[ENH] parallelization support and config for forecasting performance …

…metrics (#5813)

This PR adds the option to parallelize forecasting performance metrics
via `set_config`.

The method is simply passing on the config to `vectorize_est`, if set.

Tests are added. Docstrings do not need to be added, as apprently the
option is already inherited from `BaseObject` (mistakenly before this
PR, but now correctly).

sktime/performance_metrics/forecasting/_classes.py

@@ -273,12 +273,17 @@ def _evaluate_vectorized(self, y_true, y_pred, **kwargs):
         y_pred : VectorizedDF
         non-time-like instances of y_true, y_pred must be identical
         """
+        backend = dict()
+        backend["backend"] = self.get_config()["backend:parallel"]
+        backend["backend_params"] = self.get_config()["backend:parallel:params"]
+
         eval_result = y_true.vectorize_est(
             estimator=self.clone(),
             method="_evaluate",
             varname_of_self="y_true",
             args={**kwargs, "y_pred": y_pred},
             colname_default=self.name,
+            **backend,
         )
 
         if isinstance(self.multioutput, str) and self.multioutput == "raw_values":
@@ -305,13 +310,18 @@ def _evaluate_by_index_vectorized(self, y_true, y_pred, **kwargs):
         y_pred : VectorizedDF
         non-time-like instances of y_true, y_pred must be identical
         """
+        backend = dict()
+        backend["backend"] = self.get_config()["backend:parallel"]
+        backend["backend_params"] = self.get_config()["backend:parallel:params"]
+
         eval_result = y_true.vectorize_est(
             estimator=self.clone().set_params(**{"multilevel": "uniform_average"}),
             method="_evaluate_by_index",
             varname_of_self="y_true",
             args={**kwargs, "y_pred": y_pred},
             colname_default=self.name,
             return_type="list",
+            **backend,
         )
 
         eval_result = y_true.reconstruct(eval_result)

sktime/performance_metrics/tests/test_metrics_classes.py

@@ -14,6 +14,7 @@
 from sktime.utils._testing.hierarchical import _make_hierarchical
 from sktime.utils._testing.panel import _make_panel
 from sktime.utils._testing.series import _make_series
+from sktime.utils.parallel import _get_parallel_test_fixtures
 
 metric_classes = getmembers(_classes, isclass)
 
@@ -24,6 +25,9 @@
 
 MULTIOUTPUT = ["uniform_average", "raw_values", "numpy"]
 
+# list of parallelization backends to test
+BACKENDS = _get_parallel_test_fixtures("config")
+
 
 @pytest.mark.parametrize("n_columns", [1, 2])
 @pytest.mark.parametrize("multioutput", MULTIOUTPUT)
@@ -76,12 +80,13 @@ def test_metric_output_direct(metric, multioutput, n_columns):
     assert np.allclose(res[1], res[2])
 
 
+@pytest.mark.parametrize("backend", BACKENDS)
 @pytest.mark.parametrize("n_columns", [1, 2])
 @pytest.mark.parametrize(
     "multilevel", ["uniform_average", "uniform_average_time", "raw_values"]
 )
 @pytest.mark.parametrize("multioutput", MULTIOUTPUT)
-def test_metric_hierarchical(multioutput, multilevel, n_columns):
+def test_metric_hierarchical(multioutput, multilevel, n_columns, backend):
     """Test hierarchical input for metrics."""
     # create numpy weights based on n_columns
     if multioutput == "numpy":
@@ -94,6 +99,7 @@ def test_metric_hierarchical(multioutput, multilevel, n_columns):
     y_true = _make_hierarchical(random_state=42, n_columns=n_columns)
 
     metric = MeanSquaredError(multioutput=multioutput, multilevel=multilevel)
+    metric.set_config(**backend)
 
     res = metric(
         y_true=y_true,
@@ -183,10 +189,11 @@ def test_metric_output_by_instance(metric, multioutput, n_columns):
     assert (res.index == y_true.index).all()
 
 
+@pytest.mark.parametrize("backend", BACKENDS)
 @pytest.mark.parametrize("n_columns", [1, 2])
 @pytest.mark.parametrize("multilevel", ["uniform_average", "raw_values"])
 @pytest.mark.parametrize("multioutput", MULTIOUTPUT)
-def test_metric_hierarchical_by_index(multioutput, multilevel, n_columns):
+def test_metric_hierarchical_by_index(multioutput, multilevel, n_columns, backend):
     """Test hierarchical input for metrics."""
     # create numpy weights based on n_columns
     if multioutput == "numpy":
@@ -199,6 +206,7 @@ def test_metric_hierarchical_by_index(multioutput, multilevel, n_columns):
     y_true = _make_hierarchical(random_state=42, n_columns=n_columns)
 
     metric = MeanSquaredError(multioutput=multioutput, multilevel=multilevel)
+    metric.set_config(**backend)
 
     res = metric.evaluate_by_index(
         y_true=y_true,