def test_sample_weighted(self):
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = tf.constant([4, 8, 12, 8, 1, 3], shape=(2, 3), dtype=tf.dtypes.float32)
sample_weight = tf.constant([1.2, 3.4], shape=(2, 1))
loss = pin_obj(y_true, y_pred, sample_weight=sample_weight)
self.assertAlmostEqual(self.evaluate(loss), 40.7 / 6, 3)
def test_timestep_weighted():
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3, 1))
y_pred = tf.constant([4, 8, 12, 8, 1, 3], shape=(2, 3, 1), dtype=tf.dtypes.float32)
sample_weight = tf.constant([3, 6, 5, 0, 4, 2], shape=(2, 3))
loss = pin_obj(y_true, y_pred, sample_weight=sample_weight)
np.testing.assert_almost_equal(loss, 41.5 / 6, 3)
def test_invalid_sample_weight():
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3, 1))
y_pred = tf.constant([4, 8, 12, 8, 1, 3], shape=(2, 3, 1))
sample_weight = tf.constant([3, 6, 5, 0], shape=(2, 2))
with pytest.raises(tf.errors.InvalidArgumentError, match="Incompatible shapes"):
pin_obj(y_true, y_pred, sample_weight=sample_weight)
def test_invalid_sample_weight(self):
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3, 1))
y_pred = tf.constant([4, 8, 12, 8, 1, 3], shape=(2, 3, 1))
sample_weight = tf.constant([3, 6, 5, 0], shape=(2, 2))
with self.assertRaisesRegexp(ValueError,
"weights can not be broadcast to values"):
pin_obj(y_true, y_pred, sample_weight=sample_weight)
def test_sum_reduction():
pin_obj = quantiles.PinballLoss(reduction=tf.keras.losses.Reduction.SUM)
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = tf.constant([4, 8, 12, 8, 1, 3],
shape=(2, 3),
dtype=tf.dtypes.float32)
loss = pin_obj(y_true, y_pred, sample_weight=2.3)
np.testing.assert_almost_equal(loss, 12.65, 3)
def test_zero_weighted():
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = tf.constant([4, 8, 12, 8, 1, 3],
shape=(2, 3),
dtype=tf.dtypes.float32)
loss = pin_obj(y_true, y_pred, sample_weight=0)
np.testing.assert_almost_equal(loss, 0.0, 3)
def test_unweighted_quantile_100pc(self):
pin_obj = quantiles.PinballLoss(tau=1.0)
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = tf.constant([4, 8, 12, 8, 1, 3],
shape=(2, 3),
dtype=tf.dtypes.float32)
loss = pin_obj(y_true, y_pred)
self.assertAlmostEqual(self.evaluate(loss), 0.6666, 3)
def test_unweighted_quantile_100pc():
pin_obj = quantiles.PinballLoss(tau=1.0)
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = tf.constant([4, 8, 12, 8, 1, 3],
shape=(2, 3),
dtype=tf.dtypes.float32)
loss = pin_obj(y_true, y_pred)
np.testing.assert_almost_equal(loss, 0.6666, 3)
def test_sum_reduction(self):
pin_obj = quantiles.PinballLoss(
reduction=tf.keras.losses.Reduction.SUM)
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = tf.constant([4, 8, 12, 8, 1, 3],
shape=(2, 3),
dtype=tf.dtypes.float32)
loss = pin_obj(y_true, y_pred, sample_weight=2.3)
self.assertAlmostEqual(self.evaluate(loss), 12.65, 3)
def test_invalid_sample_weight(self):
self.skipTest("Failing. See https://github.com/tensorflow/addons/issues/1202")
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3, 1))
y_pred = tf.constant([4, 8, 12, 8, 1, 3], shape=(2, 3, 1))
sample_weight = tf.constant([3, 6, 5, 0], shape=(2, 2))
with self.assertRaisesRegexp(
ValueError, "weights can not be broadcast to values"
):
pin_obj(y_true, y_pred, sample_weight=sample_weight)
def test_no_reduction(self):
pin_obj = quantiles.PinballLoss(
reduction=tf.keras.losses.Reduction.NONE)
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3))
y_pred = tf.constant([4, 8, 12, 8, 1, 3],
shape=(2, 3),
dtype=tf.dtypes.float32)
loss = pin_obj(y_true, y_pred, sample_weight=2.3)
loss = self.evaluate(loss)
self.assertArrayNear(loss, [5.3666, 7.28333], 1e-3)
def test_timestep_weighted(self):
self.skipTest("Failing. See https://github.com/tensorflow/addons/issues/1202")
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3, 1))
y_pred = tf.constant(
[4, 8, 12, 8, 1, 3], shape=(2, 3, 1), dtype=tf.dtypes.float32
)
sample_weight = tf.constant([3, 6, 5, 0, 4, 2], shape=(2, 3))
loss = pin_obj(y_true, y_pred, sample_weight=sample_weight)
self.assertAlmostEqual(self.evaluate(loss), 41.5 / 6, 3)
def test_timestep_weighted():
pytest.skip(
"Failing. See https://github.com/tensorflow/addons/issues/1202")
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3, 1))
y_pred = tf.constant([4, 8, 12, 8, 1, 3],
shape=(2, 3, 1),
dtype=tf.dtypes.float32)
sample_weight = tf.constant([3, 6, 5, 0, 4, 2], shape=(2, 3))
loss = pin_obj(y_true, y_pred, sample_weight=sample_weight)
np.testing.assert_almost_equal(loss, 41.5 / 6, 3)
def test_invalid_sample_weight():
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([1, 9, 2, -5, -2, 6], shape=(2, 3, 1))
y_pred = tf.constant([4, 8, 12, 8, 1, 3], shape=(2, 3, 1))
sample_weight = tf.constant([3, 6, 5, 0], shape=(2, 2))
if LooseVersion(tf.__version__) >= "2.2":
with pytest.raises(tf.errors.InvalidArgumentError,
match="Incompatible shapes"):
pin_obj(y_true, y_pred, sample_weight=sample_weight)
else:
with pytest.raises(ValueError,
match="weights can not be broadcast to values"):
pin_obj(y_true, y_pred, sample_weight=sample_weight)
def test_all_correct_unweighted():
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([4, 8, 12, 8, 1, 3], shape=(2, 3))
loss = pin_obj(y_true, y_true)
assert loss == 0
def test_config():
pin_obj = quantiles.PinballLoss(
reduction=tf.keras.losses.Reduction.SUM, name="pin_1"
)
assert pin_obj.name == "pin_1"
assert pin_obj.reduction == tf.keras.losses.Reduction.SUM
def test_config(self):
pin_obj = quantiles.PinballLoss(
reduction=tf.keras.losses.Reduction.SUM, name="pin_1")
self.assertEqual(pin_obj.name, "pin_1")
self.assertEqual(pin_obj.reduction, tf.keras.losses.Reduction.SUM)
def test_all_correct_unweighted(self):
pin_obj = quantiles.PinballLoss()
y_true = tf.constant([4, 8, 12, 8, 1, 3], shape=(2, 3))
loss = pin_obj(y_true, y_true)
self.assertAlmostEqual(self.evaluate(loss), 0.0, 3)
