fmpe.cc 31 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
// transform/fmpe.cc

// Copyright 2011-2012  Yanmin Qian  Johns Hopkins University (Author: Daniel Povey)

// 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 "transform/fmpe.h"
#include "util/text-utils.h"
#include "gmm/diag-gmm-normal.h"
#include "gmm/am-diag-gmm.h"
#include "hmm/transition-model.h"

namespace kaldi {

void Fmpe::SetContexts(std::string context_str) {
  // sets the contexts_ variable.
  using std::vector;
  using std::string;
  contexts_.clear();
  vector<string> ctx_vec; // splitting context_str on ":"
  SplitStringToVector(context_str, ":", false, &ctx_vec);
  contexts_.resize(ctx_vec.size());
  for (size_t i = 0; i < ctx_vec.size(); i++) {
    vector<string> pair_vec; // splitting ctx_vec[i] on ";"
    SplitStringToVector(ctx_vec[i], ";", false, &pair_vec);
    KALDI_ASSERT(pair_vec.size() != 0 && "empty context!");
    for (size_t j = 0; j < pair_vec.size(); j++) {
      vector<string> one_pair;
      SplitStringToVector(pair_vec[j], ",", false, &one_pair);
      KALDI_ASSERT(one_pair.size() == 2 &&
                   "Mal-formed context string: bad --context-expansion option?");
      int32 pos = 0;
      BaseFloat weight = BaseFloat(0);
      bool ok = ConvertStringToInteger(one_pair[0], &pos);
      ok = ConvertStringToReal(one_pair[1], &weight) && ok;
      if (!ok)
        KALDI_ERR << "Mal-formed context string: bad --context-expansion option?";
      contexts_[i].push_back(std::make_pair(pos, weight));
    }
  }
}

void Fmpe::ComputeC() {
  KALDI_ASSERT(gmm_.NumGauss() != 0.0);
  int32 dim = gmm_.Dim();

  // Getting stats from the GMM... assume the model is
  // correct.
  SpMatrix<double> x2_stats(dim);
  Vector<double> x_stats(dim);
  double tot_count = 0.0;
  DiagGmmNormal ngmm(gmm_);
  for (int32 pdf = 0; pdf < ngmm.NumGauss(); pdf++) {
    x2_stats.AddVec2(ngmm.weights_(pdf), ngmm.means_.Row(pdf));
    x2_stats.AddDiagVec(ngmm.weights_(pdf), ngmm.vars_.Row(pdf)); // add diagonal
    // covar to diagonal elements of x2_stats.
    x_stats.AddVec(ngmm.weights_(pdf), ngmm.means_.Row(pdf));
    tot_count += ngmm.weights_(pdf);
  }
  KALDI_ASSERT(tot_count != 0.0);
  x2_stats.Scale(1.0 / tot_count);
  x_stats.Scale(1.0 / tot_count);
  x2_stats.AddVec2(-1.0, x_stats); // subtract outer product of mean,
  // to get centered covariance.
  C_.Resize(dim);
  try {
    TpMatrix<double> Ctmp(dim); Ctmp.Cholesky(x2_stats);
    C_.CopyFromTp(Ctmp);
  } catch (...) {
    KALDI_ERR << "Error initializing fMPE object: cholesky of "
        "feature variance failed.  Probably code error, or NaN/inf in model";
  }
}

void Fmpe::ComputeStddevs() {
  const Matrix<BaseFloat> &inv_vars = gmm_.inv_vars();
  stddevs_.Resize(inv_vars.NumRows(), inv_vars.NumCols());
  stddevs_.CopyFromMat(inv_vars);
  stddevs_.ApplyPow(-0.5);
}


void Fmpe::ApplyContext(const MatrixBase<BaseFloat> &intermed_feat,
                        MatrixBase<BaseFloat> *feat_out) const {
  // Applies the temporal-context part of the transformation.
  int32 dim = FeatDim(), ncontexts = NumContexts(),
      T = intermed_feat.NumRows();
  KALDI_ASSERT(intermed_feat.NumCols() == dim * ncontexts &&
               intermed_feat.NumRows() == feat_out->NumRows()
               && feat_out->NumCols() == dim);
  // note: ncontexts == contexts_.size().
  for (int32 i = 0; i < ncontexts; i++) {
    // this_intermed_feat is the chunk of the "intermediate features"
    // that corresponds to this "context"
    SubMatrix<BaseFloat> this_intermed_feat(intermed_feat, 0, T,
                                            dim*i, dim);
    for (int32 j = 0; j < static_cast<int32>(contexts_[i].size()); j++) {
      int32 t_offset = contexts_[i][j].first;
      BaseFloat weight = contexts_[i][j].second;
      // Note: we could do this more efficiently using matrix operations,
      // but this doesn't dominate the computation and I think this is
      // clearer.
      for (int32 t_out = 0; t_out < T; t_out++) { // t_out indexes the output
        int32 t_in = t_out + t_offset; // t_in indexes the input.
        if (t_in >= 0 && t_in < T) // Discard frames outside range.
          feat_out->Row(t_out).AddVec(weight, this_intermed_feat.Row(t_in));
      }
    }
  }
}

void Fmpe::ApplyContextReverse(const MatrixBase<BaseFloat> &feat_deriv,
                               MatrixBase<BaseFloat> *intermed_feat_deriv)
    const {
  // Applies the temporal-context part of the transformation,
  // in reverse, for getting derivatives for training.
  int32 dim = FeatDim(), ncontexts = NumContexts(),
      T = feat_deriv.NumRows();
  KALDI_ASSERT(intermed_feat_deriv->NumCols() == dim * ncontexts &&
               intermed_feat_deriv->NumRows() == feat_deriv.NumRows()
               && feat_deriv.NumCols() == dim);
  // note: ncontexts == contexts_.size().
  for (int32 i = 0; i < ncontexts; i++) {
    // this_intermed_feat is the chunk of the derivative of
    // "intermediate features" that corresponds to this "context"
    // (this is output, in this routine).
    SubMatrix<BaseFloat> this_intermed_feat_deriv(*intermed_feat_deriv, 0, T,
                                                  dim*i, dim);
    for (int32 j = 0; j < static_cast<int32>(contexts_[i].size()); j++) {
      int32 t_offset = contexts_[i][j].first;
      BaseFloat weight = contexts_[i][j].second;
      // Note: we could do this more efficiently using matrix operations,
      // but this doesn't dominate the computation and I think this is
      // clearer.
      for (int32 t_out = 0; t_out < T; t_out++) { // t_out indexes the output
        int32 t_in = t_out + t_offset; // t_in indexes the input.
        if (t_in >= 0 && t_in < T) // Discard frames outside range.
          this_intermed_feat_deriv.Row(t_in).AddVec(weight,
                                                    feat_deriv.Row(t_out));
        // Note: the line above is where the work happens; it's the same
        // as in ApplyContext except reversing the input and output.
      }
    }
  }
}

void Fmpe::ApplyC(MatrixBase<BaseFloat> *feat_out, bool reverse) const {
  int32 T = feat_out->NumRows();
  Vector<BaseFloat> tmp(feat_out->NumCols());
  for (int32 t = 0; t < T; t++) {
    SubVector<BaseFloat> row(*feat_out, t);
    // Next line does: tmp = C_ * row
    tmp.AddTpVec(1.0, C_, (reverse ? kTrans : kNoTrans), row, 0.0);
    row.CopyFromVec(tmp);
  }
}

// Constructs the high-dim features and applies the main projection matrix
// projT_.  This projects from dimension ngauss*(dim+1) to dim*ncontexts.  Note:
// because the input vector of size ngauss*(dim+1) is sparse in a blocky way
// (i.e. each frame only has a couple of nonzero posteriors), we deal with
// sub-matrices of the projection matrix projT_.  We actually further optimize
// the code by taking all frames in a file that had nonzero posteriors for a
// particular Gaussian, and forming a matrix out of the corresponding
// high-dimensional features; we can then use a matrix-matrix multiply rather
// than using vector-matrix operations.

void Fmpe::ApplyProjection(const MatrixBase<BaseFloat> &feat_in,
                           const std::vector<std::vector<int32> > &gselect,
                           MatrixBase<BaseFloat> *intermed_feat) const {
  int32 dim = FeatDim(), ncontexts = NumContexts();  
  
  Vector<BaseFloat> post; // will be posteriors of selected Gaussians.
  Vector<BaseFloat> input_chunk(dim+1); // will be a segment of
  // the high-dimensional features.

  // "all_posts" is a vector of ((gauss-index, time-index), gaussian
  // posterior).
  // We'll compute the posterior information, sort it, and then
  // go through it in sorted order, which maintains memory locality
  // when accessing the projection matrix.
  // Note: if we really cared we could make this use level-3 BLAS
  // (matrix-matrix multiply), but we'd need to have a temporary
  // matrix for the output and input.
  std::vector<std::pair<std::pair<int32, int32>, BaseFloat> > all_posts;
  
  for (int32 t = 0; t < feat_in.NumRows(); t++) {
    SubVector<BaseFloat> this_feat(feat_in, t);
    gmm_.LogLikelihoodsPreselect(this_feat, gselect[t], &post);
    // At this point, post will contain log-likes of the selected
    // Gaussians.
    post.ApplySoftMax(); // Now they are posteriors (which sum to one).
    for (int32 i = 0; i < post.Dim(); i++) {
      int32 gauss = gselect[t][i];
      all_posts.push_back(std::make_pair(std::make_pair(gauss, t), post(i)));
    }
  }
  std::sort(all_posts.begin(), all_posts.end());
  
  bool optimize = true;

  if (!optimize) { // Why do we keep this un-optimized code around?
    // For clarity, so you can see what's going on, and for easier
    // comparision with ApplyProjectionReverse which is similar to this
    // un-optimized segment.  Both un-optimized and optimized versions
    // should give identical transforms (up to tiny roundoff differences).
    for (size_t i = 0; i < all_posts.size(); i++) {
      int32 gauss = all_posts[i].first.first, t = all_posts[i].first.second;
      SubVector<BaseFloat> this_feat(feat_in, t);
      SubVector<BaseFloat> this_intermed_feat(*intermed_feat, t);
      BaseFloat this_post = all_posts[i].second;
      SubVector<BaseFloat> this_stddev(stddevs_, gauss);

      // The next line is equivalent to setting input_chunk to
      // -this_post * the gaussian mean / (gaussian stddev).  Note: we use
      // the fact that mean * inv_var *  stddev == mean / stddev.
      input_chunk.Range(0, dim).AddVecVec(-this_post, gmm_.means_invvars().Row(gauss),
                                          this_stddev, 0.0);
      // The next line is equivalent to adding (feat / gaussian stddev) to
      // input_chunk, so now it contains (feat - mean) / stddev, which is
      // our "normalized" feature offset.
      input_chunk.Range(0, dim).AddVecDivVec(this_post, this_feat, this_stddev,
                                             1.0);
      // The last element of this input_chunk is the posterior itself
      // (between 0 and 1).
      input_chunk(dim) = this_post * config_.post_scale;

      // this_intermed_feat += [appropriate chjunk of projT_] * input_chunk.
      this_intermed_feat.AddMatVec(1.0, projT_.Range(gauss*(dim+1), dim+1,
                                                     0, dim*ncontexts),
                                   kTrans, input_chunk, 1.0);
    }
  } else {
    size_t i = 0;
    // We process the "posts" vector in chunks, where each chunk corresponds to
    // the same Gaussian index (but different times).
    while (i < all_posts.size()) {
      int32 gauss = all_posts[i].first.first;
      SubVector<BaseFloat> this_stddev(stddevs_, gauss),
          this_mean_invvar(gmm_.means_invvars(), gauss);
      SubMatrix<BaseFloat> this_projT_chunk(projT_, gauss*(dim+1), dim+1,
                                            0, dim*ncontexts);
      int32 batch_size; // number of posteriors with same Gaussian..
      for (batch_size = 0;
           batch_size+i < static_cast<int32>(all_posts.size()) &&
               all_posts[batch_size+i].first.first == gauss;
           batch_size++); // empty loop body.
      Matrix<BaseFloat> input_chunks(batch_size, dim+1);
      Matrix<BaseFloat> intermed_temp(batch_size, dim*ncontexts);
      for (int32 j = 0; j < batch_size; j++) { // set up "input_chunks".
        // To understand this code, first examine code and comments in "non-optimized"
        // code chunk above (the other branch of the if/else statement).
        int32 t = all_posts[i+j].first.second;
        SubVector<BaseFloat> this_feat(feat_in, t);
        SubVector<BaseFloat> this_input_chunk(input_chunks, j);
        BaseFloat this_post = all_posts[i+j].second;
        this_input_chunk.Range(0, dim).AddVecVec(-this_post,
                                                 this_mean_invvar,
                                                 this_stddev, 0.0);
        this_input_chunk.Range(0, dim).AddVecDivVec(this_post, this_feat,
                                                    this_stddev, 1.0);
        this_input_chunk(dim) = this_post * config_.post_scale;
      }
      // The next line is where most of the computation will happen,
      // during the feature computation phase.  We have rearranged
      // stuff so it's a matrix-matrix operation, for greater
      // efficiency (when using optimized libraries like ATLAS).
      intermed_temp.AddMatMat(1.0, input_chunks, kNoTrans,
                              this_projT_chunk, kNoTrans, 0.0);
      for (int32 j = 0; j < batch_size; j++) { // add data from
        // intermed_temp to the output "intermed_feat"
        int32 t = all_posts[i+j].first.second;
        SubVector<BaseFloat> this_intermed_feat(*intermed_feat, t);
        SubVector<BaseFloat> this_intermed_temp(intermed_temp, j);
        // this_intermed_feat += this_intermed_temp.
        this_intermed_feat.AddVec(1.0, this_intermed_temp);
      }
      i += batch_size;
    }
  }
}      



// This function does the reverse to ApplyProjection, for the case
// where we want the derivatives w.r.t. the projection matrix.
// It stores the positive and negative parts of this separately.
void Fmpe::ApplyProjectionReverse(const MatrixBase<BaseFloat> &feat_in,
                                  const std::vector<std::vector<int32> > &gselect,
                                  const MatrixBase<BaseFloat> &intermed_feat_deriv,
                                  MatrixBase<BaseFloat> *proj_deriv_plus,
                                  MatrixBase<BaseFloat> *proj_deriv_minus) const {
  int32 dim = FeatDim(), ncontexts = NumContexts();  
  
  Vector<BaseFloat> post; // will be posteriors of selected Gaussians.
  Vector<BaseFloat> input_chunk(dim+1); // will be a segment of
  // the high-dimensional features.

  // "all_posts" is a vector of ((gauss-index, time-index), gaussian
  // posterior).
  // We'll compute the posterior information, sort it, and then
  // go through it in sorted order, which maintains memory locality
  // when accessing the projection matrix.
  std::vector<std::pair<std::pair<int32, int32>, BaseFloat> > all_posts;
  
  for (int32 t = 0; t < feat_in.NumRows(); t++) {
    SubVector<BaseFloat> this_feat(feat_in, t);
    gmm_.LogLikelihoodsPreselect(this_feat, gselect[t], &post);
    // At this point, post will contain log-likes of the selected
    // Gaussians.
    post.ApplySoftMax(); // Now they are posteriors (which sum to one).
    for (int32 i = 0; i < post.Dim(); i++) {
      // The next few lines (where we set up "input_chunk") are identical
      // to ApplyProjection.
      int32 gauss = gselect[t][i];
      all_posts.push_back(std::make_pair(std::make_pair(gauss, t), post(i)));
    }
  }
  std::sort(all_posts.begin(), all_posts.end());
  for (size_t i = 0; i < all_posts.size(); i++) {
    int32 gauss = all_posts[i].first.first, t = all_posts[i].first.second;
    BaseFloat this_post = all_posts[i].second;
    SubVector<BaseFloat> this_feat(feat_in, t);    
    SubVector<BaseFloat> this_intermed_feat_deriv(intermed_feat_deriv, t);
    SubVector<BaseFloat> this_stddev(stddevs_, gauss);
    input_chunk.Range(0, dim).AddVecVec(-this_post, gmm_.means_invvars().Row(gauss),
                                        this_stddev, 0.0);
    input_chunk.Range(0, dim).AddVecDivVec(this_post, this_feat, this_stddev,
                                           1.0);
    input_chunk(dim) = this_post * config_.post_scale;

    // If not for accumulating the + and - parts separately, we would be
    // doing something like:
    // proj_deriv_.Range(0, dim*ncontexts, gauss*(dim+1), dim+1).AddVecVec(
    //                    1.0, this_intermed_feat_deriv, input_chunk);


    SubMatrix<BaseFloat> plus_chunk(*proj_deriv_plus, 
                                    gauss*(dim+1), dim+1,
                                    0, dim*ncontexts),
        minus_chunk(*proj_deriv_minus, 
                    gauss*(dim+1), dim+1,
                    0, dim*ncontexts);
          
    // This next function takes the rank-one matrix
    //  (input_chunk * this_intermed_deriv'), and adds the positive
    // part to proj_deriv_plus, and minus the negative part to
    // proj_deriv_minus.
    AddOuterProductPlusMinus(static_cast<BaseFloat>(1.0),
                             input_chunk,
                             this_intermed_feat_deriv,
                             &plus_chunk, &minus_chunk);
  }
}      

void Fmpe::ComputeFeatures(const MatrixBase<BaseFloat> &feat_in,
                           const std::vector<std::vector<int32> > &gselect,
                           Matrix<BaseFloat> *feat_out) const {
  int32 dim = FeatDim();
  KALDI_ASSERT(feat_in.NumRows() != 0 && feat_in.NumCols() == dim);
  KALDI_ASSERT(feat_in.NumRows() == static_cast<int32>(gselect.size()));
  feat_out->Resize(feat_in.NumRows(), feat_in.NumCols()); // will zero it.
  
  // Intermediate-dimension features
  Matrix<BaseFloat> intermed_feat(feat_in.NumRows(),
                                  dim * NumContexts());

  // Apply the main projection, from high-dim to intermediate
  // dimension (dim * NumContexts()).
  ApplyProjection(feat_in, gselect, &intermed_feat);

  // Apply the temporal context and reduces from
  // dimension dim*ncontexts to dim.
  ApplyContext(intermed_feat, feat_out);

  // Lastly, apply the the "C" matrix-- linear transform on the offsets.
  ApplyC(feat_out);
}


void Fmpe::AccStats(const MatrixBase<BaseFloat> &feat_in,
                    const std::vector<std::vector<int32> > &gselect,
                    const MatrixBase<BaseFloat> &direct_feat_deriv,
                    const MatrixBase<BaseFloat> *indirect_feat_deriv, // may be NULL
                    FmpeStats *fmpe_stats) const {
  SubMatrix<BaseFloat> stats_plus(fmpe_stats->DerivPlus());
  SubMatrix<BaseFloat> stats_minus(fmpe_stats->DerivMinus());
  int32 dim = FeatDim(), ncontexts = NumContexts();
  KALDI_ASSERT(feat_in.NumRows() != 0 && feat_in.NumCols() == dim);
  KALDI_ASSERT(feat_in.NumRows() == static_cast<int32>(gselect.size()));
  KALDI_ASSERT(SameDim(stats_plus, projT_) && SameDim(stats_minus, projT_) &&
               SameDim(feat_in, direct_feat_deriv));

  if (indirect_feat_deriv != NULL)
    fmpe_stats->AccumulateChecks(feat_in, direct_feat_deriv, *indirect_feat_deriv);
  
  Matrix<BaseFloat> feat_deriv(direct_feat_deriv); // "feat_deriv" is initially direct+indirect.
  if (indirect_feat_deriv != NULL)
    feat_deriv.AddMat(1.0, *indirect_feat_deriv);
  
  // We do the "*Reverse" version of each stage now, in reverse order.
  ApplyCReverse(&feat_deriv);
  
  Matrix<BaseFloat> intermed_feat_deriv(feat_in.NumRows(), dim*ncontexts);
  ApplyContextReverse(feat_deriv, &intermed_feat_deriv);
  
  ApplyProjectionReverse(feat_in, gselect, intermed_feat_deriv,
                         &stats_plus, &stats_minus);
}


void FmpeOptions::Write(std::ostream &os, bool binary) const {
  WriteToken(os, binary, context_expansion);
  WriteBasicType(os, binary, post_scale);
}
void FmpeOptions::Read(std::istream &is, bool binary) {
  ReadToken(is, binary, &context_expansion);
  ReadBasicType(is, binary, &post_scale);
}

Fmpe::Fmpe(const DiagGmm &gmm, const FmpeOptions &config): gmm_(gmm),
                                                          config_(config) {
  SetContexts(config.context_expansion);
  ComputeC();
  ComputeStddevs();
  projT_.Resize(NumGauss() * (FeatDim()+1), FeatDim() * NumContexts());
}

BaseFloat Fmpe::Update(const FmpeUpdateOptions &config,
                       const FmpeStats &stats) {
  SubMatrix<BaseFloat> proj_deriv_plus = stats.DerivPlus(),
      proj_deriv_minus = stats.DerivMinus();
  // tot_linear_objf_impr is the change in the actual
  // objective function if it were linear, i.e.
  //   objf-gradient . parameter-change
  // Note: none of this is normalized by the #frames (we don't have
  // this info here), so that is done at the script level.
  BaseFloat tot_linear_objf_impr = 0.0;
  int32 changed = 0; // Keep track of how many elements change sign.
  KALDI_ASSERT(SameDim(proj_deriv_plus, projT_) && SameDim(proj_deriv_minus, projT_));
  KALDI_ASSERT(proj_deriv_plus.Min() >= 0);
  KALDI_ASSERT(proj_deriv_minus.Min() >= 0);
  BaseFloat learning_rate = config.learning_rate,
      l2_weight = config.l2_weight;
  
  for (int32 i = 0; i < projT_.NumRows(); i++) {
    for (int32 j = 0; j < projT_.NumCols(); j++) {
      BaseFloat p = proj_deriv_plus(i, j), n = proj_deriv_minus(i, j),
          x = projT_(i, j);
      // Suppose the basic update (before regularization) is:
      // z <-- x  +   learning_rate * (p - n) / (p + n),
      // where z is the new parameter and x is the old one.
      // Here, we view (learning_rate / (p + n)) as a parameter-specific
      // learning rate.  In fact we view this update as the maximization
      // of an auxiliary function of the form:
      //  (z-x).(p-n)    - 0.5 (z - x)^2 (p+n)/learning_rate
      // and taking the derivative w.r.t z, we get:
      // Q'(z) =  (p-n) - (z - x) (p+n) / learning_rate
      // which we set to zero and solve for z, to get z = x + learning_rate.(p-n)/(p+n)
      // At this point we add regularization, a term of the form -l2_weight * z^2.
      // Our new auxiliary function derivative is:
      // Q(z) = -2.l2_weight.z + (p-n) - (z - x) (p+n) / learning_rate
      // We can write this as:
      // Q(z) = z . (-2.l2_weight - (p+n)/learning_rate)
      //        + (p-n) + x(p+n)/learning_rate
      // solving for z, we get:
      //      z = ((p-n) + x (p+n)/learning_rate) / (2.l2_weight + (p+n)/learning_rate)

      BaseFloat z = ((p-n) + x*(p+n)/learning_rate) / (2*l2_weight + (p+n)/learning_rate);
      // z is the new parameter value.

      tot_linear_objf_impr += (z-x) * (p-n); // objf impr based on linear assumption.
      projT_(i, j) = z;
      if (z*x < 0) changed++;
    }
  }
  KALDI_LOG << "Objf impr (assuming linear) is " << tot_linear_objf_impr;
  KALDI_LOG << ((100.0*changed)/(projT_.NumRows()*projT_.NumCols()))
            << "% of matrix elements changed sign.";
  return tot_linear_objf_impr;
}

// Note: we write the GMM first, without any other header.
// This way, the gselect code can treat the form on disk as
// a normal GMM object.
void Fmpe::Write(std::ostream &os, bool binary) const {
  if (gmm_.NumGauss() == 0)
    KALDI_ERR << "Fmpe::Write, object not initialized.";
  gmm_.Write(os, binary);
  config_.Write(os, binary);
  // stddevs_ are derived, don't write them.
  projT_.Write(os, binary);
  C_.Write(os, binary);
  // contexts_ are derived from config, don't write them.
}


void Fmpe::Read(std::istream &is, bool binary) {
  gmm_.Read(is, binary);
  config_.Read(is, binary);
  ComputeStddevs(); // computed from gmm.
  projT_.Read(is, binary);
  C_.Read(is, binary);
  SetContexts(config_.context_expansion);
}


BaseFloat ComputeAmGmmFeatureDeriv(const AmDiagGmm &am_gmm,
                                   const TransitionModel &trans_model,
                                   const Posterior &posterior,
                                   const MatrixBase<BaseFloat> &features,
                                   Matrix<BaseFloat> *direct_deriv,
                                   const AccumAmDiagGmm *model_diff,
                                   Matrix<BaseFloat> *indirect_deriv) {
  KALDI_ASSERT((model_diff != NULL) == (indirect_deriv != NULL));
  BaseFloat ans = 0.0;
  KALDI_ASSERT(posterior.size() == static_cast<size_t>(features.NumRows()));
  direct_deriv->Resize(features.NumRows(), features.NumCols());
  if (indirect_deriv != NULL)
    indirect_deriv->Resize(features.NumRows(), features.NumCols());
  Vector<BaseFloat> temp_vec(features.NumCols());
  Vector<double> temp_vec_dbl(features.NumCols());
  for (size_t i = 0; i < posterior.size(); i++) {
    for (size_t j = 0; j < posterior[i].size(); j++) {
      int32 tid = posterior[i][j].first,  // transition identifier.
          pdf_id = trans_model.TransitionIdToPdf(tid);
      BaseFloat weight = posterior[i][j].second;
      const DiagGmm &gmm = am_gmm.GetPdf(pdf_id);
      Vector<BaseFloat> gauss_posteriors;
      SubVector<BaseFloat> this_feat(features, i);
      SubVector<BaseFloat> this_direct_deriv(*direct_deriv, i);
      ans += weight * 
          gmm.ComponentPosteriors(this_feat, &gauss_posteriors);
      
      gauss_posteriors.Scale(weight);
      // The next line does: to i'th row of deriv, add
      // means_invvars^T * gauss_posteriors,
      // where each row of means_invvars is the mean times
      // diagonal inverse covariance... after transposing,
      // this becomes a weighted of these rows, weighted by
      // the posteriors.  This comes from the term
      //  feat^T * inv_var * mean
      // in the objective function.
      this_direct_deriv.AddMatVec(1.0, gmm.means_invvars(), kTrans,
                                  gauss_posteriors, 1.0);      

      // next line does temp_vec == inv_vars^T * gauss_posteriors,
      // which sets temp_vec to a weighted sum of the inv_vars,
      // weighed by Gaussian posterior.
      temp_vec.AddMatVec(1.0, gmm.inv_vars(), kTrans,
                         gauss_posteriors, 0.0);
      // Add to the derivative, -(this_feat .* temp_vec),
      // which is the term that comes from the -0.5 * inv_var^T feat_sq,
      // in the objective function (where inv_var is a vector, and feat_sq
      // is a vector of squares of the feature values).
      // Note: we have to do some messing about with double-precision here
      // because the stats only come in double precision.
      this_direct_deriv.AddVecVec(-1.0, this_feat, temp_vec, 1.0);
      if (model_diff != NULL && weight > 0.0) { // We need to get the indirect diff.
        // This "weight > 0.0" checks that this is the numerator stats, as the
        // fMPE indirect diff applies only to the ML stats-- CAUTION, this
        // code will only work as-is for fMMI (and the stats should not be
        // canceled), due to the assumption that ML stats == num stats.
        Vector<double> gauss_posteriors_dbl(gauss_posteriors);
        const AccumDiagGmm &deriv_acc = model_diff->GetAcc(pdf_id);
        // part of the derivative.  Note: we could just store the direct and
        // indirect derivatives together in one matrix, but it makes it easier
        // to accumulate certain diagnostics if we store them separately.
        SubVector<BaseFloat> this_indirect_deriv(*indirect_deriv, i);
        // note: deriv_acc.mean_accumulator() contains the derivative of
        // the objective function w.r.t. the "x stats" accumulated for
        // this GMM.  variance_accumulator() is the same for the "x^2 stats".
        temp_vec_dbl.AddMatVec(1.0, deriv_acc.mean_accumulator(), kTrans,
                               gauss_posteriors_dbl, 0.0);
        this_indirect_deriv.AddVec(1.0, temp_vec_dbl);
        temp_vec_dbl.AddMatVec(1.0, deriv_acc.variance_accumulator(), kTrans,
                               gauss_posteriors_dbl, 0.0);
        temp_vec.CopyFromVec(temp_vec_dbl); // convert to float.
        // next line because d(x^2 stats for Gaussian)/d(feature) =
        // 2 * (gaussian posterior) * feature.
        this_indirect_deriv.AddVecVec(2.0, this_feat, temp_vec, 1.0);
      }
    }
  }
  return ans;
}


SubMatrix<BaseFloat> FmpeStats::DerivPlus() const { // const-ness not preserved.
  KALDI_ASSERT(deriv.NumRows() != 0);
  int32 proj_num_rows = deriv.NumRows(),
      proj_num_cols = deriv.NumCols()/2;
  return SubMatrix<BaseFloat>(deriv, 0, proj_num_rows,
                              0, proj_num_cols);
}
SubMatrix<BaseFloat> FmpeStats::DerivMinus() const { // const-ness not preserved.
  KALDI_ASSERT(deriv.NumRows() != 0);
  int32 proj_num_rows = deriv.NumRows(),
      proj_num_cols = deriv.NumCols()/2;
  return SubMatrix<BaseFloat>(deriv, 0, proj_num_rows,
                              proj_num_cols, proj_num_cols);
}

void FmpeStats::Init(const Fmpe &fmpe) {
  int32 num_rows = fmpe.ProjectionTNumRows(),
      num_cols = fmpe.ProjectionTNumCols();
  deriv.Resize(num_rows, num_cols*2);

  int32 feat_dim = fmpe.FeatDim();
  checks.Resize(8, feat_dim);
}

void FmpeStats::AccumulateChecks(const MatrixBase<BaseFloat> &feats,
                                 const MatrixBase<BaseFloat> &direct_deriv,
                                 const MatrixBase<BaseFloat> &indirect_deriv) {
  int32 T = feats.NumRows(), dim = feats.NumCols();
  KALDI_ASSERT(direct_deriv.NumRows() == T && direct_deriv.NumCols() == dim &&
               indirect_deriv.NumRows() == T && indirect_deriv.NumCols() == dim);
  KALDI_ASSERT(checks.NumRows() == 8 && checks.NumCols() == dim);
  for (int32 t = 0; t < T; t++) {
    for (int32 d = 0; d < dim; d++) {
      BaseFloat zero = 0.0;
      checks(0, d) += std::max(zero, direct_deriv(t, d));
      checks(1, d) += std::max(zero, -direct_deriv(t, d));
      checks(2, d) += std::max(zero, indirect_deriv(t, d));
      checks(3, d) += std::max(zero, -indirect_deriv(t, d));
      checks(4, d) += std::max(zero, feats(t, d)*direct_deriv(t, d));
      checks(5, d) += std::max(zero, -feats(t, d)*direct_deriv(t, d));
      checks(6, d) += std::max(zero, feats(t, d)*indirect_deriv(t, d));
      checks(7, d) += std::max(zero, -feats(t, d)*indirect_deriv(t, d));
    }
  }
}

void FmpeStats::DoChecks() {
  if (checks.IsZero()) {
    KALDI_LOG << "No checks will be done, probably indirect derivative was not used.";
    return;
  }
  int32 dim = checks.NumCols();
  Vector<double> shift_check(dim), shift_check2(dim), scale_check(dim), scale_check2(dim);
  for (int32 d = 0; d < dim; d++) {
    // shiftnumerator = direct+indirect deriv-- should be zero.
    double shift_num = checks(0, d) - checks(1, d) + checks(2, d) - checks(3, d),
        shift_den = checks(0, d) + checks(1, d) + checks(2, d) + checks(3, d),
        shift_den2 = fabs(checks(0, d) - checks(1, d)) + fabs(checks(2, d) - checks(3, d));
    shift_check(d) = shift_num / shift_den;
    shift_check2(d) = shift_num / shift_den2;
    double scale_num = checks(4, d) - checks(5, d) + checks(6, d) - checks(7, d),
        scale_den = checks(4, d) + checks(5, d) + checks(6, d) + checks(7, d),
        scale_den2 = fabs(checks(4, d) - checks(5, d)) + fabs(checks(6, d) - checks(7, d));
    scale_check(d) = scale_num / scale_den;
    scale_check2(d) = scale_num / scale_den2;
  }

  KALDI_LOG << "Shift-check is as follows (should be in range +- 0.01 or less)."
            << shift_check;
  KALDI_LOG << "Scale-check is as follows (should be in range +- 0.01 or less)."
            << scale_check;
  KALDI_LOG << "Shift-check(2) is as follows: most elements should be in range +-0.1: "
            << shift_check2;
  KALDI_LOG << "Scale-check(2) is as follows: most elements should be in range +-0.1: "
            << scale_check2;
}

void FmpeStats::Write(std::ostream &os, bool binary) const {
  deriv.Write(os, binary);
  checks.Write(os, binary);
}

void FmpeStats::Read(std::istream &is, bool binary, bool add) {
  deriv.Read(is, binary, add);
  checks.Read(is, binary, add);
}


}  // End of namespace kaldi