def test_val_grad(model):
criterion = HeldOutLogistic(idx_val, idx_val)
algo = Forward()
val_fwd, grad_fwd = criterion.get_val_grad(model,
X,
y,
log_alpha,
algo.get_beta_jac_v,
tol=tol)
criterion = HeldOutLogistic(idx_val, idx_val)
algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000)
val_imp_fwd, grad_imp_fwd = criterion.get_val_grad(model,
X,
y,
log_alpha,
algo.get_beta_jac_v,
tol=tol)
criterion = HeldOutLogistic(idx_val, idx_val)
algo = Implicit()
val_imp, grad_imp = criterion.get_val_grad(model,
X,
y,
log_alpha,
algo.get_beta_jac_v,
tol=tol)
assert np.allclose(val_fwd, val_imp_fwd, atol=1e-4)
assert np.allclose(grad_fwd, grad_imp_fwd, atol=1e-4)
assert np.allclose(val_imp_fwd, val_imp, atol=1e-4)
# for the implcit the conjugate grad does not converge
# hence the rtol=1e-2
assert np.allclose(grad_imp_fwd, grad_imp, rtol=1e-2)
def test_val_grad(model):
#######################################################################
# Not all methods computes the full Jacobian, but all
# compute the gradients
# check that the gradient returned by all methods are the same
criterion = HeldOutLogistic(X_val, y_val, model)
algo = Forward()
val_fwd, grad_fwd = criterion.get_val_grad(log_C,
algo.get_beta_jac_v,
tol=tol)
criterion = HeldOutLogistic(X_val, y_val, model)
algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=100)
val_imp_fwd, grad_imp_fwd = criterion.get_val_grad(log_C,
algo.get_beta_jac_v,
tol=tol)
criterion = HeldOutLogistic(X_val, y_val, model)
algo = Implicit()
val_imp, grad_imp = criterion.get_val_grad(log_C,
algo.get_beta_jac_v,
tol=tol)
assert np.allclose(val_fwd, val_imp_fwd)
assert np.allclose(grad_fwd, grad_imp_fwd)
assert np.allclose(val_imp_fwd, val_imp)
assert np.allclose(grad_imp_fwd, grad_imp, atol=1e-5)
def test_val_grad_custom(model, model_custom):
criterion = HeldOutLogistic(X_val, y_val, model)
algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000)
val, grad = criterion.get_val_grad(
log_alpha, algo.get_beta_jac_v, tol=tol)
criterion = HeldOutLogistic(X_val, y_val, model_custom)
algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000)
val_custom, grad_custom = criterion.get_val_grad(
log_alpha, algo.get_beta_jac_v, tol=tol)
assert np.allclose(val, val_custom)
assert np.allclose(grad, grad_custom)