-
Notifications
You must be signed in to change notification settings - Fork 0
/
NestedKMeans.py
120 lines (101 loc) · 3.74 KB
/
NestedKMeans.py
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
import numpy as np
import matplotlib.pyplot as plt
class NestedKMeans():
def __init__(self, x, k, b, rho=100, earlyStop=False):
self.x = x
self.k = k
self.b = b
self.N, self.ndim = x.shape
self.convCritirion = False
self.d = np.zeros(self.N)
self.a = np.zeros(self.N, dtype=np.int32) - 1
self.l = np.zeros([self.N, self.k]) - 1
self.rho = rho
self.earlyStop = earlyStop
def assign(self, i):
# argmin of euclidean distance
return (np.linalg.norm(self.c.reshape(-1, self.ndim) - self.x[i])).argmin()
def assignment_with_bounds(self, i):
self.d[i] = np.linalg.norm(self.c[self.a[i]] - self.x[i]).min()
for j in range(self.k):
if j != self.a[i]:
if self.l[i, j] < self.d[i]:
self.l[i, j] = np.linalg.norm(self.c[j] - self.x[i])
if self.l[i, j] < self.d[i]:
self.a[i] = j
self.d[i] = self.l[i, j]
def initialise_cSv(self):
# Algorithm 2
self.c = np.zeros([self.k, self.ndim])
self.S = np.zeros([self.k, self.ndim])
self.v = np.zeros(self.k, dtype=np.int32)
for j in range(self.k):
choice = np.random.choice(self.N)
self.c[j] = self.x[choice]
self.S[j] = self.x[choice]
self.v[j] = 1
def accumulate(self, i):
# Algorithm 3
am = self.a[i]
self.S[am] += self.x[i]
self.v[am] += 1
def train(self, max_iter):
# Algorithm 5
t = 1
M0 = 0
M1 = self.b
self.initialise_cSv()
sse = np.zeros(self.k)
p = np.zeros(self.k)
while True:
for i in range(M0):
for j in range(self.k):
self.l[i, j] -= p[j]
a_old = np.zeros(self.N, dtype=np.int32) - 1
for i in range(M0):
a_old[i] = self.a[i]
sse[a_old[i]] -= self.d[i]**2
self.S[a_old[i]] -= self.x[i]
self.v[a_old[i]] -= 1
self.assignment_with_bounds(i)
self.accumulate(i)
sse[self.a[i]] += self.d[i]**2
for i in range(M0, M1):
for j in range(self.k):
self.l[i, j] = np.linalg.norm(self.x[i] - self.c[j])
for i in range(M0, M1):
self.a[i] = self.l[i, :].argmin()
self.d[i] = self.l[i, self.a[i]]
self.accumulate(i)
sse[self.a[i]] += self.d[i]**2
c_old = np.zeros([self.k, self.ndim]) - 1
sigmaC = np.zeros(self.k) - 1
for j in range(self.k):
sigmaC[j] = np.sqrt(sse[j] / self.v[j] * (self.v[j] - 1))
c_old[j] = self.c[j]
self.c[j] = self.S[j] / self.v[j]
p[j] = np.linalg.norm(self.c[j] - c_old[j])
if (p == 0).all() and M1 == self.N:
print("Convergence Criterion Satisfied")
break
if self.earlyStop and M1 == self.N:
print("earlyStop")
break
if self.convCritirion or t > max_iter:
if M1 == self.N:
break
if np.nanmin(sigmaC / (p)) > self.rho:
M0 = M1
M1=min(2 * M1, self.N)
print("t: ", t, "M1: ", M1)
else:
M0 = M1
t += 1
def show(self, keep=False):
plt.cla()
plt.scatter(self.x[:, 0], self.x[:, 1], c=self.a, marker=".")
plt.plot(self.c[:, 0], self.c[:, 1], 'x', color="black", markersize=10)
plt.draw()
if keep:
plt.ioff()
plt.show()