-
Notifications
You must be signed in to change notification settings - Fork 2
/
backbone.py
173 lines (133 loc) · 5.65 KB
/
backbone.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
''' Implementation for ResNet '''
import torch
import torch.nn as nn
from head import Normalized_Softmax_Loss, Normalized_BCE_Loss, Unified_Cross_Entropy_Loss
from head import Sample_to_Sample_Based_Softmax_Loss, Sample_to_Sample_Based_BCE_Loss, Unified_Threshold_Integrated_Sample_to_Sample_Loss
def conv3x3(in_planes, out_planes, stride=1, groups=1):
"""3x3 convolution with padding"""
if stride==2:
return nn.Conv2d(in_planes, out_planes, kernel_size=2, stride=stride, bias=False)
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, groups=groups, bias=False)
class IRBlock(nn.Module):
def __init__(self, inplanes, planes, stride=1, downsample=None):
super(IRBlock, self).__init__()
self.downsample = downsample
self.bn0 = nn.BatchNorm2d(inplanes)
self.conv1 = conv3x3(inplanes, planes)
self.bn1 = nn.BatchNorm2d(planes)
self.prelu = nn.PReLU(planes)
self.conv2 = conv3x3(planes, planes, stride)
self.bn2 = nn.BatchNorm2d(planes)
def forward(self, x):
out = self.bn0(x)
out = self.conv1(out)
out = self.bn1(out)
out = self.prelu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample is not None:
identity = self.downsample(x)
else:
identity = x
return out + identity
class LResNetXEIR(nn.Module):
def __init__(self, block, layers, zero_init_residual=True, num_classes=10):
super(LResNetXEIR, self).__init__()
self.inplanes = 64
self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(self.inplanes)
self.prelu = nn.PReLU(self.inplanes)
self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
self.bn4 = nn.BatchNorm2d(512)
self.dropout = nn.Dropout(p=0.4)
self.fc5 = nn.Conv2d(512, 512, (7, 7), bias=False)
self.bn5 = nn.BatchNorm1d(512, affine=True)
self.bn5.weight.requires_grad = False
self.sample_to_class_loss = Normalized_Softmax_Loss(512, num_classes)
# self.sample_to_class_loss = Normalized_BCE_Loss(512, num_classes)
# self.sample_to_class_loss = Unified_Cross_Entropy_Loss(512, num_classes)
# self.sample_to_sample_loss = Sample_to_Sample_Based_Softmax_Loss(512, num_classes)
# self.sample_to_sample_loss = Sample_to_Sample_Based_BCE_Loss(512, num_classes)
self.sample_to_sample_loss = Unified_Threshold_Integrated_Sample_to_Sample_Loss(512, num_classes)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, a=1, mode='fan_in', nonlinearity='leaky_relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.constant_(m.weight, 1)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.PReLU):
nn.init.constant_(m.weight, 0.25)
elif isinstance(m, nn.BatchNorm1d) or isinstance(m, nn.BatchNorm2d):
if m.weight is not None:
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
if zero_init_residual:
for m in self.modules():
if isinstance(m, IRBlock):
nn.init.constant_(m.bn2.weight, 0)
def _make_layer(self, block, planes, blocks, stride=1):
if stride != 1:
downsample = nn.Sequential(
conv3x3(self.inplanes, planes, stride),
nn.BatchNorm2d(planes)
)
else:
downsample = None
layers = []
layers.append(block(self.inplanes, planes, stride, downsample))
self.inplanes = planes
for _ in range(1, blocks):
layers.append(block(self.inplanes, planes))
return nn.Sequential(*layers)
def restrict_weights(self):
for m in self.modules():
if isinstance(m, nn.PReLU):
m.weight.data.clamp_(min=0.01)
def forward(self, x, t=None, epoch=0, combine=True):
if not self.training:
x = torch.cat((x, x.flip(3)), 0)
x = self.conv1(x)
x = self.bn1(x)
x = self.prelu(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.bn4(x)
x = self.dropout(x)
x = self.fc5(x)
x = torch.flatten(x, 1)
feat = self.bn5(x)
if not self.training:
a, b = feat.chunk(2)
return a + b
if epoch == 0 or combine:
L1 = self.sample_to_class_loss(feat, t)
if epoch > 0:
L2 = self.sample_to_sample_loss(feat, t)
for i in range(t.size(0)):
self.sample_to_sample_loss.feat_mem[t[i]] = feat[i].data.float()
if epoch == 0:
return L1
elif combine:
return L1 + L2
else:
return L2
def LResNet50EIR(**kwargs):
model = LResNetXEIR(IRBlock, [3, 4, 14, 3], **kwargs)
return model
def LResNet100EIR(**kwargs):
model = LResNetXEIR(IRBlock, [3, 13, 30, 3], **kwargs)
return model
def LResNet200EIR(**kwargs):
model = LResNetXEIR(IRBlock, [3, 33, 60, 3], **kwargs)
return model
if __name__ == '__main__':
model = LResNet50EIR()
print(model)