ResNet(用pytorch搭建自己的网络ResNet)笔记_new_resnet包

ResNet结构种类

ResNet一共有5个变种，其网络层数分别是18,34,50,101,152。主要区别在于使用的是两层残差块还是三层残差块，以及残差块的数量。ResNet-18和ResNet-34都是使用的两层残差块，而其余三个模型使用的是三层残差块，并且第三层的输出通道数为输入通道数的4倍。

残差块

公式为y=F（x）+x，在原来输出F（x）的基础上加上输入x

代码实现

#定义两层的残差块
class Residual_2(nn.Module):
    def __init__(self, in_channels, out_channels, use_1x1conv=False, stride=1):
        super(Residual_2, self).__init__()
        #两个3*3的卷积层
        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1, stride=stride)
        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1)
        #1*1的卷积保证维度一致
        if use_1x1conv:
            self.conv3 = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride)
        else:
            self.conv3 = None
        #BN层
        self.bn1 = nn.BatchNorm2d(out_channels)
        self.bn2 = nn.BatchNorm2d(out_channels)
    def forward(self, X):
        Y = self.conv1(X)
        Y = self.bn1(Y)
        Y = torch.nn.functional.relu(Y)

        Y = self.conv2(Y)
        Y = self.bn2(Y)

        if self.conv3:
            X = self.conv3(X)

        return torch.nn.functional.relu(Y + X)

#定义三层的残差块
class Residual_3(nn.Module):
    def __init__(self, in_channels, out_channels, use_1x1conv=False, stride=1):
        super(Residual_3, self).__init__()
        #三层卷积层
        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride)
        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1)
        self.conv3 = nn.Conv2d(out_channels, out_channels*4, kernel_size=1)
        #1*1的卷积保证维度一致
        if use_1x1conv:
            self.conv4 = nn.Conv2d(in_channels, out_channels*4, kernel_size=1, stride=stride)
        else:
            self.conv4 = None
        #BN层
        self.bn1 = nn.BatchNorm2d(out_channels)
        self.bn2 = nn.BatchNorm2d(out_channels)
        self.bn3 = nn.BatchNorm2d(out_channels*4)
    def forward(self, X):
        Y = self.conv1(X)
        Y = self.bn1(Y)
        Y = torch.nn.functional.relu(Y)

        Y = self.conv2(Y)
        Y = self.bn2(Y)
        Y = torch.nn.functional.relu(Y)

        Y = self.conv3(Y)
        Y = self.bn3(Y)

        if self.conv4:
            X = self.conv4(X)

        return torch.nn.functional.relu(Y + X)
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

注意

当X与Y通道数目不同时，这里使用1x1的conv卷积层来使得最终的输入和输出的通道数达到一致

残差块的第一层会有一个参数stride，通过设置步长为2可以改变输出图片的尺寸

第一层的输入是in_channels，输出是out_channels，通过这一层之后卷积核的数量也会发生改变。其余层的输入和输出都是out_channels。特殊地，对于三层的残差块，最后一层的输出是out_channels*4

实现不同结构的ResNet

#类别数
classes=40

#平铺
class FlattenLayer(nn.Module):
def init(self):
super(FlattenLayer, self).init()

def forward(self, input):
    return input.view(input.size(0), -1)
1
2

#全局平均池化层
class GlobalAvgPool2d(nn.Module):
def init(self):
super(GlobalAvgPool2d, self).init()
def forward(self, x):
return nn.functional.avg_pool2d(x, kernel_size=x.size()[2:])

def resnet_block(in_channels, out_channels, num_residuals, basicblock=2, first_block=False):
blk = []
for i in range(num_residuals):
if basicblock == 2:
if i == 0 and first_block == False :
blk.append(Residual_2(in_channels, out_channels, use_1x1conv=True, stride=2))
else :
blk.append(Residual_2(out_channels, out_channels))
else:
if i==0:
if first_block:
blk.append(Residual_3(in_channels, out_channels, use_1x1conv=True))
else :
blk.append(Residual_3(in_channels4, out_channels, use_1x1conv=True, stride=2))
else:
blk.append(Residual_3(out_channels4, out_channels, use_1x1conv=True))

return nn.Sequential(*blk)
1

定义resnet网络

def ResNet_model(layers):
    #前两层
    net = nn.Sequential(
        # 7*7的卷积层
        nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3),
        nn.BatchNorm2d(64),
        nn.ReLU(),
        # 3*3的最大池化层
        nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
    )
    #定义不同结构的ResNet
    if layers == 18:
        basicblock=2
        num_residual=[2,2,2,2]
    elif layers == 34:
        basicblock=2
        num_residual=[3,4,6,3]
    elif layers == 50:
        basicblock=3
        num_residual=[3,4,6,3]
    elif layers == 101:
        basicblock=3
        num_residual=[3,4,23,3]
    elif layers == 152:
        basicblock=3
        num_residual=[3,8,36,3]
    else :
        exit("ResNet结构不对！")
    #添加block
    net.add_module("resnet_block1", resnet_block(64, 64, num_residual[0], basicblock, first_block=True))
    net.add_module("resnet_block2", resnet_block(64, 128, num_residual[1], basicblock))
    net.add_module("resnet_block3", resnet_block(128, 256, num_residual[2], basicblock))
    net.add_module("resnet_block4", resnet_block(256, 512, num_residual[3], basicblock))
    #添加平均池化层、全连接层
    net.add_module("global_avg_pool", GlobalAvgPool2d())
    if basicblock==2:
        net.add_module("fc", nn.Sequential(FlattenLayer(), nn.Linear(512, classes)))
    else:
        net.add_module("fc", nn.Sequential(FlattenLayer(), nn.Linear(2048, classes)))
    return net
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

网络的最开始是一个7X7的卷积层接上一个3X3的最大池化层。然后是四个block块，最后加上平均池化层和全连接层

五种ResNet模型均使用了四个block块，第一个block块不改变图片的尺寸，后面三个block块的第一个残差块的第一层均使用步长为2的卷积层来使尺寸减半。

对于三层的残差块，由于每个残差块中最后一层的输出通道数是输入通道数的4倍，所以除了第一个block的第一个残差块，其余残差块的输入通道数都要乘以4.

测试

if __name__ == '__main__':
    net = ResNet_model(152)
    X = torch.rand((16, 3, 224, 224))
    for name, layer in net.named_children():
        X = layer(X)
        print(name, ' output shape:\t', X.shape)
1
2
3
4
5
6

对ResNet_model（）函数中的参数进行修改，即可调用不同结构的ResNet模型