Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into quantize_transpiler_update

paddle/fluid/framework/details/cow_ptr.h

@@ -20,41 +20,79 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-template <class T>
+// Change it to thread safe flags if needed.
-class COWPtr {
+class ThreadUnsafeOwnershipFlags {
  public:
-  typedef std::shared_ptr<T> RefPtr;
+  explicit ThreadUnsafeOwnershipFlags(bool flag) : flag_(flag) {}
 
- private:
+  ThreadUnsafeOwnershipFlags(const ThreadUnsafeOwnershipFlags& other) = delete;
-  RefPtr m_sp;
+  ThreadUnsafeOwnershipFlags& operator=(
+      const ThreadUnsafeOwnershipFlags& other) = delete;
+  ThreadUnsafeOwnershipFlags(ThreadUnsafeOwnershipFlags&& other) = default;
 
-  void detach() {
+  void SetOwnership(bool flag) { flag_ = flag; }
-    T* tmp = m_sp.get();
+
-    if (!(tmp == nullptr || m_sp.unique())) {
+  // Invoke the callback if it is not owned.
-      m_sp = RefPtr(new T(*tmp));
+  template <typename Callback>
+  void AcquireOwnershipOnce(Callback acquire) {
+    if (!flag_) {
+      acquire();
+      flag_ = true;
     }
   }
 
- public:
+ private:
-  COWPtr() : m_sp(nullptr) {}
+  bool flag_;
-  explicit COWPtr(T* t) : m_sp(t) {}
+};
-  explicit COWPtr(const RefPtr& refptr) : m_sp(refptr) {}
 
-  const T& Data() const { return operator*(); }
+// Copy-On-Write pointer.
+// It will hold a T* pointer, and only copy once when `MutableData` is invoked.
+//
+// The template parameter OwnershipFlags should have:
+//   * a constructor takes a bool. True if own.
+//   * SetOwnership(bool flag).
+//   * AcquireOwnershipOnce(Callback). It will invoke the callback if it is not
+//     owned.
+//
+// https://en.wikipedia.org/wiki/Copy-on-write
+template <typename T, typename OwnershipFlags = ThreadUnsafeOwnershipFlags>
+class COWPtr {
+ public:
+  // Ctor from raw pointer.
+  explicit COWPtr(T* ptr) : payload_(ptr), ownership_{true} {}
 
-  T* MutableData() { return operator->(); }
+  // Move methods. Steal ownership from origin
+  COWPtr(COWPtr&& other)
+      : payload_(other.payload_), ownership_{std::move(other.ownership_)} {}
+  COWPtr& operator=(COWPtr&& origin) = default;
 
-  const T& operator*() const { return *m_sp; }
+  // Copy methods. Not own payload
-  T& operator*() {
+  COWPtr(const COWPtr& other) : payload_(other.payload_), ownership_{false} {}
-    detach();
+  COWPtr& operator=(const COWPtr& other) {
-    return *m_sp;
+    payload_ = other.payload_;
+    ownership_.SetOwnership(false);
+    return *this;
   }
-  const T* operator->() const { return m_sp.operator->(); }
+
-  T* operator->() {
+  // Access read only data.
-    detach();
+  const T& Data() const { return *payload_; }
-    return m_sp.operator->();
+
+  // Access mutable data. If the data is not owned, the data will be copied
+  // before.
+  T* MutableData() {
+    ownership_.AcquireOwnershipOnce(
+        [this] { payload_.reset(new T(*payload_)); });
+    return payload_.get();
   }
+
+ private:
+  // Actual data pointer.
+  std::shared_ptr<T> payload_;
+
+  // Ownership flag.
+  OwnershipFlags ownership_;
 };
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/cow_ptr_test.cc

@@ -30,14 +30,6 @@ TEST(COWPtr, all) {
   ASSERT_EQ(ptr2.Data(), 10);
 }
 
-TEST(COWPtr, change_old) {
-  COWPtr<int> ptr(new int{0});
-  COWPtr<int> ptr2 = ptr;
-  *ptr.MutableData() = 10;
-  ASSERT_EQ(ptr2.Data(), 0);
-  ASSERT_EQ(ptr.Data(), 10);
-}
-
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -210,43 +210,6 @@ std::vector<std::string> MultiDevSSAGraphBuilder::FindDistTrainRecvVars(
   return recv_vars;
 }
 
-bool MultiDevSSAGraphBuilder::IsDistTrainOp(
-    ir::Node *node, const std::vector<std::string> &send_vars,
-    const std::vector<std::string> &recv_vars) const {
-  if (send_vars.size() == 0 || recv_vars.size() == 0) {
-    return false;
-  }
-
-  /**
-   * Check any of opvars contains `.block` and in sendvars
-   */
-  auto checker = [](const std::vector<std::string> &opvars,
-                    const std::vector<std::string> &rpc_vars) -> bool {
-    for (auto &var : opvars) {
-      // a variable name with the suffix `.block` means it's a splited
-      // variable by (DistributeTranspiler)
-      // [python/paddle/fluid/transpiler/distribute_transpiler.py]
-      if (var.find(".block") != std::string::npos &&
-          std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) {
-        return true;
-      }
-    }
-    return false;
-  };
-
-  std::vector<std::string> input_var_names;
-  std::vector<std::string> output_var_names;
-  for (ir::Node *input : node->inputs) {
-    input_var_names.push_back(input->Name());
-  }
-  for (ir::Node *output : node->outputs) {
-    output_var_names.push_back(output->Name());
-  }
-
-  return checker(output_var_names, send_vars) ||
-         checker(input_var_names, recv_vars);
-}
-
 size_t MultiDevSSAGraphBuilder::GetAppropriateDeviceID(
     const std::vector<std::string> &var_names) const {
   int64_t numel_sum = 0;
@@ -370,7 +333,9 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
         }
       }
       is_dist_train = true;
-    } else if (IsDistTrainOp(node, send_vars, recv_vars)) {
+    } else if (boost::get<int>(node->Op()->GetAttr(
+                   OpProtoAndCheckerMaker::OpRoleAttrName())) ==
+               static_cast<int>(OpRole::kDist)) {
       int op_dev_id = CreateDistTrainOp(&result, node);
       if (node->Op()->Type() == "concat") {
         auto origin_param_name = node->Op()->OutputArgumentNames()[0];
@@ -736,6 +701,7 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
           .emplace(varname, op_dev_id);
     }
   } else {
+    LOG(ERROR) << "got unexpected dist op: " << node->Op()->Type();
     PADDLE_THROW(
         "the distribute training related op should be in [split_byref, "
         "concat].");

paddle/fluid/framework/details/multi_devices_graph_pass.h

@@ -51,12 +51,6 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
   int CreateRPCOp(ir::Graph *result, ir::Node *node) const;
   int CreateDistTrainOp(ir::Graph *result, ir::Node *node) const;
 
-  /**
-   * Is this operator as the end-point operator before/after send operator.
-   */
-  bool IsDistTrainOp(ir::Node *node, const std::vector<std::string> &send_vars,
-                     const std::vector<std::string> &recv_vars) const;
-
   std::vector<std::string> FindDistTrainSendVars(
       const std::vector<ir::Node *> &nodes) const;
 

paddle/fluid/framework/op_proto_maker.cc

@@ -120,6 +120,7 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
           {static_cast<int>(OpRole::kForward),
            static_cast<int>(OpRole::kBackward),
            static_cast<int>(OpRole::kOptimize), static_cast<int>(OpRole::kRPC),
+           static_cast<int>(OpRole::kDist), static_cast<int>(OpRole::kLRSched),
            static_cast<int>(OpRole::kLoss) | static_cast<int>(OpRole::kForward),
            static_cast<int>(OpRole::kLoss) |
                static_cast<int>(OpRole::kBackward),

paddle/fluid/framework/op_proto_maker.h

@@ -26,7 +26,13 @@ enum class OpRole {
   kForward = 0x0000,
   kBackward = 0x0001,
   kOptimize = 0x0002,
+  // RPC role is for send/recv releated op
   kRPC = 0x0003,
+  // Dist role is for split_byref/split_selected_rows/concat
+  // used for distributed training.
+  kDist = 0x0004,
+  // Tag all learning rate scheduler operators.
+  kLRSched = 0x0005,
 
   kLoss = 0x0100,
   // The default value of op's role. This should be only used for unittests and

paddle/fluid/operators/detection_map_op.h

@@ -76,8 +76,8 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
     auto ap_type = GetAPType(ctx.Attr<std::string>("ap_type"));
     int class_num = ctx.Attr<int>("class_num");
 
-    auto& label_lod = in_label->lod();
+    auto label_lod = in_label->lod();
-    auto& detect_lod = in_detect->lod();
+    auto detect_lod = in_detect->lod();
     PADDLE_ENFORCE_EQ(label_lod.size(), 1UL,
                       "Only support one level sequence now.");
     PADDLE_ENFORCE_EQ(label_lod[0].size(), detect_lod[0].size(),
@@ -166,11 +166,11 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
     auto labels = framework::EigenTensor<T, 2>::From(input_label);
     auto detect = framework::EigenTensor<T, 2>::From(input_detect);
 
-    auto& label_lod = input_label.lod();
+    auto label_lod = input_label.lod();
-    auto& detect_lod = input_detect.lod();
+    auto detect_lod = input_detect.lod();
 
     int batch_size = label_lod[0].size() - 1;
-    auto& label_index = label_lod[0];
+    auto label_index = label_lod[0];
 
     for (int n = 0; n < batch_size; ++n) {
       std::map<int, std::vector<Box>> boxes;
@@ -274,6 +274,7 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
 
     output_true_pos->set_lod(true_pos_lod);
     output_false_pos->set_lod(false_pos_lod);
+    return;
   }
 
   void GetInputPos(const framework::Tensor& input_pos_count,
@@ -291,7 +292,7 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
     auto SetData = [](const framework::LoDTensor& pos_tensor,
                       std::map<int, std::vector<std::pair<T, int>>>& pos) {
       const T* pos_data = pos_tensor.data<T>();
-      auto& pos_data_lod = pos_tensor.lod()[0];
+      auto pos_data_lod = pos_tensor.lod()[0];
       for (size_t i = 0; i < pos_data_lod.size() - 1; ++i) {
         for (size_t j = pos_data_lod[i]; j < pos_data_lod[i + 1]; ++j) {
           T score = pos_data[j * 2];
@@ -316,23 +317,20 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
       std::map<int, std::vector<std::pair<T, int>>>* false_pos) const {
     int batch_size = gt_boxes.size();
     for (int n = 0; n < batch_size; ++n) {
-      auto& image_gt_boxes = gt_boxes[n];
+      auto image_gt_boxes = gt_boxes[n];
-      for (auto& image_gt_box : image_gt_boxes) {
+      for (auto it = image_gt_boxes.begin(); it != image_gt_boxes.end(); ++it) {
         size_t count = 0;
-        auto& labeled_bboxes = image_gt_box.second;
+        auto labeled_bboxes = it->second;
         if (evaluate_difficult) {
           count = labeled_bboxes.size();
         } else {
-          for (auto& box : labeled_bboxes) {
+          for (size_t i = 0; i < labeled_bboxes.size(); ++i)
-            if (!box.is_difficult) {
+            if (!(labeled_bboxes[i].is_difficult)) ++count;
-              ++count;
-            }
-          }
         }
         if (count == 0) {
           continue;
         }
-        int label = image_gt_box.first;
+        int label = it->first;
         if (label_pos_count->find(label) == label_pos_count->end()) {
           (*label_pos_count)[label] = count;
         } else {

paddle/fluid/operators/distributed/variable_response.cc

@@ -92,9 +92,14 @@ bool VariableResponse::CopyLodTensorData(
     ::google::protobuf::io::CodedInputStream* input,
     const platform::DeviceContext& ctx, const framework::DDim& dims,
     int length) {
+  auto server_var = GetVar();
+  if (!server_var) {
+    LOG(ERROR) << "recved var should not on current server: "
+               << meta_.varname();
+    return false;
+  }
   auto* tensor = GetVar()->GetMutable<framework::LoDTensor>();
   tensor->Resize(dims);
-
   framework::LoD lod;
   for (int i = 0; i < meta_.lod_level(); ++i) {
     framework::Vector<size_t> v;
@@ -107,7 +112,6 @@ bool VariableResponse::CopyLodTensorData(
 
   void* tensor_data =
       tensor->mutable_data(ctx.GetPlace(), ToTypeIndex(meta_.data_type()));
-
   if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) {
     return false;
   }

paddle/fluid/operators/extract_rows_op.cc

@@ -50,7 +50,7 @@ class ExtractRowsOp : public framework::OperatorBase {
     auto &in = scope.FindVar(Input("X"))->Get<framework::SelectedRows>();
     auto out = scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
 
-    auto &in_rows = in.rows();
+    auto in_rows = in.rows();
     auto out_dim = framework::make_ddim(
         std::vector<int64_t>{static_cast<int64_t>(in_rows.size()), 1});
     auto dst_ptr = out->mutable_data<int64_t>(out_dim, in.place());

paddle/fluid/operators/math/selected_rows_functor.cu

@@ -60,9 +60,11 @@ struct SelectedRowsAdd<platform::CUDADeviceContext, T> {
     auto out_place = context.GetPlace();
     PADDLE_ENFORCE(platform::is_gpu_place(out_place));
 
-    memory::Copy(boost::get<platform::CUDAPlace>(out_place), out_data,
+    memory::Copy(
-                 boost::get<platform::CUDAPlace>(in1_place), in1_data,
+        boost::get<platform::CUDAPlace>(out_place), out_data,
-                 in1_value.numel() * sizeof(T), context.stream());
+        boost::get<platform::CUDAPlace>(in1_place), in1_data,
+        in1_value.numel() * sizeof(T),
+        reinterpret_cast<const platform::CUDADeviceContext&>(context).stream());
 
     auto* in2_data = in2_value.data<T>();
     memory::Copy(boost::get<platform::CUDAPlace>(out_place),
@@ -107,7 +109,7 @@ struct SelectedRowsAddTensor<platform::CUDADeviceContext, T> {
     PADDLE_ENFORCE_EQ(in1_height, out_dims[0]);
 
     auto& in1_value = input1.value();
-    framework::Vector<int64_t> in1_rows(input1.rows());
+    auto& in1_rows = input1.rows();
 
     int64_t in1_row_numel = in1_value.numel() / in1_rows.size();
     PADDLE_ENFORCE_EQ(in1_row_numel, input2.numel() / in1_height);
@@ -146,7 +148,7 @@ struct SelectedRowsAddTo<platform::CUDADeviceContext, T> {
     auto in1_height = input1.height();
     PADDLE_ENFORCE_EQ(in1_height, input2->height());
 
-    auto& in1_rows = input1.rows();
+    framework::Vector<int64_t> in1_rows(input1.rows());
     auto& in2_rows = *(input2->mutable_rows());
 
     auto& in1_value = input1.value();
@@ -206,7 +208,7 @@ struct SelectedRowsAddToTensor<platform::CUDADeviceContext, T> {
     PADDLE_ENFORCE_EQ(in1_height, in2_dims[0]);
 
     auto& in1_value = input1.value();
-    framework::Vector<int64_t> in1_rows(input1.rows());
+    auto& in1_rows = input1.rows();
 
     int64_t in1_row_numel = in1_value.numel() / in1_rows.size();
     PADDLE_ENFORCE_EQ(in1_row_numel, input2->numel() / in1_height);

paddle/fluid/operators/math/selected_rows_functor_test.cu

@@ -20,7 +20,9 @@ limitations under the License. */
 TEST(selected_rows_functor, gpu_add) {
   paddle::platform::CUDAPlace gpu_place(0);
   paddle::platform::CPUPlace cpu_place;
-  paddle::platform::CUDADeviceContext ctx(gpu_place);
+  paddle::platform::CUDADeviceContext& ctx =
+      *reinterpret_cast<paddle::platform::CUDADeviceContext*>(
+          paddle::platform::DeviceContextPool::Instance().Get(gpu_place));
   paddle::operators::math::SetConstant<paddle::platform::CUDADeviceContext,
                                        float>
       functor;
@@ -132,7 +134,9 @@ TEST(selected_rows_functor, gpu_add) {
 TEST(selected_rows_functor, gpu_add_to) {
   paddle::platform::CUDAPlace gpu_place(0);
   paddle::platform::CPUPlace cpu_place;
-  paddle::platform::CUDADeviceContext ctx(gpu_place);
+  paddle::platform::CUDADeviceContext& ctx =
+      *reinterpret_cast<paddle::platform::CUDADeviceContext*>(
+          paddle::platform::DeviceContextPool::Instance().Get(gpu_place));
   paddle::operators::math::SetConstant<paddle::platform::CUDADeviceContext,
                                        float>
       functor;

paddle/fluid/operators/sum_op.h

@@ -123,6 +123,7 @@ class SumKernel : public framework::OpKernel<T> {
 
       out_value->Resize(framework::make_ddim(in_dim));
       out_value->mutable_data<T>(context.GetPlace());
+
       // if all the input sparse vars are empty, no need to
       // merge these vars.
       if (first_dim == 0UL) {

paddle/fluid/pybind/const_value.cc

@@ -36,7 +36,9 @@ void BindConstValue(pybind11::module* m) {
       .value("Backward", framework::OpRole::kBackward)
       .value("Optimize", framework::OpRole::kOptimize)
       .value("Loss", framework::OpRole::kLoss)
-      .value("RPC", framework::OpRole::kRPC);
+      .value("RPC", framework::OpRole::kRPC)
+      .value("Dist", framework::OpRole::kDist)
+      .value("LRSched", framework::OpRole::kLRSched);
 
   op_proto_and_checker_maker.def(
       "kOpRoleAttrName", framework::OpProtoAndCheckerMaker::OpRoleAttrName);

python/paddle/fluid/framework.py

@@ -1509,6 +1509,30 @@ class Program(object):
         self._op_role_var = []
         self._current_role = OpRole.Forward
 
+    @contextlib.contextmanager
+    def _lr_schedule_guard(self):
+        """
+        A with guard to set :code:`LRSched` :code:`OpRole` and
+        :code:`OpRoleVar` automatically. The :code:`OpRoleVar` is
+        set to the target learning rate.
+
+        Notes: This is a very low level API. Users should not use it directly.
+
+
+        Examples:
+
+            >>> p, g = backward(...)
+            >>> with program.lr_schedule_guard():
+            >>>     lr = lr * decay
+        """
+        OpRole = core.op_proto_and_checker_maker.OpRole
+        self._current_role = OpRole.LRSched
+        # TODO(typhoonzero): how to set target learning rate var
+        self._op_role_var = []
+        yield
+        self._op_role_var = []
+        self._current_role = OpRole.Forward
+
     def __str__(self):
         """
         Get the protobuf debug string of this Program.

python/paddle/fluid/initializer.py

@@ -74,7 +74,7 @@ class Initializer(object):
     directly, but need to use one of its implementations.
     """
 
-    def __init_(self):
+    def __init__(self):
         pass
 
     def __call__(self, param, block):
@@ -293,7 +293,7 @@ class TruncatedNormalInitializer(Initializer):
         assert loc is not None
         assert scale is not None
         assert seed is not None
-        super(NormalInitializer, self).__init__()
+        super(TruncatedNormalInitializer, self).__init__()
         self._mean = loc
         self._std_dev = scale
         self._seed = seed

python/paddle/fluid/layers/learning_rate_scheduler.py

@@ -27,7 +27,7 @@ from . import nn
 from . import ops
 from . import tensor
 from ..initializer import init_on_cpu
-from ..framework import default_main_program, Parameter
+from ..framework import default_main_program, Parameter, unique_name
 
 __all__ = [
     'exponential_decay', 'natural_exp_decay', 'inverse_time_decay',
@@ -63,11 +63,12 @@ def noam_decay(d_model, warmup_steps):
     Returns:
         The decayed learning rate.
     """
-    global_step = _decay_step_counter(1)
+    with default_main_program()._lr_schedule_guard():
+        global_step = _decay_step_counter(1)
 
-    a = global_step**-0.5
+        a = global_step**-0.5
-    b = (warmup_steps**-1.5) * global_step
+        b = (warmup_steps**-1.5) * global_step
-    lr_value = (d_model**-0.5) * ops.elementwise_min(a, b)
+        lr_value = (d_model**-0.5) * ops.elementwise_min(a, b)
 
     return lr_value
 
@@ -108,14 +109,15 @@ def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False):
           sgd_optimizer.minimize(avg_cost)
 
     """
-    global_step = _decay_step_counter()
+    with default_main_program()._lr_schedule_guard():
+        global_step = _decay_step_counter()
 
-    div_res = global_step / decay_steps
+        div_res = global_step / decay_steps
-    if staircase:
+        if staircase:
-        div_res = ops.floor(div_res)
+            div_res = ops.floor(div_res)
-    decayed_lr = learning_rate * (decay_rate**div_res)
+        decayed_lr = learning_rate * (decay_rate**div_res)
 
-    return decayed_lr
+        return decayed_lr
 
 
 def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False):
@@ -136,14 +138,15 @@ def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False):
     Returns:
         The decayed learning rate
     """
-    global_step = _decay_step_counter()
+    with default_main_program()._lr_schedule_guard():
+        global_step = _decay_step_counter()
 
-    div_res = global_step / decay_steps
+        div_res = global_step / decay_steps
-    if staircase:
+        if staircase:
-        div_res = ops.floor(div_res)
+            div_res = ops.floor(div_res)
-    decayed_lr = learning_rate * ops.exp(-1 * decay_rate * div_res)
+        decayed_lr = learning_rate * ops.exp(-1 * decay_rate * div_res)
 
-    return decayed_lr
+        return decayed_lr
 
 
 def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False):
@@ -181,15 +184,16 @@ def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False):
                     staircase=True))
           sgd_optimizer.minimize(avg_cost)
     """
-    global_step = _decay_step_counter()
+    with default_main_program()._lr_schedule_guard():
+        global_step = _decay_step_counter()
 
-    div_res = global_step / decay_steps
+        div_res = global_step / decay_steps
-    if staircase:
+        if staircase:
-        div_res = ops.floor(div_res)
+            div_res = ops.floor(div_res)
 
-    decayed_lr = learning_rate / (1 + decay_rate * div_res)
+        decayed_lr = learning_rate / (1 + decay_rate * div_res)
 
-    return decayed_lr
+        return decayed_lr
 
 
 def polynomial_decay(learning_rate,
@@ -220,25 +224,28 @@ def polynomial_decay(learning_rate,
     Returns:
         Variable: The decayed learning rate
     """
-    global_step = _decay_step_counter()
+    with default_main_program()._lr_schedule_guard():
-
+        global_step = _decay_step_counter()
-    if cycle:
+
-        div_res = ops.ceil(global_step / decay_steps)
+        if cycle:
-        zero_var = tensor.fill_constant(shape=[1], dtype='float32', value=0.0)
+            div_res = ops.ceil(global_step / decay_steps)
-        one_var = tensor.fill_constant(shape=[1], dtype='float32', value=1.0)
+            zero_var = tensor.fill_constant(
-
+                shape=[1], dtype='float32', value=0.0)
-        with control_flow.Switch() as switch:
+            one_var = tensor.fill_constant(
-            with switch.case(global_step == zero_var):
+                shape=[1], dtype='float32', value=1.0)
-                tensor.assign(input=one_var, output=div_res)
+
-        decay_steps = decay_steps * div_res
+            with control_flow.Switch() as switch:
-    else:
+                with switch.case(global_step == zero_var):
-        decay_steps_var = tensor.fill_constant(
+                    tensor.assign(input=one_var, output=div_res)
-            shape=[1], dtype='float32', value=float(decay_steps))
+            decay_steps = decay_steps * div_res
-        global_step = ops.elementwise_min(x=global_step, y=decay_steps_var)
+        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)
 
-    decayed_lr = (learning_rate - end_learning_rate) * \
+        decayed_lr = (learning_rate - end_learning_rate) * \
-        ((1 - global_step / decay_steps) ** power) + end_learning_rate
+            ((1 - global_step / decay_steps) ** power) + end_learning_rate
-    return decayed_lr
+        return decayed_lr
 
 
 def piecewise_decay(boundaries, values):
@@ -266,34 +273,36 @@ def piecewise_decay(boundaries, values):
 
 
     """
+    with default_main_program()._lr_schedule_guard():
+        if len(values) - len(boundaries) != 1:
+            raise ValueError("len(values) - len(boundaries) should be 1")
 
-    if len(values) - len(boundaries) != 1:
+        global_step = _decay_step_counter()
-        raise ValueError("len(values) - len(boundaries) should be 1")
-
-    global_step = _decay_step_counter()
 
-    lr = tensor.create_global_var(
+        lr = tensor.create_global_var(
-        shape=[1],
+            shape=[1],
-        value=0.0,
+            value=0.0,
-        dtype='float32',
+            dtype='float32',
-        persistable=True,
+            persistable=True,
-        name="learning_rate")
+            name="learning_rate")
 
-    with control_flow.Switch() as switch:
+        with control_flow.Switch() as switch:
-        for i in range(len(boundaries)):
+            for i in range(len(boundaries)):
-            boundary_val = tensor.fill_constant(
+                boundary_val = tensor.fill_constant(
+                    shape=[1],
+                    dtype='float32',
+                    value=float(boundaries[i]),
+                    force_cpu=True)
+                value_var = tensor.fill_constant(
+                    shape=[1], dtype='float32', value=float(values[i]))
+                with switch.case(global_step < boundary_val):
+                    tensor.assign(value_var, lr)
+            last_value_var = tensor.fill_constant(
                 shape=[1],
                 dtype='float32',
-                value=float(boundaries[i]),
+                value=float(values[len(values) - 1]))
-                force_cpu=True)
+            with switch.default():
-            value_var = tensor.fill_constant(
+                tensor.assign(last_value_var, lr)
-                shape=[1], dtype='float32', value=float(values[i]))
-            with switch.case(global_step < boundary_val):
-                tensor.assign(value_var, lr)
-        last_value_var = tensor.fill_constant(
-            shape=[1], dtype='float32', value=float(values[len(values) - 1]))
-        with switch.default():
-            tensor.assign(last_value_var, lr)
 
     return lr
 

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -80,7 +80,8 @@ if(WITH_DISTRIBUTE)
         py_test_modules(test_dist_se_resnext MODULES test_dist_se_resnext SERIAL)
     endif(NOT APPLE)
     py_test_modules(test_dist_transpiler MODULES test_dist_transpiler)
-    py_test_modules(test_dist_transformer MODULES test_dist_transformer SERIAL)
+    #FIXME(gongwb): random fails.
+    #py_test_modules(test_dist_transformer MODULES test_dist_transformer SERIAL)
 endif()
 py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SERIAL)
 py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL)

python/paddle/fluid/tests/unittests/op_test.py

@@ -345,7 +345,7 @@ class OpTest(unittest.TestCase):
                         actual_t, expect_t, atol=atol, equal_nan=equal_nan),
                     "Output (" + out_name + ") has diff at " + str(place) +
                     "\nExpect " + str(expect_t) + "\n" + "But Got" +
-                    str(actual_t) + " in class " + self.__class__.__name__)
+                    str(actual_t))
                 if isinstance(expect, tuple):
                     self.assertListEqual(actual.recursive_sequence_lengths(),
                                          expect[1], "Output (" + out_name +

python/paddle/fluid/tests/unittests/test_detection_map_op.py

@@ -20,7 +20,6 @@ import six
 import sys
 import collections
 import math
-import paddle.fluid as fluid
 from op_test import OpTest
 
 
@@ -33,7 +32,7 @@ class TestDetectionMAPOp(OpTest):
         self.detect = np.array(self.detect).astype('float32')
         self.mAP = np.array(self.mAP).astype('float32')
 
-        if len(self.class_pos_count) > 0:
+        if (len(self.class_pos_count) > 0):
             self.class_pos_count = np.array(self.class_pos_count).astype(
                 'int32')
             self.true_pos = np.array(self.true_pos).astype('float32')
@@ -274,7 +273,7 @@ class TestDetectionMAPOp11Point(TestDetectionMAPOp):
 class TestDetectionMAPOpMultiBatch(TestDetectionMAPOp):
     def init_test_case(self):
         super(TestDetectionMAPOpMultiBatch, self).init_test_case()
-        self.class_pos_count = [0, 2, 1, 0]
+        self.class_pos_count = [0, 2, 1]
         self.true_pos_lod = [[0, 3, 2]]
         self.true_pos = [[0.7, 1.], [0.3, 0.], [0.2, 1.], [0.8, 0.], [0.1, 1.]]
         self.false_pos_lod = [[0, 3, 2]]

python/paddle/fluid/tests/unittests/test_dist_mnist.py

@@ -22,7 +22,7 @@ class TestDistMnist2x2(TestDistBase):
         self._sync_mode = True
         self._use_reduce = False
 
-    def test_se_resnext(self):
+    def test_dist_train(self):
         self.check_with_place("dist_mnist.py", delta=1e-7)
 
 
@@ -31,7 +31,7 @@ class TestDistMnist2x2WithMemopt(TestDistBase):
         self._sync_mode = True
         self._mem_opt = True
 
-    def test_se_resnext(self):
+    def test_dist_train(self):
         self.check_with_place("dist_mnist.py", delta=1e-7)
 
 
@@ -40,7 +40,7 @@ class TestDistMnistAsync(TestDistBase):
         self._sync_mode = False
         self._use_reduce = False
 
-    def test_se_resnext(self):
+    def test_dist_train(self):
         self.check_with_place("dist_mnist.py", delta=200)
 
 

python/paddle/fluid/tests/unittests/test_dist_se_resnext.py

@@ -21,7 +21,16 @@ class TestDistSeResneXt2x2(TestDistBase):
     def _setup_config(self):
         self._sync_mode = True
 
-    def test_se_resnext(self):
+    def test_dist_train(self):
+        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)
 
 
@@ -29,7 +38,7 @@ class TestDistSeResneXt2x2Async(TestDistBase):
     def _setup_config(self):
         self._sync_mode = False
 
-    def test_se_resnext(self):
+    def test_dist_train(self):
         self.check_with_place("dist_se_resnext.py", delta=100)
 
 

python/paddle/fluid/tests/unittests/test_dist_transformer.py

@@ -59,7 +59,7 @@ class TestDistTransformer2x2Sync(TestDistBase):
     def _setup_config(self):
         self._sync_mode = True
 
-    def test_transformer(self):
+    def test_dist_train(self):
         download_files()
         self.check_with_place("dist_transformer.py", delta=1e-5)
 
@@ -68,7 +68,7 @@ class TestDistTransformer2x2Async(TestDistBase):
     def _setup_config(self):
         self._sync_mode = False
 
-    def test_transformer(self):
+    def test_dist_train(self):
         download_files()
         self.check_with_place("dist_transformer.py", delta=1.0)
 

python/paddle/fluid/tests/unittests/test_dist_word2vec.py

@@ -17,19 +17,28 @@ import unittest
 from test_dist_base import TestDistBase
 
 
-class TestDistSeResneXt2x2(TestDistBase):
+class TestDistW2V2x2(TestDistBase):
     def _setup_config(self):
         self._sync_mode = True
 
-    def test_se_resnext(self):
+    def test_dist_train(self):
         self.check_with_place("dist_word2vec.py", delta=1e-4)
 
 
-class TestDistSeResneXt2x2Async(TestDistBase):
+class TestDistW2V2x2WithMemOpt(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = True
+        self._mem_opt = True
+
+    def test_dist_train(self):
+        self.check_with_place("dist_word2vec.py", delta=1e-4)
+
+
+class TestDistW2V2x2Async(TestDistBase):
     def _setup_config(self):
         self._sync_mode = False
 
-    def test_se_resnext(self):
+    def test_dist_train(self):
         self.check_with_place("dist_word2vec.py", delta=1)
 
 

python/paddle/fluid/transpiler/details/program_utils.py

@@ -21,13 +21,12 @@ import paddle
 
 
 def delete_ops(block, ops):
-    try:
+    for op in ops:
-        start = list(block.ops).index(ops[0])
+        try:
-        end = list(block.ops).index(ops[-1])
+            idx = list(block.ops).index(op)
-        [block._remove_op(start) for _ in six.moves.range(end - start + 1)]
+            block._remove_op(idx)
-    except Exception as e:
+        except Exception as e:
-        raise e
+            print(e)
-    block.program._sync_with_cpp()
 
 
 def find_op_by_input_arg(block, arg_name):
@@ -37,10 +36,18 @@ def find_op_by_input_arg(block, arg_name):
     return -1
 
 
-def find_op_by_output_arg(block, arg_name):
+def find_op_by_output_arg(block, arg_name, reverse=False):
-    for index, op in enumerate(block.ops):
+    if reverse:
-        if arg_name in op.output_arg_names:
+        pos = len(block.ops) - 1
-            return index
+        while pos >= 0:
+            op = block.ops[pos]
+            if arg_name in op.output_arg_names:
+                return pos
+            pos -= 1
+    else:
+        for index, op in enumerate(block.ops):
+            if arg_name in op.output_arg_names:
+                return index
     return -1