nnet-convolutional-component.h 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
// nnet3/nnet-convolutional-component.h

// Copyright      2017  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.

#ifndef KALDI_NNET3_NNET_CONVOLUTIONAL_COMPONENT_H_
#define KALDI_NNET3_NNET_CONVOLUTIONAL_COMPONENT_H_

#include "nnet3/nnet-common.h"
#include "nnet3/nnet-component-itf.h"
#include "nnet3/natural-gradient-online.h"
#include "nnet3/convolution.h"
#include <iostream>

namespace kaldi {
namespace nnet3 {

/// @file  nnet-convolutional-component.h
///
/// This file can be viewed as 'overflow' from nnet-general-component.h.
/// It contains a number of components which implement some kind of
/// convolution.


/**
   TimeHeightConvolutionComponent implements 2-dimensional convolution where one
   of the dimensions of convolution (which traditionally would be called the
   width axis) is identified with time (i.e. the 't' component of Indexes).  For
   a deeper understanding of how this works, please see convolution.h.

   The following are the parameters accepted on the config line, with examples
   of their values.


   Parameters inherited from UpdatableComponent (see comment above declaration of
   UpdadableComponent in nnet-component-itf.h for details):
       learning-rate, learning-rate-factor, max-change

   Convolution-related parameters:

     num-filters-in   E.g. num-filters-in=32.  Number of input filters (the
                      number of separate versions of the input image).  The
                      filter-dim has stride 1 in the input and output vectors,
                      i.e. we order the input as (all-filters-for-height=0,
                      all-filters-for-height=1, etc.)
     num-filters-out  E.g. num-filters-out=64. The number of output filters (the
                      number of separate versions of the output image).  As with
                      the input, the filter-dim has stride 1.
     height-in        E.g. height-in=40.  The height of the input image.  The
                      width is not specified the the model level, as it's
                      identified with "t" and is called the time axis; the width
                      is determined by how many "t" values were available at the
                      input of the network, and how many were requested at the
                      output.
     height-out       E.g. height-out=40.  The height of the output image.  Will
                      normally be <= (the input height divided by
                      height-subsample-out).
     height-subsample-out E.g. height-subsample-out=2 (defaults to 1).
                      Subsampling factor on the height axis, e.g. you might set
                      this to 2 if you are doing subsampling on this layer,
                      which would involve discarding every other height
                      increment at the output.  There is no corresponding config
                      for the time dimension, as time subsampling is determined
                      by which 't' values you request at the output, together
                      with the values of 'time-offsets' at different layers of
                      the network.
     height-offsets   E.g. height-offsets=-1,0,1 The set of height offsets that
                      contribute to each output pixel: with the values -1,0,1,
                      height 10 at the output would see data from heights
                      9,10,11 at the input.  These values will normally be
                      consecutive.  Negative values imply zero-padding on the
                      bottom of the image, since output-height 0 is always
                      defined.  Zero-padding at the top of the image is
                      determined in a similar way (e.g. if height-in==height-out
                      and height-offsets=-1,0,1, then there is 1 pixel of
                      padding at the top and bottom of the image).
     time-offsets     E.g. time-offsets=-1,0,1 The time offsets that we require
                      at the input to produce a given output; these are
                      comparable to the offsets used in TDNNs.  Note that the
                      time axis is always numbered using an absolute scheme, so
                      that if there is subsampling on the time axis, then later
                      in the network you'll see time-offsets like "-2,0,2" or
                      "-4,0,4".  Subsampling on the time axis is not explicitly
                      specified but is implicit based on tracking dependencies.
     offsets          Setting 'offsets' is an alternative to setting both
                      height-offsets and time-offsets, that is useful for
                      configurations with less regularity.  It is a semicolon-
                      separated list of pairs (time-offset,height-offset) that
                      might look like: -1,1;-1,0;-1,1;0,1;....;1,1
     required-time-offsets E.g. required-time-offsets=0 (defaults to the same
                      value as time-offsets).  This is a set of time offsets,
                      which if specified must be a nonempty subset of
                      time-offsets; it determines whether zero-padding on the
                      time axis is allowed in cases where there is insufficient
                      input.  If not specified it defaults to the same as
                      'time-offsets', meaning there is no zero-padding on the
                      time axis.  Note: for speech tasks we tend to pad on the
                      time axis with repeats of the first or last frame, rather
                      than zero; and this is handled while preparing the data
                      and not by the core components of the nnet3 framework.  So
                      for speech tasks we wouldn't normally set this value.
     max-memory-mb    Maximum amount of temporary memory, in megabytes, that may
                      be used as temporary matrices in the convolution computation.
                      default=200.0.

   Initialization parameters:
      param-stddev    Standard deviation of the linear parameters of the
                      convolution.  Defaults to sqrt(1.0 / (num-filters-in *
                      num-height-offsets * num-time-offsets)), e.g.
                      sqrt(1.0/(64*3*3)) for a 3x3 kernel with 64 input
                      filters; this value will ensure that the output has
                      unit stddev if the input has unit stddev.
      bias-stddev     Standard deviation of bias terms.  default=0.0.
      init-unit       Defaults to false.  If true, it is required that
                      num-filters-in equal num-filters-out and there should
                      exist a (height, time) offset in the model equal to (0,
                      0).  We will initialize the parameter matrix to be
                      equivalent to the identity transform.  In this case,
                      param-stddev is ignored.


   Natural-gradient related options are below; you won't normally have to
   set these.

      use-natural-gradient e.g. use-natural-gradient=false (defaults to true).
                       You can set this to false to disable the natural gradient
                       updates (you won't normally want to do this).
      rank-out        Rank used in low-rank-plus-unit estimate of the Fisher-matrix
                      factor that has the dimension (num-rows of the parameter
                      space), which equals num-filters-out.  It
                      defaults to the minimum of 80, or half of the number of
                      output filters.
      rank-in         Rank used in low-rank-plus-unit estimate of the Fisher
                      matrix factor which has the dimension (num-cols of the
                      parameter matrix), which has the dimension
                      (num-input-filters * number of time-offsets * number of
                      height-offsets + 1), e.g. num-input-filters * 3 * 3 + 1
                      for a 3x3 kernel (the +1 is for the bias term).
                      It defaults to the minimum of 80, or half the
                      num-rows of the parameter matrix.  [note: I'm considering
                      decreasing this default to e.g. 40 or 20].
      num-minibatches-history
                      This is used setting the 'num_samples_history'
                      configuration value of the natural gradient object.
                      There is no concept of samples (frames) in the
                      application of natural gradient to the convnet, because
                      we do it all on the rows and columns of the derivative.
                      default=4.0.  A larger value means the Fisher matrix is
                      averaged over more minibatches (it's an exponential-decay
                      thing).
      alpha-out       Constant that determines how much we smooth the
                      Fisher-matrix factors with the unit matrix, for the
                      space of dimension num-filters-out.  default=4.0.
      alpha-in        Constant that determines how much we smooth the
                      Fisher-matrix factors with the unit matrix, for the
                      space of dimension (num-filters-in * num-time-offsets *
                      num-height-offsets + 1).  default=4.0.


   Example of a 3x3 kernel with no subsampling, and with zero-padding on both
   the the height and time axis, and where there has previously been no
   subsampling on the time axis:

     num-filters-in=32 num-filters-out=64 height-in=28 height-out=28 \
       height-subsample-out=1 height-offsets=-1,0,1 time-offsets=-1,0,1 \
       required-time-offsets=0

   Example of a 3x3 kernel with no subsampling, without zero-padding on
   either axis, and where there has *previously* been 2-fold subsampling
   on the time axis:

     num-filters-in=32 num-filters-out=64 height-in=20 height-out=18 \
       height-subsample-out=1 height-offsets=0,1,2 time-offsets=0,2,4

   [note: above, the choice to have the time-offsets start at zero rather than
   be centered is just a choice: it assumes that at the output of the network
   you would want to request indexes with t=0, while at the input the t values
   start from zero.]

   Example of a 3x3 kernel with subsampling on the height axis,
   without zero-padding on either axis, and where there has
   previously been 2-fold subsampling on the time axis:

     num-filters-in=32 num-filters-out=64 height-in=20 height-out=9 \
       height-subsample-out=2 height-offsets=0,1,2 time-offsets=0,2,4

  [note: subsampling on the time axis is not expressed in the layer itself:
  any time you increase the distance between time-offsets, like changing
  them from 0,1,2 to 0,2,4, you are effectively subsampling the previous
  layer-- assuming you only request the output at one time value or at
  multiples of the total subsampling factor.]

  Example of a 1x1 kernel:

     num-filters-in=64 num-filters-out=64 height-in=20 height-out=20 \
       height-subsample-out=1 height-offsets=0 time-offsets=0
 */
class TimeHeightConvolutionComponent: public UpdatableComponent {
 public:

  // The use of this constructor should only precede InitFromConfig()
  TimeHeightConvolutionComponent();

  // Copy constructor
  TimeHeightConvolutionComponent(const TimeHeightConvolutionComponent &other);

  virtual int32 InputDim() const;
  virtual int32 OutputDim() const;

  virtual std::string Info() const;
  virtual void InitFromConfig(ConfigLine *cfl);
  virtual std::string Type() const { return "TimeHeightConvolutionComponent"; }
  virtual int32 Properties() const {
    return kUpdatableComponent|kReordersIndexes|kBackpropAdds|
        kBackpropNeedsInput|kInputContiguous|kOutputContiguous;
  }
  virtual void* Propagate(const ComponentPrecomputedIndexes *indexes,
                         const CuMatrixBase<BaseFloat> &in,
                         CuMatrixBase<BaseFloat> *out) const;
  virtual void 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;

  virtual void Read(std::istream &is, bool binary);
  virtual void Write(std::ostream &os, bool binary) const;
  virtual Component* Copy() const {
    return new TimeHeightConvolutionComponent(*this);
  }


  // Some functions that are only to be reimplemented for GeneralComponents.

  // This ReorderIndexes function may insert 'blank' indexes (indexes with
  // t == kNoTime) as well as reordering the indexes.  This is allowed
  // behavior of ReorderIndexes functions.
  virtual void ReorderIndexes(std::vector<Index> *input_indexes,
                              std::vector<Index> *output_indexes) const;

  virtual void GetInputIndexes(const MiscComputationInfo &misc_info,
                               const Index &output_index,
                               std::vector<Index> *desired_indexes) const;

  // This function returns true if at least one of the input indexes used to
  // compute this output index is computable.
  virtual bool IsComputable(const MiscComputationInfo &misc_info,
                            const Index &output_index,
                            const IndexSet &input_index_set,
                            std::vector<Index> *used_inputs) const;

  virtual ComponentPrecomputedIndexes* PrecomputeIndexes(
      const MiscComputationInfo &misc_info,
      const std::vector<Index> &input_indexes,
      const std::vector<Index> &output_indexes,
      bool need_backprop) const;

  // Some functions from base-class UpdatableComponent.
  virtual void Scale(BaseFloat scale);
  virtual void Add(BaseFloat alpha, const Component &other);
  virtual void PerturbParams(BaseFloat stddev);
  virtual BaseFloat DotProduct(const UpdatableComponent &other) const;
  virtual int32 NumParameters() const;
  virtual void Vectorize(VectorBase<BaseFloat> *params) const;
  virtual void UnVectorize(const VectorBase<BaseFloat> &params);
  virtual void FreezeNaturalGradient(bool freeze);


  class PrecomputedIndexes: public ComponentPrecomputedIndexes {
   public:
    PrecomputedIndexes() { }
    PrecomputedIndexes(const PrecomputedIndexes &other):
        computation(other.computation) { }
    virtual PrecomputedIndexes *Copy() const;
    virtual void Write(std::ostream &os, bool binary) const;
    virtual void Read(std::istream &os, bool binary);
    virtual std::string Type() const {
      return "TimeHeightConvolutionComponentPrecomputedIndexes";
    }
    virtual ~PrecomputedIndexes() { }

    time_height_convolution::ConvolutionComputation computation;
  };

  void ScaleLinearParams(BaseFloat alpha) { linear_params_.Scale(alpha); }

  void ConsolidateMemory();
 private:

  void Check() const;

  // computes derived parameters required_time_offsets_ and all_time_offsets_.
  void ComputeDerived();

  // Function that updates linear_params_ and bias_params_, which
  // uses the natural gradient code.
  void UpdateNaturalGradient(
      const PrecomputedIndexes &indexes,
      const CuMatrixBase<BaseFloat> &in_value,
      const CuMatrixBase<BaseFloat> &out_deriv);

  // Function that updates linear_params_ and bias_params_, which
  // does not use the natural gradient code.
  void UpdateSimple(
      const PrecomputedIndexes &indexes,
      const CuMatrixBase<BaseFloat> &in_value,
      const CuMatrixBase<BaseFloat> &out_deriv);

  // Function called to initialize linear_params_ if init-unit=true in the config
  // line.
  void InitUnit();

  time_height_convolution::ConvolutionModel model_;

  // all_time_offsets_ is a copy of the corresponding variable in
  // model, stored as a vector instead of as a set for efficiency.
  std::vector<int32> all_time_offsets_;
  // time_offset_required_ is a vector with the same dimension as
  // 'all_time_offsets_', which is true if the corresponding time-offset
  // is a member of model_.required_time_offsets_.
  std::vector<bool> time_offset_required_;

  // the linear parameters of the convolution.
  // dimension is model_.ParamRows() by model.ParamCols(),
  // which equals num-filters-out by
  // (num-filters-in * patch-rows * patch-cols),
  // a.k.a.
  // (num-filters-in * num-time-offsets * num-height-offset).
  CuMatrix<BaseFloat> linear_params_;
  // the bias parameters of the convolution, dimension is
  // model_.num_filters_out.
  CuVector<BaseFloat> bias_params_;


  // Maximum amount of temporary memory in megabytes that is allowed to be used
  // in the convolution computation.  (this is per computation, but it's
  // released immediately after it's used, so it doesn't matter how many there
  // are).
  BaseFloat max_memory_mb_;

  // Controls whether or not the natural-gradient is used.
  // Note: even if this is true, if is_gradient_ (from the
  // UpdatableComponent base class) is true, we'll do the 'simple'
  // update that doesn't include natural gradient.
  bool use_natural_gradient_;

  // Preconditioner for the input space, of dimension linear_params_.NumCols() +
  // 1 (the 1 is for the bias).  As with other natural-gradient objects, it's
  // not stored with the model on disk but is reinitialized each time we start
  // up.
  OnlineNaturalGradient preconditioner_in_;

  // Preconditioner for the output space, of dimension
  // linear_params_.NumRows().
  OnlineNaturalGradient preconditioner_out_;
};



/**
   TdnnComponent is a more memory-efficient alternative to manually splicing
   several frames of input and then using a NaturalGradientAffineComponent or
   a LinearComponent.  It does the splicing of the input itself, using
   mechanisms similar to what TimeHeightConvolutionComponent uses.  The
   implementation is in nnet-tdnn-component.cc

   Parameters inherited from UpdatableComponent (see comment above declaration of
   UpdadableComponent in nnet-component-itf.h for details):
       learning-rate, learning-rate-factor, max-change

   Important parameters:

     input-dim         The input feature dimension (before splicing).

     output-dim        The output feature dimension

     time-offsets     E.g. time-offsets=-1,0,1 or time-offsets=-3,0,3.
                      The time offsets that we require at the input to produce a given output.
                      comparable to the offsets used in TDNNs.  They
                      must be unique (no repeats).
     use-bias         Defaults to true, but set to false if you want this to
                      be linear rather than affine in its input.


  Extra parameters:
    orthonormal-constraint=0.0 If you set this to 1.0, then the linear_params_
                      matrix will be (approximately) constrained during training
                      to have orthonormal rows (or columns, whichever is
                      fewer).. it turns out the real name for this is a
                      "semi-orthogonal" matrix.  You can choose a positive
                      nonzero value different than 1.0 to have a scaled
                      semi-orthgonal matrix, i.e. with singular values at the
                      selected value (e.g. 0.5, or 2.0).  This is not enforced
                      inside the component itself; you have to call
                      ConstrainOrthonormal() from the training code to do this.
                      All this component does is return the
                      OrthonormalConstraint() value.  If you set this to a
                      negative value, it's like saying "for any value", i.e. it
                      will constrain the parameter matrix to be closer to "any
                      alpha" times a semi-orthogonal matrix, without changing
                      its overall norm.


   Initialization parameters:
      param-stddev    Standard deviation of the linear parameters of the
                      convolution.  Defaults to
                      sqrt(1.0 / (input-dim * the number of time-offsets))
      bias-stddev     Standard deviation of bias terms.  default=0.0.
                      You should not set this if you set use-bias=false.


   Natural-gradient related options are below; you won't normally have to
   set these as the defaults are reasonable.

      use-natural-gradient e.g. use-natural-gradient=false (defaults to true).
                       You can set this to false to disable the natural gradient
                       updates (you won't normally want to do this).
      rank-out         Rank used in low-rank-plus-unit estimate of the Fisher-matrix
                       factor that has the dimension (num-rows of linear_params_),
                       which equals output_dim.  It
                       defaults to the minimum of 80, or half of the output dim.
      rank-in          Rank used in low-rank-plus-unit estimate of the Fisher
                       matrix factor which has the dimension (num-cols of the
                       parameter matrix), which is input-dim times the number of
                       time offsets.  It defaults to the minimum of 20, or half the
                       num-rows of the parameter matrix.
      num-samples-history
                      This becomes the 'num_samples_history'
                      configuration value of the natural gradient objects.  The
                      default value is 2000.0.

 */
class TdnnComponent: public UpdatableComponent {
 public:

  // The use of this constructor should only precede InitFromConfig()
  TdnnComponent();

  // Copy constructor
  TdnnComponent(const TdnnComponent &other);

  virtual int32 InputDim() const {
    return linear_params_.NumCols() / static_cast<int32>(time_offsets_.size());
  }
  virtual int32 OutputDim() const { return linear_params_.NumRows(); }

  virtual std::string Info() const;
  virtual void InitFromConfig(ConfigLine *cfl);
  virtual std::string Type() const { return "TdnnComponent"; }
  virtual int32 Properties() const {
    return kUpdatableComponent|kReordersIndexes|kBackpropAdds|
        (bias_params_.Dim() == 0 ? kPropagateAdds : 0)|
        kBackpropNeedsInput;
  }
  virtual void* Propagate(const ComponentPrecomputedIndexes *indexes,
                         const CuMatrixBase<BaseFloat> &in,
                         CuMatrixBase<BaseFloat> *out) const;
  virtual void 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;

  virtual void Read(std::istream &is, bool binary);
  virtual void Write(std::ostream &os, bool binary) const;
  virtual Component* Copy() const {
    return new TdnnComponent(*this);
  }


  // Some functions that are only to be reimplemented for GeneralComponents.

  // This ReorderIndexes function may insert 'blank' indexes (indexes with
  // t == kNoTime) as well as reordering the indexes.  This is allowed
  // behavior of ReorderIndexes functions.
  virtual void ReorderIndexes(std::vector<Index> *input_indexes,
                              std::vector<Index> *output_indexes) const;

  virtual void GetInputIndexes(const MiscComputationInfo &misc_info,
                               const Index &output_index,
                               std::vector<Index> *desired_indexes) const;

  // This function returns true if at least one of the input indexes used to
  // compute this output index is computable.
  virtual bool IsComputable(const MiscComputationInfo &misc_info,
                            const Index &output_index,
                            const IndexSet &input_index_set,
                            std::vector<Index> *used_inputs) const;

  virtual ComponentPrecomputedIndexes* PrecomputeIndexes(
      const MiscComputationInfo &misc_info,
      const std::vector<Index> &input_indexes,
      const std::vector<Index> &output_indexes,
      bool need_backprop) const;

  // Some functions from base-class UpdatableComponent.
  virtual void Scale(BaseFloat scale);
  virtual void Add(BaseFloat alpha, const Component &other);
  virtual void PerturbParams(BaseFloat stddev);
  virtual BaseFloat DotProduct(const UpdatableComponent &other) const;
  virtual int32 NumParameters() const;
  virtual void Vectorize(VectorBase<BaseFloat> *params) const;
  virtual void UnVectorize(const VectorBase<BaseFloat> &params);
  virtual void FreezeNaturalGradient(bool freeze);


  class PrecomputedIndexes: public ComponentPrecomputedIndexes {
   public:
    PrecomputedIndexes() { }
    PrecomputedIndexes(const PrecomputedIndexes &other):
        row_stride(other.row_stride), row_offsets(other.row_offsets) { }
    virtual PrecomputedIndexes *Copy() const;
    virtual void Write(std::ostream &os, bool binary) const;
    virtual void Read(std::istream &os, bool binary);
    virtual std::string Type() const {
      return "TdnnComponentPrecomputedIndexes";
    }
    virtual ~PrecomputedIndexes() { }


    // input_row_stride is the stride (in number of rows) we have to take in the
    // input matrix each time we form a sub-matrix that will be part of the
    // input to the tdnn operation.  Normally this will be 1, but it may be,
    // for example, 3 in layers where we do subsampling.
    int32 row_stride;

    // 'row_offsets' is of the same dimension as time_offsets_.  Each element
    // describes the row offset (in the input matrix) of a sub-matrix, and each.
    // We will append together these sub-matrices (row-wise) to be the input to
    // the affine or linear transform.
    std::vector<int32> row_offsets;
  };

  CuMatrixBase<BaseFloat> &LinearParams() { return linear_params_; }

  // This allows you to resize the vector in order to add a bias where
  // there previously was none-- obviously this should be done carefully.
  CuVector<BaseFloat> &BiasParams() { return bias_params_; }

  BaseFloat OrthonormalConstraint() const { return orthonormal_constraint_; }

  void ConsolidateMemory();
 private:

  // This static function is a utility function that extracts a CuSubMatrix
  // representing a subset of rows of 'input_matrix'.
  // The numpy syntax would be:
  //   return input_matrix[row_offset:row_stride:num_output_rows*row_stride,:]
  static CuSubMatrix<BaseFloat> GetInputPart(
      const CuMatrixBase<BaseFloat> &input_matrix,
      int32 num_output_rows,
      int32 row_stride,
      int32 row_offset);

  // see the definition for more explanation.
  static void ModifyComputationIo(time_height_convolution::ConvolutionComputationIo *io);

  void Check() const;

  // Function that updates linear_params_, and bias_params_ if present, which
  // uses the natural gradient code.
  void UpdateNaturalGradient(
      const PrecomputedIndexes &indexes,
      const CuMatrixBase<BaseFloat> &in_value,
      const CuMatrixBase<BaseFloat> &out_deriv);

  // Function that updates linear_params_, and bias_params_ if present, which
  // does not use the natural gradient code.
  void UpdateSimple(
      const PrecomputedIndexes &indexes,
      const CuMatrixBase<BaseFloat> &in_value,
      const CuMatrixBase<BaseFloat> &out_deriv);




  // time_offsets_ is the list of time-offsets of the input that
  // we append together; it will typically be (-1,0,1) or (-3,0,3).
  std::vector<int32> time_offsets_;

  // the linear parameters of the network; its NumRows() is the output
  // dim, and its NumCols() equals the input dim times time_offsets_.size().
  CuMatrix<BaseFloat> linear_params_;

  // the bias parameters if this is an affine transform, or the empty vector if
  // this is a linear operation (i.e. use-bias == false in the config).
  CuVector<BaseFloat> bias_params_;

  // If nonzero, this controls how we apply an orthonormal constraint to the
  // parameter matrix; see docs for ConstrainOrthonormal() in nnet-utils.h.
  // This class just returns the value via the OrthonormalConstraint() function;
  // it doesn't actually do anything with it directly.
  BaseFloat orthonormal_constraint_;

  // Controls whether or not the natural-gradient is used.  Note: even if this
  // is true, if is_gradient_ (from the UpdatableComponent base class) is true,
  // we'll do the 'simple' update that doesn't include natural gradient.
  bool use_natural_gradient_;

  // Preconditioner for the input space, of dimension linear_params_.NumCols() +
  // 1 (the 1 is for the bias).  As with other natural-gradient objects, it's
  // not stored with the model on disk but is reinitialized each time we start
  // up.
  OnlineNaturalGradient preconditioner_in_;

  // Preconditioner for the output space, of dimension
  // linear_params_.NumRows().
  OnlineNaturalGradient preconditioner_out_;
};







} // namespace nnet3
} // namespace kaldi


#endif