Merge pull request #12 from Orange-OpenSource/lin-tree-fix

Fix plot for linearized trees
Orange-OpenSource · Dec 20, 2024 · ff2d499 · ff2d499
2 parents 92209ce + 75f6e18
commit ff2d499
Show file tree

Hide file tree

Showing 2 changed files with 47 additions and 34 deletions.
diff --git a/examples/plot_linearized.ipynb b/examples/plot_linearized.ipynb
diff --git a/mislabeled/probe/_linear.py b/mislabeled/probe/_linear.py
@@ -90,19 +90,17 @@ def grad_p(self, X, y):
         # gradients w.r.t. the parameters (weight, intercept)
         dl_dy = self.grad_y(X, y)
 
-        X_p = X
+        X_p = X if not sp.issparse(X) else X.toarray()
         if self.intercept is not None:
-            X_p = np.hstack((X, np.ones((X.shape[0], 1))))
-        if sp.issparse(X_p):
-            X_p = X_p.toarray()
+            X_p = np.hstack((X_p, np.ones((X_p.shape[0], 1))))
+
         return dl_dy[:, :, None] * X_p[:, None, :]
 
     def hessian(self, X, y):
-        X_p = X
+        X_p = X if not sp.issparse(X) else X.toarray()
+
         if self.intercept is not None:
-            X_p = np.hstack((X, np.ones((X.shape[0], 1))))
-        if sp.issparse(X_p):
-            X_p = X_p.toarray()
+            X_p = np.hstack((X_p, np.ones((X_p.shape[0], 1))))
 
         if self.loss == "l2":
             H = 2.0 * X_p.T @ X_p
@@ -179,7 +177,7 @@ def linearize_linear_model_ridge(estimator, X, y):
 
 @linearize.register(SGDClassifier)
 def linearize_linear_model_sgdclassifier(estimator, X, y):
-    X, y = check_X_y(X, y, accept_sparse=False, dtype=[np.float64, np.float32])
+    X, y = check_X_y(X, y, accept_sparse=True, dtype=[np.float64, np.float32])
     coef = estimator.coef_.T
     intercept = estimator.intercept_ if estimator.fit_intercept else None
     linear = LinearModel(coef, intercept, loss=estimator.loss, regul=estimator.alpha)
@@ -213,19 +211,16 @@ def linearize_trees(
     X,
     y,
     default_linear_model=dict(
-        classification=LogisticRegressionCV(
-            max_iter=1000, fit_intercept=False, n_jobs=-1
-        ),
-        regression=RidgeCV(fit_intercept=False),
+        classification=LogisticRegression(max_iter=1000),
+        regression=RidgeCV(),
     ),
 ):
     leaves = OneHotEncoder().fit_transform(estimator.apply(X).reshape(X.shape[0], -1))
     if is_classifier(estimator):
         linear = default_linear_model["classification"]
-        linear.fit(leaves, y)
     else:
         linear = default_linear_model["regression"]
-        linear.fit(leaves, y)
+    linear.fit(leaves, y)
     return linearize(linear, leaves, y)