-
Notifications
You must be signed in to change notification settings - Fork 2
/
model.py
253 lines (214 loc) · 10.4 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
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
def get_upsampling_weight(in_channels, out_channels, kernel_size):
"""Make a 2D bilinear kernel suitable for upsampling"""
factor = (kernel_size + 1) // 2
if kernel_size % 2 == 1:
center = factor - 1
else:
center = factor - 0.5
og = np.ogrid[:kernel_size, :kernel_size]
filt = (1 - abs(og[0] - center) / factor) * \
(1 - abs(og[1] - center) / factor)
weight = np.zeros((in_channels, out_channels, kernel_size, kernel_size),
dtype=np.float64)
weight[range(in_channels), range(out_channels), :, :] = filt
return torch.from_numpy(weight).float()
####################################### Focal Network ###########################################
class FocalNet(nn.Module):
def __init__(self, n_class=2, refine=True):
super(FocalNet, self).__init__()
# original image's size = 256*256*3
# conv1
self.conv1_1 = nn.Conv2d(3, 64, 3, padding=1)
self.bn1_1 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu1_1 = nn.ReLU(inplace=True)
self.conv1_2 = nn.Conv2d(64, 64, 3, padding=1)
self.bn1_2 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu1_2 = nn.ReLU(inplace=True)
self.pool1 = nn.MaxPool2d(2, stride=2, ceil_mode=True) # 1/2 2 layers
# conv2
self.conv2_1 = nn.Conv2d(64, 128, 3, padding=1)
self.bn2_1 = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
self.relu2_1 = nn.ReLU(inplace=True)
self.conv2_2 = nn.Conv2d(128, 128, 3, padding=1)
self.bn2_2 = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True)
self.relu2_2 = nn.ReLU(inplace=True)
self.pool2 = nn.MaxPool2d(2, stride=2, ceil_mode=True) # 1/4 2 layers
# conv3
self.conv3_1 = nn.Conv2d(128, 256, 3, padding=1)
self.bn3_1 = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
self.relu3_1 = nn.ReLU(inplace=True)
self.conv3_2 = nn.Conv2d(256, 256, 3, padding=1)
self.bn3_2 = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
self.relu3_2 = nn.ReLU(inplace=True)
self.conv3_3 = nn.Conv2d(256, 256, 3, padding=1)
self.bn3_3 = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
self.relu3_3 = nn.ReLU(inplace=True)
self.conv3_4 = nn.Conv2d(256, 256, 3, padding=1)
self.bn3_4 = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True)
self.relu3_4 = nn.ReLU(inplace=True)
self.pool3 = nn.MaxPool2d(2, stride=2, ceil_mode=True) # 1/8 4 layers
# conv4
self.conv4_1 = nn.Conv2d(256, 512, 3, padding=1)
self.bn4_1 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu4_1 = nn.ReLU(inplace=True)
self.conv4_2 = nn.Conv2d(512, 512, 3, padding=1)
self.bn4_2 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu4_2 = nn.ReLU(inplace=True)
self.conv4_3 = nn.Conv2d(512, 512, 3, padding=1)
self.bn4_3 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu4_3 = nn.ReLU(inplace=True)
self.conv4_4 = nn.Conv2d(512, 512, 3, padding=1)
self.bn4_4 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu4_4 = nn.ReLU(inplace=True)
self.pool4 = nn.MaxPool2d(2, stride=2, ceil_mode=True) # 1/16 4 layers
# conv5
self.conv5_1 = nn.Conv2d(512, 512, 3, padding=1)
self.bn5_1 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu5_1 = nn.ReLU(inplace=True)
self.conv5_2 = nn.Conv2d(512, 512, 3, padding=1)
self.bn5_2 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu5_2 = nn.ReLU(inplace=True)
self.conv5_3 = nn.Conv2d(512, 512, 3, padding=1)
self.bn5_3 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu5_3 = nn.ReLU(inplace=True)
self.conv5_4 = nn.Conv2d(512, 512, 3, padding=1)
self.bn5_4 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True)
self.relu5_4 = nn.ReLU(inplace=True) # 1/32 4 layers
# conv5 add
self.conv5_c = nn.Conv2d(512, 64, 3, padding=1)
# conv4 add
self.conv4_c = nn.Conv2d(512, 64, 3, padding=1)
# conv3 add
self.conv3_c = nn.Conv2d(256, 64, 3, padding=1)
# ----------------------------------- LFRM -----------------------------------#
if refine:
self.conv_refine_1 = nn.Conv2d(64, 64, 3, padding=1)
self.bn_refine_1 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine_1 = nn.PReLU()
self.conv_refine_2 = nn.Conv2d(64, 64, 3, padding=1)
self.bn_refine_2 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine_2 = nn.PReLU()
self.conv_refine_3 = nn.Conv2d(64, 64, 3, padding=1)
self.bn_refine_3 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True)
self.relu_refine_3 = nn.PReLU()
self._initialize_weights()
def _initialize_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
# m.weight.data.zero_()
nn.init.normal(m.weight.data, std=0.01)
if m.bias is not None:
m.bias.data.zero_()
if isinstance(m, nn.ConvTranspose2d):
assert m.kernel_size[0] == m.kernel_size[1]
initial_weight = get_upsampling_weight(m.in_channels, m.out_channels, m.kernel_size[0])
m.weight.data.copy_(initial_weight)
def forward(self, x):
h = x
h = self.relu1_1(self.bn1_1(self.conv1_1(h)))
h = self.relu1_2(self.bn1_2(self.conv1_2(h)))
h_nopool1 = h
h = self.pool1(h)
# pool1 = h
h = self.relu2_1(self.bn2_1(self.conv2_1(h)))
h = self.relu2_2(self.bn2_2(self.conv2_2(h)))
h_nopool2 = h
h = self.pool2(h)
# pool2 = h
h = self.relu3_1(self.bn3_1(self.conv3_1(h)))
h = self.relu3_2(self.bn3_2(self.conv3_2(h)))
h = self.relu3_3(self.bn3_3(self.conv3_3(h)))
h = self.relu3_4(self.bn3_4(self.conv3_4(h)))
h_nopool3 = h
h = self.pool3(h)
# pool3 = h
h = self.relu4_1(self.bn4_1(self.conv4_1(h)))
h = self.relu4_2(self.bn4_2(self.conv4_2(h)))
h = self.relu4_3(self.bn4_3(self.conv4_3(h)))
h = self.relu4_4(self.bn4_4(self.conv4_4(h)))
h_nopool4 = h
h = self.pool4(h)
h = self.relu5_1(self.bn5_1(self.conv5_1(h)))
h = self.relu5_2(self.bn5_2(self.conv5_2(h)))
h = self.relu5_3(self.bn5_3(self.conv5_3(h)))
h = self.relu5_4(self.bn5_4(self.conv5_4(h)))
# conv5, conv4 add
h = self.conv5_c(h)
h = F.upsample(h, scale_factor=2, mode='bilinear')
h_nopool4 = self.conv4_c(h_nopool4)
h = h + h_nopool4
# conv4, conv3 add
h = F.upsample(h, scale_factor=2, mode='bilinear')
h_nopool3 = self.conv3_c(h_nopool3)
h = h + h_nopool3
# ------------------------------ LFRM --------------------------- #
n_size, channel, weight, height = x.size()
if n_size == 24 or n_size == 12:
h = self.relu_refine_1(self.bn_refine_1(self.conv_refine_1(h)))
h = self.relu_refine_2(self.bn_refine_2(self.conv_refine_2(h)))
h = self.relu_refine_3(self.bn_refine_3(self.conv_refine_3(h)))
return h
def copy_params_from_vgg19_bn_focal(self, vgg19_bn):
features = [
self.conv1_1, self.bn1_1, self.relu1_1,
self.conv1_2, self.bn1_2, self.relu1_2,
self.pool1,
self.conv2_1, self.bn2_1, self.relu2_1,
self.conv2_2, self.bn2_2, self.relu2_2,
self.pool2,
self.conv3_1, self.bn3_1, self.relu3_1,
self.conv3_2, self.bn3_2, self.relu3_2,
self.conv3_3, self.bn3_3, self.relu3_3,
self.conv3_4, self.bn3_4, self.relu3_4,
self.pool3,
self.conv4_1, self.bn4_1, self.relu4_1,
self.conv4_2, self.bn4_2, self.relu4_2,
self.conv4_3, self.bn4_3, self.relu4_3,
self.conv4_4, self.bn4_4, self.relu4_4,
self.pool4,
self.conv5_1, self.bn5_1, self.relu5_1,
self.conv5_2, self.bn5_2, self.relu5_2,
self.conv5_3, self.bn5_3, self.relu5_3,
self.conv5_4, self.bn5_4, self.relu5_4,
# self.conv6,
# self.conv7,
]
for l1, l2 in zip(vgg19_bn.features, features):
if isinstance(l1, nn.Conv2d) and isinstance(l2, nn.Conv2d):
assert l1.weight.size() == l2.weight.size()
assert l1.bias.size() == l2.bias.size()
l2.weight.data = l1.weight.data
l2.bias.data = l1.bias.data
if isinstance(l1, nn.BatchNorm2d) and isinstance(l2, nn.BatchNorm2d):
assert l1.weight.size() == l2.weight.size()
assert l1.bias.size() == l2.bias.size()
l2.weight.data = l1.weight.data
l2.bias.data = l1.bias.data
# >>>>>>>>>>>>>>>>>>>>>>>>>>>> FocalNet_pred supervised <<<<<<<<<<<<<<<<<<<<<<<<<<<
class FocalNet_sub(nn.Module):
def __init__(self, n_class=2):
super(FocalNet_sub, self).__init__()
# conv ——> prediction
self.conv = nn.Conv2d(64, 2, 1, padding=0)
self._initialize_weights()
def _initialize_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d): # initialize conv2d's weights normal,std=0.1,bias=0
# m.weight.data.zero_()
nn.init.normal(m.weight.data, std=0.01)
if m.bias is not None:
m.bias.data.zero_()
if isinstance(m, nn.ConvTranspose2d): # Transpose 2d is equal to get_susampling_weight.
assert m.kernel_size[0] == m.kernel_size[1]
initial_weight = get_upsampling_weight(m.in_channels, m.out_channels, m.kernel_size[0])
m.weight.data.copy_(initial_weight)
def forward(self, x):
h = x
h = self.conv(h)
h = F.upsample(h, scale_factor=4, mode='bilinear')
return h
########################################## Focal_END ###############################################