勘误

本文是对“信息论中的各种熵”一文的勘误，并且提供了一个更加易用的脚本。

---

• 勘误

  • 概率越平均则熵越高
• Entropy

  • Generate data
  • All-in-One Class
  • Compute prob and jointProb

    • Display prob & jointProb
  • Compute entropy

    • Shannon entropy
    • Other entropy
  • Compare entropy

    • Select signal with minimal entropy
    • Order signals by joint entropy (ascending)
    • Order signals by mutual information (descending)
    • Order signals by cross entropy (ascending)

概率越平均则熵越高

从统计水平上说，

一个信号它的概率分布越平均，则它的熵越高。

之间的算法显然地违背了这个准则，因此在此进行勘误，并且提供了一个更加易用的脚本，它提供在 All-in-One class 脚本中。

另外，在 Compare entropy 部分中，我对联合熵、互信息和交叉熵的物理意义进行简要解释，尤其是给出了直观的例子，对它们的值的大小与信号之间“相像程度”之间的关系进行了简要解释。

Entropy

import scipy
import sklearn.preprocessing

import numpy as np
import pandas as pd
import plotly.express as px
import plotly.subplots as subplots
import plotly.graph_objects as go

from tqdm.auto import tqdm
from noise import pnoise1

Generate data

Create simulation dataset using Perlin noise. It includes signals with 1000 time points, and repeat 500 times. The 500 signals are independent, and the distributions are similar.

def generate_data(repeat=50, num=1000):

    data = []

    seeds = [e for e in range(num)]
    np.random.shuffle(seeds)
    seeds = seeds[:repeat]
    print(seeds)

    for seed in tqdm(seeds):
        octaves = 1 * 4 + 0 * np.random.randint(3, 7)
        persistence = 0 * 0.8 + 1 * np.random.uniform(0.7, 0.9)
        s = np.array([
            pnoise1(e, persistence=persistence, base=seed, octaves=octaves)
            for e in np.linspace(0, 1, num)
        ])
        data.append(s)

    data = np.array(data)

    data = sklearn.preprocessing.minmax_scale(data, axis=1)

    print(data.shape)
    return data

data = generate_data()

df = pd.DataFrame()
for j in range(5):
    df['signal-{}'.format(j)] = data[j]
px.scatter(df, title='Signals').show()

px.imshow(data, title='Signal').show()

All-in-One Class

class DataWithProb(object):
    '''
    Data with Probability

    Every entropy you need for a dataset.
    '''
    def __init__(self):
        pass

    def load(self, data):
        '''
        Load new data, the data is 2-d array,
        The 1st dimension is the signal.

        The function will compute things automatically
        - prob.: The empirical probability of every signal,
                 in 1-d array
        - joint prob.: The empirical joint probability of every two signals,
                       in 2-d array

        Since the data is inherently discrete, the discrete prob is computed.

        :param:data: The data will be computed
        '''
        self.data = data
        self.shape = data.shape
        print('Data shape is {}'.format(self.shape))
        bins = self.auto_bins()
        self.compute_prob(bins)
        self.compute_joint_prob(bins)

    def auto_bins(self, num=100):
        '''
        Compute bins by default setting.

        :param: num: A number of how many bins we will use

        :return: bins: Hhe linear seperated bins
        '''
        data = self.data
        bins = np.linspace(np.min(data), np.max(data), num)
        return bins

    def compute_prob(self, bins):
        '''
        Compute prob for every signal

        :param: bins: The bins
        :return: prob: The prob for every signal, the shape is n x (m-1),
                       where m is the count of bins, n is the count of signal
        '''
        data = self.data
        n = data.shape[0]
        m = bins.shape[0]

        prob = np.zeros((n, m-1))

        j = 0
        for d in tqdm(data, 'Prob.'):
            a, b = np.histogram(d, bins=bins)
            a = a.astype(np.float32)
            a /= np.sum(a)
            prob[j] = a
            j += 1
            pass

        self.prob = prob
        self.prob_bins = bins
        return prob

    def compute_joint_prob(self, bins):
        '''
        Compute joint prob for every two signal

        :param: bins: The bins
        :return: joint_prob: The joint_prob matrix of every two signal,
                             the shape is n x n x (m-1) x (m-1).
                             The first two n refer the signal pair;
                             the last two (m-1) refer the bins grid.
        '''
        data = self.data
        n = data.shape[0]
        m = bins.shape[0]

        joint_prob = np.zeros((n, n, m-1, m-1))

        for j in tqdm(range(n), 'Joint prob.'):
            for k in range(n):
                a, b, c = np.histogram2d(data[j], data[k], bins=bins)
                a = a.astype(np.float32)
                a /= np.sum(a)
                joint_prob[j][k] = a
                pass

        self.joint_prob = joint_prob
        self.joint_prob_bins = bins
        return joint_prob

    def shannon_entropy(self):
        '''
        Compute shannon entropy for every signal

        :return: entropy: The shannon entropy for every signal,
                          it is a 1-d array
        '''
        entropy = np.array([scipy.stats.entropy(p) for p in tqdm(self.prob, 'Shannon Envropy')])
        return entropy

    def joint_entropy(self):
        '''
        Compute joint entropy for every two signals

        :return: joint_entropy: The joint entropy for every two signal,
                                it is a 2-d array
        '''
        joint_prob = self.joint_prob
        n = joint_prob.shape[0]

        joint_entropy = np.zeros((n, n))

        for j in tqdm(range(n), 'Joint Entropy'):
            for k in range(n):
                p = joint_prob[j][k].flatten()
                e = scipy.stats.entropy(p)
                joint_entropy[j][k] = e
                pass

        return joint_entropy

    def mutual_information(self):
        '''
        Compute mutual information for every two signals

        :return: mutual_information: The mutual information for every two signals,
                                     it is a 2-d array
        '''
        prob = self.prob
        joint_prob = self.joint_prob
        n = joint_prob.shape[0]

        mutual_information = np.zeros((n, n))

        for j in tqdm(range(n), 'Mutual Information'):
            for k in range(n):
                p1 = prob[j][:, np.newaxis]
                p2 = prob[k][np.newaxis, :]
                pxy = np.matmul(p1, p2)
                pp = joint_prob[j, k]

                m = pp != 0
                pxy = pxy[m]
                pp = pp[m]

                s = pp * np.log(pxy / pp)
                mutual_information[j][k] = -np.sum(s)

        return mutual_information

    def cross_entropy(self):
        '''
        Compute cross entropy for every two signals

        :return: cross_entropy: The cross entropy for every two signals,
                                it is a 2-d array
        '''
        prob = self.prob
        n = prob.shape[0]

        cross_entropy = np.zeros((n, n))

        for j in tqdm(range(n), 'Cross Entropy'):
            for k in range(n):
                p1 = prob[j]
                p2 = prob[k]

                m = p2 != 0
                p1 = p1[m]
                p2 = p2[m]

                s = p1 * np.log(p2)
                cross_entropy[j][k] = -np.sum(s)

        return cross_entropy

Compute prob and jointProb

dwp = DataWithProb()
dwp.load(data)
dwp.prob.shape, dwp.joint_prob.shape

Display prob & jointProb

px.imshow(dwp.prob, title='Prob.', color_continuous_scale='dense').show()

d = dwp.joint_prob
bins = dwp.joint_prob_bins
px.imshow(d[0][1], title='Joint prob. (0, 1)', x=bins[:-1], y=bins[:-1], color_continuous_scale='dense').show()
px.imshow(np.mean(d, axis=0)[1], title='Joint prob. (all, 1)', x=bins[:-1], y=bins[:-1], color_continuous_scale='dense').show()

Compute entropy

Shannon entropy

shannon_entropy = dwp.shannon_entropy()
px.scatter(shannon_entropy, title='Shannon Entropy').show()

df = pd.DataFrame()

df['Max Ent. ({:0.2f})'.format(np.max(shannon_entropy))] = dwp.data[np.argmax(shannon_entropy)]
df['Min Ent. ({:0.2f})'.format(np.min(shannon_entropy))] = dwp.data[np.argmin(shannon_entropy)]

fig = subplots.make_subplots(rows=1, cols=2)

f = px.line(df)
for d in f.data:
    fig.append_trace(d, row=1, col=2)

f = px.violin(df)
for d in f.data:
    fig.append_trace(d, row=1, col=1)

fig.update_layout(title='Entropy')

fig.show()

Other entropy

joint_entropy = dwp.joint_entropy()
mutual_information = dwp.mutual_information()
cross_entropy = dwp.cross_entropy()

fig = subplots.make_subplots(rows=1, cols=3, subplot_titles=['Joint entropy', 'Mutual information', 'Cross entropy'])

f = px.imshow(joint_entropy, title='Joint Entropy')
fig.add_trace(f.data[0], row=1, col=1)

f = px.imshow(mutual_information, title='Joint Entropy')
fig.add_trace(f.data[0], row=1, col=2)

f = px.imshow(cross_entropy, title='Joint Entropy')
fig.add_trace(f.data[0], row=1, col=3)

fig.show()

Compare entropy

Select signal with minimal entropy

select = np.argmin(shannon_entropy)
print('Select the {}th signal by min entropy'.format(select))
px.line(dwp.data[select], title = 'Min entropy signal').show()

color_discrete_sequence = px.colors.sequential.PuBu_r

def draw_by_order(data, order, title, color_discrete_sequence=color_discrete_sequence):

    n = data.shape[1]

    dfs = [pd.DataFrame(dict(order=0, x=range(n), value=data[select]))]

    for j in range(5):
        dfs.append(
            pd.DataFrame(dict(order=j + 1, x=range(n), value=data[order[j]])))

    df = pd.concat(dfs, axis=0)

    fig = subplots.make_subplots(rows=2, cols=2, specs=[[{}, {}],
           [{"colspan": 2}, None]])

    # Plot waves
    f = px.line(df,
            x='x',
            y='value',
            color='order',
            color_discrete_sequence=color_discrete_sequence,
            title=title)#.show()

    for d in f.data:
        fig.add_trace(d, row=1, col=1)

    # Plot violin for prob.
    f = px.violin(df,
              y='value',
              x='order',
              color='order',
              color_discrete_sequence=color_discrete_sequence,
              title=title)#.show()

    for d in f.data:
        fig.add_trace(d, row=2, col=1)

    # Plot radar waves
    k = (2 * np.pi) / n
    df['corr_x'] = df['value'] * np.cos(df['x'] * k)
    df['corr_y'] = df['value'] * np.sin(df['x'] * k)

    f = px.line(df,
            x='corr_x',
            y='corr_y',
            color='order',
            color_discrete_sequence=color_discrete_sequence,
            title=title)
    f.update_layout(yaxis = dict(scaleanchor = 'x'))

    for d in f.data:
        fig.add_trace(d, row=1, col=2)

    fig.update_layout(height=800, width=800, title=title)
    fig.show()

    return df

Order signals by joint entropy (ascending)

The joint entropy is smaller means it requires little information added to estimate the other signal based on the signal.

Order signals by joint entropy (ascending)

order = np.argsort(joint_entropy[select])
title = 'Joint entropy (top 5)'
data = dwp.data

draw_by_order(data, order, title)

Order signals by mutual information (descending)

The mutual information is larger means the less independent between the two signals.

Order signals by mutual information (descending)

order = np.argsort(mutual_information[select])[::-1]
title = 'Mutual Information (top 5)'
data = dwp.data

draw_by_order(data, order, title)

Order signals by cross entropy (ascending)

The cross entropy is smaller means the less different between the signals.

Order signals by cross entropy (ascending)

order = np.argsort(cross_entropy[select])
title = 'Cross Entropy (top 5)'
data = dwp.data

draw_by_order(data, order, title)