dropout_op.h 3.0 KB
Newer Older
X
Xinghai Sun 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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
16
#include <random>
X
Xinghai Sun 已提交
17 18 19 20 21 22 23 24 25 26 27
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;

28
template <typename Place, typename T, typename AttrType>
Y
Yu Yang 已提交
29
class CPUDropoutKernel : public framework::OpKernel<T> {
30 31 32 33
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* x = context.Input<Tensor>("X");
    auto* y = context.Output<Tensor>("Out");
34
    const auto* x_data = x->data<T>();
35
    auto* y_data = y->mutable_data<T>(context.GetPlace());
36
    float dropout_prob = context.Attr<float>("dropout_prob");
37

38
    if (!context.Attr<bool>("is_test")) {
39 40
      auto* mask = context.Output<Tensor>("Mask");
      auto* mask_data = mask->mutable_data<T>(context.GetPlace());
41 42 43
      int seed = context.Attr<int>("seed");
      std::minstd_rand engine;
      engine.seed(seed);
44
      std::uniform_real_distribution<float> dist(0, 1);
45 46 47 48 49 50 51 52 53
      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 {
          mask_data[i] = 1;
          y_data[i] = x_data[i];
        }
54
      }
55 56 57 58 59
    } else {
      auto X = EigenMatrix<T>::Reshape(*x, 1);
      auto Y = EigenMatrix<T>::Reshape(*y, 1);
      auto place = context.GetEigenDevice<Place>();
      Y.device(place) = X * dropout_prob;
60 61 62 63
    }
  }
};

X
Xinghai Sun 已提交
64
template <typename Place, typename T>
Y
Yu Yang 已提交
65
class DropoutGradKernel : public framework::OpKernel<T> {
X
Xinghai Sun 已提交
66 67
 public:
  void Compute(const framework::ExecutionContext& context) const override {
68 69
    PADDLE_ENFORCE(!context.Attr<bool>("is_test"),
                   "GradOp is only callable when is_test is false");
70

X
Xinghai Sun 已提交
71 72 73 74 75
    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());

76 77 78
    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
79 80

    auto place = context.GetEigenDevice<Place>();
81
    dX.device(place) = dY * M;
X
Xinghai Sun 已提交
82 83 84 85 86
  }
};

}  // namespace operators
}  // namespace paddle