source: trunk/yat/classifier/KernelLookup.h @ 1133

Last change on this file since 1133 was 1133, checked in by Peter, 14 years ago

using SmartPtr? in KernelLookup? rather than doing the ref count myself

  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date ID
File size: 11.4 KB
Line 
1#ifndef _theplu_yat_classifier_kernel_lookup_
2#define _theplu_yat_classifier_kernel_lookup_
3
4// $Id$
5
6/*
7  Copyright (C) 2005 Jari Häkkinen, Peter Johansson
8  Copyright (C) 2006 Jari Häkkinen, Markus Ringnér, Peter Johansson
9  Copyright (C) 2007, 2008 Peter Johansson
10
11  This file is part of the yat library, http://trac.thep.lu.se/yat
12
13  The yat library is free software; you can redistribute it and/or
14  modify it under the terms of the GNU General Public License as
15  published by the Free Software Foundation; either version 2 of the
16  License, or (at your option) any later version.
17
18  The yat library is distributed in the hope that it will be useful,
19  but WITHOUT ANY WARRANTY; without even the implied warranty of
20  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
21  General Public License for more details.
22
23  You should have received a copy of the GNU General Public License
24  along with this program; if not, write to the Free Software
25  Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA
26  02111-1307, USA.
27*/
28
29#include "Kernel.h"
30#include "yat/utility/Container2DIterator.h"
31#include "yat/utility/iterator_traits.h"
32#include "yat/utility/SmartPtr.h"
33#include "yat/utility/StrideIterator.h"
34
35#include <vector>
36
37namespace theplu {
38namespace yat {
39namespace classifier {
40
41  class DataLookup2D;
42  class KernelFunction;
43  class MatrixLookup;
44  class MatrixLookupWeighted;
45
46  ///
47  /// @brief Lookup into Kernel
48  ///
49  /// This is the KernelLookup class to be used together with kernel
50  /// methods such as Support Vector Machines (SVM). The class does
51  /// not contain any data or values, but rather is a lookup into a
52  /// Kernel object. Each row and each column corresponds to a row and
53  /// a column in the Kernel, respectively. This design allow for fast
54  /// creation of sub-kernels, which is a common operation in most
55  /// traning/validation procedures.
56  ///
57  /// A KernelLookup can be created directly from a Kernel or from an
58  /// other KernelLookup. In the latter case, the resulting
59  /// KernelLookup is looking directly into the underlying Kernel to
60  /// avoid multiple lookups.
61  ///
62  /// There is a possibility to set the KernelLookup as owner of the
63  /// underlying kernel. This implies that underlying kernel is deleted
64  /// in destructor of MatrixLookup, but only if there is no other
65  /// owner of the underlying kernel. A reference counter is used to
66  /// keep track of number of owners. Ownership is copied in copy
67  /// constructors and assignments.
68  ///
69  class KernelLookup
70  {
71
72  public:
73    /// 'Read Only' iterator
74    typedef utility::StrideIterator<
75    utility::Container2DIterator<const KernelLookup, const double, void, 
76                                 const double> >
77    const_iterator;
78
79    /**
80       'Read only' iterator intended to iterate over a column
81     */
82    typedef const_iterator const_column_iterator;
83
84    /**
85       'Read only' iterator intended to iterate over a row
86     */
87    typedef const_iterator const_row_iterator;
88
89    ///
90    /// @brief Constructor a Lookup into a Kernel
91    ///
92    /// Constructs a KernelLookup corresponding to the Kernel @a
93    /// kernel. By default @a owner is set to false, which means
94    /// KernelLookup does not own the underlying Kernel.
95    ///
96    /// @note If underlying Kernel goes out of scope or is deleted, the
97    /// KernelLookup becomes invalid and the result of further use is
98    /// undefined.
99    ///
100    /// @note Do not construct two KernelLookups from the same @a
101    /// kernel with @a owner set to true because that will cause
102    /// multiple deletion of @a kernel.
103    ///
104    KernelLookup(const Kernel& kernel, const bool owner=false);
105
106    ///
107    /// @brief Constructing a Lookup into a subKernel
108    ///
109    /// Creating a Lookup into parts of the Kernel. In the created
110    /// Lookup the element in the \f$ i \f$ th row in the \f$ j \f$ th
111    /// column is identical to the element in row row[i] and columns
112    /// column[j] in the underlying @a kernel. If @a owner is set to
113    /// true yhe underlying @a kernel is destroyed in the destructor.
114    ///
115    /// @note If @a kernel goes out of scope or is deleted, the
116    /// returned pointer becomes invalid and the result of further use is
117    /// undefined.
118    ///
119    /// @note For training usage row index shall always be equal to
120    /// column index.
121    ///
122    KernelLookup(const Kernel& kernel, const std::vector<size_t>& row, 
123                 const std::vector<size_t>& column, const bool owner=false);
124
125    ///
126    /// @brief Copy constructor.
127    ///
128    /// A Lookup is created looking into the
129    /// same underlying Kernel as @a kl is looking into.
130    ///
131    /// If \a kl is owner of underlying data, constructed
132    /// KernelLookup will also be set as owner of underlying data.
133    ///
134    KernelLookup(const KernelLookup& kl);
135
136
137    ///
138    /// @brief Contructing a sub-KernelLookup.
139    ///
140    /// Contructor building a sub-KernelLookup from a KernelLookup
141    /// defined by row index vector and column index vector. In the
142    /// created Lookup the element in the \f$ i \f$ th row in the
143    /// \f$ j \f$ th column is identical to the element in row row[i] and
144    /// columns column[j] in the copied @a kl. The resulting
145    /// KernelLookup is independent of the old KernelLookup, but is
146    /// undefined in case underlying Kernel is destroyed.
147    ///
148    /// If \a kl is owner of underlying data, constructed
149    /// KernelLookup will also be set as owner of underlying data.
150    ///
151    /// @note For training usage row index shall always be equal to
152    /// column index.
153    ///
154    KernelLookup(const KernelLookup& kl, const std::vector<size_t>& row, 
155                 const std::vector<size_t>& column);
156
157    ///
158    /// Constructor taking the column (default) or row index vector as
159    /// input. If @a row is false the created KernelLookup will have
160    /// equally many rows as @a kernel.
161    ///
162    /// If \a kl is owner of underlying data, constructed
163    /// KernelLookup will also be set as owner of underlying data.
164    ///
165    /// @note If underlying kernel goes out of scope or is deleted, the
166    /// KernelLookup becomes invalid and the result of further use is
167    /// undefined.
168    ///
169    KernelLookup(const KernelLookup& kernel, const std::vector<size_t>&, 
170                 const bool row=false);
171
172    ///
173    /// @brief Destructor
174    ///
175    /// Deletes underlying Kernel if KernelLookup owns it and there is
176    /// no other owner.
177    ///
178    virtual ~KernelLookup(void);
179
180    /**
181       Iterator iterates along a row. When end of row is reached it
182       jumps to beginning of next row.
183
184       \return const_iterator pointing to upper-left element.
185     */
186    const_iterator begin(void) const;
187
188    /**
189       Iterator iterates along a column.
190
191       \return iterator pointing to first element of column \a i.
192     */
193    const_column_iterator begin_column(size_t) const;
194
195    /**
196       Iterator iterates along a column.
197
198       \return const_iterator pointing to first element of column \a i.
199     */
200    const_row_iterator begin_row(size_t) const;
201
202    /**
203       \return number of columns
204    */
205    size_t columns(void) const;
206
207    ///
208    /// Each column in returned DataLookup corresponds to the column
209    /// in KernelLookup.
210    ///
211    /// \return data that KernelLookup is built upon.
212    ///
213    /// @note Returns a dynamically allocated MatrixLookup, which has
214    /// to be deleted by the caller to avoid memory leaks.
215    ///
216    const DataLookup2D* data(void) const;
217
218    /**
219       Function to calculate a new Kernel element using the underlying
220       KernelFunction. The value is calculated between @a vec and the
221       data vector of the \a i th sample, in other words, the
222       sample corresponding to the \a i th row.
223    */
224    double element(const DataLookup1D& vec, size_t i) const;
225
226    /**
227       Function to calculate a new Kernel element using the underlying
228       KernelFunction. The value is calulated between @a vec and the
229       data vector of the \f$ i \f$ th sample, in other words, the
230       sample corresponding to the \f$ i \f$ th row or \f$ i \f$ th
231       column. In case KernelLookup is a sub-Kernel and not symmetric,
232       the kernel value is calculated between @a vec and the data
233       vector corresponding to \f$ i \f$ th row.
234    */
235    double element(const DataLookupWeighted1D& vec, size_t i) const;
236
237    /**
238       \return const_iterator pointing to end of matrix
239     */
240    const_iterator end(void) const;
241
242    /**
243       \return const_iterator pointing to end of column \a i
244     */
245    const_column_iterator end_column(size_t) const;
246
247    /**
248       \return const_iterator pointing to end of row \a i
249     */
250    const_row_iterator end_row(size_t) const;
251
252    /**
253       \return number of rows
254    */
255    size_t rows(void) const;
256
257    /**
258       Each element in returned KernelLookup is calculated using only
259       selected features (defined by @a index). Each element
260       corresponds to the same pair of samples as in the original
261       KernelLookup.
262
263       \note Returns a dynamically allocated KernelLookup, which has
264       to be deleted by the caller to avoid memory leaks.
265    */
266    const KernelLookup* selected(const std::vector<size_t>& index) const;
267   
268    /**
269       This function is useful when predicting on an independent data
270       set using a kernel-based classifier. In returned KernelLookup
271       column \f$ i \f$ corresponds to column \f$ i \f$ in @a
272       data. Row \f$ i \f$ in returned KernelLookup corresponds to
273       same sample as row \f$ i \f$ in @a this. In other words, this
274       function returns a KernelLookup containing the kernel elements
275       between the passed @a data and the internal underlying data @a
276       this was built from.
277   
278       @note Returns a dynamically allocated DataLookup2D, which has
279       to be deleted by the caller to avoid memory leaks.
280    */
281    const KernelLookup* test_kernel(const MatrixLookup& data) const;
282
283    /**
284       This function is useful when predicting on an independent data
285       set using a kernel-based classifier. In returned KernelLookup
286       column \f$ i \f$ corresponds to column \f$ i \f$ in @a
287       data. Row \f$ i \f$ in returned KernelLookup corresponds to
288       same sample as row \f$ i \f$ in @a this. In other words, this
289       function returns a KernelLookup containing the kernel elements
290       between the passed @a data and the internal underlying data @a
291       this was built from.
292   
293       @note Returns a dynamically allocated DataLookup2D, which has
294       to be deleted by the caller to avoid memory leaks.
295    */
296    const KernelLookup* test_kernel(const MatrixLookupWeighted& data) const;
297
298    /**
299       \brief Creates a sub-Kernel identical to the one created using
300       KernelLookup(*this, train, train).
301   
302       \return pointer to dynamically allocated sub-Lookup of the
303       KernelLookup
304   
305       \note Returns a dynamically allocated DataLookup2D, which has
306       to be deleted by the caller to avoid memory leaks.
307    */
308    const KernelLookup* training_data(const std::vector<size_t>& train) const;
309
310    /**
311       In returned kernel each row corresponds to a training sample
312       and each column corresponds to a validation sample. The created
313       sub-KernelLookup is equivalent to using KernelLooup(*this,
314       train, validation).
315   
316       \return sub-Lookup of the DataLookup2D
317   
318       \note Returns a dynamically allocated DataLookup2D, which has
319       to be deleted by the caller to avoid memory leaks.
320    */
321    const KernelLookup* 
322    validation_data(const std::vector<size_t>& train, 
323                    const std::vector<size_t>& validation) const;
324
325    /**
326       \return true if underlying Kernel is weighted
327    */
328    bool weighted(void) const;
329
330    /**
331       \return element at position (\a row, \a column) in the Kernel
332       matrix
333    */
334    double operator()(size_t row, size_t column) const;
335
336  private:
337    const KernelLookup& operator=(const KernelLookup&);
338
339    std::vector<size_t> column_index_;
340    utility::SmartPtr<const Kernel> kernel_;
341    std::vector<size_t> row_index_;
342   
343  }; // class KernelLookup
344
345}}} // of namespace classifier, yat, and theplu
346
347#endif
Note: See TracBrowser for help on using the repository browser.