File size: 10,568 Bytes
8d11d43
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py

from dataclasses import dataclass
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.models import ModelMixin
from diffusers.utils import BaseOutput
from diffusers.models.attention import FeedForward, AdaLayerNorm

from einops import rearrange, repeat


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


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,
        add_audio_layer=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)

        # 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,
                    upcast_attention=upcast_attention,
                    add_audio_layer=add_audio_layer,
                )
                for d in range(num_layers)
            ]
        )

        # 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)

    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")

        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,
        upcast_attention: bool = False,
        add_audio_layer=False,
    ):
        super().__init__()
        self.use_ada_layer_norm = num_embeds_ada_norm is not None
        self.add_audio_layer = add_audio_layer

        self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
        self.attn1 = Attention(
            query_dim=dim,
            heads=num_attention_heads,
            dim_head=attention_head_dim,
            dropout=dropout,
            bias=attention_bias,
            upcast_attention=upcast_attention,
        )

        # Cross-attn
        if add_audio_layer:
            self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
            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

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

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

        hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask) + hidden_states

        if self.attn2 is not None and encoder_hidden_states is not None:
            if encoder_hidden_states.dim() == 4:
                encoder_hidden_states = rearrange(encoder_hidden_states, "b f s d -> (b f) s d")
            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

        return hidden_states


class Attention(nn.Module):
    def __init__(
        self,
        query_dim: int,
        cross_attention_dim: Optional[int] = None,
        heads: int = 8,
        dim_head: int = 64,
        dropout: float = 0.0,
        bias=False,
        upcast_attention: bool = False,
        upcast_softmax: bool = False,
        norm_num_groups: Optional[int] = None,
    ):
        super().__init__()
        inner_dim = dim_head * heads
        cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
        self.upcast_attention = upcast_attention
        self.upcast_softmax = upcast_softmax

        self.scale = dim_head**-0.5

        self.heads = heads

        if norm_num_groups is not None:
            self.group_norm = nn.GroupNorm(num_channels=inner_dim, num_groups=norm_num_groups, eps=1e-5, affine=True)
        else:
            self.group_norm = None

        self.to_q = nn.Linear(query_dim, inner_dim, bias=bias)
        self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=bias)
        self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=bias)

        self.to_out = nn.ModuleList([])
        self.to_out.append(nn.Linear(inner_dim, query_dim))
        self.to_out.append(nn.Dropout(dropout))

    def split_heads(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        tensor = tensor.reshape(batch_size, seq_len, self.heads, dim // self.heads)
        tensor = tensor.permute(0, 2, 1, 3)
        return tensor

    def concat_heads(self, tensor):
        batch_size, heads, seq_len, head_dim = tensor.shape
        tensor = tensor.permute(0, 2, 1, 3)
        tensor = tensor.reshape(batch_size, seq_len, heads * head_dim)
        return tensor

    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
        if self.group_norm is not None:
            hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

        query = self.to_q(hidden_states)
        query = self.split_heads(query)

        encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
        key = self.to_k(encoder_hidden_states)
        value = self.to_v(encoder_hidden_states)

        key = self.split_heads(key)
        value = self.split_heads(value)

        if attention_mask is not None:
            if attention_mask.shape[-1] != query.shape[1]:
                target_length = query.shape[1]
                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
                attention_mask = attention_mask.repeat_interleave(self.heads, dim=0)

        # Use PyTorch native implementation of FlashAttention-2
        hidden_states = F.scaled_dot_product_attention(query, key, value, attn_mask=attention_mask)

        hidden_states = self.concat_heads(hidden_states)

        # linear proj
        hidden_states = self.to_out[0](hidden_states)

        # dropout
        hidden_states = self.to_out[1](hidden_states)
        return hidden_states