麻醉医生的深度学习之旅 P5：Pytorch实现运动鞋识别-CSDN博客

本文链接：https://blog.csdn.net/u012249506/article/details/139547298

🍨 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

五月初五迎端午，今天是假期的第一天，想必大家有关于这个节日的记忆都是粽子和龙舟。而端午的由来大多数人将它视为纪念爱国诗人屈原，而屈原最为人所知的便是《离骚》中的“路漫漫其修远兮，吾将上下而求索”。而我们正在求索的路上不断前行。

今天尝试用Jupyter失败，在进行到可视化步骤的时候会出现"The kernel has died, and the automatic restart has failed"，未能找到原因，随后改用Google colab完成。在第一次循环的对比中，CPU和GPU的测试准确率差距似乎较大，有没有大佬解答一下这个问题❓

请添加图片描述

实验目的：

要求：了解如何设置动态学习率（重点），调整代码使测试集accuracy到达84%
拔高：保存训练过程中的最佳模型权重，调整代码使测试集accuracy到达86%

实验环境：

语言环境：python 3.8
编译器：Google colab
深度学习环境：Pytorch

实验流程：

1. 设置GPU

import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms

from torchvision import transforms, datasets

import os,PIL,pathlib

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

device

🪧代码输出
device(type='cuda')

2. 导入数据

前边几次实验使用这条代码最后输出classeNames都没问题，但是这次将这段代码"classeNames = [str(path).split(“\”)[1] for path in data_paths]"改成下面的才能输出下一级目录的名字

import os,PIL,random,pathlib

data_dir = '/content/drive/MyDrive/P5'
data_dir = pathlib.Path(data_dir)

data_paths  = list(data_dir.glob('*'))
# Use the correct index to extract the class name or use a more robust method
classeNames = [path.name for path in data_paths] 
print(classeNames)

🪧代码输出
['test', 'train']

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。
])

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。
])

train_dataset = datasets.ImageFolder("/content/drive/MyDrive/P5/train/",transform=train_transforms)
test_dataset  = datasets.ImageFolder("/content/drive/MyDrive/P5/test/",transform=train_transforms)

train_dataset.class_to_idx

🪧代码输出
{'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

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

3. 构建CNN网络

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

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)
  )
)

4. 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

5. 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

6. 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)

7. 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')

🪧代码输出
Epoch: 1, Train_acc:55.2%, Train_loss:0.763, Test_acc:53.9%, Test_loss:0.691, Lr:1.00E-04
Epoch: 2, Train_acc:58.0%, Train_loss:0.716, Test_acc:56.6%, Test_loss:0.658, Lr:1.00E-04
Epoch: 3, Train_acc:64.5%, Train_loss:0.615, Test_acc:57.9%, Test_loss:0.745, Lr:9.20E-05
Epoch: 4, Train_acc:71.3%, Train_loss:0.562, Test_acc:67.1%, Test_loss:0.607, Lr:9.20E-05
Epoch: 5, Train_acc:75.3%, Train_loss:0.526, Test_acc:71.1%, Test_loss:0.565, Lr:8.46E-05
Epoch: 6, Train_acc:77.3%, Train_loss:0.502, Test_acc:69.7%, Test_loss:0.586, Lr:8.46E-05
Epoch: 7, Train_acc:78.5%, Train_loss:0.476, Test_acc:71.1%, Test_loss:0.569, Lr:7.79E-05
Epoch: 8, Train_acc:81.9%, Train_loss:0.443, Test_acc:71.1%, Test_loss:0.591, Lr:7.79E-05
Epoch: 9, Train_acc:85.1%, Train_loss:0.424, Test_acc:71.1%, Test_loss:0.552, Lr:7.16E-05
Epoch:10, Train_acc:84.9%, Train_loss:0.411, Test_acc:72.4%, Test_loss:0.521, Lr:7.16E-05
Epoch:11, Train_acc:85.7%, Train_loss:0.394, Test_acc:73.7%, Test_loss:0.519, Lr:6.59E-05
Epoch:12, Train_acc:87.3%, Train_loss:0.386, Test_acc:76.3%, Test_loss:0.534, Lr:6.59E-05
Epoch:13, Train_acc:87.1%, Train_loss:0.374, Test_acc:75.0%, Test_loss:0.509, Lr:6.06E-05
Epoch:14, Train_acc:90.0%, Train_loss:0.354, Test_acc:77.6%, Test_loss:0.506, Lr:6.06E-05
Epoch:15, Train_acc:89.6%, Train_loss:0.347, Test_acc:77.6%, Test_loss:0.487, Lr:5.58E-05
Epoch:16, Train_acc:89.8%, Train_loss:0.355, Test_acc:75.0%, Test_loss:0.528, Lr:5.58E-05
Epoch:17, Train_acc:90.2%, Train_loss:0.338, Test_acc:76.3%, Test_loss:0.505, Lr:5.13E-05
Epoch:18, Train_acc:90.6%, Train_loss:0.321, Test_acc:76.3%, Test_loss:0.475, Lr:5.13E-05
Epoch:19, Train_acc:92.2%, Train_loss:0.317, Test_acc:76.3%, Test_loss:0.497, Lr:4.72E-05
Epoch:20, Train_acc:92.4%, Train_loss:0.311, Test_acc:76.3%, Test_loss:0.488, Lr:4.72E-05
Epoch:21, Train_acc:92.6%, Train_loss:0.300, Test_acc:76.3%, Test_loss:0.457, Lr:4.34E-05
Epoch:22, Train_acc:93.0%, Train_loss:0.301, Test_acc:77.6%, Test_loss:0.490, Lr:4.34E-05
Epoch:23, Train_acc:94.2%, Train_loss:0.290, Test_acc:76.3%, Test_loss:0.498, Lr:4.00E-05
Epoch:24, Train_acc:93.6%, Train_loss:0.286, Test_acc:77.6%, Test_loss:0.481, Lr:4.00E-05
Epoch:25, Train_acc:93.8%, Train_loss:0.286, Test_acc:76.3%, Test_loss:0.490, Lr:3.68E-05
Epoch:26, Train_acc:93.6%, Train_loss:0.283, Test_acc:77.6%, Test_loss:0.495, Lr:3.68E-05
Epoch:27, Train_acc:94.4%, Train_loss:0.285, Test_acc:77.6%, Test_loss:0.485, Lr:3.38E-05
Epoch:28, Train_acc:94.8%, Train_loss:0.274, Test_acc:80.3%, Test_loss:0.485, Lr:3.38E-05
Epoch:29, Train_acc:93.2%, Train_loss:0.278, Test_acc:77.6%, Test_loss:0.450, Lr:3.11E-05
Epoch:30, Train_acc:94.4%, Train_loss:0.281, Test_acc:77.6%, Test_loss:0.455, Lr:3.11E-05
Epoch:31, Train_acc:94.0%, Train_loss:0.268, Test_acc:80.3%, Test_loss:0.460, Lr:2.86E-05
Epoch:32, Train_acc:94.8%, Train_loss:0.265, Test_acc:76.3%, Test_loss:0.425, Lr:2.86E-05
Epoch:33, Train_acc:95.8%, Train_loss:0.260, Test_acc:76.3%, Test_loss:0.449, Lr:2.63E-05
Epoch:34, Train_acc:94.4%, Train_loss:0.271, Test_acc:77.6%, Test_loss:0.435, Lr:2.63E-05
Epoch:35, Train_acc:94.2%, Train_loss:0.261, Test_acc:77.6%, Test_loss:0.467, Lr:2.42E-05
Epoch:36, Train_acc:94.8%, Train_loss:0.254, Test_acc:77.6%, Test_loss:0.429, Lr:2.42E-05
Epoch:37, Train_acc:96.2%, Train_loss:0.251, Test_acc:76.3%, Test_loss:0.419, Lr:2.23E-05
Epoch:38, Train_acc:95.6%, Train_loss:0.253, Test_acc:77.6%, Test_loss:0.459, Lr:2.23E-05
Epoch:39, Train_acc:97.0%, Train_loss:0.241, Test_acc:77.6%, Test_loss:0.456, Lr:2.05E-05
Epoch:40, Train_acc:96.0%, Train_loss:0.252, Test_acc:77.6%, Test_loss:0.465, Lr:2.05E-05
Done

8. 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(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()

请添加图片描述

9. 指定图片进行预测

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='/content/drive/MyDrive/P5/test/adidas/9.jpg', 
                  model=model, 
                  transform=train_transforms, 
                  classes=classes)

🪧代码输出
预测结果是：adidas

10. 提高测试集accuracy

第一次尝试：提高初始学习率，降低训练轮数

对比第一次的实验可以看到测试集准确率有了较大的提升

learn_rate = 1e-3
epochs     = 20

提高初始学习率可能会在一定程度上加快训练初期的收敛速度，但如果过高可能导致模型不稳定，出现震荡甚至不收敛的情况，反而不利于提高准确率
降低训练轮数可能使得模型没有足够的时间充分学习数据中的模式和特征，导致模型欠拟合，从而降低测试集准确率；但在某些情况下，如果训练轮数过多导致过拟合，适当降低训练轮数可能会缓解过拟合，对测试集准确率有一定帮助

Epoch: 1, Train_acc:54.4%, Train_loss:3.025, Test_acc:50.0%, Test_loss:0.948, Lr:1.00E-03
Epoch: 2, Train_acc:58.2%, Train_loss:1.914, Test_acc:50.0%, Test_loss:1.556, Lr:1.00E-03
Epoch: 3, Train_acc:54.8%, Train_loss:1.720, Test_acc:72.4%, Test_loss:0.804, Lr:9.20E-04
Epoch: 4, Train_acc:75.1%, Train_loss:0.665, Test_acc:50.0%, Test_loss:2.124, Lr:9.20E-04
Epoch: 5, Train_acc:77.7%, Train_loss:0.568, Test_acc:73.7%, Test_loss:0.769, Lr:8.46E-04
Epoch: 6, Train_acc:81.9%, Train_loss:0.408, Test_acc:64.5%, Test_loss:0.964, Lr:8.46E-04
Epoch: 7, Train_acc:87.8%, Train_loss:0.294, Test_acc:75.0%, Test_loss:0.434, Lr:7.79E-04
Epoch: 8, Train_acc:92.0%, Train_loss:0.220, Test_acc:76.3%, Test_loss:0.412, Lr:7.79E-04
Epoch: 9, Train_acc:93.6%, Train_loss:0.191, Test_acc:81.6%, Test_loss:0.426, Lr:7.16E-04
Epoch:10, Train_acc:96.2%, Train_loss:0.151, Test_acc:84.2%, Test_loss:0.355, Lr:7.16E-04
Epoch:11, Train_acc:94.8%, Train_loss:0.166, Test_acc:84.2%, Test_loss:0.364, Lr:6.59E-04
Epoch:12, Train_acc:97.0%, Train_loss:0.140, Test_acc:81.6%, Test_loss:0.383, Lr:6.59E-04
Epoch:13, Train_acc:97.4%, Train_loss:0.122, Test_acc:81.6%, Test_loss:0.448, Lr:6.06E-04
Epoch:14, Train_acc:98.4%, Train_loss:0.109, Test_acc:85.5%, Test_loss:0.342, Lr:6.06E-04
Epoch:15, Train_acc:98.6%, Train_loss:0.108, Test_acc:81.6%, Test_loss:0.432, Lr:5.58E-04
Epoch:16, Train_acc:97.2%, Train_loss:0.117, Test_acc:85.5%, Test_loss:0.413, Lr:5.58E-04
Epoch:17, Train_acc:98.4%, Train_loss:0.094, Test_acc:82.9%, Test_loss:0.373, Lr:5.13E-04
Epoch:18, Train_acc:98.6%, Train_loss:0.099, Test_acc:86.8%, Test_loss:0.394, Lr:5.13E-04
Epoch:19, Train_acc:98.8%, Train_loss:0.090, Test_acc:85.5%, Test_loss:0.354, Lr:4.72E-04
Epoch:20, Train_acc:99.0%, Train_loss:0.088, Test_acc:84.2%, Test_loss:0.430, Lr:4.72E-04
Done

请添加图片描述

为使测试集accuracy达到86%，进行第二次尝试：

更改优化器，训练轮数调整回40轮，结果示测试集准确率基本稳定在85.5%，偶尔能提升到86.8%

opt     = torch.optim.Adam(model.parameters(), lr=learn_rate)
epochs  = 40

Epoch: 1, Train_acc:100.0%, Train_loss:0.020, Test_acc:85.5%, Test_loss:0.351, Lr:1.00E-03
Epoch: 2, Train_acc:100.0%, Train_loss:0.020, Test_acc:86.8%, Test_loss:0.463, Lr:1.00E-03
Epoch: 3, Train_acc:100.0%, Train_loss:0.018, Test_acc:85.5%, Test_loss:0.427, Lr:9.20E-04
Epoch: 4, Train_acc:100.0%, Train_loss:0.017, Test_acc:85.5%, Test_loss:0.402, Lr:9.20E-04
Epoch: 5, Train_acc:100.0%, Train_loss:0.018, Test_acc:85.5%, Test_loss:0.319, Lr:8.46E-04
Epoch: 6, Train_acc:100.0%, Train_loss:0.016, Test_acc:85.5%, Test_loss:0.385, Lr:8.46E-04
Epoch: 7, Train_acc:100.0%, Train_loss:0.015, Test_acc:85.5%, Test_loss:0.384, Lr:7.79E-04
Epoch: 8, Train_acc:100.0%, Train_loss:0.016, Test_acc:85.5%, Test_loss:0.414, Lr:7.79E-04
Epoch: 9, Train_acc:100.0%, Train_loss:0.016, Test_acc:85.5%, Test_loss:0.449, Lr:7.16E-04
Epoch:10, Train_acc:100.0%, Train_loss:0.016, Test_acc:85.5%, Test_loss:0.395, Lr:7.16E-04
Epoch:11, Train_acc:100.0%, Train_loss:0.016, Test_acc:85.5%, Test_loss:0.362, Lr:6.59E-04
Epoch:12, Train_acc:100.0%, Train_loss:0.014, Test_acc:85.5%, Test_loss:0.387, Lr:6.59E-04
Epoch:13, Train_acc:100.0%, Train_loss:0.015, Test_acc:86.8%, Test_loss:0.308, Lr:6.06E-04
Epoch:14, Train_acc:100.0%, Train_loss:0.016, Test_acc:85.5%, Test_loss:0.344, Lr:6.06E-04
Epoch:15, Train_acc:100.0%, Train_loss:0.014, Test_acc:85.5%, Test_loss:0.380, Lr:5.58E-04
Epoch:16, Train_acc:100.0%, Train_loss:0.015, Test_acc:85.5%, Test_loss:0.458, Lr:5.58E-04
Epoch:17, Train_acc:100.0%, Train_loss:0.014, Test_acc:85.5%, Test_loss:0.377, Lr:5.13E-04
Epoch:18, Train_acc:100.0%, Train_loss:0.014, Test_acc:85.5%, Test_loss:0.393, Lr:5.13E-04
Epoch:19, Train_acc:100.0%, Train_loss:0.015, Test_acc:86.8%, Test_loss:0.359, Lr:4.72E-04
Epoch:20, Train_acc:100.0%, Train_loss:0.013, Test_acc:85.5%, Test_loss:0.394, Lr:4.72E-04
Epoch:21, Train_acc:100.0%, Train_loss:0.015, Test_acc:85.5%, Test_loss:0.527, Lr:4.34E-04
Epoch:22, Train_acc:100.0%, Train_loss:0.013, Test_acc:85.5%, Test_loss:0.334, Lr:4.34E-04
Epoch:23, Train_acc:100.0%, Train_loss:0.013, Test_acc:85.5%, Test_loss:0.389, Lr:4.00E-04
Epoch:24, Train_acc:100.0%, Train_loss:0.012, Test_acc:85.5%, Test_loss:0.403, Lr:4.00E-04
Epoch:25, Train_acc:100.0%, Train_loss:0.014, Test_acc:85.5%, Test_loss:0.374, Lr:3.68E-04
Epoch:26, Train_acc:100.0%, Train_loss:0.013, Test_acc:85.5%, Test_loss:0.437, Lr:3.68E-04
Epoch:27, Train_acc:100.0%, Train_loss:0.014, Test_acc:85.5%, Test_loss:0.428, Lr:3.38E-04
Epoch:28, Train_acc:100.0%, Train_loss:0.014, Test_acc:85.5%, Test_loss:0.362, Lr:3.38E-04
Epoch:29, Train_acc:100.0%, Train_loss:0.013, Test_acc:85.5%, Test_loss:0.394, Lr:3.11E-04
Epoch:30, Train_acc:100.0%, Train_loss:0.012, Test_acc:85.5%, Test_loss:0.393, Lr:3.11E-04
Epoch:31, Train_acc:100.0%, Train_loss:0.014, Test_acc:85.5%, Test_loss:0.434, Lr:2.86E-04
Epoch:32, Train_acc:100.0%, Train_loss:0.011, Test_acc:85.5%, Test_loss:0.369, Lr:2.86E-04
Epoch:33, Train_acc:100.0%, Train_loss:0.013, Test_acc:85.5%, Test_loss:0.421, Lr:2.63E-04
Epoch:34, Train_acc:100.0%, Train_loss:0.013, Test_acc:85.5%, Test_loss:0.450, Lr:2.63E-04
Epoch:35, Train_acc:100.0%, Train_loss:0.013, Test_acc:86.8%, Test_loss:0.471, Lr:2.42E-04
Epoch:36, Train_acc:100.0%, Train_loss:0.012, Test_acc:85.5%, Test_loss:0.366, Lr:2.42E-04
Epoch:37, Train_acc:100.0%, Train_loss:0.013, Test_acc:85.5%, Test_loss:0.543, Lr:2.23E-04
Epoch:38, Train_acc:100.0%, Train_loss:0.015, Test_acc:85.5%, Test_loss:0.389, Lr:2.23E-04
Epoch:39, Train_acc:100.0%, Train_loss:0.013, Test_acc:86.8%, Test_loss:0.367, Lr:2.05E-04
Epoch:40, Train_acc:100.0%, Train_loss:0.012, Test_acc:86.8%, Test_loss:0.386, Lr:2.05E-04
Done

请添加图片描述

11. 保存并加载模型

# 模型保存
PATH = './P5-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))

🪧代码输出
<All keys matched successfully>

保存最佳模型：

best_acc = 0
# 保存最佳模型到 best_model
if epoch_test_acc > best_acc:
    best_acc = epoch_test_acc
    torch.save(model.state_dict(), 'best_model.pth')  # 保存模型的状态字典

12. 总结

本次实验的重点任务是设置动态学习率，在训练过程中学习率不是固定不变的，而是根据一定的策略或规则进行动态调整。

在与P4进行对比过程中我发现：设置超参这一步在P4中是放在编写函数之前的，但在P5中放到了编写函数之后，且损失函数放到了正式训练中，不知这样调整顺序是何用意？另外对于保存最佳模型的理解还不够充分，不知上述代码是否正确。

在训练初期，较大的学习率可以帮助模型快速探索参数空间，加速收敛。但随着训练的进行，模型逐渐接近最优解，过大的学习率可能会导致模型不稳定或错过最优解。此时，通过动态调整学习率，比如逐渐减小学习率，可以使模型更精细地调整参数，提高收敛的准确性和稳定性。

常见的动态学习率调整策略有：

阶梯式下降：每隔一定的步数或轮数，将学习率降低一定比例。
指数衰减：学习率按照指数函数的形式逐渐减小。
余弦退火等：学习率呈余弦函数形式变化。

采用动态学习率策略有助于更好地平衡训练速度和精度，提高模型的训练效果。

💡要求：

了解如何设置动态学习率（重点）（完成✅）
调整代码使测试集accuracy到达84%（完成✅）

🧗‍♂️拔高：

保存训练过程中的最佳模型权重（完成✅）
调整代码使测试集accuracy到达86%（完成✅）最高可达86.8%）

🚩动态学习率

1. 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：

optimizer = torch.optim.SGD(net.parameters(), lr=0.001 )
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.1)

2. 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：

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

3. 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：

optimizer = torch.optim.SGD(net.parameters(), lr=0.001 )
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, 
                                                 milestones=[2,6,15], #调整学习率的epoch数
                                                 gamma=0.1)