3709 |
08 Nov 17 |
peter |
// $Id$ |
3709 |
08 Nov 17 |
peter |
2 |
|
3709 |
08 Nov 17 |
peter |
3 |
/* |
4207 |
26 Aug 22 |
peter |
Copyright (C) 2017, 2021, 2022 Peter Johansson |
3709 |
08 Nov 17 |
peter |
5 |
|
3709 |
08 Nov 17 |
peter |
This file is part of the yat library, http://dev.thep.lu.se/yat |
3709 |
08 Nov 17 |
peter |
7 |
|
3709 |
08 Nov 17 |
peter |
The yat library is free software; you can redistribute it and/or |
3709 |
08 Nov 17 |
peter |
modify it under the terms of the GNU General Public License as |
3709 |
08 Nov 17 |
peter |
published by the Free Software Foundation; either version 3 of the |
3709 |
08 Nov 17 |
peter |
License, or (at your option) any later version. |
3709 |
08 Nov 17 |
peter |
12 |
|
3709 |
08 Nov 17 |
peter |
The yat library is distributed in the hope that it will be useful, |
3709 |
08 Nov 17 |
peter |
but WITHOUT ANY WARRANTY; without even the implied warranty of |
3709 |
08 Nov 17 |
peter |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
3709 |
08 Nov 17 |
peter |
General Public License for more details. |
3709 |
08 Nov 17 |
peter |
17 |
|
3709 |
08 Nov 17 |
peter |
You should have received a copy of the GNU General Public License |
3709 |
08 Nov 17 |
peter |
along with yat. If not, see <http://www.gnu.org/licenses/>. |
3709 |
08 Nov 17 |
peter |
20 |
*/ |
3709 |
08 Nov 17 |
peter |
21 |
|
3709 |
08 Nov 17 |
peter |
22 |
#include <config.h> |
3709 |
08 Nov 17 |
peter |
23 |
|
3709 |
08 Nov 17 |
peter |
24 |
#include "Perceptron.h" |
3709 |
08 Nov 17 |
peter |
25 |
|
3709 |
08 Nov 17 |
peter |
26 |
#include "Target.h" |
3709 |
08 Nov 17 |
peter |
27 |
|
3709 |
08 Nov 17 |
peter |
28 |
#include "yat/utility/DiagonalMatrix.h" |
3709 |
08 Nov 17 |
peter |
29 |
#include "yat/utility/Matrix.h" |
3709 |
08 Nov 17 |
peter |
30 |
#include "yat/utility/Vector.h" |
3709 |
08 Nov 17 |
peter |
31 |
|
3709 |
08 Nov 17 |
peter |
32 |
#include <gsl/gsl_cdf.h> |
3709 |
08 Nov 17 |
peter |
33 |
|
3709 |
08 Nov 17 |
peter |
34 |
#include <cassert> |
3709 |
08 Nov 17 |
peter |
35 |
#include <cmath> |
3709 |
08 Nov 17 |
peter |
36 |
#include <cmath> |
3709 |
08 Nov 17 |
peter |
37 |
|
3709 |
08 Nov 17 |
peter |
38 |
namespace theplu { |
3709 |
08 Nov 17 |
peter |
39 |
namespace yat { |
3709 |
08 Nov 17 |
peter |
40 |
namespace classifier { |
3709 |
08 Nov 17 |
peter |
41 |
|
3709 |
08 Nov 17 |
peter |
42 |
const utility::Matrix& Perceptron::covariance(void) const |
3709 |
08 Nov 17 |
peter |
43 |
{ |
3709 |
08 Nov 17 |
peter |
44 |
return covariance_; |
3709 |
08 Nov 17 |
peter |
45 |
} |
3709 |
08 Nov 17 |
peter |
46 |
|
3709 |
08 Nov 17 |
peter |
47 |
|
3709 |
08 Nov 17 |
peter |
48 |
double Perceptron::margin(size_t i, double alpha) const |
3709 |
08 Nov 17 |
peter |
49 |
{ |
3709 |
08 Nov 17 |
peter |
50 |
return gsl_cdf_ugaussian_Qinv(alpha/2) * std::sqrt(covariance_(i, i)); |
3709 |
08 Nov 17 |
peter |
51 |
} |
3709 |
08 Nov 17 |
peter |
52 |
|
3709 |
08 Nov 17 |
peter |
53 |
|
3709 |
08 Nov 17 |
peter |
54 |
double Perceptron::oddsratio(size_t i) const |
3709 |
08 Nov 17 |
peter |
55 |
{ |
3709 |
08 Nov 17 |
peter |
56 |
return std::exp(weight_(i)); |
3709 |
08 Nov 17 |
peter |
57 |
} |
3709 |
08 Nov 17 |
peter |
58 |
|
3709 |
08 Nov 17 |
peter |
59 |
|
3709 |
08 Nov 17 |
peter |
60 |
double Perceptron::oddsratio_lower_CI(size_t i, double alpha) const |
3709 |
08 Nov 17 |
peter |
61 |
{ |
3709 |
08 Nov 17 |
peter |
62 |
return std::exp(weight_(i) - margin(i, alpha)); |
3709 |
08 Nov 17 |
peter |
63 |
} |
3709 |
08 Nov 17 |
peter |
64 |
|
3709 |
08 Nov 17 |
peter |
65 |
|
3709 |
08 Nov 17 |
peter |
66 |
double Perceptron::oddsratio_upper_CI(size_t i, double alpha) const |
3709 |
08 Nov 17 |
peter |
67 |
{ |
3709 |
08 Nov 17 |
peter |
68 |
return std::exp(weight_(i) + margin(i, alpha)); |
3709 |
08 Nov 17 |
peter |
69 |
} |
3709 |
08 Nov 17 |
peter |
70 |
|
3709 |
08 Nov 17 |
peter |
71 |
|
3709 |
08 Nov 17 |
peter |
72 |
double Perceptron::p_value(size_t i) const |
3709 |
08 Nov 17 |
peter |
73 |
{ |
3709 |
08 Nov 17 |
peter |
74 |
double z = weight_(i) / std::sqrt(covariance_(i, i)); |
3709 |
08 Nov 17 |
peter |
75 |
return 2*gsl_cdf_ugaussian_Q(std::abs(z)); |
3709 |
08 Nov 17 |
peter |
76 |
} |
3709 |
08 Nov 17 |
peter |
77 |
|
3709 |
08 Nov 17 |
peter |
78 |
|
3709 |
08 Nov 17 |
peter |
79 |
double Perceptron::predict(const utility::VectorBase& x) const |
3709 |
08 Nov 17 |
peter |
80 |
{ |
3709 |
08 Nov 17 |
peter |
81 |
assert(x.size() == weight_.size()); |
3709 |
08 Nov 17 |
peter |
82 |
const double f = weight_ * x; |
3709 |
08 Nov 17 |
peter |
83 |
return 1.0 / (1 + std::exp(-f)); |
3709 |
08 Nov 17 |
peter |
84 |
} |
3709 |
08 Nov 17 |
peter |
85 |
|
3709 |
08 Nov 17 |
peter |
86 |
|
4125 |
14 Jan 22 |
peter |
87 |
void Perceptron::train(const utility::MatrixBase& X, const Target& target) |
3709 |
08 Nov 17 |
peter |
88 |
{ |
3709 |
08 Nov 17 |
peter |
89 |
size_t n = X.rows(); |
3709 |
08 Nov 17 |
peter |
90 |
size_t p = X.columns(); |
3709 |
08 Nov 17 |
peter |
91 |
|
3709 |
08 Nov 17 |
peter |
92 |
assert(target.size() == n); |
3709 |
08 Nov 17 |
peter |
93 |
weight_.resize(p); |
3709 |
08 Nov 17 |
peter |
94 |
covariance_.resize(p, p); |
3709 |
08 Nov 17 |
peter |
95 |
|
3709 |
08 Nov 17 |
peter |
// weight vector is updated as |
3709 |
08 Nov 17 |
peter |
// w = (X'SX)^-1 X' (SXw + y - mu) |
3709 |
08 Nov 17 |
peter |
// X is n x p |
3709 |
08 Nov 17 |
peter |
// mu is vector of (trained) expected values (see predict(1)) |
3709 |
08 Nov 17 |
peter |
100 |
utility::Vector mu(n); |
3709 |
08 Nov 17 |
peter |
// S is diagonal n x n with S_ii = mu_i (1 - mu_i) |
3709 |
08 Nov 17 |
peter |
102 |
utility::DiagonalMatrix S(n, n); |
3709 |
08 Nov 17 |
peter |
// y is binary vector |
3709 |
08 Nov 17 |
peter |
104 |
utility::Vector y(n); |
3709 |
08 Nov 17 |
peter |
105 |
for (size_t i=0; i<n; ++i) |
3709 |
08 Nov 17 |
peter |
106 |
if (target.binary(i)) |
3709 |
08 Nov 17 |
peter |
107 |
y(i) = 1.0; |
3709 |
08 Nov 17 |
peter |
108 |
|
4052 |
26 Mar 21 |
peter |
// We use the Iteratively Rewighted Least Square algorithm as described |
4052 |
26 Mar 21 |
peter |
// https://en.wikipedia.org/wiki/Logistic_regression |
4052 |
26 Mar 21 |
peter |
111 |
|
3709 |
08 Nov 17 |
peter |
112 |
size_t max_epochs = 100; |
3709 |
08 Nov 17 |
peter |
113 |
double sum_squared = 1.0; // some (relatively) large number |
3709 |
08 Nov 17 |
peter |
114 |
for (size_t epoch=0; sum_squared > 1e-8 && epoch < max_epochs; ++epoch) { |
3709 |
08 Nov 17 |
peter |
115 |
for (size_t i=0; i<mu.size(); ++i) { |
3709 |
08 Nov 17 |
peter |
116 |
mu(i) = predict(X.row_const_view(i)); |
3709 |
08 Nov 17 |
peter |
117 |
S(i) = mu(i) * (1.0 - mu(i)); |
3709 |
08 Nov 17 |
peter |
118 |
} |
3709 |
08 Nov 17 |
peter |
119 |
|
3709 |
08 Nov 17 |
peter |
// w = (X'SX)^-1 X' (SXw + y - mu) |
3709 |
08 Nov 17 |
peter |
121 |
assert(X.rows() == S.rows()); |
3709 |
08 Nov 17 |
peter |
122 |
assert(S.columns() == X.rows()); |
4140 |
29 Jan 22 |
peter |
123 |
utility::inverse_svd(utility::Matrix(transpose(X)*S*X), covariance_); |
3709 |
08 Nov 17 |
peter |
124 |
|
3709 |
08 Nov 17 |
peter |
125 |
assert(y.size() == mu.size()); |
3709 |
08 Nov 17 |
peter |
126 |
utility::Vector delta = covariance_ * (transpose(X) * (y - mu)); |
3709 |
08 Nov 17 |
peter |
127 |
weight_ += delta; |
3709 |
08 Nov 17 |
peter |
128 |
sum_squared = delta * delta; |
3709 |
08 Nov 17 |
peter |
129 |
} |
3709 |
08 Nov 17 |
peter |
130 |
} |
3709 |
08 Nov 17 |
peter |
131 |
|
3709 |
08 Nov 17 |
peter |
132 |
|
3709 |
08 Nov 17 |
peter |
133 |
const utility::Vector& Perceptron::weight(void) const |
3709 |
08 Nov 17 |
peter |
134 |
{ |
3709 |
08 Nov 17 |
peter |
135 |
return weight_; |
3709 |
08 Nov 17 |
peter |
136 |
} |
3709 |
08 Nov 17 |
peter |
137 |
|
3709 |
08 Nov 17 |
peter |
138 |
|
3709 |
08 Nov 17 |
peter |
139 |
}}} |