mosesdecoder/moses/LM/IRST.cpp

// $Id$

/***********************************************************************
Moses - factored phrase-based language decoder
Copyright (C) 2006 University of Edinburgh

This library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.

This library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
Lesser General Public License for more details.

You should have received a copy of the GNU Lesser General Public
License along with this library; if not, write to the Free Software
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
***********************************************************************/

#include <limits>
#include <iostream>
#include <fstream>
#include "dictionary.h"
#include "n_gram.h"
#include "lmContainer.h"

using namespace irstlm;

#include "IRST.h"
#include "moses/LM/PointerState.h"
#include "moses/TypeDef.h"
#include "moses/Util.h"
#include "moses/FactorCollection.h"
#include "moses/Phrase.h"
#include "moses/InputFileStream.h"
#include "moses/StaticData.h"
#include "moses/TranslationTask.h"

using namespace std;

namespace Moses
{

class IRSTLMState : public PointerState
{
public:
  IRSTLMState():PointerState(NULL)  {}
  IRSTLMState(const void* lms):PointerState(lms)  {}
  IRSTLMState(const IRSTLMState& copy_from):PointerState(copy_from.lmstate)  {}

  IRSTLMState& operator=( const IRSTLMState& rhs ) {
    lmstate = rhs.lmstate;
    return *this;
  }

  const void* GetState() const {
    return lmstate;
  }
};

LanguageModelIRST::LanguageModelIRST(const std::string &line)
  :LanguageModelSingleFactor(line)
  ,m_lmtb_dub(0), m_lmtb_size(0)
{
  const StaticData &staticData = StaticData::Instance();
  int threadCount = staticData.ThreadCount();
  if (threadCount != 1) {
    throw runtime_error("Error: " + SPrint(threadCount) + " number of threads specified but IRST LM is not threadsafe.");
  }

  ReadParameters();

  VERBOSE(4, GetScoreProducerDescription() << " LanguageModelIRST::LanguageModelIRST() m_lmtb_dub:|" << m_lmtb_dub << "|" << std::endl);
  VERBOSE(4, GetScoreProducerDescription() << " LanguageModelIRST::LanguageModelIRST() m_filePath:|" << m_filePath << "|" << std::endl);
  VERBOSE(4, GetScoreProducerDescription() << " LanguageModelIRST::LanguageModelIRST() m_factorType:|" << m_factorType << "|" << std::endl);
  VERBOSE(4, GetScoreProducerDescription() << " LanguageModelIRST::LanguageModelIRST() m_lmtb_size:|" << m_lmtb_size << "|" << std::endl);
}

LanguageModelIRST::~LanguageModelIRST()
{

#ifndef WIN32
  TRACE_ERR( "reset mmap\n");
  if (m_lmtb) m_lmtb->reset_mmap();
#endif

  delete m_lmtb;
}


bool LanguageModelIRST::IsUseable(const FactorMask &mask) const
{
  bool ret = mask[m_factorType];
  return ret;
}

void LanguageModelIRST::Load(AllOptions::ptr const& opts)
{
  FactorCollection &factorCollection = FactorCollection::Instance();

  m_lmtb = m_lmtb->CreateLanguageModel(m_filePath);
  if (m_lmtb_size > 0) m_lmtb->setMaxLoadedLevel(m_lmtb_size);
  m_lmtb->load(m_filePath);
  d=m_lmtb->getDict();
  d->incflag(1);

  m_nGramOrder = m_lmtb_size = m_lmtb->maxlevel();

  // LM can be ok, just outputs warnings
  // Mauro: in the original, the following two instructions are wrongly switched:
  m_unknownId = d->oovcode(); // at the level of micro tags
  m_empty = -1; // code for an empty position

  CreateFactors(factorCollection);

  VERBOSE(1, GetScoreProducerDescription() << "  LanguageModelIRST::Load() m_unknownId=" << m_unknownId << std::endl);

  //install caches to save time (only if PS_CACHE_ENABLE is defined through compilation flags)
  m_lmtb->init_caches(m_lmtb_size>2?m_lmtb_size-1:2);

  if (m_lmtb_dub > 0) m_lmtb->setlogOOVpenalty(m_lmtb_dub);
}

void LanguageModelIRST::CreateFactors(FactorCollection &factorCollection)
{
  // add factors which have srilm id
  // code copied & paste from SRI LM class. should do template function
  std::map<size_t, int> lmIdMap;
  size_t maxFactorId = 0; // to create lookup vector later on
  m_empty = -1; // code for an empty position

  dict_entry *entry;
  dictionary_iter iter(d); // at the level of micro tags
  while ( (entry = iter.next()) != NULL) {
    size_t factorId = factorCollection.AddFactor(Output, m_factorType, entry->word)->GetId();
    lmIdMap[factorId] = entry->code;
    maxFactorId = (factorId > maxFactorId) ? factorId : maxFactorId;
  }

  size_t factorId;

  m_sentenceStart = factorCollection.AddFactor(Output, m_factorType, BOS_);
  factorId = m_sentenceStart->GetId();
  const std::string bs = BOS_;
  const std::string es = EOS_;
  m_lmtb_sentenceStart=lmIdMap[factorId] = GetLmID(BOS_);
  maxFactorId = (factorId > maxFactorId) ? factorId : maxFactorId;
  m_sentenceStartWord[m_factorType] = m_sentenceStart;

  m_sentenceEnd		= factorCollection.AddFactor(Output, m_factorType, EOS_);
  factorId = m_sentenceEnd->GetId();
  m_lmtb_sentenceEnd=lmIdMap[factorId] = GetLmID(EOS_);
  maxFactorId = (factorId > maxFactorId) ? factorId : maxFactorId;
  m_sentenceEndWord[m_factorType] = m_sentenceEnd;

  // add to lookup vector in object
  m_lmIdLookup.resize(maxFactorId+1);
  fill(m_lmIdLookup.begin(), m_lmIdLookup.end(), m_empty);

  map<size_t, int>::iterator iterMap;
  for (iterMap = lmIdMap.begin() ; iterMap != lmIdMap.end() ; ++iterMap) {
    m_lmIdLookup[iterMap->first] = iterMap->second;
  }
}

int LanguageModelIRST::GetLmID( const std::string &str ) const
{
  return d->encode( str.c_str() ); // at the level of micro tags
}

int LanguageModelIRST::GetLmID( const Word &word ) const
{
  return GetLmID( word.GetFactor(m_factorType) );
}

int LanguageModelIRST::GetLmID( const Factor *factor ) const
{
  size_t factorId = factor->GetId();

  if  ((factorId >= m_lmIdLookup.size()) || (m_lmIdLookup[factorId] == m_empty)) {
    if (d->incflag()==1) {
      std::string s = factor->GetString().as_string();
      int code = d->encode(s.c_str());

      //////////
      ///poiche' non c'e' distinzione tra i factorIDs delle parole sorgenti
      ///e delle parole target in Moses, puo' accadere che una parola target
      ///di cui non sia stato ancora calcolato il suo codice target abbia
      ///comunque un factorID noto (e quindi minore di m_lmIdLookup.size())
      ///E' necessario dunque identificare questi casi di indeterminatezza
      ///del codice target. Attualmente, questo controllo e' stato implementato
      ///impostando a    m_empty     tutti i termini che non hanno ancora
      //ricevuto un codice target effettivo
      ///////////

      ///OLD PROBLEM - SOLVED
////////////
/// IL PPROBLEMA ERA QUI
/// m_lmIdLookup.push_back(code);
/// PERCHE' USANDO PUSH_BACK IN REALTA' INSEREVIVAMO L'ELEMENTO NUOVO
/// IN POSIZIONE (factorID-1) invece che in posizione factrID dove dopo andiamo a leggerlo (vedi caso C
/// Cosi' funziona ....
/// ho un dubbio su cosa c'e' nelle prime posizioni di m_lmIdLookup
/// quindi
/// e scopro che rimane vuota una entry ogni due
/// perche' factorID cresce di due in due (perche' codifica sia source che target) "vuota" la posizione (factorID-1)
/// non da problemi di correttezza, ma solo di "spreco" di memoria
/// potremmo sostituirerendere  m_lmIdLookup una std:map invece che un std::vector,
/// ma si perde in efficienza nell'accesso perche' non e' piu' possibile quello random dei vettori
/// a te la scelta!!!!
////////////////


      if (factorId >= m_lmIdLookup.size()) {
        //resize and fill with m_empty
        //increment the array more than needed to avoid too many resizing operation.
        m_lmIdLookup.resize(factorId+10, m_empty);
      }

      //insert new code
      m_lmIdLookup[factorId] = code;
      return code;

    } else {
      return m_unknownId;
    }
  } else {
    return m_lmIdLookup[factorId];
  }
}

const FFState* LanguageModelIRST::EmptyHypothesisState(const InputType &/*input*/) const
{
  std::auto_ptr<IRSTLMState> ret(new IRSTLMState());

  return ret.release();
}

void LanguageModelIRST::CalcScore(const Phrase &phrase, float &fullScore, float &ngramScore, size_t &oovCount) const
{
  fullScore = 0;
  ngramScore = 0;
  oovCount = 0;

  if ( !phrase.GetSize() ) return;

  int _min = min(m_lmtb_size - 1, (int) phrase.GetSize());

  int codes[m_lmtb_size];
  int idx = 0;
  codes[idx] = m_lmtb_sentenceStart;
  ++idx;
  int position = 0;

  char* msp = NULL;
  float before_boundary = 0.0;
  for (; position < _min; ++position) {
    codes[idx] = GetLmID(phrase.GetWord(position));
    if (codes[idx] == m_unknownId) ++oovCount;
    before_boundary += m_lmtb->clprob(codes,idx+1,NULL,NULL,&msp);
    ++idx;
  }

  ngramScore = 0.0;
  int end_loop = (int) phrase.GetSize();

  for (; position < end_loop; ++position) {
    for (idx = 1; idx < m_lmtb_size; ++idx) {
      codes[idx-1] = codes[idx];
    }
    codes[idx-1] = GetLmID(phrase.GetWord(position));
    if (codes[idx-1] == m_unknownId) ++oovCount;
    ngramScore += m_lmtb->clprob(codes,idx,NULL,NULL,&msp);
  }
  before_boundary = TransformLMScore(before_boundary);
  ngramScore = TransformLMScore(ngramScore);
  fullScore = ngramScore + before_boundary;
}

FFState* LanguageModelIRST::EvaluateWhenApplied(const Hypothesis &hypo, const FFState *ps, ScoreComponentCollection *out) const
{
  if (!hypo.GetCurrTargetLength()) {
    std::auto_ptr<IRSTLMState> ret(new IRSTLMState(ps));
    return ret.release();
  }

  //[begin, end) in STL-like fashion.
  const int begin = (const int) hypo.GetCurrTargetWordsRange().GetStartPos();
  const int end = (const int) hypo.GetCurrTargetWordsRange().GetEndPos() + 1;
  const int adjust_end = (const int) std::min(end, begin + m_lmtb_size - 1);

  //set up context
  //fill the farthest positions with sentenceStart symbols, if "empty" positions are available
  //so that the vector looks like = "<s> <s> context_word context_word" for a two-word context and a LM of order 5
  int codes[m_lmtb_size];
  int idx=m_lmtb_size-1;
  int position = (const int) begin;
  while (position >= 0) {
    codes[idx] =  GetLmID(hypo.GetWord(position));
    --idx;
    --position;
  }
  while (idx>=0) {
    codes[idx] = m_lmtb_sentenceStart;
    --idx;
  }

  char* msp = NULL;
  float score = m_lmtb->clprob(codes,m_lmtb_size,NULL,NULL,&msp);

  position = (const int) begin+1;
  while (position < adjust_end) {
    for (idx=1; idx<m_lmtb_size; idx++) {
      codes[idx-1] = codes[idx];
    }
    codes[idx-1] =  GetLmID(hypo.GetWord(position));
    score += m_lmtb->clprob(codes,m_lmtb_size,NULL,NULL,&msp);
    ++position;
  }

  //adding probability of having sentenceEnd symbol, after this phrase;
  //this could happen only when all source words are covered
  if (hypo.IsSourceCompleted()) {
    idx=m_lmtb_size-1;
    codes[idx] = m_lmtb_sentenceEnd;
    --idx;
    position = (const int) end - 1;
    while (position >= 0 && idx >= 0) {
      codes[idx] =  GetLmID(hypo.GetWord(position));
      --idx;
      --position;
    }
    while (idx>=0) {
      codes[idx] = m_lmtb_sentenceStart;
      --idx;
    }
    score += m_lmtb->clprob(codes,m_lmtb_size,NULL,NULL,&msp);
  } else {
    // need to set the LM state

    if (adjust_end < end)   { //the LMstate of this target phrase refers to the last m_lmtb_size-1 words
      position = (const int) end - 1;
      for (idx=m_lmtb_size-1; idx>0; --idx) {
        codes[idx] =  GetLmID(hypo.GetWord(position));
      }
      codes[idx] = m_lmtb_sentenceStart;
      msp = (char *) m_lmtb->cmaxsuffptr(codes,m_lmtb_size);
    }
  }

  score = TransformLMScore(score);
  out->PlusEquals(this, score);

  std::auto_ptr<IRSTLMState> ret(new IRSTLMState(msp));

  return ret.release();
}

LMResult LanguageModelIRST::GetValue(const vector<const Word*> &contextFactor, State* finalState) const
{
  // set up context
  size_t count = contextFactor.size();
  if (count < 0) {
    cerr << "ERROR count < 0\n";
    exit(100);
  };

  // set up context
  int codes[MAX_NGRAM_SIZE];

  size_t idx=0;
  //fill the farthest positions with at most ONE sentenceEnd symbol and at most ONE sentenceEnd symbol, if "empty" positions are available
  //so that the vector looks like = "</s> <s> context_word context_word" for a two-word context and a LM of order 5
  if (count < (size_t) (m_lmtb_size-1)) codes[idx++] = m_lmtb_sentenceEnd;
  if (count < (size_t) m_lmtb_size) codes[idx++] = m_lmtb_sentenceStart;

  for (size_t i = 0 ; i < count ; i++) {
    codes[idx] =  GetLmID(*contextFactor[i]);
    ++idx;
  }

  LMResult result;
  result.unknown = (codes[idx - 1] == m_unknownId);

  char* msp = NULL;
  result.score = m_lmtb->clprob(codes,idx,NULL,NULL,&msp);

  if (finalState) *finalState=(State *) msp;

  result.score = TransformLMScore(result.score);

  return result;
}

bool LMCacheCleanup(const int sentences_done, const size_t m_lmcache_cleanup_threshold)
{
  if (sentences_done==-1) return true;
  if (m_lmcache_cleanup_threshold)
    if (sentences_done % m_lmcache_cleanup_threshold == 0)
      return true;
  return false;
}

void LanguageModelIRST::InitializeForInput(ttasksptr const& ttask)
{
  //nothing to do
#ifdef TRACE_CACHE
  m_lmtb->sentence_id++;
#endif
}

void LanguageModelIRST::CleanUpAfterSentenceProcessing(const InputType& source)
{
  const StaticData &staticData = StaticData::Instance();
  static int sentenceCount = 0;
  sentenceCount++;

  size_t lmcache_cleanup_threshold = staticData.GetLMCacheCleanupThreshold();

  if (LMCacheCleanup(sentenceCount, lmcache_cleanup_threshold)) {
    TRACE_ERR( "reset caches\n");
    m_lmtb->reset_caches();
  }
}

void LanguageModelIRST::SetParameter(const std::string& key, const std::string& value)
{
  if (key == "dub") {
    m_lmtb_dub = Scan<unsigned int>(value);
  } else {
    LanguageModelSingleFactor::SetParameter(key, value);
  }
  m_lmtb_size = m_nGramOrder;
}

}