神经网络框架搭建及模型优化

1 数据及配置

---

1.1 配置

需要安装PyTorch,下载安装torch、torchvision、torchaudio，GPU需下载cuda版本，CPU可直接下载

cuda版本较大，最后通过控制面板pip install +存储地址离线下载，
CPU版本需再下载安装VC_redist.x64.exe，可下载上述三个后运行，通过报错网址直接下载安装

1.2 数据

使用的是 torchvision.datasets.MNIST的手写数据，包括特征数据和结果类别

1.3 函数导入

import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensor

1.4 数据函数

train_data = datasets.MNIST(root='data',        # 数据集存储的根目录train=True,         # 加载训练集download=True,      # 如果数据集不存在，自动下载transform=ToTensor() # 将图像转换为张量
)

• root 指定数据集存储的根目录。如果数据集不存在，会自动下载到这个目录。
• train 决定加载训练集还是测试集。True 表示加载训练集，False 表示加载测试集。
• download 如果数据集不在 root 指定的目录中，是否自动下载数据集。True 表示自动下载。
• transform 对加载的数据进行预处理或转换。通常用于将数据转换为模型所需的格式，如将图像转换为张量。

1.5 数据打包

train_dataloader = DataLoader(train_data, batch_size=64)

• train_data, 打包数据
• batch_size=64，打包个数

代码展示：

import torch
print(torch.__version__)
import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensortrain_data = datasets.MNIST(root = 'data',train = True,download = True,transform = ToTensor()
)
test_data = datasets.MNIST(root = 'data',train = False,download = True,transform = ToTensor()
)
print(len(train_data))
print(len(test_data))
from matplotlib import pyplot as plt
figure = plt.figure()
for i in range(9):img,label = train_data[i+59000]figure.add_subplot(3,3,i+1)plt.title(label)plt.axis('off')plt.imshow(img.squeeze(),cmap='gray')a = img.squeeze()
plt.show()train_dataloader = DataLoader(train_data, batch_size=64)
test_dataloader= DataLoader(test_data, batch_size=64)

运行结果：

调试查看:

2 神经网络框架搭建

---

2.1 框架确认

在搭建神经网络框架前，需先确认建立怎样的框架，目前并没有理论的指导，凭经验建立框架如下：

输入层：输入的图像数据（28*28）个神经元。
中间层1：全连接层，128个神经元，
中间层2：全连接层，256个神经元，
输出层：全连接层，10个神经元，对应10个类别。
需注意，中间层需使用激励函数激活，对累加数进行非线性的映射，以及forward前向传播过程的函数名不可更改，

2.2 函数搭建

• nn.Flatten() ， 将输入展平为一维向量
• nn.Linear(28*28, 128) ，全连接层，需注意每个连接层的输入输出需前后对应
• torch.sigmoid(x)，对中间层的输出应用Sigmoid激活函数

# 定义一个神经网络类，继承自 nn.Module
class NeuralNetwork(nn.Module):def __init__(self):super().__init__()  # 调用父类 nn.Module 的构造函数# 定义网络层self.flatten = nn.Flatten()  # 将输入展平为一维向量，适用于将图像数据（如28x28）展平为784维self.hidden1 = nn.Linear(28*28, 128)  # 第一个全连接层，输入维度为784（28*28），输出维度为128self.hidden2 = nn.Linear(128, 256)    # 第二个全连接层，输入维度为128，输出维度为256self.out = nn.Linear(256, 10)         # 输出层，输入维度为256，输出维度为10（对应10个类别）# 定义前向传播过程def forward(self, x):x = self.flatten(x)       # 将输入数据展平x = self.hidden1(x)       # 通过第一个全连接层x = torch.sigmoid(x)      # 对第一个全连接层的输出应用Sigmoid激活函数x = self.hidden2(x)       # 通过第二个全连接层x = torch.sigmoid(x)      # 对第二个全连接层的输出应用Sigmoid激活函数x = self.out(x)           # 通过输出层return x                  # 返回最终的输出

2.3 框架上传

• device = ‘cuda’ if torch.cuda.is_available() else ‘mps’ if torch.backends.mps.is_available() else ‘cpu’，确认设备， 检查是否有可用的GPU设备，如果有则使用GPU，否则使用CPU
• model = NeuralNetwork().to(device)，框架上传到GPU/CPU

模型输出展示：

3 模型优化

---

3.1 函数理解

• optimizer = torch.optim.Adam(model.parameters(), lr=0.001)，定义优化器：
  • Adam(）使用Adam优化算法，也可为SGD等优化算法
  • model.parameters()为优化模型的参数，
  • lr为学习率/梯度下降步长为0.001
• loss_fn = nn.CrossEntropyLoss(pre,y)，定义损失函数，使用交叉熵损失函数，适用于分类任务
  • pre,预测结果
  • y,真实结果
  • loss_fn.item()，当前损失值
• model.train() ，将模型设置为训练模式，模型参数是可变的
• x, y = x.to(device), y.to(device)，将数据移动到指定设备（GPU或CPU）
• 反向传播：清零梯度，计算梯度，更新模型参数
  • optimizer.zero_grad() ，清零梯度缓存
    loss.backward()， 计算梯度
    optimizer.step() ， 更新模型参数
• model.eval()，将模型设置为评估模式，模型参数是不可变
• with torch.no_grad()，禁用梯度计算，在测试过程中不需要计算梯度

3.2 训练模型和测试模型代码

optimizer = torch.optim.Adam(model.parameters(),lr=0.001)
loss_fn = nn.CrossEntropyLoss()
def train(dataloader,model,loss_fn,optimizer):model.train()batch_size_num = 1for x,y in dataloader:x,y = x.to(device),y.to(device)pred = model.forward(x)loss = loss_fn(pred,y)optimizer.zero_grad()loss.backward()optimizer.step()loss_value = loss.item()if batch_size_num %100 ==0:print(f'loss: {loss_value:>7f}  [number: {batch_size_num}]')batch_size_num +=1train(train_dataloader,model,loss_fn,optimizer)def test(dataloader,model,loss_fn):size = len(dataloader.dataset)num_batches = len(dataloader)model.eval()test_loss,correct = 0,0with torch.no_grad():for x,y in dataloader:x,y = x.to(device),y.to(device)pred = model.forward(x)test_loss += loss_fn(pred,y).item()correct +=(pred.argmax(1) == y).type(torch.float).sum().item()a = (pred.argmax(1)==y)b = (pred.argmax(1)==y).type(torch.float)test_loss /=num_batchescorrect /= sizeprint(f'test result: \n Accuracy: {(100*correct)}%, Avg loss:{test_loss}')

4 最终代码测试

---

4.1 SGD优化算法

torch.optim.SGD(model.parameters(),lr=0.01)

代码展示：

import torchprint(torch.__version__)
import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensortrain_data = datasets.MNIST(root = 'data',train = True,download = True,transform = ToTensor()
)
test_data = datasets.MNIST(root = 'data',train = False,download = True,transform = ToTensor()
)
print(len(train_data))
print(len(test_data))
from matplotlib import pyplot as plt
figure = plt.figure()
for i in range(9):img,label = train_data[i+59000]figure.add_subplot(3,3,i+1)plt.title(label)plt.axis('off')plt.imshow(img.squeeze(),cmap='gray')a = img.squeeze()
plt.show()train_dataloader = DataLoader(train_data, batch_size=64)
test_dataloader= DataLoader(test_data, batch_size=64)
device = 'cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu'
print(f'Using {device} device')
class NeuralNetwork(nn.Module):def __init__(self):super().__init__()self.flatten = nn.Flatten()self.hidden1 = nn.Linear(28*28,128)self.hidden2 = nn.Linear(128, 256)self.out = nn.Linear(256,10)def forward(self,x):x = self.flatten(x)x = self.hidden1(x)x = torch.sigmoid(x)x = self.hidden2(x)x = torch.sigmoid(x)x = self.out(x)return x
model = NeuralNetwork().to(device)
#
print(model)
optimizer = torch.optim.SGD(model.parameters(),lr=0.01)
loss_fn = nn.CrossEntropyLoss()
def train(dataloader,model,loss_fn,optimizer):model.train()batch_size_num = 1for x,y in dataloader:x,y = x.to(device),y.to(device)pred = model.forward(x)loss = loss_fn(pred,y)optimizer.zero_grad()loss.backward()optimizer.step()loss_value = loss.item()if batch_size_num %100 ==0:print(f'loss: {loss_value:>7f}  [number: {batch_size_num}]')batch_size_num +=1def test(dataloader,model,loss_fn):size = len(dataloader.dataset)num_batches = len(dataloader)model.eval()test_loss,correct = 0,0with torch.no_grad():for x,y in dataloader:x,y = x.to(device),y.to(device)pred = model.forward(x)test_loss += loss_fn(pred,y).item()correct +=(pred.argmax(1) == y).type(torch.float).sum().item()a = (pred.argmax(1)==y)b = (pred.argmax(1)==y).type(torch.float)test_loss /=num_batchescorrect /= sizeprint(f'test result: \n Accuracy: {(100*correct)}%, Avg loss:{test_loss}')
#train(train_dataloader,model,loss_fn,optimizer)
test(test_dataloader,model,loss_fn)

运行结果：

4.2 Adam优化算法

自适应算法，torch.optim.Adam(model.parameters(),lr=0.01)

运行结果：

4.3 多次迭代

代码展示：

import torchprint(torch.__version__)
import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensortrain_data = datasets.MNIST(root = 'data',train = True,download = True,transform = ToTensor()
)
test_data = datasets.MNIST(root = 'data',train = False,download = True,transform = ToTensor()
)
print(len(train_data))
print(len(test_data))
from matplotlib import pyplot as plt
figure = plt.figure()
for i in range(9):img,label = train_data[i+59000]figure.add_subplot(3,3,i+1)plt.title(label)plt.axis('off')plt.imshow(img.squeeze(),cmap='gray')a = img.squeeze()
plt.show()train_dataloader = DataLoader(train_data, batch_size=64)
test_dataloader= DataLoader(test_data, batch_size=64)
device = 'cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu'
print(f'Using {device} device')
class NeuralNetwork(nn.Module):def __init__(self):super().__init__()self.flatten = nn.Flatten()self.hidden1 = nn.Linear(28*28,128)self.hidden2 = nn.Linear(128, 256)self.out = nn.Linear(256,10)def forward(self,x):x = self.flatten(x)x = self.hidden1(x)x = torch.sigmoid(x)x = self.hidden2(x)x = torch.sigmoid(x)x = self.out(x)return x
model = NeuralNetwork().to(device)
#
print(model)
optimizer = torch.optim.Adam(model.parameters(),lr=0.01)
loss_fn = nn.CrossEntropyLoss()
def train(dataloader,model,loss_fn,optimizer):model.train()batch_size_num = 1for x,y in dataloader:x,y = x.to(device),y.to(device)pred = model.forward(x)loss = loss_fn(pred,y)optimizer.zero_grad()loss.backward()optimizer.step()loss_value = loss.item()if batch_size_num %100 ==0:print(f'loss: {loss_value:>7f}  [number: {batch_size_num}]')batch_size_num +=1def test(dataloader,model,loss_fn):size = len(dataloader.dataset)num_batches = len(dataloader)model.eval()test_loss,correct = 0,0with torch.no_grad():for x,y in dataloader:x,y = x.to(device),y.to(device)pred = model.forward(x)test_loss += loss_fn(pred,y).item()correct +=(pred.argmax(1) == y).type(torch.float).sum().item()a = (pred.argmax(1)==y)b = (pred.argmax(1)==y).type(torch.float)test_loss /=num_batchescorrect /= sizeprint(f'test result: \n Accuracy: {(100*correct)}%, Avg loss:{test_loss}')
#train(train_dataloader,model,loss_fn,optimizer)
test(test_dataloader,model,loss_fn)
#
e = 30
for i in range(e):print(f'e: {i+1}\n------------------')train(train_dataloader, model, loss_fn, optimizer)
print('done')test(test_dataloader, model, loss_fn)

运行结果：