pool_with_index_op.cc 12.2 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
C
chengduoZH 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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. */

Y
Yi Wang 已提交
15
#include "paddle/fluid/operators/pool_with_index_op.h"
C
chengduoZH 已提交
16 17 18 19

namespace paddle {
namespace operators {

Y
Yang Yang 已提交
20
inline int MaxPoolOutputSize(int input_size, int filter_size, int padding,
C
chengduoZH 已提交
21
                             int stride) {
C
chengduoZH 已提交
22 23 24 25 26 27 28 29
  int output_size = (input_size - filter_size + 2 * padding) / stride + 1;
  return output_size;
}

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

C
fix doc  
chengduoZH 已提交
30
  void InferShape(framework::InferShapeContext *ctx) const override {
C
chengduoZH 已提交
31
    PADDLE_ENFORCE(ctx->HasInput("X"),
32
                   "Input(X) of Pooling should not be null.");
C
chengduoZH 已提交
33
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
34
                   "Output(Out) of Pooling should not be null.");
C
chengduoZH 已提交
35
    PADDLE_ENFORCE(ctx->HasOutput("Mask"),
36
                   "Output(Mask) of Pooling should not be null.");
C
chengduoZH 已提交
37 38 39 40 41 42 43 44

    auto in_x_dims = ctx->GetInputDim("X");

    std::vector<int> ksize = ctx->Attrs().Get<std::vector<int>>("ksize");
    std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
    std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");

    PADDLE_ENFORCE(in_x_dims.size() == 4 || in_x_dims.size() == 5,
C
chengduoZH 已提交
45
                   "Pooling intput should be 4-D or 5-D tensor.");
C
chengduoZH 已提交
46

C
chengduoZH 已提交
47
    if (ctx->Attrs().Get<bool>("global_pooling")) {
C
chengduoZH 已提交
48
      ksize.resize(static_cast<size_t>(in_x_dims.size()) - 2);
C
fix bug  
chengduoZH 已提交
49 50
      for (size_t i = 0; i < ksize.size(); ++i) {
        paddings[i] = 0;
C
chengduoZH 已提交
51
        ksize[i] = static_cast<int>(in_x_dims[i + 2]);
C
fix bug  
chengduoZH 已提交
52
      }
C
chengduoZH 已提交
53 54 55
    }

    PADDLE_ENFORCE(in_x_dims.size() - ksize.size() == 2U,
C
fix doc  
chengduoZH 已提交
56
                   "Input size and pooling size should be consistent.");
C
chengduoZH 已提交
57
    PADDLE_ENFORCE_EQ(ksize.size(), strides.size(),
C
chengduoZH 已提交
58
                      "Strides size and pooling size should be the same.");
C
chengduoZH 已提交
59
    PADDLE_ENFORCE_EQ(ksize.size(), paddings.size(),
C
chengduoZH 已提交
60
                      "Paddings size and pooling size should be the same.");
C
chengduoZH 已提交
61 62 63

    std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1]});
    for (size_t i = 0; i < ksize.size(); ++i) {
Y
Yang Yang 已提交
64
      output_shape.push_back(MaxPoolOutputSize(in_x_dims[i + 2], ksize[i],
C
chengduoZH 已提交
65 66 67 68 69
                                               paddings[i], strides[i]));
    }
    ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
    ctx->SetOutputDim("Mask", framework::make_ddim(output_shape));
  }
70 71

 protected:
72
  framework::OpKernelType GetExpectedKernelType(
73 74 75 76 77
      const framework::ExecutionContext &ctx) const override {
    return framework::OpKernelType(
        framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
        ctx.device_context());
  }
C
chengduoZH 已提交
78 79 80 81 82 83
};

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

C
fix doc  
chengduoZH 已提交
84
  void InferShape(framework::InferShapeContext *ctx) const override {
85
    PADDLE_ENFORCE(ctx->HasInput("Mask"), "Input(Mask) must not be null.");
C
chengduoZH 已提交
86
    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
C
chengduoZH 已提交
87 88
    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
                   "Input(X@GRAD) should not be null.");
C
chengduoZH 已提交
89 90
    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
  }
91 92

 protected:
93
  framework::OpKernelType GetExpectedKernelType(
94 95 96 97 98
      const framework::ExecutionContext &ctx) const override {
    return framework::OpKernelType(
        framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
        ctx.device_context());
  }
C
chengduoZH 已提交
99 100 101 102
};

class MaxPool2dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
103
  void Make() override {
C
chengduoZH 已提交
104 105
    AddInput(
        "X",
106 107 108 109
        "(Tensor) The input tensor of pooling operator. "
        "The format of input tensor is NCHW, where N is batch size, C is the "
        "number of channels, H is the height of the image, "
        "and W is the width of the image.");
C
chengduoZH 已提交
110
    AddOutput("Out",
111 112 113 114 115
              "(Tensor) The output tensor of pooling operator. "
              "The format of output tensor is also NCHW, "
              "where N is batch size, C is "
              "the number of channels, H is the height of the image "
              "and W is the width of the image.");
C
chengduoZH 已提交
116
    AddOutput("Mask",
117 118 119 120 121 122
              "(Tensor) The Mask tensor of pooling operator."
              "The format of output tensor is also NCHW, "
              "where N is batch size, C is the number of channels, "
              "H is the height of the image, "
              "and W is the width of the image. "
              "It represents the index in the current feature map.");
C
chengduoZH 已提交
123

C
fix bug  
chengduoZH 已提交
124
    AddAttr<std::vector<int>>("ksize",
125 126
                              "(vector<int>) The pooling window size(height, "
                              "width) of pooling operator. "
C
chengduoZH 已提交
127
                              "If global_pooling = true, ksize and paddings "
C
fix bug  
chengduoZH 已提交
128 129
                              "will be ignored.");  // TODO(Chengduo): Add
                                                    // checker. (Currently,
C
fix doc  
chengduoZH 已提交
130
    // TypedAttrChecker don't support vector type.)
C
fix bug  
chengduoZH 已提交
131
    AddAttr<bool>(
C
chengduoZH 已提交
132
        "global_pooling",
133
        "(bool, default:false) Whether to use the global pooling. "
C
chengduoZH 已提交
134
        "If global_pooling = true, ksize and paddings will be ignored.")
C
chengduoZH 已提交
135
        .SetDefault(false);
136 137 138
    AddAttr<std::vector<int>>("strides",
                              "(vector<int>, default {1, 1}), strides(height, "
                              "width) of pooling operator.")
C
chengduoZH 已提交
139
        .SetDefault({1, 1});  // TODO(Chengduo): Add checker. (Currently,
C
fix doc  
chengduoZH 已提交
140
    // TypedAttrChecker don't support vector type.)
C
chengduoZH 已提交
141 142
    AddAttr<std::vector<int>>(
        "paddings",
143
        "(vector<int>, default:{0, 0}), paddings(height, width) of pooling "
144
        "operator. "
C
chengduoZH 已提交
145
        "If global_pooling = true, paddings and will be ignored.")
C
chengduoZH 已提交
146
        .SetDefault({0, 0});  // TODO(Chengduo): Add checker. (Currently,
C
fix doc  
chengduoZH 已提交
147
    // TypedAttrChecker don't support vector type.)
C
chengduoZH 已提交
148 149

    AddComment(R"DOC(
150 151
MaxPool2d Operator.

C
chengduoZH 已提交
152
The maxPooling2d with index operation calculates the output and the mask
153 154 155 156
based on the input, ksize, strides, and paddings parameters. Input(X) and
output(Out, Mask) are in NCHW format, where N is batch size, C is the
number of channels, H is the height of the feature, 
and W is the width of the feature.
C
chengduoZH 已提交
157 158
Parameters(ksize, strides, paddings) are two elements.
These two elements represent height and width, respectively.
C
chengduoZH 已提交
159 160 161 162
The input(X) size and output(Out, Mask) size may be different.

Example:
  Input:
163
       X shape: $(N, C, H_{in}, W_{in})$
C
chengduoZH 已提交
164
  Output:
165 166
       Out shape: $(N, C, H_{out}, W_{out})$
       Mask shape: $(N, C, H_{out}, W_{out})$
167
  Where
168
       $$
169 170
       H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\
       W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1
171 172
       $$

C
chengduoZH 已提交
173 174 175 176 177 178
)DOC");
  }
};

class MaxPool3dWithIndexOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
179
  void Make() override {
180 181 182 183 184 185
    AddInput("X",
             "(Tensor) The input tensor of pooling operator. "
             "The format of input tensor is NCDHW, where N is batch size, C is "
             "the number of channels, and D, H and W are the depth, height and "
             "width of "
             "the image, respectively");
C
chengduoZH 已提交
186
    AddOutput("Out",
187 188 189 190 191
              "(Tensor) The output tensor of pooling operator. "
              "The format of output tensor is also NCDHW, "
              "where N is the batch size, C is the number of channels, "
              "and D, H and W are the depth, height and "
              "width of the image, respectively.");
C
chengduoZH 已提交
192
    AddOutput("Mask",
193 194 195 196 197 198
              "(Tensor) The Mask tensor of pooling operator. "
              "The format of output tensor is also NCDHW, "
              "where N is the batch size, C is the number of channels, and "
              "D, H and W are the depth, height and width "
              "of the image, respectively. "
              "It represents the index in the current feature map.");
C
chengduoZH 已提交
199

C
fix bug  
chengduoZH 已提交
200
    AddAttr<std::vector<int>>("ksize",
201 202
                              "(vector<int>) The pooling window size(depth, "
                              "height, width) of pooling operator. "
C
chengduoZH 已提交
203
                              "If global_pooling = true, ksize and paddings "
C
fix bug  
chengduoZH 已提交
204 205
                              "will be ignored.");  // TODO(Chengduo): Add
                                                    // checker. (Currently,
C
fix doc  
chengduoZH 已提交
206
    // TypedAttrChecker don't support vector type.)
C
fix bug  
chengduoZH 已提交
207
    AddAttr<bool>(
C
chengduoZH 已提交
208
        "global_pooling",
209
        "(bool, default false) Whether to use the global pooling. "
C
chengduoZH 已提交
210
        "If global_pooling = true, ksize and paddings will be ignored.")
C
chengduoZH 已提交
211
        .SetDefault(false);
C
fix doc  
chengduoZH 已提交
212
    AddAttr<std::vector<int>>("strides",
213
                              "(vector<int>, default {1,1,1}), strides(depth, "
C
fix doc  
chengduoZH 已提交
214
                              "height, width) of pooling operator.")
C
chengduoZH 已提交
215
        .SetDefault({1, 1, 1});  // TODO(Chengduo): Add checker. (Currently,
C
fix doc  
chengduoZH 已提交
216
    // TypedAttrChecker don't support vector type.)
C
fix bug  
chengduoZH 已提交
217 218
    AddAttr<std::vector<int>>(
        "paddings",
219
        "(vector, default {0,0,0}), paddings(depth, "
220
        "height, width) of pooling operator. "
C
chengduoZH 已提交
221
        "If global_pooling = true, paddings and ksize will be ignored.")
C
chengduoZH 已提交
222
        .SetDefault({0, 0, 0});  // TODO(Chengduo): Add checker. (Currently,
C
fix doc  
chengduoZH 已提交
223
    // TypedAttrChecker don't support vector type.)
224

C
chengduoZH 已提交
225
    AddComment(R"DOC(
226 227
MaxPool3d Operator.

C
chengduoZH 已提交
228 229
The maxpooling3d with index operation calculates the output and the mask
based on the input and ksize, strides, paddings parameters.
230 231 232 233
Input(X) and output(Out, Mask) are in NCDHW format, where N is batch
size, C is the number of channels, and D, H and W are the depth, height and
width of the feature, respectively. 
Parameters(ksize, strides, paddings) are three elements.
C
chengduoZH 已提交
234
These three elements represent depth, height and width, respectively.
C
chengduoZH 已提交
235 236 237 238
The input(X) size and output(Out, Mask) size may be different.

Example:
  Input:
239
       X shape: $(N, C, D_{in}, H_{in}, W_{in})$
C
chengduoZH 已提交
240
  Output:
241 242
       Out shape: $(N, C, D_{out}, H_{out}, W_{out})$
       Mask shape: $(N, C, D_{out}, H_{out}, W_{out})$
243
  Where
244
       $$
245 246 247
       D_{out} = \frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\
       H_{out} = \frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 \\
       W_{out} = \frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1
248 249
       $$

C
chengduoZH 已提交
250 251 252
)DOC");
  }
};
C
chengduoZH 已提交
253

C
chengduoZH 已提交
254 255 256 257 258
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

Y
Yang Yang 已提交
259 260
REGISTER_OPERATOR(max_pool2d_with_index, ops::MaxPoolWithIndexOp,
                  ops::MaxPool2dWithIndexOpMaker,
261 262
                  paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(max_pool2d_with_index_grad, ops::MaxPoolWithIndexOpGrad);
C
chengduoZH 已提交
263 264

REGISTER_OP_CPU_KERNEL(
C
chengduoZH 已提交
265
    max_pool2d_with_index,
Q
QI JUN 已提交
266 267 268
    ops::MaxPoolWithIndexKernel<paddle::platform::CPUDeviceContext, float, int>,
    ops::MaxPoolWithIndexKernel<paddle::platform::CPUDeviceContext, double,
                                int>);
C
chengduoZH 已提交
269
REGISTER_OP_CPU_KERNEL(
C
chengduoZH 已提交
270
    max_pool2d_with_index_grad,
Q
QI JUN 已提交
271 272 273
    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUDeviceContext, float,
                                    int>,
    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUDeviceContext, double,
274
                                    int>);
C
chengduoZH 已提交
275

Y
Yang Yang 已提交
276 277
REGISTER_OPERATOR(max_pool3d_with_index, ops::MaxPoolWithIndexOp,
                  ops::MaxPool3dWithIndexOpMaker,
278 279
                  paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(max_pool3d_with_index_grad, ops::MaxPoolWithIndexOpGrad);
C
chengduoZH 已提交
280 281

REGISTER_OP_CPU_KERNEL(
C
chengduoZH 已提交
282
    max_pool3d_with_index,
Q
QI JUN 已提交
283 284 285
    ops::MaxPoolWithIndexKernel<paddle::platform::CPUDeviceContext, float, int>,
    ops::MaxPoolWithIndexKernel<paddle::platform::CPUDeviceContext, double,
                                int>);
C
chengduoZH 已提交
286
REGISTER_OP_CPU_KERNEL(
C
chengduoZH 已提交
287
    max_pool3d_with_index_grad,
Q
QI JUN 已提交
288 289 290
    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUDeviceContext, float,
                                    int>,
    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUDeviceContext, double,
291
                                    int>);