def propositionalize(knowledgeGraph=None):
""" Translate a knowledge graph in a two-hot directed vector.
Returns a dataset with 'x' consisting of a two-hot directed vector that represents the existence of a subject s and
an object o, and with 'y' consisting of a one-hot vector that represents the existence of a predicate p.
Given triple <s, p, o>, mapped keys <i, j, k>, and number of entities n, input vector x[i] == 1.0 AND
x[k+n] == 1.0 AND x[!(i|k)] == 0.0. Equally, output vector y[j] == 1.0 AND y[!(j)] == 0.0.
"""
if not knowledgeGraph:
raise ValueError('Expects knowledge graph.')
pi = progressIndicator('Translating Knowledge Graph')
pi.start()
try:
n = len(knowledgeGraph.graph) # number of triples
data = {'x': np.zeros((n, 2*len(knowledgeGraph.entities)), dtype=np.float32), # single-point precision to preserve memory
'y': np.zeros((n, len(knowledgeGraph.relations)), dtype=np.float32)}
# populate data set
for (i, (triple)) in enumerate(knowledgeGraph.graph, 0):
data['x'][i, knowledgeGraph.entities[triple[0]]] = 1.0
data['y'][i, knowledgeGraph.relations[triple[1]]] = 1.0
data['x'][i, knowledgeGraph.entities[triple[2]] + len(knowledgeGraph.entities)] = 1.0
finally:
pi.stop()
return data
def run_program(dataset, model, learner, nepoch):
pi = progressIndicator("[Train]")
pi.start()
avg_error = 0.0
for epoch in range(nepoch):
error = 0.0
for fold in dataset.train:
error += learner.run(model, fold)
fold.__reset__()
dataset.train.__reset__()
avg_error += error / dataset.train.size
pi.stop()
print(" ".ljust(len(str(nepoch))+27) + "Mean Error: {:.5f}".format(avg_error/dataset.test.size))
def read(self, local_path=None, remote_path=None, format=None):
pi = progressIndicator('Reading Knowledge Graph')
pi.start()
try:
if local_path is None and remote_path is None:
raise ValueError('Path cannot be left undefined')
self.path = local_path if local_path else remote_path
if not format:
format = rdflib.util.guess_format(self.path)
self.graph.parse(self.path, format=format)
self.update()
finally:
pi.stop()
