def _approx_ndcg_loss(labels,
logits,
weights=None,
reduction=tf.compat.v1.losses.Reduction.SUM,
name=None,
alpha=10.):
"""Computes ApproxNDCG loss.
ApproxNDCG ["A general approximation framework for direct optimization of
information retrieval measures" by Qin et al.] is a smooth approximation
to NDCG. Its performance on datasets with graded relevance is competitive
with other state-of-the-art algorithms [see "Revisiting Approximate Metric
Optimization in the Age of Deep Neural Networks" by Bruch et al.].
Args:
labels: A `Tensor` of the same shape as `logits` representing graded
relevance.
logits: A `Tensor` with shape [batch_size, list_size]. Each value is the
ranking score of the corresponding item.
weights: A scalar, a `Tensor` with shape [batch_size, 1] for list-wise
weights, or a `Tensor` with shape [batch_size, list_size] for item-wise
weights. If None, the weight of a list in the mini-batch is set to the sum
of the labels of the items in that list.
reduction: One of `tf.losses.Reduction` except `NONE`. Describes how to
reduce training loss over batch.
name: A string used as the name for this loss.
alpha: The exponent in the generalized sigmoid function.
Returns:
An op for the ApproxNDCG loss.
"""
loss = losses_impl.ApproxNDCGLoss(name, params={'alpha': alpha})
with tf.compat.v1.name_scope(loss.name, 'approx_ndcg_loss',
(labels, logits, weights)):
return loss.compute(labels, logits, weights, reduction)
def __init__(self,
reduction=tf.losses.Reduction.AUTO,
name=None,
lambda_weight=None):
super(ApproxNDCGLoss, self).__init__(reduction, name)
self._loss = losses_impl.ApproxNDCGLoss(
name='{}_impl'.format(name), lambda_weight=lambda_weight)
def test_approx_ndcg_loss(self):
with tf.Graph().as_default():
scores = [[1.4, -2.8, -0.4], [0., 1.8, 10.2], [1., 1.2, -3.2]]
# ranks= [[1, 3, 2], [2, 1, 3], [2, 1, 3]]
labels = [[0., 2., 1.], [1., 0., -1.], [0., 0., 0.]]
weights = [[2.], [1.], [1.]]
example_weights = [[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]]
norm_weights = []
for weight, label in zip(example_weights, labels):
sum_label = sum(max(0, l) for l in label)
norm_weights.append(
sum(w * max(0, l) for w, l in zip(weight, label)) /
sum_label if sum_label else 0)
reduction = tf.compat.v1.losses.Reduction.SUM
with self.cached_session():
loss_fn = losses_impl.ApproxNDCGLoss(name=None,
temperature=0.1)
self.assertAlmostEqual(
loss_fn.compute(labels, scores, None, reduction).eval(),
-((1 / (3 / ln(2) + 1 / ln(3))) *
(3 / ln(4) + 1 / ln(3)) + ln(2) * (1 / ln(3))),
places=5)
self.assertAlmostEqual(
loss_fn.compute(labels, scores, weights, reduction).eval(),
-(2 * (1 / (3 / ln(2) + 1 / ln(3))) *
(3 / ln(4) + 1 / ln(3)) + 1 * ln(2) * (1 / ln(3))),
places=5)
self.assertAlmostEqual(
loss_fn.compute(labels, scores, example_weights,
reduction).eval(),
-(norm_weights[0] * (1 / (3 / ln(2) + 1 / ln(3))) *
(3 / ln(4) + 1 / ln(3)) + norm_weights[1] * ln(2) *
(1 / ln(3))),
places=5)
def __init__(self,
reduction=tf.losses.Reduction.AUTO,
name=None,
lambda_weight=None,
temperature=0.1,
ragged=False):
super().__init__(reduction, name, lambda_weight, temperature, ragged)
self._loss = losses_impl.ApproxNDCGLoss(
name='{}_impl'.format(name) if name else None,
lambda_weight=lambda_weight,
temperature=temperature,
ragged=ragged)
def test_listwise_compute_per_list(self):
with tf.Graph().as_default():
scores = [[1., 3., 2.], [1., 2., 3.]]
labels = [[0., 0., 1.], [0., 0., 2.]]
per_item_weights = [[2., 3., 4.], [1., 1., 1.]]
with self.cached_session():
# ApproxNDCGLoss is chosen as an arbitrary listwise loss to test the
# `compute_per_list` behavior.
loss_fn = losses_impl.ApproxNDCGLoss(name=None)
losses, weights = loss_fn.compute_per_list(
labels, scores, per_item_weights)
losses, weights = losses.eval(), weights.eval()
self.assertAllClose(losses, [-0.63093, -0.796248])
self.assertAllClose(weights, [4., 1.])
def make_loss_metric_fn(loss_key,
weights_feature_name=None,
lambda_weight=None,
name=None):
"""Factory method to create a metric based on a loss.
Args:
loss_key: A key in `RankingLossKey`.
weights_feature_name: A `string` specifying the name of the weights feature
in `features` dict.
lambda_weight: A `_LambdaWeight` object.
name: A `string` used as the name for this metric.
Returns:
A metric fn with the following Args:
* `labels`: A `Tensor` of the same shape as `predictions` representing
graded relevance.
* `predictions`: A `Tensor` with shape [batch_size, list_size]. Each value
is the ranking score of the corresponding example.
* `features`: A dict of `Tensor`s that contains all features.
"""
metric_dict = {
RankingLossKey.PAIRWISE_HINGE_LOSS:
losses_impl.PairwiseHingeLoss(name, lambda_weight=lambda_weight),
RankingLossKey.PAIRWISE_LOGISTIC_LOSS:
losses_impl.PairwiseLogisticLoss(name, lambda_weight=lambda_weight),
RankingLossKey.PAIRWISE_SOFT_ZERO_ONE_LOSS:
losses_impl.PairwiseSoftZeroOneLoss(
name, lambda_weight=lambda_weight),
RankingLossKey.SOFTMAX_LOSS:
losses_impl.SoftmaxLoss(name, lambda_weight=lambda_weight),
RankingLossKey.SIGMOID_CROSS_ENTROPY_LOSS:
losses_impl.SigmoidCrossEntropyLoss(name),
RankingLossKey.MEAN_SQUARED_LOSS:
losses_impl.MeanSquaredLoss(name),
RankingLossKey.LIST_MLE_LOSS:
losses_impl.ListMLELoss(name, lambda_weight=lambda_weight),
RankingLossKey.APPROX_NDCG_LOSS:
losses_impl.ApproxNDCGLoss(name),
RankingLossKey.APPROX_MRR_LOSS:
losses_impl.ApproxMRRLoss(name),
RankingLossKey.GUMBEL_APPROX_NDCG_LOSS: losses_impl.ApproxNDCGLoss(name),
}
def _get_weights(features):
"""Get weights tensor from features and reshape it to 2-D if necessary."""
weights = None
if weights_feature_name:
weights = tf.convert_to_tensor(value=features[weights_feature_name])
# Convert weights to a 2-D Tensor.
weights = utils.reshape_to_2d(weights)
return weights
def metric_fn(labels, predictions, features):
"""Defines the metric fn."""
weights = _get_weights(features)
loss = metric_dict.get(loss_key, None)
if loss is None:
raise ValueError('loss_key {} not supported.'.format(loss_key))
return loss.eval_metric(labels, predictions, weights)
return metric_fn
