def test_sample_with_hard_constraints(self):
f1 = np.concatenate([np.zeros(75, dtype=np.float32), np.ones(25, dtype=np.float32)], axis=0)
f2 = np.concatenate([np.zeros(50, dtype=np.float32), np.ones(50, dtype=np.float32)], axis=0)
f3 = np.concatenate([np.zeros(20, dtype=np.float32), np.ones(80, dtype=np.float32)], axis=0)
data = np.stack([f1, f2, f3], axis=1)
model = trepan.DiscreteModel()
model.fit(data)
constraints = [
("left", trepan.Rule(0, 0.5, trepan.Rule.SplitType.BELOW)),
("right", trepan.Rule(2, 0.5, trepan.Rule.SplitType.BELOW))
]
oracle = trepan.Oracle(lambda x: x[:, 0], trepan.Oracle.DataType.DISCRETE, 0.05, 0.05)
num_samples = 1000
samples = []
for _ in range(num_samples):
samples.append(oracle.sample_with_constraints(model, constraints))
samples = np.stack(samples)
self.assertTrue(np.all(samples[:, 0] == 0))
self.assertTrue(np.all(samples[:, 2] == 1))
p1_0 = np.sum(samples[:, 1] == 0) / samples.shape[0]
p1_1 = np.sum(samples[:, 1] == 1) / samples.shape[0]
self.assertTrue(0.4 <= p1_0 <= 0.6)
self.assertTrue(0.4 <= p1_1 <= 0.6)
def test_train_impossible(self):
data = np.random.uniform(0, 1, size=[100, 40])
labels = np.random.randint(0, 30, size=100)
oracle = trepan.Oracle(lambda x: x[:, 0], trepan.Oracle.DataType.DISCRETE, 0.05, 0.05)
tp = trepan.Trepan(data, labels, oracle, 5, 50)
tp.train()
def count_nodes(node):
internal = 0
leafs = 0
if node is not None:
if node.leaf:
leafs += 1
else:
internal += 1
x, y = count_nodes(node.left_child)
internal += x
leafs += y
x, y = count_nodes(node.right_child)
internal += x
leafs += y
return internal, leafs
internal_count, _ = count_nodes(tp.root)
self.assertEqual(internal_count, 5)
def test_sample_failure_mode(self):
# test with a specific failure mode
model = trepan.DiscreteModel()
model.distributions = [
np.array([1.], dtype=np.float32), np.array([1.], dtype=np.float32), np.array([1.], dtype=np.float32),
np.array([1.], dtype=np.float32), np.array([1.], dtype=np.float32), np.array([0.5, 0.5], dtype=np.float32),
np.array([1.], dtype=np.float32), np.array([1.], dtype=np.float32), np.array([0.5, 0.5], dtype=np.float32),
np.array([0.5, 0.5], dtype=np.float32), np.array([1.], dtype=np.float32), np.array([1.], dtype=np.float32),
np.array([1.], dtype=np.float32), np.array([1.], dtype=np.float32), np.array([1.], dtype=np.float32)
]
model.values = [
[0.0], [1.0], [1.0], [1.0], [1.0], [0.0, 1.0], [1.0], [1.0], [0.0, 1.0], [0.0, 1.0], [1.0], [1.0], [1.0],
[0.0], [1.0]
]
model.num_features = len(model.values)
rule1 = trepan.Rule(1, 0.5, trepan.Rule.SplitType.ABOVE)
rule1.add_split(5, 0.5, trepan.Rule.SplitType.BELOW)
rule1.num_required = 2
rule2 = trepan.Rule(14, 0.5, trepan.Rule.SplitType.ABOVE)
rule2.add_split(4, 0.5, trepan.Rule.SplitType.BELOW)
rule2.num_required = 2
rule3 = trepan.Rule(13, 0.5, trepan.Rule.SplitType.ABOVE)
rule3.add_split(7, 0.5, trepan.Rule.SplitType.ABOVE)
rule3.add_split(8, 0.5, trepan.Rule.SplitType.BELOW)
rule4 = trepan.Rule(3, 0.5, trepan.Rule.SplitType.ABOVE)
rule4.add_split(10, 0.5, trepan.Rule.SplitType.ABOVE)
rule4.num_required = 2
rule5 = trepan.Rule(2, 0.5, trepan.Rule.SplitType.ABOVE)
rule5.add_split(9, 0.5, trepan.Rule.SplitType.ABOVE)
constraints = [
("left", trepan.Rule(6, 0.5, trepan.Rule.SplitType.ABOVE)),
("right", trepan.Rule(0, 0.5, trepan.Rule.SplitType.ABOVE)),
("left", rule1),
("left", rule2),
("left", rule3),
("left", trepan.Rule(12, 0.5, trepan.Rule.SplitType.ABOVE)),
("left", rule4),
("left", rule5)
]
oracle = trepan.Oracle(lambda x: x[:, 0], trepan.Oracle.DataType.DISCRETE, 0.05, 0.05)
oracle.sample_with_constraints(model, constraints)
def test_step_end(self):
data = np.concatenate([np.zeros(100, dtype=np.float32), np.ones(100, dtype=np.float32)], axis=0)
labels = cp.deepcopy(data)
data = np.expand_dims(data, axis=1)
oracle = trepan.Oracle(lambda x: x[:, 0], trepan.Oracle.DataType.DISCRETE, 0.05, 0.05)
tp = trepan.Trepan(data, labels, oracle, 15, 50)
tp.step()
self.assertIsNotNone(tp.root.left_child)
self.assertIsNotNone(tp.root.left_child.parent)
self.assertIsNotNone(tp.root.right_child)
self.assertIsNotNone(tp.root.right_child.parent)
self.assertIsNone(tp.root.left_child.rule)
self.assertIsNone(tp.root.right_child.rule)
self.assertTrue(tp.queue.is_empty())
def test_sample_with_disj_constraints(self):
f1 = np.concatenate([np.zeros(75, dtype=np.float32), np.ones(25, dtype=np.float32)], axis=0)
f2 = np.concatenate([np.zeros(50, dtype=np.float32), np.ones(50, dtype=np.float32)], axis=0)
f3 = np.concatenate([np.zeros(20, dtype=np.float32), np.ones(80, dtype=np.float32)], axis=0)
data = np.stack([f1, f2, f3], axis=1)
model = trepan.DiscreteModel()
model.fit(data)
rule = trepan.Rule(0, 0.5, trepan.Rule.SplitType.BELOW)
rule.add_split(1, 0.5, trepan.Rule.SplitType.ABOVE)
rule.add_split(2, 0.5, trepan.Rule.SplitType.BELOW)
rule.num_required = 2
constraints = [
("left", rule)
]
oracle = trepan.Oracle(lambda x: x[:, 0], trepan.Oracle.DataType.DISCRETE, 0.05, 0.05)
oracle.sample_with_constraints(model, constraints)
def test_prune_tree_none(self):
p = trepan.Trepan.Node()
p_lc = trepan.Trepan.Node()
p_rc = trepan.Trepan.Node()
p_lc.majority_class = "a"
p_lc.fidelity = 1.0
p_lc.leaf = True
p_rc.majority_class = "b"
p_rc.fidelity = 1.0
p_rc.leaf = True
p.majority_class = "a"
p.fidelity = 1.0
p.leaf = False
p.left_child = p_lc
p_lc.parent = p
p.right_child = p_rc
p_rc.parent = p
data = np.random.uniform(-1, 1, size=[100, 10])
labels = np.random.randint(0, 10, size=100)
oracle = trepan.Oracle(lambda x: x[:, 0], trepan.Oracle.DataType.DISCRETE, 0.05, 0.05)
tp = trepan.Trepan(data, labels, oracle, 15, 50)
tp.root = p
tp.prune()
self.assertEqual(tp.root.majority_class, "a")
self.assertEqual(tp.root.fidelity, 1)
self.assertEqual(tp.root.left_child, p_lc)
self.assertEqual(tp.root.right_child, p_rc)
def test_step_continue(self):
data = np.concatenate([np.zeros(20, dtype=np.float32), np.ones(20, dtype=np.float32)], axis=0)
data = np.expand_dims(data, axis=1)
labels = np.concatenate([
np.zeros(10, dtype=np.float32), np.ones(10, dtype=np.float32),
np.ones(10, dtype=np.float32) + 1, np.ones(10, dtype=np.float32) + 2
], axis=0)
oracle = trepan.Oracle(lambda x: x[:, 0], trepan.Oracle.DataType.DISCRETE, 0.05, 0.05)
tp = trepan.Trepan(data, labels, oracle, 15, 50)
tp.step()
self.assertIsNotNone(tp.root.left_child)
self.assertIsNotNone(tp.root.left_child.parent)
self.assertIsNotNone(tp.root.right_child)
self.assertIsNotNone(tp.root.right_child.parent)
self.assertIsNone(tp.root.left_child.rule)
self.assertIsNone(tp.root.right_child.rule)
self.assertEqual(tp.queue.size, 2)