def lower_bound_contribution(self, gradient=False):
# Compute E[ log p(X|parents) - log q(X) ] over q(X)q(parents)
# Messages from parents
#u_parents = [parent.message_to_child() for parent in self.parents]
u_parents = self._message_from_parents()
phi = self._distribution.compute_phi_from_parents(*u_parents)
# G from parents
L = self._distribution.compute_cgf_from_parents(*u_parents)
# L = g
# G for unobserved variables (ignored variables are handled
# properly automatically)
latent_mask = np.logical_not(self.observed)
#latent_mask = np.logical_and(self.mask, np.logical_not(self.observed))
# F for observed, G for latent
L = L + np.where(self.observed, self.f, -self.g)
for (phi_p, phi_q, u_q, dims) in zip(phi, self.phi, self.u, self.dims):
# Form a mask which puts observed variables to zero and
# broadcasts properly
latent_mask_i = utils.add_trailing_axes(
utils.add_leading_axes(latent_mask,
len(self.plates) -
np.ndim(latent_mask)), len(dims))
axis_sum = tuple(range(-len(dims), 0))
# Compute the term
phi_q = np.where(latent_mask_i, phi_q, 0)
# TODO/FIXME: Use einsum here?
Z = np.sum((phi_p - phi_q) * u_q, axis=axis_sum)
L = L + Z
return (np.sum(np.where(self.mask, L, 0)) * self._plate_multiplier(
self.plates, np.shape(L), np.shape(self.mask)))
def update_u(self, u, mask=True):
# Store the computed moments u but do not change moments for
# observations, i.e., utilize the mask.
for ind in range(len(u)):
# Add axes to the mask for the variable dimensions (mask
# contains only axes for the plates).
u_mask = utils.add_trailing_axes(mask, self.ndims[ind])
# Enlarge self.u[ind] as necessary so that it can store the
# broadcasted result.
sh = utils.broadcasted_shape_from_arrays(self.u[ind], u[ind],
u_mask)
self.u[ind] = utils.repeat_to_shape(self.u[ind], sh)
# Use mask to update only unobserved plates and keep the
# observed as before
np.copyto(self.u[ind], u[ind], where=u_mask)
# Make sure u has the correct number of dimensions:
shape = self.get_shape(ind)
ndim = len(shape)
ndim_u = np.ndim(self.u[ind])
if ndim > ndim_u:
self.u[ind] = utils.add_leading_axes(u[ind], ndim - ndim_u)
elif ndim < np.ndim(self.u[ind]):
raise Exception("Weird, this shouldn't happen.. :)")
def _update_phi_from_parents(self, *u_parents):
# TODO/FIXME: Could this be combined to the function
# _update_distribution_and_lowerbound ?
# No, because some initialization methods may want to use this.
# This makes correct broadcasting
self.phi = self._distribution.compute_phi_from_parents(*u_parents)
#self.phi = self._compute_phi_from_parents(*u_parents)
self.phi = list(self.phi)
# Make sure phi has the correct number of axes. It makes life
# a bit easier elsewhere.
for i in range(len(self.phi)):
axes = len(self.plates) + self.ndims[i] - np.ndim(self.phi[i])
if axes > 0:
# Add axes
self.phi[i] = utils.add_leading_axes(self.phi[i], axes)
elif axes < 0:
# Remove extra leading axes
first = -(len(self.plates)+self.ndims[i])
sh = np.shape(self.phi[i])[first:]
self.phi[i] = np.reshape(self.phi[i], sh)
# Check that the shape is correct
if not utils.is_shape_subset(np.shape(self.phi[i]),
self.get_shape(i)):
raise ValueError("Incorrect shape of phi[%d] in node class %s. "
"Shape is %s but it should be broadcastable "
"to shape %s."
% (i,
self.__class__.__name__,
np.shape(self.phi[i]),
self.get_shape(i)))
def _update_phi_from_parents(self, *u_parents):
# TODO/FIXME: Could this be combined to the function
# _update_distribution_and_lowerbound ?
# No, because some initialization methods may want to use this.
# This makes correct broadcasting
self.phi = self._distribution.compute_phi_from_parents(*u_parents)
#self.phi = self._compute_phi_from_parents(*u_parents)
self.phi = list(self.phi)
# Make sure phi has the correct number of axes. It makes life
# a bit easier elsewhere.
for i in range(len(self.phi)):
axes = len(self.plates) + self._distribution.ndims[i] - np.ndim(
self.phi[i])
if axes > 0:
# Add axes
self.phi[i] = utils.add_leading_axes(self.phi[i], axes)
elif axes < 0:
# Remove extra leading axes
first = -(len(self.plates) + self._distribution.ndims[i])
sh = np.shape(self.phi[i])[first:]
self.phi[i] = np.reshape(self.phi[i], sh)
# Check that the shape is correct
if not utils.is_shape_subset(np.shape(self.phi[i]),
self.get_shape(i)):
raise ValueError("Incorrect shape in phi[%d]. Shape is %s but "
"it should be broadcastable to shape %s." %
(i, np.shape(self.phi[i]), self.get_shape(i)))
def update_u(self, u, mask=True):
# Store the computed moments u but do not change moments for
# observations, i.e., utilize the mask.
for ind in range(len(u)):
# Add axes to the mask for the variable dimensions (mask
# contains only axes for the plates).
u_mask = utils.add_trailing_axes(mask, self.ndims[ind])
# Enlarge self.u[ind] as necessary so that it can store the
# broadcasted result.
sh = utils.broadcasted_shape_from_arrays(self.u[ind], u[ind], u_mask)
self.u[ind] = utils.repeat_to_shape(self.u[ind], sh)
# Use mask to update only unobserved plates and keep the
# observed as before
np.copyto(self.u[ind],
u[ind],
where=u_mask)
# Make sure u has the correct number of dimensions:
shape = self.get_shape(ind)
ndim = len(shape)
ndim_u = np.ndim(self.u[ind])
if ndim > ndim_u:
self.u[ind] = utils.add_leading_axes(u[ind], ndim - ndim_u)
elif ndim < np.ndim(self.u[ind]):
raise Exception("Weird, this shouldn't happen.. :)")
def update_phi_from_parents(self, u_parents):
# This makes correct broadcasting
self.phi = self.compute_phi_from_parents(u_parents)
self.phi = list(self.phi)
# Make sure phi has the correct number of axes. It makes life
# a bit easier elsewhere.
for i in range(len(self.phi)):
axes = len(self.plates) + self.ndims[i] - np.ndim(self.phi[i])
if axes > 0:
# Add axes
self.phi[i] = utils.add_leading_axes(self.phi[i], axes)
elif axes < 0:
# Remove extra leading axes
first = -(len(self.plates)+self.ndims[i])
sh = np.shape(self.phi[i])[first:]
self.phi[i] = np.reshape(self.phi[i], sh)
def _set_moments(self, u, mask=True):
# Store the computed moments u but do not change moments for
# observations, i.e., utilize the mask.
for ind in range(len(u)):
# Add axes to the mask for the variable dimensions (mask
# contains only axes for the plates).
u_mask = utils.add_trailing_axes(mask, self._distribution.ndims[ind])
# Enlarge self.u[ind] as necessary so that it can store the
# broadcasted result.
sh = utils.broadcasted_shape_from_arrays(self.u[ind], u[ind], u_mask)
self.u[ind] = utils.repeat_to_shape(self.u[ind], sh)
# TODO/FIXME/BUG: The mask of observations is not used, observations
# may be overwritten!!! ???
# Hah, this function is used to set the observations! The caller
# should be careful what mask he uses! If you want to set only
# latent variables, then use such a mask.
# Use mask to update only unobserved plates and keep the
# observed as before
np.copyto(self.u[ind],
u[ind],
where=u_mask)
# Make sure u has the correct number of dimensions:
# TODO/FIXME: Maybe it would be good to also check that u has a
# shape that is a sub-shape of get_shape.
shape = self.get_shape(ind)
ndim = len(shape)
ndim_u = np.ndim(self.u[ind])
if ndim > ndim_u:
self.u[ind] = utils.add_leading_axes(u[ind], ndim - ndim_u)
elif ndim < ndim_u:
raise RuntimeError(
"The size of the variable %s's %s-th moment "
"array is %s which is larger than it should "
"be, that is, %s, based on the plates %s and "
"dimension %s. Check that you have provided "
"plates properly."
% (self.name,
ind,
np.shape(self.u[ind]),
shape,
self.plates,
self.dims[ind]))
def _set_moments(self, u, mask=True):
# Store the computed moments u but do not change moments for
# observations, i.e., utilize the mask.
for ind in range(len(u)):
# Add axes to the mask for the variable dimensions (mask
# contains only axes for the plates).
u_mask = utils.add_trailing_axes(mask,
self._distribution.ndims[ind])
# Enlarge self.u[ind] as necessary so that it can store the
# broadcasted result.
sh = utils.broadcasted_shape_from_arrays(self.u[ind], u[ind],
u_mask)
self.u[ind] = utils.repeat_to_shape(self.u[ind], sh)
# TODO/FIXME/BUG: The mask of observations is not used, observations
# may be overwritten!!! ???
# Hah, this function is used to set the observations! The caller
# should be careful what mask he uses! If you want to set only
# latent variables, then use such a mask.
# Use mask to update only unobserved plates and keep the
# observed as before
np.copyto(self.u[ind], u[ind], where=u_mask)
# Make sure u has the correct number of dimensions:
# TODO/FIXME: Maybe it would be good to also check that u has a
# shape that is a sub-shape of get_shape.
shape = self.get_shape(ind)
ndim = len(shape)
ndim_u = np.ndim(self.u[ind])
if ndim > ndim_u:
self.u[ind] = utils.add_leading_axes(u[ind], ndim - ndim_u)
elif ndim < ndim_u:
raise RuntimeError(
"The size of the variable %s's %s-th moment "
"array is %s which is larger than it should "
"be, that is, %s, based on the plates %s and "
"dimension %s. Check that you have provided "
"plates properly." % (self.name, ind, np.shape(
self.u[ind]), shape, self.plates, self.dims[ind]))
def _update_phi_from_parents(self, *u_parents):
# TODO/FIXME: Could this be combined to the function
# _update_distribution_and_lowerbound ?
# No, because some initialization methods may want to use this.
# This makes correct broadcasting
self.phi = self._compute_phi_from_parents(*u_parents)
self.phi = list(self.phi)
# Make sure phi has the correct number of axes. It makes life
# a bit easier elsewhere.
for i in range(len(self.phi)):
axes = len(self.plates) + self.ndims[i] - np.ndim(self.phi[i])
if axes > 0:
# Add axes
self.phi[i] = utils.add_leading_axes(self.phi[i], axes)
elif axes < 0:
# Remove extra leading axes
first = -(len(self.plates) + self.ndims[i])
sh = np.shape(self.phi[i])[first:]
self.phi[i] = np.reshape(self.phi[i], sh)
def _update_phi_from_parents(self, *u_parents):
# TODO/FIXME: Could this be combined to the function
# _update_distribution_and_lowerbound ?
# No, because some initialization methods may want to use this.
# This makes correct broadcasting
self.phi = self._compute_phi_from_parents(*u_parents)
self.phi = list(self.phi)
# Make sure phi has the correct number of axes. It makes life
# a bit easier elsewhere.
for i in range(len(self.phi)):
axes = len(self.plates) + self.ndims[i] - np.ndim(self.phi[i])
if axes > 0:
# Add axes
self.phi[i] = utils.add_leading_axes(self.phi[i], axes)
elif axes < 0:
# Remove extra leading axes
first = -(len(self.plates)+self.ndims[i])
sh = np.shape(self.phi[i])[first:]
self.phi[i] = np.reshape(self.phi[i], sh)
def lower_bound_contribution(self, gradient=False):
# Compute E[ log p(X|parents) - log q(X) ] over q(X)q(parents)
# Messages from parents
#u_parents = [parent.message_to_child() for parent in self.parents]
u_parents = self._message_from_parents()
phi = self._compute_phi_from_parents(*u_parents)
# G from parents
L = self._compute_cgf_from_parents(*u_parents)
# L = g
# G for unobserved variables (ignored variables are handled
# properly automatically)
latent_mask = np.logical_not(self.observed)
#latent_mask = np.logical_and(self.mask, np.logical_not(self.observed))
# F for observed, G for latent
L = L + np.where(self.observed, self.f, -self.g)
for (phi_p, phi_q, u_q, dims) in zip(phi, self.phi, self.u, self.dims):
# Form a mask which puts observed variables to zero and
# broadcasts properly
latent_mask_i = utils.add_trailing_axes(
utils.add_leading_axes(
latent_mask,
len(self.plates) - np.ndim(latent_mask)),
len(dims))
axis_sum = tuple(range(-len(dims),0))
# Compute the term
phi_q = np.where(latent_mask_i, phi_q, 0)
# TODO/FIXME: Use einsum here?
Z = np.sum((phi_p-phi_q) * u_q, axis=axis_sum)
L = L + Z
return (np.sum(np.where(self.mask, L, 0))
* self._plate_multiplier(self.plates,
np.shape(L),
np.shape(self.mask)))