pool_op.h 11.1 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

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

17
#include <algorithm>
18 19
#include <string>
#include <vector>
Y
Yi Wang 已提交
20 21 22 23
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/pooling.h"
24 25 26 27
namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
28 29 30 31 32 33

class PoolOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override;
34 35 36 37

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override;
38 39 40 41

  framework::OpKernelType GetKernelTypeForVar(
      const std::string& var_name, const Tensor& tensor,
      const framework::OpKernelType& expected_kernel_type) const;
42 43 44 45 46 47 48
};

class PoolOpGrad : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override;
49 50 51 52

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override;
53 54 55 56
};

class Pool2dOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
57
  void Make() override;
58 59 60 61
};

class Pool3dOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
62
  void Make() override;
63
};
64 65 66

template <typename T = int>
inline void UpdatePadding(std::vector<T>* paddings, const bool global_pooling,
67 68 69
                          const bool adaptive,
                          const std::string padding_algorithm,
                          const framework::DDim data_dims,
70 71
                          const std::vector<T>& strides,
                          const std::vector<T>& ksize) {
72
  // set padding size == data_dims.size() * 2
73
  auto data_shape = framework::vectorize<T>(data_dims);
74 75
  if (static_cast<int>(paddings->size()) == data_dims.size()) {
    for (int i = 0; i < data_dims.size(); ++i) {
76
      T copy_pad = *(paddings->begin() + 2 * i);
77 78 79 80 81 82 83 84
      paddings->insert(paddings->begin() + 2 * i + 1, copy_pad);
    }
  } else {
    PADDLE_ENFORCE_EQ(
        data_dims.size() * 2, paddings->size(),
        "Paddings size should be the same or twice as the pooling size.");
  }

85
  // when padding_algorithm is "VALID" or "SAME"
86
  if (padding_algorithm == "SAME") {
87
    for (int i = 0; i < data_dims.size(); ++i) {
88 89
      T out_size = (data_dims[i] + strides[i] - 1) / strides[i];
      T pad_sum =
90 91
          std::max((out_size - 1) * strides[i] + ksize[i] - data_shape[i],
                   static_cast<T>(0));
92 93
      T pad_0 = pad_sum / 2;
      T pad_1 = pad_sum - pad_0;
94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110
      *(paddings->begin() + i * 2) = pad_0;
      *(paddings->begin() + i * 2 + 1) = pad_1;
    }
  } else if (padding_algorithm == "VALID") {
    for (auto it = paddings->begin(); it != paddings->end(); it++) {
      *it = 0;
    }
  }

  // if global_pooling == true or adaptive == true, padding will be ignore
  if (global_pooling || adaptive) {
    for (auto it = paddings->begin(); it != paddings->end(); it++) {
      *it = 0;
    }
  }
}

111 112
template <typename T = int>
inline void UpdateKsize(std::vector<T>* ksize,
113 114 115
                        const framework::DDim data_dims) {
  ksize->resize(static_cast<size_t>(data_dims.size()));
  for (size_t i = 0; i < ksize->size(); ++i) {
116
    *(ksize->begin() + i) = static_cast<T>(data_dims[i]);
117 118
  }
}
119

Q
QI JUN 已提交
120
template <typename DeviceContext, typename T>
C
chengduoZH 已提交
121
class PoolKernel : public framework::OpKernel<T> {
122 123
 public:
  void Compute(const framework::ExecutionContext& context) const override {
C
chengduoZH 已提交
124
    const Tensor* in_x = context.Input<Tensor>("X");
125
    Tensor* out = context.Output<Tensor>("Out");
126

C
chengduoZH 已提交
127
    std::string pooling_type = context.Attr<std::string>("pooling_type");
128 129 130
    std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
131
    std::string data_format = context.Attr<std::string>("data_format");
132
    bool exclusive = context.Attr<bool>("exclusive");
133
    bool adaptive = context.Attr<bool>("adaptive");
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153
    bool global_pooling = context.Attr<bool>("global_pooling");
    std::string padding_algorithm =
        context.Attr<std::string>("padding_algorithm");

    const bool channel_last = (data_format == "NHWC" || data_format == "NDHWC");

    // update paddings
    auto in_x_dims = in_x->dims();
    framework::DDim data_dims;
    if (channel_last) {
      data_dims = framework::slice_ddim(in_x_dims, 1, in_x_dims.size() - 1);
    } else {
      data_dims = framework::slice_ddim(in_x_dims, 2, in_x_dims.size());
    }

    UpdatePadding(&paddings, global_pooling, adaptive, padding_algorithm,
                  data_dims, strides, ksize);
    if (data_dims.size() * 2 == paddings.size()) {
      for (size_t i = 0; i < data_dims.size(); ++i) {
        paddings.erase(paddings.begin() + i + 1);
154 155
      }
    }
156 157 158 159 160

    if (global_pooling) {
      UpdateKsize(&ksize, data_dims);
    }

Q
QI JUN 已提交
161
    auto& dev_ctx = context.template device_context<DeviceContext>();
162 163 164
    switch (ksize.size()) {
      case 2: {
        if (pooling_type == "max") {
C
chengduoZH 已提交
165
          paddle::operators::math::Pool2dFunctor<
Q
QI JUN 已提交
166
              DeviceContext, paddle::operators::math::MaxPool<T>, T>
167
              pool2d_forward;
168
          paddle::operators::math::MaxPool<T> pool_process;
169 170
          pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, data_format,
                         pool_process, true, false, out);
171

C
chengduoZH 已提交
172
        } else if (pooling_type == "avg") {
C
chengduoZH 已提交
173
          paddle::operators::math::Pool2dFunctor<
Q
QI JUN 已提交
174
              DeviceContext, paddle::operators::math::AvgPool<T>, T>
175
              pool2d_forward;
176
          paddle::operators::math::AvgPool<T> pool_process;
177 178
          pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, data_format,
                         pool_process, exclusive, adaptive, out);
179 180 181 182
        }
      } break;
      case 3: {
        if (pooling_type == "max") {
C
chengduoZH 已提交
183
          paddle::operators::math::Pool3dFunctor<
Q
QI JUN 已提交
184
              DeviceContext, paddle::operators::math::MaxPool<T>, T>
185
              pool3d_forward;
186
          paddle::operators::math::MaxPool<T> pool_process;
187 188 189
          pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, data_format,
                         pool_process, true, false, out);

C
chengduoZH 已提交
190
        } else if (pooling_type == "avg") {
C
chengduoZH 已提交
191
          paddle::operators::math::Pool3dFunctor<
Q
QI JUN 已提交
192
              DeviceContext, paddle::operators::math::AvgPool<T>, T>
193
              pool3d_forward;
194
          paddle::operators::math::AvgPool<T> pool_process;
195 196
          pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, data_format,
                         pool_process, exclusive, adaptive, out);
197 198
        }
      } break;
C
fix bug  
chengduoZH 已提交
199
      default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }
200 201 202 203
    }
  }
};

Q
QI JUN 已提交
204
template <typename DeviceContext, typename T>
C
chengduoZH 已提交
205
class PoolGradKernel : public framework::OpKernel<T> {
206 207
 public:
  void Compute(const framework::ExecutionContext& context) const override {
C
chengduoZH 已提交
208
    const Tensor* in_x = context.Input<Tensor>("X");
209 210 211
    const Tensor* out = context.Input<Tensor>("Out");
    const Tensor* out_grad =
        context.Input<Tensor>(framework::GradVarName("Out"));
C
chengduoZH 已提交
212
    Tensor* in_x_grad = context.Output<Tensor>(framework::GradVarName("X"));
213

C
chengduoZH 已提交
214
    std::string pooling_type = context.Attr<std::string>("pooling_type");
215 216 217
    std::vector<int> ksize = context.Attr<std::vector<int>>("ksize");
    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
218
    bool exclusive = context.Attr<bool>("exclusive");
219
    bool adaptive = context.Attr<bool>("adaptive");
220 221 222 223 224 225
    std::string data_format = context.Attr<std::string>("data_format");
    bool global_pooling = context.Attr<bool>("global_pooling");
    std::string padding_algorithm =
        context.Attr<std::string>("padding_algorithm");

    const bool channel_last = (data_format == "NHWC" || data_format == "NDHWC");
226

227 228 229 230 231 232 233 234 235 236 237 238 239
    // update paddings
    auto in_x_dims = in_x->dims();
    framework::DDim data_dims;
    if (channel_last) {
      data_dims = framework::slice_ddim(in_x_dims, 1, in_x_dims.size() - 1);
    } else {
      data_dims = framework::slice_ddim(in_x_dims, 2, in_x_dims.size());
    }
    UpdatePadding(&paddings, global_pooling, adaptive, padding_algorithm,
                  data_dims, strides, ksize);
    if (data_dims.size() * 2 == paddings.size()) {
      for (size_t i = 0; i < data_dims.size(); ++i) {
        paddings.erase(paddings.begin() + i + 1);
C
fix bug  
chengduoZH 已提交
240
      }
241
    }
242 243 244 245 246

    if (global_pooling) {
      UpdateKsize(&ksize, data_dims);
    }

Q
QI JUN 已提交
247
    auto& dev_ctx = context.template device_context<DeviceContext>();
C
chengduoZH 已提交
248 249
    if (in_x_grad) {
      in_x_grad->mutable_data<T>(context.GetPlace());
Q
QI JUN 已提交
250 251
      paddle::operators::math::SetConstant<DeviceContext, T> set_constant;
      set_constant(dev_ctx, in_x_grad, 0.0);
252 253 254 255

      switch (ksize.size()) {
        case 2: {
          if (pooling_type == "max") {
Q
QI JUN 已提交
256
            paddle::operators::math::MaxPool2dGradFunctor<DeviceContext, T>
257
                pool2d_backward;
Q
QI JUN 已提交
258
            pool2d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides,
259
                            paddings, data_format, in_x_grad);
C
chengduoZH 已提交
260
          } else if (pooling_type == "avg") {
C
chengduoZH 已提交
261
            paddle::operators::math::Pool2dGradFunctor<
Q
QI JUN 已提交
262
                DeviceContext, paddle::operators::math::AvgPoolGrad<T>, T>
263
                pool2d_backward;
264
            paddle::operators::math::AvgPoolGrad<T> pool_process;
Q
QI JUN 已提交
265
            pool2d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides,
266 267
                            paddings, data_format, pool_process, exclusive,
                            adaptive, in_x_grad);
268 269 270 271
          }
        } break;
        case 3: {
          if (pooling_type == "max") {
Q
QI JUN 已提交
272
            paddle::operators::math::MaxPool3dGradFunctor<DeviceContext, T>
273
                pool3d_backward;
Q
QI JUN 已提交
274
            pool3d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides,
275
                            paddings, data_format, in_x_grad);
C
chengduoZH 已提交
276
          } else if (pooling_type == "avg") {
C
chengduoZH 已提交
277
            paddle::operators::math::Pool3dGradFunctor<
Q
QI JUN 已提交
278
                DeviceContext, paddle::operators::math::AvgPoolGrad<T>, T>
279
                pool3d_backward;
280
            paddle::operators::math::AvgPoolGrad<T> pool_process;
Q
QI JUN 已提交
281
            pool3d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides,
282 283
                            paddings, data_format, pool_process, exclusive,
                            adaptive, in_x_grad);
284 285
          }
        } break;
C
fix bug  
chengduoZH 已提交
286
        default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); }
287 288 289 290 291 292 293
      }
    }
  }
};

}  // namespace operators
}  // namespace paddle