最近在B站看沐神的动手学深度学习视频,记录一下学习过程
查看本文的jupyter notebook格式(更加清晰美观哦!)
手动实现线性回归
%matplotlib inline
import random
import torch
from d2l import torch as d2l
def synthetic_data(w, b, num_examples):
"""生成 y = Xw+b+噪声"""
X = torch.normal(0, 1, (num_examples, len(w)))
y = torch.matmul(X, w)+b
y += torch.normal(0, 0.01, y.shape)
return X, y.reshape((-1, 1))
true_w = torch.tensor([2, -3.4])
true_b = 4.2
features, labels = synthetic_data(true_w, true_b, 1000)
print('features:',features[0],'\nlabels:', labels[0])
features: tensor([0.8068, 0.2531])
labels: tensor([4.9606])
d2l.set_figsize()
d2l.plt.scatter(features[:, 1].detach().numpy(),
labels.detach().numpy(), 1)
<matplotlib.collections.PathCollection at 0x14326e99470>
[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-i3XuW3jL-1628955277759)(output_6_1.svg)]
定义一个data_iter函数,该函数接收批量大小、特征矩阵和标签向量作为输入,生成大小为batch_size的小批量
def data_iter(batch_size, features, labels):
num_examples = len(features)
indices = list(range(num_examples))
random.shuffle(indices)
for i in range(0, num_examples, batch_size):
batch_indices = torch.tensor(
indices[i:min(i+batch_size, num_examples)])
yield features[batch_indices], labels[batch_indices]
batch_size = 10
for X, y in data_iter(batch_size, features, labels):
print(X, '\n', y)
break
tensor([[ 1.0626, -1.3819],
[ 1.3002, -1.6405],
[-0.2143, 0.1157],
[ 0.3747, 0.1398],
[ 0.7321, 0.7536],
[-1.0199, 1.0867],
[-1.1761, 1.0229],
[ 1.0091, 0.3677],
[-0.3801, -0.5643],
[-0.7545, 1.8007]])
tensor([[11.0434],
[12.3610],
[ 3.3817],
[ 4.4807],
[ 3.1153],
[-1.5422],
[-1.6426],
[ 4.9857],
[ 5.3447],
[-3.4343]])
定义 初始化模型参数
w = torch.normal(0, 0.01, size=(2,1), requires_grad=True)
b = torch.zeros(1, requires_grad = True)
定义模型
def linreg(X, w, b):
"""线性回归模型"""
return torch.matmul(X, w)+b
定义损失函数
在损失函数中并没有取均值,即没有除以batch_size,而是在优化算法中做了,二者等价
def squared_loss(y_hat, y):
"""均方损失"""
return (y_hat-y.reshape(y_hat.shape))**2/2
定义随机梯度下降(stochastic gradient descent)优化算法
def sgd(params, lr, batch_size):
"""小批量随机梯度下降"""
with torch.no_grad():
for param in params:
param -= lr*param.grad/batch_size
param.grad.zero_()
训练过程
lr = 0.03
num_epochs = 3
net = linreg
loss = squared_loss
for epoch in range(num_epochs):
for X, y in data_iter(batch_size, features, labels):
l = loss(net(X, w, b), y) # `X`和`y`的小批量损失
# 因为l形状是(`batch_size`, 1), 而不是一个标量。`l` 中的所有元素被累加到一起
# 并以此计算关于[`w`,`b`]的梯度
l.sum().backward()
sgd([w,b], lr, batch_size)
with torch.no_grad():
train_l = loss(net(features, w, b), labels)
print(f'epoch {epoch+1}, loss {float(train_l.mean()):f}')
epoch 1, loss 0.042440
epoch 2, loss 0.000161
epoch 3, loss 0.000047
print(f'w的估计误差: {true_w - w.reshape(true_w.shape)}')
print(f'b的估计误差: {true_b - b}')
w的估计误差: tensor([-0.0002, -0.0006], grad_fn=<SubBackward0>)
b的估计误差: tensor([0.0012], grad_fn=<RsubBackward1>)
线性回归的简洁实现
import numpy as np
from torch.utils import data
true_w = torch.tensor([2, -3.4])
true_b = 4.2
features, labels = d2l.synthetic_data(true_w, true_b, 100)
调用框架中现有的API来读取数据
def load_array(data_arrays, batch_size, is_train=True):
"""构造一个PyTorch数据迭代器"""
dataset = data.TensorDataset(*data_arrays)
return data.DataLoader(dataset, batch_size, shuffle = is_train)
batch_size = 10
data_iter = load_array((features, labels), batch_size)
next(iter(data_iter))
[tensor([[ 0.5283, 0.2948],
[-0.8903, -0.3965],
[-1.5611, -1.0055],
[ 1.1645, 0.5376],
[ 2.0624, -0.3454],
[ 0.1723, -1.8691],
[-0.2568, -1.4501],
[ 0.0065, -0.4836],
[-1.1152, -0.6035],
[ 0.4217, -0.7177]]),
tensor([[ 4.2522],
[ 3.7881],
[ 4.5270],
[ 4.6987],
[ 9.5021],
[10.9073],
[ 8.6155],
[ 5.8502],
[ 4.0087],
[ 7.4830]])]
使用框架预定义好的层
# `nn`是神经网络的缩写
from torch import nn
net = nn.Sequential(nn.Linear(2, 1))
初始化模型参数
net[0].weight.data.normal_(0, 0.01)
net[0].bias.data.fill_(0)
tensor([0.])
计算均方误差使用的是MSELoss类,也成为平方L2范数
loss = nn.MSELoss()
实例化SGD实例
trainer = torch.optim.SGD(net.parameters(), lr=0.03)
训练过程代码与从零开始实现时所做的非常相似
num_epochs = 3
for epoch in range(num_epochs):
for X, y in data_iter:
l = loss(net(X), y)
trainer.zero_grad() # 为什么不是在trainer.step()语句之后?
l.backward()
trainer.step()
l = loss(net(features), labels)
print(f'epoch {epoch+1}, loss {l:f}')
epoch 1, loss 9.681911
epoch 2, loss 3.371778
epoch 3, loss 1.197687
问答
为什么使用平方损失而不用绝对差值(L1范数)?
其实二者没太大区别,都可以。最初没有使用绝对差值,是因为在值为0处不容易求导。
损失为什么要求平均?
不求平均也可以,但是学习率会根据batch的大小而变化(因为此时相当于扩大了梯度,
而缩小了学习率),不容易调超参数。为了解耦,需要求平均。
backward()这里是调用PyTorch自定义的back propogation吗?
是的
每个batch计算的时候,为什么需要把梯度先清零呀?
因为Pytorch不帮我们自动清零,如果不清零的话,会在原有的梯度上自动累加。