summaryrefslogtreecommitdiff
path: root/training/lbl_model.cc
blob: def5075ab7318c87dfd08d52233869b62cf37c03 (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
#include <iostream>

#include "config.h"
#ifndef HAVE_EIGEN
  int main() { std::cerr << "Please rebuild with --with-eigen PATH\n"; return 1; }
#else

#include <cstdlib>
#include <algorithm>
#include <cmath>
#include <set>
#include <cstring> // memset
#include <ctime>

#ifdef HAVE_MPI
#include <boost/mpi/timer.hpp>
#include <boost/mpi.hpp>
namespace mpi = boost::mpi;
#endif
#include <boost/math/special_functions/fpclassify.hpp>
#include <boost/program_options.hpp>
#include <boost/program_options/variables_map.hpp>
#include <Eigen/Dense>

#include "corpus_tools.h"
#include "optimize.h"
#include "array2d.h"
#include "m.h"
#include "lattice.h"
#include "stringlib.h"
#include "filelib.h"
#include "tdict.h"

namespace po = boost::program_options;
using namespace std;

#define kDIMENSIONS 100
typedef Eigen::Matrix<double, kDIMENSIONS, 1> RVector;
typedef Eigen::Matrix<double, 1, kDIMENSIONS> RTVector;
typedef Eigen::Matrix<double, kDIMENSIONS, kDIMENSIONS> TMatrix;
vector<RVector> r_src, r_trg;

bool InitCommandLine(int argc, char** argv, po::variables_map* conf) {
  po::options_description opts("Configuration options");
  opts.add_options()
        ("input,i",po::value<string>(),"Input file")
        ("iterations,I",po::value<unsigned>()->default_value(1000),"Number of iterations of training")
        ("regularization_strength,C",po::value<double>()->default_value(0.1),"L2 regularization strength (0 for no regularization)")
        ("eta", po::value<double>()->default_value(0.1f), "Eta for SGD")
        ("source_embeddings,f", po::value<string>(), "File containing source embeddings (if unset, random vectors will be used)")
        ("target_embeddings,e", po::value<string>(), "File containing target embeddings (if unset, random vectors will be used)")
        ("random_seed,s", po::value<unsigned>(), "Random seed")
        ("diagonal_tension,T", po::value<double>()->default_value(4.0), "How sharp or flat around the diagonal is the alignment distribution (0 = uniform, >0 sharpens)")
        ("testset,x", po::value<string>(), "After training completes, compute the log likelihood of this set of sentence pairs under the learned model");
  po::options_description clo("Command line options");
  clo.add_options()
        ("config", po::value<string>(), "Configuration file")
        ("help,h", "Print this help message and exit");
  po::options_description dconfig_options, dcmdline_options;
  dconfig_options.add(opts);
  dcmdline_options.add(opts).add(clo);
  
  po::store(parse_command_line(argc, argv, dcmdline_options), *conf);
  if (conf->count("config")) {
    ifstream config((*conf)["config"].as<string>().c_str());
    po::store(po::parse_config_file(config, dconfig_options), *conf);
  }
  po::notify(*conf);

  if (argc < 2 || conf->count("help")) {
    cerr << "Usage " << argv[0] << " [OPTIONS] -i corpus.fr-en\n";
    cerr << dcmdline_options << endl;
    return false;
  }
  return true;
}

void Normalize(RVector* v) {
  double norm = v->norm();
  assert(norm > 0.0f);
  *v /= norm;
}

void Flatten(const TMatrix& m, vector<double>* v) {
  unsigned c = 0;
  v->resize(kDIMENSIONS * kDIMENSIONS);
  for (unsigned i = 0; i < kDIMENSIONS; ++i)
    for (unsigned j = 0; j < kDIMENSIONS; ++j) {
      assert(boost::math::isfinite(m(i, j)));
      (*v)[c++] = m(i,j);
    }
}

void Unflatten(const vector<double>& v, TMatrix* m) {
  unsigned c = 0;
  for (unsigned i = 0; i < kDIMENSIONS; ++i)
    for (unsigned j = 0; j < kDIMENSIONS; ++j) {
      assert(boost::math::isfinite(v[c]));
      (*m)(i, j) = v[c++];
    }
}

double ApplyRegularization(const double C,
                           const vector<double>& weights,
                           vector<double>* g) {
  assert(weights.size() == g->size());
  double reg = 0;
  for (size_t i = 0; i < weights.size(); ++i) {
    const double& w_i = weights[i];
    double& g_i = (*g)[i];
    reg += C * w_i * w_i;
    g_i += 2 * C * w_i;
  }
  return reg;
}

void LoadEmbeddings(const string& filename, vector<RVector>* pv) {
  vector<RVector>& v = *pv;
  cerr << "Reading embeddings from " << filename << " ...\n";
  ReadFile rf(filename);
  istream& in = *rf.stream();
  string line;
  unsigned lc = 0;
  while(getline(in, line)) {
    ++lc;
    size_t cur = line.find(' ');
    if (cur == string::npos || cur == 0) {
      cerr << "Parse error reading line " << lc << ":\n" << line << endl;
      abort();
    }
    WordID w = TD::Convert(line.substr(0, cur));
    if (w >= v.size()) continue;
    RVector& curv = v[w];
    line[cur] = 0;
    size_t start = cur + 1;
    cur = start + 1;
    size_t c = 0;
    while(cur < line.size()) {
      if (line[cur] == ' ') {
        line[cur] = 0;
        curv[c++] = strtod(&line[start], NULL);
        start = cur + 1;
        cur = start;
        if (c == kDIMENSIONS) break;
      }
      ++cur;
    }
    if (c < kDIMENSIONS && cur != start) {
      if (cur < line.size()) line[cur] = 0;
      curv[c++] = strtod(&line[start], NULL);
    }
    if (c != kDIMENSIONS) {
      static bool first = true;
      if (first) {
        cerr << " read " << c << " dimensions from embedding file, but built with " << kDIMENSIONS << " (filling in with random values)\n";
        first = false;
      }
      for (; c < kDIMENSIONS; ++c) curv[c] = rand();
    }
    if (c == kDIMENSIONS && cur != line.size()) {
      static bool first = true;
      if (first) {
        cerr << " embedding file contains more dimensions than configured with, truncating.\n";
        first = false;
      }
    }
  }
}

int main(int argc, char** argv) {
#ifdef HAVE_MPI
  std::cerr << "**MPI enabled.\n";
  mpi::environment env(argc, argv);
  mpi::communicator world;
  const int size = world.size(); 
  const int rank = world.rank();
#else
  std::cerr << "**MPI disabled.\n";
  const int rank = 0;
  const int size = 1;
#endif
  po::variables_map conf;
  if (!InitCommandLine(argc, argv, &conf)) return 1;
  const string fname = conf["input"].as<string>();
  const double reg_strength = conf["regularization_strength"].as<double>();
  const bool has_l2 = reg_strength;
  assert(reg_strength >= 0.0f);
  const int ITERATIONS = conf["iterations"].as<unsigned>();
  const double eta = conf["eta"].as<double>();
  const double diagonal_tension = conf["diagonal_tension"].as<double>();
  bool SGD = false;
  if (diagonal_tension < 0.0) {
    cerr << "Invalid value for diagonal_tension: must be >= 0\n";
    return 1;
  }
  string testset;
  if (conf.count("testset")) testset = conf["testset"].as<string>();

  unsigned lc = 0;
  vector<double> unnormed_a_i;
  bool flag = false;
  vector<vector<WordID> > srcs, trgs;
  vector<WordID> vocab_e;
  {
    set<WordID> svocab_e, svocab_f;
    CorpusTools::ReadFromFile(fname, &srcs, NULL, &trgs, &svocab_e, rank, size);
    copy(svocab_e.begin(), svocab_e.end(), back_inserter(vocab_e));
  }
  cerr << "Number of target word types: " << vocab_e.size() << endl;
  const double num_examples = lc;

  boost::shared_ptr<LBFGSOptimizer> lbfgs;
  if (rank == 0)
    lbfgs.reset(new LBFGSOptimizer(kDIMENSIONS * kDIMENSIONS, 100));
  r_trg.resize(TD::NumWords() + 1);
  r_src.resize(TD::NumWords() + 1);
  vector<set<unsigned> > trg_pos(TD::NumWords() + 1);

  if (conf.count("random_seed")) {
    srand(conf["random_seed"].as<unsigned>());
  } else {
    unsigned seed = time(NULL) + rank * 100;
    cerr << "Random seed: " << seed << endl;
    srand(seed);
  }
  
  TMatrix t;
  if (rank == 0) {
    t = TMatrix::Random() / 50.0;
    for (unsigned i = 1; i < r_trg.size(); ++i) {
      r_trg[i] = RVector::Random();
      r_src[i] = RVector::Random();
    }
    if (conf.count("source_embeddings"))
      LoadEmbeddings(conf["source_embeddings"].as<string>(), &r_src);
    if (conf.count("target_embeddings"))
      LoadEmbeddings(conf["target_embeddings"].as<string>(), &r_trg);
  }

  // do optimization
  TMatrix g = TMatrix::Zero();
  vector<TMatrix> exp_src;
  vector<double> z_src;
  vector<double> flat_g, flat_t;
  Flatten(t, &flat_t);
  bool converged = false;
  // TODO broadcast embeddings
  for (int iter = 0; !converged && iter < ITERATIONS; ++iter) {
#ifdef HAVE_MPI
    mpi::broadcast(world, &flat_t[0], flat_t.size(), 0);
#endif
    Unflatten(flat_t, &t);
    cerr << "ITERATION " << (iter + 1) << endl;
    double likelihood = 0;
    double denom = 0.0;
    lc = 0;
    flag = false;
    g *= 0;
    for (unsigned i = 0; i < srcs.size(); ++i) {
      const vector<WordID>& src = srcs[i];
      const vector<WordID>& trg = trgs[i];
      ++lc;
      if (rank == 0 && lc % 1000 == 0) { cerr << '.'; flag = true; }
      if (rank == 0 && lc %50000 == 0) { cerr << " [" << lc << "]\n" << flush; flag = false; }
      denom += trg.size();

      exp_src.clear(); exp_src.resize(src.size(), TMatrix::Zero());
      z_src.clear(); z_src.resize(src.size(), 0.0);
      Array2D<TMatrix> exp_refs(src.size(), trg.size(), TMatrix::Zero());
      Array2D<double> z_refs(src.size(), trg.size(), 0.0);
      for (unsigned j = 0; j < trg.size(); ++j)
        trg_pos[trg[j]].insert(j);

      for (unsigned i = 0; i < src.size(); ++i) {
        const RVector& r_s = r_src[src[i]];
        const RTVector pred = r_s.transpose() * t;
        TMatrix& exp_m = exp_src[i];
        double& z = z_src[i];
        for (unsigned k = 0; k < vocab_e.size(); ++k) {
          const WordID v_k = vocab_e[k];
          const RVector& r_t = r_trg[v_k];
          const double dot_prod = pred * r_t;
          const double u = exp(dot_prod);
          z += u;
          const TMatrix v = r_s * r_t.transpose() * u;
          exp_m += v;
          set<unsigned>& ref_locs = trg_pos[v_k];
          if (!ref_locs.empty()) {
            for (set<unsigned>::iterator it = ref_locs.begin(); it != ref_locs.end(); ++it) {
              TMatrix& exp_ref_ij = exp_refs(i, *it);
              double& z_ref_ij = z_refs(i, *it);
              z_ref_ij += u;
              exp_ref_ij += v;
            }
          }
        }
      }
      for (unsigned j = 0; j < trg.size(); ++j)
        trg_pos[trg[j]].clear();

      // model expectations for a single target generation with
      // uniform alignment prior
      // TODO: when using a non-uniform alignment, m_exp will be
      // a function of j (below)
      double m_z = 0;
      TMatrix m_exp = TMatrix::Zero();
      for (unsigned i = 0; i < src.size(); ++i) {
        m_exp += exp_src[i];
        m_z += z_src[i];
      }
      m_exp /= m_z;

      Array2D<bool> al(src.size(), trg.size(), false);
      for (unsigned j = 0; j < trg.size(); ++j) {
        double ref_z = 0;
        TMatrix ref_exp = TMatrix::Zero();
        int max_i = 0;
        double max_s = -9999999;
        for (unsigned i = 0; i < src.size(); ++i) {
          ref_exp += exp_refs(i, j);
          ref_z += z_refs(i, j);
          if (log(z_refs(i, j)) > max_s) {
            max_s = log(z_refs(i, j));
            max_i = i;
          }
          // TODO handle alignment prob
        }
        if (ref_z <= 0) { 
          cerr << "TRG=" << TD::Convert(trg[j]) << endl;
          cerr << " LINE=" << lc << " (RANK=" << rank << "/" << size << ")" << endl;
          cerr << " REF_EXP=\n" << ref_exp << endl;
          cerr << " M_EXP=\n" << m_exp << endl;
          abort();
        }
        al(max_i, j) = true;
        ref_exp /= ref_z;
        g += m_exp - ref_exp;
        likelihood += log(ref_z) - log(m_z);
        if (SGD) {
          t -= g * eta / num_examples;
          g *= 0;
        }
      }
      
      if (rank == 0 && (iter == (ITERATIONS - 1) || lc < 12)) { cerr << al << endl; }
    }
    if (flag && rank == 0) { cerr << endl; }

    double obj = 0;
    if (!SGD) {
      Flatten(g, &flat_g);
      obj = -likelihood;
      // TODO - reduce gradient
    }

    if (rank == 0) {
      double gn = 0;
      for (unsigned i = 0; i < flat_g.size(); ++i)
        gn += flat_g[i]*flat_g[i];
      const double base2_likelihood = likelihood / log(2);
      cerr << "  log_e likelihood: " << likelihood << endl;
      cerr << "  log_2 likelihood: " << base2_likelihood << endl;
      cerr << "     cross entropy: " << (-base2_likelihood / denom) << endl;
      cerr << "        perplexity: " << pow(2.0, -base2_likelihood / denom) << endl;
      cerr << "     gradient norm: " << sqrt(gn) << endl;
      if (!SGD) {
        if (has_l2) {
          const double r = ApplyRegularization(reg_strength,
                                               flat_t,
                                               &flat_g);
          obj += r;
          cerr << "    regularization: " << r << endl;
        }
        lbfgs->Optimize(obj, flat_g, &flat_t);
        converged = (lbfgs->HasConverged());
      }
    }
#ifdef HAVE_MPI
    mpi::broadcast(world, converged, 0);
#endif
  }
  cerr << "TRANSLATION MATRIX:" << endl << t << endl;
  return 0;
}

#endif