def testNormalizeL1WithZero(self):
with self.cached_session() as sess:
tensor = tf.constant(0.0, shape=[2, 3])
normalized_tensor = utils.normalize(tensor, 'l1')
expected_tensor = tf.constant(0.0, shape=[2, 3])
sess.run(normalized_tensor)
self.assertAllEqual(normalized_tensor, expected_tensor)
def testNormalizeInfWithOnes(self):
with self.cached_session() as sess:
target_tensor = tf.constant(1.0, shape=[2, 4])
normalized_tensor = utils.normalize(target_tensor, 'infinity')
expected_tensor = tf.constant(1.0, shape=[2, 4])
sess.run(normalized_tensor)
self.assertAllEqual(normalized_tensor, expected_tensor)
def testNormalizeL1(self):
# target_tensor = [[1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0]]
target_tensor = tf.constant(1.0, shape=[2, 4])
normalized_tensor = self.evaluate(utils.normalize(target_tensor, 'l1'))
# L1 norm of target_tensor (other than batch/1st dim) is [4, 4]; therefore
# normalized_tensor = [[0.25, 0.25, 0.25, 0.25], [0.25, 0.25, 0.25, 0.25]]
expected_tensor = tf.constant(0.25, shape=[2, 4])
self.assertAllEqual(normalized_tensor, expected_tensor)
def testNormalizeInf(self):
with self.cached_session() as sess:
target_tensor = tf.constant([[1.0, 2.0, -4.0], [-1.0, 5.0, -3.0]])
normalized_tensor = utils.normalize(target_tensor, 'infinity')
expected_tensor = tf.constant([[0.25, 0.5, -1.0],
[-0.2, 1.0, -0.6]])
sess.run(normalized_tensor)
self.assertAllEqual(normalized_tensor, expected_tensor)
def testNormalizeL2(self):
# target_tensor = [[1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0]]
target_tensor = tf.constant(1.0, shape=[2, 4])
normalized_tensor = self.evaluate(utils.normalize(target_tensor, 'l2'))
# L2 norm of target_tensor (other than batch/1st dim) is:
# [sqrt(1^2+1^2+1^2+1^2), sqrt(1^2+1^2+1^2+1^2)] = [2, 2], and therefore
# normalized_tensor = [[0.5, 0.5, 0.5, 0.5], [0.5, 0.5, 0.5, 0.5]]
expected_tensor = tf.constant(0.5, shape=[2, 4])
self.assertAllEqual(normalized_tensor, expected_tensor)
def testNormalizeL1(self):
with self.cached_session() as sess:
# target_tensor = [[1.0, 1.0, 1.0, 1.0], [1.0, 1.0, 1.0, 1.0]]
target_tensor = tf.constant(1.0, shape=[2, 4])
normalized_tensor = utils.normalize(target_tensor, 'l1')
# L1 norm of target_tensor (other than batch/1st dim) is [4, 4]; therefore
# target_tensor = [[0.25, 0.25, 0.25, 0.25], [0.25, 0.25, 0.25, 0.25]]
expected_tensor = tf.constant(0.25, shape=[2, 4])
sess.run(normalized_tensor)
self.assertAllEqual(normalized_tensor, expected_tensor)
def testNormalizeL1WithZero(self):
tensor = tf.constant(0.0, shape=[2, 3])
normalized_tensor = self.evaluate(utils.normalize(tensor, 'l1'))
expected_tensor = tf.constant(0.0, shape=[2, 3])
self.assertAllEqual(normalized_tensor, expected_tensor)
def testNormalizeInfWithOnes(self):
target_tensor = tf.constant(1.0, shape=[2, 4])
normalized_tensor = self.evaluate(
utils.normalize(target_tensor, 'infinity'))
expected_tensor = tf.constant(1.0, shape=[2, 4])
self.assertAllEqual(normalized_tensor, expected_tensor)
def testNormalizeInf(self):
target_tensor = tf.constant([[1.0, 2.0, -4.0], [-1.0, 5.0, -3.0]])
normalized_tensor = self.evaluate(
utils.normalize(target_tensor, 'infinity'))
expected_tensor = tf.constant([[0.25, 0.5, -1.0], [-0.2, 1.0, -0.6]])
self.assertAllEqual(normalized_tensor, expected_tensor)
def normalize_with_mask(perturbation):
perturbation = utils.apply_feature_mask(perturbation,
neighbor_config.feature_mask)
return utils.normalize(perturbation, neighbor_config.adv_grad_norm)
