diff --git a/cmake/external/pybind11.cmake b/cmake/external/pybind11.cmake
index c885877a2bcd622dc8d1f3ca2821a0bed84561bf..3a10ea945d3d1a3fa88954196905872b2668d5a0 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         "v2.1.1"
+        GIT_TAG         "v2.2.4"
         PREFIX          ${PYBIND_SOURCE_DIR}
         UPDATE_COMMAND  ""
         CONFIGURE_COMMAND ""
diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec
index 8397ae093ba1b02aa35a9ba75130cdbaef41a978..0a71f153434cc89b26d97e1e1fcfae6c2bc88566 100644
--- a/paddle/fluid/API.spec
+++ b/paddle/fluid/API.spec
@@ -26,10 +26,10 @@ paddle.fluid.release_memory ArgSpec(args=['input_program', 'skip_opt_set'], vara
 paddle.fluid.DistributeTranspilerConfig.__init__ 
 paddle.fluid.ParallelExecutor.__init__ ArgSpec(args=['self', 'use_cuda', 'loss_name', 'main_program', 'share_vars_from', 'exec_strategy', 'build_strategy', 'num_trainers', 'trainer_id', 'scope'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 1, 0, None))
 paddle.fluid.ParallelExecutor.run ArgSpec(args=['self', 'fetch_list', 'feed', 'feed_dict', 'return_numpy'], varargs=None, keywords=None, defaults=(None, None, True))
-paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ExecutionStrategy) -> None
-paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.GradientScaleStrategy, arg0: int) -> None
-paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.core.ReduceStrategy, arg0: int) -> None
-paddle.fluid.BuildStrategy.__init__ __init__(self: paddle.fluid.core.BuildStrategy) -> None
+paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.ExecutionStrategy) -> None
+paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.GradientScaleStrategy, arg0: int) -> None
+paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.ReduceStrategy, arg0: int) -> None
+paddle.fluid.BuildStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy) -> None
 paddle.fluid.create_lod_tensor ArgSpec(args=['data', 'recursive_seq_lens', 'place'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.create_random_int_lodtensor ArgSpec(args=['recursive_seq_lens', 'base_shape', 'place', 'low', 'high'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.io.save_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None))
diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc
index 586e92c2b3146d75a673d1fe326dbee7297a3bfb..0443ff3fc3d055c54527240cdf57c195e1c0acf8 100644
--- a/paddle/fluid/pybind/protobuf.cc
+++ b/paddle/fluid/pybind/protobuf.cc
@@ -30,11 +30,12 @@ namespace pybind11 {
 namespace detail {
 
 // Can be replaced by a generic lambda in C++14
-struct variant_caster_visitor : public boost::static_visitor<handle> {
+struct __attribute__((visibility("hidden"))) paddle_variant_caster_visitor
+    : public boost::static_visitor<handle> {
   return_value_policy policy;
   handle parent;
 
-  variant_caster_visitor(return_value_policy policy, handle parent)
+  paddle_variant_caster_visitor(return_value_policy policy, handle parent)
       : policy(policy), parent(parent) {}
 
   template <class T>
@@ -44,10 +45,10 @@ struct variant_caster_visitor : public boost::static_visitor<handle> {
 };
 
 template <class Variant>
-struct variant_caster;
+struct paddle_variant_caster;
 
 template <template <class...> class V, class... Ts>
-struct variant_caster<V<Ts...>> {
+struct paddle_variant_caster<V<Ts...>> {
   using Type = V<Ts...>;
 
   template <typename T>
@@ -90,7 +91,7 @@ struct variant_caster<V<Ts...>> {
 
   static handle cast(Type const &src, return_value_policy policy,
                      handle parent) {
-    variant_caster_visitor visitor(policy, parent);
+    paddle_variant_caster_visitor visitor(policy, parent);
     return boost::apply_visitor(visitor, src);
   }
 
@@ -101,7 +102,7 @@ struct variant_caster<V<Ts...>> {
 // Add specialization for concrete variant type
 template <class... Args>
 struct type_caster<boost::variant<Args...>>
-    : variant_caster<boost::variant<Args...>> {};
+    : paddle_variant_caster<boost::variant<Args...>> {};
 
 }  // namespace detail
 }  // namespace pybind11
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index 795800fd51763759c0f660e3eb60625afe669881..a2a629acdfe65ae250164dad4c2367d525887acf 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -86,12 +86,12 @@ bool IsCompiledWithDIST() {
 #endif
 }
 
-PYBIND11_PLUGIN(core) {
+PYBIND11_MODULE(core, m) {
   // Not used, just make sure cpu_info.cc is linked.
   paddle::platform::CpuTotalPhysicalMemory();
 
   paddle::memory::allocation::UseAllocatorStrategyGFlag();
-  py::module m("core", "C++ core of PaddlePaddle");
+  m.doc() = "C++ core of PaddlePaddle";
 
   // using framework in this function. Since it is inside a function, it will
   // not cause namespace pollution.
@@ -907,7 +907,6 @@ All parameter, weight, gradient are variables in Paddle.
       });
 
   BindRecordIOWriter(&m);
-  return m.ptr();
 }
 }  // namespace pybind
 }  // namespace paddle
diff --git a/paddle/fluid/pybind/tensor_py.h b/paddle/fluid/pybind/tensor_py.h
index b39323f843f8dbf5a7e4bac841c8cb8ed7efdc07..02a75236f6c7c7a64f2aa110ca7a7e3d92832fe9 100644
--- a/paddle/fluid/pybind/tensor_py.h
+++ b/paddle/fluid/pybind/tensor_py.h
@@ -21,7 +21,6 @@ limitations under the License. */
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/float16.h"
-#include "pybind11/common.h"
 #include "pybind11/numpy.h"
 #include "pybind11/pybind11.h"
 
diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh
index 9632eaec005df6cdfd289f0e4e6879ad2168cdeb..86925b26e7e894d7dba1feac68b79e66316752d7 100755
--- a/paddle/scripts/paddle_build.sh
+++ b/paddle/scripts/paddle_build.sh
@@ -149,7 +149,7 @@ function cmake_gen() {
             elif [ "$1" == "cp37-cp37m" ]; then
                 export LD_LIBRARY_PATH=/opt/_internal/cpython-3.7.0/lib/:${LD_LIBRARY_PATH}
                 export PATH=/opt/_internal/cpython-3.7.0/bin/:${PATH}
-                export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/_internal/cpython-3.7.0/bin/python3
+                export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/_internal/cpython-3.7.0/bin/python3.7
             -DPYTHON_INCLUDE_DIR:PATH=/opt/_internal/cpython-3.7.0/include/python3.7m
             -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-3.7.0/lib/libpython3.so"
            fi