File size: 4,990 Bytes
d57fabf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

from typing import Any

import torch
import torch.nn as nn


def _choose_gn_groups(channels: int, max_groups: int = 8) -> int:
    for g in range(min(max_groups, channels), 0, -1):
        if channels % g == 0:
            return g
    return 1


class _ChannelGate(nn.Module):
    def __init__(self, channels: int, reduction: int = 4) -> None:
        super().__init__()
        hidden = max(channels // reduction, 8)
        self.pool = nn.AdaptiveAvgPool3d(1)
        self.fc1 = nn.Conv3d(channels, hidden, kernel_size=1, bias=True)
        self.act = nn.GELU()
        self.fc2 = nn.Conv3d(hidden, channels, kernel_size=1, bias=True)
        self.gate = nn.Sigmoid()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        s = self.pool(x)
        s = self.fc1(s)
        s = self.act(s)
        s = self.fc2(s)
        return x * self.gate(s)


class _FastHyperBlock(nn.Module):
    """
    Efficient RF-expanding residual block.

    Each block contributes one effective k=3 receptive-field expansion stage via
    three parallel branches operating on the same expanded activation:
      - spatial depthwise (1,3,3)
      - temporal depthwise (3,1,1)
      - grouped 3D mixing (3,3,3)
    """

    def __init__(
        self,
        channels: int,
        mid_dim: int,
        mix_groups: int = 6,
        dropout_p: float = 0.02,
        gate_reduction: int = 4,
    ) -> None:
        super().__init__()
        gn1 = _choose_gn_groups(channels)
        gn2 = _choose_gn_groups(mid_dim)
        mix_groups = max(1, min(mix_groups, mid_dim))
        while mid_dim % mix_groups != 0 and mix_groups > 1:
            mix_groups -= 1

        self.pre = nn.Sequential(
            nn.GroupNorm(gn1, channels),
            nn.Conv3d(channels, mid_dim, kernel_size=1, bias=True),
            nn.GELU(),
        )
        self.spatial = nn.Sequential(
            nn.Conv3d(
                mid_dim,
                mid_dim,
                kernel_size=(1, 3, 3),
                padding=(0, 1, 1),
                groups=mid_dim,
                bias=True,
            ),
            nn.GELU(),
        )
        self.temporal = nn.Sequential(
            nn.Conv3d(
                mid_dim,
                mid_dim,
                kernel_size=(3, 1, 1),
                padding=(1, 0, 0),
                groups=mid_dim,
                bias=True,
            ),
            nn.GELU(),
        )
        self.mixed = nn.Sequential(
            nn.GroupNorm(gn2, mid_dim),
            nn.Conv3d(
                mid_dim,
                mid_dim,
                kernel_size=3,
                padding=1,
                groups=mix_groups,
                bias=True,
            ),
            nn.GELU(),
        )
        self.fuse = nn.Sequential(
            nn.Conv3d(mid_dim, channels, kernel_size=1, bias=True),
            nn.GELU(),
        )
        self.gate = _ChannelGate(channels, reduction=gate_reduction)
        self.dropout = nn.Dropout3d(dropout_p) if dropout_p > 0 else nn.Identity()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        h = self.pre(x)
        h = self.spatial(h) + self.temporal(h) + self.mixed(h)
        h = self.fuse(h)
        h = self.gate(h)
        h = self.dropout(h)
        return x + h


class PredecoderFastHyperRF13V1(nn.Module):
    """
    Faster-stronger candidate for model 6 under the public Ising-Decoding API.

    Input / output shape:
        (B, 4, T, D, D) -> (B, 4, T, D, D)
    """

    def __init__(
        self,
        input_channels: int = 4,
        out_channels: int = 4,
        hidden_dim: int = 96,
        mid_dim: int = 144,
        mix_groups: int = 6,
        num_blocks: int = 5,
        stem_kernel_size: int = 3,
        dropout_p: float = 0.02,
        gate_reduction: int = 4,
        **_: Any,
    ) -> None:
        super().__init__()
        pad = stem_kernel_size // 2
        gn = _choose_gn_groups(hidden_dim)
        self.stem = nn.Sequential(
            nn.Conv3d(
                input_channels,
                hidden_dim,
                kernel_size=stem_kernel_size,
                padding=pad,
                bias=True,
            ),
            nn.GroupNorm(gn, hidden_dim),
            nn.GELU(),
        )
        self.blocks = nn.Sequential(*[
            _FastHyperBlock(
                channels=hidden_dim,
                mid_dim=mid_dim,
                mix_groups=mix_groups,
                dropout_p=dropout_p,
                gate_reduction=gate_reduction,
            ) for _ in range(num_blocks)
        ])
        self.head = nn.Sequential(
            nn.GroupNorm(gn, hidden_dim),
            nn.Conv3d(hidden_dim, hidden_dim, kernel_size=1, bias=True),
            nn.GELU(),
            nn.Conv3d(hidden_dim, out_channels, kernel_size=1, bias=True),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.stem(x)
        x = self.blocks(x)
        x = self.head(x)
        return x