python-3.x - python中简单实现线性回归时出现溢出错误

Question

我在运行代码时遇到此错误:

ErrorValue += ((m*x + b) - y)**2 RuntimeWarning: double_scalars 中遇到溢出

有人可以解释一下我的代码有什么问题吗？如果可以对我的线性reg尝试的正确性以及改进代码的方法提出有用的建议，那就太好了。

非常感谢!

import csv
import matplotlib.pyplot as plt
from numpy import *

'''
This is a simple implementation of linear regression on correlation
hours studied by student and the marks they obtained.
'''
def run():

    points = genfromtxt("data.csv", delimiter=",")

    # x is hours studied, y is marks obtained.

    # We are applying the function: y = b + mx
    for i in range(len(points)):
        x = points[i][0]
        y = points[i][1]

    N = len(points)
    b = 0
    m = 0
    alpha = 0.001 # alpha is the learning rate
    ErrorThreshold  = 0.003
    NumberOfIterations = 1000 # We cancel the gradient descent after a number of iterations, if it still doesn't reach the threshold we want.

    sum_m = 0
    sum_b = 0

    for i in range(NumberOfIterations):
        while mean_squared_error(x,y,b,m,points) > ErrorThreshold:
            b , m  = gradient_descent(m,b,alpha,x,y,N,points)

def mean_squared_error(x,y,b,m,points):
    ErrorValue = 0
    for i in range(len(points)):
        ErrorValue += ((m*x + b) - y)**2
    return ErrorValue / len(points) 


def gradient_descent(m,b,alpha,N,x,y,points):

    #dealing with summation sign in gradient descent
    sum_m = 0
    sum_b = 0

    for i in range(len(points)):
        x = points[i][0]
        y = points[i][1]
        sum_m += m*x + b - y
        sum_b += m*x + b - y 
        #repeating just for clarification purposes.

    new_b = b - (2/N)*sum_b
    new_m = m - (((2*m)/N))*sum_m

    return new_b, new_m

if __name__ == '__main__':
    run()

Answer 1

当您跟踪(打印)中间值时，您会看到什么？例如，在您的例程中添加几行。

def mean_squared_error(x,y,b,m,points):
    print("ENTER", x, y, b, m, len(points))
    ErrorValue = 0
    for i in range(len(points)):
        ErrorValue += ((m*x + b) - y)**2
        print("TRACE", i, ErrorValue)
    return ErrorValue / len(points) 

另外，我不确定你的计算是否正确；我想你可能想要

        ErrorValue += ((m*x[i] + b) - y[i])**2

目前，您正在乘以整个向量，而不仅仅是标量，但您正在执行 len(points) 次。

最后，使用len(x)并且根本不传入points会更容易吗？除了长度之外，您没有使用点。