dropout_op.h 3.8 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>
P
phlrain 已提交
17
#include <string>
Y
Yi Wang 已提交
18

Y
Yi Wang 已提交
19 20
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
X
Xinghai Sun 已提交
21 22 23 24 25 26 27

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
28
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
X
Xinghai Sun 已提交
29

K
Kexin Zhao 已提交
30
template <typename DeviceContext, typename T>
Y
Yu Yang 已提交
31
class CPUDropoutKernel : public framework::OpKernel<T> {
32 33 34 35
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* x = context.Input<Tensor>("X");
    auto* y = context.Output<Tensor>("Out");
36
    const auto* x_data = x->data<T>();
37
    auto* y_data = y->mutable_data<T>(context.GetPlace());
38
    float dropout_prob = context.Attr<float>("dropout_prob");
39

P
phlrain 已提交
40 41
    auto dropout_implementation =
        context.Attr<std::string>("dropout_implementation");
42
    if (!context.Attr<bool>("is_test")) {
43 44
      auto* mask = context.Output<Tensor>("Mask");
      auto* mask_data = mask->mutable_data<T>(context.GetPlace());
45 46 47 48

      // NOTE: fixed seed should only be used in unittest or for debug.
      // Guarantee to use random seed in training.
      std::random_device rnd;
49
      std::minstd_rand engine;
50 51
      int seed =
          context.Attr<bool>("fix_seed") ? context.Attr<int>("seed") : rnd();
52
      engine.seed(seed);
53

54
      std::uniform_real_distribution<float> dist(0, 1);
P
phlrain 已提交
55

56 57 58 59 60 61
      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
62
          if (dropout_implementation == "upscale_in_train") {
P
phlrain 已提交
63 64 65 66 67 68
            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];
          }
69
        }
70
      }
71
    } else {
72 73
      auto X = EigenMatrix<T>::Reshape(*x, 1);
      auto Y = EigenMatrix<T>::Reshape(*y, 1);
Q
QI JUN 已提交
74 75
      auto& place =
          *context.template device_context<DeviceContext>().eigen_device();
P
phlrain 已提交
76
      if (dropout_implementation == "upscale_in_train") {
P
phlrain 已提交
77 78 79 80
        Y.device(place) = X;
      } else {
        Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
      }
81 82 83 84
    }
  }
};

Q
QI JUN 已提交
85
template <typename DeviceContext, typename T>
Y
Yu Yang 已提交
86
class DropoutGradKernel : public framework::OpKernel<T> {
X
Xinghai Sun 已提交
87 88
 public:
  void Compute(const framework::ExecutionContext& context) const override {
89 90
    PADDLE_ENFORCE(!context.Attr<bool>("is_test"),
                   "GradOp is only callable when is_test is false");
91

X
Xinghai Sun 已提交
92 93 94 95 96
    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());

97 98 99
    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
100

Q
QI JUN 已提交
101 102
    auto& place =
        *context.template device_context<DeviceContext>().eigen_device();
103
    dX.device(place) = dY * M;
X
Xinghai Sun 已提交
104 105 106 107 108
  }
};

}  // namespace operators
}  // namespace paddle