def update(self, smc):
prev_Phi = self.Phi.copy()
mq = rs.MultinomialQueue(self.prev_W)
nprop = 0
for n in range(self.N):
As = np.empty(self.Nparis, 'int')
for m in range(self.Nparis):
while True:
a = mq.dequeue(1)
nprop += 1
lp = (smc.fk.logpt(smc.t, self.prev_X[a], smc.X[n]) -
smc.fk.upper_bound_log_pt(t))
if np.log(random.rand()) < lp:
break
As[m] = a
mod_Phi = (self.prev_Phi[As] +
smc.fk.add_func(smc.t, self.prev_X[As], smc.X[n]))
self.Phi[n] = np.average(mod_Phi, axis=0)
self.nprop.append(nprop)
def _backward_sampling_ON(self, M, idx):
"""O(N) version of backward sampling.
not meant to be called directly, see backward_sampling
"""
nattempts = 0
for t in reversed(range(self.T - 1)):
where_rejected = np.arange(M)
who_rejected = self.X[t + 1][idx[t + 1, :]]
nrejected = M
gen = rs.MultinomialQueue(self.wgts[t].W, M=M)
while nrejected > 0:
nattempts += nrejected
nprop = gen.dequeue(nrejected)
lpr_acc = (
self.fk.logpt(t + 1, self.X[t][nprop], who_rejected) -
self.fk.upper_bound_trans(t + 1))
newly_accepted = np.log(random.rand(nrejected)) < lpr_acc
still_rejected = np.logical_not(newly_accepted)
idx[t, where_rejected[newly_accepted]] = nprop[newly_accepted]
where_rejected = where_rejected[still_rejected]
who_rejected = who_rejected[still_rejected]
nrejected -= sum(newly_accepted)
return (M * (self.T - 1)) / nattempts