def test_svd_ncomponents_lt_nfeatures():
# smoke test for input where a dimension is 1
rng = np.random.mtrand.RandomState(42)
model = CMF(n_components=3, solver="newton", x_init='svd', y_init='svd',
U_non_negative=False, V_non_negative=False, Z_non_negative=False,
random_state=0, max_iter=1)
X = rng.randn(6, 4)
Y = rng.randn(4, 2)
model.fit(X, Y)
def test_fit_close(solver):
rng = np.random.mtrand.RandomState(42)
# Test that the fit is not too far away
for rndm_state in [0]:
pnmf = CMF(n_components=5, solver=solver, x_init='nndsvdar', y_init='nndsvdar',
random_state=rndm_state, max_iter=1000)
X = np.abs(rng.randn(6, 5))
Y = np.abs(rng.randn(5, 6))
assert_less(pnmf.fit(X, Y).reconstruction_err_, 0.1)
def test_transform_after_fit(solver):
rng = np.random.mtrand.RandomState(36)
X = rng.randn(7, 5)
Y = rng.randn(5, 3)
fit_model = CMF(n_components=2, solver=solver, x_init='random', y_init='random',
U_non_negative=False, V_non_negative=False, Z_non_negative=False,
random_state=0, max_iter=100)
fit_transform_model = clone(fit_model)
fit_model.fit(X, Y)
U_f, V_f, Z_f = fit_model.transform(X, Y)
U_ft, V_ft, Z_ft = fit_transform_model.fit_transform(X, Y)
# the initalizations will differ, so the results may also differ slightly
assert_array_almost_equal(U_f, U_ft, decimal=2)
assert_array_almost_equal(V_f, V_ft, decimal=2)
assert_array_almost_equal(Z_f, Z_ft, decimal=2)
def test_stochastic_newton_solver():
rng = np.random.mtrand.RandomState(42)
model = CMF(n_components=5, solver="newton", x_init='svd', y_init='svd',
U_non_negative=False, V_non_negative=False, Z_non_negative=False, alpha=0.5,
sg_sample_ratio=0.5, random_state=0, max_iter=1000)
X = rng.randn(6, 5)
Y = rng.randn(5, 6)
assert_less(model.fit(X, Y).reconstruction_err_, 0.1)
def test_recover_low_rank_matrix(solver):
rng = np.random.mtrand.RandomState(42)
# Test that the fit is not too far away
pnmf = CMF(5, solver=solver, x_init='nndsvdar', y_init='nndsvdar',
random_state=0, max_iter=1000)
U = np.abs(rng.randn(10, 5))
V = np.abs(rng.randn(8, 5))
Z = np.abs(rng.randn(6, 5))
X = np.dot(U, V.T)
Y = np.dot(V, Z.T)
assert_less(pnmf.fit(X, Y).reconstruction_err_, 1.0)
def test_stochastic_newton_solver_sparse_input_close():
rng = np.random.mtrand.RandomState(42)
model = CMF(n_components=5, solver="newton", x_init='svd', y_init='svd',
U_non_negative=False, V_non_negative=False, Z_non_negative=False, alpha=0.5,
sg_sample_ratio=0.5, random_state=0, max_iter=1000)
A = rng.randn(6, 5)
B = rng.randn(5, 6)
A_sparse = csr_matrix(A)
B_sparse = csr_matrix(B)
assert_less(model.fit(A_sparse, B_sparse).reconstruction_err_, 0.1)