def convert_for_exact_inference(g, p):
"""
Converts the graph and parameters to a BBN.
:param g: Directed acyclic graph (DAG in the form of networkx).
:param p: Parameters.
:return: BBN.
"""
bbn = Bbn()
bbn_nodes = {}
for node in g.nodes:
id = node
params = p[id]['params'].flatten()
states = [
'state{}'.format(state) for state in range(p[id]['shape'][1])
]
v = Variable(id, str(id), states)
n = BbnNode(v, params)
bbn.add_node(n)
bbn_nodes[id] = n
for e in g.edges:
pa = bbn_nodes[e[0]]
ch = bbn_nodes[e[1]]
bbn.add_edge(Edge(pa, ch, EdgeType.DIRECTED))
return bbn
def build_bbn(variable_profiles, g, p):
"""
Builds a BBN from a DAG, g, and paremeters, p.
:param variable_profiles: Variable profiles.
:param g: DAG.
:param p: Parameters.
:return: BBN.
"""
bbn = Bbn()
nodes = list(g.nodes)
bbn_node_dict = {}
for idx in nodes:
name = g.nodes[idx]['name']
domain = variable_profiles[name]
cpt = p[idx]
v = Variable(idx, name, domain)
n = BbnNode(v, cpt)
bbn.add_node(n)
bbn_node_dict[idx] = n
edges = list(g.edges)
for edge in edges:
pa = bbn_node_dict[edge[0]]
ch = bbn_node_dict[edge[1]]
e = Edge(pa, ch, EdgeType.DIRECTED)
bbn.add_edge(e)
return bbn
def from_libpgm_discrete_object(bn):
"""
Converts a libpgm discrete network object into a py-bbn one.
:param bn: libpgm discrete BBN.
:return: py-bbn BBN.
"""
def get_nodes(bn, domain_spaces=True):
def get_parent_domains(name, bn):
parents = bn.Vdata[name]['parents']
domains = []
if parents is None or len(parents) == 0:
return domains
for parent in parents:
domain = bn.Vdata[parent]['vals'][:]
domains.append(domain)
return domains
def cross_product(domains):
products = []
if domains is None or len(domains) == 0:
return products
for e in itertools.product(*domains):
products.append(e)
return products
def stringify_cross_product(pa_domains, domain_spaces=True):
joiner_delim = ', ' if domain_spaces is True else ','
s = []
for pa_domain in pa_domains:
r = joiner_delim.join(
["'{}'".format(v) for v in pa_domain])
r = '[{}]'.format(r)
s.append(r)
return s
def get_cond_probs(name, bn, domain_spaces=True):
probs = []
pa_domains = stringify_cross_product(
cross_product(get_parent_domains(name, bn)), domain_spaces)
if len(pa_domains) == 0:
probs = bn.Vdata[name]['cprob'][:]
else:
for pa_domain in pa_domains:
cprob = bn.Vdata[name]['cprob'][pa_domain]
probs.extend(cprob)
return probs
nodes = {}
for name in bn.V:
domain = bn.Vdata[name]['vals'][:]
order = bn.Vdata[name]['ord']
probs = get_cond_probs(name, bn, domain_spaces)
node = BbnNode(Variable(order, name, domain), probs)
nodes[name] = node
return nodes
def get_edges(bn, nodes):
edges = []
for k, v in bn.Vdata.items():
ch = nodes[k]
if v['parents'] is not None and len(v['parents']) > 0:
parents = [nodes[pa] for pa in v['parents']]
for pa in parents:
edge = Edge(pa, ch, EdgeType.DIRECTED)
edges.append(edge)
return edges
nodes = get_nodes(bn)
edges = get_edges(bn, nodes)
bbn = Bbn()
for node in sorted(nodes.values(), key=lambda n: n.id):
bbn.add_node(node)
for e in edges:
bbn.add_edge(e)
return bbn
def from_data(structure, df):
"""
Creates a BBN.
:param structure: A dictionary where keys are names of children and values are list of parent names.
:param df: A dataframe.
:return: BBN.
"""
def get_profile(df):
profile = {}
for c in df.columns:
values = sorted(list(df[c].value_counts().index))
profile[c] = values
return profile
def get_n2i(parents):
g = nx.DiGraph()
for k in parents:
g.add_node(k)
for ch, pas in parents.items():
for pa in pas:
g.add_edge(pa, ch)
nodes = list(topological_sort(g))
return {n: i for i, n in enumerate(nodes)}
def get_cpt(name, parents, n2v, df):
parents = sorted(parents)
n2v = {k: sorted(v) for k, v in n2v.items()}
n = df.shape[0]
cpts = []
if len(parents) == 0:
for v in n2v[name]:
c = df[df[name] == v].shape[0]
p = c / n
cpts.append(p)
else:
domains = [(n, d) for n, d in n2v.items() if n in parents]
domains = sorted(domains, key=lambda tup: tup[0])
domain_names = [tup[0] for tup in domains]
domain_values = [tup[1] for tup in domains]
domains = list(product(*domain_values))
for values in domains:
probs = []
denom_q = ' and '.join(
[f'{n}=="{v}"' for n, v in zip(domain_names, values)])
for v in n2v[name]:
numer_q = f'{name}=="{v}" and {denom_q}'
numer = df.query(numer_q).shape[0] / n
denom = df.query(denom_q).shape[0] / n
prob = numer / denom
probs.append(prob)
probs = pd.Series(probs)
probs = probs / probs.sum()
probs = list(probs)
cpts.extend(probs)
return cpts
n2v = get_profile(df)
n2i = get_n2i(df)
n2c = {n: get_cpt(n, structure[n], n2v, df) for n in structure}
bbn = Bbn()
nodes = {}
for name in n2v:
idx = n2i[name]
values = n2v[name]
cpts = n2c[name]
v = Variable(idx, name, values)
node = BbnNode(v, cpts)
nodes[name] = node
bbn.add_node(node)
for ch, parents in structure.items():
ch_node = nodes[ch]
for pa in parents:
pa_node = nodes[pa]
edge = Edge(pa_node, ch_node, EdgeType.DIRECTED)
bbn.add_edge(edge)
return bbn
