Fix for balanced class weight (#1080)

* add balanced branch of _compute_class_weight

* Remove extra computation of weights

onedal/datatypes/validation.py

@@ -18,6 +18,7 @@
 import warnings
 from scipy import sparse as sp
 from scipy.sparse import issparse, dok_matrix, lil_matrix
+from sklearn.preprocessing import LabelEncoder
 from collections.abc import Sequence
 from numbers import Integral
 
@@ -57,7 +58,15 @@ def _compute_class_weight(class_weight, classes, y):
     if class_weight is None or len(class_weight) == 0:
         weight = np.ones(classes.shape[0], dtype=np.float64, order='C')
     elif class_weight == 'balanced':
-        weight = None
+        y_ = _column_or_1d(y)
+        classes, _ = np.unique(y_, return_inverse=True)
+
+        le = LabelEncoder()
+        y_ind = le.fit_transform(y_)
+        if not all(np.in1d(classes, le.classes_)):
+            raise ValueError("classes should have valid labels that are in y")
+
+        weight = len(y_) / (len(le.classes_) * np.bincount(y_ind).astype(np.float64))
     else:
         # user-defined dictionary
         weight = np.ones(classes.shape[0], dtype=np.float64, order='C')

sklearnex/svm/nusvc.py

@@ -195,11 +195,7 @@ def _onedal_fit(self, X, y, sample_weight=None, queue=None):
 
         self._onedal_estimator = onedal_NuSVC(**onedal_params)
         self._onedal_estimator.fit(X, y, sample_weight, queue=queue)
-
-        if self.class_weight == 'balanced':
-            self.class_weight_ = self._compute_balanced_class_weight(y)
-        else:
-            self.class_weight_ = self._onedal_estimator.class_weight_
+        self.class_weight_ = self._onedal_estimator.class_weight_
 
         if self.probability:
             self._fit_proba(X, y, sample_weight, queue=queue)

sklearnex/svm/svc.py

@@ -209,11 +209,7 @@ def _onedal_fit(self, X, y, sample_weight=None, queue=None):
 
         self._onedal_estimator = onedal_SVC(**onedal_params)
         self._onedal_estimator.fit(X, y, sample_weight, queue=queue)
-
-        if self.class_weight == 'balanced':
-            self.class_weight_ = self._compute_balanced_class_weight(y)
-        else:
-            self.class_weight_ = self._onedal_estimator.class_weight_
+        self.class_weight_ = self._onedal_estimator.class_weight_
 
         if self.probability:
             self._fit_proba(X, y, sample_weight, queue=queue)