def test_subsampled_find_largest_singular_value(random_svd):
s = random_svd[1]
X = generate_matrix_from_svd(*random_svd)
smax = _singular_values.find_largest_singular_value(
X, subsampling_scheme="sqrt", random_state=np.random)
assert (s[0] - smax) / max(smax, s[0]) < SUBSAMPLED_TOL
def test_find_largest_singular_value(random_svd):
s = random_svd[1]
X = generate_matrix_from_svd(*random_svd)
smax = _singular_values.find_largest_singular_value(X,
random_state=np.random)
assert abs(smax - s[0]) / max(smax, s[0]) < TOL
assert smax < s[0]
def _compute_lipschitz(self, X, y):
num_rows, num_cols = X.shape
if self.frobenius_lipchitz:
return la.norm(X, "fro")**2 / (num_rows * num_cols)
s_max = find_largest_singular_value(
X,
subsampling_scheme=self.subsampling_scheme,
random_state=self.random_state_,
)
SSE_lipschitz = 1.5 * s_max**2
return SSE_lipschitz / num_rows
def _estimate_lipschitz(self, X_aug, y):
num_rows = X_aug.shape[0]
if self.frobenius_lipchitz:
if sparse.issparse(X_aug):
return sparse.linalg.norm(X_aug, "fro") ** 2 / num_rows
return la.norm(X_aug, "fro") ** 2 / num_rows
s_max = find_largest_singular_value(
X_aug,
subsampling_scheme=self.subsampling_scheme,
random_state=self.random_state_,
)
SSE_lipschitz = 1.5 * s_max ** 2
return SSE_lipschitz / num_rows
