def __init__(self, A, X, **kwargs):
# Convert arrays to constant nodes
if utils.is_numeric(A):
A = Constant(Gaussian)(A)
if utils.is_numeric(X):
X = Constant(Gaussian)(X)
MN = A.dims[0][0]
N = X.dims[0][0]
# Check parents
if A.dims != ( (MN,), (MN, MN) ):
raise ValueError("Invalid dimensionality of the first parent.")
if X.dims != ( (N,), (N, N) ):
raise ValueError("Invalid dimensionality of the second parent.")
if (MN % N) != 0:
raise ValueError("The dimensionality of the first parent should be "
"a multiple of the second parent.")
# Dimensionality of the output
M = int(MN / N)
dims = ( (M,), (M,M) )
Node.__init__(self, A, X, dims=dims, **kwargs)
def __init__(self, A, X, **kwargs):
# Convert arrays to constant nodes
if utils.is_numeric(A):
A = Constant(Gaussian)(A)
if utils.is_numeric(X):
X = Constant(Gaussian)(X)
MN = A.dims[0][0]
N = X.dims[0][0]
# Check parents
if A.dims != ((MN, ), (MN, MN)):
raise ValueError("Invalid dimensionality of the first parent.")
if X.dims != ((N, ), (N, N)):
raise ValueError("Invalid dimensionality of the second parent.")
if (MN % N) != 0:
raise ValueError(
"The dimensionality of the first parent should be "
"a multiple of the second parent.")
# Dimensionality of the output
M = int(MN / N)
dims = ((M, ), (M, M))
Node.__init__(self, A, X, dims=dims, **kwargs)
def __init__(self, mu, Lambda, A, v, n=None, **kwargs):
"""
`mu` is the mean of x_0
`Lambda` is the precision of x_0
`A` is the dynamic matrix
`v` is the diagonal precision of the innovation
"""
self.parameter_distributions = (Gaussian, Wishart, Gaussian, Gamma)
# Check for constant mu
if utils.is_numeric(mu):
mu = Constant(Gaussian)(np.atleast_1d(mu))
# Check for constant Lambda
if utils.is_numeric(Lambda):
Lambda = Constant(Wishart)(np.atleast_2d(Lambda))
# Check for constant A
if utils.is_numeric(A):
A = Constant(Gaussian)(np.atleast_2d(A))
# Check for constant V
if utils.is_numeric(v):
v = Constant(Gamma)(np.atleast_1d(v))
# A dummy wrapper for the number of time instances.
n_A = 1
if len(A.plates) >= 2:
n_A = A.plates[-2]
n_v = 1
if len(v.plates) >= 2:
n_v = v.plates[-2]
if n_v != n_A and n_v != 1 and n_A != 1:
raise Exception("Plates of A and v are giving different number of time instances")
n_A = max(n_v, n_A)
if n is None:
if n_A == 1:
raise Exception("""The number of time instances could not be determined
automatically. Give the number of
time instances.""")
n = n_A + 1
elif n_A != 1 and n_A+1 != n:
raise Exception("The number of time instances must match "
"the number of last plates of parents: "
"%d != %d+1" % (n, n_A))
# Construct
super().__init__(mu, Lambda, A, v, n=n, **kwargs)
def __init__(self, *parents, **kwargs):
# For now, do not use plates other than from the parents,
# although it would be possible (it would mean that you'd
# create several "datasets" with identical "PCA
# distribution"). Maybe it is better to create such plates in
# children nodes?
#plates = []
parents = list(parents)
# Convert constant arrays to constant nodes
for n in range(len(parents)):
if utils.is_numeric(parents[n]):
parents[n] = Constant(Gaussian)(parents[n])
parent_dims = [parent.dims for parent in parents]
parent_plates = [parent.plates for parent in parents]
# Check that the parents have equal dimensions
for n in range(len(parent_dims) - 1):
if parent_dims[n] != parent_dims[n + 1]:
raise ValueError("Dimensions of the Gaussians do not "
"match: %s" % (parent_dims, ))
## try:
## plates = utils.broadcasted_shape(*parent_plates)
## except ValueError:
## raise ValueError("The plates of the parents are "
## "incompatible: %s" % (parent_plates,))
super().__init__(
*parents,
#plates=tuple(plates),
dims=((), ()),
**kwargs)
def __init__(self, *parents, **kwargs):
# For now, do not use plates other than from the parents,
# although it would be possible (it would mean that you'd
# create several "datasets" with identical "PCA
# distribution"). Maybe it is better to create such plates in
# children nodes?
#plates = []
parents = list(parents)
# Convert constant arrays to constant nodes
for n in range(len(parents)):
if utils.is_numeric(parents[n]):
parents[n] = Constant(Gaussian)(parents[n])
parent_dims = [parent.dims for parent in parents]
parent_plates = [parent.plates for parent in parents]
# Check that the parents have equal dimensions
for n in range(len(parent_dims)-1):
if parent_dims[n] != parent_dims[n+1]:
raise ValueError("Dimensions of the Gaussians do not "
"match: %s" % (parent_dims,))
## try:
## plates = utils.broadcasted_shape(*parent_plates)
## except ValueError:
## raise ValueError("The plates of the parents are "
## "incompatible: %s" % (parent_plates,))
super().__init__(*parents,
#plates=tuple(plates),
dims=((),()),
**kwargs)
def __init__(self, a, b, **kwargs):
self.parameter_distributions = (GammaPrior, Gamma)
# TODO: USE asarray(a)
# Check for constant a
if utils.is_numeric(a):
a = Constant(GammaPrior)(a)
# Check for constant b
if utils.is_numeric(b):
b = Constant(Gamma)(b)
#b = NodeConstant([b, np.log(b)], plates=np.shape(b), dims=[(),()])
# Construct
super().__init__(a, b, **kwargs)
def __init__(self, a, b, **kwargs):
self.parameter_distributions = (GammaPrior, Gamma)
# TODO: USE asarray(a)
# Check for constant a
if utils.is_numeric(a):
a = Constant(GammaPrior)(a)
# Check for constant b
if utils.is_numeric(b):
b = Constant(Gamma)(b)
#b = NodeConstant([b, np.log(b)], plates=np.shape(b), dims=[(),()])
# Construct
super().__init__(a, b, **kwargs)
def __init__(self, X, **kwargs):
# Check for constant n
if utils.is_numeric(X):
X = Constant(GaussianMarkovChain)(X)
# Make the time dimension a plate dimension...
#plates = X.plates + (X.dims[0][0],)
# ... and remove it from the variable dimensions
dims = ( X.dims[0][-1:], X.dims[1][-2:] )
super().__init__(X,
#plates=plates,
dims=dims,
**kwargs)
def __init__(self, alpha, **kwargs):
# Check for constant
if utils.is_numeric(alpha):
alpha = Constant(Gamma)(alpha)
# Remove the last plate...
#plates = alpha.plates[:-1]
# ... and use it as the dimensionality of the Wishart
# distribution
if len(alpha.plates) == 0:
raise Exception("Gamma variable needs to have plates in "
"order to be used as a diagonal Wishart.")
D = alpha.plates[-1]
dims = ( (D,D), () )
# Construct the node
super().__init__(alpha,
#plates=plates,
dims=dims,
**kwargs)
def __init__(self, alpha, **kwargs):
# Check for constant
if utils.is_numeric(alpha):
alpha = Constant(Gamma)(alpha)
# Remove the last plate...
#plates = alpha.plates[:-1]
# ... and use it as the dimensionality of the Wishart
# distribution
if len(alpha.plates) == 0:
raise Exception("Gamma variable needs to have plates in "
"order to be used as a diagonal Wishart.")
D = alpha.plates[-1]
dims = ( (D,D), () )
# Construct the node
super().__init__(alpha,
#plates=plates,
dims=dims,
**kwargs)
def __init__(self, mu, Lambda, B, S, v, n=None, **kwargs):
"""
`mu` is the mean of x_0, (...)x(D)
`Lambda` is the precision of x_0, (...)x(D,D)
`B` is the dynamic matrices, (...,D)x(D,K)
`S` is the temporal weights for the dynamic matrices, (...,N)x(K)
`v` is the diagonal precision of the innovation, (...,D)x()
"""
# TODO/FIXME: Check these
self.parameter_distributions = (Gaussian, Wishart, GaussianArrayARD, Gaussian, Gamma)
# Check for constant mu
if utils.is_numeric(mu):
mu = Constant(Gaussian)(np.atleast_1d(mu))
# Check for constant Lambda
if utils.is_numeric(Lambda):
Lambda = Constant(Wishart)(np.atleast_2d(Lambda))
# Check for constant S
if utils.is_numeric(S):
S = Constant(Gaussian)(np.atleast_2d(S))
# Check for constant B
if utils.is_numeric(B):
D = Lambda.dims[0][-1]
K = S.dims[0][-1]
B = Constant(_GaussianArrayARD((D,K)))(np.atleast_3d(B))
# Check for constant V
if utils.is_numeric(v):
v = Constant(Gamma)(np.atleast_1d(v))
# A dummy wrapper for the number of time instances.
n_S = 1
if len(S.plates) >= 1:
n_S = S.plates[-1]
n_v = 1
if len(v.plates) >= 2:
n_v = v.plates[-2]
if n_v != n_S and n_v != 1 and n_S != 1:
raise Exception(
"Plates of A and v are giving different number of time "
"instances")
n_S = max(n_v, n_S)
if n is None:
if n_S == 1:
raise Exception(
"The number of time instances could not be determined "
"automatically. Give the number of time instances.")
n = n_S + 1
elif n_S != 1 and n_S+1 != n:
raise Exception(
"The number of time instances must match the number of last "
"plates of parents:" "%d != %d+1"
% (n, n_S))
# Construct
super().__init__(mu, Lambda, B, S, v, n=n, **kwargs)
def __init__(self, *args, iterator_axis=None, **kwargs):
"""
SumMultiply(Node1, map1, Node2, map2, ..., NodeN, mapN [, map_out])
"""
args = list(args)
if len(args) < 2:
raise ValueError("Not enough inputs")
if iterator_axis is not None:
raise NotImplementedError("Iterator axis not implemented yet")
if iterator_axis is not None and not isinstance(iterator_axis, int):
raise ValueError("Iterator axis must be integer")
# Two different parsing methods, depends on how the arguments are given
if utils.is_string(args[0]):
# This is the format:
# SumMultiply('ik,k,kj->ij', X, Y, Z)
strings = args[0]
nodes = args[1:]
# Remove whitespace
strings = utils.remove_whitespace(strings)
# Split on '->' (should contain only one '->' or none)
strings = strings.split('->')
if len(strings) > 2:
raise ValueError('The string contains too many ->')
strings_in = strings[0]
if len(strings) == 2:
string_out = strings[1]
else:
string_out = ''
# Split former part on ',' (the number of parts should be equal to
# nodes)
strings_in = strings_in.split(',')
if len(strings_in) != len(nodes):
raise ValueError('Number of given input nodes is different '
'from the input keys in the string')
# Split strings into key lists using single character keys
keysets = [list(string_in) for string_in in strings_in]
keys_out = list(string_out)
else:
# This is the format:
# SumMultiply(X, [0,2], Y, [2], Z, [2,1], [0,1])
# If given, the output mapping is the last argument
if len(args) % 2 == 0:
keys_out = []
else:
keys_out = args.pop(-1)
# Node and axis mapping are given in turns
nodes = args[::2]
keysets = args[1::2]
# Find all the keys (store only once each)
full_keyset = []
for keyset in keysets:
full_keyset += keyset
#full_keyset += list(keyset.keys())
full_keyset = list(set(full_keyset))
#
# Check the validity of each node
#
for n in range(len(nodes)):
# Convert constant arrays to constant nodes
if utils.is_numeric(nodes[n]):
# TODO/FIXME: Use GaussianArray and infer the dimensionality
# from the length of the axis mapping!
nodes[n] = Constant(Gaussian)(nodes[n])
# Check that the maps and the size of the variable are consistent
if len(nodes[n].dims[0]) != len(keysets[n]):
raise ValueError("Wrong number of keys (%d) for the node "
"number %d with %d dimensions"
% (len(keysets[n]),
n,
len(nodes[n].dims[0])))
# Check that the keys are unique
if len(set(keysets[n])) != len(keysets[n]):
raise ValueError("Axis keys for node number %d are not unique"
% n)
# Check that the dims are proper Gaussians
if len(nodes[n].dims) != 2:
raise ValueError("Node %d is not Gaussian" % n)
if nodes[n].dims[0] + nodes[n].dims[0] != nodes[n].dims[1]:
raise ValueError("Node %d is not Gaussian" % n)
# Check the validity of output keys: each output key must be included in
# the input keys
if len(keys_out) != len(set(keys_out)):
raise ValueError("Output keys are not unique")
for key in keys_out:
if key not in full_keyset:
raise ValueError("Output key %s does not appear in any input"
% key)
# Check the validity of the nodes with respect to the key mapping.
# Check that the node dimensions map and broadcast properly, that is,
# all the nodes using the same key for axes must have equal size for
# those axes (or size 1).
broadcasted_size = {}
for key in full_keyset:
broadcasted_size[key] = 1
for (node, keyset) in zip(nodes, keysets):
try:
# Find the axis for the key
index = keyset.index(key)
except ValueError:
# OK, this node doesn't use this key for any axis
pass
else:
# Length of the axis for that key
node_size = node.dims[0][index]
if node_size != broadcasted_size[key]:
if broadcasted_size[key] == 1:
# Apply broadcasting
broadcasted_size[key] = node_size
elif node_size != 1:
# Different sizes and neither has size 1
raise ValueError("Axes using key %s do not "
"broadcast properly"
% key)
# Compute the shape of the output
dim0 = [broadcasted_size[key] for key in keys_out]
dim1 = dim0 + dim0
# Rename the keys to [0,1,...,N-1] where N is the total number of keys
self.N_keys = len(full_keyset)
self.out_keys = [full_keyset.index(key) for key in keys_out]
self.in_keys = [ [full_keyset.index(key) for key in keyset]
for keyset in keysets ]
super().__init__(*nodes,
dims=(tuple(dim0),tuple(dim1)),
**kwargs)