diff --git a/.gitignore b/.gitignore
index 90138f996cf9cacc3c1cbff0cf2600eefca3f305..fa0c8882606b76ac71b43dcde7e1df6770c46c31 100644
--- a/.gitignore
+++ b/.gitignore
@@ -28,3 +28,4 @@ third_party/
 build_*
 # clion workspace.
 cmake-build-*
+model_test
diff --git a/Dockerfile b/Dockerfile
index 738bba9bc2e1ab19709722fe04f1490b1b13bd4f..c8b9eed6d60e5d3b32fc14c0c7af80a785145d1b 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -75,14 +75,14 @@ RUN pip3 install -U wheel && \
     pip3 install -U docopt PyYAML sphinx==1.5.6 && \
     pip3 install sphinx-rtd-theme==0.1.9 recommonmark && \
     easy_install -U pip && \
-    pip install -U wheel && \
+    pip install -U pip setuptools wheel && \
     pip install -U docopt PyYAML sphinx==1.5.6 && \
     pip install sphinx-rtd-theme==0.1.9 recommonmark
 
-RUN pip3 install pre-commit 'ipython==5.3.0' && \
+RUN pip3 install 'pre-commit==1.10.4' 'ipython==5.3.0' && \
     pip3 install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
     pip3 install opencv-python && \
-    pip install pre-commit 'ipython==5.3.0' && \
+    pip install 'pre-commit==1.10.4' 'ipython==5.3.0' && \
     pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
     pip install opencv-python
 
diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec
index 3ac9fe31b4f3a4e23b181a6005ec00077f562e38..d317117bcf95fd68ae2c6136ebd471d2e80017de 100644
--- a/paddle/fluid/API.spec
+++ b/paddle/fluid/API.spec
@@ -86,7 +86,7 @@ paddle.fluid.layers.reduce_prod ArgSpec(args=['input', 'dim', 'keep_dim', 'name'
 paddle.fluid.layers.sequence_first_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.sequence_last_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.sequence_slice ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name'], varargs=None, keywords=None, defaults=(False, None, None))
+paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name', 'dropout_implementation'], varargs=None, keywords=None, defaults=(False, None, None, 'downgrade_in_infer'))
 paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.edit_distance ArgSpec(args=['input', 'label', 'normalized', 'ignored_tokens'], varargs=None, keywords=None, defaults=(True, None))
@@ -107,7 +107,7 @@ paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label',
 paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1))
-paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', 'inplace', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None))
+paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', 'inplace', 'name'], varargs=None, keywords=None, defaults=(None, None, False, None))
 paddle.fluid.layers.squeeze ArgSpec(args=['input', 'axes', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.unsqueeze ArgSpec(args=['input', 'axes', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lod_reset ArgSpec(args=['x', 'y', 'target_lod'], varargs=None, keywords=None, defaults=(None, None))
diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc
index 134fcee826715672a6e021e9bf694bb771ebb830..ebd1d644bcce0554904ee0931827ef43a6d52b11 100644
--- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc
+++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc
@@ -252,9 +252,9 @@ std::vector<ir::Node *> SortOpsAndDelayOptimizeOp(const ir::Graph &graph) {
   std::vector<ir::Node *> sorted_ret;
   for (size_t i = 0; i < ret.size(); ++i) {
     if (i < last_backward) {
-      if (boost::get<int>(ret[i]->Op()->GetAttr(
-              OpProtoAndCheckerMaker::OpRoleAttrName())) ==
-          static_cast<int>(OpRole::kOptimize)) {
+      if (static_cast<bool>(boost::get<int>(ret[i]->Op()->GetAttr(
+                                OpProtoAndCheckerMaker::OpRoleAttrName())) &
+                            static_cast<int>(OpRole::kOptimize))) {
         optimize_ops.push_back(ret[i]);
       } else {
         sorted_ret.push_back(ret[i]);
diff --git a/paddle/fluid/framework/mixed_vector.h b/paddle/fluid/framework/mixed_vector.h
index 77386f4f069489b6ff7b927a281bdc286ff816e0..e1aac6dc5a92fb616f00de5806f044b83c2f503f 100644
--- a/paddle/fluid/framework/mixed_vector.h
+++ b/paddle/fluid/framework/mixed_vector.h
@@ -542,6 +542,33 @@ class CPUVector : public std::vector<T, std::allocator<T>> {
     this->reserve(this->size() + size_t(end - begin));
     this->insert(this->end(), begin, end);
   }
+
+  const T *CUDAData(platform::Place place) const {
+    PADDLE_THROW(
+        "Vector::CUDAData() method is not supported in CPU-only version");
+  }
+
+  T *CUDAMutableData(platform::Place place) {
+    PADDLE_THROW(
+        "Vector::CUDAMutableData() method is not supported in CPU-only "
+        "version");
+  }
+
+  const T *Data(platform::Place place) const {
+    PADDLE_ENFORCE(
+        platform::is_cpu_place(place),
+        "Vector::Data() method is not supported when not in CPUPlace");
+    return this->data();
+  }
+
+  T *MutableData(platform::Place place) {
+    PADDLE_ENFORCE(
+        platform::is_cpu_place(place),
+        "Vector::MutableData() method is not supported when not in CPUPlace");
+    return this->data();
+  }
+
+  const void *Handle() const { return static_cast<const void *>(this); }
 };
 
 template <typename T>
diff --git a/paddle/fluid/framework/naive_executor.cc b/paddle/fluid/framework/naive_executor.cc
index 2840d503f1454271afb309efdd435225ab077dc0..7fb42feb95b4d54aec693228721c052f683f4d80 100644
--- a/paddle/fluid/framework/naive_executor.cc
+++ b/paddle/fluid/framework/naive_executor.cc
@@ -146,22 +146,5 @@ void NaiveExecutor::CleanFeedFetchOps() {
   ops_.swap(ops);
 }
 
-void NaiveExecutor::EnableMKLDNN(const ProgramDesc &program) {
-#ifdef PADDLE_WITH_MKLDNN
-  VLOG(3) << "use_mkldnn=True";
-  for (size_t block_id = 0; block_id < program.Size(); ++block_id) {
-    auto *block = const_cast<ProgramDesc &>(program).MutableBlock(block_id);
-    for (auto *op : block->AllOps()) {
-      if (op->HasAttr("use_mkldnn")) {
-        op->SetAttr("use_mkldnn", true);
-      }
-    }
-  }
-#else
-  LOG(WARNING)
-      << "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option";
-#endif
-}
-
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/fluid/framework/naive_executor.h b/paddle/fluid/framework/naive_executor.h
index 9374f3f4a35cc0f90e5b2d6e8b397784b8eae123..ddfa6e1f4d8b73f594fc381ab505797491cdd378 100644
--- a/paddle/fluid/framework/naive_executor.h
+++ b/paddle/fluid/framework/naive_executor.h
@@ -48,8 +48,6 @@ class NaiveExecutor {
 
   void CleanFeedFetchOps();
 
-  void EnableMKLDNN(const ProgramDesc& program);
-
  protected:
   void CreateVariables(const ProgramDesc& desc, Scope* scope, int block_id);
 
diff --git a/paddle/fluid/framework/op_proto_maker.cc b/paddle/fluid/framework/op_proto_maker.cc
index 152fc3361a733b906765f3206089c5252658c213..ca31303f77c4a30eb64c43404e214779ea78aeaf 100644
--- a/paddle/fluid/framework/op_proto_maker.cc
+++ b/paddle/fluid/framework/op_proto_maker.cc
@@ -71,6 +71,8 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
            static_cast<int>(OpRole::kLoss) | static_cast<int>(OpRole::kForward),
            static_cast<int>(OpRole::kLoss) |
                static_cast<int>(OpRole::kBackward),
+           static_cast<int>(OpRole::kOptimize) |
+               static_cast<int>(OpRole::kLRSched),
            static_cast<int>(OpRole::kNotSpecified)})
       .SetDefault(static_cast<int>(OpRole::kNotSpecified));
   AddAttr<std::vector<std::string>>(OpRoleVarAttrName(),
diff --git a/paddle/fluid/framework/op_proto_maker.h b/paddle/fluid/framework/op_proto_maker.h
index cd2471dc49503bb83245673ce543364f5f873995..5527783faad6f7ea6e0b9e776cbde3c743277f16 100644
--- a/paddle/fluid/framework/op_proto_maker.h
+++ b/paddle/fluid/framework/op_proto_maker.h
@@ -20,6 +20,9 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+//////////////////////////
+// Don't add more roles to make this too complicated!
+//////////////////////////
 enum class OpRole {
   kForward = 0x0000,
   kBackward = 0x0001,
diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc
index 3368ae2ee4cf65b85abf1fcd89dee14f43522e1f..7dad872dd04754d2866c361b74ab1236ff743da5 100644
--- a/paddle/fluid/framework/parallel_executor.cc
+++ b/paddle/fluid/framework/parallel_executor.cc
@@ -156,12 +156,6 @@ ParallelExecutor::ParallelExecutor(
                            params, member_->local_scopes_, member_->use_cuda_);
 #endif
 
-  // If the loss_var_name is given, the number of graph should be only one.
-  if (loss_var_name.size()) {
-    PADDLE_ENFORCE_EQ(ir::GraphNum(*graph), 1,
-                      "The number of graph should be only one");
-  }
-
   if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
     member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
         exec_strategy, member_->local_scopes_, places, std::move(graph)));
diff --git a/paddle/fluid/framework/threadpool.cc b/paddle/fluid/framework/threadpool.cc
index 18cdca3a658a6a89e6ab637a7f38825756acfea8..a588cb417aebe94bd4aeda02b1bc8ba07a04b960 100644
--- a/paddle/fluid/framework/threadpool.cc
+++ b/paddle/fluid/framework/threadpool.cc
@@ -25,7 +25,6 @@ DEFINE_int32(dist_threadpool_size, 0,
 
 namespace paddle {
 namespace framework {
-
 std::unique_ptr<ThreadPool> ThreadPool::threadpool_(nullptr);
 std::once_flag ThreadPool::init_flag_;
 
@@ -47,8 +46,7 @@ void ThreadPool::Init() {
   }
 }
 
-ThreadPool::ThreadPool(int num_threads)
-    : total_threads_(num_threads), idle_threads_(num_threads), running_(true) {
+ThreadPool::ThreadPool(int num_threads) : running_(true) {
   threads_.resize(num_threads);
   for (auto& thread : threads_) {
     // TODO(Yancey1989): binding the thread on the specify CPU number
@@ -59,6 +57,7 @@ ThreadPool::ThreadPool(int num_threads)
 ThreadPool::~ThreadPool() {
   {
     // notify all threads to stop running
+    std::lock_guard<std::mutex> l(mutex_);
     running_ = false;
     scheduled_.notify_all();
   }
@@ -69,36 +68,24 @@ ThreadPool::~ThreadPool() {
   }
 }
 
-void ThreadPool::Wait() {
-  std::unique_lock<std::mutex> lock(mutex_);
-  completed_.wait(lock, [=] { return Done() == true; });
-}
-
 void ThreadPool::TaskLoop() {
-  while (running_) {
+  while (true) {
     std::unique_lock<std::mutex> lock(mutex_);
-    scheduled_.wait(lock, [=] { return !tasks_.empty() || !running_; });
 
-    if (!running_) {
-      break;
+    scheduled_.wait(
+        lock, [this] { return !this->tasks_.empty() || !this->running_; });
+
+    if (!running_ || tasks_.empty()) {
+      return;
     }
+
     // pop a task from the task queue
     auto task = std::move(tasks_.front());
     tasks_.pop();
-
-    --idle_threads_;
     lock.unlock();
 
     // run the task
     task();
-
-    {
-      std::unique_lock<std::mutex> lock(mutex_);
-      ++idle_threads_;
-      if (Done()) {
-        completed_.notify_all();
-      }
-    }
   }
 }
 
diff --git a/paddle/fluid/framework/threadpool.h b/paddle/fluid/framework/threadpool.h
index 94111ee335b1a5df327b3e46d62069b4735c54f6..0687e628aaa4fb7b2e67938fa09a319c8bb35aff 100644
--- a/paddle/fluid/framework/threadpool.h
+++ b/paddle/fluid/framework/threadpool.h
@@ -57,15 +57,6 @@ class ThreadPool {
 
   ~ThreadPool();
 
-  // Returns the number of threads created by the constructor.
-  size_t Threads() const { return total_threads_; }
-
-  // Returns the number of currently idle threads.
-  size_t IdleThreads() {
-    std::unique_lock<std::mutex> lock(mutex_);
-    return idle_threads_;
-  }
-
   // Run pushes a function to the task queue and returns a std::future
   // object.  To wait for the completion of the task, call
   // std::future::wait().
@@ -94,25 +85,13 @@ class ThreadPool {
     });
     std::future<std::unique_ptr<platform::EnforceNotMet>> f = task.get_future();
     tasks_.push(std::move(task));
-    lock.unlock();
     scheduled_.notify_one();
     return f;
   }
 
-  // Wait until all the tasks are completed.
-  void Wait();
-
  private:
   DISABLE_COPY_AND_ASSIGN(ThreadPool);
 
-  // If the task queue is empty and avaialbe is equal to the number of
-  // threads, means that all tasks are completed.  Note: this function
-  // is not thread-safe.  Returns true if all tasks are completed.
-  // Note: don't delete the data member total_threads_ and use
-  // threads_.size() instead; because you'd need to lock the mutex
-  // before accessing threads_.
-  bool Done() { return tasks_.empty() && idle_threads_ == total_threads_; }
-
   // The constructor starts threads to run TaskLoop, which retrieves
   // and runs tasks from the queue.
   void TaskLoop();
@@ -125,14 +104,11 @@ class ThreadPool {
   static std::once_flag init_flag_;
 
   std::vector<std::unique_ptr<std::thread>> threads_;
-  const size_t total_threads_;
-  size_t idle_threads_;
 
   std::queue<Task> tasks_;
   std::mutex mutex_;
   bool running_;
   std::condition_variable scheduled_;
-  std::condition_variable completed_;
 };
 
 class ThreadPoolIO : ThreadPool {
diff --git a/paddle/fluid/framework/threadpool_test.cc b/paddle/fluid/framework/threadpool_test.cc
index 27a4ffd4fcbf293a3dea1744b29384d0bee0c137..884d61e23428a0ad758946295ca9c470767e93ef 100644
--- a/paddle/fluid/framework/threadpool_test.cc
+++ b/paddle/fluid/framework/threadpool_test.cc
@@ -19,10 +19,11 @@ limitations under the License. */
 
 namespace framework = paddle::framework;
 
-void do_sum(framework::ThreadPool* pool, std::atomic<int>* sum, int cnt) {
-  std::vector<std::future<void>> fs;
+void do_sum(std::vector<std::future<void>>* fs, std::mutex* mu,
+            std::atomic<int>* sum, int cnt) {
   for (int i = 0; i < cnt; ++i) {
-    fs.push_back(framework::Async([sum]() { sum->fetch_add(1); }));
+    std::lock_guard<std::mutex> l(*mu);
+    fs->push_back(framework::Async([sum]() { sum->fetch_add(1); }));
   }
 }
 
@@ -40,18 +41,21 @@ TEST(ThreadPool, ConcurrentInit) {
 }
 
 TEST(ThreadPool, ConcurrentRun) {
-  framework::ThreadPool* pool = framework::ThreadPool::GetInstance();
   std::atomic<int> sum(0);
   std::vector<std::thread> threads;
+  std::vector<std::future<void>> fs;
+  std::mutex fs_mu;
   int n = 50;
   // sum = (n * (n + 1)) / 2
   for (int i = 1; i <= n; ++i) {
-    std::thread t(do_sum, pool, &sum, i);
+    std::thread t(do_sum, &fs, &fs_mu, &sum, i);
     threads.push_back(std::move(t));
   }
   for (auto& t : threads) {
     t.join();
   }
-  pool->Wait();
+  for (auto& t : fs) {
+    t.wait();
+  }
   EXPECT_EQ(sum, ((n + 1) * n) / 2);
 }
diff --git a/paddle/fluid/operators/detection/generate_proposals_op.cc b/paddle/fluid/operators/detection/generate_proposals_op.cc
index a69d9c9a529f26b3981ca8d1ba226fda71b8820a..709c2dfc4b7c67d7d04074c58ce6da85b6e790fe 100644
--- a/paddle/fluid/operators/detection/generate_proposals_op.cc
+++ b/paddle/fluid/operators/detection/generate_proposals_op.cc
@@ -284,7 +284,7 @@ static inline Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox,
       selected_indices.push_back(idx);
       ++selected_num;
     }
-    sorted_indices.erase(sorted_indices.end());
+    sorted_indices.erase(sorted_indices.end() - 1);
     if (flag && eta < 1 && adaptive_threshold > 0.5) {
       adaptive_threshold *= eta;
     }
diff --git a/paddle/fluid/operators/dropout_op.cc b/paddle/fluid/operators/dropout_op.cc
index 07322e720f26213ea777be3cd22f2fead28507f0..3c28ef30922e6d6ba09b96282619eef15867631e 100644
--- a/paddle/fluid/operators/dropout_op.cc
+++ b/paddle/fluid/operators/dropout_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/dropout_op.h"
+#include <string>
 
 namespace paddle {
 namespace operators {
@@ -57,6 +58,29 @@ class DropoutOpMaker : public framework::OpProtoAndCheckerMaker {
                   "will be dropped.")
         .SetDefault(false);
     AddAttr<int>("seed", "Dropout random seed.").SetDefault(0);
+    AddAttr<std::string>(
+        "dropout_implementation",
+        "[\"downgrade_in_infer\"|\"upscale_in_train\"]"
+        "There are two kinds of ways to implement dropout"
+        "(the mask below is a tensor have the same shape with input"
+        "the value of mask is 0 or 1, the ratio of 0 is dropout_prob)"
+        "1. downgrade_in_infer(default), downgrade the outcome at inference "
+        "time"
+        "   train: out = input * mask"
+        "   inference: out = input * dropout_prob"
+        "2. upscale_in_train, upscale the outcome at training time, do nothing "
+        "in inference"
+        "   train: out = input * mask / ( 1.0 - dropout_prob )"
+        "   inference: out = input"
+        "   dropout op can be removed from the program. the program will be "
+        "efficient")
+        .SetDefault("downgrade_in_infer")
+        .AddCustomChecker([](const std::string& type) {
+          PADDLE_ENFORCE(
+              type == "downgrade_in_infer" || type == "upscale_in_train",
+              "dropout_implementation can only be downgrade_in_infer or "
+              "upscale_in_train");
+        });
 
     AddComment(R"DOC(
 Dropout Operator.
@@ -104,7 +128,9 @@ REGISTER_OPERATOR(dropout, ops::DropoutOp, ops::DropoutOpMaker,
                   paddle::framework::DefaultGradOpDescMaker<true>);
 REGISTER_OPERATOR(dropout_grad, ops::DropoutOpGrad);
 REGISTER_OP_CPU_KERNEL(
-    dropout, ops::CPUDropoutKernel<paddle::platform::CPUDeviceContext, float>);
+    dropout, ops::CPUDropoutKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::CPUDropoutKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     dropout_grad,
-    ops::DropoutGradKernel<paddle::platform::CPUDeviceContext, float>);
+    ops::DropoutGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::DropoutGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/fluid/operators/dropout_op.cu b/paddle/fluid/operators/dropout_op.cu
index 1dd66e0280c46c0624ff70e822cb6fa6f06b7aa9..e011f47e086183a4ef3a3373c17acd6c21b6cf7e 100644
--- a/paddle/fluid/operators/dropout_op.cu
+++ b/paddle/fluid/operators/dropout_op.cu
@@ -17,6 +17,7 @@ limitations under the License. */
 #include <thrust/iterator/counting_iterator.h>
 #include <thrust/random.h>
 #include <thrust/transform.h>
+#include <string>
 #include "paddle/fluid/operators/dropout_op.h"
 #include "paddle/fluid/platform/float16.h"
 
@@ -26,7 +27,8 @@ namespace operators {
 template <typename T>
 __global__ void RandomGenerator(const size_t n, const int seed,
                                 const float dropout_prob, const T* src,
-                                T* mask_data, T* dst) {
+                                T* mask_data, T* dst,
+                                bool is_upscale_in_train) {
   thrust::minstd_rand rng;
   rng.seed(seed);
   thrust::uniform_real_distribution<float> dist(0, 1);
@@ -47,7 +49,11 @@ __global__ void RandomGenerator(const size_t n, const int seed,
     if (dist(rng) < dropout_prob) {
       mask = static_cast<T>(0);
     } else {
-      mask = static_cast<T>(1);
+      if (is_upscale_in_train) {
+        mask = static_cast<T>(1.0f / (1.0f - dropout_prob));
+      } else {
+        mask = static_cast<T>(1);
+      }
     }
     dest = s * mask;
     mask_data[idx] = mask;
@@ -67,6 +73,8 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
     y->mutable_data<T>(context.GetPlace());
     float dropout_prob = context.Attr<float>("dropout_prob");
 
+    auto dropout_implementation =
+        context.Attr<std::string>("dropout_implementation");
     auto& place = *context.template device_context<Place>().eigen_device();
     if (!context.Attr<bool>("is_test")) {
       auto* mask = context.Output<Tensor>("Mask");
@@ -83,11 +91,16 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
       int grid = (x->numel() + threads - 1) / threads;
       RandomGenerator<
           T><<<grid, threads, 0, context.cuda_device_context().stream()>>>(
-          size, seed, dropout_prob, x_data, mask_data, y_data);
+          size, seed, dropout_prob, x_data, mask_data, y_data,
+          (dropout_implementation == "upscale_in_train"));
     } else {
       auto X = EigenMatrix<T>::Reshape(*x, 1);
       auto Y = EigenMatrix<T>::Reshape(*y, 1);
-      Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
+      if (dropout_implementation == "upscale_in_train") {
+        Y.device(place) = X;
+      } else {
+        Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
+      }
     }
   }
 };
@@ -99,6 +112,8 @@ namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 REGISTER_OP_CUDA_KERNEL(
     dropout, ops::GPUDropoutKernel<plat::CUDADeviceContext, float>,
-    ops::GPUDropoutKernel<plat::CUDADeviceContext, plat::float16>);
-REGISTER_OP_CUDA_KERNEL(dropout_grad,
-                        ops::DropoutGradKernel<plat::CUDADeviceContext, float>);
+    ops::GPUDropoutKernel<plat::CUDADeviceContext, plat::float16>,
+    ops::GPUDropoutKernel<plat::CUDADeviceContext, double>);
+REGISTER_OP_CUDA_KERNEL(
+    dropout_grad, ops::DropoutGradKernel<plat::CUDADeviceContext, float>,
+    ops::DropoutGradKernel<plat::CUDADeviceContext, double>);
diff --git a/paddle/fluid/operators/dropout_op.h b/paddle/fluid/operators/dropout_op.h
index 0628b4b826d2730a8e3fb4842e4ae550b8c00569..6c629b7b6d255828023ed25680675ca104a33e12 100644
--- a/paddle/fluid/operators/dropout_op.h
+++ b/paddle/fluid/operators/dropout_op.h
@@ -14,6 +14,7 @@ limitations under the License. */
 #pragma once
 
 #include <random>
+#include <string>
 
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
@@ -36,6 +37,8 @@ class CPUDropoutKernel : public framework::OpKernel<T> {
     auto* y_data = y->mutable_data<T>(context.GetPlace());
     float dropout_prob = context.Attr<float>("dropout_prob");
 
+    auto dropout_implementation =
+        context.Attr<std::string>("dropout_implementation");
     if (!context.Attr<bool>("is_test")) {
       auto* mask = context.Output<Tensor>("Mask");
       auto* mask_data = mask->mutable_data<T>(context.GetPlace());
@@ -49,14 +52,20 @@ class CPUDropoutKernel : public framework::OpKernel<T> {
       engine.seed(seed);
 
       std::uniform_real_distribution<float> dist(0, 1);
+
       size_t size = framework::product(mask->dims());
       for (size_t i = 0; i < size; ++i) {
         if (dist(engine) < dropout_prob) {
           mask_data[i] = 0;
           y_data[i] = 0;
         } else {
-          mask_data[i] = 1;
-          y_data[i] = x_data[i];
+          if (dropout_implementation == "upscale_in_train") {
+            mask_data[i] = 1.0f / static_cast<T>(1.0f - dropout_prob);
+            y_data[i] = x_data[i] / static_cast<T>(1.0f - dropout_prob);
+          } else {
+            mask_data[i] = 1;
+            y_data[i] = x_data[i];
+          }
         }
       }
     } else {
@@ -64,7 +73,11 @@ class CPUDropoutKernel : public framework::OpKernel<T> {
       auto Y = EigenMatrix<T>::Reshape(*y, 1);
       auto& place =
           *context.template device_context<DeviceContext>().eigen_device();
-      Y.device(place) = X * (1.0f - dropout_prob);
+      if (dropout_implementation == "upscale_in_train") {
+        Y.device(place) = X;
+      } else {
+        Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
+      }
     }
   }
 };
diff --git a/paddle/fluid/operators/math/jit_kernel_rnn.cc b/paddle/fluid/operators/math/jit_kernel_rnn.cc
index c0847f0bee415393255983468349ad314716bcb3..fab293f7d03eb923995fa4cd99af955a34faa6a4 100644
--- a/paddle/fluid/operators/math/jit_kernel_rnn.cc
+++ b/paddle/fluid/operators/math/jit_kernel_rnn.cc
@@ -136,6 +136,7 @@ static std::shared_ptr<const VActKernel<T>> GetActKernel(
   return nullptr;
 }
 
+#ifdef __AVX__
 template <jit::cpu_isa_t isa>
 static std::unique_ptr<AVXAct> GetAVXAct(const std::string& type) {
   if (type == "sigmoid") {
@@ -150,6 +151,7 @@ static std::unique_ptr<AVXAct> GetAVXAct(const std::string& type) {
   PADDLE_THROW("Not support type: %s", type);
   return nullptr;
 }
+#endif
 
 /* LSTM JitKernel */
 template <typename T, jit::cpu_isa_t isa, jit_block>
diff --git a/paddle/fluid/operators/softmax_cudnn_op.cu.cc b/paddle/fluid/operators/softmax_cudnn_op.cu.cc
index 2bdb23e999621b10799b5163f326bc4b66a437e6..f6e241af0634650f4a32be6a4547617f8ec3ee60 100644
--- a/paddle/fluid/operators/softmax_cudnn_op.cu.cc
+++ b/paddle/fluid/operators/softmax_cudnn_op.cu.cc
@@ -76,6 +76,8 @@ namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 REGISTER_OP_KERNEL(softmax, CUDNN, plat::CUDAPlace,
                    ops::SoftmaxCUDNNKernel<float>,
+                   ops::SoftmaxCUDNNKernel<double>,
                    ops::SoftmaxCUDNNKernel<plat::float16>);
 REGISTER_OP_KERNEL(softmax_grad, CUDNN, plat::CUDAPlace,
-                   ops::SoftmaxGradCUDNNKernel<float>);
+                   ops::SoftmaxGradCUDNNKernel<float>,
+                   ops::SoftmaxGradCUDNNKernel<double>);
diff --git a/paddle/fluid/operators/transpose_op.cc b/paddle/fluid/operators/transpose_op.cc
index 6a9fc6611a8f8eaa6749aefac0673ccabaebbcfe..bbd71db6062107f6ba40343c84d942b54b3958e6 100644
--- a/paddle/fluid/operators/transpose_op.cc
+++ b/paddle/fluid/operators/transpose_op.cc
@@ -210,18 +210,21 @@ REGISTER_OPERATOR(transpose, ops::TransposeOp, ops::TransposeOpMaker,
 REGISTER_OPERATOR(transpose_grad, ops::TransposeOpGrad);
 
 REGISTER_OP_CPU_KERNEL(
-    transpose, ops::TransposeKernel<paddle::platform::CPUDeviceContext, float>);
+    transpose, ops::TransposeKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::TransposeKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     transpose_grad,
-    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, float>);
+    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, double>);
 
 REGISTER_OPERATOR(transpose2, ops::Transpose2Op, ops::Transpose2OpMaker,
                   ops::Transpose2GradMaker);
 REGISTER_OPERATOR(transpose2_grad, ops::Transpose2OpGrad);
 
 REGISTER_OP_CPU_KERNEL(
-    transpose2,
-    ops::TransposeKernel<paddle::platform::CPUDeviceContext, float>);
+    transpose2, ops::TransposeKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::TransposeKernel<paddle::platform::CPUDeviceContext, double>);
 REGISTER_OP_CPU_KERNEL(
     transpose2_grad,
-    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, float>);
+    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/fluid/operators/transpose_op.cu.cc b/paddle/fluid/operators/transpose_op.cu.cc
index c1b5a8b31be243fab3af06a18c8e51986c953700..b4025350fa9f3610bde43eee91cd059f3063813f 100644
--- a/paddle/fluid/operators/transpose_op.cu.cc
+++ b/paddle/fluid/operators/transpose_op.cu.cc
@@ -16,15 +16,18 @@ limitations under the License. */
 
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
-    transpose,
-    ops::TransposeKernel<paddle::platform::CUDADeviceContext, float>);
+    transpose, ops::TransposeKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::TransposeKernel<paddle::platform::CUDADeviceContext, double>);
 REGISTER_OP_CUDA_KERNEL(
     transpose_grad,
-    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, float>);
+    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, double>);
 
 REGISTER_OP_CUDA_KERNEL(
     transpose2,
-    ops::TransposeKernel<paddle::platform::CUDADeviceContext, float>);
+    ops::TransposeKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::TransposeKernel<paddle::platform::CUDADeviceContext, double>);
 REGISTER_OP_CUDA_KERNEL(
     transpose2_grad,
-    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, float>);
+    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, double>);
diff --git a/python/paddle/fluid/clip.py b/python/paddle/fluid/clip.py
index 4c24d0d6a7069c75c7b9b8245f4567ae8bfc2742..1738afe93e99f1de28bec2fb23be8e1a309d9288 100644
--- a/python/paddle/fluid/clip.py
+++ b/python/paddle/fluid/clip.py
@@ -272,7 +272,7 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
                 )
 
         square = grad * grad
-        local_norm_var = layers.cast(layers.reduce_sum(input=square), 'float64')
+        local_norm_var = layers.reduce_sum(input=square)
         context[self.group_name].append(local_norm_var)
 
         self.context = context
@@ -282,7 +282,6 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
         if group_scale_name not in self.context:
             group_norm_var = layers.sums(input=self.context[self.group_name])
             group_norm_var = layers.sqrt(x=group_norm_var)
-            group_norm_var = layers.cast(group_norm_var, 'float32')
             clip_var = self.context[self.group_name + "_clip"]
             group_scale_var = layers.elementwise_div(
                 x=clip_var,
@@ -333,7 +332,8 @@ def append_gradient_clip_ops(param_grads):
     for p, g in param_grads:
         if g is None:
             continue
-        with p.block.program._optimized_guard([p, g]):
+        with p.block.program._optimized_guard(
+            [p, g]), framework.name_scope('append_clip'):
             clip_attr = getattr(p, 'gradient_clip_attr', NullGradientClipAttr())
             if clip_attr is None:
                 clip_attr = NullGradientClipAttr()
@@ -348,7 +348,8 @@ def append_gradient_clip_ops(param_grads):
     for p, g in param_grads:
         if g is None:
             continue
-        with p.block.program._optimized_guard([p, g]):
+        with p.block.program._optimized_guard(
+            [p, g]), framework.name_scope('append_graident_clip'):
             res.append(clip_attr._create_operators(param=p, grad=g))
 
     return res
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index b07d0131a32c3f2744854a17b180ae714d532f80..fd03dff386cad21c727ca0f266fa1b37ad65b4ad 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -1496,6 +1496,9 @@ class Program(object):
             >>> with program._optimized_guard([p,g]):
             >>>     p = p - 0.001 * g
         """
+        tmp_role = self._current_role
+        tmp_var = self._op_role_var
+
         OpRole = core.op_proto_and_checker_maker.OpRole
         self._current_role = OpRole.Optimize
         self._op_role_var = [
@@ -1503,11 +1506,11 @@ class Program(object):
             for var in param_and_grads
         ]
         yield
-        self._op_role_var = []
-        self._current_role = OpRole.Forward
+        self._op_role_var = tmp_var
+        self._current_role = tmp_role
 
     @contextlib.contextmanager
-    def _lr_schedule_guard(self):
+    def _lr_schedule_guard(self, is_with_opt=False):
         """
         A with guard to set :code:`LRSched` :code:`OpRole` and
         :code:`OpRoleVar` automatically. The :code:`OpRoleVar` is
@@ -1515,6 +1518,10 @@ class Program(object):
 
         Notes: This is a very low level API. Users should not use it directly.
 
+        Args:
+            is_with_opt: Only set to true if these ops a in the middle
+                 of a bunch of optimize ops so that it can be treated
+                 correctly. For example, sgd->lr_op->sgd->lr_op->sgd.
 
         Examples:
 
@@ -1528,6 +1535,8 @@ class Program(object):
 
         OpRole = core.op_proto_and_checker_maker.OpRole
         self._current_role = OpRole.LRSched
+        if is_with_opt:
+            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
         # TODO(typhoonzero): how to set target learning rate var
         self._op_role_var = []
         yield
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index d3b5f13b577099afb9545b5177fbf7e07f9a8e45..c762633c601950e06dbd29e20231e943054c27b3 100644
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -981,7 +981,12 @@ def cos_sim(X, Y):
     return out
 
 
-def dropout(x, dropout_prob, is_test=False, seed=None, name=None):
+def dropout(x,
+            dropout_prob,
+            is_test=False,
+            seed=None,
+            name=None,
+            dropout_implementation="downgrade_in_infer"):
     """
     Computes dropout.
 
@@ -1001,6 +1006,21 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None):
                     units will be dropped. DO NOT use a fixed seed in training.
         name (str|None): A name for this layer(optional). If set None, the layer
                          will be named automatically.
+        dropout_implementation(string): ['downgrade_in_infer'(defauld)|'upscale_in_train']
+                                        1. downgrade_in_infer(default), downgrade the outcome at inference
+                                           train: out = input * mask
+                                           inference: out = input * dropout_prob
+                                           (make is a tensor same shape with input, value is 0 or 1
+                                            ratio of 0 is dropout_prob)
+                                        2. upscale_in_train, upscale the outcome at training time
+                                           train: out = input * mask / ( 1.0 - dropout_prob )
+                                           inference: out = input
+                                           (make is a tensor same shape with input, value is 0 or 1
+                                            ratio of 0 is dropout_prob)
+                                           dropout op can be removed from the program. 
+                                           the program will be efficient
+                                        
+
 
     Returns:
         Variable: A tensor variable is the shape with `x`.
@@ -1030,7 +1050,8 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None):
             'dropout_prob': dropout_prob,
             'is_test': is_test,
             'fix_seed': seed is not None,
-            'seed': seed if seed is not None else 0
+            'seed': seed if seed is not None else 0,
+            'dropout_implementation': dropout_implementation,
         })
     return out
 
@@ -4845,7 +4866,7 @@ def autoincreased_step_counter(counter_name=None, begin=1, step=1):
     return counter
 
 
-def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
+def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None):
     """
     Gives a new shape to the input Tensor without changing its data.
 
@@ -4893,15 +4914,22 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
                                 :attr:`shape` specifying shape. That is to
                                 say :attr:`actual_shape` has a higher priority
                                 than :attr:`shape`.
-        act (str): The non-linear activation to be applied to output variable.
-        inplace(bool): If this flag is set true, the output
-                       shares data with input without copying, otherwise
-                       a new output tensor is created
-                       whose data is copied from input x.
+        act (str): The non-linear activation to be applied to the reshaped tensor
+                   variable.
+        inplace(bool): Must use :attr:`False` if :attr:`x` is used in multiple
+                       operators. If this flag is set :attr:`True`, reuse input
+                       :attr:`x` to reshape, which will change the shape of
+                       tensor variable :attr:`x` and might cause errors when
+                       :attr:`x` is used in multiple operators. If :attr:`False`,
+                       preserve the shape :attr:`x` and create a new output tensor
+                       variable whose data is copied from input x but reshaped.
         name (str): The name of this layer. It is optional.
 
     Returns:
-        Variable: The output tensor.
+        Variable: The reshaped tensor variable if :attr:`act` is None. It is a \
+                  new tensor variable if :attr:`inplace` is :attr:`False`, \
+                  otherwise it is :attr:`x`. If :attr:`act` is not None, return \
+                  the activated tensor variable.
 
     Raises:
         TypeError: if actual_shape is neither Variable nor None.
@@ -4912,7 +4940,7 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
             data = fluid.layers.data(
                 name='data', shape=[2, 4, 6], dtype='float32')
             reshaped = fluid.layers.reshape(
-                x=data, shape=[-1, 0, 3, 2], act='tanh', inplace=True)
+                x=data, shape=[-1, 0, 3, 2], inplace=True)
     """
 
     if not (isinstance(shape, list) or isinstance(shape, tuple)):
@@ -4939,7 +4967,8 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
                 "except one unknown dimension.")
 
     helper = LayerHelper("reshape2", **locals())
-    out = helper.create_variable_for_type_inference(dtype=x.dtype)
+    out = x if inplace else helper.create_variable_for_type_inference(
+        dtype=x.dtype)
     x_shape = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type="reshape2",
diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py
index 17af44afdde5cdbec082d473457ef01974695bc6..6ea280c733996436a6218cc6f759f2cf7c652ac9 100644
--- a/python/paddle/fluid/optimizer.py
+++ b/python/paddle/fluid/optimizer.py
@@ -111,7 +111,9 @@ class Optimizer(object):
             if param_lr == 1.0:
                 return self._global_learning_rate()
             else:
-                with default_main_program()._lr_schedule_guard():
+                with default_main_program()._lr_schedule_guard(
+                        is_with_opt=True), framework.name_scope(
+                            'scale_with_param_lr'):
                     return self._global_learning_rate() * param_lr
 
     def _create_accumulators(self, block, parameters):
@@ -602,7 +604,8 @@ class AdamOptimizer(Optimizer):
         for param, grad in param_and_grads:
             if grad is None:
                 continue
-            with param.block.program._optimized_guard([param, grad]):
+            with param.block.program._optimized_guard(
+                [param, grad]), name_scope("optimizer"):
                 beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                                       param)
                 beta2_pow_acc = self._get_accumulator(self._beta2_pow_acc_str,
@@ -740,7 +743,8 @@ class AdamaxOptimizer(Optimizer):
         for param, grad in parameters_and_grads:
             if grad is None:
                 continue
-            with param.block.program._optimized_guard([param, grad]):
+            with param.block.program._optimized_guard(
+                [param, grad]), name_scope('adamx'):
                 beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                                       param)
                 main_block.append_op(
@@ -1279,7 +1283,8 @@ class ModelAverage(Optimizer):
         for param, grad in self.params_grads:
             if grad is None:
                 continue
-            with param.block.program._optimized_guard([param, grad]):
+            with param.block.program._optimized_guard(
+                [param, grad]), name_scope('move_average'):
                 self._append_average_accumulate_op(param)
 
         self.apply_program = Program()
diff --git a/python/paddle/fluid/regularizer.py b/python/paddle/fluid/regularizer.py
index c151fbd17208bb6e3104e8d0f6590392c6095987..57185da4d1d38f3848994aae105411cf2844843a 100644
--- a/python/paddle/fluid/regularizer.py
+++ b/python/paddle/fluid/regularizer.py
@@ -47,7 +47,8 @@ def append_regularization_ops(parameters_and_grads, regularization=None):
         if grad is None:
             params_and_grads.append((param, grad))
             continue
-        with param.block.program._optimized_guard([param, grad]):
+        with param.block.program._optimized_guard(
+            [param, grad]), framework.name_scope('regularization'):
             regularization_term = None
             if param.regularizer is not None:
                 # Add variable for regularization term in grad block
diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt
index 68e498c6e82c6d2b389e2e08355b63db9de0bd14..cf54bc2dbe788f3757a7ef93f26156d118a0cd02 100644
--- a/python/paddle/fluid/tests/unittests/CMakeLists.txt
+++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt
@@ -78,9 +78,9 @@ if(WITH_DISTRIBUTE)
         set_tests_properties(test_dist_word2vec PROPERTIES TIMEOUT 200)
         py_test_modules(test_dist_se_resnext MODULES test_dist_se_resnext)
         set_tests_properties(test_dist_se_resnext PROPERTIES TIMEOUT 1000)
-
-        py_test_modules(test_dist_transformer MODULES test_dist_transformer)
-        set_tests_properties(test_dist_transformer PROPERTIES TIMEOUT 1000)
+        # FIXME(typhoonzero): add this back
+	#py_test_modules(test_dist_transformer MODULES test_dist_transformer)
+	#set_tests_properties(test_dist_transformer PROPERTIES TIMEOUT 1000)
     endif(NOT APPLE)
     py_test_modules(test_dist_transpiler MODULES test_dist_transpiler)
 endif()
diff --git a/python/paddle/fluid/tests/unittests/test_dropout_op.py b/python/paddle/fluid/tests/unittests/test_dropout_op.py
index 0296bc2af4e0b79478c34b4cceab32b5a8a50f2f..be3c5f3b9558ec522803ed9a5acedea75cda6ccc 100644
--- a/python/paddle/fluid/tests/unittests/test_dropout_op.py
+++ b/python/paddle/fluid/tests/unittests/test_dropout_op.py
@@ -85,6 +85,69 @@ class TestDropoutOp5(OpTest):
         self.check_output()
 
 
+class TestDropoutOp6(TestDropoutOp):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
+        self.attrs = {
+            'dropout_prob': 1.0,
+            'fix_seed': True,
+            'is_test': False,
+            'dropout_implementation': 'upscale_in_train'
+        }
+        self.outputs = {
+            'Out': np.zeros((32, 64)).astype('float32'),
+            'Mask': np.zeros((32, 64)).astype('float32')
+        }
+
+
+class TestDropoutOp7(TestDropoutOp):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64, 2)).astype("float32")}
+        self.attrs = {
+            'dropout_prob': 0.0,
+            'fix_seed': True,
+            'is_test': False,
+            'dropout_implementation': 'upscale_in_train'
+        }
+        self.outputs = {
+            'Out': self.inputs['X'],
+            'Mask': np.ones((32, 64, 2)).astype('float32')
+        }
+
+
+class TestDropoutOp8(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
+        self.attrs = {
+            'dropout_prob': 0.35,
+            'fix_seed': True,
+            'is_test': True,
+            'dropout_implementation': 'upscale_in_train'
+        }
+        self.outputs = {'Out': self.inputs['X']}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestDropoutOp9(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64, 3)).astype("float32")}
+        self.attrs = {
+            'dropout_prob': 0.75,
+            'is_test': True,
+            'dropout_implementation': 'upscale_in_train'
+        }
+        self.outputs = {'Out': self.inputs['X']}
+
+    def test_check_output(self):
+        self.check_output()
+
+
 class TestFP16DropoutOp(OpTest):
     def setUp(self):
         self.op_type = "dropout"
diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py
index 2192139f8d5950286691a77333dd8ec35505b033..28d7df8e45d4f3209fb6205b76f8d1be8e73392b 100644
--- a/python/paddle/fluid/transpiler/distribute_transpiler.py
+++ b/python/paddle/fluid/transpiler/distribute_transpiler.py
@@ -49,6 +49,7 @@ LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad"
 OP_ROLE_VAR_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleVarAttrName()
 RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName(
 )
+OPT_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.Optimize
 RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC
 DIST_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.Dist
 LR_SCHED_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.LRSched
@@ -1717,8 +1718,10 @@ to transpile() call.")
         lr_ops = []
         block = self.origin_program.global_block()
         for op in block.ops:
-            if int(op.attr(RPC_OP_ROLE_ATTR_NAME)) == int(
-                    LR_SCHED_OP_ROLE_ATTR_VALUE):
+            role_id = int(op.attr(RPC_OP_ROLE_ATTR_NAME))
+            if role_id == int(LR_SCHED_OP_ROLE_ATTR_VALUE) or \
+                role_id == int(LR_SCHED_OP_ROLE_ATTR_VALUE) | \
+                    int(OPT_OP_ROLE_ATTR_VALUE):
                 lr_ops.append(op)
                 log("append lr op: ", op.type)
         return lr_ops