nnet-normalize-component.cc 25.9 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
// nnet3/nnet-normalize-component.cc

// Copyright      2015-2017  Johns Hopkins University (author: Daniel Povey)
//                2015  Guoguo Chen
//                2015  Daniel Galvez

// 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 <iterator>
#include <sstream>
#include <algorithm>
#include <iomanip>
#include "nnet3/nnet-normalize-component.h"
#include "nnet3/nnet-parse.h"
#include "cudamatrix/cu-math.h"

namespace kaldi {
namespace nnet3 {

const BaseFloat NormalizeComponent::kSquaredNormFloor =
    pow(2.0, NormalizeComponent::kExpSquaredNormFloor);

NormalizeComponent::NormalizeComponent(const NormalizeComponent &other):
    input_dim_(other.input_dim_), block_dim_(other.block_dim_),
    target_rms_(other.target_rms_),
    add_log_stddev_(other.add_log_stddev_) { }

void NormalizeComponent::InitFromConfig(ConfigLine *cfl) {
  input_dim_ = 0;
  add_log_stddev_ = false;
  target_rms_ = 1.0;
  bool ok = cfl->GetValue("dim", &input_dim_) ||
      cfl->GetValue("input-dim", &input_dim_);
  block_dim_ = input_dim_;
  cfl->GetValue("block-dim", &block_dim_);
  cfl->GetValue("target-rms", &target_rms_);
  cfl->GetValue("add-log-stddev", &add_log_stddev_);
  if (!ok || cfl->HasUnusedValues() || input_dim_ <= 0 || target_rms_ <= 0.0 ||
      block_dim_ <= 0 || input_dim_ % block_dim_ != 0)
    KALDI_ERR << "Invalid initializer for layer of type "
              << Type() << ": \"" << cfl->WholeLine() << "\"";
}

void NormalizeComponent::Read(std::istream &is, bool binary) {
  std::string token;
  ReadToken(is, binary, &token);
  if (token == "<NormalizeComponent>") {
    ReadToken(is, binary, &token);
  }
  KALDI_ASSERT(token == "<Dim>" || token == "<InputDim>");
  ReadBasicType(is, binary, &input_dim_); // Read dimension.
  ReadToken(is, binary, &token);
  if (token == "<BlockDim>") {
    ReadBasicType(is, binary, &block_dim_);
    ReadToken(is, binary, &token);
  } else {
    block_dim_ = input_dim_;
  }
  // read target_rms_ if it is available.
  if (token == "<TargetRms>") {
    ReadBasicType(is, binary, &target_rms_);
    ReadToken(is, binary, &token);
  }
  //  Read add_log_stddev_ token, if it is available.
  if (token == "<AddLogStddev>") {
    ReadBasicType(is, binary, &add_log_stddev_);
    ReadToken(is, binary, &token);
  } else {
    add_log_stddev_ = false;
  }
  if (token == "<ValueAvg>") {
    // back-compatibility code.
    CuVector<double> temp;
    temp.Read(is, binary);
    ExpectToken(is, binary, "<DerivAvg>");
    temp.Read(is, binary);
    ExpectToken(is, binary, "<Count>");
    double count;
    ReadBasicType(is, binary, &count);
    ReadToken(is, binary, &token);
  }
  KALDI_ASSERT(token == "</NormalizeComponent>");
}

void NormalizeComponent::Write(std::ostream &os, bool binary) const {
  WriteToken(os, binary, "<NormalizeComponent>");
  WriteToken(os, binary, "<InputDim>");
  WriteBasicType(os, binary, input_dim_);
  if (block_dim_ != input_dim_) {
    WriteToken(os, binary, "<BlockDim>");
    WriteBasicType(os, binary, block_dim_);
  }
  WriteToken(os, binary, "<TargetRms>");
  WriteBasicType(os, binary, target_rms_);
  WriteToken(os, binary, "<AddLogStddev>");
  WriteBasicType(os, binary, add_log_stddev_);
  WriteToken(os, binary, "</NormalizeComponent>");
}

std::string NormalizeComponent::Info() const {
  std::ostringstream stream;
  stream << Type() << ", input-dim=" << InputDim()
         << ", output-dim=" << OutputDim() << ", target-rms=" << target_rms_
         << ", add-log-stddev=" << std::boolalpha << add_log_stddev_;
  if (block_dim_ != input_dim_)
    stream << ", block-dim=" << block_dim_;
  return stream.str();
}

// The output y_i = scale * x_i,
// and we want to RMS value of the y_i to equal target_rms,
// so y^t y = D * target_rms^2 (if y is one row of the input).
// we need to have scale = 1.0 / sqrt(x^t x / (D * target_rms^2)).
// there is also flooring involved, to avoid division-by-zero
// problems.  It's important for the backprop, that the floor's
// square root is exactly representable as float.
// If add_log_stddev_ is true, log(max(epsi, sqrt(x^t x / D)))
// is an extra dimension of the output.
void* NormalizeComponent::Propagate(const ComponentPrecomputedIndexes *indexes,
                                   const CuMatrixBase<BaseFloat> &in,
                                   CuMatrixBase<BaseFloat> *out) const {
  KALDI_ASSERT(in.NumCols() == InputDim() && out->NumCols() == OutputDim() &&
               in.NumRows() == out->NumRows());
  if (block_dim_ != input_dim_) {
    int32 num_blocks = input_dim_ / block_dim_,
        new_num_rows = in.NumRows() * num_blocks,
        output_block_dim = block_dim_ + (add_log_stddev_ ? 1 : 0);
    KALDI_ASSERT(in.Stride() == in.NumCols() && out->Stride() == out->NumCols());
    CuSubMatrix<BaseFloat> in_reshaped(in.Data(), new_num_rows,
                                       block_dim_, block_dim_),
        out_reshaped(out->Data(), new_num_rows,
                     output_block_dim, output_block_dim);
    cu::NormalizePerRow(in_reshaped, target_rms_, add_log_stddev_,
                        &out_reshaped);
  } else {
    cu::NormalizePerRow(in, target_rms_, add_log_stddev_, out);
  }
  return NULL;
}

/*
  A note on the derivative of NormalizeComponent...
  let both row_in and row_out be vectors of dimension D.
  Let p = row_in^T row_in / (D * target_rms^2), and let
  f = 1.0 / sqrt(max(kSquaredNormFloor, p)), and we compute row_out as:
  row_out = f row_in.
  Suppose we have a quantity deriv_out which is the derivative
  of the objective function w.r.t. row_out.  We want to compute
  deriv_in which is the derivative of the objective function w.r.t.
  row_in.  Let the objective function be F.  One term is obvious: we have
  deriv_in = f deriv_out + ....
  next we have to take into account the derivative that gets back-propagated
  through f.  Obviously, dF/df = deriv_out^T row_in.
  And df/dp = (p <= kSquaredNormFloor ? 0.0 : -0.5 p^{-1.5}) = (f == 1.0 / sqrt(kSquaredNormFloor) ? 0.0 : -0.5 f^3),
  and dp/d(row_in) = 2/(D * target_rms^2) row_in. [it's vector_valued].
  So this term in dF/d(row_in) equals:
  dF/df df/dp dp/d(row_in)   =    2/(D * target_rms^2) (f == 1.0 / sqrt(kSquaredNormFloor)  ? 0.0 : -0.5 f^3) (deriv_out^T row_in) row_in
  So
  deriv_in = f deriv_out + (f == 1.0 ? 0.0 : -f^3  / (D * target_rms^2) ) (deriv_out^T row_in) row_in

  if add_log_stddev_ true, the deriv_in has another term as
  dF/dx_i = dF/df . df/dx_i => df/dx_i = x_i/(x^T x)
*/
void NormalizeComponent::Backprop(const std::string &debug_info,
                                  const ComponentPrecomputedIndexes *indexes,
                                  const CuMatrixBase<BaseFloat> &in_value,
                                  const CuMatrixBase<BaseFloat> &, // out_value
                                  const CuMatrixBase<BaseFloat> &out_deriv,
                                  void *memo,
                                  Component *to_update,
                                  CuMatrixBase<BaseFloat> *in_deriv) const {
  if (!in_deriv)
    return;
  if (block_dim_ != input_dim_) {
    int32 num_blocks = input_dim_ / block_dim_,
        new_num_rows = in_value.NumRows() * num_blocks,
        output_block_dim = block_dim_ + (add_log_stddev_ ? 1 : 0);
    KALDI_ASSERT(in_value.Stride() == in_value.NumCols() &&
                 out_deriv.Stride() == out_deriv.NumCols() &&
                 in_deriv->Stride() == in_deriv->NumCols());
    CuSubMatrix<BaseFloat> in_value_reshaped(in_value.Data(), new_num_rows,
                                             block_dim_, block_dim_),
        out_deriv_reshaped(out_deriv.Data(), new_num_rows,
                           output_block_dim, output_block_dim),
        in_deriv_reshaped(in_deriv->Data(), new_num_rows,
                          block_dim_, block_dim_);
    cu::DiffNormalizePerRow(in_value_reshaped, out_deriv_reshaped, target_rms_,
                            add_log_stddev_, &in_deriv_reshaped);
  } else {
    cu::DiffNormalizePerRow(in_value, out_deriv, target_rms_, add_log_stddev_,
                            in_deriv);
  }
}

void BatchNormComponent::ComputeDerived() {
  if (!test_mode_) {
    offset_.Resize(0);
    scale_.Resize(0);
    return;
  }

  if (count_ == 0.0) {
    KALDI_WARN << "Test-mode is set but there is no data count.  "
        "Creating random counts.  This only makes sense "
        "in unit-tests (or compute_prob_*.0.log).  If you see this "
        "elsewhere, something is very wrong.";
    count_ = 1.0;
    stats_sum_.SetRandn();
    stats_sumsq_.SetRandn();
    stats_sumsq_.AddVecVec(1.0, stats_sum_, stats_sum_, 1.0);
  }

  offset_.Resize(block_dim_);
  scale_.Resize(block_dim_);
  offset_.CopyFromVec(stats_sum_);
  offset_.Scale(-1.0 / count_);
  // now offset_ is -mean.
  scale_.CopyFromVec(stats_sumsq_);
  scale_.Scale(1.0 / count_);
  scale_.AddVecVec(-1.0, offset_, offset_, 1.0);
  // now scale_ is variance.
  // Mathematically the ApplyFloor statement should be a no-op; this is in case
  // of numerical roundoff.
  scale_.ApplyFloor(0.0);
  scale_.Add(epsilon_);
  BaseFloat power = -0.5;
  scale_.ApplyPow(power);
  // now scale_ = min(variance, epsilon)^power
  // next, multiply by the target RMS (normally 1.0).
  scale_.Scale(target_rms_);
  offset_.MulElements(scale_);
  // now offset_ is -(scale*mean).
}

void BatchNormComponent::SetTestMode(bool test_mode) {
  test_mode_ = test_mode;
  ComputeDerived();
}

void BatchNormComponent::Check() const {
  KALDI_ASSERT(dim_ > 0 && block_dim_ > 0 && dim_ % block_dim_ == 0 &&
               epsilon_ > 0.0 && target_rms_ > 0.0);
}

BatchNormComponent::BatchNormComponent(const BatchNormComponent &other):
    dim_(other.dim_), block_dim_(other.block_dim_),
    epsilon_(other.epsilon_), target_rms_(other.target_rms_),
    test_mode_(other.test_mode_), count_(other.count_),
    stats_sum_(other.stats_sum_), stats_sumsq_(other.stats_sumsq_) {
  ComputeDerived();
  Check();
}


std::string BatchNormComponent::Info() const {
  std::ostringstream stream;
  stream << Type() << ", dim=" << dim_ << ", block-dim=" << block_dim_
         << ", epsilon=" << epsilon_ << ", target-rms=" << target_rms_
         << ", count=" << count_
         << ", test-mode=" << (test_mode_ ? "true" : "false");
  if (count_ > 0) {
    Vector<BaseFloat> mean(stats_sum_), var(stats_sumsq_);
    mean.Scale(1.0 / count_);
    var.Scale(1.0 / count_);
    // subtract mean^2 from var.
    var.AddVecVec(-1.0, mean, mean, 1.0);
    var.ApplyFloor(0.0);
    var.ApplyPow(0.5);  // make it the stddev.
    stream << ", data-mean=" << SummarizeVector(mean)
           << ", data-stddev=" << SummarizeVector(var);
  }
  return stream.str();
}

void BatchNormComponent::InitFromConfig(ConfigLine *cfl) {
  dim_ = -1;
  block_dim_ = -1;
  epsilon_ = 1.0e-03;
  target_rms_ = 1.0;
  test_mode_ = false;
  bool ok = cfl->GetValue("dim", &dim_);
  cfl->GetValue("block-dim", &block_dim_);
  cfl->GetValue("epsilon", &epsilon_);
  cfl->GetValue("target-rms", &target_rms_);
  cfl->GetValue("test-mode", &test_mode_);
  if (!ok || dim_ <= 0) {
    KALDI_ERR << "BatchNormComponent must have 'dim' specified, and > 0";
  }
  if (block_dim_ == -1)
    block_dim_ = dim_;
  if (!(block_dim_ > 0 && dim_ % block_dim_ == 0 &&
        epsilon_ > 0 && target_rms_ > 0))
    KALDI_ERR << "Invalid configuration in BatchNormComponent.";
  if (cfl->HasUnusedValues())
    KALDI_ERR << "Could not process these elements in initializer: "
              << cfl->UnusedValues();
  count_ = 0;
  stats_sum_.Resize(block_dim_);
  stats_sumsq_.Resize(block_dim_);
  if (test_mode_) {
    ComputeDerived();
  }
}



/*
  BATCHNORM_MATH

  This comment describes the equations involved in batch normalization, and
  derives the forward and back-propagation.

  This is all dimension-by-dimension, so we just imagine the inputs
  are scalars x(i), for i=0 .. n-1.

  FORWARD PASS:

  Let 'power' be a constant, equal to -0.5 for regular batch-norm.

  To simplify the math we (conceptually, not physically) do the normalization in
  two stages: first mean, then variance, so we have x(i) -> y(i) -> z(i).

  The name 'rscale' means 'raw scale', meaning the scale before including
  target-rms.  Later we'll define 'scale = target-rms * rscale', to make some
  of the actual computations slightly more efficient.

  Define:   mean = 1/I * sum_i x(i)
            y(i) = x(i) - mean

            var = 1/I \sum_i y(i)^2
         rscale = sqrt(var + epsilon)^power   <---- For regular batchnorm, power == -0.5.
           z(i) = target-rms * rscale * y(i)


  Most of the rest of this comment derives how to compute the derivatives.  If
  you just want the formulas, please skip to the string 'BACKWARD PASS' below.

  We'll use a notation where an apostrophe on something means (the derivative of
  the objective function w.r.t. that thing), so y'(i) is df/dy(i), and so on.
  We are given y'(i).  Propagating the derivatives backward:

    rscale' = (sum_i y(i) z'(i)) * target-rms
            = (sum_i z(i) z'(i)) / rscale

  [ note: d(rscale)/d(var) = power * (var + epsilon)^{power - 1}
                           = power * rscale^{(power-1)/power}  ]

    var' = rscale' * power * rscale^{(power-1)/power}
         = power * (\sum_i z'(i) z(i)) * rscale^{(power-1)/power - 1}
         = power * (\sum_i z'(i) z(i)) * rscale^{-1/power}

  [note: the following formula is of the form "direct term" + "indirect term"]
    y'(i) =  z'(i) * target-rms * rscale   +    2/I y(i) var'

  Now, the above is inconvenient because it contains y(i) which is an intermediate
  quantity.  We reformulate in terms of z(i), using y(i) = z(i) / (target-rms * rscale), so:

  defining
   var_deriv_mod = 2/I * var' / (target-rms * rscale)
                 = 2/I * power/target-rms * (\sum_i z'(i) z(i)) * rscale^{-(1+power)/power}
 we have:
    y'(i) =  z'(i) * target-rms * rscale   +    z(i) var_deriv_mod

 Now,
    mean' = \sum_i y'(i)
          = (target-rms * rscale * \sum_i z'(i))  +  (var_deriv_mod \sum_i z(i))
     [... and the 2nd term above is zero when summed over i, because \sum_i z(i) is zero, ...]
          = target-rms * rscale * \sum_i z(i)
 and:
    x'(i) =  z'(i) * target-rms * rscale   +    z(i) var_deriv_mod   -  1/I mean'
          =  z'(i) * target-rms * rscale   +    z(i) var_deriv_mod   -  1/I * target-rms * rscale * \sum_i z'(i)
          =  target-rms * rscale * (z'(i) - 1/I * \sum_i z'(i))  +  z(i) var_deriv_mod

    It will simplify the code if we define:

      scale = target-rms * rscale.  This way, we can write as follows:

  BACKWARD PASS (recap):

   var_deriv_mod = 2 * power * target-rms^{1/power} * (1/I \sum_i z'(i) z(i)) * scale^{-(1+power)/power}
                .. which for power = -0.5, simplifies to:
   var_deriv_mod = -1.0 / (target-rms^2) * (1/I \sum_i z'(i) z(i)) * scale

           x'(i) = scale * (z'(i) - 1/I * \sum_i z'(i))  + z(i) var_deriv_mod

  */
void* BatchNormComponent::Propagate(const ComponentPrecomputedIndexes *indexes,
                                    const CuMatrixBase<BaseFloat> &in,
                                    CuMatrixBase<BaseFloat> *out) const {
  KALDI_ASSERT(SameDim(in, *out) &&
               (in.NumCols() == dim_ || in.NumCols() == block_dim_));
  if (in.NumCols() != block_dim_) {
    // if block_dim_ != dim_, we recurse; this helps keep the main code
    // simple.
    KALDI_ASSERT(in.Stride() == in.NumCols() && out->Stride() == out->NumCols());
    int32 ratio = dim_ / block_dim_, orig_rows = in.NumRows(),
        orig_cols = in.NumCols(), new_rows = orig_rows * ratio,
        new_cols = orig_cols / ratio;
    CuSubMatrix<BaseFloat> in_reshaped(in.Data(), new_rows, new_cols, new_cols),
        out_reshaped(out->Data(), new_rows, new_cols, new_cols);
    return Propagate(indexes, in_reshaped, &out_reshaped);
  }

  // From this point, we can assume that the num-cols of 'in' and 'out'
  // equals block_dim_.

  if (!test_mode_) {
    // search in the comment above for FORWARD PASS to see what is being
    // implemented here.
    // if this takes too much time due to multiple different CUDA calls,
    // we'll consider making a single kernel for some of it.
    Memo *memo = new Memo;
    int32 num_frames = in.NumRows(), dim = block_dim_;
    memo->num_frames = num_frames;
    memo->mean_uvar_scale.Resize(5, dim);
    CuSubVector<BaseFloat> mean(memo->mean_uvar_scale, 0),
        uvar(memo->mean_uvar_scale, 1),
        scale(memo->mean_uvar_scale, 2);
    mean.AddRowSumMat(1.0 / num_frames, in, 0.0);
    uvar.AddDiagMat2(1.0 / num_frames, in, kTrans, 0.0);
    scale.CopyFromVec(uvar);

    // by applying this scale at this point, we save a multiply later on.
    BaseFloat var_scale = 1.0 / (target_rms_ * target_rms_);
    scale.AddVecVec(-var_scale, mean, mean, var_scale);
    // at this point, 'scale' contains just the variance (times target-rms^{-2}).
    scale.ApplyFloor(0.0);
    scale.Add(var_scale * epsilon_);
    // Now 'scale' contains the variance floored to zero and then with epsilon
    // added [both times 1/target-rms^2].
    scale.ApplyPow(-0.5);
    // now 'scale' is the actual scale we'll use.

    // the next command will do no work if out == in, for in-place propagation.
    out->CopyFromMat(in);
    out->AddVecToRows(-1.0, mean, 1.0);
    out->MulColsVec(scale);
    return static_cast<void*>(memo);
  } else {
    if (offset_.Dim() != block_dim_) {
      if (count_ == 0)
        KALDI_ERR << "Test mode set in BatchNormComponent, but no stats.";
      else  // why was ComputeDerived() not called?
        KALDI_ERR << "Code error in BatchNormComponent";
    }
    out->CopyFromMat(in);
    out->MulColsVec(scale_);
    out->AddVecToRows(1.0, offset_, 1.0);
    return NULL;
  }
}

void BatchNormComponent::Backprop(
    const std::string &debug_info,
    const ComponentPrecomputedIndexes *indexes,
    const CuMatrixBase<BaseFloat> &in_value,  // unused
    const CuMatrixBase<BaseFloat> &out_value,
    const CuMatrixBase<BaseFloat> &out_deriv,
    void *memo_in,
    Component *to_update,  // unused
    CuMatrixBase<BaseFloat> *in_deriv) const {

  KALDI_ASSERT(SameDim(out_value, out_deriv) &&
               SameDim(out_value, *in_deriv) &&
               (out_value.NumCols() == dim_ ||
                out_value.NumCols() == block_dim_));
  if (out_value.NumCols() != block_dim_) {
    // if block_dim_ != dim_, we recurse; this helps keep the main code
    // simple.
    KALDI_ASSERT(out_value.Stride() == out_value.NumCols() &&
                 out_deriv.Stride() == out_deriv.NumCols() &&
                 in_deriv->Stride() == in_deriv->NumCols());
    int32 ratio = dim_ / block_dim_,
        orig_rows = out_value.NumRows(),
        orig_cols = out_value.NumCols(),
        new_rows = orig_rows * ratio, new_cols = orig_cols / ratio;
    CuSubMatrix<BaseFloat> out_value_reshaped(out_value.Data(), new_rows,
                                              new_cols, new_cols),
        out_deriv_reshaped(out_deriv.Data(), new_rows, new_cols, new_cols),
        in_deriv_reshaped(in_deriv->Data(), new_rows, new_cols, new_cols);
    // we'll never use in_value, so pass it in unchanged.
    Backprop(debug_info, indexes, in_value,
             out_value_reshaped, out_deriv_reshaped,
             memo_in, to_update, &in_deriv_reshaped);
    return;
  }

  Memo *memo = static_cast<Memo*>(memo_in);

  if (!test_mode_) {
    // search above for BACKWARD PASS for a comment describing the math.
    KALDI_ASSERT(memo != NULL && "memo not passed into backprop");
    int32 num_frames = memo->num_frames;
    KALDI_ASSERT(out_value.NumRows() == num_frames);
    CuSubVector<BaseFloat>
        scale(memo->mean_uvar_scale, 2),
        var_deriv_mod(memo->mean_uvar_scale, 3),
        temp(memo->mean_uvar_scale, 4);

    // var_deriv_mod is going to contain:
    //  2 * power * target-rms^{1/power} * (1/I \sum_i z'(i) z(i)) * scale^{-(1+power)/power}
    // which for power = -0.5 simplifies to:
    // -1.0 / (target_rms * target_rms).
    // but for now we don't have the power of 'scale', we'll add that later.
    BaseFloat coeff = -1.0 / (target_rms_ * target_rms_ * num_frames);

    var_deriv_mod.AddDiagMatMat(coeff, out_value, kTrans,
                                out_deriv, kNoTrans, 0.0);
    var_deriv_mod.MulElements(scale);

    temp.AddRowSumMat(-1.0 / num_frames, out_deriv, 0.0);
    // the following statement does no work if in_deriv and out_deriv are the
    // same matrix.
    in_deriv->CopyFromMat(out_deriv);
    in_deriv->AddVecToRows(1.0, temp);
    // At this point, *in_deriv contains
    // (z'(i) - 1/I * \sum_i z'(i))
    in_deriv->MulColsVec(scale);
    // At this point, *in_deriv contains
    // scale * (z'(i) - 1/I * \sum_i z'(i))

    in_deriv->AddMatDiagVec(1.0, out_value, kNoTrans,
                            var_deriv_mod, 1.0);

    // At this point, *in_deriv contains what we described in the comment
    // starting BATCHNORM_MATH as:
    // x'(i) = scale * (z'(i) - 1/I * \sum_i z'(i))  + z(i) var_deriv_mod
  } else {
    KALDI_ASSERT(offset_.Dim() == block_dim_);
    // the next call does no work if they point to the same memory.
    in_deriv->CopyFromMat(out_deriv);
    in_deriv->MulColsVec(scale_);
  }
}

void BatchNormComponent::StoreStats(
    const CuMatrixBase<BaseFloat> &in_value,
    const CuMatrixBase<BaseFloat> &out_value,
    void *memo_in) {
  // in test mode this component does not store stats, it doesn't provide the
  // kStoresStats flag.
  KALDI_ASSERT(!test_mode_);
  KALDI_ASSERT(out_value.NumCols() == dim_ || out_value.NumCols() == block_dim_);
  if (out_value.NumCols() != block_dim_) {
    // if block_dim_ != dim_, we recurse; this helps keep the main code
    // simple.
    KALDI_ASSERT(out_value.Stride() == out_value.NumCols());
    int32 ratio = dim_ / block_dim_,
        orig_rows = out_value.NumRows(),
        orig_cols = out_value.NumCols(),
        new_rows = orig_rows * ratio, new_cols = orig_cols / ratio;
    CuSubMatrix<BaseFloat> out_value_reshaped(out_value.Data(), new_rows,
                                              new_cols, new_cols);
    // we'll never use in_value, so just pass it in unchanged.
    StoreStats(in_value, out_value_reshaped, memo_in);
    return;
  }

  Memo *memo = static_cast<Memo*>(memo_in);
  KALDI_ASSERT(out_value.NumRows() == memo->num_frames);

  CuSubVector<BaseFloat> mean(memo->mean_uvar_scale, 0),
      uvar(memo->mean_uvar_scale, 1);
  KALDI_ASSERT(mean.Dim() == block_dim_ && memo->num_frames > 0);
  BaseFloat num_frames = memo->num_frames;
  if (stats_sum_.Dim() != block_dim_) {
    stats_sum_.Resize(block_dim_);
    stats_sumsq_.Resize(block_dim_);
    KALDI_ASSERT(count_ == 0);
  }
  count_ += num_frames;
  stats_sum_.AddVec(num_frames, mean, 1.0);
  stats_sumsq_.AddVec(num_frames, uvar, 1.0);
}

void BatchNormComponent::Read(std::istream &is, bool binary) {
  ExpectOneOrTwoTokens(is, binary, "<BatchNormComponent>", "<Dim>");
  ReadBasicType(is, binary, &dim_);
  ExpectToken(is, binary, "<BlockDim>");
  ReadBasicType(is, binary, &block_dim_);
  ExpectToken(is, binary, "<Epsilon>");
  ReadBasicType(is, binary, &epsilon_);
  ExpectToken(is, binary, "<TargetRms>");
  ReadBasicType(is, binary, &target_rms_);
  ExpectToken(is, binary, "<TestMode>");
  ReadBasicType(is, binary, &test_mode_);
  ExpectToken(is, binary, "<Count>");
  ReadBasicType(is, binary, &count_);
  ExpectToken(is, binary, "<StatsMean>");
  stats_sum_.Read(is, binary);
  ExpectToken(is, binary, "<StatsVar>");
  stats_sumsq_.Read(is, binary);
  stats_sumsq_.AddVecVec(1.0, stats_sum_, stats_sum_, 1.0);
  stats_sum_.Scale(count_);
  stats_sumsq_.Scale(count_);
  ExpectToken(is, binary, "</BatchNormComponent>");
  ComputeDerived();
  Check();
}

void BatchNormComponent::Write(std::ostream &os, bool binary) const {
  Check();
  WriteToken(os, binary, "<BatchNormComponent>");
  WriteToken(os, binary, "<Dim>");
  WriteBasicType(os, binary, dim_);
  WriteToken(os, binary, "<BlockDim>");
  WriteBasicType(os, binary, block_dim_);
  WriteToken(os, binary, "<Epsilon>");
  WriteBasicType(os, binary, epsilon_);
  WriteToken(os, binary, "<TargetRms>");
  WriteBasicType(os, binary, target_rms_);
  WriteToken(os, binary, "<TestMode>");
  WriteBasicType(os, binary, test_mode_);
  WriteToken(os, binary, "<Count>");
  WriteBasicType(os, binary,  count_);
  CuVector<BaseFloat> mean(stats_sum_), var(stats_sumsq_);
  if (count_ != 0) {
    mean.Scale(1.0 / count_);
    var.Scale(1.0 / count_);
    var.AddVecVec(-1.0, mean, mean, 1.0);
  }
  WriteToken(os, binary, "<StatsMean>");
  mean.Write(os, binary);
  WriteToken(os, binary, "<StatsVar>");
  var.Write(os, binary);
  WriteToken(os, binary, "</BatchNormComponent>");
}

void BatchNormComponent::Scale(BaseFloat scale) {
  if (scale == 0) {
    count_ = 0.0;
    stats_sum_.SetZero();
    stats_sumsq_.SetZero();
  } else {
    count_ *= scale;
    stats_sum_.Scale(scale);
    stats_sumsq_.Scale(scale);
  }
}


void BatchNormComponent::Add(BaseFloat alpha, const Component &other_in) {
  const BatchNormComponent *other =
      dynamic_cast<const BatchNormComponent*>(&other_in);
  count_ += alpha * other->count_;
  stats_sum_.AddVec(alpha, other->stats_sum_);
  stats_sumsq_.AddVec(alpha, other->stats_sumsq_);
  // this operation might change offset_ and scale_, so we recompute them
  // in this instance (but not in Scale()).
  ComputeDerived();
}

void BatchNormComponent::ZeroStats() {
  // We only zero the stats if we're not in test mode.  In test mode, this would
  // be dangerous as the stats are the source for the transform, and zeroing
  // them and then calling ComputeDerived() again would remove the transform
  // parameters (offset_ and scale_).
  if (!test_mode_) {
    count_ = 0.0;
    stats_sum_.SetZero();
    stats_sumsq_.SetZero();
  }
}


} // namespace nnet3
} // namespace kaldi