第P5周：Pytorch实现运动鞋识别第P5周：Pytorch实现运动鞋识别要求：了解如何设置动态学习率（重点) 调

第P5周：Pytorch实现运动鞋识别

🍨 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

要求：
1. 了解如何设置动态学习率（重点)
2. 调整代码使测试集accuracy到达84%。
拔高（可选）：
1. 保存训练过程中的最佳模型权重
2. 调整代码使测试集accuracy到达86%。
我的环境：
- 操作系统：CentOS7
- 显卡：RTX3090
- 显卡驱动：535.154.05
- CUDA版本: 12.2
- 语言环境：Python3.10
- 编译器：Jupyter Lab
- 深度学习环境：
  - torch==12.1
  - torchvision==0.18.1

一、前期准备

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

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

device

device(type='cuda')

2. 导入数据

import os,PIL,random,pathlib

data_dir = './data/'
data_dir = pathlib.Path(data_dir)

data_paths  = list(data_dir.glob('*'))
classeNames = [str(path).split("/")[1] for path in data_paths]
classeNames

['train', 'test']

第一步：使用pathlib.Path()函数将字符串类型的文件夹路径转换为pathlib.Path对象。
第二步：使用glob()方法获取data_dir路径下的所有文件路径，并以列表形式存储在data_paths中。
第三步：通过split()函数对data_paths中的每个文件路径执行分割操作，获得各个文件所属的类别名称，并存储在classeNames中
第四步：打印classeNames列表，显示每个文件所属的类别名称。

# 关于transforms.Compose的更多介绍可以参考：https://blog.csdn.net/qq_38251616/article/details/124878863
train_transforms = transforms.Compose([
    transforms.Resize([224, 224]),  # 将输入图片resize成统一尺寸
    # transforms.RandomHorizontalFlip(), # 随机水平翻转
    transforms.ToTensor(),          # 将PIL Image或numpy.ndarray转换为tensor，并归一化到[0,1]之间
    transforms.Normalize(           # 标准化处理-->转换为标准正太分布（高斯分布），使模型更容易收敛
        mean=[0.485, 0.456, 0.406], 
        std=[0.229, 0.224, 0.225])  # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])

test_transform = transforms.Compose([
    transforms.Resize([224, 224]),  # 将输入图片resize成统一尺寸
    transforms.ToTensor(),          # 将PIL Image或numpy.ndarray转换为tensor，并归一化到[0,1]之间
    transforms.Normalize(           # 标准化处理-->转换为标准正太分布（高斯分布），使模型更容易收敛
        mean=[0.485, 0.456, 0.406], 
        std=[0.229, 0.224, 0.225])  # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])

train_dataset = datasets.ImageFolder("./data/train/",transform=train_transforms)
test_dataset  = datasets.ImageFolder("./data/test/",transform=train_transforms)

train_dataset.class_to_idx

{'adidas': 0, 'nike': 1}

batch_size = 32

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

Shape of X [N, C, H, W]:  torch.Size([32, 3, 224, 224])
Shape of y:  torch.Size([32]) torch.int64

二、构建简单的CNN网络

网络结构图：

import torch.nn.functional as F

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.conv1=nn.Sequential(
            nn.Conv2d(3, 12, kernel_size=5, padding=0), # 12*220*220
            nn.BatchNorm2d(12),
            nn.ReLU())
        
        self.conv2=nn.Sequential(
            nn.Conv2d(12, 12, kernel_size=5, padding=0), # 12*216*216
            nn.BatchNorm2d(12),
            nn.ReLU())
        
        self.pool3=nn.Sequential(
            nn.MaxPool2d(2))                              # 12*108*108
        
        self.conv4=nn.Sequential(
            nn.Conv2d(12, 24, kernel_size=5, padding=0), # 24*104*104
            nn.BatchNorm2d(24),
            nn.ReLU())
        
        self.conv5=nn.Sequential(
            nn.Conv2d(24, 24, kernel_size=5, padding=0), # 24*100*100
            nn.BatchNorm2d(24),
            nn.ReLU())
        
        self.pool6=nn.Sequential(
            nn.MaxPool2d(2))                              # 24*50*50

        self.dropout = nn.Sequential(
            nn.Dropout(0.2))
        
        self.fc=nn.Sequential(
            nn.Linear(24*50*50, len(classeNames)))
        
    def forward(self, x):
        
        batch_size = x.size(0)
        x = self.conv1(x)  # 卷积-BN-激活
        x = self.conv2(x)  # 卷积-BN-激活
        x = self.pool3(x)  # 池化
        x = self.conv4(x)  # 卷积-BN-激活
        x = self.conv5(x)  # 卷积-BN-激活
        x = self.pool6(x)  # 池化
        x = self.dropout(x)
        x = x.view(batch_size, -1)  # flatten 变成全连接网络需要的输入 (batch, 24*50*50) ==> (batch, -1), -1 此处自动算出的是24*50*50
        x = self.fc(x)
       
        return x

device = "cuda" if torch.cuda.is_available() else "cpu"
print("Using {} device".format(device))

model = Model().to(device)
model

Using cuda device

Model(
  (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()
  )
  (pool3): Sequential(
    (0): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (conv4): 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()
  )
  (conv5): 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()
  )
  (pool6): 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)
  )
)

训练函数

1. 编写训练函数

# 训练循环
def train(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)  # 训练集的大小
    num_batches = len(dataloader)   # 批次数目, (size/batch_size，向上取整)

    train_loss, train_acc = 0, 0  # 初始化训练损失和正确率
    
    for X, y in dataloader:  # 获取图片及其标签
        X, y = X.to(device), y.to(device)
        
        # 计算预测误差
        pred = model(X)          # 网络输出
        loss = loss_fn(pred, y)  # 计算网络输出和真实值之间的差距，targets为真实值，计算二者差值即为损失
        
        # 反向传播
        optimizer.zero_grad()  # grad属性归零
        loss.backward()        # 反向传播
        optimizer.step()       # 每一步自动更新
        
        # 记录acc与loss
        train_acc  += (pred.argmax(1) == y).type(torch.float).sum().item()
        train_loss += loss.item()
            
    train_acc  /= size
    train_loss /= num_batches

    return train_acc, train_loss

2. 编写测试函数

测试函数和训练函数大致相同，但是由于不进行梯度下降对网络权重进行更新，所以不需要传入优化器

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

3. 设置动态学习率

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

learn_rate = 1e-4 # 初始学习率
optimizer  = torch.optim.SGD(model.parameters(), lr=learn_rate)

调用官方动态学习率接口（与上面方法等价）

# # 调用官方动态学习率接口时使用
# lambda1 = lambda epoch: (0.92 ** (epoch // 2))
# optimizer = torch.optim.SGD(model.parameters(), lr=learn_rate)
# scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda1) #选定调整方法

4. 正式训练

loss_fn    = nn.CrossEntropyLoss() # 创建损失函数
epochs     = 40

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

for epoch in range(epochs):
    # 更新学习率（使用自定义学习率时使用）
    adjust_learning_rate(optimizer, epoch, learn_rate)
    
    model.train()
    epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)
    # scheduler.step() # 更新学习率（调用官方动态学习率接口时使用）
    
    model.eval()
    epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
    
    train_acc.append(epoch_train_acc)
    train_loss.append(epoch_train_loss)
    test_acc.append(epoch_test_acc)
    test_loss.append(epoch_test_loss)
    
    # 获取当前的学习率
    lr = optimizer.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))
print('Done')

Epoch: 1, Train_acc:48.8%, Train_loss:0.770, Test_acc:52.6%, Test_loss:0.678, Lr:1.00E-04
Epoch: 2, Train_acc:60.6%, Train_loss:0.653, Test_acc:69.7%, Test_loss:0.608, Lr:1.00E-04
Epoch: 3, Train_acc:63.1%, Train_loss:0.655, Test_acc:71.1%, Test_loss:0.553, Lr:9.20E-05
Epoch: 4, Train_acc:65.7%, Train_loss:0.617, Test_acc:71.1%, Test_loss:0.545, Lr:9.20E-05
Epoch: 5, Train_acc:70.7%, Train_loss:0.557, Test_acc:72.4%, Test_loss:0.577, Lr:8.46E-05
Epoch: 6, Train_acc:69.9%, Train_loss:0.559, Test_acc:73.7%, Test_loss:0.544, Lr:8.46E-05
Epoch: 7, Train_acc:76.1%, Train_loss:0.514, Test_acc:76.3%, Test_loss:0.512, Lr:7.79E-05
Epoch: 8, Train_acc:79.9%, Train_loss:0.471, Test_acc:77.6%, Test_loss:0.483, Lr:7.79E-05
Epoch: 9, Train_acc:80.7%, Train_loss:0.473, Test_acc:76.3%, Test_loss:0.477, Lr:7.16E-05
Epoch:10, Train_acc:81.7%, Train_loss:0.449, Test_acc:76.3%, Test_loss:0.477, Lr:7.16E-05
Epoch:11, Train_acc:83.3%, Train_loss:0.433, Test_acc:78.9%, Test_loss:0.479, Lr:6.59E-05
Epoch:12, Train_acc:85.7%, Train_loss:0.415, Test_acc:77.6%, Test_loss:0.458, Lr:6.59E-05
Epoch:13, Train_acc:88.2%, Train_loss:0.394, Test_acc:77.6%, Test_loss:0.482, Lr:6.06E-05
Epoch:14, Train_acc:86.3%, Train_loss:0.392, Test_acc:77.6%, Test_loss:0.511, Lr:6.06E-05
Epoch:15, Train_acc:87.3%, Train_loss:0.385, Test_acc:78.9%, Test_loss:0.498, Lr:5.58E-05
Epoch:16, Train_acc:89.2%, Train_loss:0.375, Test_acc:78.9%, Test_loss:0.448, Lr:5.58E-05
Epoch:17, Train_acc:86.3%, Train_loss:0.385, Test_acc:80.3%, Test_loss:0.489, Lr:5.13E-05
Epoch:18, Train_acc:89.0%, Train_loss:0.363, Test_acc:80.3%, Test_loss:0.444, Lr:5.13E-05
Epoch:19, Train_acc:89.2%, Train_loss:0.358, Test_acc:78.9%, Test_loss:0.468, Lr:4.72E-05
Epoch:20, Train_acc:89.2%, Train_loss:0.348, Test_acc:78.9%, Test_loss:0.415, Lr:4.72E-05
Epoch:21, Train_acc:88.6%, Train_loss:0.352, Test_acc:80.3%, Test_loss:0.442, Lr:4.34E-05
Epoch:22, Train_acc:91.6%, Train_loss:0.341, Test_acc:80.3%, Test_loss:0.455, Lr:4.34E-05
Epoch:23, Train_acc:90.0%, Train_loss:0.335, Test_acc:80.3%, Test_loss:0.402, Lr:4.00E-05
Epoch:24, Train_acc:90.6%, Train_loss:0.327, Test_acc:80.3%, Test_loss:0.464, Lr:4.00E-05
Epoch:25, Train_acc:91.4%, Train_loss:0.321, Test_acc:80.3%, Test_loss:0.439, Lr:3.68E-05
Epoch:26, Train_acc:91.4%, Train_loss:0.322, Test_acc:81.6%, Test_loss:0.415, Lr:3.68E-05
Epoch:27, Train_acc:91.2%, Train_loss:0.324, Test_acc:80.3%, Test_loss:0.498, Lr:3.38E-05
Epoch:28, Train_acc:92.8%, Train_loss:0.311, Test_acc:81.6%, Test_loss:0.408, Lr:3.38E-05
Epoch:29, Train_acc:92.6%, Train_loss:0.307, Test_acc:82.9%, Test_loss:0.444, Lr:3.11E-05
Epoch:30, Train_acc:92.6%, Train_loss:0.305, Test_acc:81.6%, Test_loss:0.437, Lr:3.11E-05
Epoch:31, Train_acc:94.6%, Train_loss:0.296, Test_acc:80.3%, Test_loss:0.464, Lr:2.86E-05
Epoch:32, Train_acc:92.0%, Train_loss:0.305, Test_acc:84.2%, Test_loss:0.451, Lr:2.86E-05
Epoch:33, Train_acc:93.2%, Train_loss:0.298, Test_acc:81.6%, Test_loss:0.441, Lr:2.63E-05
Epoch:34, Train_acc:94.0%, Train_loss:0.301, Test_acc:81.6%, Test_loss:0.393, Lr:2.63E-05
Epoch:35, Train_acc:93.4%, Train_loss:0.295, Test_acc:82.9%, Test_loss:0.401, Lr:2.42E-05
Epoch:36, Train_acc:95.4%, Train_loss:0.283, Test_acc:81.6%, Test_loss:0.443, Lr:2.42E-05
Epoch:37, Train_acc:92.6%, Train_loss:0.290, Test_acc:81.6%, Test_loss:0.431, Lr:2.23E-05
Epoch:38, Train_acc:94.8%, Train_loss:0.285, Test_acc:81.6%, Test_loss:0.429, Lr:2.23E-05
Epoch:39, Train_acc:94.4%, Train_loss:0.288, Test_acc:81.6%, Test_loss:0.433, Lr:2.05E-05
Epoch:40, Train_acc:95.2%, Train_loss:0.276, Test_acc:81.6%, Test_loss:0.452, Lr:2.05E-05
Done

四、结果可视化

1. Loss 与 Accurary 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(epochs)

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()
"""

def plot_acc_loss(epoch_acc, epoch_loss):
    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

    train_acc, test_acc = epoch_acc
    train_loss, test_loss = epoch_loss

    epochs_range = range(len(train_acc))

    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()

plot_acc_loss([train_acc, test_acc], [train_loss, test_loss])

2. 指定图片进行预测

torch.squeeze()详解：对数据的维度进行压缩，去掉维数为1的的维度
函数原型： torch.squeeze(input, dim=None, *, out=None)
关键参数说明：
- input (Tensor)：输入Tensor
- dim (int, optional)：如果给定，输入将只在这个维度上被压缩
实战案例：

x = torch.zeros(2, 1, 2, 1, 2)
x.size()
y = torch.squeeze(x)
y.size()
y = torch.squeeze(x, 0)
y.size()
y = torch.squeeze(x, 1)
y.size()

torch.Size([2, 2, 1, 2])

torch.unsqueeze()详解：对数据维度进行扩充。给指定位置加上维数为一的维度
函数原型： torch.unsqueeze(input, dim)
关键参数说明：
- input (Tensor)：输入Tensor
- dim (int)：插入单例维度的索引
实战案例：

x = torch.tensor([1, 2, 3, 4])
torch.unsqueeze(x, 0)
torch.unsqueeze(x, 1)

tensor([[1],
        [2],
        [3],
        [4]])

from PIL import Image 

classes = list(train_dataset.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}')

# 预测训练集中的某张照片
predict_one_image(image_path='./data/test/adidas/1.jpg', 
                  model=model, 
                  transform=train_transforms, 
                  classes=classes)

预测结果是：adidas

五、保存并加载模型

# 模型保存
PATH = './model.pth'  # 保存的参数文件名
torch.save(model.state_dict(), PATH)

# 将参数加载到model当中
model.load_state_dict(torch.load(PATH, map_location=device))

<All keys matched successfully>

六、动态学习率

torch.optim.lr_scheduler.StepLR 等间隔动态调整方法，每经过step_size个epoch，做一次学习率decay，以gamma值为缩小倍数。

函数原型： torch.optim.lr_scheduler.StepLR(optimizer, step_size, gamma=0.1, last_epoch=-1)
关键参数详解：
- optimizer(Optimizer)：是之前定义好的需要优化的优化器的实例名
- step_size(int)：是学习率衰减的周期，每经过每个epoch，做一次学习率decay
- gamma(float)：学习率衰减的乘法因子。Default：0.1
用法示例：

## 重置模型
device = "cuda" if torch.cuda.is_available() else "cpu"
model = Model().to(device)

learn_rate = 1e-3

optimizer = torch.optim.SGD(model.parameters(), lr=learn_rate)
#scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.1)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9)
loss_fn    = nn.CrossEntropyLoss() # 创建损失函数
epochs     = 40

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

for epoch in range(epochs):
    model.train()
    epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)

    model.eval()
    epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
    
    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()
    
    lr = optimizer.state_dict()['param_groups'][0]['lr']
    state_lr.append(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))
print('Done')

plot_acc_loss([train_acc, test_acc], [train_loss, test_loss])

plt.figure(figsize=(4, 3))
plt.plot(range(len(state_lr)), state_lr, label='')
plt.title('Learning Rate')
plt.show()

Epoch: 1, Train_acc:54.8%, Train_loss:3.151, Test_acc:59.2%, Test_loss:0.751, Lr:9.00E-04
Epoch: 2, Train_acc:63.1%, Train_loss:1.556, Test_acc:51.3%, Test_loss:1.699, Lr:8.10E-04
Epoch: 3, Train_acc:68.9%, Train_loss:1.068, Test_acc:59.2%, Test_loss:1.159, Lr:7.29E-04
Epoch: 4, Train_acc:81.1%, Train_loss:0.517, Test_acc:73.7%, Test_loss:0.559, Lr:6.56E-04
Epoch: 5, Train_acc:86.1%, Train_loss:0.324, Test_acc:75.0%, Test_loss:0.536, Lr:5.90E-04
Epoch: 6, Train_acc:92.2%, Train_loss:0.228, Test_acc:76.3%, Test_loss:0.520, Lr:5.31E-04
Epoch: 7, Train_acc:93.2%, Train_loss:0.208, Test_acc:75.0%, Test_loss:0.469, Lr:4.78E-04
Epoch: 8, Train_acc:94.4%, Train_loss:0.178, Test_acc:75.0%, Test_loss:0.519, Lr:4.30E-04
Epoch: 9, Train_acc:96.0%, Train_loss:0.161, Test_acc:77.6%, Test_loss:0.452, Lr:3.87E-04
Epoch:10, Train_acc:97.0%, Train_loss:0.147, Test_acc:77.6%, Test_loss:0.449, Lr:3.49E-04
Epoch:11, Train_acc:97.2%, Train_loss:0.144, Test_acc:80.3%, Test_loss:0.561, Lr:3.14E-04
Epoch:12, Train_acc:97.8%, Train_loss:0.128, Test_acc:80.3%, Test_loss:0.507, Lr:2.82E-04
Epoch:13, Train_acc:98.2%, Train_loss:0.122, Test_acc:81.6%, Test_loss:0.511, Lr:2.54E-04
Epoch:14, Train_acc:98.4%, Train_loss:0.114, Test_acc:81.6%, Test_loss:0.483, Lr:2.29E-04
Epoch:15, Train_acc:98.4%, Train_loss:0.109, Test_acc:80.3%, Test_loss:0.444, Lr:2.06E-04
Epoch:16, Train_acc:98.2%, Train_loss:0.109, Test_acc:80.3%, Test_loss:0.495, Lr:1.85E-04
Epoch:17, Train_acc:98.4%, Train_loss:0.107, Test_acc:81.6%, Test_loss:0.495, Lr:1.67E-04
Epoch:18, Train_acc:98.8%, Train_loss:0.100, Test_acc:80.3%, Test_loss:0.456, Lr:1.50E-04
Epoch:19, Train_acc:98.8%, Train_loss:0.099, Test_acc:81.6%, Test_loss:0.447, Lr:1.35E-04
Epoch:20, Train_acc:98.6%, Train_loss:0.105, Test_acc:80.3%, Test_loss:0.475, Lr:1.22E-04
Epoch:21, Train_acc:98.8%, Train_loss:0.102, Test_acc:81.6%, Test_loss:0.446, Lr:1.09E-04
Epoch:22, Train_acc:99.0%, Train_loss:0.101, Test_acc:80.3%, Test_loss:0.509, Lr:9.85E-05
Epoch:23, Train_acc:99.2%, Train_loss:0.094, Test_acc:81.6%, Test_loss:0.406, Lr:8.86E-05
Epoch:24, Train_acc:99.0%, Train_loss:0.096, Test_acc:81.6%, Test_loss:0.435, Lr:7.98E-05
Epoch:25, Train_acc:98.8%, Train_loss:0.096, Test_acc:81.6%, Test_loss:0.438, Lr:7.18E-05
Epoch:26, Train_acc:99.0%, Train_loss:0.093, Test_acc:81.6%, Test_loss:0.507, Lr:6.46E-05
Epoch:27, Train_acc:99.0%, Train_loss:0.091, Test_acc:81.6%, Test_loss:0.456, Lr:5.81E-05
Epoch:28, Train_acc:99.0%, Train_loss:0.092, Test_acc:80.3%, Test_loss:0.495, Lr:5.23E-05
Epoch:29, Train_acc:99.2%, Train_loss:0.093, Test_acc:81.6%, Test_loss:0.448, Lr:4.71E-05
Epoch:30, Train_acc:99.0%, Train_loss:0.090, Test_acc:81.6%, Test_loss:0.408, Lr:4.24E-05
Epoch:31, Train_acc:99.0%, Train_loss:0.089, Test_acc:81.6%, Test_loss:0.459, Lr:3.82E-05
Epoch:32, Train_acc:99.0%, Train_loss:0.089, Test_acc:81.6%, Test_loss:0.523, Lr:3.43E-05
Epoch:33, Train_acc:98.8%, Train_loss:0.090, Test_acc:80.3%, Test_loss:0.443, Lr:3.09E-05
Epoch:34, Train_acc:99.0%, Train_loss:0.089, Test_acc:81.6%, Test_loss:0.437, Lr:2.78E-05
Epoch:35, Train_acc:99.0%, Train_loss:0.092, Test_acc:81.6%, Test_loss:0.470, Lr:2.50E-05
Epoch:36, Train_acc:99.2%, Train_loss:0.089, Test_acc:81.6%, Test_loss:0.486, Lr:2.25E-05
Epoch:37, Train_acc:98.8%, Train_loss:0.090, Test_acc:81.6%, Test_loss:0.474, Lr:2.03E-05
Epoch:38, Train_acc:98.6%, Train_loss:0.092, Test_acc:81.6%, Test_loss:0.450, Lr:1.82E-05
Epoch:39, Train_acc:99.0%, Train_loss:0.087, Test_acc:81.6%, Test_loss:0.421, Lr:1.64E-05
Epoch:40, Train_acc:99.2%, Train_loss:0.086, Test_acc:80.3%, Test_loss:0.415, Lr:1.48E-05
Done

lr_scheduler.LambdaLR 根据自己定义的函数更新学习率。

函数原型： torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda, last_epoch=-1, verbose=False)
关键参数详解：
- optimizer(Optimizer)：是之前定义好的需要优化的优化器的实例名
- lr_lambda(function)：更新学习率的函数
用法示例：

## 重置模型
device = "cuda" if torch.cuda.is_available() else "cpu"
model = Model().to(device)

learn_rate = 1e-3

lambda1 = lambda epoch: (0.92 ** (epoch // 2)) # 第二组参数的调整方法
optimizer = torch.optim.SGD(model.parameters(), lr=learn_rate)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda1) #选定调整方法


loss_fn    = nn.CrossEntropyLoss() # 创建损失函数
epochs     = 40

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

for epoch in range(epochs):
    model.train()
    epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)

    model.eval()
    epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
    
    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()
    
    lr = optimizer.state_dict()['param_groups'][0]['lr']
    state_lr.append(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))
print('Done')

plot_acc_loss([train_acc, test_acc], [train_loss, test_loss])

plt.figure(figsize=(4, 3))
plt.plot(range(len(state_lr)), state_lr, label='')
plt.title('Learning Rate')
plt.show()

Epoch: 1, Train_acc:49.6%, Train_loss:3.356, Test_acc:55.3%, Test_loss:0.620, Lr:1.00E-03
Epoch: 2, Train_acc:59.8%, Train_loss:1.574, Test_acc:55.3%, Test_loss:1.134, Lr:9.20E-04
Epoch: 3, Train_acc:61.8%, Train_loss:1.208, Test_acc:57.9%, Test_loss:0.880, Lr:9.20E-04
Epoch: 4, Train_acc:55.2%, Train_loss:1.327, Test_acc:71.1%, Test_loss:0.687, Lr:8.46E-04
Epoch: 5, Train_acc:80.1%, Train_loss:0.438, Test_acc:61.8%, Test_loss:0.642, Lr:8.46E-04
Epoch: 6, Train_acc:86.7%, Train_loss:0.334, Test_acc:68.4%, Test_loss:0.542, Lr:7.79E-04
Epoch: 7, Train_acc:88.0%, Train_loss:0.291, Test_acc:78.9%, Test_loss:0.509, Lr:7.79E-04
Epoch: 8, Train_acc:92.4%, Train_loss:0.215, Test_acc:80.3%, Test_loss:0.476, Lr:7.16E-04
Epoch: 9, Train_acc:95.4%, Train_loss:0.181, Test_acc:76.3%, Test_loss:0.553, Lr:7.16E-04
Epoch:10, Train_acc:95.0%, Train_loss:0.168, Test_acc:77.6%, Test_loss:0.426, Lr:6.59E-04
Epoch:11, Train_acc:96.4%, Train_loss:0.144, Test_acc:81.6%, Test_loss:0.448, Lr:6.59E-04
Epoch:12, Train_acc:97.2%, Train_loss:0.137, Test_acc:81.6%, Test_loss:0.421, Lr:6.06E-04
Epoch:13, Train_acc:97.0%, Train_loss:0.131, Test_acc:82.9%, Test_loss:0.362, Lr:6.06E-04
Epoch:14, Train_acc:97.6%, Train_loss:0.124, Test_acc:81.6%, Test_loss:0.380, Lr:5.58E-04
Epoch:15, Train_acc:97.6%, Train_loss:0.118, Test_acc:81.6%, Test_loss:0.343, Lr:5.58E-04
Epoch:16, Train_acc:98.4%, Train_loss:0.111, Test_acc:82.9%, Test_loss:0.384, Lr:5.13E-04
Epoch:17, Train_acc:98.0%, Train_loss:0.103, Test_acc:84.2%, Test_loss:0.388, Lr:5.13E-04
Epoch:18, Train_acc:98.2%, Train_loss:0.103, Test_acc:81.6%, Test_loss:0.395, Lr:4.72E-04
Epoch:19, Train_acc:98.2%, Train_loss:0.105, Test_acc:81.6%, Test_loss:0.386, Lr:4.72E-04
Epoch:20, Train_acc:98.0%, Train_loss:0.097, Test_acc:81.6%, Test_loss:0.361, Lr:4.34E-04
Epoch:21, Train_acc:98.6%, Train_loss:0.093, Test_acc:81.6%, Test_loss:0.359, Lr:4.34E-04
Epoch:22, Train_acc:98.2%, Train_loss:0.085, Test_acc:82.9%, Test_loss:0.340, Lr:4.00E-04
Epoch:23, Train_acc:98.8%, Train_loss:0.089, Test_acc:82.9%, Test_loss:0.355, Lr:4.00E-04
Epoch:24, Train_acc:98.8%, Train_loss:0.088, Test_acc:82.9%, Test_loss:0.356, Lr:3.68E-04
Epoch:25, Train_acc:98.6%, Train_loss:0.082, Test_acc:82.9%, Test_loss:0.366, Lr:3.68E-04
Epoch:26, Train_acc:98.6%, Train_loss:0.080, Test_acc:85.5%, Test_loss:0.336, Lr:3.38E-04
Epoch:27, Train_acc:98.8%, Train_loss:0.077, Test_acc:82.9%, Test_loss:0.382, Lr:3.38E-04
Epoch:28, Train_acc:98.6%, Train_loss:0.076, Test_acc:82.9%, Test_loss:0.387, Lr:3.11E-04
Epoch:29, Train_acc:99.2%, Train_loss:0.071, Test_acc:82.9%, Test_loss:0.347, Lr:3.11E-04
Epoch:30, Train_acc:98.8%, Train_loss:0.074, Test_acc:86.8%, Test_loss:0.363, Lr:2.86E-04
Epoch:31, Train_acc:99.2%, Train_loss:0.069, Test_acc:82.9%, Test_loss:0.375, Lr:2.86E-04
Epoch:32, Train_acc:98.8%, Train_loss:0.071, Test_acc:86.8%, Test_loss:0.394, Lr:2.63E-04
Epoch:33, Train_acc:99.6%, Train_loss:0.069, Test_acc:84.2%, Test_loss:0.305, Lr:2.63E-04
Epoch:34, Train_acc:99.2%, Train_loss:0.066, Test_acc:84.2%, Test_loss:0.322, Lr:2.42E-04
Epoch:35, Train_acc:99.6%, Train_loss:0.068, Test_acc:82.9%, Test_loss:0.371, Lr:2.42E-04
Epoch:36, Train_acc:99.4%, Train_loss:0.067, Test_acc:84.2%, Test_loss:0.306, Lr:2.23E-04
Epoch:37, Train_acc:99.0%, Train_loss:0.063, Test_acc:85.5%, Test_loss:0.312, Lr:2.23E-04
Epoch:38, Train_acc:99.4%, Train_loss:0.064, Test_acc:84.2%, Test_loss:0.311, Lr:2.05E-04
Epoch:39, Train_acc:98.8%, Train_loss:0.064, Test_acc:85.5%, Test_loss:0.343, Lr:2.05E-04
Epoch:40, Train_acc:99.2%, Train_loss:0.062, Test_acc:84.2%, Test_loss:0.388, Lr:1.89E-04
Done

lr_scheduler.MultiStepLR 在特定的 epoch 中调整学习率

函数原型： torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones, gamma=0.1, last_epoch=-1, verbose=False)
关键参数详解：
- optimizer(Optimizer)：是之前定义好的需要优化的优化器的实例名
- milestones(list)：是一个关于epoch数值的list，表示在达到哪个epoch范围内开始变化，必须是升序排列
- gamma(float)：学习率衰减的乘法因子。Default：0.1
用法示例：

## 重置模型
device = "cuda" if torch.cuda.is_available() else "cpu"
model = Model().to(device)

learn_rate = 1e-3

optimizer = torch.optim.SGD(model.parameters(), lr=learn_rate )
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, 
                                                 milestones=[2,6,15,25,35], #调整学习率的epoch数
                                                 gamma=0.1)

loss_fn    = nn.CrossEntropyLoss() # 创建损失函数
epochs     = 40

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

for epoch in range(epochs):
    model.train()
    epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)

    model.eval()
    epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
    
    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()
    
    lr = optimizer.state_dict()['param_groups'][0]['lr']
    state_lr.append(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))
print('Done')

plot_acc_loss([train_acc, test_acc], [train_loss, test_loss])

plt.figure(figsize=(4, 3))
plt.plot(range(len(state_lr)), state_lr, label='')
plt.title('Learning Rate')
plt.show()

Epoch: 1, Train_acc:51.0%, Train_loss:3.536, Test_acc:60.5%, Test_loss:0.611, Lr:1.00E-03
Epoch: 2, Train_acc:61.2%, Train_loss:1.506, Test_acc:52.6%, Test_loss:0.921, Lr:1.00E-04
Epoch: 3, Train_acc:75.1%, Train_loss:0.545, Test_acc:69.7%, Test_loss:0.516, Lr:1.00E-04
Epoch: 4, Train_acc:82.9%, Train_loss:0.410, Test_acc:73.7%, Test_loss:0.461, Lr:1.00E-04
Epoch: 5, Train_acc:86.5%, Train_loss:0.376, Test_acc:75.0%, Test_loss:0.454, Lr:1.00E-04
Epoch: 6, Train_acc:83.1%, Train_loss:0.370, Test_acc:78.9%, Test_loss:0.499, Lr:1.00E-05
Epoch: 7, Train_acc:88.0%, Train_loss:0.338, Test_acc:78.9%, Test_loss:0.461, Lr:1.00E-05
Epoch: 8, Train_acc:87.1%, Train_loss:0.350, Test_acc:78.9%, Test_loss:0.504, Lr:1.00E-05
Epoch: 9, Train_acc:84.3%, Train_loss:0.362, Test_acc:77.6%, Test_loss:0.496, Lr:1.00E-05
Epoch:10, Train_acc:86.1%, Train_loss:0.350, Test_acc:76.3%, Test_loss:0.469, Lr:1.00E-05
Epoch:11, Train_acc:86.3%, Train_loss:0.340, Test_acc:78.9%, Test_loss:0.482, Lr:1.00E-05
Epoch:12, Train_acc:85.1%, Train_loss:0.357, Test_acc:78.9%, Test_loss:0.499, Lr:1.00E-05
Epoch:13, Train_acc:83.9%, Train_loss:0.373, Test_acc:77.6%, Test_loss:0.485, Lr:1.00E-05
Epoch:14, Train_acc:85.3%, Train_loss:0.352, Test_acc:78.9%, Test_loss:0.492, Lr:1.00E-05
Epoch:15, Train_acc:86.1%, Train_loss:0.340, Test_acc:77.6%, Test_loss:0.453, Lr:1.00E-06
Epoch:16, Train_acc:87.3%, Train_loss:0.336, Test_acc:78.9%, Test_loss:0.499, Lr:1.00E-06
Epoch:17, Train_acc:85.5%, Train_loss:0.357, Test_acc:78.9%, Test_loss:0.494, Lr:1.00E-06
Epoch:18, Train_acc:86.7%, Train_loss:0.345, Test_acc:78.9%, Test_loss:0.524, Lr:1.00E-06
Epoch:19, Train_acc:86.3%, Train_loss:0.340, Test_acc:78.9%, Test_loss:0.460, Lr:1.00E-06
Epoch:20, Train_acc:87.8%, Train_loss:0.331, Test_acc:77.6%, Test_loss:0.454, Lr:1.00E-06
Epoch:21, Train_acc:86.3%, Train_loss:0.355, Test_acc:78.9%, Test_loss:0.492, Lr:1.00E-06
Epoch:22, Train_acc:86.3%, Train_loss:0.341, Test_acc:77.6%, Test_loss:0.495, Lr:1.00E-06
Epoch:23, Train_acc:87.1%, Train_loss:0.339, Test_acc:77.6%, Test_loss:0.492, Lr:1.00E-06
Epoch:24, Train_acc:87.1%, Train_loss:0.335, Test_acc:78.9%, Test_loss:0.514, Lr:1.00E-06
Epoch:25, Train_acc:86.7%, Train_loss:0.352, Test_acc:78.9%, Test_loss:0.463, Lr:1.00E-07
Epoch:26, Train_acc:86.7%, Train_loss:0.341, Test_acc:77.6%, Test_loss:0.512, Lr:1.00E-07
Epoch:27, Train_acc:87.5%, Train_loss:0.343, Test_acc:78.9%, Test_loss:0.475, Lr:1.00E-07
Epoch:28, Train_acc:88.2%, Train_loss:0.340, Test_acc:78.9%, Test_loss:0.478, Lr:1.00E-07
Epoch:29, Train_acc:85.5%, Train_loss:0.354, Test_acc:78.9%, Test_loss:0.447, Lr:1.00E-07
Epoch:30, Train_acc:87.1%, Train_loss:0.333, Test_acc:77.6%, Test_loss:0.496, Lr:1.00E-07
Epoch:31, Train_acc:84.9%, Train_loss:0.349, Test_acc:78.9%, Test_loss:0.541, Lr:1.00E-07
Epoch:32, Train_acc:85.9%, Train_loss:0.349, Test_acc:78.9%, Test_loss:0.502, Lr:1.00E-07
Epoch:33, Train_acc:88.0%, Train_loss:0.335, Test_acc:78.9%, Test_loss:0.435, Lr:1.00E-07
Epoch:34, Train_acc:85.1%, Train_loss:0.346, Test_acc:76.3%, Test_loss:0.486, Lr:1.00E-07
Epoch:35, Train_acc:87.1%, Train_loss:0.354, Test_acc:78.9%, Test_loss:0.516, Lr:1.00E-08
Epoch:36, Train_acc:86.9%, Train_loss:0.335, Test_acc:78.9%, Test_loss:0.495, Lr:1.00E-08
Epoch:37, Train_acc:88.0%, Train_loss:0.332, Test_acc:78.9%, Test_loss:0.482, Lr:1.00E-08
Epoch:38, Train_acc:86.7%, Train_loss:0.341, Test_acc:78.9%, Test_loss:0.569, Lr:1.00E-08
Epoch:39, Train_acc:86.3%, Train_loss:0.340, Test_acc:78.9%, Test_loss:0.453, Lr:1.00E-08
Epoch:40, Train_acc:86.1%, Train_loss:0.357, Test_acc:78.9%, Test_loss:0.526, Lr:1.00E-08
Done

七、总结

一般而言，在LR不变的情况下，Batch Size越大，模型收敛效果越差。
下次可以试试更改LR和Batch Size 参数，检测模型收敛效果。
根据自己定义的函数更新学习率时，lamda=(0.92 ** (epoch // 2)), 每两个epochs下降到原来学习率的92%时，可以使测试集accuracy达到84%。

第P5周：Pytorch实现运动鞋识别

第P5周：Pytorch实现运动鞋识别

一、前期准备

1. 设置GPU

2. 导入数据

二、构建简单的CNN网络

训练函数

1. 编写训练函数

2. 编写测试函数

3. 设置动态学习率

4. 正式训练

四、 结果可视化

1. Loss 与 Accurary 图

2. 指定图片进行预测

五、保存并加载模型

六、 动态学习率

七、总结

四、结果可视化

六、动态学习率