动手学深度学习(二)——正则化(从零开始)

文章作者:Tyan
博客:noahsnail.com  |  CSDN  |  简书

注:本文为李沐大神的《动手学深度学习》的课程笔记!

高维线性回归

使用线性函数$y = 0.05 + \sum_{i = 1}^p 0.01x_i + \text{noise}$生成数据样本,噪音服从均值0和标准差为0.01的正态分布。

1
2
3
4
5
6
7
8
9
10
11
# 导入mxnet
import random
import mxnet as mx

# 设置随机种子
random.seed(2)
mx.random.seed(2)

from mxnet import gluon
from mxnet import ndarray as nd
from mxnet import autograd

生成数据集

1
2
3
4
5
6
7
8
9
10
11
12
13
14
# 训练数据数量
num_train = 20

# 测试数据数量
num_test = 100

# 输入数据特征维度
num_inputs = 200

# 实际权重
true_w = nd.ones((num_inputs, 1)) * 0.01

# 实际偏置
true_b = 0.05
1
2
3
4
5
6
7
8
9
10
# 生成数据
X = nd.random_normal(shape=(num_train + num_test, num_inputs))
y = nd.dot(X, true_w) + true_b

# 添加随机噪声
y += 0.01 * nd.random_normal(shape=y.shape)

# 训练数据和测试数据
X_train, X_test = X[:num_train, :], X[num_train:, :]
y_train, y_test = y[:num_train], y[num_train:]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# 批数据大小
batch_size = 1

# 通过yield进行数据读取
def data_iter(num_examples):
# 产生样本的索引
idx = list(range(num_examples))
# 将索引随机打乱
random.shuffle(idx)
# 迭代一个epoch
for i in range(0, num_examples, batch_size):
# 依次取出样本的索引, 这种实现方式在num_examples/batch_size不能整除时也适用
j = nd.array(idx[i:min((i + batch_size), num_examples)])
# 根据提供的索引取元素
yield nd.take(X, j), nd.take(y, j)

初始化模型参数

1
2
3
4
5
6
7
8
9
10
11
12
def init_params():
# 随机初始化权重w
w = nd.random_normal(shape=(num_inputs, 1))
# 偏置b初始化为0
b = nd.zeros((1,))
# w, b放入list里
params = [w, b]

# 需要计算反向传播, 添加自动求导
for param in params:
param.attach_grad()
return params

$L_2$ 范数正则化

在训练时最小化函数为:$\text{loss} + \lambda \sum_{p \in \textrm{params}}|p|_2^2。$

1
2
3
# L2范数
def L2_penalty(w, b):
return ((w**2).sum() + b ** 2) / 2

定义训练和测试

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
%matplotlib inline
import matplotlib as mpl
mpl.rcParams['figure.dpi']= 120
import matplotlib.pyplot as plt
import numpy as np


# 定义网络
def net(X, w, b):
return nd.dot(X, w) + b

# 损失函数
def square_loss(yhat, y):
return (yhat - y.reshape(yhat.shape)) ** 2 / 2

# 梯度下降
def sgd(params, lr, batch_size):
for param in params:
param[:] = param - lr * param.grad / batch_size

# 测试
def test(net, params, X, y):
return square_loss(net(X, *params), y).mean().asscalar()

# 训练
def train(_lambda):
# 定义训练的迭代周期
epochs = 10
# 定义学习率
learning_rate = 0.005
# 初始化参数
w, b = params = init_params()
# 训练损失
train_loss = []
# 测试损失
test_loss = []
for epoch in range(epochs):
for data, label in data_iter(num_train):
# 记录梯度
with autograd.record():
# 计算预测值
output = net(data, *params)
# 计算loss
loss = square_loss(output, label) + _lambda * L2_penalty(*params)
# 反向传播
loss.backward()
# 更新梯度
sgd(params, learning_rate, batch_size)
# 训练损失
train_loss.append(test(net, params, X_train, y_train))
# 测试损失
test_loss.append(test(net, params, X_test, y_test))

# 绘制损失图像
plt.plot(train_loss)
plt.plot(test_loss)
plt.legend(['train', 'test'])
plt.show()
return 'learned w[:10]:', w[:10].T, 'learned b:', b

观察过拟合

1
train(0)

Overfitting

('learned w[:10]:', 
 [[ 1.04817176 -0.02568593  0.86764956  0.29322267  0.01006179 -0.56152576
    0.38436398 -0.30840367 -2.32450151  0.03733355]]
 <NDArray 1x10 @cpu(0)>, 'learned b:', 
 [ 0.79914856]
 <NDArray 1 @cpu(0)>)

使用正则化

1
train(5)

Normal

('learned w[:10]:', 
 [[ 0.00107633 -0.00052574  0.00450233 -0.00110545 -0.0068391  -0.00181657
   -0.00530632  0.00512845 -0.00742549 -0.00058495]]
 <NDArray 1x10 @cpu(0)>, 'learned b:', 
 [ 0.00449432]
 <NDArray 1 @cpu(0)>)
如果有收获,可以请我喝杯咖啡!