在Python中实现Gradient Descent并收到溢出错误

Question

梯度下降和溢出错误

我目前正在python中实现矢量化梯度下降。 但是，我继续遇到溢出错误。 我的数据集中的数字并不是非常大。 我正在使用这个公式：

我选择此实现以避免使用衍生物。 有没有人对如何解决这个问题有任何建议，或者我是否实施错误？ 先感谢您！

数据集链接： https ： //www.kaggle.com/CooperUnion/anime-recommendations-database/data

## Cleaning Data ##
import math
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd

data = pd.read_csv('anime.csv')
# print(data.corr())
# print(data['members'].isnull().values.any()) # Prints False
# print(data['rating'].isnull().values.any()) # Prints True

members = [] # Corresponding fan club size for row 
ratings = [] # Corresponding rating for row

for row in data.iterrows():
    if not math.isnan(row[1]['rating']): # Checks for Null ratings
        members.append(row[1]['members'])
        ratings.append(row[1]['rating'])


plt.plot(members, ratings)
plt.savefig('scatterplot.png')

theta0 = 0.3 # Random guess
theta1 = 0.3 # Random guess
error = 0

公式的

def hypothesis(x, theta0, theta1):
    return  theta0 + theta1 * x

def costFunction(x, y, theta0, theta1, m):
    loss = 0 
    for i in range(m): # Represents summation
        loss += (hypothesis(x[i], theta0, theta1) - y[i])**2
    loss *= 1 / (2 * m) # Represents 1/2m
    return loss

def gradientDescent(x, y, theta0, theta1, alpha, m, iterations=1500):
    for i in range(iterations):
        gradient0 = 0
        gradient1 = 0
        for j in range(m):
            gradient0 += hypothesis(x[j], theta0, theta1) - y[j]
            gradient1 += (hypothesis(x[j], theta0, theta1) - y[j]) * x[j]
        gradient0 *= 1/m
        gradient1 *= 1/m
        temp0 = theta0 - alpha * gradient0
        temp1 = theta1 - alpha * gradient1
        theta0 = temp0
        theta1 = temp1
        error = costFunction(x, y, theta0, theta1, len(y))
        print("Error is:", error)
    return theta0, theta1

print(gradientDescent(members, ratings, theta0, theta1, 0.01, len(ratings)))

错误的

在几次迭代之后，我在我的gradientDescent函数中调用的costFunction给了我一个OverflowError：（34，'Result too large'）。 但是，我希望我的代码能够不断打印出一个递减的错误值。

    Error is: 1.7515692852199285e+23
    Error is: 2.012089675182454e+38
    Error is: 2.3113586742689143e+53
    Error is: 2.6551395730578252e+68
    Error is: 3.05005286756189e+83
    Error is: 3.503703756035943e+98
    Error is: 4.024828599077087e+113
    Error is: 4.623463163528686e+128
    Error is: 5.311135890211131e+143
    Error is: 6.101089907410428e+158
    Error is: 7.008538065634975e+173
    Error is: 8.050955905074458e+188
    Error is: 9.248418197694096e+203
    Error is: 1.0623985545062037e+219
    Error is: 1.220414847696018e+234
    Error is: 1.4019337603196565e+249
    Error is: 1.6104509643047377e+264
    Error is: 1.8499820618048921e+279
    Error is: 2.1251399172389593e+294
    Traceback (most recent call last):
      File "tyreeGradientDescent.py", line 54, in <module>
        print(gradientDescent(members, ratings, theta0, theta1, 0.01, len(ratings)))
      File "tyreeGradientDescent.py", line 50, in gradientDescent
        error = costFunction(x, y, theta0, theta1, len(y))
      File "tyreeGradientDescent.py", line 33, in costFunction
        loss += (hypothesis(x[i], theta0, theta1) - y[i])**2
    OverflowError: (34, 'Result too large')

Answer 1

您的数据值非常大，这使您的损失功能非常陡峭。 结果是除非将数据标准化为更小的值，否则您需要一个小的 alpha。 由于alpha值过大，您的渐变下降会在整个地方跳跃并实际发散，这就是您的错误率上升而不是下降的原因。

使用当前数据，alpha为0.0000000001将使误差收敛。 经过30次迭代，我的损失来自：

Error is: 66634985.91339202

至

Error is: 16.90452378179708

Answer 2

import numpy as np
import pandas as pd

X = [0.5, 2.5]
Y = [0.2, 0.9]

def f(w, b, x): #sigmoid with parameter w,b
    return 1.0/(1.0 * np.exp(-(w * x + b)))


def error(w, b):
    err = 0.0
    for x, y in zip(X, Y):
        fx = f(w, b, x)
        err += 0.5 * (fx - y)**2
    return err

def grad_b(w, b, x, y):
    fx = f(w, b, x)
    return (fx - y) * fx * (1 - fx)

def grad_w(w, b, x, y):
    fx = f(w, b, x)
    return (fx - y) * fx * (1 - fx) * x

def do_gradient_descent():
    w, b, eta, max_epochs = 1, 1, 0.01, 100
    for i in range(max_epochs):
        dw, db = 0, 0
        for x, y in zip(X, Y):
            dw += grad_w(w, b, x, y)
            db += grad_b(w, b, x, y)
        w = w - eta * dw
        print(w)
        b = b - eta * db
        print(b)
    er = error(w, b)
    #print(er)
    return er
##Calling Gradient Descent function
do_gradient_descent()