Logistic回归 Python实现

Logistic回归

Logistic函数

$$f(x) = \frac{1}{1+e^{-x}}$$
其函数图像为：

绘图方法

>>> import numpy as np
>>> x = np.arange(-10,10,0.1)
>>> y = 1/(1+np.exp(-x))
>>> import matplotlib.pyplot as plt
>>> plt.plot(x,y)
[<matplotlib.lines.Line2D object at 0x00000000054FFE10>]
>>> plt.show()

其在近0点陡峭的上升特点决定了它可以将一个回归问题转换为一个分类问题。

详情请见我的另外一篇博客
http://blog.csdn.net/taiji1985/article/details/50969697

在ml/dt.py中的数据生成

def createD3():
    np.random.seed(101)
    ndim = 2
    n2  =10
    a = 3+5*np.random.randn(n2,ndim)
    b = 18+4*np.random.randn(n2,ndim)
    X = np.concatenate((a,b))
    ay = np.zeros(n2)
    by = np.ones(n2)
    ylabel = np.concatenate((ay,by))
    return {'X':X,'ylabel':ylabel}

ml/fig.py中的绘图

# -*- coding: UTF-8 -*-
'''
Created on 2016-4-24

@author: Administrator
'''
import numpy as np
import operator
import matplotlib.pyplot as plt


#绘图
def plotData(ds,type='o'):
    X= ds['X']
    y=ds['ylabel']
    n = X.shape[0]
    cn = len(np.unique(y))
    cs = ['r','g']
    dd  = np.arange(n)
    for i in range(2):
        index= y == i
        xx=X[dd[index]]
        plt.plot(xx[:,0],xx[:,1],type,markerfacecolor=cs[i],markersize=14)

    xmax = np.max(X[:,0])
    xmin = np.min(X[:,0])
    ymax = np.max(X[:,1])
    ymin = np.min(X[:,1])
    print xmin,xmax,ymin,ymax
    dx = xmax - xmin
    dy = ymax - ymin
    plt.axis([xmin-dx*0.1, xmax + dx*0.1, ymin-dy*0.1, ymax +dy*0.1])

分类器代码

# -*- coding: UTF-8 -*-
'''
Logistic 回归
Created on 2016-4-26

@author: taiji1985
'''
import numpy as np
import operator
import matplotlib.pyplot as plt
def sigmoid(X):
    return 1.0/(1+np.exp(-X))
    pass

#生成该矩阵的增广矩阵，添加最后一列（全部为1）
def augment(X):
    n,ndim = X.shape
    a = np.mat(np.ones((n,1)))
    return np.concatenate((X,a),axis=1)

def classify(X,w):
    X = np.mat(X)
    w = np.mat(w)
    if w.shape[0] < w.shape[1]:
        w = w.T

    #增广
    X= augment(X)
    d = X*w
    r = np.zeros((X.shape[0],1))
    r[d>0] = 1
    return r
    pass


#梯度下降法学习
#alpha 学习因子
def learn(dataset,alpha=0.001):
    X= dataset['X']
    ylabel = dataset['ylabel']
    n,ndim = X.shape
    cls_label = np.unique(ylabel)
    cn=len(cls_label)


    X = np.mat(X)
    ylabel = np.mat(ylabel).T

    #生成增广矩阵
    X = augment(X)
    ndim += 1


    max_c = 500
    w = np.ones((ndim,1))
    i = 0
    ep= 0.0001
    cha = 1

    while cha > ep:
        #计算y = wx + b 
        ypred =X*w
        #计算 logistic
        ypred = sigmoid(ypred)
        #计算误差
        error = ylabel - ypred
        cha = alpha*X.T*error
        w = w + cha
        cha = np.abs( np.sum(cha))
        print i,w.T
        i=i+1

    return w

测试

'''
Created on 2016-4-26

@author: Administrator
'''
from ml import lsd
from ml import dt
from ml import fig
import numpy as np

import matplotlib.pyplot as plt

ds = dt.createD3()
w = lsd.learn(ds,0.01)

#draw line
lx = [0, -w[2]/w[0]]
ly = [-w[2]/w[1],0]

plt.figure()

fig.plotData(ds, 'o')

plt.plot(lx,ly)


plt.show()


ypred = lsd.classify(ds['X'],w)
ylabel = np.mat(ds['ylabel']).T

print ypred
print ylabel

dist = np.sum(ypred != ylabel)
print dist,len(ylabel), 1.0*dist/len(ylabel)