nicholasLane commited on
Commit
4505fe3
·
1 Parent(s): a141401

add neosr ref

Browse files
Files changed (1) hide show
  1. NEOSR/grdfnet_arch.py +140 -0
NEOSR/grdfnet_arch.py ADDED
@@ -0,0 +1,140 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch import nn
3
+
4
+ from neosr.archs.arch_util import net_opt
5
+ from neosr.utils.registry import ARCH_REGISTRY
6
+
7
+ upscale, __ = net_opt()
8
+
9
+
10
+ def conv3x3(in_channels, out_channels, bias=True):
11
+ return nn.Conv2d(
12
+ in_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=bias
13
+ )
14
+
15
+
16
+ class LWGRB(nn.Module):
17
+ def __init__(self, channels: int, bias: bool = True, identity=True):
18
+ super().__init__()
19
+ self.identity = identity
20
+ self.conv1 = conv3x3(channels, channels, bias)
21
+ self.act = nn.LeakyReLU(0.1, inplace=True)
22
+ self.conv2 = conv3x3(channels, channels, bias)
23
+ nn.init.zeros_(self.conv2.weight)
24
+ nn.init.zeros_(self.conv2.bias)
25
+
26
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
27
+ r = self.conv2(self.act(self.conv1(x)))
28
+ a = torch.sigmoid(r)
29
+ return x + r * a
30
+
31
+
32
+ class LWDRB(nn.Module):
33
+ def __init__(self, c, bias=True, dil=4):
34
+ super().__init__()
35
+ self.conv1 = conv3x3(c, c, bias)
36
+ self.act = nn.LeakyReLU(0.1, inplace=True)
37
+ self.conv2 = nn.Conv2d(c, c, 3, padding=dil, dilation=dil, bias=bias)
38
+ nn.init.zeros_(self.conv2.weight)
39
+ nn.init.zeros_(self.conv2.bias)
40
+
41
+ def forward(self, x):
42
+ r = self.conv2(self.act(self.conv1(x)))
43
+ return x + r
44
+
45
+
46
+ class LWGRBShuffle(nn.Module):
47
+ def __init__(self, in_ch, out_ch, scale, bias: bool = True):
48
+ super().__init__()
49
+ self.expand = conv3x3(in_ch, out_ch * scale * scale, bias)
50
+ self.up = nn.PixelShuffle(scale)
51
+ self.refine = LWGRB(out_ch, bias=bias)
52
+
53
+ def forward(self, x):
54
+ x = self.expand(x)
55
+ x = self.up(x)
56
+ return self.refine(x)
57
+
58
+
59
+ @ARCH_REGISTRY.register()
60
+ class GRDFNet(nn.Module):
61
+ """
62
+ GRDFNet (Gated + Residual Dilated Fast Network)
63
+ Configurable block stacking via integer `num_sets`.
64
+
65
+ Structure:
66
+ Stem: 3xLWGRB + 2xLWDRB
67
+ Repeated: num_sets x [LWGRB + LWDRB]
68
+ """
69
+
70
+ def __init__(
71
+ self,
72
+ num_in_ch: int = 3,
73
+ num_out_ch: int = 3,
74
+ feature_channels: int = 32,
75
+ upscale: int = upscale,
76
+ bias: bool = True,
77
+ norm: bool = False,
78
+ img_range: float = 1.0,
79
+ rgb_mean=(0.5, 0.5, 0.5),
80
+ num_sets: int = 3,
81
+ ):
82
+ super().__init__()
83
+
84
+ self.in_ch = num_in_ch
85
+ self.out_ch = num_out_ch
86
+ self.c = feature_channels
87
+ self.scale = upscale
88
+ self.img_range = img_range
89
+ self.gamma = nn.Parameter(torch.tensor(0.5))
90
+ self.num_sets = num_sets
91
+
92
+ self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
93
+ if not norm:
94
+ self.register_buffer("no_norm", torch.zeros(1))
95
+ else:
96
+ self.no_norm = None
97
+
98
+ self.head = conv3x3(self.in_ch, self.c, bias)
99
+ self.body = self._make_body(bias)
100
+ self.tail = conv3x3(self.c, self.out_ch, bias)
101
+
102
+ if self.scale == 1:
103
+ self.upsample0 = nn.Identity()
104
+ else:
105
+ self.upsample0 = LWGRBShuffle(
106
+ self.out_ch, self.out_ch, self.scale, bias=bias
107
+ )
108
+
109
+ def _make_body(self, bias: bool):
110
+ blocks = [
111
+ LWGRB(self.c, bias),
112
+ LWGRB(self.c, bias),
113
+ LWGRB(self.c, bias),
114
+ LWDRB(self.c, bias),
115
+ LWDRB(self.c, bias),
116
+ ]
117
+ for _ in range(self.num_sets):
118
+ blocks += [LWGRB(self.c, bias), LWDRB(self.c, bias)]
119
+ return nn.Sequential(*blocks)
120
+
121
+ @property
122
+ def is_norm(self) -> bool:
123
+ return getattr(self, "no_norm", None) is None
124
+
125
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
126
+ if self.is_norm:
127
+ self.mean = self.mean.type_as(x)
128
+ x = (x - self.mean) * self.img_range
129
+
130
+ feat = self.head(x)
131
+ feat = self.body(feat)
132
+ out_feat = self.tail(feat)
133
+ out_feat = (1.0 - self.gamma) * x + self.gamma * out_feat
134
+
135
+ out = self.upsample0(out_feat)
136
+
137
+ if self.is_norm:
138
+ out = out / self.img_range + self.mean
139
+
140
+ return out