python 好用的画图模块_Python实用画图方法——基于matplotlib模块

#KDD-98 experiment:using a synthetic biased sample

#(the selection probability is known)to estimate

#the generalization error of a hypothesis

import numpy as np

import random

import scipy.stats as stats

import matplotlib.pyplot as plt

#imitate KDD-98 dataset,assume it has 10000 records

m = 10000

X,loss = np.empty(m), np.empty(m)

for i in range(m):

X[i] = random.randint(0,7)

#directly create the loss l(h(x),y) for each X[i]

loss[i] = 1 + X[i] * 5 + random.gauss(0,4)

true_err = loss.mean()

print 'the estimated generalization error via unbiased sample :',true_err

#generate n biased samples,i.e. n biased testing sets

n = 1000

uncorrected_loss,corrected_loss = np.empty(n),np.empty(n)

for k in range(n):

X_biased,loss_biased = [],[]

for i in range(m):

if(X[i] in [0,1,2,3]):

s = stats.bernoulli.rvs(0.3)

else:#X[i] in [4,5,6,7]

s = stats.bernoulli.rvs(0.9)

if(1 == s):

X_biased.append(X[i])

loss_biased.append(loss[i])

uncorrected_loss[k] = np.mean(loss_biased)

m_biased = len(X_biased)

overall_selection_prob = 1.0 * m_biased / m

weighted_loss = np.empty(m_biased)

for i in range(m_biased):

if(X_biased[i] in [0,1,2,3]):

weighted_loss[i] = overall_selection_prob * loss_biased[i] / 0.3

else:

weighted_loss[i] = overall_selection_prob * loss_biased[i] / 0.9

corrected_loss[k] = weighted_loss.mean()

#obtain the bin values and the corresponding frequency values

fre, bins, patches = plt.hist(x = uncorrected_loss, bins = 10,normed = False)

frequency = fre / sum(fre)

fre1, bins1, patches1 = plt.hist(x = corrected_loss, bins = 10,normed = False)

frequency1 = fre1 / sum(fre1)

#show the middle graph which is useless

plt.show()

#=================================================================================

#using scalar true_err, numpy arrays uncorrected_loss and corrected_loss to make

#the histogram plot

fig,ax = plt.subplots(nrows = 1, ncols = 2)

num_bins = 10

#plot the histogram for the uncorrected_loss,i.e. the generalization error estimated

#by biased sample

#do the bar plot

ax[0].bar(left = bins[:-1], height = frequency, width=0.05)

ax[0].plot([true_err] * 5,np.linspace(0,frequency.max() + 0.1,5,endpoint=True), 'r--')

ax[0].set_xlabel('Uncorrected estimate of the generalization error')

ax[0].set_ylabel('Frequency')

#plot the histogram for the corrected_loss,i.e. the generalization error estimated

#by a weighted sample

ax[1].bar(left = bins1[:-1], height = frequency1, width=0.05)

ax[1].plot([true_err] * 5,np.linspace(0,frequency1.max() + 0.1,5,endpoint=True), 'r--')

ax[1].set_xlabel('Corrected estimate of the generalization error')

ax[1].set_ylabel('Frequency')

# Tweak spacing to prevent clipping of ylabel

fig.tight_layout()

#show the target graph

plt.show()