一，前言(引用)
>- **🍨 本文为[🔗365天深度学习训练营](https://mp.weixin.qq.com/s/Z9yL_wt7L8aPOr9Lqb1K3w) 中的学习记录博客**
>- **🍖 原作者：[K同学啊](https://mtyjkh.blog.csdn.net/)**二，训练准备
1.训练环境
- Framework: Pytorch, matplotlib,numpy
- Compiler：Jupyter Lab
- Cpu: AMD Ryzen 5600H
2.数据集
两组数据Adidas和Nike鞋图片，各有251张作为Training Data, 38张作为Test Data.
三，训练过程
初始化:
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")
导入数据:
data_dir = './46-data/'
data_dir = pathlib.Path(data_dir)
data_paths = list(data_dir.glob('*'))
classeNames = [str(path).split("\\")[1] for path in data_paths]['test', 'train']
- 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。 - 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
数据集：
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("./46-data/train/",transform=train_transforms)
test_dataset = datasets.ImageFolder("./46-data/test/",transform=train_transforms)对应关系
train_dataset.class_to_idx 得到{'adidas': 0, 'nike': 1}
transforms.Compose() 这个类的主要作用是串联多个图片变换的操作。Compose — Torchvision 0.19 documentation
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)- dataset（必需参数）：这是你的数据集对象，通常是
torch.utils.data.Dataset的子类，它包含了你的数据样本。 - batch_size（可选参数）：指定每个小批次中包含的样本数。默认值为 1。
- shuffle（可选参数）：如果设置为
True，则在每个 epoch 开始时对数据进行洗牌，以随机打乱样本的顺序。这对于训练数据的随机性很重要，以避免模型学习到数据的顺序性。默认值为False。 - num_workers（可选参数）：用于数据加载的子进程数量。通常，将其设置为大于 0 的值可以加快数据加载速度，特别是当数据集很大时。默认值为 0，表示在主进程中加载数据。
打印
for X, y in test_dl:
print("Shape of X [N, C, H, W]: ", X.shape)
print("Shape of y: ", y.shape, y.dtype)
breakShape of X [N, C, H, W]: torch.Size([32, 3, 224, 224])
Shape of y: torch.Size([32]) torch.int64
设计CNN训练模型:
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 输入通道3，输出通道12，卷积核大小5*5，无填充
nn.BatchNorm2d(12), # 批量归一化，用于稳定训练过程
nn.ReLU()) # ReLU 激活函数
self.conv2 = nn.Sequential(
nn.Conv2d(12, 12, kernel_size=5, padding=0), # 12*216*216 第二层卷积，保持输入和输出通道一致
nn.BatchNorm2d(12), # 批量归一化
nn.ReLU())
self.pool3 = nn.Sequential(
nn.MaxPool2d(2)) # 12*108*108 最大池化，池化核大小为2*2，减少空间维度
self.conv4 = nn.Sequential(
nn.Conv2d(12, 24, kernel_size=5, padding=0), # 24*104*104 卷积层，输入通道12，输出通道24
nn.BatchNorm2d(24), # 批量归一化
nn.ReLU())
self.conv5 = nn.Sequential(
nn.Conv2d(24, 24, kernel_size=5, padding=0), # 24*100*100 卷积层，保持通道不变
nn.BatchNorm2d(24), # 批量归一化
nn.ReLU())
self.pool6 = nn.Sequential(
nn.MaxPool2d(2)) # 24*50*50 最大池化，进一步减少空间维度
self.dropout = nn.Sequential(
nn.Dropout(0.2)) # 随机丢弃20%的神经元，防止过拟合
self.fc = nn.Sequential(
nn.Linear(24*50*50, len(classeNames))) # 全连接层，将24*50*50展平为一维，并映射到类别数的输出层
def forward(self, x):
batch_size = x.size(0) # 获取输入的batch大小
x = self.conv1(x) # 第一层卷积
x = self.conv2(x) # 第二层卷积
x = self.pool3(x) # 第一次池化
x = self.conv4(x) # 第三层卷积
x = self.conv5(x) # 第四层卷积
x = self.pool6(x) # 第二次池化
x = self.dropout(x) # 随机丢弃
x = x.view(batch_size, -1) # 展平张量，准备输入全连接层 (batch, 24*50*50) ==> (batch, -1)
x = self.fc(x) # 全连接层输出
return x
device = "cpu"
print("Using {} device".format(device))
model = Model().to(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测试循环：
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动态学习率：
学习率调度器简介
- 在训练深度学习模型时，动态调整学习率可以帮助模型更好地收敛。通常，开始时用较大的学习率进行快速学习，随着训练的进行逐渐降低学习率，以避免错过局部最优解。
torch.optim.lr_scheduler提供了一些工具来帮助动态调整学习率。
使用 LambdaLR 调度器
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)
lambda1是一个定义学习率调整方式的函数，随训练进行，学习率每 2 个 epoch 按照衰减因子0.92递减。torch.optim.SGD()是定义优化器，用于调整模型参数。这里使用的初始学习率是learn_rate = 1e-4，即0.0001。LambdaLR调度器接收optimizer以及lr_lambda这个学习率衰减函数，用来调整学习率。
工作原理：
lambda1是一个函数，定义为(0.92 ** (epoch // 2))，表示每 2 个 epoch，学习率按 0.92 的比例衰减。- 当模型每训练两个 epoch 后，学习率会被乘以 0.92。例如：
- 第 0 和 1 个 epoch，学习率不变。
- 第 2 和 3 个 epoch，学习率变为
0.92 * 初始学习率。 - 第 4 和 5 个 epoch，学习率再乘以 0.92，如此循环递减。
手动调整学习率：
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
该函数通过手动调整学习率的方式实现相同的效果：
epoch代表当前训练的 epoch 数。start_lr是初始学习率。- 该函数每 2 个 epoch 将学习率调整为之前的
92%，通过遍历优化器中的param_groups，将新的学习率赋值给每个参数组。
训练过程：
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:85.9%, Train_loss:0.388, Test_acc:77.6%, Test_loss:0.481, Lr:1.00E-04
Epoch: 2, Train_acc:89.0%, Train_loss:0.369, Test_acc:80.3%, Test_loss:0.436, Lr:1.00E-04
Epoch: 3, Train_acc:90.0%, Train_loss:0.351, Test_acc:80.3%, Test_loss:0.446, Lr:9.20E-05
Epoch: 4, Train_acc:90.4%, Train_loss:0.337, Test_acc:81.6%, Test_loss:0.463, Lr:9.20E-05
Epoch: 5, Train_acc:92.0%, Train_loss:0.323, Test_acc:81.6%, Test_loss:0.474, Lr:8.46E-05
Epoch: 6, Train_acc:91.4%, Train_loss:0.316, Test_acc:80.3%, Test_loss:0.472, Lr:8.46E-05
Epoch: 7, Train_acc:92.2%, Train_loss:0.311, Test_acc:80.3%, Test_loss:0.462, Lr:7.79E-05
Epoch: 8, Train_acc:91.6%, Train_loss:0.309, Test_acc:80.3%, Test_loss:0.450, Lr:7.79E-05
Epoch: 9, Train_acc:94.6%, Train_loss:0.288, Test_acc:82.9%, Test_loss:0.477, Lr:7.16E-05
Epoch:10, Train_acc:93.2%, Train_loss:0.291, Test_acc:80.3%, Test_loss:0.434, Lr:7.16E-05
Epoch:11, Train_acc:94.0%, Train_loss:0.279, Test_acc:80.3%, Test_loss:0.427, Lr:6.59E-05
Epoch:12, Train_acc:95.4%, Train_loss:0.272, Test_acc:82.9%, Test_loss:0.473, Lr:6.59E-05
Epoch:13, Train_acc:95.0%, Train_loss:0.266, Test_acc:80.3%, Test_loss:0.417, Lr:6.06E-05
Epoch:14, Train_acc:95.6%, Train_loss:0.254, Test_acc:80.3%, Test_loss:0.431, Lr:6.06E-05
Epoch:15, Train_acc:96.4%, Train_loss:0.242, Test_acc:80.3%, Test_loss:0.430, Lr:5.58E-05
Epoch:16, Train_acc:95.4%, Train_loss:0.250, Test_acc:81.6%, Test_loss:0.442, Lr:5.58E-05
Epoch:17, Train_acc:96.2%, Train_loss:0.239, Test_acc:82.9%, Test_loss:0.454, Lr:5.13E-05
Epoch:18, Train_acc:96.4%, Train_loss:0.245, Test_acc:80.3%, Test_loss:0.415, Lr:5.13E-05
Epoch:19, Train_acc:95.6%, Train_loss:0.238, Test_acc:80.3%, Test_loss:0.432, Lr:4.72E-05
Epoch:20, Train_acc:96.0%, Train_loss:0.236, Test_acc:80.3%, Test_loss:0.424, Lr:4.72E-05
Epoch:21, Train_acc:96.4%, Train_loss:0.222, Test_acc:81.6%, Test_loss:0.455, Lr:4.34E-05
Epoch:22, Train_acc:96.6%, Train_loss:0.227, Test_acc:80.3%, Test_loss:0.403, Lr:4.34E-05
Epoch:23, Train_acc:96.8%, Train_loss:0.227, Test_acc:81.6%, Test_loss:0.420, Lr:4.00E-05
Epoch:24, Train_acc:97.4%, Train_loss:0.222, Test_acc:81.6%, Test_loss:0.385, Lr:4.00E-05
Epoch:25, Train_acc:97.8%, Train_loss:0.207, Test_acc:81.6%, Test_loss:0.440, Lr:3.68E-05
Epoch:26, Train_acc:96.6%, Train_loss:0.212, Test_acc:81.6%, Test_loss:0.403, Lr:3.68E-05
Epoch:27, Train_acc:97.6%, Train_loss:0.206, Test_acc:81.6%, Test_loss:0.394, Lr:3.38E-05
Epoch:28, Train_acc:96.2%, Train_loss:0.211, Test_acc:81.6%, Test_loss:0.409, Lr:3.38E-05
Epoch:29, Train_acc:97.2%, Train_loss:0.210, Test_acc:81.6%, Test_loss:0.466, Lr:3.11E-05
Epoch:30, Train_acc:97.2%, Train_loss:0.206, Test_acc:81.6%, Test_loss:0.396, Lr:3.11E-05
Epoch:31, Train_acc:97.4%, Train_loss:0.207, Test_acc:81.6%, Test_loss:0.410, Lr:2.86E-05
Epoch:32, Train_acc:97.0%, Train_loss:0.207, Test_acc:81.6%, Test_loss:0.427, Lr:2.86E-05
Epoch:33, Train_acc:97.4%, Train_loss:0.211, Test_acc:81.6%, Test_loss:0.419, Lr:2.63E-05
Epoch:34, Train_acc:96.8%, Train_loss:0.205, Test_acc:81.6%, Test_loss:0.399, Lr:2.63E-05
Epoch:35, Train_acc:98.4%, Train_loss:0.199, Test_acc:84.2%, Test_loss:0.436, Lr:2.42E-05
Epoch:36, Train_acc:97.6%, Train_loss:0.199, Test_acc:81.6%, Test_loss:0.394, Lr:2.42E-05
Epoch:37, Train_acc:97.6%, Train_loss:0.194, Test_acc:82.9%, Test_loss:0.414, Lr:2.23E-05
Epoch:38, Train_acc:97.6%, Train_loss:0.194, Test_acc:82.9%, Test_loss:0.387, Lr:2.23E-05
Epoch:39, Train_acc:97.6%, Train_loss:0.193, Test_acc:81.6%, Test_loss:0.427, Lr:2.05E-05
Epoch:40, Train_acc:97.0%, Train_loss:0.197, Test_acc:81.6%, Test_loss:0.411, Lr:2.05E-05
Done
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()
验证示例：
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='./46-data/test/adidas/1.jpg',
model=model,
transform=train_transforms,
classes=classes)预测结果是：adidas
总结
本次实验中，掌握了在训练深度学习模型的过程中，如何使用 PyTorch 进行数据处理、模型构建、训练与测试的全流程操作。通过对 Adidas 和 Nike 两个鞋类图像分类任务的实践，理解了卷积神经网络（CNN）的构建方式，及其在图像分类中的作用。
在训练过程中，重点关注了超参数设置和动态学习率调整的效果。通过使用 PyTorch 提供的学习率调度器以及手动调整学习率的策略，探索了不同学习率对模型收敛速度和准确率的影响。在训练和测试阶段，通过绘制 Loss 和 Accuracy 的变化曲线，清晰地展示了模型的收敛趋势。
此外，通过增加图片预测模块，掌握了如何使用训练好的模型对新的图片进行类别预测，并成功实现了模型的保存与加载。这些操作有助于深入理解深度学习中的模型管理与优化技巧。通过实验对比学习率调整的效果，验证了动态学习率对模型最终性能的提升作用。
扫一扫
325
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
