plda.cc 29.3 KB
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
// ivector/plda.cc

// Copyright 2013     Daniel Povey
//           2015     David Snyder

// See ../../COPYING for clarification regarding multiple authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//  http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.

#include <vector>
#include "ivector/plda.h"

namespace kaldi {

void Plda::Write(std::ostream &os, bool binary) const {
  WriteToken(os, binary, "<Plda>");
  mean_.Write(os, binary);
  transform_.Write(os, binary);
  psi_.Write(os, binary);
  WriteToken(os, binary, "</Plda>");
}

void Plda::Read(std::istream &is, bool binary) {
  ExpectToken(is, binary, "<Plda>");
  mean_.Read(is, binary);
  transform_.Read(is, binary);
  psi_.Read(is, binary);
  ExpectToken(is, binary, "</Plda>");
  ComputeDerivedVars();
}

template<class Real>
/// This function computes a projection matrix that when applied makes the
/// covariance unit (i.e. all 1).
static void ComputeNormalizingTransform(const SpMatrix<Real> &covar,
                                        MatrixBase<Real> *proj) {
  int32 dim = covar.NumRows();
  TpMatrix<Real> C(dim);  // Cholesky of covar, covar = C C^T
  C.Cholesky(covar);
  C.Invert();  // The matrix that makes covar unit is C^{-1}, because
               // C^{-1} covar C^{-T} = C^{-1} C C^T C^{-T} = I.
  proj->CopyFromTp(C, kNoTrans);  // set "proj" to C^{-1}.
}


void Plda::ComputeDerivedVars() {
  KALDI_ASSERT(Dim() > 0);
  offset_.Resize(Dim());
  offset_.AddMatVec(-1.0, transform_, kNoTrans, mean_, 0.0);
}


/**
   This comment explains the thinking behind the function LogLikelihoodRatio.
   The reference is "Probabilistic Linear Discriminant Analysis" by
   Sergey Ioffe, ECCV 2006.

   I'm looking at the un-numbered equation between eqs. (4) and (5),
   that says
     P(u^p | u^g_{1...n}) =  N (u^p | \frac{n \Psi}{n \Psi + I} \bar{u}^g, I + \frac{\Psi}{n\Psi + I})

   Here, the superscript ^p refers to the "probe" example (e.g. the example
   to be classified), and u^g_1 is the first "gallery" example, i.e. the first
   training example of that class.  \psi is the between-class covariance
   matrix, assumed to be diagonalized, and I can be interpreted as the within-class
   covariance matrix which we have made unit.

   We want the likelihood ratio P(u^p | u^g_{1..n}) / P(u^p), where the
   numerator is the probability of u^p given that it's in that class, and the
   denominator is the probability of u^p with no class assumption at all
   (e.g. in its own class).

   The expression above even works for n = 0 (e.g. the denominator of the likelihood
   ratio), where it gives us
     P(u^p) = N(u^p | 0, I + \Psi)
   i.e. it's distributed with zero mean and covarance (within + between).
   The likelihood ratio we want is:
      N(u^p | \frac{n \Psi}{n \Psi + I} \bar{u}^g, I + \frac{\Psi}{n \Psi + I}) /
      N(u^p | 0, I + \Psi)
   where \bar{u}^g is the mean of the "gallery examples"; and we can expand the
   log likelihood ratio as
     - 0.5 [ (u^p - m) (I + \Psi/(n \Psi + I))^{-1} (u^p - m)  +  logdet(I + \Psi/(n \Psi + I)) ]
     + 0.5 [u^p (I + \Psi) u^p  +  logdet(I + \Psi) ]
   where m = (n \Psi)/(n \Psi + I) \bar{u}^g.

 */

double Plda::GetNormalizationFactor(
    const VectorBase<double> &transformed_ivector,
    int32 num_examples) const {
  KALDI_ASSERT(num_examples > 0);
  // Work out the normalization factor.  The covariance for an average over
  // "num_examples" training iVectors equals \Psi + I/num_examples.
  Vector<double> transformed_ivector_sq(transformed_ivector);
  transformed_ivector_sq.ApplyPow(2.0);
  // inv_covar will equal 1.0 / (\Psi + I/num_examples).
  Vector<double> inv_covar(psi_);
  inv_covar.Add(1.0 / num_examples);
  inv_covar.InvertElements();
  // "transformed_ivector" should have covariance (\Psi + I/num_examples), i.e.
  // within-class/num_examples plus between-class covariance.  So
  // transformed_ivector_sq . (I/num_examples + \Psi)^{-1} should be equal to
  //  the dimension.
  double dot_prod = VecVec(inv_covar, transformed_ivector_sq);
  return sqrt(Dim() / dot_prod);
}


double Plda::TransformIvector(const PldaConfig &config,
                              const VectorBase<double> &ivector,
                              int32 num_examples,
                              VectorBase<double> *transformed_ivector) const {
  KALDI_ASSERT(ivector.Dim() == Dim() && transformed_ivector->Dim() == Dim());
  double normalization_factor;
  transformed_ivector->CopyFromVec(offset_);
  transformed_ivector->AddMatVec(1.0, transform_, kNoTrans, ivector, 1.0);
  if (config.simple_length_norm)
    normalization_factor = sqrt(transformed_ivector->Dim())
      / transformed_ivector->Norm(2.0);
  else
    normalization_factor = GetNormalizationFactor(*transformed_ivector,
                                                  num_examples);
  if (config.normalize_length)
    transformed_ivector->Scale(normalization_factor);
  return normalization_factor;
}

// "float" version of TransformIvector.
float Plda::TransformIvector(const PldaConfig &config,
                             const VectorBase<float> &ivector,
                             int32 num_examples,
                             VectorBase<float> *transformed_ivector) const {
  Vector<double> tmp(ivector), tmp_out(ivector.Dim());
  float ans = TransformIvector(config, tmp, num_examples, &tmp_out);
  transformed_ivector->CopyFromVec(tmp_out);
  return ans;
}


// There is an extended comment within this file, referencing a paper by
// Ioffe, that may clarify what this function is doing.
double Plda::LogLikelihoodRatio(
    const VectorBase<double> &transformed_train_ivector,
    int32 n, // number of training utterances.
    const VectorBase<double> &transformed_test_ivector) const {
  int32 dim = Dim();
  double loglike_given_class, loglike_without_class;
  { // work out loglike_given_class.
    // "mean" will be the mean of the distribution if it comes from the
    // training example.  The mean is \frac{n \Psi}{n \Psi + I} \bar{u}^g
    // "variance" will be the variance of that distribution, equal to
    // I + \frac{\Psi}{n\Psi + I}.
    Vector<double> mean(dim, kUndefined);
    Vector<double> variance(dim, kUndefined);
    for (int32 i = 0; i < dim; i++) {
      mean(i) = n * psi_(i) / (n * psi_(i) + 1.0)
        * transformed_train_ivector(i);
      variance(i) = 1.0 + psi_(i) / (n * psi_(i) + 1.0);
    }
    double logdet = variance.SumLog();
    Vector<double> sqdiff(transformed_test_ivector);
    sqdiff.AddVec(-1.0, mean);
    sqdiff.ApplyPow(2.0);
    variance.InvertElements();
    loglike_given_class = -0.5 * (logdet + M_LOG_2PI * dim +
                                  VecVec(sqdiff, variance));
  }
  { // work out loglike_without_class.  Here the mean is zero and the variance
    // is I + \Psi.
    Vector<double> sqdiff(transformed_test_ivector); // there is no offset.
    sqdiff.ApplyPow(2.0);
    Vector<double> variance(psi_);
    variance.Add(1.0); // I + \Psi.
    double logdet = variance.SumLog();
    variance.InvertElements();
    loglike_without_class = -0.5 * (logdet + M_LOG_2PI * dim +
                                    VecVec(sqdiff, variance));
  }
  double loglike_ratio = loglike_given_class - loglike_without_class;
  return loglike_ratio;
}


void Plda::SmoothWithinClassCovariance(double smoothing_factor) {
  KALDI_ASSERT(smoothing_factor >= 0.0 && smoothing_factor <= 1.0);
  // smoothing_factor > 1.0 is possible but wouldn't really make sense.

  KALDI_LOG << "Smoothing within-class covariance by " << smoothing_factor
            << ", Psi is initially: " << psi_;
  Vector<double> within_class_covar(Dim());
  within_class_covar.Set(1.0); // It's now the current within-class covariance
                               // (a diagonal matrix) in the space transformed
                               // by transform_.
  within_class_covar.AddVec(smoothing_factor, psi_);
  /// We now revise our estimate of the within-class covariance to this
  /// larger value.  This means that the transform has to change to as
  /// to make this new, larger covariance unit.  And our between-class
  /// covariance in this space is now less.

  psi_.DivElements(within_class_covar);
  KALDI_LOG << "New value of Psi is " << psi_;

  within_class_covar.ApplyPow(-0.5);
  transform_.MulRowsVec(within_class_covar);

  ComputeDerivedVars();
}

void Plda::ApplyTransform(const Matrix<double> &in_transform) {
  KALDI_ASSERT(in_transform.NumRows() <= Dim()
    && in_transform.NumCols() == Dim());

  // Apply in_transform to mean_.
  Vector<double> mean_new(in_transform.NumRows());
  mean_new.AddMatVec(1.0, in_transform, kNoTrans, mean_, 0.0);
  mean_.Resize(in_transform.NumRows());
  mean_.CopyFromVec(mean_new);

  SpMatrix<double> between_var(in_transform.NumCols()),
                   within_var(in_transform.NumCols()),
                   psi_mat(in_transform.NumCols()),
                   between_var_new(Dim()),
                   within_var_new(Dim());
  Matrix<double> transform_invert(transform_);

  // Next, compute the between_var and within_var that existed
  // prior to diagonalization.
  psi_mat.AddDiagVec(1.0, psi_);
  transform_invert.Invert();
  within_var.AddMat2(1.0, transform_invert, kNoTrans, 0.0);
  between_var.AddMat2Sp(1.0, transform_invert, kNoTrans, psi_mat, 0.0);

  // Next, transform the variances using the input transformation.
  between_var_new.AddMat2Sp(1.0, in_transform, kNoTrans, between_var, 0.0);
  within_var_new.AddMat2Sp(1.0, in_transform, kNoTrans, within_var, 0.0);

  // Finally, we need to recompute psi_ and transform_. The remainder of
  // the code in this function  is a lightly modified copy of
  // PldaEstimator::GetOutput().
  Matrix<double> transform1(Dim(), Dim());
  ComputeNormalizingTransform(within_var_new, &transform1);
  // Now transform is a matrix that if we project with it,
  // within_var becomes unit.
  // between_var_proj is between_var after projecting with transform1.
  SpMatrix<double> between_var_proj(Dim());
  between_var_proj.AddMat2Sp(1.0, transform1, kNoTrans, between_var_new, 0.0);

  Matrix<double> U(Dim(), Dim());
  Vector<double> s(Dim());
  // Do symmetric eigenvalue decomposition between_var_proj = U diag(s) U^T,
  // where U is orthogonal.
  between_var_proj.Eig(&s, &U);

  KALDI_ASSERT(s.Min() >= 0.0);
  int32 n;
  s.ApplyFloor(0.0, &n);
  if (n > 0) {
    KALDI_WARN << "Floored " << n << " eigenvalues of between-class "
               << "variance to zero.";
  }
  // Sort from greatest to smallest eigenvalue.
  SortSvd(&s, &U);

  // The transform U^T will make between_var_proj diagonal with value s
  // (i.e. U^T U diag(s) U U^T = diag(s)).  The final transform that
  // makes within_var unit and between_var diagonal is U^T transform1,
  // i.e. first transform1 and then U^T.
  transform_.Resize(Dim(), Dim());
  transform_.AddMatMat(1.0, U, kTrans, transform1, kNoTrans, 0.0);
  psi_.Resize(Dim());
  psi_.CopyFromVec(s);
  ComputeDerivedVars();
}

void PldaStats::AddSamples(double weight,
                           const Matrix<double> &group) {
  if (dim_ == 0) {
    Init(group.NumCols());
  } else {
    KALDI_ASSERT(dim_ == group.NumCols());
  }
  int32 n = group.NumRows(); // number of examples for this class
  Vector<double> *mean = new Vector<double>(dim_);
  mean->AddRowSumMat(1.0 / n, group);

  offset_scatter_.AddMat2(weight, group, kTrans, 1.0);
  // the following statement has the same effect as if we
  // had first subtracted the mean from each element of
  // the group before the statement above.
  offset_scatter_.AddVec2(-n * weight, *mean);

  class_info_.push_back(ClassInfo(weight, mean, n));

  num_classes_ ++;
  num_examples_ += n;
  class_weight_ += weight;
  example_weight_ += weight * n;

  sum_.AddVec(weight, *mean);
}

PldaStats::~PldaStats() {
  for (size_t i = 0; i < class_info_.size(); i++)
    delete class_info_[i].mean;
}

bool PldaStats::IsSorted() const {
  for (size_t i = 0; i + 1 < class_info_.size(); i++)
    if (class_info_[i+1] < class_info_[i])
      return false;
  return true;
}

void PldaStats::Init(int32 dim) {
  KALDI_ASSERT(dim_ == 0);
  dim_ = dim;
  num_classes_ = 0;
  num_examples_ = 0;
  class_weight_ = 0.0;
  example_weight_ = 0.0;
  sum_.Resize(dim);
  offset_scatter_.Resize(dim);
  KALDI_ASSERT(class_info_.empty());
}


PldaEstimator::PldaEstimator(const PldaStats &stats):
    stats_(stats) {
  KALDI_ASSERT(stats.IsSorted());
  InitParameters();
}


double PldaEstimator::ComputeObjfPart1() const {
  // Returns the part of the objf relating to offsets from the class means.
  // within_class_count equals the sum over the classes, of the weight of that
  // class (normally 1) times (1 - #examples) of that class, which equals the
  // rank of the covariance we're modeling.  We imagine that we're modeling (1 -
  // #examples) separate samples, each with the within-class covariance.. the
  // argument is a little complicated and involves an orthogonal complement of a
  // matrix whose first row computes the mean.

  double within_class_count = stats_.example_weight_ - stats_.class_weight_,
      within_logdet, det_sign;
  SpMatrix<double> inv_within_var(within_var_);
  inv_within_var.Invert(&within_logdet, &det_sign);
  KALDI_ASSERT(det_sign == 1 && "Within-class covariance is singular");

  double objf = -0.5 * (within_class_count * (within_logdet + M_LOG_2PI * Dim())
                        + TraceSpSp(inv_within_var, stats_.offset_scatter_));
  return objf;
}

double PldaEstimator::ComputeObjfPart2() const {
  double tot_objf = 0.0;

  int32 n = -1; // the number of examples for the current class
  SpMatrix<double> combined_inv_var(Dim());
  // combined_inv_var = (between_var_ + within_var_ / n)^{-1}
  double combined_var_logdet;

  for (size_t i = 0; i < stats_.class_info_.size(); i++) {
    const ClassInfo &info = stats_.class_info_[i];
    if (info.num_examples != n) {
      n = info.num_examples;
      // variance of mean of n examples is between-class + 1/n * within-class
      combined_inv_var.CopyFromSp(between_var_);
      combined_inv_var.AddSp(1.0 / n, within_var_);
      combined_inv_var.Invert(&combined_var_logdet);
    }
    Vector<double> mean (*(info.mean));
    mean.AddVec(-1.0 / stats_.class_weight_, stats_.sum_);
    tot_objf += info.weight * -0.5 * (combined_var_logdet + M_LOG_2PI * Dim()
                                      + VecSpVec(mean, combined_inv_var, mean));
  }
  return tot_objf;
}

double PldaEstimator::ComputeObjf() const {
  double ans1 = ComputeObjfPart1(),
      ans2 = ComputeObjfPart2(),
      ans = ans1 + ans2,
      example_weights = stats_.example_weight_,
      normalized_ans = ans / example_weights;
  KALDI_LOG << "Within-class objf per sample is " << (ans1 / example_weights)
            << ", between-class is " << (ans2 / example_weights)
            << ", total is " << normalized_ans;
  return normalized_ans;
}

void PldaEstimator::InitParameters() {
  within_var_.Resize(Dim());
  within_var_.SetUnit();
  between_var_.Resize(Dim());
  between_var_.SetUnit();
}

void PldaEstimator::ResetPerIterStats() {
  within_var_stats_.Resize(Dim());
  within_var_count_ = 0.0;
  between_var_stats_.Resize(Dim());
  between_var_count_ = 0.0;
}

void PldaEstimator::GetStatsFromIntraClass() {
  within_var_stats_.AddSp(1.0, stats_.offset_scatter_);
  // Note: in the normal case, the expression below will be equal to the sum
  // over the classes, of (1-n), where n is the #examples for that class.  That
  // is the rank of the scatter matrix that "offset_scatter_" has for that
  // class. [if weights other than 1.0 are used, it will be different.]
  within_var_count_ += (stats_.example_weight_ - stats_.class_weight_);
}


/**
   GetStatsFromClassMeans() is the more complicated part of PLDA estimation.
   Let's suppose the mean of a particular class is m, and suppose that
   that class had n examples.  We suppose that
     m ~ N(0, between_var_ + 1/n within_var_)
   i.e. m is Gaussian-distributed with zero mean and variance equal to the
   between-class variance plus 1/n times the within-class variance.  Now, m
   is observed (as stats_.class_info_[something].mean).  We're doing an E-M
   procedure where we treat m as the sum of two variables:
     m = x + y
   where
     x ~ N(0, between_var_)
     y ~ N(0, 1/n * within_var_)
   The distribution of x will contribute to the stats of between_var_, and
   y to within_var_.  Now, y = m - x, so we can focus on working out the
   distribution of x and then we can very simply get the distribution of y.
   The following expression also includes the likelihood of y as a function of
   x.  Note: the C is different from line to line.

   log p(x) = C - 0.5 ( x^T between_var^{-1} x  + (m-x)^T (1/n within_var)^{-1) (m-x) )
            = C - 0.5 x^T (between_var^{-1} + n within_var^{-1}) x + x^T z

            where z = n within_var^{-1} m, and we can write this as:

   log p(x) = C - 0.5 (x-w)^T (between_var^{-1} + n within_var^{-1}) (x-w)

    where x^T (between_var^{-1} + n within_var^{-1}) w = x^T z, i.e.
       (between_var^{-1} + n within_var^{-1}) w = z = n within_var^{-1} m, so

       w = (between_var^{-1} + n within_var^{-1})^{-1} * n within_var^{-1} m

    We can see that the distribution over x is Gaussian, with mean w and variance
     (between_var^{-1} + n within_var^{-1})^{-1}.
    The distribution over y is Gaussian with the same variance, and mean m - w.
    So the update to the between-var stats will be:
       between-var-stats += w w^T + (between_var^{-1} + n within_var^{-1})^{-1}.
    and the update to the within-var stats will be:
       within-var-stats += n ( (m-w) (m-w)^T (between_var^{-1} + n within_var^{-1})^{-1} ).

    The drawback of this formulation is that each time we encounter a different
    value of n (number of examples) we will have to do a different matrix
    inversion.  We'll try to improve on this later using a suitable transform.
 */

void PldaEstimator::GetStatsFromClassMeans() {
  SpMatrix<double> between_var_inv(between_var_);
  between_var_inv.Invert();
  SpMatrix<double> within_var_inv(within_var_);
  within_var_inv.Invert();
  // mixed_var will equal (between_var^{-1} + n within_var^{-1})^{-1}.
  SpMatrix<double> mixed_var(Dim());
  int32 n = -1; // the current number of examples for the class.

  for (size_t i = 0; i < stats_.class_info_.size(); i++) {
    const ClassInfo &info = stats_.class_info_[i];
    double weight = info.weight;
    if (info.num_examples != n) {
      n = info.num_examples;
      mixed_var.CopyFromSp(between_var_inv);
      mixed_var.AddSp(n, within_var_inv);
      mixed_var.Invert();
    }
    Vector<double> m = *(info.mean); // the mean for this class.
    m.AddVec(-1.0 / stats_.class_weight_, stats_.sum_); // remove global mean
    Vector<double> temp(Dim()); // n within_var^{-1} m
    temp.AddSpVec(n, within_var_inv, m, 0.0);
    Vector<double> w(Dim()); // w, as defined in the comment.
    w.AddSpVec(1.0, mixed_var, temp, 0.0);
    Vector<double> m_w(m); // m - w
    m_w.AddVec(-1.0, w);
    between_var_stats_.AddSp(weight, mixed_var);
    between_var_stats_.AddVec2(weight, w);
    between_var_count_ += weight;
    within_var_stats_.AddSp(weight * n, mixed_var);
    within_var_stats_.AddVec2(weight * n, m_w);
    within_var_count_ += weight;
  }
}

void PldaEstimator::EstimateFromStats() {
  within_var_.CopyFromSp(within_var_stats_);
  within_var_.Scale(1.0 / within_var_count_);
  between_var_.CopyFromSp(between_var_stats_);
  between_var_.Scale(1.0 / between_var_count_);

  KALDI_LOG << "Trace of within-class variance is " << within_var_.Trace();
  KALDI_LOG << "Trace of between-class variance is " << between_var_.Trace();
}


void PldaEstimator::EstimateOneIter() {
  ResetPerIterStats();
  GetStatsFromIntraClass();
  GetStatsFromClassMeans();
  EstimateFromStats();
  KALDI_VLOG(2) << "Objective function is " << ComputeObjf();
}


void PldaEstimator::Estimate(const PldaEstimationConfig &config,
                             Plda *plda) {
  KALDI_ASSERT(stats_.example_weight_ > 0 && "Cannot estimate with no stats");
  for (int32 i = 0; i < config.num_em_iters; i++) {
    KALDI_LOG << "Plda estimation iteration " << i
              << " of " << config.num_em_iters;
    EstimateOneIter();
  }
  GetOutput(plda);
}


void PldaEstimator::GetOutput(Plda *plda) {
  plda->mean_ = stats_.sum_;
  plda->mean_.Scale(1.0 / stats_.class_weight_);
  KALDI_LOG << "Norm of mean of iVector distribution is "
            << plda->mean_.Norm(2.0);

  Matrix<double> transform1(Dim(), Dim());
  ComputeNormalizingTransform(within_var_, &transform1);
  // now transform is a matrix that if we project with it,
  // within_var_ becomes unit.

  // between_var_proj is between_var after projecting with transform1.
  SpMatrix<double> between_var_proj(Dim());
  between_var_proj.AddMat2Sp(1.0, transform1, kNoTrans, between_var_, 0.0);

  Matrix<double> U(Dim(), Dim());
  Vector<double> s(Dim());
  // Do symmetric eigenvalue decomposition between_var_proj = U diag(s) U^T,
  // where U is orthogonal.
  between_var_proj.Eig(&s, &U);

  KALDI_ASSERT(s.Min() >= 0.0);
  int32 n;
  s.ApplyFloor(0.0, &n);
  if (n > 0) {
    KALDI_WARN << "Floored " << n << " eigenvalues of between-class "
               << "variance to zero.";
  }
  // Sort from greatest to smallest eigenvalue.
  SortSvd(&s, &U);

  // The transform U^T will make between_var_proj diagonal with value s
  // (i.e. U^T U diag(s) U U^T = diag(s)).  The final transform that
  // makes within_var_ unit and between_var_ diagonal is U^T transform1,
  // i.e. first transform1 and then U^T.

  plda->transform_.Resize(Dim(), Dim());
  plda->transform_.AddMatMat(1.0, U, kTrans, transform1, kNoTrans, 0.0);
  plda->psi_ = s;

  KALDI_LOG << "Diagonal of between-class variance in normalized space is " << s;

  if (GetVerboseLevel() >= 2) { // at higher verbose levels, do a self-test
                                // (just tests that this function does what it
                                // should).
    SpMatrix<double> tmp_within(Dim());
    tmp_within.AddMat2Sp(1.0, plda->transform_, kNoTrans, within_var_, 0.0);
    KALDI_ASSERT(tmp_within.IsUnit(0.0001));
    SpMatrix<double> tmp_between(Dim());
    tmp_between.AddMat2Sp(1.0, plda->transform_, kNoTrans, between_var_, 0.0);
    KALDI_ASSERT(tmp_between.IsDiagonal(0.0001));
    Vector<double> psi(Dim());
    psi.CopyDiagFromSp(tmp_between);
    AssertEqual(psi, plda->psi_);
  }
  plda->ComputeDerivedVars();
}

void PldaUnsupervisedAdaptor::AddStats(double weight,
                                       const Vector<double> &ivector) {
  if (mean_stats_.Dim() == 0) {
    mean_stats_.Resize(ivector.Dim());
    variance_stats_.Resize(ivector.Dim());
  }
  KALDI_ASSERT(weight >= 0.0);
  tot_weight_ += weight;
  mean_stats_.AddVec(weight, ivector);
  variance_stats_.AddVec2(weight, ivector);
}

void PldaUnsupervisedAdaptor::AddStats(double weight,
                                       const Vector<float> &ivector) {
  Vector<double> ivector_dbl(ivector);
  this->AddStats(weight, ivector_dbl);
}

void PldaUnsupervisedAdaptor::UpdatePlda(const PldaUnsupervisedAdaptorConfig &config,
                                         Plda *plda) const {
  KALDI_ASSERT(tot_weight_ > 0.0);
  int32 dim = mean_stats_.Dim();
  KALDI_ASSERT(dim == plda->Dim());
  Vector<double> mean(mean_stats_);
  mean.Scale(1.0 / tot_weight_);
  SpMatrix<double> variance(variance_stats_);
  variance.Scale(1.0 / tot_weight_);
  variance.AddVec2(-1.0, mean);  // Make it the uncentered variance.

  // mean_diff of the adaptation data from the training data.  We optionally add
  // this to our total covariance matrix
  Vector<double> mean_diff(mean);
  mean_diff.AddVec(-1.0, plda->mean_);
  KALDI_ASSERT(config.mean_diff_scale >= 0.0);
  variance.AddVec2(config.mean_diff_scale, mean_diff);

  // update the plda's mean data-member with our adaptation-data mean.
  plda->mean_.CopyFromVec(mean);


  // transform_model_ is a row-scaled version of plda->transform_ that
  // transforms into the space where the total covariance is 1.0.  Because
  // plda->transform_ transforms into a space where the within-class covar is
  // 1.0 and the the between-class covar is diag(plda->psi_), we need to scale
  // each dimension i by 1.0 / sqrt(1.0 + plda->psi_(i))

  Matrix<double> transform_mod(plda->transform_);
  for (int32 i = 0; i < dim; i++)
    transform_mod.Row(i).Scale(1.0 / sqrt(1.0 + plda->psi_(i)));

  // project the variance of the adaptation set into this space where
  // the total covariance is unit.
  SpMatrix<double> variance_proj(dim);
  variance_proj.AddMat2Sp(1.0, transform_mod, kNoTrans,
                          variance, 0.0);

  // Do eigenvalue decomposition of variance_proj; this will tell us the
  // directions in which the adaptation-data covariance is more than
  // the training-data covariance.
  Matrix<double> P(dim, dim);
  Vector<double> s(dim);
  variance_proj.Eig(&s, &P);
  SortSvd(&s, &P);
  KALDI_LOG << "Eigenvalues of adaptation-data total-covariance in space where "
            << "training-data total-covariance is unit, is: " << s;

  // W, B are the (within,between)-class covars in the space transformed by
  // transform_mod.
  SpMatrix<double> W(dim), B(dim);
  for (int32 i = 0; i < dim; i++) {
    W(i, i) =           1.0 / (1.0 + plda->psi_(i)),
    B(i, i) = plda->psi_(i) / (1.0 + plda->psi_(i));
  }

  // OK, so variance_proj (projected by transform_mod) is P diag(s) P^T.
  // Suppose that after transform_mod we project by P^T.  Then the adaptation-data's
  // variance would be P^T P diag(s) P^T P = diag(s), and the PLDA model's
  // within class variance would be P^T W P and its between-class variance would be
  // P^T B P.  We'd still have that W+B = I in this space.
  // First let's compute these projected variances... we call the "proj2" because
  // it's after the data has been projected twice (actually, transformed, as there is no
  // dimension loss), by transform_mod and then P^T.

  SpMatrix<double> Wproj2(dim), Bproj2(dim);
  Wproj2.AddMat2Sp(1.0, P, kTrans, W, 0.0);
  Bproj2.AddMat2Sp(1.0, P, kTrans, B, 0.0);

  Matrix<double> Ptrans(P, kTrans);

  SpMatrix<double> Wproj2mod(Wproj2), Bproj2mod(Bproj2);

  for (int32 i = 0; i < dim; i++) {
    // For this eigenvalue, compute the within-class covar projected with this direction,
    // and the same for between.
    BaseFloat within = Wproj2(i, i),
        between = Bproj2(i, i);
    KALDI_LOG << "For " << i << "'th eigenvalue, value is " << s(i)
              << ", within-class covar in this direction is " << within
              << ", between-class is " << between;
    if (s(i) > 1.0) {
      double excess_eig = s(i) - 1.0;
      double excess_within_covar = excess_eig * config.within_covar_scale,
          excess_between_covar = excess_eig * config.between_covar_scale;
      Wproj2mod(i, i) += excess_within_covar;
      Bproj2mod(i, i) += excess_between_covar;
    } /*
        Below I was considering a method like below, to try to scale up
        the dimensions that had less variance than expected in our sample..
        this didn't help, and actually when I set that power to +0.2 instead
        of -0.5 it gave me an improvement on sre08.  But I'm not sure
        about this.. it just doesn't seem right.
      else {
      BaseFloat scale = pow(std::max(1.0e-10, s(i)), -0.5);
      BaseFloat max_scale = 10.0;  // I'll make this configurable later.
      scale = std::min(scale, max_scale);
      Ptrans.Row(i).Scale(scale);
      } */
  }

  // combined transform "transform_mod" and then P^T that takes us to the space
  // where {W,B}proj2{,mod} are.
  Matrix<double> combined_trans(dim, dim);
  combined_trans.AddMatMat(1.0, Ptrans, kNoTrans,
                           transform_mod, kNoTrans, 0.0);
  Matrix<double> combined_trans_inv(combined_trans);  // ... and its inverse.
  combined_trans_inv.Invert();

  // Wmod and Bmod are as Wproj2 and Bproj2 but taken back into the original
  // iVector space.
  SpMatrix<double> Wmod(dim), Bmod(dim);
  Wmod.AddMat2Sp(1.0, combined_trans_inv, kNoTrans, Wproj2mod, 0.0);
  Bmod.AddMat2Sp(1.0, combined_trans_inv, kNoTrans, Bproj2mod, 0.0);

  TpMatrix<double> C(dim);
  // Do Cholesky Wmod = C C^T.  Now if we use C^{-1} as a transform, we have
  // C^{-1} W C^{-T} = I, so it makes the within-class covar unit.
  C.Cholesky(Wmod);
  TpMatrix<double> Cinv(C);
  Cinv.Invert();

  // Bmod_proj is Bmod projected by Cinv.
  SpMatrix<double> Bmod_proj(dim);
  Bmod_proj.AddTp2Sp(1.0, Cinv, kNoTrans, Bmod, 0.0);
  Vector<double> psi_new(dim);
  Matrix<double> Q(dim, dim);
  // Do symmetric eigenvalue decomposition of Bmod_proj, so
  // Bmod_proj = Q diag(psi_new) Q^T
  Bmod_proj.Eig(&psi_new, &Q);
  SortSvd(&psi_new, &Q);
  // This means that if we use Q^T as a transform, then Q^T Bmod_proj Q =
  // diag(psi_new), hence Q^T diagonalizes Bmod_proj (while leaving the
  // within-covar unit).
  // The final transform we want, that projects from our original
  // space to our newly normalized space, is:
  // first Cinv, then Q^T, i.e. the
  // matrix Q^T Cinv.
  Matrix<double> final_transform(dim, dim);
  final_transform.AddMatTp(1.0, Q, kTrans, Cinv, kNoTrans, 0.0);

  KALDI_LOG << "Old diagonal of between-class covar was: "
            << plda->psi_ << ", new diagonal is "
            << psi_new;
  plda->transform_.CopyFromMat(final_transform);
  plda->psi_.CopyFromVec(psi_new);
}

} // namespace kaldi