/
net.py
131 lines (106 loc) · 5.28 KB
/
net.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
import torch
import torch.nn as nn
import math
import cv2
import torch.nn.functional as F
class VGG16(nn.Module):
def __init__(self, args):
super(VGG16, self).__init__()
self.stage = args.stage
self.conv1_1 = nn.Conv2d(4, 64, kernel_size=3,stride = 1, padding=1,bias=True)
self.conv1_2 = nn.Conv2d(64, 64, kernel_size=3,stride = 1, padding=1,bias=True)
self.conv2_1 = nn.Conv2d(64, 128, kernel_size=3, padding=1,bias=True)
self.conv2_2 = nn.Conv2d(128, 128, kernel_size=3, padding=1,bias=True)
self.conv3_1 = nn.Conv2d(128, 256, kernel_size=3, padding=1,bias=True)
self.conv3_2 = nn.Conv2d(256, 256, kernel_size=3, padding=1,bias=True)
self.conv3_3 = nn.Conv2d(256, 256, kernel_size=3, padding=1,bias=True)
self.conv4_1 = nn.Conv2d(256, 512, kernel_size=3, padding=1,bias=True)
self.conv4_2 = nn.Conv2d(512, 512, kernel_size=3, padding=1,bias=True)
self.conv4_3 = nn.Conv2d(512, 512, kernel_size=3, padding=1,bias=True)
self.conv5_1 = nn.Conv2d(512, 512, kernel_size=3, padding=1,bias=True)
self.conv5_2 = nn.Conv2d(512, 512, kernel_size=3, padding=1,bias=True)
self.conv5_3 = nn.Conv2d(512, 512, kernel_size=3, padding=1,bias=True)
# model released before 2019.09.09 should use kernel_size=1 & padding=0
#self.conv6_1 = nn.Conv2d(512, 512, kernel_size=1, padding=0,bias=True)
self.conv6_1 = nn.Conv2d(512, 512, kernel_size=3, padding=1,bias=True)
self.deconv6_1 = nn.Conv2d(512, 512, kernel_size=1,bias=True)
self.deconv5_1 = nn.Conv2d(512, 512, kernel_size=5, padding=2,bias=True)
self.deconv4_1 = nn.Conv2d(512, 256, kernel_size=5, padding=2,bias=True)
self.deconv3_1 = nn.Conv2d(256, 128, kernel_size=5, padding=2,bias=True)
self.deconv2_1 = nn.Conv2d(128, 64, kernel_size=5, padding=2,bias=True)
self.deconv1_1 = nn.Conv2d(64, 64, kernel_size=5, padding=2,bias=True)
self.deconv1 = nn.Conv2d(64, 1, kernel_size=5, padding=2,bias=True)
if args.stage == 2:
# for stage2 training
for p in self.parameters():
p.requires_grad=False
if self.stage == 2 or self.stage == 3:
self.refine_conv1 = nn.Conv2d(4, 64, kernel_size=3, padding=1, bias=True)
self.refine_conv2 = nn.Conv2d(64, 64, kernel_size=3, padding=1, bias=True)
self.refine_conv3 = nn.Conv2d(64, 64, kernel_size=3, padding=1, bias=True)
self.refine_pred = nn.Conv2d(64, 1, kernel_size=3, padding=1, bias=True)
def forward(self, x):
# Stage 1
x11 = F.relu(self.conv1_1(x))
x12 = F.relu(self.conv1_2(x11))
x1p, id1 = F.max_pool2d(x12,kernel_size=(2,2), stride=(2,2),return_indices=True)
# Stage 2
x21 = F.relu(self.conv2_1(x1p))
x22 = F.relu(self.conv2_2(x21))
x2p, id2 = F.max_pool2d(x22,kernel_size=(2,2), stride=(2,2),return_indices=True)
# Stage 3
x31 = F.relu(self.conv3_1(x2p))
x32 = F.relu(self.conv3_2(x31))
x33 = F.relu(self.conv3_3(x32))
x3p, id3 = F.max_pool2d(x33,kernel_size=(2,2), stride=(2,2),return_indices=True)
# Stage 4
x41 = F.relu(self.conv4_1(x3p))
x42 = F.relu(self.conv4_2(x41))
x43 = F.relu(self.conv4_3(x42))
x4p, id4 = F.max_pool2d(x43,kernel_size=(2,2), stride=(2,2),return_indices=True)
# Stage 5
x51 = F.relu(self.conv5_1(x4p))
x52 = F.relu(self.conv5_2(x51))
x53 = F.relu(self.conv5_3(x52))
x5p, id5 = F.max_pool2d(x53,kernel_size=(2,2), stride=(2,2),return_indices=True)
# Stage 6
x61 = F.relu(self.conv6_1(x5p))
# Stage 6d
x61d = F.relu(self.deconv6_1(x61))
# Stage 5d
x5d = F.max_unpool2d(x61d,id5, kernel_size=2, stride=2)
x5d = x5d + x53
x51d = F.relu(self.deconv5_1(x5d))
# Stage 4d
x4d = F.max_unpool2d(x51d, id4, kernel_size=2, stride=2)
x4d = x4d + x43
x41d = F.relu(self.deconv4_1(x4d))
# Stage 3d
x3d = F.max_unpool2d(x41d, id3, kernel_size=2, stride=2)
x3d = x3d + x33
x31d = F.relu(self.deconv3_1(x3d))
# Stage 2d
x2d = F.max_unpool2d(x31d, id2, kernel_size=2, stride=2)
x2d = x2d + x22
x21d = F.relu(self.deconv2_1(x2d))
# Stage 1d
x1d = F.max_unpool2d(x21d, id1, kernel_size=2, stride=2)
x1d = x1d + x12
x12d = F.relu(self.deconv1_1(x1d))
# Should add sigmoid? github repo add so.
raw_alpha = self.deconv1(x12d)
pred_mattes = F.sigmoid(raw_alpha)
if self.stage <= 1:
return pred_mattes, 0
# Stage2 refine conv1
refine0 = torch.cat((x[:, :3, :, :], pred_mattes), 1)
refine1 = F.relu(self.refine_conv1(refine0))
refine2 = F.relu(self.refine_conv2(refine1))
refine3 = F.relu(self.refine_conv3(refine2))
# Should add sigmoid?
# sigmoid lead to refine result all converge to 0...
#pred_refine = F.sigmoid(self.refine_pred(refine3))
pred_refine = self.refine_pred(refine3)
pred_alpha = F.sigmoid(raw_alpha + pred_refine)
#print(pred_mattes.mean(), pred_alpha.mean(), pred_refine.sum())
return pred_mattes, pred_alpha