dropout_op.h 3.7 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
X
Xinghai Sun 已提交
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
X
Xinghai Sun 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
X
Xinghai Sun 已提交
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. */
X
Xinghai Sun 已提交
14
#pragma once
Y
Yi Wang 已提交
15

16
#include <random>
Y
Yi Wang 已提交
17

Y
Yi Wang 已提交
18 19
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
X
Xinghai Sun 已提交
20 21 22 23 24 25 26

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
27
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
X
Xinghai Sun 已提交
28

K
Kexin Zhao 已提交
29
template <typename DeviceContext, typename T>
Y
Yu Yang 已提交
30
class CPUDropoutKernel : public framework::OpKernel<T> {
31 32 33 34
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* x = context.Input<Tensor>("X");
    auto* y = context.Output<Tensor>("Out");
35
    const auto* x_data = x->data<T>();
36
    auto* y_data = y->mutable_data<T>(context.GetPlace());
37
    float dropout_prob = context.Attr<float>("dropout_prob");
38

P
phlrain 已提交
39
    auto dropout_implementation = context.Attr<bool>("dropout_implementation");
40
    if (!context.Attr<bool>("is_test")) {
41 42
      auto* mask = context.Output<Tensor>("Mask");
      auto* mask_data = mask->mutable_data<T>(context.GetPlace());
43 44 45 46

      // NOTE: fixed seed should only be used in unittest or for debug.
      // Guarantee to use random seed in training.
      std::random_device rnd;
47
      std::minstd_rand engine;
48 49
      int seed =
          context.Attr<bool>("fix_seed") ? context.Attr<int>("seed") : rnd();
50
      engine.seed(seed);
51

52
      std::uniform_real_distribution<float> dist(0, 1);
P
phlrain 已提交
53

54 55 56 57 58 59
      size_t size = framework::product(mask->dims());
      for (size_t i = 0; i < size; ++i) {
        if (dist(engine) < dropout_prob) {
          mask_data[i] = 0;
          y_data[i] = 0;
        } else {
P
phlrain 已提交
60 61 62 63 64 65 66
          if (dropout_implementation) {
            mask_data[i] = 1.0f / static_cast<T>(1.0f - dropout_prob);
            y_data[i] = x_data[i] / static_cast<T>(1.0f - dropout_prob);
          } else {
            mask_data[i] = 1;
            y_data[i] = x_data[i];
          }
67
        }
68
      }
69
    } else {
70 71
      auto X = EigenMatrix<T>::Reshape(*x, 1);
      auto Y = EigenMatrix<T>::Reshape(*y, 1);
Q
QI JUN 已提交
72 73
      auto& place =
          *context.template device_context<DeviceContext>().eigen_device();
P
phlrain 已提交
74 75 76 77 78
      if (dropout_implementation) {
        Y.device(place) = X;
      } else {
        Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
      }
79 80 81 82
    }
  }
};

Q
QI JUN 已提交
83
template <typename DeviceContext, typename T>
Y
Yu Yang 已提交
84
class DropoutGradKernel : public framework::OpKernel<T> {
X
Xinghai Sun 已提交
85 86
 public:
  void Compute(const framework::ExecutionContext& context) const override {
87 88
    PADDLE_ENFORCE(!context.Attr<bool>("is_test"),
                   "GradOp is only callable when is_test is false");
89

X
Xinghai Sun 已提交
90 91 92 93 94
    auto* grad_x = context.Output<Tensor>(framework::GradVarName("X"));
    auto* grad_y = context.Input<Tensor>(framework::GradVarName("Out"));
    auto* mask = context.Input<Tensor>("Mask");
    grad_x->mutable_data<T>(context.GetPlace());

95 96 97
    auto M = EigenMatrix<T>::Reshape(*mask, 1);
    auto dX = EigenMatrix<T>::Reshape(*grad_x, 1);
    auto dY = EigenMatrix<T>::Reshape(*grad_y, 1);
X
Xinghai Sun 已提交
98

Q
QI JUN 已提交
99 100
    auto& place =
        *context.template device_context<DeviceContext>().eigen_device();
101
    dX.device(place) = dY * M;
X
Xinghai Sun 已提交
102 103 104 105 106
  }
};

}  // namespace operators
}  // namespace paddle