diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec
index 40063f158e..c4a65bb13e 100644
--- a/paddle/fluid/API.spec
+++ b/paddle/fluid/API.spec
@@ -160,7 +160,16 @@ paddle.fluid.layers.relu ArgSpec(args=['x', 'name'], varargs=None, keywords=None
 paddle.fluid.layers.log ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.elu ArgSpec(args=['x', 'alpha', 'name'], varargs=None, keywords=None, defaults=(1.0, None))
+paddle.fluid.layers.relu6 ArgSpec(args=['x', 'threshold', 'name'], varargs=None, keywords=None, defaults=(6.0, None))
+paddle.fluid.layers.pow ArgSpec(args=['x', 'factor', 'name'], varargs=None, keywords=None, defaults=(1.0, None))
+paddle.fluid.layers.stanh ArgSpec(args=['x', 'scale_a', 'scale_b', 'name'], varargs=None, keywords=None, defaults=(0.6666666666666666, 1.7159, None))
+paddle.fluid.layers.hard_sigmoid ArgSpec(args=['x', 'slope', 'offset', 'name'], varargs=None, keywords=None, defaults=(0.2, 0.5, None))
+paddle.fluid.layers.swish ArgSpec(args=['x', 'beta', 'name'], varargs=None, keywords=None, defaults=(1.0, None))
 paddle.fluid.layers.prelu ArgSpec(args=['x', 'mode', 'param_attr', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.brelu ArgSpec(args=['x', 't_min', 't_max', 'name'], varargs=None, keywords=None, defaults=(0.0, 24.0, None))
+paddle.fluid.layers.leaky_relu ArgSpec(args=['x', 'alpha', 'name'], varargs=None, keywords=None, defaults=(0.02, None))
+paddle.fluid.layers.soft_relu ArgSpec(args=['x', 'threshold', 'name'], varargs=None, keywords=None, defaults=(40.0, None))
 paddle.fluid.layers.flatten ArgSpec(args=['x', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, None))
 paddle.fluid.layers.sequence_mask ArgSpec(args=['x', 'maxlen', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 'int64', None))
 paddle.fluid.layers.stack ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=(0,))
@@ -169,6 +178,14 @@ paddle.fluid.layers.unstack ArgSpec(args=['x', 'axis', 'num'], varargs=None, key
 paddle.fluid.layers.sequence_enumerate ArgSpec(args=['input', 'win_size', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0, None))
 paddle.fluid.layers.expand ArgSpec(args=['x', 'expand_times', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_concat ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.scale ArgSpec(args=['x', 'scale', 'bias', 'bias_after_scale', 'act', 'name'], varargs=None, keywords=None, defaults=(1.0, 0.0, True, None, None))
+paddle.fluid.layers.elementwise_add ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
+paddle.fluid.layers.elementwise_div ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
+paddle.fluid.layers.elementwise_sub ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
+paddle.fluid.layers.elementwise_mul ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
+paddle.fluid.layers.elementwise_max ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
+paddle.fluid.layers.elementwise_min ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
+paddle.fluid.layers.elementwise_pow ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
@@ -233,15 +250,7 @@ paddle.fluid.layers.Print ArgSpec(args=['input', 'first_n', 'message', 'summariz
 paddle.fluid.layers.is_empty ArgSpec(args=['x', 'cond'], varargs=None, keywords='ignored', defaults=(None,))
 paddle.fluid.layers.mean ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.mul ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.scale ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.elementwise_add ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.elementwise_div ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.elementwise_sub ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.elementwise_mul ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.elementwise_max ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.elementwise_min ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.elementwise_pow ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.clip ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.clip_by_norm ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.logical_and ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
@@ -256,32 +265,23 @@ paddle.fluid.layers.sum ArgSpec(args=[], varargs='args', keywords='kwargs', defa
 paddle.fluid.layers.slice ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.shape ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.maxout ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.sigmoid ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.logsigmoid ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.exp ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.tanh ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.tanh_shrink ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.softshrink ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.sqrt ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.abs ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.ceil ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.floor ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.cos ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.sin ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.round ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.reciprocal ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.square ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.softplus ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.softsign ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.brelu ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.leaky_relu ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.soft_relu ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.elu ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.relu6 ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.pow ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.stanh ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.hard_sigmoid ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.swish ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
+paddle.fluid.layers.sigmoid ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.logsigmoid ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.exp ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.tanh ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.tanh_shrink ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.sqrt ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.abs ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.ceil ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.floor ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.cos ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.sin ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.round ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.reciprocal ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.square ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.softplus ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.softsign ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.uniform_random ArgSpec(args=['shape', 'dtype', 'min', 'max', 'seed'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.layers.hard_shrink ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.cumsum ArgSpec(args=['x', 'axis', 'exclusive', 'reverse'], varargs=None, keywords=None, defaults=(None, None, None))
diff --git a/paddle/fluid/framework/details/reference_count_op_handle.h b/paddle/fluid/framework/details/reference_count_op_handle.h
index 71db8d952f..fc479a4c4a 100644
--- a/paddle/fluid/framework/details/reference_count_op_handle.h
+++ b/paddle/fluid/framework/details/reference_count_op_handle.h
@@ -22,6 +22,7 @@
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/garbage_collector.h"
 #include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/tensor.h"
 
 namespace paddle {
@@ -46,17 +47,15 @@ class ReferenceCountOpHandle : public OpHandleBase {
                          const std::vector<std::string> &var_names,
                          GarbageCollector<Tensor> *gc,
                          AtomicReferenceCountMap *ref_cnts)
-      : OpHandleBase(node),
-        scope_(scope),
-        var_names_(var_names),
-        gc_(gc),
-        ref_cnts_(ref_cnts) {
+      : OpHandleBase(node), scope_(scope), gc_(gc), ref_cnts_(ref_cnts) {
     dev_ctx_ = static_cast<platform::CUDADeviceContext *>(
         platform::DeviceContextPool::Instance().Get(place));
     if (IsStreamGarabageCollector()) {
       PADDLE_ENFORCE(cudaSetDevice(place.device));
       PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
     }
+
+    for (auto &name : var_names) AddVar(name);
   }
 
   ~ReferenceCountOpHandle() {
@@ -69,19 +68,35 @@ class ReferenceCountOpHandle : public OpHandleBase {
 
   std::string Name() const override { return "reference_count"; }
 
+  void AddVar(const std::string &name) {
+    auto it = var_names_.find(name);
+    if (it != var_names_.end())
+      ++(it->second);
+    else
+      var_names_[name] = 1;
+  }
+
  protected:
   void RunImpl() override {
     auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
-    std::vector<LoDTensor *> tensors;
-    for (auto &name : var_names_) {
+    std::vector<Tensor *> tensors;
+    for (auto &pair : var_names_) {
+      auto &name = pair.first;
       auto it = ref_cnts_->find(name);
       if (it == ref_cnts_->end()) continue;
 
       auto *var = exec_scope->FindVar(name);
-      if (var == nullptr || !var->IsType<LoDTensor>()) continue;
-
-      if (it->second.fetch_sub(1) <= 1) {
-        tensors.emplace_back(var->GetMutable<LoDTensor>());
+      if (var == nullptr) continue;
+
+      if (var->IsType<LoDTensor>()) {
+        if (it->second.fetch_sub(pair.second) <= pair.second) {
+          tensors.emplace_back(var->GetMutable<LoDTensor>());
+        }
+      } else if (var->IsType<SelectedRows>()) {
+        if (it->second.fetch_sub(pair.second) <= pair.second) {
+          tensors.emplace_back(
+              var->GetMutable<SelectedRows>()->mutable_value());
+        }
       }
     }
 
@@ -91,7 +106,7 @@ class ReferenceCountOpHandle : public OpHandleBase {
   }
 
  private:
-  void ClearTensors(const std::vector<LoDTensor *> &tensors) {
+  void ClearTensors(const std::vector<Tensor *> &tensors) {
     auto *gc = dynamic_cast<StreamGarbageCollector<Tensor> *>(gc_);
     if (gc != nullptr) {
       auto compute_stream = dev_ctx_->stream();
@@ -112,7 +127,7 @@ class ReferenceCountOpHandle : public OpHandleBase {
 
   const Scope *scope_;
   platform::CUDADeviceContext *dev_ctx_;
-  std::vector<std::string> var_names_;
+  std::unordered_map<std::string, int> var_names_;
   GarbageCollector<Tensor> *gc_;       // not own
   AtomicReferenceCountMap *ref_cnts_;  // not own
   cudaEvent_t event_;
diff --git a/paddle/fluid/framework/details/reference_count_pass.cc b/paddle/fluid/framework/details/reference_count_pass.cc
index 344754d5a1..b1ce551ce7 100644
--- a/paddle/fluid/framework/details/reference_count_pass.cc
+++ b/paddle/fluid/framework/details/reference_count_pass.cc
@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <queue>
 #include <string>
 #include <vector>
 
@@ -23,6 +24,25 @@ namespace paddle {
 namespace framework {
 namespace details {
 
+static ComputationOpHandle *FindNextComputationOpHandle(VarHandle *var_in) {
+  std::queue<VarHandleBase *> queue;
+  queue.push(var_in);
+  do {
+    auto *var = queue.front();
+    queue.pop();
+    for (auto *op : var->PendingOps()) {
+      auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
+      if (compute_op != nullptr && compute_op->GetPlace() == var_in->place_) {
+        return compute_op;
+      }
+      for (auto *out_var : op->Outputs()) {
+        queue.push(out_var);
+      }
+    }
+  } while (!queue.empty());
+  return nullptr;
+}
+
 std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
   auto &ref_cnts = Get<DeviceReferenceCountMap>(kGlobalReferenceCount);
@@ -34,6 +54,9 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
   // Step 2: Find all variables in non-computation ops which refers to variables
   // in computation ops
   std::unordered_set<std::string> names;
+  std::unordered_map<OpHandleBase *, std::unique_ptr<ReferenceCountOpHandle>>
+      compute_ref_cnt_map;
+
   auto get_ref_cnts_from_compute_op = [&](
       const std::unique_ptr<OpHandleBase> &op,
       const std::vector<VarHandleBase *> &vars) {
@@ -54,15 +77,18 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
       VarDesc *var_desc = var_handle->Node()->Var();
       auto var_name = var_handle->Node()->Name();
 
-      // This is wierd but there is really some variables without var_desc
+      // This is weird but there is really some variables without var_desc
       // in computation_op
       if (var_desc == nullptr) {
         if (compute_op->Node()->Op()->Block()->FindVar(var_name) == nullptr)
           continue;
       } else {
-        if (var_desc->Persistable() ||
-            var_desc->Proto()->type().type() != proto::VarType::LOD_TENSOR)
+        if (var_desc->Persistable()) continue;
+        auto var_type = var_desc->Proto()->type().type();
+        if (var_type != proto::VarType::LOD_TENSOR &&
+            var_type != proto::VarType::SELECTED_ROWS) {
           continue;
+        }
       }
 
       // compute op only runs in one device
@@ -93,12 +119,33 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
       if (ref_cnts.count(place.device) &&
           ref_cnts[place.device]->count(var_name)) {
         ++(*ref_cnts[place.device])[var_name];
+
+        auto *next_compute_op = FindNextComputationOpHandle(var_handle);
+        if (next_compute_op != nullptr) {
+          if (compute_ref_cnt_map.count(next_compute_op)) {
+            compute_ref_cnt_map[next_compute_op]->AddVar(var_name);
+            VLOG(5) << "Add reference count of " << var_name << " to Operator "
+                    << next_compute_op->Name();
+          } else {
+            // Create new reference_count_op_handle
+            ir::Node *ref_cnt_node = graph->CreateEmptyNode(
+                "reference_count", ir::Node::Type::kOperation);
+            auto *ref_cnt_handle = new ReferenceCountOpHandle(
+                ref_cnt_node, next_compute_op->GetScope(), place, {var_name},
+                gcs[place.device].get(), cur_ref_cnts[place.device].get());
+            if (next_compute_op->Outputs().empty()) {
+              auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
+              next_compute_op->AddOutput(dep_var);
+              graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
+            }
+            ref_cnt_handle->AddInput(next_compute_op->Outputs().front());
+            compute_ref_cnt_map[next_compute_op].reset(ref_cnt_handle);
+          }
+        }
       }
     }
   };
 
-  std::unordered_map<OpHandleBase *, ReferenceCountOpHandle *>
-      compute_ref_cnt_map;
   auto &all_ops = graph->Get<GraphOps>(kGraphOps);
   for (auto &op : all_ops) {
     auto in_var_names = get_ref_cnts_from_compute_op(op, op->Inputs());
@@ -113,11 +160,13 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
     auto *ref_cnt_handle = new ReferenceCountOpHandle(
         ref_cnt_node, compute_op->GetScope(), place, in_var_names,
         gcs[place.device].get(), cur_ref_cnts[place.device].get());
-    auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
-    compute_op->AddOutput(dep_var);
-    ref_cnt_handle->AddInput(dep_var);
-    graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
-    compute_ref_cnt_map[compute_op] = ref_cnt_handle;
+    if (compute_op->Outputs().empty()) {
+      auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
+      compute_op->AddOutput(dep_var);
+      graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
+    }
+    ref_cnt_handle->AddInput(compute_op->Outputs().front());
+    compute_ref_cnt_map[compute_op].reset(ref_cnt_handle);
   }
 
   for (auto &op : all_ops) {
@@ -131,7 +180,11 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
     new_all_ops.emplace_back(std::move(op));
     auto it = compute_ref_cnt_map.find(new_all_ops.back().get());
     if (it != compute_ref_cnt_map.end()) {
-      new_all_ops.emplace_back(it->second);
+      // Add LeafNode to ReferenceCountOpHandle
+      auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
+      graph->Get<GraphDepVars>(kGraphDepVars).emplace(dummy_leaf);
+      it->second->AddOutput(dummy_leaf);
+      new_all_ops.emplace_back(std::move(it->second));
     }
   }
 
diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc
index 86f6147cf7..17f942571d 100644
--- a/paddle/fluid/framework/op_desc.cc
+++ b/paddle/fluid/framework/op_desc.cc
@@ -54,6 +54,10 @@ class CompileTimeInferShapeContext : public InferShapeContext {
                 size_t j = 0) const override {
     PADDLE_ENFORCE_LT(i, Inputs(in).size());
     PADDLE_ENFORCE_LT(j, Outputs(out).size());
+    PADDLE_ENFORCE(Inputs(in)[i] != framework::kEmptyVarName,
+                   "The %s[%d] is @EMPTY@", in, i);
+    PADDLE_ENFORCE(Outputs(out)[j] != framework::kEmptyVarName,
+                   "The %s[%d] is @EMPTY@", out, j);
     auto *in_var = block_.FindVarRecursive(Inputs(in)[i]);
     auto *out_var = block_.FindVarRecursive(Outputs(out)[j]);
     if (in_var->GetType() != proto::VarType::LOD_TENSOR) {
@@ -63,6 +67,7 @@ class CompileTimeInferShapeContext : public InferShapeContext {
     PADDLE_ENFORCE_EQ(in_var->GetType(), proto::VarType::LOD_TENSOR,
                       "The %d-th output of Output(%s) must be LoDTensor.", j,
                       out);
+
     out_var->SetLoDLevel(in_var->GetLoDLevel());
   }
 
diff --git a/paddle/fluid/framework/shape_inference.cc b/paddle/fluid/framework/shape_inference.cc
index ddff2c7c26..89eb00ff65 100644
--- a/paddle/fluid/framework/shape_inference.cc
+++ b/paddle/fluid/framework/shape_inference.cc
@@ -46,6 +46,16 @@ std::vector<DDim> InferShapeContext::GetReaderDims(
   return this->GetRepeatedDims(arg_names[0]);
 }
 
+void InferShapeContext::ShareLoDs(const std::string &in,
+                                  const std::string &out) const {
+  PADDLE_ENFORCE_EQ(Inputs(in).size(), Outputs(out).size(),
+                    "The number of arguments in %s and %s is not equal.", in,
+                    out);
+  for (size_t i = 0; i < in.size(); ++i) {
+    ShareLoD(in, out, i, i);
+  }
+}
+
 DDim InferShapeContext::GetInputsElementDim(const std::string &name,
                                             int idx) const {
   const std::vector<std::string> &names = Inputs(name);
diff --git a/paddle/fluid/framework/shape_inference.h b/paddle/fluid/framework/shape_inference.h
index 5f497cafa0..fd220d961a 100644
--- a/paddle/fluid/framework/shape_inference.h
+++ b/paddle/fluid/framework/shape_inference.h
@@ -56,6 +56,8 @@ class InferShapeContext {
   virtual const std::vector<std::string> &Outputs(
       const std::string &name) const = 0;
 
+  void ShareLoDs(const std::string &in, const std::string &out) const;
+
   virtual void ShareLoD(const std::string &in, const std::string &out,
                         size_t i = 0, size_t j = 0) const = 0;
 
diff --git a/paddle/fluid/inference/api/api_impl.h b/paddle/fluid/inference/api/api_impl.h
index ec801c5885..6ecc32a700 100644
--- a/paddle/fluid/inference/api/api_impl.h
+++ b/paddle/fluid/inference/api/api_impl.h
@@ -20,10 +20,9 @@
 #include <string>
 #include <vector>
 
-#include "paddle/fluid/inference/api/paddle_inference_api.h"
-
 #include "paddle/fluid/framework/ddim.h"
 #include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/io.h"
 #include "paddle/fluid/platform/init.h"
 #include "paddle/fluid/platform/profiler.h"
diff --git a/paddle/fluid/inference/tests/api/tester_helper.h b/paddle/fluid/inference/tests/api/tester_helper.h
index 384a40a3f9..05cd343433 100644
--- a/paddle/fluid/inference/tests/api/tester_helper.h
+++ b/paddle/fluid/inference/tests/api/tester_helper.h
@@ -74,8 +74,8 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
   }
 }
 
-std::unique_ptr<PaddlePredictor> GetPrediction(AnalysisConfig config,
-                                               bool use_analysis = true) {
+std::unique_ptr<PaddlePredictor> CreateTestPredictor(
+    const AnalysisConfig &config, bool use_analysis = true) {
   if (use_analysis) {
     return CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
         config);
@@ -92,7 +92,7 @@ size_t GetSize(const PaddleTensor &out) {
 
 std::unordered_map<std::string, int> GetFuseStatis(AnalysisConfig config,
                                                    int *num_ops) {
-  auto predictor = GetPrediction(config);
+  auto predictor = CreateTestPredictor(config);
   AnalysisPredictor *analysis_predictor =
       dynamic_cast<AnalysisPredictor *>(predictor.get());
   auto &fuse_statis = analysis_predictor->analysis_argument()
@@ -113,11 +113,12 @@ std::unordered_map<std::string, int> GetFuseStatis(AnalysisConfig config,
 }
 
 void TestOneThreadPrediction(
-    AnalysisConfig config, const std::vector<std::vector<PaddleTensor>> inputs,
+    const AnalysisConfig &config,
+    const std::vector<std::vector<PaddleTensor>> &inputs,
     std::vector<PaddleTensor> *outputs, bool use_analysis = true) {
   int batch_size = FLAGS_batch_size;
   int num_times = FLAGS_repeat;
-  auto predictor = GetPrediction(config, use_analysis);
+  auto predictor = CreateTestPredictor(config, use_analysis);
   Timer timer;
   timer.tic();
   for (int i = 0; i < num_times; i++) {
@@ -130,7 +131,8 @@ void TestOneThreadPrediction(
 }
 
 void TestMultiThreadPrediction(
-    AnalysisConfig config, const std::vector<std::vector<PaddleTensor>> inputs,
+    const AnalysisConfig &config,
+    const std::vector<std::vector<PaddleTensor>> &inputs,
     std::vector<PaddleTensor> *outputs, int num_threads,
     bool use_analysis = true) {
   int batch_size = FLAGS_batch_size;
@@ -140,7 +142,7 @@ void TestMultiThreadPrediction(
   // TODO(yanchunwei): Bug here, the analyzer phase can't be parallelled
   // because AttentionLSTM's hard code nodeid will be damanged.
   for (int tid = 0; tid < num_threads; ++tid) {
-    predictors.emplace_back(GetPrediction(config, use_analysis));
+    predictors.emplace_back(CreateTestPredictor(config, use_analysis));
   }
   for (int tid = 0; tid < num_threads; ++tid) {
     threads.emplace_back([&, tid]() {
@@ -164,8 +166,8 @@ void TestMultiThreadPrediction(
   }
 }
 
-void TestPrediction(AnalysisConfig config,
-                    const std::vector<std::vector<PaddleTensor>> inputs,
+void TestPrediction(const AnalysisConfig &config,
+                    const std::vector<std::vector<PaddleTensor>> &inputs,
                     std::vector<PaddleTensor> *outputs, int num_threads,
                     bool use_analysis = FLAGS_use_analysis) {
   LOG(INFO) << "use_analysis: " << use_analysis;
@@ -178,8 +180,8 @@ void TestPrediction(AnalysisConfig config,
 }
 
 void CompareNativeAndAnalysis(
-    AnalysisConfig config,
-    const std::vector<std::vector<PaddleTensor>> inputs) {
+    const AnalysisConfig &config,
+    const std::vector<std::vector<PaddleTensor>> &inputs) {
   std::vector<PaddleTensor> native_outputs, analysis_outputs;
   TestOneThreadPrediction(config, inputs, &native_outputs, false);
   TestOneThreadPrediction(config, inputs, &analysis_outputs, true);
diff --git a/paddle/fluid/operators/adam_op.h b/paddle/fluid/operators/adam_op.h
index 5b27068c9e..4cb1f3a80e 100644
--- a/paddle/fluid/operators/adam_op.h
+++ b/paddle/fluid/operators/adam_op.h
@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 #include <math.h>  // for sqrt in CPU and CUDA
 #include <Eigen/Dense>
+#include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
 #include "paddle/fluid/operators/math/selected_rows_functor.h"
@@ -306,26 +307,43 @@ class AdamOpKernel : public framework::OpKernel<T> {
         VLOG(3) << "grad row size is 0!!";
         return;
       }
-      // merge duplicated rows if any.
-      // The rows of grad_merge have been sorted inside MergeAdd functor
-      scatter::MergeAdd<DeviceContext, T> merge_func;
-      auto& grad_merge = *(ctx.scope()
-                               .NewScope()
-                               .Var("sparse_adam_grad_merge")
-                               ->GetMutable<framework::SelectedRows>());
-      merge_func(ctx.template device_context<DeviceContext>(), grad,
-                 &grad_merge);
+
+      std::vector<int64_t> cpu_rows(grad.rows().begin(), grad.rows().end());
+      bool is_strict_sorted = true;
+      for (size_t i = 1; i < cpu_rows.size(); ++i) {
+        if (cpu_rows[i - 1] >= cpu_rows[i]) {
+          is_strict_sorted = false;
+          break;
+        }
+      }
+
+      const framework::SelectedRows* grad_merge_ptr;
+      if (is_strict_sorted) {
+        grad_merge_ptr = &grad;
+      } else {
+        // merge duplicated rows if any.
+        // The rows of grad_merge have been sorted inside MergeAdd functor
+        scatter::MergeAdd<DeviceContext, T> merge_func;
+        auto* grad_merge_var = const_cast<framework::Scope&>(ctx.scope())
+                                   .Var()
+                                   ->GetMutable<framework::SelectedRows>();
+        merge_func(ctx.template device_context<DeviceContext>(), grad,
+                   grad_merge_var);
+        grad_merge_ptr = grad_merge_var;
+      }
+
+      auto& grad_merge = *grad_merge_ptr;
       auto& grad_tensor = grad_merge.value();
       const T* grad_data = grad_tensor.template data<T>();
-      int64_t* rows = nullptr;
-// When compiled without CUDA, the CUDAMutableData() interface should not be
+      const int64_t* rows = nullptr;
+// When compiled without CUDA, the CUDAData() interface should not be
 // provided.
 #if defined(PADDLE_WITH_CUDA)
       if (platform::is_gpu_place(ctx.GetPlace())) {
-        rows = grad_merge.mutable_rows()->CUDAMutableData(ctx.GetPlace());
+        rows = grad_merge.rows().CUDAData(ctx.GetPlace());
       } else {
 #endif
-        rows = grad_merge.mutable_rows()->data();
+        rows = grad_merge.rows().data();
 
 #if defined(PADDLE_WITH_CUDA)
       }
diff --git a/paddle/fluid/operators/concat_op.cc b/paddle/fluid/operators/concat_op.cc
index bc58612f9d..57817da71a 100644
--- a/paddle/fluid/operators/concat_op.cc
+++ b/paddle/fluid/operators/concat_op.cc
@@ -94,8 +94,20 @@ class ConcatOpGrad : public framework::OperatorWithKernel {
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
   void InferShape(framework::InferShapeContext *ctx) const override {
-    ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X"));
-    ctx->ShareLoD("X", framework::GradVarName("X"));
+    auto in_x = "X";
+    auto out_x_g_n = framework::GradVarName(in_x);
+    ctx->SetOutputsDim(out_x_g_n, ctx->GetInputsDim(in_x));
+    auto &in_names = ctx->Inputs(in_x);
+    auto &out_names = ctx->Outputs(out_x_g_n);
+    PADDLE_ENFORCE_EQ(
+        in_names.size(), out_names.size(),
+        "The number of arguments in %s[%d] and %s[%d] is not equal.", in_x,
+        in_names.size(), out_x_g_n, out_names.size());
+    for (size_t i = 0; i < in_names.size(); ++i) {
+      if (out_names[i] != framework::kEmptyVarName) {
+        ctx->ShareLoD(in_x, out_x_g_n, i, i);
+      }
+    }
   }
 };
 
diff --git a/paddle/fluid/operators/scale_op.cc b/paddle/fluid/operators/scale_op.cc
index 13be6c65be..bf4df4f600 100644
--- a/paddle/fluid/operators/scale_op.cc
+++ b/paddle/fluid/operators/scale_op.cc
@@ -46,9 +46,15 @@ class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(
 **Scale operator**
 
-Multiply the input tensor with a float scalar to scale the input tensor.
+Apply scaling and bias addition to the input tensor.
 
-$$Out = scale*X$$
+if bias_after_scale=True:
+
+$$Out = scale*X + bias$$
+
+else:
+
+$$Out = scale*(X + bias)$$
 )DOC");
     AddAttr<float>("scale", "The scaling factor of the scale operator.")
         .SetDefault(1.0);
diff --git a/python/paddle/fluid/clip.py b/python/paddle/fluid/clip.py
index 79904cec93..32b8f1189f 100644
--- a/python/paddle/fluid/clip.py
+++ b/python/paddle/fluid/clip.py
@@ -280,7 +280,7 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
         group_scale_name = self.group_name + "_scale"
         if group_scale_name not in self.context:
             group_norm_var = layers.sums(input=self.context[self.group_name])
-            layers.sqrt(x=group_norm_var, out=group_norm_var)
+            group_norm_var = layers.sqrt(x=group_norm_var)
             clip_var = self.context[self.group_name + "_clip"]
             group_scale_var = layers.elementwise_div(
                 x=clip_var,
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index d7e5e47048..1d3c942290 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -489,7 +489,8 @@ class OpProtoHolder(object):
     def generated_op_attr_names():
         return {
             core.op_proto_and_checker_maker.kOpRoleAttrName(),
-            core.op_proto_and_checker_maker.kOpRoleVarAttrName()
+            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
+            core.op_proto_and_checker_maker.kOpNameScopeAttrName()
         }
 
 
diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py
index 8963d74de0..8c11921d9b 100644
--- a/python/paddle/fluid/layers/layer_function_generator.py
+++ b/python/paddle/fluid/layers/layer_function_generator.py
@@ -23,7 +23,10 @@ from ..proto import framework_pb2
 from ..framework import OpProtoHolder, Variable
 from ..layer_helper import LayerHelper
 
-__all__ = ['deprecated', 'generate_layer_fn', 'autodoc', 'templatedoc']
+__all__ = [
+    'deprecated', 'generate_layer_fn', 'generate_layer_fn_noattr', 'autodoc',
+    'templatedoc'
+]
 
 
 def _convert_(name):
@@ -58,7 +61,7 @@ def escape_math(text):
                                     _two_dollar_pattern_.sub(r"!!\1!!", text)))
 
 
-def _generate_doc_string_(op_proto):
+def _generate_doc_string_(op_proto, additional_args_lines=None):
     """
     Generate docstring by OpProto
 
@@ -98,6 +101,13 @@ def _generate_doc_string_(op_proto):
         buf.write(escape_math(each_attr.comment))
         buf.write('\n')
 
+    if additional_args_lines is not None:
+        for line in additional_args_lines:
+            line = line.strip()
+            buf.write('    ')
+            buf.write(line)
+            buf.write('\n')
+
     if len(op_proto.outputs) != 0:
         buf.write('\nReturns:\n')
         buf.write('    ')
@@ -205,6 +215,29 @@ def generate_layer_fn(op_type):
     return func
 
 
+def generate_layer_fn_noattr(op_type):
+    """Register the Python layer for an Operator without Attribute.
+
+    Args:
+       op_type: The name of the operator to be created.
+
+    This function takes in the operator type (sigmoid, exp , tanh etc) and
+    creates the operator functionality.
+
+    """
+    op_proto = OpProtoHolder.instance().get_op_proto(op_type)
+
+    def func(x, name=None):
+        helper = LayerHelper(op_type, **locals())
+        output = helper.create_tmp_variable(dtype=x.dtype)
+        helper.append_op(type=op_type, inputs={"X": x}, outputs={"Out": output})
+        return output
+
+    func.__name__ = op_type
+    func.__doc__ = _generate_doc_string_(op_proto)
+    return func
+
+
 def deprecated(func_or_class):
     """
     Deprecated warning decorator. It will result a warning message.
diff --git a/python/paddle/fluid/layers/learning_rate_scheduler.py b/python/paddle/fluid/layers/learning_rate_scheduler.py
index 2b947ca9e8..dfd801a098 100644
--- a/python/paddle/fluid/layers/learning_rate_scheduler.py
+++ b/python/paddle/fluid/layers/learning_rate_scheduler.py
@@ -68,7 +68,7 @@ def noam_decay(d_model, warmup_steps):
 
         a = global_step**-0.5
         b = (warmup_steps**-1.5) * global_step
-        lr_value = (d_model**-0.5) * ops.elementwise_min(a, b)
+        lr_value = (d_model**-0.5) * nn.elementwise_min(a, b)
 
     return lr_value
 
@@ -241,7 +241,7 @@ def polynomial_decay(learning_rate,
         else:
             decay_steps_var = tensor.fill_constant(
                 shape=[1], dtype='float32', value=float(decay_steps))
-            global_step = ops.elementwise_min(x=global_step, y=decay_steps_var)
+            global_step = nn.elementwise_min(x=global_step, y=decay_steps_var)
 
         decayed_lr = (learning_rate - end_learning_rate) * \
             ((1 - global_step / decay_steps) ** power) + end_learning_rate
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index f896cfa04b..6e0f3de414 100644
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -20,9 +20,9 @@ from __future__ import print_function
 import numpy as np
 from ..layer_helper import LayerHelper
 from ..initializer import Normal, Constant
-from ..framework import Variable
+from ..framework import Variable, OpProtoHolder
 from ..param_attr import ParamAttr
-from .layer_function_generator import autodoc, templatedoc
+from .layer_function_generator import autodoc, templatedoc, _generate_doc_string_
 from .tensor import concat
 from . import utils
 from .. import unique_name
@@ -107,7 +107,16 @@ __all__ = [
     'log',
     'crop',
     'rank_loss',
+    'elu',
+    'relu6',
+    'pow',
+    'stanh',
+    'hard_sigmoid',
+    'swish',
     'prelu',
+    'brelu',
+    'leaky_relu',
+    'soft_relu',
     'flatten',
     'sequence_mask',
     'stack',
@@ -116,6 +125,14 @@ __all__ = [
     'sequence_enumerate',
     'expand',
     'sequence_concat',
+    'scale',
+    'elementwise_add',
+    'elementwise_div',
+    'elementwise_sub',
+    'elementwise_mul',
+    'elementwise_max',
+    'elementwise_min',
+    'elementwise_pow',
 ]
 
 
@@ -3605,7 +3622,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
         attrs={
             'transpose_X': transpose_x,
             'transpose_Y': transpose_y,
-            'alpha': alpha,
+            'alpha': float(alpha),
         })
     return out
 
@@ -5895,6 +5912,148 @@ def pad2d(input,
     return out
 
 
+@templatedoc()
+def elu(x, alpha=1.0, name=None):
+    """
+    ${comment}
+    Args:
+        x(${x_type}): ${x_comment}
+        alpha(${alpha_type}|1.0): ${alpha_comment}
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+
+    Returns:
+        output(${out_type}): ${out_comment}
+    """
+    helper = LayerHelper('elu', **locals())
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='elu',
+        inputs={'X': x},
+        outputs={'Out': out},
+        attrs={'alpha': alpha})
+    return out
+
+
+@templatedoc()
+def relu6(x, threshold=6.0, name=None):
+    """
+    ${comment}
+    Args:
+        x(${x_type}): ${x_comment}
+        threshold(${threshold_type}|6.0): ${threshold_comment}
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+
+    Returns:
+        output(${out_type}): ${out_comment}
+    """
+    helper = LayerHelper('relu6', **locals())
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='relu6',
+        inputs={'X': x},
+        outputs={'Out': out},
+        attrs={'threshold': threshold})
+    return out
+
+
+@templatedoc()
+def pow(x, factor=1.0, name=None):
+    """
+    ${comment}
+    Args:
+        x(${x_type}): ${x_comment}
+        factor(${factor_type}|1.0): ${factor_comment}
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+
+    Returns:
+        output(${out_type}): ${out_comment}
+    """
+    helper = LayerHelper('pow', **locals())
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='pow',
+        inputs={'X': x},
+        outputs={'Out': out},
+        attrs={'factor': factor})
+    return out
+
+
+@templatedoc()
+def stanh(x, scale_a=2.0 / 3.0, scale_b=1.7159, name=None):
+    """
+    ${comment}
+    Args:
+        x(${x_type}): ${x_comment}
+        scale_a(${scale_a_type}|2.0 / 3.0): ${scale_a_comment}
+        scale_b(${scale_b_type}|1.7159): ${scale_b_comment}
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+
+    Returns:
+        output(${out_type}): ${out_comment}
+    """
+    helper = LayerHelper('stanh', **locals())
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='stanh',
+        inputs={'X': x},
+        outputs={'Out': out},
+        attrs={'scale_a': scale_a,
+               'scale_b': scale_b})
+    return out
+
+
+@templatedoc()
+def hard_sigmoid(x, slope=0.2, offset=0.5, name=None):
+    """
+    ${comment}
+    Args:
+        x(${x_type}): ${x_comment}
+        slope(${slope_type}|0.2): ${slope_comment}
+        offset(${offset_type}|0.5): ${offset_comment}
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+
+    Returns:
+        output(${out_type}): ${out_comment}
+    """
+    helper = LayerHelper('hard_sigmoid', **locals())
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='hard_sigmoid',
+        inputs={'X': x},
+        outputs={'Out': out},
+        attrs={'slope': slope,
+               'offset': offset})
+    return out
+
+
+@templatedoc()
+def swish(x, beta=1.0, name=None):
+    """
+    ${comment}
+    Args:
+        x(${x_type}): ${x_comment}
+        beta(${beta_type}|1.0): ${beta_comment}
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+
+    Returns:
+        output(${out_type}): ${out_comment}
+    """
+    helper = LayerHelper('swish', **locals())
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='swish',
+        inputs={'X': x},
+        outputs={'Out': out},
+        attrs={'slope': beta})
+    return out
+
+
 def prelu(x, mode, param_attr=None, name=None):
     """
     Equation:
@@ -5948,6 +6107,74 @@ def prelu(x, mode, param_attr=None, name=None):
     return out
 
 
+@templatedoc()
+def brelu(x, t_min=0.0, t_max=24.0, name=None):
+    """
+    ${comment}
+    Args:
+        x(${x_type}): ${x_comment}
+        t_min(${t_min_type}|0.0): ${t_min_comment}
+        t_max(${t_max_type}|24.0): ${t_max_comment}
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+     Returns:
+        output(${out_type}): ${out_comment}
+    """
+    helper = LayerHelper('brelu', **locals())
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='brelu',
+        inputs={'X': x},
+        outputs={'Out': out},
+        attrs={'t_min': t_min,
+               't_max': t_max})
+    return out
+
+
+@templatedoc()
+def leaky_relu(x, alpha=0.02, name=None):
+    """
+    ${comment}
+    Args:
+        x(${x_type}): ${x_comment}
+        alpha(${alpha_type}|0.02): ${alpha_comment}
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+     Returns:
+        output(${out_type}): ${out_comment}
+    """
+    helper = LayerHelper('leaky_relu', **locals())
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='leaky_relu',
+        inputs={'X': x},
+        outputs={'Out': out},
+        attrs={'alpha': alpha})
+    return out
+
+
+@templatedoc()
+def soft_relu(x, threshold=40.0, name=None):
+    """
+    ${comment}
+    Args:
+        x(${x_type}): ${x_comment}
+        threshold(${threshold_type}|40.0): ${threshold_comment}
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+     Returns:
+        output(${out_type}): ${out_comment}
+    """
+    helper = LayerHelper('soft_relu', **locals())
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='soft_relu',
+        inputs={'X': x},
+        outputs={'Out': out},
+        attrs={'threshold': threshold})
+    return out
+
+
 def flatten(x, axis=1, name=None):
     """
     **Flatten layer**
@@ -6234,3 +6461,105 @@ def expand(x, expand_times, name=None):
         outputs={'Out': out},
         attrs={'expand_times': expand_times})
     return out
+
+
+def _elementwise_op(helper):
+    op_type = helper.layer_type
+    x = helper.kwargs.get('x', None)
+    y = helper.kwargs.get('y', None)
+    assert x is not None, 'x cannot be None in {}'.format(op_type)
+    assert y is not None, 'y cannot be None in {}'.format(op_type)
+    axis = helper.kwargs.get('axis', -1)
+    use_mkldnn = helper.kwargs.get('use_mkldnn', False)
+    name = helper.kwargs.get('name', None)
+    if name is None:
+        out = helper.create_tmp_variable(dtype=x.dtype)
+    else:
+        out = helper.create_variable(
+            name=name, dtype=x.dtype, persistable=False)
+
+    helper.append_op(
+        type=op_type,
+        inputs={'X': x,
+                'Y': y},
+        outputs={'Out': out},
+        attrs={'axis': axis,
+               'use_mkldnn': use_mkldnn})
+    return helper.append_activation(out)
+
+
+@templatedoc()
+def scale(x, scale=1.0, bias=0.0, bias_after_scale=True, act=None, name=None):
+    """
+    ${comment}
+
+    Args:
+        x(${x_type}): ${x_comment}
+        scale(${scale_type}): ${scale_comment}
+        bias(${bias_type}): ${bias_comment}
+        bias_after_scale(${bias_after_scale_type}): ${bias_after_scale_comment}
+        act(basestring|None): Activation applied to the output.
+        name(basestring|None): Name of the output. 
+
+    Returns:
+        out(${out_type}): ${out_comment}
+    """
+
+    helper = LayerHelper('scale', **locals())
+    if name is None:
+        out = helper.create_tmp_variable(dtype=x.dtype)
+    else:
+        out = helper.create_variable(
+            name=name, dtype=x.dtype, persistable=False)
+
+    helper.append_op(
+        type='scale',
+        inputs={'X': x},
+        outputs={'Out': out},
+        attrs={
+            'scale': float(scale),
+            'bias': float(bias),
+            'bias_after_scale': bias_after_scale
+        })
+    return helper.append_activation(out)
+
+
+def elementwise_add(x, y, axis=-1, use_mkldnn=False, act=None, name=None):
+    return _elementwise_op(LayerHelper('elementwise_add', **locals()))
+
+
+def elementwise_div(x, y, axis=-1, use_mkldnn=False, act=None, name=None):
+    return _elementwise_op(LayerHelper('elementwise_div', **locals()))
+
+
+def elementwise_sub(x, y, axis=-1, use_mkldnn=False, act=None, name=None):
+    return _elementwise_op(LayerHelper('elementwise_sub', **locals()))
+
+
+def elementwise_mul(x, y, axis=-1, use_mkldnn=False, act=None, name=None):
+    return _elementwise_op(LayerHelper('elementwise_mul', **locals()))
+
+
+def elementwise_max(x, y, axis=-1, use_mkldnn=False, act=None, name=None):
+    return _elementwise_op(LayerHelper('elementwise_max', **locals()))
+
+
+def elementwise_min(x, y, axis=-1, use_mkldnn=False, act=None, name=None):
+    return _elementwise_op(LayerHelper('elementwise_min', **locals()))
+
+
+def elementwise_pow(x, y, axis=-1, use_mkldnn=False, act=None, name=None):
+    return _elementwise_op(LayerHelper('elementwise_pow', **locals()))
+
+
+for func in [
+        elementwise_add, elementwise_div, elementwise_sub, elementwise_mul,
+        elementwise_max, elementwise_min, elementwise_pow
+]:
+    op_proto = OpProtoHolder.instance().get_op_proto(func.__name__)
+    func.__doc__ = _generate_doc_string_(
+        op_proto,
+        additional_args_lines=[
+            "act (basestring|None): Activation applied to the output.",
+            "name (basestring|None): Name of the output."
+        ])
diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py
index 129252653d..48d92c342d 100644
--- a/python/paddle/fluid/layers/ops.py
+++ b/python/paddle/fluid/layers/ops.py
@@ -13,15 +13,14 @@
 # limitations under the License.
 
 from __future__ import print_function
-from .layer_function_generator import generate_layer_fn
+from .layer_function_generator import generate_layer_fn, generate_layer_fn_noattr
 
-__activations__ = [
+__activations_noattr__ = [
     'sigmoid',
     'logsigmoid',
     'exp',
     'tanh',
     'tanh_shrink',
-    'softshrink',
     'sqrt',
     'abs',
     'ceil',
@@ -33,29 +32,12 @@ __activations__ = [
     'square',
     'softplus',
     'softsign',
-    'brelu',
-    'leaky_relu',
-    'soft_relu',
-    'elu',
-    'relu6',
-    'pow',
-    'stanh',
-    'hard_sigmoid',
-    'swish',
 ]
 
 __all__ = [
     'mean',
     'mul',
-    'scale',
     'sigmoid_cross_entropy_with_logits',
-    'elementwise_add',
-    'elementwise_div',
-    'elementwise_sub',
-    'elementwise_mul',
-    'elementwise_max',
-    'elementwise_min',
-    'elementwise_pow',
     'clip',
     'clip_by_norm',
     'logical_and',
@@ -70,11 +52,22 @@ __all__ = [
     'slice',
     'shape',
     'maxout',
-] + __activations__
+    'softshrink',
+]
 
 for _OP in set(__all__):
     globals()[_OP] = generate_layer_fn(_OP)
 
+# It is a hot fix in some unittest using:
+#   fluid.layers.scale(x=x, scale=10.0, out=out_var)
+# e.g.: test_program_code.py, test_dist_train.py
+globals()['_scale'] = generate_layer_fn('scale')
+
+__all__ += __activations_noattr__
+
+for _OP in set(__activations_noattr__):
+    globals()[_OP] = generate_layer_fn_noattr(_OP)
+
 __all__ += ["uniform_random"]
 
 _uniform_random_ = generate_layer_fn('uniform_random')
diff --git a/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py b/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py
index 7c3ed09168..c2b089694e 100644
--- a/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py
+++ b/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py
@@ -25,13 +25,14 @@ class TestDistSeResneXt2x2(TestDistBase):
         self.check_with_place("dist_se_resnext.py", delta=1e-7)
 
 
-class TestDistseResnXt2x2WithMemopt(TestDistBase):
-    def _setup_config(self):
-        self._sync_mode = True
-        self._mem_opt = True
-
-    def test_dist_train(self):
-        self.check_with_place("dist_se_resnext.py", delta=1e-7)
+# TODO(typhoonzero): fix this test
+# class TestDistseResnXt2x2WithMemopt(TestDistBase):
+#     def _setup_config(self):
+#         self._sync_mode = True
+#         self._mem_opt = True
+
+#     def test_dist_train(self):
+#         self.check_with_place("dist_se_resnext.py", delta=1e-7)
 
 
 class TestDistSeResneXt2x2Async(TestDistBase):
diff --git a/python/paddle/fluid/tests/unittests/test_dist_train.py b/python/paddle/fluid/tests/unittests/test_dist_train.py
index 083525ccf5..d0875d9ea4 100644
--- a/python/paddle/fluid/tests/unittests/test_dist_train.py
+++ b/python/paddle/fluid/tests/unittests/test_dist_train.py
@@ -27,6 +27,7 @@ import paddle.fluid.layers as layers
 from paddle.fluid.layers.io import ListenAndServ
 from paddle.fluid.layers.io import Recv
 from paddle.fluid.layers.io import Send
+import paddle.fluid.layers.ops as ops
 
 from paddle.fluid import core
 
@@ -89,7 +90,7 @@ class TestSendOp(unittest.TestCase):
                     name="X",
                     append_batch_size=False)
                 fluid.initializer.Constant(value=1.0)(x, main.global_block())
-                layers.scale(x=x, scale=10.0, out=out_var)
+                ops._scale(x=x, scale=10.0, out=out_var)
 
         self.server_exe = fluid.Executor(place)
         self.server_exe.run(main)
diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py
index 6855a0e2c0..1fe7016924 100644
--- a/python/paddle/fluid/tests/unittests/test_layers.py
+++ b/python/paddle/fluid/tests/unittests/test_layers.py
@@ -573,6 +573,158 @@ class TestBook(unittest.TestCase):
             self.assertIsNotNone(out)
         print(str(program))
 
+    def test_brelu(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.brelu(input, t_min=1.0, t_max=20.0, name='brelu')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_leaky_relu(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.leaky_relu(input, alpha=0.1, name='leaky_relu')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_soft_relu(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.soft_relu(input, threshold=30.0, name='soft_relu')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_sigmoid(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.sigmoid(input, name='sigmoid')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_logsigmoid(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.logsigmoid(input, name='logsigmoid')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_exp(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.exp(input, name='exp')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_tanh(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.tanh(input, name='tanh')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_tanh_shrink(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.tanh_shrink(input, name='tanh_shrink')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_sqrt(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.sqrt(input, name='sqrt')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_abs(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.abs(input, name='abs')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_ceil(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.ceil(input, name='ceil')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_floor(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.floor(input, name='floor')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_cos(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.cos(input, name='cos')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_sin(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.sin(input, name='sin')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_round(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.round(input, name='round')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_reciprocal(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.reciprocal(input, name='reciprocal')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_square(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.square(input, name='square')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_softplus(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.softplus(input, name='softplus')
+            self.assertIsNotNone(out)
+        print(str(program))
+
+    def test_softsign(self):
+        program = Program()
+        with program_guard(program):
+            input = layers.data(name="input", shape=[16], dtype="float32")
+            out = layers.softsign(input, name='softsign')
+            self.assertIsNotNone(out)
+        print(str(program))
+
     def test_roi_perspective_transform(self):
         program = Program()
         with program_guard(program):
diff --git a/python/paddle/fluid/tests/unittests/test_program_code.py b/python/paddle/fluid/tests/unittests/test_program_code.py
index e9c2b92861..27b22ba939 100644
--- a/python/paddle/fluid/tests/unittests/test_program_code.py
+++ b/python/paddle/fluid/tests/unittests/test_program_code.py
@@ -25,6 +25,7 @@ import paddle.fluid.layers as layers
 from paddle.fluid.layers.io import ListenAndServ
 from paddle.fluid.layers.io import Recv
 from paddle.fluid.layers.io import Send
+import paddle.fluid.layers.ops as ops
 
 from paddle.fluid.transpiler.details import program_to_code
 
@@ -52,7 +53,7 @@ class TestProgram2Code(unittest.TestCase):
                     name="X",
                     append_batch_size=False)
                 fluid.initializer.Constant(value=1.0)(x, main.global_block())
-                layers.scale(x=x, scale=10.0, out=out_var)
+                ops._scale(x=x, scale=10.0, out=out_var)
 
         program_to_code(main)