File size: 18,082 Bytes
2890711
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
from functools import partial
from typing import List, Optional, Union

from einops import rearrange

import torch
from ...modules.diffusionmodules.openaimodel import *
from ...modules.video_attention import SpatialVideoTransformer
from ...util import default
from .util import AlphaBlender


class VideoResBlock(ResBlock):
    def __init__(
        self,
        channels: int,
        emb_channels: int,
        dropout: float,
        video_kernel_size: Union[int, List[int]] = 3,
        merge_strategy: str = "fixed",
        merge_factor: float = 0.5,
        out_channels: Optional[int] = None,
        use_conv: bool = False,
        use_scale_shift_norm: bool = False,
        dims: int = 2,
        use_checkpoint: bool = False,
        up: bool = False,
        down: bool = False,
    ):
        super().__init__(
            channels,
            emb_channels,
            dropout,
            out_channels=out_channels,
            use_conv=use_conv,
            use_scale_shift_norm=use_scale_shift_norm,
            dims=dims,
            use_checkpoint=use_checkpoint,
            up=up,
            down=down,
        )

        self.time_stack = ResBlock(
            default(out_channels, channels),
            emb_channels,
            dropout=dropout,
            dims=3,
            out_channels=default(out_channels, channels),
            use_scale_shift_norm=False,
            use_conv=False,
            up=False,
            down=False,
            kernel_size=video_kernel_size,
            use_checkpoint=use_checkpoint,
            exchange_temb_dims=True,
        )
        self.time_mixer = AlphaBlender(
            alpha=merge_factor,
            merge_strategy=merge_strategy,
            rearrange_pattern="b t -> b 1 t 1 1",
        )

    def forward(
        self,
        x: th.Tensor,
        emb: th.Tensor,
        num_video_frames: int,
        image_only_indicator: Optional[th.Tensor] = None,
    ) -> th.Tensor:
        x = super().forward(x, emb)

        x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=num_video_frames)
        x = rearrange(x, "(b t) c h w -> b c t h w", t=num_video_frames)

        x = self.time_stack(
            x, rearrange(emb, "(b t) ... -> b t ...", t=num_video_frames)
        )
        x = self.time_mixer(
            x_spatial=x_mix, x_temporal=x, image_only_indicator=image_only_indicator
        )
        x = rearrange(x, "b c t h w -> (b t) c h w")
        return x


class VideoUNet(nn.Module):
    def __init__(
        self,
        in_channels: int,
        model_channels: int,
        out_channels: int,
        num_frames: int,
        num_res_blocks: int,
        attention_resolutions: int,
        dropout: float = 0.0,
        channel_mult: List[int] = (1, 2, 4, 8),
        conv_resample: bool = True,
        dims: int = 2,
        num_classes: Optional[int] = None,
        use_checkpoint: bool = False,
        num_heads: int = -1,
        num_head_channels: int = -1,
        num_heads_upsample: int = -1,
        use_scale_shift_norm: bool = False,
        resblock_updown: bool = False,
        transformer_depth: Union[List[int], int] = 1,
        transformer_depth_middle: Optional[int] = None,
        context_dim: Optional[int] = None,
        time_downup: bool = False,
        time_context_dim: Optional[int] = None,
        extra_ff_mix_layer: bool = False,
        use_spatial_context: bool = False,
        merge_strategy: str = "fixed",
        merge_factor: float = 0.5,
        spatial_transformer_attn_type: str = "softmax",
        video_kernel_size: Union[int, List[int]] = 3,
        use_linear_in_transformer: bool = False,
        adm_in_channels: Optional[int] = None,
        disable_temporal_crossattention: bool = False,
        max_ddpm_temb_period: int = 10000,
    ):
        super(VideoUNet, self).__init__()
        assert context_dim is not None

        if num_heads_upsample == -1:
            num_heads_upsample = num_heads

        if num_heads == -1:
            assert num_head_channels != -1

        if num_head_channels == -1:
            assert num_heads != -1

        self.in_channels = in_channels
        self.model_channels = model_channels
        self.out_channels = out_channels
        self.num_frames = num_frames
        if isinstance(transformer_depth, int):
            transformer_depth = len(channel_mult) * [transformer_depth]
        transformer_depth_middle = default(
            transformer_depth_middle, transformer_depth[-1]
        )

        self.num_res_blocks = num_res_blocks
        self.attention_resolutions = attention_resolutions
        self.dropout = dropout
        self.channel_mult = channel_mult
        self.conv_resample = conv_resample
        self.num_classes = num_classes
        self.use_checkpoint = use_checkpoint
        self.num_heads = num_heads
        self.num_head_channels = num_head_channels
        self.num_heads_upsample = num_heads_upsample

        time_embed_dim = model_channels * 4
        self.time_embed = nn.Sequential(
            linear(model_channels, time_embed_dim),
            nn.SiLU(),
            linear(time_embed_dim, time_embed_dim),
        )

        if self.num_classes is not None:
            if isinstance(self.num_classes, int):
                self.label_emb = nn.Embedding(num_classes, time_embed_dim)
            elif self.num_classes == "continuous":
                print("setting up linear c_adm embedding layer")
                self.label_emb = nn.Linear(1, time_embed_dim)
            elif self.num_classes == "timestep":
                self.label_emb = nn.Sequential(
                    Timestep(model_channels),
                    nn.Sequential(
                        linear(model_channels, time_embed_dim),
                        nn.SiLU(),
                        linear(time_embed_dim, time_embed_dim),
                    ),
                )
            elif self.num_classes == "sequential":
                assert adm_in_channels is not None
                self.label_emb = nn.Sequential(
                    nn.Sequential(
                        linear(adm_in_channels, time_embed_dim),
                        nn.SiLU(),
                        linear(time_embed_dim, time_embed_dim),
                    )
                )
            else:
                raise ValueError()

        self.input_blocks = nn.ModuleList(
            [
                TimestepEmbedSequential(
                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
                )
            ]
        )
        self._feature_size = model_channels
        input_block_chans = [model_channels]
        ch = model_channels
        ds = 1

        def get_attention_layer(
            ch,
            num_heads,
            dim_head,
            depth=1,
            context_dim=None,
            use_checkpoint=False,
            disabled_sa=False,
        ):
            return SpatialVideoTransformer(
                ch,
                num_heads,
                dim_head,
                depth=depth,
                context_dim=context_dim,
                time_context_dim=time_context_dim,
                dropout=dropout,
                ff_in=extra_ff_mix_layer,
                use_spatial_context=use_spatial_context,
                merge_strategy=merge_strategy,
                merge_factor=merge_factor,
                checkpoint=use_checkpoint,
                use_linear=use_linear_in_transformer,
                attn_mode=spatial_transformer_attn_type,
                disable_self_attn=disabled_sa,
                disable_temporal_crossattention=disable_temporal_crossattention,
                max_time_embed_period=max_ddpm_temb_period,
            )

        def get_resblock(
            merge_factor,
            merge_strategy,
            video_kernel_size,
            ch,
            time_embed_dim,
            dropout,
            out_ch,
            dims,
            use_checkpoint,
            use_scale_shift_norm,
            down=False,
            up=False,
        ):
            return VideoResBlock(
                merge_factor=merge_factor,
                merge_strategy=merge_strategy,
                video_kernel_size=video_kernel_size,
                channels=ch,
                emb_channels=time_embed_dim,
                dropout=dropout,
                out_channels=out_ch,
                dims=dims,
                use_checkpoint=use_checkpoint,
                use_scale_shift_norm=use_scale_shift_norm,
                down=down,
                up=up,
            )

        for level, mult in enumerate(channel_mult):
            for _ in range(num_res_blocks):
                layers = [
                    get_resblock(
                        merge_factor=merge_factor,
                        merge_strategy=merge_strategy,
                        video_kernel_size=video_kernel_size,
                        ch=ch,
                        time_embed_dim=time_embed_dim,
                        dropout=dropout,
                        out_ch=mult * model_channels,
                        dims=dims,
                        use_checkpoint=use_checkpoint,
                        use_scale_shift_norm=use_scale_shift_norm,
                    )
                ]
                ch = mult * model_channels
                if ds in attention_resolutions:
                    if num_head_channels == -1:
                        dim_head = ch // num_heads
                    else:
                        num_heads = ch // num_head_channels
                        dim_head = num_head_channels

                    layers.append(
                        get_attention_layer(
                            ch,
                            num_heads,
                            dim_head,
                            depth=transformer_depth[level],
                            context_dim=context_dim,
                            use_checkpoint=use_checkpoint,
                            disabled_sa=False,
                        )
                    )
                self.input_blocks.append(TimestepEmbedSequential(*layers))
                self._feature_size += ch
                input_block_chans.append(ch)
            if level != len(channel_mult) - 1:
                ds *= 2
                out_ch = ch
                self.input_blocks.append(
                    TimestepEmbedSequential(
                        get_resblock(
                            merge_factor=merge_factor,
                            merge_strategy=merge_strategy,
                            video_kernel_size=video_kernel_size,
                            ch=ch,
                            time_embed_dim=time_embed_dim,
                            dropout=dropout,
                            out_ch=out_ch,
                            dims=dims,
                            use_checkpoint=use_checkpoint,
                            use_scale_shift_norm=use_scale_shift_norm,
                            down=True,
                        )
                        if resblock_updown
                        else Downsample(
                            ch,
                            conv_resample,
                            dims=dims,
                            out_channels=out_ch,
                            third_down=time_downup,
                        )
                    )
                )
                ch = out_ch
                input_block_chans.append(ch)

                self._feature_size += ch

        if num_head_channels == -1:
            dim_head = ch // num_heads
        else:
            num_heads = ch // num_head_channels
            dim_head = num_head_channels

        self.middle_block = TimestepEmbedSequential(
            get_resblock(
                merge_factor=merge_factor,
                merge_strategy=merge_strategy,
                video_kernel_size=video_kernel_size,
                ch=ch,
                time_embed_dim=time_embed_dim,
                out_ch=None,
                dropout=dropout,
                dims=dims,
                use_checkpoint=use_checkpoint,
                use_scale_shift_norm=use_scale_shift_norm,
            ),
            get_attention_layer(
                ch,
                num_heads,
                dim_head,
                depth=transformer_depth_middle,
                context_dim=context_dim,
                use_checkpoint=use_checkpoint,
            ),
            get_resblock(
                merge_factor=merge_factor,
                merge_strategy=merge_strategy,
                video_kernel_size=video_kernel_size,
                ch=ch,
                out_ch=None,
                time_embed_dim=time_embed_dim,
                dropout=dropout,
                dims=dims,
                use_checkpoint=use_checkpoint,
                use_scale_shift_norm=use_scale_shift_norm,
            ),
        )
        self._feature_size += ch

        self.output_blocks = nn.ModuleList([])
        for level, mult in list(enumerate(channel_mult))[::-1]:
            for i in range(num_res_blocks + 1):
                ich = input_block_chans.pop()
                layers = [
                    get_resblock(
                        merge_factor=merge_factor,
                        merge_strategy=merge_strategy,
                        video_kernel_size=video_kernel_size,
                        ch=ch + ich,
                        time_embed_dim=time_embed_dim,
                        dropout=dropout,
                        out_ch=model_channels * mult,
                        dims=dims,
                        use_checkpoint=use_checkpoint,
                        use_scale_shift_norm=use_scale_shift_norm,
                    )
                ]
                ch = model_channels * mult
                if ds in attention_resolutions:
                    if num_head_channels == -1:
                        dim_head = ch // num_heads
                    else:
                        num_heads = ch // num_head_channels
                        dim_head = num_head_channels

                    layers.append(
                        get_attention_layer(
                            ch,
                            num_heads,
                            dim_head,
                            depth=transformer_depth[level],
                            context_dim=context_dim,
                            use_checkpoint=use_checkpoint,
                            disabled_sa=False,
                        )
                    )
                if level and i == num_res_blocks:
                    out_ch = ch
                    ds //= 2
                    layers.append(
                        get_resblock(
                            merge_factor=merge_factor,
                            merge_strategy=merge_strategy,
                            video_kernel_size=video_kernel_size,
                            ch=ch,
                            time_embed_dim=time_embed_dim,
                            dropout=dropout,
                            out_ch=out_ch,
                            dims=dims,
                            use_checkpoint=use_checkpoint,
                            use_scale_shift_norm=use_scale_shift_norm,
                            up=True,
                        )
                        if resblock_updown
                        else Upsample(
                            ch,
                            conv_resample,
                            dims=dims,
                            out_channels=out_ch,
                            third_up=time_downup,
                        )
                    )

                self.output_blocks.append(TimestepEmbedSequential(*layers))
                self._feature_size += ch

        self.out = nn.Sequential(
            normalization(ch),
            nn.SiLU(),
            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
        )

    def forward(
        self,
        x: th.Tensor,
        timesteps: th.Tensor,
        context: Optional[th.Tensor] = None,
        y: Optional[th.Tensor] = None,
        time_context: Optional[th.Tensor] = None,
        num_video_frames: Optional[int] = None,
        image_only_indicator: Optional[th.Tensor] = None,
    ):
        assert (y is not None) == (
            self.num_classes is not None
        ), "must specify y if and only if the model is class-conditional -> no, relax this TODO"
        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
        emb = self.time_embed(t_emb)

        ## tbd: check the role of "image_only_indicator"
        num_video_frames = self.num_frames
        image_only_indicator = torch.zeros(
                    x.shape[0]//num_video_frames, num_video_frames
                ).to(x.device) if image_only_indicator is None else image_only_indicator

        if self.num_classes is not None:
            assert y.shape[0] == x.shape[0]
            emb = emb + self.label_emb(y)

        ## x shape: [bt,c,h,w]
        h = x
        hs = []
        for module in self.input_blocks:
            h = module(
                h,
                emb,
                context=context,
                image_only_indicator=image_only_indicator,
                time_context=time_context,
                num_video_frames=num_video_frames,
            )
            hs.append(h)
        h = self.middle_block(
            h,
            emb,
            context=context,
            image_only_indicator=image_only_indicator,
            time_context=time_context,
            num_video_frames=num_video_frames,
        )
        for module in self.output_blocks:
            h = th.cat([h, hs.pop()], dim=1)
            h = module(
                h,
                emb,
                context=context,
                image_only_indicator=image_only_indicator,
                time_context=time_context,
                num_video_frames=num_video_frames,
            )
        h = h.type(x.dtype)
        return self.out(h)