一、原理和概念
1.回归
回归最简单的定义是,给出一个点集D,用一个函数去拟合这个点集。而且使得点集与拟合函数间的误差最小,假设这个函数曲线是一条直线,那就被称为线性回归;假设曲线是一条二次曲线,就被称为二次回归。
以下仅介绍线性回归的基本实现。
2.假设函数、误差、代价函数
参考 Machine Learning 学习笔记2 - linear regression with one variable(单变量线性回归)
最小化误差一般有两个方法:最小二乘法和梯度下降法
最小二乘法可以一步到位,直接算出未知参数,但他是有前提的。梯度下降法和最小二乘法不一样,它通过一步一步的迭代,慢慢的去靠近到那条最优直线。
平方误差:
代价函数:
(系数是为了之后求梯度的时候方便)
3.梯度下降算法
梯度下降算法是一种优化算法,它可以帮助我们找到一个函数的局部极小值,不仅用在线性回归模型中,非线性也可以。在求解损失函数的最小值时,可以通过梯度下降法来一步步的迭代求解,得到最小化的损失函数和模型参数值。反过来,如果我们需要求解损失函数的最大值,这时就需要用梯度上升法来迭代了。
下图是假设函数 h(x)、 代价函数J()和梯度下降算法:
完整的梯度下降算法:
梯度下降算法的Python实现:
# -*- coding: utf-8 -*-
# @Time : 2018/3/6 18:32
# @Author : TanRong
# @Software: PyCharm
# @File : gradient descent.py import numpy as np
import matplotlib.pyplot
import pylab # 参数含义:y=kx+b;learning_rate学习速率、步长幅度;num_iter迭代的次数 #计算梯度并更新k,b
def gradient(current_k, current_b, data, learning_rate):
k_gradient = 0
b_gradient = 0
m = float(len(data))
for i in range(0, len(data)):
x = data[i,0]
y = data[i,1]
k_gradient += (1/m)*(current_k*x + current_b - y) * x
b_gradient += (1/m)*(current_k*x + current_b - y)
update_k = current_k - learning_rate * k_gradient
update_b = current_b - learning_rate * b_gradient
return[update_k, update_b] #优化器
def optimizer(data, initial_k, initial_b, learning_rate, num_iter):
k = initial_k
b = initial_b #Gradient descent 梯度下降
for i in range(num_iter):
#更新 k、b
k,b = gradient(k, b, data, learning_rate)
return [k,b] #绘图
def plot_data(data, k, b):
x = data[:,0]
y = data[:,1]
y_predict = k * x + b
pylab.plot(x,y,'o')
pylab.plot(x,y_predict,'k-')
pylab.show() #计算平方差
def error(data, k, b):
totalError = 0;
for i in range(0, len(data)):
x = data[i,0]
y = data[i,1]
totalError += (k*x+b-y)**2
return totalError / float(len(data))
#梯度下降算法 实现线性回归
def Linear_regression():
data = np.loadtxt('train_data.csv', delimiter = ',') #训练数据
learning_rate = 0.01
initial_k = 0.0
initial_b = 0.0
num_iter = 1000 [k,b] = optimizer(data, initial_k, initial_b, learning_rate, num_iter)
print("k:", k,";b:", b)
print("平方差/代价函数:", error(data, k, b)) plot_data(data, k, b) Linear_regression()
代码和数据的下载:https://github.com/~~~
(数据用的别人的)
参考代码:
#http://blog.csdn.net/sxf1061926959/article/details/66976356?locationNum=9&fps=1 import numpy as np
import pylab def compute_error(b,m,data): totalError = 0
#Two ways to implement this
#first way
# for i in range(0,len(data)):
# x = data[i,0]
# y = data[i,1]
#
# totalError += (y-(m*x+b))**2 #second way
x = data[:,0]
y = data[:,1]
totalError = (y-m*x-b)**2
totalError = np.sum(totalError,axis=0) return totalError/float(len(data)) def optimizer(data,starting_b,starting_m,learning_rate,num_iter):
b = starting_b
m = starting_m #gradient descent
for i in range(num_iter):
#update b and m with the new more accurate b and m by performing
# thie gradient step
b,m =compute_gradient(b,m,data,learning_rate)
if i%100==0:
print 'iter {0}:error={1}'.format(i,compute_error(b,m,data))
return [b,m] def compute_gradient(b_current,m_current,data ,learning_rate): b_gradient = 0
m_gradient = 0 N = float(len(data))
#Two ways to implement this
#first way
# for i in range(0,len(data)):
# x = data[i,0]
# y = data[i,1]
#
# #computing partial derivations of our error function
# #b_gradient = -(2/N)*sum((y-(m*x+b))^2)
# #m_gradient = -(2/N)*sum(x*(y-(m*x+b))^2)
# b_gradient += -(2/N)*(y-((m_current*x)+b_current))
# m_gradient += -(2/N) * x * (y-((m_current*x)+b_current)) #Vectorization implementation
x = data[:,0]
y = data[:,1]
b_gradient = -(2/N)*(y-m_current*x-b_current)
b_gradient = np.sum(b_gradient,axis=0)
m_gradient = -(2/N)*x*(y-m_current*x-b_current)
m_gradient = np.sum(m_gradient,axis=0)
#update our b and m values using out partial derivations new_b = b_current - (learning_rate * b_gradient)
new_m = m_current - (learning_rate * m_gradient)
return [new_b,new_m] def plot_data(data,b,m): #plottting
x = data[:,0]
y = data[:,1]
y_predict = m*x+b
pylab.plot(x,y,'o')
pylab.plot(x,y_predict,'k-')
pylab.show() def Linear_regression():
# get train data
data =np.loadtxt('data.csv',delimiter=',') #define hyperparamters
#learning_rate is used for update gradient
#defint the number that will iteration
# define y =mx+b
learning_rate = 0.001
initial_b =0.0
initial_m = 0.0
num_iter = 1000 #train model
#print b m error
print 'initial variables:\n initial_b = {0}\n intial_m = {1}\n error of begin = {2} \n'\
.format(initial_b,initial_m,compute_error(initial_b,initial_m,data)) #optimizing b and m
[b ,m] = optimizer(data,initial_b,initial_m,learning_rate,num_iter) #print final b m error
print 'final formula parmaters:\n b = {1}\n m={2}\n error of end = {3} \n'.format(num_iter,b,m,compute_error(b,m,data)) #plot result
plot_data(data,b,m) if __name__ =='__main__': Linear_regression()
有一些其他方法实现某个功能,可以再看一遍
参考链接:https://www.cnblogs.com/yangykaifa/p/7261316.html
http://blog.csdn.net/sxf1061926959/article/details/66976356?locationNum=9&fps=1