def _kld(self, z, q_param, p_param=None):
"""
Computes the KL-divergence of
some element z.
KL(q||p) = -∫ q(z) log [ p(z) / q(z) ]
= -E[log p(z) - log q(z)]
:param z: sample from q-distribuion
:param q_param: (mu, log_var) of the q-distribution
:param p_param: (mu, log_var) of the p-distribution
:return: KL(q||p)
"""
(mu, log_var) = q_param
if self.flow is not None:
f_z, log_det_z = self.flow(z)
qz = log_gaussian(z, mu, log_var) - sum(log_det_z)
z = f_z
else:
qz = log_gaussian(z, mu, log_var)
if p_param is None:
pz = log_standard_gaussian(z)
else:
(mu, log_var) = p_param
pz = log_gaussian(z, mu, log_var)
kl = qz - pz
return kl
def _kld(self, z, q_param, p_param=None):
"""
Compute KL-divergence of some element z.
Inputs:
z : sample from the q distribution
q_param : (mu, log_var) of the q distribution.
p_param : (mu, log_var) of the p distribution.
Returns:
KL-divergence of q||p
"""
# Define q distribution
(mu, log_var) = q_param
qz = log_gaussian(z, mu, log_var)
# Define p distribution
if p_param is None:
pz = log_standard_gaussian(z)
else:
(mu, log_var) = p_param
pz = log_gaussian(z, mu, log_var)
kl = qz - pz
return kl
def log_gauss(x, mu, log_var):
return -log_gaussian(x, mu, log_var)