diff --git a/cmake/external/pybind11.cmake b/cmake/external/pybind11.cmake
index ab23663695c3d0f20a16a03a583f2c782e450aa7..4e87dc49d8956d1fa6dec777efc5a63c6b0f79a5 100644
--- a/cmake/external/pybind11.cmake
+++ b/cmake/external/pybind11.cmake
@@ -26,7 +26,7 @@ ExternalProject_Add(
         extern_pybind
         ${EXTERNAL_PROJECT_LOG_ARGS}
         GIT_REPOSITORY  "https://github.com/pybind/pybind11.git"
-        GIT_TAG         "v2.2.1"
+        GIT_TAG         "v2.1.1"
         PREFIX          ${PYBIND_SOURCE_DIR}
         UPDATE_COMMAND  ""
         CONFIGURE_COMMAND ""
diff --git a/paddle/pybind/protobuf.cc b/paddle/pybind/protobuf.cc
index d80f6b71e9b8b2d926c3c4f1e4c8f22a5c7b86fa..4f959481537d29c089be24f9ae306f860c196c0f 100644
--- a/paddle/pybind/protobuf.cc
+++ b/paddle/pybind/protobuf.cc
@@ -21,24 +21,74 @@ limitations under the License. */
 #include "paddle/framework/program_desc.h"
 #include "paddle/framework/var_desc.h"
 
-using boost::variant;
-
+// Cast boost::variant for PyBind.
+// Copy from
+// https://github.com/pybind/pybind11/issues/576#issuecomment-269563199
 namespace pybind11 {
 namespace detail {
 
+// Can be replaced by a generic lambda in C++14
+struct variant_caster_visitor : public boost::static_visitor<handle> {
+  return_value_policy policy;
+  handle parent;
+
+  variant_caster_visitor(return_value_policy policy, handle parent)
+      : policy(policy), parent(parent) {}
+
+  template <class T>
+  handle operator()(T const &src) const {
+    return make_caster<T>::cast(src, policy, parent);
+  }
+};
+
+template <class Variant>
+struct variant_caster;
+
+template <template <class...> class V, class... Ts>
+struct variant_caster<V<Ts...>> {
+  using Type = V<Ts...>;
+
+  template <typename T>
+  typename std::enable_if<
+      !std::is_same<T, boost::detail::variant::void_>::value, bool>::type
+  try_load(handle src, bool convert) {
+    auto caster = make_caster<T>();
+    if (!load_success_ && caster.load(src, convert)) {
+      load_success_ = true;
+      value = cast_op<T>(caster);
+      return true;
+    }
+    return false;
+  }
+
+  template <typename T>
+  typename std::enable_if<std::is_same<T, boost::detail::variant::void_>::value,
+                          bool>::type
+  try_load(handle src, bool convert) {
+    return false;
+  }
+
+  bool load(handle src, bool convert) {
+    auto unused = {false, try_load<Ts>(src, convert)...};
+    (void)(unused);
+    return load_success_;
+  }
+
+  static handle cast(Type const &src, return_value_policy policy,
+                     handle parent) {
+    variant_caster_visitor visitor(policy, parent);
+    return boost::apply_visitor(visitor, src);
+  }
+
+  PYBIND11_TYPE_CASTER(Type, _("Variant"));
+  bool load_success_{false};
+};
+
 // Add specialization for concrete variant type
 template <class... Args>
 struct type_caster<boost::variant<Args...>>
     : variant_caster<boost::variant<Args...>> {};
 
-template <>
-struct visit_helper<boost::variant> {
-  template <typename... Args>
-  static auto call(Args &&... args) -> decltype(boost::apply_visitor(args...)) {
-    return boost::apply_visitor(args...);
-  }
-};
-
 }  // namespace detail
 }  // namespace pybind11
 
diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index 29459195a8dc1e4e792c13fa56973953fd1be325..ba5d46a3d516267bf8d112954b31ca69971f3795 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/pybind/protobuf.h"
-#include "pybind11/iostream.h"
 
 #include <mutex>  // for call_once
 #include <unordered_map>
@@ -62,8 +61,8 @@ bool IsCompileGPU() {
 #endif
 }
 
-PYBIND11_MODULE(core, m) {
-  m.doc() = "C++ core of PaddlePaddle";
+PYBIND11_PLUGIN(core) {
+  py::module m("core", "C++ core of PaddlePaddle");
 
   // using framework in this function. Since it is inside a function, it will
   // not cause namespace pollution.
@@ -501,8 +500,7 @@ All parameter, weight, gradient are variables in Paddle.
   m.def("enable_profiler", platform::EnableProfiler);
   m.def("disable_profiler", platform::DisableProfiler);
   m.def("reset_profiler", platform::ResetProfiler);
-
-  py::add_ostream_redirect(m, "ostream_redirect");
+  return m.ptr();
 }
 }  // namespace pybind
 }  // namespace paddle
diff --git a/python/paddle/v2/fluid/profiler.py b/python/paddle/v2/fluid/profiler.py
index d0d3fc93e560c906be8d80aecdf1386949621fcb..e5dcf42026cd4762ccbe9a0f4474fa37530435a9 100644
--- a/python/paddle/v2/fluid/profiler.py
+++ b/python/paddle/v2/fluid/profiler.py
@@ -52,22 +52,30 @@ def cuda_profiler(output_file, output_mode=None, config=None):
 
 
 def reset_profiler():
+    """The profiler clear interface.
+    reset_profiler will clear the previous time record.
+    """
     core.reset_profiler()
 
 
 @contextmanager
 def profiler(state, sorted_key=None):
     """The profiler interface.
-    Different from cuda_profiler, this fuction can be used to profile both CPU
-    and GPU program.
+    Different from cuda_profiler, this profiler can be used to profile both CPU
+    and GPU program. By defalut, it records the CPU and GPU operator kernels,
+    if you want to profile other program, you can refer the profiling tutorial
+    to add more records.
 
     Args:
-        state (string) : The profiler state, It should be 'CPU' or 'GPU'.
-        sorted_key (string) : If None, the profiler results will be printed
-            without sorting. Otherwise, the profiler results will be sorted
-            by the this flag. This flag should be one of 'calls', 'total',
-            'max', 'min' or 'ave'.
-            The `calls` means sorting by the calling counter.
+        state (string) : The profiling state, It should be 'CPU' or 'GPU'.
+            Although users may define CPUPlace or CUDAPlace when using Fluid,
+            the profiler doesn't get the state based on this Place. Since the
+            implementation is an independent part from the Fluid.
+        sorted_key (string) : If None, the profiling results will be printed
+            in the order of first end time of events. Otherwise, the profiling
+            results will be sorted by the this flag. This flag should be one
+            of 'calls', 'total', 'max', 'min' or 'ave'.
+            The `calls` means sorting by the number of calls.
             The `total` means sorting by the total execution time.
             The `max` means sorting by the maximum execution time.
             The `min` means sorting by the minimum execution time.
@@ -92,5 +100,6 @@ def profiler(state, sorted_key=None):
         'min': core.EventSortingKey.kMin,
         'ave': core.EventSortingKey.kAve,
     }
-    with core.ostream_redirect(stdout=True, stderr=True):
-        core.disable_profiler(key_map[sorted_key])
+    # TODO(qingqing) : redirect C++ ostream to Python stream.
+    # with core.ostream_redirect(stdout=True, stderr=True):
+    core.disable_profiler(key_map[sorted_key])