def test_ridge_grad_cov():
"""Test ovk.OVKRidgeRisk gradient with finite differences."""
K = ovk.DecomposableKernel(A=eye(2))
risk = ovk.OVKRidgeRisk(0.01)
assert check_grad(lambda *args: risk.functional_grad_val(*args)[0],
lambda *args: risk.functional_grad_val(*args)[1],
randn(X.shape[0] * y.shape[1]), y.ravel(), K(X,
X)) < 1e-3
def test_rff_ridge_grad_cov():
"""Test ovk.ORFFRidgeRisk gradient with finite differences."""
K = ovk.DecomposableKernel(A=eye(2))
risk = ovk.ORFFRidgeRisk(0.01)
D = 100
assert check_grad(lambda *args: risk.functional_grad_val(*args)[0],
lambda *args: risk.functional_grad_val(*args)[1],
randn(D * y.shape[1]), y.ravel(), K.get_orff_map(
X, D), K) < 1e-3
def test_grad_val():
"""Test whether ovk.OVKRidgeRisk gradient and val are consistents."""
K = ovk.DecomposableKernel(A=eye(2))
risk = ovk.OVKRidgeRisk(0.01)
C = randn(X.shape[0] * y.shape[1])
v = risk(C, y.ravel(), K(X, X))
g = risk.functional_grad(C, y.ravel(), K(X, X))
vg = risk.functional_grad_val(C, y.ravel(), K(X, X))
assert_allclose(v, vg[0])
assert_allclose(g, vg[1])
def test_rff_ridge_hinge_grad():
"""Test ovk.ORFFRidgeRisk gradient with finite differences."""
K = ovk.DecomposableKernel(A=eye(3))
risk = ovk.ORFFRidgeRisk(0.01, 'Hinge')
D = 100
y = one_hot(randint(0, 3, X.shape[0]), 3)
vl = check_grad(lambda *args: risk.functional_grad_val(*args)[0],
lambda *args: risk.functional_grad_val(*args)[1],
rand(D * y.shape[1]), y.ravel(), K.get_orff_map(X, D), K)
assert vl < 1e-3
def test_rff_ridge_SCSVM():
K = ovk.DecomposableKernel(A=eye(2))
risk = ovk.ORFFRidgeRisk(0.01, 'SCSVM')
D = 100
y = one_hot(randint(0, 3, X.shape[0]), 3)
sc = ovk.preprocessing.SimplexCoding()
y = sc.fit_transform(y)
vl = check_grad(lambda *args: risk.functional_grad_val(*args)[0],
lambda *args: risk.functional_grad_val(*args)[1],
rand(D * y.shape[1]), y.ravel(), K.get_orff_map(X, D), K)
assert vl < 1e-3
def test_rff_grad_val_least_squares():
"""Test whether ovk.ORFFRidgeRisk least square gradient and val are
consistents."""
K = ovk.DecomposableKernel(A=eye(2))
risk = ovk.ORFFRidgeRisk(0.01)
D = 100
C = randn(D * y.shape[1])
v = risk(C, y.ravel(), K.get_orff_map(X, D), K)
g = risk.functional_grad(C, y.ravel(), K.get_orff_map(X, D), K)
vg = risk.functional_grad_val(C, y.ravel(), K.get_orff_map(X, D), K)
assert_allclose(v, vg[0])
assert_allclose(g, vg[1])
def test_rff_grad_val_SCSVM():
K = ovk.DecomposableKernel(A=eye(2))
risk = ovk.ORFFRidgeRisk(0.01, 'SCSVM')
y = one_hot(randint(0, 3, X.shape[0]), 3)
sv = ovk.preprocessing.SimplexCoding()
y = sv.fit_transform(y)
D = 100
C = rand(D * y.shape[1])
v = risk(C, y.ravel(), K.get_orff_map(X, D), K)
g = risk.functional_grad(C, y.ravel(), K.get_orff_map(X, D), K)
vg = risk.functional_grad_val(C, y.ravel(), K.get_orff_map(X, D), K)
assert_allclose(v, vg[0])
assert_allclose(g, vg[1])
def test_rff_grad_val_hinge():
"""Test whether ovk.ORFFRidgeRisk hinge gradient and val are
consistents."""
K = ovk.DecomposableKernel(A=eye(3))
risk = ovk.ORFFRidgeRisk(0.01, 'Hinge')
y = one_hot(randint(0, 3, X.shape[0]), 3)
D = 100
C = randn(D * y.shape[1])
v = risk(C, y.ravel(), K.get_orff_map(X, D), K)
g = risk.functional_grad(C, y.ravel(), K.get_orff_map(X, D), K)
vg = risk.functional_grad_val(C, y.ravel(), K.get_orff_map(X, D), K)
assert_allclose(v, vg[0])
assert_allclose(g, vg[1])