change to match new scikit-learn losses and deprecated function to available_if

Author: TimotheeMathieu

sklearn_extra/robust/robust_weighted_estimator.py

@@ -26,7 +26,7 @@
 from sklearn.cluster import MiniBatchKMeans
 from sklearn.metrics.pairwise import euclidean_distances
 from sklearn.exceptions import ConvergenceWarning
-from sklearn.utils.metaestimators import if_delegate_has_method
+from sklearn.utils.metaestimators import available_if
 
 # Tool library in which we get robust mean estimators.
 from .mean_estimators import median_of_means_blocked, block_mom, huber
@@ -48,7 +48,7 @@
 
 LOSS_FUNCTIONS = {
     "hinge": (Hinge,),
-    "log": (Log,),
+    "log_loss": (Log,),
     "squared_error": (SquaredLoss,),
     "squared_loss": (SquaredLoss,),
     "squared_hinge": (SquaredHinge,),
@@ -114,8 +114,8 @@ class _RobustWeightedEstimator(BaseEstimator):
     loss : string or callable, mandatory
         Name of the loss used, must be the same loss as the one optimized in
         base_estimator.
-        Classification losses supported : 'log', 'hinge', 'squared_hinge',
-        'modified_huber'. If 'log', then the base_estimator must support
+        Classification losses supported : 'log_loss', 'hinge', 'squared_hinge',
+        'modified_huber'. If 'log_loss', then the base_estimator must support
         predict_proba. Regression losses supported : 'squared_error', 'huber'.
         If callable, the function is used as loss function ro construct
         the weights.
@@ -501,7 +501,7 @@ def predict(self, X):
         return self.base_estimator_.predict(X)
 
     def _check_proba(self):
-        if self.loss != "log":
+        if self.loss != "log_loss":
             raise AttributeError(
                 "Probability estimates are not available for"
                 " loss=%r" % self.loss
@@ -538,7 +538,14 @@ def score(self, X, y=None):
         check_is_fitted(self, attributes=["base_estimator_"])
         return self.base_estimator_.score(X, y)
 
-    @if_delegate_has_method(delegate="base_estimator")
+
+    def _estimator_has(attr):
+        def check(self):
+            return hasattr(self.base_estimator_, attr)
+
+        return check
+        
+    @available_if(_estimator_has("decision_function"))
     def decision_function(self, X):
         """Predict using the linear model. For classifiers only.
 
@@ -607,7 +614,7 @@ class RobustWeightedClassifier(BaseEstimator, ClassifierMixin):
         (using the inter-quartile range), this tends to be conservative
         (robust).
 
-    loss : string, None or callable, default="log"
+    loss : string, None or callable, default="log_loss"
         Classification losses supported : 'log', 'hinge', 'modified_huber'.
         If 'log', then the base_estimator must support predict_proba.
 
@@ -709,7 +716,7 @@ def __init__(
         max_iter=100,
         c=None,
         k=0,
-        loss="log",
+        loss="log_loss",
         sgd_args=None,
         multi_class="ovr",
         n_jobs=1,
@@ -809,7 +816,7 @@ def predict(self, X):
         return self.base_estimator_.predict(X)
 
     def _check_proba(self):
-        if self.loss != "log":
+        if self.loss != "log_loss":
             raise AttributeError(
                 "Probability estimates are not available for"
                 " loss=%r" % self.loss

sklearn_extra/robust/tests/test_robust_weighted_estimator.py

@@ -38,7 +38,7 @@
     X_cc[f] = [10, 5] + rng.normal(size=2) * 0.1
     y_cc[f] = 0
 
-classif_losses = ["log", "hinge"]
+classif_losses = ["log_loss", "hinge"]
 weightings = ["huber", "mom"]
 multi_class = ["ovr", "ovo"]
 
@@ -167,7 +167,7 @@ def test_classif_binary(weighting):
         multi_class="binary",
         random_state=rng,
     )
-    clf_not_rob = SGDClassifier(loss="log", random_state=rng)
+    clf_not_rob = SGDClassifier(loss="log_loss", random_state=rng)
     clf.fit(X_cb, y_cb)
     clf_not_rob.fit(X_cb, y_cb)
     norm_coef1 = np.linalg.norm(np.hstack([clf.coef_.ravel(), clf.intercept_]))
@@ -201,7 +201,7 @@ def test_classif_corrupted_weights(weighting):
     assert np.mean(clf.weights_[:3]) < np.mean(clf.weights_[3:])
 
 
-# Case "log" loss, test predict_proba
+# Case "log_loss" loss, test predict_proba
 @pytest.mark.parametrize("weighting", weightings)
 def test_predict_proba(weighting):
     clf = RobustWeightedClassifier(
@@ -211,7 +211,7 @@ def test_predict_proba(weighting):
         c=1e7,
         random_state=rng,
     )
-    clf_not_rob = SGDClassifier(loss="log", random_state=rng)
+    clf_not_rob = SGDClassifier(loss="log_loss", random_state=rng)
     clf.fit(X_c, y_c)
     clf_not_rob.fit(X_c, y_c)
     pred1 = clf.base_estimator_.predict_proba(X_c)[:, 1]