File size: 20,281 Bytes
24c345c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py

from dataclasses import dataclass
from turtle import forward
from typing import Optional

import torch
import torch.nn.functional as F
from torch import nn

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.modeling_utils import ModelMixin
from diffusers.utils import BaseOutput
from diffusers.utils.import_utils import is_xformers_available
from diffusers.models.attention import CrossAttention, FeedForward, AdaLayerNorm

from einops import rearrange, repeat
from .utils import zero_module


@dataclass
class Transformer3DModelOutput(BaseOutput):
    sample: torch.FloatTensor


if is_xformers_available():
    import xformers
    import xformers.ops
else:
    xformers = None


class Transformer3DModel(ModelMixin, ConfigMixin):
    @register_to_config
    def __init__(
        self,
        num_attention_heads: int = 16,
        attention_head_dim: int = 88,
        in_channels: Optional[int] = None,
        num_layers: int = 1,
        dropout: float = 0.0,
        norm_num_groups: int = 32,
        cross_attention_dim: Optional[int] = None,
        attention_bias: bool = False,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        use_linear_projection: bool = False,
        only_cross_attention: bool = False,
        upcast_attention: bool = False,
        use_motion_module: bool = False,
        unet_use_cross_frame_attention=None,
        unet_use_temporal_attention=None,
        add_audio_layer=False,
        audio_condition_method="cross_attn",
        custom_audio_layer: bool = False,
    ):
        super().__init__()
        self.use_linear_projection = use_linear_projection
        self.num_attention_heads = num_attention_heads
        self.attention_head_dim = attention_head_dim
        inner_dim = num_attention_heads * attention_head_dim

        # Define input layers
        self.in_channels = in_channels

        self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
        if use_linear_projection:
            self.proj_in = nn.Linear(in_channels, inner_dim)
        else:
            self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)

        if not custom_audio_layer:
            # Define transformers blocks
            self.transformer_blocks = nn.ModuleList(
                [
                    BasicTransformerBlock(
                        inner_dim,
                        num_attention_heads,
                        attention_head_dim,
                        dropout=dropout,
                        cross_attention_dim=cross_attention_dim,
                        activation_fn=activation_fn,
                        num_embeds_ada_norm=num_embeds_ada_norm,
                        attention_bias=attention_bias,
                        only_cross_attention=only_cross_attention,
                        upcast_attention=upcast_attention,
                        use_motion_module=use_motion_module,
                        unet_use_cross_frame_attention=unet_use_cross_frame_attention,
                        unet_use_temporal_attention=unet_use_temporal_attention,
                        add_audio_layer=add_audio_layer,
                        custom_audio_layer=custom_audio_layer,
                        audio_condition_method=audio_condition_method,
                    )
                    for d in range(num_layers)
                ]
            )
        else:
            self.transformer_blocks = nn.ModuleList(
                [
                    AudioTransformerBlock(
                        inner_dim,
                        num_attention_heads,
                        attention_head_dim,
                        dropout=dropout,
                        cross_attention_dim=cross_attention_dim,
                        activation_fn=activation_fn,
                        num_embeds_ada_norm=num_embeds_ada_norm,
                        attention_bias=attention_bias,
                        only_cross_attention=only_cross_attention,
                        upcast_attention=upcast_attention,
                        use_motion_module=use_motion_module,
                        unet_use_cross_frame_attention=unet_use_cross_frame_attention,
                        unet_use_temporal_attention=unet_use_temporal_attention,
                        add_audio_layer=add_audio_layer,
                    )
                    for d in range(num_layers)
                ]
            )

        # 4. Define output layers
        if use_linear_projection:
            self.proj_out = nn.Linear(in_channels, inner_dim)
        else:
            self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)

        if custom_audio_layer:
            self.proj_out = zero_module(self.proj_out)

    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True):
        # Input
        assert hidden_states.dim() == 5, f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
        video_length = hidden_states.shape[2]
        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")

        # No need to do this for audio input, because different audio samples are independent
        # encoder_hidden_states = repeat(encoder_hidden_states, 'b n c -> (b f) n c', f=video_length)

        batch, channel, height, weight = hidden_states.shape
        residual = hidden_states

        hidden_states = self.norm(hidden_states)
        if not self.use_linear_projection:
            hidden_states = self.proj_in(hidden_states)
            inner_dim = hidden_states.shape[1]
            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
        else:
            inner_dim = hidden_states.shape[1]
            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
            hidden_states = self.proj_in(hidden_states)

        # Blocks
        for block in self.transformer_blocks:
            hidden_states = block(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                timestep=timestep,
                video_length=video_length,
            )

        # Output
        if not self.use_linear_projection:
            hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
            hidden_states = self.proj_out(hidden_states)
        else:
            hidden_states = self.proj_out(hidden_states)
            hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()

        output = hidden_states + residual

        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
        if not return_dict:
            return (output,)

        return Transformer3DModelOutput(sample=output)


class BasicTransformerBlock(nn.Module):
    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        dropout=0.0,
        cross_attention_dim: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        attention_bias: bool = False,
        only_cross_attention: bool = False,
        upcast_attention: bool = False,
        use_motion_module: bool = False,
        unet_use_cross_frame_attention=None,
        unet_use_temporal_attention=None,
        add_audio_layer=False,
        custom_audio_layer=False,
        audio_condition_method="cross_attn",
    ):
        super().__init__()
        self.only_cross_attention = only_cross_attention
        self.use_ada_layer_norm = num_embeds_ada_norm is not None
        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
        self.unet_use_temporal_attention = unet_use_temporal_attention
        self.use_motion_module = use_motion_module
        self.add_audio_layer = add_audio_layer

        # SC-Attn
        assert unet_use_cross_frame_attention is not None
        if unet_use_cross_frame_attention:
            raise NotImplementedError("SparseCausalAttention2D not implemented yet.")
        else:
            self.attn1 = CrossAttention(
                query_dim=dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                upcast_attention=upcast_attention,
            )
        self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)

        # Cross-Attn
        if add_audio_layer and audio_condition_method == "cross_attn" and not custom_audio_layer:
            self.audio_cross_attn = AudioCrossAttn(
                dim=dim,
                cross_attention_dim=cross_attention_dim,
                num_attention_heads=num_attention_heads,
                attention_head_dim=attention_head_dim,
                dropout=dropout,
                attention_bias=attention_bias,
                upcast_attention=upcast_attention,
                num_embeds_ada_norm=num_embeds_ada_norm,
                use_ada_layer_norm=self.use_ada_layer_norm,
                zero_proj_out=False,
            )
        else:
            self.audio_cross_attn = None

        # Feed-forward
        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
        self.norm3 = nn.LayerNorm(dim)

        # Temp-Attn
        assert unet_use_temporal_attention is not None
        if unet_use_temporal_attention:
            self.attn_temp = CrossAttention(
                query_dim=dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                upcast_attention=upcast_attention,
            )
            nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
            self.norm_temp = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)

    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
        if not is_xformers_available():
            print("Here is how to install it")
            raise ModuleNotFoundError(
                "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
                " xformers",
                name="xformers",
            )
        elif not torch.cuda.is_available():
            raise ValueError(
                "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only"
                " available for GPU "
            )
        else:
            try:
                # Make sure we can run the memory efficient attention
                _ = xformers.ops.memory_efficient_attention(
                    torch.randn((1, 2, 40), device="cuda"),
                    torch.randn((1, 2, 40), device="cuda"),
                    torch.randn((1, 2, 40), device="cuda"),
                )
            except Exception as e:
                raise e
            self.attn1._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
            if self.audio_cross_attn is not None:
                self.audio_cross_attn.attn._use_memory_efficient_attention_xformers = (
                    use_memory_efficient_attention_xformers
                )
            # self.attn_temp._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers

    def forward(
        self, hidden_states, encoder_hidden_states=None, timestep=None, attention_mask=None, video_length=None
    ):
        # SparseCausal-Attention
        norm_hidden_states = (
            self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states)
        )

        # if self.only_cross_attention:
        #     hidden_states = (
        #         self.attn1(norm_hidden_states, encoder_hidden_states, attention_mask=attention_mask) + hidden_states
        #     )
        # else:
        #     hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask, video_length=video_length) + hidden_states

        # pdb.set_trace()
        if self.unet_use_cross_frame_attention:
            hidden_states = (
                self.attn1(norm_hidden_states, attention_mask=attention_mask, video_length=video_length)
                + hidden_states
            )
        else:
            hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask) + hidden_states

        if self.audio_cross_attn is not None and encoder_hidden_states is not None:
            hidden_states = self.audio_cross_attn(
                hidden_states, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask
            )

        # Feed-forward
        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states

        # Temporal-Attention
        if self.unet_use_temporal_attention:
            d = hidden_states.shape[1]
            hidden_states = rearrange(hidden_states, "(b f) d c -> (b d) f c", f=video_length)
            norm_hidden_states = (
                self.norm_temp(hidden_states, timestep) if self.use_ada_layer_norm else self.norm_temp(hidden_states)
            )
            hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)

        return hidden_states


class AudioTransformerBlock(nn.Module):
    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        dropout=0.0,
        cross_attention_dim: Optional[int] = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: Optional[int] = None,
        attention_bias: bool = False,
        only_cross_attention: bool = False,
        upcast_attention: bool = False,
        use_motion_module: bool = False,
        unet_use_cross_frame_attention=None,
        unet_use_temporal_attention=None,
        add_audio_layer=False,
    ):
        super().__init__()
        self.only_cross_attention = only_cross_attention
        self.use_ada_layer_norm = num_embeds_ada_norm is not None
        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
        self.unet_use_temporal_attention = unet_use_temporal_attention
        self.use_motion_module = use_motion_module
        self.add_audio_layer = add_audio_layer

        # SC-Attn
        assert unet_use_cross_frame_attention is not None
        if unet_use_cross_frame_attention:
            raise NotImplementedError("SparseCausalAttention2D not implemented yet.")
        else:
            self.attn1 = CrossAttention(
                query_dim=dim,
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                upcast_attention=upcast_attention,
            )
        self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)

        self.audio_cross_attn = AudioCrossAttn(
            dim=dim,
            cross_attention_dim=cross_attention_dim,
            num_attention_heads=num_attention_heads,
            attention_head_dim=attention_head_dim,
            dropout=dropout,
            attention_bias=attention_bias,
            upcast_attention=upcast_attention,
            num_embeds_ada_norm=num_embeds_ada_norm,
            use_ada_layer_norm=self.use_ada_layer_norm,
            zero_proj_out=False,
        )

        # Feed-forward
        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
        self.norm3 = nn.LayerNorm(dim)

    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
        if not is_xformers_available():
            print("Here is how to install it")
            raise ModuleNotFoundError(
                "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
                " xformers",
                name="xformers",
            )
        elif not torch.cuda.is_available():
            raise ValueError(
                "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only"
                " available for GPU "
            )
        else:
            try:
                # Make sure we can run the memory efficient attention
                _ = xformers.ops.memory_efficient_attention(
                    torch.randn((1, 2, 40), device="cuda"),
                    torch.randn((1, 2, 40), device="cuda"),
                    torch.randn((1, 2, 40), device="cuda"),
                )
            except Exception as e:
                raise e
            self.attn1._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
            if self.audio_cross_attn is not None:
                self.audio_cross_attn.attn._use_memory_efficient_attention_xformers = (
                    use_memory_efficient_attention_xformers
                )
            # self.attn_temp._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers

    def forward(
        self, hidden_states, encoder_hidden_states=None, timestep=None, attention_mask=None, video_length=None
    ):
        # SparseCausal-Attention
        norm_hidden_states = (
            self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states)
        )

        # pdb.set_trace()
        if self.unet_use_cross_frame_attention:
            hidden_states = (
                self.attn1(norm_hidden_states, attention_mask=attention_mask, video_length=video_length)
                + hidden_states
            )
        else:
            hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask) + hidden_states

        if self.audio_cross_attn is not None and encoder_hidden_states is not None:
            hidden_states = self.audio_cross_attn(
                hidden_states, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask
            )

        # Feed-forward
        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states

        return hidden_states


class AudioCrossAttn(nn.Module):
    def __init__(
        self,
        dim,
        cross_attention_dim,
        num_attention_heads,
        attention_head_dim,
        dropout,
        attention_bias,
        upcast_attention,
        num_embeds_ada_norm,
        use_ada_layer_norm,
        zero_proj_out=False,
    ):
        super().__init__()

        self.norm = AdaLayerNorm(dim, num_embeds_ada_norm) if use_ada_layer_norm else nn.LayerNorm(dim)
        self.attn = CrossAttention(
            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,
        )

        if zero_proj_out:
            self.proj_out = zero_module(nn.Linear(dim, dim))

        self.zero_proj_out = zero_proj_out
        self.use_ada_layer_norm = use_ada_layer_norm

    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, attention_mask=None):
        previous_hidden_states = hidden_states
        hidden_states = self.norm(hidden_states, timestep) if self.use_ada_layer_norm else self.norm(hidden_states)

        if encoder_hidden_states.dim() == 4:
            encoder_hidden_states = rearrange(encoder_hidden_states, "b f n d -> (b f) n d")
            
        hidden_states = self.attn(
            hidden_states, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask
        )

        if self.zero_proj_out:
            hidden_states = self.proj_out(hidden_states)
        return hidden_states + previous_hidden_states