Feature/dynamic recurrent op forward test (#4729)

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/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.NewVar(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.NewVar("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/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/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/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_dynamic_recurrent_op.py

@@ -0,0 +1,111 @@
+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()