refine code, test=develop

paddle/fluid/operators/detection/multiclass_nms_op.cc

@@ -171,15 +171,18 @@ void SliceOneClass(const platform::DeviceContext& ctx,
   const T* items_data = items.data<T>();
   const int64_t num_item = items.dims()[0];
   const int class_num = items.dims()[1];
-  int item_size = 1;
   if (items.dims().size() == 3) {
-    item_size = items.dims()[2];
-  }
+    int item_size = items.dims()[2];
     for (int i = 0; i < num_item; ++i) {
       std::memcpy(item_data + i * item_size,
                   items_data + i * class_num * item_size + class_id * item_size,
                   sizeof(T) * item_size);
     }
+  } else {
+    for (int i = 0; i < num_item; ++i) {
+      item_data[i] = items_data[i * class_num + class_id];
+    }
+  }
 }
 
 template <typename T>