max_pooling.hpp
Go to the documentation of this file.
1 
13 #ifndef MLPACK_METHODS_ANN_LAYER_MAX_POOLING_HPP
14 #define MLPACK_METHODS_ANN_LAYER_MAX_POOLING_HPP
15 
16 #include <mlpack/prereqs.hpp>
17 
18 namespace mlpack {
19 namespace ann {
20 
21 /*
22  * The max pooling rule for convolution neural networks. Take the maximum value
23  * within the receptive block.
24  */
26 {
27  public:
28  /*
29  * Return the maximum value within the receptive block.
30  *
31  * @param input Input used to perform the pooling operation.
32  */
33  template<typename MatType>
34  size_t Pooling(const MatType& input)
35  {
36  return arma::as_scalar(arma::find(input.max() == input, 1));
37  }
38 };
39 
48 template <
49  typename InputDataType = arma::mat,
50  typename OutputDataType = arma::mat
51 >
53 {
54  public:
56  MaxPooling();
57 
67  MaxPooling(const size_t kernelWidth,
68  const size_t kernelHeight,
69  const size_t strideWidth = 1,
70  const size_t strideHeight = 1,
71  const bool floor = true);
72 
80  template<typename eT>
81  void Forward(const arma::Mat<eT>& input, arma::Mat<eT>& output);
82 
92  template<typename eT>
93  void Backward(const arma::Mat<eT>& /* input */,
94  const arma::Mat<eT>& gy,
95  arma::Mat<eT>& g);
96 
98  const OutputDataType& OutputParameter() const { return outputParameter; }
100  OutputDataType& OutputParameter() { return outputParameter; }
101 
103  const OutputDataType& Delta() const { return delta; }
105  OutputDataType& Delta() { return delta; }
106 
108  size_t InputWidth() const { return inputWidth; }
110  size_t& InputWidth() { return inputWidth; }
111 
113  size_t InputHeight() const { return inputHeight; }
115  size_t& InputHeight() { return inputHeight; }
116 
118  size_t OutputWidth() const { return outputWidth; }
120  size_t& OutputWidth() { return outputWidth; }
121 
123  size_t OutputHeight() const { return outputHeight; }
125  size_t& OutputHeight() { return outputHeight; }
126 
128  size_t InputSize() const { return inSize; }
129 
131  size_t OutputSize() const { return outSize; }
132 
134  size_t KernelWidth() const { return kernelWidth; }
136  size_t& KernelWidth() { return kernelWidth; }
137 
139  size_t KernelHeight() const { return kernelHeight; }
141  size_t& KernelHeight() { return kernelHeight; }
142 
144  size_t StrideWidth() const { return strideWidth; }
146  size_t& StrideWidth() { return strideWidth; }
147 
149  size_t StrideHeight() const { return strideHeight; }
151  size_t& StrideHeight() { return strideHeight; }
152 
154  bool Floor() const { return floor; }
156  bool& Floor() { return floor; }
157 
159  bool Deterministic() const { return deterministic; }
161  bool& Deterministic() { return deterministic; }
162 
164  size_t WeightSize() const { return 0; }
165 
169  template<typename Archive>
170  void serialize(Archive& ar, const uint32_t /* version */);
171 
172  private:
180  template<typename eT>
181  void PoolingOperation(const arma::Mat<eT>& input,
182  arma::Mat<eT>& output,
183  arma::Mat<eT>& poolingIndices)
184  {
185  for (size_t j = 0, colidx = 0; j < output.n_cols;
186  ++j, colidx += strideHeight)
187  {
188  for (size_t i = 0, rowidx = 0; i < output.n_rows;
189  ++i, rowidx += strideWidth)
190  {
191  arma::mat subInput = input(
192  arma::span(rowidx, rowidx + kernelWidth - 1 - offset),
193  arma::span(colidx, colidx + kernelHeight - 1 - offset));
194 
195  const size_t idx = pooling.Pooling(subInput);
196  output(i, j) = subInput(idx);
197 
198  if (!deterministic)
199  {
200  arma::Mat<size_t> subIndices = indices(arma::span(rowidx,
201  rowidx + kernelWidth - 1 - offset),
202  arma::span(colidx, colidx + kernelHeight - 1 - offset));
203 
204  poolingIndices(i, j) = subIndices(idx);
205  }
206  }
207  }
208  }
209 
217  template<typename eT>
218  void Unpooling(const arma::Mat<eT>& error,
219  arma::Mat<eT>& output,
220  arma::Mat<eT>& poolingIndices)
221  {
222  for (size_t i = 0; i < poolingIndices.n_elem; ++i)
223  {
224  output(poolingIndices(i)) += error(i);
225  }
226  }
227 
229  size_t kernelWidth;
230 
232  size_t kernelHeight;
233 
235  size_t strideWidth;
236 
238  size_t strideHeight;
239 
241  bool floor;
242 
244  size_t inSize;
245 
247  size_t outSize;
248 
250  bool reset;
251 
253  size_t inputWidth;
254 
256  size_t inputHeight;
257 
259  size_t outputWidth;
260 
262  size_t outputHeight;
263 
265  bool deterministic;
266 
268  size_t offset;
269 
271  size_t batchSize;
272 
274  arma::cube outputTemp;
275 
277  arma::cube inputTemp;
278 
280  arma::cube gTemp;
281 
283  MaxPoolingRule pooling;
284 
286  OutputDataType delta;
287 
289  OutputDataType gradient;
290 
292  OutputDataType outputParameter;
293 
295  arma::Mat<size_t> indices;
296 
298  arma::Col<size_t> indicesCol;
299 
301  std::vector<arma::cube> poolingIndices;
302 }; // class MaxPooling
303 
304 } // namespace ann
305 } // namespace mlpack
306 
307 // Include implementation.
308 #include "max_pooling_impl.hpp"
309 
310 #endif
bool Floor() const
Get the value of the rounding operation.
size_t InputSize() const
Get the input size.
size_t KernelWidth() const
Get the kernel width.
size_t & InputWidth()
Modify the input width.
Linear algebra utility functions, generally performed on matrices or vectors.
size_t InputHeight() const
Get the input height.
The core includes that mlpack expects; standard C++ includes and Armadillo.
size_t & OutputWidth()
Modify the output width.
size_t OutputSize() const
Get the output size.
bool & Deterministic()
Modify the value of the deterministic parameter.
size_t & StrideWidth()
Modify the stride width.
size_t & StrideHeight()
Modify the stride height.
size_t & KernelHeight()
Modify the kernel height.
size_t OutputWidth() const
Get the output width.
size_t & InputHeight()
Modify the input height.
size_t & OutputHeight()
Modify the output height.
size_t Pooling(const MatType &input)
Definition: max_pooling.hpp:34
size_t StrideHeight() const
Get the stride height.
size_t KernelHeight() const
Get the kernel height.
size_t InputWidth() const
Get the input width.
const OutputDataType & Delta() const
Get the delta.
size_t OutputHeight() const
Get the output height.
size_t WeightSize() const
Get the size of the weights.
size_t StrideWidth() const
Get the stride width.
size_t & KernelWidth()
Modify the kernel width.
bool & Floor()
Modify the value of the rounding operation.
bool Deterministic() const
Get the value of the deterministic parameter.
Implementation of the MaxPooling layer.
Definition: max_pooling.hpp:52
OutputDataType & Delta()
Modify the delta.
const OutputDataType & OutputParameter() const
Get the output parameter.
Definition: max_pooling.hpp:98
OutputDataType & OutputParameter()
Modify the output parameter.