File size: 40,117 Bytes
5a510e7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
# pylint: disable=R0801
# pylint: disable=C0303

"""
This module contains various transformer blocks for different applications, such as BasicTransformerBlock,
TemporalBasicTransformerBlock, and AudioTemporalBasicTransformerBlock. These blocks are used in various models,
such as GLIGEN, UNet, and others. The transformer blocks implement self-attention, cross-attention, feed-forward
networks, and other related functions.

Functions and classes included in this module are:
- BasicTransformerBlock: A basic transformer block with self-attention, cross-attention, and feed-forward layers.
- TemporalBasicTransformerBlock: A transformer block with additional temporal attention mechanisms for video data.
- AudioTemporalBasicTransformerBlock: A transformer block with additional audio-specific mechanisms for audio data.
- zero_module: A function to zero out the parameters of a given module.

For more information on each specific class and function, please refer to the respective docstrings.
"""

from typing import Any, Dict, List, Optional

import torch
from diffusers.models.attention import (AdaLayerNorm, AdaLayerNormZero,
                                        Attention, FeedForward)
from diffusers.models.embeddings import SinusoidalPositionalEmbedding
from einops import rearrange
from torch import nn


class GatedSelfAttentionDense(nn.Module):
    """
    A gated self-attention dense layer that combines visual features and object features.

    Parameters:
        query_dim (`int`): The number of channels in the query.
        context_dim (`int`): The number of channels in the context.
        n_heads (`int`): The number of heads to use for attention.
        d_head (`int`): The number of channels in each head.
    """

    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
        super().__init__()

        # we need a linear projection since we need cat visual feature and obj feature
        self.linear = nn.Linear(context_dim, query_dim)

        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
        self.ff = FeedForward(query_dim, activation_fn="geglu")

        self.norm1 = nn.LayerNorm(query_dim)
        self.norm2 = nn.LayerNorm(query_dim)

        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))

        self.enabled = True

    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
        """
        Apply the Gated Self-Attention mechanism to the input tensor `x` and object tensor `objs`.

        Args:
            x (torch.Tensor): The input tensor.
            objs (torch.Tensor): The object tensor.

        Returns:
            torch.Tensor: The output tensor after applying Gated Self-Attention.
        """
        if not self.enabled:
            return x

        n_visual = x.shape[1]
        objs = self.linear(objs)

        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))

        return x

class BasicTransformerBlock(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.
        only_cross_attention (`bool`, *optional*):
            Whether to use only cross-attention layers. In this case two cross attention layers are used.
        double_self_attention (`bool`, *optional*):
            Whether to use two self-attention layers. In this case no cross attention layers are used.
        upcast_attention (`bool`, *optional*):
            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
            Whether to use learnable elementwise affine parameters for normalization.
        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
        final_dropout (`bool` *optional*, defaults to False):
            Whether to apply a final dropout after the last feed-forward layer.
        attention_type (`str`, *optional*, defaults to `"default"`):
            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
        positional_embeddings (`str`, *optional*, defaults to `None`):
            The type of positional embeddings to apply to.
        num_positional_embeddings (`int`, *optional*, defaults to `None`):
            The maximum number of positional embeddings to apply.
    """

    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,
        double_self_attention: bool = False,
        upcast_attention: bool = False,
        norm_elementwise_affine: bool = True,
        # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
        norm_type: str = "layer_norm",
        norm_eps: float = 1e-5,
        final_dropout: bool = False,
        attention_type: str = "default",
        positional_embeddings: Optional[str] = None,
        num_positional_embeddings: Optional[int] = None,
    ):
        super().__init__()
        self.only_cross_attention = only_cross_attention

        self.use_ada_layer_norm_zero = (
            num_embeds_ada_norm is not None
        ) and norm_type == "ada_norm_zero"
        self.use_ada_layer_norm = (
            num_embeds_ada_norm is not None
        ) and norm_type == "ada_norm"
        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
        self.use_layer_norm = norm_type == "layer_norm"

        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
            raise ValueError(
                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
            )

        if positional_embeddings and (num_positional_embeddings is None):
            raise ValueError(
                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
            )

        if positional_embeddings == "sinusoidal":
            self.pos_embed = SinusoidalPositionalEmbedding(
                dim, max_seq_length=num_positional_embeddings
            )
        else:
            self.pos_embed = None

        # Define 3 blocks. Each block has its own normalization layer.
        # 1. Self-Attn
        if self.use_ada_layer_norm:
            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
        elif self.use_ada_layer_norm_zero:
            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
        else:
            self.norm1 = nn.LayerNorm(
                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
            )

        self.attn1 = Attention(
            query_dim=dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
            upcast_attention=upcast_attention,
        )

        # 2. Cross-Attn
        if cross_attention_dim is not None or double_self_attention:
            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
            # the second cross attention block.
            self.norm2 = (
                AdaLayerNorm(dim, num_embeds_ada_norm)
                if self.use_ada_layer_norm
                else nn.LayerNorm(
                    dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
                )
            )
            self.attn2 = Attention(
                query_dim=dim,
                cross_attention_dim=(
                    cross_attention_dim if not double_self_attention else None
                ),
                heads=num_attention_heads,
                dim_head=attention_head_dim,
                dropout=dropout,
                bias=attention_bias,
                upcast_attention=upcast_attention,
            )  # is self-attn if encoder_hidden_states is none
        else:
            self.norm2 = None
            self.attn2 = None

        # 3. Feed-forward
        if not self.use_ada_layer_norm_single:
            self.norm3 = nn.LayerNorm(
                dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
            )

        self.ff = FeedForward(
            dim,
            dropout=dropout,
            activation_fn=activation_fn,
            final_dropout=final_dropout,
        )

        # 4. Fuser
        if attention_type in {"gated", "gated-text-image"}:  # Updated line
            self.fuser = GatedSelfAttentionDense(
                dim, cross_attention_dim, num_attention_heads, attention_head_dim
            )

        # 5. Scale-shift for PixArt-Alpha.
        if self.use_ada_layer_norm_single:
            self.scale_shift_table = nn.Parameter(
                torch.randn(6, dim) / dim**0.5)

        # let chunk size default to None
        self._chunk_size = None
        self._chunk_dim = 0

    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
        """
        Sets the chunk size for feed-forward processing in the transformer block.

        Args:
            chunk_size (Optional[int]): The size of the chunks to process in feed-forward layers. 
            If None, the chunk size is set to the maximum possible value.
            dim (int, optional): The dimension along which to split the input tensor into chunks. Defaults to 0.

        Returns:
            None.
        """
        self._chunk_size = chunk_size
        self._chunk_dim = dim

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        timestep: Optional[torch.LongTensor] = None,
        cross_attention_kwargs: Dict[str, Any] = None,
        class_labels: Optional[torch.LongTensor] = None,
    ) -> torch.FloatTensor:
        """
        This function defines the forward pass of the BasicTransformerBlock.

        Args:
            self (BasicTransformerBlock):
                An instance of the BasicTransformerBlock class.
            hidden_states (torch.FloatTensor):
                A tensor containing the hidden states.
            attention_mask (Optional[torch.FloatTensor], optional):
                A tensor containing the attention mask. Defaults to None.
            encoder_hidden_states (Optional[torch.FloatTensor], optional):
                A tensor containing the encoder hidden states. Defaults to None.
            encoder_attention_mask (Optional[torch.FloatTensor], optional):
                A tensor containing the encoder attention mask. Defaults to None.
            timestep (Optional[torch.LongTensor], optional):
                A tensor containing the timesteps. Defaults to None.
            cross_attention_kwargs (Dict[str, Any], optional):
                Additional cross-attention arguments. Defaults to None.
            class_labels (Optional[torch.LongTensor], optional):
                A tensor containing the class labels. Defaults to None.

        Returns:
            torch.FloatTensor:
                A tensor containing the transformed hidden states.
        """
        # Notice that normalization is always applied before the real computation in the following blocks.
        # 0. Self-Attention
        batch_size = hidden_states.shape[0]

        gate_msa = None
        scale_mlp = None
        shift_mlp = None
        gate_mlp = None
        if self.use_ada_layer_norm:
            norm_hidden_states = self.norm1(hidden_states, timestep)
        elif self.use_ada_layer_norm_zero:
            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
            )
        elif self.use_layer_norm:
            norm_hidden_states = self.norm1(hidden_states)
        elif self.use_ada_layer_norm_single:
            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
                self.scale_shift_table[None] +
                timestep.reshape(batch_size, 6, -1)
            ).chunk(6, dim=1)
            norm_hidden_states = self.norm1(hidden_states)
            norm_hidden_states = norm_hidden_states * \
                (1 + scale_msa) + shift_msa
            norm_hidden_states = norm_hidden_states.squeeze(1)
        else:
            raise ValueError("Incorrect norm used")

        if self.pos_embed is not None:
            norm_hidden_states = self.pos_embed(norm_hidden_states)

        # 1. Retrieve lora scale.
        lora_scale = (
            cross_attention_kwargs.get("scale", 1.0)
            if cross_attention_kwargs is not None
            else 1.0
        )

        # 2. Prepare GLIGEN inputs
        cross_attention_kwargs = (
            cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
        )
        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)

        attn_output = self.attn1(
            norm_hidden_states,
            encoder_hidden_states=(
                encoder_hidden_states if self.only_cross_attention else None
            ),
            attention_mask=attention_mask,
            **cross_attention_kwargs,
        )
        if self.use_ada_layer_norm_zero:
            attn_output = gate_msa.unsqueeze(1) * attn_output
        elif self.use_ada_layer_norm_single:
            attn_output = gate_msa * attn_output

        hidden_states = attn_output + hidden_states
        if hidden_states.ndim == 4:
            hidden_states = hidden_states.squeeze(1)

        # 2.5 GLIGEN Control
        if gligen_kwargs is not None:
            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])

        # 3. Cross-Attention
        if self.attn2 is not None:
            if self.use_ada_layer_norm:
                norm_hidden_states = self.norm2(hidden_states, timestep)
            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
                norm_hidden_states = self.norm2(hidden_states)
            elif self.use_ada_layer_norm_single:
                # For PixArt norm2 isn't applied here:
                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
                norm_hidden_states = hidden_states
            else:
                raise ValueError("Incorrect norm")

            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
                norm_hidden_states = self.pos_embed(norm_hidden_states)

            attn_output = self.attn2(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                **cross_attention_kwargs,
            )
            hidden_states = attn_output + hidden_states

        # 4. Feed-forward
        if not self.use_ada_layer_norm_single:
            norm_hidden_states = self.norm3(hidden_states)

        if self.use_ada_layer_norm_zero:
            norm_hidden_states = (
                norm_hidden_states *
                (1 + scale_mlp[:, None]) + shift_mlp[:, None]
            )

        if self.use_ada_layer_norm_single:
            norm_hidden_states = self.norm2(hidden_states)
            norm_hidden_states = norm_hidden_states * \
                (1 + scale_mlp) + shift_mlp

        ff_output = self.ff(norm_hidden_states, scale=lora_scale)

        if self.use_ada_layer_norm_zero:
            ff_output = gate_mlp.unsqueeze(1) * ff_output
        elif self.use_ada_layer_norm_single:
            ff_output = gate_mlp * ff_output

        hidden_states = ff_output + hidden_states
        if hidden_states.ndim == 4:
            hidden_states = hidden_states.squeeze(1)

        return hidden_states


class TemporalBasicTransformerBlock(nn.Module):
    """
    A PyTorch module that extends the BasicTransformerBlock to include temporal attention mechanisms.
    This class is particularly useful for video-related tasks where capturing temporal information within the sequence of frames is necessary.

    Attributes:
        dim (int): The dimension of the input and output embeddings.
        num_attention_heads (int): The number of attention heads in the multi-head self-attention mechanism.
        attention_head_dim (int): The dimension of each attention head.
        dropout (float): The dropout probability for the attention scores.
        cross_attention_dim (Optional[int]): The dimension of the cross-attention mechanism.
        activation_fn (str): The activation function used in the feed-forward layer.
        num_embeds_ada_norm (Optional[int]): The number of embeddings for adaptive normalization.
        attention_bias (bool): If True, uses bias in the attention mechanism.
        only_cross_attention (bool): If True, only uses cross-attention.
        upcast_attention (bool): If True, upcasts the attention mechanism for better performance.
        unet_use_cross_frame_attention (Optional[bool]): If True, uses cross-frame attention in the UNet model.
        unet_use_temporal_attention (Optional[bool]): If True, uses temporal attention in the UNet model.
    """
    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,
        unet_use_cross_frame_attention=None,
        unet_use_temporal_attention=None,
    ):
        """
        The TemporalBasicTransformerBlock class is a PyTorch module that extends the BasicTransformerBlock to include temporal attention mechanisms. 
        This is particularly useful for video-related tasks, where the model needs to capture the temporal information within the sequence of frames. 
        The block consists of self-attention, cross-attention, feed-forward, and temporal attention mechanisms.

            dim (int): The dimension of the input and output embeddings.
            num_attention_heads (int): The number of attention heads in the multi-head self-attention mechanism.
            attention_head_dim (int): The dimension of each attention head.
            dropout (float, optional): The dropout probability for the attention scores. Defaults to 0.0.
            cross_attention_dim (int, optional): The dimension of the cross-attention mechanism. Defaults to None.
            activation_fn (str, optional): The activation function used in the feed-forward layer. Defaults to "geglu".
            num_embeds_ada_norm (int, optional): The number of embeddings for adaptive normalization. Defaults to None.
            attention_bias (bool, optional): If True, uses bias in the attention mechanism. Defaults to False.
            only_cross_attention (bool, optional): If True, only uses cross-attention. Defaults to False.
            upcast_attention (bool, optional): If True, upcasts the attention mechanism for better performance. Defaults to False.
            unet_use_cross_frame_attention (bool, optional): If True, uses cross-frame attention in the UNet model. Defaults to None.
            unet_use_temporal_attention (bool, optional): If True, uses temporal attention in the UNet model. Defaults to None.

        Forward method:
            hidden_states (torch.FloatTensor): The input hidden states.
            encoder_hidden_states (torch.FloatTensor, optional): The encoder hidden states. Defaults to None.
            timestep (torch.LongTensor, optional): The current timestep for the transformer model. Defaults to None.
            attention_mask (torch.FloatTensor, optional): The attention mask for the self-attention mechanism. Defaults to None.
            video_length (int, optional): The length of the video sequence. Defaults to None.

        Returns:
            torch.FloatTensor: The output hidden states after passing through the TemporalBasicTransformerBlock.
        """
        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

        # SC-Attn
        self.attn1 = Attention(
            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 cross_attention_dim is not None:
            self.attn2 = Attention(
                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,
            )
        else:
            self.attn2 = None

        if cross_attention_dim is not None:
            self.norm2 = (
                AdaLayerNorm(dim, num_embeds_ada_norm)
                if self.use_ada_layer_norm
                else nn.LayerNorm(dim)
            )
        else:
            self.norm2 = None

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

        # Temp-Attn
        # assert unet_use_temporal_attention is not None
        if unet_use_temporal_attention is None:
            unet_use_temporal_attention = False
        if unet_use_temporal_attention:
            self.attn_temp = Attention(
                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 forward(
        self,
        hidden_states,
        encoder_hidden_states=None,
        timestep=None,
        attention_mask=None,
        video_length=None,
    ):
        """
        Forward pass for the TemporalBasicTransformerBlock.

        Args:
            hidden_states (torch.FloatTensor): The input hidden states with shape (batch_size, seq_len, dim).
            encoder_hidden_states (torch.FloatTensor, optional): The encoder hidden states with shape (batch_size, src_seq_len, dim).
            timestep (torch.LongTensor, optional): The timestep for the transformer block.
            attention_mask (torch.FloatTensor, optional): The attention mask with shape (batch_size, seq_len, seq_len).
            video_length (int, optional): The length of the video sequence.

        Returns:
            torch.FloatTensor: The output tensor after passing through the transformer block with shape (batch_size, seq_len, dim).
        """
        norm_hidden_states = (
            self.norm1(hidden_states, timestep)
            if self.use_ada_layer_norm
            else self.norm1(hidden_states)
        )

        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.attn2 is not None:
            # Cross-Attention
            norm_hidden_states = (
                self.norm2(hidden_states, timestep)
                if self.use_ada_layer_norm
                else self.norm2(hidden_states)
            )
            hidden_states = (
                self.attn2(
                    norm_hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=attention_mask,
                )
                + hidden_states
            )

        # 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 AudioTemporalBasicTransformerBlock(nn.Module):
    """
    A PyTorch module designed to handle audio data within a transformer framework, including temporal attention mechanisms.

    Attributes:
        dim (int): The dimension of the input and output embeddings.
        num_attention_heads (int): The number of attention heads.
        attention_head_dim (int): The dimension of each attention head.
        dropout (float): The dropout probability.
        cross_attention_dim (Optional[int]): The dimension of the cross-attention mechanism.
        activation_fn (str): The activation function for the feed-forward network.
        num_embeds_ada_norm (Optional[int]): The number of embeddings for adaptive normalization.
        attention_bias (bool): If True, uses bias in the attention mechanism.
        only_cross_attention (bool): If True, only uses cross-attention.
        upcast_attention (bool): If True, upcasts the attention mechanism to float32.
        unet_use_cross_frame_attention (Optional[bool]): If True, uses cross-frame attention in UNet.
        unet_use_temporal_attention (Optional[bool]): If True, uses temporal attention in UNet.
        depth (int): The depth of the transformer block.
        unet_block_name (Optional[str]): The name of the UNet block.
        stack_enable_blocks_name (Optional[List[str]]): The list of enabled blocks in the stack.
        stack_enable_blocks_depth (Optional[List[int]]): The list of depths for the enabled blocks in the stack.
    """
    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,
        unet_use_cross_frame_attention=None,
        unet_use_temporal_attention=None,
        depth=0,
        unet_block_name=None,
        stack_enable_blocks_name: Optional[List[str]] = None,
        stack_enable_blocks_depth: Optional[List[int]] = None,
    ):  
        """
        Initializes the AudioTemporalBasicTransformerBlock module.

        Args:
           dim (int): The dimension of the input and output embeddings.
           num_attention_heads (int): The number of attention heads in the multi-head self-attention mechanism.
           attention_head_dim (int): The dimension of each attention head.
           dropout (float, optional): The dropout probability for the attention mechanism. Defaults to 0.0.
           cross_attention_dim (Optional[int], optional): The dimension of the cross-attention mechanism. Defaults to None.
           activation_fn (str, optional): The activation function to be used in the feed-forward network. Defaults to "geglu".
           num_embeds_ada_norm (Optional[int], optional): The number of embeddings for adaptive normalization. Defaults to None.
           attention_bias (bool, optional): If True, uses bias in the attention mechanism. Defaults to False.
           only_cross_attention (bool, optional): If True, only uses cross-attention. Defaults to False.
           upcast_attention (bool, optional): If True, upcasts the attention mechanism to float32. Defaults to False.
           unet_use_cross_frame_attention (Optional[bool], optional): If True, uses cross-frame attention in UNet. Defaults to None.
           unet_use_temporal_attention (Optional[bool], optional): If True, uses temporal attention in UNet. Defaults to None.
           depth (int, optional): The depth of the transformer block. Defaults to 0.
           unet_block_name (Optional[str], optional): The name of the UNet block. Defaults to None.
           stack_enable_blocks_name (Optional[List[str]], optional): The list of enabled blocks in the stack. Defaults to None.
           stack_enable_blocks_depth (Optional[List[int]], optional): The list of depths for the enabled blocks in the stack. Defaults to None.
        """
        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.unet_block_name = unet_block_name
        self.depth = depth

        zero_conv_full = nn.Conv2d(
            dim, dim, kernel_size=1)
        self.zero_conv_full = zero_module(zero_conv_full)

        zero_conv_face = nn.Conv2d(
            dim, dim, kernel_size=1)
        self.zero_conv_face = zero_module(zero_conv_face)

        zero_conv_lip = nn.Conv2d(
            dim, dim, kernel_size=1)
        self.zero_conv_lip = zero_module(zero_conv_lip)
        # SC-Attn
        self.attn1 = Attention(
            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 cross_attention_dim is not None:
            if (stack_enable_blocks_name is not None and
                stack_enable_blocks_depth is not None and
                self.unet_block_name in stack_enable_blocks_name and
                self.depth in stack_enable_blocks_depth):
                self.attn2_0 = Attention(
                    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,
                )
                self.attn2_1 = Attention(
                    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,
                )
                self.attn2_2 = Attention(
                    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,
                )
                self.attn2 = None

            else:
                self.attn2 = Attention(
                    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,
                )
                self.attn2_0=None
        else:
            self.attn2 = None
            self.attn2_0 = None

        if cross_attention_dim is not None:
            self.norm2 = (
                AdaLayerNorm(dim, num_embeds_ada_norm)
                if self.use_ada_layer_norm
                else nn.LayerNorm(dim)
            )
        else:
            self.norm2 = None

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



    def forward(
        self,
        hidden_states,
        encoder_hidden_states=None,
        timestep=None,
        attention_mask=None,
        full_mask=None,
        face_mask=None,
        lip_mask=None,
        motion_scale=None,
        video_length=None,
    ):
        """
        Forward pass for the AudioTemporalBasicTransformerBlock.

        Args:
            hidden_states (torch.FloatTensor): The input hidden states.
            encoder_hidden_states (torch.FloatTensor, optional): The encoder hidden states. Defaults to None.
            timestep (torch.LongTensor, optional): The timestep for the transformer block. Defaults to None.
            attention_mask (torch.FloatTensor, optional): The attention mask. Defaults to None.
            full_mask (torch.FloatTensor, optional): The full mask. Defaults to None.
            face_mask (torch.FloatTensor, optional): The face mask. Defaults to None.
            lip_mask (torch.FloatTensor, optional): The lip mask. Defaults to None.
            video_length (int, optional): The length of the video. Defaults to None.

        Returns:
            torch.FloatTensor: The output tensor after passing through the AudioTemporalBasicTransformerBlock.
        """
        norm_hidden_states = (
            self.norm1(hidden_states, timestep)
            if self.use_ada_layer_norm
            else self.norm1(hidden_states)
        )

        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.attn2 is not None:
            # Cross-Attention
            norm_hidden_states = (
                self.norm2(hidden_states, timestep)
                if self.use_ada_layer_norm
                else self.norm2(hidden_states)
            )
            hidden_states = self.attn2(
                norm_hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=attention_mask,
            ) + hidden_states

        elif self.attn2_0 is not None:
            norm_hidden_states = (
                self.norm2(hidden_states, timestep)
                if self.use_ada_layer_norm
                else self.norm2(hidden_states)
            )

            level = self.depth
            full_hidden_states = (
                self.attn2_0(
                    norm_hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=attention_mask,
                ) * full_mask[level][:, :, None]
            )
            bz, sz, c = full_hidden_states.shape
            sz_sqrt = int(sz ** 0.5)
            full_hidden_states = full_hidden_states.reshape(
                bz, sz_sqrt, sz_sqrt, c).permute(0, 3, 1, 2)
            full_hidden_states = self.zero_conv_full(full_hidden_states).permute(0, 2, 3, 1).reshape(bz, -1, c)

            face_hidden_state = (
                self.attn2_1(
                    norm_hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=attention_mask,
                ) * face_mask[level][:, :, None]
            )
            face_hidden_state = face_hidden_state.reshape(
                bz, sz_sqrt, sz_sqrt, c).permute(0, 3, 1, 2)
            face_hidden_state = self.zero_conv_face(
                face_hidden_state).permute(0, 2, 3, 1).reshape(bz, -1, c)

            lip_hidden_state = (
                self.attn2_2(
                    norm_hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=attention_mask,
                ) * lip_mask[level][:, :, None]

            ) # [32, 4096, 320]
            lip_hidden_state = lip_hidden_state.reshape(
                bz, sz_sqrt, sz_sqrt, c).permute(0, 3, 1, 2)
            lip_hidden_state = self.zero_conv_lip(
                lip_hidden_state).permute(0, 2, 3, 1).reshape(bz, -1, c)

            if motion_scale is not None:
                hidden_states = (
                    motion_scale[0] * full_hidden_states +
                    motion_scale[1] * face_hidden_state +
                    motion_scale[2] * lip_hidden_state + hidden_states
                )
            else:
                hidden_states = (
                    full_hidden_states +
                    face_hidden_state +
                    lip_hidden_state + hidden_states
                )
        # Feed-forward
        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states

        return hidden_states

def zero_module(module):
    """
    Zeroes out the parameters of a given module.

    Args:
        module (nn.Module): The module whose parameters need to be zeroed out.

    Returns:
        None.
    """
    for p in module.parameters():
        nn.init.zeros_(p)
    return module