apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-04-30 23:36:23 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

New difference method for laplace

Browse source
This commit is contained in:
Alan Saul 2015-10-04 12:28:07 +01:00
parent 994db9f536
commit 2dc1b14934
1 changed files with 11 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

14
GPy/inference/latent_function_inference/laplace.py
View file

@ -32,8 +32,8 @@ class Laplace(LatentFunctionInference):
""" """
self._mode_finding_tolerance = 1e-7 self._mode_finding_tolerance = 1e-4
self._mode_finding_max_iter = 60 self._mode_finding_max_iter = 30
self.bad_fhat = False self.bad_fhat = False
#Store whether it is the first run of the inference so that we can choose whether we need #Store whether it is the first run of the inference so that we can choose whether we need
#to calculate things or reuse old variables #to calculate things or reuse old variables
@ -209,9 +209,12 @@ class Laplace(LatentFunctionInference):
Ki_f_new = Ki_f + step*dKi_f Ki_f_new = Ki_f + step*dKi_f
f_new = np.dot(K, Ki_f_new) f_new = np.dot(K, Ki_f_new)
#print "new {} vs old {}".format(obj(Ki_f_new, f_new), obj(Ki_f, f)) #print "new {} vs old {}".format(obj(Ki_f_new, f_new), obj(Ki_f, f))
if obj(Ki_f_new, f_new) < obj(Ki_f, f): old_obj = obj(Ki_f, f)
new_obj = obj(Ki_f_new, f_new)
if new_obj < old_obj:
raise ValueError("Shouldn't happen, brent optimization failing") raise ValueError("Shouldn't happen, brent optimization failing")
difference = np.abs(np.sum(f_new - f)) + np.abs(np.sum(Ki_f_new - Ki_f)) difference = np.abs(new_obj - old_obj)
# difference = np.abs(np.sum(f_new - f)) + np.abs(np.sum(Ki_f_new - Ki_f))
Ki_f = Ki_f_new Ki_f = Ki_f_new
f = f_new f = f_new
iteration += 1 iteration += 1
@ -316,6 +319,9 @@ class Laplace(LatentFunctionInference):
if not log_concave: if not log_concave:
#print "Under 1e-10: {}".format(np.sum(W < 1e-6)) #print "Under 1e-10: {}".format(np.sum(W < 1e-6))
W = np.clip(W, 1e-6, 1e+30) W = np.clip(W, 1e-6, 1e+30)
# For student-T we can clip this more intelligently. If the
# objective has hardly changed, we can increase the clipping limit
# by ((v+1)/v)/sigma2
# NOTE: when setting a parameter inside parameters_changed it will allways come to closed update circles!!! # NOTE: when setting a parameter inside parameters_changed it will allways come to closed update circles!!!
#W.__setitem__(W < 1e-6, 1e-6, update=False) # FIXME-HACK: This is a hack since GPy can't handle negative variances which can occur #W.__setitem__(W < 1e-6, 1e-6, update=False) # 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 # If the likelihood is non-log-concave. We wan't to say that there is a negative variance