def test_huber_continuous(self):
with self.test_session():
epsilon = tf.constant(
1e-10, dtype=tf.float32)
predicted = tf.constant(
[1], dtype=tf.float32, name='predicted')
gold_standard = tf.constant(
[0], dtype=tf.float32, name='gold_standard')
huber_loss_inside_delta = huber_loss(
predicted + epsilon, gold_standard, delta=1.0)
huber_loss_outside_delta = huber_loss(
predicted - epsilon, gold_standard, delta=1.0)
self.assertAlmostEqual(huber_loss_inside_delta.eval(),
huber_loss_outside_delta.eval())
def test_huber_continuous(self):
with self.test_session():
epsilon = tf.constant(1e-10, dtype=tf.float32)
predicted = tf.constant([1], dtype=tf.float32, name='predicted')
gold_standard = tf.constant([0],
dtype=tf.float32,
name='gold_standard')
huber_loss_inside_delta = huber_loss(predicted + epsilon,
gold_standard,
delta=1.0)
huber_loss_outside_delta = huber_loss(predicted - epsilon,
gold_standard,
delta=1.0)
self.assertAlmostEqual(huber_loss_inside_delta.eval(),
huber_loss_outside_delta.eval())
def test_huber_loss(self):
with self.test_session():
predicted = tf.constant([[0, 10], [10, 0], [10, 0], [10, 0]],
dtype=tf.float32,
name='predicted')
gold_standard = tf.constant([[0, 10], [10, 0], [10, 0], [10, 0]],
dtype=tf.float32,
name='gold_standard')
self.assertEqual(huber_loss(predicted, gold_standard).eval(), 0.0)
def test_huber_loss(self):
with self.test_session():
predicted = tf.constant(
[[0, 10], [10, 0], [10, 0], [10, 0]],
dtype=tf.float32, name='predicted')
gold_standard = tf.constant(
[[0, 10], [10, 0], [10, 0], [10, 0]],
dtype=tf.float32, name='gold_standard')
self.assertEqual(huber_loss(predicted, gold_standard).eval(), 0.0)
def test_huber_loss_hand_example(self):
with self.test_session():
# loss should be: mean( 0.2 ** 2/ 2 + (2-0.5) ) == 1.52/2 == 0.76
predicted = tf.constant(
[1.2, 1], dtype=tf.float32, name='predicted')
gold_standard = tf.constant(
[1, 3], dtype=tf.float32, name='gold_standard')
loss = huber_loss(predicted, gold_standard, delta=1.0)
self.assertAlmostEqual(loss.eval(), .76)
def test_huber_loss_hand_example(self):
with self.test_session():
# loss should be: mean( 0.2 ** 2/ 2 + (2-0.5) ) == 1.52/2 == 0.76
predicted = tf.constant([1.2, 1],
dtype=tf.float32,
name='predicted')
gold_standard = tf.constant([1, 3],
dtype=tf.float32,
name='gold_standard')
loss = huber_loss(predicted, gold_standard, delta=1.0)
self.assertAlmostEqual(loss.eval(), .76)
