/usr/include/dlib/svm/rvm.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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 | // Copyright (C) 2008 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#ifndef DLIB_RVm_
#define DLIB_RVm_
#include "rvm_abstract.h"
#include <cmath>
#include <limits>
#include "../matrix.h"
#include "../algs.h"
#include "function.h"
#include "kernel.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
namespace rvm_helpers
{
// ------------------------------------------------------------------------------------
template <typename scalar_vector_type, typename mem_manager_type>
long find_next_best_alpha_to_update (
const scalar_vector_type& S,
const scalar_vector_type& Q,
const scalar_vector_type& alpha,
const matrix<long,0,1,mem_manager_type>& active_bases,
const bool search_all_alphas,
typename scalar_vector_type::type eps
)
/*!
ensures
- if (we can find another alpha to update) then
- returns the index of said alpha
- else
- returns -1
!*/
{
typedef typename scalar_vector_type::type scalar_type;
// now use S and Q to find next alpha to update. What
// we want to do here is select the alpha to update that gives us
// the greatest improvement in marginal likelihood.
long selected_idx = -1;
scalar_type greatest_improvement = -1;
for (long i = 0; i < S.nr(); ++i)
{
scalar_type value = -1;
// if i is currently in the active set
if (active_bases(i) >= 0)
{
const long idx = active_bases(i);
const scalar_type s = alpha(idx)*S(i)/(alpha(idx) - S(i));
const scalar_type q = alpha(idx)*Q(i)/(alpha(idx) - S(i));
if (q*q-s > 0)
{
// only update an existing alpha if this is a narrow search
if (search_all_alphas == false)
{
// choosing this sample would mean doing an update of an
// existing alpha value.
scalar_type new_alpha = s*s/(q*q-s);
scalar_type cur_alpha = alpha(idx);
new_alpha = 1/new_alpha;
cur_alpha = 1/cur_alpha;
// from equation 32 in the Tipping paper
value = Q(i)*Q(i)/(S(i) + 1/(new_alpha - cur_alpha) ) -
std::log(1 + S(i)*(new_alpha - cur_alpha));
}
}
// we only pick an alpha to remove if this is a wide search and it wasn't one of the recently added ones
else if (search_all_alphas && idx+2 < alpha.size() )
{
// choosing this sample would mean the alpha value is infinite
// so we would remove the selected sample from our model.
// from equation 37 in the Tipping paper
value = Q(i)*Q(i)/(S(i) - alpha(idx)) -
std::log(1-S(i)/alpha(idx));
}
}
else if (search_all_alphas)
{
const scalar_type s = S(i);
const scalar_type q = Q(i);
if (q*q-s > 0)
{
// choosing this sample would mean we would add the selected
// sample to our model.
// from equation 27 in the Tipping paper
value = (Q(i)*Q(i)-S(i))/S(i) + std::log(S(i)/(Q(i)*Q(i)));
}
}
if (value > greatest_improvement)
{
greatest_improvement = value;
selected_idx = i;
}
}
// If the greatest_improvement in marginal likelihood we would get is less
// than our epsilon then report that there isn't anything else to do. But
// if it is big enough then return the selected_idx.
if (greatest_improvement > eps)
return selected_idx;
else
return -1;
}
} // end namespace rvm_helpers
// ------------------------------------------------------------------------------------
template <
typename kern_type
>
class rvm_trainer
{
/*!
This is an implementation of the binary classifier version of the
relevance vector machine algorithm described in the paper:
Tipping, M. E. and A. C. Faul (2003). Fast marginal likelihood maximisation
for sparse Bayesian models. In C. M. Bishop and B. J. Frey (Eds.), Proceedings
of the Ninth International Workshop on Artificial Intelligence and Statistics,
Key West, FL, Jan 3-6.
This code mostly does what is described in the above paper with the exception
that here we use a different stopping condition as well as a modified alpha
selection rule. See the code for the exact details.
!*/
public:
typedef kern_type kernel_type;
typedef typename kernel_type::scalar_type scalar_type;
typedef typename kernel_type::sample_type sample_type;
typedef typename kernel_type::mem_manager_type mem_manager_type;
typedef decision_function<kernel_type> trained_function_type;
rvm_trainer (
) : eps(0.001), max_iterations(2000)
{
}
void set_max_iterations (
unsigned long max_iterations_
)
{
max_iterations = max_iterations_;
}
unsigned long get_max_iterations (
) const
{
return max_iterations;
}
void set_epsilon (
scalar_type eps_
)
{
// make sure requires clause is not broken
DLIB_ASSERT(eps_ > 0,
"\tvoid rvm_trainer::set_epsilon(eps_)"
<< "\n\t invalid inputs were given to this function"
<< "\n\t eps: " << eps_
);
eps = eps_;
}
const scalar_type get_epsilon (
) const
{
return eps;
}
void set_kernel (
const kernel_type& k
)
{
kernel = k;
}
const kernel_type& get_kernel (
) const
{
return kernel;
}
template <
typename in_sample_vector_type,
typename in_scalar_vector_type
>
const decision_function<kernel_type> train (
const in_sample_vector_type& x,
const in_scalar_vector_type& y
) const
{
return do_train(mat(x), mat(y));
}
void swap (
rvm_trainer& item
)
{
exchange(kernel, item.kernel);
exchange(eps, item.eps);
}
private:
// ------------------------------------------------------------------------------------
typedef matrix<scalar_type,0,1,mem_manager_type> scalar_vector_type;
typedef matrix<scalar_type,0,0,mem_manager_type> scalar_matrix_type;
template <
typename in_sample_vector_type,
typename in_scalar_vector_type
>
const decision_function<kernel_type> do_train (
const in_sample_vector_type& x,
const in_scalar_vector_type& y
) const
{
// make sure requires clause is not broken
DLIB_ASSERT(is_binary_classification_problem(x,y) == true,
"\tdecision_function rvm_trainer::train(x,y)"
<< "\n\t invalid inputs were given to this function"
<< "\n\t x.nr(): " << x.nr()
<< "\n\t y.nr(): " << y.nr()
<< "\n\t x.nc(): " << x.nc()
<< "\n\t y.nc(): " << y.nc()
<< "\n\t is_binary_classification_problem(x,y): " << ((is_binary_classification_problem(x,y))? "true":"false")
);
// make a target vector where +1 examples have value 1 and -1 examples
// have a value of 0.
scalar_vector_type t(y.size());
for (long i = 0; i < y.size(); ++i)
{
if (y(i) == 1)
t(i) = 1;
else
t(i) = 0;
}
/*! This is the convention for the active_bases variable in the function:
- if (active_bases(i) >= 0) then
- alpha(active_bases(i)) == the alpha value associated with sample x(i)
- weights(active_bases(i)) == the weight value associated with sample x(i)
- colm(phi, active_bases(i)) == the column of phi associated with sample x(i)
- colm(phi, active_bases(i)) == kernel column i (from get_kernel_colum())
- else
- the i'th sample isn't in the model and notionally has an alpha of infinity and
a weight of 0.
!*/
matrix<long,0,1,mem_manager_type> active_bases(x.nr());
scalar_matrix_type phi(x.nr(),1);
scalar_vector_type alpha(1), prev_alpha;
scalar_vector_type weights(1), prev_weights;
scalar_vector_type tempv, K_col;
// set the initial values of these guys
set_all_elements(active_bases, -1);
long first_basis = pick_initial_vector(x,t);
get_kernel_colum(first_basis, x, K_col);
active_bases(first_basis) = 0;
set_colm(phi,0) = K_col;
alpha(0) = compute_initial_alpha(phi, t);
weights(0) = 1;
// now declare a bunch of other variables we will be using below
scalar_vector_type mu, t_hat, Q, S;
scalar_matrix_type sigma;
matrix<scalar_type,1,0,mem_manager_type> tempv2, tempv3;
scalar_matrix_type tempm;
scalar_vector_type t_estimate;
scalar_vector_type beta;
Q.set_size(x.nr());
S.set_size(x.nr());
bool recompute_beta = true;
bool search_all_alphas = false;
unsigned long ticker = 0;
const unsigned long rounds_of_narrow_search = 100;
unsigned long iterations = 0;
while (iterations != max_iterations)
{
iterations++;
if (recompute_beta)
{
// calculate the current t_estimate. (this is the predicted t value for each sample according to the
// current state of the classifier)
t_estimate = phi*weights;
// calculate the current beta
beta = sigmoid(t_estimate);
beta = pointwise_multiply(beta,(uniform_matrix<scalar_type>(beta.nr(),beta.nc(),1)-beta));
recompute_beta = false;
}
// Compute optimal weights and sigma for current alpha using IRLS. This is the same
// technique documented in the paper by equations 12-14.
scalar_type weight_delta = std::numeric_limits<scalar_type>::max();
int count = 0;
while (weight_delta > 0.0001)
{
// This is a sanity check to make sure we never get stuck in this
// loop to do some degenerate numerical condition
++count;
if (count > 100)
{
// jump us to where search_all_alphas will be set to true
ticker = rounds_of_narrow_search;
break;
}
// compute the updated sigma matrix
sigma = scale_columns(trans(phi),beta)*phi;
for (long r = 0; r < alpha.nr(); ++r)
sigma(r,r) += alpha(r);
sigma = inv(sigma);
// compute the updated weights vector (t_hat = phi*mu_mp + inv(B)*(t-y))
t_hat = t_estimate + trans(scale_columns(trans(t-sigmoid(t_estimate)),reciprocal(beta)));
// mu = sigma*trans(phi)*b*t_hat
mu = sigma*tmp(trans(phi)* trans(scale_columns(trans(t_hat), beta)));
// now compute how much the weights vector changed during this iteration
// through this loop.
weight_delta = max(abs(mu-weights));
// put mu into the weights vector
mu.swap(weights);
// calculate the current t_estimate
t_estimate = phi*weights;
// calculate the current beta
beta = sigmoid(t_estimate);
beta = pointwise_multiply(beta, uniform_matrix<scalar_type>(beta.nr(),beta.nc(),1)-beta);
}
// check if we should do a full search for the best alpha to optimize
if (ticker >= rounds_of_narrow_search)
{
// if the current alpha and weights are equal to what they were
// at the last time we were about to start a wide search then
// we are done.
if (equal(prev_alpha, alpha, eps) && equal(prev_weights, weights, eps))
break;
prev_alpha = alpha;
prev_weights = weights;
search_all_alphas = true;
ticker = 0;
}
else
{
search_all_alphas = false;
}
++ticker;
// compute S and Q using equations 24 and 25 (tempv = phi*sigma*trans(phi)*B*t_hat)
tempv = phi*tmp(sigma*tmp(trans(phi)*trans(scale_columns(trans(t_hat),beta))));
for (long i = 0; i < S.size(); ++i)
{
// if we are currently limiting the search for the next alpha to update
// to the set in the active set then skip a non-active vector.
if (search_all_alphas == false && active_bases(i) == -1)
continue;
// get the column for this sample out of the kernel matrix. If it is
// something in the active set then just get it right out of phi, otherwise
// we have to calculate it.
if (active_bases(i) != -1)
K_col = colm(phi,active_bases(i));
else
get_kernel_colum(i, x, K_col);
// tempv2 = trans(phi_m)*B
tempv2 = scale_columns(trans(K_col), beta);
tempv3 = tempv2*phi;
S(i) = tempv2*K_col - tempv3*sigma*trans(tempv3);
Q(i) = tempv2*t_hat - tempv2*tempv;
}
const long selected_idx = rvm_helpers::find_next_best_alpha_to_update(S,Q,alpha,active_bases, search_all_alphas, eps);
// if find_next_best_alpha_to_update didn't find any good alpha to update
if (selected_idx == -1)
{
if (search_all_alphas == false)
{
// jump us to where search_all_alphas will be set to true and try again
ticker = rounds_of_narrow_search;
continue;
}
else
{
// we are really done so quit the main loop
break;
}
}
// next we update the selected alpha.
// if the selected alpha is in the active set
if (active_bases(selected_idx) >= 0)
{
const long idx = active_bases(selected_idx);
const scalar_type s = alpha(idx)*S(selected_idx)/(alpha(idx) - S(selected_idx));
const scalar_type q = alpha(idx)*Q(selected_idx)/(alpha(idx) - S(selected_idx));
if (q*q-s > 0)
{
// reestimate the value of alpha
alpha(idx) = s*s/(q*q-s);
}
else
{
// the new alpha value is infinite so remove the selected alpha from our model
active_bases(selected_idx) = -1;
phi = remove_col(phi, idx);
weights = remove_row(weights, idx);
alpha = remove_row(alpha, idx);
// fix the index values in active_bases
for (long i = 0; i < active_bases.size(); ++i)
{
if (active_bases(i) > idx)
{
active_bases(i) -= 1;
}
}
// we changed the number of weights so we need to remember to
// recompute the beta vector next time around the main loop.
recompute_beta = true;
}
}
else
{
const scalar_type s = S(selected_idx);
const scalar_type q = Q(selected_idx);
if (q*q-s > 0)
{
// add the selected alpha to our model
active_bases(selected_idx) = phi.nc();
// update alpha
tempv.set_size(alpha.size()+1);
set_subm(tempv, get_rect(alpha)) = alpha;
tempv(phi.nc()) = s*s/(q*q-s);
tempv.swap(alpha);
// update weights
tempv.set_size(weights.size()+1);
set_subm(tempv, get_rect(weights)) = weights;
tempv(phi.nc()) = 0;
tempv.swap(weights);
// update phi by adding the new sample's kernel matrix column in as one of phi's columns
tempm.set_size(phi.nr(), phi.nc()+1);
set_subm(tempm, get_rect(phi)) = phi;
get_kernel_colum(selected_idx, x, K_col);
set_colm(tempm, phi.nc()) = K_col;
tempm.swap(phi);
// we changed the number of weights so we need to remember to
// recompute the beta vector next time around the main loop.
recompute_beta = true;
}
}
} // end while(true). So we have converged on the final answer.
// now put everything into a decision_function object and return it
std_vector_c<sample_type> dictionary;
std_vector_c<scalar_type> final_weights;
for (long i = 0; i < active_bases.size(); ++i)
{
if (active_bases(i) >= 0)
{
dictionary.push_back(x(i));
final_weights.push_back(weights(active_bases(i)));
}
}
return decision_function<kernel_type> ( mat(final_weights),
-sum(mat(final_weights))*tau,
kernel,
mat(dictionary));
}
// ------------------------------------------------------------------------------------
template <typename M1, typename M2>
long pick_initial_vector (
const M1& x,
const M2& t
) const
{
scalar_vector_type K_col;
double max_projection = -std::numeric_limits<scalar_type>::infinity();
long max_idx = 0;
// find the row in the kernel matrix that has the biggest normalized projection onto the t vector
for (long r = 0; r < x.nr(); ++r)
{
get_kernel_colum(r,x,K_col);
double temp = trans(K_col)*t;
temp = temp*temp/length_squared(K_col);
if (temp > max_projection)
{
max_projection = temp;
max_idx = r;
}
}
return max_idx;
}
// ------------------------------------------------------------------------------------
template <typename T>
void get_kernel_colum (
long idx,
const T& x,
scalar_vector_type& col
) const
{
col.set_size(x.nr());
for (long i = 0; i < col.size(); ++i)
{
col(i) = kernel(x(idx), x(i)) + tau;
}
}
// ------------------------------------------------------------------------------------
template <typename M1, typename M2>
scalar_type compute_initial_alpha (
const M1& phi,
const M2& t
) const
{
const double temp = length_squared(phi);
const double temp2 = trans(phi)*t;
return temp/( temp2*temp2/temp + variance(t)*0.1);
}
// ------------------------------------------------------------------------------------
// private member variables
kernel_type kernel;
scalar_type eps;
unsigned long max_iterations;
const static scalar_type tau;
}; // end of class rvm_trainer
template <typename kernel_type>
const typename kernel_type::scalar_type rvm_trainer<kernel_type>::tau = static_cast<typename kernel_type::scalar_type>(0.001);
// ----------------------------------------------------------------------------------------
template <typename K>
void swap (
rvm_trainer<K>& a,
rvm_trainer<K>& b
) { a.swap(b); }
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
template <
typename kern_type
>
class rvm_regression_trainer
{
/*!
This is an implementation of the regression version of the
relevance vector machine algorithm described in the paper:
Tipping, M. E. and A. C. Faul (2003). Fast marginal likelihood maximisation
for sparse Bayesian models. In C. M. Bishop and B. J. Frey (Eds.), Proceedings
of the Ninth International Workshop on Artificial Intelligence and Statistics,
Key West, FL, Jan 3-6.
This code mostly does what is described in the above paper with the exception
that here we use a different stopping condition as well as a modified alpha
selection rule. See the code for the exact details.
!*/
public:
typedef kern_type kernel_type;
typedef typename kernel_type::scalar_type scalar_type;
typedef typename kernel_type::sample_type sample_type;
typedef typename kernel_type::mem_manager_type mem_manager_type;
typedef decision_function<kernel_type> trained_function_type;
rvm_regression_trainer (
) : eps(0.001)
{
}
void set_epsilon (
scalar_type eps_
)
{
// make sure requires clause is not broken
DLIB_ASSERT(eps_ > 0,
"\tvoid rvm_regression_trainer::set_epsilon(eps_)"
<< "\n\t invalid inputs were given to this function"
<< "\n\t eps: " << eps_
);
eps = eps_;
}
const scalar_type get_epsilon (
) const
{
return eps;
}
void set_kernel (
const kernel_type& k
)
{
kernel = k;
}
const kernel_type& get_kernel (
) const
{
return kernel;
}
template <
typename in_sample_vector_type,
typename in_scalar_vector_type
>
const decision_function<kernel_type> train (
const in_sample_vector_type& x,
const in_scalar_vector_type& t
) const
{
return do_train(mat(x), mat(t));
}
void swap (
rvm_regression_trainer& item
)
{
exchange(kernel, item.kernel);
exchange(eps, item.eps);
}
private:
// ------------------------------------------------------------------------------------
typedef matrix<scalar_type,0,1,mem_manager_type> scalar_vector_type;
typedef matrix<scalar_type,0,0,mem_manager_type> scalar_matrix_type;
template <
typename in_sample_vector_type,
typename in_scalar_vector_type
>
const decision_function<kernel_type> do_train (
const in_sample_vector_type& x,
const in_scalar_vector_type& t
) const
{
// make sure requires clause is not broken
DLIB_ASSERT(is_learning_problem(x,t) && x.size() > 0,
"\tdecision_function rvm_regression_trainer::train(x,t)"
<< "\n\t invalid inputs were given to this function"
<< "\n\t x.nr(): " << x.nr()
<< "\n\t t.nr(): " << t.nr()
<< "\n\t x.nc(): " << x.nc()
<< "\n\t t.nc(): " << t.nc()
);
/*! This is the convention for the active_bases variable in the function:
- if (active_bases(i) >= 0) then
- alpha(active_bases(i)) == the alpha value associated with sample x(i)
- weights(active_bases(i)) == the weight value associated with sample x(i)
- colm(phi, active_bases(i)) == the column of phi associated with sample x(i)
- colm(phi, active_bases(i)) == kernel column i (from get_kernel_colum())
- else
- the i'th sample isn't in the model and notionally has an alpha of infinity and
a weight of 0.
!*/
matrix<long,0,1,mem_manager_type> active_bases(x.nr());
scalar_matrix_type phi(x.nr(),1);
scalar_vector_type alpha(1), prev_alpha;
scalar_vector_type weights(1), prev_weights;
scalar_vector_type tempv, K_col;
scalar_type var = variance(t)*0.1;
// set the initial values of these guys
set_all_elements(active_bases, -1);
long first_basis = pick_initial_vector(x,t);
get_kernel_colum(first_basis, x, K_col);
active_bases(first_basis) = 0;
set_colm(phi,0) = K_col;
alpha(0) = compute_initial_alpha(phi, t, var);
weights(0) = 1;
// now declare a bunch of other variables we will be using below
scalar_vector_type Q, S;
scalar_matrix_type sigma;
matrix<scalar_type,1,0,mem_manager_type> tempv2, tempv3;
scalar_matrix_type tempm;
Q.set_size(x.nr());
S.set_size(x.nr());
bool search_all_alphas = false;
unsigned long ticker = 0;
const unsigned long rounds_of_narrow_search = 100;
while (true)
{
// Compute optimal weights and sigma for current alpha using equation 6.
sigma = trans(phi)*phi/var;
for (long r = 0; r < alpha.nr(); ++r)
sigma(r,r) += alpha(r);
sigma = inv(sigma);
weights = sigma*trans(phi)*t/var;
// check if we should do a full search for the best alpha to optimize
if (ticker == rounds_of_narrow_search)
{
// if the current alpha and weights are equal to what they were
// at the last time we were about to start a wide search then
// we are done.
if (equal(prev_alpha, alpha, eps) && equal(prev_weights, weights, eps))
break;
prev_alpha = alpha;
prev_weights = weights;
search_all_alphas = true;
ticker = 0;
}
else
{
search_all_alphas = false;
}
++ticker;
// compute S and Q using equations 24 and 25 (tempv = phi*sigma*trans(phi)*B*t)
tempv = phi*tmp(sigma*tmp(trans(phi)*t/var));
for (long i = 0; i < S.size(); ++i)
{
// if we are currently limiting the search for the next alpha to update
// to the set in the active set then skip a non-active vector.
if (search_all_alphas == false && active_bases(i) == -1)
continue;
// get the column for this sample out of the kernel matrix. If it is
// something in the active set then just get it right out of phi, otherwise
// we have to calculate it.
if (active_bases(i) != -1)
K_col = colm(phi,active_bases(i));
else
get_kernel_colum(i, x, K_col);
// tempv2 = trans(phi_m)*B
tempv2 = trans(K_col)/var;
tempv3 = tempv2*phi;
S(i) = tempv2*K_col - tempv3*sigma*trans(tempv3);
Q(i) = tempv2*t - tempv2*tempv;
}
const long selected_idx = rvm_helpers::find_next_best_alpha_to_update(S,Q,alpha,active_bases, search_all_alphas, eps);
// if find_next_best_alpha_to_update didn't find any good alpha to update
if (selected_idx == -1)
{
if (search_all_alphas == false)
{
// jump us to where search_all_alphas will be set to true and try again
ticker = rounds_of_narrow_search;
continue;
}
else
{
// we are really done so quit the main loop
break;
}
}
// recompute the variance
var = length_squared(t - phi*weights)/(x.nr() - weights.size() + trans(alpha)*diag(sigma));
// next we update the selected alpha.
// if the selected alpha is in the active set
if (active_bases(selected_idx) >= 0)
{
const long idx = active_bases(selected_idx);
const scalar_type s = alpha(idx)*S(selected_idx)/(alpha(idx) - S(selected_idx));
const scalar_type q = alpha(idx)*Q(selected_idx)/(alpha(idx) - S(selected_idx));
if (q*q-s > 0)
{
// reestimate the value of alpha
alpha(idx) = s*s/(q*q-s);
}
else
{
// the new alpha value is infinite so remove the selected alpha from our model
active_bases(selected_idx) = -1;
phi = remove_col(phi, idx);
weights = remove_row(weights, idx);
alpha = remove_row(alpha, idx);
// fix the index values in active_bases
for (long i = 0; i < active_bases.size(); ++i)
{
if (active_bases(i) > idx)
{
active_bases(i) -= 1;
}
}
}
}
else
{
const scalar_type s = S(selected_idx);
const scalar_type q = Q(selected_idx);
if (q*q-s > 0)
{
// add the selected alpha to our model
active_bases(selected_idx) = phi.nc();
// update alpha
tempv.set_size(alpha.size()+1);
set_subm(tempv, get_rect(alpha)) = alpha;
tempv(phi.nc()) = s*s/(q*q-s);
tempv.swap(alpha);
// update weights
tempv.set_size(weights.size()+1);
set_subm(tempv, get_rect(weights)) = weights;
tempv(phi.nc()) = 0;
tempv.swap(weights);
// update phi by adding the new sample's kernel matrix column in as one of phi's columns
tempm.set_size(phi.nr(), phi.nc()+1);
set_subm(tempm, get_rect(phi)) = phi;
get_kernel_colum(selected_idx, x, K_col);
set_colm(tempm, phi.nc()) = K_col;
tempm.swap(phi);
}
}
} // end while(true). So we have converged on the final answer.
// now put everything into a decision_function object and return it
std_vector_c<sample_type> dictionary;
std_vector_c<scalar_type> final_weights;
for (long i = 0; i < active_bases.size(); ++i)
{
if (active_bases(i) >= 0)
{
dictionary.push_back(x(i));
final_weights.push_back(weights(active_bases(i)));
}
}
return decision_function<kernel_type> ( mat(final_weights),
-sum(mat(final_weights))*tau,
kernel,
mat(dictionary));
}
// ------------------------------------------------------------------------------------
template <typename T>
void get_kernel_colum (
long idx,
const T& x,
scalar_vector_type& col
) const
{
col.set_size(x.nr());
for (long i = 0; i < col.size(); ++i)
{
col(i) = kernel(x(idx), x(i)) + tau;
}
}
// ------------------------------------------------------------------------------------
template <typename M1, typename M2>
scalar_type compute_initial_alpha (
const M1& phi,
const M2& t,
const scalar_type& var
) const
{
const double temp = length_squared(phi);
const double temp2 = trans(phi)*t;
return temp/( temp2*temp2/temp + var);
}
// ------------------------------------------------------------------------------------
template <typename M1, typename M2>
long pick_initial_vector (
const M1& x,
const M2& t
) const
{
scalar_vector_type K_col;
double max_projection = -std::numeric_limits<scalar_type>::infinity();
long max_idx = 0;
// find the row in the kernel matrix that has the biggest normalized projection onto the t vector
for (long r = 0; r < x.nr(); ++r)
{
get_kernel_colum(r,x,K_col);
double temp = trans(K_col)*t;
temp = temp*temp/length_squared(K_col);
if (temp > max_projection)
{
max_projection = temp;
max_idx = r;
}
}
return max_idx;
}
// ------------------------------------------------------------------------------------
// private member variables
kernel_type kernel;
scalar_type eps;
const static scalar_type tau;
}; // end of class rvm_regression_trainer
template <typename kernel_type>
const typename kernel_type::scalar_type rvm_regression_trainer<kernel_type>::tau = static_cast<typename kernel_type::scalar_type>(0.001);
// ----------------------------------------------------------------------------------------
template <typename K>
void swap (
rvm_regression_trainer<K>& a,
rvm_regression_trainer<K>& b
) { a.swap(b); }
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_RVm_
|