《动手学深度学习(PyTorch版)》笔记3.3

注:书中对代码的讲解并不详细,本文对很多细节做了详细注释。另外,书上的源代码是在Jupyter Notebook上运行的,较为分散,本文将代码集中起来,并加以完善,全部用vscode在python 3.9.18下测试通过。

Chapter3 Linear Neural Networks

3.3 Concise Implementations of Linear Regression

复制代码
import numpy as np
import torch
from torch.utils import data
from d2l import torch as d2l

true_w=torch.tensor([2,-3.4])
true_b=4.2
features,labels=d2l.synthetic_data(true_w,true_b,1000)

#构造一个pytorch数据迭代器
def load_array(data_arrays,batch_size,is_train=True): #@save
    dataset=data.TensorDataset(*data_arrays)
    #"TensorDataset" is a class provided by the torch.utils.data module which is a dataset wrapper that allows you to create a dataset from a sequence of tensors. 
    #"*data_arrays" is used to unpack the tuple into individual tensors.
    #The '*' operator is used for iterable unpacking.
    #Here, data_arrays is expected to be a tuple containing the input features and corresponding labels. The "*data_arrays" syntax is used to unpack the elements of the tuple and pass them as separate arguments.
    return data.DataLoader(dataset,batch_size,shuffle=is_train)
    #Constructs a PyTorch DataLoader object which is an iterator that provides batches of data during training or testing.
batch_size=10
data_iter=load_array([features,labels],batch_size)
print(next(iter(data_iter)))#调用next()函数时会返回迭代器的下一个项目,并更新迭代器的内部状态以便下次调用

#定义模型变量,nn是神经网络的缩写
from torch import nn
net=nn.Sequential(nn.Linear(2,1))
#Creates a sequential neural network with one linear layer.
#Input size (in_features) is 2, indicating the network expects input with 2 features.
#Output size (out_features) is 1, indicating the network produces 1 output.

#初始化模型参数
net[0].weight.data.normal_(0,0.01)#The underscore at the end (normal_) indicates that this operation is performed in-place, modifying the existing tensor in memory.
net[0].bias.data.fill_(0)

#定义均方误差损失函数,也称平方L2范数,返回所有样本损失的平均值
loss=nn.MSELoss()#MSE:mean squared error 

#定义优化算法(仍是小批量随机梯度下降)
#update the parameters of the neural network (net.parameters()) using gradients computed during backpropagation. 
trainer=torch.optim.SGD(net.parameters(),lr=0.03)#SGD:stochastic gradient descent(随机梯度下降)

#训练
num_epochs=3
for epoch in range(num_epochs):
    for X,y in data_iter:
        l=loss(net(X),y)
        trainer.zero_grad()
        l.backward()
        trainer.step()#Updates the model parameters using the computed gradients and the optimization algorithm.
    l=loss(net(features),labels)
    print(f'epoch {epoch+1},loss {l:.6f}')#{l:.f}表示将变量l格式化为小数点后有6位的浮点数。
    
w=net[0].weight.data
print('w的估计误差:',true_w-w.reshape(true_w.shape))
b=net[0].bias.data
print('b的估计误差:',true_b-b)
相关推荐
f6m274dww2cmd3 分钟前
拆一个类目第一的20个曝光词:43% 的搜索量来自第二页
人工智能
2601_9679801114 分钟前
Python中heapq模块的各种用法实例
数据结构·python·算法·应用实例·heapq模块
大胖爱学习_MJ15 分钟前
超越向量检索:用 Spring Boot + Neo4j 手搓 GraphRAG 知识图谱记忆
java·后端·python
RobinDevNotes28 分钟前
24GB Mac 做数字人:声音放本地,画面租云端
人工智能·python·macos·开发工具
吃我一窝jio1 小时前
Java Bug复盘与性能优化实战
java·开发语言·笔记
user4465117917911 小时前
Agent 也能「过安检」:AgentScope 权限系统设计拆解
人工智能
烟小栀1 小时前
Windows下的深度学习环境配置参考及配置教程
人工智能
tellmewhoisi2 小时前
深度学习0:数学知识:最大似然推导过程
人工智能·深度学习·概率论
oil欧哟2 小时前
从产品经理视角拆解 DeepSeek Super-agent:双层路由与异步调度架构
人工智能
张张123y2 小时前
个人助手型 Agent 的六个工程细节:Hermes 与 OpenClaw 的审批、委派、上下文与调度
网络·人工智能·开源·后端开发·ai agent