【python】numpy实现逻辑回归

import numpy as np

class logistic:
    def __init__(self, fileName, splitBy='\t'):
        self.lr = 1e-3
        self.epochs = 3000  # 定义学习率和训练轮数

        data = np.array([line.split(splitBy) for line in open(fileName).readlines()], dtype='float')
        self.X = np.array(data[:, :-1],dtype='float').T
        self.Y = np.array([data[:, -1]],dtype='float')   # 读取数据, X 为 21 x 299 的矩阵, Y 为 1 x 299 的矩阵

        ones_line = np.ones(shape=(1,self.Y.shape[1]),dtype='float')
        self.X = np.row_stack((self.X,ones_line))       #这里在训练数据集X的最后一行加上一行 1 x 299 的全 1 矩阵 ,目的是为了bias的计算

        self.W = np.zeros(shape=(self.X.shape[0], 1))    # 定义参数weight, 形状为 22 x 1 ,和前面的 X 相乘
                                                         # X 和 weight 的行数都加一, 这样 weight * X 等效于 weight * X + bias

    def _WX(self):      # weight * X
        return np.matmul(self.W.T,self.X)

    def _sigmod(self,x):    # 计算 sigmod 函数
        return np.array((1 / (1 + np.exp(-x))), dtype='float')

    def _Logistic(self):    # 计算输出结果
        return self._sigmod(self._WX())

    def _CrossEntropy(self):    # cost 定义为交叉熵
        p1 = np.log(self._Logistic()).dot(-self.Y.T)
        p2 = np.log(1-self._Logistic()).dot(1-self.Y.T)
        return np.array((p1-p2)/self.Y.shape[1],dtype='float')

    def _Gradient(self):        # 对交叉熵求导得到梯度
        return self.X.dot((self._Logistic() - self.Y).T) / self.Y.shape[1]

    def train(self):            # 进行训练
        for i in range(self.epochs):
            self.W -= self.lr*self._Gradient()
            print('第{}轮训练,交叉熵为:{}'.format(i+1,float(self._CrossEntropy())))

    def predict(self,x):        # 进行预测
        '''
        因为 W 的形状为 22 x 1 所以输入 x 需要在最后一行添加一行全 1 矩阵
        '''
        return self._sigmod(np.matmul(self.W.T,x))

    def test(self, fileName=None, splitBy='\t'):    # 进行测试
        if fileName == None:
            y = self.predict(self.X)[0]
            for i in range(y.__len__()):
                if y[i]>=0.5:
                    y[i]=1
                else:
                    y[i]=0
            y_ = self.Y[0]

            true = 0
            for j in range(y_.__len__()):
                if y[j] == y_[j]:
                    true += 1
            print('训练集上正确率:{}'.format(true/y_.__len__()))
        else:
            data = np.array([line.split(splitBy) for line in open(fileName).readlines()], dtype='float')
            x = np.array(data[:, :-1],dtype='float').T
            y_ = np.array([data[:, -1]],dtype='float')[0]
            ones_line = np.ones(shape=(1, y_.shape[0]), dtype='float')
            x = np.row_stack((x,ones_line))
            y = self.predict(x)[0]
            for i in range(y.__len__()):
                if y[i] >= 0.5:
                    y[i] = 1
                else:
                    y[i] = 0
            true = 0
            for j in range(y_.__len__()):
                if y[j] == y_[j]:
                    true += 1
            print('测试集上正确率:{}'.format(true / y_.__len__()))

if __name__ == '__main__':
    lr = logistic('./logistic/logisticTraining.txt')
    lr.train()
    lr.test()
    lr.test(fileName='./logistic/logisticTest.txt')

数据集链接：https://pan.baidu.com/s/10pI9Y7w3MCYLhz29RhdZEg
提取码：3ajp