def test_omni_unconnected():
np.random.seed(4)
n = 100
m = 50
graphs = [er_nm(n, m) for _ in range(2)]
omni = OmnibusEmbed()
with pytest.warns(UserWarning):
omni.fit(graphs)
def test_diag_aug():
np.random.seed(5)
n = 100
p = 0.25
graphs_list = [er_np(n, p) for _ in range(2)]
graphs_arr = np.array(graphs_list)
# Test that array and list inputs results in same embeddings
omni_arr = OmnibusEmbed(diag_aug=True).fit_transform(graphs_arr)
omni_list = OmnibusEmbed(diag_aug=True).fit_transform(graphs_list)
assert array_equal(omni_list, omni_arr)
def test_omni_embed():
"""
We compare the difference of norms of OmniBar and ABar.
ABar is the lowest variance estimate of the latent positions X.
OmniBar should be reasonablly close to ABar when n_vertices is high.
"""
def compute_bar(arr):
n = arr.shape[0] // 2
return (arr[:n] + arr[n:]) / 2
X, A1, A2 = generate_data(1000)
Abar = (A1 + A2) / 2
np.random.seed(11)
omni = OmnibusEmbed(n_components=3)
OmniBar = compute_bar(omni.fit_transform([A1, A2]))
omni = OmnibusEmbed(n_components=3)
ABar = compute_bar(omni.fit_transform([Abar, Abar]))
tol = 1.0e-2
assert allclose(norm(OmniBar, axis=1),
norm(ABar, axis=1),
rtol=tol,
atol=tol)
def run(diag_aug):
X, A1, A2 = generate_data(1000, seed=2)
Abar = (A1 + A2) / 2
omni = OmnibusEmbed(n_components=3, diag_aug=diag_aug)
OmniBar = compute_bar(omni.fit_transform([A1, A2]))
omni = OmnibusEmbed(n_components=3, diag_aug=diag_aug)
ABar = compute_bar(omni.fit_transform([Abar, Abar]))
tol = 1.0e-2
assert allclose(norm(OmniBar, axis=1),
norm(ABar, axis=1),
rtol=tol,
atol=tol)
def test_omni_matrix_invalid_inputs():
with pytest.raises(ValueError):
empty_list = []
omni = OmnibusEmbed(n_components=2)
omni.fit(empty_list)
with pytest.raises(ValueError):
wrong_shapes = [np.ones((10, 10)), np.ones((20, 20))]
omni = OmnibusEmbed(n_components=2)
omni.fit(wrong_shapes)
with pytest.raises(TypeError):
wrong_dtypes = [1, 2, 3]
omni = OmnibusEmbed(n_components=2)
omni.fit(wrong_dtypes)