Merge branch 'develop' into fix/scope

CMakeLists.txt

@@ -105,6 +105,12 @@ if (WITH_C_API AND WITH_PYTHON)
     "different Python interpreter from compiling.")
 endif()
 
+if(MOBILE_INFERENCE)
+    set(THIRD_PARTY_BUILD_TYPE MinSizeRel)
+else()
+    set(THIRD_PARTY_BUILD_TYPE Release)
+endif()
+
 ########################################################################################
 
 include(external/mklml)     # download mklml package

cmake/external/eigen.cmake

@@ -8,7 +8,7 @@ ExternalProject_Add(
     extern_eigen3
     ${EXTERNAL_PROJECT_LOG_ARGS}
     GIT_REPOSITORY  "https://github.com/RLovelett/eigen.git"
-    GIT_TAG         "master"
+    GIT_TAG         4e79cb69b9425f5f8c3a84be4350d4ab75b5fd9d
     PREFIX          ${EIGEN_SOURCE_DIR}
     UPDATE_COMMAND  ""
     CONFIGURE_COMMAND ""

cmake/external/gflags.cmake

@@ -36,6 +36,7 @@ ExternalProject_Add(
     # change this back to the official Github repo once my PR is
     # merged.
     GIT_REPOSITORY  "https://github.com/wangkuiyi/gflags.git"
+    GIT_TAG         986964c07427ecb9cdb5bd73f73ebbd40e54dadb
     PREFIX          ${GFLAGS_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
@@ -45,11 +46,11 @@ ExternalProject_Add(
                     -DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR}
                     -DCMAKE_POSITION_INDEPENDENT_CODE=ON
                     -DBUILD_TESTING=OFF
-                    -DCMAKE_BUILD_TYPE=Release
+                    -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
                     ${EXTERNAL_OPTIONAL_ARGS}
     CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GFLAGS_INSTALL_DIR}
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-                     -DCMAKE_BUILD_TYPE:STRING=Release
+                     -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
 )
 
 ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL)

cmake/external/glog.cmake

@@ -31,6 +31,7 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS gflags
     GIT_REPOSITORY  "https://github.com/google/glog.git"
+    GIT_TAG         v0.3.5
     PREFIX          ${GLOG_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
@@ -43,12 +44,12 @@ ExternalProject_Add(
                     -DWITH_GFLAGS=ON
                     -Dgflags_DIR=${GFLAGS_INSTALL_DIR}/lib/cmake/gflags
                     -DBUILD_TESTING=OFF
-                    -DCMAKE_BUILD_TYPE=Release
+                    -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
                     ${EXTERNAL_OPTIONAL_ARGS}
     CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GLOG_INSTALL_DIR}
                      -DCMAKE_INSTALL_LIBDIR:PATH=${GLOG_INSTALL_DIR}/lib
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-                     -DCMAKE_BUILD_TYPE:STRING=Release
+                     -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
 )
 
 ADD_LIBRARY(glog STATIC IMPORTED GLOBAL)

cmake/external/gtest.cmake

@@ -56,11 +56,11 @@ IF(WITH_TESTING)
                         -DBUILD_GMOCK=ON
                         -Dgtest_disable_pthreads=ON
                         -Dgtest_force_shared_crt=ON
-                        -DCMAKE_BUILD_TYPE=Release
+                        -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
                         ${EXTERNAL_OPTIONAL_ARGS}
         CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GTEST_INSTALL_DIR}
                          -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-                         -DCMAKE_BUILD_TYPE:STRING=Release
+                         -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
     )
 
     ADD_LIBRARY(gtest STATIC IMPORTED GLOBAL)

cmake/external/protobuf.cmake

@@ -191,12 +191,12 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
             ${OPTIONAL_ARGS}
             -Dprotobuf_BUILD_TESTS=OFF
             -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-            -DCMAKE_BUILD_TYPE=Release
+            -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
             -DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR}
             -DCMAKE_INSTALL_LIBDIR=lib
         CMAKE_CACHE_ARGS
             -DCMAKE_INSTALL_PREFIX:PATH=${PROTOBUF_INSTALL_DIR}
-            -DCMAKE_BUILD_TYPE:STRING=Release
+            -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
             -DCMAKE_VERBOSE_MAKEFILE:BOOL=OFF
             -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
             ${OPTIONAL_CACHE_ARGS}

cmake/external/warpctc.cmake

@@ -35,6 +35,7 @@ ExternalProject_Add(
     extern_warpctc
     ${EXTERNAL_PROJECT_LOG_ARGS}
     GIT_REPOSITORY  "https://github.com/gangliao/warp-ctc.git"
+    GIT_TAG         b63a0644654a3e0ed624c85a1767bc8193aead09
     PREFIX          ${WARPCTC_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
@@ -48,9 +49,9 @@ ExternalProject_Add(
                     -DCMAKE_DISABLE_FIND_PACKAGE_Torch=ON
                     -DBUILD_SHARED=ON
                     -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-                    -DCMAKE_BUILD_TYPE=Release
+                    -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
                     ${EXTERNAL_OPTIONAL_ARGS}
-    CMAKE_CACHE_ARGS -DCMAKE_BUILD_TYPE:STRING=Release
+    CMAKE_CACHE_ARGS -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_INSTALL_PREFIX:PATH=${WARPCTC_INSTALL_DIR}
 )

cmake/external/zlib.cmake

@@ -42,11 +42,11 @@ ExternalProject_Add(
                     -DBUILD_SHARED_LIBS=OFF
                     -DCMAKE_POSITION_INDEPENDENT_CODE=ON
                     -DCMAKE_MACOSX_RPATH=ON
-                    -DCMAKE_BUILD_TYPE=Release
+                    -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
                     ${EXTERNAL_OPTIONAL_ARGS}
     CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ZLIB_INSTALL_DIR}
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-                     -DCMAKE_BUILD_TYPE:STRING=Release
+                     -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
 )
 
 LIST(APPEND external_project_dependencies zlib)

doc/design/register_grad_op.md

@@ -3,15 +3,17 @@
 
 ## The Problem Posed
 
-In our current operator registration mechanism, for each operator, the programmer should register a *gradient operator creator* function, which takes a C++ operator instance, and returns the corresponding gradient instance.
+Currently, for each C++ operator class definition, there registers a *gradient operator creator* function, which takes a C++ operator instance and returns the corresponding gradient operator instance.
 
-However, as we decided to separate the *compilation* and *execution* of DL models, we need to reshape the creator to take a protobuf `OpDesc` message, and returns a corresponding message.
+However, we noticed two problems with the current deisgn:
 
-More than that, the new registration mechanism need to support the fact that an operators' gradient computation might be a composition of operators.
+1. As we decided to separate the *compilation* and *execution* phases, we need to change the creator to take an `OpDesc` protobuf message in a `ProgramDesc` and inserts corresponding `OpDesc` messages into the `ProgramDesc` message.
 
-## Current Implementation
+1. Some operator's gradient computation requires more than one gradient operators.  For example, the gradient of *minus* consists of two operators -- an identity operaotr and a scale operator.  So we need to make the registration mechanism to support the mapping from an operator to a set of operators for gradient computation.
 
-OpInfos store in a association map which key is the operator type. The `grad_op_type` indicate associated gradient operator type. Operator can create gradient operator by `OpInfo::creator_` of gradient. The pseudo code is
+## The Current Implementation
+
+The C++ class `OpInfos` store in a association map which key is the operator type. The `grad_op_type` indicate associated gradient operator type. Operator can create gradient operator by `OpInfo::creator_` of gradient. The pseudo code is
 
 ```cpp
 struct OpInfo {

doc/design/var_desc.md

@@ -16,16 +16,23 @@ The computation graph is constructed by Data Node and Operation Node. The concep
 
 ## Definition of VarDesc
 
-A VarDesc should have a name and value, in PaddlePaddle, the value will always be a tensor. Since we use LoDTensor most of the time. We add a LoDTesnorDesc to represent it.
+A VarDesc should have a name, and value. The are two kinds of variable type in compile time, they are `LoDTensor` and `SelectedRows`. 
 
 ```proto
 message VarDesc {
   required string name = 1;
-  optional LoDTensorDesc lod_tensor = 2;
+  enum VarType {
+    LOD_TENSOR = 0;
+    SELECTED_ROWS = 1;
+  }
+  required VarType type = 2;
+  optional LoDTensorDesc lod_desc = 3;
+  optional TensorDesc selected_rows_desc = 4;
+  optional bool persistable = 5 [ default = false ];
 }
 ```
 
-## Definition of LodTensorDesc
+## Definition of TensorDesc
 
 ```proto
 enum DataType {
@@ -38,87 +45,25 @@ enum DataType {
   FP64 = 6;
 }
 
-message LoDTensorDesc {
+message TensorDesc {
   required DataType data_type = 1;
-  repeated int32 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
-  optional int32 lod_level = 3 [default=0];
+  repeated int64 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
 }
 ```
 
-## Definition of Variable in Python
-
-In Python API, layer will take Variable as Input, and return Variable as Output. There should be a class `Variable` in python to help create and manage Variable.
-
-```python
-image = Variable(dims=[-1, 640, 480])
-# fc1 and fc2 are both Variable
-fc1 = layer.fc(input=image, output_size=10)
-fc2 = layer.fc(input=fc1, output_size=20)
-```
-### what should class `Variable` Have
-1. `name`.a name of string type is used to mark the value of the Variable.
-1. `initializer`. Since our Tensor does not have value. we will always use some Operator to fullfill it when run. So we should have a initialize method to help add the init operator.
-1. `operator`. Variable should record which operator produce itself. The reaon is:
-  - we use pd.eval(targets=[var1, var2]) to run the related ops to get the value of var1 and var2. var.op is used to trace the dependency of the current variable.
-
-In PaddlePaddle, we use Block to describe Computation Graph, so in the code we will use Block but not Graph.
-
-```python
-import VarDesc
-import LoDTensorDesc
-import framework
-
-def AddInitialOperator(variable, initializer):
-	# add an initialize Operator to block to init this Variable
-
-class Variable(object):
-   def __init__(self, name, dims, type, initializer):
-      self._block = get_default_block()
-      self._name = name
-      self.op = None
-
-      tensor_desc = LoDTensorDesc(data_type=type, dims=dims)
-      _var_desc = VarDesc(name=name, lod_tensor=tensor_desc)
-      self._var = framework.CreateVar(_var_desc)
-      self._block.add_var(self)
+A TensorDesc describes `SelectedRows` and `LoDTensor`. For details of `SelectedRows`, please reference [`SelectedRows`](./selected_rows.md).
 
-      # add initial op according to initializer
-      if initializer is not None:
-          AddInitialOperator(self, initializer)
-
-   def dims(self):
-      return self._var.dims()
-
-   def data_type(self):
-       return self._var.data_type()
+## Definition of LodTensorDesc
 
-   def to_proto(self):
-       pass
+```proto
+message LoDTensorDesc {
+  required TensorDesc tensor = 1;
+  optional int lod_level = 2;
+}
 ```
 
-Then we can use this Variable to create a fc layer in Python.
+A LoDTensorDesc contains a tensor and a lod_level.
 
-```python
-import paddle as pd
-
-def flatten_size(X, num_flatten_dims):
-  prod = 1 # of last num_flatten_dims
-  for i in xrange(num_flatten_dims):
-    prod = prod * X.dims[-i-1]
-  return prod
-
-def layer.fc(X, output_size, num_flatten_dims):
-  W = Variable(pd.random_uniform(), type=FP32, dims=[flatten_size(X, num_flatten_dims), output_size])
-  b = Variable(pd.random_uniform(), type=FP32, dims=[output_size])
-  out = Variable(type=FP32)
-  y = operator.fc(X, W, b, output=out) # fc will put fc op input into out
-  pd.InferShape(y)
-  return out
-
-x = Variable(dims=[-1, 640, 480])
-y = layer.fc(x, output_size=100)
-z = layer.fc(y, output_size=200)
+## Definition of Variable in Python
 
-paddle.eval(targets=[z], ...)
-print(z)
-```
+For Variable in Python, please reference [`Python API`](./python_api.md).

paddle/framework/CMakeLists.txt

@@ -19,10 +19,10 @@ 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(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS attribute ddim)
+cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS attribute ddim op_info)
 cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
-cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto proto_desc)
+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 proto_desc)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 
@@ -53,3 +53,6 @@ endif()
 
 cc_library(tensor_array SRCS tensor_array.cc DEPS lod_tensor)
 cc_test(tensor_array_test SRCS tensor_array_test.cc DEPS tensor_array place)
+
+cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
+cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)

paddle/framework/backward.cc

@@ -28,14 +28,15 @@ namespace paddle {
 namespace framework {
 
 static inline std::unique_ptr<OperatorBase> CreateGradOp(
-    const OperatorBase& op) {
+    const OperatorBase& op, const std::unordered_set<std::string>& no_grad_set,
+    std::unordered_map<std::string, std::string>* grad_to_var) {
   OpDescBind op_desc;
   op_desc.SetInputMap(op.Inputs());
   op_desc.SetOutputMap(op.Outputs());
   op_desc.SetType(op.Type());
   op_desc.SetAttrMap(op.Attrs());
   auto& info = OpInfoMap::Instance().Get(op.Type());
-  auto grad_descs = info.GradOpMaker()(op_desc);
+  auto grad_descs = info.GradOpMaker()(op_desc, no_grad_set, grad_to_var);
   std::vector<std::unique_ptr<OperatorBase>> grad_ops;
   grad_ops.reserve(grad_descs.size());
   std::transform(grad_descs.begin(), grad_descs.end(),
@@ -98,7 +99,9 @@ static std::unique_ptr<OperatorBase> NOP() {
 //  See Backward.h for details
 static std::unique_ptr<OperatorBase> BackwardRecursive(
     const OperatorBase& forwardOp,
-    std::unordered_set<std::string>& no_grad_names, size_t& uniq_id) {
+    std::unordered_set<std::string>& no_grad_names,
+    std::unordered_map<std::string, std::string>* grad_to_var,
+    size_t& uniq_id) {
   //  If all input gradients of forwarding operator do not need to calculate,
   //  just return an NOP. Not return null ptr because NOP does not take
   //  too much time for calculation, but it is useful for simplifying logic.
@@ -136,7 +139,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
     for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend();
          ++it, ++local_op_id) {
       auto& fwd = *it;
-      auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id);
+      auto bwd = BackwardRecursive(*fwd, no_grad_names, grad_to_var, uniq_id);
       ForEachVarName(bwd->Outputs(),
                      [&dup_output_ops, local_op_id](const std::string& out) {
                        dup_output_ops[out].emplace_back(local_op_id);
@@ -187,7 +190,8 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
       net->InsertOp(pos.first + 1, std::move(pos.second));
     }
   } else {
-    std::unique_ptr<OperatorBase> grad_op(CreateGradOp(forwardOp));
+    std::unique_ptr<OperatorBase> grad_op(
+        CreateGradOp(forwardOp, no_grad_names, grad_to_var));
 
     ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
                                           const std::string& grad_input) {
@@ -226,7 +230,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
           *static_cast<const OperatorBase*>(&rnnop.stepnet());
       // create stepnet's gradient op
       rnn_grad_op->set_stepnet(
-          BackwardRecursive(stepnet_op, no_grad_names, uniq_id));
+          BackwardRecursive(stepnet_op, no_grad_names, grad_to_var, uniq_id));
     }
 
     if (net->ops_.empty()) {  // Current no aux op is added to network
@@ -253,7 +257,8 @@ std::unique_ptr<OperatorBase> Backward(
     no_grad_names.insert(name + kGradVarSuffix);
   }
   size_t uid = 0;
-  return BackwardRecursive(forwardOp, no_grad_names, uid);
+  std::unordered_map<std::string, std::string> grad_to_var;
+  return BackwardRecursive(forwardOp, no_grad_names, &grad_to_var, uid);
 }
 
 // ====================================  //
@@ -270,28 +275,31 @@ static bool AllGradInSet(const std::vector<std::string>& names,
 
 std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
     const std::unique_ptr<OpDescBind>& op_desc,
-    std::unordered_set<std::string>& no_grad_vars) {
+    std::unordered_set<std::string>* no_grad_vars,
+    std::unordered_map<std::string, std::string>* grad_to_var) {
   std::vector<std::unique_ptr<OpDescBind>> grad_op_descs;
-  // All input gradients of forwarding operator do not need to calculat.
+  // All input gradients of forwarding operator do not need to calculate.
   const std::vector<std::string>& inputs = op_desc->InputArgumentNames();
-  if (AllGradInSet(inputs, no_grad_vars)) {
+  if (AllGradInSet(inputs, *no_grad_vars)) {
     return grad_op_descs;  // empty vector
   }
   // All output gradients of forwarding operator do not need to calculate.
   const std::vector<std::string>& outputs = op_desc->OutputArgumentNames();
-  if (AllGradInSet(outputs, no_grad_vars)) {
+  if (AllGradInSet(outputs, *no_grad_vars)) {
     for (const std::string& name : inputs) {
-      no_grad_vars.insert(GradVarName(name));
+      no_grad_vars->insert(GradVarName(name));
     }
     return grad_op_descs;  // empty vector
   }
 
-  grad_op_descs = OpRegistry::CreateGradOpDescs(op_desc.get());
+  grad_op_descs = OpInfoMap::Instance()
+                      .Get(op_desc->Type())
+                      .GradOpMaker()(*op_desc, *no_grad_vars, grad_to_var);
 
   std::list<std::unique_ptr<OpDescBind>> pending_fill_zeros_ops;
   for (auto& desc : grad_op_descs) {
     for (const std::string& in_name : desc->InputArgumentNames()) {
-      if (no_grad_vars.count(in_name)) {
+      if (no_grad_vars->count(in_name)) {
         std::string prefix = in_name.substr(
             0, in_name.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
         std::string new_name = prefix + kZeroVarSuffix;
@@ -301,11 +309,6 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
         pending_fill_zeros_ops.push_back(std::move(fill_zeros_op));
       }
     }
-    for (const std::string& out_name : desc->OutputArgumentNames()) {
-      if (no_grad_vars.count(out_name)) {
-        desc->Rename(out_name, kEmptyVarName);
-      }
-    }
   }
 
   for (auto& p : pending_fill_zeros_ops) {
@@ -316,7 +319,8 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
 
 std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
     ProgramDescBind& program_desc, int block_idx,
-    std::unordered_set<std::string>& no_grad_vars) {
+    std::unordered_set<std::string>* no_grad_vars,
+    std::unordered_map<std::string, std::string>* grad_to_var) {
   BlockDescBind* cur_block = program_desc.Block(block_idx);
   std::deque<std::unique_ptr<OpDescBind>>& op_descs = cur_block->ops_;
   std::unordered_map<std::string, std::vector<size_t>> dup_out_ops;
@@ -324,15 +328,15 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
   std::vector<std::unique_ptr<OpDescBind>> backward_descs;
   for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) {
     std::vector<std::unique_ptr<OpDescBind>> op_grads =
-        MakeOpGrad(*it, no_grad_vars);
+        MakeOpGrad(*it, no_grad_vars, grad_to_var);
 
     if ((*it)->Type() == "recurrent") {
       PADDLE_ENFORCE_EQ(
           op_grads.size(), size_t(1),
           "rnn_op's gradient process should contain only one op.");
       int step_block_idx = (*it)->GetBlockAttr("stop_block");
-      auto backward_block_op_descs =
-          MakeBlockBackward(program_desc, step_block_idx, no_grad_vars);
+      auto backward_block_op_descs = MakeBlockBackward(
+          program_desc, step_block_idx, no_grad_vars, grad_to_var);
       BlockDescBind* backward_block = program_desc.AppendBlock(*cur_block);
       for (auto& ptr : backward_block_op_descs) {
         backward_block->ops_.push_back(std::move(ptr));
@@ -388,8 +392,9 @@ void AppendBackward(ProgramDescBind& program_desc,
     no_grad_var_names.insert(GradVarName(name));
   }
   const int root_block_idx = 0;
-  auto backward_op_descs =
-      MakeBlockBackward(program_desc, root_block_idx, no_grad_var_names);
+  std::unordered_map<std::string, std::string> grad_to_var;
+  auto backward_op_descs = MakeBlockBackward(program_desc, root_block_idx,
+                                             &no_grad_var_names, &grad_to_var);
   auto& forw_op_descs = program_desc.Block(root_block_idx)->ops_;
   for (auto& ptr : backward_op_descs) {
     forw_op_descs.push_back(std::move(ptr));

paddle/framework/backward_test.cc

@@ -169,6 +169,45 @@ class MultInOutOpMaker : public OpProtoAndCheckerMaker {
   }
 };
 
+class MinusGradOpDescMaker : public GradOpDescMakerBase {
+ public:
+  using GradOpDescMakerBase::GradOpDescMakerBase;
+
+  std::vector<std::unique_ptr<OpDescBind>> operator()() const override {
+    std::vector<std::unique_ptr<OpDescBind>> retv;
+    auto x_g = InputGrad("X");
+    if (!x_g.empty()) {
+      auto *op_desc = new OpDescBind();
+      op_desc->SetType("scale");
+      op_desc->SetInput("X", OutputGrad("Out"));
+      op_desc->SetOutput("Out", x_g);
+      op_desc->SetAttr("scale", 1.0f);
+      retv.emplace_back(op_desc);
+    }
+
+    auto y_g = InputGrad("Y");
+    if (!y_g.empty()) {
+      auto *op_desc = new OpDescBind();
+      op_desc->SetType("scale");
+      op_desc->SetInput("X", OutputGrad("Out"));
+      op_desc->SetOutput("Out", y_g);
+      op_desc->SetAttr("scale", -1.0f);
+      retv.emplace_back(op_desc);
+    }
+    return retv;
+  }
+};
+
+class MinusOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  MinusOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "");
+    AddInput("Y", "");
+    AddOutput("Out", "");
+    AddComment("minus for unittest");
+  }
+};
 }  // namespace framework
 }  // namespace paddle
 
@@ -187,6 +226,7 @@ REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
 REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
             f::NOP);
 REGISTER_OP(mult_in_out, f::NOP, f::MultInOutOpMaker, mult_in_out_grad, f::NOP);
+REGISTER_OPERATOR(minus, f::NOP, f::MinusOpMaker, f::MinusGradOpDescMaker);
 
 TEST(Backward, simple_op_not_need_grad) {
   auto fwd = f::OpRegistry::CreateOp(
@@ -395,12 +435,13 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
             2UL       /* external input number */
                 + 1UL /* external output number*/
                 + 1UL /* number of gradient of external output*/
-                + 2U /* internal variable number*/);
+                + 2UL /* internal variable number*/
+            );
   EXPECT_EQ(grad_fc.Outputs(all).size(),
             2UL       /* input number of mul*/
-                + 2UL /* input number of rowwise_add
-                       */
-                + 1UL /* input number of sigmod */);
+                + 2UL /* input number of rowwise_add*/
+                + 1UL /* input number of sigmod */
+                - 1UL /* out2 is not needed*/);
   EXPECT_EQ(bwd_net->ops_[1]->Inputs(all).size(), 0UL);
   EXPECT_EQ(bwd_net->ops_[1]->Outputs(all).size(), 0UL);
   EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL);
@@ -580,8 +621,7 @@ TEST(Backward, intermedia_var_no_grad) {
             std::vector<std::string>({f::GradVarName("out4")}));
   EXPECT_EQ(grad_op4->Output(f::GradVarName("X")),
             std::vector<std::string>({f::GradVarName("out1")}));
-  EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")),
-            std::vector<std::string>({f::kEmptyVarName}));
+  EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")), std::vector<std::string>());
 }
 
 TEST(Backward, var_no_grad) {
@@ -619,8 +659,7 @@ TEST(Backward, var_no_grad) {
             std::vector<std::string>({f::GradVarName("z2")}));
   EXPECT_EQ(grad_op2->Output(f::GradVarName("X")),
             std::vector<std::string>({f::GradVarName("y1")}));
-  EXPECT_EQ(grad_op2->Output(f::GradVarName("H")),
-            std::vector<std::string>({f::kEmptyVarName}));
+  EXPECT_EQ(grad_op2->Output(f::GradVarName("H")), std::vector<std::string>());
 
   f::OpDescBind *fill_zero_op = block->AllOps()[3];
   ASSERT_EQ(fill_zero_op->Type(), "fill_zeros_like");
@@ -718,4 +757,19 @@ TEST(Backward, shared_var) {
             std::vector<std::string>({f::GradVarName("x1")}));
   EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
             std::vector<std::string>({f::GradVarName("b1")}));
+}
+
+TEST(Backward, half_backward) {
+  f::ProgramDesc *program_desc = GetNewProgramDesc();
+  f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
+  f::BlockDescBind *block = program.Block(0);
+  auto *op1 = block->AppendOp();
+  op1->SetType("minus");
+  op1->SetInput("X", {"a"});
+  op1->SetInput("Y", {"b"});
+  op1->SetOutput("Out", {"out"});
+
+  AppendBackward(program, {"b"});
+  auto ops = block->AllOps();
+  ASSERT_EQ(2UL, ops.size());
 }
\ No newline at end of file

paddle/framework/block_desc.cc

@@ -66,7 +66,7 @@ std::vector<OpDescBind *> BlockDescBind::AllOps() const {
   return res;
 }
 
-void BlockDescBind::Sync() {
+void BlockDescBind::Flush() {
   if (need_update_) {
     auto &op_field = *this->desc_->mutable_ops();
     op_field.Clear();
@@ -91,5 +91,10 @@ BlockDescBind *BlockDescBind::ParentBlock() const {
   return prog_->Block(static_cast<size_t>(this->desc_->parent_idx()));
 }
 
+BlockDesc *BlockDescBind::Proto() {
+  Flush();
+  return desc_;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/block_desc.h

@@ -35,7 +35,8 @@ class BlockDescBind {
  public:
   friend std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
       ProgramDescBind &program_desc, int block_idx,
-      std::unordered_set<std::string> &no_grad_vars);
+      std::unordered_set<std::string> *no_grad_vars,
+      std::unordered_map<std::string, std::string> *grad_to_var);
 
   friend void AppendBackward(
       ProgramDescBind &program_desc,
@@ -64,9 +65,9 @@ class BlockDescBind {
 
   std::vector<OpDescBind *> AllOps() const;
 
-  void Sync();
+  void Flush();
 
-  BlockDesc *RawPtr() { return desc_; }
+  BlockDesc *Proto();
 
  private:
   ProgramDescBind *prog_;  // not_own

paddle/framework/details/op_registry.h

@@ -97,8 +97,11 @@ struct OpInfoFiller<T, kOpProtoAndCheckerMaker> {
 template <typename T>
 struct OpInfoFiller<T, kGradOpDescMaker> {
   void operator()(const char* op_type, OpInfo* info) const {
-    info->grad_op_maker_ = [](const OpDescBind& fwd_op) {
-      T maker(fwd_op);
+    info->grad_op_maker_ = [](
+        const OpDescBind& fwd_op,
+        const std::unordered_set<std::string>& no_grad_set,
+        std::unordered_map<std::string, std::string>* grad_to_var) {
+      T maker(fwd_op, no_grad_set, grad_to_var);
       return maker();
     };
   }

paddle/framework/executor_test.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <memory>
 #include <vector>
 
+#include "gflags/gflags.h"
 #include "gtest/gtest.h"
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/backward.h"
@@ -45,6 +46,7 @@ void AddOp(const std::string& type, const VariableNameMap& inputs,
   for (auto kv : outputs) {
     for (auto v : kv.second) {
       auto var = block->Var(v);
+      var->SetType(VarDesc::LOD_TENSOR);
       var->SetDataType(paddle::framework::DataType::FP32);
     }
   }
@@ -316,4 +318,14 @@ TEST_F(ExecutorTesterFeedAndFetch, GPU) {
     }
   }
 }
+
+DECLARE_double(fraction_of_gpu_memory_to_use);
+
+int main(int argc, char** argv) {
+  testing::InitGoogleTest(&argc, argv);
+  // Use less GPU memory for unittest.
+  FLAGS_fraction_of_gpu_memory_to_use = 0.25;
+  return RUN_ALL_TESTS();
+}
+
 #endif

paddle/framework/framework.proto

@@ -97,16 +97,26 @@ enum DataType {
   FP64 = 6;
 }
 
-message LoDTensorDesc {
+message TensorDesc {
   required DataType data_type = 1;
   repeated int64 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
-  optional int32 lod_level = 3 [ default = 0 ];
+}
+
+message LoDTensorDesc {
+  required TensorDesc tensor = 1;
+  optional int32 lod_level = 2 [ default = 0 ];
 }
 
 message VarDesc {
+  enum VarType {
+    LOD_TENSOR = 1;
+    SELECTED_ROWS = 2;
+  }
   required string name = 1;
-  optional LoDTensorDesc lod_tensor = 2;
-  optional bool persistable = 3 [ default = false ];
+  required VarType type = 2;
+  optional LoDTensorDesc lod_tensor = 3;
+  optional TensorDesc selected_rows = 4;
+  optional bool persistable = 5 [ default = false ];
 }
 
 message BlockDesc {

paddle/framework/grad_op_desc_maker.h

@@ -13,6 +13,8 @@
    limitations under the License. */
 
 #pragma once
+#include <string>
+#include <unordered_set>
 #include "paddle/framework/op_desc.h"
 #include "paddle/framework/operator.h"
 
@@ -21,27 +23,50 @@ namespace framework {
 
 class GradOpDescMakerBase {
  public:
-  explicit GradOpDescMakerBase(const OpDescBind& fwd_op) : fwd_op_(fwd_op) {}
+  explicit GradOpDescMakerBase(
+      const OpDescBind& fwd_op,
+      const std::unordered_set<std::string>& no_grad_set,
+      std::unordered_map<std::string, std::string>* grad_to_var)
+      : fwd_op_(fwd_op), no_grad_set_(no_grad_set), grad_to_var_(grad_to_var) {}
 
   virtual ~GradOpDescMakerBase() = default;
   virtual std::vector<std::unique_ptr<OpDescBind>> operator()() const = 0;
 
  protected:
-  static std::vector<std::string> ToGradNames(
-      const std::vector<std::string>& var_names) {
+  std::vector<std::string> InputGrad(const std::string& name,
+                                     bool drop_empty_grad = true) const {
     std::vector<std::string> ret_val;
+    auto var_names = this->Input(name);
     ret_val.reserve(var_names.size());
     std::transform(var_names.begin(), var_names.end(),
-                   std::back_inserter(ret_val), GradVarName);
-    return ret_val;
-  }
-
-  std::vector<std::string> InputGrad(const std::string& name) const {
-    return ToGradNames(fwd_op_.Input(name));
+                   std::back_inserter(ret_val),
+                   [this](const std::string& fwd_var_name) -> std::string {
+                     auto g_name = GradVarName(fwd_var_name);
+                     if (no_grad_set_.count(g_name)) {
+                       return kEmptyVarName;
+                     } else {
+                       (*this->grad_to_var_)[g_name] = fwd_var_name;
+                       return g_name;
+                     }
+                   });
+    if (!drop_empty_grad) {
+      return ret_val;
+    }
+    std::vector<std::string> dropped_ret_val;
+    dropped_ret_val.reserve(ret_val.size());
+    std::copy_if(ret_val.begin(), ret_val.end(),
+                 std::back_inserter(dropped_ret_val),
+                 [](const std::string& str) { return str != kEmptyVarName; });
+    return dropped_ret_val;
   }
 
   std::vector<std::string> OutputGrad(const std::string& name) const {
-    return ToGradNames(fwd_op_.Output(name));
+    std::vector<std::string> ret_val;
+    auto onames = this->Output(name);
+    ret_val.reserve(onames.size());
+    std::transform(onames.begin(), onames.end(), std::back_inserter(ret_val),
+                   GradVarName);
+    return ret_val;
   }
 
   std::vector<std::string> InputNames() const {
@@ -75,6 +100,8 @@ class GradOpDescMakerBase {
 
  private:
   const OpDescBind& fwd_op_;
+  const std::unordered_set<std::string>& no_grad_set_;
+  std::unordered_map<std::string, std::string>* grad_to_var_;
 };
 
 class SingleGradOpDescMaker : public GradOpDescMakerBase {
@@ -91,6 +118,7 @@ class SingleGradOpDescMaker : public GradOpDescMakerBase {
   virtual std::unique_ptr<OpDescBind> Apply() const = 0;
 };
 
+template <bool DropEmptyIG = true>
 class DefaultGradOpDescMaker : public SingleGradOpDescMaker {
  public:
   using SingleGradOpDescMaker::SingleGradOpDescMaker;
@@ -102,7 +130,8 @@ class DefaultGradOpDescMaker : public SingleGradOpDescMaker {
 
     for (auto& input_param : this->InputNames()) {
       grad->SetInput(input_param, this->Input(input_param));
-      grad->SetOutput(GradVarName(input_param), this->InputGrad(input_param));
+      grad->SetOutput(GradVarName(input_param),
+                      this->InputGrad(input_param, DropEmptyIG));
     }
 
     for (auto& output_param : this->OutputNames()) {

paddle/framework/op_desc.cc

@@ -32,7 +32,7 @@ OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs,
 }
 
 OpDesc *OpDescBind::Proto() {
-  Sync();
+  Flush();
   return &op_desc_;
 }
 
@@ -101,7 +101,7 @@ void OpDescBind::SetAttr(const std::string &name, const Attribute &v) {
 }
 
 void OpDescBind::SetBlockAttr(const std::string &name, BlockDescBind &block) {
-  BlockDesc *desc = block.RawPtr();
+  BlockDesc *desc = block.Proto();
   this->attrs_[name] = desc;
   need_update_ = true;
 }
@@ -165,7 +165,7 @@ struct SetAttrDescVisitor : public boost::static_visitor<void> {
   void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); }
 };
 
-void OpDescBind::Sync() {
+void OpDescBind::Flush() {
   if (need_update_) {
     this->op_desc_.mutable_inputs()->Clear();
     for (auto &ipt : inputs_) {

paddle/framework/op_desc.h

@@ -89,8 +89,6 @@ class OpDescBind {
     this->need_update_ = true;
   }
 
-  void Sync();
-
   const VariableNameMap &Inputs() const { return inputs_; }
 
   const VariableNameMap &Outputs() const { return outputs_; }
@@ -104,6 +102,8 @@ class OpDescBind {
 
   void InferShape(const BlockDescBind &block) const;
 
+  void Flush();
+
  private:
   template <typename MapType>
   static std::vector<typename MapType::key_type> MapKeys(const MapType &map) {

paddle/framework/op_registry.cc

@@ -59,11 +59,5 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDescBind& op_desc) {
                   op_desc.GetAttrMap());
 }
 
-std::vector<std::unique_ptr<OpDescBind>> OpRegistry::CreateGradOpDescs(
-    OpDescBind* op_desc) {
-  auto& info = OpInfoMap::Instance().Get(op_desc->Type());
-  return info.grad_op_maker_(*op_desc);
-}
-
 }  // namespace framework
 }  // namespace paddle

paddle/framework/op_registry.h

@@ -79,9 +79,6 @@ class OpRegistry {
 
   static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
-  static std::vector<std::unique_ptr<OpDescBind>> CreateGradOpDescs(
-      OpDescBind* op_desc);
-
   static std::unique_ptr<OperatorBase> CreateOp(const OpDescBind& op_desc);
 };
 
@@ -160,17 +157,18 @@ class OpKernelRegistrar : public Registrar {
 /**
  * Macro to register Operator.
  */
-#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type,           \
-                    grad_op_class)                                             \
-  REGISTER_OPERATOR(grad_op_type, grad_op_class);                              \
-  class _GradOpDescMaker_##grad_op_type##_                                     \
-      : public ::paddle::framework::DefaultGradOpDescMaker {                   \
-    using ::paddle::framework::DefaultGradOpDescMaker::DefaultGradOpDescMaker; \
-                                                                               \
-   protected:                                                                  \
-    virtual std::string GradOpType() const { return #grad_op_type; }           \
-  };                                                                           \
-  REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_,     \
+#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type,       \
+                    grad_op_class)                                         \
+  REGISTER_OPERATOR(grad_op_type, grad_op_class);                          \
+  class _GradOpDescMaker_##grad_op_type##_                                 \
+      : public ::paddle::framework::DefaultGradOpDescMaker<true> {         \
+    using ::paddle::framework::DefaultGradOpDescMaker<                     \
+        true>::DefaultGradOpDescMaker;                                     \
+                                                                           \
+   protected:                                                              \
+    virtual std::string GradOpType() const { return #grad_op_type; }       \
+  };                                                                       \
+  REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_, \
                     op_maker_class);
 
 #define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \

paddle/framework/program_desc.cc

@@ -45,7 +45,7 @@ BlockDescBind *ProgramDescBind::AppendBlock(const BlockDescBind &parent) {
 
 ProgramDesc *ProgramDescBind::Proto() {
   for (auto &block : blocks_) {
-    block->Sync();
+    block->Flush();
   }
   return prog_;
 }

paddle/framework/selected_rows.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/selected_rows.h"
+
+namespace paddle {
+namespace framework {}  // namespace framework
+}  // namespace paddle

paddle/framework/selected_rows.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 "paddle/framework/tensor.h"
+
+namespace paddle {
+namespace framework {
+
+class SelectedRows {
+ public:
+  SelectedRows(const std::vector<int64_t>& rows, const int64_t& height)
+      : rows_(rows), height_(height) {
+    value_.reset(new Tensor());
+  }
+
+  SelectedRows() { value_.reset(new Tensor()); }
+
+  platform::Place place() const { return value_->place(); }
+
+  const Tensor& value() const { return *value_; }
+
+  Tensor* mutable_value() { return value_.get(); }
+
+  int64_t height() const { return height_; }
+
+  void set_height(int64_t height) { height_ = height; }
+
+  const std::vector<int64_t>& rows() const { return rows_; }
+
+  void set_rows(const std::vector<int64_t>& rows) { rows_ = rows; }
+
+  DDim GetCompleteDims() const {
+    std::vector<int64_t> dims = vectorize(value_->dims());
+    dims[0] = height_;
+    return make_ddim(dims);
+  }
+
+ private:
+  std::vector<int64_t> rows_;
+  std::unique_ptr<Tensor> value_{nullptr};
+  int64_t height_;
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/selected_rows_test.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/selected_rows.h"
+#include "gtest/gtest.h"
+
+namespace paddle {
+namespace framework {
+
+class SelectedRowsTester : public ::testing::Test {
+ public:
+  virtual void SetUp() override {
+    std::vector<int64_t> rows{0, 4, 7};
+    int64_t height = 10;
+    int64_t row_numel = 100;
+    selected_rows_.reset(new SelectedRows(rows, height));
+
+    Tensor* value = selected_rows_->mutable_value();
+    value->mutable_data<float>(
+        make_ddim({static_cast<int64_t>(rows.size()), row_numel}), place_);
+  }
+
+ protected:
+  platform::CPUPlace place_;
+  std::unique_ptr<SelectedRows> selected_rows_{nullptr};
+};
+
+TEST_F(SelectedRowsTester, height) { ASSERT_EQ(selected_rows_->height(), 10); }
+
+TEST_F(SelectedRowsTester, dims) {
+  ASSERT_EQ(selected_rows_->value().dims(), make_ddim({3, 100}));
+}
+
+TEST_F(SelectedRowsTester, complete_dims) {
+  ASSERT_EQ(selected_rows_->GetCompleteDims(), make_ddim({10, 100}));
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/tensor_array.cc

@@ -76,6 +76,17 @@ LoDTensor PackDynamicBatch(const std::vector<LoDTensor>& source,
                            const std::vector<DySeqMeta>& meta, const LoD& lod,
                            size_t level);
 
+std::vector<size_t> GenDyBatchIndice(const DySeqMetaBatch& meta, int batch_id) {
+  // collect indice need to copy to the batch
+  std::vector<size_t> indice;
+  for (const auto& seq : meta) {
+    size_t id = seq.begin + batch_id;
+    if (id >= seq.end) break;
+    indice.push_back(id);
+  }
+  return indice;
+}
+
 }  // namespace detail
 
 const LoDTensor& TensorArray::Read(size_t index) const {
@@ -113,8 +124,8 @@ LoDTensor TensorArray::Pack(size_t level, const std::vector<DySeqMeta>& meta,
   return detail::PackDynamicBatch(values_, meta, lod, level);
 }
 
-std::vector<DySeqMeta> TensorArray::Unpack(const LoDTensor& source, int level,
-                                           bool length_desend) {
+DySeqMetaBatch TensorArray::Unpack(const LoDTensor& source, int level,
+                                   bool length_desend) {
   detail::DynamicBatchUnpacker unpacker(source, level,
                                         length_desend /*descend*/);
 
@@ -129,6 +140,7 @@ std::vector<DySeqMeta> TensorArray::Unpack(const LoDTensor& source, int level,
     Write(batch_id, unpacker.GetBatch(batch_id));
   }
 
+  PADDLE_ENFORCE(!unpacker.meta.empty());
   return unpacker.meta;
 }
 
@@ -218,13 +230,7 @@ LoDTensor DynamicBatchUnpacker::GetBatch(size_t index) {
   PADDLE_ENFORCE(!meta.empty(), "should build meta first");
   LoDTensor result;
 
-  // collect indice need to copy to the batch
-  std::vector<size_t> indice;
-  for (const auto& seq : meta) {
-    size_t id = seq.begin + index;
-    if (id >= seq.end) break;
-    indice.push_back(id);
-  }
+  auto indice = detail::GenDyBatchIndice(meta, index);
   PADDLE_ENFORCE(!indice.empty(), "invalid batch at %d", index);
 
   // copy the indice of records in LoDTensor
@@ -237,9 +243,9 @@ LoDTensor DynamicBatchUnpacker::GetBatch(size_t index) {
   for (size_t i = 0; i < indice.size(); i++) {
     auto index = indice[i];
     auto target = result.Slice<value_type>(i, i + 1);
-    auto source_ = source->Slice<value_type>(index, index + 1);
+    auto slice = source->Slice<value_type>(index, index + 1);
 
-    target.CopyFrom<value_type>(source_, platform::CPUPlace(),
+    target.CopyFrom<value_type>(slice, platform::CPUPlace(),
                                 platform::CPUDeviceContext());
   }
 

paddle/framework/tensor_array.h

@@ -34,6 +34,13 @@ struct DySeqMeta {
   size_t ori_idx;
 };
 
+using DySeqMetaBatch = std::vector<DySeqMeta>;
+
+/*
+ * Extract the indices of instances.
+ */
+std::vector<size_t> GenDyBatchIndice(const DySeqMetaBatch &metas, int batch_id);
+
 /*
  * TensorArray is a C-array-like array of tensors, it is meant to be used with
  * dynamic iteration primitives such as while_loop. It is used to segment inputs
@@ -69,7 +76,7 @@ class TensorArray {
    * Recover the original LoD-arranged LoDTensor with the `values`, `level` and
    * `indice_map`.
    */
-  LoDTensor Pack(size_t level, const std::vector<DySeqMeta> &meta,
+  LoDTensor Pack(size_t level, const DySeqMetaBatch &meta,
                  const LoD &lod) const;
 
   /*
@@ -77,8 +84,7 @@ class TensorArray {
    * `values`, if set `desend`, will sort by length in descending order else in
    * ascending order.
    */
-  std::vector<DySeqMeta> Unpack(const LoDTensor &source, int level,
-                                bool length_desend);
+  DySeqMetaBatch Unpack(const LoDTensor &source, int level, bool length_desend);
 
   /*
    * Pack the values into a tensor with rank one higher than each tensor in

paddle/framework/type_defs.h

@@ -36,8 +36,9 @@ using OpCreator = std::function<OperatorBase*(
     const std::string& /*type*/, const VariableNameMap& /*inputs*/,
     const VariableNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
 
-using GradOpMakerFN =
-    std::function<std::vector<std::unique_ptr<OpDescBind>>(const OpDescBind&)>;
+using GradOpMakerFN = std::function<std::vector<std::unique_ptr<OpDescBind>>(
+    const OpDescBind&, const std::unordered_set<std::string>& /*no_grad_set*/,
+    std::unordered_map<std::string, std::string>* /*grad_to_var*/)>;
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/var_desc.cc

@@ -13,32 +13,58 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/framework/var_desc.h"
+#include "paddle/platform/enforce.h"
 
 namespace paddle {
 namespace framework {
 
 void VarDescBind::SetShape(const std::vector<int64_t> &dims) {
-  VectorToRepeated(dims, desc_.mutable_lod_tensor()->mutable_dims());
+  VectorToRepeated(dims, mutable_tensor_desc()->mutable_dims());
 }
 
 void VarDescBind::SetDataType(DataType data_type) {
-  desc_.mutable_lod_tensor()->set_data_type(data_type);
+  mutable_tensor_desc()->set_data_type(data_type);
 }
 
 std::vector<int64_t> VarDescBind::Shape() const {
-  return RepeatedToVector(desc_.lod_tensor().dims());
+  return RepeatedToVector(tensor_desc().dims());
 }
 
-DataType VarDescBind::GetDataType() const {
-  return desc_.lod_tensor().data_type();
-}
+DataType VarDescBind::GetDataType() const { return tensor_desc().data_type(); }
 
 void VarDescBind::SetLoDLevel(int32_t lod_level) {
+  PADDLE_ENFORCE(desc_.type() == VarDesc::LOD_TENSOR);
   desc_.mutable_lod_tensor()->set_lod_level(lod_level);
 }
 
 int32_t VarDescBind::GetLodLevel() const {
+  PADDLE_ENFORCE(desc_.type() == VarDesc::LOD_TENSOR);
   return desc_.lod_tensor().lod_level();
 }
+
+const TensorDesc &VarDescBind::tensor_desc() const {
+  PADDLE_ENFORCE(desc_.has_type(), "invoke TensorDesc must after set type");
+  switch (desc_.type()) {
+    case VarDesc::SELECTED_ROWS:
+      return desc_.selected_rows();
+    case VarDesc::LOD_TENSOR:
+      return desc_.lod_tensor().tensor();
+    default:
+      PADDLE_THROW("Unexpected branch.");
+  }
+}
+
+TensorDesc *VarDescBind::mutable_tensor_desc() {
+  PADDLE_ENFORCE(desc_.has_type(),
+                 "invoke MutableTensorDesc must after set type");
+  switch (desc_.type()) {
+    case VarDesc::SELECTED_ROWS:
+      return desc_.mutable_selected_rows();
+    case VarDesc::LOD_TENSOR:
+      return desc_.mutable_lod_tensor()->mutable_tensor();
+    default:
+      PADDLE_THROW("Unexpected branch.");
+  }
+}
 }  // namespace framework
 }  // namespace paddle

paddle/framework/var_desc.h

@@ -72,7 +72,14 @@ class VarDescBind {
 
   int32_t GetLodLevel() const;
 
+  VarDesc::VarType GetType() const { return desc_.type(); }
+
+  void SetType(VarDesc::VarType type) { desc_.set_type(type); }
+
  private:
+  const TensorDesc &tensor_desc() const;
+  TensorDesc *mutable_tensor_desc();
+
   VarDesc desc_;
 };
 }  // namespace framework

paddle/operators/CMakeLists.txt

@@ -84,8 +84,9 @@ function(op_library TARGET)
     endif()
 
     # pybind USE_NO_KERNEL_OP
+    # HACK: if REGISTER_OP_CPU_KERNEL presents the operator must have kernel
     file(READ ${TARGET}.cc TARGET_CONTENT)
-    string(REGEX MATCH "OperatorWithKernel" regex_result "${TARGET_CONTENT}")
+    string(REGEX MATCH "REGISTER_OP_CPU_KERNEL" regex_result "${TARGET_CONTENT}")
     string(REPLACE "_op" "" TARGET "${TARGET}")
     if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "")
         file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(${TARGET});\n")

paddle/operators/activation_op.cc

@@ -338,6 +338,38 @@ class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker {
   }
 };
 
+template <typename AttrType>
+class HardSigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  HardSigmoidOpMaker(framework::OpProto *proto,
+                     framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "Input of HardSigmoid operator");
+    AddOutput("Y", "Output of HardSigmoid operator");
+    AddComment(R"DOC(
+Hard Sigmoid activation operator.
+
+Segment-wise linear approximation of sigmoid[1].
+This is much faster than sigmoid.
+
+hard_sigmoid = max(0, min(1, slope * x + shift))
+
+The slope should be positive. The offset can be either positive or negative.
+The default slope and shift are set from [1].
+It is recommended to use the defaults for this activation.
+
+References:
+  [1] Noisy Activation Functions
+      (https://arxiv.org/abs/1603.00391)
+
+    )DOC");
+    AddAttr<AttrType>("slope", "Slope for linear approximation of sigmoid")
+        .SetDefault(static_cast<AttrType>(0.2));
+    AddAttr<AttrType>("offset", "Offset for linear approximation of sigmoid")
+        .SetDefault(static_cast<AttrType>(0.5));
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -413,6 +445,9 @@ REGISTER_OP(thresholded_relu, ops::ActivationOp,
             ops::ThresholdedReluOpMaker<float>, thresholded_relu_grad,
             ops::ActivationOpGrad);
 
+REGISTER_OP(hard_sigmoid, ops::ActivationOp, ops::HardSigmoidOpMaker<float>,
+            hard_sigmoid_grad, ops::ActivationOpGrad);
+
 #define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor)        \
   REGISTER_OP_CPU_KERNEL(                                                      \
       act_type,                                                                \

paddle/operators/activation_op.h

@@ -616,30 +616,63 @@ struct ThresholdedReluGradFunctor : public BaseActivationFunctor<T> {
   }
 };
 
+template <typename T>
+struct HardSigmoidFunctor : public BaseActivationFunctor<T> {
+  float slope;
+  float offset;
+  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
+    return {{"slope", &slope}, {"offset", &offset}};
+  }
+
+  template <typename Device, typename X, typename Y>
+  void operator()(Device d, X x, Y y) const {
+    auto temp = x * static_cast<T>(slope) + static_cast<T>(offset);
+    y.device(d) = temp.cwiseMax(static_cast<T>(0)).cwiseMin(static_cast<T>(1));
+  }
+};
+
+template <typename T>
+struct HardSigmoidGradFunctor : public BaseActivationFunctor<T> {
+  float slope;
+  float offset;
+  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
+    return {{"slope", &slope}, {"offset", &offset}};
+  }
+
+  template <typename Device, typename X, typename Y, typename dY, typename dX>
+  void operator()(Device d, X x, Y y, dY dy, dX dx) const {
+    dx.device(d) =
+        dy *
+        ((y > static_cast<T>(0)) * (y < static_cast<T>(1))).template cast<T>() *
+        static_cast<T>(slope);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
-#define FOR_EACH_KERNEL_FUNCTOR(__macro)                          \
-  __macro(sigmoid, SigmoidFunctor, SigmoidGradFunctor);           \
-  __macro(logsigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor);  \
-  __macro(exp, ExpFunctor, ExpGradFunctor);                       \
-  __macro(relu, ReluFunctor, ReluGradFunctor);                    \
-  __macro(tanh, TanhFunctor, TanhGradFunctor);                    \
-  __macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor);  \
-  __macro(sqrt, SqrtFunctor, SqrtGradFunctor);                    \
-  __macro(abs, AbsFunctor, AbsGradFunctor);                       \
-  __macro(reciprocal, ReciprocalFunctor, ReciprocalGradFunctor);  \
-  __macro(log, LogFunctor, LogGradFunctor);                       \
-  __macro(square, SquareFunctor, SquareGradFunctor);              \
-  __macro(brelu, BReluFunctor, BReluGradFunctor);                 \
-  __macro(soft_relu, SoftReluFunctor, SoftReluGradFunctor);       \
-  __macro(pow, PowFunctor, PowGradFunctor);                       \
-  __macro(stanh, STanhFunctor, STanhGradFunctor);                 \
-  __macro(softplus, SoftplusFunctor, SoftplusGradFunctor);        \
-  __macro(softsign, SoftsignFunctor, SoftsignGradFunctor);        \
-  __macro(relu6, Relu6Functor, Relu6GradFunctor);                 \
-  __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor);    \
-  __macro(tanh_shrink, TanhShrinkFunctor, TanhShrinkGradFunctor); \
-  __macro(elu, ELUFunctor, ELUGradFunctor);                       \
-  __macro(hard_shrink, HardShrinkFunctor, HardShrinkGradFunctor); \
+#define FOR_EACH_KERNEL_FUNCTOR(__macro)                             \
+  __macro(sigmoid, SigmoidFunctor, SigmoidGradFunctor);              \
+  __macro(logsigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor);     \
+  __macro(exp, ExpFunctor, ExpGradFunctor);                          \
+  __macro(relu, ReluFunctor, ReluGradFunctor);                       \
+  __macro(tanh, TanhFunctor, TanhGradFunctor);                       \
+  __macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor);     \
+  __macro(sqrt, SqrtFunctor, SqrtGradFunctor);                       \
+  __macro(abs, AbsFunctor, AbsGradFunctor);                          \
+  __macro(reciprocal, ReciprocalFunctor, ReciprocalGradFunctor);     \
+  __macro(log, LogFunctor, LogGradFunctor);                          \
+  __macro(square, SquareFunctor, SquareGradFunctor);                 \
+  __macro(brelu, BReluFunctor, BReluGradFunctor);                    \
+  __macro(soft_relu, SoftReluFunctor, SoftReluGradFunctor);          \
+  __macro(pow, PowFunctor, PowGradFunctor);                          \
+  __macro(stanh, STanhFunctor, STanhGradFunctor);                    \
+  __macro(softplus, SoftplusFunctor, SoftplusGradFunctor);           \
+  __macro(softsign, SoftsignFunctor, SoftsignGradFunctor);           \
+  __macro(relu6, Relu6Functor, Relu6GradFunctor);                    \
+  __macro(leaky_relu, LeakyReluFunctor, LeakyReluGradFunctor);       \
+  __macro(tanh_shrink, TanhShrinkFunctor, TanhShrinkGradFunctor);    \
+  __macro(elu, ELUFunctor, ELUGradFunctor);                          \
+  __macro(hard_shrink, HardShrinkFunctor, HardShrinkGradFunctor);    \
+  __macro(hard_sigmoid, HardSigmoidFunctor, HardSigmoidGradFunctor); \
   __macro(thresholded_relu, ThresholdedReluFunctor, ThresholdedReluGradFunctor);

paddle/operators/adam_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/adam_op.h"
+
+namespace paddle {
+namespace operators {
+
+class AdamOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Param"),
+                   "Input(Param) of AdamOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Grad"),
+                   "Input(Grad) of AdamOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Moment1"),
+                   "Input(Moment1) of AdamOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Moment2"),
+                   "Input(Moment2) of AdamOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
+                   "Input(LearningRate) of AdamOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Beta1Pow"),
+                   "Input(Beta1Pow) of AdamOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Beta2Pow"),
+                   "Input(Beta2Pow) of AdamOp should not be null.");
+
+    PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
+                   "Output(ParamOut) of AdamOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Moment1Out"),
+                   "Output(Moment1Out) of AdamOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Moment2Out"),
+                   "Output(Moment2Out) of AdamOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Beta1PowOut"),
+                   "Output(Beta1PowOut) of AdamOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Beta2PowOut"),
+                   "Output(Beta2PowOut) of AdamOp should not be null.");
+
+    auto lr_dims = ctx->GetInputDim("LearningRate");
+    PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1,
+                      "Learning rate should have 1 dimension");
+    auto beta1_pow_dims = ctx->GetInputDim("Beta1Pow");
+    PADDLE_ENFORCE_EQ(framework::product(beta1_pow_dims), 1,
+                      "Beta1 power accumulator should have 1 dimension");
+    auto beta2_pow_dims = ctx->GetInputDim("Beta2Pow");
+    PADDLE_ENFORCE_EQ(framework::product(beta1_pow_dims), 1,
+                      "Beta1 power accumulator should have 1 dimension");
+
+    auto param_dims = ctx->GetInputDim("Param");
+    PADDLE_ENFORCE_EQ(
+        param_dims, ctx->GetInputDim("Grad"),
+        "Param and Grad input of AdamOp should have same dimension");
+    PADDLE_ENFORCE_EQ(
+        param_dims, ctx->GetInputDim("Moment1"),
+        "Param and Moment input of AdamOp should have same dimension");
+    PADDLE_ENFORCE_EQ(
+        param_dims, ctx->GetInputDim("Moment2"),
+        "Param and InfNorm input of AdamOp should have same dimension");
+
+    ctx->SetOutputDim("ParamOut", param_dims);
+    ctx->SetOutputDim("Moment1Out", param_dims);
+    ctx->SetOutputDim("Moment2Out", param_dims);
+    ctx->SetOutputDim("Beta1PowOut", beta1_pow_dims);
+    ctx->SetOutputDim("Beta2PowOut", beta2_pow_dims);
+  }
+};
+
+class AdamOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  AdamOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Param", "(Tensor) Input parameter");
+    AddInput("Grad", "(Tensor) Input gradient");
+    AddInput("LearningRate", "(Tensor) Learning rate");
+    AddInput("Moment1", "(Tensor) Input first moment");
+    AddInput("Moment2", "(Tensor) Input second moment");
+    AddInput("Beta1Pow", "(Tensor) Input beta1 power accumulator");
+    AddInput("Beta2Pow", "(Tensor) Input beta2 power accumulator");
+
+    AddOutput("ParamOut", "(Tensor) Output parameter");
+    AddOutput("Moment1Out", "(Tensor) Output first moment");
+    AddOutput("Moment2Out", "(Tensor) Output second moment");
+    AddOutput("Beta1PowOut", "(Tensor) Output beta1 power accumulator");
+    AddOutput("Beta2PowOut", "(Tensor) Output beta2 power accumulator");
+
+    AddAttr<float>("beta1",
+                   "(float, default 0.9) "
+                   "Exponential decay rate for the "
+                   "first moment estimates.")
+        .SetDefault(0.9f);
+    AddAttr<float>("beta2",
+                   "(float, default 0.999) "
+                   "exponential decay rate for the "
+                   "second moment estimates.")
+        .SetDefault(0.999f);
+    AddAttr<float>("epsilon",
+                   "(float, default 1.0e-8) "
+                   "Constant for numerical stability")
+        .SetDefault(1.0e-8f);
+
+    AddComment(R"DOC(
+Adam Updates Operator.
+
+This implements the Adam optimizer from Section 2 of the Adam
+paper[1]. Adam is a first-order gradient-based optimization
+method based on adaptive estimates of lower-order moments.
+
+Adam updates:
+
+moment1_out = beta1 * moment1 + (1 − beta1) * grad
+moment2_out = beta2 * moment2 + (1 − beta2) * grad * grad
+beta1_pow_out = beta1_pow * beta1
+beta2_pow_out = beta2_pow * beta2
+learning_rate_t = learning_rate_t *
+                  sqrt(1 - beta2_pow_out) / (1 - beta1_pow_out)
+param_out = param - learning_rate_t * moment1/ (sqrt(moment2) + epsilon)
+
+References:
+  [1] Adam: A Method for Stochastic Optimization
+      (https://arxiv.org/abs/1412.6980)
+
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(adam, ops::AdamOp, ops::AdamOpMaker);
+REGISTER_OP_CPU_KERNEL(adam,
+                       ops::AdamOpKernel<paddle::platform::CPUPlace, float>);

paddle/operators/adam_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/adam_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(adam,
+                       ops::AdamOpKernel<paddle::platform::GPUPlace, float>);

paddle/operators/adam_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 "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class AdamOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
+    auto moment1_out_tensor = ctx.Output<framework::Tensor>("Moment1Out");
+    auto moment2_out_tensor = ctx.Output<framework::Tensor>("Moment2Out");
+    auto beta1_pow_out_tensor = ctx.Output<framework::Tensor>("Beta1PowOut");
+    auto beta2_pow_out_tensor = ctx.Output<framework::Tensor>("Beta2PowOut");
+
+    param_out_tensor->mutable_data<T>(ctx.GetPlace());
+    moment1_out_tensor->mutable_data<T>(ctx.GetPlace());
+    moment2_out_tensor->mutable_data<T>(ctx.GetPlace());
+    beta1_pow_out_tensor->mutable_data<T>(ctx.GetPlace());
+    beta2_pow_out_tensor->mutable_data<T>(ctx.GetPlace());
+
+    float beta1 = ctx.Attr<float>("beta1");
+    float beta2 = ctx.Attr<float>("beta2");
+    float epsilon = ctx.Attr<float>("epsilon");
+
+    auto param = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("Param"));
+    auto grad = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("Grad"));
+    auto moment1 = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("Moment1"));
+    auto moment2 = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("Moment2"));
+    auto lr = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("LearningRate"));
+    auto beta1_pow = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("Beta1Pow"));
+    auto beta2_pow = framework::EigenVector<T>::Flatten(
+        *ctx.Input<framework::Tensor>("Beta2Pow"));
+    auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
+    auto moment1_out = framework::EigenVector<T>::Flatten(*moment1_out_tensor);
+    auto moment2_out = framework::EigenVector<T>::Flatten(*moment2_out_tensor);
+    auto beta1_pow_out =
+        framework::EigenVector<T>::Flatten(*beta1_pow_out_tensor);
+    auto beta2_pow_out =
+        framework::EigenVector<T>::Flatten(*beta2_pow_out_tensor);
+    auto place = ctx.GetEigenDevice<Place>();
+
+    moment1_out.device(place) = beta1 * moment1 + (1 - beta1) * grad;
+    moment2_out.device(place) = beta2 * moment2 + (1 - beta2) * grad.square();
+    beta1_pow_out.device(place) = beta1_pow * beta1;
+    beta2_pow_out.device(place) = beta2_pow * beta2;
+    // All of these are tensors of 1 element
+    auto lr_t = lr * (1 - beta2_pow_out).sqrt() / (1 - beta1_pow_out);
+    // Eigen does not support automatic broadcast
+    // Get dimensions of moment vector to broadcast lr_t
+    Eigen::DSizes<int, 1> m_dsize(moment1_out_tensor->numel());
+    param_out.device(place) =
+        param -
+        lr_t.broadcast(m_dsize) *
+            (moment1_out / (moment2_out.sqrt() + epsilon));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/dynamic_recurrent_op.cc

@@ -23,6 +23,7 @@ using framework::Scope;
 using framework::TensorArray;
 using framework::LoDTensor;
 using framework::Variable;
+using framework::DySeqMetaBatch;
 
 namespace detail {
 
@@ -33,6 +34,29 @@ inline void CreateVariables(Scope& scope,
   }
 }
 
+/*
+ * The inputs with sequence should be reordered when they are split, so the
+ * boot_states should be reordered in the same order.
+ *
+ * NOTE This may require that the `pre_state` of the first time step should just
+ * copy the `boot_state` rather than reference it, for that the content should
+ * be reordered, but the RNN op should not change the `boot_state` as an input
+ * variable's content.
+ */
+template <typename T>
+inline void ReorderBootState(const DySeqMetaBatch& metas,
+                             const LoDTensor& boot_state, LoDTensor* tensor,
+                             const platform::Place& dst_place) {
+  for (size_t seq_id = 0; seq_id < metas.size(); seq_id++) {
+    auto slice = tensor->Slice<T>(seq_id, seq_id + 1);
+    auto boot_slice =
+        boot_state.Slice<T>(metas[seq_id].ori_idx, metas[seq_id].ori_idx + 1);
+    // TODO(superjom) pass in device context as an argument
+    slice.template CopyFrom<T>(boot_slice, dst_place,
+                               platform::CPUDeviceContext());
+  }
+}
+
 }  // namespace detail
 
 class DynamicRecurrentOpProtoAndCheckerMaker
@@ -69,6 +93,7 @@ void DynamicRecurrentOp::Run(const Scope& scope,
   CreateScopes();
   WriteStepInputs();
   InitStates();
+  WriteStepOutputs();
 
   // call stepnet in all the time steps
   for (size_t step = 0; step < cache_.num_steps; step++) {
@@ -76,7 +101,6 @@ void DynamicRecurrentOp::Run(const Scope& scope,
     stepnet_->Run(step_scope, dev_ctx);
   }
 
-  WriteStepOutputs();
   ConcatOutputs();
 }
 
@@ -84,11 +108,11 @@ void DynamicRecurrentOp::SplitInputs() const {
   // TODO(superjom) make level a config
   // TODO(superjom) check all the inputs has the same LoD
   int level = 0;
-  const auto& inlinks = cache_.inlinks;
-  for (const auto& item : inlinks) {
+  for (const auto& item : cache_.inlinks) {
     const auto& var = item.second;
     const auto& tensor = var->Get<LoDTensor>();
     TensorArray& ta = step_inputs_[item.first];
+
     dy_seq_metas_[item.first] =
         ta.Unpack(tensor, level, true /*length_descend*/);
 
@@ -120,17 +144,11 @@ void DynamicRecurrentOp::WriteStepInputs() const {
 }
 
 void DynamicRecurrentOp::WriteStepOutputs() const {
-  for (size_t step = 0; step < cache_.scopes->size(); step++) {
-    auto& scope = cache_.GetScope(step);
-    for (auto& item : step_outputs_) {
-      auto* var = scope.FindVar(item.first);
-      if (var == nullptr) {
-        var = scope.Var(item.first);
-      }
-      auto* tensor = var->GetMutable<LoDTensor>();
-      item.second.WriteShared(step, *tensor);
-    }
+  // initialize step outputs
+  for (const auto& item : cache_.outlinks) {
+    step_outputs_.emplace(item.first, TensorArray());
   }
+  PADDLE_ENFORCE_GT(step_outputs_.size(), 0UL);
 }
 
 void DynamicRecurrentOp::CreateScopes() const {
@@ -145,12 +163,18 @@ void DynamicRecurrentOp::CreateScopes() const {
   PADDLE_ENFORCE_NOT_NULL(stepnet_, "stepnet should be set first");
   std::vector<std::string> memories;
   std::vector<std::string> pre_memories;
+  std::vector<std::string> stepnet_outputs;
   std::transform(arg_.memories.begin(), arg_.memories.end(),
                  std::back_inserter(memories),
                  [](const rnn::MemoryAttr& m) { return m.var; });
   std::transform(arg_.memories.begin(), arg_.memories.end(),
                  std::back_inserter(pre_memories),
                  [](const rnn::MemoryAttr& m) { return m.pre_var; });
+  for (const auto& item : stepnet_->Outputs()) {
+    for (const auto& var : item.second) {
+      stepnet_outputs.push_back(var);
+    }
+  }
 
   for (size_t step = 0; step < cache_.num_steps; step++) {
     auto& scope = cache_.GetScope(step);
@@ -158,60 +182,88 @@ void DynamicRecurrentOp::CreateScopes() const {
     detail::CreateVariables(scope, arg_.outlinks);
     detail::CreateVariables(scope, memories);
     detail::CreateVariables(scope, pre_memories);
+    detail::CreateVariables(scope, stepnet_outputs);
   }
 }
 
 void DynamicRecurrentOp::ConcatOutputs() const {
   // TODO(superjom) transform this to a config
   int level = 0;
-  // TODO(superjom) pass in some lod
-  // just a placeholder
-  framework::LoD lod;
+  for (size_t step = 0; step < cache_.num_steps; step++) {
+    auto& scope = cache_.GetScope(step);
+    for (auto& item : step_outputs_) {
+      auto* var = scope.FindVar(item.first);
+      PADDLE_ENFORCE_NOT_NULL(var);
+      auto* tensor = var->GetMutable<LoDTensor>();
+      tensor->mutable_data<value_type>(platform::CPUPlace());
+      item.second.WriteShared(step, *tensor);
+    }
+  }
+  // the inlinks' lods should be the same, so randomly get one lod.
+  const auto& some_lod =
+      cache_.scope->FindVar(arg_.inlinks.front())->Get<LoDTensor>().lod();
+  const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()];
   for (auto& item : step_outputs_) {
-    auto tensor = item.second.Pack(level, dy_seq_metas_[item.first], lod);
-    auto& output = cache_.outlinks[item.first]->Get<LoDTensor>();
-    const_cast<LoDTensor*>(&output)->ShareDataWith<value_type>(tensor);
+    auto tensor = item.second.Pack(level, some_meta, some_lod);
+    auto* output = cache_.outlinks[item.first]->GetMutable<LoDTensor>();
+    const_cast<LoDTensor*>(output)->ShareDataWith<value_type>(tensor);
   }
 }
 
 void DynamicRecurrentOp::InitStates() const {
-  // init the first state
-  // TODO(superjom) parepare the scenerio that boot state not exists
-  for (auto memory : arg_.memories) {
-    auto* boot_state_var = cache_.scope->FindVar(memory.boot_var);
-    PADDLE_ENFORCE_NOT_NULL(boot_state_var);
-    auto& boot_state = boot_state_var->Get<LoDTensor>();
-    const auto& dims = boot_state.dims();
-
-    for (size_t step = 0; step < cache_.num_steps; step++) {
-      auto& cur_scope = cache_.GetScope(step);
-      // link pre-state to boot_state
-      // init state and pre-state
-      auto* pre_state = cur_scope.FindVar(memory.pre_var);
-      PADDLE_ENFORCE_NOT_NULL(pre_state);
-      pre_state->GetMutable<LoDTensor>();
-
-      auto* state = cur_scope.FindVar(memory.var);
-      PADDLE_ENFORCE_NOT_NULL(state);
-      state->GetMutable<LoDTensor>()->Resize(dims);
-      state->GetMutable<LoDTensor>()->mutable_data<value_type>(
-          platform::CPUPlace());
-
-      if (step == 0) {
-        auto* pre_state_tensor = pre_state->GetMutable<LoDTensor>();
-        pre_state_tensor->Resize(boot_state.dims());
-        pre_state_tensor->ShareDataWith<value_type>(boot_state);
-      } else {
-        auto& pre_scope = cache_.GetScope(step - 1);
-        auto* state_pre = pre_scope.FindVar(memory.var);
-        PADDLE_ENFORCE_NOT_NULL(state_pre);
-        pre_state->GetMutable<LoDTensor>()->ShareDataWith<value_type>(
-            *state_pre->GetMutable<LoDTensor>());
-      }
+  for (size_t step = 0; step < cache_.num_steps; step++) {
+    for (const auto& memory : arg_.memories) {
+      CreateState(memory, step);
+      LinkState(memory, step);
     }
   }
 }
 
+void DynamicRecurrentOp::CreateState(const rnn::MemoryAttr& memory,
+                                     size_t step) const {
+  auto& scope = cache_.GetScope(step);
+  auto& state = *cache_.GetTensor(scope, memory.var);
+  auto& boot_state = *cache_.GetTensor(*cache_.scope, memory.boot_var);
+
+  size_t num_instances =
+      step_inputs_[arg_.inlinks.front()].Read(step).dims()[0];
+  auto dims = boot_state.dims();
+  dims[0] = num_instances;
+
+  state.Resize(dims);
+  state.mutable_data<value_type>(platform::CPUPlace());
+  states_[memory.var].WriteShared(step, state);
+}
+
+void DynamicRecurrentOp::LinkState(const rnn::MemoryAttr& memory,
+                                   size_t step) const {
+  auto& scope = cache_.GetScope(step);
+  auto& state_pre = *cache_.GetTensor(scope, memory.pre_var);
+
+  // all the step_inputs' metas should be the same, just randomly select one
+  // and get the dyseq meta.
+  const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()];
+  size_t num_instances =
+      step_inputs_[arg_.inlinks.front()].Read(step).dims()[0];
+
+  LoDTensor* pre_state{nullptr};
+  if (step == 0) {
+    pre_state = cache_.GetTensor(*cache_.scope, memory.boot_var);
+    pre_state->mutable_data<float>(platform::CPUPlace());
+    // allocate memory
+    state_pre.Resize(pre_state->dims());
+    state_pre.mutable_data<value_type>(platform::CPUPlace());
+    detail::ReorderBootState<value_type>(some_meta, *pre_state, &state_pre,
+                                         pre_state->place());
+  } else {
+    pre_state = cache_.GetTensor(cache_.GetScope(step - 1), memory.var);
+  }
+
+  // shink and share from previous state
+  auto shrinked_pre_state = pre_state->Slice<value_type>(0, num_instances);
+  state_pre.ShareDataWith<value_type>(shrinked_pre_state);
+}
+
 void DynamicRecurrentOp::ArgCache::Init(
     const rnn::ArgumentName& name, const paddle::framework::OperatorBase& op,
     const paddle::framework::Scope& scope, rnn::Argument* arg) {
@@ -261,6 +313,12 @@ Variable* DynamicRecurrentOp::ArgCache::GetVariable(const Scope& scope,
   return var;
 }
 
+LoDTensor* DynamicRecurrentOp::ArgCache::GetTensor(
+    const framework::Scope& scope, const std::string& name) {
+  auto* var = GetVariable(scope, name);
+  return var->GetMutable<LoDTensor>();
+}
+
 const rnn::ArgumentName DynamicRecurrentOp::kArgName{
     "step_net", "step_scopes",  "inlinks",      "outlinks",
     "memories", "pre_memories", "boot_memories"};

paddle/operators/dynamic_recurrent_op.h

@@ -77,6 +77,17 @@ class DynamicRecurrentOp : public framework::OperatorBase {
    */
   void InitStates() const;
 
+  /*
+   * Create state variables for each time step.
+   */
+  void CreateState(const rnn::MemoryAttr& memory, size_t step) const;
+
+  /*
+   * Link pre-state variable in current scope to the state variable in the
+   * previous time step (scope).
+   */
+  void LinkState(const rnn::MemoryAttr& memory, size_t step) const;
+
   /*
    * Concatenate outputs in each time step and generate a LoDTensor.
    */
@@ -91,6 +102,16 @@ class DynamicRecurrentOp : public framework::OperatorBase {
   }
   const OperatorBase& GetStepNet() const { return *stepnet_; }
 
+  const framework::TensorArray& state(const std::string& name) const {
+    return states_[name];
+  }
+  const framework::TensorArray& step_input(const std::string& name) const {
+    return step_inputs_[name];
+  }
+  const framework::TensorArray& step_output(const std::string& name) const {
+    return step_outputs_[name];
+  }
+
  protected:
   struct ArgCache {
     framework::Scope const* scope;
@@ -108,6 +129,9 @@ class DynamicRecurrentOp : public framework::OperatorBase {
       return *scopes->at(index);
     }
 
+    framework::LoDTensor* GetTensor(const framework::Scope& scope,
+                                    const std::string& name);
+
    private:
     void InitArgument(const rnn::ArgumentName& name, const OperatorBase& op,
                       rnn::Argument* arg);
@@ -122,7 +146,7 @@ class DynamicRecurrentOp : public framework::OperatorBase {
 
  private:
   std::unique_ptr<OperatorBase> stepnet_;
-  mutable framework::TensorArray states_;
+  mutable std::map<std::string, framework::TensorArray> states_;
   mutable std::map<std::string, framework::TensorArray> step_inputs_;
   mutable std::map<std::string, framework::TensorArray> step_outputs_;
   mutable std::map<std::string, std::vector<framework::DySeqMeta>>

paddle/operators/dynamic_recurrent_op_test.cc

@@ -87,7 +87,6 @@ class DynamicRecurrentOpTestHelper : public ::testing::Test {
     platform::CPUPlace place;
     scope.Var("step_scopes");
     CreateVar(scope, "boot_mem", framework::make_ddim({10, 20}), place);
-    // auto* out0 =
     CreateVar(scope, "out0", framework::make_ddim({10, 20}), place);
     auto* in0 = CreateVar(scope, "in0", framework::make_ddim({10, 8}), place);
     // 10 instanes with 4 sentences, length is 4, 3, 2, 1 respectively.

paddle/operators/gru_unit_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/gru_unit_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+class GRUUnitOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(%s) of GRUUnitOp should not be null.", "Input");
+    PADDLE_ENFORCE(ctx->HasInput("HiddenPrev"),
+                   "Input(%s) of GRUUnitOp should not be null.", "HiddenPrev");
+    PADDLE_ENFORCE(ctx->HasInput("Weight"),
+                   "Input(%s) of GRUUnitOp should not be null.", "Weight");
+    PADDLE_ENFORCE(ctx->HasOutput("Gate"),
+                   "Output(%s) of GRUUnitOp should not be null.", "Gate");
+    PADDLE_ENFORCE(ctx->HasOutput("ResetHiddenPrev"),
+                   "Output(%s) of GRUUnitOp should not be null.",
+                   "ResetHiddenPrev");
+    PADDLE_ENFORCE(ctx->HasOutput("Hidden"),
+                   "Output(%s) of GRUUnitOp should not be null.", "Hidden");
+    auto input_dims = ctx->GetInputDim("Input");
+    auto hidden_prev_dims = ctx->GetInputDim("HiddenPrev");
+    auto weight_dims = ctx->GetInputDim("Weight");
+    int batch_size = input_dims[0];
+    int input_size = input_dims[1];
+    int frame_size = hidden_prev_dims[1];
+    int weight_height = weight_dims[0];
+    int weight_width = weight_dims[1];
+    PADDLE_ENFORCE_EQ(
+        input_size, frame_size * 3,
+        "The input_size must be 3 times of frame_size in GRUUnitOp.");
+    PADDLE_ENFORCE_EQ(
+        weight_height, frame_size,
+        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
+    PADDLE_ENFORCE_EQ(
+        weight_width, frame_size * 3,
+        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
+    auto bias = Input("Bias");
+    if (bias != framework::kEmptyVarName) {
+      auto bias_dims = ctx->GetInputDim("Bias");
+      int bias_height = bias_dims[0];
+      int bias_width = bias_dims[1];
+      PADDLE_ENFORCE_EQ(bias_height, 1,
+                        "The shape of Bias must be [1, frame_size * 3].");
+      PADDLE_ENFORCE_EQ(bias_width, frame_size * 3,
+                        "The shape of Bias must be [1, frame_size * 3].");
+    }
+    ctx->SetOutputDim("Gate", {batch_size, frame_size * 3});
+    ctx->SetOutputDim("ResetHiddenPrev", {batch_size, frame_size});
+    ctx->SetOutputDim("Hidden", {batch_size, frame_size});
+  }
+};
+
+class GRUUnitOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  GRUUnitOpMaker(framework::OpProto* proto,
+                 framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Input",
+             "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the "
+             "input.");
+    AddInput("HiddenPrev",
+             "(Tensor) Matrix with shape [batch_size, frame_size] for the "
+             "states of previous time step.");
+    AddInput("Weight",
+             "(Tensor) Weight matrix with shape [frame_size, frame_size * 3]. "
+             "The elements continuous in memory can be divided into two parts. "
+             "The first part are weights of the update gate and reset gate "
+             "with shape [frame_size, frame_size * 2], and the second part are "
+             "weights of output candidate with shape [frame_size, frame_size]");
+    AddInput("Bias",
+             "(Tensor) Bias vector with shape [1, frame_size * 3] concating "
+             "bias of the update gate, reset gate and output candidate.");
+    AddOutput("Gate",
+              "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the "
+              "output of update gate, reset gate and output candidate")
+        .AsIntermediate();
+    AddOutput("ResetHiddenPrev",
+              "(Tensor) Matrix with shape [batch_size, frame_size] for the "
+              "reseted hidden state of previous time step.")
+        .AsIntermediate();
+    AddOutput("Hidden",
+              "(Tensor) The GRU hidden state of the current time step "
+              "with shape [batch_size, frame_size].");
+    AddAttr<int>("activation",
+                 "(enum int, default tanh) "
+                 "The activation type used for output candidate {h}_t.")
+        .SetDefault(tanh)
+        .InEnum({identity, sigmoid, tanh, relu});
+    AddAttr<int>("gate_activation",
+                 "(enum int, default sigmoid) "
+                 "The activation type used in update gate and reset gate.")
+        .SetDefault(sigmoid)
+        .InEnum({identity, sigmoid, tanh, relu});
+    AddComment(R"DOC(
+GRUUnitOp implements part calculations of the GRU unit as following:
+
+\f[
+update \ gate: u_t = actGate(xu_t + W_u * hidden_prev + bias_u) \\
+reset \ gate: r_t = actGate(xr_t + W_r * hidden_prev + bias_r)  \\
+output \ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, hidden_prev) + bias_c) \\
+output: h_t = dot((1-u_t), {h}_t) + dot(u_t, hidden_prev)
+\f]
+
+The rest of GRU unit can be completed by using FCOp's output as the input of GRUUnitOp.
+)DOC");
+  }
+};
+
+class GRUUnitGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(%s) of GRUUnitGradOp should not be null.", "Input");
+    PADDLE_ENFORCE(ctx->HasInput("HiddenPrev"),
+                   "Input(%s) of GRUUnitGradOp should not be null.",
+                   "HiddenPrev");
+    PADDLE_ENFORCE(ctx->HasInput("Weight"),
+                   "Input(%s) of GRUUnitGradOp should not be null.", "Weight");
+    PADDLE_ENFORCE(ctx->HasInput("Gate"),
+                   "Input(%s) of GRUUnitGradOp should not be null.", "Gate");
+    PADDLE_ENFORCE(ctx->HasInput("ResetHiddenPrev"),
+                   "Input(%s) of GRUUnitGradOp should not be null.",
+                   "ResetHiddenPrev");
+    PADDLE_ENFORCE(ctx->HasInput("Hidden"),
+                   "Input(%s) of GRUUnitGradOp should not be null.", "Hidden");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Gate")),
+                   "Input(%s@GRAD) of GRUUnitGradOp should not be null.",
+                   "Gate");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("ResetHiddenPrev")),
+                   "Input(%s@GRAD) of GRUUnitGradOp should not be null.",
+                   "ResetHiddenPrev");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Hidden")),
+                   "Input(%s@GRAD) of GRUUnitGradOp should not be null.",
+                   "Hidden");
+    auto input_dims = ctx->GetInputDim("Input");
+    auto hidden_prev_dims = ctx->GetInputDim("HiddenPrev");
+    auto weight_dims = ctx->GetInputDim("Weight");
+    // int batch_size = input_dims[0];
+    int input_size = input_dims[1];
+    int frame_size = hidden_prev_dims[1];
+    int weight_height = weight_dims[0];
+    int weight_width = weight_dims[1];
+    PADDLE_ENFORCE_EQ(
+        input_size, frame_size * 3,
+        "The input_size must be 3 times of frame_size in GRUUnitOp.");
+    PADDLE_ENFORCE_EQ(
+        weight_height, frame_size,
+        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
+    PADDLE_ENFORCE_EQ(
+        weight_width, frame_size * 3,
+        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
+    auto bias = Input("Bias");
+    if (bias != framework::kEmptyVarName) {
+      auto bias_dims = ctx->GetInputDim("Bias");
+      int bias_height = bias_dims[0];
+      int bias_width = bias_dims[1];
+      PADDLE_ENFORCE_EQ(bias_height, 1,
+                        "The shape of Bias must be [1, frame_size * 3].");
+      PADDLE_ENFORCE_EQ(bias_width, frame_size * 3,
+                        "The shape of Bias must be [1, frame_size * 3].");
+      auto bias_grad_name = framework::GradVarName("Bias");
+      if (ctx->HasOutput(bias_grad_name))
+        ctx->SetOutputDim(bias_grad_name, bias_dims);
+    }
+    auto input_grad_name = framework::GradVarName("Input");
+    if (ctx->HasOutput(input_grad_name))
+      ctx->SetOutputDim(input_grad_name, input_dims);
+    auto hidden_prev_grad_name = framework::GradVarName("HiddenPrev");
+    if (ctx->HasOutput(hidden_prev_grad_name))
+      ctx->SetOutputDim(hidden_prev_grad_name, hidden_prev_dims);
+    auto weight_grad_name = framework::GradVarName("Weight");
+    if (ctx->HasOutput(weight_grad_name))
+      ctx->SetOutputDim(weight_grad_name, weight_dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(gru_unit, ops::GRUUnitOp, ops::GRUUnitOpMaker, gru_unit_grad,
+            ops::GRUUnitGradOp);
+REGISTER_OP_CPU_KERNEL(gru_unit,
+                       ops::GRUUnitKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    gru_unit_grad, ops::GRUUnitGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/gru_unit_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/gru_unit_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(gru_unit,
+                       ops::GRUUnitKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    gru_unit_grad, ops::GRUUnitGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/gru_unit_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 "paddle/operators/activation_op.h"
+#include "paddle/operators/math/math_function.h"
+
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+enum GRUActivationType { identity = 0, sigmoid = 1, tanh = 2, relu = 3 };
+
+template <typename Place, typename T>
+class GRUUnitKernel : public framework::OpKernel<T> {
+ public:
+  template <typename Device, typename X, typename Y>
+  void ActCompute(const int act_type, const Device& d, X x, Y y) const {
+    if (act_type == identity)
+      y.device(d) = x;
+    else if (act_type == sigmoid)
+      SigmoidFunctor<T>()(d, x, y);
+    else if (act_type == tanh)
+      TanhFunctor<T>()(d, x, y);
+    else if (act_type == relu)
+      ReluFunctor<T>()(d, x, y);
+    else
+      PADDLE_THROW("unsupported activation type");
+  }
+
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* input = context.Input<Tensor>("Input");
+    auto* hidden_prev = context.Input<Tensor>("HiddenPrev");
+    auto* weight = context.Input<Tensor>("Weight");
+    auto* bias = context.Input<Tensor>("Bias");
+    auto* gate = context.Output<Tensor>("Gate");
+    gate->mutable_data<T>(context.GetPlace());
+    auto* reset_hidden_prev = context.Output<Tensor>("ResetHiddenPrev");
+    reset_hidden_prev->mutable_data<T>(context.GetPlace());
+    auto* hidden = context.Output<Tensor>("Hidden");
+    hidden->mutable_data<T>(context.GetPlace());
+
+    int batch_size = input->dims()[0];
+    int frame_size = hidden_prev->dims()[1];
+
+    auto x = EigenMatrix<T>::From(*input);
+    auto h_p = EigenMatrix<T>::From(*hidden_prev);
+    auto g = EigenMatrix<T>::From(*gate);
+    auto r_h_p = EigenMatrix<T>::From(*reset_hidden_prev);
+    auto h = EigenMatrix<T>::From(*hidden);
+    auto place = context.GetEigenDevice<Place>();
+
+    // calculate unactivated gate outputs
+    if (bias) {
+      auto b = EigenMatrix<T>::From(*bias);
+      g.device(place) = x +
+                        b.reshape(Eigen::array<int, 2>({{1, frame_size * 3}}))
+                            .broadcast(Eigen::array<int, 2>({{batch_size, 1}}));
+    } else {
+      g.device(place) = x;
+    }
+    const T* hidden_prev_data = hidden_prev->data<T>();
+    const T* weight_data = weight->data<T>();
+    T* gate_data = gate->data<T>();
+    T* reset_hidden_prev_data = reset_hidden_prev->data<T>();
+    math::gemm<Place, T>(context.device_context(), false, false, batch_size,
+                         2 * frame_size, frame_size, 1, hidden_prev_data,
+                         frame_size, weight_data, frame_size * 2, 1, gate_data,
+                         frame_size * 3);
+
+    // calculate activited gate
+    Eigen::array<int, 2> extents({{batch_size, frame_size}});
+    Eigen::array<int, 2> u_offsets({{0, 0}});
+    ActCompute(context.Attr<int>("gate_activation"), place,
+               g.slice(u_offsets, extents), g.slice(u_offsets, extents));
+    auto u = g.slice(u_offsets, extents);  // update gate
+    Eigen::array<int, 2> r_offsets({{0, frame_size}});
+    ActCompute(context.Attr<int>("gate_activation"), place,
+               g.slice(r_offsets, extents), g.slice(r_offsets, extents));
+    auto r = g.slice(r_offsets, extents);  // reset gate
+    r_h_p.device(place) = r * h_p;         // reset previous hidden state
+    math::gemm<Place, T>(context.device_context(), false, false, batch_size,
+                         frame_size, frame_size, 1, reset_hidden_prev_data,
+                         frame_size, weight_data + frame_size * frame_size * 2,
+                         frame_size, 1, gate_data + frame_size * 2,
+                         frame_size * 3);
+
+    Eigen::array<int, 2> c_offsets({{0, frame_size * 2}});
+    ActCompute(context.Attr<int>("activation"), place,
+               g.slice(c_offsets, extents), g.slice(c_offsets, extents));
+    auto c = g.slice(c_offsets, extents);  // output candidate
+
+    // calculate final output
+    h.device(place) = u * (h_p - c) + c;
+  }
+};
+
+template <typename Place, typename T>
+class GRUUnitGradKernel : public framework::OpKernel<T> {
+ public:
+  template <typename Device, typename X, typename Y, typename DX, typename DY>
+  void ActGradCompute(const int act_type, const Device& d, X x, Y y, DX dx,
+                      DY dy) const {
+    // x is dummy and won't be used even in Relu(use y instead)
+    if (act_type == identity)
+      dx.device(d) = dy;
+    else if (act_type == sigmoid)
+      SigmoidGradFunctor<T>()(d, x, y, dy, dx);
+    else if (act_type == tanh)
+      TanhGradFunctor<T>()(d, x, y, dy, dx);
+    else if (act_type == relu)
+      ReluGradFunctor<T>()(d, x, y, dy, dx);
+    else
+      PADDLE_THROW("unsupported activation type");
+  }
+
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* input = context.Input<Tensor>("Input");
+    auto* hidden_prev = context.Input<Tensor>("HiddenPrev");
+    auto* weight = context.Input<Tensor>("Weight");
+    auto* gate = context.Input<Tensor>("Gate");
+    auto* reset_hidden_prev = context.Input<Tensor>("ResetHiddenPrev");
+    auto* hidden_grad = context.Input<Tensor>(framework::GradVarName("Hidden"));
+    auto* input_grad = context.Output<Tensor>(framework::GradVarName("Input"));
+    auto* hidden_prev_grad =
+        context.Output<Tensor>(framework::GradVarName("HiddenPrev"));
+    auto* weight_grad =
+        context.Output<Tensor>(framework::GradVarName("Weight"));
+    auto* bias_grad = context.Output<Tensor>(framework::GradVarName("Bias"));
+    input_grad->mutable_data<T>(context.GetPlace());
+    hidden_prev_grad->mutable_data<T>(context.GetPlace());
+    weight_grad->mutable_data<T>(context.GetPlace());
+    Tensor gate_grad;
+    gate_grad.mutable_data<T>(input->dims(), context.GetPlace());
+    Tensor reset_hidden_prev_grad;
+    reset_hidden_prev_grad.mutable_data<T>(reset_hidden_prev->dims(),
+                                           context.GetPlace());
+
+    int batch_size = input->dims()[0];
+    int frame_size = hidden_prev->dims()[1];
+
+    const T* hidden_prev_data = hidden_prev->data<T>();
+    T* hidden_prev_grad_data = hidden_prev_grad->data<T>();
+    const T* weight_data = weight->data<T>();
+    T* weight_grad_data = weight_grad->data<T>();
+    T* gate_grad_data = gate_grad.data<T>();
+    const T* reset_hidden_prev_data = reset_hidden_prev->data<T>();
+    T* reset_hidden_prev_grad_data = reset_hidden_prev_grad.data<T>();
+
+    auto h_p = EigenMatrix<T>::From(*hidden_prev);
+    auto g = EigenMatrix<T>::From(*gate);
+    auto d_h = EigenMatrix<T>::From(*hidden_grad);
+    auto d_x = EigenMatrix<T>::From(*input_grad);
+    auto d_h_p = EigenMatrix<T>::From(*hidden_prev_grad);
+    auto d_g = EigenMatrix<T>::From(gate_grad);
+    auto d_r_h_p = EigenMatrix<T>::From(reset_hidden_prev_grad);
+    auto place = context.GetEigenDevice<Place>();
+
+    Eigen::array<int, 2> extents({{batch_size, frame_size}});
+    Eigen::array<int, 2> u_offsets({{0, 0}});
+    auto u = g.slice(u_offsets, extents);  // update gate
+    Eigen::array<int, 2> r_offsets({{0, frame_size}});
+    auto r = g.slice(r_offsets, extents);  // reset gate
+    Eigen::array<int, 2> c_offsets({{0, frame_size * 2}});
+    auto c = g.slice(c_offsets, extents);  // output candidate
+
+    // backward for unactivated update gate
+    ActGradCompute(context.Attr<int>("gate_activation"), place, u, u,
+                   d_g.slice(u_offsets, extents), d_h * (h_p - c));
+    // backward for unactivated output candidate
+    ActGradCompute(context.Attr<int>("activation"), place, c, c,
+                   d_g.slice(c_offsets, extents), d_h * (u.constant(T(1)) - u));
+    // backward for reset_hidden_prev
+    math::gemm<Place, T>(context.device_context(), false, true, batch_size,
+                         frame_size, frame_size, 1,
+                         gate_grad_data + frame_size * 2, frame_size * 3,
+                         weight_data + frame_size * frame_size * 2, frame_size,
+                         0, reset_hidden_prev_grad_data, frame_size);
+    // backward for state_weight
+    math::gemm<Place, T>(
+        context.device_context(), true, false, frame_size, frame_size,
+        batch_size, 1, reset_hidden_prev_data, frame_size,
+        gate_grad_data + frame_size * 2, frame_size * 3, 0,
+        weight_grad_data + frame_size * frame_size * 2, frame_size);
+    // backward for unactivated reset gate
+    ActGradCompute(context.Attr<int>("gate_activation"), place, r, r,
+                   d_g.slice(r_offsets, extents), d_r_h_p * h_p);
+    // backward for update_gate_weight and reset_gate_weight
+    math::gemm<Place, T>(context.device_context(), true, false, frame_size,
+                         frame_size * 2, batch_size, 1, hidden_prev_data,
+                         frame_size, gate_grad_data, frame_size * 3, 0,
+                         weight_grad_data, frame_size * 2);
+    // backward for hidden_prev
+    d_h_p.device(place) = d_r_h_p * r + d_h * u;
+    math::gemm<Place, T>(context.device_context(), false, true, batch_size,
+                         frame_size, frame_size * 2, 1, gate_grad_data,
+                         frame_size * 3, weight_data, frame_size * 2, 1,
+                         hidden_prev_grad_data, frame_size);
+    // backward for input
+    d_x.device(place) = d_g;
+    // backward for bias
+    if (bias_grad) {
+      bias_grad->mutable_data<T>(context.GetPlace());
+      auto d_b = EigenMatrix<T>::From(*bias_grad);
+      d_b.device(place) = d_g.sum(Eigen::array<int, 1>({{0}}));
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/multiplex_op.cc

@@ -115,8 +115,9 @@ class MultiplexGradOp : public framework::OperatorWithKernel {
 }  // namespace paddle
 namespace ops = paddle::operators;
 
-REGISTER_OP(multiplex, ops::MultiplexOp, ops::MultiplexOpMaker, multiplex_grad,
-            ops::MultiplexGradOp);
+REGISTER_OPERATOR(multiplex, ops::MultiplexOp, ops::MultiplexOpMaker,
+                  paddle::framework::DefaultGradOpDescMaker<false>);
+REGISTER_OPERATOR(multiplex_grad, ops::MultiplexGradOp);
 REGISTER_OP_CPU_KERNEL(
     multiplex, ops::MultiplexCPUKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(

paddle/operators/sum_op.cc

@@ -34,7 +34,7 @@ class SumOp : public framework::OperatorWithKernel {
     auto in_dim = x_dims[0];
     for (size_t i = 1; i < N; i++) {
       auto dim = x_dims[i];
-      PADDLE_ENFORCE(in_dim == dim, "Input tensors must have same shape");
+      PADDLE_ENFORCE_EQ(in_dim, dim, "Input tensors must have same shape");
     }
     ctx->SetOutputDim("Out", in_dim);
     ctx->ShareLoD("X", /*->*/ "Out");

paddle/pybind/protobuf.cc

@@ -123,7 +123,18 @@ void BindProgramDesc(py::module &m) {
              AppendBackward(program_desc, no_grad_vars);
            })
       .def("block", &ProgramDescBind::Block, py::return_value_policy::reference)
-      .def("num_blocks", &ProgramDescBind::Size);
+      .def("num_blocks", &ProgramDescBind::Size)
+      .def("serialize_to_string",
+           [](ProgramDescBind &program_desc) -> py::bytes {
+             const ProgramDesc *desc = program_desc.Proto();
+             PADDLE_ENFORCE(desc->IsInitialized(),
+                            "ProgramDesc has not been initialized.");
+             std::string res;
+             PADDLE_ENFORCE(
+                 desc->SerializeToString(&res),
+                 "Serialize ProgramDesc Error. This could be a bug of Paddle.");
+             return res;
+           });
 }
 
 void BindBlockDesc(py::module &m) {
@@ -149,7 +160,17 @@ void BindBlockDesc(py::module &m) {
       .def("all_vars", &BlockDescBind::AllVars,
            py::return_value_policy::reference)
       .def("all_ops", &BlockDescBind::AllOps,
-           py::return_value_policy::reference);
+           py::return_value_policy::reference)
+      .def("serialize_to_string", [](BlockDescBind &block_desc) -> py::bytes {
+        const BlockDesc *desc = block_desc.Proto();
+        PADDLE_ENFORCE(desc->IsInitialized(),
+                       "BlockDesc has not been initialized.");
+        std::string res;
+        PADDLE_ENFORCE(
+            desc->SerializeToString(&res),
+            "Serialize BlockDesc Error. This could be a bug of Paddle.");
+        return res;
+      });
 }
 
 void BindVarDsec(py::module &m) {
@@ -162,7 +183,8 @@ void BindVarDsec(py::module &m) {
       .value("FP32", DataType::FP32)
       .value("FP64", DataType::FP64);
 
-  py::class_<VarDescBind>(m, "VarDesc", "")
+  py::class_<VarDescBind> var_desc(m, "VarDesc", "");
+  var_desc
       .def("name",
            [](const VarDescBind &self) {
              py::bytes name = self.Name();
@@ -174,7 +196,23 @@ void BindVarDsec(py::module &m) {
       .def("shape", &VarDescBind::Shape, py::return_value_policy::reference)
       .def("data_type", &VarDescBind::GetDataType)
       .def("lod_level", &VarDescBind::GetLodLevel)
-      .def("set_lod_level", &VarDescBind::SetLoDLevel);
+      .def("set_lod_level", &VarDescBind::SetLoDLevel)
+      .def("type", &VarDescBind::GetType)
+      .def("set_type", &VarDescBind::SetType)
+      .def("serialize_to_string", [](VarDescBind &var_desc) -> py::bytes {
+        const VarDesc *desc = var_desc.Proto();
+        PADDLE_ENFORCE(desc->IsInitialized(),
+                       "VarDesc has not been initialized.");
+        std::string res;
+        PADDLE_ENFORCE(
+            desc->SerializeToString(&res),
+            "Serialize VarDesc Error. This could be a bug of Paddle.");
+        return res;
+      });
+
+  py::enum_<VarDesc::VarType>(var_desc, "VarType", "")
+      .value("LOD_TENSOR", VarDesc::LOD_TENSOR)
+      .value("SELECTED_ROWS", VarDesc::SELECTED_ROWS);
 }
 
 void BindOpDesc(py::module &m) {
@@ -206,7 +244,17 @@ void BindOpDesc(py::module &m) {
       .def("set_block_attr", &OpDescBind::SetBlockAttr)
       .def("block_attr", &OpDescBind::GetBlockAttr)
       .def("check_attrs", &OpDescBind::CheckAttrs)
-      .def("infer_shape", &OpDescBind::InferShape);
+      .def("infer_shape", &OpDescBind::InferShape)
+      .def("serialize_to_string", [](OpDescBind &op_desc) -> py::bytes {
+        const OpDesc *desc = op_desc.Proto();
+        PADDLE_ENFORCE(desc->IsInitialized(),
+                       "OpDesc has not been initialized.");
+        std::string res;
+        PADDLE_ENFORCE(
+            desc->SerializeToString(&res),
+            "Serialize OpDesc Error. This could be a bug of Paddle.");
+        return res;
+      });
 }
 
 }  // namespace pybind

paddle/pybind/pybind.cc

@@ -18,6 +18,7 @@ limitations under the License. */
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/tensor_array.h"
 #include "paddle/operators/cond_op.h"
+#include "paddle/operators/dynamic_recurrent_op.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
 #include "paddle/platform/enforce.h"
@@ -341,6 +342,33 @@ All parameter, weight, gradient are variables in Paddle.
         self.set_stepnet(net.Clone());
       });
 
+  py::class_<operators::DynamicRecurrentOp, OperatorBase>(m,
+                                                          "DynamicRecurrentOp")
+      .def_static("create",
+                  [](py::bytes protobin) -> operators::DynamicRecurrentOp * {
+                    OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    auto rnn_op = OpRegistry::CreateOp(desc);
+                    return static_cast<operators::DynamicRecurrentOp *>(
+                        rnn_op.release());
+                  })
+      .def("set_stepnet",
+           [](operators::DynamicRecurrentOp &self, const operators::NetOp &net)
+               -> void { self.SetStepNet(net.Clone()); })
+      .def("get_state",
+           [](operators::DynamicRecurrentOp &self, const std::string &name)
+               -> const TensorArray & { return self.state(name); })
+      .def("get_step_input",
+           [](operators::DynamicRecurrentOp &self, const std::string &name)
+               -> const TensorArray & { return self.step_input(name); })
+      .def("get_step_output",
+           [](operators::DynamicRecurrentOp &self, const std::string &name)
+               -> const TensorArray & { return self.step_output(name); });
+
   // cond_op
   py::class_<operators::CondOp, OperatorBase>(m, "CondOp")
       .def_static("create",

python/paddle/v2/framework/framework.py

@@ -10,6 +10,7 @@ __all__ = ['Block', 'Variable', 'Program', 'Operator']
 class Variable(object):
     def __init__(self,
                  block,
+                 type=core.VarDesc.VarType.LOD_TENSOR,
                  name=None,
                  shape=None,
                  dtype=None,
@@ -26,6 +27,14 @@ class Variable(object):
             self.desc = self.block.desc.var(name)
             is_new_var = True
 
+        if is_new_var:
+            self.desc.set_type(type)
+        elif self.desc.type() != type:
+            raise ValueError("Variable {0} has been created before. The "
+                             "previous type is {1}; the new type is {2}. They"
+                             " are not matched".format(self.name,
+                                                       self.desc.type(), type))
+
         if shape is not None:
             if is_new_var:
                 self.desc.set_shape(shape)
@@ -64,6 +73,13 @@ class Variable(object):
         self.block.vars[name] = self
         self.op = None
 
+    def __str__(self):
+        protostr = self.desc.serialize_to_string()
+        proto = framework_pb2.VarDesc.FromString(str(protostr))
+        return proto.__str__()
+
+    __repr__ = __str__
+
     @property
     def name(self):
         return self.desc.name()
@@ -201,6 +217,13 @@ class Operator(object):
         self.desc.check_attrs()
         self.desc.infer_shape(self.block.desc)
 
+    def __str__(self):
+        protostr = self.desc.serialize_to_string()
+        proto = framework_pb2.OpDesc.FromString(str(protostr))
+        return proto.__str__()
+
+    __repr__ = __str__
+
     @property
     def type(self):
         return self.desc.type()
@@ -243,6 +266,13 @@ class Block(object):
         self.ops = collections.deque()  # operator list
         self.program = program
 
+    def __str__(self):
+        protostr = self.desc.serialize_to_string()
+        proto = framework_pb2.BlockDesc.FromString(str(protostr))
+        return proto.__str__()
+
+    __repr__ = __str__
+
     @property
     def parent_idx(self):
         return self.desc.parent
@@ -287,6 +317,13 @@ class Program(object):
         self.blocks = [Block(self, 0)]
         self.current_block_idx = 0
 
+    def __str__(self):
+        protostr = self.desc.serialize_to_string()
+        proto = framework_pb2.ProgramDesc.FromString(str(protostr))
+        return proto.__str__()
+
+    __repr__ = __str__
+
     def global_block(self):
         return self.blocks[0]
 

python/paddle/v2/framework/op.py

@@ -219,6 +219,27 @@ class __RecurrentOp__(object):
         return core.RecurrentOp.create(proto.SerializeToString())
 
 
+class __DynamicRecurrentOp__(object):
+    __proto__ = None
+    type = "dynamic_recurrent"
+
+    def __init__(self):
+        # cache recurrent_op's proto
+        if self.__proto__ is None:
+            for op_proto in get_all_op_protos():
+                if op_proto.type == self.type:
+                    self.__proto__ = op_proto
+
+    def __call__(self, *args, **kwargs):
+        if self.type not in args and "type" not in kwargs:
+            kwargs["type"] = self.type
+        # create proto
+        create_method = OpDescCreationMethod(self.__proto__)
+        proto = create_method(*args, **kwargs)
+        # create rnnop
+        return core.DynamicRecurrentOp.create(proto.SerializeToString())
+
+
 class __CondOp__(object):
     __proto__ = None
     type = "cond"
@@ -242,4 +263,5 @@ class __CondOp__(object):
 
 Operator = OperatorFactory()  # The default global factory
 RecurrentOp = __RecurrentOp__()
+DynamicRecurrentOp = __DynamicRecurrentOp__()
 CondOp = __CondOp__()

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

@@ -384,5 +384,33 @@ class TestThresholdedRelu(OpTest):
         self.check_grad(['X'], 'Y', max_relative_error=self.relative_error)
 
 
+class TestHardSigmoid(OpTest):
+    def setUp(self):
+        self.op_type = "hard_sigmoid"
+        self.relative_error = 0.002
+
+        X = np.random.uniform(-5, 5, [2, 2]).astype("float32")
+        slope = 0.2
+        offset = 0.5
+        lower_threshold = -offset / slope
+        upper_threshold = (1 - offset) / slope
+
+        self.inputs = {'X': X}
+        # Same reason as TestAbs
+        X[np.abs(X - lower_threshold) < self.relative_error] = \
+            lower_threshold + 0.2
+        X[np.abs(X - upper_threshold) < self.relative_error] = \
+            upper_threshold - 0.2
+
+        temp = X * slope + offset
+        self.outputs = {'Y': np.maximum(0.0, np.minimum(1.0, temp))}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y', max_relative_error=0.002)
+
+
 if __name__ == "__main__":
     unittest.main()

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

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestAdamOp1(OpTest):
+    def setUp(self):
+        '''Test Adam Op with supplied attributes
+        '''
+        self.op_type = "adam"
+        param = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        grad = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        moment1 = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        # The second moment is positive
+        moment2 = np.random.random((102, 105)).astype("float32")
+
+        learning_rate = 0.004
+        beta1 = 0.78
+        beta2 = 0.836
+        epsilon = 1e-4
+        beta1_pow = beta1**10
+        beta2_pow = beta2**10
+
+        self.inputs = {
+            'Param': param,
+            'Grad': grad,
+            'Moment1': moment1,
+            'Moment2': moment2,
+            'LearningRate': np.array([learning_rate]).astype("float32"),
+            'Beta1Pow': np.array([beta1_pow]).astype("float32"),
+            'Beta2Pow': np.array([beta2_pow]).astype("float32")
+        }
+
+        self.attrs = {'epsilon': epsilon, 'beta1': beta1, 'beta2': beta2}
+
+        param_out, moment1_out, moment2_out, beta1_pow_out, \
+            beta2_pow_out = adam_step(self.inputs, self.attrs)
+
+        self.outputs = {
+            'Moment1Out': moment1_out,
+            'Moment2Out': moment2_out,
+            'Beta1PowOut': beta1_pow_out,
+            'Beta2PowOut': beta2_pow_out,
+            'ParamOut': param_out
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestAdamOp2(OpTest):
+    def setUp(self):
+        '''Test Adam Op with supplied attributes
+        '''
+        self.op_type = "adam"
+        param = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        grad = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        moment1 = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        # The second moment is positive
+        moment2 = np.random.random((102, 105)).astype("float32")
+
+        learning_rate = 0.001
+        beta1 = 0.9
+        beta2 = 0.999
+        epsilon = 1e-8
+        beta1_pow = beta1**10
+        beta2_pow = beta2**10
+
+        self.inputs = {
+            'Param': param,
+            'Grad': grad,
+            'Moment1': moment1,
+            'Moment2': moment2,
+            'LearningRate': np.array([learning_rate]).astype("float32"),
+            'Beta1Pow': np.array([beta1_pow]).astype("float32"),
+            'Beta2Pow': np.array([beta2_pow]).astype("float32")
+        }
+
+        attributes = {'epsilon': epsilon, 'beta1': beta1, 'beta2': beta2}
+
+        param_out, moment1_out, moment2_out, beta1_pow_out, \
+            beta2_pow_out = adam_step(self.inputs, attributes)
+
+        self.outputs = {
+            'Moment1Out': moment1_out,
+            'Moment2Out': moment2_out,
+            'Beta1PowOut': beta1_pow_out,
+            'Beta2PowOut': beta2_pow_out,
+            'ParamOut': param_out
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestAdamOpMultipleSteps(OpTest):
+    def setUp(self):
+        '''Test Adam Operator with supplied attributes
+        '''
+        self.op_type = "adam"
+        self.num_steps = 10
+
+        param = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        grad = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        moment1 = np.random.uniform(-1, 1, (102, 105)).astype("float32")
+        # The second moment is positive
+        moment2 = np.random.random((102, 105)).astype("float32")
+
+        learning_rate = 0.001
+        beta1 = 0.9
+        beta2 = 0.999
+        epsilon = 1e-8
+        beta1_pow = beta1**10
+        beta2_pow = beta2**10
+
+        self.inputs = {
+            'Param': param,
+            'Grad': grad,
+            'Moment1': moment1,
+            'Moment2': moment2,
+            'LearningRate': np.array([learning_rate]).astype("float32"),
+            'Beta1Pow': np.array([beta1_pow]).astype("float32"),
+            'Beta2Pow': np.array([beta2_pow]).astype("float32")
+        }
+
+        self.attrs = {'epsilon': epsilon, 'beta1': beta1, 'beta2': beta2}
+
+    def test_check_output(self):
+        for _ in range(self.num_steps):
+            param_out, moment1_out, moment2_out, beta1_pow_out, \
+                beta2_pow_out = adam_step(self.inputs, self.attrs)
+
+            self.outputs = {
+                'Moment1Out': moment1_out,
+                'Moment2Out': moment2_out,
+                'Beta1PowOut': beta1_pow_out,
+                'Beta2PowOut': beta2_pow_out,
+                'ParamOut': param_out
+            }
+
+            # Verify output for this step
+            self.check_output()
+
+            # Output of this step becomes input for next step
+            self.inputs['Param'] = param_out
+            self.inputs['Moment1'] = moment1_out
+            self.inputs['Moment2'] = moment2_out
+            self.inputs['Beta1Pow'] = beta1_pow_out
+            self.inputs['Beta2Pow'] = beta2_pow_out
+
+            # Randomize gradient for next step
+            self.inputs['Grad'] = np.random.uniform(
+                -1, 1, (102, 105)).astype("float32")
+
+
+def adam_step(inputs, attributes):
+    '''
+    Simulate one step of the adam optimizer
+    :param inputs: dict of inputs
+    :param attributes: dict of attributes
+    :return tuple: tuple of output param, moment1, moment2,
+    beta1 power accumulator and beta2 power accumulator
+    '''
+    param = inputs['Param']
+    grad = inputs['Grad']
+    moment1 = inputs['Moment1']
+    moment2 = inputs['Moment2']
+    lr = inputs['LearningRate']
+    beta1_pow = inputs['Beta1Pow']
+    beta2_pow = inputs['Beta2Pow']
+
+    beta1 = attributes['beta1']
+    beta2 = attributes['beta2']
+    epsilon = attributes['epsilon']
+
+    moment1_out = beta1 * moment1 + (1 - beta1) * grad
+    moment2_out = beta2 * moment2 + (1 - beta2) * np.square(grad)
+    beta1_pow_out = beta1_pow * beta1
+    beta2_pow_out = beta2_pow * beta2
+    lr_t = lr * np.sqrt(1 - beta2_pow_out) / (1 - beta1_pow_out)
+    param_out = param - lr_t * (moment1_out / (np.sqrt(moment2_out) + epsilon))
+    return param_out, moment1_out, moment2_out, beta1_pow_out, beta2_pow_out
+
+
+if __name__ == "__main__":
+    unittest.main()

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

+import logging
+import paddle.v2.framework.core as core
+import unittest
+from paddle.v2.framework.op import Operator, DynamicRecurrentOp
+import numpy as np
+
+
+def create_tensor(scope, name, shape, np_data):
+    tensor = scope.new_var(name).get_tensor()
+    tensor.set_dims(shape)
+    tensor.set(np_data, core.CPUPlace())
+    return tensor
+
+
+class DynamicRecurrentOpTest(unittest.TestCase):
+    '''
+    Test RNNOp
+
+    equation:
+        h_t = \sigma (W x_t + U h_{t-1})
+    weights:
+        - W
+        - U
+    vars:
+        - x
+    memories:
+        - h
+    outputs:
+       - h
+    '''
+
+    # for siplicity, just one level LoD
+    lod_py = [[0, 4, 7, 9, 10]]
+    input_dim = 30
+    num_sents = len(lod_py[0]) - 1
+    weight_dim = 15
+
+    def forward(self):
+        self.scope = core.Scope()
+        self.create_global_variables()
+        self.create_rnn_op()
+        self.create_step_net()
+        ctx = core.DeviceContext.create(core.CPUPlace())
+        self.rnnop.run(self.scope, ctx)
+        state = self.rnnop.get_state("h@mem")
+        print 'state size: ', state.size()
+
+        step_inputs = self.rnnop.get_step_input("x")
+        print "x size ", step_inputs.size()
+        for i in range(step_inputs.size()):
+            print "x %d" % i, np.array(step_inputs.read(i).get_dims())
+        step_outputs = self.rnnop.get_step_output('h@mem')
+        print 'step_outputs.size ', step_outputs.size()
+        output = self.scope.find_var("h@mem").get_tensor()
+
+        print 'output', np.array(output).shape
+
+    def create_global_variables(self):
+        x = np.random.normal(size=(self.lod_py[0][-1],
+                                   self.input_dim)).astype("float32")
+        W = np.random.normal(size=(self.input_dim,
+                                   self.input_dim)).astype("float32")
+        U = np.random.normal(size=(self.input_dim,
+                                   self.input_dim)).astype("float32")
+        h_boot = np.random.normal(size=(self.num_sents,
+                                        self.input_dim)).astype("float32")
+        # create inlink
+        x_tensor = create_tensor(self.scope, "x",
+                                 [self.num_sents, self.input_dim], x)
+        x_tensor.set_lod(self.lod_py)
+        create_tensor(self.scope, "W", [self.input_dim, self.input_dim], W)
+        create_tensor(self.scope, "U", [self.input_dim, self.input_dim], U)
+        create_tensor(self.scope, "h_boot", [self.num_sents, self.input_dim],
+                      h_boot)
+        self.scope.new_var("step_scopes")
+        self.scope.new_var("h@mem")
+
+    def create_rnn_op(self):
+        # create RNNOp
+        self.rnnop = DynamicRecurrentOp(
+            # inputs
+            inlinks=["x"],
+            boot_memories=["h_boot"],
+            step_net="stepnet",
+            # outputs
+            outlinks=["h@mem"],
+            step_scopes="step_scopes",
+            # attributes
+            pre_memories=["h@pre"],
+            memories=["h@mem"])
+
+    def create_step_net(self):
+        stepnet = core.Net.create()
+        x_fc_op = Operator("mul", X="x", Y="W", Out="Wx")
+        h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh")
+        sum_op = Operator("sum", X=["Wx", "Uh"], Out="sum")
+        sig_op = Operator("sigmoid", X="sum", Y="h@mem")
+
+        for op in [x_fc_op, h_fc_op, sum_op, sig_op]:
+            stepnet.append_op(op)
+        stepnet.complete_add_op(True)
+        self.rnnop.set_stepnet(stepnet)
+
+    def test_forward(self):
+        print 'test recurrent op forward'
+        pd_output = self.forward()
+        print 'pd_output', pd_output
+
+
+if __name__ == '__main__':
+    unittest.main()

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

+import math
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class GRUActivationType(OpTest):
+    identity = 0
+    sigmoid = 1
+    tanh = 2
+    relu = 3
+
+
+def identity(x):
+    return x
+
+
+def sigmoid(x):
+    return 1. / (1. + np.exp(-x))
+
+
+def tanh(x):
+    return 2. * sigmoid(2. * x) - 1.
+
+
+def relu(x):
+    return np.maximum(x, 0)
+
+
+class TestGRUUnitOp(OpTest):
+    batch_size = 3
+    frame_size = 5
+    activate = {
+        GRUActivationType.identity: identity,
+        GRUActivationType.sigmoid: sigmoid,
+        GRUActivationType.tanh: tanh,
+        GRUActivationType.relu: relu,
+    }
+
+    def set_inputs(self):
+        batch_size = self.batch_size
+        frame_size = self.frame_size
+        self.op_type = 'gru_unit'
+        self.inputs = {
+            'Input': np.random.uniform(
+                -0.1, 0.1, (batch_size, frame_size * 3)).astype('float32'),
+            'HiddenPrev': np.random.uniform(
+                -0.1, 0.1, (batch_size, frame_size)).astype('float32'),
+            'Weight': np.random.uniform(
+                -1. / math.sqrt(frame_size), 1. / math.sqrt(frame_size),
+                (frame_size, frame_size * 3)).astype('float32'),
+        }
+        self.attrs = {
+            'activation': GRUActivationType.tanh,
+            'gate_activation': GRUActivationType.sigmoid
+        }
+
+    def set_outputs(self):
+        # GRU calculations
+        batch_size = self.batch_size
+        frame_size = self.frame_size
+        x = self.inputs['Input']
+        h_p = self.inputs['HiddenPrev']
+        w = self.inputs['Weight']
+        b = self.inputs['Bias'] if self.inputs.has_key('Bias') else np.zeros(
+            (1, frame_size * 3))
+        g = x + np.tile(b, (batch_size, 1))
+        w_u_r = w.flatten()[:frame_size * frame_size * 2].reshape(
+            (frame_size, frame_size * 2))
+        u_r = self.activate[self.attrs['gate_activation']](np.dot(
+            h_p, w_u_r) + g[:, :frame_size * 2])
+        u = u_r[:, :frame_size]
+        r = u_r[:, frame_size:frame_size * 2]
+        r_h_p = r * h_p
+        w_c = w.flatten()[frame_size * frame_size * 2:].reshape(
+            (frame_size, frame_size))
+        c = self.activate[self.attrs['activation']](np.dot(r_h_p, w_c) +
+                                                    g[:, frame_size * 2:])
+        g = np.hstack((u_r, c))
+        h = u * h_p + (1 - u) * c
+        self.outputs = {'Gate': g, 'ResetHiddenPrev': r_h_p, 'Hidden': h}
+
+    def setUp(self):
+        self.set_inputs()
+        self.set_outputs()
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(
+            ['Input', 'HiddenPrev', 'Weight'], ['Hidden'],
+            max_relative_error=0.007)
+
+
+class TestGRUUnitOpWithBias(TestGRUUnitOp):
+    def set_inputs(self):
+        batch_size = self.batch_size
+        frame_size = self.frame_size
+        super(TestGRUUnitOpWithBias, self).set_inputs()
+        self.inputs['Bias'] = np.random.uniform(
+            -0.1, 0.1, (1, frame_size * 3)).astype('float32')
+        self.attrs = {
+            'activation': GRUActivationType.identity,
+            'gate_activation': GRUActivationType.sigmoid
+        }
+
+    def test_check_grad(self):
+        self.check_grad(
+            ['Input', 'HiddenPrev', 'Weight', 'Bias'], ['Hidden'],
+            max_relative_error=0.007)
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -14,11 +14,14 @@ class TestInferShape(unittest.TestCase):
 
         # prepare input/output
         x1 = block.var("x1")
+        x1.set_type(core.VarDesc.VarType.LOD_TENSOR)
         x1.set_shape(shape)
         x2 = block.var("x2")
+        x2.set_type(core.VarDesc.VarType.LOD_TENSOR)
         x2.set_shape(shape)
 
         out = block.var("out")
+        out.set_type(core.VarDesc.VarType.LOD_TENSOR)
 
         # prepare the operator
         sum_op_desc = block.append_op()
@@ -40,11 +43,14 @@ class TestInferShape(unittest.TestCase):
 
         # prepare input/output
         x1 = block.var("x")
+        x1.set_type(core.VarDesc.VarType.LOD_TENSOR)
         x1.set_shape(x_shape)
         x2 = block.var("y")
+        x2.set_type(core.VarDesc.VarType.LOD_TENSOR)
         x2.set_shape(y_shape)
 
         out = block.var("out")
+        out.set_type(core.VarDesc.VarType.LOD_TENSOR)
 
         # prepare the operator
         mul_op_desc = block.append_op()

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

@@ -34,6 +34,8 @@ class TestOperator(unittest.TestCase):
                     "Y": mul_y},
             outputs={"Out": [mul_out]},
             attrs={"x_num_col_dims": 1})
+
+        self.assertNotEqual(str(mul_op), "")
         self.assertEqual(mul_op.type, "mul")
         self.assertEqual(mul_op.input_names, ["X", "Y"])
         self.assertEqual(mul_op.input("X"), ["mul.x"])

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

@@ -94,17 +94,21 @@ class TestVarDesc(unittest.TestCase):
         program_desc = core.ProgramDesc.__create_program_desc__()
         block = program_desc.block(0)
         var = block.var('my_var')
+        var.set_type(core.VarDesc.VarType.SELECTED_ROWS)
         src_shape = [3, 2, 10, 8]
         var.set_shape(src_shape)
         res_shape = var.shape()
         self.assertEqual(src_shape, res_shape)
+        self.assertEqual(core.VarDesc.VarType.SELECTED_ROWS, var.type())
 
     def test_data_type(self):
         program_desc = core.ProgramDesc.__create_program_desc__()
         block = program_desc.block(0)
         var = block.var('my_var')
+        var.set_type(core.VarDesc.VarType.LOD_TENSOR)
         var.set_data_type(core.DataType.INT32)
         self.assertEqual(core.DataType.INT32, var.data_type())
+        self.assertEqual(core.VarDesc.VarType.LOD_TENSOR, var.type())
 
 
 class TestBlockDesc(unittest.TestCase):

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

@@ -21,6 +21,7 @@ class TestVariable(unittest.TestCase):
         b = g_program.current_block()
         w = b.create_var(
             dtype="float64", shape=[784, 100], lod_level=0, name="fc.w")
+        self.assertNotEqual(str(w), "")
         self.assertEqual(core.DataType.FP64, w.data_type)
         self.assertEqual((784, 100), w.shape)
         self.assertEqual("fc.w", w.name)