Added tests for GAR with explainability

Signed-off-by: Stefano Savare <[email protected]>

gtime/explainability/explainer.py

@@ -60,6 +60,7 @@ class _LimeExplainer(_RegressorExplainer):
     >>> explainer.explanations_[0]
     {'d': -0.10406889434277307, 'c': 0.07973507022816899, 'b': 0.02312395991550859, 'a': 0.006403509251399996, 'e': 0.006272607738125953}
     """
+
     def fit(
         self, model: RegressorMixin, X: np.ndarray, feature_names: List[str] = None
     ):

gtime/forecasting/gar.py

@@ -13,7 +13,9 @@
 from gtime.regressors.multi_output import MultiFeatureMultiOutputRegressor
 
 
-def initialize_estimator(estimator: RegressorMixin, explainer_type: Optional[str]) -> RegressorMixin:
+def initialize_estimator(
+    estimator: RegressorMixin, explainer_type: Optional[str]
+) -> RegressorMixin:
     if explainer_type is None:
         return estimator
     else:

gtime/forecasting/tests/test_gar.py

@@ -1,3 +1,4 @@
+import itertools
 import random
 from typing import List
 
@@ -47,6 +48,34 @@
 )
 
 
+forecasters = [GAR, GARFF, MultiFeatureGAR]
+explainers = [
+    "shap",
+]  # "lime"] for speed reason
+
+
+@pytest.mark.parametrize(
+    "forecaster,explainer", itertools.product(forecasters, explainers)
+)
+@given(
+    X_y=X_y_matrices(
+        horizon=4,
+        df_transformer=df_transformer,
+        min_length=10,
+        allow_nan_infinity=False,
+    )
+)
+def test_predict_has_explainers(forecaster, explainer, X_y):
+    X, y = X_y
+    X_train, y_train, X_test, y_test = FeatureSplitter().transform(X, y)
+    model = forecaster(LinearRegression(), explainer_type=explainer)
+    model.fit(X_train, y_train)
+    model.predict(X_test.iloc[:1, :])
+    assert len(model.estimators_) == y_test.shape[1]
+    for estimator in model.estimators_:
+        assert len(estimator.explainer_.explanations_) == 1
+
+
 @pytest.fixture
 def time_series():
     testing.N, testing.K = 200, 1
@@ -131,7 +160,7 @@ def test_initialize_estimator(estimator):
 
 @given(models())
 def test_initialize_estimator_explainable(estimator):
-    explainable_estimator = initialize_estimator(estimator, explainer_type='shap')
+    explainable_estimator = initialize_estimator(estimator, explainer_type="shap")
     assert isinstance(explainable_estimator, ExplainableRegressor)
     assert isinstance(explainable_estimator.explainer, _ShapExplainer)
 

gtime/regressors/explainable.py

@@ -53,8 +53,8 @@ def __init__(self, estimator: RegressorMixin, explainer_type: str):
         self.explainer = self._initialize_explainer()
 
     def _check_estimator(self, estimator: RegressorMixin) -> RegressorMixin:
-        if not hasattr(estimator, 'fit') or not hasattr(estimator, 'predict'):
-            raise TypeError(f'Estimator not compatible: {estimator}')
+        if not hasattr(estimator, "fit") or not hasattr(estimator, "predict"):
+            raise TypeError(f"Estimator not compatible: {estimator}")
         return estimator
 
     def _initialize_explainer(self) -> Union[_LimeExplainer, _ShapExplainer]:
@@ -82,7 +82,9 @@ def fit(self, X: np.ndarray, y: np.ndarray, feature_names: List[str] = None):
         Fitted `ExplainableRegressor`
         """
         self.estimator_ = self.estimator.fit(X, y)
-        self.explainer_ = self.explainer.fit(self.estimator_, X, feature_names=feature_names)
+        self.explainer_ = self.explainer.fit(
+            self.estimator_, X, feature_names=feature_names
+        )
         return self
 
     def predict(self, X: np.ndarray):

gtime/regressors/tests/test_explainable.py

@@ -39,7 +39,7 @@ def test_constructor(self, estimator, explainer_type):
     @given(estimator=regressors())
     def test_constructor_bad_explainer(self, estimator):
         with pytest.raises(ValueError):
-            ExplainableRegressor(estimator, 'bad')
+            ExplainableRegressor(estimator, "bad")
 
     @pytest.mark.parametrize("explainer_type", ["lime", "shap"])
     @given(bad_estimator=bad_regressors())
@@ -84,12 +84,11 @@ def test_fit_values(self, estimator, explainer_type, X_y):
     )
     def test_predict_values(self, estimator, explainer_type, X_y):
         X, y = X_y
-        X_test = X[:2, :]
+        X_test = X[:1, :]
         regressor = ExplainableRegressor(estimator, explainer_type)
         regressor_predictions = regressor.fit(X, y).predict(X_test)
 
         cloned_estimator = clone(estimator)
         estimator_predictions = cloned_estimator.fit(X, y).predict(X_test)
 
         assert regressor_predictions.shape == estimator_predictions.shape
-

gtime/regressors/tests/test_multi_output.py

@@ -68,7 +68,11 @@ def test_constructor(self, estimator):
 
     @given(
         data=data(),
-        X_y=numpy_X_y_matrices(X_y_shapes=shape_X_y_matrices(y_as_vector=False), min_value=-10000, max_value=10000),
+        X_y=numpy_X_y_matrices(
+            X_y_shapes=shape_X_y_matrices(y_as_vector=False),
+            min_value=-10000,
+            max_value=10000,
+        ),
     )
     def test_fit_bad_y(self, data, estimator, X_y):
         X, y = X_y
@@ -84,7 +88,13 @@ def test_fit_bad_y(self, data, estimator, X_y):
                 X, y, target_to_features_dict=target_to_feature_dict
             )
 
-    @given(X_y=numpy_X_y_matrices(X_y_shapes=shape_X_y_matrices(y_as_vector=False), min_value=-10000, max_value=10000))
+    @given(
+        X_y=numpy_X_y_matrices(
+            X_y_shapes=shape_X_y_matrices(y_as_vector=False),
+            min_value=-10000,
+            max_value=10000,
+        )
+    )
     def test_fit_as_multi_output_regressor_if_target_to_feature_none(
         self, estimator, X_y
     ):
@@ -110,7 +120,11 @@ def test_error_predict_with_no_fit(self, estimator, X):
 
     @given(
         data=data(),
-        X_y=numpy_X_y_matrices(X_y_shapes=shape_X_y_matrices(y_as_vector=False), min_value=-10000, max_value=10000),
+        X_y=numpy_X_y_matrices(
+            X_y_shapes=shape_X_y_matrices(y_as_vector=False),
+            min_value=-10000,
+            max_value=10000,
+        ),
     )
     def test_fit_target_to_feature_dict_working(self, data, X_y, estimator):
         X, y = X_y
@@ -126,7 +140,11 @@ def test_fit_target_to_feature_dict_working(self, data, X_y, estimator):
 
     @given(
         data=data(),
-        X_y=numpy_X_y_matrices(X_y_shapes=shape_X_y_matrices(y_as_vector=False), min_value=-10000, max_value=10000),
+        X_y=numpy_X_y_matrices(
+            X_y_shapes=shape_X_y_matrices(y_as_vector=False),
+            min_value=-10000,
+            max_value=10000,
+        ),
     )
     def test_fit_target_to_feature_dict_consistent(self, data, X_y, estimator):
         X, y = X_y
@@ -147,7 +165,11 @@ def test_fit_target_to_feature_dict_consistent(self, data, X_y, estimator):
 
     @given(
         data=data(),
-        X_y=numpy_X_y_matrices(X_y_shapes=shape_X_y_matrices(y_as_vector=False), min_value=-10000, max_value=10000),
+        X_y=numpy_X_y_matrices(
+            X_y_shapes=shape_X_y_matrices(y_as_vector=False),
+            min_value=-10000,
+            max_value=10000,
+        ),
     )
     def test_predict_target_to_feature_dict(self, data, X_y, estimator):
         X, y = X_y
@@ -165,7 +187,11 @@ def test_predict_target_to_feature_dict(self, data, X_y, estimator):
 
     @given(
         data=data(),
-        X_y=numpy_X_y_matrices(X_y_shapes=shape_X_y_matrices(y_as_vector=False), min_value=-10000, max_value=10000),
+        X_y=numpy_X_y_matrices(
+            X_y_shapes=shape_X_y_matrices(y_as_vector=False),
+            min_value=-10000,
+            max_value=10000,
+        ),
     )
     def test_error_predict_target_to_feature_dict_wrong_X_shape(
         self, data, X_y, estimator

gtime/utils/hypothesis/general_strategies.py

@@ -22,9 +22,7 @@ def ordered_pair(min_value: int, max_value: int):
 
 
 def shape_vector(min_shape=30, max_shape=200):
-    return tuples(
-        integers(min_shape, max_shape)
-    )
+    return tuples(integers(min_shape, max_shape))
 
 
 def shape_matrix(min_shape_0=30, max_shape_0=200, min_shape_1=5, max_shape_1=10):

requirements.txt

@@ -1,6 +1,7 @@
-lime
 pandas>=0.25.3
 workalendar>=7.1.1
 scipy>=0.17.0
 scikit-learn>=0.22.0
-matplotlib>=3.1.0
+matplotlib>=3.1.0
+lime>=0.2.0.0
+shap>=0.35