P5训练营：PyTorch运动鞋图片识别

第五周：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

我的环境

• 语言环境: 3.10.12
• 编译器：Google Colab（T4）
• 深度学习环境：PyTorch - torch:2.3.0+cu121; torchvision:0.18.0+cu121

一、前期准备

设置GPU

无gpu则使用cpu

from google.colab import drive
drive.mount("/content/drive/")

%cd "/content/drive/MyDrive/Colab Notebooks/jupyter notebook/data/P5"

Mounted at /content/drive/
/content/drive/Othercomputers/My laptop/jupyter notebook/data/P5

import torch
import torch.nn as nn
import torchvision.transforms as transforms
import torchvision
from torchvision import transforms, datasets, models

import os,PIL,pathlib,random

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

device

device(type='cuda')

导入数据

import random

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

data_paths  = list(data_dir.glob('*'))
classNames = [str(path).split("\\")[0] for path in data_paths]

classNames

['test', 'train']

# 关于transforms.Compose的更多介绍可以参考：https://blog.csdn.net/qq_38251616/article/details/124878863
train_transform = 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("./train/",transform=train_transform)
test_dataset  = datasets.ImageFolder("./test/",transform=test_transform)

print(train_dataset)

Dataset ImageFolder
        Number of datapoints: 502
        Root location: ./train/
        StandardTransform
    Transform: Compose(
                   Resize(size=[224, 224], interpolation=bilinear, max_size=None, antialias=True)
                   ToTensor()
                   Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
               )

print(test_dataset)

   Dataset ImageFolder
        Number of datapoints: 76
        Root location: ./test/
        StandardTransform
    Transform: Compose(
                   Resize(size=[224, 224], interpolation=bilinear, max_size=None, antialias=True)
                   ToTensor()
                   Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
               )

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 if y:", y.shape, y.dtype)
  break

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

二、构建简单的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(classNames)))

    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)
      )
    )

三、训练模型

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

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

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

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:53.4%, Train_loss:0.731, Test_acc:57.9%, Test_loss:0.679, Lr:1.00E-04
    Epoch: 2, Train_acc:62.4%, Train_loss:0.691, Test_acc:64.5%, Test_loss:0.608, Lr:1.00E-04
    Epoch: 3, Train_acc:63.1%, Train_loss:0.654, Test_acc:64.5%, Test_loss:0.583, Lr:9.20E-05
    Epoch: 4, Train_acc:68.5%, Train_loss:0.574, Test_acc:67.1%, Test_loss:0.623, Lr:9.20E-05
    ... ...
    Epoch:36, Train_acc:94.0%, Train_loss:0.286, Test_acc:85.5%, Test_loss:0.469, Lr:2.42E-05
    Epoch:37, Train_acc:95.0%, Train_loss:0.287, Test_acc:82.9%, Test_loss:0.476, Lr:2.23E-05
    Epoch:38, Train_acc:93.0%, Train_loss:0.289, Test_acc:82.9%, Test_loss:0.458, Lr:2.23E-05
    Epoch:39, Train_acc:94.4%, Train_loss:0.281, Test_acc:84.2%, Test_loss:0.438, Lr:2.05E-05
    Epoch:40, Train_acc:94.0%, Train_loss:0.286, Test_acc:84.2%, Test_loss:0.430, Lr:2.05E-05
    Done

1.model.train()

model.train()的作用是启用Batch Normalization和Dropout。

如果模型中有BN层(Batch Normalization）和Dropout，需要在训练时添加model.train()。model.train()是保证BN层能够用到每一批数据的均值和方差。对于Dropout，model.train()是随机取一部分网络连接来训练更新参数。

2.model.eval()

model.eval()的作用是不启用 Batch Normalization 和 Dropout。

model.eval()是保证BN层能够用全部训练数据的均值和方差，即测试过程中要保证BN层的均值和方差不变。对于Dropout，model.eval()是利用到了所有网络连接，即不进行随机舍弃神经元。

训练完train样本后，生成的模型model要用来测试样本。在model(test)之前，需要加上model.eval()，否则的话，有输入数据，即使不训练，它也会改变权值。这是model中含有BN层和Dropout所带来的的性质。

四、结果可视化

Loss和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()

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)
    # 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='./test/nike/33.jpg',
                  model=model,
                  transform=train_transform,
                  classes=classes)

预测结果是：nike

五、保存并加载模型

（经过修改后达到满足条件test acc的模型）

# 模型保存
PATH = './model_5.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、设置动态学习率

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

• 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

#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

• optimizer(Optimizer):是之前定义好的需要优化的优化器实例名
• lr_lambda(function):更新学习率的函数

# 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

• 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

#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)

# 调用官方接口示例：
model = [Parameter(torch.randn(2, 2, requires_grad=True))]
optimizer = SGD(model, 0.1)
scheduler = ExponentialLR(optimizer, gamma=0.9)

for epoch in range(20):
    for input, target in dataset:
        optimizer.zero_grad()
        output = model(input)
        loss = loss_fn(output, target)
        loss.backward()
        optimizer.step()
    scheduler.step()

六、修改并优化

1. 使用transforms.RandomHorizontalFlip()对训练数据进行随机水平翻转（数据增强）；在训练时，每获取一个train_dl就会操作一次，这可以使模型学习更多的特征并提升泛化能力
2. 将原模型中最后的nn.Dropout(0.2)层修改为两个nn.Dropout(0.1)
3. 优化器SGD改为Adam
4. 初始学习率改为4e-4，设置每3个epoch衰减0.92
5. 打印模型accuracy和loss并且保存40 epoch中最优结果的模型
   部分代码如下：

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.conv1=nn.Sequential(
            nn.Conv2d(3, 12, kernel_size=5, padding=0), 
            nn.BatchNorm2d(12),
            nn.ReLU())

        self.conv2=nn.Sequential(
            nn.Conv2d(12, 12, kernel_size=5, padding=0),
            nn.BatchNorm2d(12),
            nn.ReLU())

        self.pool3=nn.Sequential(
            nn.MaxPool2d(2))             
        
        self.dropout1 = nn.Dropout(0.1)

        self.conv4=nn.Sequential(
            nn.Conv2d(12, 24, kernel_size=5, padding=0),
            nn.BatchNorm2d(24),
            nn.ReLU())

        self.conv5=nn.Sequential(
            nn.Conv2d(24, 24, kernel_size=5, padding=0), 
            nn.BatchNorm2d(24),
            nn.ReLU())

        self.pool6=nn.Sequential(
            nn.MaxPool2d(2))     # 24*50*50

        self.dropout2 = nn.Sequential(
            nn.Dropout(0.1))

        self.fc=nn.Sequential(
            nn.Linear(24*50*50, len(classNames)))

    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.dropout1(x)
      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.dropout2(x)
      x = x.view(batch_size, -1) 
      x = self.fc(x)
      
      return x

# 调用官方动态学习率接口时使用
learn_rate = 4e-4
lambda1 = lambda epoch: (0.92 ** (epoch // 3))
optimizer = torch.optim.Adam(model.parameters(), lr=learn_rate)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda1) #选定调整方法

if epoch_test_acc > min_test_acc and epoch_train_acc > min_train_acc:
      min_test_acc = epoch_test_acc
      min_train_acc = epoch_train_acc
      print("save model")
      # 保存模型语句
      PATH = '/content/drive/Othercomputers/My laptop/jupyter notebook/xunlianying/5-model.pth'  # 保存的参数文件名
      torch.save(model.state_dict(), PATH)
# 版权声明：摘自重邮研究森原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
# 原文链接：https://blog.csdn.net/m0_60524373/article/details/127752401

训练结果：

save model
Epoch: 5, Train_acc:65.5%, Train_loss:1.074, Test_acc:73.7%, Test_loss:0.897, Lr:3.68E-04
save model
Epoch: 6, Train_acc:76.7%, Train_loss:0.671, Test_acc:78.9%, Test_loss:0.624, Lr:3.39E-04
save model
Epoch: 7, Train_acc:83.3%, Train_loss:0.388, Test_acc:82.9%, Test_loss:0.413, Lr:3.39E-04
... ...
save model
Epoch:29, Train_acc:99.8%, Train_loss:0.023, Test_acc:89.5%, Test_loss:0.280, Lr:1.89E-04
Epoch:30, Train_acc:100.0%, Train_loss:0.025, Test_acc:86.8%, Test_loss:0.461, Lr:1.74E-04
... ...
Epoch:36, Train_acc:100.0%, Train_loss:0.019, Test_acc:86.8%, Test_loss:0.320, Lr:1.47E-04
Epoch:37, Train_acc:100.0%, Train_loss:0.015, Test_acc:88.2%, Test_loss:0.342, Lr:1.47E-04
Epoch:38, Train_acc:100.0%, Train_loss:0.013, Test_acc:86.8%, Test_loss:0.352, Lr:1.47E-04
Epoch:39, Train_acc:100.0%, Train_loss:0.016, Test_acc:85.5%, Test_loss:0.371, Lr:1.35E-04
Epoch:40, Train_acc:100.0%, Train_loss:0.014, Test_acc:85.5%, Test_loss:0.311, Lr:1.35E-04
Done

最终保存的模型训练正确率达到99.8%，测试正确率达到89%

总结：

• 学习了动态学习率的设置方法-动态学习率可以帮助模型更快地收敛。在训练初期，大的学习率可以快速接近最优解；在训练后期，小的学习率可以更精细地优化模型
• 优化识别图像CNN模型train acc达到99%，test acc达到86%以上；

本菜鸡的感想：
根本上因为打好基础，对模型训练的各种条件和背后作用的原理了解不足，导致调整参数结构没有策略（无头苍蝇）；

这次大部分时间耗在优化模型上，首先，初始模型性能较好时再优化应该多微调而不是一下子改很多地方，经常导致“坏上加坏”（或者效果抵消了导致没啥变化）；
一开始只顾着复杂化模型，但不是越多特征和层数效果越好，得结合数据特征，使用的数据集小时不应用太复杂的网络结构（容易过拟合）。

调试过程中一开始还尝试使训练集的drop_last = True结果导致训练集的正确率一直上不去，我的推测是数据集本身就小如果刨除余下的不满batch的数量会导致每epoch学的数量更少，这种不适用数据集本身小的情况（不知道还有没有别的原因）

模型训练中训练集和测试集的正确率差不少，可能因为两者的数据量差距大（502vs76）