def E_step(self):
alphal = HMMStatesEigen._messages_forwards_log(self.hmm_trans_matrix, self.pi_0, self.aBl)
betal = HMMStatesEigen._messages_backwards_log(self.hmm_trans_matrix, self.aBl)
self.expected_states, self.expected_transcounts, self._normalizer = HMMStatesPython._expected_statistics_from_messages(
self.hmm_trans_matrix, self.aBl, alphal, betal
)
# using these is untested!
self._expected_ns = np.diag(self.expected_transcounts).copy()
self._expected_tots = self.expected_transcounts.sum(1)
self.expected_transcounts.flat[:: self.expected_transcounts.shape[0] + 1] = 0.0
def meanfieldupdate_Estep(self):
# TODO bug in here? it's not as good as sampling
num_r_samples = self.model.mf_num_samples \
if hasattr(self.model,'mf_num_samples') else 10
num_stateseq_samples_per_r = self.model.mf_num_stateseq_samples_per_r \
if hasattr(self.model,'mf_num_stateseq_samples_per_r') else 1
self.expected_states = np.zeros((self.T, self.num_states))
self.expected_transcounts = np.zeros(
(self.num_states, self.num_states))
self.expected_durations = np.zeros((self.num_states, self.T))
mf_aBl = self.mf_aBl
for i in xrange(num_r_samples):
for d in self.dur_distns:
d._resample_r_from_mf()
self.clear_caches()
trans = self.mf_bwd_trans_matrix # TODO check this
init = self.hmm_mf_bwd_pi_0
aBl = mf_aBl.repeat(self.rs, axis=1)
hmm_alphal, hmm_betal = HMMStatesEigen._messages_log(
self, trans, init, aBl)
# collect stateseq and transitions statistics from messages
hmm_expected_states, hmm_expected_transcounts, normalizer = \
HMMStatesPython._expected_statistics_from_messages(
trans,aBl,hmm_alphal,hmm_betal)
expected_states, expected_transcounts, _ \
= self._hmm_stats_to_hsmm_stats(
hmm_expected_states, hmm_expected_transcounts, normalizer)
self.expected_states += expected_states / num_r_samples
self.expected_transcounts += expected_transcounts / num_r_samples
# collect duration statistics by sampling from messages
for j in xrange(num_stateseq_samples_per_r):
self._resample_from_mf(trans, init, aBl, hmm_alphal, hmm_betal)
for state in xrange(self.num_states):
self.expected_durations[state] += \
np.bincount(
self.durations_censored[self.stateseq_norep == state],
minlength=self.T)[:self.T].astype(np.double) \
/(num_r_samples*num_stateseq_samples_per_r)
def E_step(self):
alphal = HMMStatesEigen._messages_forwards_log(self.hmm_trans_matrix,
self.pi_0, self.aBl)
betal = HMMStatesEigen._messages_backwards_log(self.hmm_trans_matrix,
self.aBl)
self.expected_states, self.expected_transcounts, self._normalizer = \
HMMStatesPython._expected_statistics_from_messages(
self.hmm_trans_matrix,
self.aBl,
alphal,
betal)
# using these is untested!
self._expected_ns = np.diag(self.expected_transcounts).copy()
self._expected_tots = self.expected_transcounts.sum(1)
self.expected_transcounts.flat[::self.expected_transcounts.shape[0] +
1] = 0.
def meanfieldupdate_Estep(self):
# TODO bug in here? it's not as good as sampling
num_r_samples = self.model.mf_num_samples \
if hasattr(self.model,'mf_num_samples') else 10
num_stateseq_samples_per_r = self.model.mf_num_stateseq_samples_per_r \
if hasattr(self.model,'mf_num_stateseq_samples_per_r') else 1
self.expected_states = np.zeros((self.T,self.num_states))
self.expected_transcounts = np.zeros((self.num_states,self.num_states))
self.expected_durations = np.zeros((self.num_states,self.T))
mf_aBl = self.mf_aBl
for i in xrange(num_r_samples):
for d in self.dur_distns:
d._resample_r_from_mf()
self.clear_caches()
trans = self.mf_bwd_trans_matrix # TODO check this
init = self.hmm_mf_bwd_pi_0
aBl = mf_aBl.repeat(self.rs,axis=1)
hmm_alphal, hmm_betal = HMMStatesEigen._messages_log(self,trans,init,aBl)
# collect stateseq and transitions statistics from messages
hmm_expected_states, hmm_expected_transcounts, normalizer = \
HMMStatesPython._expected_statistics_from_messages(
trans,aBl,hmm_alphal,hmm_betal)
expected_states, expected_transcounts, _ \
= self._hmm_stats_to_hsmm_stats(
hmm_expected_states, hmm_expected_transcounts, normalizer)
self.expected_states += expected_states / num_r_samples
self.expected_transcounts += expected_transcounts / num_r_samples
# collect duration statistics by sampling from messages
for j in xrange(num_stateseq_samples_per_r):
self._resample_from_mf(trans,init,aBl,hmm_alphal,hmm_betal)
for state in xrange(self.num_states):
self.expected_durations[state] += \
np.bincount(
self.durations_censored[self.stateseq_norep == state],
minlength=self.T)[:self.T].astype(np.double) \
/(num_r_samples*num_stateseq_samples_per_r)
def messages_backwards_log_hmm(self):
return HMMStatesPython._messages_backwards_log(
self.hmm_backwards_trans_matrix,self.hmm_aBl)
def messages_backwards_normalized_hmm(self):
return HMMStatesPython._messages_backwards_normalized(
self.hmm_backwards_trans_matrix,self.hmm_backwards_pi_0,self.hmm_aBl)
