forked from glennhickey/teHmm
-
Notifications
You must be signed in to change notification settings - Fork 0
/
_basehmm.pyx
173 lines (136 loc) · 6.2 KB
/
_basehmm.pyx
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
"""
We keep the original Scikit-Learn cython routines here mostly for consistency
checking and unit tests. The HMM uses the sped-up versions in _hmm.pyx, but
it's handy to have the original versions here to check the output's the same.
As it stands, MultinomialHMM, used only in unit tests, uses the methods below
while MultitrackHMM, used for everything else, uses the fast versions in
_hmm.pyx whenever possible. tests/hmmTests.py compares the output of
the two versions to make sure they remain the same.
-- Glenn Hickey, 2014
Derived from scikit-learn/sklearn/_hmmc.pyx
See below:
Copyright (c) 2007-2014 the scikit-learn developers.
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
a. Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
b. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
c. Neither the name of the Scikit-learn Developers nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR
ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
DAMAGE.
"""
from libc.math cimport exp, log
import numpy as np
cimport numpy as np
cimport cython
np.import_array()
ctypedef np.float64_t dtype_t
cdef dtype_t _NINF = -np.inf
@cython.boundscheck(False)
cdef dtype_t _max(dtype_t[:] values):
# find maximum value (builtin 'max' is unrolled for speed)
cdef dtype_t value
cdef dtype_t vmax = _NINF
for i in range(values.shape[0]):
value = values[i]
if value > vmax:
vmax = value
return vmax
@cython.boundscheck(False)
cpdef dtype_t _logsum(dtype_t[:] X):
cdef dtype_t vmax = _max(X)
cdef dtype_t power_sum = 0
for i in range(X.shape[0]):
power_sum += exp(X[i]-vmax)
return log(power_sum) + vmax
@cython.boundscheck(False)
def _forward(int n_observations, int n_components,
np.ndarray[dtype_t, ndim=1] log_startprob,
np.ndarray[dtype_t, ndim=2] log_transmat,
np.ndarray[dtype_t, ndim=2] framelogprob,
np.ndarray[dtype_t, ndim=2] fwdlattice):
cdef int t, i, j
cdef double logprob
cdef np.ndarray[dtype_t, ndim = 1] work_buffer
work_buffer = np.zeros(n_components)
for i in range(n_components):
fwdlattice[0, i] = log_startprob[i] + framelogprob[0, i]
for t in range(1, n_observations):
for j in range(n_components):
for i in range(n_components):
work_buffer[i] = fwdlattice[t - 1, i] + log_transmat[i, j]
fwdlattice[t, j] = _logsum(work_buffer) + framelogprob[t, j]
@cython.boundscheck(False)
def _backward(int n_observations, int n_components,
np.ndarray[dtype_t, ndim=1] log_startprob,
np.ndarray[dtype_t, ndim=2] log_transmat,
np.ndarray[dtype_t, ndim=2] framelogprob,
np.ndarray[dtype_t, ndim=2] bwdlattice):
cdef int t, i, j
cdef double logprob
cdef np.ndarray[dtype_t, ndim = 1] work_buffer
work_buffer = np.zeros(n_components)
for i in range(n_components):
bwdlattice[n_observations - 1, i] = 0.0
for t in range(n_observations - 2, -1, -1):
for i in range(n_components):
for j in range(n_components):
work_buffer[j] = log_transmat[i, j] + framelogprob[t + 1, j] \
+ bwdlattice[t + 1, j]
bwdlattice[t, i] = _logsum(work_buffer)
@cython.boundscheck(False)
def _compute_lneta(int n_observations, int n_components,
np.ndarray[dtype_t, ndim=2] fwdlattice,
np.ndarray[dtype_t, ndim=2] log_transmat,
np.ndarray[dtype_t, ndim=2] bwdlattice,
np.ndarray[dtype_t, ndim=2] framelogprob,
double logprob,
np.ndarray[dtype_t, ndim=3] lneta):
cdef int i, j, t
for t in range(n_observations - 1):
for i in range(n_components):
for j in range(n_components):
lneta[t, i, j] = fwdlattice[t, i] + log_transmat[i, j] \
+ framelogprob[t + 1, j] + bwdlattice[t + 1, j] - logprob
@cython.boundscheck(False)
def _viterbi(int n_observations, int n_components,
np.ndarray[dtype_t, ndim=1] log_startprob,
np.ndarray[dtype_t, ndim=2] log_transmat,
np.ndarray[dtype_t, ndim=2] framelogprob):
cdef int t, max_pos
cdef np.ndarray[dtype_t, ndim = 2] viterbi_lattice
cdef np.ndarray[np.int_t, ndim = 1] state_sequence
cdef dtype_t logprob
cdef np.ndarray[dtype_t, ndim = 2] work_buffer
# Initialization
state_sequence = np.empty(n_observations, dtype=np.int)
viterbi_lattice = np.zeros((n_observations, n_components))
viterbi_lattice[0] = log_startprob + framelogprob[0]
# Induction
for t in range(1, n_observations):
work_buffer = viterbi_lattice[t-1] + log_transmat.T
viterbi_lattice[t] = np.max(work_buffer, axis=1) + framelogprob[t]
# Observation traceback
max_pos = np.argmax(viterbi_lattice[n_observations - 1, :])
state_sequence[n_observations - 1] = max_pos
logprob = viterbi_lattice[n_observations - 1, max_pos]
for t in range(n_observations - 2, -1, -1):
max_pos = np.argmax(viterbi_lattice[t, :] \
+ log_transmat[:, state_sequence[t + 1]])
state_sequence[t] = max_pos
return state_sequence, logprob