def test_vectors_and_arrays(self):
X = np.random.normal(size=(10, 3))
Y = np.random.normal(size=(10, 2))
model = SelectiveRegularization(unpenalized_inds=[0],
penalized_model=Ridge(),
fit_intercept=True)
self.assertEqual(model.fit(X, Y).coef_.shape, (2, 3))
self.assertEqual(model.fit(X, Y[:, 0]).coef_.shape, (3, ))
def test_can_clone_selective_regularization(self):
X = np.random.normal(size=(10, 3))
y = np.random.normal(size=(10, ))
model = SelectiveRegularization(unpenalized_inds=[0],
penalized_model=LassoCV(),
fit_intercept=True)
model.cv = 2
model2 = clone(model, safe=False)
assert model2.cv == 2
def test_can_slice(self):
n = 100 + np.random.choice(500)
d = 5 + np.random.choice(20)
alpha = np.random.uniform(0.5, 1.5)
X = np.random.normal(size=(n, d))
y = np.random.normal(size=(n, ))
coef = SelectiveRegularization(unpenalized_inds=slice(2, None),
penalized_model=Lasso(),
fit_intercept=True).fit(X, y).coef_
X_perm = np.hstack((X[:, 1:], X[:, :1]))
coef2 = SelectiveRegularization(unpenalized_inds=slice(1, -1),
penalized_model=Lasso(),
fit_intercept=True).fit(X_perm,
y).coef_
np.testing.assert_allclose(coef2, np.hstack((coef[1:], coef[:1])))
def test_can_use_index_lambda(self):
n = 100 + np.random.choice(500)
d = 5 + np.random.choice(20)
alpha = np.random.uniform(0.5, 1.5)
X = np.random.normal(size=(n, d))
y = np.random.normal(size=(n, ))
coef = SelectiveRegularization(unpenalized_inds=slice(2, None),
penalized_model=Lasso(),
fit_intercept=True).fit(X, y).coef_
def index_lambda(X, y):
# instead of a slice, explicitly return an array of indices
return np.arange(2, X.shape[1])
coef2 = SelectiveRegularization(unpenalized_inds=index_lambda,
penalized_model=Lasso(),
fit_intercept=True).fit(X, y).coef_
np.testing.assert_allclose(coef, coef2)
def test_can_use_sample_weights(self):
n = 100 + np.random.choice(500)
d = 5 + np.random.choice(20)
n_inds = np.random.choice(np.arange(2, d - 2))
inds = np.random.choice(d, n_inds, replace=False)
alpha = np.random.uniform(0.5, 1.5)
X = np.random.normal(size=(n, d))
y = np.random.normal(size=(n, ))
sample_weight = np.random.choice([1, 2], n)
# create an extra copy of rows with weight 2
X_aug = X[sample_weight == 2, :]
y_aug = y[sample_weight == 2]
model = SelectiveRegularization(unpenalized_inds=inds,
penalized_model=Ridge(),
fit_intercept=True)
coef = model.fit(X, y, sample_weight=sample_weight).coef_
coef2 = model.fit(np.vstack((X, X_aug)), np.concatenate(
(y, y_aug))).coef_
np.testing.assert_allclose(coef, coef2)
def test_can_pass_through_attributes(self):
X = np.random.normal(size=(10, 3))
y = np.random.normal(size=(10, ))
model = SelectiveRegularization(unpenalized_inds=[0],
penalized_model=LassoCV(),
fit_intercept=True)
# _penalized_inds is only set during fitting
with self.assertRaises(AttributeError):
inds = model._penalized_inds
# cv exists on penalized model
old_cv = model.cv
model.cv = 2
model.fit(X, y)
# now we can access _penalized_inds
assert np.array_equal(model._penalized_inds, [1, 2])
# check that we can read the cv attribute back out from the underlying model
assert model.cv == 2
def test_against_ridge_ground_truth(self):
for _ in range(10):
n = 100 + np.random.choice(500)
d = 5 + np.random.choice(20)
n_inds = np.random.choice(np.arange(2, d - 2))
inds = np.random.choice(d, n_inds, replace=False)
alpha = np.random.uniform(0.5, 1.5)
X = np.random.normal(size=(n, d))
y = np.random.normal(size=(n, ))
coef = SelectiveRegularization(
unpenalized_inds=np.delete(np.arange(d), inds),
penalized_model=Ridge(alpha=alpha, fit_intercept=False),
fit_intercept=False).fit(X, y).coef_
X_aug = np.zeros((n_inds, d))
X_aug[np.arange(n_inds), inds] = np.sqrt(alpha)
y_aug = np.zeros((n_inds, ))
coefs = LinearRegression(fit_intercept=False).fit(
np.vstack((X, X_aug)), np.concatenate((y, y_aug))).coef_
np.testing.assert_allclose(coef, coefs)
def test_intercept(self):
for _ in range(10):
n = 100 + np.random.choice(500)
d = 5 + np.random.choice(20)
n_inds = np.random.choice(np.arange(2, d - 2))
inds = np.random.choice(d, n_inds, replace=False)
unpenalized_inds = np.delete(np.arange(d), inds)
alpha = np.random.uniform(0.5, 1.5)
X = np.random.normal(size=(n, d))
y = np.random.normal(size=(n, ))
models = [
SelectiveRegularization(
unpenalized_inds=unpenalized_inds,
penalized_model=Lasso(alpha=alpha,
fit_intercept=inner_intercept),
fit_intercept=True) for inner_intercept in [False, True]
]
for model in models:
model.fit(X, y)
np.testing.assert_allclose(models[0].coef_, models[1].coef_)
np.testing.assert_allclose(models[0].intercept_,
models[1].intercept_)