def test_mean_squared_loss_with_invalid_labels(self):
scores = [[1., 3., 2.]]
labels = [[0., -1., 1.]]
with self.cached_session():
self.assertAlmostEqual(ranking_losses._mean_squared_loss(
labels, scores).eval(), (1. + 1.) / 2,
places=5)
def test_mean_squared_loss(self):
scores = [[0.2, 0.5, 0.3], [0.2, 0.3, 0.5], [0.2, 0.3, 0.5]]
labels = [[0., 0., 1.], [0., 0., 2.], [0., 0., 0.]]
weights = [[2.], [1.], [1.]]
with self.cached_session():
self.assertAlmostEqual(
ranking_losses._mean_squared_loss(labels, scores).eval(),
(_mean_squared_error(labels[0], scores[0]) +
_mean_squared_error(labels[1], scores[1]) +
_mean_squared_error(labels[2], scores[2])) / 9.,
places=5)
self.assertAlmostEqual(
ranking_losses._mean_squared_loss(labels, scores, weights).eval(),
(_mean_squared_error(labels[0], scores[0]) * 2.0 +
_mean_squared_error(labels[1], scores[1]) +
_mean_squared_error(labels[2], scores[2])) / 9.,
places=5)
def test_make_loss_fn(self):
scores = [[0.2, 0.5, 0.3], [0.2, 0.3, 0.5]]
labels = [[0., 0., 1.], [0., 0., 1.]]
weights = [[2.], [1.]]
weights_1d = [2., 1.]
weights_3d = [[[2.], [1.], [0.]], [[0.], [1.], [2.]]]
weights_feature_name = 'weights'
weights_1d_feature_name = 'weights_1d'
weights_3d_feature_name = 'weights_3d'
features = {
weights_feature_name: weights,
weights_1d_feature_name: weights_1d,
weights_3d_feature_name: weights_3d
}
with self.cached_session():
pairwise_hinge_loss = ranking_losses._pairwise_hinge_loss(labels,
scores).eval()
pairwise_hinge_loss_weighted = ranking_losses._pairwise_hinge_loss(
labels, scores, weights=weights).eval()
pairwise_hinge_loss_itemwise_weighted = (
ranking_losses._pairwise_hinge_loss(
labels, scores, weights=tf.squeeze(weights_3d)).eval())
mean_squared_loss = ranking_losses._mean_squared_loss(labels,
scores).eval()
mean_squared_loss_weighted = ranking_losses._mean_squared_loss(
labels, scores, weights=weights).eval()
mean_squared_loss_itemwise_weighted = ranking_losses._mean_squared_loss(
labels, scores, weights=tf.squeeze(weights_3d)).eval()
loss_keys = [
ranking_losses.RankingLossKey.PAIRWISE_HINGE_LOSS,
ranking_losses.RankingLossKey.MEAN_SQUARED_LOSS
]
loss_fn_simple = ranking_losses.make_loss_fn(loss_keys)
self.assertAlmostEqual(
loss_fn_simple(labels, scores, features).eval(),
pairwise_hinge_loss + mean_squared_loss,
places=5)
# With 2-d list-wise weighted examples.
loss_fn_weighted_example = ranking_losses.make_loss_fn(
loss_keys, weights_feature_name=weights_feature_name)
self.assertAlmostEqual(
loss_fn_weighted_example(labels, scores, features).eval(),
pairwise_hinge_loss_weighted + mean_squared_loss_weighted,
places=5)
# With 1-d list-wise weighted examples.
loss_fn_weighted_example = ranking_losses.make_loss_fn(
loss_keys, weights_feature_name=weights_1d_feature_name)
self.assertAlmostEqual(
loss_fn_weighted_example(labels, scores, features).eval(),
pairwise_hinge_loss_weighted + mean_squared_loss_weighted,
places=5)
# With 3-d item-wise weighted examples.
loss_fn_weighted_example = ranking_losses.make_loss_fn(
loss_keys, weights_feature_name=weights_3d_feature_name)
self.assertAlmostEqual(
loss_fn_weighted_example(labels, scores, features).eval(),
pairwise_hinge_loss_itemwise_weighted +
mean_squared_loss_itemwise_weighted,
places=5)
# With both weighted loss and weighted examples.
loss_weights = [3., 2.]
weighted_loss_fn_weighted_example = ranking_losses.make_loss_fn(
loss_keys, loss_weights, weights_feature_name=weights_feature_name)
self.assertAlmostEqual(
weighted_loss_fn_weighted_example(labels, scores, features).eval(),
pairwise_hinge_loss_weighted * loss_weights[0] +
mean_squared_loss_weighted * loss_weights[1],
places=5)
# Test loss reduction method.
# Two reduction methods should return different loss values.
loss_fn_1 = ranking_losses.make_loss_fn(
loss_keys, reduction=tf.compat.v1.losses.Reduction.SUM)
loss_fn_2 = ranking_losses.make_loss_fn(
loss_keys, reduction=tf.compat.v1.losses.Reduction.MEAN)
self.assertNotAlmostEqual(
loss_fn_1(labels, scores, features).eval(),
loss_fn_2(labels, scores, features).eval())
# Test invalid inputs.
with self.assertRaisesRegexp(ValueError,
r'loss_keys cannot be None or empty.'):
ranking_losses.make_loss_fn([])
with self.assertRaisesRegexp(ValueError,
r'loss_keys cannot be None or empty.'):
ranking_losses.make_loss_fn('')
with self.assertRaisesRegexp(
ValueError, r'loss_keys and loss_weights must have the same size.'):
ranking_losses.make_loss_fn(loss_keys, [2.0])
invalid_loss_fn = ranking_losses.make_loss_fn(['invalid_key'])
with self.assertRaisesRegexp(ValueError,
r'Invalid loss_key: invalid_key.'):
invalid_loss_fn(labels, scores, features).eval()