RecurrentGradientMachine.cpp 45.7 KB
Newer Older
Z
zhangjinchao01 已提交
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
/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.

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

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/utils/Stat.h"
#include "paddle/utils/Util.h"
#include "paddle/utils/Flags.h"
#include <algorithm>
#include <functional>
#include <dlfcn.h>
#include <limits>
#include <cmath>
#include "RecurrentGradientMachine.h"
#include "NeuralNetwork.h"
#include "paddle/gserver/layers/AgentLayer.h"

P_DEFINE_string(diy_beam_search_prob_so, "", "the diy beam search cost so");

static const char* DIY_CALC_PROB_SYMBOL_NAME = "calc_prob";
static const char* DIY_START_CALC_PROB_SYMBOL_NAME = "start_calc_prob";
static const char* DIY_FINISH_CALC_PROB_SYMBOL_NAME = "finish_calc_prob";

namespace paddle {

/**
 * Start Custom Calculate Probability callback type.
 *
 * @param nNode, nodes: the path will be explored. nNodes is array size.
 *                      nodes is array elements.
 *
 * @return: A custom handler id that will passed to another callback.
 */
typedef int (*DiyStartCalcProbCallback)(size_t nNodes, int* nodes);

/**
 * Doing Custom Calculation of Probability callback type.
 *
 * @param handler: User custom handler. The return value from start calc prob.
 * @param nNode, nodes: Array. The current path.
 * @param curProb: The current log probability that neural network returns.
 *
 * @return: Log probability which user calculated, it will be updated to this
 *          path.
 * @NOTE: Return -INFINITY will DROP this path IMMEDIATELY!!
 */
typedef real (*DiyCalcProbCallback)(int handler, size_t nNodes, int* nodes,
                                    real curProb, bool atEos);

/**
 * Finish Custom Calculation of Probability callback type.
 *
 * @param handler: User custom handler. The return value from start calc prob.
 */
typedef void (*DiyStopCalcProbCallback)(int handler);

static DiyCalcProbCallback gDiyProbMethod = nullptr;
static DiyStartCalcProbCallback gDiyProbStart = nullptr;
static DiyStopCalcProbCallback gDiyProbStop = nullptr;
static void* gDiyProbHandle = nullptr;

static void exit_diy_prob() { dlclose(gDiyProbHandle); }

template <typename SymbolType>
static inline SymbolType loadDiySymbol(const char* symbolName) {
  void* sym = dlsym(gDiyProbHandle, symbolName);
  CHECK(sym) << "Cannot load symbol " << symbolName << " from "
             << FLAGS_diy_beam_search_prob_so;
  return reinterpret_cast<SymbolType>(sym);
}

static InitFunction __init__diy_prob_method([] {
  std::string soName = FLAGS_diy_beam_search_prob_so;
  if (!soName.empty()) {
    gDiyProbHandle = dlopen(soName.c_str(), RTLD_LAZY);
    CHECK(gDiyProbHandle) << "Cannot Open DIY Prob So " << soName;
    atexit(exit_diy_prob);
    gDiyProbMethod =
        loadDiySymbol<decltype(gDiyProbMethod)>(DIY_CALC_PROB_SYMBOL_NAME);
    gDiyProbStart =
        loadDiySymbol<decltype(gDiyProbStart)>(DIY_START_CALC_PROB_SYMBOL_NAME);
    gDiyProbStop =
        loadDiySymbol<decltype(gDiyProbStop)>(DIY_FINISH_CALC_PROB_SYMBOL_NAME);
  }
}, std::numeric_limits<int>::max());

class BeamSearchControlCallbacks {
public:
  RecurrentGradientMachine::BeamSearchCandidatesAdjustCallback
      beamSearchCandidateAdjust;
  RecurrentGradientMachine::NormOrDropNodeCallback normOrDropNode;
  RecurrentGradientMachine::DropCallback stopDetermineCandidates;

  //! for gcc46 aggregate initialization is not very well, so we need to
  //! explicit
  BeamSearchControlCallbacks(
      const RecurrentGradientMachine::BeamSearchCandidatesAdjustCallback&
          candidateAdjust,
      const RecurrentGradientMachine::NormOrDropNodeCallback& norm,
      const RecurrentGradientMachine::DropCallback& stop)
      : beamSearchCandidateAdjust(candidateAdjust),
        normOrDropNode(norm),
        stopDetermineCandidates(stop) {}
};

class BeamSearchStatisticsCallbacks {
public:
  RecurrentGradientMachine::EachStepCallback onEachStepStarted;
  RecurrentGradientMachine::EachStepCallback onEachStepStoped;

  BeamSearchStatisticsCallbacks(
      const RecurrentGradientMachine::EachStepCallback& start,
      const RecurrentGradientMachine::EachStepCallback& stop)
      : onEachStepStarted(start), onEachStepStoped(stop) {}
};

RecurrentGradientMachine::RecurrentGradientMachine(
    const std::string& subModelName, NeuralNetwork* rootNetwork)
    : NeuralNetwork(subModelName),
      rootNetwork_(rootNetwork),
      beamSearchCtrlCallbacks_(nullptr),
      beamSearchStatistics_(nullptr) {
  CHECK(!subModelName_.empty());
}

/**
 * bias layer, as input of memory frame 0 will give vector of zeros
 * if bias parameter is not set.
 *
 * boot bias layer create directly in recurrent gradient machine, because:
 *
 * 1. It is only one frame, so it should not be placed in layer group,
 *    which is one instance for every one frame.
 *
 * 2. It is no input layer, so it need resetHeight() before forward(),
 *    and resetHeight() must be called in recurrent gradient machine,
 *    so it's should not be placed in root network.
 */
class BootBiasLayer : public Layer {
protected:
  std::unique_ptr<Weight> biases_;
  IVectorPtr cpuIds_;

public:
  explicit BootBiasLayer(const LayerConfig& config) : Layer(config) {}

  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
    if (!Layer::init(layerMap, parameterMap)) return false;

    if (biasParameter_) {
      biases_ =
          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_));
    }
    return true;
  }

  void resetHeight(int height) {
    if (config_.has_bos_id()) {  // used as a constant id layerConfig
      IVector::resizeOrCreate(output_.ids, height, useGpu_);
      output_.ids->reset((int)config_.bos_id());
    } else {
      resetOutput(height, getSize());
    }
  }

  virtual void forward(PassType passType) {
    if (biases_) {
      MatrixPtr outV = getOutputValue();
      outV->addBias(*(biases_->getW()), 1);
      forwardActivation();
    }
  }

  virtual void backward(const UpdateCallback& callback) {
    if (biases_) {
      backwardActivation();
      biases_->getWGrad()->collectBias(*getOutputGrad(), 1);
      biases_->getParameterPtr()->incUpdate(callback);
    }
  }
};

void RecurrentGradientMachine::init(
    const ModelConfig& config, ParamInitCallback callback,
    const std::vector<ParameterType>& parameterTypes, bool useGpu) {
  NeuralNetwork::init(config, callback, parameterTypes, useGpu);
  useGpu_ = useGpu;

  auto subModelConfig =
      std::find_if(config.sub_models().begin(), config.sub_models().end(),
                   [this](const SubModelConfig& sub_model) {
                     return sub_model.name() == this->subModelName_;
                   });
  CHECK(subModelConfig != config.sub_models().end());
  reversed_ = subModelConfig->reversed();

  inFrameLines_.resize(subModelConfig->in_links_size());
  for (size_t i = 0; i < inFrameLines_.size(); ++i) {
    inFrameLines_[i].linkName = subModelConfig->in_links(i).link_name();
    inFrameLines_[i].inLayer =
        rootNetwork_->getLayer(subModelConfig->in_links(i).layer_name());
    inFrameLines_[i].hasSubseq = subModelConfig->in_links(i).has_subseq();
  }

  outFrameLines_.resize(subModelConfig->out_links_size());
  for (size_t i = 0; i < outFrameLines_.size(); ++i) {
    auto& linkPair = subModelConfig->out_links(i);
    outFrameLines_[i].layerName = linkPair.layer_name();
    outFrameLines_[i].agentLayer = rootNetwork_->getLayer(linkPair.link_name());
  }

  memoryFrameLines_.resize(subModelConfig->memories_size());
  for (size_t i = 0; i < memoryFrameLines_.size(); ++i) {
    auto& memoryConfig = subModelConfig->memories(i);
    memoryFrameLines_[i].layerName = memoryConfig.layer_name();
    memoryFrameLines_[i].linkName = memoryConfig.link_name();
    auto agentConfig =
        std::find_if(config.layers().begin(), config.layers().end(),
                     [&memoryConfig](const LayerConfig& layerConfig) {
                       return layerConfig.name() == memoryConfig.link_name();
                     });
    CHECK(agentConfig != config.layers().end());
    if (memoryConfig.has_boot_layer_name()) {
      memoryFrameLines_[i].rootLayer =
          rootNetwork_->getLayer(memoryConfig.boot_layer_name());

      LayerConfig scatterConfig = *agentConfig;
      memoryFrameLines_[i].is_sequence = memoryConfig.is_sequence();
      memoryFrameLines_[i].rootAgent.reset(
          memoryConfig.is_sequence()
              ? new SequenceScatterAgentLayer(scatterConfig)
              : new ScatterAgentLayer(scatterConfig));
      memoryFrameLines_[i].rootAgent->init(LayerMap(), parameterMap_);

      memoryFrameLines_[i].bootLayer = memoryFrameLines_[i].rootAgent;
    } else {
      LayerConfig biasConfig = *agentConfig;
      if (memoryConfig.has_boot_bias_parameter_name()) {
        biasConfig.set_bias_parameter_name(
            memoryConfig.boot_bias_parameter_name());
        biasConfig.set_active_type(memoryConfig.boot_bias_active_type());
      } else if (memoryConfig.has_boot_with_const_id()) {
        biasConfig.set_bos_id(memoryConfig.boot_with_const_id());
      }
      memoryFrameLines_[i].biasLayer.reset(new BootBiasLayer(biasConfig));
      memoryFrameLines_[i].biasLayer->init(LayerMap(), parameterMap_);

      memoryFrameLines_[i].bootLayer = memoryFrameLines_[i].biasLayer;
    }

    if (subModelConfig->has_generator()) {
      memoryFrameLines_[i].scatterAgents.resize(2);
      for (auto& agent : memoryFrameLines_[i].scatterAgents) {
        agent.reset(memoryConfig.is_sequence()
                        ? new SequenceScatterAgentLayer(*agentConfig)
                        : new ScatterAgentLayer(*agentConfig));
        agent->init(LayerMap(), parameterMap_);
      }
    }
  }

  if (subModelConfig->has_generator()) {
    generator_.config = subModelConfig->generator();
    eosFrameLine_.reset(new EosFrameLine);
    maxSequenceLength_ = generator_.config.max_num_frames();
  }

  // get parameters actually used by this Layer Group
  resizeOrCreateFrames(1);
  for (auto& para : frames_[0]->getParameters()) {
    if (para->getSharedCount() > 0) {
      parameterIds_.push_back(para->getID());
    }
  }
  for (auto& para : parameters_) {  // bias layer parameters
    if (para->getSharedCount() > 0) {
      parameterIds_.push_back(para->getID());
    }
  }

  if (subModelConfig->evaluator_names_size() > 0) {
    evaluator_.reset(frames_[0]->makeEvaluator());
  }
293 294

  targetInfoInlinkId_ = subModelConfig->target_inlinkid();
Z
zhangjinchao01 已提交
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
}

void RecurrentGradientMachine::resizeOrCreateFrames(int numFrames) {
  if ((size_t)numFrames <= frames_.size()) {
    return;
  }

  frames_.reserve(numFrames);
  for (auto& inFrameLine : inFrameLines_) {
    inFrameLine.agents.reserve(numFrames);
  }
  for (auto& outFrameLine : outFrameLines_) {
    outFrameLine.frames.reserve(numFrames);
  }
  for (auto& memoryFrameLine : memoryFrameLines_) {
    memoryFrameLine.frames.reserve(numFrames);
    memoryFrameLine.agents.reserve(numFrames);
  }
  if (eosFrameLine_) {
    eosFrameLine_->layers.reserve(numFrames);
  }

  ParamInitCallback subParamInitCb = [this](int paramId, Parameter* para) {
    para->enableSharedType(PARAMETER_VALUE,
                           this->parameters_[paramId]->getBuf(PARAMETER_VALUE),
                           this->parameters_[paramId]->getMat(PARAMETER_VALUE));
    para->enableSharedType(
        PARAMETER_GRADIENT,
        this->parameters_[paramId]->getBuf(PARAMETER_GRADIENT),
        this->parameters_[paramId]->getMat(PARAMETER_GRADIENT));
  };

  for (int i = frames_.size(); i < numFrames; ++i) {
    std::unique_ptr<NeuralNetwork> frame(
329
        NeuralNetwork::newNeuralNetwork(subModelName_));
Z
zhangjinchao01 已提交
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
    frame->init(config_, subParamInitCb);

    for (auto& inFrameLine : inFrameLines_) {
      inFrameLine.agents.push_back(frame->getLayer(inFrameLine.linkName));
    }

    for (auto& outFrameLine : outFrameLines_) {
      outFrameLine.frames.push_back(frame->getLayer(outFrameLine.layerName));
    }
    for (auto& memoryFrameLine : memoryFrameLines_) {
      memoryFrameLine.frames.push_back(
          frame->getLayer(memoryFrameLine.layerName));
      memoryFrameLine.agents.push_back(
          frame->getLayer(memoryFrameLine.linkName));
    }
    if (eosFrameLine_) {
      eosFrameLine_->layers.push_back(
          frame->getLayer(generator_.config.eos_layer_name()));
    }

    frames_.emplace_back(std::move(frame));
  }
}

void RecurrentGradientMachine::resizeBootFrame(int numSequences) {
  for (auto& memoryFrameLine : memoryFrameLines_) {
    if (memoryFrameLine.biasLayer) {
      auto biasLayer =
          dynamic_cast<BootBiasLayer*>(memoryFrameLine.biasLayer.get());
      CHECK_NOTNULL(biasLayer);
      biasLayer->resetHeight(numSequences);
    } else {  // check input root layer height
      CHECK_EQ(numSequences,
               memoryFrameLine.rootLayer->getOutput().getNumSequences());
    }
  }
}

void RecurrentGradientMachine::prefetch(const std::vector<Argument>& inArgs) {
  LOG(FATAL) << "should not use this function";
}

void RecurrentGradientMachine::forward(const std::vector<Argument>& inArgs,
                                       std::vector<Argument>* outArgs,
                                       PassType passType) {
  if (inFrameLines_.empty() && passType == PASS_TEST) {
    generateSequence();
    return;
  }  // else forward..

  const Argument& input = inFrameLines_[0].inLayer->getOutput();
  CHECK(input.sequenceStartPositions);
  int batchSize = input.getBatchSize();
  size_t numSequences = input.getNumSequences();
  const int* starts = input.sequenceStartPositions->getData(false);
  bool hasSubseq = input.hasSubseq();
386 387 388 389 390 391 392 393 394 395

  // In case of !hasSubseq or targetInfoInlinkId_ == -1, all inlinks share the
  // same inframe info
  bool shareInlinkInfo = !hasSubseq || targetInfoInlinkId_ == -1;

  // Defaultly, share info with the first inlink
  if (shareInlinkInfo) {
    targetInfoInlinkId_ = 0;
  }

Z
zhangjinchao01 已提交
396 397 398 399 400 401 402 403 404 405 406 407
  // check hasSubseq in both config and input are the same
  CHECK_EQ(hasSubseq, inFrameLines_[0].hasSubseq);

  CHECK_EQ(starts[numSequences], batchSize);
  CHECK(input.sequenceStartPositions);

  // check other inputs has same sequence length and start
  for (size_t i = 1; i < inFrameLines_.size(); ++i) {
    const Argument& input1 = inFrameLines_[i].inLayer->getOutput();
    CHECK_EQ((size_t)input1.getNumSequences(), numSequences);
    // check all inputs should have same hasSubseq flag
    CHECK_EQ(input.hasSubseq(), inFrameLines_[0].hasSubseq);
408 409 410 411 412 413 414 415 416 417 418

    // if shareInlinkInfo, checks:
    // 1. all inlinks have same number of total tokens
    // 2. all inlinks have same number of tokens for each sentence of each
    //    sample. If hasSubseq, one sample has multiple sentence, else, one
    //    sample is one sentence
    if (shareInlinkInfo) {
      CHECK_EQ(input1.getBatchSize(), batchSize);
      CHECK(std::equal(starts, starts + numSequences + 1,
                       input1.sequenceStartPositions->getData(false)));
    }
Z
zhangjinchao01 已提交
419 420 421 422 423 424 425 426 427 428 429
  }

  if (hasSubseq) {
    CHECK(input.subSequenceStartPositions);
    size_t numSubSequences = input.getNumSubSequences();
    const int* subStarts = input.subSequenceStartPositions->getData(false);
    CHECK_EQ(subStarts[numSubSequences], batchSize);
    // if hasSubseq, check other inputs has same sub-sequence and sub-start
    for (size_t i = 1; i < inFrameLines_.size(); ++i) {
      const Argument& input1 = inFrameLines_[i].inLayer->getOutput();
      CHECK_EQ((size_t)input1.getNumSubSequences(), numSubSequences);
430 431 432 433
      if (shareInlinkInfo) {
        CHECK(std::equal(subStarts, subStarts + numSubSequences + 1,
                         input1.subSequenceStartPositions->getData(false)));
      }
Z
zhangjinchao01 已提交
434 435 436
    }
  }

437 438
  info_.clear();
  info_.resize(inFrameLines_.size());
439 440 441

  seqInfos_.clear();
  seqInfos_.resize(inFrameLines_.size());
442 443 444 445 446 447

  {
    AsyncGpuBlock asyncGpuBlock;
    // if shareInlinkInfo, only calculate info of the first inlink
    // else, calculate info for each inlink
    if (shareInlinkInfo) {
448 449
      input.getSeqInfo(&seqInfos_[0]);
      maxSequenceLength_ = seqInfos_[0][0].topLevelLength;
450 451 452 453
      createInFrameInfo(0, input, passType);
    } else {
      for (size_t i = 0; i < inFrameLines_.size(); i++) {
        const Argument& input1 = inFrameLines_[i].inLayer->getOutput();
454 455
        input1.getSeqInfo(&seqInfos_[i]);
        maxSequenceLength_ = seqInfos_[i][0].topLevelLength;
456 457 458 459 460 461 462 463 464 465 466
        createInFrameInfo(i, input1, passType);
      }
    }

    // inFrameLine select rows in real layer one time
    for (size_t i = 0; i < inFrameLines_.size(); i++) {
      int curInlinkId = shareInlinkInfo ? 0 : i;
      selectRowsOneTime(inFrameLines_[i].inLayer, info_[curInlinkId].allIds,
                        &(inFrameLines_[i].outArg), passType);
    }
  }
Z
zhangjinchao01 已提交
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
  resizeOrCreateFrames(maxSequenceLength_);
  resizeBootFrame(numSequences);

  for (auto& memoryFrameLine : memoryFrameLines_) {
    if (memoryFrameLine.rootAgent) {
      auto scatterAgent =
          dynamic_cast<ScatterAgentLayer*>(memoryFrameLine.rootAgent.get());
      createMemoryFrameInfo(&memoryFrameLine, passType);
      scatterAgent->setRealLayerAndOutput(
          memoryFrameLine.rootLayer, memoryFrameLine.outArg,
          memoryFrameLine.allIds,
          /* idIndex */ 0, memoryFrameLine.allIds->getSize());
      if (memoryFrameLine.is_sequence) {  // memoryConfig is sequence
        int size = memoryFrameLine.sequenceStartPositions->getSize();
        scatterAgent->setSequenceStartPositions(
            memoryFrameLine.sequenceStartPositions,
            /* seqStartPosIndex */ 0, size);
      }
    }
  }

  for (auto& outFrameLine : outFrameLines_) {
    auto gatherAgent =
        dynamic_cast<GatherAgentLayer*>(outFrameLine.agentLayer.get());
    CHECK_NOTNULL(gatherAgent);
492 493
    gatherAgent->copyIdAndSequenceInfo(input, info_[targetInfoInlinkId_].allIds,
                                       info_[targetInfoInlinkId_].idIndex);
Z
zhangjinchao01 已提交
494 495 496
  }

  for (int i = 0; i < maxSequenceLength_; ++i) {
497
    int idSize = 0;
Z
zhangjinchao01 已提交
498
    // connect in_links
499 500 501 502
    for (size_t j = 0; j < inFrameLines_.size(); ++j) {
      // idSize denotes the sum number of tokens in each length i
      idSize = info_[j].idIndex[i + 1] - info_[j].idIndex[i];
      InFrameLine inFrameLine = inFrameLines_[j];
Z
zhangjinchao01 已提交
503 504 505
      auto scatterAgent =
          dynamic_cast<ScatterAgentLayer*>(inFrameLine.agents[i].get());
      scatterAgent->setRealLayerAndOutput(inFrameLine.inLayer,
506 507
                                          inFrameLine.outArg, info_[j].allIds,
                                          info_[j].idIndex[i], idSize);
Z
zhangjinchao01 已提交
508
      if (hasSubseq) {
509 510 511 512 513 514
        // size: the length of subsequence
        int size =
            info_[j].seqStartPosIndex[i + 1] - info_[j].seqStartPosIndex[i];
        scatterAgent->setSequenceStartPositions(info_[j].sequenceStartPositions,
                                                info_[j].seqStartPosIndex[i],
                                                size);
Z
zhangjinchao01 已提交
515 516 517 518 519 520 521 522 523 524
      }
    }

    // connect out_links
    for (auto& outFrameLine : outFrameLines_) {
      auto gatherAgent =
          dynamic_cast<GatherAgentLayer*>(outFrameLine.agentLayer.get());
      gatherAgent->addRealLayer(outFrameLine.frames[i]);
    }
    // connect memory links
525 526 527 528
    // Adopt info_[0].idIndex because seq which has_subseq=True
    // doesn't support Memory with !hasSubseq bootlayer;
    // And inlinks that !hasSubSeq must have same inlink length.
    idSize = info_[0].idIndex[i + 1] - info_[0].idIndex[i];
Z
zhangjinchao01 已提交
529 530 531 532
    for (auto& memoryFrameLine : memoryFrameLines_) {
      NeuralNetwork::connect(
          memoryFrameLine.agents[i],
          i == 0 ? memoryFrameLine.bootLayer : memoryFrameLine.frames[i - 1],
533
          numSeqs_[i] /*height of agent*/);
Z
zhangjinchao01 已提交
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
    }
  }

  REGISTER_TIMER_INFO("RecurrentFwTime", "RecurrentFwTime");
  // forward
  for (auto& memoryFrameLine : memoryFrameLines_) {
    memoryFrameLine.bootLayer->forward(passType);
  }
  for (int i = 0; i < maxSequenceLength_; ++i) {
    const std::vector<Argument> inArgs;
    std::vector<Argument> outArgs;
    frames_[i]->forward(inArgs, &outArgs, passType);
  }
  if (evaluator_ && passType == PASS_TEST) {
    this->eval(evaluator_.get());
  }
}

void RecurrentGradientMachine::backward(const UpdateCallback& callback) {
  REGISTER_TIMER_INFO("RecurrentBwTime", "RecurrentBwTime");
  AsyncGpuBlock asyncGpuBlock;
  for (int i = maxSequenceLength_ - 1; i >= 0; --i) {
    frames_[i]->backward(nullptr);
  }
  for (auto& memoryFrameLine : memoryFrameLines_) {
    memoryFrameLine.bootLayer->backward(nullptr);
  }

  // call printers here so the gradient can be printed
  if (evaluator_) {
    this->eval(evaluator_.get());
  }
}

void RecurrentGradientMachine::forwardBackward(
    const std::vector<Argument>& inArgs, std::vector<Argument>* outArgs,
    PassType passType, const UpdateCallback& callback) {
  LOG(FATAL) << "should not use this function";
}

void RecurrentGradientMachine::eval(Evaluator* evaluator) {
  // call printers frame by frame
  for (int i = 0; i < maxSequenceLength_; ++i) {
    LOG(INFO) << "Recurrent Layer Group eval frame " << i << " begin";
    evaluator->eval(*(frames_[i].get()));
    LOG(INFO) << "Recurrent Layer Group eval frame " << i << " end";
  }
}

void RecurrentGradientMachine::registerBeamSearchControlCallbacks(
    const BeamSearchCandidatesAdjustCallback& adjustBeamSearch,
    const NormOrDropNodeCallback& normOrDropNode,
    const DropCallback& stopBeamSearch) {
  this->removeBeamSearchControlCallbacks();
  //! for gcc 46, aggregate initialization is not supported. TAT
  this->beamSearchCtrlCallbacks_ = new BeamSearchControlCallbacks(
      adjustBeamSearch, normOrDropNode, stopBeamSearch);
}

void RecurrentGradientMachine::removeBeamSearchControlCallbacks() {
  if (this->beamSearchCtrlCallbacks_) {
    delete this->beamSearchCtrlCallbacks_;
    this->beamSearchCtrlCallbacks_ = nullptr;
  }
}

void RecurrentGradientMachine::registerBeamSearchStatisticsCallbacks(
    const EachStepCallback& onEachStepStarted,
    const EachStepCallback& onEachStepStoped) {
  this->removeBeamSearchStatisticsCallbacks();
  this->beamSearchStatistics_ =
      new BeamSearchStatisticsCallbacks(onEachStepStarted, onEachStepStoped);
}

void RecurrentGradientMachine::removeBeamSearchStatisticsCallbacks() {
  if (this->beamSearchStatistics_) {
    delete this->beamSearchStatistics_;
    this->beamSearchStatistics_ = nullptr;
  }
}
/* create scattered id infomation for all realLayer of inFrameLines one time.
 * If hasSubseq, will also create scattered sequenceStartPositions infomation
 * for all realLayer of inFrameLines one time.
*/
618

619
void RecurrentGradientMachine::createInFrameInfo(int inlinkId,
620
                                                 const Argument& input,
Z
zhangjinchao01 已提交
621 622
                                                 PassType passType) {
  bool hasSubseq = input.hasSubseq();
623
  // numSequences: # samples(sequences) in a batch
Z
zhangjinchao01 已提交
624 625
  size_t numSequences = input.getNumSequences();
  std::vector<int> allIds;
626

627 628
  auto& seqInfo = seqInfos_[inlinkId];

629
  numSeqs_.clear();
630 631 632 633 634 635 636
  Info* inlinkInfo = &info_[inlinkId];
  inlinkInfo->idIndex.clear();
  inlinkInfo->idIndex.push_back(0);  // first idIndex = 0

  std::vector<int> sequenceStartPositions;
  const int* subSequenceStartPositions = nullptr;

637
  if (hasSubseq) {                    // for sequenceScatterAgentLayer
638 639 640 641 642 643 644 645
    subSequenceStartPositions =
        input.subSequenceStartPositions->getData(false);
    inlinkInfo->seqStartPosIndex.clear();
    inlinkInfo->seqStartPosIndex.push_back(0);  // first seqStartPosIndex = 0
  }
  // maxSequenceLength_: max topLevelLength in allsamples
  for (int i = 0; i < maxSequenceLength_; ++i) {
    if (hasSubseq) {
646
      sequenceStartPositions.push_back(0);            // first element = 0
Z
zhangjinchao01 已提交
647
    }
648 649 650 651 652 653 654 655 656 657 658 659 660
    int numSeqs = 0;
    for (size_t j = 0; j < numSequences; ++j) {
      int seqLength = seqInfo[j].topLevelLength;
      if (i >= seqLength) {
        break;
      }
      ++numSeqs;
      if (hasSubseq) {
        int subSeqStart = subSequenceStartPositions[seqInfo[j].subSeqStart + i];
        int subSeqEnd =
            subSequenceStartPositions[seqInfo[j].subSeqStart + i + 1];
        for (int k = subSeqStart; k < subSeqEnd; ++k) {
          allIds.push_back(k);
Z
zhangjinchao01 已提交
661
        }
662 663 664 665
        sequenceStartPositions.push_back(sequenceStartPositions.back() +
                                         subSeqEnd - subSeqStart);
      } else {
        int seqStart = seqInfo[j].seqStart;
Z
zhangjinchao01 已提交
666 667 668 669
        allIds.push_back(reversed_ ? (seqStart + seqLength - 1 - i)
                                   : (seqStart + i));
      }
    }
670 671 672 673 674 675 676 677 678
    inlinkInfo->idIndex.push_back(allIds.size());
    numSeqs_.push_back(numSeqs);
    if (hasSubseq) {
      inlinkInfo->seqStartPosIndex.push_back(sequenceStartPositions.size());
    }
  }
  if (hasSubseq) {
    // inFrameLine create sequenceStartPositions one time
    CHECK_EQ(sequenceStartPositions.size(),
679 680
             static_cast<size_t>(maxSequenceLength_ +
                                 input.getNumSubSequences()));
681 682 683
    CHECK_EQ(inlinkInfo->seqStartPosIndex.size(),
             static_cast<size_t>(maxSequenceLength_ + 1));
    createSeqPos(sequenceStartPositions, &inlinkInfo->sequenceStartPositions);
Z
zhangjinchao01 已提交
684
  }
685

Z
zhangjinchao01 已提交
686
  // copy and check scatterId
687 688
  copyScattedId(allIds, &inlinkInfo->allIds, input.getBatchSize());
  CHECK_EQ(inlinkInfo->idIndex.size(),
689
           static_cast<size_t>(maxSequenceLength_ + 1));
Z
zhangjinchao01 已提交
690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706
}

/* like createInFrameInfo, but for all realLayer of memoryFrameLines*/
void RecurrentGradientMachine::createMemoryFrameInfo(
    MemoryFrameLine* memoryFrameLine, PassType passType) {
  const Argument& input = (*memoryFrameLine).rootLayer->getOutput();
  size_t numSequences = input.getNumSequences();
  std::vector<int> allIds;
  bool seqFlag = (*memoryFrameLine).is_sequence;

  if (seqFlag) {  // for sequenceScatterAgentLayer
    CHECK(input.sequenceStartPositions)
        << "boot layer must be a sequence when is_sequence = true";
    std::vector<int> sequenceStartPositions;
    sequenceStartPositions.push_back(0);  // first element = 0
    const int* starts = input.sequenceStartPositions->getData(false);
    for (size_t i = 0; i < numSequences; ++i) {
707
      // memory info adopt info of inlinks[0]
708
      int seqId = seqInfos_[0][i].seqId;
Z
zhangjinchao01 已提交
709 710 711 712
      for (int k = starts[seqId]; k < starts[seqId + 1]; ++k) {
        allIds.push_back(k);
      }
      sequenceStartPositions.push_back(sequenceStartPositions.back() +
713
                                       starts[seqId + 1] - starts[seqId]);
Z
zhangjinchao01 已提交
714 715 716 717 718 719
    }
    createSeqPos(sequenceStartPositions,
                 &(*memoryFrameLine).sequenceStartPositions);

  } else {  // for scatterAgentLayer
    for (size_t i = 0; i < numSequences; ++i) {
720
      allIds.push_back(seqInfos_[0][i].seqId);
Z
zhangjinchao01 已提交
721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773
    }
  }
  // copy and check scatterId
  copyScattedId(allIds, &(*memoryFrameLine).allIds, input.getBatchSize());
  // memoryFrameLine select rows in real layer one time
  selectRowsOneTime((*memoryFrameLine).rootLayer, (*memoryFrameLine).allIds,
                    &(*memoryFrameLine).outArg, passType);
}

void RecurrentGradientMachine::copyScattedId(std::vector<int>& srcIds,
                                             IVectorPtr* dstIds, int size) {
  int idSize = srcIds.size();
  CHECK_EQ(idSize, size);
  IVector::resizeOrCreate(*dstIds, idSize, useGpu_);
  (*dstIds)->copyFrom(srcIds.data(), idSize);
  // check
  std::sort(srcIds.begin(), srcIds.end());
  for (int i = 0; i < idSize; ++i) {
    CHECK_EQ(srcIds[i], i);
  }
}

void RecurrentGradientMachine::selectRowsOneTime(LayerPtr layer,
                                                 const IVectorPtr& allIds,
                                                 Argument* arg,
                                                 PassType passType) {
  const MatrixPtr& realV = layer->getOutputValue();
  int height = realV->getHeight();
  int width = realV->getWidth();
  Matrix::resizeOrCreate(arg->value, height, width, /* trans */ false, useGpu_);
  arg->value->zeroMem();
  arg->value->selectRows(*realV, *allIds);
  if (passType != PASS_TEST) {
    Matrix::resizeOrCreate(arg->grad, height, width, /* trans */ false,
                           useGpu_);
    arg->grad->zeroMem();
  }
}

void RecurrentGradientMachine::createSeqPos(
    const std::vector<int>& sequenceStartPosition,
    ICpuGpuVectorPtr* sequenceStartPositions) {
  int size = sequenceStartPosition.size();
  const int* data = sequenceStartPosition.data();
  ICpuGpuVector::resizeOrCreate(*sequenceStartPositions, size, false);
  (*sequenceStartPositions)->copyFrom(data, size, false);
}

size_t RecurrentGradientMachine::getGenBatchSize() {
  size_t numSequences = 0;
  for (auto& memoryFrameLine : memoryFrameLines_) {
    if (!memoryFrameLine.rootLayer) continue;
    Argument& bootArg = memoryFrameLine.rootLayer->getOutput();
774 775
    size_t batchSize = memoryFrameLine.is_sequence ? bootArg.getNumSequences()
                                                   : bootArg.getBatchSize();
Z
zhangjinchao01 已提交
776 777 778 779 780 781
    if (numSequences) {
      CHECK_EQ(numSequences, batchSize);
    } else {
      numSequences = batchSize;
    }
  }
782 783 784 785 786
  CHECK(numSequences)
      << "Fail to get batch size in generation. "
         "At least one of the Memory layer MUST have a layer that is NOT in "
         "the layer group to boot it, and this boot layer is used to "
         "decide batch_size in generation process.";
Z
zhangjinchao01 已提交
787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807
  return numSequences;
}

void RecurrentGradientMachine::generateSequence() {
  CHECK_NOTNULL(eosFrameLine_.get());
  CHECK_GE(outFrameLines_.size(), 1UL);
  size_t numSequences = getGenBatchSize();

  resizeBootFrame(numSequences);
  // We create only two sub-network in generation for alternate use.
  // Thus, we can reduce total memory of output_ in layer forward.
  resizeOrCreateFrames(2);

  // outFrameLines_.size() > 1UL
  dataArgsSize_ = outFrameLines_.size() - 1;
  dataArgs_.resize(dataArgsSize_);
  dataArgsFrame_.clear();
  dataArgsFrame_.resize(dataArgsSize_);

  // connect boot frame memory links
  std::vector<int> ids(numSequences);
808 809 810
  for (size_t i = 0; i < numSequences; ++i) {
    ids[i] = i;
  }
Z
zhangjinchao01 已提交
811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833
  for (auto& memoryFrameLine : memoryFrameLines_) {
    if (memoryFrameLine.rootAgent) {
      auto scatterAgent =
          dynamic_cast<ScatterAgentLayer*>(memoryFrameLine.rootAgent.get());
      bool seqFlag = memoryFrameLine.is_sequence;
      scatterAgent->setRealLayer(memoryFrameLine.rootLayer, ids, seqFlag);
      if (seqFlag) {
        CHECK(memoryFrameLine.rootLayer->getOutput().sequenceStartPositions)
            << "boot layer must be a sequence when is_sequence = true";
      }
    }
    NeuralNetwork::connect(memoryFrameLine.agents[0], memoryFrameLine.bootLayer,
                           ids.size());
  }

  // boot layer forward
  AsyncGpuBlock asyncGpuBlock;
  for (auto& memoryFrameLine : memoryFrameLines_) {
    memoryFrameLine.bootLayer->forward(PASS_TEST);
  }

  // init outArg
  size_t resultNum = generator_.config.num_results_per_sample();
834 835
  IVector::resizeOrCreate(
      generator_.outArg.ids,
Z
zhangjinchao01 已提交
836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925
      generator_.config.max_num_frames() * numSequences * resultNum, false);
  if (resultNum > 1) {
    CHECK_LE(resultNum, static_cast<size_t>(generator_.config.beam_size()));
    Matrix::resizeOrCreate(generator_.outArg.in, /* height */ numSequences,
                           /* width */ resultNum, false, /* useGpu */ false);
  }
  ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions,
                                numSequences + 1, /* useGpu */ false);
  if (getBeamSize() > 1) {
    beamSearch(numSequences);
  } else {
    oneWaySearch(numSequences);
  }
  if (dataArgsSize_) createDataOutlink(batchMachineIdVec_);

  size_t size = generator_.ids.size();
  generator_.outArg.ids->resize(size);
  generator_.outArg.ids->copyFrom(generator_.ids.data(), size);

  OutFrameLine& outFrameLine = outFrameLines_[0];
  auto dataAgent = dynamic_cast<DataLayer*>(outFrameLine.agentLayer.get());
  CHECK_NOTNULL(dataAgent);
  dataAgent->setData(generator_.outArg);
  dataAgent->prefetch();
}

void RecurrentGradientMachine::oneWaySearch(size_t batchSize) {
  OutFrameLine& outFrameLine = outFrameLines_[0];

  // finalPaths_[0] stores the generated results of the
  // entire batch, so its size exactly equals to batchSize.
  finalPaths_.clear();
  finalPaths_.resize(1);
  std::vector<Path>& finalPaths = finalPaths_[0];
  finalPaths.resize(batchSize);

  seqIds_.resize(batchSize);
  std::vector<int> scatterIds;
  for (size_t i = 0; i < batchSize; ++i) {
    finalPaths[i].seqId = i;
    seqIds_[i] = i;
  }

  // forward
  for (int i = 0; i < maxSequenceLength_; ++i) {
    if (i && scatterIds.empty()) break;
    int machineCur = i % 2;
    int machinePrev = (i - 1) % 2;
    // connect memory links
    if (i) {
      seqIds_.clear();
      for (size_t j = 0; j < batchSize; ++j) {
        if (finalPaths[j].seqId != -1) seqIds_.push_back(j);
      }

      for (auto& memoryFrameLine : memoryFrameLines_) {
        auto scatterAgent = dynamic_cast<ScatterAgentLayer*>(
            memoryFrameLine.scatterAgents[machineCur].get());
        scatterAgent->setRealLayer(memoryFrameLine.frames[machinePrev],
                                   scatterIds, memoryFrameLine.is_sequence);
        scatterAgent->forward(PASS_TEST);
        NeuralNetwork::connect(memoryFrameLine.agents[machineCur],
                               memoryFrameLine.scatterAgents[machineCur]);
      }
    }
    const std::vector<Argument> inArgs;
    std::vector<Argument> outArgs;
    frames_[machineCur]->forward(inArgs, &outArgs, PASS_TEST);

    const IVectorPtr& idVec = outFrameLine.frames[machineCur]->getOutput().ids;
    for (size_t j = 0; j < seqIds_.size(); ++j) {
      finalPaths[seqIds_[j]].ids.push_back(idVec->getElement(j));
      finalPaths[seqIds_[j]].machineIdVec.push_back(j);
    }

    copyDataOutlinkFrame(machineCur);

    // call value printer
    if (evaluator_) {
      evaluator_->eval(*(frames_[machineCur].get()));
    }
    // check eos
    const IVectorPtr& eosVec =
        eosFrameLine_->layers[machineCur]->getOutput().ids;
    scatterIds.clear();
    for (size_t j = 0; j < seqIds_.size(); ++j) {
      if (eosVec->getElement(j) == 1U) {
        // path.seqId = -1 indicates end of generation
        // of an input sequence
        finalPaths[seqIds_[j]].seqId = -1;
926 927 928
      } else {
        scatterIds.push_back(j);
      }
Z
zhangjinchao01 已提交
929 930 931 932 933 934 935 936
    }
  }

  batchMachineIdVec_.clear();
  int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false);
  starts[0] = 0;
  generator_.ids.clear();
  for (size_t i = 0; i < batchSize; ++i) {
937
    generator_.ids.insert(generator_.ids.end(), finalPaths[i].ids.begin(),
Z
zhangjinchao01 已提交
938 939 940
                          finalPaths[i].ids.end());
    starts[i + 1] = generator_.ids.size();
    batchMachineIdVec_.insert(batchMachineIdVec_.end(),
941 942
                              finalPaths[i].machineIdVec.begin(),
                              finalPaths[i].machineIdVec.end());
Z
zhangjinchao01 已提交
943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999
  }
}

void RecurrentGradientMachine::connectPrevFrame(int stepId,
                                                std::vector<Path>& paths) {
  int machineCur = stepId % 2;
  int machinePrev = (stepId - 1) % 2;
  int beam = getBeamSize();
  machineIds_.clear();
  topIds_.clear();
  seqIds_.clear();

  for (size_t j = 0; j < paths.size(); ++j) {
    machineIds_.push_back(paths[j].machineId);
    topIds_.push_back(paths[j].machineId * beam + paths[j].topIndex);
    seqIds_.push_back(paths[j].seqId);
  }

  for (auto& memoryFrameLine : memoryFrameLines_) {
    bool isOutIds = (memoryFrameLine.layerName == outFrameLines_[0].layerName);
    auto scatterAgent = dynamic_cast<ScatterAgentLayer*>(
        memoryFrameLine.scatterAgents[machineCur].get());
    scatterAgent->setRealLayer(memoryFrameLine.frames[machinePrev],
                               isOutIds ? topIds_ : machineIds_,
                               memoryFrameLine.is_sequence);
    scatterAgent->forward(PASS_TEST);
    NeuralNetwork::connect(memoryFrameLine.agents[machineCur],
                           memoryFrameLine.scatterAgents[machineCur]);
  }
}

void RecurrentGradientMachine::forwardFrame(int machineCur) {
  // forward
  const std::vector<Argument> inArgs;
  std::vector<Argument> outArgs;
  frames_[machineCur]->forward(inArgs, &outArgs, PASS_TEST);

  copyDataOutlinkFrame(machineCur);

  IVectorPtr& ids = outFrameLines_[0].frames[machineCur]->getOutput().ids;
  MatrixPtr in = outFrameLines_[0].frames[machineCur]->getOutput().in;
  IVectorPtr& eos = eosFrameLine_->layers[machineCur]->getOutput().ids;
  if (useGpu_) {
    IVector::resizeOrCreate(cpuId_, ids->getSize(), false /* useGpu */);
    cpuId_->copyFrom(*ids);
    Matrix::resizeOrCreate(cpuProb_, in->getHeight(), in->getWidth(),
                           false /* trans */, false /* useGpu */);
    cpuProb_->copyFrom(*in);
    IVector::resizeOrCreate(cpuEos_, eos->getSize(), false /* useGpu */);
    cpuEos_->copyFrom(*eos);
  } else {
    cpuId_ = ids;
    cpuProb_ = in;
    cpuEos_ = eos;
  }
}

1000 1001 1002
void RecurrentGradientMachine::singlePathExpand(Path& curPath, size_t curPathId,
                                                std::vector<Path>& newPaths,
                                                size_t expandWidth) {
Z
zhangjinchao01 已提交
1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025
  int calc_id =
      gDiyProbStart ? gDiyProbStart(curPath.ids.size(), curPath.ids.data()) : 0;

  const int* idVec = cpuId_->getData();
  const real* probMat = cpuProb_->getData();
  const int* eosVec = cpuEos_->getData();

  for (size_t k = 0; k < expandWidth; k++) {
    int index = curPathId * expandWidth + k;
    int id = idVec[index];
    real prob = probMat[index];
    /*
     * Ordinarily, beam search greedily expands the most promising expandWidth
     * paths that currently are ALWAYS returned by MaxIdLayer.
     * In one condition, if user customizes the beam search procedure by
     * restricting the expansion within a user defined subset,
     * as a result, MaxIdLayer possibly COULD NOT return expandWidth
     * vaild expansions, and it will use -1 to indicate the end of valid
     * expansion candidates.
     */
    if (id == -1) break;

    real newLogProb = generator_.config.log_prob() ? std::log(prob) : prob;
1026 1027
    Path newPath(curPath, id, newLogProb, curPathId /*machineId*/,
                 k /*topIndex*/);
Z
zhangjinchao01 已提交
1028 1029
    if (this->beamSearchCtrlCallbacks_) {
      if (beamSearchCtrlCallbacks_->stopDetermineCandidates(
1030 1031
              newPath.seqId, newPath.ids, newPath.probHistory))
        return;
Z
zhangjinchao01 已提交
1032 1033 1034 1035 1036 1037
    }
    // outFrameLines_.size() > 1UL
    if (dataArgsSize_) {
      newPath.machineIdVec = curPath.machineIdVec;
      newPath.machineIdVec.push_back(curPathId);
    }
1038 1039
    bool atEos =
        eosVec[index] == 1U || newPath.ids.size() >= (size_t)maxSequenceLength_;
Z
zhangjinchao01 已提交
1040 1041 1042 1043 1044 1045 1046
    // adjustNewPath
    newPath.adjustProb(calc_id, atEos);
    if (this->beamSearchCtrlCallbacks_) {
      this->beamSearchCtrlCallbacks_->normOrDropNode(
          newPath.seqId, newPath.ids, newPath.probHistory, &newPath.logProb);
    }
    if (!newPath.isDropable()) {
1047 1048
      atEos ? finalPaths_[curPath.seqId].push_back(newPath)
            : newPaths.push_back(newPath);
Z
zhangjinchao01 已提交
1049 1050 1051
    }
  }  // for expandWidth

1052 1053 1054
  if (gDiyProbStop) {
    gDiyProbStop(calc_id);
  }
Z
zhangjinchao01 已提交
1055 1056
}

1057 1058
void RecurrentGradientMachine::beamExpand(std::vector<Path>& paths,
                                          std::vector<Path>& newPaths) {
Z
zhangjinchao01 已提交
1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070
  size_t candidatePathCount = paths.size();
  // idVec.size() could be larger than candidatePathCount * beam,
  // so user can drop some node customly.
  CHECK_EQ(cpuId_->getSize() % candidatePathCount, 0UL);
  size_t expandWidth = cpuId_->getSize() / candidatePathCount;

  // iterate over each sequence
  size_t totalExpandCount = 0;
  int prevSeqId = -1;
  int curSeqId = 0;
  for (size_t j = 0; j <= candidatePathCount; j++) {
    // expansions of a single sequence are all processed
1071
    curSeqId = (j < candidatePathCount ? paths[j].seqId : curSeqId + 1);
Z
zhangjinchao01 已提交
1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082
    if (prevSeqId != -1 && curSeqId != prevSeqId) {
      totalExpandCount += beamShrink(newPaths, prevSeqId, totalExpandCount);
    }
    if (j == candidatePathCount) return;
    singlePathExpand(paths[j], j, newPaths, expandWidth);

    prevSeqId = paths[j].seqId;
  }  // for paths
}

// Drop extra nodes to beam size.
1083 1084 1085 1086 1087 1088 1089 1090
size_t RecurrentGradientMachine::beamShrink(std::vector<Path>& newPaths,
                                            size_t seqId,
                                            size_t totalExpandCount) {
  size_t minNewPathSize =
      std::min(getBeamSize(), newPaths.size() - totalExpandCount);
  if (!minNewPathSize) {
    return 0;
  }
Z
zhangjinchao01 已提交
1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102
  std::nth_element(newPaths.begin() + totalExpandCount,
                   newPaths.begin() + totalExpandCount + minNewPathSize,
                   newPaths.end(), Path::greaterPath);
  newPaths.resize(totalExpandCount + minNewPathSize);

  real minPathLogProb = std::min_element(newPaths.end() - minNewPathSize,
                                         newPaths.end())->logProb;
  real maxPathLogProb = std::max_element(newPaths.end() - minNewPathSize,
                                         newPaths.end())->logProb;

  // Remove the already formed paths that are relatively short
  finalPaths_[seqId].erase(
1103 1104
      std::remove_if(finalPaths_[seqId].begin(), finalPaths_[seqId].end(),
                     [&](Path& p) { return p.logProb < minPathLogProb; }),
Z
zhangjinchao01 已提交
1105 1106 1107 1108 1109 1110 1111 1112
      finalPaths_[seqId].end());
  for (auto p : finalPaths_[seqId]) {
    if (minFinalPathLogProb_[seqId] > p.logProb) {
      minFinalPathLogProb_[seqId] = p.logProb;
    }
  }

  if (finalPaths_[seqId].size() >= getBeamSize() &&
1113
      minFinalPathLogProb_[seqId] >= maxPathLogProb) {
Z
zhangjinchao01 已提交
1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149
    newPaths.resize(totalExpandCount);
    return 0;
  }
  return minNewPathSize;
}

void RecurrentGradientMachine::fillGenOutputs() {
  size_t numResults = generator_.config.num_results_per_sample();
  for (size_t i = 0; i < finalPaths_.size(); ++i) {
    size_t minFinalPathsSize = std::min(numResults, finalPaths_[i].size());
    std::partial_sort(finalPaths_[i].begin(),
                      finalPaths_[i].begin() + minFinalPathsSize,
                      finalPaths_[i].end(), Path::greaterPath);
    finalPaths_[i].resize(minFinalPathsSize);
  }

  batchMachineIdVec_.clear();
  generator_.ids.clear();
  if (numResults > 1) {
    real* probs = generator_.outArg.in->getData();
    int* starts =
        generator_.outArg.sequenceStartPositions->getMutableData(false);
    starts[0] = 0;
    for (size_t i = 0; i < finalPaths_.size(); ++i) {
      for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
        Path& path = finalPaths_[i][j];
        generator_.ids.push_back(path.ids.size());  // sequence size
        generator_.ids.insert(generator_.ids.end(), path.ids.begin(),
                              path.ids.end());
        generator_.ids.push_back(-1);  // end of sequence
        probs[i * numResults + j] = path.logProb;

        if (!j && dataArgsSize_) {
          // in beam search, here only reserved the top 1 generated result
          // for out_links that are not the generated word indices.
          batchMachineIdVec_.insert(batchMachineIdVec_.end(),
1150 1151
                                    path.machineIdVec.begin(),
                                    path.machineIdVec.end());
Z
zhangjinchao01 已提交
1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174
        }
      }
      starts[i + 1] = generator_.ids.size();
    }
  } else {
    for (size_t i = 0; i < finalPaths_.size(); ++i) {
      CHECK(!finalPaths_[i].empty());
      generator_.ids = finalPaths_[i][0].ids;
    }
  }
}

void RecurrentGradientMachine::copyDataOutlinkFrame(size_t machineCur) {
  for (size_t i = 0; i < dataArgsSize_; i++) {
    Argument outFrame;
    outFrame.resizeAndCopyFrom(
        outFrameLines_[i + 1].frames[machineCur]->getOutput(), useGpu_);
    dataArgsFrame_[i].emplace_back(outFrame);
  }
}

void RecurrentGradientMachine::createDataOutlink(
    std::vector<int>& machineIdVec) {
1175 1176
  size_t seqNum =
      getBeamSize() > 1UL ? finalPaths_.size() : finalPaths_[0].size();
Z
zhangjinchao01 已提交
1177 1178
  std::vector<int> starts(seqNum + 1, 0);
  for (size_t i = 0; i < seqNum; ++i) {
1179 1180
    size_t seqLen = getBeamSize() > 1UL ? finalPaths_[i][0].ids.size()
                                        : finalPaths_[0][i].ids.size();
Z
zhangjinchao01 已提交
1181 1182 1183 1184
    starts[i + 1] = starts[i] + seqLen;
  }

  for (size_t i = 0; i < dataArgsSize_; i++) {
1185 1186
    dataArgs_[i].concat(dataArgsFrame_[i], machineIdVec, starts, useGpu_,
                        HPPL_STREAM_1, PASS_TEST);
Z
zhangjinchao01 已提交
1187

1188 1189
    auto dataAgent =
        dynamic_cast<DataLayer*>(outFrameLines_[i + 1].agentLayer.get());
Z
zhangjinchao01 已提交
1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255
    CHECK_NOTNULL(dataAgent);
    dataAgent->setData(dataArgs_[i]);
  }
}

void RecurrentGradientMachine::beamSearch(size_t batchSize) {
  finalPaths_.clear();
  finalPaths_.resize(batchSize);
  seqIds_.resize(batchSize);
  minFinalPathLogProb_.clear();
  minFinalPathLogProb_.resize(batchSize, 0);

  std::vector<Path> paths;
  std::vector<Path> newPaths;
  for (size_t i = 0; i < batchSize; ++i) {
    paths.push_back(Path(i));
    if (this->beamSearchCtrlCallbacks_) {
      paths.back().recordHistory();
    }
  }

  // restart beam search
  stopBeamSearch_ = false;
  for (int i = 0; i < maxSequenceLength_; ++i) {
    int machineCur = i % 2;
    std::unique_ptr<
        ScopedCallbacks<const RecurrentGradientMachine::EachStepCallback&, int>>
        statisticsBlock;
    if (this->beamSearchStatistics_) {
      auto ptr =
          new ScopedCallbacks<const RecurrentGradientMachine::EachStepCallback&,
                              int>(beamSearchStatistics_->onEachStepStarted,
                                   beamSearchStatistics_->onEachStepStoped, i);
      statisticsBlock.reset(ptr);
    }
    if (stopBeamSearch_) break;

    if (i) connectPrevFrame(i, paths);

    if (this->beamSearchCtrlCallbacks_) {
      std::vector<std::vector<int>*> prefixes;
      prefixes.resize(paths.size());
      std::transform(
          paths.begin(), paths.end(), prefixes.begin(),
          [](const Path& p) { return const_cast<std::vector<int>*>(&p.ids); });
      beamSearchCtrlCallbacks_->beamSearchCandidateAdjust(
          prefixes, frames_[machineCur].get(), i);
    }

    forwardFrame(machineCur);
    beamExpand(paths, newPaths);
    if (newPaths.empty()) break;

    paths = newPaths;
    newPaths.clear();
  }  // end for machineCur
  fillGenOutputs();
}

void RecurrentGradientMachine::Path::adjustProb(int calc_id, bool atEos) {
  if (gDiyProbMethod) {
    logProb = gDiyProbMethod(calc_id, ids.size(), ids.data(), logProb, atEos);
  }
}

}  // namespace paddle