File size: 11,783 Bytes
d711508
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import math
from dataclasses import dataclass
from typing import Callable, Optional

import torch
import torch.nn.functional as F
from diffusers.models.attention import Attention, FeedForward
from diffusers.utils import BaseOutput
from diffusers.utils.torch_utils import maybe_allow_in_graph
from einops import rearrange, repeat
from torch import Tensor, nn


def zero_module(module):
    # Zero out the parameters of a module and return it.
    for p in module.parameters():
        p.detach().zero_()
    return module


@dataclass
class TemporalTransformer3DModelOutput(BaseOutput):
    sample: torch.FloatTensor


def get_motion_module(in_channels, motion_module_type: str, motion_module_kwargs: dict):
    if motion_module_type == "Vanilla":
        return VanillaTemporalModule(
            in_channels=in_channels,
            **motion_module_kwargs,
        )
    else:
        raise ValueError


class VanillaTemporalModule(nn.Module):
    def __init__(
        self,
        in_channels,
        num_attention_heads=8,
        num_transformer_block=2,
        attention_block_types=("Temporal_Self", "Temporal_Self"),
        cross_frame_attention_mode=None,
        temporal_position_encoding=False,
        temporal_position_encoding_max_len=24,
        temporal_attention_dim_div=1,
        zero_initialize=True,
    ):
        super().__init__()

        self.temporal_transformer = TemporalTransformer3DModel(
            in_channels=in_channels,
            num_attention_heads=num_attention_heads,
            attention_head_dim=in_channels
            // num_attention_heads
            // temporal_attention_dim_div,
            num_layers=num_transformer_block,
            attention_block_types=attention_block_types,
            cross_frame_attention_mode=cross_frame_attention_mode,
            temporal_position_encoding=temporal_position_encoding,
            temporal_position_encoding_max_len=temporal_position_encoding_max_len,
        )

        if zero_initialize:
            self.temporal_transformer.proj_out = zero_module(
                self.temporal_transformer.proj_out
            )
        self.skip_temporal_layers = False  # Whether to skip temporal layer

    def forward(
        self,
        input_tensor,
        temb,
        encoder_hidden_states,
        attention_mask=None,
        anchor_frame_idx=None,
    ):
        if self.skip_temporal_layers is True:
            return input_tensor

        hidden_states = input_tensor
        hidden_states = self.temporal_transformer(
            hidden_states, encoder_hidden_states, attention_mask
        )

        output = hidden_states
        return output


@maybe_allow_in_graph
class TemporalTransformer3DModel(nn.Module):
    def __init__(
        self,
        in_channels,
        num_attention_heads,
        attention_head_dim,
        num_layers,
        attention_block_types=(
            "Temporal_Self",
            "Temporal_Self",
        ),
        dropout=0.0,
        norm_num_groups=32,
        cross_attention_dim=768,
        activation_fn="geglu",
        attention_bias=False,
        upcast_attention=False,
        cross_frame_attention_mode=None,
        temporal_position_encoding=False,
        temporal_position_encoding_max_len=24,
    ):
        super().__init__()

        inner_dim = num_attention_heads * attention_head_dim

        self.norm = torch.nn.GroupNorm(
            num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
        )
        self.proj_in = nn.Linear(in_channels, inner_dim)

        self.transformer_blocks = nn.ModuleList(
            [
                TemporalTransformerBlock(
                    dim=inner_dim,
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=attention_head_dim,
                    attention_block_types=attention_block_types,
                    dropout=dropout,
                    norm_num_groups=norm_num_groups,
                    cross_attention_dim=cross_attention_dim,
                    activation_fn=activation_fn,
                    attention_bias=attention_bias,
                    upcast_attention=upcast_attention,
                    cross_frame_attention_mode=cross_frame_attention_mode,
                    temporal_position_encoding=temporal_position_encoding,
                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
                )
                for d in range(num_layers)
            ]
        )
        self.proj_out = nn.Linear(inner_dim, in_channels)

    def forward(
        self,
        hidden_states: Tensor,
        encoder_hidden_states: Optional[Tensor] = None,
        attention_mask: Optional[Tensor] = None,
    ):
        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)
        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)

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

        # output
        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)

        return output


@maybe_allow_in_graph
class TemporalTransformerBlock(nn.Module):
    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        attention_block_types=(
            "Temporal_Self",
            "Temporal_Self",
        ),
        dropout=0.0,
        norm_num_groups: int = 32,
        cross_attention_dim: int = 768,
        activation_fn: str = "geglu",
        attention_bias: bool = False,
        upcast_attention: bool = False,
        cross_frame_attention_mode=None,
        temporal_position_encoding: bool = False,
        temporal_position_encoding_max_len: int = 24,
    ):
        super().__init__()

        attention_blocks = []
        norms = []

        for block_name in attention_block_types:
            attention_blocks.append(
                VersatileAttention(
                    attention_mode=block_name.split("_")[0],
                    cross_attention_dim=(
                        cross_attention_dim if block_name.endswith("_Cross") else None
                    ),
                    query_dim=dim,
                    heads=num_attention_heads,
                    dim_head=attention_head_dim,
                    dropout=dropout,
                    bias=attention_bias,
                    upcast_attention=upcast_attention,
                    cross_frame_attention_mode=cross_frame_attention_mode,
                    temporal_position_encoding=temporal_position_encoding,
                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
                )
            )
            norms.append(nn.LayerNorm(dim))

        self.attention_blocks = nn.ModuleList(attention_blocks)
        self.norms = nn.ModuleList(norms)

        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
        self.ff_norm = nn.LayerNorm(dim)

    def forward(
        self,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
        video_length=None,
    ):
        for attention_block, norm in zip(self.attention_blocks, self.norms):
            norm_hidden_states = norm(hidden_states)
            hidden_states = (
                attention_block(
                    norm_hidden_states,
                    encoder_hidden_states=(
                        encoder_hidden_states
                        if attention_block.is_cross_attention
                        else None
                    ),
                    video_length=video_length,
                )
                + hidden_states
            )

        hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states

        output = hidden_states
        return output


class PositionalEncoding(nn.Module):
    def __init__(self, d_model, dropout: float = 0.0, max_len: int = 24):
        super().__init__()
        self.dropout: nn.Module = nn.Dropout(p=dropout)
        position = torch.arange(max_len).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model)
        )
        pe: Tensor = torch.zeros(1, max_len, d_model)
        pe[0, :, 0::2] = torch.sin(position * div_term)
        pe[0, :, 1::2] = torch.cos(position * div_term)
        self.register_buffer("pe", pe)

    def forward(self, x: Tensor):
        x = x + self.pe[:, : x.size(1)]
        return self.dropout(x)


@maybe_allow_in_graph
class VersatileAttention(Attention):
    def __init__(
        self,
        attention_mode: str = None,
        cross_frame_attention_mode: Optional[str] = None,
        temporal_position_encoding: bool = False,
        temporal_position_encoding_max_len: int = 24,
        *args,
        **kwargs,
    ):
        super().__init__(*args, **kwargs)
        if attention_mode.lower() != "temporal":
            raise ValueError(f"Attention mode {attention_mode} is not supported.")

        self.attention_mode = attention_mode
        self.is_cross_attention = kwargs["cross_attention_dim"] is not None

        self.pos_encoder = (
            PositionalEncoding(
                kwargs["query_dim"],
                dropout=0.0,
                max_len=temporal_position_encoding_max_len,
            )
            if (temporal_position_encoding and attention_mode == "Temporal")
            else None
        )

    def extra_repr(self):
        return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}"

    def forward(
        self,
        hidden_states: Tensor,
        encoder_hidden_states=None,
        attention_mask=None,
        video_length=None,
    ):
        if self.attention_mode == "Temporal":
            d = hidden_states.shape[1]
            hidden_states = rearrange(
                hidden_states, "(b f) d c -> (b d) f c", f=video_length
            )

            if self.pos_encoder is not None:
                hidden_states = self.pos_encoder(hidden_states)

            encoder_hidden_states = (
                repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d)
                if encoder_hidden_states is not None
                else encoder_hidden_states
            )
        else:
            raise NotImplementedError

        # attention processor makes this easy so that's nice
        hidden_states = self.processor(
            self, hidden_states, encoder_hidden_states, attention_mask
        )

        if self.attention_mode == "Temporal":
            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)

        return hidden_states

    def set_use_memory_efficient_attention_xformers(
        self,
        use_memory_efficient_attention_xformers: bool,
        attention_op: Optional[Callable] = None,
    ):
        return None