File size: 8,188 Bytes
a342aa8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from dataclasses import dataclass, field

import torch
import torch.nn as nn

from seva.modules.layers import (
    Downsample,
    GroupNorm32,
    ResBlock,
    TimestepEmbedSequential,
    Upsample,
    timestep_embedding,
)
from seva.modules.transformer import MultiviewTransformer


@dataclass
class SevaParams(object):
    in_channels: int = 11
    model_channels: int = 320
    out_channels: int = 4
    num_frames: int = 21
    num_res_blocks: int = 2
    attention_resolutions: list[int] = field(default_factory=lambda: [4, 2, 1])
    channel_mult: list[int] = field(default_factory=lambda: [1, 2, 4, 4])
    num_head_channels: int = 64
    transformer_depth: list[int] = field(default_factory=lambda: [1, 1, 1, 1])
    context_dim: int = 1024
    dense_in_channels: int = 6
    dropout: float = 0.0
    unflatten_names: list[str] = field(
        default_factory=lambda: ["middle_ds8", "output_ds4", "output_ds2"]
    )

    def __post_init__(self):
        assert len(self.channel_mult) == len(self.transformer_depth)


class Seva(nn.Module):
    def __init__(self, params: SevaParams) -> None:
        super().__init__()
        self.params = params
        self.model_channels = params.model_channels
        self.out_channels = params.out_channels
        self.num_head_channels = params.num_head_channels

        time_embed_dim = params.model_channels * 4
        self.time_embed = nn.Sequential(
            nn.Linear(params.model_channels, time_embed_dim),
            nn.SiLU(),
            nn.Linear(time_embed_dim, time_embed_dim),
        )

        self.input_blocks = nn.ModuleList(
            [
                TimestepEmbedSequential(
                    nn.Conv2d(params.in_channels, params.model_channels, 3, padding=1)
                )
            ]
        )
        self._feature_size = params.model_channels
        input_block_chans = [params.model_channels]
        ch = params.model_channels
        ds = 1
        for level, mult in enumerate(params.channel_mult):
            for _ in range(params.num_res_blocks):
                input_layers: list[ResBlock | MultiviewTransformer | Downsample] = [
                    ResBlock(
                        channels=ch,
                        emb_channels=time_embed_dim,
                        out_channels=mult * params.model_channels,
                        dense_in_channels=params.dense_in_channels,
                        dropout=params.dropout,
                    )
                ]
                ch = mult * params.model_channels
                if ds in params.attention_resolutions:
                    num_heads = ch // params.num_head_channels
                    dim_head = params.num_head_channels
                    input_layers.append(
                        MultiviewTransformer(
                            ch,
                            num_heads,
                            dim_head,
                            name=f"input_ds{ds}",
                            depth=params.transformer_depth[level],
                            context_dim=params.context_dim,
                            unflatten_names=params.unflatten_names,
                        )
                    )
                self.input_blocks.append(TimestepEmbedSequential(*input_layers))
                self._feature_size += ch
                input_block_chans.append(ch)
            if level != len(params.channel_mult) - 1:
                ds *= 2
                out_ch = ch
                self.input_blocks.append(
                    TimestepEmbedSequential(Downsample(ch, out_channels=out_ch))
                )
                ch = out_ch
                input_block_chans.append(ch)
                self._feature_size += ch

        num_heads = ch // params.num_head_channels
        dim_head = params.num_head_channels

        self.middle_block = TimestepEmbedSequential(
            ResBlock(
                channels=ch,
                emb_channels=time_embed_dim,
                out_channels=None,
                dense_in_channels=params.dense_in_channels,
                dropout=params.dropout,
            ),
            MultiviewTransformer(
                ch,
                num_heads,
                dim_head,
                name=f"middle_ds{ds}",
                depth=params.transformer_depth[-1],
                context_dim=params.context_dim,
                unflatten_names=params.unflatten_names,
            ),
            ResBlock(
                channels=ch,
                emb_channels=time_embed_dim,
                out_channels=None,
                dense_in_channels=params.dense_in_channels,
                dropout=params.dropout,
            ),
        )
        self._feature_size += ch

        self.output_blocks = nn.ModuleList([])
        for level, mult in list(enumerate(params.channel_mult))[::-1]:
            for i in range(params.num_res_blocks + 1):
                ich = input_block_chans.pop()
                output_layers: list[ResBlock | MultiviewTransformer | Upsample] = [
                    ResBlock(
                        channels=ch + ich,
                        emb_channels=time_embed_dim,
                        out_channels=params.model_channels * mult,
                        dense_in_channels=params.dense_in_channels,
                        dropout=params.dropout,
                    )
                ]
                ch = params.model_channels * mult
                if ds in params.attention_resolutions:
                    num_heads = ch // params.num_head_channels
                    dim_head = params.num_head_channels

                    output_layers.append(
                        MultiviewTransformer(
                            ch,
                            num_heads,
                            dim_head,
                            name=f"output_ds{ds}",
                            depth=params.transformer_depth[level],
                            context_dim=params.context_dim,
                            unflatten_names=params.unflatten_names,
                        )
                    )
                if level and i == params.num_res_blocks:
                    out_ch = ch
                    ds //= 2
                    output_layers.append(Upsample(ch, out_ch))
                self.output_blocks.append(TimestepEmbedSequential(*output_layers))
                self._feature_size += ch

        self.out = nn.Sequential(
            GroupNorm32(32, ch),
            nn.SiLU(),
            nn.Conv2d(self.model_channels, params.out_channels, 3, padding=1),
        )

    def forward(
        self,
        x: torch.Tensor,
        t: torch.Tensor,
        y: torch.Tensor,
        dense_y: torch.Tensor,
        num_frames: int | None = None,
    ) -> torch.Tensor:
        num_frames = num_frames or self.params.num_frames
        t_emb = timestep_embedding(t, self.model_channels)
        t_emb = self.time_embed(t_emb)

        hs = []
        h = x
        for module in self.input_blocks:
            h = module(
                h,
                emb=t_emb,
                context=y,
                dense_emb=dense_y,
                num_frames=num_frames,
            )
            hs.append(h)
        h = self.middle_block(
            h,
            emb=t_emb,
            context=y,
            dense_emb=dense_y,
            num_frames=num_frames,
        )
        for module in self.output_blocks:
            h = torch.cat([h, hs.pop()], dim=1)
            h = module(
                h,
                emb=t_emb,
                context=y,
                dense_emb=dense_y,
                num_frames=num_frames,
            )
        h = h.type(x.dtype)
        return self.out(h)


class SGMWrapper(nn.Module):
    def __init__(self, module: Seva):
        super().__init__()
        self.module = module

    def forward(
        self, x: torch.Tensor, t: torch.Tensor, c: dict, **kwargs
    ) -> torch.Tensor:
        x = torch.cat((x, c.get("concat", torch.Tensor([]).type_as(x))), dim=1)
        return self.module(
            x,
            t=t,
            y=c["crossattn"],
            dense_y=c["dense_vector"],
            **kwargs,
        )