def testDtypeFlexibility(self):
"""Tests the loss on inputs of varying data types."""
shape = [20, 3]
logits = np.random.randn(*shape)
targets = tf.truncated_normal(shape)
positive_weights = tf.constant(3, dtype=tf.int64)
negative_weights = 1
loss = util.weighted_sigmoid_cross_entropy_with_logits(
targets, logits, positive_weights, negative_weights)
with self.test_session():
self.assertEqual(loss.eval().dtype, np.float)
def testDtypeFlexibility(self):
"""Tests the loss on inputs of varying data types."""
shape = [20, 3]
logits = np.random.randn(*shape)
targets = tf.truncated_normal(shape)
positive_weights = tf.constant(3, dtype=tf.int64)
negative_weights = 1
loss = util.weighted_sigmoid_cross_entropy_with_logits(
targets, logits, positive_weights, negative_weights)
with self.test_session():
self.assertEqual(loss.eval().dtype, np.float)
def testWeightedSigmoidCrossEntropy(self, batch_size, num_labels, weight):
"""Tests thats the tf and numpy functions agree on many instances."""
x_shape = [batch_size, num_labels]
targets = np.random.random_sample(x_shape).astype(np.float32)
logits = np.random.randn(*x_shape).astype(np.float32)
with self.test_session():
loss = util.weighted_sigmoid_cross_entropy_with_logits(
targets,
logits,
weight,
1.0,
name='weighted-loss')
expected = weighted_sigmoid_cross_entropy(targets, logits, weight)
self.assertAllClose(expected, loss.eval(), atol=0.000001)
def testTrivialCompatibilityWithSigmoidCrossEntropy(self):
"""Tests compatibility with unweighted function with weight 1.0."""
x_shape = [300, 10]
targets = np.random.random_sample(x_shape).astype(np.float32)
logits = np.random.randn(*x_shape).astype(np.float32)
weighted_loss = util.weighted_sigmoid_cross_entropy_with_logits(
targets,
logits)
expected_loss = (
tf.contrib.nn.deprecated_flipped_sigmoid_cross_entropy_with_logits(
logits, targets))
with self.test_session():
self.assertAllClose(expected_loss.eval(),
weighted_loss.eval(),
atol=0.000001)
def testWeightedSigmoidCrossEntropy(self, batch_size, num_labels, weight):
"""Tests thats the tf and numpy functions agree on many instances."""
x_shape = [batch_size, num_labels]
targets = np.random.random_sample(x_shape).astype(np.float32)
logits = np.random.randn(*x_shape).astype(np.float32)
with self.test_session():
loss = util.weighted_sigmoid_cross_entropy_with_logits(
targets,
logits,
weight,
1.0,
name='weighted-loss')
expected = weighted_sigmoid_cross_entropy(targets, logits, weight)
self.assertAllClose(expected, loss.eval(), atol=0.000001)
def testTrivialCompatibilityWithSigmoidCrossEntropy(self):
"""Tests compatibility with unweighted function with weight 1.0."""
x_shape = [300, 10]
targets = np.random.random_sample(x_shape).astype(np.float32)
logits = np.random.randn(*x_shape).astype(np.float32)
weighted_loss = util.weighted_sigmoid_cross_entropy_with_logits(
targets,
logits)
expected_loss = (
tf.contrib.nn.deprecated_flipped_sigmoid_cross_entropy_with_logits(
logits, targets))
with self.test_session():
self.assertAllClose(expected_loss.eval(),
weighted_loss.eval(),
atol=0.000001)
def testNonTrivialCompatibilityWithSigmoidCrossEntropy(self):
"""Tests use of an arbitrary weight (4.12)."""
x_shape = [300, 10]
targets = np.random.random_sample(x_shape).astype(np.float32)
logits = np.random.randn(*x_shape).astype(np.float32)
weight = 4.12
weighted_loss = util.weighted_sigmoid_cross_entropy_with_logits(
targets,
logits,
weight,
weight)
expected_loss = (
weight *
tf.contrib.nn.deprecated_flipped_sigmoid_cross_entropy_with_logits(
logits, targets))
with self.test_session():
self.assertAllClose(expected_loss.eval(),
weighted_loss.eval(),
atol=0.000001)
def testNonTrivialCompatibilityWithSigmoidCrossEntropy(self):
"""Tests use of an arbitrary weight (4.12)."""
x_shape = [300, 10]
targets = np.random.random_sample(x_shape).astype(np.float32)
logits = np.random.randn(*x_shape).astype(np.float32)
weight = 4.12
weighted_loss = util.weighted_sigmoid_cross_entropy_with_logits(
targets,
logits,
weight,
weight)
expected_loss = (
weight *
tf.contrib.nn.deprecated_flipped_sigmoid_cross_entropy_with_logits(
logits, targets))
with self.test_session():
self.assertAllClose(expected_loss.eval(),
weighted_loss.eval(),
atol=0.000001)
def testDifferentSizeWeightedSigmoidCrossEntropy(self):
"""Tests correctness on 3D tensors.
Tests that the function works as expected when logits is a 3D tensor and
targets is a 2D tensor.
"""
targets_shape = [30, 4]
logits_shape = [targets_shape[0], targets_shape[1], 3]
targets = np.random.random_sample(targets_shape).astype(np.float32)
logits = np.random.randn(*logits_shape).astype(np.float32)
weight_vector = [2.0, 3.0, 13.0]
loss = util.weighted_sigmoid_cross_entropy_with_logits(targets,
logits,
weight_vector)
with self.test_session():
loss = loss.eval()
for i in range(0, len(weight_vector)):
expected = weighted_sigmoid_cross_entropy(targets, logits[:, :, i],
weight_vector[i])
self.assertAllClose(loss[:, :, i], expected, atol=0.000001)
def testGradients(self):
"""Tests that weighted loss gradients behave as expected."""
dummy_tensor = tf.constant(1.0)
positives_shape = [10, 1]
positives_logits = dummy_tensor * tf.Variable(
tf.random_normal(positives_shape) + 1.0)
positives_targets = tf.ones(positives_shape)
positives_weight = 4.6
positives_loss = (
tf.contrib.nn.deprecated_flipped_sigmoid_cross_entropy_with_logits(
positives_logits, positives_targets) * positives_weight)
negatives_shape = [190, 1]
negatives_logits = dummy_tensor * tf.Variable(
tf.random_normal(negatives_shape))
negatives_targets = tf.zeros(negatives_shape)
negatives_weight = 0.9
negatives_loss = (
tf.contrib.nn.deprecated_flipped_sigmoid_cross_entropy_with_logits(
negatives_logits, negatives_targets) * negatives_weight)
all_logits = tf.concat([positives_logits, negatives_logits], 0)
all_targets = tf.concat([positives_targets, negatives_targets], 0)
weighted_loss = tf.reduce_sum(
util.weighted_sigmoid_cross_entropy_with_logits(
all_targets, all_logits, positives_weight, negatives_weight))
weighted_gradients = tf.gradients(weighted_loss, dummy_tensor)
expected_loss = tf.add(
tf.reduce_sum(positives_loss),
tf.reduce_sum(negatives_loss))
expected_gradients = tf.gradients(expected_loss, dummy_tensor)
with tf.Session() as session:
tf.global_variables_initializer().run()
grad, expected_grad = session.run(
[weighted_gradients, expected_gradients])
self.assertAllClose(grad, expected_grad)
def testDifferentSizeWeightedSigmoidCrossEntropy(self):
"""Tests correctness on 3D tensors.
Tests that the function works as expected when logits is a 3D tensor and
targets is a 2D tensor.
"""
targets_shape = [30, 4]
logits_shape = [targets_shape[0], targets_shape[1], 3]
targets = np.random.random_sample(targets_shape).astype(np.float32)
logits = np.random.randn(*logits_shape).astype(np.float32)
weight_vector = [2.0, 3.0, 13.0]
loss = util.weighted_sigmoid_cross_entropy_with_logits(targets,
logits,
weight_vector)
with self.test_session():
loss = loss.eval()
for i in range(0, len(weight_vector)):
expected = weighted_sigmoid_cross_entropy(targets, logits[:, :, i],
weight_vector[i])
self.assertAllClose(loss[:, :, i], expected, atol=0.000001)
def testGradients(self):
"""Tests that weighted loss gradients behave as expected."""
dummy_tensor = tf.constant(1.0)
positives_shape = [10, 1]
positives_logits = dummy_tensor * tf.Variable(
tf.random_normal(positives_shape) + 1.0)
positives_targets = tf.ones(positives_shape)
positives_weight = 4.6
positives_loss = (
tf.contrib.nn.deprecated_flipped_sigmoid_cross_entropy_with_logits(
positives_logits, positives_targets) * positives_weight)
negatives_shape = [190, 1]
negatives_logits = dummy_tensor * tf.Variable(
tf.random_normal(negatives_shape))
negatives_targets = tf.zeros(negatives_shape)
negatives_weight = 0.9
negatives_loss = (
tf.contrib.nn.deprecated_flipped_sigmoid_cross_entropy_with_logits(
negatives_logits, negatives_targets) * negatives_weight)
all_logits = tf.concat([positives_logits, negatives_logits], 0)
all_targets = tf.concat([positives_targets, negatives_targets], 0)
weighted_loss = tf.reduce_sum(
util.weighted_sigmoid_cross_entropy_with_logits(
all_targets, all_logits, positives_weight, negatives_weight))
weighted_gradients = tf.gradients(weighted_loss, dummy_tensor)
expected_loss = tf.add(
tf.reduce_sum(positives_loss),
tf.reduce_sum(negatives_loss))
expected_gradients = tf.gradients(expected_loss, dummy_tensor)
with tf.Session() as session:
tf.global_variables_initializer().run()
grad, expected_grad = session.run(
[weighted_gradients, expected_gradients])
self.assertAllClose(grad, expected_grad)
