apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-06-11 15:15:15 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Set warnings for truncated hessian, it has been noted that that by

Browse source 
truncating we can have incorrect posteriors, though at convergence
this should not be a problem, could be fixed by not using Cholsky
as the decomposition as it cannot handle non-positive definite mats
This commit is contained in:
Alan Saul 2013-12-03 15:21:45 +00:00
parent 4a751fd2da
commit cb6c1dd0d2
1 changed files with 10 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

17
GPy/likelihoods/laplace.py
View file

@ -250,8 +250,11 @@ class Laplace(likelihood):
self.W = -self.noise_model.d2logpdf_df2(self.f_hat, self.data, extra_data=self.extra_data)
if not self.noise_model.log_concave:
#print "Under 1e-10: {}".format(np.sum(self.W < 1e-6))
self.W[self.W < 1e-6] = 1e-6 # FIXME-HACK: This is a hack since GPy can't handle negative variances which can occur
i = self.W < 1e-6
if np.any(i):
warnings.warn('truncating non log-concave likelihood curvature')
# FIXME-HACK: This is a hack since GPy can't handle negative variances which can occur
self.W[i] = 1e-6
self.W12BiW12, self.ln_B_det = self._compute_B_statistics(self.K, self.W, np.eye(self.N))
@ -270,14 +273,14 @@ class Laplace(likelihood):
:type W: Vector of diagonal values of hessian (1xN)
:param a: Matrix to calculate W12BiW12a
:type a: Matrix NxN
:returns: (W12BiW12, ln_B_det)
:returns: (W12BiW12a, ln_B_det)
"""
if not self.noise_model.log_concave:
#print "Under 1e-10: {}".format(np.sum(W < 1e-6))
W[W < 1e-6] = 1e-6 # FIXME-HACK: This is a hack since GPy can't handle negative variances which can occur
# If the likelihood is non-log-concave. We wan't to say that there is a negative variance
# To cause the posterior to become less certain than the prior and likelihood,
# This is a property only held by non-log-concave likelihoods
W[W < 1e-10] = 1e-10 # FIXME-HACK: This is a hack since GPy can't handle negative variances which can occur
# If the likelihood is non-log-concave. We wan't to say that there is a negative variance
# To cause the posterior to become less certain than the prior and likelihood,
# This is a property only held by non-log-concave likelihoods
#W is diagonal so its sqrt is just the sqrt of the diagonal elements