def resample_states(self,**kwargs):
# NOTE: kwargs is just to absorb any multiprocessing stuff
# TODO only use this when the number/size of sequences warrant it
from messages import resample_arhmm
assert self.obs_distns[0].D_out > 1
if len(self.states_list) > 0:
stateseqs = [np.empty(s.T,dtype='int32') for s in self.states_list]
params, normalizers = map(np.array,zip(*[self._param_matrix(o) for o in self.obs_distns]))
params, normalizers = params.repeat(s.rs,axis=0), normalizers.repeat(s.rs,axis=0)
stats, _, loglikes = resample_arhmm(
[s.hmm_pi_0.astype(self.dtype) for s in self.states_list],
[s.hmm_trans_matrix.astype(self.dtype) for s in self.states_list],
params.astype(self.dtype), normalizers.astype(self.dtype),
[undo_AR_striding(s.data,self.nlags) for s in self.states_list],
stateseqs,
[np.random.uniform(size=s.T).astype(self.dtype) for s in self.states_list],
self.alphans)
for s, stateseq, loglike in zip(self.states_list,stateseqs,loglikes):
s.stateseq = stateseq
s._map_states()
s._normalizer = loglike
starts, ends = cumsum(s.rs,strict=True), cumsum(s.rs,strict=False)
stats = map(np.array,stats)
stats = [sum(stats[start:end]) for start, end in zip(starts,ends)]
self._obs_stats = stats
else:
self._obs_stats = None
def slices_from_indicators(indseq):
indseq = np.asarray(indseq)
if not indseq.any():
return []
else:
vals, durs = rle(indseq)
starts, ends = cumsum(durs, strict=True), cumsum(durs, strict=False)
return [
slice(start, end) for val, start, end in zip(vals, starts, ends)
if val
]
def hmm_pi_0(self):
delays = self.delays
starts = cumsum(delays,strict=True)
pi_0 = np.zeros(delays.sum())
pi_0[starts] = self.pi_0
return pi_0
def hmm_pi_0(self):
delays = self.delays
starts = cumsum(delays,strict=True)
pi_0 = np.zeros(delays.sum())
pi_0[starts] = self.pi_0
return pi_0
def hmm_trans_matrix(self):
# NOTE: more general version, allows different delays, o/w we could
# construct with np.kron
if self._hmm_trans_matrix is None:
ps, delays = map(np.array,zip(*[(d.p,d.delay) for d in self.dur_distns]))
starts, ends = cumsum(delays,strict=True), cumsum(delays,strict=False)
trans_matrix = self._hmm_trans_matrix = np.zeros((ends[-1],ends[-1]))
for (i,j), Aij in np.ndenumerate(self.trans_matrix):
block = trans_matrix[starts[i]:ends[i],starts[j]:ends[j]]
if i == j:
block[:-1,1:] = np.eye(block.shape[0]-1)
block[-1,-1] = 1-ps[i]
else:
block[-1,0] = ps[j]*Aij
return self._hmm_trans_matrix
def hmm_trans_matrix(self):
# NOTE: more general version, allows different delays, o/w we could
# construct with np.kron
if self._hmm_trans_matrix is None:
ps, delays = map(np.array,zip(*[(d.p,d.delay) for d in self.dur_distns]))
starts, ends = cumsum(delays,strict=True), cumsum(delays,strict=False)
trans_matrix = self._hmm_trans_matrix = np.zeros((ends[-1],ends[-1]))
for (i,j), Aij in np.ndenumerate(self.trans_matrix):
block = trans_matrix[starts[i]:ends[i],starts[j]:ends[j]]
if i == j:
block[:-1,1:] = np.eye(block.shape[0]-1)
block[-1,-1] = 1-ps[i]
else:
block[-1,0] = ps[j]*Aij
return self._hmm_trans_matrix
def _beta_vlb(self):
return np.log(self.beta).sum() + self.gamma*np.log(1-cumsum(self.beta)).sum() \
- 3*np.log(1-cumsum(self.beta,strict=True)).sum()
def _grad_log_p_beta(beta,alpha):
# NOTE: switched argument name gamma <-> alpha
return -(alpha-1)*rcumsum(1./(1-cumsum(beta))) \
+ 2*rcumsum(1./(1-cumsum(beta,strict=True)),strict=True)
def _beta_vlb(self):
return np.log(self.beta).sum() + self.gamma*np.log(1-cumsum(self.beta)).sum() \
- 3*np.log(1-cumsum(self.beta,strict=True)).sum()
def _grad_log_p_beta(beta,alpha):
# NOTE: switched argument name gamma <-> alpha
return -(alpha-1)*rcumsum(1./(1-cumsum(beta))) \
+ 2*rcumsum(1./(1-cumsum(beta,strict=True)),strict=True)
def feas(pt):
return (pt > 0.).all() and pt.sum() < 1. and not np.isinf(
1. / (1 - cumsum(pt))).any()
def feas(pt):
return (pt>0.).all() and pt.sum() < 1. and not np.isinf(1./(1-cumsum(pt))).any()
