File size: 6,105 Bytes
03a856a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from dataclasses import dataclass
from typing import Optional

import torch
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models import ModelMixin
from diffusers.utils import BaseOutput
from diffusers.utils.import_utils import is_xformers_available
from einops import rearrange, repeat
from torch import nn

from .attention import TemporalBasicTransformerBlock


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


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


class Transformer3DModel(ModelMixin, ConfigMixin):
    _supports_gradient_checkpointing = True

    @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,
        unet_use_cross_frame_attention=None,
        unet_use_temporal_attention=None,
    ):
        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
            )

        # Define transformers blocks
        self.transformer_blocks = nn.ModuleList(
            [
                TemporalBasicTransformerBlock(
                    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,
                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
                    unet_use_temporal_attention=unet_use_temporal_attention,
                )
                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
            )

        self.gradient_checkpointing = False

    def _set_gradient_checkpointing(self, module, value=False):
        if hasattr(module, "gradient_checkpointing"):
            module.gradient_checkpointing = value

    def forward(
        self,
        hidden_states,
        encoder_hidden_states=None,
        audio_cond_fea=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")
        if audio_cond_fea.shape[0] != hidden_states.shape[0]:
            if len(audio_cond_fea.shape) == 3:
                audio_cond_fea = rearrange(
                    audio_cond_fea, "b f c -> (b f) 1 c"
                )
            elif len(audio_cond_fea.shape) == 4:
                audio_cond_fea = rearrange(
                    audio_cond_fea, "b f n c -> (b f) n c"
                )
        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 i, block in enumerate(self.transformer_blocks):
            hidden_states = block(
                hidden_states,
                encoder_hidden_states=encoder_hidden_states,
                audio_cond_fea=audio_cond_fea,
                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)