Merge pull request #3068 from reyoung/feature/backward

Feature/backward

paddle/framework/CMakeLists.txt

@@ -32,4 +32,7 @@ add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch
 add_dependencies(framework_py_proto framework_py_proto_init)
 
 cc_library(net SRCS net.cc DEPS op_registry)
-cc_test(net_op_test SRCS net_op_test.cc DEPS net add_op mul_op sigmoid_op softmax_op fc_op)
+cc_test(net_op_test SRCS net_op_test.cc DEPS net)
+
+cc_library(backward SRCS backward.cc DEPS net)
+cc_test(backward_test SRCS backward_test.cc DEPS backward)

paddle/framework/backward.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/backward.h"
+#include <list>
+#include "paddle/framework/net.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace framework {
+
+static bool AllInSet(const std::vector<std::string>& names,
+                     const std::string& suffix,
+                     const std::unordered_set<std::string>& set) {
+  for (auto& name : names) {
+    if (set.find(name + suffix) == set.end()) {
+      return false;
+    }
+  }
+  return true;
+}
+
+static std::shared_ptr<OperatorBase> NOP() {
+  auto net_op = std::make_shared<NetOp>();
+  net_op->type_ = "@NOP@";
+  net_op->CompleteAddOp();
+  return net_op;
+}
+
+//  Get backward operator from a forward operator, recursively implementation.
+//
+//  no_grad_names the gradient variable names without gradient calculating.
+//
+//  uniq_id is a unique index used inside recursively calling BackwardRecursive.
+//  use `uid = uniq_id++;` to get the unique index, and pass `uniq_id` through
+//  recursive calling.
+//
+//  returns The backward operator. For simple situation, it is a simple
+//  operator. For complex situation, it is a NetOp.
+//
+//  See Backward.h for details
+static std::shared_ptr<OperatorBase> BackwardRecursive(
+    const OperatorBase& forwardOp,
+    std::unordered_set<std::string>& no_grad_names, size_t& uniq_id);
+std::shared_ptr<OperatorBase> BackwardRecursive(
+    const OperatorBase& forwardOp,
+    std::unordered_set<std::string>& no_grad_names, 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.
+  if (AllInSet(forwardOp.inputs_, OperatorBase::GRAD_VAR_SUFFIX(),
+               no_grad_names)) {
+    return NOP();
+  }
+
+  //  All output gradients of forwarding operator do not need to calculate. Then
+  //  all input gradients cannot be computed at all, and we put them into
+  //  `no_grad_names` set. Return an NOP.
+  if (AllInSet(forwardOp.outputs_, OperatorBase::GRAD_VAR_SUFFIX(),
+               no_grad_names)) {
+    for (auto& name : forwardOp.inputs_) {
+      // Mark all input is not need
+      no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX());
+    }
+    return NOP();
+  }
+
+  // Returned gradient network
+  auto net = std::make_shared<NetOp>();
+
+  if (forwardOp.IsNetOp()) {
+    // Because forwardOp is a net op, it can static_cast.
+    auto& forwardNet = static_cast<const NetOp&>(forwardOp);
+
+    // Map from output gradient variable name to operator's indices in backward
+    // net. That operator generates that variable.
+    std::unordered_map<std::string, std::vector<size_t>> dup_output_ops;
+
+    size_t local_op_id = 0;
+    // reversely travel forwardNet
+    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);
+      net->AddOp(bwd);
+      for (auto& out : bwd->outputs_) {
+        dup_output_ops[out].emplace_back(local_op_id);
+      }
+    }
+    // Get unique ID for this method.
+    auto uid = uniq_id++;
+    // TODO(dzh): more comment
+    using Pos = std::pair<size_t, std::shared_ptr<OperatorBase>>;
+    std::list<Pos> insert_position;
+    for (auto& dup_output_op : dup_output_ops) {
+      const std::string& name = dup_output_op.first;
+      auto& dup_op = dup_output_op.second;
+      if (dup_op.size() == 1) continue;
+      std::vector<std::string> dup_outputs;
+
+      for (size_t i = 0; i < dup_op.size(); ++i) {
+        auto op_offset = dup_op[i];
+        dup_outputs.push_back(name + "@RENAME@" + std::to_string(uid) + "@" +
+                              std::to_string(i));
+        net->ops_[op_offset]->Rename(name, dup_outputs.back());
+      }
+      insert_position.push_back(
+          {dup_op.back(),
+           OpRegistry::CreateOp(
+               "add", {dup_outputs}, {name},
+               {{"input_format",
+                 std::vector<int>{0, static_cast<int>(dup_outputs.size())}}})});
+    }
+
+    insert_position.sort(
+        [](const Pos& l, const Pos& r) { return l.first > r.first; });
+
+    for (auto& pos : insert_position) {
+      net->InsertOp(pos.first + 1, pos.second);
+    }
+
+  } else {
+    std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
+    for (std::string& grad_input : grad_op->inputs_) {
+      if (no_grad_names.count(grad_input)) {
+        std::string prefix = grad_input.substr(
+            0, grad_input.size() - OperatorBase::GRAD_VAR_SUFFIX().size());
+        grad_input = prefix + OperatorBase::ZERO_VAR_SUFFIX();
+
+        // If part of input gradient of that operator is not calculated, fill
+        // zero variables to that input gradient.
+        net->AddOp(OpRegistry::CreateOp("fill_zeros_like", {prefix},
+                                        {grad_input}, {}));
+      }
+    }
+
+    for (std::string& grad_output : grad_op->outputs_) {
+      if (no_grad_names.count(grad_output)) {
+        grad_output = OperatorBase::EMPTY_VAR_NAME();
+      }
+    }
+
+    if (net->ops_.empty()) {  // Current no aux op is added to network
+      return grad_op;
+    }
+    net->AddOp(grad_op);
+  }
+  net->type_ = "@GENERATED_BACKWARD@";
+  net->CompleteAddOp();
+  return net;
+}
+
+// See header for comments
+std::shared_ptr<OperatorBase> Backward(
+    const OperatorBase& forwardOp,
+    const std::unordered_set<std::string>& no_grad_vars) {
+  std::unordered_set<std::string> no_grad_names;
+  no_grad_names.reserve(no_grad_vars.size());
+
+  for (auto& name : no_grad_vars) {
+    no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX());
+  }
+  size_t uid = 0;
+  return BackwardRecursive(forwardOp, no_grad_names, uid);
+}
+}  // namespace framework
+}  // namespace paddle

paddle/framework/backward.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 <unordered_set>
+#include "operator.h"
+namespace paddle {
+namespace framework {
+
+// Create the backward operator from a forward operator.
+// TODO(yuyang18): Add more API reference comment.
+extern std::shared_ptr<OperatorBase> Backward(
+    const OperatorBase& forwardOp,
+    const std::unordered_set<std::string>& no_grad_vars);
+}  // namespace framework
+}  // namespace paddle

paddle/framework/backward.md

+## Operator/expression 's Backward
+
+### Motivation
+
+In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation lineage, the operator/ expression's Backward feature will generate the backward pass respect to forward pass.
+
+### Implement : gradient operator registry
+
+|                        | forward operator | backward operator                |
+| ---------------------- | ---------------- | -------------------------------- |
+| **Operator::inputs_**  | Inputs           | Inputs, Outputs, OutputGradients |
+| **Operator::outputs_** | Outputs          | InputGradients                   |
+
+Inputs/Outputs means the input/output of the operator,  InputGradients/OutputGradients is the gradient respect to forward opeartor. Forward operator and Backward operator are isomorphic, save their corresponding needs into member attribute.
+
+We use a global hash map record the gradient operators available, follow the philosophy  of minimum core, make operator pluggable unit. Each gradient is an operator and it needs to regist itself. 
+
+grad_op_builder(fengjiayi)
+
+### Implement : Backward network
+
+given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`.
+
+1. bla bla bla (yuyang)
+
+2. NetOp 
+
+   when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively and ensure them done. During the process, we need to collect the `OutputGradients` name.
+
+   We share variable in the same scope, as a result, duplicate operator `OutputGradients` will overwirte then duplicate variable.  
+
+   ![./images/duplicate_op]()
+
+    Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator instead. 
+
+![./images/duplicate_op2]()
+
+​	Then collect the sub graph OutputGradients/InputGradients as the NetOp's and return it.

paddle/framework/backward_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/backward.h"
+
+#include <gtest/gtest.h>
+#include "paddle/framework/net.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace framework {
+
+class EmptyOp : public OperatorBase {
+ public:
+  void InferShape(const Scope &scope) const override {}
+  void Run(const Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {}
+};
+
+class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "Input X of Add").IgnoreGradient();
+    AddInput("b", "Bias of Add").IgnoreGradient();
+    AddOutput("Out", "Out of Add").IgnoreGradient();
+    AddComment("Add Op");
+  }
+};
+
+class MulOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("A", "A");
+    AddInput("B", "B");
+    AddOutput("Out", "Out");
+    AddComment("Mul");
+  }
+};
+
+class SigmoidOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "X");
+    AddOutput("Y", "Y");
+    AddComment("Sigmoid");
+  }
+};
+
+class NoGradOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  NoGradOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "X input");
+    AddOutput("Y", "Y output");
+    AddComment("NoGradOp, same input output. no Grad");
+  }
+};
+
+class FcOp : public NetOp {
+ public:
+  void Init() override {
+    AddOp(OpRegistry::CreateOp("mul", {Input("X"), Input("W")},
+                               {Output("mul_result")}, {}));
+    auto b_name = Input("b");
+    std::string before_act = "mul_result";
+    if (b_name != EMPTY_VAR_NAME()) {
+      AddOp(OpRegistry::CreateOp("rowwise_add", {Output("mul_result"), b_name},
+                                 {Output("add_result")}, {}));
+      before_act = "add_result";
+    } else {
+      auto out_varname = Output("add_result");
+      if (out_varname != EMPTY_VAR_NAME()) {
+        this->Rename(out_varname, EMPTY_VAR_NAME());
+      }
+    }
+
+    AddOp(OpRegistry::CreateOp("sigmoid", {Output(before_act)}, {Output("Out")},
+                               {}));
+    CompleteAddOp(false);
+  }
+};
+
+class FcOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  FcOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "x");
+    AddInput("W", "w");
+    AddInput("b", "b");
+    AddOutput("mul_result", "").SetTemporary();
+    AddOutput("add_result", "").SetTemporary();
+    AddOutput("Out", "");
+    AddComment("");
+  }
+};
+
+class ManyOutputOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  ManyOutputOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("x", "x");
+    AddOutput("y", "y");
+    AddOutput("z", "z");
+    AddComment("");
+  }
+};
+
+class FillZeroOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  FillZeroOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("x", "x");
+    AddOutput("out", "out");
+    AddComment("");
+  }
+};
+
+class AddOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "x").SetMultiple();
+    AddOutput("Y", "y");
+    AddComment("");
+  }
+};
+}  // namespace framework
+}  // namespace paddle
+
+namespace f = paddle::framework;
+using EnforceNotMet = paddle::platform::EnforceNotMet;
+REGISTER_OP(rowwise_add, f::EmptyOp, f::RowWiseAddOpMaker);
+REGISTER_GRADIENT_OP(rowwise_add, rowwise_add_grad, f::EmptyOp);
+REGISTER_OP(mul, f::EmptyOp, f::MulOpMaker);
+REGISTER_GRADIENT_OP(mul, mul_grad, f::EmptyOp);
+REGISTER_OP(sigmoid, f::EmptyOp, f::SigmoidOpMaker);
+REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, f::EmptyOp);
+REGISTER_OP(nograd, f::EmptyOp, f::NoGradOpMaker);
+REGISTER_OP(fill_zeros_like, f::EmptyOp, f::FillZeroOpMaker);
+REGISTER_OP(add, f::EmptyOp, f::AddOpMaker);
+REGISTER_GRADIENT_OP(add, add_grad, f::EmptyOp);
+REGISTER_OP(fc, f::FcOp, f::FcOpMaker);
+REGISTER_OP(many_output_op, f::EmptyOp, f::ManyOutputOpMaker);
+REGISTER_GRADIENT_OP(many_output_op, many_output_op_grad, f::EmptyOp);
+
+TEST(Backward, simple_op_grad) {
+  auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
+  ASSERT_NE(fwd, nullptr);
+  auto gop = f::OpRegistry::CreateGradOp(*fwd);
+  ASSERT_EQ(1UL, gop->inputs_.size());
+  ASSERT_EQ("Out" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->inputs_[0]);
+  ASSERT_EQ("rowwise_add_grad", gop->type_);
+  ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->outputs_[0]);
+  ASSERT_EQ("b" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->outputs_[1]);
+
+  ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+            gop->Output("X" + f::OperatorBase::GRAD_VAR_SUFFIX()));
+}
+
+TEST(Backward, simple_op_not_need_grad) {
+  auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
+  ASSERT_NE(fwd, nullptr);
+  auto gop = f::Backward(*fwd, {"X"});
+  ASSERT_EQ(std::find(gop->outputs_.begin(), gop->outputs_.end(),
+                      "X" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            gop->outputs_.end());
+
+  auto no_input_gop = f::Backward(*fwd, {"X", "b"});
+  ASSERT_NE(no_input_gop, nullptr);
+  ASSERT_TRUE(no_input_gop->IsNetOp());
+  ASSERT_EQ(0UL, std::static_pointer_cast<f::NetOp>(no_input_gop)->ops_.size());
+}
+
+TEST(Backward, net_fc_backward_normal) {
+  std::shared_ptr<f::OperatorBase> fwd = f::OpRegistry::CreateOp(
+      "fc", {"X", "w", "b"}, {"mul_result", "add_result", "out"}, {});
+  ASSERT_NE(fwd, nullptr);
+  std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
+  ASSERT_TRUE(gop->IsNetOp());
+  auto net = static_cast<f::NetOp *>(gop.get());
+
+  ASSERT_NO_THROW(net->DebugString());
+
+  ASSERT_EQ(3UL, net->ops_.size());
+
+  f::OperatorBase &d_sigmoid = *net->ops_[0];
+  ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);
+
+  f::OperatorBase &d_add = *net->ops_[1];
+  ASSERT_EQ("rowwise_add_grad", d_add.type_);
+
+  f::OperatorBase &d_mul = *net->ops_[2];
+  ASSERT_EQ("mul_grad", d_mul.type_);
+}
+
+TEST(Backward, net_fc_backward_not_have_b) {
+  std::shared_ptr<f::OperatorBase> fwd = f::OpRegistry::CreateOp(
+      "fc", {"X", "w", f::OperatorBase::EMPTY_VAR_NAME()},
+      {"mul_result", "add_result", "tmp"}, {});
+  ASSERT_NE(fwd, nullptr);
+  std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
+  ASSERT_TRUE(gop->IsNetOp());
+  auto net = static_cast<f::NetOp *>(gop.get());
+
+  ASSERT_NO_THROW(net->DebugString());
+
+  ASSERT_EQ(2UL, net->ops_.size());
+
+  f::OperatorBase &d_sigmoid = *net->ops_[0];
+  ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);
+
+  f::OperatorBase &d_mul = *net->ops_[1];
+  ASSERT_EQ("mul_grad", d_mul.type_);
+}
+
+TEST(Backward, net_input_of_network_not_need_grad) {
+  f::NetOp net;
+  net.AddOp(f::OpRegistry::CreateOp("fc", {"X", "W1", "b1"},
+                                    {"mul_tmp_0", "add_tmp_0", "hidden0"}, {}));
+  net.AddOp(f::OpRegistry::CreateOp("fc", {"hidden0", "W2", "b2"},
+                                    {"mul_tmp_1", "add_tmp_1", "hidden1"}, {}));
+  net.CompleteAddOp();
+  auto bwd = Backward(net, {"X"});  // X@GRAD is not need.
+  ASSERT_TRUE(bwd->IsNetOp());
+  auto bwd_net = static_cast<f::NetOp *>(bwd.get());
+
+  std::unordered_set<std::string> all_output = std::unordered_set<std::string>(
+      bwd_net->outputs_.begin(), bwd_net->outputs_.end());
+  all_output.erase(f::OperatorBase::EMPTY_VAR_NAME());
+
+  for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) {
+    ASSERT_NE(all_output.find(out + f::OperatorBase::GRAD_VAR_SUFFIX()),
+              all_output.end());
+  }
+
+  // Not Generated X
+  ASSERT_EQ(all_output.find("X" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            all_output.end());
+
+  ASSERT_EQ(2UL, bwd_net->ops_.size());
+  ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp());
+  auto first_fc_grad = static_cast<f::NetOp *>(bwd_net->ops_[1].get());
+  ASSERT_EQ(3UL, first_fc_grad->ops_.size());
+  ASSERT_EQ(
+      f::OperatorBase::EMPTY_VAR_NAME(),
+      first_fc_grad->ops_[2]->Output("A" + f::OperatorBase::GRAD_VAR_SUFFIX()));
+}
+
+TEST(Backward, net_shared_weight) {
+  f::NetOp net;
+  net.AddOp(f::OpRegistry::CreateOp("mul", {"X", "W"}, {"Out"}, {}));
+  net.AddOp(f::OpRegistry::CreateOp("mul", {"Out", "W"}, {"FinalOut"}, {}));
+  net.CompleteAddOp();
+
+  auto bwd = f::Backward(net, {});
+  ASSERT_TRUE(bwd->IsNetOp());
+  auto bwd_net = static_cast<f::NetOp *>(bwd.get());
+  ASSERT_EQ(3UL, bwd_net->ops_.size());
+  ASSERT_EQ("add", bwd_net->ops_[2]->type_);
+}
+
+TEST(Backward, op_register_grad_not_for_network) {
+  auto fwd = f::OpRegistry::CreateOp(
+      "fc", {"X", "W", "b"}, {"mul_out", "add_out", "out1"},
+      {{"temporary_index", std::vector<int>{0, 1}}});
+
+  ASSERT_THROW(f::OpRegistry::CreateGradOp(*fwd), EnforceNotMet);
+}
+
+TEST(Backward, op_all_input_are_not_need) {
+  auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
+  auto backward = f::Backward(*fwd, {"X", "b"});
+  ASSERT_TRUE(backward->IsNetOp());
+  auto net = static_cast<f::NetOp *>(backward.get());
+  ASSERT_TRUE(net->ops_.empty());
+}
+
+TEST(Backward, op_all_output_are_not_need) {
+  auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
+  auto backward = f::Backward(*fwd, {"Out"});
+  ASSERT_TRUE(backward->IsNetOp());
+  auto net = static_cast<f::NetOp *>(backward.get());
+  ASSERT_TRUE(net->ops_.empty());
+}
+
+TEST(Backward, op_part_of_output_are_not_need) {
+  auto fwd = f::OpRegistry::CreateOp("many_output_op", {"X"}, {"Y", "Z"}, {});
+  auto backward = f::Backward(*fwd, {"Z"});
+  ASSERT_TRUE(backward->IsNetOp());
+  auto net = static_cast<f::NetOp *>(backward.get());
+  ASSERT_EQ(net->ops_.size(), 2UL);
+
+  auto &fill_zero = *net->ops_[0];
+  ASSERT_EQ("fill_zeros_like", fill_zero.type_);
+  ASSERT_EQ(1UL, fill_zero.inputs_.size());
+  ASSERT_EQ("Z", fill_zero.inputs_[0]);
+  ASSERT_EQ(1UL, fill_zero.outputs_.size());
+  ASSERT_EQ("Z" + f::OperatorBase::ZERO_VAR_SUFFIX(), fill_zero.outputs_[0]);
+
+  auto &d_many_out = *net->ops_[1];
+  ASSERT_EQ("many_output_op_grad", d_many_out.type_);
+  ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size());  // I/O/OG
+  ASSERT_EQ("Z" + f::OperatorBase::ZERO_VAR_SUFFIX(),
+            d_many_out.Input("z" + f::OperatorBase::GRAD_VAR_SUFFIX()));
+  ASSERT_EQ("Y" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+            d_many_out.Input("y" + f::OperatorBase::GRAD_VAR_SUFFIX()));
+  ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+            d_many_out.Output("x" + f::OperatorBase::GRAD_VAR_SUFFIX()));
+}
+
+TEST(Backward, op_part_of_input_are_not_need) {
+  auto fwd = f::OpRegistry::CreateOp("mul", {"a", "b"}, {"out"}, {});
+  auto backward = f::Backward(*fwd, {"a"});
+  auto &grad_mul = *backward;
+  ASSERT_EQ(grad_mul.type_, "mul_grad");
+  ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL);
+  ASSERT_EQ(grad_mul.outputs_.size(), 2UL);
+  ASSERT_EQ(grad_mul.Output("A" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            f::OperatorBase::EMPTY_VAR_NAME());
+  ASSERT_EQ(grad_mul.Output("B" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            "b" + f::OperatorBase::GRAD_VAR_SUFFIX());
+  ASSERT_EQ(grad_mul.Input("Out" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            "out" + f::OperatorBase::GRAD_VAR_SUFFIX());
+  ASSERT_EQ(grad_mul.Input("A"), "a");
+  ASSERT_EQ(grad_mul.Input("B"), "b");
+  ASSERT_EQ(grad_mul.Input("Out"), "out");
+}
+
+TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
+  f::NetOp net;
+  net.AddOp(f::OpRegistry::CreateOp("fc", {"x1", "w1", "b1"},
+                                    {"mul_out1", "add_out1", "out1"}, {}));
+  net.AddOp(f::OpRegistry::CreateOp("fc", {"out1", "w2", "b2"},
+                                    {"mul_out2", "tmp_out2", "out2"}, {}));
+  net.AddOp(f::OpRegistry::CreateOp("fc", {"out2", "w3", "b3"},
+                                    {"mul_out3", "tmp_out3", "out3"}, {}));
+  net.CompleteAddOp();
+  auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"});
+  ASSERT_TRUE(backward->IsNetOp());
+  auto bwd_net = static_cast<f::NetOp *>(backward.get());
+  ASSERT_EQ(bwd_net->ops_.size(), 3UL);
+  auto &grad_fc = *bwd_net->ops_[0];
+  EXPECT_EQ(grad_fc.inputs_.size(),
+            3UL       /* external input number */
+                + 1UL /* external output number*/
+                + 1UL /* number of gradient of external output*/
+                - 1UL /*ignoreGradient varable number*/
+                + 2U /* internal variable number*/);
+  EXPECT_EQ(grad_fc.outputs_.size(), 2UL       /* input number of mul*/
+                                         + 2UL /* input number of rowwise_add */
+                                         + 1UL /* input number of sigmod */);
+  EXPECT_EQ(bwd_net->ops_[1]->inputs_.size(), 0UL);
+  EXPECT_EQ(bwd_net->ops_[1]->outputs_.size(), 0UL);
+  EXPECT_EQ(bwd_net->ops_[2]->inputs_.size(), 0UL);
+  EXPECT_EQ(bwd_net->ops_[2]->outputs_.size(), 0UL);
+
+  /*
+    EXPECT_EQ(grad_fc.Output("X" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+              f::OperatorBase::EMPTY_VAR_NAME());
+  EXPECT_EQ(grad_fc.Output("W" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+    "w3" + f::OperatorBase::GRAD_VAR_SUFFIX());
+  EXPECT_EQ(grad_fc.Output("b" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+    "b3" + f::OperatorBase::GRAD_VAR_SUFFIX());
+  EXPECT_EQ(grad_fc.Output("mul_result" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+  "mul_out3" + f::OperatorBase::GRAD_VAR_SUFFIX());
+
+  EXPECT_EQ(grad_fc.Input("Out" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+  "out3" + f::OperatorBase::GRAD_VAR_SUFFIX());
+  EXPECT_EQ(grad_fc.Input("X"), "out2");
+  EXPECT_EQ(grad_fc.Input("W"), "w3");
+  EXPECT_EQ(grad_fc.Input("mul_result"), "mul_out3");
+  EXPECT_EQ(grad_fc.Input("add_result"), "tmp_out3");
+  EXPECT_EQ(grad_fc.Input("Out"), "out3");
+  */
+}

paddle/framework/grad_op_builder.cc

@@ -20,7 +20,7 @@ namespace framework {
 
 OperatorBase* GradOpBuilder::Build() {
   BuildOpInOutArgList();
-  std::string grad_op_type = OpRegistry::grad_ops().at(op_->type_);
+  std::string grad_op_type = OpRegistry::grad_ops().at(op_.type_);
   OperatorBase* grad_op = OpRegistry::op_creators().at(grad_op_type)();
   grad_op->type_ = grad_op_type;
   CompleteGradOp(grad_op);
@@ -39,15 +39,15 @@ OpInOutArg* GradOpBuilder::BuildArg(const VarProto& var,
 }
 
 void GradOpBuilder::BuildOpInOutArgList() {
-  const OpProto& op_proto = OpRegistry::protos().at(op_->type_);
-  const auto& var_map = *(OpRegistry::VarIndexMaps().at(op_->type_));
+  const OpProto& op_proto = OpRegistry::protos().at(op_.type_);
+  const auto& var_map = *(OpRegistry::VarIndexMaps().at(op_.type_));
   const std::vector<int>& in_format =
-      op_->attrs_.count("input_format")
-          ? op_->GetAttr<std::vector<int>>("input_format")
+      op_.attrs_.count("input_format")
+          ? op_.GetAttr<std::vector<int>>("input_format")
           : std::vector<int>();
   const std::vector<int>& out_format =
-      op_->attrs_.count("output_format")
-          ? op_->GetAttr<std::vector<int>>("output_format")
+      op_.attrs_.count("output_format")
+          ? op_.GetAttr<std::vector<int>>("output_format")
           : std::vector<int>();
   for (const auto& var : op_proto.inputs()) {
     arg_list_.emplace_back(
@@ -70,8 +70,7 @@ void GradOpBuilder::AddArgIntoGradOp(const OpInOutArg* arg,
   }
   (*varmap)[var_name] = idx++;
   size_t pre_sz = in_out.size();
-  auto base_it =
-      arg->type_ == IN ? op_->inputs_.begin() : op_->outputs_.begin();
+  auto base_it = arg->type_ == IN ? op_.inputs_.begin() : op_.outputs_.begin();
   std::copy(base_it + arg->begin_idx_, base_it + arg->end_idx_,
             std::back_inserter(in_out));
   if (is_grad) {
@@ -83,7 +82,7 @@ void GradOpBuilder::AddArgIntoGradOp(const OpInOutArg* arg,
 }
 
 void GradOpBuilder::CompleteGradOp(OperatorBase* grad_op) const {
-  grad_op->attrs_ = op_->attrs_;
+  grad_op->attrs_ = op_.attrs_;
   grad_op->attrs_.erase("input_format");
   grad_op->attrs_.erase("output_format");
   VarIndexMap* grad_varmap = new VarIndexMap();

paddle/framework/grad_op_builder.h

@@ -29,7 +29,7 @@ class GradOpBuilder {
   using VarIndexMap = std::unordered_map<std::string, int>;
 
  public:
-  GradOpBuilder(const OperatorBase* op) : op_(op) {}
+  GradOpBuilder(const OperatorBase& op) : op_(op) {}
   OperatorBase* Build();
 
  private:
@@ -40,7 +40,7 @@ class GradOpBuilder {
                         std::vector<int>& format, VarIndexMap* varmap, int& idx,
                         bool is_grad) const;
   void CompleteGradOp(OperatorBase* grad_op) const;
-  const OperatorBase* op_;
+  const OperatorBase& op_;
   std::vector<std::shared_ptr<OpInOutArg>> arg_list_;
 };
 

paddle/framework/grad_op_builder_test.cc

@@ -11,7 +11,7 @@ namespace framework {
 TEST(GradOpBuilder, AddTwo) {
   std::shared_ptr<OperatorBase> add_op(
       OpRegistry::CreateOp("add_two", {"x", "y"}, {"out"}, {}));
-  std::shared_ptr<OperatorBase> grad_add_op = OpRegistry::CreateGradOp(add_op);
+  std::shared_ptr<OperatorBase> grad_add_op = OpRegistry::CreateGradOp(*add_op);
   EXPECT_EQ(static_cast<int>(grad_add_op->inputs_.size()), 4);
   EXPECT_EQ(static_cast<int>(grad_add_op->outputs_.size()), 2);
   EXPECT_EQ(grad_add_op->Input("X"), "x");

paddle/framework/net.h

@@ -68,9 +68,18 @@ class NetOp : public OperatorBase {
    */
   void AddOp(const std::shared_ptr<OperatorBase>& op) {
     PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed");
+    PADDLE_ENFORCE(op != nullptr, "Cannot Insert Null op");
     ops_.push_back(op);
   }
 
+  void InsertOp(size_t pos, const std::shared_ptr<OperatorBase>& op) {
+    PADDLE_ENFORCE(!add_op_done_,
+                   "Cannot InsertOp when this network is sealed");
+    PADDLE_ENFORCE(op != nullptr, "Cannot Insert Null op");
+    PADDLE_ENFORCE(pos <= ops_.size(), "Out of range");
+    ops_.insert(ops_.begin() + pos, op);
+  }
+
   void CompleteAddOp(bool calculate = true);
 
   std::string DebugString() const override;

paddle/framework/net_op_test.cc

@@ -3,11 +3,6 @@
 #include <paddle/framework/op_registry.h>
 #include <paddle/framework/operator.h>
 
-USE_OP(add_two);
-USE_OP(mul);
-USE_OP(sigmoid);
-USE_OP(softmax);
-
 namespace paddle {
 namespace framework {
 
@@ -25,6 +20,13 @@ class TestOp : public OperatorBase {
   }
 };
 
+class EmptyOp : public OperatorBase {
+ public:
+  void InferShape(const Scope& scope) const override {}
+  void Run(const Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {}
+};
+
 template <typename T>
 void AssertSameVectorWithoutOrder(const std::vector<T>& expected,
                                   const std::vector<T>& actual) {
@@ -71,20 +73,17 @@ TEST(OpKernel, all) {
   ASSERT_THROW(net->AddOp(op2), paddle::platform::EnforceNotMet);
 }
 
-//! TODO(yuyang18): Refine Backward Op.
-// TEST(AddBackwardOp, TestGradOp) {
-//  auto net = std::make_shared<NetOp>();
-//  ASSERT_NE(net, nullptr);
-//  net->AddOp(framework::OpRegistry::CreateOp("mul", {"X", "Y"}, {"Out"}, {}));
-//  net->AddOp(
-//      framework::OpRegistry::CreateOp("add_two", {"X", "Y"}, {"Out"}, {}));
-//  net->AddOp(framework::OpRegistry::CreateOp("add_two", {"X", "Y"}, {""},
-//  {}));
-//  auto grad_ops = AddBackwardOp(net);
-//  for (auto& op : grad_ops->ops_) {
-//    op->DebugString();
-//  }
-//}
+TEST(Net, insert_op) {
+  NetOp net;
+  auto op1 = std::make_shared<EmptyOp>();
+  op1->inputs_ = {"x", "w1", "b1"};
+  op1->outputs_ = {"y"};
+  net.AddOp(op1);
+  net.InsertOp(0, op1);
+  ASSERT_EQ(2UL, net.ops_.size());
+  net.InsertOp(2, op1);
+  ASSERT_EQ(3UL, net.ops_.size());
+}
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/op_registry.h

@@ -86,43 +86,46 @@ class OpProtoAndCheckerMaker {
   }
 
  protected:
-  void AddInput(const std::string& name, const std::string& comment,
-                bool multiple = false, bool ignore_gradient = false) {
+  struct VariableBuilder {
+    VarProto* var_;
+    std::function<void()> on_multiple_;
+    std::function<void()> on_temporary_;
+
+    VariableBuilder& SetMultiple() {
+      var_->set_multiple(true);
+      on_multiple_();
+      return *this;
+    }
+
+    VariableBuilder& SetTemporary() {
+      PADDLE_ENFORCE(bool(on_temporary_), "Cannot set temporary");
+      var_->set_temporary(true);
+      on_temporary_();
+      return *this;
+    }
+
+    VariableBuilder& IgnoreGradient() {
+      var_->set_ignore_gradient(true);
+      return *this;
+    }
+  };
+
+  VariableBuilder AddInput(const std::string& name,
+                           const std::string& comment) {
     auto input = proto_->mutable_inputs()->Add();
     *input->mutable_name() = name;
     *input->mutable_comment() = comment;
-    input->set_ignore_gradient(ignore_gradient);
-    input->set_multiple(multiple);
-    if (multiple) {
-      SetHasMultipleInput();
-    }
-  }
-
-  void AddInputs(const std::string& name, const std::string& comment,
-                 bool ignore_gradient = false) {
-    AddInput(name, comment, true, ignore_gradient);
+    return VariableBuilder{input, [=] { this->SetHasMultipleInput(); },
+                           nullptr};
   }
 
-  void AddOutput(const std::string& name, const std::string& comment,
-                 bool temporary = false, bool multiple = false,
-                 bool ignore_gradient = false) {
+  VariableBuilder AddOutput(const std::string& name,
+                            const std::string& comment) {
     auto output = proto_->mutable_outputs()->Add();
     *output->mutable_name() = name;
     *output->mutable_comment() = comment;
-    output->set_ignore_gradient(ignore_gradient);
-    output->set_multiple(multiple);
-    if (multiple) {
-      SetHasMultipleOutput();
-    }
-    output->set_temporary(temporary);
-    if (temporary) {
-      SetHasTemporaryOutput();
-    }
-  }
-
-  void AddOutputs(const std::string& name, const std::string& comment,
-                  bool temporary = false, bool ignore_gradient = false) {
-    AddOutput(name, comment, temporary, true, ignore_gradient);
+    return VariableBuilder{output, [=] { this->SetHasMultipleOutput(); },
+                           [=] { this->SetHasTemporaryOutput(); }};
   }
 
   template <typename T>
@@ -300,9 +303,10 @@ class OpRegistry {
     return CreateOp(op_desc.type(), inputs, outputs, attrs);
   }
 
-  static std::shared_ptr<OperatorBase> CreateGradOp(
-      std::shared_ptr<OperatorBase> op) {
-    GradOpBuilder builder(op.get());
+  static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op) {
+    PADDLE_ENFORCE(!op.IsNetOp(),
+                   "Use framework::Backward to get backward ops");
+    GradOpBuilder builder(op);
     std::shared_ptr<OperatorBase> grad_op(builder.Build());
     grad_op->Init();
     return grad_op;

paddle/framework/op_registry_test.cc

@@ -36,9 +36,8 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
   MyTestOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInputs("input", "input of cosine op");
-    AddOutput("output", "output of cosine op",
-              /*temporary*/ true);
+    AddInput("input", "input of cosine op").SetMultiple();
+    AddOutput("output", "output of cosine op").SetTemporary();
     auto my_checker = [](int i) {
       PADDLE_ENFORCE(i % 2 == 0, "'test_attr' must be even!");
     };

paddle/framework/operator.cc

@@ -52,7 +52,7 @@ std::vector<std::string> OperatorBase::Inputs(const std::string& name) const {
   PADDLE_ENFORCE(in_out_idxs_ != nullptr, "IO Idx could not be nullptr");
   auto input_format = GetAttr<std::vector<int>>("input_format");
   auto offset = in_out_idxs_->at(name);
-  PADDLE_ENFORCE(input_format.at((size_t)offset + 1) <= inputs_.size(),
+  PADDLE_ENFORCE(input_format.at((size_t)offset + 1) <= (int)inputs_.size(),
                  "Input Out Of Range");
 
   return std::vector<std::string>{
@@ -78,7 +78,7 @@ std::vector<std::string> OperatorBase::Outputs(const std::string& name) const {
   PADDLE_ENFORCE(in_out_idxs_ != nullptr, "InOut Indice could not be nullptr");
   auto output_format = GetAttr<std::vector<int>>("output_format");
   auto offset = in_out_idxs_->at(name);
-  PADDLE_ENFORCE(output_format.at((size_t)offset + 1) <= outputs_.size(),
+  PADDLE_ENFORCE(output_format.at((size_t)offset + 1) <= (int)outputs_.size(),
                  "Output Out of Range");
   return std::vector<std::string>{
       outputs_.begin() + output_format.at(offset),
@@ -105,5 +105,11 @@ std::string OperatorBase::DebugString() const {
   return ss.str();
 }
 
+void OperatorBase::Rename(const std::string& old_name,
+                          const std::string& new_name) {
+  std::replace(inputs_.begin(), inputs_.end(), old_name, new_name);
+  std::replace(outputs_.begin(), outputs_.end(), old_name, new_name);
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/operator.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <algorithm>
 #include <boost/variant.hpp>
 #include <string>
 #include <unordered_map>
@@ -54,6 +55,9 @@ class OperatorBase {
   /// e.g. Variable "x@GRAD" is the gradient of varibale "x".
   static std::string GRAD_VAR_SUFFIX() { return "@GRAD"; }
 
+  /// Variables with this suffix are supposed to be filled up with zeros.
+  static std::string ZERO_VAR_SUFFIX() { return "@ZERO"; }
+
   virtual ~OperatorBase() {}
 
   template <typename T>
@@ -79,8 +83,12 @@ class OperatorBase {
 
   virtual bool IsNetOp() const { return false; }
 
+  /// rename inputs outputs name
+  void Rename(const std::string& old_name, const std::string& new_name);
+
   //! Get a input with argument's name described in `op_proto`
   const std::string& Input(const std::string& name) const;
+
   //! Get a input which has multiple variables.
   //! TODO add a vector_view to prevent memory copy.
   std::vector<std::string> Inputs(const std::string& name) const;
@@ -92,7 +100,13 @@ class OperatorBase {
 
  public:
   std::string type_;
+  // NOTE: in case of OpGrad, inputs_ contains:
+  // I (Inputs)
+  // O (Outputs)
+  // OG (Output Gradients)
   std::vector<std::string> inputs_;
+  // NOTE: in case of OpGrad, outputs_ contains
+  // IG (Inputs Gradients)
   std::vector<std::string> outputs_;
   AttributeMap attrs_;
   // store the arguments' offset described in op_desc.

paddle/framework/operator_test.cc

@@ -137,9 +137,9 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker
   OpKernelTestMultiInputsProtoAndCheckerMaker(OpProto* proto,
                                               OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInputs("xs", "inputs of test op");
+    AddInput("xs", "inputs of test op").SetMultiple();
     AddInput("k", "input of test op");
-    AddOutputs("ys", "outputs of test op");
+    AddOutput("ys", "outputs of test op").SetMultiple();
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
         .LargerThan(0.0);

paddle/operators/CMakeLists.txt

@@ -49,6 +49,7 @@ op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc)
 op_library(sigmoid_op SRCS sigmoid_op.cu sigmoid_op.cc)
 op_library(softmax_op SRCS softmax_op.cc softmax_op.cu)
 op_library(cross_entropy_op SRCS cross_entropy_op.cc cross_entropy_op.cu)
+op_library(fill_zeros_like_op SRCS fill_zeros_like_op.cc fill_zeros_like_op.cu)
 
 op_library(fc_op SRCS fc_op.cc DEPS mul_op rowwise_add_op sigmoid_op
         softmax_op net)

paddle/operators/fc_op.cc

@@ -50,8 +50,8 @@ public:
     AddInput("b", "the bias of fc operator");
 
     AddOutput("Y", "the output of fc operator");
-    AddOutput(
-        "before_act", "the before activation output of fc operator", true);
+    AddOutput("before_act", "the before activation output of fc operator")
+        .SetTemporary();
     AddAttr<std::string>("activation", "The activation key for fc layer")
         .SetDefault("sigmoid")
         .InEnum({"sigmoid", "softmax"});

paddle/operators/fill_zeros_like_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/fill_zeros_like_op.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/tensor.h"
+
+namespace paddle {
+namespace operators {
+
+class FillZerosLikeOp : public framework::OperatorWithKernel {
+protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE(ctx.InputSize() == 1UL,
+                   "Input size of FillZerosLikeOp must be one.");
+    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
+                   "Output size of AddOp must be one.");
+    PADDLE_ENFORCE(ctx.InputVar(0) != nullptr,
+                   "Input of FillZerosLikeOp must be set.");
+    PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr,
+                   "Output of FillZerosLikeOp must be set.");
+    ctx.Output<framework::Tensor>(0)->Resize(
+        ctx.Input<framework::Tensor>(0)->dims());
+  }
+};
+
+class FillZerosLikeOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  FillZerosLikeOpMaker(framework::OpProto *proto,
+                       framework::OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Src", "The input of fill-zeros-like op.");
+    AddOutput("Dst", "The varibale will be filled up with zeros.");
+    AddComment(R"DOC(
+Fill up a vriable with zeros.
+
+The output will have the same size with input.
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP(fill_zeros_like,
+            paddle::operators::FillZerosLikeOp,
+            paddle::operators::FillZerosLikeOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    fill_zeros_like,
+    paddle::operators::FillZerosLikeKernel<paddle::platform::CPUPlace, float>);

paddle/operators/fill_zeros_like_op.cu

+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/fill_zeros_like_op.h"
+
+REGISTER_OP_GPU_KERNEL(
+    fill_zeros_like,
+    paddle::operators::FillZerosLikeKernel<paddle::platform::GPUPlace, float>);
\ No newline at end of file

paddle/operators/fill_zeros_like_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 "glog/logging.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/operator.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class FillZerosLikeKernel : public framework::OpKernel {
+public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* output = context.Output<framework::Tensor>(0);
+    output->mutable_data<T>(context.GetPlace());
+    framework::EigenVector<T>::Flatten(*output).setZero();
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/recurrent_network_op.cc

@@ -307,13 +307,14 @@ public:
       : OpProtoAndCheckerMaker(proto, op_checker) {
     const auto& name = RecurrentOp::kArgName;
     // inputs and outputs stored in proto
-    AddInputs(name.inlinks,
-              "the input that need to be segmented for each step.");
-    AddInputs(name.boot_memories, "variables to initialize memories.");
+    AddInput(name.inlinks, "the input that need to be segmented for each step.")
+        .SetMultiple();
+    AddInput(name.boot_memories, "variables to initialize memories.")
+        .SetMultiple();
     AddInput(name.step_net, "network shared by all steps.");
 
-    AddOutputs(name.outlinks,
-               "the output that need to concated for all steps.");
+    AddOutput(name.outlinks, "the output that need to concated for all steps.")
+        .SetMultiple();
     AddOutput(name.step_scopes, "step scopes");
 
     // Attributes stored in AttributeMap