def _do_forward_pass(self, framelogprob):
# Based on hmmlearn's _BaseHMM
safe_startmat = self.startprob_ + np.finfo(float).eps
safe_transmat = self.transmat_ + np.finfo(float).eps
n_samples, n_components = framelogprob.shape
fwdlattice = np.zeros((n_samples, n_components))
_hmmc._forward(n_samples, n_components, np.log(safe_startmat),
np.log(safe_transmat), framelogprob, fwdlattice)
return logsumexp(fwdlattice[-1]), fwdlattice
def _do_forward_pass(self, framelogprob):
# Based on hmmlearn's _BaseHMM
safe_startmat = self.startprob_ + np.finfo(float).eps
safe_transmat = self.transmat_ + np.finfo(float).eps
n_samples, n_components = framelogprob.shape
fwdlattice = np.zeros((n_samples, n_components))
_hmmc._forward(n_samples, n_components,
np.log(safe_startmat),
np.log(safe_transmat),
framelogprob, fwdlattice)
return logsumexp(fwdlattice[-1]), fwdlattice
def _do_score_samples(self, data, lengths=None): # adapted hmmlearn
# TODO: Support lengths arguement
framelogprob = self._compute_log_likelihood(data)
logprob, fwdlattice = self._do_forward_pass(framelogprob)
bwdlattice = self._do_backward_pass(framelogprob)
gamma = fwdlattice + bwdlattice
# gamma is guaranteed to be correctly normalized by logprob at
# all frames, unless we do approximate inference using pruning.
# So, we will normalize each frame explicitly in case we
# pruned too aggressively.
posteriors = np.exp(gamma.T - logsumexp(gamma, axis=1)).T
posteriors += np.finfo(np.float64).eps
posteriors /= np.sum(posteriors, axis=1).reshape((-1, 1))
return logprob, posteriors
def _do_score_samples(self, data, lengths=None): # adapted hmmlearn
# TODO: Support lengths arguement
framelogprob = self._compute_log_likelihood(data)
logprob, fwdlattice = self._do_forward_pass(framelogprob)
bwdlattice = self._do_backward_pass(framelogprob)
gamma = fwdlattice + bwdlattice
# gamma is guaranteed to be correctly normalized by logprob at
# all frames, unless we do approximate inference using pruning.
# So, we will normalize each frame explicitly in case we
# pruned too aggressively.
posteriors = np.exp(gamma.T - logsumexp(gamma, axis=1)).T
posteriors += np.finfo(np.float64).eps
posteriors /= np.sum(posteriors, axis=1).reshape((-1, 1))
return logprob, posteriors
def _decode_map(self, data): # adapted hmmlearn
framelogprob = self._compute_log_likelihood(data)
logprob, fwdlattice = self._do_forward_pass(framelogprob)
bwdlattice = self._do_backward_pass(framelogprob)
gamma = fwdlattice + bwdlattice
# gamma is guaranteed to be correctly normalized by logprob at
# all frames, unless we do approximate inference using pruning.
# So, we will normalize each frame explicitly in case we
# pruned too aggressively.
posteriors = np.exp(gamma.T - logsumexp(gamma, axis=1)).T
posteriors += np.finfo(np.float64).eps
posteriors /= np.sum(posteriors, axis=1).reshape((-1, 1))
state_sequence = np.argmax(posteriors, axis=1)
map_logprob = np.max(posteriors, axis=1).sum()
return map_logprob, state_sequence
def _decode_map(self, data): # adapted hmmlearn
framelogprob = self._compute_log_likelihood(data)
logprob, fwdlattice = self._do_forward_pass(framelogprob)
bwdlattice = self._do_backward_pass(framelogprob)
gamma = fwdlattice + bwdlattice
# gamma is guaranteed to be correctly normalized by logprob at
# all frames, unless we do approximate inference using pruning.
# So, we will normalize each frame explicitly in case we
# pruned too aggressively.
posteriors = np.exp(gamma.T - logsumexp(gamma, axis=1)).T
posteriors += np.finfo(np.float64).eps
posteriors /= np.sum(posteriors, axis=1).reshape((-1, 1))
state_sequence = np.argmax(posteriors, axis=1)
map_logprob = np.max(posteriors, axis=1).sum()
return map_logprob, state_sequence
def _accumulate_sufficient_statistics(self, stats, X, framelogprob,
posteriors, fwdlattice, bwdlattice):
"""Updates sufficient statistics from a given sample.
Parameters
----------
stats : dict
Sufficient statistics as returned by
:meth:`~base._BaseHMM._initialize_sufficient_statistics`.
X : array, shape (n_samples, n_features)
Sample sequence.
framelogprob : array, shape (n_samples, n_components)
Log-probabilities of each sample under each of the model states.
posteriors : array, shape (n_samples, n_components)
Posterior probabilities of each sample being generated by each
of the model states.
fwdlattice, bwdlattice : array, shape (n_samples, n_components)
Log-forward and log-backward probabilities.
"""
# Based on hmmlearn's _BaseHMM
safe_transmat = self.transmat_ + np.finfo(float).eps
stats['nobs'] += 1
if 's' in self.params:
stats['start'] += posteriors[0]
if 't' in self.params:
n_samples, n_components = framelogprob.shape
# when the sample is of length 1, it contains no transitions
# so there is no reason to update our trans. matrix estimate
if n_samples <= 1:
return
lneta = np.zeros((n_samples - 1, n_components, n_components))
_hmmc._compute_lneta(n_samples, n_components, fwdlattice,
np.log(safe_transmat), bwdlattice,
framelogprob, lneta)
stats['trans'] += np.exp(logsumexp(lneta, axis=0))
# stats['trans'] = np.round(stats['trans'])
# if np.sum(stats['trans']) != X.shape[0]-1:
# warnings.warn("transmat counts != n_samples", RuntimeWarning)
# import pdb; pdb.set_trace()
stats['trans'][np.where(stats['trans'] < 0.01)] = 0.0
def _accumulate_sufficient_statistics(self, stats, X, framelogprob,
posteriors, fwdlattice, bwdlattice):
"""Updates sufficient statistics from a given sample.
Parameters
----------
stats : dict
Sufficient statistics as returned by
:meth:`~base._BaseHMM._initialize_sufficient_statistics`.
X : array, shape (n_samples, n_features)
Sample sequence.
framelogprob : array, shape (n_samples, n_components)
Log-probabilities of each sample under each of the model states.
posteriors : array, shape (n_samples, n_components)
Posterior probabilities of each sample being generated by each
of the model states.
fwdlattice, bwdlattice : array, shape (n_samples, n_components)
Log-forward and log-backward probabilities.
"""
# Based on hmmlearn's _BaseHMM
safe_transmat = self.transmat_ + np.finfo(float).eps
stats['nobs'] += 1
if 's' in self.params:
stats['start'] += posteriors[0]
if 't' in self.params:
n_samples, n_components = framelogprob.shape
# when the sample is of length 1, it contains no transitions
# so there is no reason to update our trans. matrix estimate
if n_samples <= 1:
return
lneta = np.zeros((n_samples - 1, n_components, n_components))
_hmmc._compute_lneta(n_samples, n_components, fwdlattice,
np.log(safe_transmat),
bwdlattice, framelogprob, lneta)
stats['trans'] += np.exp(logsumexp(lneta, axis=0))
# stats['trans'] = np.round(stats['trans'])
# if np.sum(stats['trans']) != X.shape[0]-1:
# warnings.warn("transmat counts != n_samples", RuntimeWarning)
# import pdb; pdb.set_trace()
stats['trans'][np.where(stats['trans'] < 0.01)] = 0.0