def testName(self):
# Default name.
loss_fn = loss_lib.MMDLoss()
self.assertEqual(loss_fn.name, "mmd_loss")
# Custom name.
loss_fn = loss_lib.MMDLoss(name="custom_loss")
self.assertEqual(loss_fn.name, "custom_loss")
def testForMMDLoss(self):
loss = utils._get_loss('mmd')
self.assertIsInstance(loss, mmd_loss.MMDLoss)
loss = utils._get_loss('mmd_loss')
self.assertIsInstance(loss, mmd_loss.MMDLoss)
loss = utils._get_loss(mmd_loss.MMDLoss())
self.assertIsInstance(loss, mmd_loss.MMDLoss)
loss_name = 'custom_name'
loss = utils._get_loss(mmd_loss.MMDLoss(name=loss_name))
self.assertIsInstance(loss, mmd_loss.MMDLoss)
self.assertEqual(loss.name, loss_name)
def testRaisesExpectedErrors(self):
kernel = gaussian_kernel.GaussianKernel()
loss_lib.MMDLoss(kernel)
bad_kernel = lambda x: kernel(x) # pylint:disable=unnecessary-lambda
with self.assertRaisesRegex(
TypeError,
"predictions_kernel.*must be.*MinDiffKernel.*string.*4.*int"):
loss_lib.MMDLoss(4)
with self.assertRaisesRegex(
TypeError,
"predictions_kernel.*must be.*MinDiffKernel.*string.*lambda"):
loss_lib.MMDLoss(bad_kernel)
def testGaussianKernelNegativeCorrelationNoWeights(self):
loss_fn = loss_lib.MMDLoss(predictions_transform=tf.sigmoid)
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
predictions = tf.constant([[0.12], [0.7], [0.2], [0.86], [0.32]])
loss_value = loss_fn(membership, predictions)
self.assertAllClose(1.464054, loss_value)
def testLaplacianKernelNegativeCorrelationNoWeights(self):
loss_fn = mmd_lib.MMDLoss("laplace", predictions_transform=tf.sigmoid)
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
predictions = tf.constant([[0.12], [0.7], [0.2], [0.86], [0.32]])
loss_value = loss_fn(membership, predictions)
self.assertAllClose(1.119343, loss_value)
def testLaplacianKernelNoWeights(self):
loss_fn = loss_lib.MMDLoss("laplace", predictions_transform=tf.sigmoid)
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
predictions = tf.constant([[0.3], [0.1], [0.86], [0.06], [0.75]])
loss_value = loss_fn(membership, predictions)
self.assertAllClose(0.970659, loss_value)
def testGaussianKernelNoWeights(self):
loss_fn = loss_lib.MMDLoss(predictions_transform=tf.sigmoid)
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
predictions = tf.constant([[0.3], [0.1], [0.86], [0.06], [0.75]])
loss_value = loss_fn(membership, predictions)
self.assertAllClose(1.083264, loss_value)
def testGaussianKernelWithWeights(self):
loss_fn = loss_lib.MMDLoss(predictions_transform=tf.sigmoid)
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
predictions = tf.constant([[0.3], [0.1], [0.86], [0.06], [0.75]])
sample_weights = [[1.0], [2.0], [2.5], [1.2], [0.9]]
loss_value = loss_fn(membership, predictions, sample_weights)
self.assertAllClose(1.293555, loss_value)
def testGaussianKernelNegativeCorrelationWithWeights(self):
loss_fn = loss_lib.MMDLoss(predictions_transform=tf.sigmoid)
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
predictions = tf.constant([[0.12], [0.7], [0.2], [0.86], [0.32]])
sample_weights = [[1.0], [2.0], [2.5], [1.2], [0.9]]
loss_value = loss_fn(membership, predictions, sample_weights)
self.assertAllClose(1.481506, loss_value)
def testGaussianKernelAllZeroWeights(self):
loss_fn = loss_lib.MMDLoss()
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
predictions = tf.constant([[0.12], [0.7], [0.2], [0.86], [0.32]])
sample_weights = [[0.0], [0.0], [0.0], [0.0], [0.0]]
loss_value = loss_fn(membership, predictions, sample_weights)
self.assertAllClose(0, loss_value)
def testLaplacianKernelWithWeights(self):
loss_fn = mmd_lib.MMDLoss("laplace", predictions_transform=tf.sigmoid)
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
predictions = tf.constant([[0.3], [0.1], [0.86], [0.06], [0.75]])
sample_weights = tf.constant([[1.0], [2.0], [2.5], [1.2], [0.9]])
loss_value = loss_fn(membership, predictions, sample_weights)
self.assertAllClose(1.122741, loss_value)
def testGaussianKernelMultiDimTensor(self):
loss_fn = loss_lib.MMDLoss(predictions_transform=tf.sigmoid)
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0], [1.0], [0.0]])
predictions = tf.constant([[0.0, 0.5], [1.0, 0.0], [0.8, 0.0], [0.0, 0.8],
[0.75, 0.8], [0.2, 0.4], [0.5, 0.1]])
sample_weights = [[1.0], [1.0], [1.0], [1.0], [0.0], [0.0], [0.0]]
loss_value = loss_fn(membership, predictions, sample_weights)
self.assertAllClose(0.262881, loss_value)
def testSerialization(self):
loss = loss_lib.MMDLoss()
serialized_loss = tf.keras.utils.serialize_keras_object(loss)
deserialized_loss = tf.keras.utils.deserialize_keras_object(serialized_loss)
self.assertIsInstance(deserialized_loss, loss_lib.MMDLoss)
self.assertIsNone(deserialized_loss.predictions_transform)
self.assertIsInstance(deserialized_loss.predictions_kernel,
gaussian_kernel.GaussianKernel)
self.assertEqual(deserialized_loss.name, loss.name)
def testGaussianKernelSomeNegativeWeights(self):
loss_fn = loss_lib.MMDLoss()
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
predictions = tf.constant([[0.12], [0.7], [0.2], [0.86], [0.32]])
sample_weights = [[1.0], [2.0], [-2.5], [1.2], [0.9]]
with self.assertRaisesRegex(
tf.errors.InvalidArgumentError,
"sample_weight.*cannot contain any negative weights"):
loss_value = loss_fn(membership, predictions, sample_weights)
if not tf.executing_eagerly():
with self.cached_session() as sess:
sess.run(loss_value)
def testSerializationWithTransformAndKernel(self):
predictions_fn = lambda x: x * 5.1 # Arbitrary operation.
loss = loss_lib.MMDLoss(
predictions_transform=predictions_fn, kernel="laplacian")
serialized_loss = tf.keras.utils.serialize_keras_object(loss)
deserialized_loss = tf.keras.utils.deserialize_keras_object(serialized_loss)
self.assertIsInstance(deserialized_loss, loss_lib.MMDLoss)
val = 7 # Arbitrary value.
self.assertEqual(
deserialized_loss.predictions_transform(val), predictions_fn(val))
self.assertIsInstance(deserialized_loss.predictions_kernel,
laplacian_kernel.LaplacianKernel)
self.assertEqual(deserialized_loss.name, loss.name)
def testLaplacianGradientsAllZeroWeights(self):
loss_fn = mmd_lib.MMDLoss("laplace", predictions_transform=tf.sigmoid)
variables = tf.constant([[0.1], [0.3], [0.5], [0.7], [0.9]])
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
sample_weights = tf.constant([[0.0], [0.0], [0.0], [0.0], [0.0]])
with tf.GradientTape() as tape:
tape.watch(variables)
predictions = variables * 3 # arbitrary linear operation.
loss_value = loss_fn(membership, predictions, sample_weights)
gradients = tape.gradient(loss_value, variables)
# Gradients should all be 0 for weights that are all 0.
expected_gradients = [[0.0], [0.0], [0.0], [0.0], [0.0]]
self.assertAllClose(expected_gradients, gradients)
def testGaussianGradients(self):
loss_fn = mmd_lib.MMDLoss(predictions_transform=tf.sigmoid)
variables = tf.constant([[0.1], [0.3], [0.5], [0.7], [0.9]])
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
sample_weights = tf.constant([[1.0], [2.0], [2.5], [1.2], [0.9]])
with tf.GradientTape() as tape:
tape.watch(variables)
predictions = variables * 3 # arbitrary linear operation.
loss_value = loss_fn(membership, predictions, sample_weights)
gradients = tape.gradient(loss_value, variables)
# Assert that gradient computations are non trivial and do not change based
# on loss implementation.
expected_gradients = [[-0.85786223], [-1.8886726], [1.1220325],
[0.5379708], [-0.02113436]]
self.assertAllClose(expected_gradients, gradients)
def testLaplacianGradients(self):
loss_fn = mmd_lib.MMDLoss("laplace", predictions_transform=tf.sigmoid)
variables = tf.constant([[0.1], [0.3], [0.5], [0.7], [0.9]])
membership = tf.constant([[1.0], [0.0], [1.0], [0.0], [1.0]])
sample_weights = tf.constant([[1.0], [2.0], [2.5], [1.2], [0.9]])
with tf.GradientTape() as tape:
tape.watch(variables)
predictions = variables * 3 # arbitrary linear operation.
loss_value = loss_fn(membership, predictions, sample_weights)
gradients = tape.gradient(loss_value, variables)
# Assert that gradient computations are non trivial and do not change based
# on loss implementation.
expected_gradients = [[-0.40014654], [-0.7467264], [0.39164954],
[0.10974818], [0.08942482]]
self.assertAllClose(expected_gradients, gradients)
def testEnableSummaryHistogram(self):
loss = mmd_lib.MMDLoss()
self.assertTrue(loss.enable_summary_histogram)
loss = mmd_lib.MMDLoss(enable_summary_histogram=False)
self.assertFalse(loss.enable_summary_histogram)
