nnet-nnet.cc 28 KB
Newer Older
1
// nnet2/nnet-nnet.cc
2 3

// Copyright 2011-2012  Karel Vesely
4
//           2012-2014  Johns Hopkins University (author: Daniel Povey)
5

6 7
// See ../../COPYING for clarification regarding multiple authors
//
8 9 10 11 12 13 14 15 16 17 18 19 20
// 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.

21 22 23
#include <algorithm>
#include <set>
#include <string>
24
#include "nnet2/nnet-nnet.h"
25 26 27
#include "util/stl-utils.h"

namespace kaldi {
28
namespace nnet2 {
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44


int32 Nnet::OutputDim() const {
  KALDI_ASSERT(!components_.empty());
  return components_.back()->OutputDim();
}

int32 Nnet::InputDim() const {
  KALDI_ASSERT(!components_.empty());
  return components_.front()->InputDim();
}


int32 Nnet::LeftContext() const {
  KALDI_ASSERT(!components_.empty());
  int32 ans = 0;
45 46 47 48 49 50 51 52 53 54
  for (size_t i = 0; i < components_.size(); i++) {
    ans += components_[i]->Context().front();
  }
  return -1*ans;
  // nnet-components return left context as a non-positive integer
  // however the nnet-update, nnet-compute, train-nnet-perturbed expect a
  // non-negative left context. In addition, the NnetExample also stores data
  // left context as positive integer. To be compatible with these other classes
  // Nnet::LeftContext() returns a non-negative left context.
              
55 56 57 58 59
}

int32 Nnet::RightContext() const {
  KALDI_ASSERT(!components_.empty());
  int32 ans = 0;
60 61 62
  for (size_t i = 0; i < components_.size(); i++) {
    ans += components_[i]->Context().back();
  }
63 64 65
  return ans;
}

66
void Nnet::ComputeChunkInfo(int32 input_chunk_size,
67 68 69 70 71 72
                            int32 num_chunks,
                            std::vector<ChunkInfo> *chunk_info_out) const {
  // First compute the output-chunk indices for the last component in the network.
  // we assume that the numbering of the input starts from zero.
  int32 output_chunk_size = input_chunk_size - LeftContext() - RightContext();
  KALDI_ASSERT(output_chunk_size > 0);
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
  std::vector<int32> current_output_inds;
  for (int32 i = 0; i < output_chunk_size; i++)
    current_output_inds.push_back(i + LeftContext());

  (*chunk_info_out).resize(NumComponents() + 1);

  // indexes for last component is empty, since the last component's chunk is
  // always contiguous
  // component's output is always contiguous
  (*chunk_info_out)[NumComponents()] = ChunkInfo(
      GetComponent(NumComponents() - 1).OutputDim(),
      num_chunks, current_output_inds.front(),
      current_output_inds.back());

  std::vector<int32> current_input_inds;
  for (int32 i = NumComponents() - 1; i >= 0; i--) {
    std::vector<int32> current_context = GetComponent(i).Context();
    std::set<int32> current_input_ind_set;
    for (size_t j = 0; j < current_context.size(); j++) 
      for (size_t k = 0; k < current_output_inds.size(); k++)
        current_input_ind_set.insert(current_context[j] +
                                     current_output_inds[k]);
    current_output_inds.resize(current_input_ind_set.size());
    std::copy(current_input_ind_set.begin(),
              current_input_ind_set.end(),
              current_output_inds.begin());

    // checking if the vector has contiguous data
    // assign indexes only if the data is not contiguous
    if (current_output_inds.size() !=
        current_output_inds.back() - current_output_inds.front() + 1) {
      (*chunk_info_out)[i] = ChunkInfo(GetComponent(i).InputDim(),
                                       num_chunks,
                                       current_output_inds);
    } else  {
      (*chunk_info_out)[i] = ChunkInfo(GetComponent(i).InputDim(),
                                       num_chunks,
                                       current_output_inds.front(),
                                       current_output_inds.back());
    }
  }

  // TODO: Make a set of components which can deal with data rearrangement.
  // Define this set in an appropriate place so that
  // users adding new components can simply update the list.
  const char *dinit[] = {"SpliceComponent", "SpliceMaxComponent"};
  std::vector< std::string > data_rearrange_components(dinit, dinit + 2);

  // Ensuring that all components until the first component capable of data
  // rearrangement (e.g. SpliceComponent|SpliceMaxComponent) operate on
  // contiguous chunks at the input
  for (size_t i = 0 ; i < NumComponents() ; i++) {
      (*chunk_info_out)[i].MakeOffsetsContiguous();
      // Check if the current component is present in the set of components
      // capable of data rearrangement.
      if (std::find(data_rearrange_components.begin(),
                    data_rearrange_components.end(),
                    components_[i]->Type())
          != data_rearrange_components.end())
          break;
  }

  // sanity testing for chunk_info_out vector
  for (size_t i = 0; i < chunk_info_out->size(); i++) {
    (*chunk_info_out)[i].Check();
    // (*chunk_info_out)[i].ToString();
  }

}

143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162
const Component& Nnet::GetComponent(int32 component) const {
  KALDI_ASSERT(static_cast<size_t>(component) < components_.size());
  return *(components_[component]);
}

Component& Nnet::GetComponent(int32 component) {
  KALDI_ASSERT(static_cast<size_t>(component) < components_.size());
  return *(components_[component]);
}

void Nnet::SetZero(bool treat_as_gradient) {
  for (size_t i = 0; i < components_.size(); i++) {
    UpdatableComponent *uc = dynamic_cast<UpdatableComponent*>(components_[i]);
    if (uc != NULL) uc->SetZero(treat_as_gradient);
    NonlinearComponent *nc = dynamic_cast<NonlinearComponent*>(components_[i]);
    if (nc != NULL) nc->Scale(0.0);
  }
}

void Nnet::Write(std::ostream &os, bool binary) const {
163
  Check();
164 165 166 167 168 169 170 171 172 173
  WriteToken(os, binary, "<Nnet>");
  int32 num_components = components_.size();
  WriteToken(os, binary, "<NumComponents>");
  WriteBasicType(os, binary, num_components);
  WriteToken(os, binary, "<Components>");
  for (int32 c = 0; c < num_components; c++) {
    components_[c]->Write(os, binary);
    if (!binary) os << std::endl;
  }
  WriteToken(os, binary, "</Components>");
174
  WriteToken(os, binary, "</Nnet>");
175 176 177 178 179 180 181 182 183 184
}

void Nnet::Read(std::istream &is, bool binary) {
  Destroy();
  ExpectToken(is, binary, "<Nnet>");
  int32 num_components;
  ExpectToken(is, binary, "<NumComponents>");
  ReadBasicType(is, binary, &num_components);
  ExpectToken(is, binary, "<Components>");
  components_.resize(num_components);
185
  for (int32 c = 0; c < num_components; c++)
186 187
    components_[c] = Component::ReadNew(is, binary);
  ExpectToken(is, binary, "</Components>");
188 189
  ExpectToken(is, binary, "</Nnet>");
  SetIndexes();
190
  Check();
191 192 193 194 195 196 197 198
}


void Nnet::ZeroStats() {
  for (size_t i = 0; i < components_.size(); i++) {
    NonlinearComponent *nonlinear_component =
        dynamic_cast<NonlinearComponent*>(components_[i]);
    if (nonlinear_component != NULL)
199
      nonlinear_component->Scale(0.0);  // Zero the stats this way.
200 201 202 203 204 205 206 207 208 209 210 211
  }
}
void Nnet::Destroy() {
  while (!components_.empty()) {
    delete components_.back();
    components_.pop_back();
  }
}

void Nnet::ComponentDotProducts(
    const Nnet &other,
    VectorBase<BaseFloat> *dot_prod) const {
212 213
  KALDI_ASSERT(dot_prod->Dim() == NumUpdatableComponents());
  int32 index = 0;
214 215 216 217 218
  for (size_t i = 0; i < components_.size(); i++) {
    UpdatableComponent *uc1 = dynamic_cast<UpdatableComponent*>(components_[i]);
    const UpdatableComponent *uc2 = dynamic_cast<const UpdatableComponent*>(
        &(other.GetComponent(i)));
    KALDI_ASSERT((uc1 != NULL) == (uc2 != NULL));
219 220 221 222
    if (uc1 != NULL) {
      (*dot_prod)(index) = uc1->DotProduct(*uc2);
      index++;
    }
223
  }
224
  KALDI_ASSERT(index == NumUpdatableComponents());
225 226 227 228 229 230
}


Nnet::Nnet(const Nnet &other): components_(other.components_.size()) {
  for (size_t i = 0; i < other.components_.size(); i++)
    components_[i] = other.components_[i]->Copy();
231 232 233 234 235 236 237 238 239 240 241 242 243 244 245
  SetIndexes();
  Check();
}

Nnet::Nnet(const Nnet &other1, const Nnet &other2) {
  int32 dim1 = other1.OutputDim(), dim2 = other2.InputDim();
  if (dim1 != dim2)
    KALDI_ERR << "Concatenating neural nets: dimension mismatch "
              << dim1 << " vs. " << dim2;
  for (size_t i = 0; i < other1.components_.size(); i++)
    components_.push_back(other1.components_[i]->Copy());
  for (size_t i = 0; i < other2.components_.size(); i++)
    components_.push_back(other2.components_[i]->Copy());
  SetIndexes();
  Check();
246 247 248 249 250 251 252 253
}


Nnet &Nnet::operator = (const Nnet &other) {
  Destroy();
  components_.resize(other.components_.size());
  for (size_t i = 0; i < other.components_.size(); i++)
    components_[i] = other.components_[i]->Copy();
254 255
  SetIndexes();
  Check();
256 257 258 259 260 261
  return *this;
}

std::string Nnet::Info() const {
  std::ostringstream ostr;
  ostr << "num-components " << NumComponents() << std::endl;
262
  ostr << "num-updatable-components " << NumUpdatableComponents() << std::endl;
263 264 265 266
  ostr << "left-context " << LeftContext() << std::endl;
  ostr << "right-context " << RightContext() << std::endl;
  ostr << "input-dim " << InputDim() << std::endl;
  ostr << "output-dim " << OutputDim() << std::endl;
267
  ostr << "parameter-dim " << GetParameterDim() << std::endl;
268
  for (int32 i = 0; i < NumComponents(); i++)
269 270 271 272 273 274 275 276 277 278
    ostr << "component " << i << " : " << components_[i]->Info() << std::endl;
  return ostr.str();
}

void Nnet::Check() const {
  for (size_t i = 0; i + 1 < components_.size(); i++) {
    KALDI_ASSERT(components_[i] != NULL);
    int32 output_dim = components_[i]->OutputDim(),
      next_input_dim = components_[i+1]->InputDim();
    KALDI_ASSERT(output_dim == next_input_dim);
279
    KALDI_ASSERT(components_[i]->Index() == static_cast<int32>(i));
280 281 282 283 284 285 286 287 288 289
  }
}

void Nnet::Init(std::istream &is) {
  Destroy();
  std::string line;
  /* example config file as follows.  The things in brackets specify the context
     splicing for each layer, and after that is the info about the actual layer.
     Imagine the input dim is 13, and the speaker dim is 40, so (13 x 9) + 40 =  527.
     The config file might be as follows; the lines beginning with # are comments.
290

291 292 293 294 295 296
     # layer-type layer-options
     AffineLayer 0.01 0.001 527 1000 0.04356
  */
  components_.clear();
  while (getline(is, line)) {
    std::istringstream line_is(line);
297 298
    line_is >> std::ws;  // Eat up whitespace.
    if (line_is.peek() == '#' || line_is.eof()) continue;  // Comment or empty.
299 300
    Component *c = Component::NewFromString(line);
    KALDI_ASSERT(c != NULL);
301
    components_.push_back(c);
302
  }
303
  SetIndexes();
304 305 306 307 308 309
  Check();
}

void Nnet::Init(std::vector<Component*> *components) {
  Destroy();
  components_.swap(*components);
310
  SetIndexes();
311 312 313 314 315 316 317 318
  Check();
}


void Nnet::ScaleLearningRates(BaseFloat factor) {
  std::ostringstream ostr;
  for (int32 c = 0; c < NumComponents(); c++) {
    UpdatableComponent *uc = dynamic_cast<UpdatableComponent*>(components_[c]);
319
    if (uc != NULL) {  // Updatable component...
320 321 322 323 324 325 326 327 328
      uc->SetLearningRate(uc->LearningRate() * factor);
      ostr << uc->LearningRate() << " ";
    }
  }
  KALDI_LOG << "Scaled learning rates by " << factor
            << ", new learning rates are "
            << ostr.str();
}

329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347
void Nnet::ScaleLearningRates(std::map<std::string, BaseFloat> scale_factors) {
  std::ostringstream ostr;
  for (int32 c = 0; c < NumComponents(); c++) {
    UpdatableComponent *uc = dynamic_cast<UpdatableComponent*>(components_[c]);
    if (uc != NULL) {  // Updatable component...
      // check if scaling factor was specified for a component of this type
      std::map<std::string, BaseFloat>::const_iterator lr_iterator =
        scale_factors.find(uc->Type());
      if (lr_iterator != scale_factors.end())  {
        uc->SetLearningRate(uc->LearningRate() * lr_iterator->second);
        ostr << uc->LearningRate() << " ";
      }
    }
  }
  KALDI_LOG << "Scaled learning rates by component-type specific factor, "
            << "new learning rates are "
            << ostr.str();
}

348 349 350
void Nnet::SetLearningRates(BaseFloat learning_rate) {
  for (int32 c = 0; c < NumComponents(); c++) {
    UpdatableComponent *uc = dynamic_cast<UpdatableComponent*>(components_[c]);
351
    if (uc != NULL) {  // Updatable component...
352 353 354 355 356 357
      uc->SetLearningRate(learning_rate);
    }
  }
  KALDI_LOG << "Set learning rates to " << learning_rate;
}

Dan Povey's avatar
Dan Povey committed
358 359 360 361
void Nnet::ResizeOutputLayer(int32 new_num_pdfs) {
  KALDI_ASSERT(new_num_pdfs > 0);
  KALDI_ASSERT(NumComponents() > 2);
  int32 nc = NumComponents();
362
  SumGroupComponent *sgc = dynamic_cast<SumGroupComponent*>(components_[nc - 1]);
Dan Povey's avatar
Dan Povey committed
363 364 365
  if (sgc != NULL) {
    // Remove it.  We'll resize things later.
    delete sgc;
366 367
    components_.erase(components_.begin() + nc - 1,
                      components_.begin() + nc);
Dan Povey's avatar
Dan Povey committed
368
    nc--;
369
  }
370 371 372 373
  
  SoftmaxComponent *sc;
  if ((sc = dynamic_cast<SoftmaxComponent*>(components_[nc - 1])) == NULL)
    KALDI_ERR << "Expected last component to be SoftmaxComponent.";
374

375 376 377 378 379 380 381 382 383 384 385 386 387 388
  // check if nc-1 has a FixedScaleComponent
  bool has_fixed_scale_component = false;
  int32 fixed_scale_component_index = -1;
  int32 final_affine_component_index = nc - 2;
  int32 softmax_component_index = nc - 1;
  FixedScaleComponent *fsc =
      dynamic_cast<FixedScaleComponent*>(
          components_[final_affine_component_index]);
  if (fsc != NULL)  {
    has_fixed_scale_component = true; 
    fixed_scale_component_index = nc - 2;
    final_affine_component_index = nc - 3;
  }
    
Dan Povey's avatar
Dan Povey committed
389
  // note: it could be child class of AffineComponent.
390 391
  AffineComponent *ac = dynamic_cast<AffineComponent*>(
      components_[final_affine_component_index]);
Dan Povey's avatar
Dan Povey committed
392 393 394
  if (ac == NULL)
    KALDI_ERR << "Network doesn't have expected structure (didn't find final "
              << "AffineComponent).";
395 396 397 398 399 400 401 402 403 404 405 406 407
  if (has_fixed_scale_component)  {
    // collapse the fixed_scale_component with the affine_component before it
    AffineComponent *ac_new = dynamic_cast<AffineComponent*>(ac->CollapseWithNext(*fsc));
    KALDI_ASSERT(ac_new != NULL);
    delete fsc;
    delete ac;
    components_.erase(components_.begin() + fixed_scale_component_index,
                      components_.begin() + (fixed_scale_component_index + 1));
    components_[final_affine_component_index] = ac_new;
    ac = ac_new;
    softmax_component_index = softmax_component_index - 1;
  }
   
Dan Povey's avatar
Dan Povey committed
408 409
  ac->Resize(ac->InputDim(), new_num_pdfs);
  // Remove the softmax component, and replace it with a new one
410 411
  delete components_[softmax_component_index];
  components_[softmax_component_index] = new SoftmaxComponent(new_num_pdfs);
Dan Povey's avatar
Dan Povey committed
412 413
  this->Check();
}
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

int32 Nnet::NumUpdatableComponents() const {
  int32 ans = 0;
  for (int32 i = 0; i < NumComponents(); i++)
    if (dynamic_cast<const UpdatableComponent*>(&(GetComponent(i))) != NULL)
      ans++;
  return ans;
}

void Nnet::ScaleComponents(const VectorBase<BaseFloat> &scale_params) {
  KALDI_ASSERT(scale_params.Dim() == this->NumUpdatableComponents());
  int32 i = 0;
  for (int32 j = 0; j < NumComponents(); j++) {
    UpdatableComponent *uc =
        dynamic_cast<UpdatableComponent*>(&(GetComponent(j)));
    if (uc!= NULL) {
      uc->Scale(scale_params(i));
      i++;
    }
  }
  KALDI_ASSERT(i == scale_params.Dim());
}

// Scales all UpdatableComponents and all NonlinearComponents.
void Nnet::Scale(BaseFloat scale) {
  for (int32 i = 0; i < NumComponents(); i++) {
    UpdatableComponent *uc =
        dynamic_cast<UpdatableComponent*>(&(GetComponent(i)));
    if (uc != NULL) uc->Scale(scale);
    NonlinearComponent *nc =
        dynamic_cast<NonlinearComponent*>(&(GetComponent(i)));
    if (nc != NULL) nc->Scale(scale);
  }
}

void Nnet::CopyStatsFrom(const Nnet &other) {
  KALDI_ASSERT(NumComponents() == other.NumComponents());
  for (int32 i = 0; i < NumComponents(); i++) {
    NonlinearComponent *nc =
        dynamic_cast<NonlinearComponent*>(&(GetComponent(i)));
    const NonlinearComponent *nc_other =
        dynamic_cast<const NonlinearComponent*>(&(other.GetComponent(i)));
    if (nc != NULL) {
      nc->Scale(0.0);
      nc->Add(1.0, *nc_other);
    }
  }
}

void Nnet::SetLearningRates(const VectorBase<BaseFloat> &learning_rates) {
  KALDI_ASSERT(learning_rates.Dim() == this->NumUpdatableComponents());
465
  KALDI_ASSERT(learning_rates.Min() >= 0.0);  // we allow zero learning rate.
466 467 468 469 470 471 472 473 474 475 476 477
  int32 i = 0;
  for (int32 j = 0; j < NumComponents(); j++) {
    UpdatableComponent *uc =
        dynamic_cast<UpdatableComponent*>(&(GetComponent(j)));
    if (uc!= NULL) {
      uc->SetLearningRate(learning_rates(i));
      i++;
    }
  }
  KALDI_ASSERT(i == learning_rates.Dim());
}

478 479 480 481 482 483 484 485 486 487 488 489 490 491
void Nnet::GetLearningRates(VectorBase<BaseFloat> *learning_rates) const {
  KALDI_ASSERT(learning_rates->Dim() == this->NumUpdatableComponents());
  int32 i = 0;
  for (int32 j = 0; j < NumComponents(); j++) {
    const UpdatableComponent *uc =
        dynamic_cast<const UpdatableComponent*>(&(GetComponent(j)));
    if (uc!= NULL) {
      (*learning_rates)(i) = uc->LearningRate();
      i++;
    }
  }
  KALDI_ASSERT(i == learning_rates->Dim());
}

492 493 494 495 496 497 498 499 500 501 502 503
void Nnet::Resize(int32 new_size) {
  KALDI_ASSERT(new_size <= static_cast<int32>(components_.size()));
  for (size_t i = new_size; i < components_.size(); i++)
    delete components_[i];
  components_.resize(new_size);
}

void Nnet::RemoveDropout() {
  std::vector<Component*> components;
  int32 removed = 0;
  for (size_t i = 0; i < components_.size(); i++) {
    if (dynamic_cast<DropoutComponent*>(components_[i]) != NULL ||
504
        dynamic_cast<AdditiveNoiseComponent*>(components_[i]) != NULL) {
505 506 507 508 509 510 511 512 513
      delete components_[i];
      removed++;
    } else {
      components.push_back(components_[i]);
    }
  }
  components_ = components;
  if (removed > 0)
    KALDI_LOG << "Removed " << removed << " dropout components.";
514 515
  SetIndexes();
  Check();
516 517
}

518 519 520 521 522 523 524 525 526 527 528 529
void Nnet::SetDropoutScale(BaseFloat scale) {
  size_t n_set = 0;
  for (size_t i = 0; i < components_.size(); i++) {
    DropoutComponent *dc =
        dynamic_cast<DropoutComponent*>(components_[i]);
    if (dc != NULL) {
      dc->SetDropoutScale(scale);
      n_set++;
    }
  }
  KALDI_LOG << "Set dropout scale to " << scale
            << " for " << n_set << " components.";
530
}
531 532


533 534 535 536 537 538 539
void Nnet::RemovePreconditioning() {
  for (size_t i = 0; i < components_.size(); i++) {
    if (dynamic_cast<AffineComponentPreconditioned*>(components_[i]) != NULL) {
      AffineComponent *ac = new AffineComponent(
          *(dynamic_cast<AffineComponent*>(components_[i])));
      delete components_[i];
      components_[i] = ac;
540 541 542 543 544 545
    } else if (dynamic_cast<AffineComponentPreconditionedOnline*>(
        components_[i]) != NULL) {
      AffineComponent *ac = new AffineComponent(
          *(dynamic_cast<AffineComponent*>(components_[i])));
      delete components_[i];
      components_[i] = ac;
546 547
    }
  }
548 549
  SetIndexes();
  Check();
550 551
}

552

553 554 555 556
void Nnet::SwitchToOnlinePreconditioning(int32 rank_in, int32 rank_out,
                                         int32 update_period,
                                         BaseFloat num_samples_history,
                                         BaseFloat alpha) {
557 558
  int32 switched = 0;
  for (size_t i = 0; i < components_.size(); i++) {
559
    if (dynamic_cast<AffineComponent*>(components_[i]) != NULL) {
560 561
      AffineComponentPreconditionedOnline *ac =
          new AffineComponentPreconditionedOnline(
562
              *(dynamic_cast<AffineComponent*>(components_[i])),
563
              rank_in, rank_out, update_period, num_samples_history, alpha);
564 565 566 567 568 569
      delete components_[i];
      components_[i] = ac;
      switched++;
    }
  }
  KALDI_LOG << "Switched " << switched << " components to use online "
570 571 572
            << "preconditioning, with (input, output) rank = "
            << rank_in << ", " << rank_out << " and num_samples_history = "
            << num_samples_history;
573 574 575 576 577
  SetIndexes();
  Check();
}


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
void Nnet::AddNnet(const VectorBase<BaseFloat> &scale_params,
                   const Nnet &other) {
  KALDI_ASSERT(scale_params.Dim() == this->NumUpdatableComponents());
  int32 i = 0;
  for (int32 j = 0; j < NumComponents(); j++) {
    UpdatableComponent *uc =
        dynamic_cast<UpdatableComponent*>(&(GetComponent(j)));
    const UpdatableComponent *uc_other =
        dynamic_cast<const UpdatableComponent*>(&(other.GetComponent(j)));
    if (uc != NULL) {
      KALDI_ASSERT(uc_other != NULL);
      BaseFloat alpha = scale_params(i);
      uc->Add(alpha, *uc_other);
      i++;
    }
  }
  KALDI_ASSERT(i == scale_params.Dim());
}

void Nnet::AddNnet(BaseFloat alpha,
                   const Nnet &other) {
  for (int32 i = 0; i < NumComponents(); i++) {
    UpdatableComponent *uc =
        dynamic_cast<UpdatableComponent*>(&(GetComponent(i)));
    const UpdatableComponent *uc_other =
        dynamic_cast<const UpdatableComponent*>(&(other.GetComponent(i)));
    if (uc != NULL) {
      KALDI_ASSERT(uc_other != NULL);
      uc->Add(alpha, *uc_other);
    }
    NonlinearComponent *nc =
        dynamic_cast<NonlinearComponent*>(&(GetComponent(i)));
    const NonlinearComponent *nc_other =
        dynamic_cast<const NonlinearComponent*>(&(other.GetComponent(i)));
    if (nc != NULL) {
      KALDI_ASSERT(nc_other != NULL);
      nc->Add(alpha, *nc_other);
    }
  }
}

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
void Nnet::AddNnet(BaseFloat alpha,
                   Nnet *other,
                   BaseFloat beta) {
  for (int32 i = 0; i < NumComponents(); i++) {
    UpdatableComponent *uc =
        dynamic_cast<UpdatableComponent*>(&(GetComponent(i)));
    UpdatableComponent *uc_other =
        dynamic_cast<UpdatableComponent*>(&(other->GetComponent(i)));
    if (uc != NULL) {
      KALDI_ASSERT(uc_other != NULL);
      uc->Add(alpha, *uc_other);
      uc_other->Scale(beta);
    }
    NonlinearComponent *nc =
        dynamic_cast<NonlinearComponent*>(&(GetComponent(i)));
    NonlinearComponent *nc_other =
        dynamic_cast<NonlinearComponent*>(&(other->GetComponent(i)));
    if (nc != NULL) {
      KALDI_ASSERT(nc_other != NULL);
      nc->Add(alpha, *nc_other);
      nc_other->Scale(beta);
    }
  }
}


645 646
void Nnet::Append(Component *new_component) {
  components_.push_back(new_component);
647
  SetIndexes();
648 649 650 651 652 653 654
  Check();
}

void Nnet::SetComponent(int32 c, Component *component) {
  KALDI_ASSERT(static_cast<size_t>(c) < components_.size());
  delete components_[c];
  components_[c] = component;
655
  SetIndexes();
656
  Check();  // Check that all the dimensions still match up.
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
}

int32 Nnet::GetParameterDim() const {
  int32 ans = 0;
  for (int32 c = 0; c < NumComponents(); c++) {
    const UpdatableComponent *uc = dynamic_cast<const UpdatableComponent*>(
        &(GetComponent(c)));
    if (uc != NULL)
      ans += uc->GetParameterDim();
  }
  return ans;
}

void Nnet::Vectorize(VectorBase<BaseFloat> *params) const {
  int32 offset = 0;
  for (int32 c = 0; c < NumComponents(); c++) {
    const UpdatableComponent *uc = dynamic_cast<const UpdatableComponent*>(
        &(GetComponent(c)));
    if (uc != NULL) {
      int32 size = uc->GetParameterDim();
      SubVector<BaseFloat> temp(*params, offset, size);
      uc->Vectorize(&temp);
      offset += size;
    }
  }
  KALDI_ASSERT(offset == GetParameterDim());
}

685 686 687 688 689 690 691 692 693 694
void Nnet::ResetGenerators() { // resets random-number generators for all random
                               // components.
  for (int32 c = 0; c < NumComponents(); c++) {
    RandomComponent *rc = dynamic_cast<RandomComponent*>(
        &(GetComponent(c)));
    if (rc != NULL)
      rc->ResetGenerator();
  }
}

695 696 697 698 699 700 701 702 703 704 705 706 707 708
void Nnet::UnVectorize(const VectorBase<BaseFloat> &params) {
  int32 offset = 0;
  for (int32 c = 0; c < NumComponents(); c++) {
    UpdatableComponent *uc = dynamic_cast<UpdatableComponent*>(
        &(GetComponent(c)));
    if (uc != NULL) {
      int32 size = uc->GetParameterDim();
      uc->UnVectorize(params.Range(offset, size));
      offset += size;
    }
  }
  KALDI_ASSERT(offset == GetParameterDim());
}

709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767
void Nnet::LimitRankOfLastLayer(int32 dim) {
  for (int32 i = components_.size() - 1; i >= 0; i--) {
    AffineComponent *a = NULL, *b = NULL,
        *c = dynamic_cast<AffineComponent*>(components_[i]);
    if (c != NULL) {
      c->LimitRank(dim, &a, &b);
      delete c;
      components_[i] = a;
      components_.insert(components_.begin() + i + 1, b);
      this->SetIndexes();
      this->Check();
      return;
    }
  }
  KALDI_ERR << "No affine component found in neural net.";
}

void Nnet::SetIndexes() {
  for (size_t i = 0; i < components_.size(); i++)
    components_[i]->SetIndex(i);
}

void Nnet::Collapse(bool match_updatableness) {
  int32 num_collapsed = 0;
  bool changed = true;
  while (changed) {
    changed = false;
    for (size_t i = 0; i + 1 < components_.size(); i++) {
      AffineComponent *a1 = dynamic_cast<AffineComponent*>(components_[i]),
          *a2 = dynamic_cast<AffineComponent*>(components_[i + 1]);
      FixedAffineComponent
          *f1 = dynamic_cast<FixedAffineComponent*>(components_[i]),
          *f2 = dynamic_cast<FixedAffineComponent*>(components_[i + 1]);
      Component *c = NULL;
      if (a1 != NULL && a2 != NULL) {
        c = a1->CollapseWithNext(*a2);
      } else if (a1 != NULL && f2 != NULL && !match_updatableness) {
        c = a1->CollapseWithNext(*f2);
      } else if (f1 != NULL && a2 != NULL && !match_updatableness) {
        c = a2->CollapseWithPrevious(*f1);
      }
      if (c != NULL) {
        delete components_[i];
        delete components_[i + 1];
        components_[i] = c;
        // This was causing valgrind errors, so doing it differently.  Either
        // a standard-library bug or I misunderstood something.
        // components_.erase(components_.begin() + i + i,
        //                   components_.begin() + i + 2);
        for (size_t j = i + 1; j + 1 < components_.size(); j++)
          components_[j] = components_[j + 1];
        components_.pop_back();
        changed = true;
        num_collapsed++;
      }
    }
  }
  this->SetIndexes();
  this->Check();
768 769 770
  KALDI_LOG << "Collapsed " << num_collapsed << " components."
            << (num_collapsed == 0 && match_updatableness == true ?
                "  Try --match-updatableness=false." : "");
771 772
}

773 774 775 776
Nnet *GenRandomNnet(int32 input_dim,
                    int32 output_dim) {
  std::vector<Component*> components;
  int32 cur_dim = input_dim;
777 778
  // have up to 10 layers before the final one.
  for (size_t i = 0; i < 10; i++) {
779 780 781 782 783 784 785 786 787 788 789 790 791 792
    if (rand() % 2 == 0) {
      // add an affine component.
      int32 next_dim = 50 + rand() % 100;
      BaseFloat learning_rate = 0.0001, param_stddev = 0.001,
          bias_stddev = 0.1;
      AffineComponent *component = new AffineComponent();
      component->Init(learning_rate, cur_dim, next_dim,
                      param_stddev, bias_stddev);
      components.push_back(component);
      cur_dim = next_dim;
    } else if (rand() % 2 == 0) {
      components.push_back(new SigmoidComponent(cur_dim));
    } else if (rand() % 2 == 0 && cur_dim < 200) {
      SpliceComponent *component = new SpliceComponent();
793 794 795 796 797 798 799 800 801 802 803
      std::vector<int32> context;
      while (true) {
        context.clear();
        for (int32 i = -3; i <= 3; i++) {
          if (rand() % 3 == 0)
            context.push_back(i);
        }
        if (!context.empty() && context.front() <= 0 &&
            context.back() >= 0)
          break;
      }
804
      component->Init(cur_dim, context);
805
      components.push_back(component);
806
      cur_dim = cur_dim * context.size();
807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828
    } else {
      break;
    }
  }

  {
    AffineComponent *component = new AffineComponent();
    BaseFloat learning_rate = 0.0001, param_stddev = 0.001,
        bias_stddev = 0.1;
    component->Init(learning_rate, cur_dim, output_dim,
                    param_stddev, bias_stddev);
    components.push_back(component);
    cur_dim = output_dim;
  }

  components.push_back(new SoftmaxComponent(cur_dim));

  Nnet *ans = new Nnet();
  ans->Init(&components);
  return ans;
}

Dan Povey's avatar
Dan Povey committed
829 830 831 832 833 834 835
int32 Nnet::FirstUpdatableComponent() const {
  for (int32 i = 0; i < NumComponents(); i++) {
    if (dynamic_cast<UpdatableComponent*>(components_[i]) != NULL)
      return i;
  }
  return NumComponents();
}
836 837


838 839 840
int32 Nnet::LastUpdatableComponent() const {
  for (int32 i = NumComponents() - 1; i >= 0; i--)
    if (dynamic_cast<UpdatableComponent*>(components_[i]) != NULL)
Dan Povey's avatar
Dan Povey committed
841 842
      return i;
  return -1;
843 844
}

845 846
}  // namespace nnet2
}  // namespace kaldi
847