-
Notifications
You must be signed in to change notification settings - Fork 736
/
hmm.py
985 lines (807 loc) · 39.4 KB
/
hmm.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
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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
# Hidden Markov Models
#
# Author: Ron Weiss <ronweiss@gmail.com>
# Shiqiao Du <lucidfrontier.45@gmail.com>
# API changes: Jaques Grobler <jaquesgrobler@gmail.com>
# Modifications to create of the HMMLearn module: Gael Varoquaux
# More API changes: Sergei Lebedev <superbobry@gmail.com>
"""
The :mod:`hmmlearn.hmm` module implements hidden Markov models.
"""
import numpy as np
from scipy.special import logsumexp
from sklearn import cluster
from sklearn.utils import check_random_state
from . import _utils
from .stats import log_multivariate_normal_density
from .base import _BaseHMM
from .utils import iter_from_X_lengths, normalize, fill_covars
__all__ = ["GMMHMM", "GaussianHMM", "MultinomialHMM"]
COVARIANCE_TYPES = frozenset(("spherical", "diag", "full", "tied"))
class GaussianHMM(_BaseHMM):
r"""Hidden Markov Model with Gaussian emissions.
Parameters
----------
n_components : int
Number of states.
covariance_type : string, optional
String describing the type of covariance parameters to
use. Must be one of
* "spherical" --- each state uses a single variance value that
applies to all features.
* "diag" --- each state uses a diagonal covariance matrix.
* "full" --- each state uses a full (i.e. unrestricted)
covariance matrix.
* "tied" --- all states use **the same** full covariance matrix.
Defaults to "diag".
min_covar : float, optional
Floor on the diagonal of the covariance matrix to prevent
overfitting. Defaults to 1e-3.
startprob_prior : array, shape (n_components, ), optional
Parameters of the Dirichlet prior distribution for
:attr:`startprob_`.
transmat_prior : array, shape (n_components, n_components), optional
Parameters of the Dirichlet prior distribution for each row
of the transition probabilities :attr:`transmat_`.
means_prior, means_weight : array, shape (n_components, ), optional
Mean and precision of the Normal prior distribtion for
:attr:`means_`.
covars_prior, covars_weight : array, shape (n_components, ), optional
Parameters of the prior distribution for the covariance matrix
:attr:`covars_`.
If :attr:`covariance_type` is "spherical" or "diag" the prior is
the inverse gamma distribution, otherwise --- the inverse Wishart
distribution.
algorithm : string, optional
Decoder algorithm. Must be one of "viterbi" or`"map".
Defaults to "viterbi".
random_state: RandomState or an int seed, optional
A random number generator instance.
n_iter : int, optional
Maximum number of iterations to perform.
tol : float, optional
Convergence threshold. EM will stop if the gain in log-likelihood
is below this value.
verbose : bool, optional
When ``True`` per-iteration convergence reports are printed
to :data:`sys.stderr`. You can diagnose convergence via the
:attr:`monitor_` attribute.
params : string, optional
Controls which parameters are updated in the training
process. Can contain any combination of 's' for startprob,
't' for transmat, 'm' for means and 'c' for covars. Defaults
to all parameters.
init_params : string, optional
Controls which parameters are initialized prior to
training. Can contain any combination of 's' for
startprob, 't' for transmat, 'm' for means and 'c' for covars.
Defaults to all parameters.
Attributes
----------
n_features : int
Dimensionality of the Gaussian emissions.
monitor\_ : ConvergenceMonitor
Monitor object used to check the convergence of EM.
transmat\_ : array, shape (n_components, n_components)
Matrix of transition probabilities between states.
startprob\_ : array, shape (n_components, )
Initial state occupation distribution.
means\_ : array, shape (n_components, n_features)
Mean parameters for each state.
covars\_ : array
Covariance parameters for each state.
The shape depends on :attr:`covariance_type`::
(n_components, ) if "spherical",
(n_features, n_features) if "tied",
(n_components, n_features) if "diag",
(n_components, n_features, n_features) if "full"
Examples
--------
>>> from hmmlearn.hmm import GaussianHMM
>>> GaussianHMM(n_components=2)
... #doctest: +ELLIPSIS +NORMALIZE_WHITESPACE
GaussianHMM(algorithm='viterbi',...
"""
def __init__(self, n_components=1, covariance_type='diag',
min_covar=1e-3,
startprob_prior=1.0, transmat_prior=1.0,
means_prior=0, means_weight=0,
covars_prior=1e-2, covars_weight=1,
algorithm="viterbi", random_state=None,
n_iter=10, tol=1e-2, verbose=False,
params="stmc", init_params="stmc"):
_BaseHMM.__init__(self, n_components,
startprob_prior=startprob_prior,
transmat_prior=transmat_prior, algorithm=algorithm,
random_state=random_state, n_iter=n_iter,
tol=tol, params=params, verbose=verbose,
init_params=init_params)
self.covariance_type = covariance_type
self.min_covar = min_covar
self.means_prior = means_prior
self.means_weight = means_weight
self.covars_prior = covars_prior
self.covars_weight = covars_weight
@property
def covars_(self):
"""Return covars as a full matrix."""
return fill_covars(self._covars_, self.covariance_type,
self.n_components, self.n_features)
@covars_.setter
def covars_(self, covars):
self._covars_ = np.asarray(covars).copy()
def _check(self):
super(GaussianHMM, self)._check()
self.means_ = np.asarray(self.means_)
self.n_features = self.means_.shape[1]
if self.covariance_type not in COVARIANCE_TYPES:
raise ValueError('covariance_type must be one of {}'
.format(COVARIANCE_TYPES))
_utils._validate_covars(self._covars_, self.covariance_type,
self.n_components)
def _init(self, X, lengths=None):
super(GaussianHMM, self)._init(X, lengths=lengths)
_, n_features = X.shape
if hasattr(self, 'n_features') and self.n_features != n_features:
raise ValueError('Unexpected number of dimensions, got %s but '
'expected %s' % (n_features, self.n_features))
self.n_features = n_features
if 'm' in self.init_params or not hasattr(self, "means_"):
kmeans = cluster.KMeans(n_clusters=self.n_components,
random_state=self.random_state)
kmeans.fit(X)
self.means_ = kmeans.cluster_centers_
if 'c' in self.init_params or not hasattr(self, "covars_"):
cv = np.cov(X.T) + self.min_covar * np.eye(X.shape[1])
if not cv.shape:
cv.shape = (1, 1)
self._covars_ = \
_utils.distribute_covar_matrix_to_match_covariance_type(
cv, self.covariance_type, self.n_components).copy()
def _compute_log_likelihood(self, X):
return log_multivariate_normal_density(
X, self.means_, self._covars_, self.covariance_type)
def _generate_sample_from_state(self, state, random_state=None):
random_state = check_random_state(random_state)
return random_state.multivariate_normal(
self.means_[state], self.covars_[state]
)
def _initialize_sufficient_statistics(self):
stats = super(GaussianHMM, self)._initialize_sufficient_statistics()
stats['post'] = np.zeros(self.n_components)
stats['obs'] = np.zeros((self.n_components, self.n_features))
stats['obs**2'] = np.zeros((self.n_components, self.n_features))
if self.covariance_type in ('tied', 'full'):
stats['obs*obs.T'] = np.zeros((self.n_components, self.n_features,
self.n_features))
return stats
def _accumulate_sufficient_statistics(self, stats, obs, framelogprob,
posteriors, fwdlattice, bwdlattice):
super(GaussianHMM, self)._accumulate_sufficient_statistics(
stats, obs, framelogprob, posteriors, fwdlattice, bwdlattice)
if 'm' in self.params or 'c' in self.params:
stats['post'] += posteriors.sum(axis=0)
stats['obs'] += np.dot(posteriors.T, obs)
if 'c' in self.params:
if self.covariance_type in ('spherical', 'diag'):
stats['obs**2'] += np.dot(posteriors.T, obs ** 2)
elif self.covariance_type in ('tied', 'full'):
# posteriors: (nt, nc); obs: (nt, nf); obs: (nt, nf)
# -> (nc, nf, nf)
stats['obs*obs.T'] += np.einsum(
'ij,ik,il->jkl', posteriors, obs, obs)
def _do_mstep(self, stats):
super(GaussianHMM, self)._do_mstep(stats)
means_prior = self.means_prior
means_weight = self.means_weight
# TODO: find a proper reference for estimates for different
# covariance models.
# Based on Huang, Acero, Hon, "Spoken Language Processing",
# p. 443 - 445
denom = stats['post'][:, np.newaxis]
if 'm' in self.params:
self.means_ = ((means_weight * means_prior + stats['obs'])
/ (means_weight + denom))
if 'c' in self.params:
covars_prior = self.covars_prior
covars_weight = self.covars_weight
meandiff = self.means_ - means_prior
if self.covariance_type in ('spherical', 'diag'):
cv_num = (means_weight * meandiff**2
+ stats['obs**2']
- 2 * self.means_ * stats['obs']
+ self.means_**2 * denom)
cv_den = max(covars_weight - 1, 0) + denom
self._covars_ = \
(covars_prior + cv_num) / np.maximum(cv_den, 1e-5)
if self.covariance_type == 'spherical':
self._covars_ = np.tile(
self._covars_.mean(1)[:, np.newaxis],
(1, self._covars_.shape[1]))
elif self.covariance_type in ('tied', 'full'):
cv_num = np.empty((self.n_components, self.n_features,
self.n_features))
for c in range(self.n_components):
obsmean = np.outer(stats['obs'][c], self.means_[c])
cv_num[c] = (means_weight * np.outer(meandiff[c],
meandiff[c])
+ stats['obs*obs.T'][c]
- obsmean - obsmean.T
+ np.outer(self.means_[c], self.means_[c])
* stats['post'][c])
cvweight = max(covars_weight - self.n_features, 0)
if self.covariance_type == 'tied':
self._covars_ = ((covars_prior + cv_num.sum(axis=0)) /
(cvweight + stats['post'].sum()))
elif self.covariance_type == 'full':
self._covars_ = ((covars_prior + cv_num) /
(cvweight + stats['post'][:, None, None]))
class MultinomialHMM(_BaseHMM):
r"""Hidden Markov Model with multinomial (discrete) emissions
Parameters
----------
n_components : int
Number of states.
startprob_prior : array, shape (n_components, ), optional
Parameters of the Dirichlet prior distribution for
:attr:`startprob_`.
transmat_prior : array, shape (n_components, n_components), optional
Parameters of the Dirichlet prior distribution for each row
of the transition probabilities :attr:`transmat_`.
algorithm : string, optional
Decoder algorithm. Must be one of "viterbi" or "map".
Defaults to "viterbi".
random_state: RandomState or an int seed, optional
A random number generator instance.
n_iter : int, optional
Maximum number of iterations to perform.
tol : float, optional
Convergence threshold. EM will stop if the gain in log-likelihood
is below this value.
verbose : bool, optional
When ``True`` per-iteration convergence reports are printed
to :data:`sys.stderr`. You can diagnose convergence via the
:attr:`monitor_` attribute.
params : string, optional
Controls which parameters are updated in the training
process. Can contain any combination of 's' for startprob,
't' for transmat, 'e' for emissionprob.
Defaults to all parameters.
init_params : string, optional
Controls which parameters are initialized prior to
training. Can contain any combination of 's' for
startprob, 't' for transmat, 'e' for emissionprob.
Defaults to all parameters.
Attributes
----------
n_features : int
Number of possible symbols emitted by the model (in the samples).
monitor\_ : ConvergenceMonitor
Monitor object used to check the convergence of EM.
transmat\_ : array, shape (n_components, n_components)
Matrix of transition probabilities between states.
startprob\_ : array, shape (n_components, )
Initial state occupation distribution.
emissionprob\_ : array, shape (n_components, n_features)
Probability of emitting a given symbol when in each state.
Examples
--------
>>> from hmmlearn.hmm import MultinomialHMM
>>> MultinomialHMM(n_components=2)
... #doctest: +ELLIPSIS +NORMALIZE_WHITESPACE
MultinomialHMM(algorithm='viterbi',...
"""
# TODO: accept the prior on emissionprob_ for consistency.
def __init__(self, n_components=1,
startprob_prior=1.0, transmat_prior=1.0,
algorithm="viterbi", random_state=None,
n_iter=10, tol=1e-2, verbose=False,
params="ste", init_params="ste"):
_BaseHMM.__init__(self, n_components,
startprob_prior=startprob_prior,
transmat_prior=transmat_prior,
algorithm=algorithm,
random_state=random_state,
n_iter=n_iter, tol=tol, verbose=verbose,
params=params, init_params=init_params)
def _init(self, X, lengths=None):
if not self._check_input_symbols(X):
raise ValueError("expected a sample from "
"a Multinomial distribution.")
super(MultinomialHMM, self)._init(X, lengths=lengths)
self.random_state = check_random_state(self.random_state)
if 'e' in self.init_params:
if not hasattr(self, "n_features"):
symbols = set()
for i, j in iter_from_X_lengths(X, lengths):
symbols |= set(X[i:j].flatten())
self.n_features = len(symbols)
self.emissionprob_ = self.random_state \
.rand(self.n_components, self.n_features)
normalize(self.emissionprob_, axis=1)
def _check(self):
super(MultinomialHMM, self)._check()
self.emissionprob_ = np.atleast_2d(self.emissionprob_)
n_features = getattr(self, "n_features", self.emissionprob_.shape[1])
if self.emissionprob_.shape != (self.n_components, n_features):
raise ValueError(
"emissionprob_ must have shape (n_components, n_features)")
else:
self.n_features = n_features
def _compute_log_likelihood(self, X):
return np.log(self.emissionprob_)[:, np.concatenate(X)].T
def _generate_sample_from_state(self, state, random_state=None):
cdf = np.cumsum(self.emissionprob_[state, :])
random_state = check_random_state(random_state)
return [(cdf > random_state.rand()).argmax()]
def _initialize_sufficient_statistics(self):
stats = super(MultinomialHMM, self)._initialize_sufficient_statistics()
stats['obs'] = np.zeros((self.n_components, self.n_features))
return stats
def _accumulate_sufficient_statistics(self, stats, X, framelogprob,
posteriors, fwdlattice, bwdlattice):
super(MultinomialHMM, self)._accumulate_sufficient_statistics(
stats, X, framelogprob, posteriors, fwdlattice, bwdlattice)
if 'e' in self.params:
for t, symbol in enumerate(np.concatenate(X)):
stats['obs'][:, symbol] += posteriors[t]
def _do_mstep(self, stats):
super(MultinomialHMM, self)._do_mstep(stats)
if 'e' in self.params:
self.emissionprob_ = (stats['obs']
/ stats['obs'].sum(axis=1)[:, np.newaxis])
def _check_input_symbols(self, X):
"""Check if ``X`` is a sample from a Multinomial distribution.
That is ``X`` should be an array of non-negative integers from
range ``[min(X), max(X)]``, such that each integer from the range
occurs in ``X`` at least once.
For example ``[0, 0, 2, 1, 3, 1, 1]`` is a valid sample from a
Multinomial distribution, while ``[0, 0, 3, 5, 10]`` is not.
"""
symbols = np.concatenate(X)
if (len(symbols) == 1 # not enough data
or not np.issubdtype(symbols.dtype, np.integer) # not an integer
or (symbols < 0).any()): # not positive
return False
u = np.unique(symbols)
return u[0] == 0 and u[-1] == len(u) - 1
class GMMHMM(_BaseHMM):
r"""Hidden Markov Model with Gaussian mixture emissions.
Parameters
----------
n_components : int
Number of states in the model.
n_mix : int
Number of states in the GMM.
covariance_type : string, optional
String describing the type of covariance parameters to
use. Must be one of
* "spherical" --- each state uses a single variance value that
applies to all features.
* "diag" --- each state uses a diagonal covariance matrix.
* "full" --- each state uses a full (i.e. unrestricted)
covariance matrix.
* "tied" --- all states use **the same** full covariance matrix.
Defaults to "diag".
min_covar : float, optional
Floor on the diagonal of the covariance matrix to prevent
overfitting. Defaults to 1e-3.
startprob_prior : array, shape (n_components, ), optional
Parameters of the Dirichlet prior distribution for
:attr:`startprob_`.
transmat_prior : array, shape (n_components, n_components), optional
Parameters of the Dirichlet prior distribution for each row
of the transition probabilities :attr:`transmat_`.
weights_prior : array, shape (n_mix, ), optional
Parameters of the Dirichlet prior distribution for
:attr:`weights_`.
means_prior, means_weight : array, shape (n_mix, ), optional
Mean and precision of the Normal prior distribtion for
:attr:`means_`.
covars_prior, covars_weight : array, shape (n_mix, ), optional
Parameters of the prior distribution for the covariance matrix
:attr:`covars_`.
If :attr:`covariance_type` is "spherical" or "diag" the prior is
the inverse gamma distribution, otherwise --- the inverse Wishart
distribution.
algorithm : string, optional
Decoder algorithm. Must be one of "viterbi" or "map".
Defaults to "viterbi".
random_state: RandomState or an int seed, optional
A random number generator instance.
n_iter : int, optional
Maximum number of iterations to perform.
tol : float, optional
Convergence threshold. EM will stop if the gain in log-likelihood
is below this value.
verbose : bool, optional
When ``True`` per-iteration convergence reports are printed
to :data:`sys.stderr`. You can diagnose convergence via the
:attr:`monitor_` attribute.
init_params : string, optional
Controls which parameters are initialized prior to training. Can
contain any combination of 's' for startprob, 't' for transmat, 'm'
for means, 'c' for covars, and 'w' for GMM mixing weights.
Defaults to all parameters.
params : string, optional
Controls which parameters are updated in the training process. Can
contain any combination of 's' for startprob, 't' for transmat, 'm' for
means, and 'c' for covars, and 'w' for GMM mixing weights.
Defaults to all parameters.
Attributes
----------
monitor\_ : ConvergenceMonitor
Monitor object used to check the convergence of EM.
startprob\_ : array, shape (n_components, )
Initial state occupation distribution.
transmat\_ : array, shape (n_components, n_components)
Matrix of transition probabilities between states.
weights\_ : array, shape (n_components, n_mix)
Mixture weights for each state.
means\_ : array, shape (n_components, n_mix)
Mean parameters for each mixture component in each state.
covars\_ : array
Covariance parameters for each mixture components in each state.
The shape depends on :attr:`covariance_type`::
(n_components, n_mix) if "spherical",
(n_components, n_features, n_features) if "tied",
(n_components, n_mix, n_features) if "diag",
(n_components, n_mix, n_features, n_features) if "full"
"""
def __init__(self, n_components=1, n_mix=1,
min_covar=1e-3, startprob_prior=1.0, transmat_prior=1.0,
weights_prior=1.0, means_prior=0.0, means_weight=0.0,
covars_prior=None, covars_weight=None,
algorithm="viterbi", covariance_type="diag",
random_state=None, n_iter=10, tol=1e-2,
verbose=False, params="stmcw",
init_params="stmcw"):
_BaseHMM.__init__(self, n_components,
startprob_prior=startprob_prior,
transmat_prior=transmat_prior,
algorithm=algorithm, random_state=random_state,
n_iter=n_iter, tol=tol, verbose=verbose,
params=params, init_params=init_params)
self.covariance_type = covariance_type
self.min_covar = min_covar
self.n_mix = n_mix
self.weights_prior = weights_prior
self.means_prior = means_prior
self.means_weight = means_weight
self.covars_prior = covars_prior
self.covars_weight = covars_weight
def _init(self, X, lengths=None):
super(GMMHMM, self)._init(X, lengths=lengths)
_n_samples, self.n_features = X.shape
# Default values for covariance prior parameters
self._init_covar_priors()
self._fix_priors_shape()
main_kmeans = cluster.KMeans(n_clusters=self.n_components,
random_state=self.random_state)
labels = main_kmeans.fit_predict(X)
kmeanses = []
for label in range(self.n_components):
kmeans = cluster.KMeans(n_clusters=self.n_mix,
random_state=self.random_state)
kmeans.fit(X[np.where(labels == label)])
kmeanses.append(kmeans)
if 'w' in self.init_params or not hasattr(self, "weights_"):
self.weights_ = (np.ones((self.n_components, self.n_mix)) /
(np.ones((self.n_components, 1)) * self.n_mix))
if 'm' in self.init_params or not hasattr(self, "means_"):
self.means_ = np.zeros((self.n_components, self.n_mix,
self.n_features))
for i, kmeans in enumerate(kmeanses):
self.means_[i] = kmeans.cluster_centers_
if 'c' in self.init_params or not hasattr(self, "covars_"):
cv = np.cov(X.T) + self.min_covar * np.eye(self.n_features)
if not cv.shape:
cv.shape = (1, 1)
if self.covariance_type == 'tied':
self.covars_ = np.zeros((self.n_components,
self.n_features, self.n_features))
self.covars_[:] = cv
elif self.covariance_type == 'full':
self.covars_ = np.zeros((self.n_components, self.n_mix,
self.n_features, self.n_features))
self.covars_[:] = cv
elif self.covariance_type == 'diag':
self.covars_ = np.zeros((self.n_components, self.n_mix,
self.n_features))
self.covars_[:] = np.diag(cv)
elif self.covariance_type == 'spherical':
self.covars_ = np.zeros((self.n_components, self.n_mix))
self.covars_[:] = cv.mean()
def _init_covar_priors(self):
if self.covariance_type == "full":
if self.covars_prior is None:
self.covars_prior = 0.0
if self.covars_weight is None:
self.covars_weight = -(1.0 + self.n_features + 1.0)
elif self.covariance_type == "tied":
if self.covars_prior is None:
self.covars_prior = 0.0
if self.covars_weight is None:
self.covars_weight = -(self.n_mix + self.n_features + 1.0)
elif self.covariance_type == "diag":
if self.covars_prior is None:
self.covars_prior = -1.5
if self.covars_weight is None:
self.covars_weight = 0.0
elif self.covariance_type == "spherical":
if self.covars_prior is None:
self.covars_prior = -(self.n_mix + 2.0) / 2.0
if self.covars_weight is None:
self.covars_weight = 0.0
def _fix_priors_shape(self):
# If priors are numbers, this function will make them into a
# matrix of proper shape
self.weights_prior = np.broadcast_to(
self.weights_prior, (self.n_components, self.n_mix)).copy()
self.means_prior = np.broadcast_to(
self.means_prior,
(self.n_components, self.n_mix, self.n_features)).copy()
self.means_weight = np.broadcast_to(
self.means_weight,
(self.n_components, self.n_mix)).copy()
if self.covariance_type == "full":
self.covars_prior = np.broadcast_to(
self.covars_prior,
(self.n_components, self.n_mix,
self.n_features, self.n_features)).copy()
self.covars_weight = np.broadcast_to(
self.covars_weight, (self.n_components, self.n_mix)).copy()
elif self.covariance_type == "tied":
self.covars_prior = np.broadcast_to(
self.covars_prior,
(self.n_components, self.n_features, self.n_features)).copy()
self.covars_weight = np.broadcast_to(
self.covars_weight, self.n_components).copy()
elif self.covariance_type == "diag":
self.covars_prior = np.broadcast_to(
self.covars_prior,
(self.n_components, self.n_mix, self.n_features)).copy()
self.covars_weight = np.broadcast_to(
self.covars_weight,
(self.n_components, self.n_mix, self.n_features)).copy()
elif self.covariance_type == "spherical":
self.covars_prior = np.broadcast_to(
self.covars_prior, (self.n_components, self.n_mix)).copy()
self.covars_weight = np.broadcast_to(
self.covars_weight, (self.n_components, self.n_mix)).copy()
def _check(self):
super(GMMHMM, self)._check()
if not hasattr(self, "n_features"):
self.n_features = self.means_.shape[2]
self._init_covar_priors()
self._fix_priors_shape()
# Checking covariance type
if self.covariance_type not in COVARIANCE_TYPES:
raise ValueError("covariance_type must be one of {}"
.format(COVARIANCE_TYPES))
self.weights_ = np.array(self.weights_)
# Checking mixture weights' shape
if self.weights_.shape != (self.n_components, self.n_mix):
raise ValueError("mixture weights must have shape "
"(n_components, n_mix), actual shape: {}"
.format(self.weights_.shape))
# Checking mixture weights' mathematical correctness
if not np.allclose(np.sum(self.weights_, axis=1),
np.ones(self.n_components)):
raise ValueError("mixture weights must sum up to 1")
# Checking means' shape
self.means_ = np.array(self.means_)
if self.means_.shape != (self.n_components, self.n_mix,
self.n_features):
raise ValueError("mixture means must have shape "
"(n_components, n_mix, n_features), "
"actual shape: {}".format(self.means_.shape))
# Checking covariances' shape
self.covars_ = np.array(self.covars_)
covars_shape = self.covars_.shape
needed_shapes = {
"spherical": (self.n_components, self.n_mix),
"tied": (self.n_components, self.n_features, self.n_features),
"diag": (self.n_components, self.n_mix, self.n_features),
"full": (self.n_components, self.n_mix,
self.n_features, self.n_features)
}
needed_shape = needed_shapes[self.covariance_type]
if covars_shape != needed_shape:
raise ValueError("{!r} mixture covars must have shape {}, "
"actual shape: {}"
.format(self.covariance_type,
needed_shape, covars_shape))
# Checking covariances' mathematical correctness
from scipy import linalg
if (self.covariance_type == "spherical" or
self.covariance_type == "diag"):
if np.any(self.covars_ <= 0):
raise ValueError("{!r} mixture covars must be non-negative"
.format(self.covariance_type))
elif self.covariance_type == "tied":
for i, covar in enumerate(self.covars_):
if (not np.allclose(covar, covar.T) or
np.any(linalg.eigvalsh(covar) <= 0)):
raise ValueError("'tied' mixture covars must be "
"symmetric, positive-definite")
elif self.covariance_type == "full":
for i, mix_covars in enumerate(self.covars_):
for j, covar in enumerate(mix_covars):
if (not np.allclose(covar, covar.T) or
np.any(linalg.eigvalsh(covar) <= 0)):
raise ValueError(
"'full' covariance matrix of mixture {} of "
"component {} must be symmetric, positive-definite"
.format(j, i))
def _generate_sample_from_state(self, state, random_state=None):
if random_state is None:
random_state = self.random_state
random_state = check_random_state(random_state)
cur_weights = self.weights_[state]
i_gauss = random_state.choice(self.n_mix, p=cur_weights)
if self.covariance_type == 'tied':
# self.covars_.shape == (n_components, n_features, n_features)
# shouldn't that be (n_mix, ...)?
covs = self.covars_
else:
covs = self.covars_[:, i_gauss]
covs = fill_covars(covs, self.covariance_type,
self.n_components, self.n_features)
return random_state.multivariate_normal(
self.means_[state, i_gauss], covs[state]
)
def _compute_log_weighted_gaussian_densities(self, X, i_comp):
cur_means = self.means_[i_comp]
cur_covs = self.covars_[i_comp]
if self.covariance_type == 'spherical':
cur_covs = cur_covs[:, np.newaxis]
log_cur_weights = np.log(self.weights_[i_comp])
return log_multivariate_normal_density(
X, cur_means, cur_covs, self.covariance_type
) + log_cur_weights
def _compute_log_likelihood(self, X):
n_samples, _ = X.shape
res = np.zeros((n_samples, self.n_components))
for i in range(self.n_components):
log_denses = self._compute_log_weighted_gaussian_densities(X, i)
with np.errstate(under="ignore"):
res[:, i] = logsumexp(log_denses, axis=1)
return res
def _initialize_sufficient_statistics(self):
stats = super(GMMHMM, self)._initialize_sufficient_statistics()
stats['n_samples'] = 0
stats['post_comp_mix'] = None
stats['post_mix_sum'] = np.zeros((self.n_components, self.n_mix))
stats['post_sum'] = np.zeros(self.n_components)
stats['samples'] = None
stats['centered'] = None
return stats
def _accumulate_sufficient_statistics(self, stats, X, framelogprob,
post_comp, fwdlattice, bwdlattice):
# TODO: support multiple frames
super(GMMHMM, self)._accumulate_sufficient_statistics(
stats, X, framelogprob, post_comp, fwdlattice, bwdlattice
)
n_samples, _ = X.shape
stats['n_samples'] = n_samples
stats['samples'] = X
prob_mix = np.zeros((n_samples, self.n_components, self.n_mix))
for p in range(self.n_components):
log_denses = self._compute_log_weighted_gaussian_densities(X, p)
with np.errstate(under="ignore"):
prob_mix[:, p, :] = np.exp(log_denses) + np.finfo(np.float).eps
prob_mix_sum = np.sum(prob_mix, axis=2)
post_mix = prob_mix / prob_mix_sum[:, :, np.newaxis]
post_comp_mix = post_comp[:, :, np.newaxis] * post_mix
stats['post_comp_mix'] = post_comp_mix
stats['post_mix_sum'] = np.sum(post_comp_mix, axis=0)
stats['post_sum'] = np.sum(post_comp, axis=0)
stats['centered'] = X[:, np.newaxis, np.newaxis, :] - self.means_
def _do_mstep(self, stats):
super(GMMHMM, self)._do_mstep(stats)
n_samples = stats['n_samples']
n_features = self.n_features
# Maximizing weights
alphas_minus_one = self.weights_prior - 1
new_weights_numer = stats['post_mix_sum'] + alphas_minus_one
new_weights_denom = (
stats['post_sum'] + np.sum(alphas_minus_one, axis=1)
)[:, np.newaxis]
new_weights = new_weights_numer / new_weights_denom
# Maximizing means
lambdas, mus = self.means_weight, self.means_prior
new_means_numer = np.einsum(
'ijk,il->jkl',
stats['post_comp_mix'], stats['samples']
) + lambdas[:, :, np.newaxis] * mus
new_means_denom = (stats['post_mix_sum'] + lambdas)[:, :, np.newaxis]
new_means = new_means_numer / new_means_denom
# Maximizing covariances
centered_means = self.means_ - mus
if self.covariance_type == 'full':
centered = stats['centered'].reshape((
n_samples, self.n_components, self.n_mix, self.n_features, 1
))
centered_t = stats['centered'].reshape((
n_samples, self.n_components, self.n_mix, 1, self.n_features
))
centered_dots = centered * centered_t
psis_t = np.transpose(self.covars_prior, axes=(0, 1, 3, 2))
nus = self.covars_weight
centr_means_resh = centered_means.reshape((
self.n_components, self.n_mix, self.n_features, 1
))
centr_means_resh_t = centered_means.reshape((
self.n_components, self.n_mix, 1, self.n_features
))
centered_means_dots = centr_means_resh * centr_means_resh_t
new_cov_numer = np.einsum(
'ijk,ijklm->jklm',
stats['post_comp_mix'], centered_dots
) + psis_t + (lambdas[:, :, np.newaxis, np.newaxis] *
centered_means_dots)
new_cov_denom = (
stats['post_mix_sum'] + 1 + nus + self.n_features + 1
)[:, :, np.newaxis, np.newaxis]
new_cov = new_cov_numer / new_cov_denom
elif self.covariance_type == 'diag':
centered2 = stats['centered'] ** 2
centered_means2 = centered_means ** 2
alphas = self.covars_prior
betas = self.covars_weight
new_cov_numer = np.einsum(
'ijk,ijkl->jkl',
stats['post_comp_mix'], centered2
) + lambdas[:, :, np.newaxis] * centered_means2 + 2 * betas
new_cov_denom = (
stats['post_mix_sum'][:, :, np.newaxis] + 1 + 2 * (alphas + 1)
)
new_cov = new_cov_numer / new_cov_denom
elif self.covariance_type == 'spherical':
centered_norm2 = np.sum(stats['centered'] ** 2, axis=-1)
alphas = self.covars_prior
betas = self.covars_weight
centered_means_norm2 = np.sum(centered_means ** 2, axis=-1)
new_cov_numer = np.einsum(
'ijk,ijk->jk',
stats['post_comp_mix'], centered_norm2
) + lambdas * centered_means_norm2 + 2 * betas
new_cov_denom = (
n_features * stats['post_mix_sum'] + n_features +
2 * (alphas + 1)
)
new_cov = new_cov_numer / new_cov_denom
elif self.covariance_type == 'tied':
centered = stats['centered'].reshape((
n_samples, self.n_components, self.n_mix, self.n_features, 1
))
centered_t = stats['centered'].reshape((
n_samples, self.n_components, self.n_mix, 1, self.n_features
))
centered_dots = centered * centered_t
psis_t = np.transpose(self.covars_prior, axes=(0, 2, 1))
nus = self.covars_weight
centr_means_resh = centered_means.reshape((
self.n_components, self.n_mix, self.n_features, 1
))
centr_means_resh_t = centered_means.reshape((
self.n_components, self.n_mix, 1, self.n_features
))
centered_means_dots = centr_means_resh * centr_means_resh_t
lambdas_cmdots_prod_sum = np.einsum(
'ij,ijkl->ikl',
lambdas, centered_means_dots
)
new_cov_numer = np.einsum(
'ijk,ijklm->jlm',
stats['post_comp_mix'], centered_dots
) + lambdas_cmdots_prod_sum + psis_t
new_cov_denom = (
stats['post_sum'] + self.n_mix + nus + self.n_features + 1
)[:, np.newaxis, np.newaxis]
new_cov = new_cov_numer / new_cov_denom
# Assigning new values to class members
self.weights_ = new_weights
self.means_ = new_means
self.covars_ = new_cov