def step(self):
self._index += 1
if self._index % self.sleep_interval == 0:
v = pm.value(self.v)
m = pm.value(self.m)
val = self.stochastic.value
lp = pm.logp_of_set(self.other_children)
# Choose a direction along which to step.
dirvec = np.random.normal(size=self.n)
dirvec /= np.sqrt(np.sum(dirvec**2))
# Orthogonalize
orthoproj = gramschmidt(dirvec)
scaled_orthoproj = v*orthoproj.T
pck = np.dot(dirvec, scaled_orthoproj.T)
kck = np.linalg.inv(np.dot(scaled_orthoproj,orthoproj))
pckkck = np.dot(pck,kck)
# Figure out conditional variance
condvar = np.dot(dirvec, dirvec*v) - np.dot(pck, pckkck)
# condmean = np.dot(dirvec, m) + np.dot(pckkck, np.dot(orthoproj.T, (val-m)))
# Compute slice of log-probability surface
tries = np.linspace(-4*np.sqrt(condvar), 4*np.sqrt(condvar), 501)
lps = 0*tries
for i in xrange(len(tries)):
new_val = val + tries[i]*dirvec
self.stochastic.value = new_val
try:
lps[i] = self.f_fr.logp + self.stochastic.logp
except:
lps[i] = -np.inf
if np.all(np.isinf(lps)):
raise ValueError, 'All -inf.'
lps -= pm.flib.logsum(lps[True-np.isinf(lps)])
ps = np.exp(lps)
index = pm.rcategorical(ps)
new_val = val + tries[index]*dirvec
self.stochastic.value = new_val
try:
lpp = pm.logp_of_set(self.other_children)
if np.log(np.random.random()) < lpp - lp:
self.accepted += 1
else:
self.stochastic.value = val
self.rejected += 1
except pm.ZeroProbability:
self.stochastic.value = val
self.rejected += 1
self.logp_plus_loglike
def get_logp_plus_loglike(self):
return pm.logp_of_set(self.markov_blanket_no_f)
def get_logp_plus_loglike(self):
return pm.logp_of_set(self.markov_blanket_no_f)
def pymc_logp():
return logp_of_set(model)
def _get_logp_plus_loglike(self):
return pm.logp_of_set(self.children)
def _get_logp_plus_loglike(self):
return pm.logp_of_set(self.children)
def pymc_logp():
return logp_of_set(model)
def get_logp_plus_loglike(self):
return pm.logp_of_set(self.mb_for_logp) + self.evidence.value
def approximate_evidence(gmrf, Mc, ppc):
"The INLA approximation \pi(y|\theta)"
pygx = pm.logp_of_set(gmrf.extended_children)
px = gmrf.logp
pxgy = gmrf.parents['backend'].mvn_logp(gmrf.value, Mc, **ppc)
return pygx + px - pxgy
