fixing modelbuilder benchmark for covtype dataset (#125)

Author: Alexsandruss

modelbuilders_bench/lgbm_mb.py

@@ -99,7 +99,11 @@
     if 'cudf' in str(type(y_train)):
         params.n_classes = y_train[y_train.columns[0]].nunique()
     else:
-        params.n_classes = len(np.unique(y_train))
+        unique_y_train = np.unique(y_train)
+        params.n_classes = len(unique_y_train)
+        if max(unique_y_train) != len(unique_y_train) - 1:
+            params.n_classes = int(max(unique_y_train)) + 1
+
     if params.n_classes > 2:
         lgbm_params['num_class'] = params.n_classes
 

modelbuilders_bench/xgb_mb.py

@@ -30,7 +30,7 @@ def convert_xgb_predictions(y_pred, objective):
     if objective == 'multi:softprob':
         y_pred = convert_probs_to_classes(y_pred)
     elif objective == 'binary:logistic':
-        y_pred = y_pred.astype(np.int32)
+        y_pred = (y_pred >= 0.5).astype(np.int32)
     return y_pred