深度学习基础案例5--运用动态学习率构建CNN卷积神经网络实现的运动鞋识别(测试集的准确率84%)

• 🍨 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

前言

• 前几天一直很忙，一直在数学建模中，没有来得及更新，接下来将恢复正常
• 这一次的案例很有意思：在学习动态调整学习率的时候，本来想着记录训练过程中的训练集中损失率最低的学习率，记录下来了发现是：0.001(是初始值），然后用0.001去训练，发现出现了过拟合，哈哈哈哈哈。

目标

• 学习动态调整学习率
• 使测试集的准确率达到84%

结果

• 达到了84%

1、数据预处理

数据文件夹说明(data): 分为训练集和测试集，每个文件夹里面都含有不同品牌的运动鞋分类，分类单独一个文件

1、导入库

import torch 
import torch.nn as nn
import torchvision 
import numpy as np 
import os, PIL, pathlib 

device = ('cuda' if torch.cuda.is_available() else 'cup')
device

输出：

'cuda'

2、数据导入与展示

# 查看数据数据文件夹内容
data_dir = './data/train/'
data_dir = pathlib.Path(data_dir)

# 获取该文件夹内内容
data_path = data_dir.glob('*')   # 获取绝对路径
classNames = [str(path).split('\\')[2] for path in data_path]
classNames

输出：

['adidas', 'nike']

---

# 数据展示
import matplotlib.pyplot as plt 
from PIL import Image

# 获取文件名称
data_path_name = './data/train/nike/'
data_path_list = [f for f in os.listdir(data_path_name) if f.endswith(('jpg', 'png'))]

# 创建画板
fig, axes = plt.subplots(2, 8, figsize=(16, 6))  # fig：画板，ases子图

# 展示
for ax, img_file in zip(axes.flat, data_path_list):
    path_name = os.path.join(data_path_name, img_file)
    img = Image.open(path_name)
    ax.imshow(img)
    ax.axis('off')
    
plt.show()

​

​

3、数据处理

# 将所有的数据图片统一格式
from torchvision import transforms, datasets 

train_path = './data/train/'
test_path = './data/test/'

# 定义训练集、测试集图片标准
train_transforms = transforms.Compose([
    transforms.Resize([224, 224]),  # 统一图片大小
    transforms.ToTensor(),           # 统一规格
    transforms.Normalize(           # 数据标准化
        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] 
    )
])

# 数据处理
train_data = datasets.ImageFolder(root=train_path, transform=train_transforms)
test_data = datasets.ImageFolder(root=test_path, transform=test_transforms)

4、加载与划分动态数据

batch_size = 32

train_dl = torch.utils.data.DataLoader(train_data,
                                       batch_size=batch_size,
                                       shuffle=True,
                                       num_workers=1)

test_dl = torch.utils.data.DataLoader(test_data,
                                      batch_size=batch_size,
                                      shuffle=True,
                                      num_workers=1)

# 展示图像参数
for param, data in train_dl:
    print("image(N, C, H, W): ", param.shape)
    print("data: ", data)
    break

image(N, C, H, W):  torch.Size([32, 3, 224, 224])
data:  tensor([0, 0, 1, 0, 0, 1, 1, 0, 0, 0, 0, 1, 1, 1, 0, 1, 1, 0, 0, 0, 1, 1, 0, 0,
        0, 1, 1, 0, 1, 0, 0, 0])

2、构建CNN神经网络

卷积：--> 12*220*220 --> 12*216*216
池化：--> 12*108*108
卷积: --> 24*104*104 --> 24*100*100
池化：--> 24*50*50 --> 25*50*2 

import torch.nn.functional as F 

class Net_work(nn.Module):
    def __init__(self):
        super(Net_work, self).__init__()   # 父类信息构建
        
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels=3, out_channels=12, kernel_size=5, padding=0),
            nn.BatchNorm2d(12),  # 第一个参数：特征数量
            nn.ReLU()
        )
        
        self.conv2 = nn.Sequential(
            nn.Conv2d(in_channels=12, out_channels=12, kernel_size=5, padding=0),
            nn.BatchNorm2d(12),
            nn.ReLU()
        )
        
        self.pool1 = nn.Sequential(
            nn.MaxPool2d(2, 2)
        )
        
        self.conv3 = nn.Sequential(
            nn.Conv2d(in_channels=12, out_channels=24, kernel_size=5, padding=0),
            nn.BatchNorm2d(24),
            nn.ReLU()
        )
        
        self.conv4 = nn.Sequential(
            nn.Conv2d(in_channels=24, out_channels=24, kernel_size=5, padding=0),
            nn.BatchNorm2d(24),
            nn.ReLU()
        )
        
        self.pool2 = nn.Sequential(
            nn.MaxPool2d(2, 2)
        )
        
        self.dropout = nn.Sequential(
            nn.Dropout(0.2)
        )
        
        self.fc = nn.Sequential(
            nn.Linear(24*50*50, len(classNames))
        )
        
    def forward(self, x):
        
        batch_size = x.size(0)  # 每一次训练的批次大小，N,C,H,W
        x = self.conv1(x)   # 卷积-->NB-->激活
        x = self.conv2(x)   # 卷积-->NB-->激活
        x = self.pool1(x)   # 池化
        x = self.conv3(x)   # 卷积-->NB-->激活
        x = self.conv4(x)   # 卷积-->NB-->激活
        x = self.pool2(x)   # 池化
        x = x.view(batch_size, -1)  # -1，代表自动展示，将24*50*50展开
        x = self.fc(x)
        
        return x

# 将网络结构导入GPU
model = Net_work().to(device)
model

输出：

Net_work(
  (conv1): Sequential(
    (0): Conv2d(3, 12, kernel_size=(5, 5), stride=(1, 1))
    (1): BatchNorm2d(12, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (conv2): Sequential(
    (0): Conv2d(12, 12, kernel_size=(5, 5), stride=(1, 1))
    (1): BatchNorm2d(12, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (pool1): Sequential(
    (0): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (conv3): Sequential(
    (0): Conv2d(12, 24, kernel_size=(5, 5), stride=(1, 1))
    (1): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (conv4): Sequential(
    (0): Conv2d(24, 24, kernel_size=(5, 5), stride=(1, 1))
    (1): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU()
  )
  (pool2): Sequential(
    (0): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (dropout): Sequential(
    (0): Dropout(p=0.2, inplace=False)
  )
  (fc): Sequential(
    (0): Linear(in_features=60000, out_features=2, bias=True)
  )
)

3、模型的训练准备

1、设置超参数

# 创建损失函数
loss_fn = nn.CrossEntropyLoss()
# 初始化学习率
lr = 1e-4
# 创建梯度下降法
optimizer = torch.optim.SGD(model.parameters(), lr=lr)

2、创建训练函数

def train(dataloader, model, loss_fn, optimizer):
    # 总大小
    size = len(dataloader.dataset)
    # 批次大小
    batch_size = len(dataloader)
    
    # 准确率和损失
    trian_acc, train_loss = 0, 0
    
    # 训练
    for X, y in dataloader:
        X, y = X.to(device), y.to(device)
        
        # 模型训练与误差评分
        pred = model(X)
        loss = loss_fn(pred, y)
        
        # 梯度清零
        optimizer.zero_grad()  # 梯度上更新
        # 方向传播
        loss.backward()
        # 梯度更新
        optimizer.step()
        
        # 记录损失和准确率
        train_loss += loss.item()
        trian_acc += (pred.argmax(1) == y).type(torch.float64).sum().item()
    
    # 计算损失和准确率
    trian_acc /= size
    train_loss /= batch_size
    
    return trian_acc, train_loss

3、创建测试函数

def test(dataloader, model, loss_fn):
    size = len(dataloader.dataset)
    batch_size = len(dataloader)
    
    test_acc, test_loss = 0, 0
    
    with torch.no_grad():
        for X, y in dataloader:
            X, y = X.to(device), y.to(device)
            
            pred = model(X)
            loss = loss_fn(pred, y)
            
            test_loss += loss.item()
            test_acc += (pred.argmax(1) == y).type(torch.float64).sum().item()
            
    test_acc /= size 
    test_loss /= batch_size
    
    return test_acc, test_loss

4、动态调整学习率

def adjust_learning_rate(optimizer, epoch, start_lr):
    # 调整规则：每 2 次都衰减到原来的 0.92
    lr = start_lr * (0.95 ** (epoch // 2))
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

4、正式训练

train_acc = []
train_loss = []
test_acc = []
test_loss = []

# 定义训练次数
epoches = 40

for epoch in range(epoches):
    
    # 动态调整学习率
    adjust_learning_rate(optimizer, epoches, lr)
    
    # 训练
    model.train()
    epoch_trian_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)
    
    # 测试
    model.eval()
    epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
    
    # 记录
    train_acc.append(epoch_trian_acc)
    train_loss.append(epoch_train_loss)
    test_acc.append(epoch_test_acc)
    test_loss.append(epoch_test_loss)
    
    template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}')
    print(template.format(epoch+1, epoch_trian_acc*100, epoch_train_loss, epoch_test_acc*100, epoch_test_loss))

Epoch: 1, Train_acc:52.0%, Train_loss:0.736, Test_acc:52.6%, Test_loss:0.695
Epoch: 2, Train_acc:59.2%, Train_loss:0.685, Test_acc:61.8%, Test_loss:0.672
Epoch: 3, Train_acc:65.1%, Train_loss:0.644, Test_acc:72.4%, Test_loss:0.593
Epoch: 4, Train_acc:66.7%, Train_loss:0.616, Test_acc:65.8%, Test_loss:0.571
Epoch: 5, Train_acc:69.3%, Train_loss:0.602, Test_acc:68.4%, Test_loss:0.555
Epoch: 6, Train_acc:69.7%, Train_loss:0.580, Test_acc:69.7%, Test_loss:0.525
Epoch: 7, Train_acc:73.3%, Train_loss:0.559, Test_acc:77.6%, Test_loss:0.530
Epoch: 8, Train_acc:75.9%, Train_loss:0.544, Test_acc:71.1%, Test_loss:0.513
Epoch: 9, Train_acc:75.7%, Train_loss:0.531, Test_acc:75.0%, Test_loss:0.495
Epoch:10, Train_acc:78.7%, Train_loss:0.517, Test_acc:81.6%, Test_loss:0.508
Epoch:11, Train_acc:78.9%, Train_loss:0.499, Test_acc:78.9%, Test_loss:0.503
Epoch:12, Train_acc:81.3%, Train_loss:0.486, Test_acc:77.6%, Test_loss:0.482
Epoch:13, Train_acc:80.5%, Train_loss:0.480, Test_acc:76.3%, Test_loss:0.476
Epoch:14, Train_acc:83.3%, Train_loss:0.468, Test_acc:81.6%, Test_loss:0.497
Epoch:15, Train_acc:81.7%, Train_loss:0.459, Test_acc:81.6%, Test_loss:0.518
Epoch:16, Train_acc:83.9%, Train_loss:0.461, Test_acc:77.6%, Test_loss:0.465
Epoch:17, Train_acc:85.9%, Train_loss:0.443, Test_acc:78.9%, Test_loss:0.497
Epoch:18, Train_acc:84.9%, Train_loss:0.430, Test_acc:78.9%, Test_loss:0.485
Epoch:19, Train_acc:85.9%, Train_loss:0.428, Test_acc:78.9%, Test_loss:0.504
Epoch:20, Train_acc:86.5%, Train_loss:0.418, Test_acc:82.9%, Test_loss:0.446
Epoch:21, Train_acc:87.5%, Train_loss:0.406, Test_acc:82.9%, Test_loss:0.464
Epoch:22, Train_acc:87.5%, Train_loss:0.403, Test_acc:78.9%, Test_loss:0.486
Epoch:23, Train_acc:87.6%, Train_loss:0.393, Test_acc:81.6%, Test_loss:0.443
Epoch:24, Train_acc:88.6%, Train_loss:0.391, Test_acc:84.2%, Test_loss:0.435
Epoch:25, Train_acc:89.8%, Train_loss:0.374, Test_acc:78.9%, Test_loss:0.421
Epoch:26, Train_acc:89.8%, Train_loss:0.371, Test_acc:81.6%, Test_loss:0.444
Epoch:27, Train_acc:90.2%, Train_loss:0.372, Test_acc:82.9%, Test_loss:0.435
Epoch:28, Train_acc:90.8%, Train_loss:0.360, Test_acc:80.3%, Test_loss:0.431
Epoch:29, Train_acc:89.8%, Train_loss:0.356, Test_acc:80.3%, Test_loss:0.423
Epoch:30, Train_acc:91.8%, Train_loss:0.346, Test_acc:78.9%, Test_loss:0.447
Epoch:31, Train_acc:91.2%, Train_loss:0.343, Test_acc:84.2%, Test_loss:0.420
Epoch:32, Train_acc:92.4%, Train_loss:0.338, Test_acc:82.9%, Test_loss:0.455
Epoch:33, Train_acc:92.8%, Train_loss:0.333, Test_acc:80.3%, Test_loss:0.469
Epoch:34, Train_acc:92.4%, Train_loss:0.326, Test_acc:80.3%, Test_loss:0.432
Epoch:35, Train_acc:93.0%, Train_loss:0.321, Test_acc:82.9%, Test_loss:0.429
Epoch:36, Train_acc:92.4%, Train_loss:0.323, Test_acc:77.6%, Test_loss:0.459
Epoch:37, Train_acc:93.8%, Train_loss:0.312, Test_acc:84.2%, Test_loss:0.458
Epoch:38, Train_acc:94.0%, Train_loss:0.312, Test_acc:84.2%, Test_loss:0.437
Epoch:39, Train_acc:94.8%, Train_loss:0.306, Test_acc:81.6%, Test_loss:0.434
Epoch:40, Train_acc:93.6%, Train_loss:0.304, Test_acc:84.2%, Test_loss:0.421

5、结果可视化

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

epoch_length = range(epoches)

plt.figure(figsize=(12, 3))

plt.subplot(1, 2, 1)
plt.plot(epoch_length, train_acc, label='Train Accuaray')
plt.plot(epoch_length, test_acc, label='Test Accuaray')
plt.legend(loc='lower right')
plt.title('Accurary')

plt.subplot(1, 2, 2)
plt.plot(epoch_length, train_loss, label='Train Loss')
plt.plot(epoch_length, test_loss, label='Test Loss')
plt.legend(loc='upper right')
plt.title('Loss')

plt.show()

模型评价

• 准确率：

  • 训练集稳定逐步上升
  • 测试集不太稳定，但是总体趋向上升

• 损失率：

  • 训练集和测试集总体趋于下降
  • 训练和测试的差距后面，大于0.1，继续训练可能会出现过拟合的现象

6、预测

from PIL import Image

# 获取类型
classes = list(train_data.class_to_idx)

# 需要参数：路径、模型、类别
def predict_one_image(image_path, model, transform, classes):
    test_img = Image.open(image_path).convert('RGB')  # 以GRB颜色打开
    # 展示
    plt.imshow(test_img)
    
    test_img = transform(test_img)   # 统一规格
    # 压缩
    img = test_img.to(device).unsqueeze(0)  # 去掉第一个 1
    
    # 预测
    model.eval()
    output = model(img)
    
    _, pred = torch.max(output, 1)
    pred_class = classes[pred]
    print(f'预测结果是: {pred_class}')
    
# 预测
predict_one_image("./data/test/adidas/10.jpg", model, train_transforms, classes)

预测结果是: adidas

7、模型保存

path = './model.pth'
torch.save(model.state_dict(), path)  # 保存模型状态

model.load_state_dict(torch.load(path, map_location=device))  # 报错模型参数

输出：

<All keys matched successfully>

8、总结

准确率和损失率：

• 理想：测试集损失率底，且测试集准确率高

• 过拟合：训练集准确率高，而测试集准确率比较低，比如在这个案例中，如果学习率直接设置固定值，会发现到后面的时候，准确率上升，甚至达到了98%，但是测试集准确率却一直在78%徘徊，故如训练次数多的时候，训练集准确度一般会一直上升(有梯度下降法优化)，但是测试集可能会在后一个地方一直徘徊，甚至出现下降的现象，从而出现过拟合的现象