def test_dtype_match_cholesky():
# Test different alphas in cholesky solver to ensure full coverage.
# This test is separated from test_dtype_match for clarity.
rng = np.random.RandomState(0)
alpha = (1.0, 0.5)
n_samples, n_features, n_target = 6, 7, 2
X_64 = rng.randn(n_samples, n_features)
y_64 = rng.randn(n_samples, n_target)
X_32 = X_64.astype(np.float32)
y_32 = y_64.astype(np.float32)
# Check type consistency 32bits
ridge_32 = Ridge(alpha=alpha, solver='cholesky')
ridge_32.fit(X_32, y_32)
coef_32 = ridge_32.coef_
# Check type consistency 64 bits
ridge_64 = Ridge(alpha=alpha, solver='cholesky')
ridge_64.fit(X_64, y_64)
coef_64 = ridge_64.coef_
# Do all the checks at once, like this is easier to debug
assert coef_32.dtype == X_32.dtype
assert coef_64.dtype == X_64.dtype
assert ridge_32.predict(X_32).dtype == X_32.dtype
assert ridge_64.predict(X_64).dtype == X_64.dtype
assert_almost_equal(ridge_32.coef_, ridge_64.coef_, decimal=5)
def test_ridge_individual_penalties():
# Tests the ridge object using individual penalties
rng = np.random.RandomState(42)
n_samples, n_features, n_targets = 20, 10, 5
X = rng.randn(n_samples, n_features)
y = rng.randn(n_samples, n_targets)
penalties = np.arange(n_targets)
coef_cholesky = np.array([
Ridge(alpha=alpha, solver="cholesky").fit(X, target).coef_
for alpha, target in zip(penalties, y.T)
])
coefs_indiv_pen = [
Ridge(alpha=penalties, solver=solver, tol=1e-8).fit(X, y).coef_
for solver in ['svd', 'sparse_cg', 'lsqr', 'cholesky', 'sag', 'saga']
]
for coef_indiv_pen in coefs_indiv_pen:
assert_array_almost_equal(coef_cholesky, coef_indiv_pen)
# Test error is raised when number of targets and penalties do not match.
ridge = Ridge(alpha=penalties[:-1])
assert_raises(ValueError, ridge.fit, X, y)
def test_ridge_fit_intercept_sparse_error(solver):
X, y = _make_sparse_offset_regression(n_features=20, random_state=0)
X_csr = sp.csr_matrix(X)
sparse_ridge = Ridge(solver=solver)
err_msg = "solver='{}' does not support".format(solver)
with pytest.raises(ValueError, match=err_msg):
sparse_ridge.fit(X_csr, y)
def test_dtype_match(solver):
rng = np.random.RandomState(0)
alpha = 1.0
n_samples, n_features = 6, 5
X_64 = rng.randn(n_samples, n_features)
y_64 = rng.randn(n_samples)
X_32 = X_64.astype(np.float32)
y_32 = y_64.astype(np.float32)
tol = 2 * np.finfo(np.float32).resolution
# Check type consistency 32bits
ridge_32 = Ridge(alpha=alpha, solver=solver, max_iter=500, tol=tol)
ridge_32.fit(X_32, y_32)
coef_32 = ridge_32.coef_
# Check type consistency 64 bits
ridge_64 = Ridge(alpha=alpha, solver=solver, max_iter=500, tol=tol)
ridge_64.fit(X_64, y_64)
coef_64 = ridge_64.coef_
# Do the actual checks at once for easier debug
assert coef_32.dtype == X_32.dtype
assert coef_64.dtype == X_64.dtype
assert ridge_32.predict(X_32).dtype == X_32.dtype
assert ridge_64.predict(X_64).dtype == X_64.dtype
assert_allclose(ridge_32.coef_, ridge_64.coef_, rtol=1e-4, atol=5e-4)
def test_solver_consistency(solver, proportion_nonzero, n_samples, dtype,
sparse_X, seed):
alpha = 1.
noise = 50. if proportion_nonzero > .9 else 500.
X, y = _make_sparse_offset_regression(
bias=10,
n_features=30,
proportion_nonzero=proportion_nonzero,
noise=noise,
random_state=seed,
n_samples=n_samples)
svd_ridge = Ridge(solver='svd', normalize=True, alpha=alpha).fit(X, y)
X = X.astype(dtype, copy=False)
y = y.astype(dtype, copy=False)
if sparse_X:
X = sp.csr_matrix(X)
if solver == 'ridgecv':
ridge = RidgeCV(alphas=[alpha], normalize=True)
else:
ridge = Ridge(solver=solver, tol=1e-10, normalize=True, alpha=alpha)
ridge.fit(X, y)
assert_allclose(ridge.coef_, svd_ridge.coef_, atol=1e-3, rtol=1e-3)
assert_allclose(ridge.intercept_,
svd_ridge.intercept_,
atol=1e-3,
rtol=1e-3)
def _test_tolerance(filter_):
ridge = Ridge(tol=1e-5, fit_intercept=False)
ridge.fit(filter_(X_diabetes), y_diabetes)
score = ridge.score(filter_(X_diabetes), y_diabetes)
ridge2 = Ridge(tol=1e-3, fit_intercept=False)
ridge2.fit(filter_(X_diabetes), y_diabetes)
score2 = ridge2.score(filter_(X_diabetes), y_diabetes)
assert score >= score2
def test_ridge_singular():
# test on a singular matrix
rng = np.random.RandomState(0)
n_samples, n_features = 6, 6
y = rng.randn(n_samples // 2)
y = np.concatenate((y, y))
X = rng.randn(n_samples // 2, n_features)
X = np.concatenate((X, X), axis=0)
ridge = Ridge(alpha=0)
ridge.fit(X, y)
assert ridge.score(X, y) > 0.9
def test_sparse_design_with_sample_weights():
# Sample weights must work with sparse matrices
n_sampless = [2, 3]
n_featuress = [3, 2]
rng = np.random.RandomState(42)
sparse_matrix_converters = [
sp.coo_matrix, sp.csr_matrix, sp.csc_matrix, sp.lil_matrix,
sp.dok_matrix
]
sparse_ridge = Ridge(alpha=1., fit_intercept=False)
dense_ridge = Ridge(alpha=1., fit_intercept=False)
for n_samples, n_features in zip(n_sampless, n_featuress):
X = rng.randn(n_samples, n_features)
y = rng.randn(n_samples)
sample_weights = rng.randn(n_samples)**2 + 1
for sparse_converter in sparse_matrix_converters:
X_sparse = sparse_converter(X)
sparse_ridge.fit(X_sparse, y, sample_weight=sample_weights)
dense_ridge.fit(X, y, sample_weight=sample_weights)
assert_array_almost_equal(sparse_ridge.coef_,
dense_ridge.coef_,
decimal=6)
def test_ridge_sample_weights():
# TODO: loop over sparse data as well
# Note: parametrizing this test with pytest results in failed
# assertions, meaning that is is not extremely robust
rng = np.random.RandomState(0)
param_grid = product((1.0, 1e-2), (True, False),
('svd', 'cholesky', 'lsqr', 'sparse_cg'))
for n_samples, n_features in ((6, 5), (5, 10)):
y = rng.randn(n_samples)
X = rng.randn(n_samples, n_features)
sample_weight = 1.0 + rng.rand(n_samples)
for (alpha, intercept, solver) in param_grid:
# Ridge with explicit sample_weight
est = Ridge(alpha=alpha,
fit_intercept=intercept,
solver=solver,
tol=1e-6)
est.fit(X, y, sample_weight=sample_weight)
coefs = est.coef_
inter = est.intercept_
# Closed form of the weighted regularized least square
# theta = (X^T W X + alpha I)^(-1) * X^T W y
W = np.diag(sample_weight)
if intercept is False:
X_aug = X
I = np.eye(n_features)
else:
dummy_column = np.ones(shape=(n_samples, 1))
X_aug = np.concatenate((dummy_column, X), axis=1)
I = np.eye(n_features + 1)
I[0, 0] = 0
cf_coefs = linalg.solve(
X_aug.T.dot(W).dot(X_aug) + alpha * I,
X_aug.T.dot(W).dot(y))
if intercept is False:
assert_array_almost_equal(coefs, cf_coefs)
else:
assert_array_almost_equal(coefs, cf_coefs[1:])
assert_almost_equal(inter, cf_coefs[0])
def test_ridge_sag_with_X_fortran():
# check that Fortran array are converted when using SAG solver
X, y = make_regression(random_state=42)
# for the order of X and y to not be C-ordered arrays
X = np.asfortranarray(X)
X = X[::2, :]
y = y[::2]
Ridge(solver='sag').fit(X, y)
def _test_ridge_cv_normalize(filter_):
ridge_cv = RidgeCV(normalize=True, cv=3)
ridge_cv.fit(filter_(10. * X_diabetes), y_diabetes)
gs = GridSearchCV(Ridge(normalize=True, solver='sparse_cg'),
cv=3,
param_grid={'alpha': ridge_cv.alphas})
gs.fit(filter_(10. * X_diabetes), y_diabetes)
assert gs.best_estimator_.alpha == ridge_cv.alpha_
def test_ridge_gcv_sample_weights(gcv_mode, X_constructor, fit_intercept,
n_features, y_shape, noise):
alphas = [1e-3, .1, 1., 10., 1e3]
rng = np.random.RandomState(0)
n_targets = y_shape[-1] if len(y_shape) == 2 else 1
X, y = _make_sparse_offset_regression(n_samples=11,
n_features=n_features,
n_targets=n_targets,
random_state=0,
shuffle=False,
noise=noise)
y = y.reshape(y_shape)
sample_weight = 3 * rng.randn(len(X))
sample_weight = (sample_weight - sample_weight.min() + 1).astype(int)
indices = np.repeat(np.arange(X.shape[0]), sample_weight)
sample_weight = sample_weight.astype(float)
X_tiled, y_tiled = X[indices], y[indices]
cv = GroupKFold(n_splits=X.shape[0])
splits = cv.split(X_tiled, y_tiled, groups=indices)
kfold = RidgeCV(alphas=alphas,
cv=splits,
scoring='neg_mean_squared_error',
fit_intercept=fit_intercept)
# ignore warning from GridSearchCV: DeprecationWarning: The default of the
# `iid` parameter will change from True to False in version 0.22 and will
# be removed in 0.24
with ignore_warnings(category=DeprecationWarning):
kfold.fit(X_tiled, y_tiled)
ridge_reg = Ridge(alpha=kfold.alpha_, fit_intercept=fit_intercept)
splits = cv.split(X_tiled, y_tiled, groups=indices)
predictions = cross_val_predict(ridge_reg, X_tiled, y_tiled, cv=splits)
kfold_errors = (y_tiled - predictions)**2
kfold_errors = [
np.sum(kfold_errors[indices == i], axis=0)
for i in np.arange(X.shape[0])
]
kfold_errors = np.asarray(kfold_errors)
X_gcv = X_constructor(X)
gcv_ridge = RidgeCV(alphas=alphas,
store_cv_values=True,
gcv_mode=gcv_mode,
fit_intercept=fit_intercept)
gcv_ridge.fit(X_gcv, y, sample_weight=sample_weight)
if len(y_shape) == 2:
gcv_errors = gcv_ridge.cv_values_[:, :, alphas.index(kfold.alpha_)]
else:
gcv_errors = gcv_ridge.cv_values_[:, alphas.index(kfold.alpha_)]
assert kfold.alpha_ == pytest.approx(gcv_ridge.alpha_)
assert_allclose(gcv_errors, kfold_errors, rtol=1e-3)
assert_allclose(gcv_ridge.coef_, kfold.coef_, rtol=1e-3)
assert_allclose(gcv_ridge.intercept_, kfold.intercept_, rtol=1e-3)
def _test_multi_ridge_diabetes(filter_):
# simulate several responses
Y = np.vstack((y_diabetes, y_diabetes)).T
n_features = X_diabetes.shape[1]
ridge = Ridge(fit_intercept=False)
ridge.fit(filter_(X_diabetes), Y)
assert ridge.coef_.shape == (2, n_features)
Y_pred = ridge.predict(filter_(X_diabetes))
ridge.fit(filter_(X_diabetes), y_diabetes)
y_pred = ridge.predict(filter_(X_diabetes))
assert_array_almost_equal(np.vstack((y_pred, y_pred)).T, Y_pred, decimal=3)
def test_n_iter():
# Test that self.n_iter_ is correct.
n_targets = 2
X, y = X_diabetes, y_diabetes
y_n = np.tile(y, (n_targets, 1)).T
for max_iter in range(1, 4):
for solver in ('sag', 'saga', 'lsqr'):
reg = Ridge(solver=solver, max_iter=max_iter, tol=1e-12)
reg.fit(X, y_n)
assert_array_equal(reg.n_iter_, np.tile(max_iter, n_targets))
for solver in ('sparse_cg', 'svd', 'cholesky'):
reg = Ridge(solver=solver, max_iter=1, tol=1e-1)
reg.fit(X, y_n)
assert reg.n_iter_ is None
def test_ridge_fit_intercept_sparse_sag():
X, y = _make_sparse_offset_regression(n_features=5,
n_samples=20,
random_state=0,
X_offset=5.)
X_csr = sp.csr_matrix(X)
params = dict(alpha=1.,
solver='sag',
fit_intercept=True,
tol=1e-10,
max_iter=100000)
dense_ridge = Ridge(**params)
sparse_ridge = Ridge(**params)
dense_ridge.fit(X, y)
with pytest.warns(None) as record:
sparse_ridge.fit(X_csr, y)
assert len(record) == 0
assert np.allclose(dense_ridge.intercept_,
sparse_ridge.intercept_,
rtol=1e-4)
assert np.allclose(dense_ridge.coef_, sparse_ridge.coef_, rtol=1e-4)
with pytest.warns(UserWarning, match='"sag" solver requires.*'):
Ridge(solver='sag').fit(X_csr, y)
def test_ridge_shapes():
# Test shape of coef_ and intercept_
rng = np.random.RandomState(0)
n_samples, n_features = 5, 10
X = rng.randn(n_samples, n_features)
y = rng.randn(n_samples)
Y1 = y[:, np.newaxis]
Y = np.c_[y, 1 + y]
ridge = Ridge()
ridge.fit(X, y)
assert ridge.coef_.shape == (n_features, )
assert ridge.intercept_.shape == ()
ridge.fit(X, Y1)
assert ridge.coef_.shape == (1, n_features)
assert ridge.intercept_.shape == (1, )
ridge.fit(X, Y)
assert ridge.coef_.shape == (2, n_features)
assert ridge.intercept_.shape == (2, )
def test_ridge_intercept():
# Test intercept with multiple targets GH issue #708
rng = np.random.RandomState(0)
n_samples, n_features = 5, 10
X = rng.randn(n_samples, n_features)
y = rng.randn(n_samples)
Y = np.c_[y, 1. + y]
ridge = Ridge()
ridge.fit(X, y)
intercept = ridge.intercept_
ridge.fit(X, Y)
assert_almost_equal(ridge.intercept_[0], intercept)
assert_almost_equal(ridge.intercept_[1], intercept + 1.)
def test_toy_ridge_object():
# Test BayesianRegression ridge classifier
# TODO: test also n_samples > n_features
X = np.array([[1], [2]])
Y = np.array([1, 2])
reg = Ridge(alpha=0.0)
reg.fit(X, Y)
X_test = [[1], [2], [3], [4]]
assert_almost_equal(reg.predict(X_test), [1., 2, 3, 4])
assert len(reg.coef_.shape) == 1
assert type(reg.intercept_) == np.float64
Y = np.vstack((Y, Y)).T
reg.fit(X, Y)
X_test = [[1], [2], [3], [4]]
assert len(reg.coef_.shape) == 2
assert type(reg.intercept_) == np.ndarray
def test_ridgecv_sample_weight():
rng = np.random.RandomState(0)
alphas = (0.1, 1.0, 10.0)
# There are different algorithms for n_samples > n_features
# and the opposite, so test them both.
for n_samples, n_features in ((6, 5), (5, 10)):
y = rng.randn(n_samples)
X = rng.randn(n_samples, n_features)
sample_weight = 1.0 + rng.rand(n_samples)
cv = KFold(5)
ridgecv = RidgeCV(alphas=alphas, cv=cv)
ridgecv.fit(X, y, sample_weight=sample_weight)
# Check using GridSearchCV directly
parameters = {'alpha': alphas}
gs = GridSearchCV(Ridge(), parameters, cv=cv)
gs.fit(X, y, sample_weight=sample_weight)
assert ridgecv.alpha_ == gs.best_estimator_.alpha
assert_array_almost_equal(ridgecv.coef_, gs.best_estimator_.coef_)
def test_ridge_fit_intercept_sparse(solver):
X, y = _make_sparse_offset_regression(n_features=20, random_state=0)
X_csr = sp.csr_matrix(X)
# for now only sparse_cg can correctly fit an intercept with sparse X with
# default tol and max_iter.
# sag is tested separately in test_ridge_fit_intercept_sparse_sag
# because it requires more iterations and should raise a warning if default
# max_iter is used.
# other solvers raise an exception, as checked in
# test_ridge_fit_intercept_sparse_error
#
# "auto" should switch to "sparse_cg" when X is sparse
# so the reference we use for both ("auto" and "sparse_cg") is
# Ridge(solver="sparse_cg"), fitted using the dense representation (note
# that "sparse_cg" can fit sparse or dense data)
dense_ridge = Ridge(solver='sparse_cg')
sparse_ridge = Ridge(solver=solver)
dense_ridge.fit(X, y)
with pytest.warns(None) as record:
sparse_ridge.fit(X_csr, y)
assert len(record) == 0
assert np.allclose(dense_ridge.intercept_, sparse_ridge.intercept_)
assert np.allclose(dense_ridge.coef_, sparse_ridge.coef_)
def test_ridge_vs_lstsq():
# On alpha=0., Ridge and OLS yield the same solution.
rng = np.random.RandomState(0)
# we need more samples than features
n_samples, n_features = 5, 4
y = rng.randn(n_samples)
X = rng.randn(n_samples, n_features)
ridge = Ridge(alpha=0., fit_intercept=False)
ols = LinearRegression(fit_intercept=False)
ridge.fit(X, y)
ols.fit(X, y)
assert_almost_equal(ridge.coef_, ols.coef_)
ridge.fit(X, y)
ols.fit(X, y)
assert_almost_equal(ridge.coef_, ols.coef_)
def test_raises_value_error_if_sample_weights_greater_than_1d():
# Sample weights must be either scalar or 1D
n_sampless = [2, 3]
n_featuress = [3, 2]
rng = np.random.RandomState(42)
for n_samples, n_features in zip(n_sampless, n_featuress):
X = rng.randn(n_samples, n_features)
y = rng.randn(n_samples)
sample_weights_OK = rng.randn(n_samples)**2 + 1
sample_weights_OK_1 = 1.
sample_weights_OK_2 = 2.
sample_weights_not_OK = sample_weights_OK[:, np.newaxis]
sample_weights_not_OK_2 = sample_weights_OK[np.newaxis, :]
ridge = Ridge(alpha=1)
# make sure the "OK" sample weights actually work
ridge.fit(X, y, sample_weights_OK)
ridge.fit(X, y, sample_weights_OK_1)
ridge.fit(X, y, sample_weights_OK_2)
def fit_ridge_not_ok():
ridge.fit(X, y, sample_weights_not_OK)
def fit_ridge_not_ok_2():
ridge.fit(X, y, sample_weights_not_OK_2)
assert_raise_message(ValueError,
"Sample weights must be 1D array or scalar",
fit_ridge_not_ok)
assert_raise_message(ValueError,
"Sample weights must be 1D array or scalar",
fit_ridge_not_ok_2)
def test_ridge_sparse_svd():
X = sp.csc_matrix(rng.rand(100, 10))
y = rng.rand(100)
ridge = Ridge(solver='svd', fit_intercept=False)
assert_raises(TypeError, ridge.fit, X, y)
def test_ridge(solver):
# Ridge regression convergence test using score
# TODO: for this test to be robust, we should use a dataset instead
# of np.random.
rng = np.random.RandomState(0)
alpha = 1.0
# With more samples than features
n_samples, n_features = 6, 5
y = rng.randn(n_samples)
X = rng.randn(n_samples, n_features)
ridge = Ridge(alpha=alpha, solver=solver)
ridge.fit(X, y)
assert ridge.coef_.shape == (X.shape[1], )
assert ridge.score(X, y) > 0.47
if solver in ("cholesky", "sag"):
# Currently the only solvers to support sample_weight.
ridge.fit(X, y, sample_weight=np.ones(n_samples))
assert ridge.score(X, y) > 0.47
# With more features than samples
n_samples, n_features = 5, 10
y = rng.randn(n_samples)
X = rng.randn(n_samples, n_features)
ridge = Ridge(alpha=alpha, solver=solver)
ridge.fit(X, y)
assert ridge.score(X, y) > .9
if solver in ("cholesky", "sag"):
# Currently the only solvers to support sample_weight.
ridge.fit(X, y, sample_weight=np.ones(n_samples))
assert ridge.score(X, y) > 0.9
def _test_ridge_diabetes(filter_):
ridge = Ridge(fit_intercept=False)
ridge.fit(filter_(X_diabetes), y_diabetes)
return np.round(ridge.score(filter_(X_diabetes), y_diabetes), 5)
},
{
'name': 'RandomForest',
'instance': RandomForestRegressor(),
'complexity_label': 'estimators',
'complexity_computer': lambda clf: clf.n_estimators
},
{
'name': 'SVR',
'instance': SVR(kernel='rbf'),
'complexity_label': 'support vectors',
'complexity_computer': lambda clf: len(clf.support_vectors_)
},
]
}
benchmark(configuration)
# benchmark n_features influence on prediction speed
percentile = 90
percentiles = n_feature_influence({'ridge': Ridge()}, configuration['n_train'],
configuration['n_test'], [100, 250, 500],
percentile)
plot_n_features_influence(percentiles, percentile)
# benchmark throughput
throughputs = benchmark_throughputs(configuration)
plot_benchmark_throughput(throughputs, configuration)
stop_time = time.time()
print("example run in %.2fs" % (stop_time - start_time))
def test_sparse_cg_max_iter():
reg = Ridge(solver="sparse_cg", max_iter=1)
reg.fit(X_diabetes, y_diabetes)
assert reg.coef_.shape[0] == X_diabetes.shape[1]