def score()

in tensorflow_ranking/python/keras/canned/gam.py [0:0]

• 45 lines of code
• 13 McCabe index (conditional complexity)

  def score(self, context_features=None, example_features=None, training=True):
    """Univariate scoring of context and one example to generate a score.

    Args:
      context_features: (dict) context feature names to 2D tensors of shape
        [batch_size, ...].
      example_features: (dict) example feature names to 2D tensors of shape
        [batch_size, ...].
      training: (bool) whether in training or inference mode.

    Returns:
      (tf.Tensor) A score tensor of shape [batch_size, 1].
    """
    example_feature_names = sorted(list(self.example_feature_columns.keys()))
    context_feature_names = sorted(list(self.context_feature_columns.keys()))
    context_input = [
        tf.keras.layers.Flatten()(context_features[name])
        for name in context_feature_names
    ]
    example_input = [
        tf.keras.layers.Flatten()(example_features[name])
        for name in example_feature_names
    ]

    # Construct a tower for each example feature.
    sub_logits_list = []
    with tf.name_scope('example_feature_towers'):
      for name, input_tensor in zip(example_feature_names, example_input):
        with tf.name_scope('{}_tower'.format(name)):
          cur = input_tensor
          layers = self._per_example_feature_layers[name]
          for layer in layers:
            cur = layer(cur, training=training)
          sub_logits = tf.identity(
              cur, name='{}_{}'.format(name, _SUBSCORE_POSTFIX))
          sub_logits_list.append(sub_logits)

    sub_weights_list = []
    if context_input:
      # Construct a tower for each context feature.
      with tf.name_scope('context_feature_towers'):
        for name, input_tensor in zip(context_feature_names, context_input):
          with tf.name_scope('{}_tower'.format(name)):
            cur = input_tensor
            layers = self._per_context_feature_layers[name]
            for layer in layers:
              cur = layer(cur, training=training)
            cur = tf.keras.layers.Softmax()(cur)
            sub_weights = tf.identity(
                cur, name='{}_{}'.format(name, _SUBWEIGHT_POSTFIX))
            sub_weights_list.append(sub_weights)

    # Construct an additive model from the outputs of all example feature towers
    # weighted by outputs of all context feature towers.
    # Note that these layers do not have any trainable variables, hence we
    # are not defining them in init but defining them here, similar to Flatten.
    if sub_weights_list:
      sub_logits = tf.keras.layers.Concatenate(axis=-1)(sub_logits_list)
      sub_weights = (
          tf.keras.layers.Add()(sub_weights_list)
          if len(sub_weights_list) > 1 else sub_weights_list[0])
      logits = tf.keras.backend.sum(sub_logits * sub_weights, axis=-1)
    else:
      logits = tf.keras.layers.Add()(
          sub_logits_list) if len(sub_logits_list) > 1 else sub_logits_list[0]
    return logits