import numpy as np
from sklearn.utils.optimize import _newton_cg
from scipy.optimize import fmin_ncg
from sklearn.utils._testing import assert_array_almost_equal
def test_newton_cg():
# Test that newton_cg gives same result as scipy's fmin_ncg
rng = np.random.RandomState(0)
A = rng.normal(size=(10, 10))
x0 = np.ones(10)
def func(x):
Ax = A.dot(x)
return .5 * (Ax).dot(Ax)
def grad(x):
return A.T.dot(A.dot(x))
def hess(x, p):
return p.dot(A.T.dot(A.dot(x.all())))
def grad_hess(x):
return grad(x), lambda x: A.T.dot(A.dot(x))
assert_array_almost_equal(
_newton_cg(grad_hess, func, grad, x0, tol=1e-10)[0],
fmin_ncg(f=func, x0=x0, fprime=grad, fhess_p=hess)
)