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

paddle/framework/backward.cc

@@ -270,6 +270,19 @@ static bool AllGradInSet(const std::vector<std::string>& names,
       return false;
     }
   }
+  if (VLOG_IS_ON(10)) {
+    std::ostringstream sout;
+    sout << "All input {";
+    for (auto& name : names) {
+      sout << name << ",";
+    }
+    sout << "} is in {";
+    for (auto& name : set) {
+      sout << name << ",";
+    }
+    sout << "}";
+    VLOG(10) << sout.str();
+  }
   return true;
 }
 
@@ -290,14 +303,12 @@ static void CreateGradVarInBlock(
   auto ops = block_desc->AllOps();
   for (size_t op_index = grad_op_start_index; op_index < ops.size();
        ++op_index) {
-    bool need_infer_shape = false;
     std::unordered_set<std::string> new_vars;
     ForEachVarName(ops[op_index]->Outputs(),
                    [&](const std::string& grad_var_name) {
                      if (block_desc->HasVar(grad_var_name)) {
                        return false;
                      }
-                     need_infer_shape = true;
                      auto var = block_desc->Var(grad_var_name);
                      new_vars.insert(var->Name());
                      auto it = param_name_map.find(grad_var_name);
@@ -311,7 +322,6 @@ static void CreateGradVarInBlock(
                      grad_record.op_idx_ = static_cast<int>(op_index);
                      return false; /* not break */
                    });
-    if (need_infer_shape) {
     ops[op_index]->InferVarType(block_desc);
     for (auto& arg : ops[op_index]->OutputArgumentNames()) {
       if (new_vars.find(arg) == new_vars.end()) {
@@ -328,7 +338,6 @@ static void CreateGradVarInBlock(
     }
     ops[op_index]->InferShape(*block_desc);
   }
-  }
 }
 
 std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
@@ -387,6 +396,7 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
     ProgramDescBind& program_desc, int block_idx,
     std::unordered_set<std::string>* no_grad_vars,
     std::unordered_map<std::string, std::string>* grad_to_var) {
+  VLOG(5) << "MakeBlockBackward";
   BlockDescBind* cur_block = program_desc.MutableBlock(block_idx);
   std::vector<OpDescBind*> op_descs = cur_block->AllOps();
   std::unordered_map<std::string, std::vector<size_t>> dup_out_ops;
@@ -394,9 +404,10 @@ 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) {
+    VLOG(5) << "Making backward " << (*it)->Type() << " op";
     std::vector<std::unique_ptr<OpDescBind>> op_grads;
 
-    if ((*it)->Type() == "recurrent") {
+    if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") {
       int step_block_idx = (*it)->GetBlockAttr("step_block");
       BlockDescBind* backward_block = CreateStepBlock(
           program_desc, no_grad_vars, grad_to_var, step_block_idx);
@@ -410,6 +421,15 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var);
     }
 
+    if (VLOG_IS_ON(10)) {
+      std::ostringstream sout;
+      sout << "Made ";
+      for (auto& op_grad : op_grads) {
+        sout << op_grad->Type() << " ";
+      }
+      VLOG(10) << sout.str();
+    }
+
     for (const auto& desc : op_grads) {
       for (const std::string& out_name : desc->OutputArgumentNames()) {
         if (out_name.find("@GRAD") == std::string::npos) {
@@ -425,6 +445,8 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
         op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs),
         [](std::unique_ptr<OpDescBind>& ptr) { return std::move(ptr); });
   }
+
+  VLOG(5) << "Appending Sums";
   // Check whether some variables are written more than once
   std::list<std::pair<size_t, std::unique_ptr<OpDescBind>>> pending_sum_ops;
   for (const auto& dup : dup_out_ops) {
@@ -432,16 +454,22 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
     const std::vector<size_t> dup_op = dup.second;
     if (out_name != kEmptyVarName && dup_op.size() > 1) {
       std::vector<std::string> sum_op_inputs;
+      std::string next_g_name = out_name;
       for (size_t i = 0; i < dup_op.size(); ++i) {
+        VLOG(10) << backward_descs[dup_op[i]]->Type() << " has " << out_name
+                 << " duplicated";
         std::string new_name = out_name + "@RENAME@" + std::to_string(i);
-        backward_descs[dup_op[i]]->Rename(out_name, new_name);
+        backward_descs[dup_op[i]]->RenameOutput(out_name, new_name);
+        backward_descs[dup_op[i]]->RenameInput(out_name, next_g_name);
         sum_op_inputs.emplace_back(new_name);
+        next_g_name = sum_op_inputs.back();
       }
       std::unique_ptr<OpDescBind> sum_op(new OpDescBind(
           "sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}}, {}));
       pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)});
     }
   }
+
   pending_sum_ops.sort(
       [](const std::pair<size_t, std::unique_ptr<OpDescBind>>& a,
          const std::pair<size_t, std::unique_ptr<OpDescBind>>& b) {
@@ -452,6 +480,8 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
                           std::move(p.second));
   }
 
+  VLOG(5) << "MakeBlockBackward Finished";
+
   return backward_descs;
 }
 

paddle/framework/data_type.h

@@ -29,6 +29,8 @@ inline DataType ToDataType(std::type_index type) {
     return DataType::INT32;
   } else if (typeid(int64_t).hash_code() == type.hash_code()) {
     return DataType::INT64;
+  } else if (typeid(bool).hash_code() == type.hash_code()) {
+    return DataType::BOOL;
   } else {
     PADDLE_THROW("Not supported");
   }

paddle/framework/ddim.cc

@@ -60,8 +60,7 @@ void make_ddim(DDim& ddim, const int64_t* dims, int n) {
       ddim = make_dim<9>(dims);
       break;
     default:
-      throw std::invalid_argument(
-          "Dynamic dimensions must have between [1, 9] dimensions.");
+      PADDLE_THROW("Dynamic dimensions must have between [1, 9] dimensions.");
   }
 }
 

paddle/framework/executor.cc

@@ -120,6 +120,7 @@ void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
+    VLOG(10) << op->DebugString();
     op->Run(*local_scope, *device);
   }
   if (create_local_scope) {

paddle/framework/op_desc.cc

@@ -235,6 +235,23 @@ void OpDescBind::Rename(const std::string &old_name,
   need_update_ = true;
 }
 
+void OpDescBind::RenameOutput(const std::string &old_name,
+                              const std::string &new_name) {
+  for (auto &output : outputs_) {
+    std::replace(output.second.begin(), output.second.end(), old_name,
+                 new_name);
+  }
+  need_update_ = true;
+}
+
+void OpDescBind::RenameInput(const std::string &old_name,
+                             const std::string &new_name) {
+  for (auto &input : inputs_) {
+    std::replace(input.second.begin(), input.second.end(), old_name, new_name);
+  }
+  need_update_ = true;
+}
+
 struct SetAttrDescVisitor : public boost::static_visitor<void> {
   explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {}
   mutable OpDesc::Attr *attr_;
@@ -448,7 +465,12 @@ const std::vector<std::string> &CompileTimeInferShapeContext::Outputs(
 DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const {
   auto var = block_.FindVarRecursive(name);
   PADDLE_ENFORCE(var != nullptr, "Cannot find variable %s", name);
+  try {
     return framework::make_ddim(var->Shape());
+  } catch (...) {
+    VLOG(5) << "GetDim of variable " << name << " error";
+    std::rethrow_exception(std::current_exception());
+  }
 }
 
 void CompileTimeInferShapeContext::SetDim(const std::string &name,

paddle/framework/op_desc.h

@@ -73,6 +73,10 @@ class OpDescBind {
 
   void Rename(const std::string &old_name, const std::string &new_name);
 
+  void RenameOutput(const std::string &old_name, const std::string &new_name);
+
+  void RenameInput(const std::string &old_name, const std::string &new_name);
+
   // Only be used in C++
   const AttributeMap &GetAttrMap() const;
 

paddle/framework/operator.cc

@@ -403,19 +403,6 @@ class RuntimeInferShapeContext : public InferShapeContext {
 
 void OperatorWithKernel::Run(const Scope& scope,
                              const platform::DeviceContext& dev_ctx) const {
-  if (VLOG_IS_ON(1)) {
-    auto inputs = this->InputVars();
-    auto outputs = this->OutputVars(true);
-    std::ostringstream sout;
-    sout << "Run operator " << this->Type() << " From [";
-    std::ostream_iterator<std::string> out_it(sout, ",");
-    std::copy(inputs.begin(), inputs.end(), out_it);
-    sout << "] to [";
-    std::copy(outputs.begin(), outputs.end(), out_it);
-    sout << "]";
-    VLOG(1) << sout.str();
-  }
-
   RuntimeInferShapeContext infer_shape_ctx(*this, scope);
   this->InferShape(&infer_shape_ctx);
 

paddle/framework/scope.cc

@@ -38,11 +38,12 @@ Scope& Scope::NewScope() const {
 Variable* Scope::Var(const std::string& name) {
   auto iter = vars_.find(name);
   if (iter != vars_.end()) {
+    VLOG(3) << "Get existing variable " << name;
     return iter->second;
   }
   Variable* v = new Variable();
   vars_[name] = v;
-  VLOG(3) << "Create variable " << name << " on scope";
+  VLOG(3) << "Create variable " << name;
   v->name_ = &(vars_.find(name)->first);
   return v;
 }

paddle/framework/shape_inference.h

@@ -53,6 +53,10 @@ class InferShapeContext {
 
   virtual bool IsRuntime() const = 0;
 
+  // Note: In while op, we need this to be public
+  void SetDims(const std::vector<std::string> &names,
+               const std::vector<framework::DDim> &dims);
+
  protected:
   virtual framework::DDim GetDim(const std::string &name) const = 0;
   virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0;
@@ -60,9 +64,6 @@ class InferShapeContext {
   std::vector<framework::DDim> GetDims(
       const std::vector<std::string> &names) const;
 
-  void SetDims(const std::vector<std::string> &names,
-               const std::vector<framework::DDim> &dims);
-
   std::vector<VarDesc::VarType> GetVarTypes(
       const std::vector<std::string> &names) const;
 

paddle/operators/array_operator.h

@@ -42,6 +42,7 @@ class ArrayOp : public framework::OperatorBase {
     } else {
       offset = static_cast<size_t>(*i_tensor.data<int64_t>());
     }
+    VLOG(10) << " Offset = " << offset;
     return offset;
   }
 };

paddle/operators/bilinear_tensor_product_op.h

@@ -174,7 +174,7 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
     // Caculate the gradient of Input(Bias).
     if (d_bias) {
       d_bias->mutable_data<T>(ctx.GetPlace());
-      auto d_bias_mat = EigenMatrix<T>::From(*d_bias);
+      auto d_bias_mat = framework::EigenVector<T>::Flatten(*d_bias);
       d_bias_mat.device(place) = d_out_mat.sum(Eigen::DSizes<int, 1>(0));
     }
   }

paddle/operators/cos_sim_op.h

@@ -132,7 +132,7 @@ class CosSimGradKernel : public framework::OpKernel<T> {
       // compute dy
       if (out_grad_y) {
         out_grad_y->mutable_data<T>(context.GetPlace());
-        auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
+        auto dy = EigenVector<T>::Flatten(*out_grad_y);
         auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast;
         dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({{0}}));
       }

paddle/operators/detail/safe_ref.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   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
+
+namespace paddle {
+namespace operators {
+namespace detail {
+/**
+ * Get Reference From Pointer with check. The error message is printf format,
+ * and passed by `args`
+ */
+template <typename T, typename... ARGS>
+inline T &Ref(T *ptr, ARGS &&... args) {
+  PADDLE_ENFORCE(ptr != nullptr, args...);
+  return *ptr;
+}
+}  // namespace detail
+}  // namespace operators
+}  // namespace paddle

paddle/operators/fill_constant_batch_size_like_op.cc

@@ -101,4 +101,7 @@ REGISTER_OPERATOR(fill_constant_batch_size_like,
 REGISTER_OP_CPU_KERNEL(
     fill_constant_batch_size_like,
     ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, float>,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, double>);
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, double>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace, int>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::CPUPlace,
+                                           int64_t>);

paddle/operators/fill_constant_batch_size_like_op.cu.cc

@@ -19,4 +19,7 @@ namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(
     fill_constant_batch_size_like,
     ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, float>,
-    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, double>);
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, double>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace, int>,
+    ops::FillConstantBatchSizeLikeOpKernel<paddle::platform::GPUPlace,
+                                           int64_t>);

paddle/operators/fill_zeros_like_op.cc

@@ -54,5 +54,8 @@ namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, ops::FillZerosLikeOp,
                              ops::FillZerosLikeOpMaker);
 REGISTER_OP_CPU_KERNEL(
-    fill_zeros_like,
-    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, float>);
+    fill_zeros_like, ops::FillZerosLikeKernel<paddle::platform::CPUPlace, int>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, int64_t>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, float>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, double>,
+    ops::FillZerosLikeKernel<paddle::platform::CPUPlace, bool>);

paddle/operators/fill_zeros_like_op.cu.cc

@@ -17,5 +17,8 @@
 
 namespace ops = paddle::operators;
 REGISTER_OP_GPU_KERNEL(
-    fill_zeros_like,
-    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, float>);
+    fill_zeros_like, ops::FillZerosLikeKernel<paddle::platform::GPUPlace, int>,
+    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, int64_t>,
+    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, float>,
+    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, double>,
+    ops::FillZerosLikeKernel<paddle::platform::GPUPlace, bool>);

paddle/operators/math/math_function.cc

@@ -250,6 +250,8 @@ void axpy<platform::CPUPlace, double>(const platform::DeviceContext& context,
 template struct SetConstant<platform::CPUPlace, float>;
 template struct SetConstant<platform::CPUPlace, double>;
 template struct SetConstant<platform::CPUPlace, int>;
+template struct SetConstant<platform::CPUPlace, int64_t>;
+template struct SetConstant<platform::CPUPlace, bool>;
 
 #define DEFINE_CPU_TRANS(RANK)                                \
   template struct Transpose<platform::CPUPlace, float, RANK>; \

paddle/operators/math/math_function.cu

@@ -256,6 +256,8 @@ void axpy<platform::GPUPlace, double>(const platform::DeviceContext& context,
 template struct SetConstant<platform::GPUPlace, float>;
 template struct SetConstant<platform::GPUPlace, double>;
 template struct SetConstant<platform::GPUPlace, int>;
+template struct SetConstant<platform::GPUPlace, int64_t>;
+template struct SetConstant<platform::GPUPlace, bool>;
 
 #define DEFINE_GPU_TRANS(RANK)                                \
   template struct Transpose<platform::GPUPlace, float, RANK>; \

paddle/operators/sum_op.cc

@@ -12,6 +12,7 @@ limitations under the License. */
 #include "paddle/operators/sum_op.h"
 #include <vector>
 #include "paddle/framework/var_type_inference.h"
+#include "paddle/operators/detail/safe_ref.h"
 
 namespace paddle {
 namespace operators {
@@ -59,7 +60,8 @@ class SumOp : public framework::OperatorWithKernel {
               x_vars[0]->Get<framework::SelectedRows>().value().type()),
           ctx.device_context());
     } else if (x_vars[0]->IsType<framework::LoDTensorArray>()) {
-      auto& array = x_vars[0]->Get<framework::LoDTensorArray>();
+      for (auto& x_var : x_vars) {
+        auto& array = x_var->Get<framework::LoDTensorArray>();
         for (auto& each : array) {
           if (each.numel() != 0) {
             return framework::OpKernelType(framework::ToDataType(each.type()),
@@ -67,6 +69,8 @@ class SumOp : public framework::OperatorWithKernel {
           }
         }
       }
+      PADDLE_THROW("Cannot find the input data type by all input data");
+    }
     PADDLE_THROW("Unexpected branch. Input type is %s",
                  x_vars[0]->Type().name());
   }
@@ -96,6 +100,11 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
     auto& inputs = op_desc.Input("X");
     auto var_type = framework::VarDesc::SELECTED_ROWS;
 
+    for (auto& name : op_desc.Input("X")) {
+      VLOG(10) << name << " "
+               << block->FindRecursiveOrCreateVar(name)->GetType();
+    }
+
     bool any_input_is_lod_tensor = std::any_of(
         inputs.begin(), inputs.end(), [block](const std::string& name) {
           return block->FindRecursiveOrCreateVar(name)->GetType() ==
@@ -103,7 +112,7 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
         });
 
     auto is_tensor_array = [block](const std::string& name) {
-      return block->FindRecursiveOrCreateVar(name)->GetType() ==
+      return detail::Ref(block->FindRecursiveOrCreateVar(name)).GetType() ==
              framework::VarDesc::LOD_TENSOR_ARRAY;
     };
 
@@ -113,14 +122,26 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
         std::all_of(inputs.begin(), inputs.end(), is_tensor_array);
 
     if (any_input_is_tensor_array) {
-      PADDLE_ENFORCE(all_inputs_are_tensor_array);
+      if (!all_inputs_are_tensor_array) {
+        std::ostringstream os;
+        for (auto& each : inputs) {
+          os << "    " << each << " type is "
+             << detail::Ref(block->FindRecursiveOrCreateVar(each)).GetType()
+             << "\n";
+        }
+        PADDLE_ENFORCE(all_inputs_are_tensor_array,
+                       "Not all inputs are tensor array:\n%s", os.str());
+      }
       var_type = framework::VarDesc::LOD_TENSOR_ARRAY;
     } else if (any_input_is_lod_tensor) {
       var_type = framework::VarDesc::LOD_TENSOR;
     }
 
     auto out_var_name = op_desc.Output("Out").front();
-    block->FindRecursiveOrCreateVar(out_var_name)->SetType(var_type);
+    auto& out_var = detail::Ref(block->FindRecursiveOrCreateVar(out_var_name));
+    out_var.SetType(var_type);
+    auto& in_var = detail::Ref(block->FindVarRecursive(inputs.front()));
+    out_var.SetDataType(in_var.GetDataType());
   }
 };
 

paddle/operators/tensor_array_read_write_op.cc

@@ -12,7 +12,7 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 #include "paddle/operators/array_operator.h"
-
+#include "paddle/operators/detail/safe_ref.h"
 namespace paddle {
 namespace operators {
 
@@ -33,6 +33,8 @@ class WriteToArrayOp : public ArrayOp {
     auto *out =
         scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensorArray>();
     if (offset >= out->size()) {
+      VLOG(10) << "Resize " << Output("Out") << " from " << out->size()
+               << " to " << offset + 1;
       out->resize(offset + 1);
     }
     auto *out_tensor = &out->at(offset);
@@ -85,11 +87,15 @@ class WriteToArrayInferVarType : public framework::VarTypeInference {
  public:
   void operator()(const framework::OpDescBind &op_desc,
                   framework::BlockDescBind *block) const override {
-    for (auto &out_var : op_desc.OutputArgumentNames()) {
-      VLOG(10) << "Set Variable " << out_var << " as LOD_TENSOR_ARRAY";
-      block->FindRecursiveOrCreateVar(out_var)->SetType(
-          framework::VarDesc::LOD_TENSOR_ARRAY);
-    }
+    auto x_name = op_desc.Input("X")[0];
+    auto out_name = op_desc.Output("Out")[0];
+    VLOG(10) << "Set Variable " << out_name << " as LOD_TENSOR_ARRAY";
+    auto &out = detail::Ref(block->FindRecursiveOrCreateVar(out_name),
+                            "Cannot found %s", out_name);
+    out.SetType(framework::VarDesc::LOD_TENSOR_ARRAY);
+    auto &x =
+        detail::Ref(block->FindVarRecursive(x_name), "Cannot found %s", x_name);
+    out.SetDataType(x.GetDataType());
   }
 };
 
@@ -107,11 +113,11 @@ class ReadFromArrayOp : public ArrayOp {
     auto &x_array = x->Get<framework::LoDTensorArray>();
     auto *out = scope.FindVar(Output("Out"));
     PADDLE_ENFORCE(out != nullptr, "Out must be set");
-    auto *out_tesnor = out->GetMutable<framework::LoDTensor>();
+    auto *out_tensor = out->GetMutable<framework::LoDTensor>();
     size_t offset = GetOffset(scope, dev_ctx);
     PADDLE_ENFORCE_LT(offset, x_array.size());
-    out_tesnor->CopyFrom(x_array[offset], dev_ctx.GetPlace(), dev_ctx);
-    out_tesnor->set_lod(x_array[offset].lod());
+    out_tensor->CopyFrom(x_array[offset], dev_ctx.GetPlace(), dev_ctx);
+    out_tensor->set_lod(x_array[offset].lod());
   }
 };
 

paddle/operators/while_op.cc

@@ -14,8 +14,10 @@
 
 #include <vector>
 #include "paddle/framework/executor.h"
+#include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
+#include "paddle/operators/detail/safe_ref.h"
 
 namespace paddle {
 namespace operators {
@@ -26,8 +28,9 @@ using LoDTensor = framework::LoDTensor;
 constexpr char kStepBlock[] = "step_block";
 constexpr char kCondition[] = "Condition";
 constexpr char kStepScopes[] = "StepScopes";
-constexpr char kParamGrads[] = "X@Grad";
 constexpr char kParameters[] = "X";
+constexpr char kParamGrads[] = "X@GRAD";
+constexpr char kOutputs[] = "Out";
 
 class WhileOp : public framework::OperatorBase {
  public:
@@ -71,9 +74,9 @@ class WhileOpMaker : public framework::OpProtoAndCheckerMaker {
         kCondition,
         "(Bool) An scalar. When it's False, the While Op will be terminated.")
         .AsDuplicable();
-    AddOutput("Out",
+    AddOutput(kOutputs,
               "A set of variables, which will be assigned with values "
-              "generated by perators inside the block of While Op.")
+              "generated by the operators inside the block of While Op.")
         .AsDuplicable();
     AddOutput(kStepScopes,
               "(StepScopeVar) A vector of local scope, which size equals the "
@@ -104,17 +107,64 @@ class WhileGradOp : public framework::OperatorBase {
     auto *step_scopes =
         scope.FindVar(Input(kStepScopes))->GetMutable<StepScopeVar>();
 
+    auto outside_og_names = Inputs(framework::GradVarName(kOutputs));
+    auto inside_og_names =
+        Attr<std::vector<std::string>>("original_output_grad");
+
+    PADDLE_ENFORCE_EQ(outside_og_names.size(), inside_og_names.size());
+
     for (auto cur_scope_iter = step_scopes->rbegin();
          cur_scope_iter != step_scopes->rend(); ++cur_scope_iter) {
+      VLOG(3) << "Start backward at time_step "
+              << cur_scope_iter - step_scopes->rbegin();
+      framework::Scope &cur_scope = **cur_scope_iter;
+      // Link OG from outside to inside
+      for (size_t i = 0; i < outside_og_names.size(); ++i) {
+        auto outside_og_name = outside_og_names[i];
+        auto inside_og_name = inside_og_names[i];
+        VLOG(10) << "Linking outside " << outside_og_name << " --> inside "
+                 << inside_og_name;
+        auto &og_outside = detail::Ref(scope.FindVar(outside_og_name));
+        auto &og_inside = detail::Ref(cur_scope.Var(inside_og_name));
+        if (og_outside.Type().hash_code() ==
+            typeid(framework::LoDTensor).hash_code()) {
+          auto &outside_tensor = og_outside.Get<framework::LoDTensor>();
+          auto &inside_tensor =
+              detail::Ref(og_inside.GetMutable<framework::LoDTensor>());
+          inside_tensor.set_lod(outside_tensor.lod());
+          inside_tensor.ShareDataWith(outside_tensor);
+        } else if (og_outside.Type().hash_code() ==
+                   typeid(framework::LoDTensorArray).hash_code()) {
+          auto &outside_array = og_outside.Get<framework::LoDTensorArray>();
+          auto &inside_array =
+              detail::Ref(og_inside.GetMutable<framework::LoDTensorArray>());
+          VLOG(10) << outside_og_name << " size = " << outside_array.size();
+          inside_array.resize(outside_array.size());
+
+          for (size_t j = 0; j < inside_array.size(); ++j) {
+            VLOG(10) << j << " " << outside_array[j].numel();
+            if (outside_array[j].numel() != 0) {
+              inside_array[j].set_lod(outside_array[j].lod());
+              inside_array[j].ShareDataWith(outside_array[j]);
+            } else {
+              PADDLE_ENFORCE_EQ(inside_array[j].numel(), 0);
+            }
+          }
+        }
+      }
+
       executor.Run(*program, *cur_scope_iter, block->ID(), false);
 
       auto &pg_names = Outputs(kParamGrads);
       auto &p_names = Inputs(kParameters);
       PADDLE_ENFORCE_EQ(pg_names.size(), p_names.size());
-      for (size_t prog_id = 0; prog_id < pg_names.size(); ++prog_id) {
-        auto inside_grad_name = framework::GradVarName(p_names[prog_id]);
+      for (size_t param_id = 0; param_id < pg_names.size(); ++param_id) {
+        if (pg_names[param_id] == framework::kEmptyVarName) {
+          continue;  // iterator doesn't have gradient
+        }
+        auto inside_grad_name = framework::GradVarName(p_names[param_id]);
 
-        //  // TODO(tonyyang-savil: Not sure we need the following
+        //  // TODO(tonyyang-svail): Not sure we need the following
         //  // If does not compute gradient of that variable inside rnn,
         //  just
         //  // continue
@@ -126,7 +176,7 @@ class WhileGradOp : public framework::OperatorBase {
         // zero gradient variable in step 0
         if (cur_scope_iter == step_scopes->rbegin()) {
           auto *var = (*cur_scope_iter)->FindVar(inside_grad_name);
-          PADDLE_ENFORCE_NOT_NULL(var);
+          PADDLE_ENFORCE_NOT_NULL(var, "Can not find var %s", inside_grad_name);
           if (var->IsType<LoDTensor>()) {
             auto &inside_tensor = var->Get<framework::LoDTensor>();
             framework::AttributeMap attrs;
@@ -135,27 +185,18 @@ class WhileGradOp : public framework::OperatorBase {
             attrs["value"] = 0.0f;
 
             auto zero_op = framework::OpRegistry::CreateOp(
-                "fill_constant", {}, {{"Out", {pg_names[prog_id]}}}, attrs);
+                "fill_constant", {}, {{"Out", {pg_names[param_id]}}}, attrs);
             zero_op->Run(scope, dev_ctx);
           }
         }
 
         // sum gradient
-        auto *outside_var = scope.FindVar(pg_names[prog_id]);
-        PADDLE_ENFORCE_NOT_NULL(outside_var);
-        auto &outside_tensor = *outside_var->GetMutable<framework::LoDTensor>();
-
-        std::string result_var_name;
-        auto *local_result_var = (*cur_scope_iter)->Var(&result_var_name);
-        auto &local_result_tensor =
-            *local_result_var->GetMutable<framework::LoDTensor>();
-
-        local_result_tensor.ShareDataWith(outside_tensor);
-
+        auto new_inside_name = cur_scope.Rename(inside_grad_name);
         auto sum_op = framework::OpRegistry::CreateOp(
-            "sum", {{"X", {result_var_name, inside_grad_name}}},
-            {{"Out", {result_var_name}}}, {});
-        sum_op->Run(**cur_scope_iter, dev_ctx);
+            "sum", {{"X", {pg_names[param_id], new_inside_name}}},
+            {{"Out", {pg_names[param_id]}}}, {});
+        sum_op->Run(cur_scope, dev_ctx);
+        cur_scope.Rename(new_inside_name, inside_grad_name);
       }
     }
   }
@@ -169,29 +210,110 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker {
   virtual std::unique_ptr<framework::OpDescBind> Apply() const {
     auto *grad = new framework::OpDescBind();
     grad->SetType("while_grad");
-    for (auto &input_param : this->InputNames()) {
-      grad->SetInput(input_param, this->Input(input_param));
-      grad->SetOutput(framework::GradVarName(input_param),
-                      this->InputGrad(input_param));
+    grad->SetInput(kParameters, Input(kParameters));
+    grad->SetOutput(
+        framework::GradVarName(kParameters),
+        InputGrad(kParameters, /*do not drop empty gradient*/ false));
+    grad->SetInput(kOutputs, Output(kOutputs));
+
+    // OG should be re-calculated by step blocks, since many outputs of while op
+    // do not need to calculate gradients.
+    std::unordered_set<std::string> block_ins;
+    {
+      for (auto &p : Input(kParameters)) {
+        block_ins.insert(p);
+      }
+      for (auto &o : Output(kOutputs)) {
+        block_ins.insert(o);
+      }
+    }
+    std::unordered_set<std::string> extra_inputs;
+    for (size_t i = 0; i < grad_block_[0]->OpSize(); ++i) {
+      for (auto &input_name : grad_block_[0]->Op(i)->InputArgumentNames()) {
+        if (block_ins.find(input_name) != block_ins.end()) {
+          continue;
+        }
+        extra_inputs.insert(input_name);
       }
 
-    for (auto &output_param : this->OutputNames()) {
-      grad->SetInput(output_param, this->Output(output_param));
-      if (output_param != kStepScopes) {
-        grad->SetInput(framework::GradVarName(output_param),
-                       this->OutputGrad(output_param));
+      for (auto &output_name : grad_block_[0]->Op(i)->OutputArgumentNames()) {
+        block_ins.insert(output_name);
       }
     }
+
+    std::vector<std::string> extra_inputs_list;
+    extra_inputs_list.resize(extra_inputs.size());
+    std::copy(extra_inputs.begin(), extra_inputs.end(),
+              extra_inputs_list.begin());
+    grad->SetInput(framework::GradVarName(kOutputs), extra_inputs_list);
+    grad->SetInput(kStepScopes, Output(kStepScopes));
     grad->SetAttrMap(this->Attrs());
     grad->SetBlockAttr(kStepBlock, *grad_block_[0]);
+    // record the original output gradient names, since the gradient name of
+    // while operator could be renamed.
+    grad->SetAttr("original_output_grad", extra_inputs_list);
 
     return std::unique_ptr<framework::OpDescBind>(grad);
   }
 };
 
+class WhileGradOpVarTypeInference : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDescBind &op_desc,
+                  framework::BlockDescBind *block) const override {
+    auto p_names = op_desc.Input(kParameters);
+    auto pg_names = op_desc.Output(framework::GradVarName(kParameters));
+
+    for (size_t i = 0; i < p_names.size(); ++i) {
+      auto &p_var = detail::Ref(block->FindVarRecursive(p_names[i]));
+      auto *g_var = block->FindVarRecursive(pg_names[i]);
+      if (g_var != nullptr) {  // Gradient could be @EMPTY@
+        VLOG(5) << "Setting " << pg_names[i] << " following " << p_names[i]
+                << " type: " << p_var.GetType();
+        g_var->SetType(p_var.GetType());
+        g_var->SetDataType(p_var.GetDataType());
+      }
+    }
+  }
+};
+
+class WhileGradOpShapeInference : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *ctx) const override {
+    ctx->HasInputs(kParameters);
+    ctx->HasOutputs(framework::GradVarName(kParameters));
+    ctx->HasInputs(kOutputs);
+    ctx->HasInputs(framework::GradVarName(kOutputs));
+
+    auto p_names = ctx->Inputs(kParameters);
+    auto pg_names = ctx->Outputs(kParamGrads);
+    auto dims = ctx->GetInputsDim(kParameters);
+    auto var_types = ctx->GetInputsVarType(kParameters);
+    std::vector<std::string> names_to_set;
+    std::vector<framework::DDim> dims_to_set;
+    for (size_t i = 0; i < p_names.size(); ++i) {
+      if (pg_names[i] == framework::kEmptyVarName) {
+        continue;
+      }
+      if (var_types[i] == framework::VarDesc::LOD_TENSOR) {
+        names_to_set.push_back(pg_names[i]);
+        dims_to_set.push_back(dims[i]);
+      } else if (var_types[i] == framework::VarDesc::LOD_TENSOR_ARRAY) {
+        // not sure how to set the dim of LOD_TENSOR_ARRAY
+        names_to_set.push_back(pg_names[i]);
+        dims_to_set.push_back(dims[i]);
+      }
+    }
+    ctx->SetDims(names_to_set, dims_to_set);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 REGISTER_OPERATOR(while, paddle::operators::WhileOp,
                   paddle::operators::WhileOpMaker,
                   paddle::operators::WhileGradOpDescMaker);
+REGISTER_OPERATOR(while_grad, paddle::operators::WhileGradOp,
+                  paddle::operators::WhileGradOpShapeInference,
+                  paddle::operators::WhileGradOpVarTypeInference);

python/paddle/v2/fluid/framework.py

@@ -12,9 +12,9 @@ def unique_name(prefix):
     return "_".join([prefix, str(uid)])
 
 
-def _debug_string_(proto):
+def _debug_string_(proto, throw_on_error=True):
     error_fields = list()
-    if not proto.IsInitialized(error_fields):
+    if not proto.IsInitialized(error_fields) and throw_on_error:
         raise ValueError("{0} are not initialized\nThe message is {1}".format(
             error_fields, proto))
     return proto.__str__()
@@ -101,9 +101,12 @@ class Variable(object):
         self.stop_gradient = stop_gradient
 
     def __str__(self):
+        return self.to_string(True)
+
+    def to_string(self, throw_on_error):
         protostr = self.desc.serialize_to_string()
         proto = framework_pb2.VarDesc.FromString(str(protostr))
-        return _debug_string_(proto)
+        return _debug_string_(proto, throw_on_error)
 
     __repr__ = __str__
 
@@ -291,10 +294,13 @@ class Operator(object):
             self.desc.infer_var_type(self.block.desc)
             self.desc.infer_shape(self.block.desc)
 
-    def __str__(self):
+    def to_string(self, throw_on_error):
         protostr = self.desc.serialize_to_string()
         proto = framework_pb2.OpDesc.FromString(str(protostr))
-        return _debug_string_(proto)
+        return _debug_string_(proto, throw_on_error)
+
+    def __str__(self):
+        return self.to_string(True)
 
     __repr__ = __str__
 
@@ -349,9 +355,12 @@ class Block(object):
         self.program = program
 
     def __str__(self):
+        return self.to_string(True)
+
+    def to_string(self, throw_on_error):
         protostr = self.desc.serialize_to_string()
         proto = framework_pb2.BlockDesc.FromString(str(protostr))
-        return _debug_string_(proto)
+        return _debug_string_(proto, throw_on_error)
 
     __repr__ = __str__
 
@@ -454,9 +463,12 @@ class Program(object):
         self.current_block_idx = 0
 
     def __str__(self):
+        return self.to_string(True)
+
+    def to_string(self, throw_on_error):
         protostr = self.desc.serialize_to_string()
         proto = framework_pb2.ProgramDesc.FromString(str(protostr))
-        return _debug_string_(proto)
+        return _debug_string_(proto, throw_on_error)
 
     def clone(self):
         p = Program()
@@ -512,7 +524,14 @@ class Program(object):
         assert isinstance(target, Variable)
         if no_grad_set is None:
             no_grad_set = set()
-        param_to_grad_info = self.desc.append_backward(target.desc, no_grad_set)
+        try:
+            param_to_grad_info = self.desc.append_backward(target.desc,
+                                                           no_grad_set)
+        except Exception as e:
+            raise core.EnforceNotMet(
+                str(e) + "\nCurrent protobuf is\n{0}".format(
+                    self.to_string(False)))
+
         self.sync_with_cpp()
         return param_to_grad_info
 

python/paddle/v2/fluid/net_drawer.py

@@ -66,10 +66,13 @@ def parse_graph(program, graph, var_dict, **kwargs):
             if not var_dict.has_key(var):
                 var_dict[var] = "Feed"
 
+    temp_id = 0
     proto = framework_pb2.ProgramDesc.FromString(
         program.desc.serialize_to_string())
     for block in proto.blocks:
         for op in block.ops:
+            op.type = op.type + "_" + str(temp_id)
+            temp_id += 1
             graph.node(**draw_node(op))
             for o in op.outputs:
                 for arg in o.arguments:
@@ -78,6 +81,7 @@ def parse_graph(program, graph, var_dict, **kwargs):
                 for arg in e.arguments:
                     if var_dict.has_key(arg):
                         graph.edge(**draw_edge(var_dict, op, e, arg))
+        break  # only plot the first block
 
 
 def draw_graph(startup_program, main_program, **kwargs):

python/paddle/v2/fluid/tests/test_while_op.py

@@ -2,6 +2,7 @@ import unittest
 import paddle.v2.fluid.layers as layers
 from paddle.v2.fluid.executor import Executor
 import paddle.v2.fluid.core as core
+from paddle.v2.fluid.backward import append_backward_ops
 import numpy
 
 
@@ -16,7 +17,7 @@ class TestWhileOp(unittest.TestCase):
         i = layers.zeros(shape=[1], dtype='int64')
         i.stop_gradient = True
         init = layers.zeros(shape=[10], dtype='float32')
-        mem_array = layers.array_write(init, i=i)
+        mem_array = layers.array_write(x=init, i=i)
         data_array = layers.array_write(x=d0, i=i)
 
         i = layers.increment(i)
@@ -29,17 +30,23 @@ class TestWhileOp(unittest.TestCase):
         i.stop_gradient = True
 
         array_len = layers.fill_constant(shape=[1], dtype='int64', value=3)
+        array_len.stop_gradient = True
         cond = layers.less_than(x=i, y=array_len)
 
         while_op = layers.While(cond=cond)
         with while_op.block():
             d = layers.array_read(array=data_array, i=i)
             prev = layers.array_read(array=mem_array, i=i)
-            i = layers.increment(x=i, in_place=True)
             result = layers.sums(input=[d, prev])
+
+            i = layers.increment(x=i, in_place=True)
             layers.array_write(result, i=i, array=mem_array)
             layers.less_than(x=i, y=array_len, cond=cond)
-        sum_result = layers.array_read(mem_array, i=array_len)
+
+        sum_result = layers.array_read(array=mem_array, i=i)
+        loss = layers.mean(x=sum_result)
+
+        append_backward_ops(loss)
 
         cpu = core.CPUPlace()
         exe = Executor(cpu)