-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathmodel.py
196 lines (146 loc) · 8.47 KB
/
model.py
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
import torch
import torch.nn as nn
import pdb
class BasicConv3d(nn.Module):
def __init__(self, wn, in_channel, out_channel, kernel_size, stride, padding=(0,0,0)):
super(BasicConv3d, self).__init__()
self.conv = wn(nn.Conv3d(in_channel, out_channel,
kernel_size=kernel_size, stride=stride,
padding=padding))
self.relu = nn.ReLU(inplace=True)
def forward(self, x):
x = self.conv(x)
x = self.relu(x)
return x
class S3Dblock(nn.Module):
def __init__(self, wn, n_feats):
super(S3Dblock, self).__init__()
self.conv = nn.Sequential(
BasicConv3d(wn, n_feats, n_feats, kernel_size=(1,3,3), stride=1, padding=(0,1,1)),
BasicConv3d(wn, n_feats, n_feats, kernel_size=(3,1,1), stride=1, padding=(1,0,0))
)
def forward(self, x):
return self.conv(x)
def _to_4d_tensor(x, depth_stride=None):
"""Converts a 5d tensor to 4d by stackin
the batch and depth dimensions."""
x = x.transpose(0, 2) # swap batch and depth dimensions: NxCxDxHxW => DxCxNxHxW
if depth_stride:
x = x[::depth_stride] # downsample feature maps along depth dimension
depth = x.size()[0]
x = x.permute(2, 0, 1, 3, 4) # DxCxNxHxW => NxDxCxHxW
x = torch.split(x, 1, dim=0) # split along batch dimension: NxDxCxHxW => N*[1xDxCxHxW]
x = torch.cat(x, 1) # concatenate along depth dimension: N*[1xDxCxHxW] => 1x(N*D)xCxHxW
x = x.squeeze(0) # 1x(N*D)xCxHxW => (N*D)xCxHxW
return x, depth
def _to_5d_tensor(x, depth):
"""Converts a 4d tensor back to 5d by splitting
the batch dimension to restore the depth dimension."""
x = torch.split(x, depth) # (N*D)xCxHxW => N*[DxCxHxW]
x = torch.stack(x, dim=0) # re-instate the batch dimension: NxDxCxHxW
x = x.transpose(1, 2) # swap back depth and channel dimensions: NxDxCxHxW => NxCxDxHxW
return x
class Block(nn.Module):
def __init__(self, wn, n_feats, n_conv):
super(Block, self).__init__()
self.relu = nn.ReLU(inplace=True)
Block1 = []
for i in range(n_conv):
Block1.append(S3Dblock(wn, n_feats))
self.Block1 = nn.Sequential(*Block1)
Block2 = []
for i in range(n_conv):
Block2.append(S3Dblock(wn, n_feats))
self.Block2 = nn.Sequential(*Block2)
Block3 = []
for i in range(n_conv):
Block3.append(S3Dblock(wn, n_feats))
self.Block3 = nn.Sequential(*Block3)
self.reduceF = BasicConv3d(wn, n_feats*3, n_feats, kernel_size=1, stride=1)
self.Conv = S3Dblock(wn, n_feats)
self.gamma = nn.Parameter(torch.ones(3))
conv1 = []
conv1.append(wn(nn.Conv2d(n_feats, n_feats, kernel_size=(3,3), stride=1, padding=(1,1))))
conv1.append(self.relu)
conv1.append(wn(nn.Conv2d(n_feats, n_feats, kernel_size=(3,3), stride=1, padding=(1,1))))
self.conv1 = nn.Sequential(*conv1)
conv2 = []
conv2.append(wn(nn.Conv2d(n_feats, n_feats, kernel_size=(3,3), stride=1, padding=(1,1))))
conv2.append(self.relu)
conv2.append(wn(nn.Conv2d(n_feats, n_feats, kernel_size=(3,3), stride=1, padding=(1,1))))
self.conv2 = nn.Sequential(*conv2)
conv3 = []
conv3.append(wn(nn.Conv2d(n_feats, n_feats, kernel_size=(3,3), stride=1, padding=(1,1))))
conv3.append(self.relu)
conv3.append(wn(nn.Conv2d(n_feats, n_feats, kernel_size=(3,3), stride=1, padding=(1,1))))
self.conv3 = nn.Sequential(*conv3)
def forward(self, x):
res = x
x1 = self.Block1(x) + x
x2 = self.Block2(x1) + x1
x3 = self.Block3(x2) + x2
x1, depth = _to_4d_tensor(x1, depth_stride=1)
x1 = self.conv1(x1)
x1 = _to_5d_tensor(x1, depth)
x2, depth = _to_4d_tensor(x2, depth_stride=1)
x2 = self.conv2(x2)
x2 = _to_5d_tensor(x2, depth)
x3, depth = _to_4d_tensor(x3, depth_stride=1)
x3 = self.conv3(x3)
x3 = _to_5d_tensor(x3, depth)
x = torch.cat([self.gamma[0]*x1, self.gamma[1]*x2, self.gamma[2]*x3], 1)
x = self.reduceF(x)
x = self.relu(x)
x = x + res
x = self.Conv(x)
return x
class MCNet(nn.Module):
def __init__(self, args):
super(MCNet, self).__init__()
scale = args.upscale_factor
n_colors = args.n_colors
n_feats = args.n_feats
n_conv = args.n_conv
kernel_size = 3
band_mean = (0.0939, 0.0950, 0.0869, 0.0839, 0.0850, 0.0809, 0.0769, 0.0762, 0.0788, 0.0790, 0.0834,
0.0894, 0.0944, 0.0956, 0.0939, 0.1187, 0.0903, 0.0928, 0.0985, 0.1046, 0.1121, 0.1194,
0.1240, 0.1256, 0.1259, 0.1272, 0.1291, 0.1300, 0.1352, 0.1428, 0.1541) #CAVE
# band_mean = (0.0100, 0.0137, 0.0219, 0.0285, 0.0376, 0.0424, 0.0512, 0.0651, 0.0694, 0.0723, 0.0816,
# 0.0950, 0.1338, 0.1525, 0.1217, 0.1187, 0.1337, 0.1481, 0.1601, 0.1817, 0.1752, 0.1445,
# 0.1450, 0.1378, 0.1343, 0.1328, 0.1303, 0.1299, 0.1456, 0.1433, 0.1303) #Hararvd
# band_mean = (0.0944, 0.1143, 0.1297, 0.1368, 0.1599, 0.1853, 0.2029, 0.2149, 0.2278, 0.2275, 0.2311,
# 0.2331, 0.2265, 0.2347, 0.2384, 0.1187, 0.2425, 0.2441, 0.2471, 0.2453, 0.2494, 0.2584,
# 0.2597, 0.2547, 0.2552, 0.2434, 0.2386, 0.2385, 0.2326, 0.2112, 0.2227) #ICVL
# band_mean = (0.0483, 0.0400, 0.0363, 0.0373, 0.0425, 0.0520, 0.0559, 0.0539, 0.0568, 0.0564, 0.0591,
# 0.0678, 0.0797, 0.0927, 0.0986, 0.1086, 0.1086, 0.1015, 0.0994, 0.0947, 0.0980, 0.0973,
# 0.0925, 0.0873, 0.0887, 0.0854, 0.0844, 0.0833, 0.0823, 0.0866, 0.1171, 0.1538, 0.1535) #Foster
# band_mean = (0.0595, 0.0600, 0.0651, 0.0639, 0.0641, 0.0637, 0.0646, 0.0618, 0.0679, 0.0641, 0.0677,
# 0.0650, 0.0671, 0.0687, 0.0693, 0.0687, 0.0688, 0.0677, 0.0689, 0.0736, 0.0735, 0.0728, 0.0713, 0.0734,
# 0.0726, 0.0722, 0.074, 0.0742, 0.0794, 0.0892, 0.1005) #Foster2002
wn = lambda x: torch.nn.utils.weight_norm(x)
self.band_mean = torch.autograd.Variable(torch.FloatTensor(band_mean)).view([1, n_colors, 1, 1])
self.head = wn(nn.Conv3d(1, n_feats, kernel_size, padding=kernel_size//2))
self.SSRM1 = Block(wn, n_feats, n_conv)
self.SSRM2 = Block(wn, n_feats, n_conv)
self.SSRM3 = Block(wn, n_feats, n_conv)
self.SSRM4 = Block(wn, n_feats, n_conv)
tail = []
tail.append(wn(nn.ConvTranspose3d(n_feats, n_feats, kernel_size=(3,2+scale,2+scale), stride=(1,scale,scale), padding=(1,1,1))))
tail.append(wn(nn.Conv3d(n_feats, 1, kernel_size, padding=kernel_size//2)))
self.tail = nn.Sequential(*tail)
def forward(self, x):
x = x - self.band_mean.cuda()
x = x.unsqueeze(1)
T = self.head(x)
x = self.SSRM1(T)
x = torch.add(x, T)
x = self.SSRM2(x)
x = torch.add(x, T)
x = self.SSRM3(x)
x = torch.add(x, T)
x = self.SSRM4(x)
x = torch.add(x, T)
x = self.tail(x)
x = x.squeeze(1)
x = x + self.band_mean.cuda()
return x