def test_toy_ex_lmnn(X, y, loss):
"""Test that the loss give the right result on a toy example"""
L = np.array([[1]])
lmnn = LMNN(k=1, regularization=0.5)
k = lmnn.k
reg = lmnn.regularization
X, y = lmnn._prepare_inputs(X, y, dtype=float, ensure_min_samples=2)
num_pts, n_components = X.shape
unique_labels, label_inds = np.unique(y, return_inverse=True)
lmnn.labels_ = np.arange(len(unique_labels))
lmnn.components_ = np.eye(n_components)
target_neighbors = lmnn._select_targets(X, label_inds)
# sum outer products
dfG = _sum_outer_products(X, target_neighbors.flatten(),
np.repeat(np.arange(X.shape[0]), k))
# storage
a1 = [None] * k
a2 = [None] * k
for nn_idx in range(k):
a1[nn_idx] = np.array([])
a2[nn_idx] = np.array([])
# assert that the loss equals the one computed by hand
assert lmnn._loss_grad(X, L.reshape(-1, X.shape[1]), dfG, k, reg,
target_neighbors, label_inds)[1] == loss
def test_loss_grad_lbfgs(self):
"""Test gradient of loss function
Assert that the gradient is almost equal to its finite differences
approximation.
"""
rng = np.random.RandomState(42)
X, y = make_classification(random_state=rng)
L = rng.randn(rng.randint(1, X.shape[1] + 1), X.shape[1])
lmnn = LMNN()
k = lmnn.k
reg = lmnn.regularization
X, y = lmnn._prepare_inputs(X, y, dtype=float,
ensure_min_samples=2)
num_pts, n_components = X.shape
unique_labels, label_inds = np.unique(y, return_inverse=True)
lmnn.labels_ = np.arange(len(unique_labels))
lmnn.components_ = np.eye(n_components)
target_neighbors = lmnn._select_targets(X, label_inds)
# sum outer products
dfG = _sum_outer_products(X, target_neighbors.flatten(),
np.repeat(np.arange(X.shape[0]), k))
# initialize L
def loss_grad(flat_L):
return lmnn._loss_grad(X, flat_L.reshape(-1, X.shape[1]), dfG,
k, reg, target_neighbors, label_inds)
def fun(x):
return loss_grad(x)[1]
def grad(x):
return loss_grad(x)[0].ravel()
# compute relative error
epsilon = np.sqrt(np.finfo(float).eps)
rel_diff = (check_grad(fun, grad, L.ravel()) /
np.linalg.norm(approx_fprime(L.ravel(), fun, epsilon)))
np.testing.assert_almost_equal(rel_diff, 0., decimal=5)
