = output_states + (hidden_states,)
907
908
909

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
910
911
912
913
                if self.downsample_type == "resnet":
                    hidden_states = downsampler(hidden_states, temb=temb)
                else:
                    hidden_states = downsampler(hidden_states)
914
915
916
917
918
919

            output_states += (hidden_states,)

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
920
class CrossAttnDownBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
921
922
923
924
925
926
927
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
928
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
929
930
931
932
933
        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,
934
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
935
936
937
938
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        downsample_padding=1,
        add_downsample=True,
939
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
940
        use_linear_projection=False,
941
        only_cross_attention=False,
942
        upcast_attention=False,
943
        attention_type="default",
Patrick von Platen's avatar
Patrick von Platen committed
944
945
946
947
948
    ):
        super().__init__()
        resnets = []
        attentions = []

949
        self.has_cross_attention = True
950
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
951
952
953
954

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
955
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
956
957
958
959
960
961
962
963
964
965
966
967
                    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,
                )
            )
968
969
970
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
971
972
                        num_attention_heads,
                        out_channels // num_attention_heads,
973
                        in_channels=out_channels,
974
                        num_layers=transformer_layers_per_block,
975
976
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
977
                        use_linear_projection=use_linear_projection,
978
                        only_cross_attention=only_cross_attention,
979
                        upcast_attention=upcast_attention,
980
                        attention_type=attention_type,
981
982
983
984
985
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
986
987
                        num_attention_heads,
                        out_channels // num_attention_heads,
988
989
990
991
992
                        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
993
994
995
996
997
998
999
1000
                )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
1001
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
Patrick von Platen's avatar
Patrick von Platen committed
1002
1003
1004
1005
1006
1007
                    )
                ]
            )
        else:
            self.downsamplers = None

1008
1009
        self.gradient_checkpointing = False

1010
    def forward(
1011
1012
1013
1014
1015
1016
1017
        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
1018
        additional_residuals=None,
1019
    ):
Patrick von Platen's avatar
Patrick von Platen committed
1020
1021
        output_states = ()

Will Berman's avatar
Will Berman committed
1022
1023
1024
        blocks = list(zip(self.resnets, self.attentions))

        for i, (resnet, attn) in enumerate(blocks):
1025
1026
            if self.training and self.gradient_checkpointing:

Will Berman's avatar
Will Berman committed
1027
                def create_custom_forward(module, return_dict=None):
1028
                    def custom_forward(*inputs):
Will Berman's avatar
Will Berman committed
1029
1030
1031
1032
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)
1033
1034
1035

                    return custom_forward

1036
1037
1038
1039
1040
1041
1042
                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,
                )
ethansmith2000's avatar
ethansmith2000 committed
1043
                hidden_states = attn(
1044
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
1045
1046
1047
1048
1049
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
1050
                )[0]
1051
1052
            else:
                hidden_states = resnet(hidden_states, temb)
1053
1054
1055
1056
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
1057
1058
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
1059
1060
                    return_dict=False,
                )[0]
1061

Will Berman's avatar
Will Berman committed
1062
1063
1064
1065
            # 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

1066
            output_states = output_states + (hidden_states,)
Patrick von Platen's avatar
Patrick von Platen committed
1067
1068
1069
1070
1071

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

1072
            output_states = output_states + (hidden_states,)
Patrick von Platen's avatar
Patrick von Platen committed
1073
1074
1075
1076

        return hidden_states, output_states


Patrick von Platen's avatar
Patrick von Platen committed
1077
class DownBlock2D(nn.Module):
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
    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
1092
        downsample_padding=1,
1093
1094
1095
1096
1097
1098
1099
    ):
        super().__init__()
        resnets = []

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1100
                ResnetBlock2D(
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
                    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
1118
1119
                [
                    Downsample2D(
1120
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
Patrick von Platen's avatar
Patrick von Platen committed
1121
1122
                    )
                ]
1123
1124
1125
1126
            )
        else:
            self.downsamplers = None

1127
1128
        self.gradient_checkpointing = False

1129
1130
1131
1132
    def forward(self, hidden_states, temb=None):
        output_states = ()

        for resnet in self.resnets:
1133
1134
1135
1136
1137
1138
1139
1140
            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

1141
1142
1143
1144
1145
1146
1147
1148
                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
                    )
1149
1150
1151
            else:
                hidden_states = resnet(hidden_states, temb)

1152
            output_states = output_states + (hidden_states,)
1153
1154
1155
1156
1157

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

1158
            output_states = output_states + (hidden_states,)
1159
1160
1161
1162

        return hidden_states, output_states


1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
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(
1185
                ResnetBlock2D(
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
                    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(
1205
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
                    )
                ]
            )
        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


1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
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,
1235
        attention_head_dim=1,
1236
1237
1238
1239
1240
1241
1242
1243
        output_scale_factor=1.0,
        add_downsample=True,
        downsample_padding=1,
    ):
        super().__init__()
        resnets = []
        attentions = []

1244
1245
1246
1247
1248
1249
        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

1250
1251
1252
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            resnets.append(
1253
                ResnetBlock2D(
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
                    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(
1267
                Attention(
1268
                    out_channels,
1269
1270
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1271
1272
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
1273
                    norm_num_groups=resnet_groups,
1274
1275
1276
1277
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
                )
            )

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

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
1288
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
                    )
                ]
            )
        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
1307
class AttnSkipDownBlock2D(nn.Module):
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
    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,
1319
        attention_head_dim=1,
1320
1321
1322
1323
1324
1325
1326
        output_scale_factor=np.sqrt(2.0),
        add_downsample=True,
    ):
        super().__init__()
        self.attentions = nn.ModuleList([])
        self.resnets = nn.ModuleList([])

1327
1328
1329
1330
1331
1332
        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

1333
1334
1335
        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels
            self.resnets.append(
1336
                ResnetBlock2D(
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
                    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(
1351
                Attention(
1352
                    out_channels,
1353
1354
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
1355
1356
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
1357
1358
1359
1360
1361
                    norm_num_groups=32,
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
1362
1363
1364
1365
                )
            )

        if add_downsample:
1366
            self.resnet_down = ResnetBlock2D(
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
                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,
1377
                use_in_shortcut=True,
1378
1379
1380
                down=True,
                kernel="fir",
            )
1381
            self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
            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
1408
class SkipDownBlock2D(nn.Module):
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
    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(
1430
                ResnetBlock2D(
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
                    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:
1446
            self.resnet_down = ResnetBlock2D(
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
                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,
1457
                use_in_shortcut=True,
1458
1459
1460
                down=True,
                kernel="fir",
            )
1461
            self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
            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
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
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,
1502
        skip_time_act=False,
Will Berman's avatar
Will Berman committed
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
    ):
        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,
1521
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
                )
            )

        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,
1541
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
                        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

1563
1564
1565
1566
1567
1568
1569
1570
                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
1571
1572
1573
            else:
                hidden_states = resnet(hidden_states, temb)

1574
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1575
1576
1577
1578
1579

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

1580
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597

        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,
1598
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
1599
1600
1601
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_downsample=True,
1602
        skip_time_act=False,
1603
        only_cross_attention=False,
1604
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
1605
1606
1607
1608
1609
1610
1611
1612
    ):
        super().__init__()

        self.has_cross_attention = True

        resnets = []
        attentions = []

1613
1614
        self.attention_head_dim = attention_head_dim
        self.num_heads = out_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629

        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,
1630
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1631
1632
                )
            )
1633
1634
1635
1636
1637

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

Will Berman's avatar
Will Berman committed
1638
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
1639
                Attention(
Will Berman's avatar
Will Berman committed
1640
1641
1642
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
1643
                    dim_head=attention_head_dim,
Will Berman's avatar
Will Berman committed
1644
1645
1646
1647
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
1648
                    only_cross_attention=only_cross_attention,
1649
                    cross_attention_norm=cross_attention_norm,
1650
                    processor=processor,
Will Berman's avatar
Will Berman committed
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
                )
            )
        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,
1670
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
1671
1672
1673
1674
1675
1676
1677
1678
1679
                        down=True,
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

1680
    def forward(
1681
1682
1683
1684
1685
1686
1687
        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,
1688
    ):
Will Berman's avatar
Will Berman committed
1689
        output_states = ()
1690
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
Will Berman's avatar
Will Berman committed
1691

1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
        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
1703
        for resnet, attn in zip(self.resnets, self.attentions):
1704
            if self.training and self.gradient_checkpointing:
Will Berman's avatar
Will Berman committed
1705

1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
                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)
ethansmith2000's avatar
ethansmith2000 committed
1716
                hidden_states = attn(
1717
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
1718
1719
1720
1721
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=mask,
                    **cross_attention_kwargs,
                )
1722
1723
1724
1725
1726
1727
            else:
                hidden_states = resnet(hidden_states, temb)

                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
1728
                    attention_mask=mask,
1729
1730
                    **cross_attention_kwargs,
                )
Will Berman's avatar
Will Berman committed
1731

1732
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1733
1734
1735
1736
1737

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

1738
            output_states = output_states + (hidden_states,)
Will Berman's avatar
Will Berman committed
1739
1740
1741
1742

        return hidden_states, output_states


1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
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

1801
1802
1803
1804
1805
1806
1807
1808
                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
                    )
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
            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,
1832
        attention_head_dim: int = 64,
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
        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,
1865
1866
                    out_channels // attention_head_dim,
                    attention_head_dim,
1867
1868
1869
1870
                    cross_attention_dim=cross_attention_dim,
                    temb_channels=temb_channels,
                    attention_bias=True,
                    add_self_attention=add_self_attention,
1871
                    cross_attention_norm="layer_norm",
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
                    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(
1887
1888
1889
1890
1891
1892
1893
        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,
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
    ):
        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

1909
1910
1911
1912
1913
1914
1915
                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,
                )
ethansmith2000's avatar
ethansmith2000 committed
1916
                hidden_states = attn(
1917
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
1918
1919
1920
1921
1922
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                    encoder_attention_mask=encoder_attention_mask,
1923
                )
1924
1925
1926
1927
1928
1929
1930
1931
            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,
1932
                    encoder_attention_mask=encoder_attention_mask,
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
                )

            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
1947
class AttnUpBlock2D(nn.Module):
1948
1949
1950
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
1951
1952
        prev_output_channel: int,
        out_channels: int,
1953
1954
1955
1956
1957
1958
1959
1960
        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,
1961
        attention_head_dim=1,
1962
        output_scale_factor=1.0,
1963
        upsample_type="conv",
1964
1965
1966
1967
1968
    ):
        super().__init__()
        resnets = []
        attentions = []

1969
1970
        self.upsample_type = upsample_type

1971
1972
1973
1974
1975
1976
        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

1977
        for i in range(num_layers):
Patrick von Platen's avatar
Patrick von Platen committed
1978
1979
1980
            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

1981
            resnets.append(
1982
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
1983
1984
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
                    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(
1996
                Attention(
Patrick von Platen's avatar
Patrick von Platen committed
1997
                    out_channels,
1998
1999
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
2000
                    rescale_output_factor=output_scale_factor,
Patrick von Platen's avatar
Patrick von Platen committed
2001
                    eps=resnet_eps,
Will Berman's avatar
Will Berman committed
2002
                    norm_num_groups=resnet_groups,
2003
2004
2005
2006
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
2007
2008
2009
2010
2011
2012
                )
            )

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

2013
        if upsample_type == "conv":
Patrick von Platen's avatar
Patrick von Platen committed
2014
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
        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,
                    )
                ]
            )
2033
2034
2035
        else:
            self.upsamplers = None

2036
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
        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:
2048
2049
2050
2051
                if self.upsample_type == "resnet":
                    hidden_states = upsampler(hidden_states, temb=temb)
                else:
                    hidden_states = upsampler(hidden_states)
2052
2053
2054
2055

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2056
class CrossAttnUpBlock2D(nn.Module):
Patrick von Platen's avatar
Patrick von Platen committed
2057
2058
2059
2060
2061
2062
2063
2064
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
2065
        transformer_layers_per_block: int = 1,
Patrick von Platen's avatar
Patrick von Platen committed
2066
2067
2068
2069
2070
        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,
2071
        num_attention_heads=1,
Patrick von Platen's avatar
Patrick von Platen committed
2072
2073
2074
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
2075
        dual_cross_attention=False,
Suraj Patil's avatar
Suraj Patil committed
2076
        use_linear_projection=False,
2077
        only_cross_attention=False,
2078
        upcast_attention=False,
2079
        attention_type="default",
Patrick von Platen's avatar
Patrick von Platen committed
2080
2081
2082
2083
2084
    ):
        super().__init__()
        resnets = []
        attentions = []

2085
        self.has_cross_attention = True
2086
        self.num_attention_heads = num_attention_heads
Patrick von Platen's avatar
Patrick von Platen committed
2087
2088
2089
2090
2091
2092

        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(
2093
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
                    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,
                )
            )
2106
2107
2108
            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
2109
2110
                        num_attention_heads,
                        out_channels // num_attention_heads,
2111
                        in_channels=out_channels,
2112
                        num_layers=transformer_layers_per_block,
2113
2114
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
Suraj Patil's avatar
Suraj Patil committed
2115
                        use_linear_projection=use_linear_projection,
2116
                        only_cross_attention=only_cross_attention,
2117
                        upcast_attention=upcast_attention,
2118
                        attention_type=attention_type,
2119
2120
2121
2122
2123
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
2124
2125
                        num_attention_heads,
                        out_channels // num_attention_heads,
2126
2127
2128
2129
2130
                        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
2131
2132
2133
2134
2135
2136
2137
2138
2139
                )
        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

2140
2141
2142
2143
        self.gradient_checkpointing = False

    def forward(
        self,
2144
2145
2146
2147
2148
2149
2150
2151
        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,
2152
    ):
Patrick von Platen's avatar
Patrick von Platen committed
2153
2154
2155
2156
2157
2158
        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)

2159
2160
            if self.training and self.gradient_checkpointing:

Will Berman's avatar
Will Berman committed
2161
                def create_custom_forward(module, return_dict=None):
2162
                    def custom_forward(*inputs):
Will Berman's avatar
Will Berman committed
2163
2164
2165
2166
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)
2167
2168
2169

                    return custom_forward

2170
2171
2172
2173
2174
2175
2176
                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,
                )
ethansmith2000's avatar
ethansmith2000 committed
2177
                hidden_states = attn(
2178
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
2179
2180
2181
2182
2183
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
2184
                )[0]
2185
2186
            else:
                hidden_states = resnet(hidden_states, temb)
2187
2188
2189
2190
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
2191
2192
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
2193
2194
                    return_dict=False,
                )[0]
Patrick von Platen's avatar
Patrick von Platen committed
2195
2196
2197

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2198
                hidden_states = upsampler(hidden_states, upsample_size)
Patrick von Platen's avatar
Patrick von Platen committed
2199
2200
2201
2202

        return hidden_states


Patrick von Platen's avatar
Patrick von Platen committed
2203
class UpBlock2D(nn.Module):
2204
2205
2206
    def __init__(
        self,
        in_channels: int,
Patrick von Platen's avatar
Patrick von Platen committed
2207
2208
        prev_output_channel: int,
        out_channels: int,
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
        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
2224
2225
2226
            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

2227
            resnets.append(
2228
                ResnetBlock2D(
Patrick von Platen's avatar
Patrick von Platen committed
2229
2230
                    in_channels=resnet_in_channels + res_skip_channels,
                    out_channels=out_channels,
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
                    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
2245
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
2246
2247
2248
        else:
            self.upsamplers = None

2249
2250
        self.gradient_checkpointing = False

2251
    def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
2252
2253
2254
2255
2256
2257
        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)

2258
2259
2260
2261
2262
2263
2264
2265
            if self.training and self.gradient_checkpointing:

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

                    return custom_forward

2266
2267
2268
2269
2270
2271
2272
2273
                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
                    )
2274
2275
            else:
                hidden_states = resnet(hidden_states, temb)
2276
2277
2278

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
2279
                hidden_states = upsampler(hidden_states, upsample_size)
2280
2281

        return hidden_states
2282
2283


2284
2285
2286
2287
2288
2289
2290
2291
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
2292
        resnet_time_scale_shift: str = "default",  # default, spatial
2293
2294
2295
2296
2297
        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
2298
        temb_channels=None,
2299
2300
2301
2302
2303
2304
2305
2306
    ):
        super().__init__()
        resnets = []

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

            resnets.append(
2307
                ResnetBlock2D(
2308
2309
                    in_channels=input_channels,
                    out_channels=out_channels,
YiYi Xu's avatar
YiYi Xu committed
2310
                    temb_channels=temb_channels,
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
                    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
2328
    def forward(self, hidden_states, temb=None):
2329
        for resnet in self.resnets:
YiYi Xu's avatar
YiYi Xu committed
2330
            hidden_states = resnet(hidden_states, temb=temb)
2331
2332
2333
2334
2335
2336
2337
2338

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

        return hidden_states


2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
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,
2351
        attention_head_dim=1,
2352
2353
        output_scale_factor=1.0,
        add_upsample=True,
YiYi Xu's avatar
YiYi Xu committed
2354
        temb_channels=None,
2355
2356
2357
2358
2359
    ):
        super().__init__()
        resnets = []
        attentions = []

2360
2361
2362
2363
2364
2365
        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

2366
2367
2368
2369
        for i in range(num_layers):
            input_channels = in_channels if i == 0 else out_channels

            resnets.append(
2370
                ResnetBlock2D(
2371
2372
                    in_channels=input_channels,
                    out_channels=out_channels,
YiYi Xu's avatar
YiYi Xu committed
2373
                    temb_channels=temb_channels,
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
                    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(
2384
                Attention(
2385
                    out_channels,
2386
2387
                    heads=out_channels // attention_head_dim,
                    dim_head=attention_head_dim,
2388
2389
                    rescale_output_factor=output_scale_factor,
                    eps=resnet_eps,
2390
                    norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None,
YiYi Xu's avatar
YiYi Xu committed
2391
                    spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
2392
2393
2394
2395
                    residual_connection=True,
                    bias=True,
                    upcast_softmax=True,
                    _from_deprecated_attn_block=True,
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
                )
            )

        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
2407
    def forward(self, hidden_states, temb=None):
2408
        for resnet, attn in zip(self.resnets, self.attentions):
YiYi Xu's avatar
YiYi Xu committed
2409
2410
            hidden_states = resnet(hidden_states, temb=temb)
            hidden_states = attn(hidden_states, temb=temb)
2411
2412
2413
2414
2415
2416
2417
2418

        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
2419
class AttnSkipUpBlock2D(nn.Module):
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
    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,
2432
        attention_head_dim=1,
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
        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(
2445
                ResnetBlock2D(
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
                    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,
                )
            )

2460
2461
2462
2463
2464
2465
        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

2466
        self.attentions.append(
2467
            Attention(
2468
                out_channels,
2469
2470
                heads=out_channels // attention_head_dim,
                dim_head=attention_head_dim,
2471
2472
                rescale_output_factor=output_scale_factor,
                eps=resnet_eps,
2473
2474
2475
2476
2477
                norm_num_groups=32,
                residual_connection=True,
                bias=True,
                upcast_softmax=True,
                _from_deprecated_attn_block=True,
2478
2479
2480
2481
2482
            )
        )

        self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
        if add_upsample:
2483
            self.resnet_up = ResnetBlock2D(
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
                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,
2495
                use_in_shortcut=True,
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
                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
2538
class SkipUpBlock2D(nn.Module):
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
    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(
2563
                ResnetBlock2D(
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
                    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:
2580
            self.resnet_up = ResnetBlock2D(
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
                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,
2592
                use_in_shortcut=True,
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
                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
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648


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,
2649
        skip_time_act=False,
Will Berman's avatar
Will Berman committed
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
    ):
        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,
2670
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
                )
            )

        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,
2690
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
                        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

2715
2716
2717
2718
2719
2720
2721
2722
                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
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
            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,
2747
        attention_head_dim=1,
Will Berman's avatar
Will Berman committed
2748
2749
2750
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
2751
        skip_time_act=False,
2752
        only_cross_attention=False,
2753
        cross_attention_norm=None,
Will Berman's avatar
Will Berman committed
2754
2755
2756
2757
2758
2759
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
2760
        self.attention_head_dim = attention_head_dim
Will Berman's avatar
Will Berman committed
2761

2762
        self.num_heads = out_channels // self.attention_head_dim
Will Berman's avatar
Will Berman committed
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779

        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,
2780
                    skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2781
2782
                )
            )
2783
2784
2785
2786
2787

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

Will Berman's avatar
Will Berman committed
2788
            attentions.append(
Patrick von Platen's avatar
Patrick von Platen committed
2789
                Attention(
Will Berman's avatar
Will Berman committed
2790
2791
2792
                    query_dim=out_channels,
                    cross_attention_dim=out_channels,
                    heads=self.num_heads,
2793
                    dim_head=self.attention_head_dim,
Will Berman's avatar
Will Berman committed
2794
2795
2796
2797
                    added_kv_proj_dim=cross_attention_dim,
                    norm_num_groups=resnet_groups,
                    bias=True,
                    upcast_softmax=True,
2798
                    only_cross_attention=only_cross_attention,
2799
                    cross_attention_norm=cross_attention_norm,
2800
                    processor=processor,
Will Berman's avatar
Will Berman committed
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
                )
            )
        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,
2820
                        skip_time_act=skip_time_act,
Will Berman's avatar
Will Berman committed
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
                        up=True,
                    )
                ]
            )
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
2832
2833
2834
2835
2836
2837
2838
2839
        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
2840
    ):
2841
        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853

        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
2854
2855
2856
2857
2858
2859
2860
        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)

2861
            if self.training and self.gradient_checkpointing:
Will Berman's avatar
Will Berman committed
2862

2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
                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)
ethansmith2000's avatar
ethansmith2000 committed
2873
                hidden_states = attn(
2874
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
2875
2876
2877
2878
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=mask,
                    **cross_attention_kwargs,
                )
2879
2880
2881
2882
2883
2884
            else:
                hidden_states = resnet(hidden_states, temb)

                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
2885
                    attention_mask=mask,
2886
2887
                    **cross_attention_kwargs,
                )
Will Berman's avatar
Will Berman committed
2888
2889
2890
2891
2892
2893

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

        return hidden_states
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957


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

2958
2959
2960
2961
2962
2963
2964
2965
                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
                    )
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
            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,
2987
        attention_head_dim=1,  # attention dim_head
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
        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
3001
        self.attention_head_dim = attention_head_dim
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036

        # 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,
3037
                    k_out_channels // attention_head_dim
3038
                    if (i == num_layers - 1)
3039
3040
                    else out_channels // attention_head_dim,
                    attention_head_dim,
3041
3042
3043
3044
                    cross_attention_dim=cross_attention_dim,
                    temb_channels=temb_channels,
                    attention_bias=True,
                    add_self_attention=add_self_attention,
3045
                    cross_attention_norm="layer_norm",
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
                    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,
3062
3063
3064
3065
3066
3067
3068
3069
        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,
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
    ):
        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

3087
3088
3089
3090
3091
3092
3093
                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,
                )
ethansmith2000's avatar
ethansmith2000 committed
3094
                hidden_states = attn(
3095
                    hidden_states,
ethansmith2000's avatar
ethansmith2000 committed
3096
3097
3098
3099
3100
3101
                    encoder_hidden_states=encoder_hidden_states,
                    emb=temb,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                    encoder_attention_mask=encoder_attention_mask,
                )
3102
3103
3104
3105
3106
3107
3108
3109
            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,
3110
                    encoder_attention_mask=encoder_attention_mask,
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
                )

        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,
3149
        cross_attention_norm: Optional[str] = None,
3150
3151
3152
3153
3154
3155
3156
3157
        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
3158
            self.attn1 = Attention(
3159
3160
3161
3162
3163
3164
                query_dim=dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                cross_attention_dim=None,
3165
                cross_attention_norm=None,
3166
3167
3168
3169
            )

        # 2. Cross-Attn
        self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
Patrick von Platen's avatar
Patrick von Platen committed
3170
        self.attn2 = Attention(
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
            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,
3189
3190
3191
3192
3193
3194
3195
3196
        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,
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
    ):
        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,
3210
                attention_mask=attention_mask,
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
                **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,
3225
            attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask,
3226
3227
3228
3229
3230
3231
3232
            **cross_attention_kwargs,
        )
        attn_output = self._to_4d(attn_output, height, weight)

        hidden_states = attn_output + hidden_states

        return hidden_states