def test_smooth_fraction(self):
"""For the weights using absolute rank."""
sorted_labels = [[2.0, 1.0, 0.0]]
lambda_weight = ranking_losses.DCGLambdaWeight(smooth_fraction=1.0)
with self.cached_session():
self.assertAllClose(
lambda_weight.pair_weights(sorted_labels).eval(),
[[[0., 1. / 2., 2. * 2. / 3.], [1. / 2., 0., 1. / 6.],
[2. * 2. / 3., 1. / 6., 0.]]])
lambda_weight = ranking_losses.DCGLambdaWeight(topn=1, smooth_fraction=1.0)
with self.cached_session():
self.assertAllClose(
lambda_weight.pair_weights(sorted_labels).eval(),
[[[0., 1., 2.], [1., 0., 0.], [2., 0., 0.]]])
def test_invalid_labels(self):
sorted_labels = [[2.0, 1.0, -1.0]]
lambda_weight = ranking_losses.DCGLambdaWeight()
with self.cached_session():
self.assertAllClose(
lambda_weight.pair_weights(sorted_labels).eval(),
[[[0., 1. / 2., 0.], [1. / 2., 0., 0.], [0., 0., 0.]]])
def test_softmax_loss(self):
with tf.Graph().as_default():
scores = [[1., 3., 2.], [1., 2., 3.], [1., 2., 3.]]
labels = [[0., 0., 1.], [0., 0., 2.], [0., 0., 0.]]
weights = [[2.], [1.], [1.]]
with self.cached_session():
self.assertAlmostEqual(
ranking_losses._softmax_loss(labels, scores).eval(),
-(math.log(_softmax(scores[0])[2]) +
math.log(_softmax(scores[1])[2]) * 2.) / 2.,
places=5)
self.assertAlmostEqual(
ranking_losses._softmax_loss(labels, scores,
weights).eval(),
-(math.log(_softmax(scores[0])[2]) * 2. +
math.log(_softmax(scores[1])[2]) * 2. * 1.) / 2.,
places=5)
# Test LambdaWeight.
lambda_weight = ranking_losses.DCGLambdaWeight(
rank_discount_fn=lambda r: 1. / tf.math.log1p(r))
self.assertAlmostEqual(
ranking_losses._softmax_loss(
labels, scores, lambda_weight=lambda_weight).eval(),
-(math.log(_softmax(scores[0])[2]) / math.log(1. + 2.) +
math.log(_softmax(scores[1])[2]) * 2. /
math.log(1. + 1.)) / 2.,
places=5)
def test_individual_weights(self):
sorted_labels = [[1.0, 2.0]]
lambda_weight = ranking_losses.DCGLambdaWeight(normalized=True)
max_dcg = 2.5
with self.cached_session():
self.assertAllClose(
lambda_weight.individual_weights(sorted_labels).eval(),
[[1. / max_dcg / 1., 2. / max_dcg / 2.]])
def test_normalized(self):
sorted_labels = [[1.0, 2.0]]
lambda_weight = ranking_losses.DCGLambdaWeight(normalized=True)
max_dcg = 2.5
with self.cached_session():
self.assertAllClose(
lambda_weight.pair_weights(sorted_labels).eval(),
[[[0., 1. / 2. / max_dcg], [1. / 2. / max_dcg, 0.]]])
def test_default(self):
"""For the weight using rank diff."""
sorted_labels = [[2.0, 1.0, 0.0]]
lambda_weight = ranking_losses.DCGLambdaWeight()
with self.cached_session():
self.assertAllClose(
lambda_weight.pair_weights(sorted_labels).eval(),
[[[0., 1. / 2., 2. * 1. / 6.], [1. / 2., 0., 1. / 2.],
[2. * 1. / 6., 1. / 2., 0.]]])
def test_topn(self):
with tf.Graph().as_default():
sorted_labels = [[2.0, 1.0, 0.0]]
lambda_weight = ranking_losses.DCGLambdaWeight(topn=1)
with self.cached_session():
self.assertAllClose(
lambda_weight.pair_weights(sorted_labels).eval(),
[[[0., 1. / 2., 2. * 1. / 2.], [1. / 2., 0., 0.],
[2. * 1. / 2., 0., 0.]]])
def test_gain_and_discount(self):
sorted_labels = [[2.0, 1.0]]
lambda_weight = ranking_losses.DCGLambdaWeight(
gain_fn=lambda x: tf.pow(2., x) - 1.,
rank_discount_fn=lambda r: 1. / tf.math.log1p(r))
with self.cached_session():
self.assertAllClose(
lambda_weight.pair_weights(sorted_labels).eval(),
[[[0., 2. * (1. / math.log(2.) - 1. / math.log(3.))],
[2. * (1. / math.log(2.) - 1. / math.log(3.)), 0.]]])
def _check_make_pairwise_loss(self, loss_key):
"""Helper function to test `make_loss_fn`."""
scores = [[1., 3., 2.], [1., 2., 3.]]
labels = [[0., 0., 1.], [0., 0., 2.]]
listwise_weights = [[2.], [1.]]
listwise_weights_expanded = [[2.] * 3, [1.] * 3]
itemwise_weights = [[2., 3., 4.], [1., 1., 1.]]
default_weights = [1.] * 3
weights_feature_name = 'weights'
list_size = 3.
features = {}
loss_fn = ranking_losses.make_loss_fn(loss_key)
with self.cached_session():
# Individual lists.
self.assertAlmostEqual(
loss_fn([labels[0]], [scores[0]], features).eval(),
_batch_aggregation(
[_pairwise_loss(labels[0], scores[0], default_weights,
loss_key)]),
places=5)
self.assertAlmostEqual(
loss_fn([labels[1]], [scores[1]], features).eval(),
_batch_aggregation(
[_pairwise_loss(labels[1], scores[1], default_weights,
loss_key)]),
places=5)
# Itemwise weights.
loss_fn = ranking_losses.make_loss_fn(
loss_key, weights_feature_name=weights_feature_name)
features[weights_feature_name] = [itemwise_weights[0]]
self.assertAlmostEqual(
loss_fn([labels[0]], [scores[0]], features).eval(),
_batch_aggregation([
_pairwise_loss(labels[0], scores[0], itemwise_weights[0],
loss_key)
]),
places=5)
features[weights_feature_name] = [itemwise_weights[1]]
self.assertAlmostEqual(
loss_fn([labels[1]], [scores[1]], features).eval(),
_batch_aggregation([
_pairwise_loss(labels[1], scores[1], itemwise_weights[1],
loss_key)
]),
places=5)
# Multiple lists.
features[weights_feature_name] = listwise_weights
self.assertAlmostEqual(
loss_fn(labels, scores, features).eval(),
_batch_aggregation([
_pairwise_loss(labels[0], scores[0], listwise_weights_expanded[0],
loss_key),
_pairwise_loss(labels[1], scores[1], listwise_weights_expanded[1],
loss_key)
]),
places=5)
# Test LambdaWeight.
lambda_weight = ranking_losses.DCGLambdaWeight(
rank_discount_fn=lambda r: 1. / tf.math.log1p(r), smooth_fraction=1.)
loss_fn = ranking_losses.make_loss_fn(
loss_key,
weights_feature_name=weights_feature_name,
lambda_weight=lambda_weight)
self.assertAlmostEqual(
loss_fn(labels, scores, features).eval(),
_batch_aggregation([
_pairwise_loss(
labels[0],
scores[0],
listwise_weights_expanded[0],
loss_key,
rank_discount_form='LOG'),
_pairwise_loss(
labels[1],
scores[1],
listwise_weights_expanded[1],
loss_key,
rank_discount_form='LOG')
]) * list_size,
places=5)
# Test loss reduction method.
# Two reduction methods should return different loss values.
loss_fn_1 = ranking_losses.make_loss_fn(
loss_key, reduction=tf.compat.v1.losses.Reduction.SUM)
loss_fn_2 = ranking_losses.make_loss_fn(
loss_key, reduction=tf.compat.v1.losses.Reduction.MEAN)
self.assertNotAlmostEqual(
loss_fn_1(labels, scores, features).eval(),
loss_fn_2(labels, scores, features).eval())
def _check_pairwise_loss(self, loss_fn):
"""Helper function to test `loss_fn`."""
scores = [[1., 3., 2.], [1., 2., 3.]]
labels = [[0., 0., 1.], [0., 0., 2.]]
listwise_weights = [[2.], [1.]]
listwise_weights_expanded = [[2.] * 3, [1.] * 3]
itemwise_weights = [[2., 3., 4.], [1., 1., 1.]]
default_weights = [1.] * 3
list_size = 3.
loss_form_dict = {
ranking_losses._pairwise_hinge_loss:
ranking_losses.RankingLossKey.PAIRWISE_HINGE_LOSS,
ranking_losses._pairwise_logistic_loss:
ranking_losses.RankingLossKey.PAIRWISE_LOGISTIC_LOSS,
ranking_losses._pairwise_soft_zero_one_loss:
ranking_losses.RankingLossKey.PAIRWISE_SOFT_ZERO_ONE_LOSS,
}
loss_form = loss_form_dict[loss_fn]
with self.cached_session():
# Individual lists.
self.assertAlmostEqual(
loss_fn([labels[0]], [scores[0]]).eval(),
_batch_aggregation([
_pairwise_loss(labels[0], scores[0], default_weights, loss_form)
]),
places=5)
self.assertAlmostEqual(
loss_fn([labels[1]], [scores[1]]).eval(),
_batch_aggregation([
_pairwise_loss(labels[1], scores[1], default_weights, loss_form)
]),
places=5)
# Itemwise weights.
self.assertAlmostEqual(
loss_fn([labels[0]], [scores[0]],
weights=[itemwise_weights[0]]).eval(),
_batch_aggregation([
_pairwise_loss(labels[0], scores[0], itemwise_weights[0],
loss_form)
]),
places=5)
self.assertAlmostEqual(
loss_fn([labels[1]], [scores[1]],
weights=[itemwise_weights[1]]).eval(),
_batch_aggregation([
_pairwise_loss(labels[1], scores[1], itemwise_weights[1],
loss_form)
]),
places=5)
# Multiple lists.
self.assertAlmostEqual(
loss_fn(labels, scores, weights=listwise_weights).eval(),
_batch_aggregation([
_pairwise_loss(labels[0], scores[0], listwise_weights_expanded[0],
loss_form),
_pairwise_loss(labels[1], scores[1], listwise_weights_expanded[1],
loss_form)
]),
places=5)
# Test LambdaWeight.
lambda_weight = ranking_losses.DCGLambdaWeight(
rank_discount_fn=lambda r: 1. / tf.math.log1p(r), smooth_fraction=1.)
self.assertAlmostEqual(
loss_fn(
labels,
scores,
weights=listwise_weights,
lambda_weight=lambda_weight).eval(),
_batch_aggregation([
_pairwise_loss(
labels[0],
scores[0],
listwise_weights_expanded[0],
loss_form,
rank_discount_form='LOG'),
_pairwise_loss(
labels[1],
scores[1],
listwise_weights_expanded[1],
loss_form,
rank_discount_form='LOG')
]) * list_size,
places=5)
# Test loss reduction method.
# Two reduction methods should return different loss values.
reduced_1 = loss_fn(
labels, scores, reduction=tf.compat.v1.losses.Reduction.SUM).eval()
reduced_2 = loss_fn(
labels, scores, reduction=tf.compat.v1.losses.Reduction.MEAN).eval()
self.assertNotAlmostEqual(reduced_1, reduced_2)
