def build_prior_KL(self):
# whitening of priors can be implemented here
"""
This gives KL divergence between inducing points priors and approximated posteriors
KL(q(u_g)||p(u_g)) + KL(q(u_f)||p(u_f))
q(u_f) = N(u_f|u_fm,u_fs)
p(u_f) = N(u_f|0,Kfmm)
q(u_g) = N(u_g|u_gm,u_gs)
p(u_g) = N(u_g|0,Kgmm)
"""
if self.whiten:
if self.q_diag:
KL = kullback_leiblers.gauss_kl_white_diag(self.u_fm, self.u_fs_sqrt) + \
kullback_leiblers.gauss_kl_white_diag(self.u_gm, self.u_gs_sqrt)
else:
KL = kullback_leiblers.gauss_kl_white(self.u_fm, self.u_fs_sqrt) + \
kullback_leiblers.gauss_kl_white(self.u_gm, self.u_gs_sqrt)
else:
Kfmm = self.kernf.K(self.Zf) + tf.eye(
self.num_inducing_f,
dtype=float_type) * settings.numerics.jitter_level
Kgmm = self.kerng.K(self.Zg) + tf.eye(
self.num_inducing_g,
dtype=float_type) * settings.numerics.jitter_level
if self.q_diag:
KL = kullback_leiblers.gauss_kl_diag(self.u_fm, self.u_fs_sqrt, Kfmm) + \
kullback_leiblers.gauss_kl_diag(self.u_gm, self.u_gs_sqrt, Kgmm)
else:
KL = kullback_leiblers.gauss_kl(self.u_fm, self.u_fs_sqrt, Kfmm) + \
kullback_leiblers.gauss_kl(self.u_gm, self.u_gs_sqrt, Kgmm)
return KL
def build_prior_KL(self):
if self.whiten:
if self.q_diag:
KL = kullback_leiblers.gauss_kl_white_diag(
self.q_mu, self.q_sqrt)
else:
KL = kullback_leiblers.gauss_kl_white(self.q_mu, self.q_sqrt)
else:
K = self.kern.K(self.Z) + tf.eye(self.num_inducing,
dtype=float_type) * settings.numerics.jitter_level
if self.q_diag:
KL = kullback_leiblers.gauss_kl_diag(self.q_mu, self.q_sqrt, K)
else:
KL = kullback_leiblers.gauss_kl(self.q_mu, self.q_sqrt, K)
return KL
def KL(self):
return gauss_kl_white(self.q_mu, self.q_sqrt)
def KL(self):
all_mu = tf.concat([n.q_mu for n in self.nodes], 1)
all_sqrt = tf.concat([n.q_sqrt for n in self.nodes], 2)
return gauss_kl_white(all_mu, all_sqrt)