def __init__(self,
topn=None,
alpha=0.5,
rank_discount_fn=_DEFAULT_RANK_DISCOUNT_FN,
seed=None,
dtype=None,
name="alpha_dcg_metric",
**kwargs):
"""Construct the ranking metric class for alpha-DCG.
Args:
topn: A cutoff for how many examples to consider for this metric.
alpha: A float between 0 and 1, parameter used in definition of alpha-DCG.
Introduced as an assessor error in judging whether a document is
covering a subtopic of the query.
rank_discount_fn: A function of rank discounts. Default is set to
discount = 1 / log2(rank+1).
seed: The ops-level random seed used in shuffle ties in `sort_by_scores`.
dtype: Data type of the metric output. See `tf.keras.metrics.Metric`.
name: A string used as the name for this metric.
**kwargs: Other keyward arguments used in `tf.keras.metrics.Metric`.
"""
super(AlphaDCGMetric, self).__init__(name=name, dtype=dtype, **kwargs)
self._topn = topn
self._alpha = alpha
self._rank_discount_fn = rank_discount_fn
self._seed = seed
self._metric = metrics_impl.AlphaDCGMetric(
name=name,
topn=topn,
alpha=alpha,
rank_discount_fn=rank_discount_fn,
seed=seed)
def alpha_discounted_cumulative_gain(
labels,
predictions,
weights=None,
topn=None,
name=None,
rank_discount_fn=_DEFAULT_RANK_DISCOUNT_FN,
alpha=0.5,
seed=None):
"""Computes alpha discounted cumulative gain (alpha-DCG).
Args:
labels: A `Tensor` with shape [batch_size, list_size, subtopic_size]. Each
value represents graded relevance to a subtopic: 1 for relevent subtopic,
0 for irrelevant, and -1 for paddings. When the actual subtopic number
of a query is smaller than the `subtopic_size`, `labels` will be padded
to `subtopic_size` with -1, similar to the paddings used for queries
with doc number less then list_size.
predictions: A `Tensor` with shape [batch_size, list_size]. Each value is
the ranking score of the corresponding example.
weights: A `Tensor` of shape [batch_size, list_size] or [batch_size, 1].
They are per-example and per-list, respectively.
topn: A cutoff for how many examples to consider for this metric.
name: A string used as the name for this metric.
rank_discount_fn: A function of rank discounts. Default is set to
discount = 1 / log2(rank+1).
alpha: A float between 0 and 1. Originally introduced as an assessor error
in judging whether a document is covering a subtopic of the query. It
can also be interpreted as the inverse number of documents covering the
same subtopic reader needs to get and confirm the subtopic information
of a query.
seed: The ops-level random seed used in shuffle ties in `sort_by_scores`.
Returns:
A metric for the weighted alpha discounted cumulative gain of the batch.
"""
metric = metrics_impl.AlphaDCGMetric(name,
topn,
alpha=alpha,
rank_discount_fn=rank_discount_fn,
seed=seed)
with tf.compat.v1.name_scope(name, 'alpha_discounted_cumulative_gain',
(labels, predictions, weights)):
# TODO: Add mask argument for metric.compute() call
alpha_dcg, per_list_weights = metric.compute(labels, predictions,
weights)
return tf.compat.v1.metrics.mean(alpha_dcg, per_list_weights)
def __init__(self,
name="alpha_dcg_metric",
topn=None,
alpha=0.5,
rank_discount_fn=None,
seed=None,
dtype=None,
ragged=False,
**kwargs):
"""Construct the ranking metric class for alpha-DCG.
Args:
name: A string used as the name for this metric.
topn: A cutoff for how many examples to consider for this metric.
alpha: A float between 0 and 1, parameter used in definition of alpha-DCG.
Introduced as an assessor error in judging whether a document is
covering a subtopic of the query.
rank_discount_fn: A function of rank discounts. Default is set to
`1 / log2(rank+1)`. The `rank_discount_fn` should be keras serializable.
Please see the `log2_inverse` above as an example when defining user
customized functions.
seed: The ops-level random seed used in shuffle ties in `sort_by_scores`.
dtype: Data type of the metric output. See `tf.keras.metrics.Metric`.
ragged: A bool indicating whether the supplied tensors are ragged. If
True y_true, y_pred and sample_weight (if providing per-example weights)
need to be ragged tensors with compatible shapes.
**kwargs: Other keyward arguments used in `tf.keras.metrics.Metric`.
"""
super(AlphaDCGMetric, self).__init__(name=name,
dtype=dtype,
ragged=ragged,
**kwargs)
self._topn = topn
self._alpha = alpha
self._rank_discount_fn = rank_discount_fn or utils.log2_inverse
self._seed = seed
self._metric = metrics_impl.AlphaDCGMetric(
name=name,
topn=topn,
alpha=alpha,
rank_discount_fn=self._rank_discount_fn,
seed=seed,
ragged=ragged)