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

CMakeLists.txt

@@ -137,9 +137,9 @@ set(EXTERNAL_LIBS
 )
 
 if(WITH_GPU)
-    list(APPEND EXTERNAL_LIB ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
+    list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
     if(NOT WITH_DSO)
-        list(APPEND EXTERNAL_LIB ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
+        list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
     endif(NOT WITH_DSO)
 endif(WITH_GPU)
 

paddle/framework/backward.cc

@@ -15,6 +15,8 @@
 #include "paddle/framework/backward.h"
 
 #include <list>
+#include <memory>
+
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
@@ -43,11 +45,11 @@ static bool AllInSet(
   return all_in_set;
 }
 
-static std::shared_ptr<OperatorBase> NOP() {
-  auto net_op = std::make_shared<operators::NetOp>();
+static std::unique_ptr<OperatorBase> NOP() {
+  auto net_op = new operators::NetOp();
   net_op->SetType("@NOP@");
   net_op->CompleteAddOp();
-  return net_op;
+  return std::unique_ptr<OperatorBase>(net_op);
 }
 
 //  Get backward operator from a forward operator, a recursive implementation.
@@ -62,11 +64,7 @@ static std::shared_ptr<OperatorBase> NOP() {
 //  operator, in a complex situation, it maybe a NetOp.
 //
 //  See Backward.h for details
-static std::shared_ptr<OperatorBase> BackwardRecursive(
-    const OperatorBase& forwardOp,
-    std::unordered_set<std::string>& no_grad_names, size_t& uniq_id);
-
-std::shared_ptr<OperatorBase> BackwardRecursive(
+static std::unique_ptr<OperatorBase> BackwardRecursive(
     const OperatorBase& forwardOp,
     std::unordered_set<std::string>& no_grad_names, size_t& uniq_id) {
   //  If all input gradients of forwarding operator do not need to calculate,
@@ -91,7 +89,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
   }
 
   // Returned gradient network
-  auto net = std::make_shared<operators::NetOp>();
+  auto net = std::unique_ptr<operators::NetOp>(new operators::NetOp());
 
   if (forwardOp.IsNetOp()) {
     // Because forwardOp is a net op, it can static_cast.
@@ -105,14 +103,14 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     // reversely travel forwardNet and collect all duplicate outputs.
     for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend();
          ++it, ++local_op_id) {
-      auto fwd = *it;
+      auto& fwd = *it;
       auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id);
-      net->AddOp(bwd);
       ForEachVarName(bwd->Outputs(),
                      [&dup_output_ops, local_op_id](const std::string& out) {
                        dup_output_ops[out].emplace_back(local_op_id);
                        return false;
                      });
+      net->AddOp(std::move(bwd));
     }
     // Get unique ID for this method.
     auto uid = uniq_id++;
@@ -122,7 +120,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     // to handle this case. For each duplicate output, rename it to an alias
     // (original name with a offset), append an `add` op for its operator,
     // and finally sum all the alias variable to the final output variable y.
-    using Pos = std::pair<size_t, std::shared_ptr<OperatorBase>>;
+    using Pos = std::pair<size_t, std::unique_ptr<OperatorBase>>;
     std::list<Pos> insert_position;
     for (auto& dup_output_op : dup_output_ops) {
       const std::string& name = dup_output_op.first;
@@ -150,13 +148,13 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
         [](const Pos& l, const Pos& r) { return l.first > r.first; });
 
     for (auto& pos : insert_position) {
-      net->InsertOp(pos.first + 1, pos.second);
+      net->InsertOp(pos.first + 1, std::move(pos.second));
     }
   } else {
-    std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
+    std::unique_ptr<OperatorBase> grad_op(OpRegistry::CreateGradOp(forwardOp));
 
-    ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net,
-                                       grad_op](const std::string& grad_input) {
+    ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
+                                          const std::string& grad_input) {
       if (no_grad_names.count(grad_input)) {
         // +1 for \0
         std::string prefix = grad_input.substr(
@@ -190,23 +188,23 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
       const auto& stepnet_op =
           *static_cast<const OperatorBase*>(&rnnop.stepnet());
       // create stepnet's gradient op
-      auto grad_stepnet = BackwardRecursive(stepnet_op, no_grad_names, uniq_id);
       rnn_grad_op->set_stepnet(
-          std::static_pointer_cast<operators::NetOp>(grad_stepnet));
+          BackwardRecursive(stepnet_op, no_grad_names, uniq_id));
     }
 
     if (net->ops_.empty()) {  // Current no aux op is added to network
       return grad_op;
     }
-    net->AddOp(grad_op);
+    net->AddOp(std::move(grad_op));
   }
   net->SetType("@GENERATED_BACKWARD@");
   net->CompleteAddOp();
-  return net;
-}  // namespace framework
+  return std::unique_ptr<OperatorBase>(
+      static_cast<OperatorBase*>(net.release()));
+}
 
 // See header for comments
-std::shared_ptr<OperatorBase> Backward(
+std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars) {
   std::unordered_set<std::string> no_grad_names;

paddle/framework/backward.h

@@ -20,7 +20,7 @@ namespace framework {
 
 // Create the backward operator from a forward operator.
 // TODO(yuyang18): Add more API reference comment.
-extern std::shared_ptr<OperatorBase> Backward(
+extern std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars);
 }  // namespace framework

paddle/framework/backward_test.cc

@@ -32,9 +32,9 @@ class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
  public:
   RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "Input X of Add").AsNoGradient();
-    AddInput("b", "Bias of Add").AsNoGradient();
-    AddOutput("Out", "Out of Add").AsNoGradient();
+    AddInput("X", "Input X of Add").NotInGradient();
+    AddInput("b", "Bias of Add").NotInGradient();
+    AddOutput("Out", "Out of Add").NotInGradient();
     AddComment("Add Op");
   }
 };
@@ -180,8 +180,7 @@ TEST(Backward, simple_op_not_need_grad) {
   auto no_input_gop = f::Backward(*fwd, {"x", "b"});
   ASSERT_NE(no_input_gop, nullptr);
   ASSERT_TRUE(no_input_gop->IsNetOp());
-  ASSERT_EQ(0UL,
-            std::static_pointer_cast<ops::NetOp>(no_input_gop)->ops_.size());
+  ASSERT_EQ(0UL, static_cast<ops::NetOp *>(no_input_gop.get())->ops_.size());
 }
 
 TEST(Backward, net_fc_backward_normal) {

paddle/framework/framework.proto

@@ -60,7 +60,7 @@ message OpProto {
 
     optional bool duplicable = 3 [ default = false ];
     optional bool intermediate = 4 [ default = false ];
-    optional bool no_gradient = 5 [ default = false ];
+    optional bool not_in_gradient = 5 [ default = false ];
   }
 
   // AttrProto describes the C++ type Attribute.

paddle/framework/grad_op_builder.cc

@@ -28,7 +28,7 @@ static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
   const auto& src_arg_list =
       src_type == OpArgType::IN ? proto->inputs() : proto->outputs();
   for (const auto& arg : src_arg_list) {
-    if (arg.no_gradient() && !is_grad) continue;
+    if (arg.not_in_gradient() && !is_grad) continue;
     const std::string src_name = arg.name();
     std::string dst_name = is_grad ? GradVarName(src_name) : src_name;
     dst_inout[dst_name].reserve(src_inout.at(src_name).size());

paddle/framework/grad_op_builder_test.cc

@@ -26,10 +26,10 @@ class IOIgnoredOpMaker : public OpProtoAndCheckerMaker {
   IOIgnoredOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("In1", "a single input");
-    AddInput("In2_mult", "a multiple input").AsDuplicable().AsNoGradient();
+    AddInput("In2_mult", "a multiple input").AsDuplicable().NotInGradient();
     AddInput("In3_mult", "another multiple input").AsDuplicable();
     AddOutput("Out1_mult", "a multiple output").AsDuplicable();
-    AddOutput("Out2", "a single output").AsNoGradient();
+    AddOutput("Out2", "a single output").NotInGradient();
     AddComment("op with inputs and outputs ignored in gradient calculating");
   }
 };

paddle/framework/op_registry.cc

@@ -19,7 +19,7 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-std::shared_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
                                                    const VarNameMap& inputs,
                                                    const VarNameMap& outputs,
                                                    AttributeMap attrs) {
@@ -28,10 +28,10 @@ std::shared_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
                  "Operator '%s' has not been registered.", type);
   it->second.checker_->Check(attrs);
   auto op = it->second.creator_(type, inputs, outputs, attrs);
-  return std::shared_ptr<OperatorBase>(op);
+  return std::unique_ptr<OperatorBase>(op);
 }
 
-std::shared_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
   VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
   VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
   AttributeMap attrs;
@@ -55,10 +55,9 @@ OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
   return ret_val;
 }
 
-std::shared_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
+std::unique_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
   PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
-  std::shared_ptr<OperatorBase> grad_op(BuildGradOp(&op));
-  return grad_op;
+  return std::unique_ptr<OperatorBase>(BuildGradOp(&op));
 }
 
 }  // namespace framework

paddle/framework/op_registry.h

@@ -77,17 +77,17 @@ class OpRegistry {
     }
   }
 
-  static std::shared_ptr<OperatorBase> CreateOp(const std::string& type,
+  static std::unique_ptr<OperatorBase> CreateOp(const std::string& type,
                                                 const VarNameMap& inputs,
                                                 const VarNameMap& outputs,
                                                 AttributeMap attrs);
 
-  static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
+  static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
   static VarNameMap ConvertOpDescVarsToVarNameMap(
       const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars);
 
-  static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op);
+  static std::unique_ptr<OperatorBase> CreateGradOp(const OperatorBase& op);
 
   static std::unordered_map<std::string, const OpInfo>& op_info_map() {
     static std::unordered_map<std::string, const OpInfo> op_info_map_;
@@ -144,8 +144,18 @@ class OpKernelRegistrar : public Registrar {
                     grad_op_class)                                            \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                             \
       __reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \
-  static ::paddle::framework::OpRegistrar<op_class, op_maker_class,           \
-                                          grad_op_class>                      \
+  class _OpClass_##op_type##_ : public op_class {                             \
+   public:                                                                    \
+    DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_);                            \
+    DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class);                   \
+  };                                                                          \
+  class _OpGradClass_##op_type##_ : public grad_op_class {                    \
+   public:                                                                    \
+    DEFINE_OP_CLONE_METHOD(_OpGradClass_##op_type##_);                        \
+    DEFINE_OP_CONSTRUCTOR(_OpGradClass_##op_type##_, grad_op_class);          \
+  };                                                                          \
+  static ::paddle::framework::OpRegistrar<                                    \
+      _OpClass_##op_type##_, op_maker_class, _OpGradClass_##op_type##_>       \
       __op_registrar_##op_type##__(#op_type, #grad_op_type);                  \
   int TouchOpRegistrar_##op_type() {                                          \
     __op_registrar_##op_type##__.Touch();                                     \
@@ -176,7 +186,8 @@ class OpKernelRegistrar : public Registrar {
   REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
 
 /**
- * Macro to mark what Operator and Kernel we will use and tell the compiler to
+ * Macro to mark what Operator and Kernel
+ * we will use and tell the compiler to
  * link them into target.
  */
 #define USE_OP_ITSELF(op_type)                                    \
@@ -196,7 +207,8 @@ class OpKernelRegistrar : public Registrar {
       __attribute__((unused)) =                                  \
           TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE()
 
-// TODO(fengjiayi): The following macros seems ugly, do we have better method?
+// TODO(fengjiayi): The following macros
+// seems ugly, do we have better method?
 
 #ifdef PADDLE_ONLY_CPU
 #define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU)

paddle/framework/op_registry_test.cc

@@ -76,8 +76,7 @@ TEST(OpRegistry, CreateOp) {
   attr->set_type(paddle::framework::AttrType::FLOAT);
   attr->set_f(scale);
 
-  std::shared_ptr<paddle::framework::OperatorBase> op =
-      paddle::framework::OpRegistry::CreateOp(op_desc);
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
@@ -118,8 +117,7 @@ TEST(OpRegistry, DefaultValue) {
 
   ASSERT_TRUE(op_desc.IsInitialized());
 
-  std::shared_ptr<paddle::framework::OperatorBase> op =
-      paddle::framework::OpRegistry::CreateOp(op_desc);
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);

paddle/framework/operator.h

@@ -67,10 +67,6 @@ class OperatorBase {
   OperatorBase(const std::string& type, const VarNameMap& inputs,
                const VarNameMap& outputs, const AttributeMap& attrs);
 
-  OperatorBase(const OperatorBase& o) = delete;
-  OperatorBase& operator=(const OperatorBase& o) = delete;
-  OperatorBase(OperatorBase&& o) = delete;
-
   virtual ~OperatorBase() {}
 
   template <typename T>
@@ -116,10 +112,14 @@ class OperatorBase {
   void SetType(const std::string& type) { type_ = type; }
   const AttributeMap& Attrs() const { return attrs_; }
 
+  // Return a new operator instance, which is as same as this.
+  // Use unique_ptr to prevent caller forget to delete this pointer.
+  virtual std::unique_ptr<OperatorBase> Clone() const = 0;
+
  protected:
   std::string type_;
   // NOTE: in case of OpGrad, inputs_ contains:
-  // I (Inputs)
+  // I (Inputs)opear
   // O (Outputs)
   // OG (Output Gradients)
   VarNameMap inputs_;
@@ -130,12 +130,32 @@ class OperatorBase {
   AttributeMap attrs_;
 };
 
+// Macro for define a clone method.
+// If you are writing an kernel operator, `Clone` will be defined when you
+// register it. i.e. `Clone` method is not needed to define by yourself.
+#define DEFINE_OP_CLONE_METHOD(CLS)                       \
+  std::unique_ptr<OperatorBase> Clone() const final {     \
+    return std::unique_ptr<OperatorBase>(new CLS(*this)); \
+  }
+
+// Macro for define a default constructor for Operator.
+// You can also use
+//   using PARENT_CLASS::PARENT_CLASS;
+// to use parent's constructor.
+#define DEFINE_OP_CONSTRUCTOR(CLS, PARENT_CLS)                                 \
+  CLS(const std::string& type, const VarNameMap& inputs,                       \
+      const VarNameMap& outputs, const paddle::framework::AttributeMap& attrs) \
+      : PARENT_CLS(type, inputs, outputs, attrs) {}
+
 class NOP : public OperatorBase {
  public:
   using OperatorBase::OperatorBase;
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {}
+  std::unique_ptr<OperatorBase> Clone() const override {
+    return std::unique_ptr<OperatorBase>(new NOP(*this));
+  }
 };
 
 // this class not only make proto but also init attribute checkers.
@@ -164,11 +184,8 @@ class OpProtoAndCheckerMaker {
       return *this;
     }
 
-    // TODO(FengJiayi, yuyang18): `AsNoGradient` is a very bad name, because it
-    // means that input/output is not needed when calculate gradient. It does
-    // not mean no gradient when backward. It should be changed soon.
-    VariableBuilder& AsNoGradient() {
-      var_->set_no_gradient(true);
+    VariableBuilder& NotInGradient() {
+      var_->set_not_in_gradient(true);
       return *this;
     }
   };

paddle/framework/operator_test.cc

@@ -245,3 +245,21 @@ TEST(OpKernel, multi_inputs) {
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   op->Run(scope, cpu_device_context);
 }
+
+class OperatorClone : public paddle::framework::OperatorBase {
+ public:
+  DEFINE_OP_CLONE_METHOD(OperatorClone);
+  OperatorClone(const std::string& type, const VarNameMap& inputs,
+                const VarNameMap& outputs,
+                const paddle::framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void InferShape(const paddle::framework::Scope& scope) const override {}
+  void Run(const paddle::framework::Scope& scope,
+           const paddle::platform::DeviceContext& dev_ctx) const override {}
+};
+
+TEST(Operator, Clone) {
+  OperatorClone a("ABC", {}, {}, {});
+  auto b = a.Clone();
+  ASSERT_EQ(a.Type(), b->Type());
+}

paddle/framework/pybind.cc

@@ -49,29 +49,6 @@ namespace framework {
 
 using Tensor = framework::Tensor;
 
-template <typename ClassType>
-void ExposeOperator(ClassType &m) {
-  m.def("infer_shape", &ClassType::type::InferShape)
-      .def("run", &ClassType::type::Run)
-      .def("type",
-           [](const typename ClassType::type &op) -> std::string {
-             return op.Type();
-           })
-      .def("outputs",
-           [](const typename ClassType::type &op)
-               -> std::map<std::string, std::vector<std::string>> {
-                 return op.Outputs();
-               })
-      .def("inputs",
-           [](const typename ClassType::type &op) { return op.Inputs(); })
-      .def("__str__", &ClassType::type::DebugString)
-      .def("no_intermediate_outputs",
-           [](const typename ClassType::type &op) {
-             return op.OutputVars(false);
-           })
-      .def("support_gpu", &ClassType::type::SupportGPU);
-}
-
 static size_t UniqueIntegerGenerator() {
   static std::atomic<size_t> generator;
   return generator.fetch_add(1);
@@ -208,75 +185,69 @@ All parameter, weight, gradient are variables in Paddle.
       .def(py::init<>())
       .def("__str__", string::to_string<const platform::CPUPlace &>);
 
-  py::class_<OperatorBase, std::shared_ptr<OperatorBase>> operator_base(
-      m, "Operator");
-
-  operator_base.def_static("create", [](py::bytes protobin) {
-    OpDesc desc;
-    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
-                   "Cannot parse user input to OpDesc");
-    PADDLE_ENFORCE(desc.IsInitialized(),
-                   "User OpDesc is not initialized, reason %s",
-                   desc.InitializationErrorString());
-    return OpRegistry::CreateOp(desc);
-  });
-
-  operator_base.def("backward",
-                    [](const OperatorBase &forwardOp,
-                       const std::unordered_set<std::string> &no_grad_vars) {
-                      return Backward(forwardOp, no_grad_vars);
-                    });
-
-  ExposeOperator(operator_base);
-
-  py::class_<operators::NetOp, std::shared_ptr<operators::NetOp>> net(m, "Net");
-
-  net.def_static("create",
-                 []() -> std::shared_ptr<operators::NetOp> {
-                   auto retv = std::make_shared<operators::NetOp>();
-                   retv->SetType("plain_net");
-                   return retv;
-                 })
-      .def("add_op", &operators::NetOp::AddOp)
-      .def("add_op",
-           [](operators::NetOp &self,
-              const std::shared_ptr<operators::NetOp> &net) -> void {
-             self.AddOp(std::static_pointer_cast<OperatorBase>(net));
-           })
-      .def("add_op",
-           [](operators::NetOp &self,
-              const std::shared_ptr<operators::RecurrentOp> &rnn) -> void {
-             self.AddOp(std::static_pointer_cast<OperatorBase>(rnn));
+  py::class_<OperatorBase>(m, "Operator")
+      .def_static("create",
+                  [](py::bytes protobin) {
+                    OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    return OpRegistry::CreateOp(desc);
+                  })
+      .def("backward",
+           [](const OperatorBase &forwardOp,
+              const std::unordered_set<std::string> &no_grad_vars) {
+             return Backward(forwardOp, no_grad_vars).release();
            })
+      .def("infer_shape", &OperatorBase::InferShape)
+      .def("run", &OperatorBase::Run)
+      .def("type",
+           [](const OperatorBase &op) -> std::string { return op.Type(); })
+      .def("outputs",
+           [](const OperatorBase &op)
+               -> std::map<std::string, std::vector<std::string>> {
+                 return op.Outputs();
+               })
+      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
+      .def("__str__", &OperatorBase::DebugString)
+      .def("no_intermediate_outputs",
+           [](const OperatorBase &op) { return op.OutputVars(false); })
+      .def("support_gpu", &OperatorBase::SupportGPU);
+
+  py::class_<operators::NetOp, OperatorBase>(m, "Net")
+      .def_static("create",
+                  []() -> operators::NetOp * {
+                    auto *retv = new operators::NetOp;
+                    retv->SetType("plain_net");
+                    return retv;
+                  })
+      .def("add_op", [](operators::NetOp &self,
+                        const OperatorBase &op) { self.AddOp(op); })
       .def("complete_add_op", &operators::NetOp::CompleteAddOp)
       .def("complete_add_op", [](std::shared_ptr<operators::NetOp> &self) {
         self->CompleteAddOp();
       });
 
-  ExposeOperator(net);
-
   // recurrent_op
-  py::class_<operators::RecurrentOp, std::shared_ptr<operators::RecurrentOp>>
-      rnn(m, "RecurrentOp");
-
-  rnn.def_static(
-         "create",
-         [](py::bytes protobin) -> std::shared_ptr<operators::RecurrentOp> {
-           OpDesc desc;
-           PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
-                          "Cannot parse user input to OpDesc");
-           PADDLE_ENFORCE(desc.IsInitialized(),
-                          "User OpDesc is not initialized, reason %s",
-                          desc.InitializationErrorString());
-           auto rnn_op = OpRegistry::CreateOp(desc);
-           return std::dynamic_pointer_cast<operators::RecurrentOp>(rnn_op);
-         })
-      .def("set_stepnet",
-           [](operators::RecurrentOp &self,
-              const std::shared_ptr<operators::NetOp> &net) -> void {
-             self.set_stepnet(net);
-           });
-  ExposeOperator(rnn);
+  py::class_<operators::RecurrentOp, OperatorBase>(m, "RecurrentOp")
+      .def_static(
+          "create",
+          [](py::bytes protobin) -> operators::RecurrentOp * {
+            OpDesc desc;
+            PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                           "Cannot parse user input to OpDesc");
+            PADDLE_ENFORCE(desc.IsInitialized(),
+                           "User OpDesc is not initialized, reason %s",
+                           desc.InitializationErrorString());
+            auto rnn_op = OpRegistry::CreateOp(desc);
+            return static_cast<operators::RecurrentOp *>(rnn_op.release());
+          })
+      .def("set_stepnet", [](operators::RecurrentOp &self,
+                             const operators::NetOp &net) -> void {
+        self.set_stepnet(net.Clone());
+      });
 
   m.def("unique_integer", UniqueIntegerGenerator);
 

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -57,11 +57,14 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap,
 }
 
 void MKLDNNFcLayer::convertWeightsFromPaddle() {
-  if (FLAGS_use_mkldnn_wgt) {
+  if (hasInitedWgt_) {
     return;
   }
 
-  if (hasInitedWgt_) {
+  // TODO(TJ): dst format should get from wgtVal_
+  int dstFmt = PARAM_FORMAT_MKLDNN_OI;
+  int srcFmt = weight_->getParameterPtr()->getHeaderFormat();
+  if (srcFmt == dstFmt) {
     return;
   }
 
@@ -78,6 +81,7 @@ void MKLDNNFcLayer::convertWeightsFromPaddle() {
   MatrixPtr paddleWgtT;
   paddleWgt->transpose(paddleWgtT, true);
   weight_->getW()->copyFrom(*paddleWgtT);
+  weight_->getParameterPtr()->setHeaderFormat(dstFmt);
   hasInitedWgt_ = true;
 }
 

paddle/gserver/tests/MKLDNNTester.cpp

@@ -330,9 +330,7 @@ void MKLDNNTester::run(const TestConfig& dnn,
   log_ = log;
   lvl_ = level;
 
-  // Firstly test FLAGS_use_mkldnn_wgt = false
-  FLAGS_use_mkldnn_wgt = false;
-  // reset and run once
+  // Firstly test mkldnn init from PARAM_FORMAT_ORIGINAL weight
   reset(dnn, ref, batchSize);
   randomWgtDatas();
   clearWgtDiffs();
@@ -342,17 +340,32 @@ void MKLDNNTester::run(const TestConfig& dnn,
     runOnce();
   }
 
-  // Then test FLAGS_use_mkldnn_wgt = true
-  FLAGS_use_mkldnn_wgt = true;
-  // after run once the mkldnn weight has been stored in dnnlayer
+  if (parameters_[DNN].empty()) {
+    // has no paramters
+    return;
+  }
+
+  // After run some iterations, the mkldnn weight has been stored in dnnLayer
+  // and we can also get the mkldnn weight parameter header format.
+  // Weight parameter should always be index 0 (and bias index 1).
+  // TODO(TJ): should also consider mean and var format when batchnorm ready
+  int dnnWgtFmt = parameters_[DNN][0]->getHeaderFormat();
+  int refWgtFmt = parameters_[REF][0]->getHeaderFormat();
+  if (dnnWgtFmt == refWgtFmt) {
+    // weight format are equal, so no need check more
+    return;
+  }
+
   // then save the weights and restart again
   vector<VectorPtr> dnnWgts, refWgts;
   CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size());
   saveWgt(parameters_[DNN], dnnWgts);
   saveWgt(parameters_[REF], refWgts);
 
-  // restart again with flag true
+  // restart again with dnn weight format
   reset(dnn, ref, batchSize);
+  // TODO(TJ): should also considerate mean and var format when batchnorm ready
+  parameters_[DNN][0]->setHeaderFormat(dnnWgtFmt);
 
   // restore wgt
   restoreWgt(dnnWgts, parameters_[DNN]);

paddle/gserver/tests/MKLDNNTester.h

@@ -108,7 +108,7 @@ private:
    * if many(>failRate) wrong(abs(dnn-ref)/abs(ref)>thres) points return the
    * max(diff/ref)
    * else return sum(abs(a-b)) / sum(abs(b))
-   * The return value should smaller than eps when passing.
+   * The return value should be smaller than eps when passing.
    */
   double getDelta(const real* d1,
                   const real* d2,

paddle/memory/detail/system_allocator.cc

@@ -27,7 +27,7 @@ limitations under the License. */
 // between host and device.  Allocates too much would reduce the amount
 // of memory available to the system for paging.  So, by default, we
 // should set false to use_pinned_memory.
-DEFINE_bool(use_pinned_memory, false, "If set, allocate cpu pinned memory.");
+DEFINE_bool(use_pinned_memory, true, "If set, allocate cpu pinned memory.");
 
 namespace paddle {
 namespace memory {

paddle/memory/memory.cc

@@ -13,22 +13,38 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/memory/memory.h"
+
+#include <algorithm>  // for transform
+#include <cstring>    // for memcpy
+#include <memory>     // for unique_ptr
+#include <mutex>      // for call_once
+
+#include "glog/logging.h"
+
 #include "paddle/memory/detail/buddy_allocator.h"
 #include "paddle/memory/detail/system_allocator.h"
+#include "paddle/platform/gpu_info.h"
 
-#include <cstring>  // for memcpy
+DECLARE_double(fraction_of_gpu_memory_to_use);
 
 namespace paddle {
 namespace memory {
 
-detail::BuddyAllocator* GetCPUBuddyAllocator() {
-  static detail::BuddyAllocator* a = nullptr;
-  if (a == nullptr) {
-    a = new detail::BuddyAllocator(new detail::CPUAllocator,
-                                   platform::CpuMinChunkSize(),
-                                   platform::CpuMaxChunkSize());
-  }
-  return a;
+using BuddyAllocator = detail::BuddyAllocator;
+
+std::once_flag cpu_allocator_flag;
+std::once_flag gpu_allocator_flag;
+
+BuddyAllocator* GetCPUBuddyAllocator() {
+  static std::unique_ptr<BuddyAllocator> a{nullptr};
+
+  std::call_once(cpu_allocator_flag, [&]() {
+    a.reset(new BuddyAllocator(new detail::CPUAllocator,
+                               platform::CpuMinChunkSize(),
+                               platform::CpuMaxChunkSize()));
+  });
+
+  return a.get();
 }
 
 template <>
@@ -48,20 +64,36 @@ size_t Used<platform::CPUPlace>(platform::CPUPlace place) {
 
 #ifndef PADDLE_ONLY_CPU
 
-detail::BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
-  static detail::BuddyAllocator** as = NULL;
-  if (as == NULL) {
+BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
+  using BuddyAllocVec = std::vector<BuddyAllocator*>;
+  static std::unique_ptr<BuddyAllocVec, void (*)(BuddyAllocVec * p)> as{
+      new BuddyAllocVec, [](BuddyAllocVec* p) {
+        std::for_each(p->begin(), p->end(),
+                      [](BuddyAllocator* p) { delete p; });
+      }};
+
+  // GPU buddy allocators
+  auto& allocators = *as.get();
+
+  // GPU buddy allocator initialization
+  std::call_once(gpu_allocator_flag, [&]() {
     int gpu_num = platform::GetDeviceCount();
-    as = new detail::BuddyAllocator*[gpu_num];
+    allocators.reserve(gpu_num);
     for (int gpu = 0; gpu < gpu_num; gpu++) {
       platform::SetDeviceId(gpu);
-      as[gpu] = new detail::BuddyAllocator(new detail::GPUAllocator,
-                                           platform::GpuMinChunkSize(),
-                                           platform::GpuMaxChunkSize());
+      allocators.emplace_back(new BuddyAllocator(new detail::GPUAllocator,
+                                                 platform::GpuMinChunkSize(),
+                                                 platform::GpuMaxChunkSize()));
     }
-  }
+    VLOG(3) << "\n\nNOTE: each GPU device use "
+            << FLAGS_fraction_of_gpu_memory_to_use * 100 << "% of GPU memory.\n"
+            << "You can set environment variable '"
+            << platform::kEnvFractionGpuMemoryToUse
+            << "' to change the fraction of GPU usage.\n\n";
+  });
+
   platform::SetDeviceId(gpu_id);
-  return as[gpu_id];
+  return allocators[gpu_id];
 }
 
 template <>

paddle/memory/memory.h

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/platform/gpu_info.h"
 #include "paddle/platform/place.h"
 
 namespace paddle {

paddle/operators/gather_test.cc

@@ -45,4 +45,8 @@ TEST(Gather, GatherData) {
 
   for (int i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4);
   for (int i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4);
+
+  delete src;
+  delete index;
+  delete output;
 }

paddle/operators/mean_op.cc

@@ -34,7 +34,7 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
   MeanOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input of mean op");
-    AddOutput("Out", "The output of mean op").AsNoGradient();
+    AddOutput("Out", "The output of mean op").NotInGradient();
     AddComment("Mean Operator");
   }
 };

paddle/operators/mean_op.h

@@ -55,9 +55,10 @@ class MeanGradKernel : public framework::OpKernel {
     IG->mutable_data<T>(context.GetPlace());
 
     T ig_size = (T)framework::product(IG->dims());
+    Eigen::DSizes<int, 1> bcast(ig_size);
 
     EigenVector<T>::Flatten(*IG).device(context.GetEigenDevice<Place>()) =
-        EigenScalar<T>::From(*OG) / ig_size;
+        (EigenVector<T>::From(*OG) / ig_size).broadcast(bcast);
   }
 };
 

paddle/operators/net_op.cc

@@ -85,7 +85,14 @@ NetOp::NetOp(const std::string& type,
              const framework::OperatorBase::VarNameMap& inputs,
              const framework::OperatorBase::VarNameMap& outputs,
              const framework::AttributeMap& attrs)
-    : OperatorBase(type, inputs, outputs, attrs) {}
+    : framework::OperatorBase(type, inputs, outputs, attrs) {}
+
+std::unique_ptr<framework::OperatorBase> NetOp::Clone() const {
+  PADDLE_ENFORCE(
+      add_op_done_,
+      "Must clone a sealed NetOp, invoke Net::CompleteAddOp before clone");
+  return std::unique_ptr<OperatorBase>(new NetOp(*this));
+}
 
 }  // namespace operators
 }  // namespace paddle

paddle/operators/net_op.h

@@ -41,6 +41,16 @@ class NetOp : public framework::OperatorBase {
   NetOp(const std::string& type, const VarNameMap& inputs,
         const VarNameMap& outputs, const framework::AttributeMap& attrs);
 
+  NetOp(const NetOp& o) : framework::OperatorBase(o.type_, {}, {}, o.attrs_) {
+    this->ops_.reserve(o.ops_.size());
+    std::transform(
+        o.ops_.begin(), o.ops_.end(), std::back_inserter(this->ops_),
+        [](const std::unique_ptr<framework::OperatorBase>& op) {
+          return std::unique_ptr<framework::OperatorBase>(op->Clone());
+        });
+    this->CompleteAddOp();
+  }
+
   /**
    * Infer all the operators' input and output variables' shapes, will be called
    * before every mini-batch
@@ -74,21 +84,27 @@ class NetOp : public framework::OperatorBase {
     return true;
   }
 
+  void AddOp(const framework::OperatorBase& op) { AddOp(op.Clone()); }
+
   /**
    * @brief Add an operator by ptr
    */
-  void AddOp(const std::shared_ptr<OperatorBase>& op) {
+  void AddOp(std::unique_ptr<framework::OperatorBase> op) {
     PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed");
     PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
-    ops_.push_back(op);
+    ops_.push_back(std::move(op));
   }
 
-  void InsertOp(size_t pos, const std::shared_ptr<OperatorBase>& op) {
+  void InsertOp(size_t pos, std::unique_ptr<framework::OperatorBase> op) {
     PADDLE_ENFORCE(!add_op_done_,
                    "Cannot InsertOp when this network is sealed");
     PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
     PADDLE_ENFORCE_LE(pos, ops_.size(), "Out of range");
-    ops_.insert(ops_.begin() + pos, op);
+    ops_.insert(ops_.begin() + pos, std::move(op));
+  }
+
+  void InsertOp(size_t pos, const framework::OperatorBase& op) {
+    InsertOp(pos, op.Clone());
   }
 
   void CompleteAddOp(bool calculate = true);
@@ -98,7 +114,9 @@ class NetOp : public framework::OperatorBase {
   bool IsNetOp() const override;
   std::vector<std::string> OutputVars(bool has_intermediate) const override;
 
-  std::vector<std::shared_ptr<OperatorBase>> ops_;
+  std::unique_ptr<framework::OperatorBase> Clone() const override;
+
+  std::vector<std::unique_ptr<framework::OperatorBase>> ops_;
 
  private:
   bool add_op_done_{false};

paddle/operators/net_op_test.cc

@@ -13,6 +13,7 @@ static int run_cnt = 0;
 class TestOp : public framework::OperatorBase {
  public:
   using framework::OperatorBase::OperatorBase;
+  DEFINE_OP_CLONE_METHOD(TestOp);
   void InferShape(const Scope& scope) const override { ++infer_shape_cnt; }
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {
@@ -37,15 +38,12 @@ TEST(OpKernel, all) {
   auto net = std::make_shared<NetOp>();
   ASSERT_NE(net, nullptr);
 
-  auto op1 = std::shared_ptr<TestOp>(
+  net->AddOp(std::unique_ptr<TestOp>(
       new TestOp("test", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
-                 {{"Out", {"y"}}}, {}));
-  net->AddOp(op1);
-
-  auto op2 = std::shared_ptr<TestOp>(
+                 {{"Out", {"y"}}}, {})));
+  net->AddOp(std::unique_ptr<TestOp>(
       new TestOp("test", {{"X", {"y"}}, {"W", {"w2"}}, {"b", {"b2"}}},
-                 {{"Out", {"z"}}}, {}));
-  net->AddOp(op2);
+                 {{"Out", {"z"}}}, {})));
 
   net->CompleteAddOp();
   AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"},
@@ -60,15 +58,31 @@ TEST(OpKernel, all) {
 
 TEST(NetOp, insert_op) {
   NetOp net;
-  auto op1 = std::shared_ptr<framework::NOP>(
+  auto op1 = std::unique_ptr<framework::NOP>(
       new framework::NOP("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
                          {{"Out", {"y"}}}, {}));
-  net.AddOp(op1);
-  net.InsertOp(0, op1);
+  net.AddOp(*op1);
+  net.InsertOp(0, *op1);
   ASSERT_EQ(2UL, net.ops_.size());
-  net.InsertOp(2, op1);
+  net.InsertOp(2, std::move(op1));
   ASSERT_EQ(3UL, net.ops_.size());
 }
 
+TEST(NetOp, Clone) {
+  NetOp net;
+  net.AddOp(
+      std::unique_ptr<framework::NOP>(new framework::NOP{"empty", {}, {}, {}}));
+  net.AddOp(std::unique_ptr<framework::NOP>(
+      new framework::NOP{"empty2", {}, {}, {}}));
+  net.CompleteAddOp(true);
+  auto new_net_op = net.Clone();
+  ASSERT_NE(new_net_op, nullptr);
+  ASSERT_TRUE(new_net_op->IsNetOp());
+  auto* new_net = static_cast<NetOp*>(new_net_op.get());
+  ASSERT_EQ(2, new_net->ops_.size());
+  ASSERT_EQ(new_net->ops_[0]->Type(), "empty");
+  ASSERT_EQ(new_net->ops_[1]->Type(), "empty2");
+}
+
 }  // namespace operators
 }  // namespace paddle