7. 神经网络训练MNIST数据集的简单实现

1. 实现全连接神经网络搭建(第3节内容)

在简单神经网络框架中，增加ReLU激活函数

import torch.nn as nn
##########################Step 1: 全连接神经网络搭建#############################
class NeuralNet(nn.Module):
    def __init__(self, in_features, hidden_features, out_features):
        super(NeuralNet, self).__init__()
        #  输入层 ->  隐藏层
        self.layer1 = nn.Linear(in_features, hidden_features)
        #  隐藏层 -> 输出层
        self.layer2 = nn.Linear(hidden_features, out_features)
 
    def forward(self, x):
        y = self.layer1(x)          # 第一层参数传递
        y = nn.functional.relu(y)   # ReLU激活函数
        y = self.layer2(y)          # 第二层参数传递
        return y

2. 完成数据集下载(第6节内容)

from torchvision.transforms import ToTensor
from torchvision.datasets import MNIST
############################Step 2: 数据集下载##################################
trainData = MNIST(root = "./",            
                  train = True,          
                  transform=ToTensor(), 
                  download = True)       
testData = MNIST(root = "./",
                  train = False,
                  transform=ToTensor(),
                  download = True)

3. 完成数据集加载(第6节内容)

from torch.utils.data import DataLoader
############################Step 3: 数据集加载##################################
batch_size = 64
trainData_loader = DataLoader(dataset = trainData,
                              batch_size = batch_size, 
                              shuffle = True)  
testData_loader = DataLoader(dataset = testData,
                             batch_size = batch_size,        
                             shuffle = True)

4. 样本标签的独热编码（One-hot encoding）

import torch
#######################  Step 4: 独热编码   ################################
def one_hot(label,depth=10):
    out = torch.zeros(label.size(0),depth)
    idx = torch.LongTensor(label).view(-1,1)
    out.scatter_(dim=1,index=idx,value=1)
    return out

在本案例中，并不是将问题当做 数据拟合，而是当做 数字分类 来处理。基于独热编码格式，将标量标签0~9编码成一个10维的向量。

5. 基于已搭建的网络结构来训练样本

简单思路： 将一张28X28的张量数据，通过简单全连接神经网络结构，进行学习分类，输出一个10的向量对应0~9数字。

实际运行：每次打包64个数据样本图片，调用NeuralNet网络框架进行参数传递，得到输出y_pred。与实际标签y对比，计算损失函数，调用最小梯度算法，计算新的神经元权重，更新NeuralNet网络框架。

in_features = 28*28
hidden_features = 1000
out_features = 10

NN_model = NeuralNet(in_features, hidden_features, out_features) # 定义网络具体参数
L = nn.CrossEntropyLoss()                                        # 交叉熵损失函数

from torch.optim.sgd import SGD
optimizer = SGD(NN_model.parameters(),lr=0.1)                    # 最小梯度优化算法

progress = []
loss_degree = []
#######################    Step 5: 机器学习     ################################
for idx,(data,label) in enumerate(trainData_loader):
    # Size([batch_size,1,28,28]) -> Size([batch_size,28*28])
    data = data.view(data.size(0),28*28)
    optimizer.zero_grad()        #梯度清零
    y_pred = NN_model(data)      #正向学习
    y = one_hot(label)
    loss = L(y_pred,y)           #计算损失
    loss.backward()              #反向传播，计算参数
    optimizer.step()             #参数更新

    progress.append(100. * idx / len(trainData_loader))
    loss_degree.append(loss.item())

 6. 训练过程中数据的可视化（方便理解）

import matplotlib.pyplot as plt
plt.figure(dpi=300)    
plt.plot(progress,loss_degree)
plt.xlabel("Progess (%)")
plt.ylabel("CrossEntropy Loss")
plt.xlim([0,100])
plt.show()

 6. 将NN_model用于预测测试集数据

progress = []
accuracy = []
#######################    Step 5: 模型评估     ################################
for idx,(data,label) in enumerate(testData_loader):
    data = data.view(data.size(0),28*28)
    y_pred = NN_model(data)
    _, y_pred = torch.max(y_pred, dim=1)  # dim = 1 列是第0个维度，行是第1个维度 
    progress.append(100. * idx / len(testData_loader))
    accuracy.append( ((y_pred == label).sum().item())/ data.size(0) )

plt.figure(dpi=300)    
plt.plot(progress,accuracy)
plt.xlabel("Progess (%)")
plt.ylabel("Accuracy (%)")
plt.xlim([0,100])
plt.show()

7. 训练模型的保存和加载

torch.save(NN_model,"./NN_model.pth")   # 模型保存：torch.save(net,PATH)
NN_model = torch.load("./NN_model.pth") # 模型加载：net = torch.load(PATH)    

在当前路径下生成NN_model.pth文件，即训练的模型文件，用于后续加载。

附录：完整代码

import torch
import torch.nn as nn
from torch.optim.sgd import SGD
from torchvision.transforms import ToTensor
from torchvision.datasets import MNIST
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt
##########################Step 0: 超参数的初始化################################
batch_size = 64
in_features = 28*28
hidden_features =256
out_features = 10
##########################Step 1: 全连接神经网络搭建#############################
class NeuralNet(nn.Module):
    def __init__(self, in_features, hidden_features, out_features):
        super(NeuralNet, self).__init__()
        self.layer1 = nn.Linear(in_features, hidden_features)
        self.layer2 = nn.Linear(hidden_features, out_features)
 
    def forward(self, x):
        y = self.layer1(x)
        y = nn.functional.relu(y)
        y = self.layer2(y)
        return y
############################Step 2: 数据集下载##################################
trainData = MNIST(root = "./",            
                  train = True,          
                  transform=ToTensor(), 
                  download = True)       
testData = MNIST(root = "./",
                  train = False,
                  transform=ToTensor(),
                  download = True)
############################Step 3: 数据集加载##################################
trainData_loader = DataLoader(dataset = trainData,
                              batch_size = batch_size, 
                              shuffle = True)  
testData_loader = DataLoader(dataset = testData,
                             batch_size = batch_size,        
                             shuffle = True)
#######################Step 4: 标签编码#########################################
def one_hot(label,depth=10):
    out = torch.zeros(label.size(0),depth)
    idx = torch.LongTensor(label).view(-1,1)
    out.scatter_(dim=1,index=idx,value=1)
    return out
###############################################################################
NN_model = NeuralNet(in_features, hidden_features, out_features) # 定义网络具体参数
L = nn.CrossEntropyLoss()                                        # 交叉熵损失函数
optimizer = SGD(NN_model.parameters(),lr=0.1)                   # 最小梯度优化算法

progress = []
loss_degree = []
#######################    Step 5: 机器学习     ################################
for idx,(data,label) in enumerate(trainData_loader):
    # Size([batch_size,1,28,28]) -> Size([batch_size,28*28])
    data = data.view(data.size(0),28*28)
    optimizer.zero_grad()        #梯度清零
    y_pred = NN_model(data)      #正向学习
    y = one_hot(label)
    loss = L(y_pred,y)           #计算误差
    loss.backward()              #反向传播，计算参数
    optimizer.step()             #参数更新
 
    progress.append(100. * idx / len(trainData_loader))
    loss_degree.append(loss.item())


plt.figure(dpi=300)    
plt.plot(progress,loss_degree)
plt.xlabel("Progess (%)")
plt.ylabel("CrossEntropy Loss")
plt.xlim([0,100])
plt.show()

progress = []
accuracy = []
#######################    Step 5: 模型评估     ################################
for idx,(data,label) in enumerate(testData_loader):
    data = data.view(data.size(0),28*28)
    y_pred = NN_model(data)
    _, y_pred = torch.max(y_pred, dim=1)  # dim = 1 列是第0个维度，行是第1个维度 
    progress.append(100. * idx / len(testData_loader))
    accuracy.append( ((y_pred == label).sum().item())/ data.size(0) )

plt.figure(dpi=300)    
plt.plot(progress,accuracy)
plt.xlabel("Progess (%)")
plt.ylabel("Accuracy (%)")
plt.xlim([0,100])
plt.show()