diff --git a/python/src/main.cpp b/python/src/main.cpp
index 1fd9473d3..10b28f5b6 100644
--- a/python/src/main.cpp
+++ b/python/src/main.cpp
@@ -884,7 +884,7 @@ PYBIND11_MODULE(ncnn, m)
 #endif // NCNN_STDIO
 
     .def("clear", &Net::clear)
-    .def("create_extractor", &Net::create_extractor)
+    .def("create_extractor", &Net::create_extractor, py::keep_alive<0, 1>()) //net should be kept alive until retuned ex is freed by gc
     .def("blobs", &Net::blobs, py::return_value_policy::reference_internal)
     .def("layers", &Net::layers, py::return_value_policy::reference_internal);
 
@@ -1113,8 +1113,8 @@ PYBIND11_MODULE(ncnn, m)
     m.def("destroy_gpu_instance", &destroy_gpu_instance);
     m.def("get_gpu_count", &get_gpu_count);
     m.def("get_default_gpu_index", &get_default_gpu_index);
-    m.def("get_gpu_info", &get_gpu_info, py::arg("device_index") = 0, py::return_value_policy::reference_internal);
-    m.def("get_gpu_device", &get_gpu_device, py::arg("device_index") = 0, py::return_value_policy::reference_internal);
+    m.def("get_gpu_info", &get_gpu_info, py::arg("device_index") = 0, py::return_value_policy::reference);
+    m.def("get_gpu_device", &get_gpu_device, py::arg("device_index") = 0, py::return_value_policy::reference);
 
     py::class_<VkBufferMemory>(m, "VkBufferMemory")
     .def_readwrite("offset", &VkBufferMemory::offset)
diff --git a/python/tests/benchmark.py b/python/tests/benchmark.py
index 21e5804ba..6f0bbfdf7 100644
--- a/python/tests/benchmark.py
+++ b/python/tests/benchmark.py
@@ -56,6 +56,7 @@ def benchmark(comment, _in, opt):
 
     # warm up
     for i in range(g_warmup_loop_count):
+        # test with statement
         with net.create_extractor() as ex:
             ex.input("data", _in)
             ex.extract("output")
@@ -67,9 +68,10 @@ def benchmark(comment, _in, opt):
     for i in range(g_loop_count):
         start = time.time()
 
-        with net.create_extractor() as ex:
-            ex.input("data", _in)
-            ex.extract("output")
+        # test net keep alive until ex freed
+        ex = net.create_extractor()
+        ex.input("data", _in)
+        ex.extract("output")
 
         end = time.time()