def test_chi_square_model_reject(self):
data1 = np.stack([
np.concatenate([np.ones(50), np.zeros(50)], axis=0),
np.concatenate([np.ones(50), np.zeros(50)], axis=0)
], axis=1)
data2 = np.stack([
np.concatenate([np.ones(30), np.zeros(70)], axis=0),
np.concatenate([np.ones(50), np.zeros(50)], axis=0)
], axis=1)
data3 = np.stack([
np.concatenate([np.ones(50), np.zeros(50)], axis=0),
np.concatenate([np.ones(1), np.zeros(99)], axis=0)
], axis=1)
model1 = trepan.DiscreteModel()
model1.fit(data1)
model2 = trepan.DiscreteModel()
model2.fit(data2)
model3 = trepan.DiscreteModel()
model3.fit(data3)
self.assertFalse(trepan.chi_square_model(model1, model2, set()))
self.assertFalse(trepan.chi_square_model(model1, model3, set()))
def test_chi_square_model_accept(self):
data1 = np.stack([
np.concatenate([np.ones(50), np.zeros(50)], axis=0),
np.concatenate([np.ones(50), np.zeros(50)], axis=0)
], axis=1)
data2 = np.stack([
np.concatenate([np.ones(50), np.zeros(50)], axis=0),
np.concatenate([np.ones(50), np.zeros(50)], axis=0)
], axis=1)
data3 = np.stack([
np.concatenate([np.ones(40), np.zeros(60)], axis=0),
np.concatenate([np.ones(50), np.zeros(50)], axis=0)
], axis=1)
model1 = trepan.DiscreteModel()
model1.fit(data1)
model2 = trepan.DiscreteModel()
model2.fit(data2)
model3 = trepan.DiscreteModel()
model3.fit(data3)
self.assertTrue(trepan.chi_square_model(model1, model2, set()))
self.assertTrue(trepan.chi_square_model(model1, model3, set()))
def test_sample(self):
f1 = np.concatenate([np.zeros(75, dtype=np.float32), np.ones(25, dtype=np.float32)], axis=0)
f2 = np.concatenate([np.zeros(1, dtype=np.float32), np.ones(99, dtype=np.float32)], axis=0)
data = np.stack([f1, f2], axis=1)
model = trepan.DiscreteModel()
model.fit(data)
np.random.seed(2018)
num_samples = 1000
samples = []
for _ in range(num_samples):
samples.append(model.sample())
samples = np.stack(samples, axis=0)
p0_0 = np.sum(samples[:, 0] == 0) / samples.shape[0]
p0_1 = np.sum(samples[:, 0] == 1) / samples.shape[0]
p1_0 = np.sum(samples[:, 1] == 0) / samples.shape[0]
p1_1 = np.sum(samples[:, 1] == 1) / samples.shape[0]
self.assertTrue(0.7 <= p0_0 <= 0.8)
self.assertTrue(0.2 <= p0_1 <= 0.3)
self.assertTrue(0.001 <= p1_0 <= 0.1)
self.assertTrue(0.8 <= p1_1 <= 1.0)
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_fit(self):
f1 = np.concatenate([np.zeros(75, dtype=np.float32), np.ones(25, dtype=np.float32)], axis=0)
f2 = np.concatenate([np.zeros(1, dtype=np.float32), np.ones(99, dtype=np.float32)], axis=0)
data = np.stack([f1, f2], axis=1)
model = trepan.DiscreteModel()
model.fit(data)
np.testing.assert_array_almost_equal(np.array([0.75, 0.25], dtype=np.float32), model.distributions[0])
np.testing.assert_array_almost_equal(np.array([0.01, 0.99], dtype=np.float32), model.distributions[1])
self.assertEqual([[0.0, 1.0], [0.0, 1.0]], model.values)
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_split_probability(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)
data = np.stack([f1, f2], axis=1)
model = trepan.DiscreteModel()
model.fit(data)
split1 = (0, 0.5, trepan.Rule.SplitType.BELOW)
split2 = (1, 0.5, trepan.Rule.SplitType.ABOVE)
self.assertEqual(model.split_probability(*split1), 0.75)
self.assertEqual(model.split_probability(*split2), 0.5)
def test_set_zero(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)
data = np.stack([f1, f2], axis=1)
model = trepan.DiscreteModel()
model.fit(data)
model.set_zero(0, 1)
model.set_zero(1, 0)
for _ in range(100):
np.testing.assert_array_almost_equal(model.sample(), [0., 1.])
def test_zero_by_split(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)
data = np.stack([f1, f2], axis=1)
model = trepan.DiscreteModel()
model.fit(data)
split1 = (0, 0.5, trepan.Rule.SplitType.BELOW)
split2 = (1, 0.5, trepan.Rule.SplitType.ABOVE)
model.zero_by_split(*split1)
model.zero_by_split(*split2)
for _ in range(100):
np.testing.assert_array_almost_equal(model.sample(), [0., 1.])
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)