def test_create_markov():
mtd = MTD(3, 2, verbose=0)
mtd.fit(np.array([i for i in range(27)]))
assert mtd.transition_matrix.max() <= 1.0
assert mtd.transition_matrix.min() >= 0.0
assert np.isclose(mtd.transition_matrix.sum(1).max(), 1.0)
assert mtd.transition_matrix.shape == (9, 3)
def test_one_fit_random():
n_dimensions = 2
order = 3
mtd = MTD(n_dimensions, order, verbose=0)
x = np.array([[100, 900], [100, 900], [900, 100], [900, 100], [100, 900],
[100, 900], [900, 100], [900, 100]]).reshape(-1, 1).ravel()
n_direct = np.array([[[2000., 2000.], [2000., 2000.]],
[[400., 3600.], [3600., 400.]],
[[2000., 2000.], [2000., 2000.]]])
log_likelihood, lambdas, transition_matrices = mtd._fit_one(
x, mtd.indexes_, order, n_dimensions, 0.1, 100, 0, 'random', n_direct)
assert lambdas[0] < lambdas[1]
assert lambdas[2] < lambdas[1]
def test_ex_max():
for seed in range(100):
np.random.seed(seed)
mtd = MTD(3, 2, verbose=0, number_of_initiations=1)
mtd.fit(np.random.randint(
0,
100,
27,
))
assert mtd.lambdas.shape == (2, )
assert np.isclose(sum(mtd.lambdas), 1.0)
assert max(mtd.lambdas) <= 1
assert min(mtd.lambdas) >= 0
assert np.isclose(sum(mtd.transition_matrices[0, 0, :]), 1.0)
assert mtd.transition_matrices.shape == (2, 3, 3)
assert mtd.transition_matrices.min() >= 0
assert mtd.transition_matrices.max() <= 1
def test_final_estimates():
np.random.seed(42)
n_dimensions = 2
order = 3
mtd = MTD(n_dimensions, order, n_jobs=1, max_iter=1000, verbose=0)
x = np.array([[200, 800], [200, 800], [900, 100], [900, 100], [200, 800],
[200, 800], [900, 100], [900, 100]]).reshape(-1,
1).ravel() * 100
mtd.fit(x)
assert np.isclose(mtd.lambdas[1], 1., atol=0.01)
assert np.isclose(mtd.transition_matrices[1][0, 0], 0.2, atol=0.01)
assert np.isclose(mtd.transition_matrices[1][1, 1], 0.1, atol=0.01)
assert np.isclose(mtd.transition_matrices[1][1, 0], 0.9, atol=0.01)
assert np.isclose(mtd.transition_matrices[1][0, 1], 0.8, atol=0.01)
def test_create_indexes():
mtd = MTD(4, 3)
assert len(mtd.indexes_) == 256
def test_criterion():
n_dimensions = 2
order = 3
mtd1 = MTD(n_dimensions, order, verbose=0)
x = np.array([[100, 900], [100, 900], [900, 100], [900, 100], [
100, 900
], [100, 900], [900, 100], [
900, 100
]]).reshape(
-1,
1).ravel() # this is generated by MTD model with lambdas = [0, 1, 0]
mtd1.fit(x)
mtd1._calculate_aic()
order = 2
mtd2 = MTD(n_dimensions, order, verbose=0)
x = x.reshape(2, -1).sum(0)
mtd2.fit(x)
mtd2._calculate_aic()
order = 1
mtd3 = MTD(n_dimensions, order, verbose=0)
x = x.reshape(2, -1).sum(0)
mtd3.fit(x)
mtd3._calculate_aic()
assert mtd3.aic > mtd1.aic > mtd2.aic
def test_aic():
mtd1 = MTD(3, 2, verbose=0)
mtd1.fit(np.array([i for i in range(27)]))
mtd1._calculate_aic()
mtd2 = MTD(3, 3, verbose=0)
mtd2.fit(np.array([i for i in range(81)]))
mtd2._calculate_aic()
assert mtd1.aic < mtd2.aic
def test_input_len_no_error():
mtd = MTD(3, 2, verbose=0)
mtd.fit(np.array([i for i in range(27)]))
def test_input_len_error():
with pytest.raises(ValueError):
mtd = MTD(4, 3, verbose=0)
mtd.fit(np.array([1, 2, 3]))
def test_init_method_error():
with pytest.raises(ValueError):
mtd = MTD(4, 3, init_method='a')
# ## Encode paths pe = PathEncoder(3) pe.fit(x, y) pe.label_dict x_tr3, y_tr3 = pe.transform(x, y) x_tr3[:5] y_tr3[:5] # ## Fitting model model = MTD(order=3) model.fit(x_tr3, y_tr3) # ## Information criteria model.aic model.bic # ## Trained parameters model.lambdas.round(3) model.transition_matrices.round(3)
bics.append(model.bic) print(model.aic.round(1), model.bic.round(1)) # + order = 1 sc = SequenceCutter(order) x, y = sc.transform(df.Change_enc.values) pe = PathEncoder(order) pe.fit(x, y) x_tr, y_tr = pe.transform(x, y) model = MTD(order=order, n_jobs=-1, number_of_initiations=100) model.fit(x_tr, y_tr) aics.append(model.aic) bics.append(model.bic) print(model.aic.round(1), model.bic.round(1)) # + order = 2 sc = SequenceCutter(order) x, y = sc.transform(df.Change_enc.values) pe = PathEncoder(order) pe.fit(x, y)