Merge branch 'develop' of github.com:baidu/Paddle into feature/identity_op

CMakeLists.txt

@@ -55,6 +55,7 @@ option(WITH_C_API       "Compile PaddlePaddle with C-API(Prediction)"   OFF)
 option(WITH_GOLANG      "Compile PaddlePaddle with GOLANG"              OFF)
 option(GLIDE_INSTALL    "Download and install go dependencies "         ON)
 option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
+option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
 
 # CMAKE_BUILD_TYPE
 if(NOT CMAKE_BUILD_TYPE)

cmake/configure.cmake

@@ -28,6 +28,10 @@ if(NOT WITH_TIMER)
     add_definitions(-DPADDLE_DISABLE_TIMER)
 endif(NOT WITH_TIMER)
 
+if(USE_EIGEN_FOR_BLAS)
+    add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS)
+endif(USE_EIGEN_FOR_BLAS)
+
 if(NOT WITH_PROFILER)
     add_definitions(-DPADDLE_DISABLE_PROFILER)
 endif(NOT WITH_PROFILER)

doc/api/v2/config/layer.rst

@@ -362,6 +362,11 @@ trans
 ..  autoclass:: paddle.v2.layer.trans
     :noindex:
 
+scale_shift
+-----------
+..  autoclass:: paddle.v2.layer.scale_shift
+    :noindex:
+
 Sampling Layers
 ===============
 

paddle/framework/CMakeLists.txt

@@ -18,8 +18,8 @@ cc_test(scope_test SRCS scope_test.cc DEPS scope)
 proto_library(framework_proto SRCS framework.proto)
 
 cc_library(attribute SRCS attribute.cc DEPS framework_proto)
-
-cc_library(operator SRCS operator.cc DEPS framework_proto device_context tensor scope attribute)
+cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
+cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 
 cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator)
@@ -45,6 +45,7 @@ cc_library(paddle_pybind SHARED
     SRCS pybind.cc
     DEPS pybind python backward
     sgd_op
+    gather_op
     add_op
     mul_op
     rowwise_add_op

paddle/framework/backward.cc

@@ -110,7 +110,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
                        dup_output_ops[out].emplace_back(local_op_id);
                        return false;
                      });
-      net->AddOp(std::move(bwd));
+      net->AppendOp(std::move(bwd));
     }
     // Get unique ID for this method.
     auto uid = uniq_id++;
@@ -163,7 +163,8 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
 
         // If part of input gradient of that operator is not calculated, fill
         // zero variables to that input gradient.
-        net->AddOp(OpRegistry::CreateOp("fill_zeros_like", {{"Src", {prefix}}},
+        net->AppendOp(OpRegistry::CreateOp("fill_zeros_like",
+                                           {{"Src", {prefix}}},
                                            {{"Dst", {grad_input}}}, {}));
       }
       return false;
@@ -195,7 +196,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
     if (net->ops_.empty()) {  // Current no aux op is added to network
       return grad_op;
     }
-    net->AddOp(std::move(grad_op));
+    net->AppendOp(std::move(grad_op));
   }
   net->SetType("@GENERATED_BACKWARD@");
   net->CompleteAddOp();

paddle/framework/backward_test.cc

@@ -72,16 +72,16 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker {
 
 class FcOp : public operators::NetOp {
  public:
-  FcOp(const std::string &type, const VarNameMap &inputs,
-       const VarNameMap &outputs, const AttributeMap &attrs)
+  FcOp(const std::string &type, const VariableNameMap &inputs,
+       const VariableNameMap &outputs, const AttributeMap &attrs)
       : NetOp(type, inputs, outputs, attrs) {
-    AddOp(OpRegistry::CreateOp("mul",
+    AppendOp(OpRegistry::CreateOp("mul",
                                   {{"X", {Input("X")}}, {"Y", {Input("W")}}},
                                   {{"Out", {Output("mul_result")}}}, {}));
     auto input_b = Inputs("b");
     std::string before_act = "mul_result";
     if (input_b.size() != 0) {
-      AddOp(OpRegistry::CreateOp(
+      AppendOp(OpRegistry::CreateOp(
           "rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}},
           {{"Out", {Output("add_result")}}}, {}));
       before_act = "add_result";
@@ -92,7 +92,7 @@ class FcOp : public operators::NetOp {
       }
     }
 
-    AddOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
+    AppendOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
                                   {{"Out", {Output("Out")}}}, {}));
     CompleteAddOp(false);
   }
@@ -234,13 +234,13 @@ TEST(Backward, net_fc_backward_not_have_b) {
 
 TEST(Backward, net_input_of_network_not_need_grad) {
   ops::NetOp net;
-  net.AddOp(f::OpRegistry::CreateOp(
+  net.AppendOp(f::OpRegistry::CreateOp(
       "fc", {{"X", {"x"}}, {"W", {"W1"}}, {"b", {"b1"}}},
       {{"mul_result", {"mul_tmp_0"}},
        {"add_result", {"add_tmp_0"}},
        {"Out", {"hidden0"}}},
       {}));
-  net.AddOp(f::OpRegistry::CreateOp(
+  net.AppendOp(f::OpRegistry::CreateOp(
       "fc", {{"X", {"hidden0"}}, {"W", {"W2"}}, {"b", {"b2"}}},
       {{"mul_result", {"mul_tmp_1"}},
        {"add_result", {"add_tmp_1"}},
@@ -273,9 +273,9 @@ TEST(Backward, net_input_of_network_not_need_grad) {
 
 TEST(Backward, net_shared_weight) {
   ops::NetOp net;
-  net.AddOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}},
+  net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}},
                                        {{"Out", {"out"}}}, {}));
-  net.AddOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}},
+  net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}},
                                        {{"Out", {"FinalOut"}}}, {}));
   net.CompleteAddOp();
 
@@ -357,19 +357,19 @@ TEST(Backward, op_part_of_input_are_not_need) {
 
 TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
   ops::NetOp net;
-  net.AddOp(f::OpRegistry::CreateOp(
+  net.AppendOp(f::OpRegistry::CreateOp(
       "fc", {{"X", {"x1"}}, {"W", {"w1"}}, {"b", {"b1"}}},
       {{"mul_result", {"mul_out1"}},
        {"add_result", {"add_out1"}},
        {"Out", {"out1"}}},
       {}));
-  net.AddOp(f::OpRegistry::CreateOp(
+  net.AppendOp(f::OpRegistry::CreateOp(
       "fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}},
       {{"mul_result", {"mul_out2"}},
        {"add_result", {"tmp_out2"}},
        {"Out", {"out2"}}},
       {}));
-  net.AddOp(f::OpRegistry::CreateOp(
+  net.AppendOp(f::OpRegistry::CreateOp(
       "fc", {{"X", {"out2"}}, {"W", {"w3"}}, {"b", {"b3"}}},
       {{"mul_result", {"mul_out3"}},
        {"add_result", {"tmp_out3"}},

paddle/framework/grad_op_builder.cc

@@ -20,13 +20,13 @@ namespace framework {
 enum class OpArgType { IN, OUT };
 
 static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
-                       bool is_grad, OperatorBase::VarNameMap* vars) {
+                       bool is_grad, VariableNameMap* vars) {
   const auto& src_inout =
       src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs();
   auto& dst_inout = *vars;
-  const OpProto* proto = OpRegistry::op_info_map().at(src_op->Type()).proto_;
+  auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto();
   const auto& src_arg_list =
-      src_type == OpArgType::IN ? proto->inputs() : proto->outputs();
+      src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
   for (const auto& arg : src_arg_list) {
     if (arg.not_in_gradient() && !is_grad) continue;
     const std::string src_name = arg.name();
@@ -40,26 +40,18 @@ static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
 }
 
 OperatorBase* BuildGradOp(const OperatorBase* op) {
-  auto it = OpRegistry::op_info_map().find(op->Type());
-  PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
-                 "'%s' has not been registered.", op->Type());
-  PADDLE_ENFORCE(it->second.proto_ != nullptr, "'%s' has no OpProto.",
-                 op->Type());
-  std::string grad_op_type = it->second.grad_op_type_;
-  PADDLE_ENFORCE(!grad_op_type.empty(), "'%s' has no gradient operator.",
-                 op->Type());
+  auto& info = OpInfoMap::Instance().Get(op->Type());
+  PADDLE_ENFORCE(info.HasGradientOp());
 
-  OperatorBase::VarNameMap inputs;
-  OperatorBase::VarNameMap outputs;
+  VariableNameMap inputs;
+  VariableNameMap outputs;
   TransOpArg(op, OpArgType::IN, false, &inputs);   // I
   TransOpArg(op, OpArgType::OUT, false, &inputs);  // O
   TransOpArg(op, OpArgType::OUT, true, &inputs);   // OG
   TransOpArg(op, OpArgType::IN, true, &outputs);   // IG
 
-  it = OpRegistry::op_info_map().find(grad_op_type);
-  PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
-                 "'%s' has not been registered.", grad_op_type);
-  return it->second.creator_(grad_op_type, inputs, outputs, op->Attrs());
+  auto& grad_info = OpInfoMap::Instance().Get(info.grad_op_type_);
+  return grad_info.Creator()(info.grad_op_type_, inputs, outputs, op->Attrs());
 }
 
 }  // namespace framework

paddle/framework/op_info.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/framework/op_info.h"
+
+namespace paddle {
+namespace framework {
+
+static OpInfoMap* g_op_info_map = nullptr;
+
+OpInfoMap& OpInfoMap::Instance() {
+  if (g_op_info_map == nullptr) {
+    g_op_info_map = new OpInfoMap();
+  }
+  return *g_op_info_map;
+}
+}  // namespace framework
+}  // namespace paddle

paddle/framework/op_info.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include <functional>
+#include <map>
+#include <string>
+#include <unordered_map>
+
+#include "paddle/framework/attribute.h"
+
+namespace paddle {
+namespace framework {
+class OperatorBase;
+using VariableNameMap = std::map<std::string, std::vector<std::string>>;
+
+using OpCreator = std::function<OperatorBase*(
+    const std::string& /*type*/, const VariableNameMap& /*inputs*/,
+    const VariableNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
+
+struct OpInfo {
+  OpCreator creator_;
+  std::string grad_op_type_;
+  OpProto* proto_;
+  OpAttrChecker* checker_;
+
+  bool HasOpProtoAndChecker() const {
+    return proto_ != nullptr && checker_ != nullptr;
+  }
+
+  const OpProto& Proto() const {
+    PADDLE_ENFORCE_NOT_NULL(proto_, "Operator Proto has not been registered");
+    PADDLE_ENFORCE(proto_->IsInitialized(),
+                   "Operator Proto must be initialized in op info");
+    return *proto_;
+  }
+
+  const OpAttrChecker& Checker() const {
+    PADDLE_ENFORCE_NOT_NULL(checker_,
+                            "Operator Checker has not been registered");
+    return *checker_;
+  }
+
+  const OpCreator& Creator() const {
+    PADDLE_ENFORCE_NOT_NULL(creator_,
+                            "Operator Creator has not been registered");
+    return creator_;
+  }
+
+  bool HasGradientOp() const { return !grad_op_type_.empty(); }
+};
+
+class OpInfoMap {
+ public:
+  static OpInfoMap& Instance();
+
+  OpInfoMap(const OpInfoMap& o) = delete;
+  OpInfoMap(OpInfoMap&& o) = delete;
+  OpInfoMap& operator=(const OpInfoMap& o) = delete;
+  OpInfoMap& operator=(OpInfoMap&& o) = delete;
+
+  bool Has(const std::string& op_type) const {
+    return map_.find(op_type) != map_.end();
+  }
+
+  void Insert(const std::string& type, const OpInfo& info) {
+    PADDLE_ENFORCE(!Has(type), "Operator %s has been registered", type);
+    map_.insert({type, info});
+  }
+
+  const OpInfo& Get(const std::string& type) const {
+    auto it = map_.find(type);
+    PADDLE_ENFORCE(it != map_.end(), "Operator %s are not found", type);
+    return it->second;
+  }
+
+  template <typename Callback>
+  void IterAllInfo(Callback callback) {
+    for (auto& it : map_) {
+      callback(it.first, it.second);
+    }
+  }
+
+ private:
+  OpInfoMap() = default;
+  std::unordered_map<std::string, const OpInfo> map_;
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/op_registry.cc

@@ -19,32 +19,18 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
-                                                   const VarNameMap& inputs,
-                                                   const VarNameMap& outputs,
-                                                   AttributeMap attrs) {
-  auto it = op_info_map().find(type);
-  PADDLE_ENFORCE(it != op_info_map().end(),
-                 "Operator '%s' has not been registered.", type);
-  it->second.checker_->Check(attrs);
-  auto op = it->second.creator_(type, inputs, outputs, attrs);
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
+    const std::string& type, const VariableNameMap& inputs,
+    const VariableNameMap& outputs, AttributeMap attrs) {
+  auto& info = OpInfoMap::Instance().Get(type);
+  info.Checker().Check(attrs);
+  auto op = info.Creator()(type, inputs, outputs, attrs);
   return std::unique_ptr<OperatorBase>(op);
 }
 
-std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
-  VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
-  VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
-  AttributeMap attrs;
-  for (auto& attr : op_desc.attrs()) {
-    attrs[attr.name()] = GetAttrValue(attr);
-  }
-
-  return CreateOp(op_desc.type(), inputs, outputs, attrs);
-}
-
-OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
+static VariableNameMap ConvertOpDescVarsToVarNameMap(
     const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
-  VarNameMap ret_val;
+  VariableNameMap ret_val;
   for (auto& var : op_desc_vars) {
     auto& var_names = ret_val[var.parameter()];
     auto& var_names_in_proto = var.arguments();
@@ -55,6 +41,17 @@ OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
   return ret_val;
 }
 
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
+  VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
+  VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
+  AttributeMap attrs;
+  for (auto& attr : op_desc.attrs()) {
+    attrs[attr.name()] = GetAttrValue(attr);
+  }
+
+  return CreateOp(op_desc.type(), inputs, outputs, attrs);
+}
+
 std::unique_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
   PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
   return std::unique_ptr<OperatorBase>(BuildGradOp(&op));

paddle/framework/op_registry.h

@@ -23,6 +23,7 @@ limitations under the License. */
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/grad_op_builder.h"
+#include "paddle/framework/op_info.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
 
@@ -30,28 +31,16 @@ namespace paddle {
 namespace framework {
 
 class OpRegistry {
-  using VarNameMap = OperatorBase::VarNameMap;
-  using OpCreator = std::function<OperatorBase*(
-      const std::string& /*type*/, const VarNameMap& /*inputs*/,
-      const VarNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
-
  public:
-  struct OpInfo {
-    OpCreator creator_;
-    std::string grad_op_type_;
-    OpProto* proto_;
-    OpAttrChecker* checker_;
-  };
-
   template <typename OpType, typename ProtoMakerType, typename GradOpType>
   static void RegisterOp(const std::string& op_type,
                          const std::string& grad_op_type) {
-    PADDLE_ENFORCE(op_info_map().count(op_type) == 0,
+    PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type),
                    "'%s' is registered more than once.", op_type);
     OpInfo op_info;
-    op_info.creator_ = [](const std::string& type, const VarNameMap& inputs,
-                          const VarNameMap& outputs,
-                          const AttributeMap& attrs) {
+    op_info.creator_ = [](
+        const std::string& type, const VariableNameMap& inputs,
+        const VariableNameMap& outputs, const AttributeMap& attrs) {
       return new OpType(type, inputs, outputs, attrs);
     };
     op_info.grad_op_type_ = grad_op_type;
@@ -70,7 +59,7 @@ class OpRegistry {
       op_info.proto_ = nullptr;
       op_info.checker_ = nullptr;
     }
-    op_info_map().insert(std::make_pair(op_type, op_info));
+    OpInfoMap::Instance().Insert(op_type, op_info);
     // register gradient op
     if (!grad_op_type.empty()) {
       RegisterOp<GradOpType, NOPMaker, NOP>(grad_op_type, "");
@@ -78,21 +67,13 @@ class OpRegistry {
   }
 
   static std::unique_ptr<OperatorBase> CreateOp(const std::string& type,
-                                                const VarNameMap& inputs,
-                                                const VarNameMap& outputs,
+                                                const VariableNameMap& inputs,
+                                                const VariableNameMap& outputs,
                                                 AttributeMap attrs);
 
   static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
-  static VarNameMap ConvertOpDescVarsToVarNameMap(
-      const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars);
-
   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_;
-    return op_info_map_;
-  }
 };
 
 class Registrar {

paddle/framework/operator.cc

@@ -115,8 +115,8 @@ void OperatorBase::Rename(const std::string& old_name,
 }
 
 OperatorBase::OperatorBase(const std::string& type,
-                           const OperatorBase::VarNameMap& inputs,
-                           const OperatorBase::VarNameMap& outputs,
+                           const VariableNameMap& inputs,
+                           const VariableNameMap& outputs,
                            const AttributeMap& attrs)
     : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) {
   static std::atomic<size_t> gUniqId(0UL);
@@ -141,18 +141,10 @@ std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
     }
     return ret_val;
   }
-  auto it = OpRegistry::op_info_map().find(type_);
-  PADDLE_ENFORCE(
-      it != OpRegistry::op_info_map().end(),
-      "Operator %s not registered, cannot figure out intermediate outputs",
-      type_);
-  PADDLE_ENFORCE(
-      it->second.proto_ != nullptr,
-      "Operator %s has no OpProto, cannot figure out intermediate outputs",
-      type_);
+  auto& info = OpInfoMap::Instance().Get(Type());
 
   // get all OpProto::Var for outputs
-  for (auto& o : it->second.proto_->outputs()) {
+  for (auto& o : info.Proto().outputs()) {
     // ignore all intermediate output
     if (o.intermediate()) continue;
     auto out = outputs_.find(o.name());

paddle/framework/operator.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <unordered_map>
 #include <vector>
 
+#include "op_info.h"
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/scope.h"
@@ -62,10 +63,8 @@ class ExecutionContext;
  */
 class OperatorBase {
  public:
-  using VarNameMap = std::map<std::string, std::vector<std::string>>;
-
-  OperatorBase(const std::string& type, const VarNameMap& inputs,
-               const VarNameMap& outputs, const AttributeMap& attrs);
+  OperatorBase(const std::string& type, const VariableNameMap& inputs,
+               const VariableNameMap& outputs, const AttributeMap& attrs);
 
   virtual ~OperatorBase() {}
 
@@ -93,8 +92,8 @@ class OperatorBase {
   /// rename inputs outputs name
   void Rename(const std::string& old_name, const std::string& new_name);
 
-  const VarNameMap& Inputs() const { return inputs_; }
-  const VarNameMap& Outputs() const { return outputs_; }
+  const VariableNameMap& Inputs() const { return inputs_; }
+  const VariableNameMap& Outputs() const { return outputs_; }
   //! Get a input with argument's name described in `op_proto`
   const std::string& Input(const std::string& name) const;
   //! Get a input which has multiple variables.
@@ -122,30 +121,32 @@ class OperatorBase {
   // I (Inputs)opear
   // O (Outputs)
   // OG (Output Gradients)
-  VarNameMap inputs_;
+  VariableNameMap inputs_;
 
   // NOTE: in case of OpGrad, outputs_ contains
   // IG (Inputs Gradients)
-  VarNameMap outputs_;
+  VariableNameMap outputs_;
   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)                       \
+#define DEFINE_OP_CLONE_METHOD(cls)                       \
   std::unique_ptr<OperatorBase> Clone() const final {     \
-    return std::unique_ptr<OperatorBase>(new CLS(*this)); \
+    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) {}
+#define DEFINE_OP_CONSTRUCTOR(cls, parent_cls)             \
+  cls(const std::string& type,                             \
+      const ::paddle::framework::VariableNameMap& inputs,  \
+      const ::paddle::framework::VariableNameMap& outputs, \
+      const paddle::framework::AttributeMap& attrs)        \
+      : parent_cls(type, inputs, outputs, attrs) {}
 
 class NOP : public OperatorBase {
  public:
@@ -389,8 +390,8 @@ class OperatorWithKernel : public OperatorBase {
   using OpKernelMap =
       std::unordered_map<OpKernelKey, std::unique_ptr<OpKernel>, OpKernelHash>;
 
-  OperatorWithKernel(const std::string& type, const VarNameMap& inputs,
-                     const VarNameMap& outputs, const AttributeMap& attrs)
+  OperatorWithKernel(const std::string& type, const VariableNameMap& inputs,
+                     const VariableNameMap& outputs, const AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
 
   void InferShape(const Scope& scope) const override {

paddle/framework/operator_test.cc

@@ -23,8 +23,8 @@ static int op_run_num = 0;
 
 class OpWithoutKernelTest : public OperatorBase {
  public:
-  OpWithoutKernelTest(const std::string& type, const VarNameMap& inputs,
-                      const VarNameMap& outputs, const AttributeMap& attrs)
+  OpWithoutKernelTest(const std::string& type, const VariableNameMap& inputs,
+                      const VariableNameMap& outputs, const AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs), x(1) {}
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
@@ -249,8 +249,9 @@ TEST(OpKernel, multi_inputs) {
 class OperatorClone : public paddle::framework::OperatorBase {
  public:
   DEFINE_OP_CLONE_METHOD(OperatorClone);
-  OperatorClone(const std::string& type, const VarNameMap& inputs,
-                const VarNameMap& outputs,
+  OperatorClone(const std::string& type,
+                const paddle::framework::VariableNameMap& inputs,
+                const paddle::framework::VariableNameMap& outputs,
                 const paddle::framework::AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void InferShape(const paddle::framework::Scope& scope) const override {}

paddle/framework/pybind.cc

@@ -31,7 +31,7 @@ limitations under the License. */
 namespace py = pybind11;
 
 USE_OP(add_two);
-USE_CPU_ONLY_OP(onehot_cross_entropy);
+USE_OP(onehot_cross_entropy);
 USE_OP(sgd);
 USE_OP(mul);
 USE_OP(mean);
@@ -44,6 +44,7 @@ USE_OP(gaussian_random);
 USE_OP(uniform_random);
 USE_OP(scale);
 USE_OP_ITSELF(identity);
+USE_CPU_ONLY_OP(gather);
 
 namespace paddle {
 namespace framework {
@@ -140,19 +141,16 @@ All parameter, weight, gradient are variables in Paddle.
   //! @note: Be careful! PyBind will return std::string as an unicode, not
   //! Python str. If you want a str object, you should cast them in Python.
   m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
-    auto &op_info_map = OpRegistry::op_info_map();
     std::vector<py::bytes> ret_values;
-    for (auto it = op_info_map.begin(); it != op_info_map.end(); ++it) {
-      const OpProto *proto = it->second.proto_;
-      if (proto == nullptr) {
-        continue;
-      }
-      PADDLE_ENFORCE(proto->IsInitialized(), "OpProto must all be initialized");
+
+    OpInfoMap::Instance().IterAllInfo([&ret_values](const std::string &type,
+                                                    const OpInfo &info) {
+      if (!info.HasOpProtoAndChecker()) return;
       std::string str;
-      PADDLE_ENFORCE(proto->SerializeToString(&str),
+      PADDLE_ENFORCE(info.Proto().SerializeToString(&str),
                      "Serialize OpProto Error. This could be a bug of Paddle.");
-      ret_values.push_back(py::bytes(str));
-    }
+      ret_values.emplace_back(str);
+    });
     return ret_values;
   });
   m.def_submodule(
@@ -224,8 +222,8 @@ All parameter, weight, gradient are variables in Paddle.
                     retv->SetType("plain_net");
                     return retv;
                   })
-      .def("add_op", [](operators::NetOp &self,
-                        const OperatorBase &op) { self.AddOp(op); })
+      .def("append_op", [](operators::NetOp &self,
+                           const OperatorBase &op) { self.AppendOp(op); })
       .def("complete_add_op", &operators::NetOp::CompleteAddOp)
       .def("complete_add_op", [](std::shared_ptr<operators::NetOp> &self) {
         self->CompleteAddOp();

paddle/function/CMakeLists.txt

@@ -4,6 +4,10 @@ file(GLOB cpp_files . *Op.cpp)
 list(APPEND h_files Function.h)
 list(APPEND cpp_files Function.cpp)
 list(APPEND cpp_files BufferArg.cpp)
+list(APPEND cpp_files GemmFunctor.cpp)
+if(USE_EIGEN_FOR_BLAS)
+  list(APPEND cpp_files EigenGemm.cpp)
+endif(USE_EIGEN_FOR_BLAS)
 
 if(WITH_GPU)
     file(GLOB cu_files . *OpGpu.cu)

paddle/function/DepthwiseConvOp.cpp

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #include "DepthwiseConvOp.h"
 #include "ConvOp.h"
-#include "GemmFunctor.h"
 
 namespace paddle {
 

paddle/function/DepthwiseConvOpGpu.cu

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "DepthwiseConvOp.h"
-#include "GemmFunctor.h"
 #include "paddle/math/BaseMatrix.h"
 
 namespace paddle {

paddle/function/EigenGemm.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <glog/logging.h>
+#include "unsupported/Eigen/CXX11/Tensor"
+
+namespace paddle {
+
+template <class T>
+struct EigenBlasGemm {
+  typedef Eigen::TensorMap<Eigen::Tensor<T, 2, Eigen::RowMajor, int>,
+                           Eigen::Aligned>
+      Matrix;
+
+  static void compute(const bool transA,
+                      const bool transB,
+                      const int M,
+                      const int N,
+                      const int K,
+                      const T alpha,
+                      const T* A,
+                      const int lda,
+                      const T* B,
+                      const int ldb,
+                      const T beta,
+                      T* C,
+                      const int ldc) {
+    Eigen::array<int, 2> sizeA;
+    if (transA) {
+      sizeA[0] = K;
+      sizeA[1] = M;
+      CHECK_EQ(M, lda);
+    } else {
+      sizeA[0] = M;
+      sizeA[1] = K;
+      CHECK_EQ(K, lda);
+    }
+    Eigen::array<int, 2> sizeB;
+    if (transB) {
+      sizeB[0] = N;
+      sizeB[1] = K;
+      CHECK_EQ(K, ldb);
+    } else {
+      sizeB[0] = K;
+      sizeB[1] = N;
+      CHECK_EQ(N, ldb);
+    }
+    Eigen::array<int, 2> sizeC;
+    sizeC[0] = M;
+    sizeC[1] = N;
+    CHECK_EQ(N, ldc);
+
+    const Matrix a(const_cast<T*>(A), sizeA);
+    const Matrix b(const_cast<T*>(B), sizeB);
+    Matrix c(C, sizeC);
+
+    typedef typename Eigen::Tensor<T, 2>::DimensionPair DimPair;
+    Eigen::array<DimPair, 1> dims;
+    dims[0] = DimPair(1, 0);
+    dims[0].first = transA ? 0 : 1;
+    dims[0].second = transB ? 1 : 0;
+
+    Eigen::DefaultDevice device;
+    if (alpha == T(1) && beta == T(0)) {
+      c.device(device) = a.contract(b, dims);
+    } else if (alpha == T(1) && beta == T(1)) {
+      c.device(device) += a.contract(b, dims);
+    } else {
+      c.device(device) = alpha * a.contract(b, dims) + beta * c;
+    }
+  }
+};
+
+#ifdef PADDLE_TYPE_DOUBLE
+template class EigenBlasGemm<double>;
+#else
+template class EigenBlasGemm<float>;
+#endif
+
+}  // namespace paddle

paddle/function/GemmConvOp.cpp

@@ -85,7 +85,6 @@ public:
     }
 
     Im2ColFunctor<kCFO, Device, real> im2col;
-    GemmFunctor<Device, real> gemm;
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -108,8 +107,8 @@ public:
         int M = outputChannels / groups_;
         int N = outputHeight * outputWidth;
         int K = inputChannels / groups_ * filterHeight * filterWidth;
-        gemm(CblasNoTrans,
-             CblasNoTrans,
+        BlasGemm<Device, real>::compute(false,
+                                        false,
                                         M,
                                         N,
                                         K,
@@ -188,8 +187,6 @@ public:
     }
 
     Col2ImFunctor<kCFO, Device, real> col2im;
-    GemmFunctor<Device, real> gemm;
-
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -205,8 +202,8 @@ public:
           colData = inputGrad + g * inputOffset;
           scale = 1.0f;
         }
-        gemm(CblasTrans,
-             CblasNoTrans,
+        BlasGemm<Device, real>::compute(true,
+                                        false,
                                         M,
                                         N,
                                         K,
@@ -299,7 +296,6 @@ public:
     }
 
     Im2ColFunctor<kCFO, Device, real> im2col;
-    GemmFunctor<Device, real> gemm;
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -321,8 +317,8 @@ public:
         int M = outputChannels / groups_;
         int K = outputHeight * outputWidth;
         int N = inputChannels / groups_ * filterHeight * filterWidth;
-        gemm(CblasNoTrans,
-             CblasTrans,
+        BlasGemm<Device, real>::compute(false,
+                                        true,
                                         M,
                                         N,
                                         K,

paddle/function/GemmFunctor.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "GemmFunctor.h"
+#include "paddle/math/MathFunctions.h"
+
+namespace paddle {
+
+template <class T>
+struct BlasGemm<DEVICE_TYPE_CPU, T> {
+  static void compute(const bool transA,
+                      const bool transB,
+                      const int M,
+                      const int N,
+                      const int K,
+                      const T alpha,
+                      const T* A,
+                      const int lda,
+                      const T* B,
+                      const int ldb,
+                      const T beta,
+                      T* C,
+                      const int ldc) {
+#ifdef PADDLE_USE_EIGEN_FOR_BLAS
+    EigenBlasGemm<T>::compute(
+        transA, transB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc);
+#else
+    gemm<T>(transA == false ? CblasNoTrans : CblasTrans,
+            transB == false ? CblasNoTrans : CblasTrans,
+            M,
+            N,
+            K,
+            alpha,
+            A,
+            lda,
+            B,
+            ldb,
+            beta,
+            C,
+            ldc);
+#endif
+  }
+};
+
+template <class T>
+struct BlasGemm<DEVICE_TYPE_GPU, T> {
+  static void compute(const bool transA,
+                      const bool transB,
+                      const int M,
+                      const int N,
+                      const int K,
+                      const T alpha,
+                      const T* A,
+                      const int lda,
+                      const T* B,
+                      const int ldb,
+                      const T beta,
+                      T* C,
+                      const int ldc) {
+    hl_matrix_mul((T*)A,
+                  transA == false ? HPPL_OP_N : HPPL_OP_T,
+                  (T*)B,
+                  transB == false ? HPPL_OP_N : HPPL_OP_T,
+                  C,
+                  M,
+                  N,
+                  K,
+                  alpha,
+                  beta,
+                  lda,
+                  ldb,
+                  ldc);
+  }
+};
+
+template struct BlasGemm<DEVICE_TYPE_CPU, real>;
+template struct BlasGemm<DEVICE_TYPE_GPU, real>;
+
+}  // namespace paddle

paddle/function/GemmFunctor.h

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/math/MathFunctions.h"
+#include "TensorType.h"
 
 namespace paddle {
 
@@ -24,10 +24,9 @@ namespace paddle {
 // of MatMulFunction, we need to consider the reconstruction of hl_matrix_mul
 // interface.
 template <DeviceType Device, class T>
-class GemmFunctor {
-public:
-  void operator()(const CBLAS_TRANSPOSE transA,
-                  const CBLAS_TRANSPOSE TransB,
+struct BlasGemm {
+  static void compute(const bool transA,
+                      const bool transB,
                       const int M,
                       const int N,
                       const int K,
@@ -41,11 +40,15 @@ public:
                       const int ldc);
 };
 
+// TODO(hedaoyuan): Since the definition of the real type in the Paddle
+// conflicts with the Eigen library, so compile the Eigen code can not
+// include the Paddle header file. And need an EigenBlasGemm template class
+// that does not contain the DeviceType parameter.
+// I will fix this problem and merge BlasGemm and EigenBlasGemm into one.
 template <class T>
-class GemmFunctor<DEVICE_TYPE_CPU, T> {
-public:
-  void operator()(const CBLAS_TRANSPOSE transA,
-                  const CBLAS_TRANSPOSE TransB,
+struct EigenBlasGemm {
+  static void compute(const bool transA,
+                      const bool transB,
                       const int M,
                       const int N,
                       const int K,
@@ -56,41 +59,7 @@ public:
                       const int ldb,
                       const T beta,
                       T* C,
-                  const int ldc) {
-    gemm<T>(transA, TransB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc);
-  }
-};
-
-template <class T>
-class GemmFunctor<DEVICE_TYPE_GPU, T> {
-public:
-  void operator()(const CBLAS_TRANSPOSE transA,
-                  const CBLAS_TRANSPOSE TransB,
-                  const int M,
-                  const int N,
-                  const int K,
-                  const T alpha,
-                  const T* A,
-                  const int lda,
-                  const T* B,
-                  const int ldb,
-                  const T beta,
-                  T* C,
-                  const int ldc) {
-    hl_matrix_mul((T*)A,
-                  transA == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T,
-                  (T*)B,
-                  TransB == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T,
-                  C,
-                  M,
-                  N,
-                  K,
-                  alpha,
-                  beta,
-                  lda,
-                  ldb,
-                  ldc);
-  }
+                      const int ldc);
 };
 
 }  // namespace paddle

paddle/gserver/gradientmachines/NeuralNetwork.cpp

@@ -202,7 +202,7 @@ void NeuralNetwork::prefetch(const std::vector<Argument>& inArgs) {
         auto mat = dynamic_cast<SparsePrefetchRowCpuMatrix*>(
             para->getMat(PARAMETER_VALUE).get());
         para->clearGradient();
-        mat->clearIndices();
+        if (mat) mat->clearIndices();
       }
     }
   }

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -184,7 +184,7 @@ public:
   }
 
   void backward(const UpdateCallback& callback) override {
-    if (biases_) {
+    if (biases_ && biases_->getWGrad()) {
       backwardActivation();
       biases_->getWGrad()->collectBias(*getOutputGrad(), 1);
       biases_->getParameterPtr()->incUpdate(callback);
@@ -1012,11 +1012,6 @@ void RecurrentGradientMachine::generateSequence() {
                            /* width */ resultNum,
                            false,
                            /* useGpu */ false);
-    Matrix::resizeOrCreate(generator_.outArg.value,
-                           /* height */ maxGenWordCount,
-                           /* width */ 1,
-                           false,
-                           /* useGpu */ false);
   }
   ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions,
                                 numSequences + 1,
@@ -1026,7 +1021,7 @@ void RecurrentGradientMachine::generateSequence() {
   } else {
     oneWaySearch(numSequences);
   }
-  if (dataArgsSize_) createDataOutlink(batchMachineIdVec_);
+  if (dataArgsSize_) createDataOutlink();
 
   size_t size = generator_.ids.size();
   generator_.outArg.ids->resize(size);
@@ -1106,6 +1101,7 @@ void RecurrentGradientMachine::oneWaySearch(size_t batchSize) {
   }
 
   batchMachineIdVec_.clear();
+  batchMachineStartPos_.clear();
   int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false);
   starts[0] = 0;
   generator_.ids.clear();
@@ -1312,13 +1308,20 @@ void RecurrentGradientMachine::fillGenOutputs() {
     finalPaths_[i].resize(minFinalPathsSize);
   }
 
-  batchMachineIdVec_.clear();
   generator_.ids.clear();
   int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false);
   starts[0] = 0;
   if (numResults > 1) {
-    real* probs = generator_.outArg.in->getData();
+    int idsProbSaveSize = 0;
+    for (auto inSeq : finalPaths_) {
+      for (auto path : inSeq) idsProbSaveSize += path.ids.size();
+      idsProbSaveSize += inSeq.size();
+    }
+    Matrix::resizeOrCreate(
+        generator_.outArg.value, idsProbSaveSize, 1, false, false);
     real* idsProb = generator_.outArg.value->getData();
+
+    real* probs = generator_.outArg.in->getData();
     size_t curPos = 0;
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
@@ -1333,24 +1336,16 @@ void RecurrentGradientMachine::fillGenOutputs() {
         curPos += genLen;
         idsProb[curPos++] = -1.0;
         probs[i * numResults + j] = path.logProb;
-
-        if (!j && dataArgsSize_) {
-          // in beam search, here only reserved the top 1 generated result
-          // for out_links that are not the generated word indices.
-          batchMachineIdVec_.insert(batchMachineIdVec_.end(),
-                                    path.machineIdVec.begin(),
-                                    path.machineIdVec.end());
-        }
       }
       starts[i + 1] = generator_.ids.size();
     }
   } else {
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       CHECK(!finalPaths_[i].empty());
-      generator_.ids.insert(generator_.ids.begin(),
-                            finalPaths_[i][0].ids.begin(),
-                            finalPaths_[i][0].ids.end());
-      starts[i + 1] = starts[i] + finalPaths_[i][0].ids.size();
+      Path& path = finalPaths_[i][0];
+      generator_.ids.insert(
+          generator_.ids.begin(), path.ids.begin(), path.ids.end());
+      starts[i + 1] = starts[i] + path.ids.size();
     }
   }
 }
@@ -1364,25 +1359,76 @@ void RecurrentGradientMachine::copyDataOutlinkFrame(size_t machineCur) {
   }
 }
 
-void RecurrentGradientMachine::createDataOutlink(
-    std::vector<int>& machineIdVec) {
-  size_t seqNum =
-      getBeamSize() > 1UL ? finalPaths_.size() : finalPaths_[0].size();
-  std::vector<int> starts(seqNum + 1, 0);
-  for (size_t i = 0; i < seqNum; ++i) {
-    size_t seqLen = getBeamSize() > 1UL ? finalPaths_[i][0].ids.size()
-                                        : finalPaths_[0][i].ids.size();
-    starts[i + 1] = starts[i] + seqLen;
+void RecurrentGradientMachine::createDataOutlinkSelRowsInfo(
+    bool isSeq, std::vector<Argument>& outArgs) {
+  batchMachineIdVec_.clear();
+
+  size_t seqIdx = 0;
+  for (size_t i = 0; i < finalPaths_.size(); ++i) {
+    for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
+      std::vector<int>& machineIdVec = finalPaths_[i][j].machineIdVec;
+      if (isSeq) {
+        for (size_t i = 0; i < machineIdVec.size(); ++i) {
+          size_t rowId = machineIdVec[i];
+          int* seqPos =
+              outArgs[i].sequenceStartPositions->getMutableData(false);
+          batchMachineIdVec_.push_back(seqPos[rowId]);
+        }
+      } else {
+        batchMachineIdVec_.insert(
+            batchMachineIdVec_.end(), machineIdVec.begin(), machineIdVec.end());
+      }
+      seqIdx++;
+    }
+  }
+}
+
+void RecurrentGradientMachine::createDataOutlinkCopySizeInfo(
+    bool isSeq, std::vector<Argument>& outArgs, std::vector<int>& copySize) {
+  size_t totalSeqNum = std::accumulate(
+      finalPaths_.begin(),
+      finalPaths_.end(),
+      0UL,
+      [](size_t a, const std::vector<Path>& b) { return a + b.size(); });
+  copySize.resize(totalSeqNum, 1);
+
+  batchMachineStartPos_.resize(totalSeqNum + 1, 0);
+  if (isSeq) {
+    ICpuGpuVectorPtr inputSeqStartPos = outArgs[0].sequenceStartPositions;
+    CHECK_EQ(static_cast<size_t>(inputSeqStartPos->getSize() - 1),
+             getBeamSize() > 1 ? finalPaths_.size() : finalPaths_[0].size());
+    int* starts = inputSeqStartPos->getMutableData(false);
+    int seqId = 0;
+    for (int i = 0; i < finalPaths_.size(); ++i) {
+      for (int j = 0; j < finalPaths_[i].size(); ++j) {
+        copySize[seqId] = getBeamSize() > 1 ? starts[i + 1] - starts[i]
+                                            : starts[j + 1] - starts[j];
+        batchMachineStartPos_[seqId + 1] =
+            batchMachineStartPos_[seqId] + finalPaths_[i][j].ids.size();
+        seqId++;
+      }
+    }
+  } else {
+    for (size_t i = 0; i < finalPaths_[0].size(); ++i)
+      batchMachineStartPos_[i + 1] =
+          batchMachineStartPos_[i] + finalPaths_[0][i].ids.size();
   }
+}
 
+void RecurrentGradientMachine::createDataOutlink() {
   for (size_t i = 0; i < dataArgsSize_; i++) {
+    bool isSeq = dataArgsFrame_[i][0].hasSeq();
+    std::vector<int> copySize;
+    createDataOutlinkCopySizeInfo(isSeq, dataArgsFrame_[i], copySize);
+    createDataOutlinkSelRowsInfo(isSeq, dataArgsFrame_[i]);
+
     dataArgs_[i].concat(dataArgsFrame_[i],
-                        machineIdVec,
-                        starts,
+                        batchMachineIdVec_,
+                        batchMachineStartPos_,
+                        copySize,
                         useGpu_,
                         HPPL_STREAM_1,
                         PASS_TEST);
-
     auto dataAgent =
         dynamic_cast<DataLayer*>(outFrameLines_[i + 1].agentLayer.get());
     CHECK_NOTNULL(dataAgent);

paddle/gserver/gradientmachines/RecurrentGradientMachine.h

@@ -190,7 +190,7 @@ public:
     std::vector<int> ids;
 
     /**
-     * @brief idsProb, log probability of each generated words.
+     * @brief idsProb, log probability of each generated word.
      */
     std::vector<real> idsProb;
 
@@ -472,15 +472,43 @@ private:
   void copyDataOutlinkFrame(size_t machineCur);
 
   /*
-   * @brief In generation, if the layer group has more than 1 outlink, outlinks
-   * except the first one are data outlinks. This function creates the data
-   * outlinks.
-   * @note In beam search, only one generated sequence with the hightest log
-   * probabilites are retained.
-   * @param machineIdVec : select a row of output matrix in each frame
-   * that the generation process expanded.
+   * @brief In generation, if the layer group has more than 1 outlink, outlink
+   * except the first one is a data outlink. In RecurrentLayerGroup, each time
+   * step is a separate Network, outputs of a layer inside the
+   * RecurrentLayerGroup are stored in separate Arguments. If one layer is
+   * specified as an outlink of RecurrentLayerGroup. This function will
+   * collect outputs in each time step of each generated sequence which are
+   * dispersed in separate Arguments to form a new single Argument as output of
+   * RecurrentLayerGroup.
    */
-  void createDataOutlink(std::vector<int>& machineIdVec);
+  void createDataOutlink();
+
+  /*
+   * @brief decide to select how many rows from the Matrix stored the forward
+   * pass results from a start position.
+   *
+   * @param isSeq: a flag indicating whetehr the layer to be output of the
+   * RecurrentGradientMachine is a sequence or not
+   * @param outArgs: all of the the returned Arguments of the forward pass
+   * during the generation process.
+   * @param copySize: the returned result, number of rows to select from the
+   * Matrix stored the forward pass results from a start position.
+   */
+  void createDataOutlinkCopySizeInfo(bool isSeq,
+                                     std::vector<Argument>& outArgs,
+                                     std::vector<int>& copySize);
+
+  /*
+   * @brief decide index of the start row for each time step of a generated
+   * sequence in Matrix stored the entire beam search batch's forward pass
+   * results.
+   *
+   * @param isSeq: a flag indicating whether the layer to be output of the
+   * RecurrentGradientMachine is a sequence or not
+   * @param outArgs: all of the returned Arguments of the forward pass
+   * during the generation process.
+   */
+  void createDataOutlinkSelRowsInfo(bool isSeq, std::vector<Argument>& outArgs);
 
   /*
    * @brief used in beam search, connect previous frame to form recurrent link
@@ -543,6 +571,7 @@ private:
   std::vector<int> topIds_;
   std::vector<int> seqIds_;
   std::vector<int> batchMachineIdVec_;
+  std::vector<int> batchMachineStartPos_;
   std::vector<std::vector<Path>> finalPaths_;
   std::vector<real> minFinalPathLogProb_;
   BeamSearchControlCallbacks* beamSearchCtrlCallbacks_;

paddle/gserver/layers/ScaleShiftLayer.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "Layer.h"
+
+namespace paddle {
+
+/**
+ * A layer applies a linear transformation to each element in each row of
+ * the input matrix. For each element, the layer first re-scale it and then
+ * adds a bias to it.
+ *
+ * \f[
+ *    y = wx + b
+ * \f]
+ *
+ * Here, w is the scale and b is the bias. Both w and b are trainable scalars.
+ *
+ */
+
+class ScaleShiftLayer : public Layer {
+protected:
+  std::unique_ptr<Weight> scale_;
+  std::unique_ptr<Weight> offset_;
+
+public:
+  explicit ScaleShiftLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+
+REGISTER_LAYER(scale_shift, ScaleShiftLayer);
+
+bool ScaleShiftLayer::init(const LayerMap& layerMap,
+                           const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK_EQ(inputLayers_.size(), 1U);
+  scale_.reset(new Weight(1, 1, parameters_[0]));
+  if (biasParameter_.get() != NULL) {
+    offset_ = std::unique_ptr<Weight>(new Weight(1, 1, biasParameter_));
+  }
+  return true;
+}
+
+void ScaleShiftLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  MatrixPtr inV = getInputValue(0);
+  resetOutput(inV->getHeight(), inV->getWidth());
+  MatrixPtr outV = getOutputValue();
+  real scaleValue = scale_->getW()->getElement(0, 0);
+  outV->mulScalar(*inV, scaleValue);
+  if (offset_) {
+    real offsetValue = offset_->getW()->getElement(0, 0);
+    outV->add(offsetValue);
+  }
+}
+
+void ScaleShiftLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inV = getInputValue(0);
+  MatrixPtr inG = getInputGrad(0);
+  MatrixPtr outV = getOutputValue();
+  MatrixPtr outG = getOutputGrad();
+
+  /* Calculate the parameter gradient for the current layer */
+  if (scale_->getWGrad()) {
+    MatrixPtr rowSumMtx;
+    Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_);
+    // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ij} * c_{ij}
+    rowSumMtx->sumOfProducts(
+        /* b= */ *inV, /* c= */ *outG, /* scaleSum= */ 1, /* scaleDest= */ 0.);
+    // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ji}
+    scale_->getWGrad()->sumCols(
+        /* b= */ *rowSumMtx, /* scaleSum= */ 1., /* scaleDest= */ 1.);
+    scale_->getParameterPtr()->incUpdate(callback);
+  }
+  if (offset_ && offset_->getWGrad()) {
+    MatrixPtr rowSumMtx;
+    Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_);
+    rowSumMtx->sumRows(*outG, 1., 0.);
+    offset_->getWGrad()->sumCols(*rowSumMtx, 1., 1.);
+    offset_->getParameterPtr()->incUpdate(callback);
+  }
+
+  /* Calculate the input layers error */
+  if (inG) {
+    real scaleValue = scale_->getW()->getElement(0, 0);
+    inG->add(*outG, scaleValue);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -2007,6 +2007,21 @@ TEST(Layer, RowL2NormLayer) {
   }
 }
 
+TEST(Layer, ScaleShiftLayer) {
+  const size_t batchSize = 16;
+  const size_t size = 32;
+  TestConfig config;
+  config.layerConfig.set_type("scale_shift");
+  config.layerConfig.set_size(size);
+  config.biasSize = 1;
+  config.inputDefs.push_back(
+      {INPUT_DATA, "input", /* dim= */ size, /* paraSize= */ 1});
+  config.layerConfig.add_inputs();
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "scale_shift", batchSize, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/gserver/tests/test_NetworkCompare.cpp

@@ -269,7 +269,8 @@ TEST(Compare, img_conv2) {
   bool useGpu = FLAGS_use_gpu;
   double eps = FLAGS_checkgrad_eps;
   FLAGS_use_gpu = true;
-  FLAGS_checkgrad_eps = 1e-2;
+  // Sometimes, this unit test will fail with 1e-2
+  FLAGS_checkgrad_eps = 4e-2;
   compareNetwork(config_file_a, config_file_b);
   FLAGS_use_gpu = useGpu;
   FLAGS_checkgrad_eps = eps;

paddle/operators/CMakeLists.txt

@@ -43,6 +43,7 @@ endfunction()
 
 add_subdirectory(math)
 cc_test(gather_test SRCS gather_test.cc DEPS tensor)
+op_library(gather_op SRCS gather_op.cc gather_op.cu)
 
 cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
 

paddle/operators/cross_entropy_op.cc

@@ -39,11 +39,10 @@ class OnehotCrossEntropyGradientOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto X_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto X = ctx.Input<Tensor>("X");
 
-    // TODO(superjom) add enforce here after helper functions ready
-    X_grad->Resize(X->dims());
+    dX->Resize(X->dims());
   }
 };
 
@@ -70,9 +69,7 @@ namespace ops = paddle::operators;
 REGISTER_OP(onehot_cross_entropy, ops::OnehotCrossEntropyOp,
             ops::OnehotCrossEntropyOpMaker, onehot_cross_entropy_grad,
             ops::OnehotCrossEntropyGradientOp);
-REGISTER_OP_CPU_KERNEL(
-    onehot_cross_entropy,
-    ops::OnehotCrossEntropyOpKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(
-    onehot_cross_entropy_grad,
-    ops::OnehotCrossEntropyGradientOpKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(onehot_cross_entropy,
+                       ops::OnehotCrossEntropyOpKernel<float>);
+REGISTER_OP_CPU_KERNEL(onehot_cross_entropy_grad,
+                       ops::OnehotCrossEntropyGradientOpKernel<float>);

paddle/operators/cross_entropy_op.cu

@@ -12,10 +12,122 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#define EIGEN_USE_GPU
-#include "paddle/operators/cross_entropy_op.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/platform/assert.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T>
+__host__ __device__ T clipping_log(const T x) {
+  PADDLE_ASSERT(std::is_floating_point<T>::value);
+  const T kApproInf = 1e20;
+  T v = log(x);
+  if (v == INFINITY) {
+    return kApproInf;
+  }
+  if (v == -INFINITY) {
+    return -kApproInf;
+  }
+  return v;
+}
+
+template <typename T>
+__global__ void CrossEntropyKernel(T* Y, const T* X, const int* label,
+                                   const int N, const int D) {
+  // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file.
+  // CUDA_1D_KERNEL_LOOP(i, N) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
+       i += blockDim.x * gridDim.x) {
+    PADDLE_ASSERT(label[i] >= 0 && label[i] < D);
+    Y[i] = -clipping_log(X[i * D + label[i]]);
+  }
+}
+
+// TODO(qingqing): make zero setting an common function.
+template <typename T>
+__global__ void zero(T* X, const int N) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
+       i += blockDim.x * gridDim.x) {
+    X[i] = 0.0;
+  }
+}
+
+template <typename T>
+__global__ void CrossEntropyGradientKernel(T* dX, const T* dY, const T* X,
+                                           const int* label, const int N,
+                                           const int D) {
+  // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file.
+  // CUDA_1D_KERNEL_LOOP(i, N) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
+       i += blockDim.x * gridDim.x) {
+    int idx = i * D + label[i];
+    dX[idx] = -dY[i] / X[idx];
+  }
+}
+
+template <typename T>
+class OnehotCrossEntropyOpCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use GPUPlace.");
+
+    auto X = ctx.Input<Tensor>("X");
+    const T* Xdata = X->data<T>();
+    const int* label_data = ctx.Input<Tensor>("label")->data<int>();
+    auto Y = ctx.Output<Tensor>("Y");
+    Y->mutable_data<T>(ctx.GetPlace());
+    T* Ydata = Y->data<T>();
+
+    int N = X->dims()[0];
+    int D = X->dims()[1];
+    int block = 512;
+    int grid = (N + block - 1) / block;
+    // TODO(qingqing) launch kernel on specified stream
+    // base on ExecutionContext.
+    CrossEntropyKernel<T><<<grid, block>>>(Ydata, Xdata, label_data, N, D);
+  }
+};
+
+template <typename T>
+class OnehotCrossEntropyGradientOpCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use GPUPlace.");
+
+    auto X = ctx.Input<Tensor>("X");
+    auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto dY = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    auto label = ctx.Input<Tensor>("label");
+
+    auto* dXdata = dX->template mutable_data<T>(ctx.GetPlace());
+    auto* dYdata = dY->template data<T>();
+    auto* Xdata = X->template data<T>();
+    auto* label_data = label->data<int>();
+
+    int N = X->dims()[0];
+    int D = X->dims()[1];
+    int block = 512;
+    int grid = (N * D + block - 1) / block;
+    zero<T><<<grid, block>>>(dXdata, N * D);
+
+    grid = (N + block - 1) / block;
+    // TODO(qingqing): launch kernel on specified stream
+    // base on ExecutionContext.
+    CrossEntropyGradientKernel<T><<<grid, block>>>(dXdata, dYdata, Xdata,
+                                                   label_data, N, D);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(
-    onehot_cross_entropy,
-    ops::OnehotCrossEntropyOpKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(onehot_cross_entropy,
+                       ops::OnehotCrossEntropyOpCUDAKernel<float>);
+REGISTER_OP_GPU_KERNEL(onehot_cross_entropy_grad,
+                       ops::OnehotCrossEntropyGradientOpCUDAKernel<float>);

paddle/operators/cross_entropy_op.h

@@ -21,7 +21,7 @@ namespace operators {
 using Tensor = framework::Tensor;
 
 template <typename T>
-T tolerable_value(T x) {
+inline T tolerable_value(const T x) {
   static_assert(std::is_floating_point<T>::value,
                 "tolerable_value works only on float, "
                 "double and double double.");
@@ -39,10 +39,13 @@ T tolerable_value(T x) {
   return x;
 }
 
-template <typename Place, typename T>
+template <typename T>
 class OnehotCrossEntropyOpKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
+                   "It must use CPUPlace.");
+
     auto X = ctx.Input<Tensor>("X");
     const T* Xdata = X->data<T>();
     const int* label_data = ctx.Input<Tensor>("label")->data<int>();
@@ -62,10 +65,13 @@ class OnehotCrossEntropyOpKernel : public framework::OpKernel {
   }
 };
 
-template <typename Place, typename T>
+template <typename T>
 class OnehotCrossEntropyGradientOpKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
+                   "It must use CPUPlace.");
+
     auto X = ctx.Input<Tensor>("X");
     auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto dY = ctx.Input<Tensor>(framework::GradVarName("Y"));
@@ -79,6 +85,8 @@ class OnehotCrossEntropyGradientOpKernel : public framework::OpKernel {
     const int batch_size = X->dims()[0];
     const int class_num = X->dims()[1];
 
+    // TODO(qingqing): make zero setting an common function.
+    memset(dXdata, 0, sizeof(T) * batch_size * class_num);
     for (int i = 0; i < batch_size; ++i) {
       int index = i * class_num + label_data[i];
       dXdata[index] = -tolerable_value(dYdata[i] / Xdata[index]);

paddle/operators/gather.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <cstring>
 
 #include "paddle/framework/ddim.h"
+#include "paddle/framework/eigen.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/place.h"
 
@@ -25,13 +26,13 @@ namespace operators {
 
 // Implementation of CPU copy
 template <typename T>
-void CPUGather(const T* params, const int* indices, const int slice_size,
+void CPUGather(const T* src, const int* indices, const int slice_size,
                const int index_size, T* output) {
   const size_t slice_bytes = slice_size * sizeof(T);
 
   for (int i = 0; i < index_size; ++i) {
     int index_ = indices[i];
-    memcpy(output + i * slice_size, params + index_ * slice_size, slice_bytes);
+    memcpy(output + i * slice_size, src + index_ * slice_size, slice_bytes);
   }
 }
 
@@ -55,7 +56,7 @@ void Gather(const platform::Place& place, const paddle::framework::Tensor* src,
   int index_size = index->dims()[0];
 
   auto src_dims = src->dims();
-  paddle::framework::DDim output_dims(src_dims);
+  framework::DDim output_dims(src_dims);
   output_dims[0] = index_size;
 
   // slice size

paddle/operators/gather_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/gather_op.h"
+#include "paddle/framework/ddim.h"
+
+namespace paddle {
+namespace operators {
+
+class GatherOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    int batch_size = ctx.Input<Tensor>("Index")->dims()[0];
+    PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0");
+    framework::DDim output_dims(ctx.Input<Tensor>("X")->dims());
+    output_dims[0] = batch_size;
+    ctx.Output<Tensor>("Out")->Resize(output_dims);
+  }
+};
+
+class GatherGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto X_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto X = ctx.Input<Tensor>("X");
+
+    X_grad->Resize(X->dims());
+  }
+};
+
+class GatherOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  GatherOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The source input of gather op");
+    AddInput("Index", "The index input of gather op");
+    AddOutput("Out", "The output of add op");
+    AddComment(R"DOC(
+Gather Operator by selecting from the first axis, 
+
+Out = X[Index]
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(gather, ops::GatherOp, ops::GatherOpMaker, gather_grad,
+            ops::GatherGradOp);
+REGISTER_OP_CPU_KERNEL(gather,
+                       ops::GatherOpKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    gather_grad,
+    ops::GatherGradientOpKernel<paddle::platform::CPUPlace, float>);

paddle/operators/gather_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/gather_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(gather,
+                       ops::GatherOpKernel<paddle::platform::GPUPlace, float>);

paddle/operators/gather_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "gather.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "scatter.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename Place, typename T>
+class GatherOpKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto *X = ctx.Input<Tensor>("X");
+    auto *Index = ctx.Input<Tensor>("Index");
+    auto *Y = ctx.Output<Tensor>("Out");
+
+    Y->mutable_data<T>(ctx.GetPlace());
+    Gather<T>(ctx.GetPlace(), X, Index, Y);
+  }
+};
+
+template <typename Place, typename T>
+class GatherGradientOpKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto *Index = ctx.Input<Tensor>("Index");
+    auto *dX = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto *dO = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+    dX->mutable_data<T>(ctx.GetPlace());
+    ScatterUpdate<T>(ctx.GetPlace(), dO, Index, dX);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/gaussian_random_op.cc

 /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.
    You may obtain a copy of the License at
-
    http://www.apache.org/licenses/LICENSE-2.0
-
    Unless required by applicable law or agreed to in writing, software
    distributed under the License is distributed on an "AS IS" BASIS,
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
@@ -19,25 +16,25 @@ namespace paddle {
 namespace operators {
 
 template <typename T>
-class GaussianRandomKernel : public framework::OpKernel {
+class CPUGaussianRandomKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     float mean = context.op_.GetAttr<float>("mean");
     float std = context.op_.GetAttr<float>("std");
-    auto* tensor = context.Output<framework::Tensor>(0);
+    auto* tensor = context.Output<framework::Tensor>("Out");
     T* data = tensor->mutable_data<T>(context.GetPlace());
 
-    // TODO(dzh): attribute does not support unsigned int.
-    // And we need a global random seed configuration.
-    int seed = context.op_.GetAttr<int>("seed");
+    unsigned int seed =
+        static_cast<unsigned int>(context.op_.GetAttr<int>("seed"));
+    std::minstd_rand engine;
     if (seed == 0) {
       seed = std::random_device()();
     }
-    std::mt19937 g(seed);
-    std::normal_distribution<T> distribution(mean, std);
+    engine.seed(seed);
+    std::normal_distribution<T> dist(mean, std);
     ssize_t size = framework::product(tensor->dims());
-    for (int i = 0; i < size; ++i) {
-      data[i] = distribution(g);
+    for (ssize_t i = 0; i < size; ++i) {
+      data[i] = dist(engine);
     }
   }
 };
@@ -48,7 +45,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& context) const override {
-    auto* tensor = context.Output<framework::Tensor>(0);
+    auto* tensor = context.Output<framework::Tensor>("Out");
     auto dims = GetAttr<std::vector<int>>("dims");
     PADDLE_ENFORCE(dims.size() > 0UL,
                    "dims can be one int or array. dims must be set.");
@@ -68,8 +65,8 @@ Use to initialize tensor with gaussian random generator.
 )DOC");
 
     AddAttr<std::vector<int>>("dims", "The dimension of random tensor.");
-    AddAttr<float>("mean", "mean value of random.").SetDefault(.0f);
-    AddAttr<float>("std", "minimum value of random value.").SetDefault(1.0f);
+    AddAttr<float>("mean", "mean of random tensor.").SetDefault(.0f);
+    AddAttr<float>("std", "std of random tensor.").SetDefault(1.0f);
     AddAttr<int>("seed",
                  "Random seed of generator."
                  "0 means use system wide seed")
@@ -83,4 +80,4 @@ Use to initialize tensor with gaussian random generator.
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(gaussian_random, ops::GaussianRandomOp,
                              ops::GaussianRandomOpMaker);
-REGISTER_OP_CPU_KERNEL(gaussian_random, ops::GaussianRandomKernel<float>);
+REGISTER_OP_CPU_KERNEL(gaussian_random, ops::CPUGaussianRandomKernel<float>);

paddle/operators/gaussian_random_op.cu

 /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.
    You may obtain a copy of the License at
-
    http://www.apache.org/licenses/LICENSE-2.0
-
    Unless required by applicable law or agreed to in writing, software
    distributed under the License is distributed on an "AS IS" BASIS,
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include <memory>
-#include <random>
-#include "paddle/platform/dynload/curand.h"
-#include "paddle/platform/gpu_info.h"
-
+#include <thrust/device_ptr.h>
+#include <thrust/iterator/counting_iterator.h>
+#include <thrust/random.h>
+#include <thrust/transform.h>
 #include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
 
 namespace paddle {
 namespace operators {
 
 template <typename T>
-class GaussianRandomKernel : public framework::OpKernel {
+struct GaussianGenerator {
+  T mean_, std_;
+  unsigned int seed_;
+
+  __host__ __device__ GaussianGenerator(T mean, T std, int seed)
+      : mean_(mean), std_(std), seed_(seed) {}
+
+  __host__ __device__ T operator()(const unsigned int n) const {
+    thrust::minstd_rand rng;
+    rng.seed(seed_);
+    thrust::normal_distribution<T> dist(mean_, std_);
+    rng.discard(n);
+    return dist(rng);
+  }
+};
+
+template <typename T>
+class GPUGaussianRandomKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    float mean = context.op_.GetAttr<float>("mean");
-    float std = context.op_.GetAttr<float>("std");
-    auto* tensor = context.Output<framework::Tensor>(0);
+    auto* tensor = context.Output<framework::Tensor>("Out");
     T* data = tensor->mutable_data<T>(context.GetPlace());
-
-    int seed = context.op_.GetAttr<int>("seed");
+    unsigned int seed =
+        static_cast<unsigned int>(context.op_.GetAttr<int>("seed"));
     if (seed == 0) {
       std::random_device rd;
       seed = rd();
     }
-    curandGenerator_t g;
-    PADDLE_ENFORCE(platform::dynload::curandCreateGenerator(
-        &g, CURAND_RNG_PSEUDO_DEFAULT));
-    PADDLE_ENFORCE(
-        platform::dynload::curandSetPseudoRandomGeneratorSeed(g, seed));
-    platform::dynload::curandGenerateNormal(
-        g, data, framework::product(tensor->dims()), mean, std);
+    T mean = static_cast<T>(context.op_.GetAttr<float>("mean"));
+    T std = static_cast<T>(context.op_.GetAttr<float>("std"));
+    thrust::counting_iterator<unsigned int> index_sequence_begin(0);
+    ssize_t N = framework::product(tensor->dims());
+    thrust::transform(index_sequence_begin, index_sequence_begin + N,
+                      thrust::device_ptr<T>(data),
+                      GaussianGenerator<T>(mean, std, seed));
   }
 };
 
 }  // namespace operators
 }  // namespace paddle
 
-namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(gaussian_random, ops::GaussianRandomKernel<float>);
+REGISTER_OP_GPU_KERNEL(gaussian_random,
+                       paddle::operators::GPUGaussianRandomKernel<float>);

paddle/operators/mul_op.cc

@@ -13,7 +13,6 @@
    limitations under the License. */
 
 #include "paddle/operators/mul_op.h"
-#include "paddle/operators/math/math_function.h"
 
 namespace paddle {
 namespace operators {

paddle/operators/net_op.cc

@@ -86,9 +86,8 @@ std::vector<std::string> NetOp::OutputVars(bool has_intermediate) const {
   return ret_val;
 }
 
-NetOp::NetOp(const std::string& type,
-             const framework::OperatorBase::VarNameMap& inputs,
-             const framework::OperatorBase::VarNameMap& outputs,
+NetOp::NetOp(const std::string& type, const framework::VariableNameMap& inputs,
+             const framework::VariableNameMap& outputs,
              const framework::AttributeMap& attrs)
     : framework::OperatorBase(type, inputs, outputs, attrs) {}
 

paddle/operators/net_op.h

@@ -38,8 +38,10 @@ class NetOp : public framework::OperatorBase {
  public:
   static const char kAll[];
   NetOp() : framework::OperatorBase("plain_net", {}, {}, {}) {}
-  NetOp(const std::string& type, const VarNameMap& inputs,
-        const VarNameMap& outputs, const framework::AttributeMap& attrs);
+
+  NetOp(const std::string& type, const framework::VariableNameMap& inputs,
+        const framework::VariableNameMap& outputs,
+        const framework::AttributeMap& attrs);
 
   NetOp(const NetOp& o) : framework::OperatorBase(o.type_, {}, {}, o.attrs_) {
     this->ops_.reserve(o.ops_.size());
@@ -84,13 +86,14 @@ class NetOp : public framework::OperatorBase {
     return true;
   }
 
-  void AddOp(const framework::OperatorBase& op) { AddOp(op.Clone()); }
+  void AppendOp(const framework::OperatorBase& op) { AppendOp(op.Clone()); }
 
   /**
    * @brief Add an operator by ptr
    */
-  void AddOp(std::unique_ptr<framework::OperatorBase> op) {
-    PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed");
+  void AppendOp(std::unique_ptr<framework::OperatorBase> op) {
+    PADDLE_ENFORCE(!add_op_done_,
+                   "Cannot AppendOp when this network is sealed");
     PADDLE_ENFORCE_NOT_NULL(op, "Cannot Insert Null op");
     ops_.push_back(std::move(op));
   }

paddle/operators/net_op_test.cc

@@ -38,10 +38,10 @@ TEST(OpKernel, all) {
   auto net = std::make_shared<NetOp>();
   ASSERT_NE(net, nullptr);
 
-  net->AddOp(std::unique_ptr<TestOp>(
+  net->AppendOp(std::unique_ptr<TestOp>(
       new TestOp("test", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
                  {{"Out", {"y"}}}, {})));
-  net->AddOp(std::unique_ptr<TestOp>(
+  net->AppendOp(std::unique_ptr<TestOp>(
       new TestOp("test", {{"X", {"y"}}, {"W", {"w2"}}, {"b", {"b2"}}},
                  {{"Out", {"z"}}}, {})));
 
@@ -61,7 +61,7 @@ TEST(NetOp, insert_op) {
   auto op1 = std::unique_ptr<framework::NOP>(
       new framework::NOP("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
                          {{"Out", {"y"}}}, {}));
-  net.AddOp(*op1);
+  net.AppendOp(*op1);
   net.InsertOp(0, *op1);
   ASSERT_EQ(2UL, net.ops_.size());
   net.InsertOp(2, std::move(op1));
@@ -70,9 +70,9 @@ TEST(NetOp, insert_op) {
 
 TEST(NetOp, Clone) {
   NetOp net;
-  net.AddOp(
+  net.AppendOp(
       std::unique_ptr<framework::NOP>(new framework::NOP{"empty", {}, {}, {}}));
-  net.AddOp(std::unique_ptr<framework::NOP>(
+  net.AppendOp(std::unique_ptr<framework::NOP>(
       new framework::NOP{"empty2", {}, {}, {}}));
   net.CompleteAddOp(true);
   auto new_net_op = net.Clone();

paddle/operators/recurrent_op.cc

@@ -131,8 +131,8 @@ const rnn::ArgumentName RecurrentGradientOp::kArgName{
     "memories",    "pre_memories", "boot_memories@grad"};
 
 RecurrentOp::RecurrentOp(const std::string& type,
-                         const framework::OperatorBase::VarNameMap& inputs,
-                         const framework::OperatorBase::VarNameMap& outputs,
+                         const framework::VariableNameMap& inputs,
+                         const framework::VariableNameMap& outputs,
                          const framework::AttributeMap& attrs)
     : OperatorBase(type, inputs, outputs, attrs) {
   rnn::InitArgument(kArgName, &arg_, *this);
@@ -223,8 +223,8 @@ void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const {
 }
 
 RecurrentGradientOp::RecurrentGradientOp(
-    const std::string& type, const framework::OperatorBase::VarNameMap& inputs,
-    const framework::OperatorBase::VarNameMap& outputs,
+    const std::string& type, const framework::VariableNameMap& inputs,
+    const framework::VariableNameMap& outputs,
     const framework::AttributeMap& attrs)
     : OperatorBase(type, inputs, outputs, attrs) {
   rnn::InitArgument(kArgName, &arg_, *this);

paddle/operators/recurrent_op.h

@@ -114,8 +114,9 @@ class RecurrentGradientAlgorithm {
 
 class RecurrentOp : public framework::OperatorBase {
  public:
-  RecurrentOp(const std::string& type, const VarNameMap& inputs,
-              const VarNameMap& outputs, const framework::AttributeMap& attrs);
+  RecurrentOp(const std::string& type, const framework::VariableNameMap& inputs,
+              const framework::VariableNameMap& outputs,
+              const framework::AttributeMap& attrs);
 
   RecurrentOp(const RecurrentOp& o)
       : framework::OperatorBase(
@@ -150,8 +151,9 @@ class RecurrentOp : public framework::OperatorBase {
 
 class RecurrentGradientOp : public framework::OperatorBase {
  public:
-  RecurrentGradientOp(const std::string& type, const VarNameMap& inputs,
-                      const VarNameMap& outputs,
+  RecurrentGradientOp(const std::string& type,
+                      const framework::VariableNameMap& inputs,
+                      const framework::VariableNameMap& outputs,
                       const framework::AttributeMap& attrs);
 
   RecurrentGradientOp(const RecurrentGradientOp& o)

paddle/operators/rowwise_add_op.h

 /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
 
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
     http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License. */
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
 
 #pragma once
 #include "paddle/framework/eigen.h"
@@ -63,7 +63,7 @@ class RowwiseAddGradKernel : public framework::OpKernel {
 
     // https://eigen.tuxfamily.org/dox/unsupported/TensorBase_8h_source.html
     // colwise add
-    Eigen::array<int, 1> dims{{1}}; /* dimension to reduce */
+    Eigen::array<int, 1> dims{{0}}; /* dimension to reduce */
     EigenVector<T>::Flatten(*db).device(place) = OutGrad.sum(dims);
   }
 };

paddle/operators/scale_op.cc

@@ -20,8 +20,9 @@ namespace operators {
 
 class ScaleOp : public framework::OperatorWithKernel {
  public:
-  ScaleOp(const std::string &type, const VarNameMap &inputs,
-          const VarNameMap &outputs, const framework::AttributeMap &attrs)
+  ScaleOp(const std::string &type, const framework::VariableNameMap &inputs,
+          const framework::VariableNameMap &outputs,
+          const framework::AttributeMap &attrs)
       : OperatorWithKernel(type, inputs, outputs, attrs) {}
 
  protected:
@@ -52,10 +53,11 @@ The equation is: Out = scale*X
 template <typename AttrType>
 class ScaleGradOp : public NetOp {
  public:
-  ScaleGradOp(const std::string &type, const VarNameMap &inputs,
-              const VarNameMap &outputs, const framework::AttributeMap &attrs)
+  ScaleGradOp(const std::string &type, const framework::VariableNameMap &inputs,
+              const framework::VariableNameMap &outputs,
+              const framework::AttributeMap &attrs)
       : NetOp(type, inputs, outputs, attrs) {
-    AddOp(framework::OpRegistry::CreateOp(
+    AppendOp(framework::OpRegistry::CreateOp(
         "scale", {{"X", {Input(framework::GradVarName("Out"))}}},
         {{"Out", {Output(framework::GradVarName("X"))}}},
         {{"scale", GetAttr<AttrType>("scale")}}));
@@ -80,10 +82,11 @@ class IdentityOpMaker : public framework::OpProtoAndCheckerMaker {
 template <typename AttrType>
 class IdentityOp : public NetOp {
  public:
-  IdentityOp(const std::string &type, const VarNameMap &inputs,
-             const VarNameMap &outputs, const framework::AttributeMap &attrs)
+  IdentityOp(const std::string &type, const framework::VariableNameMap &inputs,
+             const framework::VariableNameMap &outputs,
+             const framework::AttributeMap &attrs)
       : NetOp(type, inputs, outputs, attrs) {
-    AddOp(framework::OpRegistry::CreateOp(
+    AppendOp(framework::OpRegistry::CreateOp(
         "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}},
         {{"scale", static_cast<AttrType>(1)}}));
   }

paddle/operators/uniform_random_op.cc

 /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.
    You may obtain a copy of the License at
-
    http://www.apache.org/licenses/LICENSE-2.0
-
    Unless required by applicable law or agreed to in writing, software
    distributed under the License is distributed on an "AS IS" BASIS,
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
@@ -39,7 +36,8 @@ class CPUUniformRandomKernel : public framework::OpKernel {
     std::uniform_real_distribution<T> dist(
         static_cast<T>(context.op_.GetAttr<float>("min")),
         static_cast<T>(context.op_.GetAttr<float>("max")));
-    for (ssize_t i = 0; i < framework::product(tensor->dims()); ++i) {
+    ssize_t size = framework::product(tensor->dims());
+    for (ssize_t i = 0; i < size; ++i) {
       data[i] = dist(engine);
     }
   }
@@ -66,7 +64,6 @@ class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker {
       : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddOutput("Out", "The output tensor of uniform random op");
     AddComment(R"DOC(Uniform random operator.
-
 Used to initialize tensor with uniform random generator.
 )DOC");
     AddAttr<std::vector<int>>("dims", "the dimension of random tensor");

paddle/operators/uniform_random_op.cu

 /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.
    You may obtain a copy of the License at
-
    http://www.apache.org/licenses/LICENSE-2.0
-
    Unless required by applicable law or agreed to in writing, software
    distributed under the License is distributed on an "AS IS" BASIS,
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

paddle/parameter/Argument.cpp

@@ -276,17 +276,21 @@ int32_t Argument::resizeAndCopyFrom(const Argument& src,
 void Argument::concat(const std::vector<Argument>& args,
                       const std::vector<int>& selectRows,
                       const std::vector<int>& seqStartPos,
+                      const std::vector<int>& copySize,
                       bool useGpu,
                       hl_stream_t stream,
                       PassType passType) {
   CHECK(!subSequenceStartPositions)
       << "undefined behavior for subsequence positions";
 
-  size_t batchSize = selectRows.size();
+  size_t batchSize = 0;
+  for (size_t i = 0; i < copySize.size(); ++i)
+    batchSize += copySize[i] * (seqStartPos[i + 1] - seqStartPos[i]);
+
   auto copyArg = [batchSize, stream](MatrixPtr& dst,
                                      MatrixPtr src,
-                                     int startRow,
-                                     int pos,
+                                     int desStartRow,
+                                     int srcStartRow,
                                      int size,
                                      bool useGpu) {
     if (!src) {
@@ -300,14 +304,14 @@ void Argument::concat(const std::vector<Argument>& args,
       dst->resize(batchSize, width);
     }
 
-    MatrixPtr tmpMatrix = dst->subMatrix(startRow, size);
-    tmpMatrix->copyFrom(*src->subMatrix(pos, size), stream);
+    MatrixPtr tmpMatrix = dst->subMatrix(desStartRow, size);
+    tmpMatrix->copyFrom(*src->subMatrix(srcStartRow, size), stream);
   };
 
   auto copyIds = [batchSize, stream](IVectorPtr& dst,
                                      const IVectorPtr& src,
-                                     int startRow,
-                                     int pos,
+                                     int desStartRow,
+                                     int srcStartRow,
                                      int size,
                                      bool useGpu) {
     if (!src) {
@@ -315,13 +319,14 @@ void Argument::concat(const std::vector<Argument>& args,
       return;
     }
     IVector::resizeOrCreate(dst, batchSize, useGpu);
-    dst->subVec(startRow, size)->copyFrom(*src->subVec(pos, size), stream);
+    dst->subVec(desStartRow, size)
+        ->copyFrom(*src->subVec(srcStartRow, size), stream);
   };
 
   auto copyStrs = [batchSize, stream](SVectorPtr& dst,
                                       const SVectorPtr& src,
-                                      int startRow,
-                                      int pos,
+                                      int desStartRow,
+                                      int srcStartRow,
                                       int size,
                                       bool useGpu) {
     if (!src) {
@@ -333,30 +338,31 @@ void Argument::concat(const std::vector<Argument>& args,
     } else {
       dst->resize(batchSize);
     }
-    std::copy(
-        src->begin() + pos, src->begin() + pos + size, dst->begin() + startRow);
+    std::copy(src->begin() + srcStartRow,
+              src->begin() + srcStartRow + size,
+              dst->begin() + desStartRow);
   };
 
   dataId = args[0].dataId;
   CHECK_NE(seqStartPos.size(), 0UL);
-  size_t sampleNum = seqStartPos.size() - 1;
-  for (size_t i = 0; i < sampleNum; ++i) {
+  int desStartRow = 0;
+  for (size_t i = 0; i < copySize.size(); ++i) {
     int startPos = seqStartPos[i];
     int endPos = seqStartPos[i + 1];
     CHECK_GE(args.size(), static_cast<size_t>(endPos - startPos));
     for (int j = startPos; j < endPos; ++j) {
       const Argument& arg = args[j - startPos];
-      CHECK_EQ(arg.dataId, dataId) << "Arguments in concat should have"
-                                   << " same dataId";
-      const int copySize = 1;
-      const int rowIdx = selectRows[j];
-      copyArg(in, arg.in, j, rowIdx, copySize, useGpu);
-      copyArg(value, arg.value, j, rowIdx, copySize, useGpu);
+      CHECK_EQ(arg.dataId, dataId) << "Arguments to concatenate should have "
+                                   << "the same dataId.";
+      const int srcStartRow = selectRows[j];
+      copyArg(in, arg.in, desStartRow, srcStartRow, copySize[i], useGpu);
+      copyArg(value, arg.value, desStartRow, srcStartRow, copySize[i], useGpu);
       if (passType != PASS_TEST) {
-        copyArg(grad, arg.grad, j, rowIdx, copySize, useGpu);
+        copyArg(grad, arg.grad, desStartRow, srcStartRow, copySize[i], useGpu);
       }
-      copyIds(ids, arg.ids, j, rowIdx, copySize, useGpu);
-      copyStrs(strs, arg.strs, j, rowIdx, copySize, useGpu);
+      copyIds(ids, arg.ids, desStartRow, srcStartRow, copySize[i], useGpu);
+      copyStrs(strs, arg.strs, desStartRow, srcStartRow, copySize[i], useGpu);
+      desStartRow += copySize[i];
     }
   }
   ICpuGpuVector::resizeOrCreate(

paddle/parameter/Argument.h

@@ -240,6 +240,7 @@ struct Argument {
   void concat(const std::vector<Argument>& args,
               const std::vector<int>& selectRows,
               const std::vector<int>& seqStartPos,
+              const std::vector<int>& copySize,
               bool useGpu,
               hl_stream_t stream,
               PassType passType);

paddle/parameter/Parameter.h

@@ -65,8 +65,11 @@ public:
   size_t getSize() const { return config_.size(); }
 
   bool isFullSize() const {
+    if (bufs_[PARAMETER_VALUE]) {
       return this->getSize() == bufs_[PARAMETER_VALUE]->getSize();
     }
+    return false;
+  }
 
   inline bool useGpu() const { return useGpu_; }
 

paddle/platform/device_context.cc

@@ -114,9 +114,6 @@ CUDADeviceContext::~CUDADeviceContext() {
     PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_));
   }
 
-  if (curand_generator_) {
-    PADDLE_ENFORCE(dynload::curandDestroyGenerator(curand_generator_));
-  }
   eigen_stream_.reset();
   eigen_device_.reset();
   PADDLE_ENFORCE(cudaStreamDestroy(stream_));
@@ -152,19 +149,6 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() {
 
 cudaStream_t CUDADeviceContext::stream() { return stream_; }
 
-curandGenerator_t CUDADeviceContext::curand_generator() {
-  if (!curand_generator_) {
-    SetDeviceId(place_.device);
-    PADDLE_ENFORCE(dynload::curandCreateGenerator(&curand_generator_,
-                                                  CURAND_RNG_PSEUDO_DEFAULT));
-    PADDLE_ENFORCE(
-        dynload::curandSetPseudoRandomGeneratorSeed(curand_generator_, seed_));
-
-    PADDLE_ENFORCE(dynload::curandSetStream(curand_generator_, stream_));
-  }
-  return curand_generator_;
-}
-
 #endif  // PADDLE_ONLY_CPU
 
 }  // namespace platform

paddle/platform/device_context.h

@@ -17,7 +17,6 @@ limitations under the License. */
 #ifndef PADDLE_ONLY_CPU
 #include "paddle/platform/dynload/cublas.h"
 #include "paddle/platform/dynload/cudnn.h"
-#include "paddle/platform/dynload/curand.h"
 #include "paddle/platform/gpu_info.h"
 #define EIGEN_USE_GPU
 #endif
@@ -40,7 +39,7 @@ class DeviceContext {
 class CPUDeviceContext : public DeviceContext {
  public:
   CPUDeviceContext();
-  explicit CPUDeviceContext(CPUPlace);
+  explicit CPUDeviceContext(CPUPlace place);
   virtual ~CPUDeviceContext() {}
 
   Eigen::DefaultDevice* eigen_device() const;
@@ -56,7 +55,7 @@ class EigenCudaStreamDevice;
 
 class CUDADeviceContext : public DeviceContext {
  public:
-  explicit CUDADeviceContext(GPUPlace);
+  explicit CUDADeviceContext(GPUPlace place);
   virtual ~CUDADeviceContext();
 
   /*! \brief  Wait for all operations completion in the stream. */
@@ -75,9 +74,6 @@ class CUDADeviceContext : public DeviceContext {
   /*! \brief  Return cudnn  handle in the device context. */
   cudnnHandle_t     cudnn_handle();
 
-  /*! \brief  Return curand handle in the device context. */
-  curandGenerator_t curand_generator();
-
   /*! \brief  Return cuda stream in the device context. */
   cudaStream_t      stream();
   // clang-format on
@@ -85,18 +81,13 @@ class CUDADeviceContext : public DeviceContext {
  private:
   GPUPlace place_;
 
- private:
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
   std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
 
- private:
-  uint64_t seed_;
-
   // clang-format off
   cudaStream_t       stream_{nullptr};
   cudnnHandle_t      cudnn_handle_{nullptr};
   cublasHandle_t     cublas_handle_{nullptr};
-  curandGenerator_t  curand_generator_{nullptr};
   // clang-format on
 };
 

paddle/platform/device_context_test.cc

@@ -43,8 +43,6 @@ TEST(Device, CUDADeviceContext) {
     ASSERT_NE(nullptr, cudnn_handle);
     cublasHandle_t cublas_handle = device_context->cublas_handle();
     ASSERT_NE(nullptr, cublas_handle);
-    curandGenerator_t curand_handle = device_context->curand_generator();
-    ASSERT_NE(nullptr, curand_handle);
     ASSERT_NE(nullptr, device_context->stream());
     delete device_context;
   }

paddle/pserver/ParameterClient2.cpp

@@ -65,7 +65,6 @@ void ParameterClient2::initThreads() {
     LOG(INFO) << "parallel_thread_num dosent need to set";
   }
   syncThreadPool_.reset(new SyncThreadPool(threadNum_));
-
   startThreads();
 }
 
@@ -224,6 +223,14 @@ void ParameterClient2::prepareSendData(
     request.set_cost(cost);
     request.set_batch_status(batchStatus);
     CHECK_EQ(request.blocks_size(), 0);
+    VLOG(10) << "request: trainer_id: " << request.trainer_id()
+             << " update_mode" << request.update_mode()
+             << " send_back_parameter: " << request.send_back_parameter()
+             << " send_back_parameter_type: "
+             << request.send_back_parameter_type()
+             << " num_samples: " << request.num_samples()
+             << " cost: " << request.cost()
+             << " batch_status: " << request.batch_status();
   }
   for (const auto& segments : parameterSegments) {
     const auto it = parameterMap_.find(segments.id);
@@ -251,11 +258,17 @@ void ParameterClient2::prepareSendData(
       CHECK(sendMat != nullptr) << "sendMat is nullptr";
 
       syncThreadPool_->exec([&](int tid, size_t numThreads) {
+        std::lock_guard<std::mutex> guard(sparseAutoGrowthMutex_);
         const auto& localIndices = prefetchMat->getLocalIndices();
         /// num of sparse rows
         size_t nLocalBlocks = localIndices.size();
         uint64_t beginDim = 0;
         uint64_t endDim = 0;
+
+        // FIXME(typhoonzero): let it resize first
+        prefetchMat->getLocalRow(nLocalBlocks + 1);
+        sendMat->getLocalRow(nLocalBlocks + 1);
+
         for (size_t row = 0; row < nLocalBlocks; ++row) {
           int64_t blockId = localIndices[row];  // local row -> sparse row
           int serverId = std::abs((blockId + nameHash) % serviceNum_);
@@ -275,7 +288,6 @@ void ParameterClient2::prepareSendData(
           block->set_begin_pos(row * blockSize);
           /// block len
           block->set_block_size(endDim - beginDim);
-
           if (sendingPara) {
             sendJob->parallelInputIovs[serverId].push_back(
                 {sendMat->getLocalRow(row), sizeof(real) * (size_t)blockSize});

paddle/pserver/ParameterClient2.h

@@ -583,6 +583,7 @@ protected:
 #ifndef PADDLE_DISABLE_TIMER
   uint64_t forwardbackwordTime_;
 #endif
+  std::mutex sparseAutoGrowthMutex_;
 
   /// map id to parameter used for decoding protobuf data
   std::unordered_map<size_t, ParameterPtr> parameterMap_;

python/paddle/trainer/config_parser.py

@@ -338,7 +338,8 @@ def RecurrentLayerGroupWithoutOutLinksBegin(name,
         in_links_count += 1
         layer_name = MakeLayerNameInParentSubmodel(name)
         layer = g_layer_map[layer_name]
-        ScatterAgentLayer(name=name, size=layer.size)
+        ScatterAgentLayer(
+            name=name, size=layer.size, width=layer.width, height=layer.height)
 
         pair = g_current_submodel.in_links.add()
         pair.layer_name = layer_name
@@ -2197,8 +2198,8 @@ class MaxOutLayer(LayerBase):
         maxout_conf = self.config.inputs[0].maxout_conf
         parse_maxout(self.inputs[0].maxout, input_layer.name, maxout_conf)
         out_channels = maxout_conf.image_conf.channels / maxout_conf.groups
-        self.set_cnn_layer(name, g_layer_map[input_layer.name].height,
-                           g_layer_map[input_layer.name].width, out_channels)
+        self.set_cnn_layer(name, maxout_conf.image_conf.img_size_y,
+                           maxout_conf.image_conf.img_size, out_channels)
 
 
 @config_layer('row_conv')
@@ -2232,6 +2233,20 @@ class ClipLayer(LayerBase):
         self.config.inputs[0].clip_conf.max = max
 
 
+@config_layer('scale_shift')
+class ScaleShiftLayer(LayerBase):
+    def __init__(self, name, inputs, bias=True, **xargs):
+        super(ScaleShiftLayer, self).__init__(
+            name, 'scale_shift', 0, inputs=inputs, **xargs)
+        config_assert(
+            len(self.inputs) == 1,
+            'ScaleShiftLayer must have one and only one input.')
+        input_layer = self.get_input_layer(0)
+        self.set_layer_size(input_layer.size)
+        self.create_input_parameter(0, 1, [1, 1])
+        self.create_bias_parameter(bias, 1)
+
+
 # key: cost type
 # value: cost class
 g_cost_map = {}
@@ -2391,9 +2406,11 @@ class GatherAgentLayer(LayerBase):
 
 @config_layer('scatter_agent')
 class ScatterAgentLayer(LayerBase):
-    def __init__(self, name, size, device=None):
+    def __init__(self, name, size, width=None, height=None, device=None):
         super(ScatterAgentLayer, self).__init__(
             name, 'scatter_agent', size, inputs=[], device=device)
+        if height and width:
+            self.set_layer_height_width(height, width)
 
 
 @config_layer('multiplex')

python/paddle/trainer_config_helpers/layers.py

@@ -16,11 +16,13 @@ import functools
 import collections
 import inspect
 
+import paddle.trainer.config_parser as cp
 from paddle.trainer.config_parser import *
 from .activations import LinearActivation, SigmoidActivation, TanhActivation, \
     ReluActivation, IdentityActivation, SoftmaxActivation, BaseActivation
 from .evaluators import *
-from .poolings import MaxPooling, AvgPooling, BasePoolingType
+from .poolings import MaxPooling, AvgPooling, BasePoolingType, \
+    CudnnAvgPooling, CudnnMaxPooling
 from .attrs import *
 from .default_decorators import *
 
@@ -133,6 +135,7 @@ __all__ = [
     'clip_layer',
     'slice_projection',
     'kmax_sequence_score_layer',
+    'scale_shift_layer',
 ]
 
 
@@ -230,6 +233,7 @@ class LayerType(object):
     CLIP_LAYER = 'clip'
 
     KMAX_SEQ_SCORE = 'kmax_seq_score'
+    SCALE_SHIFT_LAYER = 'scale_shift'
 
     @staticmethod
     def is_layer_type(type_name):
@@ -328,6 +332,14 @@ class LayerOutput(object):
         self.outputs = outputs
         self.reverse = reverse
 
+    @property
+    def width(self):
+        return cp.g_layer_map[self.full_name].width
+
+    @property
+    def height(self):
+        return cp.g_layer_map[self.full_name].height
+
     def set_input(self, input):
         """
         Set the input for a memory layer. Can only be used for memory layer
@@ -909,7 +921,13 @@ def data_layer(name, size, height=None, width=None, layer_attr=None):
         width=width,
         **ExtraLayerAttribute.to_kwargs(layer_attr))
 
-    return LayerOutput(name, LayerType.DATA, size=size)
+    num_filters = None
+    if height is not None and width is not None:
+        num_filters = size / (width * height)
+        assert num_filters * width * height == size, \
+            "size=%s width=%s height=%s" % (size, width, height)
+
+    return LayerOutput(name, LayerType.DATA, size=size, num_filters=num_filters)
 
 
 @wrap_name_default("embedding")
@@ -2569,6 +2587,10 @@ def img_pool_layer(input,
         assert input.num_filters is not None
         num_channels = input.num_filters
 
+    assert type(pool_type) in [AvgPooling, MaxPooling, CudnnAvgPooling,
+                               CudnnMaxPooling], \
+        "only (Cudnn)AvgPooling, (Cudnn)MaxPooling are supported"
+
     if pool_type is None:
         pool_type = MaxPooling()
     elif isinstance(pool_type, AvgPooling):
@@ -2578,7 +2600,6 @@ def img_pool_layer(input,
         if (
         isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \
         else pool_type.name
-
     pool_size_y = pool_size if pool_size_y is None else pool_size_y
     stride_y = stride if stride_y is None else stride_y
     padding_y = padding if padding_y is None else padding_y
@@ -4202,8 +4223,7 @@ def conv_operator(img,
         num_channels = img.num_filters
 
     assert isinstance(filter, LayerOutput)
-    if filter.size is not None:
-        filter.size = filter_size * filter_size_y * num_filters * num_channels
+    assert filter.size is not None
 
     opCls = ConvTransOperator if trans else ConvOperator
 
@@ -4914,7 +4934,6 @@ def maxout_layer(input, groups, num_channels=None, name=None, layer_attr=None):
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
-    assert input.layer_type == LayerType.CONV_LAYER
     assert isinstance(input.activation, LinearActivation)
     assert groups > 1
     if num_channels is None:
@@ -6210,3 +6229,43 @@ def kmax_sequence_score_layer(input, name=None, beam_size=1):
 
     return LayerOutput(
         name, LayerType.KMAX_SEQ_SCORE, parents=[input], size=input.size)
+
+
+@wrap_name_default("scale_shift")
+@wrap_param_attr_default()
+@wrap_bias_attr_default()
+def scale_shift_layer(input, name=None, param_attr=None, bias_attr=None):
+    """
+    A layer applies a linear transformation to each element in each row of
+    the input matrix. For each element, the layer first re-scale it and then
+    adds a bias to it.
+
+    This layer is very like the SlopeInterceptLayer, except the scale and
+    bias are trainable.
+
+    .. math::
+
+        y = w * x + b
+
+    .. code-block:: python
+
+        scale_shift = scale_shift_layer(input=input_layer, bias_attr=False)
+
+    :param name: The Layer Name.
+    :type name: basestring
+    :param input: The input layer.
+    :type input: LayerOutput.
+    :param param_attr: The parameter attribute of scaling.
+    :type param_attr: ParameterAttribute
+    :param bias_attr: The parameter attribute of shifting.
+    :type bias_attr: ParameterAttribute
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    Layer(
+        name=name,
+        type=LayerType.SCALE_SHIFT_LAYER,
+        inputs=Input(input.name, **param_attr.attr),
+        bias=ParamAttr.to_bias(bias_attr))
+    return LayerOutput(
+        name, LayerType.SCALE_SHIFT_LAYER, parents=[input], size=input.size)

python/paddle/trainer_config_helpers/tests/configs/file_list.sh

@@ -8,6 +8,6 @@ test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops
 test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer
 test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer
 test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
-test_kmax_seq_socre_layer test_seq_select_layers)
+test_kmax_seq_socre_layer test_seq_select_layers test_scale_shift_layer)
 
 export whole_configs=(test_split_datasource)

python/paddle/trainer_config_helpers/tests/configs/protostr/test_scale_shift_layer.protostr

+type: "nn"
+layers {
+  name: "data"
+  type: "data"
+  size: 100
+  active_type: ""
+}
+layers {
+  name: "__scale_shift_0__"
+  type: "scale_shift"
+  size: 100
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    input_parameter_name: "___scale_shift_0__.w0"
+  }
+}
+layers {
+  name: "__scale_shift_1__"
+  type: "scale_shift"
+  size: 100
+  active_type: ""
+  inputs {
+    input_layer_name: "data"
+    input_parameter_name: "___scale_shift_1__.w0"
+  }
+  bias_parameter_name: "___scale_shift_1__.wbias"
+}
+parameters {
+  name: "___scale_shift_0__.w0"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 1.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: true
+}
+parameters {
+  name: "___scale_shift_1__.w0"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 1.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: true
+}
+parameters {
+  name: "___scale_shift_1__.wbias"
+  size: 1
+  initial_mean: 0.0
+  initial_std: 0.0
+  dims: 1
+  dims: 1
+  initial_strategy: 0
+  initial_smart: false
+}
+input_layer_names: "data"
+output_layer_names: "__scale_shift_0__"
+output_layer_names: "__scale_shift_1__"
+sub_models {
+  name: "root"
+  layer_names: "data"
+  layer_names: "__scale_shift_0__"
+  layer_names: "__scale_shift_1__"
+  input_layer_names: "data"
+  output_layer_names: "__scale_shift_0__"
+  output_layer_names: "__scale_shift_1__"
+  is_recurrent_layer_group: false
+}
+

python/paddle/trainer_config_helpers/tests/configs/test_scale_shift_layer.py

+from paddle.trainer_config_helpers import *
+
+data = data_layer(name='data', size=100)
+
+scale = scale_shift_layer(input=data, bias_attr=False)
+
+scale_shift = scale_shift_layer(input=data)
+
+outputs(scale, scale_shift)

python/paddle/v2/framework/tests/CMakeLists.txt

@@ -13,6 +13,7 @@ py_test(test_add_two_op SRCS test_add_two_op.py)
 py_test(test_sigmoid_op SRCS test_sigmoid_op.py)
 py_test(test_softmax_op SRCS test_softmax_op.py)
 py_test(test_cross_entropy_op SRCS test_cross_entropy_op.py)
+py_test(test_gather_op SRCS test_gather_op.py)
 py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py)
 
 py_test(gradient_checker SRCS gradient_checker.py)
@@ -22,7 +23,7 @@ py_test(test_rowwise_add_op SRCS test_rowwise_add_op.py)
 py_test(test_default_scope_funcs SRCS test_default_scope_funcs.py)
 
 py_test(test_operator SRCS test_operator.py)
-# py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
+py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
 py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
 py_test(test_recurrent_op SRCS test_recurrent_op.py)
 py_test(test_sgd_op SRCS test_sgd_op.py)

python/paddle/v2/framework/tests/op_test_util.py

@@ -64,7 +64,8 @@ class OpTestMeta(type):
                     actual = numpy.array(scope.find_var(out_name).get_tensor())
                     expect = self.outputs[out_name]
                     self.assertTrue(
-                        numpy.allclose(actual, expect),
+                        numpy.allclose(
+                            actual, expect, atol=1e-05),
                         "output name: " + out_name + "has diff")
 
         obj.test_all = test_all

python/paddle/v2/framework/tests/test_cross_entropy_op.py

@@ -8,9 +8,8 @@ class TestCrossEntropy(unittest.TestCase):
     __metaclass__ = OpTestMeta
 
     def setUp(self):
-        # TODO this unit test is not passed
         self.type = "onehot_cross_entropy"
-        batch_size = 100
+        batch_size = 30
         class_num = 10
         X = numpy.random.random((batch_size, class_num)).astype("float32")
         label = 5 * numpy.ones(batch_size).astype("int32")
@@ -22,9 +21,9 @@ class TestCrossEntropy(unittest.TestCase):
 
 
 class CrossEntropyGradOpTest(GradientChecker):
-    def test_softmax_grad(self):
+    def test_check_grad(self):
         op = create_op("onehot_cross_entropy")
-        batch_size = 100
+        batch_size = 30
         class_num = 10
         inputs = {
             "X": numpy.random.uniform(

python/paddle/v2/framework/tests/test_gather_op.py

+import unittest
+from op_test_util import OpTestMeta
+from gradient_checker import GradientChecker, create_op
+import numpy
+import paddle.v2.framework.core as core
+from paddle.v2.framework.op import Operator
+
+
+class TestGatherOp(unittest.TestCase):
+    __metaclass__ = OpTestMeta
+
+    def setUp(self):
+        self.type = "gather"
+        xnp = numpy.random.random((10, 20)).astype("float32")
+        self.inputs = {
+            'X': xnp,
+            'Index': numpy.array([1, 3, 5]).astype("int32")
+        }
+        self.outputs = {'Out': self.inputs['X'][self.inputs['Index']]}
+
+
+class TestGatherGradOp(GradientChecker):
+    def test_gather_grad(self):
+        print 'creating op'
+        op = create_op("gather")
+        print 'creating op done'
+        xnp = numpy.random.random((10, 20)).astype("float32")
+        inputs = {'X': xnp, 'Index': numpy.array([1, 3, 5]).astype("int32")}
+        print 'correct before check gradient'
+        self.check_grad(op, inputs, set("X"), "Out")
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/v2/framework/tests/test_net.py

@@ -6,8 +6,8 @@ import unittest
 def fc(X, W, Y):
     ret_v = core.Net.create()
 
-    ret_v.add_op(Operator("mul", X="X", Y="W", Out="pre_activation"))
-    ret_v.add_op(Operator("sigmoid", X="pre_activation", Y=Y))
+    ret_v.append_op(Operator("mul", X="X", Y="W", Out="pre_activation"))
+    ret_v.append_op(Operator("sigmoid", X="pre_activation", Y=Y))
     ret_v.complete_add_op(True)
     return ret_v
 
@@ -16,12 +16,12 @@ class TestNet(unittest.TestCase):
     def test_net_all(self):
         net = core.Net.create()
         op1 = Operator("add_two", X="X", Y="Y", Out="Out")
-        net.add_op(op1)
+        net.append_op(op1)
 
         net2 = core.Net.create()
-        net2.add_op(fc(X="X", W="w", Y="fc.out"))
+        net2.append_op(fc(X="X", W="w", Y="fc.out"))
         net2.complete_add_op(True)
-        net.add_op(net2)
+        net.append_op(net2)
         net.complete_add_op(True)
 
         expected = '''

python/paddle/v2/framework/tests/test_recurrent_op.py

@@ -150,7 +150,7 @@ class TestRecurrentOp(unittest.TestCase):
         sig_op = Operator("sigmoid", X="sum", Y="h@alias")
 
         for op in [x_fc_op, h_fc_op, sum_op, sig_op]:
-            stepnet.add_op(op)
+            stepnet.append_op(op)
         stepnet.complete_add_op(True)
         self.rnnop.set_stepnet(stepnet)
 

python/paddle/v2/framework/tests/test_rowwise_add_op.py

@@ -20,7 +20,7 @@ class RowwiseAddGradOpTest(GradientChecker):
     def test_rowwise_add(self):
         op = create_op("rowwise_add")
         inputs = {
-            "X": np.random.uniform(0.1, 1, [10, 10]).astype("float32"),
+            "X": np.random.uniform(0.1, 1, [5, 10]).astype("float32"),
             "b": np.random.uniform(0.1, 1, [10]).astype("float32")
         }
         self.check_grad(op, inputs, set(["X", "b"]), "Out")