resnet.py 7.56 KB
Newer Older
1
2
3
4
5
6
7
8
9
import torch.nn as nn
import torch.utils.model_zoo as model_zoo


__all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101',
           'resnet152']


model_urls = {
10
11
12
13
14
    'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
    'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
    'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
    'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth',
    'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth',
15
16
17
18
}


def conv3x3(in_planes, out_planes, stride=1):
19
    """3x3 convolution with padding"""
20
21
22
23
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=1, bias=False)


24
25
26
27
28
def conv1x1(in_planes, out_planes, stride=1):
    """1x1 convolution"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)


Soumith Chintala's avatar
Soumith Chintala committed
29
class BasicBlock(nn.Module):
30
31
    expansion = 1

32
    def __init__(self, inplanes, planes, stride=1, downsample=None, norm_layer=None):
33
        super(BasicBlock, self).__init__()
34
35
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
36
        # Both self.conv1 and self.downsample layers downsample the input when stride != 1
37
        self.conv1 = conv3x3(inplanes, planes, stride)
38
        self.bn1 = norm_layer(planes)
39
40
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes)
41
        self.bn2 = norm_layer(planes)
42
43
44
45
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
46
        identity = x
47
48
49
50
51
52
53
54
55

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        if self.downsample is not None:
56
            identity = self.downsample(x)
57

58
        out += identity
59
60
61
62
63
        out = self.relu(out)

        return out


Soumith Chintala's avatar
Soumith Chintala committed
64
class Bottleneck(nn.Module):
65
66
    expansion = 4

67
    def __init__(self, inplanes, planes, stride=1, downsample=None, norm_layer=None):
68
        super(Bottleneck, self).__init__()
69
70
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
71
        # Both self.conv2 and self.downsample layers downsample the input when stride != 1
72
        self.conv1 = conv1x1(inplanes, planes)
73
        self.bn1 = norm_layer(planes)
74
        self.conv2 = conv3x3(planes, planes, stride)
75
        self.bn2 = norm_layer(planes)
76
        self.conv3 = conv1x1(planes, planes * self.expansion)
77
        self.bn3 = norm_layer(planes * self.expansion)
78
79
80
81
82
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
83
        identity = x
84
85
86
87
88
89
90
91
92
93
94
95
96

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
97
            identity = self.downsample(x)
98

99
        out += identity
100
101
102
103
104
        out = self.relu(out)

        return out


Soumith Chintala's avatar
Soumith Chintala committed
105
class ResNet(nn.Module):
106

107
    def __init__(self, block, layers, num_classes=1000, zero_init_residual=False, norm_layer=None):
108
        super(ResNet, self).__init__()
109
110
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
Amir Arsalan Soltani's avatar
Amir Arsalan Soltani committed
111
        self.inplanes = 64
112
113
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
114
        self.bn1 = norm_layer(64)
115
116
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
117
118
119
120
        self.layer1 = self._make_layer(block, 64, layers[0], norm_layer=norm_layer)
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, norm_layer=norm_layer)
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, norm_layer=norm_layer)
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, norm_layer=norm_layer)
121
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
122
123
124
125
        self.fc = nn.Linear(512 * block.expansion, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
126
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
127
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
128
129
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)
130

131
132
133
134
135
136
137
138
139
140
        # Zero-initialize the last BN in each residual branch,
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)

141
142
143
    def _make_layer(self, block, planes, blocks, stride=1, norm_layer=None):
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
144
145
146
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
147
                conv1x1(self.inplanes, planes * block.expansion, stride),
148
                norm_layer(planes * block.expansion),
149
150
151
            )

        layers = []
152
        layers.append(block(self.inplanes, planes, stride, downsample, norm_layer))
153
        self.inplanes = planes * block.expansion
154
        for _ in range(1, blocks):
155
            layers.append(block(self.inplanes, planes, norm_layer=norm_layer))
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.fc(x)

        return x


177
def resnet18(pretrained=False, **kwargs):
178
179
180
181
182
    """Constructs a ResNet-18 model.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
183
    model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
184
185
186
187
188
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet18']))
    return model


189
def resnet34(pretrained=False, **kwargs):
190
191
192
193
194
    """Constructs a ResNet-34 model.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
195
    model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs)
196
197
198
199
200
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet34']))
    return model


201
def resnet50(pretrained=False, **kwargs):
202
203
204
205
206
    """Constructs a ResNet-50 model.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
207
    model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
208
209
210
211
212
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet50']))
    return model


213
def resnet101(pretrained=False, **kwargs):
214
215
216
217
218
    """Constructs a ResNet-101 model.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
219
    model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs)
Sam Gross's avatar
Sam Gross committed
220
221
222
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet101']))
    return model
223
224


225
def resnet152(pretrained=False, **kwargs):
226
227
228
229
230
    """Constructs a ResNet-152 model.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
231
    model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs)
232
233
234
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet152']))
    return model