def learning(sample, method, parameters):
if method == "cpc":
binNumber, alpha = parameters
learner = otagr.ContinuousPC(sample, binNumber, alpha)
ndag = learner.learnDAG()
TTest = otagr.ContinuousTTest(sample, alpha)
jointDistributions = []
for i in range(ndag.getSize()):
d = 1 + ndag.getParents(i).getSize()
if d == 1:
bernsteinCopula = ot.Uniform(0.0, 1.0)
else:
K = TTest.GetK(len(sample), d)
indices = [int(n) for n in ndag.getParents(i)]
indices = [i] + indices
bernsteinCopula = ot.EmpiricalBernsteinCopula(
sample.getMarginal(indices), K, False)
jointDistributions.append(bernsteinCopula)
bn = named_dag_to_bn(ndag)
elif method == "elidan":
#print(sample.getDescription())
max_parents, n_restart_hc = parameters
copula, dag = hc.hill_climbing(sample, max_parents, n_restart_hc)[0:2]
#bn = dag_to_bn(dag, Tstruct.names())
bn = dag_to_bn(dag, sample.getDescription())
else:
print("Wrong entry for method argument !")
return bn
def learning(self, sample):
if self.method == "cpc":
learner = otagr.ContinuousPC(sample, self.parameters['binNumber'],
self.parameters['alpha'])
start = time.time()
ndag = learner.learnDAG()
end = time.time()
# TTest = otagr.ContinuousTTest(sample, self.parameters['alpha'])
# jointDistributions = []
# for i in range(ndag.getSize()):
# d = 1+ndag.getParents(i).getSize()
# if d == 1:
# bernsteinCopula = ot.Uniform(0.0, 1.0)
# else:
# K = TTest.GetK(len(sample), d)
# indices = [int(n) for n in ndag.getParents(i)]
# indices = [i] + indices
# bernsteinCopula = ot.EmpiricalBernsteinCopula(sample.getMarginal(indices), K, False)
# jointDistributions.append(bernsteinCopula)
elif self.method == "cbic":
#print(sample.getDescription())
max_parents = self.parameters['max_parents']
n_restart_hc = self.parameters['hc_restart']
cmode = self.parameters['cmode']
learner = otagr.TabuList(sample, max_parents, n_restart_hc, 5)
learner.setCMode(cmode)
start = time.time()
ndag = learner.learnDAG()
end = time.time()
#bn = dag_to_bn(dag, Tstruct.names())
elif self.method == "cmiic":
cmode = self.parameters['cmode']
kmode = self.parameters['kmode']
learner = otagr.ContinuousMIIC(sample)
learner.setCMode(cmode)
learner.setKMode(kmode)
learner.setAlpha(self.kalpha)
# learner.setBeta(self.kbeta)
start = time.time()
ndag = learner.learnDAG()
end = time.time()
# bn = gu.named_dag_to_bn(ndag)
elif self.method == "dmiic":
# learner.setBeta(self.kbeta)
ndag, start, end = dsc.learnDAG(sample)
# bn = gu.named_dag_to_bn(ndag)
elif self.method == "lgbn":
start = time.time()
end = time.time()
else:
print("Wrong entry for method argument !")
return ndag, end - start
def CBN_PC(data, result_structure_path):
print("CBN with PC")
skeleton_path = result_structure_path.joinpath("skeleton")
skeleton_path.mkdir(parents=True, exist_ok=True)
pdag_path = result_structure_path.joinpath("pdag")
pdag_path.mkdir(parents=True, exist_ok=True)
dag_path = result_structure_path.joinpath("dag")
dag_path.mkdir(parents=True, exist_ok=True)
skeleton_file_name = "skeleton_" + str(size).zfill(7) + ".dot"
skeleton_done = skeleton_path.joinpath(skeleton_file_name).exists()
pdag_file_name = "pdag_" + str(size).zfill(7) + ".dot"
pdag_done = pdag_path.joinpath(pdag_file_name).exists()
dag_file_name = "dag_" + str(size).zfill(7) + ".dot"
dag_done = dag_path.joinpath(dag_file_name).exists()
alpha = 0.01
conditioningSet = 4
learner = otagr.ContinuousPC(data, conditioningSet, alpha)
learner.setVerbosity(True)
if not skeleton_done:
skel = learner.learnSkeleton()
gu.write_graph(
skel,
skeleton_path.joinpath("skeleton_" + str(size).zfill(7) + ".dot"))
if not pdag_done:
pdag = learner.learnPDAG()
gu.write_graph(
pdag, pdag_path.joinpath("pdag_" + str(size).zfill(7) + ".dot"))
if not dag_done:
dag = learner.learnDAG()
gu.write_graph(dag,
dag_path.joinpath("dag_" + str(size).zfill(7) + ".dot"))
else:
dag, names = gu.read_graph(
dag_path.joinpath("dag_" + str(size).zfill(7) + ".dot"))
dag = otagr.NamedDAG(dag, names)
print("Learning parameters")
factories = [
ot.KernelSmoothing(ot.Epanechnikov()),
ot.BernsteinCopulaFactory()
]
ot.Log.SetFile("log")
ot.Log.Show(ot.Log.INFO)
model = otagr.ContinuousBayesianNetworkFactory(factories, dag, alpha,
conditioningSet,
False).build(data)
ot.Log.Show(ot.Log.INFO)
return model
def testAsiaDirichlet():
data = ot.Sample.ImportFromTextFile(
os.path.join(os.path.dirname(__file__), "asia_dirichlet_5000.csv"),
",")
alpha = 0.1
binNumber = 3
learner = otagrum.ContinuousPC(data, binNumber, alpha)
learner.setVerbosity(True)
pdag = learner.learnPDAG()
# print(pdag)
print(learner.PDAGtoDot(pdag))
dag = learner.learnDAG()
print(dag.toDot())
def testSpecificInstance():
size = 1000
data = generateDataForSpecificInstance(size)
alpha = 0.1
binNumber = 3
learner = otagrum.ContinuousPC(data, binNumber, alpha)
# skel = learner.learnSkeleton()
# print(skel.toDot())
dag = learner.learnDAG()
print(dag.toDot())
sys.stdout.flush()
def CPC_learning(data, maxCondSet=5, alpha=0.1):
# Try an estimation of the coefficients distribution using
# univariate kernel smoothing for the marginals and PC to learn the structure
# of dependence parameterized by Bernstein copula
dimension = data.getDimension()
print("Build CPC coefficients distribution")
t0 = time()
print(" Learning structure")
t1 = time()
learner = otagrum.ContinuousPC(data, maxCondSet, alpha)
dag = learner.learnDAG()
with open("dags/new_dag_CPC_{}.dot".format(alpha), "w") as f:
f.write(dag.toDot())
print(" t=", time() - t1, "s")
cbn = CBN_parameter_learning(data, dag)
print("t=", time() - t0, "s")
# distribution = ot.ComposedDistribution(marginals, cbn)
return cbn
def learn(data, alpha=0.1, verbose=False):
with timer("Initiating"):
learn = otagr.ContinuousPC(data, alpha=alpha)
learn.setVerbosity(verbose)
learn.setOptimalPolicy(True)
with timer("Learning skeleton"):
sk = learn.inferSkeleton()
print("Nodes : {} , Edges : {}".format(sk.size(), sk.sizeEdges()))
if sk.size() < 40:
gnb.showDot(learn.skeletonToDot(sk), size="20", format="png")
else:
print(sk.edges())
with timer("Learning VStructures"):
mg = learn.inferPDAG(sk)
if mg.size() < 40:
gnb.showDot(mg.toDot(), size="20", format="png")
else:
print(sk.edges())
with timer("Learning complete DAG"):
dag = learn.learnDAG()
return dag
n_nodes = []
times_cpc = []
times_elidan = []
for i in range(start_size, end_size + 1, step):
print("Number of node :", i, flush=True)
n_nodes.append(i)
n_arc = int(density * (i - 1))
bn = generator.generate(i, n_arc)
TNdag = otagr.NamedDAG(bn.dag(), bn.names())
data = ut.generate_gaussian_data(TNdag, sample_size,
float(args.correlation))
learner = otagr.ContinuousPC(data, mcss, alpha)
start = time.time()
LNdagCPC = learner.learnDAG()
end = time.time()
times_cpc.append(end - start)
start = time.time()
LNdagElidan = hc.hill_climbing(data, max_parents, n_restart_hc)[1]
end = time.time()
times_elidan.append(end - start)
#LNdagCPC = [[ut.named_dag_to_bn(LNdagCPC)]]
#LNdagElidan = [[ut.dag_to_bn(LNdagElidan, data.getDescription())]]
#cpc_scores = ut.structural_scores(ut.named_dag_to_bn(TNdag), LNdagCPC)
#elidan_scores = ut.structural_scores(ut.named_dag_to_bn(TNdag), LNdagElidan)
f = figure_path.joinpath("pairs_ref_" + dataset_name + ".pdf")
pairs(data_draw, f)
#####################LEARNING CBN MODEL############################
# LEARNING STRUCTURE #
# CBIC Algorithm
learner = otagr.TabuList(data_ref, 2, 10, 2) # Using CPC algorithm
cbic_dag = learner.learnDAG() # Learning DAG
write_graph(cbic_dag, structure_path.joinpath("cbic_dag_" + dataset_name + ".dot"))
# CPC Algorithm
learner = otagr.ContinuousPC(data_ref, 4, 0.05) # Using CPC algorithm
cpc_dag = learner.learnDAG() # Learning DAG
write_graph(cpc_dag, structure_path.joinpath("cpc_dag_" + dataset_name + ".dot"))
# CMIIC ALGORITHM
alphas = np.arange(10, 501, 5)/1000
# alphas = [0.04, 0.05]
fig, ax = plt.subplots()
x_major_ticks = np.arange(0, 0.5, 0.05)
x_minor_ticks = np.arange(0, 0.5, 0.01)
y_major_ticks = np.arange(0, 25, 5)
y_minor_ticks = np.arange(0, 25, 1)
from __future__ import print_function from time import time import openturns as ot import pyAgrum as gum import otagrum def generateDataForSpecificInstance(size): R = ot.CorrelationMatrix(3) R[0, 1] = 0.5 R[0, 2] = 0.45 collection = [ot.FrankCopula(3.0), ot.NormalCopula(R), ot.ClaytonCopula(2.0)] copula = ot.ComposedCopula(collection) return copula.getSample(size) size = 1000 data = generateDataForSpecificInstance(size) alpha = 0.9 binNumber = 3 t0 = time() learner = otagrum.ContinuousPC(data, binNumber, alpha) t1 = time() - t0 #skel = learner.getSkeleton()
import openturns as ot
import otagrum as otagr
X = ot.Sample.ImportFromCSVFile("advised.csv")
binNumber = 2
alpha = 0.9
learner = otagr.ContinuousPC(X, binNumber, alpha)
res = learner.learnDAG()
print(res.toDot())
import pyAgrum as gum
import openturns as ot
import otagrum as otagr
print('Importing data')
data = ot.Sample.ImportFromTextFile(
'../data/Standard_coefficients_0100000.csv', ';')
data = data[0:20000]
data = data.getMarginal(range(0, 12))
print('Initializing the learners')
learners = {
'cbic': otagr.TabuList(data, 3, 1, 2),
'cpc': otagr.ContinuousPC(data, 4, 0.01),
'cmiic': otagr.ContinuousMIIC(data)
}
dags = {}
for (name, learner) in learners.items():
print('Learning with ', name)
dags[name] = learner.learnDAG()
for (name, dag) in dags.items():
dot = dag.toDot()
with open("dag_{}.dot".format(name), "w") as f:
f.write(dot)
Tstruct_file_path = "data/cbn2/struct.txt"
data_set_name = data_set_path.split('/')[-1].split('.')[0]
Tstruct_file_name = Tstruct_file_path.split('/')[-1].split('.')[0]
with open(Tstruct_file_path, 'r') as file:
arcs = file.read().replace('\n', '')
Tstruct = gum.fastBN(arcs)
data = np.loadtxt(data_set_path, delimiter=',', skiprows=1)
sizes = np.linspace(1000, len(data), 10, dtype=int)
alpha = 0.1
binNumber = 3
list_g = []
for size in sizes:
print(size)
sample = data[np.random.randint(0, len(data), size=size)]
sample = ot.Sample(sample)
learner = otagr.ContinuousPC(sample, binNumber, alpha)
list_g.append(learner.learnDAG())
#bn = gum.BayesNet()
#for name in Tstruct.names():
# bn.add(gum.LabelizedVariable(name))
#for arc in g.arcs():
# bn.addArc(arc[0], arc[1])
#comparison = GraphicalBNComparator(Tstruct, bn)
#print(comparison.scores())
# %% # Having a CBN, we can now sample from it. # %% ot.RandomGenerator.SetSeed(10) # Set random seed sample = cbn.getSample(1000) train = sample[:-100] test = sample[-100:] # %% # Learning the structure with continuous PC: # Now that we have data, we can use it to learn the structure with the continuous PC algorithm. # %% learner = otagrum.ContinuousPC(sample, maxConditioningSetSize=5, alpha=0.1) # %% # We first learn the skeleton, that is the undirected structure. # %% skeleton = learner.learnSkeleton() # %% skeleton # %% # Then we look for the v-structures, leading to a Partially Directed Acyclic Graph (PDAG) # %% pdag = learner.learnPDAG()
