Score fixed

autoPyTorch/api/base_task.py

@@ -1781,7 +1781,8 @@ def predict(
     def score(
         self,
         y_pred: np.ndarray,
-        y_test: Union[np.ndarray, pd.DataFrame]
+        y_test: Union[np.ndarray, pd.DataFrame],
+        **score_kwargs: Any,
     ) -> Dict[str, float]:
         """Calculate the score on the test set.
         Calculate the evaluation measure on the test set.
@@ -1791,6 +1792,8 @@ def score(
                 The test predictions
             y_test (np.ndarray):
                 The test ground truth labels.
+            score_kwargs: Any
+                additional arguments for computing the scores. Some metrics might require special arguments
 
         Returns:
             Dict[str, float]:
@@ -1804,7 +1807,8 @@ def score(
                              "Please check the log file for related errors. ")
         return calculate_score(target=y_test, prediction=y_pred,
                                task_type=STRING_TO_TASK_TYPES[self.task_type],
-                               metrics=[self._metric])
+                               metrics=[self._metric],
+                               **score_kwargs)
 
     def __getstate__(self) -> Dict[str, Any]:
         # Cannot serialize a client!

autoPyTorch/api/time_series_forecasting.py

@@ -1,4 +1,5 @@
 from typing import Any, Callable, Dict, List, Mapping, Optional, Tuple, Union
+import warnings
 
 import numpy as np
 
@@ -590,3 +591,65 @@ def update_sliding_window_size(self, n_prediction_steps: int) -> None:
             ],
             default_value=int(np.ceil(1.25 * base_window_size)),
         )
+
+    def score(
+        self,
+        y_pred: np.ndarray,
+        y_test: Union[np.ndarray, pd.DataFrame],
+        y_test_past: Optional[List[Union[np.ndarray, pd.DataFrame]]] = None,
+        sp: Optional[int] = None,
+        n_prediction_steps: Optional[int] = None,
+        **score_kwargs: Any,
+    ) -> Dict[str, float]:
+        """Calculate the score on the test set.
+        Calculate the evaluation measure on the test set. As forecasting metrics might require some additional metrics,
+        they can  be either given by the users or by default given by the api.
+
+        NOTE: MASE metric requires to scale the predicted values with the previous data. However, given that we do not
+        have access to the previous data within this function, users need to manually compute the coefficient with the
+        function autoPyTorch.pipeline.components.training.metrics.metrics.compute_mase_coefficient and multiple both
+        y_pred and y_test with that coefficient
+
+        Args:
+            y_pred (np.ndarray):
+                The test predictions
+            y_test (np.ndarray):
+                The test ground truth labels.
+            y_test_past:
+            sp (Optional[int]):
+
+            score_kwargs: Any
+                additional arguments for computing the scores. Some metrics might require special arguments
+
+        Returns:
+            Dict[str, float]:
+                Value of the evaluation metric calculated on the test set.
+        """
+        y_pred = np.asarray(y_pred)
+        y_test = np.asarray(y_test)
+        assert np.all(y_pred.shape == y_test.shape), f"y_pred and y_test must have the same shape! But they are " \
+                                                     f"{y_pred.shape} and {y_test.shape}!"
+        score_kwargs_forecasting = {}
+        if sp is None:
+            sp = self.dataset.seasonality
+        score_kwargs_forecasting['sp'] = sp
+        if n_prediction_steps is None:
+            n_prediction_steps = self.dataset.n_prediction_steps
+        score_kwargs_forecasting['n_prediction_steps'] = n_prediction_steps
+        if y_test_past is not None:
+            assert len(y_test_past) == len(y_test), f'The length of y_test_past must be equal to the length of ' \
+                                                    f'y_test.But they are {len(y_test_past)} and {len(y_test)}'
+            from autoPyTorch.pipeline.components.training.metrics.metrics import compute_mase_coefficient
+            mase_coefficient = np.asarray([compute_mase_coefficient(y_past, sp) for y_past in y_test_past])
+            if len(y_pred.shape) > 2:
+                mase_coefficient = np.expand_dims(mase_coefficient, 1)
+            # Match the shape of mase_coefficient and y_test
+            score_kwargs_forecasting['mase_coefficient'] = mase_coefficient
+        else:
+            if self._metric.name.endswith('MASE_Forecasting'):
+                warnings.warn("To compute MASE losses, the past target values must be provided. Here we simply ignore "
+                              "the scaling coefficient and the loss degenerate to a MAE loss")
+
+        return super(TimeSeriesForecastingTask, self).score(np.asarray(y_pred), np.asarray(y_test),
+                                                            **score_kwargs_forecasting,
+                                                            **score_kwargs)

autoPyTorch/evaluation/time_series_forecasting_train_evaluator.py

@@ -146,10 +146,7 @@ def __init__(self, backend: Backend, queue: Queue,
         self.num_sequences = self.datamanager.num_sequences
         self.num_targets = self.datamanager.num_targets
         self.seq_length_min = np.min(self.num_sequences)
-        seasonality = SEASONALITY_MAP.get(self.datamanager.freq, 1)
-        if isinstance(seasonality, list):
-            seasonality = min(seasonality)  # Use to calculate MASE
-        self.seasonality = int(seasonality)  # type: ignore[call-overload]
+        self.seasonality = self.datamanager.seasonality
 
         self.max_budget = max_budget
         self.min_num_test_instances = min_num_test_instances

examples/20_basics/example_time_series_forecasting.py

@@ -12,6 +12,8 @@
 import warnings
 import copy
 
+import numpy as np
+
 os.environ['JOBLIB_TEMP_FOLDER'] = tmp.gettempdir()
 os.environ['OMP_NUM_THREADS'] = '1'
 os.environ['OPENBLAS_NUM_THREADS'] = '1'
@@ -91,3 +93,6 @@
 # test_sets2 = api.dataset.generate_test_seqs()
 
 pred = api.predict(test_sets)
+# To compute the scores with AutoPyTorch, the ground truth value must be of shape [n_seq, seq_length, n_output]
+# or [n_seq * seq_length, n_output]
+score = api.score(np.expand_dims(pred, -1), np.expand_dims(np.asarray(y_test), -1))

test/test_api/test_api.py

@@ -39,8 +39,7 @@
 from autoPyTorch.optimizer.smbo import AutoMLSMBO
 from autoPyTorch.pipeline.base_pipeline import BasePipeline
 from autoPyTorch.pipeline.components.setup.traditional_ml.traditional_learner import _traditional_learners
-from autoPyTorch.pipeline.components.training.metrics.metrics import accuracy
-
+from autoPyTorch.pipeline.components.training.metrics.metrics import accuracy, mean_MASE_forecasting
 
 CV_NUM_SPLITS = 2
 HOLDOUT_NUM_SPLITS = 1
@@ -578,13 +577,23 @@ def test_time_series_forecasting(forecasting_toy_dataset, resampling_strategy, b
     y_pred = estimator.predict(X_test)
 
     assert np.shape(y_pred) == np.shape(y_test)
+    score_mse = estimator.score(np.expand_dims(np.asarray(y_pred), -1),
+                            np.expand_dims(np.asarray(y_test), -1))
+    assert isinstance(score_mse, dict)
+
+    estimator._metric = mean_MASE_forecasting
+    score_mase = estimator.score(np.expand_dims(np.asarray(y_pred), -1),
+                            np.expand_dims(np.asarray(y_test), -1),
+                            y_test_past=y_train)
 
     # Test refit on dummy data
     estimator.refit(dataset=backend.load_datamanager())
     # Make sure that a configuration space is stored in the estimator
     assert isinstance(estimator.get_search_space(), CS.ConfigurationSpace)
 
 
+
+
 @pytest.mark.parametrize('openml_id', (
     1590,  # Adult to test NaN in categorical columns
 ))