def get_vlb(self):
E_Sigmainv, E_Sigmainv_A, E_AT_Sigmainv_A, E_logdetSigmainv = \
mniw_expectedstats(*self._natural_to_standard(self.mf_natural_hypparam))
A, B, C, d = self.natural_hypparam - self.mf_natural_hypparam
bilinear_term = -1./2 * np.trace(A.dot(E_Sigmainv)) \
+ np.trace(B.T.dot(E_Sigmainv_A)) \
- 1./2 * np.trace(C.dot(E_AT_Sigmainv_A)) \
+ 1./2 * d * E_logdetSigmainv
# log normalizer term
Z = mniw_log_partitionfunction(
*self._natural_to_standard(self.natural_hypparam))
Z_mf = mniw_log_partitionfunction(
*self._natural_to_standard(self.mf_natural_hypparam))
return bilinear_term - (Z - Z_mf)
def get_vlb(self):
E_Sigmainv, E_Sigmainv_A, E_AT_Sigmainv_A, E_logdetSigmainv = \
mniw_expectedstats(*self._natural_to_standard(self.mf_natural_hypparam))
A, B, C, d = self.natural_hypparam - self.mf_natural_hypparam
bilinear_term = -1./2 * np.trace(A.dot(E_Sigmainv)) \
+ np.trace(B.T.dot(E_Sigmainv_A)) \
- 1./2 * np.trace(C.dot(E_AT_Sigmainv_A)) \
+ 1./2 * d * E_logdetSigmainv
# log normalizer term
Z = mniw_log_partitionfunction(*self._natural_to_standard(
self.natural_hypparam))
Z_mf = mniw_log_partitionfunction(*self._natural_to_standard(
self.mf_natural_hypparam))
return bilinear_term - (Z - Z_mf)
def dense_regression_logprior(regression):
A = regression.A
Sigmainv = np.linalg.inv(regression.sigma)
Sigmainv_A = Sigmainv.dot(A)
AT_Sigmainv_A = A.T.dot(Sigmainv_A)
logdetSigmainv = np.linalg.slogdet(Sigmainv)[1]
A, B, C, d = regression.natural_hypparam
bilinear_term = -1./2 * np.trace(A.dot(Sigmainv)) \
+ np.trace(B.T.dot(Sigmainv_A)) \
- 1./2 * np.trace(C.dot(AT_Sigmainv_A)) \
+ 1./2 * d * logdetSigmainv
# log normalizer term
from pybasicbayes.util.stats import mniw_log_partitionfunction
Z = mniw_log_partitionfunction(
*regression._natural_to_standard(regression.natural_hypparam))
return bilinear_term - Z
def dense_regression_logprior(regression):
A = regression.A
Sigmainv = np.linalg.inv(regression.sigma)
Sigmainv_A = Sigmainv.dot(A)
AT_Sigmainv_A = A.T.dot(Sigmainv_A)
logdetSigmainv = np.linalg.slogdet(Sigmainv)[1]
A, B, C, d = regression.natural_hypparam
bilinear_term = -1./2 * np.trace(A.dot(Sigmainv)) \
+ np.trace(B.T.dot(Sigmainv_A)) \
- 1./2 * np.trace(C.dot(AT_Sigmainv_A)) \
+ 1./2 * d * logdetSigmainv
# log normalizer term
from pybasicbayes.util.stats import mniw_log_partitionfunction
Z = mniw_log_partitionfunction(
*regression._natural_to_standard(regression.natural_hypparam))
return bilinear_term - Z