sequence_pool_op.h 5.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
/* 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
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
L
Luo Tao 已提交
18
#include "paddle/operators/math/math_function.h"
19 20 21 22 23 24

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
25 26 27
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
28 29 30 31
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;

32 33 34 35 36 37 38 39 40
enum SeqPoolType {
  AVERAGE = 0,
  SUM = 1,
  SQRT = 2,  // square_root_n
  MAX = 3,
  LAST = 4,
  FIRST = 5
};

41
template <typename Place, typename T>
Y
Yu Yang 已提交
42
class SequencePoolKernel : public framework::OpKernel<T> {
43 44 45 46
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* in = context.Input<LoDTensor>("X");
    auto* out = context.Output<LoDTensor>("Out");
47
    int strategy = context.Attr<int>("strategy");
48 49

    auto dims = in->dims();
Q
Qiao Longfei 已提交
50
    auto lod = in->lod();
51 52
    int64_t w = in->numel() / dims[0];

Q
Qiao Longfei 已提交
53 54 55 56 57 58 59 60 61 62 63
    // InferShape by lod
    PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
    PADDLE_ENFORCE_GE(
        dims[0],
        /*batch size = */ static_cast<int64_t>(lod[0].size() - 1),
        "The first dimension of Input(X) must be large than batch size.");
    dims[0] = lod[0].size() - 1;
    out->Resize({dims});

    auto lod_level_0 = lod[0];

64 65
    out->mutable_data<T>(context.GetPlace());
    auto place = context.GetEigenDevice<Place>();
Q
Qiao Longfei 已提交
66 67 68
    for (int i = 0; i < static_cast<int>(lod_level_0.size()) - 1; ++i) {
      Tensor in_t = in->Slice<T>(static_cast<int>(lod_level_0[i]),
                                 static_cast<int>(lod_level_0[i + 1]));
69
      Tensor out_t = out->Slice<T>(i, i + 1);
Q
Qiao Longfei 已提交
70
      int64_t h = static_cast<int64_t>(lod_level_0[i + 1] - lod_level_0[i]);
71 72
      auto in_e = EigenMatrix<T>::From(in_t, framework::make_ddim({h, w}));
      auto out_e = EigenVector<T>::Flatten(out_t);
73 74 75 76 77 78 79 80

      switch (strategy) {
        case AVERAGE:
          out_e.device(place) = in_e.mean(Eigen::array<int, 1>({{0}}));
          break;
        case SUM:
          out_e.device(place) = in_e.sum(Eigen::array<int, 1>({{0}}));
          break;
L
Luo Tao 已提交
81 82 83 84
        case SQRT:
          out_e.device(place) = in_e.sum(Eigen::array<int, 1>({{0}})) /
                                std::sqrt(static_cast<T>(h));
          break;
L
Luo Tao 已提交
85 86 87 88 89 90
        case LAST:
          out_e.device(place) = in_e.chip(h - 1, 0);
          break;
        case FIRST:
          out_e.device(place) = in_e.chip(0, 0);
          break;
91
        default:
L
Luo Tao 已提交
92
          PADDLE_THROW("unsupported pooling strategy");
93
      }
94 95 96 97 98
    }
  }
};

template <typename Place, typename T>
Y
Yu Yang 已提交
99
class SequencePoolGradKernel : public framework::OpKernel<T> {
100 101
 public:
  void Compute(const framework::ExecutionContext& context) const override {
102
    auto* in = context.Input<LoDTensor>("X");
103
    auto* out_g = context.Input<LoDTensor>(framework::GradVarName("Out"));
104
    auto* in_g = context.Output<LoDTensor>(framework::GradVarName("X"));
105
    int strategy = context.Attr<int>("strategy");
106 107

    auto dims = in->dims();
108
    auto lod = in->lod()[0];
109 110 111
    int64_t w = in->numel() / dims[0];

    in_g->mutable_data<T>(context.GetPlace());
L
Luo Tao 已提交
112 113
    if (strategy == LAST || strategy == FIRST) {
      // set X@Grad be zero at first when strategy is LAST/FIRST
Q
qijun 已提交
114 115
      math::SetConstant<Place, T> functor;
      functor(context.device_context(), in_g, 0);
L
Luo Tao 已提交
116
    }
117
    auto place = context.GetEigenDevice<Place>();
118 119 120
    for (int i = 0; i < static_cast<int>(lod.size()) - 1; ++i) {
      auto in_g_t = in_g->Slice<T>(static_cast<int>(lod[i]),
                                   static_cast<int>(lod[i + 1]));
121
      auto out_g_t = out_g->Slice<T>(i, i + 1);
122
      int64_t h = static_cast<int64_t>(lod[i + 1] - lod[i]);
123 124
      auto in_g_e = EigenMatrix<T>::From(in_g_t, {h, w});
      auto out_g_e = EigenMatrix<T>::From(out_g_t, {1, w});
125
      Eigen::DSizes<int, 2> bcast(h, 1);
126 127 128 129 130 131 132 133

      switch (strategy) {
        case AVERAGE:
          in_g_e.device(place) = (out_g_e / static_cast<T>(h)).broadcast(bcast);
          break;
        case SUM:
          in_g_e.device(place) = (out_g_e).broadcast(bcast);
          break;
L
Luo Tao 已提交
134 135 136 137
        case SQRT:
          in_g_e.device(place) =
              (out_g_e / std::sqrt(static_cast<T>(h))).broadcast(bcast);
          break;
L
Luo Tao 已提交
138 139 140 141 142 143
        case LAST:
          in_g_e.chip(h - 1, 0).device(place) = out_g_e;
          break;
        case FIRST:
          in_g_e.chip(0, 0).device(place) = out_g_e;
          break;
144
        default:
L
Luo Tao 已提交
145
          PADDLE_THROW("unsupported pooling strategy");
146
      }
147 148 149 150 151 152
    }
  }
};

}  // namespace operators
}  // namespace paddle