summaryrefslogtreecommitdiff
path: root/decoder/ff_lm.cc
blob: 21c05cf2d922e8de297f48f51a9780c2b353690a (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
//TODO: allow features to reorder by heuristic*weight the rules' terminal phrases (or of hyperedges').  if first pass has pruning, then compute over whole ruleset as part of heuristic

#include "ff_lm.h"

#include <sstream>
#include <unistd.h>
#include <sys/socket.h>
#include <sys/types.h>
#include <netinet/in.h>
#include <netdb.h>

#include <boost/shared_ptr.hpp>

#include "tdict.h"
#include "Vocab.h"
#include "Ngram.h"
#include "hg.h"
#include "stringlib.h"

#ifdef HAVE_RANDLM
// http://randlm.sourceforge.net/
#include "RandLM.h"
#endif

using namespace std;

namespace NgramCache {
  struct Cache {
    map<WordID, Cache> tree;
    float prob;
    Cache() : prob() {}
  };
  static Cache cache_;
  void Clear() { cache_.tree.clear(); }
}

struct LMClient {

  LMClient(const char* host) : port(6666) {
    strcpy(request_buffer, "prob ");
    s = const_cast<char*>(strchr(host, ':'));  // TODO fix const_cast
    if (s != NULL) {
      *s = '\0';
      ++s;
      port = atoi(s);
    }
    sock = socket(AF_INET, SOCK_STREAM, 0);
    hp = gethostbyname(host);
    if (hp == NULL) {
      cerr << "unknown host " << host << endl;
      abort();
    }
    bzero((char *)&server, sizeof(server));
    bcopy(hp->h_addr, (char *)&server.sin_addr, hp->h_length);
    server.sin_family = hp->h_addrtype;
    server.sin_port = htons(port);

    int errors = 0;
    while (connect(sock, (struct sockaddr *)&server, sizeof(server)) < 0) {
      cerr << "Error: connect()\n";
      sleep(1);
      errors++;
      if (errors > 3) exit(1);
    }
    cerr << "Connected to LM on " << host << " on port " << port << endl;
  }

  float wordProb(int word, int* context) {
    NgramCache::Cache* cur = &NgramCache::cache_;
    int i = 0;
    while (context[i] > 0) {
      cur = &cur->tree[context[i++]];
    }
    cur = &cur->tree[word];
    if (cur->prob) { return cur->prob; }

    i = 0;
    int pos = TD::AppendString(word, 5, 16000, request_buffer);
    while (context[i] > 0) {
      assert(pos < 15995);
      request_buffer[pos] = ' ';
      ++pos;
      pos = TD::AppendString(context[i], pos, 16000, request_buffer);
      ++i;
    }
    assert(pos < 15999);
    request_buffer[pos] = '\n';
    ++pos;
    request_buffer[pos] = 0;
    write(sock, request_buffer, pos);
    int r = read(sock, res, 6);
    int errors = 0;
    int cnt = 0;
    while (1) {
      if (r < 0) {
        errors++; sleep(1);
        cerr << "Error: read()\n";
        if (errors > 5) exit(1);
      } else if (r==0 || res[cnt] == '\n') { break; }
      else {
        cnt += r;
        if (cnt==6) break;
        read(sock, &res[cnt], 6-cnt);
      }
    }
    cur->prob = *reinterpret_cast<float*>(res);
    return cur->prob;
  }

 private:
  int sock, port;
  char *s;
  struct hostent *hp;
  struct sockaddr_in server;
  char res[8];
  char request_buffer[16000];
};

class LanguageModelImpl {
 public:
  explicit LanguageModelImpl(int order) :
      ngram_(*TD::dict_, order), buffer_(), order_(order), state_size_(OrderToStateSize(order) - 1),
      floor_(-100.0),
      client_(),
      kSTART(TD::Convert("<s>")),
      kSTOP(TD::Convert("</s>")),
      kUNKNOWN(TD::Convert("<unk>")),
      kNONE(-1),
      kSTAR(TD::Convert("<{STAR}>")) {}

  LanguageModelImpl(int order, const string& f) :
      ngram_(*TD::dict_, order), buffer_(), order_(order), state_size_(OrderToStateSize(order) - 1),
      floor_(-100.0),
      client_(NULL),
      kSTART(TD::Convert("<s>")),
      kSTOP(TD::Convert("</s>")),
      kUNKNOWN(TD::Convert("<unk>")),
      kNONE(-1),
      kSTAR(TD::Convert("<{STAR}>")) {
    if (f.find("lm://") == 0) {
      client_ = new LMClient(f.substr(5).c_str());
    } else {
      File file(f.c_str(), "r", 0);
      assert(file);
      cerr << "Reading " << order_ << "-gram LM from " << f << endl;
      ngram_.read(file, false);
    }
  }

  virtual ~LanguageModelImpl() {
    delete client_;
  }

  inline int StateSize(const void* state) const {
    return *(static_cast<const char*>(state) + state_size_);
  }

  inline void SetStateSize(int size, void* state) const {
    *(static_cast<char*>(state) + state_size_) = size;
  }

  virtual double WordProb(int word, int* context) {
    return client_ ?
          client_->wordProb(word, context)
        : ngram_.wordProb(word, (VocabIndex*)context);
  }

  inline double LookupProbForBufferContents(int i) {
//    int k = i; cerr << "P("; while(buffer_[k] > 0) { std::cerr << TD::Convert(buffer_[k++]) << " "; }
    double p = WordProb(buffer_[i], &buffer_[i+1]);
    if (p < floor_) p = floor_;
//    cerr << ")=" << p << endl;
    return p;
  }

  string DebugStateToString(const void* state) const {
    int len = StateSize(state);
    const int* astate = reinterpret_cast<const int*>(state);
    string res = "[";
    for (int i = 0; i < len; ++i) {
      res += " ";
      res += TD::Convert(astate[i]);
    }
    res += " ]";
    return res;
  }

  inline double ProbNoRemnant(int i, int len) {
    int edge = len;
    bool flag = true;
    double sum = 0.0;
    while (i >= 0) {
      if (buffer_[i] == kSTAR) {
        edge = i;
        flag = false;
      } else if (buffer_[i] <= 0) {
        edge = i;
        flag = true;
      } else {
        if ((edge-i >= order_) || (flag && !(i == (len-1) && buffer_[i] == kSTART)))
          sum += LookupProbForBufferContents(i);
      }
      --i;
    }
    return sum;
  }

  double EstimateProb(const vector<WordID>& phrase) {
    int len = phrase.size();
    buffer_.resize(len + 1);
    buffer_[len] = kNONE;
    int i = len - 1;
    for (int j = 0; j < len; ++j,--i)
      buffer_[i] = phrase[j];
    return ProbNoRemnant(len - 1, len);
  }

  double EstimateProb(const void* state) {
    int len = StateSize(state);
    // cerr << "residual len: " << len << endl;
    buffer_.resize(len + 1);
    buffer_[len] = kNONE;
    const int* astate = reinterpret_cast<const int*>(state);
    int i = len - 1;
    for (int j = 0; j < len; ++j,--i)
      buffer_[i] = astate[j];
    return ProbNoRemnant(len - 1, len);
  }

  double FinalTraversalCost(const void* state) {
    int slen = StateSize(state);
    int len = slen + 2;
    // cerr << "residual len: " << len << endl;
    buffer_.resize(len + 1);
    buffer_[len] = kNONE;
    buffer_[len-1] = kSTART;
    const int* astate = reinterpret_cast<const int*>(state);
    int i = len - 2;
    for (int j = 0; j < slen; ++j,--i)
      buffer_[i] = astate[j];
    buffer_[i] = kSTOP;
    assert(i == 0);
    return ProbNoRemnant(len - 1, len);
  }

  double LookupWords(const TRule& rule, const vector<const void*>& ant_states, void* vstate) {
    int len = rule.ELength() - rule.Arity();
    for (int i = 0; i < ant_states.size(); ++i)
      len += StateSize(ant_states[i]);
    buffer_.resize(len + 1);
    buffer_[len] = kNONE;
    int i = len - 1;
    const vector<WordID>& e = rule.e();
    for (int j = 0; j < e.size(); ++j) {
      if (e[j] < 1) {
        const int* astate = reinterpret_cast<const int*>(ant_states[-e[j]]);
        int slen = StateSize(astate);
        for (int k = 0; k < slen; ++k)
          buffer_[i--] = astate[k];
      } else {
        buffer_[i--] = e[j];
      }
    }

    double sum = 0.0;
    int* remnant = reinterpret_cast<int*>(vstate);
    int j = 0;
    i = len - 1;
    int edge = len;

    while (i >= 0) {
      if (buffer_[i] == kSTAR) {
        edge = i;
      } else if (edge-i >= order_) {
        sum += LookupProbForBufferContents(i);
      } else if (edge == len && remnant) {
        remnant[j++] = buffer_[i];
      }
      --i;
    }
    if (!remnant) return sum;

    if (edge != len || len >= order_) {
      remnant[j++] = kSTAR;
      if (order_-1 < edge) edge = order_-1;
      for (int i = edge-1; i >= 0; --i)
        remnant[j++] = buffer_[i];
    }

    SetStateSize(j, vstate);
    return sum;
  }

  static int OrderToStateSize(int order) {
    return ((order-1) * 2 + 1) * sizeof(WordID) + 1;
  }

 protected:
  Ngram ngram_;
  vector<WordID> buffer_;
  const int order_;
  const int state_size_;
  const double floor_;
 private:
  LMClient* client_;

 public:
  const WordID kSTART;
  const WordID kSTOP;
  const WordID kUNKNOWN;
  const WordID kNONE;
  const WordID kSTAR;
};

LanguageModel::LanguageModel(const string& param) :
    fid_(FD::Convert("LanguageModel")) {
  vector<string> argv;
  int argc = SplitOnWhitespace(param, &argv);
  int order = 3;
  // TODO add support for -n FeatureName
  string filename;
  if (argc < 1) { cerr << "LanguageModel requires a filename, minimally!\n"; abort(); }
  else if (argc == 1) { filename = argv[0]; }
  else if (argc == 2 || argc > 3) { cerr << "Don't understand 'LanguageModel " << param << "'\n"; }
  else if (argc == 3) {
    if (argv[0] == "-o") {
      order = atoi(argv[1].c_str());
      filename = argv[2];
    } else if (argv[1] == "-o") {
      order = atoi(argv[2].c_str());
      filename = argv[0];
    }
  }
  SetStateSize(LanguageModelImpl::OrderToStateSize(order));
  pimpl_ = new LanguageModelImpl(order, filename);
}

LanguageModel::~LanguageModel() {
  delete pimpl_;
}

string LanguageModel::DebugStateToString(const void* state) const{
  return pimpl_->DebugStateToString(state);
}

void LanguageModel::TraversalFeaturesImpl(const SentenceMetadata& smeta,
                                          const Hypergraph::Edge& edge,
                                          const vector<const void*>& ant_states,
                                          SparseVector<double>* features,
                                          SparseVector<double>* estimated_features,
                                          void* state) const {
  (void) smeta;
  features->set_value(fid_, pimpl_->LookupWords(*edge.rule_, ant_states, state));
  estimated_features->set_value(fid_, pimpl_->EstimateProb(state));
}

void LanguageModel::FinalTraversalFeatures(const void* ant_state,
                                           SparseVector<double>* features) const {
  features->set_value(fid_, pimpl_->FinalTraversalCost(ant_state));
}

#ifdef HAVE_RANDLM
struct RandLMImpl : public LanguageModelImpl {
  RandLMImpl(int order, randlm::RandLM* rlm) :
      LanguageModelImpl(order),
      rlm_(rlm),
      oov_(rlm->getWordID(rlm->getOOV())),
      rb_(1000, oov_) {
    map<int, randlm::WordID> map_cdec2randlm;
    int max_wordid = 0;
    for(map<randlm::Word, randlm::WordID>::const_iterator it = rlm->vocabStart();
        it != rlm->vocabEnd(); ++it) {
      const int cur = TD::Convert(it->first);
      map_cdec2randlm[TD::Convert(it->first)] = it->second;
      if (cur > max_wordid) max_wordid = cur;
    }
    cdec2randlm_.resize(max_wordid + 1, oov_);
    for (map<int, randlm::WordID>::iterator it = map_cdec2randlm.begin();
         it != map_cdec2randlm.end(); ++it)
      cdec2randlm_[it->first] = it->second;
    map_cdec2randlm.clear();
  }

  inline randlm::WordID Convert2RandLM(int w) {
    return (w < cdec2randlm_.size() ? cdec2randlm_[w] : oov_);
  }

  virtual double WordProb(int word, int* context) {
    int i = order_;
    int c = 1;
    rb_[i] = Convert2RandLM(word);
    while (i > 1 && *context > 0) {
      --i;
      rb_[i] = Convert2RandLM(*context);
      ++context;
      ++c;
    }
    const void* finalState = 0;
    int found;
    //cerr << "I = " << i << endl;
    return rlm_->getProb(&rb_[i], c, &found, &finalState);
  }
 private:
  boost::shared_ptr<randlm::RandLM> rlm_;
  randlm::WordID oov_;
  vector<randlm::WordID> cdec2randlm_;
  vector<randlm::WordID> rb_;
};

LanguageModelRandLM::LanguageModelRandLM(const string& param) :
    fid_(FD::Convert("RandLM")) {
  vector<string> argv;
  int argc = SplitOnWhitespace(param, &argv);
  int order = 3;
  // TODO add support for -n FeatureName
  string filename;
  if (argc < 1) { cerr << "RandLM requires a filename, minimally!\n"; abort(); }
  else if (argc == 1) { filename = argv[0]; }
  else if (argc == 2 || argc > 3) { cerr << "Don't understand 'RandLM " << param << "'\n"; }
  else if (argc == 3) {
    if (argv[0] == "-o") {
      order = atoi(argv[1].c_str());
      filename = argv[2];
    } else if (argv[1] == "-o") {
      order = atoi(argv[2].c_str());
      filename = argv[0];
    }
  }
  SetStateSize(LanguageModelImpl::OrderToStateSize(order));
  int cache_MB = 200; // increase cache size
  randlm::RandLM* rlm = randlm::RandLM::initRandLM(filename, order, cache_MB);
  assert(rlm != NULL);
  pimpl_ = new RandLMImpl(order, rlm);
}

LanguageModelRandLM::~LanguageModelRandLM() {
  delete pimpl_;
}

void LanguageModelRandLM::TraversalFeaturesImpl(const SentenceMetadata& smeta,
                                          const Hypergraph::Edge& edge,
                                          const vector<const void*>& ant_states,
                                          SparseVector<double>* features,
                                          SparseVector<double>* estimated_features,
                                          void* state) const {
  (void) smeta;
  features->set_value(fid_, pimpl_->LookupWords(*edge.rule_, ant_states, state));
  estimated_features->set_value(fid_, pimpl_->EstimateProb(state));
}

void LanguageModelRandLM::FinalTraversalFeatures(const void* ant_state,
                                           SparseVector<double>* features) const {
  features->set_value(fid_, pimpl_->FinalTraversalCost(ant_state));
}

#endif