grad_op_desc_maker.h 6.5 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2

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
#include <algorithm>
M
minqiyang 已提交
17
#include <memory>
18
#include <string>
M
minqiyang 已提交
19
#include <unordered_map>
20
#include <unordered_set>
Y
Yu Yang 已提交
21
#include <vector>
Y
Yi Wang 已提交
22 23
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/operator.h"
24 25 26 27

namespace paddle {
namespace framework {

28 29 30 31 32 33 34 35
/*
  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.
 */
36 37
class GradOpDescMakerBase {
 public:
38
  explicit GradOpDescMakerBase(
Y
Yu Yang 已提交
39
      const OpDesc& fwd_op, const std::unordered_set<std::string>& no_grad_set,
Y
Yu Yang 已提交
40
      std::unordered_map<std::string, std::string>* grad_to_var,
Y
Yu Yang 已提交
41
      const std::vector<BlockDesc*>& grad_block = std::vector<BlockDesc*>())
Y
Yu Yang 已提交
42 43 44 45
      : fwd_op_(fwd_op),
        no_grad_set_(no_grad_set),
        grad_to_var_(grad_to_var),
        grad_block_(grad_block) {}
46 47

  virtual ~GradOpDescMakerBase() = default;
Y
Yu Yang 已提交
48
  virtual std::vector<std::unique_ptr<OpDesc>> operator()() const = 0;
49 50

 protected:
51 52
  std::vector<std::string> InputGrad(const std::string& name,
                                     bool drop_empty_grad = true) const {
53
    std::vector<std::string> ret_val;
54
    auto var_names = this->Input(name);
55
    ret_val.reserve(var_names.size());
56 57 58 59
    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);
M
minqiyang 已提交
60
                     if (no_grad_set_.empty() || !no_grad_set_.count(g_name)) {
M
minqiyang 已提交
61 62
                       (*this->grad_to_var_)[g_name] = fwd_var_name;
                       return g_name;
63
                     } else {
M
minqiyang 已提交
64
                       return kEmptyVarName;
65 66
                     }
                   });
67 68 69
    if (!drop_empty_grad) {
      return ret_val;
    }
70 71 72 73 74
    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"
75
                      " ambiguous."
76 77 78
                      " Op type %s",
                      fwd_op_.Type());

79 80 81 82 83 84
    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;
85 86 87
  }

  std::vector<std::string> OutputGrad(const std::string& name) const {
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),
92 93 94 95 96
                   [this](const std::string& fwd_var_name) -> std::string {
                     auto g_name = GradVarName(fwd_var_name);
                     (*this->grad_to_var_)[g_name] = fwd_var_name;
                     return g_name;
                   });
97
    return ret_val;
98 99
  }

Y
Yu Yang 已提交
100 101
  std::vector<std::string> InputNames() const {
    return this->fwd_op_.InputNames();
102 103
  }

Y
Yu Yang 已提交
104 105
  std::vector<std::string> OutputNames() const {
    return this->fwd_op_.OutputNames();
106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126
  }

  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;
  }

E
emailweixu 已提交
127 128 129 130 131
  template <typename T>
  inline const T& Attr(const std::string& name) const {
    return boost::get<T>(GetAttr(name));
  }

132 133
  std::string ForwardOpType() const { return this->fwd_op_.Type(); }

S
sneaxiy 已提交
134 135 136
 protected:
  const OpDesc& ForwardOp() const { return fwd_op_; }

137
 private:
Y
Yu Yang 已提交
138
  const OpDesc& fwd_op_;
139
  const std::unordered_set<std::string>& no_grad_set_;
140
  std::unordered_map<std::string, std::string>* grad_to_var_;
Y
Yu Yang 已提交
141 142

 protected:
Y
Yu Yang 已提交
143
  std::vector<BlockDesc*> grad_block_;
144 145 146 147
};

class SingleGradOpDescMaker : public GradOpDescMakerBase {
 public:
Y
Yu Yang 已提交
148 149
  using GradOpDescMakerBase::GradOpDescMakerBase;

S
sneaxiy 已提交
150
  std::vector<std::unique_ptr<OpDesc>> operator()() const final {
Y
Yu Yang 已提交
151
    std::vector<std::unique_ptr<OpDesc>> retv;
Y
Yu Yang 已提交
152 153 154
    retv.emplace_back(this->Apply());
    return retv;
  }
155 156

 protected:
Y
Yu Yang 已提交
157
  virtual std::unique_ptr<OpDesc> Apply() const = 0;
158 159
};

160
template <bool DropEmptyIG = true>
S
sneaxiy 已提交
161
class DefaultGradOpDescMaker final : public SingleGradOpDescMaker {
Y
Yu Yang 已提交
162 163 164
 public:
  using SingleGradOpDescMaker::SingleGradOpDescMaker;

165
 protected:
S
sneaxiy 已提交
166
  std::unique_ptr<OpDesc> Apply() const final {
Y
Yu Yang 已提交
167
    auto* grad = new OpDesc();
S
sneaxiy 已提交
168
    grad->SetType(this->ForwardOpType() + "_grad");
169

Y
Yu Yang 已提交
170
    for (auto& input_param : this->InputNames()) {
Y
Yu Yang 已提交
171
      grad->SetInput(input_param, this->Input(input_param));
172 173
      grad->SetOutput(GradVarName(input_param),
                      this->InputGrad(input_param, DropEmptyIG));
174 175
    }

Y
Yu Yang 已提交
176
    for (auto& output_param : this->OutputNames()) {
Y
Yu Yang 已提交
177 178
      grad->SetInput(output_param, this->Output(output_param));
      grad->SetInput(GradVarName(output_param), this->OutputGrad(output_param));
179 180
    }

Y
Yu Yang 已提交
181
    grad->SetAttrMap(this->Attrs());
182

Y
Yu Yang 已提交
183
    return std::unique_ptr<OpDesc>(grad);
184 185 186
  }
};

S
sneaxiy 已提交
187
class EmptyGradOpMaker final : public GradOpDescMakerBase {
Y
Yu Yang 已提交
188 189
 public:
  using GradOpDescMakerBase::GradOpDescMakerBase;
S
sneaxiy 已提交
190
  std::vector<std::unique_ptr<OpDesc>> operator()() const final { return {}; }
Y
Yu Yang 已提交
191 192
};

193 194
}  // namespace framework
}  // namespace paddle