python 实现逻辑斯谛回归（logistic regression）

有用请点赞，没用请差评。

欢迎分享本文，转载请保留出处。

 

logistic regression模型的原理可以参考这两篇博客：

https://www.jianshu.com/p/4cf34bf158a1

https://blog.csdn.net/c406495762/article/details/77723333

这两篇博客在训练模型时采用的公式是：

，也就是代码中的实现方法二，另外自己也根据李航《统计学习方法》上面的对数似然函数推导了梯度公式，如下：

 采用的数据集是只有两类label(0,1)的MNIST数据集：https://download.csdn.net/download/big_pai/11166917

# -*- coding:utf-8 -*-
# logistic regression，逻辑斯谛回归，极大似然估计模型。算法参考李航《统计学习方法》
#author:Tomator

import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split


class Logistic_regression(object):
    def __init__(self,learn_rate,max_interation):
        self.learn_rate=learn_rate
        self.max_interation=max_interation
        self.w=None

    # 第一种实现方法，参考李航《统计学习方法》
    def train(self,train_vector,train_label):
        train_nums,train_feture_nums=train_vector.shape
        print("train_data",train_nums,train_feture_nums)
        train_label=train_label.reshape(train_nums,1)
        train_label = np.tile(train_label, (1, train_feture_nums))
        self.w=np.zeros(train_feture_nums)
        interations=1
        # while interations < self.max_interation:
        while interations < self.max_interation:
            alpha_y=np.exp(np.dot(self.w,train_vector.T)) / (1 + np.exp(np.dot(self.w,train_vector.T)))
            alpha_y=alpha_y.reshape((train_nums,1))
            alpha_y=np.tile(alpha_y,(1,train_feture_nums))
            # print(alpha_y.shape,train_label.shape)
            alpha_w=np.sum(train_label*train_vector-train_vector*alpha_y)
            self.w=self.w+alpha_w
            # print(self.w.shape)
            interations+=1

        self.w=self.w.reshape((1,train_feture_nums))
        return self.w

    def predict(self, vector):
        # vector=vector.reshape((784,1))
        exp_wx=np.exp(np.dot(self.w,vector.T))
        predict_1=exp_wx/(1+exp_wx)
        predict_2=1/(1+exp_wx)
        # print("predict",predict_1,predict_2)
        if predict_1 > predict_2:
            return 0
        else:
            return 1

    # 第二种实现方法
    def sigmoid(self,z):
        return 1.0 / (1 + np.exp(-z))

    def train2(self, train_vector, train_label):
        train_nums, train_feture_nums = train_vector.shape
        print("train_data", train_nums, train_feture_nums)
        self.w = np.zeros(train_feture_nums)
        print(self.w.shape)
        interations = 1
        while interations < self.max_interation:
            v = np.dot(self.w, train_vector.T)
            error = train_label - self.sigmoid(v)
            # print(train_label.shape,error.shape)
            self.w += self.learn_rate * np.dot(error, train_vector)
            # print(self.w.shape)
            interations += 1
        return self.w

    def predict2(self, x):
        PT = self.sigmoid(np.dot(self.w,x.T))
        if PT > 1 - PT:
            return 1
        else:
            return 0


if __name__ == "__main__":

    np.seterr(divide='ignore', invalid='ignore')
    """
    调用pandas的read_csv()方法时，默认使用C engine作为parser engine，而当文件名中含有中文的时候，用C engine在部分情况下就会出错。所以在调用read_csv()方法时指定engine为Python就可以解决问题了。
    """
    data=pd.read_csv("D:\\python3_anaconda3\\学习\\机器学习\\机器学习数据集\\MNIST数据集——两类label\\train_binary.csv",skiprows=1,engine='python')
    data = data.values[1:,:]
    data=np.array(data)

    # 划分训练集和测试集
    train_data,test_data = train_test_split(data,test_size=0.3, random_state=0)
    # CSV文件的第一行为表头，转换成numpy数组时要剔除
    train_vector=train_data[:,1:]
    train_label=train_data[:,0]
    test_vector=test_data[:,1:]
    test_label=test_data[:,0]

    print("test_label.shape",test_label.shape)
    logistic=Logistic_regression(learn_rate=0.00001,max_interation=5000)
    logistic.train(train_vector,train_label)

    score=0
    for vector,label in zip(test_vector,test_label):
        if logistic.predict(vector) == label:
            score+=1
    print(score/len(test_label))

 

 第一种实现方法的输出结果，训练时间较慢。

  第二种实现方法的输出结果，速度较快。