分类模型--ResNet系列--ResNet50

发布时间：2023-03-01 13:00

ResNet是什么？解决了？

Residual net（残差网络）：将靠前若干层的某一层数据输出直接跳过多层引入到后面的数据层的输入部分；表明了后面的特征层的内容会有一部分由前面的某一层线性贡献。结构如下：

分类模型--ResNet系列--ResNet50_第1张图片

ResNet网络解决深度网络退化问题。采用ResNet网络结构的网络层数越深，性能越佳；如下图一显示，56层的深度网络的训练误差和测试误差反而比20层网络更大，这就是深度网络退化；而图二展示了使用ResNet结构前后，34层网络与18层网络的训练结果对比；

分类模型--ResNet系列--ResNet50_第2张图片

分类模型--ResNet系列--ResNet50_第3张图片

 ResNet原理及结构

假设我们想要网络块学习到的映射为H(x)，而直接学习H(x)是很难学习到的。若我们学习另一个残差函数F(x) = H(x) - x可以很容易学习，因为此时网络块的训练目标是将F(x)逼近于0，而不是某一特定映射。因此，最后的映射H(x)就是将F(x)和x相加，H(x) = F(x) + x，如图所示。

分类模型--ResNet系列--ResNet50_第4张图片

 因此，这个网络块的输出y为

y=F(x,{W_{i}})+x.

 为了相加，必须保证加号左右侧维度相同，因此可写成通式如下，Ws用于匹配维度；

y=F(x,W_{i})+W_{s} \: x.

文中提到两种维度匹配的方式:(A)用zero-paddiing增加维度；（B）用1*1卷积增加维度；

在ResNet网络中，有两种基础块，分别是BasicBlock和BotteNeck；前者用于ResNet34以下的网络，后者用于ResNet50及以上的网络；

ResNet50

其包含两个基本块，分别叫做 Conv Block 和 Identity Block，其中Conv块输入和输出维度是不一样的，所以不能连续串联，其用于改变网络的维度；Identity块输入维度和输出维度相同，可以串联，用于加深网络；

Conv Block

在这里插入图片描述

 Identity Block

在这里插入图片描述

 整体结构如下：

在这里插入图片描述

代码

resnet50.py

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


class ResNet50BasicBlock(nn.Module):
    def __init__(self, in_channel, outs, kernerl_size, stride, padding):
        super(ResNet50BasicBlock, self).__init__()
        self.conv1 = nn.Conv2d(in_channel, outs[0], kernel_size=kernerl_size[0], stride=stride[0], padding=padding[0])
        self.bn1 = nn.BatchNorm2d(outs[0])
        self.conv2 = nn.Conv2d(outs[0], outs[1], kernel_size=kernerl_size[1], stride=stride[0], padding=padding[1])
        self.bn2 = nn.BatchNorm2d(outs[1])
        self.conv3 = nn.Conv2d(outs[1], outs[2], kernel_size=kernerl_size[2], stride=stride[0], padding=padding[2])
        self.bn3 = nn.BatchNorm2d(outs[2])

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

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

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

        return F.relu(out + x)


class ResNet50DownBlock(nn.Module):
    def __init__(self, in_channel, outs, kernel_size, stride, padding):
        super(ResNet50DownBlock, self).__init__()
        # out1, out2, out3 = outs
        # print(outs)
        self.conv1 = nn.Conv2d(in_channel, outs[0], kernel_size=kernel_size[0], stride=stride[0], padding=padding[0])
        self.bn1 = nn.BatchNorm2d(outs[0])
        self.conv2 = nn.Conv2d(outs[0], outs[1], kernel_size=kernel_size[1], stride=stride[1], padding=padding[1])
        self.bn2 = nn.BatchNorm2d(outs[1])
        self.conv3 = nn.Conv2d(outs[1], outs[2], kernel_size=kernel_size[2], stride=stride[2], padding=padding[2])
        self.bn3 = nn.BatchNorm2d(outs[2])

        self.extra = nn.Sequential(
            nn.Conv2d(in_channel, outs[2], kernel_size=1, stride=stride[3], padding=0),
            nn.BatchNorm2d(outs[2])
        )

    def forward(self, x):
        x_shortcut = self.extra(x)
        out = self.conv1(x)
        out = self.bn1(out)
        out = F.relu(out)

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

        out = self.conv3(out)
        out = self.bn3(out)
        return F.relu(x_shortcut + out)


class ResNet50(nn.Module):
    def __init__(self):
        super(ResNet50, self).__init__()
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

        self.layer1 = nn.Sequential(
            ResNet50DownBlock(64, outs=[64, 64, 256], kernel_size=[1, 3, 1], stride=[1, 1, 1, 1], padding=[0, 1, 0]),
            ResNet50BasicBlock(256, outs=[64, 64, 256], kernerl_size=[1, 3, 1], stride=[1, 1, 1, 1], padding=[0, 1, 0]),
            ResNet50BasicBlock(256, outs=[64, 64, 256], kernerl_size=[1, 3, 1], stride=[1, 1, 1, 1], padding=[0, 1, 0]),
        )

        self.layer2 = nn.Sequential(
            ResNet50DownBlock(256, outs=[128, 128, 512], kernel_size=[1, 3, 1], stride=[1, 2, 1, 2], padding=[0, 1, 0]),
            ResNet50BasicBlock(512, outs=[128, 128, 512], kernerl_size=[1, 3, 1], stride=[1, 1, 1, 1], padding=[0, 1, 0]),
            ResNet50BasicBlock(512, outs=[128, 128, 512], kernerl_size=[1, 3, 1], stride=[1, 1, 1, 1], padding=[0, 1, 0]),
            ResNet50DownBlock(512, outs=[128, 128, 512], kernel_size=[1, 3, 1], stride=[1, 1, 1, 1], padding=[0, 1, 0])
        )

        self.layer3 = nn.Sequential(
            ResNet50DownBlock(512, outs=[256, 256, 1024], kernel_size=[1, 3, 1], stride=[1, 2, 1, 2], padding=[0, 1, 0]),
            ResNet50BasicBlock(1024, outs=[256, 256, 1024], kernerl_size=[1, 3, 1], stride=[1, 1, 1, 1],
                               padding=[0, 1, 0]),
            ResNet50BasicBlock(1024, outs=[256, 256, 1024], kernerl_size=[1, 3, 1], stride=[1, 1, 1, 1],
                               padding=[0, 1, 0]),
            ResNet50DownBlock(1024, outs=[256, 256, 1024], kernel_size=[1, 3, 1], stride=[1, 1, 1, 1],
                              padding=[0, 1, 0]),
            ResNet50DownBlock(1024, outs=[256, 256, 1024], kernel_size=[1, 3, 1], stride=[1, 1, 1, 1],
                              padding=[0, 1, 0]),
            ResNet50DownBlock(1024, outs=[256, 256, 1024], kernel_size=[1, 3, 1], stride=[1, 1, 1, 1],
                              padding=[0, 1, 0])
        )

        self.layer4 = nn.Sequential(
            ResNet50DownBlock(1024, outs=[512, 512, 2048], kernel_size=[1, 3, 1], stride=[1, 2, 1, 2],
                              padding=[0, 1, 0]),
            ResNet50DownBlock(2048, outs=[512, 512, 2048], kernel_size=[1, 3, 1], stride=[1, 1, 1, 1],
                              padding=[0, 1, 0]),
            ResNet50DownBlock(2048, outs=[512, 512, 2048], kernel_size=[1, 3, 1], stride=[1, 1, 1, 1],
                              padding=[0, 1, 0])
        )

        self.avgpool = nn.AdaptiveAvgPool2d(output_size=(1, 1))

        self.fc = nn.Linear(2048, 10)

    def forward(self, x):
        out = self.conv1(x)
        out = self.maxpool(out)
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = self.avgpool(out)
        out = out.reshape(x.shape[0], -1)
        out = self.fc(out)
        return out


if __name__ == '__main__':
    x = torch.randn(2, 3, 224, 224)
    net = ResNet50()
    out = net(x)
    print('out.shape: ', out.shape)
    print(out)

main.py

import torch
from torch import nn, optim
import torchvision.transforms as transforms
from torchvision import datasets
from torch.utils.data import DataLoader
from resnet50 import ResNet50


#  用CIFAR-10 数据集进行实验

def main():
    batchsz = 128

    cifar_train = datasets.CIFAR10('cifar', True, transform=transforms.Compose([
        transforms.Resize((32, 32)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406],
                             std=[0.229, 0.224, 0.225])
    ]), download=True)
    cifar_train = DataLoader(cifar_train, batch_size=batchsz, shuffle=True)

    cifar_test = datasets.CIFAR10('cifar', False, transform=transforms.Compose([
        transforms.Resize((32, 32)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406],
                             std=[0.229, 0.224, 0.225])
    ]), download=True)
    cifar_test = DataLoader(cifar_test, batch_size=batchsz, shuffle=True)

    x, label = iter(cifar_train).next()
    print('x:', x.shape, 'label:', label.shape)

    device = torch.device('cuda')
    # model = Lenet5().to(device)
    model = ResNet50().to(device)

    criteon = nn.CrossEntropyLoss().to(device)
    optimizer = optim.Adam(model.parameters(), lr=1e-3)
    # print(model)

    for epoch in range(1000):

        model.train()
        for batchidx, (x, label) in enumerate(cifar_train):
            # [b, 3, 32, 32]
            # [b]
            x, label = x.to(device), label.to(device)

            logits = model(x)
            # logits: [b, 10]
            # label:  [b]
            # loss: tensor scalar
            loss = criteon(logits, label)

            # backprop
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

        print(epoch, 'loss:', loss.item())

        model.eval()
        with torch.no_grad():
            # test
            total_correct = 0
            total_num = 0
            for x, label in cifar_test:
                # [b, 3, 32, 32]
                # [b]
                x, label = x.to(device), label.to(device)

                # [b, 10]
                logits = model(x)
                # [b]
                pred = logits.argmax(dim=1)
                # [b] vs [b] => scalar tensor
                correct = torch.eq(pred, label).float().sum().item()
                total_correct += correct
                total_num += x.size(0)
                # print(correct)

            acc = total_correct / total_num
            print(epoch, 'test acc:', acc)


if __name__ == '__main__':
    main()

结果展示

分类模型--ResNet系列--ResNet50_第5张图片

跑了100个Epoch，标号从0-99.最后的AC结果为0.7841

分类模型--ResNet系列--ResNet50_第6张图片

相关推荐

纯干货分享 | 考PMP留下来的一些重要资料（收藏下载）

JavaScript中array.reduce()数组方法的四种使用实例

动物识别专家系统

奋斗丨我还年轻，但是我却没有时间再去荒废了！

PyTorch学习—6.PyTorch数据读取机制Dataloader与Dataset

内嵌物理知识神经网络（PINN）是个坑吗？

图像处理——mixup

科技云报道：数字化时代 云原生安全为何变得如此重要？

跨域浏览器设置解决前端跨域问题

Spring Cloud详解实现声明式微服务调用OpenFeign方法

Vue复刻华为官网 (一)

51单片机+DHT11温湿度传感器+LCD1602显示（最好用的程序，没有之一）

Python3，为了给女神暗送秋波，我默默的写了一个图片字符画生成器，真香。

vue设计与实现中的cleanup

神经网络调参经验总结

浏览器链接跳转进入小程序指定页面（适用安卓和iOS）

java实现求 的近似值_Java 蒙特卡洛算法求圆周率近似值实例详解

吴恩达《机器学习》课后测试Ex1：线性回归（详细Python代码注解）

什么是“企业级”低代码？成为企业级低代码必须具备的5种能力

Java实现ArrayList排序的方法详解

ItVuer - 免责声明 - 关于我们 - 联系我们

本网站信息来源于互联网，如有侵权请联系：561261067@qq.com

桂ICP备16001015号