File size: 6,511 Bytes
6e6f7d1
 
b4bbb92
 
 
 
6e6f7d1
 
b4bbb92
6e6f7d1
b4bbb92
 
6e6f7d1
b4bbb92
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6e6f7d1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b4bbb92
6e6f7d1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import random
from typing import Literal, Sequence

import torch
import torch.nn as nn
import torch.nn.functional as F
from accel_hydra.utils.torch import create_mask_from_length
from x_transformers.x_transformers import Encoder

from ..mmdit.modules import ChannelLastConv1d, ConvMLP, MLP


class CLIPEncoder(nn.Module):
    def __init__(
        self,
        video_feat_dim: int,
        embed_dim: int,
    ):
        super().__init__()
        self.mlp = nn.Linear(video_feat_dim, embed_dim)
        self.init_weights()

    def init_weights(self):
        def _init_weights(module):
            if isinstance(module, nn.Linear):
                nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0.)

        self.apply(_init_weights)

    def forward(
        self, clip: torch.Tensor, clip_lengths: Sequence[int], **kwargs
    ):
        device = clip.device
        x = F.normalize(clip, p=2, dim=-1)
        x = self.mlp(x)
        mask = create_mask_from_length(clip_lengths).to(device)
        return {
            "ta_output": x,
            "ta_mask": mask,
            "nta_output": None,
            "nta_mask": None
        }


class CLIPInterpolateVideoEncoder(nn.Module):
    def __init__(
        self,
        video_feat_dim: int,
        embed_dim: int,
        latent_token_rate: int,
        transformer_cfg: dict | None = None,
    ):
        super().__init__()
        self.mlp = nn.Linear(video_feat_dim, embed_dim)
        self.latent_token_rate = latent_token_rate
        self.transformer_enhancer = None
        if transformer_cfg is not None:
            self.transformer_enhancer = Encoder(
                dim=embed_dim,
                **transformer_cfg,
            )
        self.init_weights()

    def init_weights(self):
        def _init_weights(module):
            if isinstance(module, nn.Linear):
                nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0.)

        self.apply(_init_weights)

    def forward(self, clip: torch.Tensor, duration: Sequence[float], **kwargs):
        device = clip.device
        x = F.normalize(clip, p=2, dim=-1)
        x = self.mlp(x)

        target_len = int(max(duration) * self.latent_token_rate)

        x = x.transpose(1, 2)
        x = F.interpolate(x, target_len, mode='nearest-exact')
        x = x.transpose(1, 2)

        latent_length = torch.as_tensor(duration) * self.latent_token_rate
        mask = create_mask_from_length(latent_length).to(device)
        if self.transformer_enhancer is not None:
            x = self.transformer_enhancer(x, mask=mask)

        return {
            "ta_output": x,
            "ta_mask": mask,
            "nta_output": None,
            "nta_mask": None
        }


class CLIPSyncEncoder(nn.Module):
    def __init__(
        self,
        clip_dim: int,
        sync_dim: int,
        embed_dim: int,
        latent_token_rate: int,
        fusion: Literal["add", "concat"] = "add",
        sync_drop_ratio: float = 0.0,
    ) -> None:
        super().__init__()

        self.fusion = fusion
        self.sync_drop_ratio = sync_drop_ratio

        self.clip_proj = nn.Linear(clip_dim, embed_dim)
        self.sync_proj = nn.Linear(sync_dim, embed_dim)
        self.sync_pos_emb = nn.Parameter(torch.zeros((1, 1, 8, sync_dim)))
        if fusion == "concat":
            self.fusion_proj = nn.Linear(embed_dim * 2, embed_dim)

        self.sync_frame_rate = 25
        self.sync_num_frames_per_segment = 16
        self.sync_step_size = 8
        self.sync_downsample_rate = 2
        self.latent_token_rate = latent_token_rate

        self.initialize_weights()

    def initialize_weights(self):
        def _basic_init(module):
            if isinstance(module, nn.Linear):
                torch.nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)

        self.apply(_basic_init)

    def get_sync_seq_len(self, duration):
        num_frames = duration * self.sync_frame_rate
        num_segments = (
            num_frames - self.sync_num_frames_per_segment
        ) // self.sync_step_size + 1
        sync_seq_len = int(
            num_segments * self.sync_num_frames_per_segment /
            self.sync_downsample_rate
        )
        return sync_seq_len

    def forward(
        self, clip: torch.Tensor, sync: torch.Tensor,
        duration: Sequence[float], clip_lengths: Sequence[int],
        sync_lengths: Sequence[int], **kwargs
    ):

        clip = F.normalize(clip, p=2, dim=-1)
        sync = F.normalize(sync, p=2, dim=-1)

        latent_length = (torch.as_tensor(duration) *
                         self.latent_token_rate).int()
        ta_mask = create_mask_from_length(latent_length).to(sync.device)

        bs = clip.shape[0]
        # target_len = int(max(duration) * self.latent_token_rate)
        target_len = latent_length.max()
        sync_seq_len = self.get_sync_seq_len(max(duration))
        # B * num_segments (24) * 8 * 768
        num_sync_segments = sync_seq_len // 8
        sync = sync.view(bs, num_sync_segments, 8, -1) + self.sync_pos_emb
        sync = sync.flatten(1, 2)  # (B, VN, D)

        # extend vf to match x
        clip = self.clip_proj(clip)
        sync = self.sync_proj(sync)  # (B, D, VN)

        # upsample the sync features to match the audio
        sync = sync.transpose(1, 2)  # (B, D, VN)
        sync = F.interpolate(sync, size=target_len, mode='nearest-exact')
        sync = sync.transpose(1, 2)  # (B, N, D)

        if self.training:
            if self.sync_drop_ratio > 0:
                if random.random() < self.sync_drop_ratio:
                    sync = torch.zeros_like(sync)

        # upsample the clip features to match the audio
        clip = clip.transpose(1, 2)  # (B, D, VN)
        clip = F.interpolate(clip, size=target_len, mode='nearest-exact')
        clip = clip.transpose(1, 2)  # (B, N, D)

        if self.fusion == "add":
            ta_output = clip + sync
        elif self.fusion == "concat":
            ta_output = self.fusion_proj(torch.cat([clip, sync], dim=-1))
        else:
            raise ValueError(f"Invalid fusion method: {self.fusion}")

        return {
            "ta_output": ta_output,
            "ta_mask": ta_mask,
            "nta_output": None,
            "nta_mask": None
        }