@_export('my_metric')
class _MyMetric(_PostExportMetric):
   def __init__(self,
                target_prediction_keys: Optional[List[Text]] = None,
                labels_key: Optional[Text] = None,
                metric_tag: Optional[Text] = None):
      self._target_prediction_keys = target_prediction_keys
      self._label_keys = label_keys
      self._metric_tag = metric_tag
      self._metric_key = 'my_metric_key'

   def check_compatibility(self, features_dict:types.TensorTypeMaybeDict,
                           predictions_dict: types.TensorTypeMaybeDict,
                           labels_dict: types.TensorTypeMaybeDict) -> None:
       # Add compatibility check needed for the metric here.

   def get_metric_ops(self, features_dict: types.TensorTypeMaybeDict,
                      predictions_dict: types.TensorTypeMaybeDict,
                      labels_dict: types.TensorTypeMaybeDict
                     ) -> Dict[bytes, Tuple[types.TensorType,
                     types.TensorType]]:
        # Metric computation logic here.
        # Define value and update ops.
        value_op = compute_metric_value(...)
        update_op = create_update_op(... )
        return {self._metric_key: (value_op, update_op)}

   def populate_stats_and_pop(
       self, slice_key: slicer.SliceKeyType, combined_metrics: Dict[Text, Any],
       output_metrics: Dict[Text, metrics_pb2.MetricValue]) -> None:
       # Parses the metric and converts it into required metric format.
       metric_result = combined_metrics[self._metric_key]
       output_metrics[self._metric_key].double_value.value = metric_result

# Custom metric callback
custom_metric_callback = my_metric(
    labels_key='label',
    target_prediction_keys=['prediction'])

fairness_indicators_callback =
   post_export_metrics.fairness_indicators(
        thresholds=[0.1, 0.3, 0.5, 0.7, 0.9], labels_key=label)

add_metrics_callbacks = [custom_metric_callback,
   fairness_indicators_callback]

eval_shared_model = tfma.default_eval_shared_model(
    eval_saved_model_path=eval_saved_model_path,
    add_metrics_callbacks=add_metrics_callbacks)

eval_config = tfma.EvalConfig(...)

# Run evaluation
tfma.run_model_analysis(
    eval_config=eval_config, eval_shared_model=eval_shared_model)