measures.py
7.12 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
'''
This module is a part of my library.
It aims to compute some measures for clustering.
'''
import numpy as np
def disequilibrium_(matrix1, matrix2, isGlobal=False, mod=None):
'''
Compute disequilibrium for all the clusters.
The disequilibrium is compute from the difference
between two clustering sets.
isGlobal permet à l'utilisateur de choisir le dénominateur de
la fonction :
- True : divise la valeur par le nombre d'élément du cluster
- False : divise la valeur par le nombre d'élément total
withPower permet à l'utilisateur de décider d'appliquer un carré 2 ou
une valeur absolue.
'''
def divide_line(a, divider):
'''
Sub function used for dividing matrix by a vector line by line.
'''
return np.divide(a, divider, out=np.zeros_like(a), where=divider!=0)
dividers1 = 0
dividers2 = 0
if isGlobal:
dividers1 = matrix1.sum()
dividers2 = matrix2.sum()
else:
dividers1 = matrix1.sum(axis=1)
dividers2 = matrix2.sum(axis=1)
matrix1_divided = np.apply_along_axis(divide_line, 0, np.asarray(matrix1, dtype=np.float), dividers1)
matrix2_divided = np.apply_along_axis(divide_line, 0, np.asarray(matrix2, dtype=np.float), dividers2)
diff = matrix1_divided - matrix2_divided
mask = np.logical_not(np.logical_and(matrix2==0, matrix1==0))
result = diff
if mod != None or mod == "":
for word in mod.split(" "):
if word == "power":
result = np.power(result,2)
elif word == "human":
result = result * 100
elif word == "abs":
result = np.absolute(result)
else:
raise Exception("Need to specify an accepted mod of the disequilibrium (\"power\", \"human\" or \"abs\"")
return (mask, result)
def disequilibrium_mean_by_cluster(mask, matrix):
'''
Mean of disequilibrium
matrix is the disequilibrium calculated
from number of occurences belonging to a class,
for each cluster.
'''
nb_k = len(matrix)
results = np.zeros((nb_k))
for i in range(nb_k):
results[i] = matrix[i].sum() / mask[i].sum()
return results
def disequilibrium(matrix1, matrix2, isGlobal=False):
'''
Disequilibrium matrix
And Disequilibrium value
'''
mask, result = disequilibrium_(matrix1, matrix2, isGlobal)
result_human = result * 100
result_power = np.power(result, 2)
return (
mask,
result_human,
disequilibrium_mean_by_cluster(mask, result_power).sum()/matrix1.shape[0]
)
def compute_count_matrix(y_truth, y_hat):
'''
Check the size of the lists with assertion
'''
# Check size of the lists
assert len(y_hat) == len(y_truth), f"Matrices should have the same length y_hat: {len(y_hat)}, y_truth: {len(y_truth)}"
# Build count matrix
count_matrix = np.zeros((max(y_hat+1), max(y_truth+1)))
for i in range(len(y_hat)):
count_matrix[y_hat[i]][y_truth[i]] += 1
return count_matrix
def entropy_score(y_truth, y_hat):
'''
Need to use label encoder before givin y_hat and y_truth
Don't use one hot labels
Return a tuple with:
- result_matrix : the matrix with the log multiplied probabilities (P(x) * log(P(x)))
- result_vector : the vector avec summing entropy of each class. Each value corresponds to a cluster.
- result : the final entropy measure of the clustering
'''
def divide_line(a, divider):
'''
Sub function used for dividing matrix by a vector line by line.
'''
return np.divide(a, divider, out=np.zeros_like(a), where=divider!=0)
# Build count matrix
count_matrix = compute_count_matrix(y_truth, y_hat)
# Build dividers vector
dividers = count_matrix.sum(axis=1)
matrix_divided = np.apply_along_axis(divide_line, 0, np.asarray(count_matrix, dtype=np.float), dividers)
log_matrix = np.zeros(matrix_divided.shape)
np.log2(matrix_divided, out=log_matrix, where=count_matrix != 0)
result_matrix = -1 * np.multiply(matrix_divided, log_matrix)
result_vector = result_matrix.sum(axis=1)
result_vector.sum()
if np.isnan(np.sum(result_vector)):
print("COUNT MATRIX")
print(count_matrix)
print("MATRIX DIVIDED")
print(matrix_divided)
print("RESULT MATRIX")
print(result_matrix)
print("VECTOR MATRIX")
print(result_vector)
print("An error occured due to nan value, some values are printed before")
exit(1)
result = result_vector * dividers / dividers.sum()
result = result.sum()
return result
def purity_score(y_truth, y_hat):
'''
Return three values in a dictionary:
- purity_class_score: the purity score of the class (asp)
- purity_cluster_score: the purity score of the cluster (acp)
- K: the overall evaluation criterion (sqrt(asp * acp))
This function is based on the following article:
Unknown-multiple speaker clustering using HMM, J. Ajmera, H. Bourlard, I. Lapidot, I. McCowan
'''
def divide_line(a, divider):
'''
Sub function used for dividing matrix by a vector line by line.
'''
return np.divide(a, divider, out=np.zeros_like(a), where=divider!=0)
def compute_purity_score(count_matrix, axis=0):
if axis==0:
other_axis = 1
else:
other_axis = 0
count_per_row = count_matrix.sum(axis=axis)
dividers = np.square(count_per_row)
count_matrix_squared = np.square(count_matrix)
matrix_divided = np.apply_along_axis(divide_line, other_axis, np.asarray(count_matrix_squared, dtype=np.float), dividers)
vector_purity = np.sum(matrix_divided, axis=axis)
scalar_purity = np.average(vector_purity, weights=count_per_row)
return scalar_purity
count_matrix = compute_count_matrix(y_truth, y_hat)
purity_cluster_score = compute_purity_score(count_matrix, 1)
purity_class_score = compute_purity_score(count_matrix, 0)
K = np.sqrt(purity_cluster_score * purity_class_score)
for i in range(count_matrix.shape[0]):
for j in range(count_matrix.shape[1]):
count_matrix[i][j]
count_matrix[i]
return {
"purity_class_score": purity_class_score,
"purity_cluster_score": purity_cluster_score,
"K": K
}
if __name__ == "__main__":
print("Purity test #1")
# Hypothesis
y_hat = np.asarray([0, 1, 2, 0, 1, 0, 3, 2, 2, 3, 3, 0])
# Truth
y = np.asarray([0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3])
(result_matrix, result_vector, result) = entropy_score(y, y_hat)
print(purity_score(y, y_hat))
exit(1)
print("Purity test #2")
# Hypothesis
y_hat = np.asarray([0, 1, 2, 0, 1, 0, 3, 2, 2, 3, 3, 0, 4, 4, 4])
# Truth
y = np.asarray([0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 0, 3, 3, 3])
(result_matrix, result_vector, result) = entropy_score(y, y_hat)
exit(1)
print("Result matrix: ")
print(result_matrix)
print("Result vector: ")
print(result_vector)
print("Result: ", result)