def test_label_consistency_w_different_seeds(self):
test_cases = [
# seed1, seed2, n_features, label order, expected
(0, 1, 5, [0, 1, 2]),
(3, 2, 5, [0, 1, 2, 1, 0]),
(0, 1, 5, [0, 1, 2, 3, 0, 1, 2, 3]),
(0, 9, 5, [0, 1, 1, 0, 2, 2, 2, 0]),
]
for s1, s2, n, labels in test_cases:
rdata = RandomData(0, n, window_size=5)
k = len(set(labels)) # Num clusters
t = 200 * k * len(labels) # total ts length
breaks = [(i) * t // len(labels) for i, _ in enumerate(labels, 1)]
X, _ = rdata.generate_points(labels, breaks)
ticc1 = TICC(n_clusters=k,
window_size=5,
n_jobs=4,
random_state=s1)
ticc2 = TICC(n_clusters=k,
window_size=5,
n_jobs=4,
random_state=s2)
y1 = ticc1.fit_predict(X)
y2 = ticc2.fit_predict(X)
np.testing.assert_array_equal(y1, y2)
def test_pass_if_consistent_on_similar_random_data(self):
test_cases = [
(5, [0, 1, 0], 5),
(5, [0, 1, 2, 1, 0], 5),
]
for n, labels, repeats in test_cases:
rdata = RandomData(seed=0, n_features=n, window_size=5)
k = len(set(labels))
t = 200 * k * len(labels)
breaks = [(i) * t // len(labels) for i, _ in enumerate(labels, 1)]
rdata.generate_cluster_params(k)
# Reuse same cluster parameters for each dataset
data = [
rdata.generate_points(labels, breaks, True)[0]
for i in range(repeats)
]
ticc = TICC(n_clusters=k,
window_size=5,
beta=300,
n_jobs=4,
random_state=0,
cluster_reassignment=0.3,
verbose=True)
y_preds = [ticc.fit_predict(X) for X in data]
for y1, y2 in combinations(y_preds, 2):
result = np.sum(np.not_equal(y1, y2)) / t
assert result < 0.02
def test_pass_if_y_size_correct(self):
test_cases = [
([0, 1, 0], [20, 40, 60]),
([0, 1, 0, 1], [20, 40, 500, 600]),
]
for seg, b in test_cases:
rd = RandomData(window_size=5, n_features=5)
_, result = rd.generate_points(seg, b)
assert result.shape == (b[-1], )
def test_pass_if_y_labels_correct(self):
test_cases = [
([0, 1, 0], [3, 7, 10], [0, 0, 0, 1, 1, 1, 1, 0, 0, 0]),
([1, 0, 1, 0], [5, 10, 15,
20], [1] * 5 + [0] * 5 + [1] * 5 + [0] * 5),
]
for seg, b, expected in test_cases:
rd = RandomData(window_size=5, n_features=5)
_, y = rd.generate_points(seg, b)
assert y.tolist() == expected
def test_pass_if_consistent_with_same_seed(self):
test_cases = [
([0, 1, 0], [20, 40, 60], 0),
([0, 1, 0], [20, 40, 60], 1),
([0, 1, 0], [20, 40, 60], 10),
]
for seg, b, seed in test_cases:
X1, y1 = RandomData(seed).generate_points(seg, b)
X2, y2 = RandomData(seed).generate_points(seg, b)
np.testing.assert_array_equal(X1, X2)
np.testing.assert_array_equal(y1, y2)
def test_split_theta(self):
test_cases = [(5, 8)]
for n, w in test_cases:
rdata = RandomData(0, n_features=n, window_size=w)
cluster = _TICCluster(1)
cluster.MRF_ = rdata.block_toeplitz()
blocks = cluster.split_theta(w)
assert len(blocks) == w
for i, B in enumerate(blocks):
assert B.shape == (n, n)
A = cluster.MRF_[i * n:(i + 1) * n, :n]
assert_array_equal(B, A)
def test_pass_if_block_dims_correct(self):
test_cases = [
(5, 10),
(10, 10),
(10, 1),
(1, 1),
(1, 10),
]
for n, w in test_cases:
rdata = RandomData(n_features=n, window_size=w)
result = rdata.block_toeplitz().shape
assert result == (n * w, n * w)
def test_pass_if_generates_diff_points_each_call(self):
test_cases = [
([0, 1, 0], [20, 40, 60], 0),
([0, 1, 0], [20, 40, 60], 10),
([0, 1, 0], [20, 40, 60], 100),
]
for seg, b, seed in test_cases:
rd = RandomData(seed)
X1, y1 = rd.generate_points(seg, b)
X2, y2 = rd.generate_points(seg, b)
np.testing.assert_raises(AssertionError,
np.testing.assert_array_equal, X1, X2)
# We want different points but same labels!
np.testing.assert_array_equal(y1, y2)
def test_fit_predict(self):
rdata = RandomData(0, 5, 5)
labels = [0, 1, 2, 3, 4, 5]
breaks = [(i) * 1200 // len(labels) for i, _ in enumerate(labels, 1)]
X, _ = rdata.generate_points(labels, breaks)
ticc = TICC(n_clusters=6,
window_size=5,
beta=0,
n_jobs=4,
verbose=True,
random_state=0)
A = ticc.fit(X).predict(X)
B = ticc.fit_predict(X)
np.testing.assert_array_equal(A, B)
def test_score_increases_with_f1(self):
rdata = RandomData(0, 5, 5)
samples_per_segment = 300
labels = [0, 1, 2, 3] # , 1, 0, 1]
k = len(set(labels)) # Num clusters
t = samples_per_segment * k * len(labels) # total ts length
breaks = [(i) * t // len(labels) for i, _ in enumerate(labels, 1)]
X, y_true = rdata.generate_points(labels, breaks)
params = ParameterGrid({
'beta': [0, 500, 1000],
'window_size': [1, 5, 10],
})
models = (TICC(**p, random_state=0) for p in params)
models = Parallel(n_jobs=-1)(delayed(lambda x: x.fit(X))(m)
for m in models)
scores = [model.score(X) for model in models]
f1_scores = [
best_f1(y_true, y) for y in map(lambda x: x.predict(X), models)
]
assert scores.index(max(scores)) == f1_scores.index(max(f1_scores))
def test_fail_if_input_error(self):
test_cases = [
([0, 1], [100, 300, 900]),
([0, 1], [100, 300, 900]),
([0, 1, 3, 5], [100, 300, 450]),
]
for seg, b in test_cases:
try:
RandomData().generate_points(seg, b)
except ValueError:
assert True
def test_pass_if_positive_definite(self):
test_cases = [
(5, 10),
(10, 10),
(3, 4),
(5, 1),
]
for n, w in test_cases:
# Block Toeplitz Matrix Theta
Theta = RandomData(None, n, w).block_toeplitz()
assert bool(np.linalg.cholesky(Theta).shape)
def test_matches_original_paper_macro_F1(self):
test_cases = [
# n_features, label order, macro f1 to-beat
(5, [0, 1, 0], 0.9),
(5, [0, 1, 2, 1, 0], 0.9),
(5, [0, 1, 2, 3, 0, 1, 2, 3], 0.9),
(5, [0, 1, 1, 0, 2, 2, 2, 0], 0.9),
]
for n, labels, expected in test_cases:
rdata = RandomData(0, n, window_size=5)
# Original paper code performs at 100p/cluster!
samples_per_segment = 120
k = len(set(labels)) # Num clusters
t = samples_per_segment * k * len(labels) # total ts length
breaks = [(i) * t // len(labels) for i, _ in enumerate(labels, 1)]
X, y_tru = rdata.generate_points(labels, breaks)
ticc = TICC(n_clusters=k, window_size=5, n_jobs=4, random_state=0)
y = ticc.fit_predict(X)
# We use best_f1 because label:segment assignments are arbitrary
result = best_f1(y_tru, y, average='macro')
assert result > expected
def test_pass_if_diags_symmetric(self):
test_cases = [
(5, 10),
(10, 10),
(3, 4),
(5, 1),
]
for n, w in test_cases:
# Block Toeplitz Matrix Theta
Theta = RandomData(None, n, w).block_toeplitz()
block_chunked = [ # list of rows/cols of block matrix
np.split(col, w) for col in np.split(Theta, w, axis=1)
]
for i in range(w):
A = block_chunked[i][i]
np.testing.assert_array_equal(A.T, A)
def test_pass_if_block_toeplitz(self):
test_cases = [
(5, 10),
(10, 10),
(3, 4),
(5, 1),
]
for n, w in test_cases:
# Block Toeplitz Matrix Theta
Theta = RandomData(None, n, w).block_toeplitz()
# List of w columns of n*n blocks
block_chunked = [ # list of rows/cols of block matrix
np.split(col, w) for col in np.split(Theta, w, axis=1)
]
for i in range(w - 1):
for j in range(w - 1):
A = block_chunked[i][j]
B = block_chunked[i + 1][j + 1]
np.testing.assert_array_equal(A, B)
def test_recycling_clusters_between_calls(self):
test_cases = [
([0, 1, 0], [20, 40, 60], 0),
([0, 1, 0], [20, 40, 60], 10),
([0, 1, 0], [20, 40, 60], 100),
]
for seg, b, seed in test_cases:
rdata = RandomData(seed)
rdata.generate_cluster_params(len(set(seg)))
X1, y1 = rdata.generate_points(seg, b, True)
C1 = rdata.clusters
X2, y2 = rdata.generate_points(seg, b, True)
C2 = rdata.clusters
assert C1 == C2
fig.show()
# %% Generate data
n_features = 5
label_seq = [0, 1, 2, 0, 2, 1]
samples_per_segment = 250
window_size = 8
# Derived from above params
k = len(set(label_seq)) # Num clusters
t = samples_per_segment * len(label_seq) # total ts length
breaks = [i * t // len(label_seq) for i in range(1, len(label_seq) + 1)]
palette = {n: c['color'] for n, c in zip(range(n_features), colors)}
randomdata = RandomData(seed=1234,
n_features=n_features,
window_size=window_size)
X, y_true = randomdata.generate_points(label_seq, breaks)
# Plot Synthetic Data
plot_synthetic_data(X, breaks)
# %% Fit TICC and GMM to data
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)
ticc = TICC(n_clusters=k, window_size=window_size, random_state=1234, beta=200)
gmm = GaussianMixture(n_components=k, random_state=1234)
X_stacked = ticc.stack_data(X_scaled)
y_ticc = ticc.fit_predict(X)
y_gmm = gmm.fit_predict(X_stacked)