def testForLaplacianKernel(self):
kernel = utils._get_kernel('laplace')
self.assertIsInstance(kernel, laplacian_kernel.LaplacianKernel)
kernel = utils._get_kernel('laplace_Kernel')
self.assertIsInstance(kernel, laplacian_kernel.LaplacianKernel)
kernel = utils._get_kernel('laplacian')
self.assertIsInstance(kernel, laplacian_kernel.LaplacianKernel)
kernel = utils._get_kernel('laplacian_kernel')
self.assertIsInstance(kernel, laplacian_kernel.LaplacianKernel)
kernel = utils._get_kernel(laplacian_kernel.LaplacianKernel())
self.assertIsInstance(kernel, laplacian_kernel.LaplacianKernel)
kernel_length = 3
kernel = utils._get_kernel(laplacian_kernel.LaplacianKernel(kernel_length))
self.assertIsInstance(kernel, laplacian_kernel.LaplacianKernel)
self.assertEqual(kernel.kernel_length, kernel_length)
def testLaplacianKernel(self):
# Under certain conditions, the kernel output resembles the identity
# (within the tolerance of the checks).
# Specifically, this is true when the minimum distance, d, between any two
# elements in the input is greater than D for D = -ln(tol) * kernel_length
# where:
# - tol: assertion tolerance (these tests ues 1e-6)
# - d: minimum{norm(x_i - y_i)
# for x_i = x[i, :], y_j = y[j, :] for all i neq j}
# - kernel_length: argument to kernel initialization
#
# In the test below, we use max(kernel_length) == 0.1 which gives us
# max(D)~=1.38. Given this bound, we sometines use d=2 (meaning that each
# element is at least 2 greater or smaller than every other element in the
# input). When we do this, we expect the output to be close to the identity.
for kernel_length in [0.05, 0.075, 0.1]:
laplacian_kernel_fn = laplacian_kernel.LaplacianKernel(
kernel_length)
kernel_val = laplacian_kernel_fn(tf.constant([[2.0], [0.0]]))
self.assertAllClose(kernel_val, tf.eye(2))
kernel_val = laplacian_kernel_fn(tf.constant([[1.0], [8.0],
[3.0]]))
self.assertAllClose(kernel_val, tf.eye(3))
kernel_val = laplacian_kernel_fn(
tf.constant([[1.0, 3.0], [5.0, 7.0], [9.0, 11.0]]))
self.assertAllClose(kernel_val, tf.eye(3))
# Vector where all elements are equal should result in matrix of all 1s.
for val in [0.2, 1.0, 2.79]: # Arbitrary values.
kernel_val = laplacian_kernel_fn(
tf.constant([[val, val], [val, val]]))
# self.assertAllClose(kernel_val, tf.eye(3))
self.assertAllClose(kernel_val, [[1.0, 1.0], [1.0, 1.0]])
# If the two tensors are equal, it should be the same as a single tensor.
# Note that we pick d < 1 so that the result is not the identity.
tensor = tf.constant([[1.0], [1.01]])
single_kernel_val = laplacian_kernel_fn(tensor)
# Assert that the result is nontrivial (not the identity).
self.assertNotAllClose(tf.eye(2), single_kernel_val)
double_kernel_val = laplacian_kernel_fn(tensor, tensor)
self.assertAllClose(single_kernel_val, double_kernel_val)
self.assertIsNot(single_kernel_val, double_kernel_val)
# If the delta is the same, then the result should be the same.
# Note that we pick d < 1 so that the result is not the identity.
tensor1 = tf.constant([[1.0], [7.0]])
tensor2 = tf.constant([[1.0], [5.3]])
delta = tf.constant([[0.0], [0.2]])
kernel_val1 = laplacian_kernel_fn(tensor1, tensor1 + delta)
# Assert that the result is nontrivial (not the identity).
self.assertNotAllClose(tf.eye(2), kernel_val1)
kernel_val2 = laplacian_kernel_fn(tensor2, tensor2 + delta)
self.assertAllClose(kernel_val1, kernel_val2)
def testGetAndFromConfig(self):
kernel_length = 5 # Arbitrary value.
kernel = laplacian_kernel.LaplacianKernel(kernel_length)
kernel_from_config = laplacian_kernel.LaplacianKernel.from_config(
kernel.get_config())
self.assertIsInstance(kernel_from_config, laplacian_kernel.LaplacianKernel)
self.assertEqual(kernel_from_config.kernel_length, kernel_length)
def testSerialization(self):
kernel_length = 5 # Arbitrary value.
kernel = laplacian_kernel.LaplacianKernel(kernel_length)
serialized_kernel = tf.keras.utils.serialize_keras_object(kernel)
deserialized_kernel = tf.keras.utils.deserialize_keras_object(
serialized_kernel)
self.assertIsInstance(deserialized_kernel, laplacian_kernel.LaplacianKernel)
self.assertEqual(deserialized_kernel.kernel_length, kernel_length)
def testkernelAttributes(self):
loss = CustomLoss(membership_kernel='gauss',
predictions_kernel='laplace')
self.assertIsInstance(loss.membership_kernel,
gaussian_kernel.GaussianKernel)
self.assertIsInstance(loss.predictions_kernel,
laplacian_kernel.LaplacianKernel)
kernel = gaussian_kernel.GaussianKernel()
loss = CustomLoss(predictions_kernel=kernel)
self.assertIs(loss.predictions_kernel, kernel)
self.assertIsNone(loss.membership_kernel)
kernel = laplacian_kernel.LaplacianKernel()
loss = CustomLoss(membership_kernel=kernel)
self.assertIs(loss.membership_kernel, kernel)
self.assertIsNone(loss.predictions_kernel)
