Reset50和ResNeXt50网络图

Reset50 101 152 pytorch代码复现

import torch
import torch.nn as nn
import torchvision
import numpy as np

print("PyTorch Version: ",torch.__version__)
print("Torchvision Version: ",torchvision.__version__)

__all__ = ['ResNet50', 'ResNet101','ResNet152']

def Conv1(in_planes, places, stride=2):
    return nn.Sequential(
        nn.Conv2d(in_channels=in_planes,out_channels=places,kernel_size=7,stride=stride,padding=3, bias=False),
        nn.BatchNorm2d(places),
        nn.ReLU(inplace=True),
        nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
    )

class Bottleneck(nn.Module):
    def __init__(self,in_places,places, stride=1,downsampling=False, expansion = 4):
        super(Bottleneck,self).__init__()
        self.expansion = expansion
        self.downsampling = downsampling

        self.bottleneck = nn.Sequential(
            nn.Conv2d(in_channels=in_places,out_channels=places,kernel_size=1,stride=1, bias=False),
            nn.BatchNorm2d(places),
            nn.ReLU(inplace=True),
            nn.Conv2d(in_channels=places, out_channels=places, kernel_size=3, stride=stride, padding=1, bias=False),
            nn.BatchNorm2d(places),
            nn.ReLU(inplace=True),
            nn.Conv2d(in_channels=places, out_channels=places*self.expansion, kernel_size=1, stride=1, bias=False),
            nn.BatchNorm2d(places*self.expansion),
        )

        if self.downsampling:
            self.downsample = nn.Sequential(
                nn.Conv2d(in_channels=in_places, out_channels=places*self.expansion, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(places*self.expansion)
            )
        self.relu = nn.ReLU(inplace=True)
    def forward(self, x):
        residual = x
        out = self.bottleneck(x)

        if self.downsampling:
            residual = self.downsample(x)

        out += residual
        out = self.relu(out)
        return out

class ResNet(nn.Module):
    def __init__(self,blocks, num_classes=1000, expansion = 4):
        super(ResNet,self).__init__()
        self.expansion = expansion

        self.conv1 = Conv1(in_planes = 3, places= 64)

        self.layer1 = self.make_layer(in_places = 64, places= 64, block=blocks[0], stride=1)
        self.layer2 = self.make_layer(in_places = 256,places=128, block=blocks[1], stride=2)
        self.layer3 = self.make_layer(in_places=512,places=256, block=blocks[2], stride=2)
        self.layer4 = self.make_layer(in_places=1024,places=512, block=blocks[3], stride=2)

        self.avgpool = nn.AvgPool2d(7, stride=1)
        self.fc = nn.Linear(2048,num_classes)

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

    def make_layer(self, in_places, places, block, stride):
        layers = []
        layers.append(Bottleneck(in_places, places,stride, downsampling =True))
        for i in range(1, block):
            layers.append(Bottleneck(places*self.expansion, places))

        return nn.Sequential(*layers)


    def forward(self, x):
        x = self.conv1(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

def ResNet50():
    return ResNet([3, 4, 6, 3])

def ResNet101():
    return ResNet([3, 4, 23, 3])

def ResNet152():
    return ResNet([3, 8, 36, 3])


if __name__=='__main__':
    #model = torchvision.models.resnet50()
    model = ResNet50()
    print(model)

    input = torch.randn(1, 3, 224, 224)
    out = model(input)
    print(out.shape)

ResNeXt50 pytorch代码 

import torch
import torch.nn as nn

class Block(nn.Module):
    def __init__(self,in_channels, out_channels, stride=1, is_shortcut=False):
        super(Block,self).__init__()
        self.relu = nn.ReLU(inplace=True)
        self.is_shortcut = is_shortcut
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels, out_channels // 2, kernel_size=1,stride=stride,bias=False),
            nn.BatchNorm2d(out_channels // 2),
            nn.ReLU()
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(out_channels // 2, out_channels // 2, kernel_size=3, stride=1, padding=1, groups=32,
                                   bias=False),
            nn.BatchNorm2d(out_channels // 2),
            nn.ReLU()
        )
        self.conv3 = nn.Sequential(
            nn.Conv2d(out_channels // 2, out_channels, kernel_size=1,stride=1,bias=False),
            nn.BatchNorm2d(out_channels),
        )
        if is_shortcut:
            self.shortcut = nn.Sequential(
            nn.Conv2d(in_channels,out_channels,kernel_size=1,stride=stride,bias=1),
            nn.BatchNorm2d(out_channels)
        )
    def forward(self, x):
        x_shortcut = x
        x = self.conv1(x)
        x = self.conv2(x)
        x = self.conv3(x)
        if self.is_shortcut:
            x_shortcut = self.shortcut(x_shortcut)
        x = x + x_shortcut
        x = self.relu(x)
        return x

class Resnext(nn.Module):
    def __init__(self,num_classes,layer=[3,4,6,3]):
        super(Resnext,self).__init__()
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        )
        self.conv2 = self._make_layer(64,256,1,num=layer[0])
        self.conv3 = self._make_layer(256,512,2,num=layer[1])
        self.conv4 = self._make_layer(512,1024,2,num=layer[2])
        self.conv5 = self._make_layer(1024,2048,2,num=layer[3])
        self.global_average_pool = nn.AvgPool2d(kernel_size=7, stride=1)
        self.fc = nn.Linear(2048,num_classes)
    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = self.conv3(x)
        x = self.conv4(x)
        x = self.conv5(x)
        x = self.global_average_pool(x)
        x = torch.flatten(x,1)
        x = self.fc(x)
        return x
    def _make_layer(self,in_channels,out_channels,stride,num):
        layers = []
        block_1=Block(in_channels, out_channels,stride=stride,is_shortcut=True)
        layers.append(block_1)
        for i in range(1, num):
            layers.append(Block(out_channels,out_channels,stride=1,is_shortcut=False))
        return nn.Sequential(*layers)


net = Resnext(10)
x = torch.rand((10, 3, 224, 224))
for name,layer in net.named_children():
    if name != "fc":
        x = layer(x)
        print(name, 'output shaoe:', x.shape)
    else:
        x = x.view(x.size(0), -1)
        x = layer(x)
        print(name, 'output shaoe:', x.shape)

查看更多关于复现ResNeXt50和Reset50 pytorch代码的详细内容...