pytorch实现resnet网络结构

ResNet结构和pytorch实现

resnet的网络结构都是经过5个不同数量的残差块+最后一个全连接分类完成的。

在resnet50以后，由于层数的增加残差块发生了变化，从原来3x3卷积变为三层卷积，卷积核分别为1x1、3x3、1x1，减少了网络参数。主要通过两种方式：1.用zero-padding去增加维度 2.用1x1卷积来增加维度

这是我之前做的读书笔记，忘记看的什么书了，就不加引用了，抱歉

from torch import nn
import torch as t
from torch.nn import functional as F
import cv2

class ResdiualBlock(nn.Module):
    """
    实现子module：Residual Block
    """
    def __init__(self, inchannel, outchannel, stride=1, shortcut=None):
        super(ResdiualBlock, self).__init__()
        self.left = nn.Sequential(
            nn.Conv2d(inchannel, outchannel, 3, stride, 1, bias=False),
            nn.BatchNorm2d(outchannel),
            nn.ReLU(inplace=True),
            nn.Conv2d(outchannel, outchannel, 3, 1, 1, bias=False),
            nn.BatchNorm2d(outchannel)
        )

        self.right = shortcut

    def forward(self, x):
        out = self.left(x)
        residual = x if self.right is None else self.right(x)
        out += residual
        return F.relu(out)


class ResNet(nn.Module):
    """
    实现主Module：ResNet34
    ResNet34包含多个layer， 每个layer又包含多个residual block
    用子module实现residual block， 用_make_layer函数实现layer
    """

    def __init__(self, num_classes=1000):
        super(ResNet, self).__init__()
        # 图像转换
        self.pre = nn.Sequential(
            # in_channel, out_channel, kernel_size, stride, padding
            nn.Conv2d(3, 64, 7, 2, 3, bias=False, ),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(3, 2, 1)
        )

        # 重复的layer， 分别有3，4，6，3 个residual block
        self.layer1 = self._make_layer(64, 128, 3)
        self.layer2 = self._make_layer(128, 256, 4, stride=2)
        self.layer3 = self._make_layer(256, 512, 6, stride=2)
        self.layer4 = self._make_layer(512, 512, 3, stride=2)

        # 全连接分类
        self.fc = nn.Linear(512, num_classes)

    def _make_layer(self, inchannel, outchannel, block_num, stride=1):
        """
        构建residual block
        """
        shortcut = nn.Sequential(
            nn.Conv2d(inchannel, outchannel, 1, stride, bias=False),
            nn.BatchNorm2d(outchannel)
        )

        layers = []
        layers.append(ResdiualBlock(inchannel, outchannel, stride, shortcut))
        for i in range(1, block_num):
            layers.append(ResdiualBlock(outchannel, outchannel))
        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.pre(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = F.avg_pool2d(x, 7)
        x = x.view(x.size(0), -1)
        return self.fc(x)


model = ResNet()
# input = t.autograd.Variable(t.randn(1,2,244,244))
# out = model(input)

print(model)

查看pytorch提供的resnet的网络结构

import torch
from torchvision import models
from torchsummary import summary

resnet = models.resnet101()
print(resnet)
summary(resnet, (3, 224, 224), device='cpu')

下面进行用pytorch实现resnet101网络，resnet50和152只是残差块数量不同，其他一致。

代码有点乱，望多多包含

import torch
from torchvision import models
from torchsummary import summary
from torch import nn
from torch.nn import functional as F

resnet = models.resnet101()
print(resnet)
summary(resnet, (3, 224, 224), device='cpu')


class ResNet101(nn.Module):
    def __init__(self, in_channels, num_class):
        super(ResNet101, self).__init__()
        self.in_channels = in_channels
        self.out_channels = num_class

        self.pre = nn.Sequential(
            nn.Conv2d(in_channels, 64, kernel_size=(7, 7), stride=2, padding=3),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(3, stride=2, padding=1)
        )

        self.conv2 = self._make_layer(64, 64, 3, stride=1)
        self.conv3 = self._make_layer(256, 128, 4, stride=2)
        self.conv4 = self._make_layer(512, 256, 23, stride=2)
        self.conv5 = self._make_layer(1024, 512, 3, stride=2)

        self.pool = nn.AvgPool2d(7, stride=1)
        self.linear = nn.Linear(2048, num_class)

    def _make_layer(self, in_channels, out_channels, block_num, stride=1, expansion=4):
        layer = []
        layer.append(Bottleneck(in_channels, out_channels, stride=stride, sample=True))
        for i in range(1, block_num):
            layer.append((Bottleneck(out_channels * expansion, out_channels)))
        return nn.Sequential(*layer)

    def forward(self, x):
        x = self.pre(x)

        x = self.conv2(x)
        x = self.conv3(x)
        x = self.conv4(x)
        x = self.conv5(x)
        print(x.size())
        x = self.pool(x)

        x = x.view(x.size(0), -1)

        x = self.linear(x)
        return x


class Bottleneck(nn.Module):
    def __init__(self, in_channels, out_channels, stride=1, sample=False, expansion=4):
        super(Bottleneck, self).__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.expansion = expansion
        self.sample = sample
        self.block = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, kernel_size=1,stride=1),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(),
            nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=stride, padding=1),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(),
            nn.Conv2d(out_channels, out_channels * expansion, kernel_size=1, stride=1),
            nn.BatchNorm2d(out_channels * expansion),
            nn.ReLU()
        )
        if self.sample:
            self.downsample = nn.Sequential(
                nn.Conv2d(in_channels, out_channels * expansion, kernel_size=1, stride=stride),
                nn.BatchNorm2d(out_channels * expansion)
            )

    def forward(self, x):
        out = self.block(x)
        print(out.size())
        residual = self.downsample(x) if self.sample is not False else x
        out += residual
        return F.relu(out)


if __name__ == "__main__":
    res = ResNet101(3, 10)
    print(res)
    input = torch.randn(1, 3, 224, 224)
    out = res(input)
    print(out.shape)
    # summary(res, (3, 512, 512), device='cpu')    #太占用内存

介绍我在实现过程中遇到的问题，期间参考的博客在下方提供链接。

问题1.在101_layer中conv2_x中最后一层256通道的1x1卷积在conv3_x怎么变成了128。

下图说明其卷积操作的过程

第二个问题是输出的feature map是通过设置卷积的stride来逐倍缩小的。

参考：

https://blog.csdn.net/shanglianlm/article/details/86376627