("joe", "jennifer"),
("janice", "juliet"),
("joe", "juliet"),
("janice", "joey"),
("joe", "joey")]
ltnw.predicate("ancestor",embedding_size*2)
ltnw.predicate("parent",embedding_size*2)
for l in entities:
ltnw.constant(l, min_value=[0.] * embedding_size, max_value=[1.] * embedding_size)
for a,c in parents:
string = "parent(" + a + "," + c + ")"
logging.info(string)
ltnw.axiom(string)
ltnw.variable("a",tf.concat(list(ltnw.CONSTANTS.values()),axis=0))
ltnw.variable("b",tf.concat(list(ltnw.CONSTANTS.values()),axis=0))
ltnw.variable("c",tf.concat(list(ltnw.CONSTANTS.values()),axis=0))
ltnw.axiom("forall a,b: parent(a,b) -> ancestor(a,b)")
ltnw.axiom("forall a,b,c: (ancestor(a,b) & parent(b,c)) -> ancestor(a,c)")
ltnw.axiom("forall a: ~parent(a,a)")
ltnw.axiom("forall a: ~ancestor(a,a)")
ltnw.axiom("forall a,b: parent(a,b) -> ~parent(b,a)")
ltnw.axiom("forall a,b: ancestor(a,b) -> ~ancestor(b,a)")
ltnw.initialize_knowledgebase(
optimizer=tf.train.RMSPropOptimizer(learning_rate=0.01, decay=0.9),
("Fish", "Animal"), ("Shark", "Fish"), ("Squirrel",
"Mammal"))
all_relationships = list(itertools.product(entities, repeat=2))
ltnw.predicate("SubClass", embedding_size * 2)
for l in entities:
ltnw.constant(l,
min_value=[0.] * embedding_size,
max_value=[1.] * embedding_size)
for a, c in relationships:
string = "SubClass(" + a + "," + c + ")"
logging.info(string)
ltnw.axiom(string)
ltnw.variable("a", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
ltnw.variable("b", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
ltnw.variable("c", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
ltnw.axiom(
"forall a,b,c: (SubClass(a,b) & SubClass(b,c)) -> SubClass(a,c)")
ltnw.axiom("forall a: ~SubClass(a,a)")
ltnw.axiom("forall a,b: SubClass(a,b) -> ~SubClass(b,a)")
ltnw.initialize_knowledgebase(
optimizer=tf.train.RMSPropOptimizer(learning_rate=lr, decay=dc),
formula_aggregator=lambda *x: tf.reduce_min(tf.concat(x, axis=0)))
# Train the KB
def ltnsnet():
ltnw.predicate("ancestor", embedding_size * 2)
ltnw.predicate("parent", embedding_size * 2)
for l in entities:
ltnw.constant(l,
min_value=[0.] * embedding_size,
max_value=[1.] * embedding_size)
for index, row in training_pa.iterrows():
if row["type"] == 1:
string = "parent(" + row["first"] + "," + row["second"] + ")"
else:
string = "~parent(" + row["first"] + "," + row["second"] + ")"
ltnw.axiom(string)
for index, row in training_an.iterrows():
if row["type"] == 1:
string = "ancestor(" + row["first"] + "," + row["second"] + ")"
else:
string = "~ancestor(" + row["first"] + "," + row["second"] + ")"
ltnw.axiom(string)
ltnw.variable("a", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
ltnw.variable("b", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
ltnw.variable("c", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
ltnw.axiom("forall a,b: parent(a,b) -> ancestor(a,b)")
ltnw.axiom("forall a,b,c: (ancestor(a,b) & parent(b,c)) -> ancestor(a,c)")
ltnw.axiom("forall a: ~parent(a,a)")
ltnw.axiom("forall a: ~ancestor(a,a)")
ltnw.axiom("forall a,b: parent(a,b) -> ~parent(b,a)")
ltnw.axiom("forall a,b: ancestor(a,b) -> ~ancestor(b,a)")
ltnw.initialize_knowledgebase(
optimizer=tf.train.RMSPropOptimizer(learning_rate=0.01, decay=0.9),
formula_aggregator=lambda *x: tf.reduce_min(tf.concat(x, axis=0)))
# Train the KB
sat_level = ltnw.train(max_epochs=10000)
print(sat_level)
file_name_an = "an_prediction"
file_name_pa = "pa_prediction"
with open(file_name_an, "w") as resutls_file:
resutls_file.write(str(sat_level) + "\n")
print("inferencing an")
for index, row in test_an.iterrows():
resutls_file.write(row["first"] + "," + row["second"] + "," +
np.array_str(
ltnw.ask("ancestor(" + row["first"] + "," +
row["second"] + ")").squeeze()) +
"\n")
with open(file_name_pa, "w") as resutls_file:
resutls_file.write(str(sat_level) + "\n")
print("inferencing pa")
for index, row in test_pa.iterrows():
resutls_file.write(row["first"] + "," + row["second"] + "," +
np.array_str(
ltnw.ask("parent(" + row["first"] + "," +
row["second"] + ")").squeeze()) +
"\n")
predicates = list(map(lambda x: pred_name + str(x), range(1, 100)))
constants = list(map(lambda x: const_name + str(x), range(1, 100)))
for l in constants[:const]:
print(l)
ltnw.constant(l,
min_value=[0.] * embedding_size,
max_value=[1.] * embedding_size)
if card == 1:
ltnw.variable("x", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
for k in predicates[:pred]:
ltnw.predicate(k, embedding_size)
string = "forall x : " + k + "(x)"
print(string)
ltnw.axiom(string)
elif card == 2:
ltnw.variable("x", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
ltnw.variable("y", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
for k in predicates[:pred]:
ltnw.predicate(k, embedding_size * 2)
string = "forall x,y : " + k + "(x,y)"
print(string)
ltnw.axiom(string)
else:
ltnw.variable("x", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
ltnw.variable("y", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
ltnw.variable("z", tf.concat(list(ltnw.CONSTANTS.values()), axis=0))
for k in predicates[:pred]:
ltnw.predicate(k, embedding_size * 3)
string = "forall x,y,z : " + k + "(x,y,z)"