summaryrefslogtreecommitdiff
path: root/lib/jxl/enc_optimize.h
blob: 9da523f8ef2e48ed8c133b38d515f33bd57ce056 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
// Copyright (c) the JPEG XL Project Authors. All rights reserved.
//
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

// Utility functions for optimizing multi-dimensional nonlinear functions.

#ifndef LIB_JXL_OPTIMIZE_H_
#define LIB_JXL_OPTIMIZE_H_

#include <cmath>
#include <cstdio>
#include <functional>
#include <vector>

#include "lib/jxl/base/status.h"

namespace jxl {
namespace optimize {

// An array type of numeric values that supports math operations with operator-,
// operator+, etc.
template <typename T, size_t N>
class Array {
 public:
  Array() = default;
  explicit Array(T v) {
    for (size_t i = 0; i < N; i++) v_[i] = v;
  }

  size_t size() const { return N; }

  T& operator[](size_t index) {
    JXL_DASSERT(index < N);
    return v_[index];
  }
  T operator[](size_t index) const {
    JXL_DASSERT(index < N);
    return v_[index];
  }

 private:
  // The values used by this Array.
  T v_[N];
};

template <typename T, size_t N>
Array<T, N> operator+(const Array<T, N>& x, const Array<T, N>& y) {
  Array<T, N> z;
  for (size_t i = 0; i < N; ++i) {
    z[i] = x[i] + y[i];
  }
  return z;
}

template <typename T, size_t N>
Array<T, N> operator-(const Array<T, N>& x, const Array<T, N>& y) {
  Array<T, N> z;
  for (size_t i = 0; i < N; ++i) {
    z[i] = x[i] - y[i];
  }
  return z;
}

template <typename T, size_t N>
Array<T, N> operator*(T v, const Array<T, N>& x) {
  Array<T, N> y;
  for (size_t i = 0; i < N; ++i) {
    y[i] = v * x[i];
  }
  return y;
}

template <typename T, size_t N>
T operator*(const Array<T, N>& x, const Array<T, N>& y) {
  T r = 0.0;
  for (size_t i = 0; i < N; ++i) {
    r += x[i] * y[i];
  }
  return r;
}

// Runs Nelder-Mead like optimization. Runs for max_iterations times,
// fun gets called with a vector of size dim as argument, and returns the score
// based on those parameters (lower is better). Returns a vector of dim+1
// dimensions, where the first value is the optimal value of the function and
// the rest is the argmin value. Use init to pass an initial guess or where
// the optimal value is.
//
// Usage example:
//
// RunSimplex(2, 0.1, 100, [](const vector<float>& v) {
//   return (v[0] - 5) * (v[0] - 5) + (v[1] - 7) * (v[1] - 7);
// });
//
// Returns (0.0, 5, 7)
std::vector<double> RunSimplex(
    int dim, double amount, int max_iterations,
    const std::function<double(const std::vector<double>&)>& fun);
std::vector<double> RunSimplex(
    int dim, double amount, int max_iterations, const std::vector<double>& init,
    const std::function<double(const std::vector<double>&)>& fun);

// Implementation of the Scaled Conjugate Gradient method described in the
// following paper:
//   Moller, M. "A Scaled Conjugate Gradient Algorithm for Fast Supervised
//   Learning", Neural Networks, Vol. 6. pp. 525-533, 1993
//   http://sci2s.ugr.es/keel/pdf/algorithm/articulo/moller1990.pdf
//
// The Function template parameter is a class that has the following method:
//
//   // Returns the value of the function at point w and sets *df to be the
//   // negative gradient vector of the function at point w.
//   double Compute(const optimize::Array<T, N>& w,
//                  optimize::Array<T, N>* df) const;
//
// Returns a vector w, such that |df(w)| < grad_norm_threshold.
template <typename T, size_t N, typename Function>
Array<T, N> OptimizeWithScaledConjugateGradientMethod(
    const Function& f, const Array<T, N>& w0, const T grad_norm_threshold,
    size_t max_iters) {
  const size_t n = w0.size();
  const T rsq_threshold = grad_norm_threshold * grad_norm_threshold;
  const T sigma0 = static_cast<T>(0.0001);
  const T l_min = static_cast<T>(1.0e-15);
  const T l_max = static_cast<T>(1.0e15);

  Array<T, N> w = w0;
  Array<T, N> wp;
  Array<T, N> r;
  Array<T, N> rt;
  Array<T, N> e;
  Array<T, N> p;
  T psq;
  T fp;
  T D;
  T d;
  T m;
  T a;
  T b;
  T s;
  T t;

  T fw = f.Compute(w, &r);
  T rsq = r * r;
  e = r;
  p = r;
  T l = static_cast<T>(1.0);
  bool success = true;
  size_t n_success = 0;
  size_t k = 0;

  while (k++ < max_iters) {
    if (success) {
      m = -(p * r);
      if (m >= 0) {
        p = r;
        m = -(p * r);
      }
      psq = p * p;
      s = sigma0 / std::sqrt(psq);
      f.Compute(w + (s * p), &rt);
      t = (p * (r - rt)) / s;
    }

    d = t + l * psq;
    if (d <= 0) {
      d = l * psq;
      l = l - t / psq;
    }

    a = -m / d;
    wp = w + a * p;
    fp = f.Compute(wp, &rt);

    D = 2.0 * (fp - fw) / (a * m);
    if (D >= 0.0) {
      success = true;
      n_success++;
      w = wp;
    } else {
      success = false;
    }

    if (success) {
      e = r;
      r = rt;
      rsq = r * r;
      fw = fp;
      if (rsq <= rsq_threshold) {
        break;
      }
    }

    if (D < 0.25) {
      l = std::min(4.0 * l, l_max);
    } else if (D > 0.75) {
      l = std::max(0.25 * l, l_min);
    }

    if ((n_success % n) == 0) {
      p = r;
      l = 1.0;
    } else if (success) {
      b = ((e - r) * r) / m;
      p = b * p + r;
    }
  }

  return w;
}

}  // namespace optimize
}  // namespace jxl

#endif  // LIB_JXL_OPTIMIZE_H_