def _meanfieldupdate_sigma(self, stats, prob=1.0, stepsize=1.0):
E_ysq, E_yxT, E_xxT, E_n = stats / prob
E_A, E_AAT, _, _ = self.mf_expectations
alpha = self.alpha_0 + E_n / 2.0
beta = self.beta_0
beta += 0.5 * E_ysq
beta += -1.0 * np.sum(E_yxT * E_A, axis=1)
beta += 0.5 * np.sum(E_AAT * E_xxT, axis=(1, 2))
# Set the invgamma meanfield parameters
self.mf_alpha = update_param(self.mf_alpha, alpha, stepsize)
self.mf_beta = update_param(self.mf_beta, beta, stepsize)
def _meanfieldupdate_sigma(self, stats, prob=1.0, stepsize=1.0):
E_ysq, E_yxT, E_xxT, E_n = stats / prob
E_A, E_AAT, _, _ = self.mf_expectations
alpha = self.alpha_0 + E_n / 2.0
beta = self.beta_0
beta += 0.5 * E_ysq
beta += -1.0 * np.sum(E_yxT * E_A, axis=1)
beta += 0.5 * np.sum(E_AAT * E_xxT, axis=(1,2))
# Set the invgamma meanfield parameters
self.mf_alpha = update_param(self.mf_alpha, alpha, stepsize)
self.mf_beta = update_param(self.mf_beta, beta, stepsize)
def _meanfieldupdate_A(self, stats, prob=1.0, stepsize=1.0):
E_sigmasq_inv = self.mf_alpha / self.mf_beta
_, E_yxT, E_xxT, _ = stats / prob
# Update statistics each row of A
for d in range(self.D_out):
Jd = self.J_0 + (E_xxT[d] * E_sigmasq_inv[d])
hd = self.h_0 + (E_yxT[d] * E_sigmasq_inv[d])
# Update the mean field natural parameters
self.mf_J_A[d] = update_param(self.mf_J_A[d], Jd, stepsize)
self.mf_h_A[d] = update_param(self.mf_h_A[d], hd, stepsize)
# Clear the cache
self._mf_A_cache = {}
def _meanfieldupdate_A(self, stats, prob=1.0, stepsize=1.0):
E_sigmasq_inv = self.mf_alpha / self.mf_beta
_, E_yxT, E_xxT, _ = stats / prob
# Update statistics each row of A
for d in range(self.D_out):
Jd = self.J_0 + (E_xxT[d] * E_sigmasq_inv[d])
hd = self.h_0 + (E_yxT[d] * E_sigmasq_inv[d])
# Update the mean field natural parameters
self.mf_J_A[d] = update_param(self.mf_J_A[d], Jd, stepsize)
self.mf_h_A[d] = update_param(self.mf_h_A[d], hd, stepsize)
# Clear the cache
self._mf_A_cache = {}
