-
Notifications
You must be signed in to change notification settings - Fork 0
/
states.py
356 lines (284 loc) · 16.3 KB
/
states.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
from __future__ import division
import numpy as np
from numpy import newaxis as na
from pyhsmm.util.general import rcumsum
from pyhsmm.util.profiling import line_profiled
from pyhsmm.internals.hsmm_states import HSMMStatesPython, HSMMStatesPossibleChangepoints, \
hsmm_messages_forwards_log, hsmm_messages_backwards_log
PROFILING=False
TRUNC = 3
class HSMMSubHMMStates(HSMMStatesPython):
# NOTE: can't extend the eigen version because its sample_forwards depends
# on aBl being iid (doesnt call the sub-methods)
def __init__(self,model,substateseqs=None,**kwargs):
self.model = model
if substateseqs is not None:
raise NotImplementedError
super(HSMMSubHMMStates,self).__init__(model,**kwargs)
self.data = self.data.astype('float32',copy=False) if self.data is not None else None
def generate_states(self):
self._generate_superstates()
self._generate_substates()
def _generate_superstates(self):
super(HSMMSubHMMStates,self).generate_states()
def _generate_substates(self):
self.substates_list = []
for state, dur in zip(self.stateseq_norep,self.durations_censored):
self.model.HMMs[state].generate(dur)
self.substates_list.append(self.model.HMMs[state].states_list[-1])
def generate_obs(self):
obs = []
for subseq in self.substates_list:
obs.append(subseq.data)
obs = np.concatenate(obs)
assert len(obs) == self.T
return obs
@property
def aBls(self):
if self._aBls is None:
self._aBls = [hmm.get_aBl(self.data) for hmm in self.model.HMMs]
return self._aBls
@property
def mf_aBls(self):
if self._mf_aBls is None:
self._mf_aBls = [hmm.get_mf_aBl(self.data) for hmm in self.model.HMMs]
return self._mf_aBls
def clear_caches(self):
for hmm in self.model.HMMs:
hmm._clear_message_caches() # NOTE: this is REALLY important!
self._aBls = self._mf_aBls = None
super(HSMMSubHMMStates,self).clear_caches()
def resample(self,temp=None):
self._remove_substates_from_subHMMs()
super(HSMMSubHMMStates,self).resample() # resamples superstates
self._resample_substates()
def _resample_substates(self):
assert not hasattr(self,'substates_list') or len(self.substates_list) == 0
self.substates_list = []
indices = np.concatenate(((0,),np.cumsum(self.durations_censored[:-1])))
for state, startidx, dur in zip(self.stateseq_norep,indices,self.durations_censored):
self.model.HMMs[state].add_data(
self.data[startidx:startidx+dur],initialize_from_prior=False)
self.substates_list.append(self.model.HMMs[state].states_list[-1])
def _remove_substates_from_subHMMs(self):
if hasattr(self,'substates_list') and len(self.substates_list) > 0:
for superstate, states_obj in zip(self.stateseq_norep, self.substates_list):
self.model.HMMs[superstate].states_list.remove(states_obj)
self.substates_list = []
def set_stateseq(self,superstateseq,substateseqs):
self.stateseq = superstateseq
indices = np.concatenate(((0,),np.cumsum(self.durations_censored[:-1])))
for state, startidx, dur, substateseq in zip(self.stateseq_norep,indices,
self.durations_censored,substateseqs):
self.model.HMMs[state].add_data(
self.data[startidx:startidx+dur],stateseq=substateseq)
self.substates_list.append(self.model.HMMs[state].states_list[-1])
### NEW
def cumulative_obs_potentials(self,t):
return np.hstack([hmm.cumulative_obs_potentials(self.aBls[state][t:])[:,na]
for state, hmm in enumerate(self.model.HMMs)]), np.zeros(self.num_states)
# return np.hstack([np.logaddexp.reduce(
# hmm.messages_forwards(self.aBls[state][t:]),axis=1)[:,na]
# for state, hmm in enumerate(self.model.HMMs)])
def reverse_cumulative_obs_potentials(self,t):
return np.hstack([hmm.reverse_cumulative_obs_potentials(self.aBls[state][:t+1])[:,na]
for state, hmm in enumerate(self.model.HMMs)]), np.zeros(self.num_states)
# return np.hstack([np.logaddexp.reduce(
# np.log(hmm.init_state_distn.pi_0) + # could cache this
# hmm.messages_backwards(self.aBls[state][:t+1]),axis=1)[:,na]
# for state, hmm in enumerate(self.model.HMMs)])
def mf_cumulative_obs_potentials(self,t):
return np.hstck([hmm.mf_cumulative_obs_potentials(self.mf_aBls[state][t:])[:,na]
for state, hmm in enumerate(self.model.HMMs)])
def mf_reverse_cumulative_obs_potentials(self,t):
return np.hstack([hmm.mf_reverse_cumulative_obs_potentials(self.mf_aBls[state][:t+1])[:,na]
for state, hmm in enumerate(self.model.HMMs)])
### OLD
def cumulative_likelihood_state(self,start,stop,state):
return np.logaddexp.reduce(self.model.HMMs[state].messages_forwards(self.aBls[state][start:stop]),axis=1)
def cumulative_likelihoods(self,start,stop):
return np.hstack([self.cumulative_likelihood_state(start,stop,state)[:,na]
for state in range(self.num_states)])
def likelihood_block_state(self,start,stop,state):
return self.model.HMMs[state].cumulative_obs_potentials(self.aBls[state][start:stop])[-1]
# return np.logaddexp.reduce(self.model.HMMs[state].messages_forwards(self.aBls[state][start:stop])[-1])
def likelihood_block(self,start,stop):
return np.array([self.likelihood_block_state(start,stop,state)
for state in range(self.num_states)])
class HSMMSubHMMStatesPossibleChangepoints(HSMMSubHMMStates,HSMMStatesPossibleChangepoints):
# need this method as long as we don't have a general sample forwards (which
# we probably don't want until we get the backwards normalization right...)
def clear_caches(self):
super(HSMMSubHMMStatesPossibleChangepoints,self).clear_caches()
for hmm in self.model.HMMs:
hmm._clear_message_caches()
def sample_forwards(self,betal,betastarl):
return HSMMStatesPossibleChangepoints.sample_forwards(self,betal,betastarl)
def generate(self):
raise NotImplementedError
def cumulative_obs_potentials(self,tblock):
t = self.segmentstarts[tblock]
possible_durations = self.segmentlens[tblock:].cumsum()[:TRUNC]
return np.hstack([hmm.cumulative_obs_potentials(self.aBls[state][t:],self,t)\
[possible_durations -1][:,na]
for state, hmm in enumerate(self.model.HMMs)]), np.zeros(self.num_states)
def reverse_cumulative_obs_potentials(self,tblock):
t = self.segmentstarts[tblock] + self.segmentlens[tblock]
possible_durations = rcumsum(self.segmentlens[:tblock+1])[-TRUNC if TRUNC is not None else None:]
return np.hstack([hmm.reverse_cumulative_obs_potentials(self.aBls[state][:t],self,t)\
[-possible_durations][:,na]
# [possible_durations -1][:,na]
for state, hmm in enumerate(self.model.HMMs)])
def mf_cumulative_obs_potentials(self,tblock):
t = self.segmentstarts[tblock]
possible_durations = self.segmentlens[tblock:].cumsum()[:TRUNC]
return np.hstack([hmm.mf_cumulative_obs_potentials(self.mf_aBls[state][t:],self,t)\
[possible_durations -1][:,na]
for state, hmm in enumerate(self.model.HMMs)]), np.zeros(self.num_states)
def mf_reverse_cumulative_obs_potentials(self,tblock):
t = self.segmentstarts[tblock] + self.segmentlens[tblock]
possible_durations = rcumsum(self.segmentlens[:tblock+1])[-TRUNC if TRUNC is not None else None:]
return np.hstack([hmm.mf_reverse_cumulative_obs_potentials(self.mf_aBls[state][:t],self,t)\
[-possible_durations][:,na]
for state, hmm in enumerate(self.model.HMMs)])
# TODO TODO the following are only in here for the hard-coded truncation
def dur_potentials(self,tblock):
possible_durations = self.segmentlens[tblock:].cumsum()[:TRUNC]
return self.aDl[possible_durations -1]
def reverse_dur_potentials(self,tblock):
possible_durations = rcumsum(self.segmentlens[:tblock+1])[-TRUNC if TRUNC is not None else None:]
return self.aDl[possible_durations -1]
def dur_survival_potentials(self,tblock):
max_dur = self.segmentlens[tblock:].cumsum()[:TRUNC][-1]
return self.aDsl[max_dur -1]
def reverse_dur_survival_potentials(self,tblock):
max_dur = rcumsum(self.segmentlens[:tblock+1])[-TRUNC if TRUNC is not None else None:][0]
return self.aDsl[max_dur -1]
def mf_dur_potentials(self,tblock):
possible_durations = self.segmentlens[tblock:].cumsum()[:TRUNC]
return self.mf_aDl[possible_durations -1]
def mf_reverse_dur_potentials(self,tblock):
possible_durations = rcumsum(self.segmentlens[:tblock+1])[-TRUNC if TRUNC is not None else None:]
return self.mf_aDl[possible_durations -1]
def mf_dur_survival_potentials(self,tblock):
max_dur = self.segmentlens[tblock:].cumsum()[:TRUNC][-1]
return self.mf_aDsl[max_dur -1]
def mf_reverse_dur_survival_potentials(self,tblock):
max_dur = rcumsum(self.segmentlens[:tblock+1])[-TRUNC if TRUNC is not None else None:][0]
return self.mf_aDsl[max_dur -1]
### lots of code copying here, unfortunately TODO
@property
def all_expected_stats(self):
return self.expected_states, self.expected_transcounts, self.expected_durations, \
self._normalizer, self.subhmm_stats
@all_expected_stats.setter
def all_expected_stats(self,vals):
self.expected_states, self.expected_transcounts, self.expected_durations, \
self._normalizer, self.subhmm_stats = vals
def E_step(self):
# NOTE: this method differs from parent because it passes in self.aBls
self.clear_caches()
for hmm in self.model.HMMs:
assert len(hmm._cache) == 0 and len(hmm._reverse_cache) == 0
self.all_expected_stats = self._expected_statistics(
self.trans_potentials, np.log(self.pi_0),
self.cumulative_obs_potentials, self.reverse_cumulative_obs_potentials,
self.dur_potentials, self.reverse_dur_potentials,
self.dur_survival_potentials, self.reverse_dur_survival_potentials,
self.aBls) # here's the difference
self.stateseq = self.expected_states.argmax(1) # for plotting
@line_profiled
def meanfieldupdate(self):
# NOTE: this method differs from parent because it passes in self.aBls
self.clear_caches()
for hmm in self.model.HMMs:
assert len(hmm._cache) == 0 and len(hmm._reverse_cache) == 0
self.all_expected_stats = self._expected_statistics(
self.mf_trans_potentials, np.log(self.mf_pi_0),
self.mf_cumulative_obs_potentials, self.mf_reverse_cumulative_obs_potentials,
self.mf_dur_potentials, self.mf_reverse_dur_potentials,
self.mf_dur_survival_potentials, self.mf_reverse_dur_survival_potentials,
self.mf_aBls)
self.stateseq = self.expected_states.argmax(1) # for plotting
@line_profiled
def _expected_statistics(self,
trans_potentials, initial_state_potential,
cumulative_obs_potentials, reverse_cumulative_obs_potentials,
dur_potentials, reverse_dur_potentials,
dur_survival_potentials, reverse_dur_survival_potentials,
aBls):
# NOTE: this method differs from parent because it gets self.aBls
alphal, alphastarl, _ = hsmm_messages_forwards_log(
trans_potentials,
initial_state_potential,
reverse_cumulative_obs_potentials,
reverse_dur_potentials,
reverse_dur_survival_potentials,
np.empty((self.T,self.num_states)),np.empty((self.T,self.num_states)))
betal, betastarl, normalizer = hsmm_messages_backwards_log(
trans_potentials,
initial_state_potential,
cumulative_obs_potentials,
dur_potentials,
dur_survival_potentials,
np.empty((self.T,self.num_states)),np.empty((self.T,self.num_states)),
right_censoring=False)
expected_states = self._expected_states(
alphal, betal, alphastarl, betastarl, normalizer)
expected_transitions = self._expected_transitions(
alphal, betastarl, trans_potentials, normalizer) # TODO assumes homog trans
expected_durations = self._expected_durations(
dur_potentials,cumulative_obs_potentials,
alphastarl, betal, normalizer)
### here's the different bit!
# also compute subhmm expected stats, using aBls
subhmm_expected_states = [np.zeros((self.Tfull,hmm.num_states)) for hmm in self.model.HMMs]
subhmm_expected_trans = [np.zeros((hmm.num_states,hmm.num_states)) for hmm in self.model.HMMs]
for tblock in xrange(self.Tblock):
for tblockend, obs, dur in zip(
xrange(tblock,min(self.Tblock,tblock+TRUNC)),
cumulative_obs_potentials(tblock)[0],dur_potentials(tblock)):
tstart = self.segmentstarts[tblock]
tend = self.segmentstarts[tblockend] + self.segmentlens[tblockend]
weights = np.exp(alphastarl[tblock] + betal[tblockend] + obs + dur - normalizer)
if tblockend == self.Tblock-1:
weights += np.exp(
alphastarl[tblock] + betal[tblockend] + obs +
dur_survival_potentials(tblockend) - normalizer)
for state, (hmm, weight) in enumerate(zip(self.model.HMMs,weights)):
if weight > 0:
states, trans, _ = hmm.mf_expected_statistics( # NOTE: calls mf version!
aBls[state][tstart:tend],self,tstart,tend) # here's where aBls are used
subhmm_expected_states[state][tstart:tend] += weight*states
subhmm_expected_trans[state] += weight*trans
subhmm_stats = [[states, trans, self.data]
for states, trans in zip(subhmm_expected_states,subhmm_expected_trans)]
return expected_states, expected_transitions, expected_durations, normalizer, subhmm_stats
def _expected_durations(self,
dur_potentials,cumulative_obs_potentials,
alphastarl,betal,normalizer):
logpmfs = -np.inf*np.ones((self.Tfull,alphastarl.shape[1]))
errs = np.seterr(invalid='ignore') # logaddexp(-inf,-inf)
# TODO censoring not handled correctly here
for tblock in xrange(self.Tblock):
possible_durations = self.segmentlens[tblock:].cumsum()[:TRUNC]
obs_potentials, _ = cumulative_obs_potentials(tblock)
logpmfs[possible_durations -1] = np.logaddexp(
dur_potentials(tblock) + alphastarl[tblock]
+ betal[tblock:tblock+TRUNC if TRUNC is not None else None]
+ obs_potentials - normalizer,
logpmfs[possible_durations -1])
np.seterr(**errs)
return np.exp(logpmfs.T)
### OLD
# def block_cumulative_likelihoods(self,startblock,stopblock,possible_durations):
# # could recompute possible_durations given startblock, stopblock,
# # trunc/truncblock, and self.segmentlens, but why redo that effort?
# return np.vstack([self.block_cumulative_likelihood_state(startblock,stopblock,state,possible_durations) for state in range(self.num_states)]).T
# keep this one for forawrd sampling. maybe reimplement it?
def block_cumulative_likelihood_state(self,startblock,stopblock,state,possible_durations):
start = self.segmentstarts[startblock]
stop = self.segmentstarts[stopblock] if stopblock < len(self.segmentstarts) else None
return self.model.HMMs[state].cumulative_obs_potentials(self.aBls[state][start:])\
[possible_durations -1]
# return np.logaddexp.reduce(self.model.HMMs[state].messages_forwards(self.aBls[state][start:stop])[possible_durations-1],axis=1)