diff --git a/SCLUDA_UP2PC.py b/SCLUDA_UP2PC.py
index 491b001..5ebbead 100644
--- a/SCLUDA_UP2PC.py
+++ b/SCLUDA_UP2PC.py
@@ -74,18 +74,13 @@ def clean_sampling_epoch(labels, probabilities, output):
         sorted_index_method='normalized_margin',
     )
 
-    true_labels_idx = []  # 置信样本的索引
-    all_labels_idx = []  # 所有标签的索引
+    # 亲测高效，老的版本不知道崩了多少次^_^
+    all_labels_idx = np.arange(len(labels))
+    error_set = set(ordered_label_errors)
+    true_labels_idx = [i for i in all_indices if i not in error_set]
+    
     print('len of all_lables', len(labels))
     print('len of errors_lables', len(ordered_label_errors))
-    for i in range(len(labels)):
-        all_labels_idx.append(i)
-    if len(ordered_label_errors) == 0:
-        true_labels_idx = all_labels_idx
-    else:
-        for j in range(len(ordered_label_errors)):
-            all_labels_idx.remove(ordered_label_errors[j])
-            true_labels_idx = all_labels_idx
     print('len of true_lables', len(true_labels_idx))
     # weights
     orig_class_count = np.bincount(labels, minlength=CLASS_NUM)