NormalizeComponentsZCA.h
Go to the documentation of this file.
1//===========================================================================
2/*!
3 *
4 *
5 * \brief Data normalization to zero mean, unit variance and zero covariance while keping the original coordinate system
6 *
7 *
8 *
9 *
10 * \author T. Glasmachers
11 * \date 2010
12 *
13 *
14 * \par Copyright 1995-2017 Shark Development Team
15 *
16 * <BR><HR>
17 * This file is part of Shark.
18 * <https://shark-ml.github.io/Shark/>
19 *
20 * Shark is free software: you can redistribute it and/or modify
21 * it under the terms of the GNU Lesser General Public License as published
22 * by the Free Software Foundation, either version 3 of the License, or
23 * (at your option) any later version.
24 *
25 * Shark is distributed in the hope that it will be useful,
26 * but WITHOUT ANY WARRANTY; without even the implied warranty of
27 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
28 * GNU Lesser General Public License for more details.
29 *
30 * You should have received a copy of the GNU Lesser General Public License
31 * along with Shark. If not, see <http://www.gnu.org/licenses/>.
32 *
33 */
34//===========================================================================
35
36
37#ifndef SHARK_ALGORITHMS_TRAINERS_NORMALIZECOMPONENTSZCA_H
38#define SHARK_ALGORITHMS_TRAINERS_NORMALIZECOMPONENTSZCA_H
39
40
41#include <shark/Models/LinearModel.h>
42#include <shark/Algorithms/Trainers/AbstractTrainer.h>
43#include <shark/Data/Statistics.h>
44
45namespace shark {
46
47
48/// \brief Train a linear model to whiten the data
49///
50/// ZCA does whitening in the sense that it sets the mean to zero and the covariance to the Identity.
51/// However in contrast to NormalizeComponentsWhitening it makes sure that the initial and end coordinate
52/// system are the same and just rescales the data. The effect is, that image data still resembles images
53/// after applying ZCA in contrast to other methods which rotate the data randomly.
54/// \ingroup unsupervised_trainer
55class NormalizeComponentsZCA : public AbstractUnsupervisedTrainer<LinearModel<RealVector> >
56{
57 typedef AbstractUnsupervisedTrainer<LinearModel<RealVector> > base_type;
58public:
59
60 double m_targetVariance;
61 NormalizeComponentsZCA(double targetVariance = 1.0){
62 SHARK_RUNTIME_CHECK(targetVariance > 0.0, "Target variance must be positive");
63 m_targetVariance = targetVariance;
64 }
65
66 /// \brief From INameable: return the class name.
67 std::string name() const
68 { return "NormalizeComponentsZCA"; }
69
70 void train(ModelType& model, UnlabeledData<RealVector> const& input){
71 std::size_t dc = dataDimension(input);
72 SHARK_RUNTIME_CHECK(input.numberOfElements() >= dc + 1, "Input needs to contain more points than there are input dimensions");
73
74 // dense model with bias having input and output dimension equal to data dimension
75 model.setStructure(dc, dc, true);
76
77 RealVector mean;
78 RealMatrix covariance;
79 meanvar(input, mean, covariance);
80
81 blas::symm_eigenvalue_decomposition<RealMatrix> eigen(covariance);
82 covariance=RealMatrix(); //covariance not needed anymore
83
84
85 RealMatrix ZCAMatrix = eigen.Q() % to_diagonal(elem_inv(sqrt(eigen.D()))) % trans(eigen.Q());
86 ZCAMatrix *= std::sqrt(m_targetVariance);
87
88 RealVector offset = -prod(ZCAMatrix,mean);
89
90 model.setStructure(ZCAMatrix, offset);
91 }
92};
93
94
95}
96#endif