进击J2：ResNet50V2算法实战与解析

• 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
• 🍖 原作者：K同学啊

实验目的：

1.根据 TensorFlow 代码，编写出相应的 Pytorch 代码，使用J1的数据测试模型是否构建正确
2.了解ResNetV2与ResNetV的区别
3.改进思路是否可以迁移到其他地方（自由探索）

实验环境：

• 语言环境：python 3.8
• 编译器：Jupyter notebook
• 深度学习环境：Pytorch
  • torch==2.4.0+cu124
  • torchvision==0.19.0+cu124

论文解读

《Identity Mappings in Deep Residual Networks》

1. ResNetV2结构与ResNet结构对比

• 实线表示测试误差(右边的y轴)，虚线表示训练损失(左边的y轴)，Iterations 表示迭代次数

🧲 改进点：(a) original 表示原始的 ResNet 的残差结构，(b) proposed 表示新的 ResNet 的残差结构。主要差别就是(a)结构先卷积后进行 BN 和激活函数计算，最后执行 addition 后再进行ReLU 计算； (b) 结构先进行 BN 和激活函数计算后卷积，把 addition 后的 ReLU 计算放到了残差结构内部。

📌 改进结果：作者使用这两种不同的结构在 CIFAR-10 数据集上做测试，模型用的是 1001层的 ResNet 模型。从图中结果我们可以看出，(b) proposed 的测试集错误率明显更低一些，达到了 4.92%的错误率，(a) original 的测试集错误率是 7.61%。

2. 关于残差结构的不同尝试

(b-f) 中的快捷连接被不同的组件阻碍。为了简化插图，我们不显示BN层，这里所有单位均采用权值层之后的BN层。图中(a-f) 都是作者对残差结构的 shortcut 部分进行的不同尝试 ，作者对不同 shortcut 结构的尝试结果如下表所示 。

作者用不同 shortcut 结构的 ResNet-110 在 CIFAR-10 数据集上做测试，发现最原始的(a)original 结构是最好的，也就是 identity mapping 恒等映射是最好的。

3. 关于激活的尝试

• 最好的结果是(e) full pre-activation，其次到(a) original

模型复现

1. 设置GPU

import torch
import torch.nn as nn
import torchvision.transforms as transforms
import torchvision
from torchvision import transforms, datasets
import os, PIL, pathlib, warnings
import torch.nn.functional as F

warnings.filterwarnings("ignore")  # 忽略警告信息

# 设置CPU/GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)

2. 导入数据

train_transforms = transforms.Compose([
    transforms.Resize([224, 224]),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )
])

test_transforms = transforms.Compose([
    transforms.Resize([224, 224]),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )
])

# 导入数据
data_dir = './bird_photos/'
data_dir = pathlib.Path(data_dir)

data_paths = list(data_dir.glob('*'))
classNames = [str(path).split('\\')[-1] for path in data_paths]
print(classNames)

num_classes = len(classNames)
print(num_classes)

total_data = datasets.ImageFolder(data_dir,transform = train_transforms)
print(total_data)

3. 划分数据集

# 划分数据集
train_size = int(0.8 * len(total_data))
test_size = len(total_data) - train_size
train_dataset, test_dataset = torch.utils.data.random_split(total_data, [train_size, test_size])
print(train_dataset)
print(test_dataset)

batch_size = 8

train_dl = torch.utils.data.DataLoader(train_dataset,
                                       batch_size=batch_size,
                                       shuffle=True,
                                       # num_workers=1
                                       )
test_dl = torch.utils.data.DataLoader(test_dataset,
                                      batch_size=batch_size,
                                      shuffle=True,
                                      # num_workers=1
                                      )

for X, y in test_dl:
    print("Shape of X [N, C, H, W]: ", X.shape)
    print("Shape of y: ", y.shape, y.dtype)
    break

4. Residual Block

class Block2(nn.Module):
    def __init__(self, in_channel, filters, kernel_size=3, stride=1, conv_shortcut=False):
        super(Block2, self).__init__()
        self.preact = nn.Sequential(
            nn.BatchNorm2d(in_channel),
            nn.ReLU(True)
        )

        self.shortcut = conv_shortcut
        if self.shortcut:
            self.short = nn.Conv2d(in_channel, 4 * filters, 1, stride=stride, padding=0, bias=False)
        elif stride > 1:
            self.short = nn.MaxPool2d(kernel_size=1, stride=stride, padding=0)
        else:
            self.short = nn.Identity()

        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channel, filters, 1, stride=1, bias=False),
            nn.BatchNorm2d(filters),
            nn.ReLU(True)
        )

        self.conv2 = nn.Sequential(
            nn.Conv2d(filters, filters, kernel_size, stride=stride, padding=1, bias=False),
            nn.BatchNorm2d(filters),
            nn.ReLU(True)
        )
        self.conv3 = nn.Conv2d(filters, 4 * filters, 1, stride=1, bias=False)

    def forward(self, x):
        x1 = self.preact(x)
        if self.shortcut:
            x2 = self.short(x1)
        else:
            x2 = self.short(x)
        x1 = self.conv1(x1)
        x1 = self.conv2(x1)
        x1 = self.conv3(x1)
        x = x1 + x2

        return x

5.构建ResNet50V2

class ResNet50V2(nn.Module):
    def __init__(self,
                 include_top=True,  # 是否包含位于网络顶部的全链接层
                 preact=True,  # 是否使用预激活
                 use_bias=True,  # 是否对卷积层使用偏置
                 input_shape=[224, 224, 3],
                 classes=1000,
                 pooling=None):  # 用于分类图像的可选类数
        super(ResNet50V2, self).__init__()

        self.conv1 = nn.Sequential()
        self.conv1.add_module('conv', nn.Conv2d(3, 64, 7, stride=2, padding=3, bias=use_bias, padding_mode='zeros'))
        if not preact:
            self.conv1.add_module('bn', nn.BatchNorm2d(64))
            self.conv1.add_module('relu', nn.ReLU())
        self.conv1.add_module('max_pool', nn.MaxPool2d(kernel_size=3, stride=2, padding=1))

        self.conv2 = Stack2(64, 64, 3)
        self.conv3 = Stack2(256, 128, 4)
        self.conv4 = Stack2(512, 256, 6)
        self.conv5 = Stack2(1024, 512, 3, stride=1)

        self.post = nn.Sequential()
        if preact:
            self.post.add_module('bn', nn.BatchNorm2d(2048))
            self.post.add_module('relu', nn.ReLU())
        if include_top:
            self.post.add_module('avg_pool', nn.AdaptiveAvgPool2d((1, 1)))
            self.post.add_module('flatten', nn.Flatten())
            self.post.add_module('fc', nn.Linear(2048, classes))
        else:
            if pooling == 'avg':
                self.post.add_module('avg_pool', nn.AdaptiveAvgPool2d((1, 1)))
            elif pooling == 'max':
                self.post.add_module('max_pool', nn.AdaptiveMaxPool2d((1, 1)))

    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.post(x)
        return x

# 模型实例化
model = ResNet50V2().to(device)
print(model)

模型打印输出，部分截图：

# 统计模型参数量以及其他指标
import torchsummary as summary
summary.summary(model,(3,224,224))

部分截图：

6. 编写训练函数

# 编写训练函数
def train(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)
    num_batches = len(dataloader)

    train_acc, train_loss = 0, 0

    for X, y in dataloader:
        X, y = X.to(device), y.to(device)

        pred = model(X)
        loss = loss_fn(pred, y)

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

        train_loss += loss.item()
        train_acc += (pred.argmax(1) == y).type(torch.float).sum().item()

    train_loss /= num_batches
    train_acc /= size

    return train_acc, train_loss

7. 编写测试函数

# 编写测试函数
def test(dataloader, model, loss_fn):
    size = len(dataloader.dataset)  # 测试集的大小
    num_batches = len(dataloader)  # 批次数目, (size/batch_size，向上取整)
    test_loss, test_acc = 0, 0

    # 当不进行训练时，停止梯度更新，节省计算内存消耗
    with torch.no_grad():
        for imgs, target in dataloader:
            imgs, target = imgs.to(device), target.to(device)

            # 计算loss
            target_pred = model(imgs)
            loss = loss_fn(target_pred, target)

            test_loss += loss.item()
            test_acc += (target_pred.argmax(1) == target).type(torch.float).sum().item()

    test_acc /= size
    test_loss /= num_batches

    return test_acc, test_loss

8. 设置损失函数和学习率

import copy

loss_fn = nn.CrossEntropyLoss()
learn_rate = 1e-4
# SGD与Adam优化器，选择其中一个
# opt = torch.optim.SGD(model.parameters(),lr=learn_rate)
opt = torch.optim.Adam(model.parameters(), lr=learn_rate)

scheduler = torch.optim.lr_scheduler.StepLR(opt, step_size=1, gamma=0.9)  # 定义学习率高度器

epochs = 100  # 设置训练模型的最大轮数为100，但可能到不了100
patience = 10  # 早停的耐心值，即如果模型连续10个周期没有准确率提升，则跳出训练

train_loss = []
train_acc = []
test_loss = []
test_acc = []
best_acc = 0  # 设置一个最佳的准确率，作为最佳模型的判别指标
no_improve_epoch = 0  # 用于跟踪准确率是否提升的计数器
epoch = 0  # 用于统计最终的训练模型的轮数，这里设置初始值为0；为绘图作准备，这里的绘图范围不是epochs = 100

9. 正式训练

# 开始训练
for epoch in range(epochs):

    model.train()
    epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, opt)

    model.eval()
    epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)

    if epoch_test_acc > best_acc:
        best_acc = epoch_test_acc
        best_model = copy.deepcopy(model)
        no_improve_epoch = 0  # 重置计数器
        # 保存最佳模型的检查点
        PATH = './J2_best_model.pth'
        torch.save({
            'epoch': epoch,
            'model_state_dict': best_model.state_dict(),
            'optimizer_state_dict': opt.state_dict(),
            'loss': epoch_test_loss,
        }, PATH)
    else:
        no_improve_epoch += 1

    if no_improve_epoch >= patience:
        print(f"Early stoping triggered at epoch {epoch + 1}")
        break  # 早停

    train_acc.append(epoch_train_acc)
    train_loss.append(epoch_train_loss)
    test_acc.append(epoch_test_acc)
    test_loss.append(epoch_test_loss)

    scheduler.step()  # 更新学习率
    lr = opt.state_dict()['param_groups'][0]['lr']

    template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}, Lr:{:.2E}')
    print(
        template.format(epoch + 1, epoch_train_acc * 100, epoch_train_loss, epoch_test_acc * 100, epoch_test_loss, lr))

代码输出：

Epoch: 1, Train_acc:48.9%, Train_loss:3.001, Test_acc:61.9%, Test_loss:1.048, Lr:9.00E-05
Epoch: 2, Train_acc:71.2%, Train_loss:0.829, Test_acc:73.5%, Test_loss:0.565, Lr:8.10E-05
Epoch: 3, Train_acc:81.2%, Train_loss:0.553, Test_acc:79.6%, Test_loss:0.485, Lr:7.29E-05
Epoch: 4, Train_acc:84.5%, Train_loss:0.450, Test_acc:75.2%, Test_loss:0.747, Lr:6.56E-05
Epoch: 5, Train_acc:85.8%, Train_loss:0.388, Test_acc:87.6%, Test_loss:0.477, Lr:5.90E-05
Epoch: 6, Train_acc:91.4%, Train_loss:0.244, Test_acc:76.1%, Test_loss:0.705, Lr:5.31E-05
Epoch: 7, Train_acc:96.0%, Train_loss:0.180, Test_acc:78.8%, Test_loss:0.762, Lr:4.78E-05
Epoch: 8, Train_acc:94.2%, Train_loss:0.177, Test_acc:84.1%, Test_loss:0.586, Lr:4.30E-05
Epoch: 9, Train_acc:96.9%, Train_loss:0.123, Test_acc:80.5%, Test_loss:0.643, Lr:3.87E-05
Epoch:10, Train_acc:95.4%, Train_loss:0.132, Test_acc:78.8%, Test_loss:0.671, Lr:3.49E-05
Epoch:11, Train_acc:97.8%, Train_loss:0.105, Test_acc:87.6%, Test_loss:0.525, Lr:3.14E-05
Epoch:12, Train_acc:98.5%, Train_loss:0.079, Test_acc:88.5%, Test_loss:0.460, Lr:2.82E-05
Epoch:13, Train_acc:98.7%, Train_loss:0.083, Test_acc:86.7%, Test_loss:0.470, Lr:2.54E-05
Epoch:14, Train_acc:99.1%, Train_loss:0.072, Test_acc:87.6%, Test_loss:0.455, Lr:2.29E-05
Epoch:15, Train_acc:99.3%, Train_loss:0.043, Test_acc:86.7%, Test_loss:0.523, Lr:2.06E-05
Epoch:16, Train_acc:99.6%, Train_loss:0.036, Test_acc:85.8%, Test_loss:0.399, Lr:1.85E-05
Epoch:17, Train_acc:99.8%, Train_loss:0.030, Test_acc:85.0%, Test_loss:0.464, Lr:1.67E-05
Epoch:18, Train_acc:99.3%, Train_loss:0.049, Test_acc:87.6%, Test_loss:0.353, Lr:1.50E-05
Epoch:19, Train_acc:100.0%, Train_loss:0.024, Test_acc:87.6%, Test_loss:0.361, Lr:1.35E-05
Epoch:20, Train_acc:99.8%, Train_loss:0.021, Test_acc:87.6%, Test_loss:0.411, Lr:1.22E-05
Epoch:21, Train_acc:99.6%, Train_loss:0.028, Test_acc:87.6%, Test_loss:0.324, Lr:1.09E-05
Epoch:22, Train_acc:99.6%, Train_loss:0.023, Test_acc:89.4%, Test_loss:0.338, Lr:9.85E-06
Epoch:23, Train_acc:100.0%, Train_loss:0.023, Test_acc:87.6%, Test_loss:0.422, Lr:8.86E-06
Epoch:24, Train_acc:99.6%, Train_loss:0.026, Test_acc:86.7%, Test_loss:0.403, Lr:7.98E-06
Epoch:25, Train_acc:99.1%, Train_loss:0.039, Test_acc:86.7%, Test_loss:0.381, Lr:7.18E-06
Epoch:26, Train_acc:99.6%, Train_loss:0.020, Test_acc:90.3%, Test_loss:0.563, Lr:6.46E-06
Epoch:27, Train_acc:99.3%, Train_loss:0.021, Test_acc:91.2%, Test_loss:0.323, Lr:5.81E-06
Epoch:28, Train_acc:99.8%, Train_loss:0.029, Test_acc:89.4%, Test_loss:0.386, Lr:5.23E-06
Epoch:29, Train_acc:100.0%, Train_loss:0.021, Test_acc:90.3%, Test_loss:0.368, Lr:4.71E-06
Epoch:30, Train_acc:100.0%, Train_loss:0.012, Test_acc:90.3%, Test_loss:0.335, Lr:4.24E-06
Epoch:31, Train_acc:99.8%, Train_loss:0.018, Test_acc:86.7%, Test_loss:0.422, Lr:3.82E-06
Epoch:32, Train_acc:99.8%, Train_loss:0.012, Test_acc:88.5%, Test_loss:0.792, Lr:3.43E-06
Epoch:33, Train_acc:100.0%, Train_loss:0.013, Test_acc:87.6%, Test_loss:0.418, Lr:3.09E-06
Epoch:34, Train_acc:99.8%, Train_loss:0.024, Test_acc:88.5%, Test_loss:0.402, Lr:2.78E-06
Epoch:35, Train_acc:100.0%, Train_loss:0.011, Test_acc:88.5%, Test_loss:0.491, Lr:2.50E-06
Epoch:36, Train_acc:100.0%, Train_loss:0.014, Test_acc:90.3%, Test_loss:0.335, Lr:2.25E-06
Early stoping triggered at epoch 37

# 保存最佳模型到文件中
PATH = './J2_save_best_model.pth'  # 保存的参数文件名
torch.save(best_model.state_dict(), PATH)

print('Done')
print(epoch)
print('no_improve_epoch:', no_improve_epoch)

10. 结果可视化

# 结果可视化
# Loss与Accuracy图

import matplotlib.pyplot as plt
# 隐藏警告
import warnings
warnings.filterwarnings("ignore")  # 忽略警告信息
plt.rcParams['font.sans-serif'] = ['SimHei']  # 用来正常显示中文标签
plt.rcParams['axes.unicode_minus'] = False  # 用来正常显示负号
plt.rcParams['figure.dpi'] = 100  # 分辨率

epochs_range = range(epoch)

plt.figure(figsize=(12, 3))
plt.subplot(1, 2, 1)

plt.plot(epochs_range, train_acc, label='Training Accuracy')
plt.plot(epochs_range, test_acc, label='Test Accuracy')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')

plt.subplot(1, 2, 2)
plt.plot(epochs_range, train_loss, label='Training Loss')
plt.plot(epochs_range, test_loss, label='Test Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.show()

11. 预测

from PIL import Image

classes = list(total_data.class_to_idx)


def predict_one_image(image_path, model, transform, classes):
    test_img = Image.open(image_path).convert('RGB')
    plt.imshow(test_img)  # 展示预测的图片

    test_img = transform(test_img)
    img = test_img.to(device).unsqueeze(0)

    model.eval()
    output = model(img)

    _, pred = torch.max(output, 1)
    pred_class = classes[pred]
    print(f'预测结果是：{pred_class}')
    
import os
from pathlib import Path
import random

#从所有的图片的随机选择一张图片

image=[]
def image_path(data_dir):
    file_list=os.listdir(data_dir)                       #列出四个分类标签
    data_file_dir=file_list                              #从四个分类标签中随机选择一个
    data_dir=Path(data_dir)
    for i in data_file_dir:
        i=Path(i)
        image_file_path=data_dir.joinpath(i)            #拼接路径
        data_file_paths=image_file_path.iterdir()       #罗列文件夹的内容
        data_file_paths=list(data_file_paths)           #要转换为列表
        image.append(data_file_paths)
    file=random.choice(image)                           #从所有的图像中随机选择一类
    file=random.choice(file)                            #从选择的类中随机选择一张图片
    return file

data_dir='./bird_photos'
image_path=image_path(data_dir)
image_path

# 预测训练集中的某张照片
predict_one_image(image_path=image_path,
                  model=model,
                  transform=train_transforms,
                  classes=classes)

总结

上一次实验我们熟悉了 ResNet 的基本结构与应用，而本次实验的目的在于了解 ResNetV2 与 ResNetV 的区别，同时复现 ResNet50V2。

ResNetV2 和 ResNetV1（即ResNet）都是深度残差网络（ResNet）的变体，ResNetV2 是在 ResNetV1 的基础上进行改进的一种版本，旨在提高模型的性能和稳定性。ResNetV2 相对于 ResNetV1 最大的提升在于其采用了“预激活”残差单元的设计，即将激活函数（ReLU和BN）移到权值层（Conv）之前，形成一种“预激活”的方式，而不是常规的“后激活”方式。这种设计有助于信息流动，特别是在深层网络中，使得 ResNetV2 在深层网络（≥110layers）的实验中表现更优，比 ResNetV1 更易于训练并且精度也更高。

因为课题定下来的时间比较晚，这两周一直疲于写标书，昨天凌晨终于完成了，只要写完就是胜利✌️，8.8是个好兆头，勇敢地幻想一下美好的结果吧。J1和J2的练习都有些仓促了，下周期望自己满血归来！

ResNet50V2模型结构大图：