diff --git a/paddle/framework/attribute.cc b/paddle/framework/attribute.cc
index 4c5790693b7e48396e945d09f4fdc72b86aa5978..9eb07acdff1d00dd926f1cee9c24f9f151006d7e 100644
--- a/paddle/framework/attribute.cc
+++ b/paddle/framework/attribute.cc
@@ -44,7 +44,7 @@ AttrType AttrTypeID<std::vector<std::string>>() {
   return STRINGS;
 }
 
-Attribute GetAttrValue(const AttrDesc& attr_desc) {
+Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
   switch (attr_desc.type()) {
     case paddle::framework::AttrType::INT: {
       return attr_desc.i();
diff --git a/paddle/framework/attribute.h b/paddle/framework/attribute.h
index 3a5820e9c60539e3c771df5da4e82f6c1cae688f..d0419f07ba4dca32236128a2fa7afdc1460f9468 100644
--- a/paddle/framework/attribute.h
+++ b/paddle/framework/attribute.h
@@ -21,8 +21,7 @@ limitations under the License. */
 #include <unordered_set>
 #include <vector>
 
-#include "paddle/framework/attribute.pb.h"
-#include "paddle/framework/op_desc.pb.h"
+#include "paddle/framework/framework.pb.h"
 #include "paddle/platform/enforce.h"
 
 namespace paddle {
@@ -37,7 +36,7 @@ typedef std::unordered_map<std::string, Attribute> AttributeMap;
 template <typename T>
 AttrType AttrTypeID();
 
-Attribute GetAttrValue(const AttrDesc& attr_desc);
+Attribute GetAttrValue(const OpDesc::Attr& attr_desc);
 
 // check whether a value(attribute) fit a certain limit
 template <typename T>
diff --git a/paddle/framework/ddim.cc b/paddle/framework/ddim.cc
index 545c1dcc2a1682839d90194002fdbb748d85e808..0b76a4fdb701ea7435d0da865e928b8da15335c9 100644
--- a/paddle/framework/ddim.cc
+++ b/paddle/framework/ddim.cc
@@ -284,5 +284,11 @@ DDim::DDim(std::initializer_list<int> init_list) {
   *this = make_ddim(init_list);
 }
 
+std::string DDim::DebugString() const {
+  std::ostringstream ss;
+  ss << *this;
+  return ss.str();
+}
+
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/ddim.h b/paddle/framework/ddim.h
index 5aa5af0c19be5a209c760282cb1a090fc57a53ad..3ea3b499e5b1015a5b0b107850ac0608735ba729 100644
--- a/paddle/framework/ddim.h
+++ b/paddle/framework/ddim.h
@@ -73,6 +73,8 @@ struct DDim {
   DDim operator*(DDim d) const;
 
   ssize_t size() const;
+
+  std::string DebugString() const;
 };
 
 /**
diff --git a/paddle/framework/framework.proto b/paddle/framework/framework.proto
index 4b6dfec5cbc35c8cfc20c525f965cbec292dc124..490d7bd91bfbd27010c7007b52e713171a8904ac 100644
--- a/paddle/framework/framework.proto
+++ b/paddle/framework/framework.proto
@@ -40,8 +40,8 @@ message OpDesc {
   };
 
   message Var {
-    required string name;                 // e.g. "X"
-    optional int dup = 2 [ default = 0 ]; // e.g., "1"
+    required string op_proto_name = 1;
+    repeated string var_names = 2;
   };
 
   required string type = 3;
@@ -57,7 +57,7 @@ message OpProto {
   message Var {
     required string name = 1;
     required string comment = 2;
-    // OpDesc::Var::dup indices the duplica.
+
     optional bool duplicable = 3 [ default = false ];
     optional bool intermediate = 4 [ default = false ];
     optional bool no_gradient = 5 [ default = false ];
diff --git a/paddle/framework/grad_op_builder.cc b/paddle/framework/grad_op_builder.cc
index 6d032fb78f099f5142d64e531d1a03c10ed5e68e..da9613e7763692d5d63eb7eecd350f0b211dc1bc 100644
--- a/paddle/framework/grad_op_builder.cc
+++ b/paddle/framework/grad_op_builder.cc
@@ -13,12 +13,12 @@ express or implied. See the License for the specific language governing
 permissions and limitations under the License. */
 
 #include "paddle/framework/grad_op_builder.h"
-#include "paddle/framework/op_proto.pb.h"
+#include "paddle/framework/framework.pb.h"
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
 namespace framework {
-
+/**
 class OpRegistry;
 
 using VarIndexMap = std::unordered_map<std::string, int>;
@@ -98,6 +98,7 @@ OperatorBase* BuildGradOp(const OperatorBase* op) {
   TransOpArg(op, grad_op, OpArgType::IN, OpArgType::OUT, out_idx, true);  // IG
   return grad_op;
 }
-
+**/
+OperatorBase* BuildGradOp(const OperatorBase* op) { return nullptr; }
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h
index b2813da83d9e4c525e66bb1f79b28769627eaec2..9123e9b56fd068d7df59160e8987bc51bbf511df 100644
--- a/paddle/framework/op_registry.h
+++ b/paddle/framework/op_registry.h
@@ -20,8 +20,8 @@ limitations under the License. */
 #include <unordered_map>
 #include <unordered_set>
 #include "paddle/framework/attribute.h"
+#include "paddle/framework/framework.pb.h"
 #include "paddle/framework/grad_op_builder.h"
-#include "paddle/framework/op_desc.pb.h"
 #include "paddle/framework/scope.h"
 
 namespace paddle {
@@ -44,25 +44,20 @@ class OpProtoAndCheckerMaker {
 
  protected:
   struct VariableBuilder {
-    VarProto* var_;
-    std::function<void()> on_multiple_;
-    std::function<void()> on_temporary_;
+    OpProto::Var* var_;
 
     VariableBuilder& SetMultiple() {
-      var_->set_multiple(true);
-      on_multiple_();
+      var_->set_duplicable(true);
       return *this;
     }
 
     VariableBuilder& SetTemporary() {
-      PADDLE_ENFORCE(bool(on_temporary_), "Cannot set temporary");
-      var_->set_temporary(true);
-      on_temporary_();
+      var_->set_intermediate(true);
       return *this;
     }
 
     VariableBuilder& IgnoreGradient() {
-      var_->set_ignore_gradient(true);
+      var_->set_no_gradient(true);
       return *this;
     }
   };
@@ -72,8 +67,7 @@ class OpProtoAndCheckerMaker {
     auto input = proto_->mutable_inputs()->Add();
     *input->mutable_name() = name;
     *input->mutable_comment() = comment;
-    return VariableBuilder{input, [=] { this->SetHasMultipleInput(); },
-                           nullptr};
+    return VariableBuilder{input};
   }
 
   VariableBuilder AddOutput(const std::string& name,
@@ -81,8 +75,7 @@ class OpProtoAndCheckerMaker {
     auto output = proto_->mutable_outputs()->Add();
     *output->mutable_name() = name;
     *output->mutable_comment() = comment;
-    return VariableBuilder{output, [=] { this->SetHasMultipleOutput(); },
-                           [=] { this->SetHasTemporaryOutput(); }};
+    return VariableBuilder{output};
   }
 
   template <typename T>
@@ -102,53 +95,6 @@ class OpProtoAndCheckerMaker {
   }
 
  private:
-  void SetHasMultiple(const std::string& in_out, bool* flag) {
-    if (!*flag) {
-      AddAttr<std::vector<int>>(in_out + "_format",
-                                "The multiple index of " + in_out +
-                                    "\n"
-                                    R"DOC(
-This attribute is used by Paddle core framework. Paddle's Op support each input
-or output could be a list of variable. This attribute is used to show how that
-list organized.
-
-e.g.
-  input = ["a", "b", "c", "d", "e", "f"]
-  input_format = [0, 4, 5, 6]
-
-means
-  The number of all input variables this op is six, and they are segmented into
-  three inputs.
-
-  The first input is input[0:4], second is input[4:5], third is input[5:6].
-)DOC",
-                                /*generated*/ true);
-      *flag = true;
-    }
-  }
-
-  void SetHasMultipleInput() { SetHasMultiple("input", &has_multiple_input_); }
-  void SetHasMultipleOutput() {
-    SetHasMultiple("output", &has_multiple_output_);
-  }
-
-  void SetHasTemporaryOutput() {
-    if (!has_temporary_output_) {
-      AddAttr<std::vector<int>>("temporary_index",
-                                R"DOC(The temporary index of output.
-
-Not all output of Paddle Op is used by user. For faster computation, each op
-could output some its internal state to other op, other op could take that
-output to make compute faster.
-
-Add a mark to which output is temporary is helpful for future optimization.
-)DOC",
-                                /*generated*/ true)
-          .SetDefault(std::vector<int>());
-      has_temporary_output_ = true;
-    }
-  }
-
   void CheckNoDuplicatedInOutAttrs() {
     std::unordered_set<std::string> names;
     auto checker = [&](const std::string& name) {
@@ -169,15 +115,12 @@ Add a mark to which output is temporary is helpful for future optimization.
   OpProto* proto_;
   OpAttrChecker* op_checker_;
   bool validated_{false};
-  bool has_multiple_input_{false};
-  bool has_multiple_output_{false};
-  bool has_temporary_output_{false};
 };
 
 class OpRegistry {
   using OpCreator = std::function<OperatorBase*()>;
   using VarIndexMap = std::unordered_map<std::string, int>;
-  using VarNameList = std::vector<std::string>;
+  using VarNameMap = std::unordered_map<std::string, std::vector<std::string>>;
 
  public:
   template <typename OpType, typename ProtoMakerType>
@@ -213,8 +156,8 @@ class OpRegistry {
   }
 
   static std::shared_ptr<OperatorBase> CreateOp(const std::string& type,
-                                                const VarNameList& inputs,
-                                                const VarNameList& outputs,
+                                                const VarNameMap& inputs,
+                                                const VarNameMap& outputs,
                                                 const AttributeMap& attrs) {
     auto op_create_it = op_creators().find(type);
     PADDLE_ENFORCE(op_create_it != op_creators().end(),
@@ -230,27 +173,28 @@ class OpRegistry {
 
     GenerateTempVariableName(op);
 
-    {
-      auto var_index_it = VarIndexMaps().find(type);
-      if (var_index_it != VarIndexMaps().end()) {
-        op->in_out_idxs_ = var_index_it->second;
-      }
-    }
-
     op->Init();
     return std::shared_ptr<OperatorBase>(op);
   }
 
   static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc) {
-    std::vector<std::string> inputs;
-    inputs.reserve((size_t)op_desc.inputs_size());
-    std::copy(op_desc.inputs().begin(), op_desc.inputs().end(),
-              std::back_inserter(inputs));
+    VarNameMap inputs;
+    for (auto& input : op_desc.inputs()) {
+      auto& var_names = inputs[input.op_proto_name()];
+      auto& var_names_in_proto = input.var_names();
+      var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
+      std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
+                std::back_inserter(var_names));
+    }
 
-    std::vector<std::string> outputs;
-    outputs.reserve((size_t)op_desc.outputs_size());
-    std::copy(op_desc.outputs().begin(), op_desc.outputs().end(),
-              std::back_inserter(outputs));
+    VarNameMap outputs;
+    for (auto& output : op_desc.outputs()) {
+      auto& var_names = outputs[output.op_proto_name()];
+      auto& var_names_in_proto = output.var_names();
+      var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
+      std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
+                std::back_inserter(var_names));
+    }
 
     AttributeMap attrs;
     for (auto& attr : op_desc.attrs()) {
@@ -303,11 +247,13 @@ class OpRegistry {
 
   static void GenerateTempVariableName(OperatorBase* op) {
     static std::atomic<size_t> gUniqId(0UL);
-    for (auto& outname : op->outputs_) {
-      if (outname == kTempVarName) {
-        outname += op->type_;
-        outname += "@";
-        outname += std::to_string(gUniqId.fetch_add(1));
+    for (auto& output : op->outputs_) {
+      for (auto& output_name : output.second) {
+        if (output_name == kTempVarName) {
+          output_name += op->type_;
+          output_name += "@";
+          output_name += std::to_string(gUniqId.fetch_add(1));
+        }
       }
     }
   }
diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc
index beb6793289812cfaa6991d28379126ff29fa2547..e69db305b49b8bc77e16b0d5d9f24a25b8a28c29 100644
--- a/paddle/framework/operator.cc
+++ b/paddle/framework/operator.cc
@@ -34,83 +34,72 @@ ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
 #endif
 
 const std::string& OperatorBase::Input(const std::string& name) const {
-  PADDLE_ENFORCE(in_out_idxs_ != nullptr,
-                 "Input Output Indices could not be nullptr");
-  auto it = in_out_idxs_->find(name);
-  PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_",
+  auto it = inputs_.find(name);
+  PADDLE_ENFORCE(it != inputs_.end(), "Op %s does not have output %s", type_,
                  name);
-  if (attrs_.count("input_format") == 0) {
-    return inputs_.at((size_t)it->second);
-  } else {
-    const auto& input_format = GetAttr<std::vector<int>>("input_format");
-    int idx = input_format[it->second];
-    return inputs_.at((size_t)idx);
-  }
+  PADDLE_ENFORCE_EQ(it->second.size(), 1UL,
+                    "Op %s input %s should contain only one variable", type_,
+                    name);
+  return it->second[0];
 }
 
-std::vector<std::string> OperatorBase::Inputs(const std::string& name) const {
-  PADDLE_ENFORCE(in_out_idxs_ != nullptr, "IO Idx could not be nullptr");
-  auto input_format = GetAttr<std::vector<int>>("input_format");
-  auto offset = in_out_idxs_->at(name);
-  PADDLE_ENFORCE(input_format.at(static_cast<size_t>(offset) + 1) <=
-                     static_cast<int>(inputs_.size()),
-                 "Input Out Of Range");
-
-  return std::vector<std::string>{
-      inputs_.begin() + input_format.at(offset),
-      inputs_.begin() + input_format.at(offset + 1)};
+const std::vector<std::string>& OperatorBase::Inputs(
+    const std::string& name) const {
+  return inputs_.at(name);
 }
 
 const std::string& OperatorBase::Output(const std::string& name) const {
-  PADDLE_ENFORCE(in_out_idxs_ != nullptr, "InOut Indice could not be nullptr");
-  auto it = in_out_idxs_->find(name);
-  PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_",
+  auto it = outputs_.find(name);
+  PADDLE_ENFORCE(it != outputs_.end(), "Op %s does not have output %s", type_,
                  name);
-  if (attrs_.count("output_format") == 0) {
-    return outputs_.at((size_t)it->second);
-  } else {
-    const auto& output_format = GetAttr<std::vector<int>>("output_format");
-    int idx = output_format[it->second];
-    return outputs_.at((size_t)idx);
-  }
+  PADDLE_ENFORCE_EQ(it->second.size(), 1UL,
+                    "Op %s input %s should contain only one variable", type_,
+                    name);
+  return it->second[0];
 }
 
-std::vector<std::string> OperatorBase::Outputs(const std::string& name) const {
-  PADDLE_ENFORCE(in_out_idxs_ != nullptr, "InOut Indice could not be nullptr");
-  auto output_format = GetAttr<std::vector<int>>("output_format");
-  auto offset = in_out_idxs_->at(name);
-  PADDLE_ENFORCE(output_format.at(static_cast<size_t>(offset) + 1) <=
-                     static_cast<int>(outputs_.size()),
-                 "Output Out of Range");
-  return std::vector<std::string>{
-      outputs_.begin() + output_format.at(offset),
-      outputs_.begin() + output_format.at(offset + 1)};
+const std::vector<std::string>& OperatorBase::Outputs(
+    const std::string& name) const {
+  return outputs_.at(name);
 }
 
 std::string OperatorBase::DebugString() const {
   std::stringstream ss;
-  ss << "Op(" << type_ << "), inputs:(";
-  for (size_t i = 0; i < inputs_.size(); ++i) {
-    ss << inputs_[i];
-    if (i != inputs_.size() - 1) {
-      ss << ", ";
+  ss << "Op(" << type_ << "), inputs:{";
+  for (auto& input : inputs_) {
+    ss << input.first << "[";
+    for (size_t i = 0; i < input.second.size(); ++i) {
+      ss << input.second[i];
+      if (i != input.second.size() - 1) {
+        ss << ", ";
+      }
     }
+    ss << "]";
   }
-  ss << "), outputs:(";
-  for (size_t i = 0; i < outputs_.size(); ++i) {
-    ss << outputs_[i];
-    if (i != outputs_.size() - 1) {
-      ss << ", ";
+  ss << "}, outputs:{";
+  for (auto& output : outputs_) {
+    ss << output.first << "[";
+    for (size_t i = 0; i < output.second.size(); ++i) {
+      ss << output.second[i];
+      if (i != output.second.size() - 1) {
+        ss << ", ";
+      }
     }
+    ss << "]";
   }
-  ss << ").";
+  ss << "}.";
   return ss.str();
 }
 
 void OperatorBase::Rename(const std::string& old_name,
                           const std::string& new_name) {
-  std::replace(inputs_.begin(), inputs_.end(), old_name, new_name);
-  std::replace(outputs_.begin(), outputs_.end(), old_name, new_name);
+  for (auto& input : inputs_) {
+    std::replace(input.second.begin(), input.second.end(), old_name, new_name);
+  }
+  for (auto& output : outputs_) {
+    std::replace(output.second.begin(), output.second.end(), old_name,
+                 new_name);
+  }
 }
 
 }  // namespace framework
diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h
index 9672492d1c2c6a82c37e0a840a4ca9c111de06d8..ec498ce3bd8b5d40c3c3ad081c217c4c7e8dd593 100644
--- a/paddle/framework/operator.h
+++ b/paddle/framework/operator.h
@@ -21,8 +21,7 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/framework/attribute.h"
-#include "paddle/framework/op_desc.pb.h"
-#include "paddle/framework/op_proto.pb.h"
+#include "paddle/framework/framework.pb.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/device_context.h"
@@ -95,13 +94,12 @@ class OperatorBase {
   const std::string& Input(const std::string& name) const;
 
   //! Get a input which has multiple variables.
-  //! TODO add a vector_view to prevent memory copy.
-  std::vector<std::string> Inputs(const std::string& name) const;
+  const std::vector<std::string>& Inputs(const std::string& name) const;
   //! Get a output with argument's name described in `op_proto`
   const std::string& Output(const std::string& name) const;
   //! Get an output which has multiple variables.
   //! TODO add a vector_view to prevent memory copy.
-  std::vector<std::string> Outputs(const std::string& name) const;
+  const std::vector<std::string>& Outputs(const std::string& name) const;
 
  public:
   std::string type_;
@@ -109,13 +107,12 @@ class OperatorBase {
   // I (Inputs)
   // O (Outputs)
   // OG (Output Gradients)
-  std::vector<std::string> inputs_;
+  std::unordered_map<std::string, std::vector<std::string>> inputs_;
+
   // NOTE: in case of OpGrad, outputs_ contains
   // IG (Inputs Gradients)
-  std::vector<std::string> outputs_;
+  std::unordered_map<std::string, std::vector<std::string>> outputs_;
   AttributeMap attrs_;
-  // store the arguments' offset described in op_desc.
-  std::shared_ptr<std::unordered_map<std::string, int>> in_out_idxs_;
 };
 
 class OperatorContext {
@@ -123,16 +120,12 @@ class OperatorContext {
   OperatorContext(const OperatorBase* op, const Scope& scope)
       : op_(*op), scope_(scope) {}
 
-  size_t InputSize() const { return op_.inputs_.size(); }
-
-  size_t OutputSize() const { return op_.outputs_.size(); }
-
-  const Variable* InputVar(const size_t index) const {
-    return scope_.FindVar(op_.inputs_.at(index));
+  size_t InputSize(const std::string& name) const {
+    return op_.inputs_.at(name).size();
   }
 
-  Variable* OutputVar(const size_t index) const {
-    return scope_.FindVar(op_.outputs_.at(index));
+  size_t OutputSize(const std::string& name) const {
+    return op_.outputs_.at(name).size();
   }
 
   const Variable* InputVar(const std::string& name) const {
@@ -164,24 +157,6 @@ class OperatorContext {
     return res;
   }
 
-  template <typename T>
-  const T* Input(const size_t index) const {
-    auto var = InputVar(index);
-    PADDLE_ENFORCE(var != nullptr, "Input(%d) should not be nullptr", index);
-    return &var->Get<T>();
-  }
-
-  template <typename T>
-  T* Output(const size_t index) const {
-    auto var = OutputVar(index);
-    PADDLE_ENFORCE(
-        var != nullptr,
-        "Output(%d) not be nullptr, which means variable [%s] does not "
-        "exist in scope",
-        index, op_.outputs_[index]);
-    return var->GetMutable<T>();
-  }
-
   template <typename T>
   const T* Input(const std::string& name) const {
     auto var = InputVar(name);
diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc
index d4c05ed483ca56a31dd8ee4d81b54551ae6da0d1..29943002acab6db42dfdad2bc32f4c8725901136 100644
--- a/paddle/operators/add_op.cc
+++ b/paddle/operators/add_op.cc
@@ -20,15 +20,10 @@ namespace operators {
 class AddOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_EQ(ctx.InputSize(), 2);
-    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1);
-    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr && ctx.InputVar(1) != nullptr,
-                   "Inputs of AddOp must all be set");
-    PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr,
-                   "Outputs of AddOp must all be set");
-    PADDLE_ENFORCE(ctx.Input<Tensor>(0)->dims() == ctx.Input<Tensor>(1)->dims(),
-                   "Two input of Add Op's dimension must be same.");
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
+                      ctx.Input<Tensor>("Y")->dims(),
+                      "Two input of Add Op's dimension must be same.");
+    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
 
diff --git a/paddle/operators/add_op.h b/paddle/operators/add_op.h
index 9db19a61381fdb11350276d51d3ebbf083672022..9310c1f7edfd2059e5f31486388a17a04359be63 100644
--- a/paddle/operators/add_op.h
+++ b/paddle/operators/add_op.h
@@ -22,9 +22,9 @@ template <typename Place, typename T>
 class AddKernel : public OpKernel {
  public:
   void Compute(const ExecutionContext& context) const override {
-    auto input0 = context.Input<Tensor>(0);
-    auto input1 = context.Input<Tensor>(1);
-    auto output = context.Output<Tensor>(0);
+    auto* input0 = context.Input<Tensor>("X");
+    auto* input1 = context.Input<Tensor>("Y");
+    auto* output = context.Output<Tensor>("Out");
 
     output->mutable_data<T>(context.GetPlace());
 
diff --git a/paddle/operators/cross_entropy_op.cc b/paddle/operators/cross_entropy_op.cc
index b0e1b8e41a5320aa14e316a56dbfd01e43c6816b..77c8271fd4ca490afba11abc109f5c4296f0d1fd 100644
--- a/paddle/operators/cross_entropy_op.cc
+++ b/paddle/operators/cross_entropy_op.cc
@@ -20,19 +20,13 @@ namespace operators {
 class OnehotCrossEntropyOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 2,
-                   "Input size of OnehotCrossEntropyOp must be two");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1,
-                   "Output size of OnehotCrossEntropyOp must be one");
-    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr && ctx.InputVar(1) != nullptr,
-                   "Inputs of OnehotCrossEntropyOp must all be set");
-    PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr,
-                   "Outputs of OnehotCrossEntropyOp must all be set");
-    PADDLE_ENFORCE(ctx.Input<Tensor>(0)->dims().size() == 2,
-                   "X's dimension must be 2.");
-    PADDLE_ENFORCE(ctx.Output<Tensor>(0)->dims().size() == 1,
-                   "label's dimension must be 1.");
-    ctx.Output<Tensor>(0)->Resize({ctx.Input<Tensor>(0)->dims()[0]});
+    auto *X = ctx.Input<Tensor>("X");
+    auto *label = ctx.Input<Tensor>("label");
+
+    PADDLE_ENFORCE_EQ(X->dims().size(), 2, "X's dimension must be 2.");
+    PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label's dimension must be 1.");
+    PADDLE_ENFORCE_EQ(X->dims()[0], label->dims()[0]);
+    ctx.Output<Tensor>("Y")->Resize({X->dims()[0]});
   }
 };
 
diff --git a/paddle/operators/cross_entropy_op.h b/paddle/operators/cross_entropy_op.h
index e02e3e2945af13fe283f95f7faa03b2a76d06125..d5e3f29332809a63908ecc896b33d2adff6abe45 100644
--- a/paddle/operators/cross_entropy_op.h
+++ b/paddle/operators/cross_entropy_op.h
@@ -43,7 +43,7 @@ class OnehotCrossEntropyOpKernel : public OpKernel {
   void Compute(const ExecutionContext& ctx) const override {
     auto X = ctx.Input<Tensor>("X");
     const T* Xdata = X->data<T>();
-    const int* label_data = ctx.Input<Tensor>(1)->data<int>();
+    const int* label_data = ctx.Input<Tensor>("label")->data<int>();
     auto Y = ctx.Output<Tensor>("Y");
 
     Y->mutable_data<T>(ctx.GetPlace());
diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc
index 198b4576c887122fdac1a3fbff5a248a3d9fa0a3..405ed219f0145ec1c3d8a315f39a134d3d9d7643 100644
--- a/paddle/operators/fill_zeros_like_op.cc
+++ b/paddle/operators/fill_zeros_like_op.cc
@@ -20,16 +20,8 @@ namespace operators {
 class FillZerosLikeOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 1UL,
-                   "Input size of FillZerosLikeOp must be one.");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
-                   "Output size of AddOp must be one.");
-    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr,
-                   "Input of FillZerosLikeOp must be set.");
-    PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr,
-                   "Output of FillZerosLikeOp must be set.");
-    ctx.Output<framework::Tensor>(0)->Resize(
-        ctx.Input<framework::Tensor>(0)->dims());
+    ctx.Output<framework::Tensor>("Dst")->Resize(
+        ctx.Input<framework::Tensor>("Src")->dims());
   }
 };
 
diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc
index 8a4981c7be7587a0cc5f72cabe71e05702112ac3..aa5479ceaff379d99a5f7322edac0c5e69816974 100644
--- a/paddle/operators/mean_op.cc
+++ b/paddle/operators/mean_op.cc
@@ -20,11 +20,9 @@ namespace operators {
 class MeanOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 1, "Input size of AddOp must be one");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1, "Output size of AddOp must be one");
-    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr && ctx.OutputVar(0) != nullptr,
-                   "Input/Output of MeanOp must be initialized.");
-    ctx.Output<Tensor>(0)->Resize(framework::make_ddim({1}));
+    PADDLE_ENFORCE(ctx.InputVar("X") != nullptr,
+                   "Input of MeanOp must be initialized.");
+    ctx.Output<Tensor>("Out")->Resize({1});
   }
 };
 
diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc
index ccab9a994cc7aa9e389bd259e4c7365a06e93aa1..b9099ad4e3d000df45a8df1b6f8f2b27b197d744 100644
--- a/paddle/operators/mul_op.cc
+++ b/paddle/operators/mul_op.cc
@@ -20,9 +20,8 @@ namespace operators {
 class MulOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 2, "The mul op must take two inputs");
-    auto dim0 = ctx.Input<Tensor>(0)->dims();
-    auto dim1 = ctx.Input<Tensor>(1)->dims();
+    auto dim0 = ctx.Input<Tensor>("X")->dims();
+    auto dim1 = ctx.Input<Tensor>("Y")->dims();
     PADDLE_ENFORCE_EQ(dim0.size(), 2,
                       "input X(%s) should be a tensor with 2 dims, a matrix",
                       ctx.op_.Input("X"));
@@ -32,8 +31,7 @@ class MulOp : public OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         dim0[1], dim1[0],
         "First matrix's width must be equal with second matrix's height.");
-    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "The mul op takes only one output");
-    ctx.Output<Tensor>(0)->Resize({dim0[0], dim1[1]});
+    ctx.Output<Tensor>("Out")->Resize({dim0[0], dim1[1]});
   }
 };
 
diff --git a/paddle/operators/net_op.cc b/paddle/operators/net_op.cc
index fbc98e09923bda7f3baee04e02df9076247bff0b..b0746883d0497ec06a46cb7d3aa886b848a2204d 100644
--- a/paddle/operators/net_op.cc
+++ b/paddle/operators/net_op.cc
@@ -15,6 +15,7 @@
 */
 
 #include "paddle/operators/net_op.h"
+#include <set>
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
@@ -23,36 +24,39 @@ namespace operators {
 void NetOp::CompleteAddOp(bool calc) {
   add_op_done_ = true;
   if (!calc) return;
-  std::unordered_set<std::string> input_set;
-  std::unordered_set<std::string> output_set;
-  std::unordered_set<std::string> temp_output;
+  std::set<std::string> input_set;
+  std::set<std::string> output_set;
+  std::set<std::string> temp_output;
   for (auto& op : ops_) {
     for (auto& ipt : op->inputs_) {
-      if (!Contains(output_set, ipt)) {  // Not other op's output
-        input_set.insert(ipt);
-      } else {
-        temp_output.insert(ipt);
+      for (auto& var_name : ipt.second) {
+        if (!Contains(output_set, var_name)) {  // Not other op's output
+          input_set.insert(var_name);
+        } else {
+          temp_output.insert(var_name);
+        }
       }
     }
 
     for (auto& opt : op->outputs_) {
-      output_set.insert(opt);
+      for (auto& var_name : opt.second) {
+        output_set.insert(var_name);
+      }
     }
   }
+  auto& inputs = inputs_["all"];
+  inputs.reserve(input_set.size());
+  std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs));
+  auto& outputs = outputs_["all"];
+  outputs.reserve(output_set.size());
+  std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs));
 
-  inputs_.reserve(input_set.size());
-  std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs_));
-  std::sort(inputs_.begin(), inputs_.end());
-
-  outputs_.reserve(output_set.size());
-  std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs_));
-  std::sort(outputs_.begin(), outputs_.end());
-
+  //! TODO figure out how to generate temporary_index in Network.
   std::vector<int> tmp_index;
   tmp_index.reserve(temp_output.size());
-  int output_len = static_cast<int>(outputs_.size());
+  int output_len = static_cast<int>(outputs.size());
   for (int i = 0; i < output_len; ++i) {
-    if (Contains(temp_output, outputs_[i])) {
+    if (Contains(temp_output, outputs[i])) {
       tmp_index.push_back(i);
     }
   }
diff --git a/paddle/operators/net_op.h b/paddle/operators/net_op.h
index 6e7af7f02ae23ec65459dfd15d950a43e96fec4d..0342cf4adb831e394935083f77673611d26e2e55 100644
--- a/paddle/operators/net_op.h
+++ b/paddle/operators/net_op.h
@@ -14,8 +14,7 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/framework/op_desc.pb.h"
-#include "paddle/framework/op_proto.pb.h"
+#include "paddle/framework/framework.pb.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc
index 5e9c15ca0e6a7c56611a0fadda6c3c0839f309e6..43c9aa72cd724a792c05501330d1ffee34cde632 100644
--- a/paddle/operators/recurrent_op.cc
+++ b/paddle/operators/recurrent_op.cc
@@ -89,12 +89,17 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
       // create step net's temp inputs
       for (auto& input : net_op->inputs_) {
         // the weight are located in parent scope
-        if (!step_scope.FindVar(input))
-          step_scope.NewVar(input)->GetMutable<Tensor>();
+        for (auto& var_name : input.second) {
+          if (!step_scope.FindVar(var_name)) {
+            step_scope.NewVar(var_name)->GetMutable<Tensor>();
+          }
+        }
       }
       // create stepnet's outputs
       for (const auto& output : net_op->outputs_) {
-        step_scope.NewVar(output);
+        for (auto& var_name : output.second) {
+          step_scope.NewVar(var_name);
+        }
       }
       step_scopes->emplace_back(&step_scope);
     }
diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc
index 8d1a36f2b332faad516ced012a409ca428bbf689..c6a1f082138f5a75e2d9cebbe7b547911e952e17 100644
--- a/paddle/operators/rowwise_add_op.cc
+++ b/paddle/operators/rowwise_add_op.cc
@@ -19,16 +19,14 @@ namespace operators {
 class RowWiseAddOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 2UL,
-                   "Two inputs is needed by rowwise add");
-    auto dim0 = ctx.Input<Tensor>(0)->dims();
-    auto dim1 = ctx.Input<Tensor>(1)->dims();
+    auto dim0 = ctx.Input<Tensor>("X")->dims();
+    auto dim1 = ctx.Input<Tensor>("b")->dims();
 
     PADDLE_ENFORCE(dim0.size() == 2, "Input 0 must be matrix");
     PADDLE_ENFORCE(dim1.size() == 1, "The second input must be vector");
     PADDLE_ENFORCE(dim0[1] == dim1[0], "The width of two input must be same");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1, "The output size must be 1");
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    PADDLE_ENFORCE(ctx.OutputSize("Out") == 1, "The output size must be 1");
+    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
 
diff --git a/paddle/operators/rowwise_add_op.h b/paddle/operators/rowwise_add_op.h
index b52524c47c7b80d8ddc6a94a4a6d03db8034088d..9e9f9d110c30452439cf4108d4a574b20a581c16 100644
--- a/paddle/operators/rowwise_add_op.h
+++ b/paddle/operators/rowwise_add_op.h
@@ -25,8 +25,8 @@ class RowWiseAddKernel : public OpKernel {
     auto out = context.Output<Tensor>(0);
     out->mutable_data<T>(context.GetPlace());
 
-    auto input = EigenMatrix<T>::From(*context.Input<Tensor>(0));
-    auto bias = EigenVector<T>::From(*context.Input<Tensor>(1));
+    auto input = EigenMatrix<T>::From(*context.Input<Tensor>("X"));
+    auto bias = EigenVector<T>::From(*context.Input<Tensor>("b"));
     auto output = EigenMatrix<T>::From(*out);
 
     const int bias_size = bias.dimension(0);
diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc
index 6307583f4ee3f185845690d0e378945d066eae75..659cb41d989489d0e0f8d3cea3f5fb06094d23dc 100644
--- a/paddle/operators/sgd_op.cc
+++ b/paddle/operators/sgd_op.cc
@@ -20,14 +20,10 @@ namespace operators {
 class SGDOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 2, "Input size of SGDOp must be two");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1, "Output size of SGDOp must be one");
-    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr, "inputs[0] mast be set");
-    PADDLE_ENFORCE(ctx.InputVar(1) != nullptr, "inputs[1] mast be set");
-    PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr, "outputs[0] mast be set");
-    PADDLE_ENFORCE(ctx.Input<Tensor>(0)->dims() == ctx.Input<Tensor>(1)->dims(),
-                   "Two input of SGD Op's dimension must be same.");
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    PADDLE_ENFORCE(
+        ctx.Input<Tensor>("param")->dims() == ctx.Input<Tensor>("grad")->dims(),
+        "Two input of SGD Op's dimension must be same.");
+    ctx.Output<Tensor>("param_out")->Resize(ctx.Input<Tensor>("param")->dims());
   }
 };
 
diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc
index 1eb795faa858796f7a34aa495b43d043fdb5dd43..27904ea0c3c21ec760dfed1b822bd45b586a933d 100644
--- a/paddle/operators/sigmoid_op.cc
+++ b/paddle/operators/sigmoid_op.cc
@@ -19,9 +19,7 @@ namespace operators {
 class SigmoidOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 1, "Sigmoid Op only have one input");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1, "Sigmoid Op only have one output");
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
 
diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc
index a070458f5e55cf47253ab0df5af7a1163b4f8092..836bce2294f52af3062fe2bde82fb1df2207f54c 100644
--- a/paddle/operators/softmax_op.cc
+++ b/paddle/operators/softmax_op.cc
@@ -20,12 +20,8 @@ namespace operators {
 class SoftmaxOp : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 1UL,
-                   "Only one input is need for softmax");
     PADDLE_ENFORCE(ctx.Input<Tensor>("X")->dims().size() == 2UL,
                    "The input of softmax op must be matrix");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
-                   "Only one output is need for softmax");
     ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
@@ -43,10 +39,6 @@ class SoftmaxOpMaker : public OpProtoAndCheckerMaker {
 class SoftmaxOpGrad : public OperatorWithKernel {
  protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 3UL,
-                   "Input of SoftmaxOpGrad should be 3, X, Y, YG");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
-                   "Output of SoftmaxOpGrad should be 1");
     PADDLE_ENFORCE(ctx.InputVar("Y") != nullptr, "Input(Y) should not be null");
     PADDLE_ENFORCE(ctx.InputVar(framework::GradVarName("Y")) != nullptr,
                    "Input(Y@GRAD) should not be null");
diff --git a/paddle/platform/enforce.h b/paddle/platform/enforce.h
index bc0715656a7d61774d53d4a0643ec1c105706085..60ce5822d3ade7ea6fa940db2f002a89263a34ca 100644
--- a/paddle/platform/enforce.h
+++ b/paddle/platform/enforce.h
@@ -195,12 +195,28 @@ struct CompatibleType {
   typedef typename std::conditional<t1_to_t2, T2, T1>::type type;
 };
 
+template <typename T>
+inline std::string enforce_to_string(const T& val) {
+  std::ostringstream sout;
+  sout << val;
+  return sout.str();
+}
+template <>
+inline std::string enforce_to_string(const std::string& val) {
+  return val;
+}
+template <>
+inline std::string enforce_to_string(const char* const& val) {
+  return std::string(val);
+}
+
 #define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...)        \
   PADDLE_ENFORCE(__COMPATIBLE_TYPE(__VAL0, __VAL1, __VAL0)                    \
                      __CMP __COMPATIBLE_TYPE(__VAL0, __VAL1, __VAL1),         \
                  "enforce %s " #__CMP " %s failed, %s " #__INV_CMP " %s\n%s", \
-                 #__VAL0, #__VAL1, std::to_string(__VAL0),                    \
-                 std::to_string(__VAL1),                                      \
+                 #__VAL0, #__VAL1,                                            \
+                 paddle::platform::enforce_to_string(__VAL0),                 \
+                 paddle::platform::enforce_to_string(__VAL1),                 \
                  paddle::string::Sprintf("" __VA_ARGS__));
 
 #define __COMPATIBLE_TYPE(__VAL0, __VAL1, __VAL)              \