def _do_forward_pass(self, framelogprob):
n_samples, n_components = framelogprob.shape
fwdlattice = np.zeros((n_samples, n_components))
_hmmc._forward(n_samples, n_components, log_mask_zero(self.startprob_),
log_mask_zero(self.transmat_), framelogprob, fwdlattice)
with np.errstate(under="ignore"):
return logsumexp(fwdlattice[-1]), fwdlattice
def _compute_logprob(self, X, startprob, transmat, **kwargs):
logprobX = self._logprob_X(X, **kwargs)
n_samples, n_components = logprobX.shape
fwdlattice = np.zeros((n_samples, n_components))
_forward(n_samples, n_components, np.log(startprob), np.log(transmat),
logprobX, fwdlattice)
return logsumexp(fwdlattice[-1])
def _do_forward_pass(log_startprob, log_transmat, framelogprob):
n_samples, n_components = framelogprob.shape
fwdlattice = np.zeros((n_samples, n_components))
_hmmc._forward(
n_samples, n_components, log_startprob, log_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_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):
n_samples, n_components = framelogprob.shape
# archived numpy version
# fwdlattice = _routines._forward(n_samples, n_components,
# log_mask_zero(self.startprob_),
# log_mask_zero(self.transmat_),
# framelogprob)
fwdlattice = np.zeros((n_samples, n_components))
_hmmc._forward(n_samples, n_components,
log_mask_zero(self.startprob_),
log_mask_zero(self.transmat_),
framelogprob, fwdlattice)
with np.errstate(under="ignore"):
return special.logsumexp(fwdlattice[-1]), fwdlattice
print("hmm transition probability: \n{}".format(transmat))
data, mask = batch_data(n_components)
print("batch_framelogprob is: \n{}".format(data))
print("batch_mask is: \n{}".format(mask))
# 2. forward comparison
# 2.1 forward by hmmlearn method:
hmmlearn_fwdlattice = []
hmmlearn_logprob = []
for idx, framelogprob in enumerate(data):
fwdlattice = np.zeros((mask[idx].sum(), n_components))
_hmmc._forward(int(mask[idx].sum()), \
int(n_components), \
np.log(startprob), \
np.log(transmat), \
framelogprob[mask[idx]>0], \
fwdlattice)
hmmlearn_fwdlattice.append(fwdlattice)
hmmlearn_logprob.append(logsumexp(fwdlattice[-1]))
print("hmmlearn forward lattice: \n {}\n".format(hmmlearn_fwdlattice))
print("hmmlearn logporb: {}\n".format(hmmlearn_logprob))
# 2.2 batch forward in PyTorch
torch_logprob, torch_fwdlattice = _forward(int(data.shape[1]), \
int(n_components), \
torch.from_numpy(np.log(startprob)), \
torch.from_numpy(np.log(transmat)), \
torch.from_numpy(data), \
torch.from_numpy(mask))
print("torch batch forward: \n {}\n".format(torch_fwdlattice))
def _do_forward_pass(self, framelogprob):
n_observations, n_components = framelogprob.shape
fwdlattice = np.zeros((n_observations, n_components))
_hmmc._forward(n_observations, n_components, self._log_startprob,
self._log_transmat, framelogprob, fwdlattice)
return logsumexp(fwdlattice[-1]), fwdlattice