最近在跟着Andrew NG 的机器学习课程学习，当跟到逻辑回归的时候，有一个练习，跟着做，遇到了这个问题。
调用minimize函数返回结果result，输出result.x发现与初始theta一样，然后就打印了一下result
e:\FileData\xx\算法\machineLearning\machine-learning-andrew\exercise2\sigmoid.py:6: RuntimeWarning: overflow encountered in exp
return 1/(1+np.exp(-data))
e:\FileData\xx\算法\machineLearning\machine-learning-andrew\exercise2\costFunction.py:11: RuntimeWarning: divide by zero encountered in log
ln_h = np.log(h_x) # np.log() is the Natural Log
fun: array([[nan]])
jac: array([ -0.6 , -44.83135362, -44.37384125])
message: 'Linear search failed'
nfev: 56
nit: 0
status: 4
success: False
x: array([ -0.1 , -12.00921659, -11.26284221])
我的函数
main.py
import dataPlot
import numpy as np
import sigmoid
import costFunction
import scipy.optimize as op
#========================read datas========================
with open('./ex2/ex2data1.txt','r') as file:
lines = file.readlines()
m = len(lines)
feature_number = len(lines[0].strip().split(',')) - 1
x_data = np.zeros((m,feature_number))
y_data = np.zeros((m,1))
for i in range(m):
line_temp = lines[i].strip().split(',')
for j in range(len(line_temp)):
if j != len(line_temp) - 1:
x_data[i,j] = line_temp[j]
else:
y_data[i,0] = line_temp[j]
# dataPlot.Plot(x_data, y_data)
#========================initialize========================
x_data = np.column_stack((np.ones((m, 1)), x_data))
theta = np.zeros((feature_number+1, 1)).flatten() # must be a vector
#print(sigmoid.hyposisFunction([1,2],[0,0]))
# ideal ouput 0.5
# mine is 0.5
#========================compute cost and gradient========================
cost = costFunction.cost_function(theta, x_data,y_data)
print("initial_theta cost is {} (approx)".format(cost))
print("expected cost is 0.693")
# my output
# initial_theta cost is [[0.69314718]] (approx)
# expected cost is 0.693
initial_theta = costFunction.gradient(theta, x_data, y_data)
print("initial_theta theta is {}".format(initial_theta))
# myoutput
# initial_theta theta is [ -0.1 -12.00921659 -11.26284221]
# print('Expected gradients (approx):\n -0.1000\n -12.0092\n -11.2628\n')
#========================Optimizing using fminunc(matlab)/scipy(python)========================
result = op.minimize(costFunction.cost_function, initial_theta.flatten(), args=(x_data, y_data), method="TNC", jac=costFunction.gradient)
print(result)
# result
# e:\FileData\xx\算法\machineLearning\machine-learning-andrew\exercise2\sigmoid.py:6: RuntimeWarning: overflow encountered in exp
# return 1/(1+np.exp(-data))
# e:\FileData\xx\算法\machineLearning\machine-learning-andrew\exercise2\costFunction.py:11: RuntimeWarning: divide by zero encountered in log
# ln_h = np.log(h_x) # np.log() is the Natural Log
# fun: array([[nan]])
# jac: array([ -0.6 , -44.83135362, -44.37384125])
# message: 'Linear search failed'
# nfev: 56
# nit: 0
# status: 4
# success: False
# x: array([ -0.1 , -12.00921659, -11.26284221])
sigmoid.py
import numpy as np
# compute the sigmoid of data, data can be a vector, matrix or scalar
def sigmoid(data):
return 1/(1+np.exp(-data))
costFunction.py
import numpy as np
import sigmoid
# compute the cost and the partial derivations of theta, x,y can be a vector, matrix, scalar
def cost_function(theta, x_data, y_data):
m = y_data.shape[0]
theta = theta.reshape((-1, 1))
# compute cost
h_x = sigmoid.sigmoid(np.dot(x_data, theta))
ln_h = np.log(h_x) # np.log() is the Natural Log
part1 = -np.dot(ln_h.T, y_data) / m
ln_1h = np.log(1-h_x)
part2 = -np.dot(ln_1h.T, 1-y_data) / m
cost = part1+part2
# compute the partial derivations of theta
return cost
# because scipy.op.minimize is different from fminunc, we need a method returns the gradient independently
def gradient(theta, x_data, y_data):
theta = theta.reshape((-1, 1))
m = y_data.shape[0]
# compute cost
h_x = sigmoid.sigmoid(np.dot(x_data, theta))
# compute the partial derivations of theta
grad = np.dot(x_data.T, h_x - y_data) / m
return grad.flatten() # must be a vector
已解决，main.py中38行，调用的gradient将initial_theta给覆盖了