grad_op_desc_maker.h 6.2 KB
Newer Older
1 2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

L
Luo Tao 已提交
3 4 5
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
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
8

L
Luo Tao 已提交
9 10 11 12 13
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. */
14 15

#pragma once
16 17
#include <string>
#include <unordered_set>
Y
Yu Yang 已提交
18
#include <vector>
19 20 21 22 23 24
#include "paddle/framework/op_desc.h"
#include "paddle/framework/operator.h"

namespace paddle {
namespace framework {

25 26 27 28 29 30 31 32
/*
  This functor class is responsible for creating the gradient ops for the given
  operator fwd_op. After it is called (through operator()), the pairs of
  (gradient variable, corresponding input variable of fwd_op) will be added to
  grad_to_var. If an input variable of fwd_op is contained in no_grad_set, its
  gradient varialbe will be ignored or kEmptyVarName depending on the template
  argument DropEmptyIG in the derived classes.
 */
33 34
class GradOpDescMakerBase {
 public:
35
  explicit GradOpDescMakerBase(
Y
Yu Yang 已提交
36
      const OpDesc& fwd_op, const std::unordered_set<std::string>& no_grad_set,
Y
Yu Yang 已提交
37
      std::unordered_map<std::string, std::string>* grad_to_var,
Y
Yu Yang 已提交
38
      const std::vector<BlockDesc*>& grad_block = std::vector<BlockDesc*>())
Y
Yu Yang 已提交
39 40 41 42
      : fwd_op_(fwd_op),
        no_grad_set_(no_grad_set),
        grad_to_var_(grad_to_var),
        grad_block_(grad_block) {}
43 44

  virtual ~GradOpDescMakerBase() = default;
Y
Yu Yang 已提交
45
  virtual std::vector<std::unique_ptr<OpDesc>> operator()() const = 0;
46 47

 protected:
48 49
  std::vector<std::string> InputGrad(const std::string& name,
                                     bool drop_empty_grad = true) const {
50
    std::vector<std::string> ret_val;
51
    auto var_names = this->Input(name);
52
    ret_val.reserve(var_names.size());
53 54 55 56 57 58 59 60 61 62 63
    std::transform(var_names.begin(), var_names.end(),
                   std::back_inserter(ret_val),
                   [this](const std::string& fwd_var_name) -> std::string {
                     auto g_name = GradVarName(fwd_var_name);
                     if (no_grad_set_.count(g_name)) {
                       return kEmptyVarName;
                     } else {
                       (*this->grad_to_var_)[g_name] = fwd_var_name;
                       return g_name;
                     }
                   });
64 65 66
    if (!drop_empty_grad) {
      return ret_val;
    }
67 68 69 70 71 72 73 74 75 76
    PADDLE_ENFORCE_LE(var_names.size(), 1UL,
                      "BUG from operator developer:"
                      " for input argument with a list of variables, "
                      " drop_empty_grad is not allowed because it makes"
                      " the correspondence bewteen a variable and its gradient"
                      " ambiguous. Use REGISTER_OP_EX to register the op"
                      " or call InputGrad(?,false) in GradOpDescMaker."
                      " Op type %s",
                      fwd_op_.Type());

77 78 79 80 81 82
    std::vector<std::string> dropped_ret_val;
    dropped_ret_val.reserve(ret_val.size());
    std::copy_if(ret_val.begin(), ret_val.end(),
                 std::back_inserter(dropped_ret_val),
                 [](const std::string& str) { return str != kEmptyVarName; });
    return dropped_ret_val;
83 84 85
  }

  std::vector<std::string> OutputGrad(const std::string& name) const {
86 87 88 89 90 91
    std::vector<std::string> ret_val;
    auto onames = this->Output(name);
    ret_val.reserve(onames.size());
    std::transform(onames.begin(), onames.end(), std::back_inserter(ret_val),
                   GradVarName);
    return ret_val;
92 93
  }

Y
Yu Yang 已提交
94 95
  std::vector<std::string> InputNames() const {
    return this->fwd_op_.InputNames();
96 97
  }

Y
Yu Yang 已提交
98 99
  std::vector<std::string> OutputNames() const {
    return this->fwd_op_.OutputNames();
100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123
  }

  std::vector<std::string> Input(const std::string& name) const {
    return fwd_op_.Input(name);
  }

  std::vector<std::string> Output(const std::string& name) const {
    return fwd_op_.Output(name);
  }

  const std::unordered_map<std::string, Attribute>& Attrs() const {
    return fwd_op_.GetAttrMap();
  }

  const Attribute& GetAttr(const std::string& name) const {
    auto& map = fwd_op_.GetAttrMap();
    auto it = map.find(name);
    PADDLE_ENFORCE(it != map.end(), "Cannot find attribute %s", name);
    return it->second;
  }

  std::string ForwardOpType() const { return this->fwd_op_.Type(); }

 private:
Y
Yu Yang 已提交
124
  const OpDesc& fwd_op_;
125
  const std::unordered_set<std::string>& no_grad_set_;
126
  std::unordered_map<std::string, std::string>* grad_to_var_;
Y
Yu Yang 已提交
127 128

 protected:
Y
Yu Yang 已提交
129
  std::vector<BlockDesc*> grad_block_;
130 131 132 133
};

class SingleGradOpDescMaker : public GradOpDescMakerBase {
 public:
Y
Yu Yang 已提交
134 135
  using GradOpDescMakerBase::GradOpDescMakerBase;

Y
Yu Yang 已提交
136 137
  std::vector<std::unique_ptr<OpDesc>> operator()() const {
    std::vector<std::unique_ptr<OpDesc>> retv;
Y
Yu Yang 已提交
138 139 140
    retv.emplace_back(this->Apply());
    return retv;
  }
141 142

 protected:
Y
Yu Yang 已提交
143
  virtual std::unique_ptr<OpDesc> Apply() const = 0;
144 145
};

146
template <bool DropEmptyIG = true>
147
class DefaultGradOpDescMaker : public SingleGradOpDescMaker {
Y
Yu Yang 已提交
148 149 150
 public:
  using SingleGradOpDescMaker::SingleGradOpDescMaker;

151
 protected:
Y
Yu Yang 已提交
152 153
  virtual std::unique_ptr<OpDesc> Apply() const {
    auto* grad = new OpDesc();
Y
Yu Yang 已提交
154
    grad->SetType(this->GradOpType());
155

Y
Yu Yang 已提交
156
    for (auto& input_param : this->InputNames()) {
Y
Yu Yang 已提交
157
      grad->SetInput(input_param, this->Input(input_param));
158 159
      grad->SetOutput(GradVarName(input_param),
                      this->InputGrad(input_param, DropEmptyIG));
160 161
    }

Y
Yu Yang 已提交
162
    for (auto& output_param : this->OutputNames()) {
Y
Yu Yang 已提交
163 164
      grad->SetInput(output_param, this->Output(output_param));
      grad->SetInput(GradVarName(output_param), this->OutputGrad(output_param));
165 166
    }

Y
Yu Yang 已提交
167
    grad->SetAttrMap(this->Attrs());
168

Y
Yu Yang 已提交
169
    return std::unique_ptr<OpDesc>(grad);
170 171 172 173 174 175 176
  }

  virtual std::string GradOpType() const {
    return this->ForwardOpType() + "_grad";
  }
};

Y
Yu Yang 已提交
177 178 179
class EmptyGradOpMaker : public GradOpDescMakerBase {
 public:
  using GradOpDescMakerBase::GradOpDescMakerBase;
Y
Yu Yang 已提交
180
  std::vector<std::unique_ptr<OpDesc>> operator()() const override {
Y
Yu Yang 已提交
181 182 183 184
    return {};
  }
};

185 186
}  // namespace framework
}  // namespace paddle