-
Notifications
You must be signed in to change notification settings - Fork 0
/
HiddenMarkovModel.py
1254 lines (984 loc) · 46 KB
/
HiddenMarkovModel.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
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Natural Language Toolkit: Hidden Markov Model
#
# Copyright (C) 2001-2016 NLTK Project
# Author: Trevor Cohn <tacohn@csse.unimelb.edu.au>
# Philip Blunsom <pcbl@csse.unimelb.edu.au>
# Tiago Tresoldi <tiago@tresoldi.pro.br> (fixes)
# Steven Bird <stevenbird1@gmail.com> (fixes)
# Joseph Frazee <jfrazee@mail.utexas.edu> (fixes)
# Steven Xu <xxu@student.unimelb.edu.au> (fixes)
# URL: <http://nltk.org/>
# For license information, see LICENSE.TXT
#MOST of this is NLTK source but I made some modifications -JPC
import itertools
import pickle
try:
import numpy as np
except ImportError:
pass
from customizedProbability import (ConditionalProbDist, DictionaryProbDist, DictionaryConditionalProbDist,
LidstoneProbDist, MutableProbDist,
RandomProbDist, WittenBellProbDist, LaplaceProbDist,
SimpleGoodTuringProbDist, Lidstone_03, Lidstone_15,
Lidstone_003, Lidstone_07,ELEProbDist)
from collections import OrderedDict
from nltk.metrics import accuracy
from nltk.util import LazyMap, unique_list
from nltk.compat import python_2_unicode_compatible, izip, imap
from nltk.tag.api import TaggerI
import math
import globalModelParameters
_TEXT = 0 # index of text in a tuple
_TAG = 1 # index of tag in a tuple
def _identity( labeled_symbols ):
return labeled_symbols
@python_2_unicode_compatible
class HiddenMarkovModelTagger( TaggerI ):
"""
Hidden Markov model class, a generative model for labelling sequence data.
These models define the joint probability of a sequence of symbols and
their labels (state transitions) as the product of the starting state
probability, the probability of each state transition, and the probability
of each observation being generated from each state. This is described in
more detail in the module documentation.
This implementation is based on the HMM description in Chapter 8, Huang,
Acero and Hon, Spoken Language Processing and includes an extension for
training shallow HMM parsers or specialized HMMs as in Molina et.
al, 2002. A specialized HMM modifies training data by applying a
specialization function to create a new training set that is more
appropriate for sequential tagging with an HMM. A typical use case is
chunking.
:param symbols: the set of output symbols (alphabet)
:type symbols: seq of any
:param states: a set of states representing state space
:type states: seq of any
:param transitions: transition probabilities; Pr(s_i | s_j) is the
probability of transition from state i given the model is in
state_j
:type transitions: ConditionalProbDistI
:param outputs: output probabilities; Pr(o_k | s_i) is the probability
of emitting symbol k when entering state i
:type outputs: ConditionalProbDistI
:param priors: initial state distribution; Pr(s_i) is the probability
of starting in state i
:type priors: ProbDistI
:param transform: an optional function for transforming training
instances, defaults to the identity function.
:type transform: callable
"""
def __init__( self, symbols, states, transitions, outputs, priors,
transform = _identity ):
self._symbols = unique_list( symbols )
self._states = unique_list( states )
self._transitions = transitions
self._outputs = outputs
self._priors = priors
self._cache = None
self._transform = transform
@classmethod
def _train( cls, labeled_sequence, test_sequence = None,
unlabeled_sequence = None, transform = _identity,
estimator = None, **kwargs ):
if estimator is None:
def estimator( fd, bins ):
return LidstoneProbDist( fd, 0.1, bins )
labeled_sequence = LazyMap( transform, labeled_sequence )
symbols = unique_list( word for sent in labeled_sequence
for word, tag in sent )
tag_set = unique_list( tag for sent in labeled_sequence
for word, tag in sent )
trainer = HiddenMarkovModelTrainer( tag_set, symbols )
hmm = trainer.train_supervised( labeled_sequence, estimator = estimator )
hmm = cls( hmm._symbols, hmm._states, hmm._transitions, hmm._outputs,
hmm._priors, transform = transform )
if test_sequence:
hmm.test( test_sequence, verbose = kwargs.get( 'verbose', False ) )
if unlabeled_sequence:
max_iterations = kwargs.get( 'max_iterations', 5 )
hmm = trainer.train_unsupervised( unlabeled_sequence, model = hmm,
max_iterations = max_iterations )
if test_sequence:
hmm.test( test_sequence, verbose = kwargs.get( 'verbose', False ) )
return hmm
@classmethod
def train( cls, labeled_sequence, test_sequence = None,
unlabeled_sequence = None, **kwargs ):
"""
Train a new HiddenMarkovModelTagger using the given labeled and
unlabeled training instances. Testing will be performed if test
instances are provided.
:return: a hidden markov model tagger
:rtype: HiddenMarkovModelTagger
:param labeled_sequence: a sequence of labeled training instances,
i.e. a list of sentences represented as tuples
:type labeled_sequence: list(list)
:param test_sequence: a sequence of labeled test instances
:type test_sequence: list(list)
:param unlabeled_sequence: a sequence of unlabeled training instances,
i.e. a list of sentences represented as words
:type unlabeled_sequence: list(list)
:param transform: an optional function for transforming training
instances, defaults to the identity function, see ``transform()``
:type transform: function
:param estimator: an optional function or class that maps a
condition's frequency distribution to its probability
distribution, defaults to a Lidstone distribution with gamma = 0.1
:type estimator: class or function
:param verbose: boolean flag indicating whether training should be
verbose or include printed output
:type verbose: bool
:param max_iterations: number of Baum-Welch interations to perform
:type max_iterations: int
"""
return cls._train( labeled_sequence, test_sequence,
unlabeled_sequence, **kwargs )
def probability( self, sequence, transitions_weight = None, outputs_weight = 1 ):
"""
Returns the probability of the given symbol sequence. If the sequence
is labelled, then returns the joint probability of the symbol, state
sequence. Otherwise, uses the forward algorithm to find the
probability over all label sequences.
:return: the probability of the sequence
:rtype: float
:param sequence: the sequence of symbols which must contain the TEXT
property, and optionally the TAG property
:type sequence: Token
"""
return 2 ** (self.log_probability( self._transform( sequence ), transitions_weight, outputs_weight ))
def update_outputsModel( self ):
self._outputs = ConditionalProbDist( globalModelParameters.outputsModel, WittenBellProbDist,
size_output_alphabet )
def log_probability( self, sequence, transitions_weight = None, outputs_weight = 1 ):
"""
Returns the log-probability of the given symbol sequence. If the
sequence is labelled, then returns the joint log-probability of the
symbol, state sequence. Otherwise, uses the forward algorithm to find
the log-probability over all label sequences.
:return: the log-probability of the sequence
:rtype: float
:param sequence: the sequence of symbols which must contain the TEXT
property, and optionally the TAG property
:type sequence: Token
"""
if transitions_weight is None:
transitions_weight = 1
sequence = self._transform( sequence )
T = len( sequence )
EPSILON = ''
channelModel = 0
sourceModel = 0
if T > 0 and sequence[ 0 ][ _TAG ] is not None:
last_state = sequence[ 0 ][ _TAG ]
if last_state != EPSILON:
if transitions_weight:
sourceModel += transitions_weight * self._priors.logprob( last_state )
else:
if transitions_weight:
sourceModel += transitions_weight * math.log2( globalModelParameters.EpsilonTransition )
channelModel += outputs_weight * self._output_logprob( last_state, sequence[ 0 ][ _TEXT ] )
for t in range( 1, T ):
state = sequence[ t ][ _TAG ]
if last_state != EPSILON:
if state != EPSILON:
if transitions_weight:
sourceModel += transitions_weight * self._transitions[ last_state ].logprob( state )
else:
if transitions_weight:
sourceModel += transitions_weight * math.log2( globalModelParameters.EpsilonTransition )
else:
# check if last_state is epsilon; if so then transition with probability of Epsilon
if transitions_weight:
sourceModel += transitions_weight * math.log2( globalModelParameters.EpsilonTransition )
channelModel += outputs_weight * self._output_logprob( state, sequence[ t ][ _TEXT ] )
last_state = state
# FIXME changed exponentiation
return { 'HMMtotal': (sourceModel + channelModel),
'HMMchannel': channelModel,
'HMMsource': sourceModel,
'sequence': sequence}
def tag( self, unlabeled_sequence ):
"""
Tags the sequence with the highest probability state sequence. This
uses the best_path method to find the Viterbi path.
:return: a labelled sequence of symbols
:rtype: list
:param unlabeled_sequence: the sequence of unlabeled symbols
:type unlabeled_sequence: list
"""
unlabeled_sequence = self._transform( unlabeled_sequence )
return self._tag( unlabeled_sequence )
def _tag( self, unlabeled_sequence ):
path = self._best_path( unlabeled_sequence )
return list( izip( unlabeled_sequence, path ) )
def _output_logprob( self, state, symbol ):
"""
:return: the log probability of the symbol being observed in the given
state
:rtype: float
"""
return self._outputs[ state ].logprob( symbol )
def _create_cache( self ):
"""
The cache is a tuple (P, O, X, S) where:
- S maps symbols to integers. I.e., it is the inverse
mapping from self._symbols; for each symbol s in
self._symbols, the following is true::
self._symbols[S[s]] == s
- O is the log output probabilities::
O[i,k] = log( P(token[t]=sym[k]|tag[t]=state[i]) )
- X is the log transition probabilities::
X[i,j] = log( P(tag[t]=state[j]|tag[t-1]=state[i]) )
- P is the log prior probabilities::
P[i] = log( P(tag[0]=state[i]) )
"""
if not self._cache:
N = len( self._states )
M = len( self._symbols )
P = np.zeros( N, np.float32 )
X = np.zeros( (N, N), np.float32 )
O = np.zeros( (N, M), np.float32 )
for i in range( N ):
si = self._states[ i ]
P[ i ] = self._priors.logprob( si )
for j in range( N ):
X[ i, j ] = self._transitions[ si ].logprob( self._states[ j ] )
for k in range( M ):
O[ i, k ] = self._output_logprob( si, self._symbols[ k ] )
S = { }
for k in range( M ):
S[ self._symbols[ k ] ] = k
self._cache = (P, O, X, S)
def _update_cache( self, symbols ):
# add new symbols to the symbol table and repopulate the output
# probabilities and symbol table mapping
if symbols:
self._create_cache( )
P, O, X, S = self._cache
for symbol in symbols:
if symbol not in self._symbols:
self._cache = None
self._symbols.append( symbol )
# don't bother with the work if there aren't any new symbols
if not self._cache:
N = len( self._states )
M = len( self._symbols )
Q = O.shape[ 1 ]
# add new columns to the output probability table without
# destroying the old probabilities
O = np.hstack( [ O, np.zeros( (N, M - Q), np.float32 ) ] )
for i in range( N ):
si = self._states[ i ]
# only calculate probabilities for new symbols
for k in range( Q, M ):
O[ i, k ] = self._output_logprob( si, self._symbols[ k ] )
# only create symbol mappings for new symbols
for k in range( Q, M ):
S[ self._symbols[ k ] ] = k
self._cache = (P, O, X, S)
def reset_cache( self ):
self._cache = None
def best_path( self, unlabeled_sequence ):
"""
Returns the state sequence of the optimal (most probable) path through
the HMM. Uses the Viterbi algorithm to calculate this part by dynamic
programming.
:return: the state sequence
:rtype: sequence of any
:param unlabeled_sequence: the sequence of unlabeled symbols
:type unlabeled_sequence: list
"""
unlabeled_sequence = self._transform( unlabeled_sequence )
return self._best_path( unlabeled_sequence )
def _best_path( self, unlabeled_sequence ):
T = len( unlabeled_sequence )
N = len( self._states )
self._create_cache( )
self._update_cache( unlabeled_sequence )
P, O, X, S = self._cache
V = np.zeros( (T, N), np.float32 )
B = -np.ones( (T, N), np.int )
V[ 0 ] = P + O[ :, S[ unlabeled_sequence[ 0 ] ] ]
for t in range( 1, T ):
for j in range( N ):
vs = V[ t - 1, : ] + X[ :, j ]
best = np.argmax( vs )
V[ t, j ] = vs[ best ] + O[ j, S[ unlabeled_sequence[ t ] ] ]
B[ t, j ] = best
current = np.argmax( V[ T - 1, : ] )
sequence = [ current ]
for t in range( T - 1, 0, -1 ):
last = B[ t, current ]
sequence.append( last )
current = last
sequence.reverse( )
return list( map( self._states.__getitem__, sequence ) )
def best_path_simple( self, unlabeled_sequence ):
"""
Returns the state sequence of the optimal (most probable) path through
the HMM. Uses the Viterbi algorithm to calculate this part by dynamic
programming. This uses a simple, direct method, and is included for
teaching purposes.
:return: the state sequence
:rtype: sequence of any
:param unlabeled_sequence: the sequence of unlabeled symbols
:type unlabeled_sequence: list
"""
unlabeled_sequence = self._transform( unlabeled_sequence )
return self._best_path_simple( unlabeled_sequence )
def _best_path_simple( self, unlabeled_sequence ):
T = len( unlabeled_sequence )
N = len( self._states )
V = np.zeros( (T, N), np.float64 )
B = { }
# find the starting log probabilities for each state
symbol = unlabeled_sequence[ 0 ]
for i, state in enumerate( self._states ):
V[ 0, i ] = self._priors.logprob( state ) + \
self._output_logprob( state, symbol )
B[ 0, state ] = None
# find the maximum log probabilities for reaching each state at time t
for t in range( 1, T ):
symbol = unlabeled_sequence[ t ]
for j in range( N ):
sj = self._states[ j ]
best = None
for i in range( N ):
si = self._states[ i ]
va = V[ t - 1, i ] + self._transitions[ si ].logprob( sj )
if not best or va > best[ 0 ]:
best = (va, si)
V[ t, j ] = best[ 0 ] + self._output_logprob( sj, symbol )
B[ t, sj ] = best[ 1 ]
# find the highest probability final state
best = None
for i in range( N ):
val = V[ T - 1, i ]
if not best or val > best[ 0 ]:
best = (val, self._states[ i ])
# traverse the back-pointers B to find the state sequence
current = best[ 1 ]
sequence = [ current ]
for t in range( T - 1, 0, -1 ):
last = B[ t, current ]
sequence.append( last )
current = last
sequence.reverse( )
return sequence
def random_sample( self, rng, length ):
"""
Randomly sample the HMM to generate a sentence of a given length. This
samples the prior distribution then the observation distribution and
transition distribution for each subsequent observation and state.
This will mostly generate unintelligible garbage, but can provide some
amusement.
:return: the randomly created state/observation sequence,
generated according to the HMM's probability
distributions. The SUBTOKENS have TEXT and TAG
properties containing the observation and state
respectively.
:rtype: list
:param rng: random number generator
:type rng: Random (or any object with a random() method)
:param length: desired output length
:type length: int
"""
# sample the starting state and symbol prob dists
tokens = [ ]
state = self._sample_probdist( self._priors, rng.random( ), self._states )
symbol = self._sample_probdist( self._outputs[ state ],
rng.random( ), self._symbols )
tokens.append( (symbol, state) )
for i in range( 1, length ):
# sample the state transition and symbol prob dists
state = self._sample_probdist( self._transitions[ state ],
rng.random( ), self._states )
symbol = self._sample_probdist( self._outputs[ state ],
rng.random( ), self._symbols )
tokens.append( (symbol, state) )
return tokens
def _sample_probdist( self, probdist, p, samples ):
cum_p = 0
for sample in samples:
add_p = probdist.prob( sample )
if cum_p <= p <= cum_p + add_p:
return sample
cum_p += add_p
raise Exception( 'Invalid probability distribution - '
'does not sum to one' )
def entropy( self, unlabeled_sequence ):
"""
Returns the entropy over labellings of the given sequence. This is
given by::
H(O) = - sum_S Pr(S | O) log Pr(S | O)
where the summation ranges over all state sequences, S. Let
*Z = Pr(O) = sum_S Pr(S, O)}* where the summation ranges over all state
sequences and O is the observation sequence. As such the entropy can
be re-expressed as::
H = - sum_S Pr(S | O) log [ Pr(S, O) / Z ]
= log Z - sum_S Pr(S | O) log Pr(S, 0)
= log Z - sum_S Pr(S | O) [ log Pr(S_0) + sum_t Pr(S_t | S_{t-1}) + sum_t Pr(O_t | S_t) ]
The order of summation for the log terms can be flipped, allowing
dynamic programming to be used to calculate the entropy. Specifically,
we use the forward and backward probabilities (alpha, beta) giving::
H = log Z - sum_s0 alpha_0(s0) beta_0(s0) / Z * log Pr(s0)
+ sum_t,si,sj alpha_t(si) Pr(sj | si) Pr(O_t+1 | sj) beta_t(sj) / Z * log Pr(sj | si)
+ sum_t,st alpha_t(st) beta_t(st) / Z * log Pr(O_t | st)
This simply uses alpha and beta to find the probabilities of partial
sequences, constrained to include the given state(s) at some point in
time.
"""
unlabeled_sequence = self._transform( unlabeled_sequence )
T = len( unlabeled_sequence )
N = len( self._states )
alpha = self._forward_probability( unlabeled_sequence )
beta = self._backward_probability( unlabeled_sequence )
normalisation = logsumexp2( alpha[ T - 1 ] )
entropy = normalisation
# starting state, t = 0
for i, state in enumerate( self._states ):
p = 2 ** (alpha[ 0, i ] + beta[ 0, i ] - normalisation)
entropy -= p * self._priors.logprob( state )
# print 'p(s_0 = %s) =' % state, p
# state transitions
for t0 in range( T - 1 ):
t1 = t0 + 1
for i0, s0 in enumerate( self._states ):
for i1, s1 in enumerate( self._states ):
p = 2 ** (alpha[ t0, i0 ] + self._transitions[ s0 ].logprob( s1 ) +
self._outputs[ s1 ].logprob(
unlabeled_sequence[ t1 ][ _TEXT ] ) +
beta[ t1, i1 ] - normalisation)
entropy -= p * self._transitions[ s0 ].logprob( s1 )
# print 'p(s_%d = %s, s_%d = %s) =' % (t0, s0, t1, s1), p
# symbol emissions
for t in range( T ):
for i, state in enumerate( self._states ):
p = 2 ** (alpha[ t, i ] + beta[ t, i ] - normalisation)
entropy -= p * self._outputs[ state ].logprob(
unlabeled_sequence[ t ][ _TEXT ] )
# print 'p(s_%d = %s) =' % (t, state), p
return entropy
def point_entropy( self, unlabeled_sequence ):
"""
Returns the pointwise entropy over the possible states at each
position in the chain, given the observation sequence.
"""
unlabeled_sequence = self._transform( unlabeled_sequence )
T = len( unlabeled_sequence )
N = len( self._states )
alpha = self._forward_probability( unlabeled_sequence )
beta = self._backward_probability( unlabeled_sequence )
normalisation = logsumexp2( alpha[ T - 1 ] )
entropies = np.zeros( T, np.float64 )
probs = np.zeros( N, np.float64 )
for t in range( T ):
for s in range( N ):
probs[ s ] = alpha[ t, s ] + beta[ t, s ] - normalisation
for s in range( N ):
entropies[ t ] -= 2 ** (probs[ s ]) * probs[ s ]
return entropies
def _exhaustive_entropy( self, unlabeled_sequence ):
unlabeled_sequence = self._transform( unlabeled_sequence )
T = len( unlabeled_sequence )
N = len( self._states )
labellings = [ [ state ] for state in self._states ]
for t in range( T - 1 ):
current = labellings
labellings = [ ]
for labelling in current:
for state in self._states:
labellings.append( labelling + [ state ] )
log_probs = [ ]
for labelling in labellings:
labeled_sequence = unlabeled_sequence[ : ]
for t, label in enumerate( labelling ):
labeled_sequence[ t ] = (labeled_sequence[ t ][ _TEXT ], label)
lp = self.log_probability( labeled_sequence )
log_probs.append( lp )
normalisation = _log_add( *log_probs )
# ps = zeros((T, N), float64)
# for labelling, lp in zip(labellings, log_probs):
# for t in range(T):
# ps[t, self._states.index(labelling[t])] += \
# 2**(lp - normalisation)
#for t in range(T):
#print 'prob[%d] =' % t, ps[t]
entropy = 0
for lp in log_probs:
lp -= normalisation
entropy -= 2 ** (lp) * lp
return entropy
def _exhaustive_point_entropy( self, unlabeled_sequence ):
unlabeled_sequence = self._transform( unlabeled_sequence )
T = len( unlabeled_sequence )
N = len( self._states )
labellings = [ [ state ] for state in self._states ]
for t in range( T - 1 ):
current = labellings
labellings = [ ]
for labelling in current:
for state in self._states:
labellings.append( labelling + [ state ] )
log_probs = [ ]
for labelling in labellings:
labelled_sequence = unlabeled_sequence[ : ]
for t, label in enumerate( labelling ):
labelled_sequence[ t ] = (labelled_sequence[ t ][ _TEXT ], label)
lp = self.log_probability( labelled_sequence )
log_probs.append( lp )
normalisation = _log_add( *log_probs )
probabilities = _ninf_array( (T, N) )
for labelling, lp in zip( labellings, log_probs ):
lp -= normalisation
for t, label in enumerate( labelling ):
index = self._states.index( label )
probabilities[ t, index ] = _log_add( probabilities[ t, index ], lp )
entropies = np.zeros( T, np.float64 )
for t in range( T ):
for s in range( N ):
entropies[ t ] -= 2 ** (probabilities[ t, s ]) * probabilities[ t, s ]
return entropies
def _transitions_matrix( self ):
""" Return a matrix of transition log probabilities. """
trans_iter = (self._transitions[ sj ].logprob( si )
for sj in self._states
for si in self._states)
transitions_logprob = np.fromiter( trans_iter, dtype = np.float64 )
N = len( self._states )
return transitions_logprob.reshape( (N, N) ).T
def _outputs_vector( self, symbol ):
"""
Return a vector with log probabilities of emitting a symbol
when entering states.
"""
out_iter = (self._output_logprob( sj, symbol ) for sj in self._states)
return np.fromiter( out_iter, dtype = np.float64 )
def _forward_probability( self, unlabeled_sequence ):
"""
Return the forward probability matrix, a T by N array of
log-probabilities, where T is the length of the sequence and N is the
number of states. Each entry (t, s) gives the probability of being in
state s at time t after observing the partial symbol sequence up to
and including t.
:param unlabeled_sequence: the sequence of unlabeled symbols
:type unlabeled_sequence: list
:return: the forward log probability matrix
:rtype: array
"""
T = len( unlabeled_sequence )
N = len( self._states )
alpha = _ninf_array( (T, N) )
transitions_logprob = self._transitions_matrix( )
# Initialization
symbol = unlabeled_sequence[ 0 ][ _TEXT ]
for i, state in enumerate( self._states ):
alpha[ 0, i ] = self._priors.logprob( state ) + \
self._output_logprob( state, symbol )
# Induction
for t in range( 1, T ):
symbol = unlabeled_sequence[ t ][ _TEXT ]
output_logprob = self._outputs_vector( symbol )
for i in range( N ):
summand = alpha[ t - 1 ] + transitions_logprob[ i ]
alpha[ t, i ] = logsumexp2( summand ) + output_logprob[ i ]
return alpha
def _backward_probability( self, unlabeled_sequence ):
"""
Return the backward probability matrix, a T by N array of
log-probabilities, where T is the length of the sequence and N is the
number of states. Each entry (t, s) gives the probability of being in
state s at time t after observing the partial symbol sequence from t
.. T.
:return: the backward log probability matrix
:rtype: array
:param unlabeled_sequence: the sequence of unlabeled symbols
:type unlabeled_sequence: list
"""
T = len( unlabeled_sequence )
N = len( self._states )
beta = _ninf_array( (T, N) )
transitions_logprob = self._transitions_matrix( ).T
# initialise the backward values;
# "1" is an arbitrarily chosen value from Rabiner tutorial
beta[ T - 1, : ] = np.log2( 1 )
# inductively calculate remaining backward values
for t in range( T - 2, -1, -1 ):
symbol = unlabeled_sequence[ t + 1 ][ _TEXT ]
outputs = self._outputs_vector( symbol )
for i in range( N ):
summand = transitions_logprob[ i ] + beta[ t + 1 ] + outputs
beta[ t, i ] = logsumexp2( summand )
return beta
def test( self, test_sequence, verbose = False, **kwargs ):
"""
Tests the HiddenMarkovModelTagger instance.
:param test_sequence: a sequence of labeled test instances
:type test_sequence: list(list)
:param verbose: boolean flag indicating whether training should be
verbose or include printed output
:type verbose: bool
"""
def words( sent ):
return [ word for (word, tag) in sent ]
def tags( sent ):
return [ tag for (word, tag) in sent ]
def flatten( seq ):
return list( itertools.chain( *seq ) )
test_sequence = self._transform( test_sequence )
predicted_sequence = list( imap( self._tag, imap( words, test_sequence ) ) )
if verbose:
for test_sent, predicted_sent in izip( test_sequence, predicted_sequence ):
print( 'Test:',
' '.join( '%s/%s' % (token, tag)
for (token, tag) in test_sent ) )
print( )
print( 'Untagged:',
' '.join( "%s" % token for (token, tag) in test_sent ) )
print( )
print( 'HMM-tagged:',
' '.join( '%s/%s' % (token, tag)
for (token, tag) in predicted_sent ) )
print( )
print( 'Entropy:',
self.entropy( [ (token, None) for
(token, tag) in predicted_sent ] ) )
print( )
print( '-' * 60 )
test_tags = flatten( imap( tags, test_sequence ) )
predicted_tags = flatten( imap( tags, predicted_sequence ) )
acc = accuracy( test_tags, predicted_tags )
count = sum( len( sent ) for sent in test_sequence )
print( 'accuracy over %d tokens: %.2f' % (count, acc * 100) )
def __repr__( self ):
return ('<HiddenMarkovModelTagger %d states and %d output symbols>'
% (len( self._states ), len( self._symbols )))
class HiddenMarkovModelTrainer( object ):
"""
Algorithms for learning HMM parameters from training data. These include
both supervised learning (MLE) and unsupervised learning (Baum-Welch).
Creates an HMM trainer to induce an HMM with the given states and
output symbol alphabet. A supervised and unsupervised training
method may be used. If either of the states or symbols are not given,
these may be derived from supervised training.
:param states: the set of state labels
:type states: sequence of any
:param symbols: the set of observation symbols
:type symbols: sequence of any
"""
def __init__( self, states = None, symbols = None ):
self._states = (states if states else [ ])
self._symbols = (symbols if symbols else [ ])
def train( self, labeled_sequences = None, unlabeled_sequences = None,
**kwargs ):
"""
Trains the HMM using both (or either of) supervised and unsupervised
techniques.
:return: the trained model
:rtype: HiddenMarkovModelTagger
:param labelled_sequences: the supervised training data, a set of
labelled sequences of observations
:type labelled_sequences: list
:param unlabeled_sequences: the unsupervised training data, a set of
sequences of observations
:type unlabeled_sequences: list
:param kwargs: additional arguments to pass to the training methods
"""
assert labeled_sequences or unlabeled_sequences
model = None
if labeled_sequences:
model = self.train_supervised( labeled_sequences, **kwargs )
if unlabeled_sequences:
if model:
kwargs[ 'model' ] = model
model = self.train_unsupervised( unlabeled_sequences, **kwargs )
return model
def _baum_welch_step( self, sequence, model, symbol_to_number ):
N = len( model._states )
M = len( model._symbols )
T = len( sequence )
# compute forward and backward probabilities
alpha = model._forward_probability( sequence )
beta = model._backward_probability( sequence )
# find the log probability of the sequence
lpk = logsumexp2( alpha[ T - 1 ] )
A_numer = _ninf_array( (N, N) )
B_numer = _ninf_array( (N, M) )
A_denom = _ninf_array( N )
B_denom = _ninf_array( N )
transitions_logprob = model._transitions_matrix( ).T
for t in range( T ):
symbol = sequence[ t ][ _TEXT ] # not found? FIXME
next_symbol = None
if t < T - 1:
next_symbol = sequence[ t + 1 ][ _TEXT ] # not found? FIXME
xi = symbol_to_number[ symbol ]
next_outputs_logprob = model._outputs_vector( next_symbol )
alpha_plus_beta = alpha[ t ] + beta[ t ]
if t < T - 1:
numer_add = transitions_logprob + next_outputs_logprob + \
beta[ t + 1 ] + alpha[ t ].reshape( N, 1 )
A_numer = np.logaddexp2( A_numer, numer_add )
A_denom = np.logaddexp2( A_denom, alpha_plus_beta )
else:
B_denom = np.logaddexp2( A_denom, alpha_plus_beta )
B_numer[ :, xi ] = np.logaddexp2( B_numer[ :, xi ], alpha_plus_beta )
return lpk, A_numer, A_denom, B_numer, B_denom
def train_unsupervised( self, unlabeled_sequences, update_outputs = True,
**kwargs ):
"""
Trains the HMM using the Baum-Welch algorithm to maximise the
probability of the data sequence. This is a variant of the EM
algorithm, and is unsupervised in that it doesn't need the state
sequences for the symbols. The code is based on 'A Tutorial on Hidden
Markov Models and Selected Applications in Speech Recognition',
Lawrence Rabiner, IEEE, 1989.
:return: the trained model
:rtype: HiddenMarkovModelTagger
:param unlabeled_sequences: the training data, a set of
sequences of observations
:type unlabeled_sequences: list
kwargs may include following parameters:
:param model: a HiddenMarkovModelTagger instance used to begin
the Baum-Welch algorithm
:param max_iterations: the maximum number of EM iterations
:param convergence_logprob: the maximum change in log probability to
allow convergence
"""
# create a uniform HMM, which will be iteratively refined, unless
# given an existing model
model = kwargs.get( 'model' )
if not model:
priors = RandomProbDist( self._states )
transitions = DictionaryConditionalProbDist(
dict( (state, RandomProbDist( self._states ))
for state in self._states ) )
outputs = DictionaryConditionalProbDist(
dict( (state, RandomProbDist( self._symbols ))
for state in self._states ) )
model = HiddenMarkovModelTagger( self._symbols, self._states,
transitions, outputs, priors )
self._states = model._states
self._symbols = model._symbols
N = len( self._states )
M = len( self._symbols )
symbol_numbers = dict( (sym, i) for i, sym in enumerate( self._symbols ) )
# update model prob dists so that they can be modified
model._priors = MutableProbDist( model._priors, self._states )
model._transitions = DictionaryConditionalProbDist(
dict( (s, MutableProbDist( model._transitions[ s ], self._states ))
for s in self._states ) )
if update_outputs:
model._outputs = DictionaryConditionalProbDist(
dict( (s, MutableProbDist( model._outputs[ s ], self._symbols ))
for s in self._states ) )
model.reset_cache( )
# iterate until convergence
converged = False
last_logprob = None
iteration = 0
max_iterations = kwargs.get( 'max_iterations', 1000 )
epsilon = kwargs.get( 'convergence_logprob', 1e-6 )
while not converged and iteration < max_iterations:
A_numer = _ninf_array( (N, N) )
B_numer = _ninf_array( (N, M) )
A_denom = _ninf_array( N )
B_denom = _ninf_array( N )
logprob = 0
for sequence in unlabeled_sequences:
sequence = list( sequence )
if not sequence:
continue
(lpk, seq_A_numer, seq_A_denom,
seq_B_numer, seq_B_denom) = self._baum_welch_step( sequence, model, symbol_numbers )