def test_unified_plastic_net_real(N=500, D=1000, tol=1e-12, max_iter=10000):
r"""Test :meth:`plasticnet.classes.Regression.fit_unified_plastic_net` against sklearn ElasticNet with transformed variables."""
X, y, beta_true = make_regression(
n_samples=N, n_features=D, n_informative=N // 10, coef=True
)
X, y = scale(X), scale(y)
lambda_total = np.random.exponential()
alpha = np.random.rand()
xi = np.random.randn(D).astype(np.float64)
X_prime = X
y_prime = y - np.dot(X, xi)
lm = linear_model.ElasticNet(
alpha=lambda_total, l1_ratio=alpha, tol=tol, max_iter=max_iter
)
lm.fit(X_prime, y_prime)
beta_lm = lm.coef_ + xi
lm_pnet = Regression(X, y)
lm_pnet.xi = xi
lm_pnet.fit_unified_plastic_net(
lambda_total=lambda_total, alpha=alpha, tol=tol, max_iter=max_iter
)
np.testing.assert_almost_equal(beta_lm, lm_pnet.beta, decimal=4)
def test_unified_plastic_net_trivial(N=500, D=1000, tol=1e-12, max_iter=10000):
r"""Test unified plastic net (:math:`\xi=0` in :meth:`plasticnet.classes.Regression.fit_unified_plastic_net`) against sklearn ElasticNet."""
X, y, beta_true = make_regression(
n_samples=N, n_features=D, n_informative=N // 10, coef=True
)
X, y = scale(X), scale(y)
lambda_total = np.random.exponential()
alpha = np.random.rand()
xi = np.zeros(D, dtype=np.float64)
lm = linear_model.ElasticNet(
alpha=lambda_total, l1_ratio=alpha, tol=tol, max_iter=max_iter
)
lm.fit(X, y)
beta_lm = lm.coef_
lm_pnet = Regression(X, y)
lm_pnet.xi = xi
lm_pnet.fit_unified_plastic_net(
lambda_total=lambda_total, alpha=alpha, tol=tol, max_iter=max_iter
)
np.testing.assert_almost_equal(beta_lm, lm_pnet.beta, decimal=4)
def test_ordinary_least_squares_explicit(N=1500, D=1000, tol=1e-12, max_iter=10000):
r"""Test explicitly coded special case OLS numba code in :meth:`plasticnet.classes.Regression.fit_ordinary_least_squares` against sklearn LinearRegression"""
X, y, beta_true = make_regression(
n_samples=N, n_features=D, n_informative=N, coef=True
)
X, y = scale(X), scale(y)
lm = linear_model.LinearRegression()
lm.fit(X, y)
lm_pnet = Regression(X, y)
lm_pnet.fit_ordinary_least_squares(tol=tol, max_iter=max_iter)
np.testing.assert_almost_equal(lm.coef_, lm_pnet.beta, decimal=4)
def test_set_beta(N=500, D=1000, tol=1e-12, max_iter=10000):
r"""Test beta property"""
X, y, beta_true = make_regression(
n_samples=N, n_features=D, n_informative=N, coef=True
)
X, y = scale(X), scale(y)
lm_pnet = Regression(X, y)
new_beta = np.random.randn(D)
new_r = y - np.dot(X, new_beta)
lm_pnet.beta = new_beta
np.testing.assert_almost_equal(lm_pnet.beta, new_beta, decimal=4)
np.testing.assert_almost_equal(lm_pnet._r, new_r, decimal=4)
def test_ridge_explicit(N=500, D=1000, tol=1e-12, max_iter=10000):
r"""Test explicitly coded special case ridge numba code in :meth:`plasticnet.classes.Regression.fit_ridge` against sklearn elastic net with l1_ratio=0"""
X, y, beta_true = make_regression(
n_samples=N, n_features=D, n_informative=N // 10, coef=True
)
X, y = scale(X), scale(y)
lambda_total = np.random.exponential()
lm = linear_model.Ridge(alpha=lambda_total * N, tol=tol, max_iter=max_iter)
lm.fit(X, y)
lm_pnet = Regression(X, y)
lm_pnet.fit_ridge(lambda_total=lambda_total, tol=tol, max_iter=max_iter)
np.testing.assert_almost_equal(lm.coef_, lm_pnet.beta, decimal=4)
def test_soft_plastic_net_limiting_cases(N=500, D=1000, tol=1e-12, max_iter=10000):
r"""Test :meth:`plasticnet.classes.Regression.fit_soft_plastic_net` against sklearn ElasticNet in limiting cases."""
X, y, beta_true = make_regression(
n_samples=N, n_features=D, n_informative=N // 10, coef=True
)
X, y = scale(X), scale(y)
lambda_total = np.random.exponential()
zeta = np.random.randn(D).astype(np.float64)
alpha = 1.0
lm = linear_model.ElasticNet(
alpha=lambda_total, l1_ratio=alpha, tol=tol, max_iter=max_iter
)
lm.fit(X, y)
beta_lm = lm.coef_
lm_pnet = Regression(X, y)
lm_pnet.zeta = zeta
lm_pnet.fit_soft_plastic_net(
lambda_total=lambda_total, alpha=alpha, tol=tol, max_iter=max_iter
)
np.testing.assert_almost_equal(beta_lm, lm_pnet.beta, decimal=4)
alpha = 0.0
X_prime = X
y_prime = y - np.dot(X, zeta)
lm = linear_model.Ridge(alpha=lambda_total * N, tol=tol, max_iter=max_iter)
lm.fit(X_prime, y_prime)
beta_lm = lm.coef_ + zeta
lm_pnet = Regression(X, y)
lm_pnet.zeta = zeta
lm_pnet.fit_soft_plastic_net(
lambda_total=lambda_total, alpha=alpha, tol=tol, max_iter=max_iter
)
np.testing.assert_almost_equal(beta_lm, lm_pnet.beta, decimal=4)
def test_plastic_ridge_trivial(N=500, D=1000, tol=1e-12, max_iter=10000):
r"""Test plastic ridge(:math:`\zeta=0` in :meth:`plasticnet.classes.Regression.fit_plastic_ridge`) against sklearn ElasticNet."""
X, y, beta_true = make_regression(
n_samples=N, n_features=D, n_informative=N // 10, coef=True
)
X, y = scale(X), scale(y)
lambda_total = np.random.exponential()
zeta = np.zeros(D, dtype=np.float64)
lm = linear_model.Ridge(alpha=lambda_total * N, tol=tol, max_iter=max_iter)
lm.fit(X, y)
beta_lm = lm.coef_
lm_pnet = Regression(X, y)
lm_pnet.zeta = zeta
lm_pnet.fit_plastic_ridge(lambda_total=lambda_total, tol=tol, max_iter=max_iter)
np.testing.assert_almost_equal(beta_lm, lm_pnet.beta, decimal=4)