official/recommendation/uplift/metrics/label_mean.py

# Copyright 2023 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Keras metric for computing the label mean sliced by treatment group."""

import tensorflow as tf, tf_keras

from official.recommendation.uplift import types
from official.recommendation.uplift.metrics import treatment_sliced_metric


@tf_keras.utils.register_keras_serializable(package="Uplift")
class LabelMean(tf_keras.metrics.Metric):
  """Computes the overall and treatment sliced label mean.

  Note that the prediction tensor is expected to be of type
  `TwoTowerTrainingOutputs`.

  Example standalone usage:

  >>> label_mean = LabelMean()
  >>> y_true = tf.constant([[1], [2], [3], [4]])
  >>> y_pred = types.TwoTowerTrainingOutputs(
  ...     control_logits=tf.zeros((4, 1)),
  ...     treatment_logits=tf.zeros((4, 1)),
  ...     true_logits=tf.zeros((4, 1)),
  ...     is_treatment=tf.constant([[True], [False], [True], [False]]),
  ... )
  >>> label_mean(y_true=y_true, y_pred=y_pred)
  {
      "label/mean": 2.5,
      "label/mean/control": 3.0
      "label/mean/treatment": 2.0
  }

  Example usage with the `compile()` API:

  >>> model.compile(
  ...     optimizer="sgd",
  ...     loss=TrueLogitsLoss(tf_keras.losses.mean_squared_error),
  ...     metrics=[LabelMean()]
  ... )
  """

  def __init__(self, name: str = "label/mean", **kwargs):
    """Initializes a `LabelMean` instance.

    Args:
      name: name for the overall label mean metric result. The control and
        treatment label means will have "/control" and "/treatment" appended to
        the result name.
      **kwargs: other base metric keyword arguments.
    """
    super().__init__(name=name, **kwargs)
    self._sliced_mean = treatment_sliced_metric.TreatmentSlicedMetric(
        metric=tf_keras.metrics.Mean(name=name, **kwargs)
    )

  def update_state(
      self,
      y_true: tf.Tensor,
      y_pred: types.TwoTowerTrainingOutputs,
      sample_weight: tf.Tensor | None = None,
  ):
    """Updates the overall, control and treatment label means.

    Args:
      y_true: tensor labels.
      y_pred: prediction logits. The treatment indicator tensor is used to slice
        the labels into control and treatment groups.
      sample_weight: optional sample weight to compute weighted label means. If
        given, the sample weight will also be sliced by the treatment indicator
        tensor to compute the weighted control and treatment label means.

    Raises:
      TypeError: if y_pred is not of type `TwoTowerTrainingOutputs`.
    """
    if not isinstance(y_pred, types.TwoTowerTrainingOutputs):
      raise TypeError(
          "y_pred must be of type `TwoTowerTrainingOutputs` but got type"
          f" {type(y_pred)} instead."
      )

    self._sliced_mean.update_state(
        values=y_true,
        is_treatment=y_pred.is_treatment,
        sample_weight=sample_weight,
    )

  def result(self) -> dict[str, tf.Tensor]:
    return self._sliced_mean.result()