reduce_op.h 7.0 KB
Newer Older
G
guosheng 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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

D
Dong Zhihong 已提交
17
#include "glog/logging.h"
G
guosheng 已提交
18 19 20 21 22 23 24 25 26 27 28 29
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
using DDim = framework::DDim;
template <typename T, size_t D, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;

D
Dong Zhihong 已提交
30 31 32 33
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;

G
guosheng 已提交
34
struct SumFunctor {
35 36 37
  template <typename Place, typename X, typename Y, typename Dim>
  void operator()(const Place& place, X& x, Y& y, const Dim& dim) {
    y.device(place) = x.sum(dim);
G
guosheng 已提交
38 39 40 41
  }
};

struct SumGradFunctor {
42
  template <typename Place, typename X, typename Y, typename DX, typename DY,
G
guosheng 已提交
43
            typename Dim>
44 45 46
  void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy,
                  const Dim& dim, int size) {
    dx.device(place) = dy.broadcast(dim);
G
guosheng 已提交
47 48 49 50
  }
};

struct MeanFunctor {
51 52 53
  template <typename Place, typename X, typename Y, typename Dim>
  void operator()(const Place& place, X& x, Y& y, const Dim& dim) {
    y.device(place) = x.mean(dim);
G
guosheng 已提交
54 55 56 57
  }
};

struct MeanGradFunctor {
58
  template <typename Place, typename X, typename Y, typename DX, typename DY,
G
guosheng 已提交
59
            typename Dim>
60 61 62
  void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy,
                  const Dim& dim, int size) {
    dx.device(place) = dy.broadcast(dim) / dx.constant(size);
G
guosheng 已提交
63 64 65 66
  }
};

struct MaxFunctor {
67 68 69
  template <typename Place, typename X, typename Y, typename Dim>
  void operator()(const Place& place, X& x, Y& y, const Dim& dim) {
    y.device(place) = x.maximum(dim);
G
guosheng 已提交
70 71 72 73
  }
};

struct MinFunctor {
74 75 76
  template <typename Place, typename X, typename Y, typename Dim>
  void operator()(const Place& place, X& x, Y& y, const Dim& dim) {
    y.device(place) = x.minimum(dim);
G
guosheng 已提交
77 78 79 80
  }
};

struct MaxOrMinGradFunctor {
81
  template <typename Place, typename X, typename Y, typename DX, typename DY,
G
guosheng 已提交
82
            typename Dim>
83 84 85 86 87
  void operator()(const Place& place, X& x, Y& y, DX& dx, DY& dy,
                  const Dim& dim, int size) {
    auto equals = x == y.broadcast(dim);
    auto ones = dx.constant(1);
    auto zeros = dx.constant(0);
88 89
    // If there are multiple minimum or maximum elements, the subgradient of
    // each is the set [0, 1], and we pass gradient to all of them here.
90
    dx.device(place) = dy.broadcast(dim) * equals.select(ones, zeros);
G
guosheng 已提交
91 92 93 94
  }
};

template <typename Place, typename T, typename Functor>
Y
Yu Yang 已提交
95
class ReduceKernel : public framework::OpKernel<T> {
G
guosheng 已提交
96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    int rank = context.Input<Tensor>("X")->dims().size();
    switch (rank) {
      case 1:
        ReduceCompute<1>(context);
        break;
      case 2:
        ReduceCompute<2>(context);
        break;
      case 3:
        ReduceCompute<3>(context);
        break;
      case 4:
        ReduceCompute<4>(context);
        break;
      case 5:
        ReduceCompute<5>(context);
        break;
      case 6:
        ReduceCompute<6>(context);
        break;
    }
  }

 private:
  template <size_t D>
  void ReduceCompute(const framework::ExecutionContext& context) const {
    auto* input = context.Input<Tensor>("X");
    auto* output = context.Output<Tensor>("Out");
    output->mutable_data<T>(context.GetPlace());

    auto x = EigenTensor<T, D>::From(*input);
    auto x_rank = static_cast<int>(x.dimensions().size());
    int dim = static_cast<int>(context.Attr<int>("dim"));
    if (dim < 0) dim = x_rank + dim;
    auto reduce_dim = Eigen::array<int, 1>({{dim}});
    // construct the squeezed output tensor
134
    bool keep_dim = context.Attr<bool>("keep_dim");
G
guosheng 已提交
135 136 137 138 139 140
    DDim dims = output->dims();
    auto dims_vector = vectorize(dims);
    if (keep_dim && x_rank > 1) {
      dims_vector.erase(dims_vector.begin() + dim);
      dims = framework::make_ddim(dims_vector);
    }
D
Dong Zhihong 已提交
141

G
guosheng 已提交
142 143
    auto& place = context.GetEigenDevice<Place>();
    Functor functor;
D
Dong Zhihong 已提交
144 145 146 147 148 149 150 151

    if (D == 1) {
      auto out = EigenScalar<T>::From(*output);
      functor(place, x, out, reduce_dim);
    } else {
      auto out = EigenTensor<T, (D - 1)>::From(*output, dims);
      functor(place, x, out, reduce_dim);
    }
G
guosheng 已提交
152 153 154 155
  }
};

template <typename Place, typename T, typename Functor>
Y
Yu Yang 已提交
156
class ReduceGradKernel : public framework::OpKernel<T> {
G
guosheng 已提交
157 158 159 160 161
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    int rank = context.Input<Tensor>("X")->dims().size();
    switch (rank) {
      case 1:
162
        ReduceGradCompute<1>(context);
G
guosheng 已提交
163 164
        break;
      case 2:
165
        ReduceGradCompute<2>(context);
G
guosheng 已提交
166 167
        break;
      case 3:
168
        ReduceGradCompute<3>(context);
G
guosheng 已提交
169 170
        break;
      case 4:
171
        ReduceGradCompute<4>(context);
G
guosheng 已提交
172 173
        break;
      case 5:
174
        ReduceGradCompute<5>(context);
G
guosheng 已提交
175 176
        break;
      case 6:
177
        ReduceGradCompute<6>(context);
G
guosheng 已提交
178 179 180 181 182 183
        break;
    }
  }

 private:
  template <size_t D>
184
  void ReduceGradCompute(const framework::ExecutionContext& context) const {
G
guosheng 已提交
185 186 187 188 189
    auto* input0 = context.Input<Tensor>("X");
    auto* input1 = context.Input<Tensor>("Out");
    auto* input2 = context.Input<Tensor>(framework::GradVarName("Out"));
    auto* output = context.Output<Tensor>(framework::GradVarName("X"));

190 191 192 193 194 195 196 197 198 199 200
    output->mutable_data<T>(context.GetPlace());
    auto x = EigenTensor<T, D>::From(*input0);
    auto x_grad = EigenTensor<T, D>::From(*output);
    auto x_rank = static_cast<int>(x.dimensions().size());
    int dim = static_cast<int>(context.Attr<int>("dim"));
    if (dim < 0) dim = x_rank + dim;
    DDim dims = input0->dims();
    dims[dim] = 1;
    auto x_reduce = EigenTensor<T, D>::From(*input1, dims);
    auto x_reduce_grad = EigenTensor<T, D>::From(*input2, dims);

D
Dong Zhihong 已提交
201 202 203
    Eigen::array<int, D> broadcast_dim;
    for (size_t i = 0; i < D; ++i) broadcast_dim[i] = 1;
    broadcast_dim[dim] = input0->dims()[dim];
204 205
    auto& place = context.GetEigenDevice<Place>();
    Functor functor;
D
Dong Zhihong 已提交
206 207
    functor(place, x, x_reduce, x_grad, x_reduce_grad, broadcast_dim,
            broadcast_dim[dim]);
G
guosheng 已提交
208 209 210 211 212
  }
};

}  // namespace operators
}  // namespace paddle
213 214 215 216 217 218

#define FOR_EACH_KERNEL_FUNCTOR(__macro)                \
  __macro(reduce_sum, SumFunctor, SumGradFunctor);      \
  __macro(reduce_mean, MeanFunctor, MeanGradFunctor);   \
  __macro(reduce_max, MaxFunctor, MaxOrMinGradFunctor); \
  __macro(reduce_min, MinFunctor, MaxOrMinGradFunctor);