Python中对数下降曲线上的梯度下降

Question

我希望在对数下降曲线上运行梯度下降，如下所示：

y = y0 - a * ln(b + x)。

本例中我的 y0：800

我试图使用关于 a 和 b 的偏导数来做到这一点，但是虽然这显然可以最小化平方误差，但它并没有收敛。我知道这不是矢量化的，我可能完全采用了错误的方法。我是在犯任何简单的错误，还是完全放弃了这个问题？

import numpy as np

# constants my gradient descent model should find:
a = 4
b = 4

# function to fit on!
def function(x, a, b):
    y0 = 800
    return y0 - a * np.log(b + x)

# Generates data
def gen_data(numpoints):
    a = 4
    b = 4
    x = np.array(range(0, numpoints))
    y = function(x, a, b)
    return x, y
x, y = gen_data(600)

def grad_model(x, y, iterations):
    converged = False

    # length of dataset
    m = len(x)

    # guess   a ,  b
    theta = [0.1, 0.1]
    alpha = 0.001

    # initial error
    e = np.sum((np.square(function(x, theta[0], theta[1])) - y))

    for iteration in range(iterations):
        hypothesis = function(x, theta[0], theta[1])
        loss = hypothesis - y

        # compute partial deritaves to find slope to "fall" into
        theta0_grad = (np.mean(np.sum(-np.log(x + y)))) / (m)
        theta1_grad = (np.mean((((np.log(theta[1] + x)) / theta[0]) - (x*(np.log(theta[1] + x)) / theta[0])))) / (2*m)

        theta0 = theta[0] - (alpha * theta0_grad)
        theta1 = theta[1] - (alpha * theta1_grad)

        theta[1] = theta1
        theta[0] = theta0

        new_e = np.sum(np.square((function(x, theta[0], theta[1])) - y))
        if new_e > e:
            print "AHHHH!"
            print "Iteration: "+ str(iteration)
            break
        print theta
    return theta[0], theta[1]

Answer 1

我在您的代码中发现了一些错误。线

e = np.sum((np.square(function(x, theta[0], theta[1])) - y))

不正确，应替换为

e = np.sum((np.square(function(x, theta[0], theta[1]) - y)))

new_e 的公式包含同样的错误。

另外，梯度公式是错误的。你的损失函数是 $L(a,b) = \sum_{i=1}^N y_0 - a \log(b + x_i)$, 所以你必须计算 $L$ 对 $a$ 和 $b$ 的偏导数。 （LaTeX 真的不能在 stackoverflow 上工作吗？） 最后一点是梯度下降法有步长限制，所以我们的步长一定不能太大。这是您的代码版本更好：

import numpy as np
import matplotlib.pyplot as plt

# constants my gradient descent model should find:
a = 4.0
b = 4.0
y0 = 800.0

# function to fit on!
def function(x, a, b):
    # y0 = 800
    return y0 - a * np.log(b + x)

# Generates data
def gen_data(numpoints):
    # a = 4
    # b = 4
    x = np.array(range(0, numpoints))
    y = function(x, a, b)
    return x, y
x, y = gen_data(600)

def grad_model(x, y, iterations):
    converged = False

    # length of dataset
    m = len(x)

    # guess   a ,  b
    theta = [0.1, 0.1]
    alpha = 0.00001

    # initial error
    # e = np.sum((np.square(function(x, theta[0], theta[1])) - y))    #  This was a bug
    e = np.sum((np.square(function(x, theta[0], theta[1]) - y)))

    costs = np.zeros(iterations)

    for iteration in range(iterations):
        hypothesis = function(x, theta[0], theta[1])
        loss = hypothesis - y

        # compute partial deritaves to find slope to "fall" into
        # theta0_grad = (np.mean(np.sum(-np.log(x + y)))) / (m)
        # theta1_grad = (np.mean((((np.log(theta[1] + x)) / theta[0]) - (x*(np.log(theta[1] + x)) / theta[0])))) / (2*m)
        theta0_grad = 2*np.sum((y0 - theta[0]*np.log(theta[1] + x) - y)*(-np.log(theta[1] + x)))
        theta1_grad = 2*np.sum((y0 - theta[0]*np.log(theta[1] + x) - y)*(-theta[0]/(b + x)))

        theta0 = theta[0] - (alpha * theta0_grad)
        theta1 = theta[1] - (alpha * theta1_grad)

        theta[1] = theta1
        theta[0] = theta0

        # new_e = np.sum(np.square((function(x, theta[0], theta[1])) - y)) # This was a bug
        new_e = np.sum(np.square((function(x, theta[0], theta[1]) - y)))
        costs[iteration] = new_e
        if new_e > e:
            print "AHHHH!"
            print "Iteration: "+ str(iteration)
            # break
        print theta
    return theta[0], theta[1], costs

(theta0,theta1,costs) = grad_model(x,y,100000)
plt.semilogy(costs)