def hmm_messages_backwards(self):
betal = HMMStatesEigen._messages_backwards_log(
self.hmm_bwd_trans_matrix,
self.hmm_aBl)
self._normalizer = np.logaddexp.reduce(
np.log(self.hmm_bwd_pi_0) + betal[0] + self.hmm_aBl[0])
return betal
def hmm_messages_backwards(self):
betal = HMMStatesEigen._messages_backwards_log(
self.hmm_bwd_trans_matrix,
self.hmm_aBl)
self._normalizer = np.logaddexp.reduce(
np.log(self.hmm_bwd_pi_0) + betal[0] + self.hmm_aBl[0])
return betal
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 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 resample_hmm(self):
betal = HMMStatesEigen._messages_backwards_log(
self.hmm_bwd_trans_matrix,self.hmm_aBl)
self.stateseq = HMMStatesEigen._sample_forwards_log(
betal,self.hmm_bwd_trans_matrix,self.hmm_bwd_pi_0,self.hmm_aBl)
self._map_states()
