def test_l1_regularization(solver):
n_components = 3
rng = np.random.mtrand.RandomState(42)
X = np.abs(rng.randn(6, 5))
Y = np.abs(rng.randn(5, 4))
# L1 regularization should increase the number of zeros
l1_reg = 2.
reg = CMF(n_components=n_components, solver=solver,
l1_reg=l1_reg, random_state=42)
model = CMF(n_components=n_components, solver=solver,
l1_reg=0., random_state=42)
U_reg, V_reg, Z_reg = reg.fit_transform(X, Y)
U_model, V_model, Z_model = model.fit_transform(X, Y)
U_reg_n_zeros = U_reg[U_reg == 0].size
V_reg_n_zeros = V_reg[V_reg == 0].size
Z_reg_n_zeros = Z_reg[Z_reg == 0].size
U_model_n_zeros = U_model[U_model == 0].size
V_model_n_zeros = V_model[V_model == 0].size
Z_model_n_zeros = Z_model[Z_model == 0].size
msg = "solver: {}".format(solver)
# If one matrix is full of zeros,
# it might make sense for the other matrices to reduce the number of zeros
# Therefore, we compare the total number of zeros
assert_greater(U_reg_n_zeros + V_reg_n_zeros + Z_reg_n_zeros,
U_model_n_zeros + V_model_n_zeros + Z_model_n_zeros, msg)
def test_analysis():
# smoke test to see that analysis works
rng = np.random.mtrand.RandomState(36)
model = CMF(n_components=2, solver="newton", max_iter=1)
c = CountVectorizer()
X_ = c.fit_transform(["hello world",
"goodbye world",
"hello goodbye"])
X_ = csr_matrix(X_)
Y = np.abs(rng.randn(3, 1))
model.fit_transform(X_.T, Y)
model.print_topic_terms(c, importances=False)
model.print_topic_terms(c, importances=True)
def test_transform_custom_init():
# Smoke test that checks if CMF.fit_transform works with custom initialization
random_state = np.random.RandomState(0)
X = np.abs(random_state.randn(6, 5))
Y = np.abs(random_state.randn(5, 1))
n_components = 4
avg = np.sqrt(X.mean() / n_components)
U_init = np.abs(avg * random_state.randn(6, n_components))
V_init = np.abs(avg * random_state.randn(5, n_components))
avg = np.sqrt(Y.mean() / n_components)
Z_init = np.abs(avg * random_state.randn(1, n_components))
m = CMF(solver='newton', n_components=n_components,
x_init='custom', y_init='custom',
random_state=0)
m.fit_transform(X, Y, U=U_init, V=V_init, Z=Z_init)
def test_input_method_compatibility():
# Smoke test for combinations between different init methods
rng = np.random.mtrand.RandomState(0)
X = np.abs(rng.randn(6, 5))
Y = np.abs(rng.randn(5, 6))
n_components = 4
avg = np.sqrt(X.mean() / n_components)
U_init = np.abs(avg * rng.randn(6, n_components))
V_init = np.abs(avg * rng.randn(5, n_components))
avg = np.sqrt(Y.mean() / n_components)
Z_init = np.abs(avg * rng.randn(6, n_components))
inits = [None, 'random', 'nndsvd', 'nndsvda', 'nndsvdar', 'custom']
for x_init, y_init in itertools.product(inits, inits):
pnmf = CMF(n_components=n_components, solver='mu',
x_init=x_init, y_init=y_init,
random_state=0, max_iter=1)
pnmf.fit_transform(X, Y, U=U_init, V=V_init, Z=Z_init)
def sparse_cmf_benchmark(solver):
rng = np.random.mtrand.RandomState(42)
X = np.abs(rng.randn(2000, 150))
X[:1000, 2 * np.arange(10) + 100] = 0
X[1000:, 2 * np.arange(10)] = 0
X_sparse = SP(X)
Y = np.abs(rng.randn(150, 10))
model = CMF(n_components=10, solver=solver, random_state=42, max_iter=10)
U, V, Z = model.fit_transform(X_sparse, Y)
def test_auto_compute_alpha():
rng = np.random.mtrand.RandomState(36)
X = rng.randn(10, 10)
Y = rng.randn(10, 5)
x_emphasis_model = CMF(n_components=2, 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=100, alpha=0.5)
# automatic = weight * number_of_elements is constant for both X and Y
y_emphasis_model = CMF(n_components=2, 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=100, alpha="auto")
U1, V1, Z1 = x_emphasis_model.fit_transform(X, Y)
U2, V2, Z2 = y_emphasis_model.fit_transform(X, Y)
assert_greater(np.linalg.norm(np.dot(U2, V2.T) - X), np.linalg.norm(np.dot(U1, V1.T) - X))
assert_greater(np.linalg.norm(np.dot(V1, Z1.T) - Y), np.linalg.norm(np.dot(V2, Z2.T) - Y))
def test_l2_regularization(solver):
n_components = 3
rng = np.random.mtrand.RandomState(42)
X = np.abs(rng.randn(6, 5))
Y = np.abs(rng.randn(5, 4))
# L2 regularization should decrease the mean of the coefficients
l2_reg = 2.
model = CMF(n_components=n_components, solver=solver,
l2_reg=0., random_state=42)
reg = CMF(n_components=n_components, solver=solver,
l2_reg=l2_reg, random_state=42)
U_reg, V_reg, Z_reg = reg.fit_transform(X, Y)
U_model, V_model, Z_model = model.fit_transform(X, Y)
msg = "solver: {}".format(solver)
assert_greater(U_model.mean(), U_reg.mean(), msg)
assert_greater(V_model.mean(), V_reg.mean(), msg)
assert_greater(Z_model.mean(), Z_reg.mean(), msg)
def test_logit_link_optimization():
n_components = 5
rng = np.random.mtrand.RandomState(42)
X = 1 / (1 + np.exp(-rng.randn(6, 5)))
Y = 1 / (1 + np.exp(-rng.randn(5, 4)))
model = CMF(n_components=n_components, solver="newton",
l2_reg=0., random_state=42, x_link="logit", y_link="logit",
U_non_negative=False, V_non_negative=False, Z_non_negative=False)
U, V, Z = model.fit_transform(X, Y)
assert_less(model.reconstruction_err_, 0.1)
def sparse_cmf_with_logits_benchmark(sample_ratio):
rng = np.random.mtrand.RandomState(42)
X = np.abs(rng.randn(2000, 150))
X[:1000, 2 * np.arange(10) + 100] = 0
X[1000:, 2 * np.arange(10)] = 0
X_sparse = SP(X)
Y = expit(rng.randn(150, 10))
model = CMF(n_components=10,
solver="newton",
random_state=42,
sg_sample_ratio=sample_ratio,
max_iter=10)
U, V, Z = model.fit_transform(X_sparse, Y)
def test_transform_after_fit_no_labels(solver):
rng = np.random.mtrand.RandomState(36)
X = rng.randn(7, 5)
Y = rng.randn(5, 3)
X_new = rng.randn(15, 5)
model = CMF(n_components=2, solver=solver, x_init='svd', y_init='svd',
U_non_negative=False, V_non_negative=False, Z_non_negative=False,
random_state=0, max_iter=100)
U_ft, V_ft, Z_ft = model.fit_transform(X, Y)
U_t, V_t, Z_t = model.transform(X_new, None)
assert_array_equal(V_t, V_ft)
def test_fit_nn_output(solver):
# Test that the decomposition does not contain negative values
X = np.c_[5 * np.ones(5) - np.arange(1, 6),
5 * np.ones(5) + np.arange(1, 6)]
Y = np.c_[5 * np.ones(5) - np.arange(1, 6),
5 * np.ones(5) + np.arange(1, 6)].T
for init in (None, 'nndsvd', 'nndsvda', 'nndsvdar', 'random'):
model = CMF(n_components=2, solver=solver, x_init=init, y_init=init,
random_state=0)
U, V, Z = model.fit_transform(X, Y)
assert_false((U < 0).any() or
(V < 0).any() or
(Z < 0).any())
def test_logit_link_non_negative_optimization():
# Test if the logit link function works with a non-negative counterpart
n_components = 5
rng = np.random.mtrand.RandomState(42)
X = rng.randn(6, 5)
X[X < 0] = 0
Y = 1 / (1 + np.exp(-rng.randn(5, 4)))
model = CMF(n_components=n_components, solver="newton",
l2_reg=0., random_state=42, y_link="logit",
U_non_negative=True, V_non_negative=True, Z_non_negative=False,
hessian_pertubation=0.2, max_iter=1000)
U, V, Z = model.fit_transform(X, Y)
assert_less(model.reconstruction_err_, 0.1)
def test_nonnegative_condition_for_newton_solver():
n_components = 3
rng = np.random.mtrand.RandomState(42)
X = np.abs(rng.randn(6, 5))
Y = np.abs(rng.randn(5, 4))
model = CMF(n_components=n_components, solver="newton",
l2_reg=0., random_state=42,
U_non_negative=False, V_non_negative=False, Z_non_negative=False)
U, V, Z = model.fit_transform(X, Y)
# if one value is negative in any matrix, since X and Y are non-negative,
# all the other matrices will need to have negative values
assert_less(np.min(U), 0)
assert_less(np.min(V), 0)
assert_less(np.min(Z), 0)
def test_stochastic_newton_solver_sparse_input():
rng = np.random.mtrand.RandomState(36)
A = np.abs(rng.randn(10, 10))
A[:, 2 * np.arange(5)] = 0
A_sparse = csr_matrix(A)
B = np.abs(rng.randn(10, 5))
B[2 * np.arange(5), :] = 0
B_sparse = csr_matrix(B)
est1 = CMF(n_components=5, solver="newton", x_init='svd', y_init='svd',
U_non_negative=False, V_non_negative=False, Z_non_negative=False,
sg_sample_ratio=0.5, random_state=0, max_iter=1000)
est2 = clone(est1)
U1, V1, Z1 = est1.fit_transform(A, B)
U2, V2, Z2 = est2.fit_transform(A_sparse, B_sparse)
assert_array_almost_equal(U1, U2)
assert_array_almost_equal(V1, V2)
assert_array_almost_equal(Z1, Z2)
def test_sparse_input(solver):
# Test that sparse matrices are accepted as input
rng = np.random.mtrand.RandomState(42)
A = np.abs(rng.randn(10, 10))
A[:, 2 * np.arange(5)] = 0
A_sparse = csr_matrix(A)
B = np.abs(rng.randn(10, 5))
B[2 * np.arange(5), :] = 0
B_sparse = csr_matrix(B)
est1 = CMF(solver=solver, n_components=5,
x_init='random', y_init='random',
random_state=0, tol=1e-2)
est2 = clone(est1)
U1, V1, Z1 = est1.fit_transform(A, B)
U2, V2, Z2 = est2.fit_transform(A_sparse, B_sparse)
assert_array_almost_equal(U1, U2)
assert_array_almost_equal(V1, V2)
assert_array_almost_equal(Z1, Z2)
def dense_cmf_with_logits_benchmark():
rng = np.random.mtrand.RandomState(42)
X = np.abs(rng.randn(2000, 150))
Y = np.abs(rng.randn(150, 10))
model = CMF(n_components=10, solver="newton", random_state=42, max_iter=10)
U, V, Z = model.fit_transform(X, Y)