/usr/include/dlib/optimization/optimization.h is in libdlib-dev 18.18-1.
This file is owned by root:root, with mode 0o644.
The actual contents of the file can be viewed below.
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 | // Copyright (C) 2008 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#ifndef DLIB_OPTIMIZATIOn_H_
#define DLIB_OPTIMIZATIOn_H_
#include <cmath>
#include <limits>
#include "optimization_abstract.h"
#include "optimization_search_strategies.h"
#include "optimization_stop_strategies.h"
#include "optimization_line_search.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// Functions that transform other functions
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
template <typename funct>
class central_differences
{
public:
central_differences(const funct& f_, double eps_ = 1e-7) : f(f_), eps(eps_){}
template <typename T>
typename T::matrix_type operator()(const T& x) const
{
// T must be some sort of dlib matrix
COMPILE_TIME_ASSERT(is_matrix<T>::value);
typename T::matrix_type der(x.size());
typename T::matrix_type e(x);
for (long i = 0; i < x.size(); ++i)
{
const double old_val = e(i);
e(i) += eps;
const double delta_plus = f(e);
e(i) = old_val - eps;
const double delta_minus = f(e);
der(i) = (delta_plus - delta_minus)/(2*eps);
// and finally restore the old value of this element
e(i) = old_val;
}
return der;
}
template <typename T, typename U>
typename U::matrix_type operator()(const T& item, const U& x) const
{
// U must be some sort of dlib matrix
COMPILE_TIME_ASSERT(is_matrix<U>::value);
typename U::matrix_type der(x.size());
typename U::matrix_type e(x);
for (long i = 0; i < x.size(); ++i)
{
const double old_val = e(i);
e(i) += eps;
const double delta_plus = f(item,e);
e(i) = old_val - eps;
const double delta_minus = f(item,e);
der(i) = (delta_plus - delta_minus)/(2*eps);
// and finally restore the old value of this element
e(i) = old_val;
}
return der;
}
double operator()(const double& x) const
{
return (f(x+eps)-f(x-eps))/(2*eps);
}
private:
const funct& f;
const double eps;
};
template <typename funct>
const central_differences<funct> derivative(const funct& f) { return central_differences<funct>(f); }
template <typename funct>
const central_differences<funct> derivative(const funct& f, double eps)
{
DLIB_ASSERT (
eps > 0,
"\tcentral_differences derivative(f,eps)"
<< "\n\tYou must give an epsilon > 0"
<< "\n\teps: " << eps
);
return central_differences<funct>(f,eps);
}
// ----------------------------------------------------------------------------------------
template <typename funct, typename EXP1, typename EXP2>
struct clamped_function_object
{
clamped_function_object(
const funct& f_,
const matrix_exp<EXP1>& x_lower_,
const matrix_exp<EXP2>& x_upper_
) : f(f_), x_lower(x_lower_), x_upper(x_upper_)
{
}
template <typename T>
double operator() (
const T& x
) const
{
return f(clamp(x,x_lower,x_upper));
}
const funct& f;
const matrix_exp<EXP1>& x_lower;
const matrix_exp<EXP2>& x_upper;
};
template <typename funct, typename EXP1, typename EXP2>
clamped_function_object<funct,EXP1,EXP2> clamp_function(
const funct& f,
const matrix_exp<EXP1>& x_lower,
const matrix_exp<EXP2>& x_upper
) { return clamped_function_object<funct,EXP1,EXP2>(f,x_lower,x_upper); }
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// Functions that perform unconstrained optimization
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
template <
typename search_strategy_type,
typename stop_strategy_type,
typename funct,
typename funct_der,
typename T
>
double find_min (
search_strategy_type search_strategy,
stop_strategy_type stop_strategy,
const funct& f,
const funct_der& der,
T& x,
double min_f
)
{
COMPILE_TIME_ASSERT(is_matrix<T>::value);
// The starting point (i.e. x) must be a column vector.
COMPILE_TIME_ASSERT(T::NC <= 1);
DLIB_CASSERT (
is_col_vector(x),
"\tdouble find_min()"
<< "\n\tYou have to supply column vectors to this function"
<< "\n\tx.nc(): " << x.nc()
);
T g, s;
double f_value = f(x);
g = der(x);
if (!is_finite(f_value))
throw error("The objective function generated non-finite outputs");
if (!is_finite(g))
throw error("The objective function generated non-finite outputs");
while(stop_strategy.should_continue_search(x, f_value, g) && f_value > min_f)
{
s = search_strategy.get_next_direction(x, f_value, g);
double alpha = line_search(
make_line_search_function(f,x,s, f_value),
f_value,
make_line_search_function(der,x,s, g),
dot(g,s), // compute initial gradient for the line search
search_strategy.get_wolfe_rho(), search_strategy.get_wolfe_sigma(), min_f,
search_strategy.get_max_line_search_iterations());
// Take the search step indicated by the above line search
x += alpha*s;
if (!is_finite(f_value))
throw error("The objective function generated non-finite outputs");
if (!is_finite(g))
throw error("The objective function generated non-finite outputs");
}
return f_value;
}
// ----------------------------------------------------------------------------------------
template <
typename search_strategy_type,
typename stop_strategy_type,
typename funct,
typename funct_der,
typename T
>
double find_max (
search_strategy_type search_strategy,
stop_strategy_type stop_strategy,
const funct& f,
const funct_der& der,
T& x,
double max_f
)
{
COMPILE_TIME_ASSERT(is_matrix<T>::value);
// The starting point (i.e. x) must be a column vector.
COMPILE_TIME_ASSERT(T::NC <= 1);
DLIB_CASSERT (
is_col_vector(x),
"\tdouble find_max()"
<< "\n\tYou have to supply column vectors to this function"
<< "\n\tx.nc(): " << x.nc()
);
T g, s;
// This function is basically just a copy of find_min() but with - put in the right places
// to flip things around so that it ends up looking for the max rather than the min.
double f_value = -f(x);
g = -der(x);
if (!is_finite(f_value))
throw error("The objective function generated non-finite outputs");
if (!is_finite(g))
throw error("The objective function generated non-finite outputs");
while(stop_strategy.should_continue_search(x, f_value, g) && f_value > -max_f)
{
s = search_strategy.get_next_direction(x, f_value, g);
double alpha = line_search(
negate_function(make_line_search_function(f,x,s, f_value)),
f_value,
negate_function(make_line_search_function(der,x,s, g)),
dot(g,s), // compute initial gradient for the line search
search_strategy.get_wolfe_rho(), search_strategy.get_wolfe_sigma(), -max_f,
search_strategy.get_max_line_search_iterations()
);
// Take the search step indicated by the above line search
x += alpha*s;
// Don't forget to negate these outputs from the line search since they are
// from the unnegated versions of f() and der()
g *= -1;
f_value *= -1;
if (!is_finite(f_value))
throw error("The objective function generated non-finite outputs");
if (!is_finite(g))
throw error("The objective function generated non-finite outputs");
}
return -f_value;
}
// ----------------------------------------------------------------------------------------
template <
typename search_strategy_type,
typename stop_strategy_type,
typename funct,
typename T
>
double find_min_using_approximate_derivatives (
search_strategy_type search_strategy,
stop_strategy_type stop_strategy,
const funct& f,
T& x,
double min_f,
double derivative_eps = 1e-7
)
{
COMPILE_TIME_ASSERT(is_matrix<T>::value);
// The starting point (i.e. x) must be a column vector.
COMPILE_TIME_ASSERT(T::NC <= 1);
DLIB_CASSERT (
is_col_vector(x) && derivative_eps > 0,
"\tdouble find_min_using_approximate_derivatives()"
<< "\n\tYou have to supply column vectors to this function"
<< "\n\tx.nc(): " << x.nc()
<< "\n\tderivative_eps: " << derivative_eps
);
T g, s;
double f_value = f(x);
g = derivative(f,derivative_eps)(x);
if (!is_finite(f_value))
throw error("The objective function generated non-finite outputs");
if (!is_finite(g))
throw error("The objective function generated non-finite outputs");
while(stop_strategy.should_continue_search(x, f_value, g) && f_value > min_f)
{
s = search_strategy.get_next_direction(x, f_value, g);
double alpha = line_search(
make_line_search_function(f,x,s,f_value),
f_value,
derivative(make_line_search_function(f,x,s),derivative_eps),
dot(g,s), // Sometimes the following line is a better way of determining the initial gradient.
//derivative(make_line_search_function(f,x,s),derivative_eps)(0),
search_strategy.get_wolfe_rho(), search_strategy.get_wolfe_sigma(), min_f,
search_strategy.get_max_line_search_iterations()
);
// Take the search step indicated by the above line search
x += alpha*s;
g = derivative(f,derivative_eps)(x);
if (!is_finite(f_value))
throw error("The objective function generated non-finite outputs");
if (!is_finite(g))
throw error("The objective function generated non-finite outputs");
}
return f_value;
}
// ----------------------------------------------------------------------------------------
template <
typename search_strategy_type,
typename stop_strategy_type,
typename funct,
typename T
>
double find_max_using_approximate_derivatives (
search_strategy_type search_strategy,
stop_strategy_type stop_strategy,
const funct& f,
T& x,
double max_f,
double derivative_eps = 1e-7
)
{
COMPILE_TIME_ASSERT(is_matrix<T>::value);
// The starting point (i.e. x) must be a column vector.
COMPILE_TIME_ASSERT(T::NC <= 1);
DLIB_CASSERT (
is_col_vector(x) && derivative_eps > 0,
"\tdouble find_max_using_approximate_derivatives()"
<< "\n\tYou have to supply column vectors to this function"
<< "\n\tx.nc(): " << x.nc()
<< "\n\tderivative_eps: " << derivative_eps
);
// Just negate the necessary things and call the find_min version of this function.
return -find_min_using_approximate_derivatives(
search_strategy,
stop_strategy,
negate_function(f),
x,
-max_f,
derivative_eps
);
}
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// Functions for box constrained optimization
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
template <typename T, typename U, typename V>
T zero_bounded_variables (
const double eps,
T vect,
const T& x,
const T& gradient,
const U& x_lower,
const V& x_upper
)
{
for (long i = 0; i < gradient.size(); ++i)
{
const double tol = eps*std::abs(x(i));
// if x(i) is an active bound constraint
if (x_lower(i)+tol >= x(i) && gradient(i) > 0)
vect(i) = 0;
else if (x_upper(i)-tol <= x(i) && gradient(i) < 0)
vect(i) = 0;
}
return vect;
}
// ----------------------------------------------------------------------------------------
template <typename T, typename U, typename V>
T gap_step_assign_bounded_variables (
const double eps,
T vect,
const T& x,
const T& gradient,
const U& x_lower,
const V& x_upper
)
{
for (long i = 0; i < gradient.size(); ++i)
{
const double tol = eps*std::abs(x(i));
// if x(i) is an active bound constraint then we should set it's search
// direction such that a single step along the direction either does nothing or
// closes the gap of size tol before hitting the bound exactly.
if (x_lower(i)+tol >= x(i) && gradient(i) > 0)
vect(i) = x_lower(i)-x(i);
else if (x_upper(i)-tol <= x(i) && gradient(i) < 0)
vect(i) = x_upper(i)-x(i);
}
return vect;
}
// ----------------------------------------------------------------------------------------
template <
typename search_strategy_type,
typename stop_strategy_type,
typename funct,
typename funct_der,
typename T,
typename EXP1,
typename EXP2
>
double find_min_box_constrained (
search_strategy_type search_strategy,
stop_strategy_type stop_strategy,
const funct& f,
const funct_der& der,
T& x,
const matrix_exp<EXP1>& x_lower,
const matrix_exp<EXP2>& x_upper
)
{
/*
The implementation of this function is more or less based on the discussion in
the paper Projected Newton-type Methods in Machine Learning by Mark Schmidt, et al.
*/
// make sure the requires clause is not violated
COMPILE_TIME_ASSERT(is_matrix<T>::value);
// The starting point (i.e. x) must be a column vector.
COMPILE_TIME_ASSERT(T::NC <= 1);
DLIB_CASSERT (
is_col_vector(x) && is_col_vector(x_lower) && is_col_vector(x_upper) &&
x.size() == x_lower.size() && x.size() == x_upper.size(),
"\tdouble find_min_box_constrained()"
<< "\n\t The inputs to this function must be equal length column vectors."
<< "\n\t is_col_vector(x): " << is_col_vector(x)
<< "\n\t is_col_vector(x_upper): " << is_col_vector(x_upper)
<< "\n\t is_col_vector(x_upper): " << is_col_vector(x_upper)
<< "\n\t x.size(): " << x.size()
<< "\n\t x_lower.size(): " << x_lower.size()
<< "\n\t x_upper.size(): " << x_upper.size()
);
DLIB_ASSERT (
min(x_upper-x_lower) > 0,
"\tdouble find_min_box_constrained()"
<< "\n\t You have to supply proper box constraints to this function."
<< "\n\r min(x_upper-x_lower): " << min(x_upper-x_lower)
);
T g, s;
double f_value = f(x);
g = der(x);
if (!is_finite(f_value))
throw error("The objective function generated non-finite outputs");
if (!is_finite(g))
throw error("The objective function generated non-finite outputs");
// gap_eps determines how close we have to get to a bound constraint before we
// start basically dropping it from the optimization and consider it to be an
// active constraint.
const double gap_eps = 1e-8;
double last_alpha = 1;
while(stop_strategy.should_continue_search(x, f_value, g))
{
s = search_strategy.get_next_direction(x, f_value, zero_bounded_variables(gap_eps, g, x, g, x_lower, x_upper));
s = gap_step_assign_bounded_variables(gap_eps, s, x, g, x_lower, x_upper);
double alpha = backtracking_line_search(
make_line_search_function(clamp_function(f,x_lower,x_upper), x, s, f_value),
f_value,
dot(g,s), // compute gradient for the line search
last_alpha,
search_strategy.get_wolfe_rho(),
search_strategy.get_max_line_search_iterations());
// Do a trust region style thing for alpha. The idea is that if we take a
// small step then we are likely to take another small step. So we reuse the
// alpha from the last iteration unless the line search didn't shrink alpha at
// all, in that case, we start with a bigger alpha next time.
if (alpha == last_alpha)
last_alpha = std::min(last_alpha*10,1.0);
else
last_alpha = alpha;
// Take the search step indicated by the above line search
x = clamp(x + alpha*s, x_lower, x_upper);
g = der(x);
if (!is_finite(f_value))
throw error("The objective function generated non-finite outputs");
if (!is_finite(g))
throw error("The objective function generated non-finite outputs");
}
return f_value;
}
// ----------------------------------------------------------------------------------------
template <
typename search_strategy_type,
typename stop_strategy_type,
typename funct,
typename funct_der,
typename T
>
double find_min_box_constrained (
search_strategy_type search_strategy,
stop_strategy_type stop_strategy,
const funct& f,
const funct_der& der,
T& x,
double x_lower,
double x_upper
)
{
// The starting point (i.e. x) must be a column vector.
COMPILE_TIME_ASSERT(T::NC <= 1);
typedef typename T::type scalar_type;
return find_min_box_constrained(search_strategy,
stop_strategy,
f,
der,
x,
uniform_matrix<scalar_type>(x.size(),1,x_lower),
uniform_matrix<scalar_type>(x.size(),1,x_upper) );
}
// ----------------------------------------------------------------------------------------
template <
typename search_strategy_type,
typename stop_strategy_type,
typename funct,
typename funct_der,
typename T,
typename EXP1,
typename EXP2
>
double find_max_box_constrained (
search_strategy_type search_strategy,
stop_strategy_type stop_strategy,
const funct& f,
const funct_der& der,
T& x,
const matrix_exp<EXP1>& x_lower,
const matrix_exp<EXP2>& x_upper
)
{
// make sure the requires clause is not violated
COMPILE_TIME_ASSERT(is_matrix<T>::value);
// The starting point (i.e. x) must be a column vector.
COMPILE_TIME_ASSERT(T::NC <= 1);
DLIB_CASSERT (
is_col_vector(x) && is_col_vector(x_lower) && is_col_vector(x_upper) &&
x.size() == x_lower.size() && x.size() == x_upper.size(),
"\tdouble find_max_box_constrained()"
<< "\n\t The inputs to this function must be equal length column vectors."
<< "\n\t is_col_vector(x): " << is_col_vector(x)
<< "\n\t is_col_vector(x_upper): " << is_col_vector(x_upper)
<< "\n\t is_col_vector(x_upper): " << is_col_vector(x_upper)
<< "\n\t x.size(): " << x.size()
<< "\n\t x_lower.size(): " << x_lower.size()
<< "\n\t x_upper.size(): " << x_upper.size()
);
DLIB_ASSERT (
min(x_upper-x_lower) > 0,
"\tdouble find_max_box_constrained()"
<< "\n\t You have to supply proper box constraints to this function."
<< "\n\r min(x_upper-x_lower): " << min(x_upper-x_lower)
);
// This function is basically just a copy of find_min_box_constrained() but with - put
// in the right places to flip things around so that it ends up looking for the max
// rather than the min.
T g, s;
double f_value = -f(x);
g = -der(x);
if (!is_finite(f_value))
throw error("The objective function generated non-finite outputs");
if (!is_finite(g))
throw error("The objective function generated non-finite outputs");
// gap_eps determines how close we have to get to a bound constraint before we
// start basically dropping it from the optimization and consider it to be an
// active constraint.
const double gap_eps = 1e-8;
double last_alpha = 1;
while(stop_strategy.should_continue_search(x, f_value, g))
{
s = search_strategy.get_next_direction(x, f_value, zero_bounded_variables(gap_eps, g, x, g, x_lower, x_upper));
s = gap_step_assign_bounded_variables(gap_eps, s, x, g, x_lower, x_upper);
double alpha = backtracking_line_search(
negate_function(make_line_search_function(clamp_function(f,x_lower,x_upper), x, s, f_value)),
f_value,
dot(g,s), // compute gradient for the line search
last_alpha,
search_strategy.get_wolfe_rho(),
search_strategy.get_max_line_search_iterations());
// Do a trust region style thing for alpha. The idea is that if we take a
// small step then we are likely to take another small step. So we reuse the
// alpha from the last iteration unless the line search didn't shrink alpha at
// all, in that case, we start with a bigger alpha next time.
if (alpha == last_alpha)
last_alpha = std::min(last_alpha*10,1.0);
else
last_alpha = alpha;
// Take the search step indicated by the above line search
x = clamp(x + alpha*s, x_lower, x_upper);
g = -der(x);
// Don't forget to negate the output from the line search since it is from the
// unnegated version of f()
f_value *= -1;
if (!is_finite(f_value))
throw error("The objective function generated non-finite outputs");
if (!is_finite(g))
throw error("The objective function generated non-finite outputs");
}
return -f_value;
}
// ----------------------------------------------------------------------------------------
template <
typename search_strategy_type,
typename stop_strategy_type,
typename funct,
typename funct_der,
typename T
>
double find_max_box_constrained (
search_strategy_type search_strategy,
stop_strategy_type stop_strategy,
const funct& f,
const funct_der& der,
T& x,
double x_lower,
double x_upper
)
{
// The starting point (i.e. x) must be a column vector.
COMPILE_TIME_ASSERT(T::NC <= 1);
typedef typename T::type scalar_type;
return find_max_box_constrained(search_strategy,
stop_strategy,
f,
der,
x,
uniform_matrix<scalar_type>(x.size(),1,x_lower),
uniform_matrix<scalar_type>(x.size(),1,x_upper) );
}
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_OPTIMIZATIOn_H_
|