-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathmodel_ddfn.py
108 lines (86 loc) · 4.32 KB
/
model_ddfn.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
from __future__ import absolute_import
from __future__ import print_function
from __future__ import division
import torch
import torch.nn as nn
import torch.nn.init as init
import torch.nn.functional as F
class MsFeat(nn.Module):
def __init__(self, in_channels):
super(MsFeat, self).__init__()
self.conv1 = nn.Sequential(nn.Conv2d(in_channels, 16, 3, padding=1, dilation=1, bias=True), nn.ReLU(inplace=True))
init.kaiming_normal_(self.conv1[0].weight, 0, 'fan_in', 'relu'); init.constant_(self.conv1[0].bias, 0.0)
self.conv2 = nn.Sequential(nn.Conv2d(in_channels, 16, 3, padding=2, dilation=2, bias=True), nn.ReLU(inplace=True))
init.kaiming_normal_(self.conv2[0].weight, 0, 'fan_in', 'relu'); init.constant_(self.conv2[0].bias, 0.0)
self.conv3 = nn.Sequential(nn.Conv2d(16, 16, 3, padding=1, dilation=1, bias=True), nn.ReLU(inplace=True))
init.kaiming_normal_(self.conv3[0].weight, 0, 'fan_in', 'relu'); init.constant_(self.conv3[0].bias, 0.0)
self.conv4 = nn.Sequential(nn.Conv2d(16, 16, 3, padding=2, dilation=2, bias=True), nn.ReLU(inplace=True))
init.kaiming_normal_(self.conv4[0].weight, 0, 'fan_in', 'relu'); init.constant_(self.conv4[0].bias, 0.0)
def forward(self, inputs):
conv1 = self.conv1(inputs)
conv2 = self.conv2(inputs)
conv3 = self.conv3(conv2)
conv4 = self.conv4(conv1)
return torch.cat((conv1, conv2, conv3, conv4), 1)
class Block(nn.Module):
def __init__(self, in_channels):
super(Block, self).__init__()
self.conv1 = nn.Sequential(nn.Conv2d(in_channels, 64, 1, padding=0, dilation=1, bias=True), nn.ReLU(inplace=True))
init.kaiming_normal_(self.conv1[0].weight, 0, 'fan_in', 'relu'); init.constant_(self.conv1[0].bias, 0.0)
self.feat1 = nn.Sequential(nn.Conv2d(64, 16, 3, padding=1, dilation=1, bias=True), nn.ReLU(inplace=True))
init.kaiming_normal_(self.feat1[0].weight, 0, 'fan_in', 'relu'); init.constant_(self.feat1[0].bias, 0.0)
self.feat15 = nn.Sequential(nn.Conv2d(16, 8, 3, padding=2, dilation=2, bias=True), nn.ReLU(inplace=True))
init.kaiming_normal_(self.feat15[0].weight, 0, 'fan_in', 'relu'); init.constant_(self.feat15[0].bias, 0.0)
self.feat2 = nn.Sequential(nn.Conv2d(64, 16, 3, padding=2, dilation=2, bias=True), nn.ReLU(inplace=True))
init.kaiming_normal_(self.feat2[0].weight, 0, 'fan_in', 'relu'); init.constant_(self.feat2[0].bias, 0.0)
self.feat25 = nn.Sequential(nn.Conv2d(16, 8, 3, padding=1, dilation=1, bias=True), nn.ReLU(inplace=True))
init.kaiming_normal_(self.feat25[0].weight, 0, 'fan_in', 'relu'); init.constant_(self.feat25[0].bias, 0.0)
self.feat = nn.Sequential(nn.Conv2d(48, 16, 1, padding=0, dilation=1, bias=True), nn.ReLU(inplace=True))
init.kaiming_normal_(self.feat[0].weight, 0, 'fan_in', 'relu'); init.constant_(self.feat[0].bias, 0.0)
def forward(self, inputs):
conv1 = self.conv1(inputs)
feat1 = self.feat1(conv1)
feat15 = self.feat15(feat1)
feat2 = self.feat2(conv1)
feat25 = self.feat25(feat2)
feat = self.feat(torch.cat((feat1, feat15, feat2, feat25), 1))
return torch.cat((inputs, feat), 1)
class DDFN(nn.Module):
def __init__(self, in_channels=3):
super(DDFN, self).__init__()
base = 64
grow = 16
self.msfeat = MsFeat(in_channels)
self.dfus_block0 = Block(base)
self.dfus_block1 = Block(base + grow * 1)
self.dfus_block2 = Block(base + grow * 2)
self.dfus_block3 = Block(base + grow * 3)
self.dfus_block4 = Block(base + grow * 4)
self.dfus_block5 = Block(base + grow * 5)
self.dfus_block6 = Block(base + grow * 6)
self.dfus_block7 = Block(base + grow * 7)
self.convr = nn.Conv2d(base + grow * 8, in_channels, 1, padding=0, dilation=1, bias=False)
init.normal_(self.convr.weight, mean=0.0, std=0.001)
def forward(self, inputs):
msfeat = self.msfeat(inputs)
b0 = self.dfus_block0(msfeat)
b1 = self.dfus_block1(b0)
b2 = self.dfus_block2(b1)
b3 = self.dfus_block3(b2)
b4 = self.dfus_block4(b3)
b5 = self.dfus_block5(b4)
b6 = self.dfus_block6(b5)
b7 = self.dfus_block7(b6)
convr = self.convr(b7)
return convr
if __name__ == '__main__':
""" example of weight sharing """
#self.convs1_siamese = Conv3x3Stack(1, 12, negative_slope)
#self.convs1_siamese[0].weight = self.convs1[0].weight
import numpy as np
model = DDFN().cuda()
x = torch.tensor(np.random.random((1, 3, 1500, 1000)).astype(np.float32)).cuda()
with torch.no_grad():
out = model(x)
print(out.shape)
print('break')