def _update_log_varpar_assignment(self, children_label_np):
digamma_gamma1plus2 = sps.digamma(np.sum(self.varpar_stick, axis=1))
log_varpar_assignment = (
np.append(
sps.digamma(self.varpar_stick[:, 0]) - digamma_gamma1plus2,
0)[np.newaxis, :] + np.append(
0,
np.cumsum(
sps.digamma(self.varpar_stick[:, 1]) -
digamma_gamma1plus2))[np.newaxis, :] +
np.matmul(
np.exp(children_label_np, ),
np.transpose(
np.matmul(np.exp(self.log_varpar_label), error_mat))))
# denom=np.sum(self.log_varpar_assignment, axis=-1)[:,np.newaxis]
# self.log_varpar_assignment /= denom+(denom==0).astype(np.float64)
# self.log_varpar_assignment = self.log_varpar_assignment.reshape(self.num_children, self.num_tables_child, self.T)
# self.log_varpar_assignment-=spm.logsumexp(self.log_varpar_assignment, axis=-1)[:,np.newaxis]
log_varpar_assignment -= spm.logsumexp(log_varpar_assignment,
axis=-1)[:, np.newaxis]
self.log_varpar_assignment = (log_varpar_assignment).reshape(
self.num_children, self.num_tables_child, self.T)
self.phi_x_tau_x_error = np.matmul(
np.exp(lla.logmatmul(log_varpar_assignment,
self.log_varpar_label)),
error_mat).reshape(self.num_children, self.num_tables_child,
num_symbols)
def _update_log_varpar_label(self, children_label_np):
log_varpar_label = (
# np.matmul(
# self.mother.log_varpar_assignment[self.id], # Txmother_T
# np.matmul(
# self.mother.log_varpar_label, # mother_T x |∑|
# error_mat # |∑|x|∑|
# )
# )
self.mother.phi_x_tau_x_error[self.id] + np.matmul(
np.exp(
lla.logmatmul(
np.transpose(
self.log_varpar_assignment.reshape(
self.num_children * self.num_tables_child,
self.T)), # Tx(num_children*num_tables_child)
children_label_np # (num_children*num_tables_child)x|∑|
)),
error_mat # |∑|x|∑|
))
# denom=np.sum(self.log_varpar_label, axis=-1)[:, np.newaxis]
# self.log_varpar_label /= denom+(denom==0).astype(np.float64)
self.log_varpar_label = (
log_varpar_label -
spm.logsumexp(log_varpar_label, axis=-1)[:, np.newaxis])
def set_log_varpar_assignment(self):
self.log_varpar_assignment = np.log(
np.random.dirichlet(np.ones(self.T), len(
self.customers))) # phi in Blei and Jordan.
self.phi_x_tau_x_error = np.matmul(
np.exp(
lla.logmatmul(self.log_varpar_assignment,
self.log_varpar_label)), error_mat)
if not self.log_varpar_assignment.size:
self.log_varpar_label = np.zeros(self.log_varpar_label.shape)
def update_varpars(self):
if self.log_varpar_assignment.size:
self._update_varpar_stick()
assert not np.any(np.isnan(self.varpar_stick)), ('stick\n',
self.varpar_stick)
self._update_log_varpar_label()
assert not np.any(np.isnan(
self.log_varpar_label)), ('label\n', self.log_varpar_label)
self._update_log_varpar_assignment()
assert not np.any(np.isnan(
self.log_varpar_assignment)), ('assignment\n',
self.log_varpar_assignment)
self.phi_x_tau_x_error = np.matmul(
np.exp(
lla.logmatmul(self.log_varpar_assignment,
self.log_varpar_label)), error_mat)
assert not np.any(np.isnan(
self.phi_x_tau_x_error)), ('phi_x_tau_x_error\n',
self.phi_x_tau_x_error)
def set_log_varpar_assignment(self):
# self.children = children # Represented by a list of children restaurants.
# self.children_label = [child.log_varpar_label for child in self.children]
self.num_children = len(self.children)
self.num_tables_child = self.children[0].log_varpar_label.shape[
0] # T for children.
self.log_varpar_assignment = np.log(
np.random.dirichlet(np.ones(
self.T), (self.num_children,
self.num_tables_child))) # phi in Blei and Jordan.
# self.tau_x_error = np.matmul(
# self.log_varpar_label,
# error_mat
# )
self.phi_x_tau_x_error = np.matmul(
np.exp(
lla.logmatmul(
self.log_varpar_assignment.reshape(
self.num_children * self.num_tables_child, self.T),
self.log_varpar_label)),
error_mat).reshape(self.num_children, self.num_tables_child,
num_symbols)