class BayesNetStructure(Space, PreBayesSpace):
"""Space for Bayesian network structure search.
Built on top of a variable space that determines the variables of the
Bayes net. For example, a Euclidean space, say $\mathbb{R}^d$, would have
$d$ :class:`GaussianVariable`s. Or, a Binary space would have
:class:`BinaryVariable`s.
The default choice of variable is controlled by the ``_mapping`` class
variable based on the ``type`` field of the variable space. To have mixed
networks, use a :class:`Product` space.
:param space: The space of variables
:type space: :class:`Space`
"""
_mappings = { # map space types to bayesian variables
np.ndarray: GaussianVariable,
TernaryString: BinaryVariable,
}
def __init__(self, space):
super(BayesNetStructure, self).__init__(BayesNet)
self.space = space
self.sampler = DAGSampler()
if hasattr(space, 'dim'):
self.numVariables = space.dim
else:
raise ValueError("Cannot determine the number of dimensions "
"for BayesNetStructure space based on given "
"space; expected space to have property dim "
"indicating the number of required variables. ")
self.config = Config(numVariables=self.numVariables,
variableGenerator=self.variable,
randomizer=self.randomize,
sampler=self.sample)
self.proposal = StructureProposal(**self.config.__properties__)
self.config.structureGenerator = self.proposal
self.edges = []
self.nedges = []
def in_bounds(self, network, **kwargs):
c = network.config
return (isinstance(network, BayesNet)
and len(network.variables) == self.numVariables and np.all([
v.__class__ == self.variable(v.index, c).__class__
for v in network.variables
]))
def proportion(self, smaller, larger, index):
return 2.0**((len(smaller.edges) + len(smaller.nedges)) -
(len(larger.edges) + len(larger.nedges)))
def area(self, **kwargs):
return 1.0
def random(self):
network = BayesNet(**self.config.__properties__)
for i in xrange(5):
self.proposal.search(network)
return network
def extent(self):
return self.nedges, self.edges
#raise NotImplementedError("Bayesian networks do not have a "
# "well-defined notion of spatial extent; "
# "If you have something in mind, please "
# "subclass the space or submit a pull "
# "request.")
def hash(self, point):
point.computeEdgeStatistics()
return str(point.edges)
class BayesNetStructure(Space, PreBayesSpace):
"""Space for Bayesian network structure search.
Built on top of a variable space that determines the variables of the
Bayes net. For example, a Euclidean space, say $\mathbb{R}^d$, would have
$d$ :class:`GaussianVariable`s. Or, a Binary space would have
:class:`BinaryVariable`s.
The default choice of variable is controlled by the ``_mapping`` class
variable based on the ``type`` field of the variable space. To have mixed
networks, use a :class:`Product` space.
:param space: The space of variables
:type space: :class:`Space`
"""
_mappings = { # map space types to bayesian variables
np.ndarray: GaussianVariable,
TernaryString: BinaryVariable,
}
def __init__(self, space):
super(BayesNetStructure, self).__init__(BayesNet)
self.space = space
self.sampler = DAGSampler()
if hasattr(space, 'dim'):
self.numVariables = space.dim
else:
raise ValueError("Cannot determine the number of dimensions "
"for BayesNetStructure space based on given "
"space; expected space to have property dim "
"indicating the number of required variables. ")
self.config = Config(numVariables=self.numVariables,
variableGenerator=self.variable,
randomizer=self.randomize,
sampler=self.sample)
self.proposal = StructureProposal(**self.config.__properties__)
self.config.structureGenerator = self.proposal
self.edges = []
self.nedges = []
def in_bounds(self, network, **kwargs):
c = network.config
return (isinstance(network, BayesNet) and
len(network.variables) == self.numVariables and
np.all([v.__class__ == self.variable(v.index,c).__class__
for v in network.variables]))
def proportion(self, smaller, larger, index):
return 2.0 ** ((len(smaller.edges) + len(smaller.nedges)) -
(len(larger.edges) + len(larger.nedges)))
def area(self, **kwargs):
return 1.0
def random(self):
network = BayesNet(**self.config.__properties__)
for i in xrange(5):
self.proposal.search(network)
return network
def extent(self):
return self.nedges, self.edges
#raise NotImplementedError("Bayesian networks do not have a "
# "well-defined notion of spatial extent; "
# "If you have something in mind, please "
# "subclass the space or submit a pull "
# "request.")
def hash(self, point):
point.computeEdgeStatistics()
return str(point.edges)
