神经网络入门：从零开始构建你的第一个深度学习模型

✨道路是曲折的，前途是光明的！

📝 专注C/C++、Linux编程与人工智能领域，分享学习笔记！

🌟 感谢各位小伙伴的长期陪伴与支持，欢迎文末添加好友一起交流！

• • 前言
  • 什么是神经网络？
  • • 基本结构
  • 核心概念
  • • 1. 神经元
    • 2. 激活函数
    • 3. 前向传播
  • PyTorch实战代码
  • • 1. 定义神经网络
    • 2. 训练循环
  • 训练流程图
  • 损失函数的作用
  • 反向传播与梯度下降
  • 完整训练示例
  • 总结
  • 进一步学习方向

前言

神经网络是深度学习的基石，也是人工智能领域最激动人心的技术之一。本文将带你从零开始，理解神经网络的核心概念，并用代码实现一个简单的神经网络模型。

什么是神经网络？

神经网络是一种受人脑神经元结构启发的机器学习模型。它由相互连接的节点（神经元）组成，通过层层传递和处理信息来完成学习任务。

基本结构

#mermaid-svg-vXqK8ycg4m3bhJxp{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-vXqK8ycg4m3bhJxp .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-vXqK8ycg4m3bhJxp .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-vXqK8ycg4m3bhJxp .error-icon{fill:#552222;}#mermaid-svg-vXqK8ycg4m3bhJxp .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-vXqK8ycg4m3bhJxp .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-vXqK8ycg4m3bhJxp .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-vXqK8ycg4m3bhJxp .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-vXqK8ycg4m3bhJxp .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-vXqK8ycg4m3bhJxp .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-vXqK8ycg4m3bhJxp .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-vXqK8ycg4m3bhJxp .marker{fill:#333333;stroke:#333333;}#mermaid-svg-vXqK8ycg4m3bhJxp .marker.cross{stroke:#333333;}#mermaid-svg-vXqK8ycg4m3bhJxp svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-vXqK8ycg4m3bhJxp p{margin:0;}#mermaid-svg-vXqK8ycg4m3bhJxp .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-vXqK8ycg4m3bhJxp .cluster-label text{fill:#333;}#mermaid-svg-vXqK8ycg4m3bhJxp .cluster-label span{color:#333;}#mermaid-svg-vXqK8ycg4m3bhJxp .cluster-label span p{background-color:transparent;}#mermaid-svg-vXqK8ycg4m3bhJxp .label text,#mermaid-svg-vXqK8ycg4m3bhJxp span{fill:#333;color:#333;}#mermaid-svg-vXqK8ycg4m3bhJxp .node rect,#mermaid-svg-vXqK8ycg4m3bhJxp .node circle,#mermaid-svg-vXqK8ycg4m3bhJxp .node ellipse,#mermaid-svg-vXqK8ycg4m3bhJxp .node polygon,#mermaid-svg-vXqK8ycg4m3bhJxp .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-vXqK8ycg4m3bhJxp .rough-node .label text,#mermaid-svg-vXqK8ycg4m3bhJxp .node .label text,#mermaid-svg-vXqK8ycg4m3bhJxp .image-shape .label,#mermaid-svg-vXqK8ycg4m3bhJxp .icon-shape .label{text-anchor:middle;}#mermaid-svg-vXqK8ycg4m3bhJxp .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-vXqK8ycg4m3bhJxp .rough-node .label,#mermaid-svg-vXqK8ycg4m3bhJxp .node .label,#mermaid-svg-vXqK8ycg4m3bhJxp .image-shape .label,#mermaid-svg-vXqK8ycg4m3bhJxp .icon-shape .label{text-align:center;}#mermaid-svg-vXqK8ycg4m3bhJxp .node.clickable{cursor:pointer;}#mermaid-svg-vXqK8ycg4m3bhJxp .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-vXqK8ycg4m3bhJxp .arrowheadPath{fill:#333333;}#mermaid-svg-vXqK8ycg4m3bhJxp .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-vXqK8ycg4m3bhJxp .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-vXqK8ycg4m3bhJxp .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-vXqK8ycg4m3bhJxp .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-vXqK8ycg4m3bhJxp .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-vXqK8ycg4m3bhJxp .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-vXqK8ycg4m3bhJxp .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-vXqK8ycg4m3bhJxp .cluster text{fill:#333;}#mermaid-svg-vXqK8ycg4m3bhJxp .cluster span{color:#333;}#mermaid-svg-vXqK8ycg4m3bhJxp div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-vXqK8ycg4m3bhJxp .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-vXqK8ycg4m3bhJxp rect.text{fill:none;stroke-width:0;}#mermaid-svg-vXqK8ycg4m3bhJxp .icon-shape,#mermaid-svg-vXqK8ycg4m3bhJxp .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-vXqK8ycg4m3bhJxp .icon-shape p,#mermaid-svg-vXqK8ycg4m3bhJxp .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-vXqK8ycg4m3bhJxp .icon-shape rect,#mermaid-svg-vXqK8ycg4m3bhJxp .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-vXqK8ycg4m3bhJxp .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-vXqK8ycg4m3bhJxp .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-vXqK8ycg4m3bhJxp :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

Output Layer

Hidden Layer

Input Layer

输入特征1

输入特征2

输入特征3

神经元1

神经元2

神经元3

神经元4

输出1

输出2

核心概念

1. 神经元

每个神经元接收输入，通过权重和偏置计算，最后通过激活函数产生输出：

output = activation(Σ(input × weight) + bias)

2. 激活函数

激活函数引入非线性，使网络能够学习复杂的模式。

3. 前向传播

前向传播是数据从输入层流向输出层的过程。

#mermaid-svg-1DlyhlBLlEocWnCQ{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-1DlyhlBLlEocWnCQ .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-1DlyhlBLlEocWnCQ .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-1DlyhlBLlEocWnCQ .error-icon{fill:#552222;}#mermaid-svg-1DlyhlBLlEocWnCQ .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-1DlyhlBLlEocWnCQ .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-1DlyhlBLlEocWnCQ .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-1DlyhlBLlEocWnCQ .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-1DlyhlBLlEocWnCQ .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-1DlyhlBLlEocWnCQ .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-1DlyhlBLlEocWnCQ .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-1DlyhlBLlEocWnCQ .marker{fill:#333333;stroke:#333333;}#mermaid-svg-1DlyhlBLlEocWnCQ .marker.cross{stroke:#333333;}#mermaid-svg-1DlyhlBLlEocWnCQ svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-1DlyhlBLlEocWnCQ p{margin:0;}#mermaid-svg-1DlyhlBLlEocWnCQ .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-1DlyhlBLlEocWnCQ .cluster-label text{fill:#333;}#mermaid-svg-1DlyhlBLlEocWnCQ .cluster-label span{color:#333;}#mermaid-svg-1DlyhlBLlEocWnCQ .cluster-label span p{background-color:transparent;}#mermaid-svg-1DlyhlBLlEocWnCQ .label text,#mermaid-svg-1DlyhlBLlEocWnCQ span{fill:#333;color:#333;}#mermaid-svg-1DlyhlBLlEocWnCQ .node rect,#mermaid-svg-1DlyhlBLlEocWnCQ .node circle,#mermaid-svg-1DlyhlBLlEocWnCQ .node ellipse,#mermaid-svg-1DlyhlBLlEocWnCQ .node polygon,#mermaid-svg-1DlyhlBLlEocWnCQ .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-1DlyhlBLlEocWnCQ .rough-node .label text,#mermaid-svg-1DlyhlBLlEocWnCQ .node .label text,#mermaid-svg-1DlyhlBLlEocWnCQ .image-shape .label,#mermaid-svg-1DlyhlBLlEocWnCQ .icon-shape .label{text-anchor:middle;}#mermaid-svg-1DlyhlBLlEocWnCQ .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-1DlyhlBLlEocWnCQ .rough-node .label,#mermaid-svg-1DlyhlBLlEocWnCQ .node .label,#mermaid-svg-1DlyhlBLlEocWnCQ .image-shape .label,#mermaid-svg-1DlyhlBLlEocWnCQ .icon-shape .label{text-align:center;}#mermaid-svg-1DlyhlBLlEocWnCQ .node.clickable{cursor:pointer;}#mermaid-svg-1DlyhlBLlEocWnCQ .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-1DlyhlBLlEocWnCQ .arrowheadPath{fill:#333333;}#mermaid-svg-1DlyhlBLlEocWnCQ .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-1DlyhlBLlEocWnCQ .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-1DlyhlBLlEocWnCQ .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-1DlyhlBLlEocWnCQ .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-1DlyhlBLlEocWnCQ .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-1DlyhlBLlEocWnCQ .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-1DlyhlBLlEocWnCQ .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-1DlyhlBLlEocWnCQ .cluster text{fill:#333;}#mermaid-svg-1DlyhlBLlEocWnCQ .cluster span{color:#333;}#mermaid-svg-1DlyhlBLlEocWnCQ div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-1DlyhlBLlEocWnCQ .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-1DlyhlBLlEocWnCQ rect.text{fill:none;stroke-width:0;}#mermaid-svg-1DlyhlBLlEocWnCQ .icon-shape,#mermaid-svg-1DlyhlBLlEocWnCQ .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-1DlyhlBLlEocWnCQ .icon-shape p,#mermaid-svg-1DlyhlBLlEocWnCQ .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-1DlyhlBLlEocWnCQ .icon-shape rect,#mermaid-svg-1DlyhlBLlEocWnCQ .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-1DlyhlBLlEocWnCQ .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-1DlyhlBLlEocWnCQ .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-1DlyhlBLlEocWnCQ :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

输入数据

输入层

隐藏层1

隐藏层2

输出层

预测结果

PyTorch实战代码

1. 定义神经网络

import torch
import torch.nn as nn
import torch.optim as optim

# 定义一个简单的全连接神经网络
class SimpleNet(nn.Module):
def __init__(self, input_size, hidden_size, num_classes):
super(SimpleNet, self).__init__()
# 第一层：输入层到隐藏层
self.layer1 = nn.Linear(input_size, hidden_size)
self.relu = nn.ReLU()
# 第二层：隐藏层到输出层
self.layer2 = nn.Linear(hidden_size, num_classes)

def forward(self, x):
out = self.layer1(x)
out = self.relu(out)
out = self.layer2(out)
return out

# 初始化模型
model = SimpleNet(input_size=784, hidden_size=128, num_classes=10)
print(model)

2. 训练循环

# 定义损失函数和优化器
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)

# 训练循环
num_epochs = 10

for epoch in range(num_epochs):
for batch_idx, (data, targets) in enumerate(train_loader):
# 前向传播
outputs = model(data)
loss = criterion(outputs, targets)

# 反向传播和优化
optimizer.zero_grad()
loss.backward()
optimizer.step()

print(f'Epoch [{epoch+1}/{num_epochs}], Loss: {loss.item():.4f}')

训练流程图

#mermaid-svg-HdfPwgvyPMoxvfa2{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-HdfPwgvyPMoxvfa2 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-HdfPwgvyPMoxvfa2 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-HdfPwgvyPMoxvfa2 .error-icon{fill:#552222;}#mermaid-svg-HdfPwgvyPMoxvfa2 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-HdfPwgvyPMoxvfa2 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-HdfPwgvyPMoxvfa2 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-HdfPwgvyPMoxvfa2 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-HdfPwgvyPMoxvfa2 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-HdfPwgvyPMoxvfa2 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-HdfPwgvyPMoxvfa2 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-HdfPwgvyPMoxvfa2 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-HdfPwgvyPMoxvfa2 .marker.cross{stroke:#333333;}#mermaid-svg-HdfPwgvyPMoxvfa2 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-HdfPwgvyPMoxvfa2 p{margin:0;}#mermaid-svg-HdfPwgvyPMoxvfa2 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-HdfPwgvyPMoxvfa2 .cluster-label text{fill:#333;}#mermaid-svg-HdfPwgvyPMoxvfa2 .cluster-label span{color:#333;}#mermaid-svg-HdfPwgvyPMoxvfa2 .cluster-label span p{background-color:transparent;}#mermaid-svg-HdfPwgvyPMoxvfa2 .label text,#mermaid-svg-HdfPwgvyPMoxvfa2 span{fill:#333;color:#333;}#mermaid-svg-HdfPwgvyPMoxvfa2 .node rect,#mermaid-svg-HdfPwgvyPMoxvfa2 .node circle,#mermaid-svg-HdfPwgvyPMoxvfa2 .node ellipse,#mermaid-svg-HdfPwgvyPMoxvfa2 .node polygon,#mermaid-svg-HdfPwgvyPMoxvfa2 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-HdfPwgvyPMoxvfa2 .rough-node .label text,#mermaid-svg-HdfPwgvyPMoxvfa2 .node .label text,#mermaid-svg-HdfPwgvyPMoxvfa2 .image-shape .label,#mermaid-svg-HdfPwgvyPMoxvfa2 .icon-shape .label{text-anchor:middle;}#mermaid-svg-HdfPwgvyPMoxvfa2 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-HdfPwgvyPMoxvfa2 .rough-node .label,#mermaid-svg-HdfPwgvyPMoxvfa2 .node .label,#mermaid-svg-HdfPwgvyPMoxvfa2 .image-shape .label,#mermaid-svg-HdfPwgvyPMoxvfa2 .icon-shape .label{text-align:center;}#mermaid-svg-HdfPwgvyPMoxvfa2 .node.clickable{cursor:pointer;}#mermaid-svg-HdfPwgvyPMoxvfa2 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-HdfPwgvyPMoxvfa2 .arrowheadPath{fill:#333333;}#mermaid-svg-HdfPwgvyPMoxvfa2 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-HdfPwgvyPMoxvfa2 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-HdfPwgvyPMoxvfa2 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-HdfPwgvyPMoxvfa2 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-HdfPwgvyPMoxvfa2 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-HdfPwgvyPMoxvfa2 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-HdfPwgvyPMoxvfa2 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-HdfPwgvyPMoxvfa2 .cluster text{fill:#333;}#mermaid-svg-HdfPwgvyPMoxvfa2 .cluster span{color:#333;}#mermaid-svg-HdfPwgvyPMoxvfa2 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-HdfPwgvyPMoxvfa2 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-HdfPwgvyPMoxvfa2 rect.text{fill:none;stroke-width:0;}#mermaid-svg-HdfPwgvyPMoxvfa2 .icon-shape,#mermaid-svg-HdfPwgvyPMoxvfa2 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-HdfPwgvyPMoxvfa2 .icon-shape p,#mermaid-svg-HdfPwgvyPMoxvfa2 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-HdfPwgvyPMoxvfa2 .icon-shape rect,#mermaid-svg-HdfPwgvyPMoxvfa2 .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-HdfPwgvyPMoxvfa2 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-HdfPwgvyPMoxvfa2 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-HdfPwgvyPMoxvfa2 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

否

是

开始

初始化模型参数

获取训练数据批次

前向传播计算预测

计算损失

反向传播计算梯度

更新权重参数

是否完成所有epoch?

训练完成

损失函数的作用

损失函数衡量模型预测与真实值的差距：

# 常用损失函数示例

# 1. 均方误差（回归任务）
mse_loss = nn.MSELoss()

# 2. 交叉熵（分类任务）
cross_entropy = nn.CrossEntropyLoss()

# 3. 二元交叉熵（二分类）
bce_loss = nn.BCELoss()

反向传播与梯度下降

反向传播是训练神经网络的核心算法：

#mermaid-svg-Kgn9XRDWkr2MymDs{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-Kgn9XRDWkr2MymDs .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-Kgn9XRDWkr2MymDs .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-Kgn9XRDWkr2MymDs .error-icon{fill:#552222;}#mermaid-svg-Kgn9XRDWkr2MymDs .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-Kgn9XRDWkr2MymDs .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-Kgn9XRDWkr2MymDs .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-Kgn9XRDWkr2MymDs .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-Kgn9XRDWkr2MymDs .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-Kgn9XRDWkr2MymDs .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-Kgn9XRDWkr2MymDs .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-Kgn9XRDWkr2MymDs .marker{fill:#333333;stroke:#333333;}#mermaid-svg-Kgn9XRDWkr2MymDs .marker.cross{stroke:#333333;}#mermaid-svg-Kgn9XRDWkr2MymDs svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-Kgn9XRDWkr2MymDs p{margin:0;}#mermaid-svg-Kgn9XRDWkr2MymDs .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-Kgn9XRDWkr2MymDs .cluster-label text{fill:#333;}#mermaid-svg-Kgn9XRDWkr2MymDs .cluster-label span{color:#333;}#mermaid-svg-Kgn9XRDWkr2MymDs .cluster-label span p{background-color:transparent;}#mermaid-svg-Kgn9XRDWkr2MymDs .label text,#mermaid-svg-Kgn9XRDWkr2MymDs span{fill:#333;color:#333;}#mermaid-svg-Kgn9XRDWkr2MymDs .node rect,#mermaid-svg-Kgn9XRDWkr2MymDs .node circle,#mermaid-svg-Kgn9XRDWkr2MymDs .node ellipse,#mermaid-svg-Kgn9XRDWkr2MymDs .node polygon,#mermaid-svg-Kgn9XRDWkr2MymDs .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-Kgn9XRDWkr2MymDs .rough-node .label text,#mermaid-svg-Kgn9XRDWkr2MymDs .node .label text,#mermaid-svg-Kgn9XRDWkr2MymDs .image-shape .label,#mermaid-svg-Kgn9XRDWkr2MymDs .icon-shape .label{text-anchor:middle;}#mermaid-svg-Kgn9XRDWkr2MymDs .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-Kgn9XRDWkr2MymDs .rough-node .label,#mermaid-svg-Kgn9XRDWkr2MymDs .node .label,#mermaid-svg-Kgn9XRDWkr2MymDs .image-shape .label,#mermaid-svg-Kgn9XRDWkr2MymDs .icon-shape .label{text-align:center;}#mermaid-svg-Kgn9XRDWkr2MymDs .node.clickable{cursor:pointer;}#mermaid-svg-Kgn9XRDWkr2MymDs .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-Kgn9XRDWkr2MymDs .arrowheadPath{fill:#333333;}#mermaid-svg-Kgn9XRDWkr2MymDs .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-Kgn9XRDWkr2MymDs .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-Kgn9XRDWkr2MymDs .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-Kgn9XRDWkr2MymDs .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-Kgn9XRDWkr2MymDs .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-Kgn9XRDWkr2MymDs .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-Kgn9XRDWkr2MymDs .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-Kgn9XRDWkr2MymDs .cluster text{fill:#333;}#mermaid-svg-Kgn9XRDWkr2MymDs .cluster span{color:#333;}#mermaid-svg-Kgn9XRDWkr2MymDs div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-Kgn9XRDWkr2MymDs .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-Kgn9XRDWkr2MymDs rect.text{fill:none;stroke-width:0;}#mermaid-svg-Kgn9XRDWkr2MymDs .icon-shape,#mermaid-svg-Kgn9XRDWkr2MymDs .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-Kgn9XRDWkr2MymDs .icon-shape p,#mermaid-svg-Kgn9XRDWkr2MymDs .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-Kgn9XRDWkr2MymDs .icon-shape rect,#mermaid-svg-Kgn9XRDWkr2MymDs .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-Kgn9XRDWkr2MymDs .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-Kgn9XRDWkr2MymDs .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-Kgn9XRDWkr2MymDs :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

计算损失

计算输出层梯度

计算隐藏层梯度

计算输入层梯度

更新所有权重

完整训练示例

import torch
import torch.nn as nn
from torch.utils.data import DataLoader, TensorDataset

# 创建示例数据
X = torch.randn(1000, 20) # 1000个样本，20个特征
y = torch.randint(0, 2, (1000,)) # 二分类标签

# 创建数据加载器
dataset = TensorDataset(X, y)
train_loader = DataLoader(dataset, batch_size=32, shuffle=True)

# 定义模型
class BinaryClassifier(nn.Module):
def __init__(self):
super().__init__()
self.network = nn.Sequential(
nn.Linear(20, 64),
nn.ReLU(),
nn.Dropout(0.2),
nn.Linear(64, 32),
nn.ReLU(),
nn.Linear(32, 1),
nn.Sigmoid()
)

def forward(self, x):
return self.network(x)

# 训练
model = BinaryClassifier()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
criterion = nn.BCELoss()

for epoch in range(5):
epoch_loss = 0
for X_batch, y_batch in train_loader:
optimizer.zero_grad()
predictions = model(X_batch).squeeze()
loss = criterion(predictions, y_batch.float())
loss.backward()
optimizer.step()
epoch_loss += loss.item()
print(f'Epoch {epoch+1}, Average Loss: {epoch_loss/len(train_loader):.4f}')

总结

通过本文，我们学习了：

进一步学习方向

• CNN（卷积神经网络）：图像处理领域
• RNN/LSTM：序列数据处理
• Transformer：自然语言处理
• GAN（生成对抗网络）：图像生成

神经网络的世界非常广阔，继续探索吧！

参考资料：

• PyTorch官方文档
• 《深度学习》- Ian Goodfellow
• CS231n: 卷积神经网络课程

✍️ 坚持用 清晰易懂的图解 + 可落地的代码，让每个知识点都 简单直观！

💡 座右铭：“道路是曲折的，前途是光明的！”