-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
134 lines (116 loc) · 5.96 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
import torch
import torch.nn as nn
import torch.nn.functional as F
from ptflops import get_model_complexity_info
class SMP(nn.Module):
def __init__(self, num_channels,kernel_size):
super(SMP, self).__init__()
self.num_layers = 4
self.in_channel = num_channels
self.kernel_size = kernel_size
self.padding=(self.kernel_size-1)//2
self.num_filters = 64
self.layer_in = nn.Conv2d(in_channels=self.in_channel, out_channels=self.num_filters,kernel_size=self.kernel_size, padding=self.padding, stride=1, bias=False)
nn.init.xavier_uniform_(self.layer_in.weight.data)
self.lam_in = nn.Parameter(0.01 * torch.ones(1,self.num_filters,1,1))
self.lam_i = []
self.layer_down = []
self.layer_up = []
for i in range(self.num_layers):
down_conv = 'down_conv_{}'.format(i)
up_conv = 'up_conv_{}'.format(i)
lam_id = 'lam_{}'.format(i)
layer_2 = nn.Conv2d(in_channels=self.num_filters, out_channels=self.in_channel,kernel_size=self.kernel_size, padding=self.padding, stride=1, bias=False)
nn.init.xavier_uniform_(layer_2.weight.data)
setattr(self, down_conv, layer_2)
self.layer_down.append(getattr(self, down_conv))
layer_3 = nn.Conv2d(in_channels=self.in_channel, out_channels=self.num_filters,kernel_size=self.kernel_size, padding=self.padding, stride=1, bias=False)
nn.init.xavier_uniform_(layer_3.weight.data)
setattr(self, up_conv, layer_3)
self.layer_up.append(getattr(self, up_conv))
lam_ = nn.Parameter(0.01 * torch.ones(1,self.num_filters,1,1))
setattr(self, lam_id, lam_)
self.lam_i.append(getattr(self, lam_id))
def forward(self, mod):
p1 = self.layer_in(mod)
tensor = torch.mul(torch.sign(p1), F.relu(torch.abs(p1) - self.lam_in))
for i in range(self.num_layers):
p3 = self.layer_down[i](tensor)
p4 = self.layer_up[i](p3)
p5 = tensor - p4
p6 = torch.add(p1, p5)
tensor = torch.mul(torch.sign(p6), F.relu(torch.abs(p6) - self.lam_i[i]))
return tensor
class decoder(nn.Module):
def __init__(self):
super(decoder, self).__init__()
self.channel = 3
self.filters = 64
self.decoconv1 = nn.Conv2d(in_channels=self.filters, out_channels=self.channel, kernel_size=7,stride=1, padding=3, bias=False)
nn.init.xavier_uniform_(self.decoconv1.weight.data)
self.decoconv2 = nn.Conv2d(in_channels=self.filters, out_channels=self.channel, kernel_size=5,stride=1, padding=2, bias=False)
nn.init.xavier_uniform_(self.decoconv2.weight.data)
self.decoconv3 = nn.Conv2d(in_channels=self.filters, out_channels=self.channel, kernel_size=3,stride=1, padding=1, bias=False)
nn.init.xavier_uniform_(self.decoconv3.weight.data)
self.delam1 = nn.Parameter(torch.ones(1))
self.delam2 = nn.Parameter(torch.ones(1))
self.delam3 = nn.Parameter(torch.ones(1))
def forward(self, z1,z2,z3):
rec_x1 = self.decoconv1(z1)
rec_x2 = self.decoconv2(z2)
rec_x3 = self.decoconv3(z3)
rec_x = rec_x1 + rec_x2 + rec_x3
return rec_x,rec_x1,rec_x2,rec_x3
class MCSCNet(nn.Module):
def __init__(self):
super(MCSCNet, self).__init__()
self.channel = 3
self.num_filters = 64
self.kernel_size = 7
self.PN1 = SMP(num_channels=self.channel,kernel_size=7)
self.conv1 = nn.Conv2d(in_channels=self.num_filters,out_channels=self.channel,kernel_size=7,stride=1,padding=3,bias=False)
self.PN2 = SMP(num_channels=self.channel, kernel_size=5)
self.conv2 = nn.Conv2d(in_channels=self.num_filters, out_channels=self.channel, kernel_size=5, stride=1,padding=2, bias=False)
self.PN3 = SMP(num_channels=self.channel, kernel_size=3)
self.conv3 = nn.Conv2d(in_channels=self.num_filters, out_channels=self.channel, kernel_size=3, stride=1,padding=1, bias=False)
nn.init.xavier_uniform_(self.conv1.weight.data)
nn.init.xavier_uniform_(self.conv2.weight.data)
nn.init.xavier_uniform_(self.conv3.weight.data)
self.PN4 = SMP(num_channels=self.channel, kernel_size=7)
self.conv4 = nn.Conv2d(in_channels=self.num_filters, out_channels=self.channel, kernel_size=7, stride=1,padding=3, bias=False)
self.PN5 = SMP(num_channels=self.channel, kernel_size=5)
self.conv5 = nn.Conv2d(in_channels=self.num_filters, out_channels=self.channel, kernel_size=5, stride=1,padding=2, bias=False)
self.PN6 = SMP(num_channels=self.channel, kernel_size=3)
nn.init.xavier_uniform_(self.conv4.weight.data)
nn.init.xavier_uniform_(self.conv5.weight.data)
self.decoder=decoder()
def forward(self, x):
# The first round
z1_first=self.PN1(x)
x1_first=self.conv1(z1_first)
x2hat_first=x-x1_first
z2_first=self.PN2(x2hat_first)
x2_first=self.conv2(z2_first)
x3hat_first=x2hat_first-x2_first
z3_first=self.PN3(x3hat_first)
x3_first=self.conv3(z3_first)
x1_hat=x-x3_first-x2_first
# The second round
z1=self.PN4(x1_hat)
x1=self.conv4(z1)
x2hat=x-x1
z2=self.PN5(x2hat)
x2=self.conv5(z2)
x3hat=x2hat-x2
z3=self.PN6(x3hat)
f_pred,x1_pred,x2_pred,x3_pred=self.decoder(z1,z2,z3)
return f_pred,z1,z2,z3
if __name__ == '__main__':
model = MCSCNet()
macs, params = get_model_complexity_info(model, (3, 112,112), as_strings=True,print_per_layer_stat=True, verbose=True)
print('{:<30} {:<8}'.format('Computational complexity: ', macs))
print('{:<30} {:<8}'.format('Number of parameters: ', params))
# print(MS)
# a = torch.rand([10, 3, 64, 64])
# a_out=MS(a)
# print(a_out.shape)