sequence_pool_op.h 5.6 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
#include "paddle/operators/math/sequence_pooling.h"
20 21 22 23 24 25

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
26 27 28
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
29 30 31 32 33
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;

template <typename Place, typename T>
Y
Yu Yang 已提交
34
class SequencePoolKernel : public framework::OpKernel<T> {
35 36 37
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* in = context.Input<LoDTensor>("X");
38
    auto* out = context.Output<Tensor>("Out");
D
dzhwinter 已提交
39
    std::string pooltype = context.Attr<std::string>("pooltype");
40 41

    auto dims = in->dims();
Q
Qiao Longfei 已提交
42
    auto lod = in->lod();
43 44
    int64_t w = in->numel() / dims[0];

Q
Qiao Longfei 已提交
45 46 47 48 49 50 51 52 53 54 55
    // 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];

56
    out->mutable_data<T>(context.GetPlace());
57 58 59 60 61 62 63 64 65 66

    if (pooltype == "MAX") {
      math::MaxSeqPoolFunctor<Place, T> max_pool;
      auto* index = context.Output<Tensor>("MaxIndex");
      index->Resize({dims});
      index->mutable_data<int>(context.GetPlace());
      max_pool(context.device_context(), *in, out, index);
      return;
    }

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

D
dzhwinter 已提交
76 77 78 79 80 81 82 83 84 85 86 87 88
      if (pooltype == "AVERAGE") {
        out_e.device(place) = in_e.mean(Eigen::array<int, 1>({{0}}));
      } else if (pooltype == "SUM") {
        out_e.device(place) = in_e.sum(Eigen::array<int, 1>({{0}}));
      } else if (pooltype == "SQRT") {
        out_e.device(place) = in_e.sum(Eigen::array<int, 1>({{0}})) /
                              std::sqrt(static_cast<T>(h));
      } else if (pooltype == "LAST") {
        out_e.device(place) = in_e.chip(h - 1, 0);
      } else if (pooltype == "FIRST") {
        out_e.device(place) = in_e.chip(0, 0);
      } else {
        PADDLE_THROW("unsupported pooling pooltype");
89
      }
90 91 92 93 94
    }
  }
};

template <typename Place, typename T>
Y
Yu Yang 已提交
95
class SequencePoolGradKernel : public framework::OpKernel<T> {
96 97
 public:
  void Compute(const framework::ExecutionContext& context) const override {
98
    auto* in = context.Input<LoDTensor>("X");
99
    auto* out_g = context.Input<Tensor>(framework::GradVarName("Out"));
100
    auto* in_g = context.Output<LoDTensor>(framework::GradVarName("X"));
D
dzhwinter 已提交
101
    std::string pooltype = context.Attr<std::string>("pooltype");
102 103

    auto dims = in->dims();
104
    auto lod = in->lod()[0];
105 106 107
    int64_t w = in->numel() / dims[0];

    in_g->mutable_data<T>(context.GetPlace());
108 109 110 111 112 113 114 115

    if (pooltype == "MAX") {
      math::MaxSeqPoolGradFunctor<Place, T> max_pool_grad;
      auto* index = context.Input<Tensor>("MaxIndex");
      max_pool_grad(context.device_context(), *out_g, *index, in_g);
      return;
    }

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

D
dzhwinter 已提交
131 132 133 134 135 136 137 138 139 140 141 142 143
      if (pooltype == "AVERAGE") {
        in_g_e.device(place) = (out_g_e / static_cast<T>(h)).broadcast(bcast);
      } else if (pooltype == "SUM") {
        in_g_e.device(place) = (out_g_e).broadcast(bcast);
      } else if (pooltype == "SQRT") {
        in_g_e.device(place) =
            (out_g_e / std::sqrt(static_cast<T>(h))).broadcast(bcast);
      } else if (pooltype == "LAST") {
        in_g_e.chip(h - 1, 0).device(place) = out_g_e;
      } else if (pooltype == "FIRST") {
        in_g_e.chip(0, 0).device(place) = out_g_e;
      } else {
        PADDLE_THROW("unsupported pooling pooltype");
144
      }
145 146 147 148 149 150
    }
  }
};

}  // namespace operators
}  // namespace paddle