diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc
index c226e4e3d2a58d1a647e204c4cd26f4eb6bcd968..9d30887224fe0020ff5665f362e7403bf5c724ee 100644
--- a/paddle/framework/backward.cc
+++ b/paddle/framework/backward.cc
@@ -15,6 +15,8 @@
 #include "paddle/framework/backward.h"
 
 #include <list>
+#include <memory>
+
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
@@ -43,11 +45,11 @@ static bool AllInSet(
   return all_in_set;
 }
 
-static std::shared_ptr<OperatorBase> NOP() {
-  auto net_op = std::make_shared<operators::NetOp>();
+static std::unique_ptr<OperatorBase> NOP() {
+  auto net_op = new operators::NetOp();
   net_op->SetType("@NOP@");
   net_op->CompleteAddOp();
-  return net_op;
+  return std::unique_ptr<OperatorBase>(net_op);
 }
 
 //  Get backward operator from a forward operator, a recursive implementation.
@@ -62,11 +64,7 @@ static std::shared_ptr<OperatorBase> NOP() {
 //  operator, in a complex situation, it maybe a NetOp.
 //
 //  See Backward.h for details
-static std::shared_ptr<OperatorBase> BackwardRecursive(
-    const OperatorBase& forwardOp,
-    std::unordered_set<std::string>& no_grad_names, size_t& uniq_id);
-
-std::shared_ptr<OperatorBase> BackwardRecursive(
+static std::unique_ptr<OperatorBase> BackwardRecursive(
     const OperatorBase& forwardOp,
     std::unordered_set<std::string>& no_grad_names, size_t& uniq_id) {
   //  If all input gradients of forwarding operator do not need to calculate,
@@ -91,7 +89,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
   }
 
   // Returned gradient network
-  auto net = std::make_shared<operators::NetOp>();
+  auto net = std::unique_ptr<operators::NetOp>(new operators::NetOp());
 
   if (forwardOp.IsNetOp()) {
     // Because forwardOp is a net op, it can static_cast.
@@ -105,14 +103,14 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     // reversely travel forwardNet and collect all duplicate outputs.
     for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend();
          ++it, ++local_op_id) {
-      auto fwd = *it;
+      auto& fwd = *it;
       auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id);
-      net->AddOp(bwd);
       ForEachVarName(bwd->Outputs(),
                      [&dup_output_ops, local_op_id](const std::string& out) {
                        dup_output_ops[out].emplace_back(local_op_id);
                        return false;
                      });
+      net->AddOp(std::move(bwd));
     }
     // Get unique ID for this method.
     auto uid = uniq_id++;
@@ -122,7 +120,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     // to handle this case. For each duplicate output, rename it to an alias
     // (original name with a offset), append an `add` op for its operator,
     // and finally sum all the alias variable to the final output variable y.
-    using Pos = std::pair<size_t, std::shared_ptr<OperatorBase>>;
+    using Pos = std::pair<size_t, std::unique_ptr<OperatorBase>>;
     std::list<Pos> insert_position;
     for (auto& dup_output_op : dup_output_ops) {
       const std::string& name = dup_output_op.first;
@@ -150,13 +148,13 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
         [](const Pos& l, const Pos& r) { return l.first > r.first; });
 
     for (auto& pos : insert_position) {
-      net->InsertOp(pos.first + 1, pos.second);
+      net->InsertOp(pos.first + 1, std::move(pos.second));
     }
   } else {
-    std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
+    std::unique_ptr<OperatorBase> grad_op(OpRegistry::CreateGradOp(forwardOp));
 
-    ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net,
-                                       grad_op](const std::string& grad_input) {
+    ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
+                                          const std::string& grad_input) {
       if (no_grad_names.count(grad_input)) {
         // +1 for \0
         std::string prefix = grad_input.substr(
@@ -190,23 +188,23 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
       const auto& stepnet_op =
           *static_cast<const OperatorBase*>(&rnnop.stepnet());
       // create stepnet's gradient op
-      auto grad_stepnet = BackwardRecursive(stepnet_op, no_grad_names, uniq_id);
       rnn_grad_op->set_stepnet(
-          std::static_pointer_cast<operators::NetOp>(grad_stepnet));
+          BackwardRecursive(stepnet_op, no_grad_names, uniq_id));
     }
 
     if (net->ops_.empty()) {  // Current no aux op is added to network
       return grad_op;
     }
-    net->AddOp(grad_op);
+    net->AddOp(std::move(grad_op));
   }
   net->SetType("@GENERATED_BACKWARD@");
   net->CompleteAddOp();
-  return net;
-}  // namespace framework
+  return std::unique_ptr<OperatorBase>(
+      static_cast<OperatorBase*>(net.release()));
+}
 
 // See header for comments
-std::shared_ptr<OperatorBase> Backward(
+std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars) {
   std::unordered_set<std::string> no_grad_names;
diff --git a/paddle/framework/backward.h b/paddle/framework/backward.h
index c181919dc165cf0b49362f85e22ceb4131bbd387..1ecf69881b3126c2904920b9f4b77bfcccc9cf86 100644
--- a/paddle/framework/backward.h
+++ b/paddle/framework/backward.h
@@ -20,7 +20,7 @@ namespace framework {
 
 // Create the backward operator from a forward operator.
 // TODO(yuyang18): Add more API reference comment.
-extern std::shared_ptr<OperatorBase> Backward(
+extern std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars);
 }  // namespace framework
diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc
index d942604bf05998ab9e1ee147b28586e7e4e9777d..1003b1ccd83c85c54f07c8d2b84f993203408941 100644
--- a/paddle/framework/backward_test.cc
+++ b/paddle/framework/backward_test.cc
@@ -180,8 +180,7 @@ TEST(Backward, simple_op_not_need_grad) {
   auto no_input_gop = f::Backward(*fwd, {"x", "b"});
   ASSERT_NE(no_input_gop, nullptr);
   ASSERT_TRUE(no_input_gop->IsNetOp());
-  ASSERT_EQ(0UL,
-            std::static_pointer_cast<ops::NetOp>(no_input_gop)->ops_.size());
+  ASSERT_EQ(0UL, static_cast<ops::NetOp *>(no_input_gop.get())->ops_.size());
 }
 
 TEST(Backward, net_fc_backward_normal) {
diff --git a/paddle/framework/op_registry.cc b/paddle/framework/op_registry.cc
index f801f970f2576452dfa440af78443edf7cfd6be0..8eae86e9605da74cdc37caeb9569e7500aac2a63 100644
--- a/paddle/framework/op_registry.cc
+++ b/paddle/framework/op_registry.cc
@@ -19,7 +19,7 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-std::shared_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
                                                    const VarNameMap& inputs,
                                                    const VarNameMap& outputs,
                                                    AttributeMap attrs) {
@@ -28,10 +28,10 @@ std::shared_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
                  "Operator '%s' has not been registered.", type);
   it->second.checker_->Check(attrs);
   auto op = it->second.creator_(type, inputs, outputs, attrs);
-  return std::shared_ptr<OperatorBase>(op);
+  return std::unique_ptr<OperatorBase>(op);
 }
 
-std::shared_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
   VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
   VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
   AttributeMap attrs;
@@ -55,10 +55,9 @@ OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
   return ret_val;
 }
 
-std::shared_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
+std::unique_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
   PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
-  std::shared_ptr<OperatorBase> grad_op(BuildGradOp(&op));
-  return grad_op;
+  return std::unique_ptr<OperatorBase>(BuildGradOp(&op));
 }
 
 }  // namespace framework
diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h
index a76a4d60b4107dfa3b2a3cea09a443d7b136bcdc..4c2d13d639005d2d2710c19f63988333d89bce13 100644
--- a/paddle/framework/op_registry.h
+++ b/paddle/framework/op_registry.h
@@ -77,17 +77,17 @@ class OpRegistry {
     }
   }
 
-  static std::shared_ptr<OperatorBase> CreateOp(const std::string& type,
+  static std::unique_ptr<OperatorBase> CreateOp(const std::string& type,
                                                 const VarNameMap& inputs,
                                                 const VarNameMap& outputs,
                                                 AttributeMap attrs);
 
-  static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
+  static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
   static VarNameMap ConvertOpDescVarsToVarNameMap(
       const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars);
 
-  static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op);
+  static std::unique_ptr<OperatorBase> CreateGradOp(const OperatorBase& op);
 
   static std::unordered_map<std::string, const OpInfo>& op_info_map() {
     static std::unordered_map<std::string, const OpInfo> op_info_map_;
diff --git a/paddle/framework/op_registry_test.cc b/paddle/framework/op_registry_test.cc
index 1a85d568350dc04ca1df28129de19cd45b5204b8..50c45919c53af22665feeeebe753da283ded2b0c 100644
--- a/paddle/framework/op_registry_test.cc
+++ b/paddle/framework/op_registry_test.cc
@@ -76,8 +76,7 @@ TEST(OpRegistry, CreateOp) {
   attr->set_type(paddle::framework::AttrType::FLOAT);
   attr->set_f(scale);
 
-  std::shared_ptr<paddle::framework::OperatorBase> op =
-      paddle::framework::OpRegistry::CreateOp(op_desc);
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
@@ -118,8 +117,7 @@ TEST(OpRegistry, DefaultValue) {
 
   ASSERT_TRUE(op_desc.IsInitialized());
 
-  std::shared_ptr<paddle::framework::OperatorBase> op =
-      paddle::framework::OpRegistry::CreateOp(op_desc);
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
diff --git a/paddle/framework/pybind.cc b/paddle/framework/pybind.cc
index fe0c87bc570825014222807cb90a3bb341b44e8e..f0114b9e4908d65b3fddb493230777f9e500b4e1 100644
--- a/paddle/framework/pybind.cc
+++ b/paddle/framework/pybind.cc
@@ -48,29 +48,6 @@ namespace framework {
 
 using Tensor = framework::Tensor;
 
-template <typename ClassType>
-void ExposeOperator(ClassType &m) {
-  m.def("infer_shape", &ClassType::type::InferShape)
-      .def("run", &ClassType::type::Run)
-      .def("type",
-           [](const typename ClassType::type &op) -> std::string {
-             return op.Type();
-           })
-      .def("outputs",
-           [](const typename ClassType::type &op)
-               -> std::map<std::string, std::vector<std::string>> {
-                 return op.Outputs();
-               })
-      .def("inputs",
-           [](const typename ClassType::type &op) { return op.Inputs(); })
-      .def("__str__", &ClassType::type::DebugString)
-      .def("no_intermediate_outputs",
-           [](const typename ClassType::type &op) {
-             return op.OutputVars(false);
-           })
-      .def("support_gpu", &ClassType::type::SupportGPU);
-}
-
 static size_t UniqueIntegerGenerator() {
   static std::atomic<size_t> generator;
   return generator.fetch_add(1);
@@ -207,75 +184,69 @@ All parameter, weight, gradient are variables in Paddle.
       .def(py::init<>())
       .def("__str__", string::to_string<const platform::CPUPlace &>);
 
-  py::class_<OperatorBase, std::shared_ptr<OperatorBase>> operator_base(
-      m, "Operator");
-
-  operator_base.def_static("create", [](py::bytes protobin) {
-    OpDesc desc;
-    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
-                   "Cannot parse user input to OpDesc");
-    PADDLE_ENFORCE(desc.IsInitialized(),
-                   "User OpDesc is not initialized, reason %s",
-                   desc.InitializationErrorString());
-    return OpRegistry::CreateOp(desc);
-  });
-
-  operator_base.def("backward",
-                    [](const OperatorBase &forwardOp,
-                       const std::unordered_set<std::string> &no_grad_vars) {
-                      return Backward(forwardOp, no_grad_vars);
-                    });
-
-  ExposeOperator(operator_base);
-
-  py::class_<operators::NetOp, std::shared_ptr<operators::NetOp>> net(m, "Net");
-
-  net.def_static("create",
-                 []() -> std::shared_ptr<operators::NetOp> {
-                   auto retv = std::make_shared<operators::NetOp>();
-                   retv->SetType("plain_net");
-                   return retv;
-                 })
-      .def("add_op", &operators::NetOp::AddOp)
-      .def("add_op",
-           [](operators::NetOp &self,
-              const std::shared_ptr<operators::NetOp> &net) -> void {
-             self.AddOp(std::static_pointer_cast<OperatorBase>(net));
-           })
-      .def("add_op",
-           [](operators::NetOp &self,
-              const std::shared_ptr<operators::RecurrentOp> &rnn) -> void {
-             self.AddOp(std::static_pointer_cast<OperatorBase>(rnn));
+  py::class_<OperatorBase>(m, "Operator")
+      .def_static("create",
+                  [](py::bytes protobin) {
+                    OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    return OpRegistry::CreateOp(desc);
+                  })
+      .def("backward",
+           [](const OperatorBase &forwardOp,
+              const std::unordered_set<std::string> &no_grad_vars) {
+             return Backward(forwardOp, no_grad_vars).release();
            })
+      .def("infer_shape", &OperatorBase::InferShape)
+      .def("run", &OperatorBase::Run)
+      .def("type",
+           [](const OperatorBase &op) -> std::string { return op.Type(); })
+      .def("outputs",
+           [](const OperatorBase &op)
+               -> std::map<std::string, std::vector<std::string>> {
+                 return op.Outputs();
+               })
+      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
+      .def("__str__", &OperatorBase::DebugString)
+      .def("no_intermediate_outputs",
+           [](const OperatorBase &op) { return op.OutputVars(false); })
+      .def("support_gpu", &OperatorBase::SupportGPU);
+
+  py::class_<operators::NetOp, OperatorBase>(m, "Net")
+      .def_static("create",
+                  []() -> operators::NetOp * {
+                    auto *retv = new operators::NetOp;
+                    retv->SetType("plain_net");
+                    return retv;
+                  })
+      .def("add_op", [](operators::NetOp &self,
+                        const OperatorBase &op) { self.AddOp(op); })
       .def("complete_add_op", &operators::NetOp::CompleteAddOp)
       .def("complete_add_op", [](std::shared_ptr<operators::NetOp> &self) {
         self->CompleteAddOp();
       });
 
-  ExposeOperator(net);
-
   // recurrent_op
-  py::class_<operators::RecurrentOp, std::shared_ptr<operators::RecurrentOp>>
-      rnn(m, "RecurrentOp");
-
-  rnn.def_static(
-         "create",
-         [](py::bytes protobin) -> std::shared_ptr<operators::RecurrentOp> {
-           OpDesc desc;
-           PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
-                          "Cannot parse user input to OpDesc");
-           PADDLE_ENFORCE(desc.IsInitialized(),
-                          "User OpDesc is not initialized, reason %s",
-                          desc.InitializationErrorString());
-           auto rnn_op = OpRegistry::CreateOp(desc);
-           return std::dynamic_pointer_cast<operators::RecurrentOp>(rnn_op);
-         })
-      .def("set_stepnet",
-           [](operators::RecurrentOp &self,
-              const std::shared_ptr<operators::NetOp> &net) -> void {
-             self.set_stepnet(net);
-           });
-  ExposeOperator(rnn);
+  py::class_<operators::RecurrentOp, OperatorBase>(m, "RecurrentOp")
+      .def_static(
+          "create",
+          [](py::bytes protobin) -> operators::RecurrentOp * {
+            OpDesc desc;
+            PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                           "Cannot parse user input to OpDesc");
+            PADDLE_ENFORCE(desc.IsInitialized(),
+                           "User OpDesc is not initialized, reason %s",
+                           desc.InitializationErrorString());
+            auto rnn_op = OpRegistry::CreateOp(desc);
+            return static_cast<operators::RecurrentOp *>(rnn_op.release());
+          })
+      .def("set_stepnet", [](operators::RecurrentOp &self,
+                             const operators::NetOp &net) -> void {
+        self.set_stepnet(net.Clone());
+      });
 
   m.def("unique_integer", UniqueIntegerGenerator);
 
diff --git a/paddle/memory/detail/system_allocator.cc b/paddle/memory/detail/system_allocator.cc
index f61e67a32906083881dd7f47433521876be9b355..a270bd59581520859d43cddd2fc0cfa72080f46d 100644
--- a/paddle/memory/detail/system_allocator.cc
+++ b/paddle/memory/detail/system_allocator.cc
@@ -27,7 +27,7 @@ limitations under the License. */
 // between host and device.  Allocates too much would reduce the amount
 // of memory available to the system for paging.  So, by default, we
 // should set false to use_pinned_memory.
-DEFINE_bool(use_pinned_memory, false, "If set, allocate cpu pinned memory.");
+DEFINE_bool(use_pinned_memory, true, "If set, allocate cpu pinned memory.");
 
 namespace paddle {
 namespace memory {
diff --git a/paddle/memory/memory.cc b/paddle/memory/memory.cc
index 207025f9b1c64f0f8943f9fae5edefc9328a1d26..0266bf4f7d65c7aafd4242af41cbd1c71f44bff8 100644
--- a/paddle/memory/memory.cc
+++ b/paddle/memory/memory.cc
@@ -13,22 +13,33 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/memory/memory.h"
+
+#include <algorithm>  // for transform
+#include <cstring>    // for memcpy
+#include <memory>     // for unique_ptr
+#include <mutex>      // for call_once
+
 #include "paddle/memory/detail/buddy_allocator.h"
 #include "paddle/memory/detail/system_allocator.h"
 
-#include <cstring>  // for memcpy
-
 namespace paddle {
 namespace memory {
 
-detail::BuddyAllocator* GetCPUBuddyAllocator() {
-  static detail::BuddyAllocator* a = nullptr;
-  if (a == nullptr) {
-    a = new detail::BuddyAllocator(new detail::CPUAllocator,
-                                   platform::CpuMinChunkSize(),
-                                   platform::CpuMaxChunkSize());
-  }
-  return a;
+using BuddyAllocator = detail::BuddyAllocator;
+
+std::once_flag cpu_allocator_flag;
+std::once_flag gpu_allocator_flag;
+
+BuddyAllocator* GetCPUBuddyAllocator() {
+  static std::unique_ptr<BuddyAllocator> a{nullptr};
+
+  std::call_once(cpu_allocator_flag, [&]() {
+    a.reset(new BuddyAllocator(new detail::CPUAllocator,
+                               platform::CpuMinChunkSize(),
+                               platform::CpuMaxChunkSize()));
+  });
+
+  return a.get();
 }
 
 template <>
@@ -48,20 +59,31 @@ size_t Used<platform::CPUPlace>(platform::CPUPlace place) {
 
 #ifndef PADDLE_ONLY_CPU
 
-detail::BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
-  static detail::BuddyAllocator** as = NULL;
-  if (as == NULL) {
+BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
+  using BuddyAllocVec = std::vector<BuddyAllocator*>;
+  static std::unique_ptr<BuddyAllocVec, void (*)(BuddyAllocVec * p)> as{
+      new BuddyAllocVec, [](BuddyAllocVec* p) {
+        std::for_each(p->begin(), p->end(),
+                      [](BuddyAllocator* p) { delete p; });
+      }};
+
+  // GPU buddy allocators
+  auto& allocators = *as.get();
+
+  // GPU buddy allocator initialization
+  std::call_once(gpu_allocator_flag, [&]() {
     int gpu_num = platform::GetDeviceCount();
-    as = new detail::BuddyAllocator*[gpu_num];
+    allocators.reserve(gpu_num);
     for (int gpu = 0; gpu < gpu_num; gpu++) {
       platform::SetDeviceId(gpu);
-      as[gpu] = new detail::BuddyAllocator(new detail::GPUAllocator,
-                                           platform::GpuMinChunkSize(),
-                                           platform::GpuMaxChunkSize());
+      allocators.emplace_back(new BuddyAllocator(new detail::GPUAllocator,
+                                                 platform::GpuMinChunkSize(),
+                                                 platform::GpuMaxChunkSize()));
     }
-  }
+  });
+
   platform::SetDeviceId(gpu_id);
-  return as[gpu_id];
+  return allocators[gpu_id];
 }
 
 template <>
diff --git a/paddle/operators/gather_test.cc b/paddle/operators/gather_test.cc
index d24d83f299fdb071e60fa3cc7b223c0228cb29af..0ae1e99452973feb6d085dd6ef51e2afca988f59 100644
--- a/paddle/operators/gather_test.cc
+++ b/paddle/operators/gather_test.cc
@@ -45,4 +45,8 @@ TEST(Gather, GatherData) {
 
   for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4);
   for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4);
+
+  delete src;
+  delete index;
+  delete output;
 }
diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h
index fcb703e63bd5a82f9ffac2bf64e06fd0218dbdaa..9848af280b62729bef9243052ceae0b7d8f4c6f5 100644
--- a/paddle/operators/mean_op.h
+++ b/paddle/operators/mean_op.h
@@ -55,9 +55,10 @@ class MeanGradKernel : public framework::OpKernel {
     IG->mutable_data<T>(context.GetPlace());
 
     T ig_size = (T)framework::product(IG->dims());
+    Eigen::DSizes<int, 1> bcast(ig_size);
 
     EigenVector<T>::Flatten(*IG).device(context.GetEigenDevice<Place>()) =
-        EigenScalar<T>::From(*OG) / ig_size;
+        (EigenVector<T>::From(*OG) / ig_size).broadcast(bcast);
   }
 };
 
diff --git a/paddle/operators/net_op.h b/paddle/operators/net_op.h
index 743f0e67dbeaab2de97a6cf635aad0ee90b2cef1..885ac6eeca65998dea62c1db40b9261cceb97805 100644
--- a/paddle/operators/net_op.h
+++ b/paddle/operators/net_op.h
@@ -41,15 +41,13 @@ class NetOp : public framework::OperatorBase {
   NetOp(const std::string& type, const VarNameMap& inputs,
         const VarNameMap& outputs, const framework::AttributeMap& attrs);
 
-  NetOp(const NetOp& o)
-      : framework::OperatorBase(
-            static_cast<const framework::OperatorBase&>(o)) {
+  NetOp(const NetOp& o) : framework::OperatorBase(o.type_, {}, {}, o.attrs_) {
     this->ops_.reserve(o.ops_.size());
-    std::transform(o.ops_.begin(), o.ops_.end(), std::back_inserter(this->ops_),
-                   [](const std::shared_ptr<OperatorBase>& op)
-                       -> std::shared_ptr<OperatorBase> {
-                         return std::shared_ptr<OperatorBase>(op->Clone());
-                       });
+    std::transform(
+        o.ops_.begin(), o.ops_.end(), std::back_inserter(this->ops_),
+        [](const std::unique_ptr<framework::OperatorBase>& op) {
+          return std::unique_ptr<framework::OperatorBase>(op->Clone());
+        });
     this->CompleteAddOp();
   }
 
@@ -86,21 +84,27 @@ class NetOp : public framework::OperatorBase {
     return true;
   }
 
+  void AddOp(const framework::OperatorBase& op) { AddOp(op.Clone()); }
+
   /**
    * @brief Add an operator by ptr
    */
-  void AddOp(const std::shared_ptr<OperatorBase>& op) {
+  void AddOp(std::unique_ptr<framework::OperatorBase> op) {
     PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed");
     PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
-    ops_.push_back(op);
+    ops_.push_back(std::move(op));
   }
 
-  void InsertOp(size_t pos, const std::shared_ptr<OperatorBase>& op) {
+  void InsertOp(size_t pos, std::unique_ptr<framework::OperatorBase> op) {
     PADDLE_ENFORCE(!add_op_done_,
                    "Cannot InsertOp when this network is sealed");
     PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
     PADDLE_ENFORCE_LE(pos, ops_.size(), "Out of range");
-    ops_.insert(ops_.begin() + pos, op);
+    ops_.insert(ops_.begin() + pos, std::move(op));
+  }
+
+  void InsertOp(size_t pos, const framework::OperatorBase& op) {
+    InsertOp(pos, op.Clone());
   }
 
   void CompleteAddOp(bool calculate = true);
@@ -112,7 +116,7 @@ class NetOp : public framework::OperatorBase {
 
   std::unique_ptr<framework::OperatorBase> Clone() const override;
 
-  std::vector<std::shared_ptr<OperatorBase>> ops_;
+  std::vector<std::unique_ptr<framework::OperatorBase>> ops_;
 
  private:
   bool add_op_done_{false};
diff --git a/paddle/operators/net_op_test.cc b/paddle/operators/net_op_test.cc
index e28d4df6a570968205851c2e5b630a14c0492535..e9598610c0a74e08a613a397109ad65994821498 100644
--- a/paddle/operators/net_op_test.cc
+++ b/paddle/operators/net_op_test.cc
@@ -38,15 +38,12 @@ TEST(OpKernel, all) {
   auto net = std::make_shared<NetOp>();
   ASSERT_NE(net, nullptr);
 
-  auto op1 = std::shared_ptr<TestOp>(
+  net->AddOp(std::unique_ptr<TestOp>(
       new TestOp("test", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
-                 {{"Out", {"y"}}}, {}));
-  net->AddOp(op1);
-
-  auto op2 = std::shared_ptr<TestOp>(
+                 {{"Out", {"y"}}}, {})));
+  net->AddOp(std::unique_ptr<TestOp>(
       new TestOp("test", {{"X", {"y"}}, {"W", {"w2"}}, {"b", {"b2"}}},
-                 {{"Out", {"z"}}}, {}));
-  net->AddOp(op2);
+                 {{"Out", {"z"}}}, {})));
 
   net->CompleteAddOp();
   AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"},
@@ -61,21 +58,21 @@ TEST(OpKernel, all) {
 
 TEST(NetOp, insert_op) {
   NetOp net;
-  auto op1 = std::shared_ptr<framework::NOP>(
+  auto op1 = std::unique_ptr<framework::NOP>(
       new framework::NOP("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
                          {{"Out", {"y"}}}, {}));
-  net.AddOp(op1);
-  net.InsertOp(0, op1);
+  net.AddOp(*op1);
+  net.InsertOp(0, *op1);
   ASSERT_EQ(2UL, net.ops_.size());
-  net.InsertOp(2, op1);
+  net.InsertOp(2, std::move(op1));
   ASSERT_EQ(3UL, net.ops_.size());
 }
 
 TEST(NetOp, Clone) {
   NetOp net;
   net.AddOp(
-      std::shared_ptr<framework::NOP>(new framework::NOP{"empty", {}, {}, {}}));
-  net.AddOp(std::shared_ptr<framework::NOP>(
+      std::unique_ptr<framework::NOP>(new framework::NOP{"empty", {}, {}, {}}));
+  net.AddOp(std::unique_ptr<framework::NOP>(
       new framework::NOP{"empty2", {}, {}, {}}));
   net.CompleteAddOp(true);
   auto new_net_op = net.Clone();
diff --git a/paddle/operators/recurrent_op.h b/paddle/operators/recurrent_op.h
index 1d8a6973955cf0b4ab372412fbb5428ff2622a0a..bcfa817de8242153b164fa091309f19a6ad8a246 100644
--- a/paddle/operators/recurrent_op.h
+++ b/paddle/operators/recurrent_op.h
@@ -34,7 +34,8 @@ class RecurrentAlgorithm {
   void Run(const framework::Scope& scope,
            const platform::DeviceContext& dev_ctx) const;
 
-  void Init(rnn::Argument* arg, std::shared_ptr<NetOp>* stepnet) {
+  void Init(rnn::Argument* arg,
+            std::unique_ptr<framework::OperatorBase>* stepnet) {
     PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before.");
     arg_ = arg;
     stepnet_ = stepnet;
@@ -63,7 +64,7 @@ class RecurrentAlgorithm {
   void InitMemories(framework::Scope* step_scopes, bool infer_shape_mode) const;
 
  private:
-  std::shared_ptr<NetOp>* stepnet_;
+  std::unique_ptr<framework::OperatorBase>* stepnet_;
   rnn::Argument* arg_;
   mutable size_t seq_len_;
 };
@@ -80,7 +81,8 @@ class RecurrentGradientAlgorithm {
    * operator.
    */
  public:
-  void Init(rnn::Argument* arg, std::shared_ptr<NetOp>* stepnet) {
+  void Init(rnn::Argument* arg,
+            std::unique_ptr<framework::OperatorBase>* stepnet) {
     PADDLE_ENFORCE_NOT_NULL(stepnet, "stepnet should be set before.");
     arg_ = std::move(arg);
     stepnet_ = stepnet;
@@ -107,7 +109,7 @@ class RecurrentGradientAlgorithm {
  private:
   rnn::Argument* arg_;
   mutable size_t seq_len_;
-  std::shared_ptr<NetOp>* stepnet_;
+  std::unique_ptr<framework::OperatorBase>* stepnet_;
 };
 
 class RecurrentOp : public framework::OperatorBase {
@@ -133,15 +135,17 @@ class RecurrentOp : public framework::OperatorBase {
     alg_.Run(scope, dev_ctx);
   }
 
-  void set_stepnet(std::shared_ptr<NetOp> net) { stepnet_ = net; }
-  const NetOp& stepnet() const { return *stepnet_; }
+  void set_stepnet(std::unique_ptr<OperatorBase> net) {
+    stepnet_ = std::move(net);
+  }
+  const OperatorBase& stepnet() const { return *stepnet_; }
 
   static const rnn::ArgumentName kArgName;
 
  private:
   RecurrentAlgorithm alg_;
   rnn::Argument arg_;
-  std::shared_ptr<NetOp> stepnet_;
+  std::unique_ptr<OperatorBase> stepnet_;
 };
 
 class RecurrentGradientOp : public framework::OperatorBase {
@@ -171,12 +175,14 @@ class RecurrentGradientOp : public framework::OperatorBase {
 
   static const rnn::ArgumentName kArgName;
 
-  void set_stepnet(const std::shared_ptr<NetOp>& net) { stepnet_ = net; }
-  const NetOp& stepnet() const { return *stepnet_; }
+  void set_stepnet(std::unique_ptr<OperatorBase> net) {
+    stepnet_ = std::move(net);
+  }
+  const OperatorBase& stepnet() const { return *stepnet_; }
 
  private:
   RecurrentGradientAlgorithm alg_;
-  std::shared_ptr<NetOp> stepnet_;
+  std::unique_ptr<OperatorBase> stepnet_;
   rnn::Argument arg_;
 };
 
diff --git a/paddle/operators/scatter_test.cc b/paddle/operators/scatter_test.cc
index 4449ce6564396f1971506efb7458c00c834db19f..26fdaff1460a297fa638181641991f732533fe52 100644
--- a/paddle/operators/scatter_test.cc
+++ b/paddle/operators/scatter_test.cc
@@ -49,4 +49,8 @@ TEST(scatter, ScatterUpdate) {
     EXPECT_EQ(output->data<float>()[i], float(i - 4));
   for (size_t i = 8; i < 16; ++i) EXPECT_EQ(p_output[i], float(0));
   for (size_t i = 8; i < 16; ++i) EXPECT_EQ(output->data<float>()[i], float(0));
+
+  delete src;
+  delete index;
+  delete output;
 }
diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc
index d773a4f2d50e82146a729b1cda085ce86ade89cc..761c6de8d4d2150b30b97b58da95da3d5f33db63 100644
--- a/paddle/operators/sigmoid_op.cc
+++ b/paddle/operators/sigmoid_op.cc
@@ -44,7 +44,8 @@ class SigmoidOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    ctx.Output<Tensor>(framework::GradVarName("X"))
+        ->Resize(ctx.Input<Tensor>("Y")->dims());
   }
 };
 
diff --git a/paddle/operators/sigmoid_op.h b/paddle/operators/sigmoid_op.h
index 11ab923eb346c1f8de3a6bbebdfa874b6530004a..b01a9b3f23283471f8846325075719ba0e75ed35 100644
--- a/paddle/operators/sigmoid_op.h
+++ b/paddle/operators/sigmoid_op.h
@@ -37,7 +37,7 @@ class SigmoidKernel : public framework::OpKernel {
     auto Y = EigenVector<T>::Flatten(*output);
     auto place = context.GetEigenDevice<Place>();
 
-    Y.device(place) = 1.0 / (1.0 + (-1.0 * X).exp());
+    Y.device(place) = 1. / (1. + (-X).exp());
   }
 };
 
diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh
index 7c12664aed7957d72c0b2d12ca2afe448eac5a98..2941662f349baf57d1fe8188e88ce21d5de07750 100644
--- a/paddle/scripts/docker/build.sh
+++ b/paddle/scripts/docker/build.sh
@@ -146,7 +146,8 @@ RUN apt-get update &&\
     pip install /*.whl; apt-get install -f -y && \
     apt-get clean -y && \
     rm -f /*.whl && \
-    paddle version
+    paddle version && \
+    ldconfig
 ${DOCKERFILE_CUDNN_DSO}
 ${DOCKERFILE_GPU_ENV}
 ADD go/cmd/pserver/pserver /usr/bin/
diff --git a/python/paddle/v2/framework/tests/CMakeLists.txt b/python/paddle/v2/framework/tests/CMakeLists.txt
index 96fad9b42e04a88fdcbda093683b57451b2a3e41..4c088e7612a93be1b52bc015babee382bbd9026d 100644
--- a/python/paddle/v2/framework/tests/CMakeLists.txt
+++ b/python/paddle/v2/framework/tests/CMakeLists.txt
@@ -25,3 +25,4 @@ py_test(test_operator SRCS test_operator.py)
 # py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
 py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
 py_test(test_recurrent_op SRCS test_recurrent_op.py)
+py_test(test_gradient_checker SRCS test_gradient_checker.py)
diff --git a/python/paddle/v2/framework/tests/gradient_checker.py b/python/paddle/v2/framework/tests/gradient_checker.py
index 501cf6110ff745b8a6022b463bc9cc3a70145c60..8b8e2f444be1169c23784321721c5d8154541fcf 100644
--- a/python/paddle/v2/framework/tests/gradient_checker.py
+++ b/python/paddle/v2/framework/tests/gradient_checker.py
@@ -1,6 +1,7 @@
 import unittest
 
 import numpy
+import itertools
 import paddle.v2.framework.core as core
 from paddle.v2.framework.op import Operator
 
@@ -8,6 +9,7 @@ __all__ = ['get_numeric_gradient']
 
 
 def create_op(op_type):
+    # TODO need to set attrs
     kwargs = dict()
     for in_name in Operator.get_op_input_names(op_type):
         kwargs[in_name] = in_name
@@ -66,7 +68,6 @@ def get_numeric_gradient(op,
             local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace(
             ))
 
-    # TODO(yuyang18): Only CPU is support now.
     cpu_ctx = core.DeviceContext.create(core.CPUPlace())
 
     def get_output():
@@ -109,12 +110,110 @@ def get_numeric_gradient(op,
 
 
 class GradientChecker(unittest.TestCase):
-    def assert_is_close(self, numeric_grads, scope, max_relative_error,
-                        msg_prefix):
-        for name in numeric_grads:
-            b = numpy.array(scope.find_var(grad_var_name(name)).get_tensor())
-            a = numeric_grads[name]
+    def __get_gradient(self, forward_op, backward_op, input_value, grad_names,
+                       place):
+        """Get the input gradients after running forward and backward operators
+        on the given places.
+
+        :param forward_op: forward operator
+        :type forward_op: Operator
+        :param backward_op: backward operator
+        :type backward_op: Operator
+        :param input_value: input values.
+        :type input_value: dict{string:numpy.array}
+        :param grad_names: the names of returned input gradients.
+        :type input_value: a list of string
+        :param place: the device type.
+        :type place: CPUPlace or GPUPlace
+        :return: the input grdients of given grad_names.
+        :rtype: a list of numpy.array
+        """
+        scope = core.Scope()
+        ctx = core.DeviceContext.create(place)
+
+        inputs = forward_op.inputs()
+        in_names = [item for k in inputs for item in inputs[k]]
+        outputs = forward_op.outputs()
+        out_names = [item for k in outputs for item in outputs[k]]
+
+        # create input var and set value
+        for name, value in input_value.iteritems():
+            if name not in in_names:
+                raise ValueError(name + "does not exist in Op's inputs.")
+            var = scope.new_var(name).get_tensor()
+            var.set_dims(value.shape)
+            var.set(value, place)
+
+        # run forward op
+        for out_name in out_names:
+            scope.new_var(out_name)
+        forward_op.infer_shape(scope)
+        forward_op.run(scope, ctx)
+
+        # set output var's shape
+        # set output grad to ones
+        for name in out_names:
+            out_tensor = scope.find_var(name).get_tensor()
+            grad_tensor = scope.new_var(grad_var_name(name)).get_tensor()
+            grad_tensor.set_dims(out_tensor.shape())
+            data = numpy.ones(out_tensor.shape(), dtype=numpy.float32)
+            grad_tensor.set(data, place)
+
+        # run backward op
+        for name in backward_op.outputs():
+            scope.new_var(name)
+        backward_op.infer_shape(scope)
+        backward_op.run(scope, ctx)
+
+        outs = [
+            numpy.array(scope.find_var(name).get_tensor())
+            for name in grad_names
+        ]
+        return outs
+
+    def compare_grad(self, forward_op, input_value):
+        """ Compare the input gradients between CPU and GPU for the given forward
+        operator.
+
+        :param forward_op: forward operator
+        :type forward_op: Operator
+        :param input_value: input values.
+        :type input_value: dict{string:numpy.array}
+        :raises: AssertionError, there is different gradient value.
+        """
+        backward_op = core.Operator.backward(forward_op, set())
+        # return if not compile with GPU or not implementing GPU kernel
+        if not (core.is_compile_gpu() and backward_op.support_gpu()):
+            return
 
+        outputs = backward_op.outputs()
+        out_names = [item for k in outputs for item in outputs[k]]
+        cpu_grads = self.__get_gradient(forward_op, backward_op, input_value,
+                                        out_names, core.CPUPlace())
+        gpu_grads = self.__get_gradient(forward_op, backward_op, input_value,
+                                        out_names, core.GPUPlace(0))
+
+        for c_grad, g_grad, name in itertools.izip(cpu_grads, gpu_grads,
+                                                   out_names):
+            self.assertTrue(
+                numpy.allclose(
+                    c_grad, g_grad, atol=1e-4),
+                "output name: " + name + " has diff")
+
+    def __assert_is_close(self, numeric_grads, analytic_grads, names,
+                          max_relative_error, msg_prefix):
+        """Use relative error for the comparison.
+
+        :param numeric_grads: the numerical graidents.
+        :type numeric_grads: a list of numpy.array 
+        :param analytic_grads: the analytical graidents.
+        :type analytic_grads: a list of numpy.array 
+        :param name: the names of gradients, used to print for debug.
+        :type names: a list of string
+        :param msg_prefix: string info, used to print for debug.
+        :type msf_prefix: string
+        """
+        for a, b, name in itertools.izip(numeric_grads, analytic_grads, names):
             abs_a = numpy.abs(a)
             # if abs_a is nearly zero, then use abs error for a, not relative
             # error.
@@ -159,106 +258,26 @@ class GradientChecker(unittest.TestCase):
 
         inputs = forward_op.inputs()
         in_names = [item for k in inputs for item in inputs[k]]
-        outputs = forward_op.outputs()
-        out_names = [item for k in outputs for item in outputs[k]]
-
         for no_grad in no_grad_set:
             if no_grad not in in_names:
                 raise ValueError("no_grad should be in in_names")
-
         backward_op = core.Operator.backward(forward_op, no_grad_set)
 
-        bwd_outputs = backward_op.outputs()
-        bwd_out_names = [item for k in bwd_outputs for item in bwd_outputs[k]]
-
         places = [core.CPUPlace()]
         if not only_cpu and core.is_compile_gpu() and backward_op.support_gpu():
             places.append(core.GPUPlace(0))
 
-        numeric_grad = dict()
-        # get numeric gradient
-        for check_name in inputs_to_check:
-            numeric_grad[check_name] = \
-                get_numeric_gradient(forward_op, input_vars, output_name,
-                                     check_name)
+        # get numerical gradients
+        numeric_grads = [
+            get_numeric_gradient(forward_op, input_vars, output_name, name)
+            for name in inputs_to_check
+        ]
 
-        # get operator gradient according to different device
+        check_names = [grad_var_name(name) for name in inputs_to_check]
         for place in places:
-            scope = core.Scope()
-            ctx = core.DeviceContext.create(place)
-
-            # create input var and set value
-            for name, value in input_vars.iteritems():
-                if name not in in_names:
-                    raise ValueError(name + " not in op.inputs_")
-                var = scope.new_var(name).get_tensor()
-                var.set_dims(value.shape)
-                var.set(value, place)
-
-            # create output var
-            for out_name in out_names:
-                scope.new_var(out_name).get_tensor()
-
-            # infer the shape of output var and compute/set value of output var
-            forward_op.infer_shape(scope)
-            forward_op.run(scope, ctx)
-
-            # create output grad var
-            # set shape as the output var
-            # set value of this grad to ones
-            for name in out_names:
-                out_tensor = scope.find_var(name).get_tensor()
-                grad_tensor = scope.new_var(grad_var_name(name)).get_tensor()
-                grad_tensor.set_dims(out_tensor.shape())
-                data = 1.0 * numpy.ones(out_tensor.shape())
-                grad_tensor.set(data, place)
-
-            # create input grad var
-            for name in bwd_out_names:
-                scope.new_var(name).get_tensor()
-
-            # infer the shape of input gradient var and compute/set it's value
-            # with backward op
-            backward_op.infer_shape(scope)
-            backward_op.run(scope, ctx)
-
-            self.assert_is_close(numeric_grad, scope, max_relative_error,
-                                 "Gradient Check On %s" % str(place))
-
-
-if __name__ == '__main__':
-
-    class GetNumericGradientTest(unittest.TestCase):
-        def test_add_op(self):
-            add_op = Operator('add_two', X="X", Y="Y", Out="Z")
-            x = numpy.random.random((10, 1)).astype("float32")
-            y = numpy.random.random((10, 1)).astype("float32")
-
-            arr = get_numeric_gradient(add_op, {'X': x, "Y": y}, 'Z', 'X')
-            self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-2)
-
-        def test_softmax_op(self):
-            def stable_softmax(x):
-                """Compute the softmax of vector x in a numerically stable way."""
-                shiftx = x - numpy.max(x)
-                exps = numpy.exp(shiftx)
-                return exps / numpy.sum(exps)
-
-            def label_softmax_grad(Y, dY):
-                dX = Y * 0.0
-                for i in range(Y.shape[0]):
-                    d = numpy.dot(Y[i, :], dY[i, :])
-                    dX[i, :] = Y[i, :] * (dY[i, :] - d)
-                return dX
-
-            softmax_op = Operator("softmax", X="X", Y="Y")
-
-            X = numpy.random.random((2, 2)).astype("float32")
-            Y = numpy.apply_along_axis(stable_softmax, 1, X)
-            dY = numpy.ones(Y.shape)
-            dX = label_softmax_grad(Y, dY)
-
-            arr = get_numeric_gradient(softmax_op, {"X": X}, 'Y', 'X')
-            numpy.testing.assert_almost_equal(arr, dX, decimal=1e-2)
-
-    unittest.main()
+            # get analytical gradients according to different device
+            analytic_grads = self.__get_gradient(forward_op, backward_op,
+                                                 input_vars, check_names, place)
+            self.__assert_is_close(numeric_grads, analytic_grads, check_names,
+                                   max_relative_error,
+                                   "Gradient Check On %s" % str(place))
diff --git a/python/paddle/v2/framework/tests/test_gradient_checker.py b/python/paddle/v2/framework/tests/test_gradient_checker.py
new file mode 100644
index 0000000000000000000000000000000000000000..e0b315120862bea284e067070492dcdfbb661081
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_gradient_checker.py
@@ -0,0 +1,43 @@
+import unittest
+import numpy
+from paddle.v2.framework.op import Operator
+from gradient_checker import GradientChecker
+from gradient_checker import get_numeric_gradient
+
+
+class GetNumericGradientTest(unittest.TestCase):
+    def test_add_op(self):
+        add_op = Operator('add_two', X="X", Y="Y", Out="Z")
+        x = numpy.random.random((10, 1)).astype("float32")
+        y = numpy.random.random((10, 1)).astype("float32")
+
+        arr = get_numeric_gradient(add_op, {'X': x, "Y": y}, 'Z', 'X')
+        self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-4)
+
+    def test_softmax_op(self):
+        def stable_softmax(x):
+            """Compute the softmax of vector x in a numerically stable way."""
+            shiftx = x - numpy.max(x)
+            exps = numpy.exp(shiftx)
+            return exps / numpy.sum(exps)
+
+        def label_softmax_grad(Y, dY):
+            dX = Y * 0.0
+            for i in range(Y.shape[0]):
+                d = numpy.dot(Y[i, :], dY[i, :])
+                dX[i, :] = Y[i, :] * (dY[i, :] - d)
+            return dX
+
+        softmax_op = Operator("softmax", X="X", Y="Y")
+
+        X = numpy.random.random((2, 2)).astype("float32")
+        Y = numpy.apply_along_axis(stable_softmax, 1, X)
+        dY = numpy.ones(Y.shape)
+        dX = label_softmax_grad(Y, dY)
+
+        arr = get_numeric_gradient(softmax_op, {"X": X}, 'Y', 'X')
+        numpy.testing.assert_almost_equal(arr, dX, decimal=1e-2)
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_mean_op.py b/python/paddle/v2/framework/tests/test_mean_op.py
index b5d52b90567bcd0c9f376147145d8638049f7bab..f32b3160d651a290823223c46c45bb3b6950a505 100644
--- a/python/paddle/v2/framework/tests/test_mean_op.py
+++ b/python/paddle/v2/framework/tests/test_mean_op.py
@@ -1,5 +1,6 @@
 import unittest
 from op_test_util import OpTestMeta
+from gradient_checker import GradientChecker, create_op
 import numpy as np
 
 
@@ -12,5 +13,12 @@ class TestMeanOp(unittest.TestCase):
         self.outputs = {'Out': np.mean(self.inputs['X'])}
 
 
+class MeanGradOpTest(GradientChecker):
+    def test_normal(self):
+        op = create_op("mean")
+        inputs = {"X": np.random.random((10, 10)).astype("float32")}
+        self.check_grad(op, inputs, set("X"), "Out")
+
+
 if __name__ == '__main__':
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_sigmoid_op.py b/python/paddle/v2/framework/tests/test_sigmoid_op.py
index 2a57a41ed8b718fd420062ba68e853a4861b7359..273c2e5ab1a84d12621fe9568c4cf22073b6aed4 100644
--- a/python/paddle/v2/framework/tests/test_sigmoid_op.py
+++ b/python/paddle/v2/framework/tests/test_sigmoid_op.py
@@ -1,6 +1,7 @@
 import unittest
-from op_test_util import OpTestMeta
 import numpy as np
+from op_test_util import OpTestMeta
+from gradient_checker import GradientChecker, create_op
 
 
 class TestSigmoidOp(unittest.TestCase):
@@ -8,12 +9,20 @@ class TestSigmoidOp(unittest.TestCase):
 
     def setUp(self):
         self.type = "sigmoid"
-        self.inputs = {'X': np.random.random((32, 100)).astype("float32")}
+        self.inputs = {'X': np.random.random((15, 31)).astype("float32")}
         self.outputs = {'Y': 1 / (1 + np.exp(-self.inputs['X']))}
 
 
-#class TestSigmoidGradOp(unittest.TestCase):
-#TODO(qingqing) add unit test
+class TestSigmoidGradOp(GradientChecker):
+    def test_grad(self):
+        op = create_op("sigmoid")
+        inputs = {"X": np.random.uniform(0.1, 1, [11, 17]).astype("float32")}
+        # compare gpu and cpu results for backward op.
+        # this test will be skiped if only compiling CPU version.
+        self.compare_grad(op, inputs)
+        # check gradients 
+        self.check_grad(op, inputs, set("X"), "Y", max_relative_error=0.007)
+
 
 if __name__ == '__main__':
     unittest.main()