def resample_with_censoring(self,data=[],censored_data=[]):
'''
censored_data is full of observations that were censored, meaning a
value of x really could have been anything >= x, so this method samples
them out to be at least that large
'''
filled_in = self._uncensor_data(censored_data)
return self.resample(data=combinedata((data,filled_in)))
def resample_with_censoring(self, data=[], censored_data=[]):
'''
censored_data is full of observations that were censored, meaning a
value of x really could have been anything >= x, so this method samples
them out to be at least that large
'''
filled_in = self._uncensor_data(censored_data)
return self.resample(data=combinedata((data, filled_in)))
def _generate(self,T):
alpha = self.alpha_0
betavec = self.beta.betavec
model = self.model
self.stateseq = np.array([])
ks = list(model._occupied()) + [None]
firststateidx = sample_discrete(np.arange(len(ks)))
if firststateidx == len(ks)-1:
firststate = self._new_label(ks)
else:
firststate = ks[firststateidx]
self.dur.resample(combinedata((model._durs_withlabel(firststate),self._durs_withlabel(firststate))))
firststate_dur = self.dur.rvs()
self.stateseq = np.ones(firststate_dur,dtype=int)*firststate
t = firststate_dur
# run a family-CRF (CRF with durations) forwards
while t < T:
ks = list(model._occupied() | self._occupied())
betarest = 1-sum(betavec[k] for k in ks)
fromto_counts = np.array([model._counts_fromto(self.stateseq[t-1],k)
+ self._counts_fromto(self.stateseq[t-1],k)
for k in ks])
scores = np.array([(alpha*betavec[k] + ft if k != self.stateseq[t-1] else 0)
for k,ft in zip(ks,fromto_counts)]
+ [alpha*(1-betavec[self.stateseq[t-1]])*betarest])
nextstateidx = sample_discrete(scores)
if nextstateidx == scores.shape[0]-1:
nextstate = self._new_label(ks)
else:
nextstate = ks[nextstateidx]
# now get the duration of nextstate!
self.dur.resample(combinedata((model._durs_withlabel(nextstate),self._durs_withlabel(nextstate))))
nextstate_dur = self.dur.rvs()
self.stateseq = np.concatenate((self.stateseq,np.ones(nextstate_dur,dtype=int)*nextstate))
t += nextstate_dur
self.T = len(self.stateseq)
def resample_with_censoring_and_truncation(self,data=[],censored_data=[],left_truncation_level=None):
filled_in = self._uncensor_data(censored_data)
if left_truncation_level is not None and left_truncation_level > 1:
norm = self.pmf(np.arange(1,left_truncation_level)).sum()
num_rejected = np.random.geometric(1-norm)-1
rejected_observations = self.rvs_given_less_than(left_truncation_level,num_rejected) \
if num_rejected > 0 else []
else:
rejected_observations = []
self.resample(data=combinedata((data,filled_in,rejected_observations)))
def resample_with_censoring_and_truncation(self,
data=[],
censored_data=[],
left_truncation_level=None):
filled_in = self._uncensor_data(censored_data)
if left_truncation_level is not None and left_truncation_level > 1:
norm = self.pmf(np.arange(1, left_truncation_level)).sum()
num_rejected = np.random.geometric(1 - norm) - 1
rejected_observations = self.rvs_given_less_than(left_truncation_level,num_rejected) \
if num_rejected > 0 else []
else:
rejected_observations = []
self.resample(data=combinedata((data, filled_in,
rejected_observations)))
