summaryrefslogtreecommitdiff
path: root/training/mpi_online_optimize.cc
blob: 325ba030e6bb353d0d932d29db354ba9caa69efe (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
#include <sstream>
#include <iostream>
#include <fstream>
#include <vector>
#include <cassert>
#include <cmath>
#include <tr1/memory>

#include <boost/program_options.hpp>
#include <boost/program_options/variables_map.hpp>

#include "verbose.h"
#include "hg.h"
#include "prob.h"
#include "inside_outside.h"
#include "ff_register.h"
#include "decoder.h"
#include "filelib.h"
#include "online_optimizer.h"
#include "fdict.h"
#include "weights.h"
#include "sparse_vector.h"
#include "sampler.h"

#ifdef HAVE_MPI
#include <boost/mpi/timer.hpp>
#include <boost/mpi.hpp>
namespace mpi = boost::mpi;
#endif

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

void SanityCheck(const vector<double>& w) {
  for (int i = 0; i < w.size(); ++i) {
    assert(!isnan(w[i]));
    assert(!isinf(w[i]));
  }
}

struct FComp {
  const vector<double>& w_;
  FComp(const vector<double>& w) : w_(w) {}
  bool operator()(int a, int b) const {
    return fabs(w_[a]) > fabs(w_[b]);
  }
};

void ShowLargestFeatures(const vector<double>& w) {
  vector<int> fnums(w.size());
  for (int i = 0; i < w.size(); ++i)
    fnums[i] = i;
  vector<int>::iterator mid = fnums.begin();
  mid += (w.size() > 10 ? 10 : w.size());
  partial_sort(fnums.begin(), mid, fnums.end(), FComp(w));
  cerr << "TOP FEATURES:";
  for (vector<int>::iterator i = fnums.begin(); i != mid; ++i) {
    cerr << ' ' << FD::Convert(*i) << '=' << w[*i];
  }
  cerr << endl;
}

bool InitCommandLine(int argc, char** argv, po::variables_map* conf) {
  po::options_description opts("Configuration options");
  opts.add_options()
        ("input_weights,w",po::value<string>(),"Input feature weights file")
        ("training_data,t",po::value<string>(),"Training data corpus")
        ("training_agenda,a",po::value<string>(), "Text file listing a series of configuration files and the number of iterations to train using each configuration successively")
        ("minibatch_size_per_proc,s", po::value<unsigned>()->default_value(5), "Number of training instances evaluated per processor in each minibatch")
        ("optimization_method,m", po::value<string>()->default_value("sgd"), "Optimization method (sgd)")
        ("random_seed,S", po::value<uint32_t>(), "Random seed (if not specified, /dev/random will be used)")
        ("eta_0,e", po::value<double>()->default_value(0.2), "Initial learning rate for SGD (eta_0)")
        ("L1,1","Use L1 regularization")
        ("regularization_strength,C", po::value<double>()->default_value(1.0), "Regularization strength (C)");
  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 (conf->count("help") || !conf->count("training_data") || !conf->count("training_agenda")) {
    cerr << dcmdline_options << endl;
    return false;
  }
  return true;
}

void ReadTrainingCorpus(const string& fname, int rank, int size, vector<string>* c, vector<int>* order) {
  ReadFile rf(fname);
  istream& in = *rf.stream();
  string line;
  int id = 0;
  while(in) {
    getline(in, line);
    if (!in) break;
    if (id % size == rank) {
      c->push_back(line);
      order->push_back(id);
    }
    ++id;
  }
}

static const double kMINUS_EPSILON = -1e-6;

struct TrainingObserver : public DecoderObserver {
  void Reset() {
    acc_grad.clear();
    acc_obj = 0;
    total_complete = 0;
  } 

  void SetLocalGradientAndObjective(vector<double>* g, double* o) const {
    *o = acc_obj;
    for (SparseVector<prob_t>::const_iterator it = acc_grad.begin(); it != acc_grad.end(); ++it)
      (*g)[it->first] = it->second;
  }

  virtual void NotifyDecodingStart(const SentenceMetadata& smeta) {
    cur_model_exp.clear();
    cur_obj = 0;
    state = 1;
  }

  // compute model expectations, denominator of objective
  virtual void NotifyTranslationForest(const SentenceMetadata& smeta, Hypergraph* hg) {
    assert(state == 1);
    state = 2;
    const prob_t z = InsideOutside<prob_t,
                                   EdgeProb,
                                   SparseVector<prob_t>,
                                   EdgeFeaturesAndProbWeightFunction>(*hg, &cur_model_exp);
    cur_obj = log(z);
    cur_model_exp /= z;
  }

  // compute "empirical" expectations, numerator of objective
  virtual void NotifyAlignmentForest(const SentenceMetadata& smeta, Hypergraph* hg) {
    assert(state == 2);
    state = 3;
    SparseVector<prob_t> ref_exp;
    const prob_t ref_z = InsideOutside<prob_t,
                                       EdgeProb,
                                       SparseVector<prob_t>,
                                       EdgeFeaturesAndProbWeightFunction>(*hg, &ref_exp);
    ref_exp /= ref_z;

    double log_ref_z;
#if 0
    if (crf_uniform_empirical) {
      log_ref_z = ref_exp.dot(feature_weights);
    } else {
      log_ref_z = log(ref_z);
    }
#else
    log_ref_z = log(ref_z);
#endif

    // rounding errors means that <0 is too strict
    if ((cur_obj - log_ref_z) < kMINUS_EPSILON) {
      cerr << "DIFF. ERR! log_model_z < log_ref_z: " << cur_obj << " " << log_ref_z << endl;
      exit(1);
    }
    assert(!isnan(log_ref_z));
    ref_exp -= cur_model_exp;
    acc_grad += ref_exp;
    acc_obj += (cur_obj - log_ref_z);
  }

  virtual void NotifyDecodingComplete(const SentenceMetadata& smeta) {
    if (state == 3) {
      ++total_complete;
    } else {
    }
  }

  void GetGradient(SparseVector<double>* g) const {
    g->clear();
    for (SparseVector<prob_t>::const_iterator it = acc_grad.begin(); it != acc_grad.end(); ++it)
      g->set_value(it->first, it->second);
  }

  int total_complete;
  SparseVector<prob_t> cur_model_exp;
  SparseVector<prob_t> acc_grad;
  double acc_obj;
  double cur_obj;
  int state;
};

#ifdef HAVE_MPI
namespace boost { namespace mpi {
  template<>
  struct is_commutative<std::plus<SparseVector<double> >, SparseVector<double> > 
    : mpl::true_ { };
} } // end namespace boost::mpi
#endif

bool LoadAgenda(const string& file, vector<pair<string, int> >* a) {
  ReadFile rf(file);
  istream& in = *rf.stream();
  string line;
  while(in) {
    getline(in, line);
    if (!in) break;
    if (line.empty()) continue;
    if (line[0] == '#') continue;
    int sc = 0;
    if (line.size() < 3) return false;
    for (int i = 0; i < line.size(); ++i) { if (line[i] == ' ') ++sc; }
    if (sc != 1) { cerr << "Too many spaces in line: " << line << endl; return false; }
    size_t d = line.find(" ");
    pair<string, int> x;
    x.first = line.substr(0,d);
    x.second = atoi(line.substr(d+1).c_str());
    a->push_back(x);
    if (!FileExists(x.first)) {
      cerr << "Can't find file " << x.first << endl;
      return false;
    }
  }
  return true;
}

int main(int argc, char** argv) {
#ifdef HAVE_MPI
  mpi::environment env(argc, argv);
  mpi::communicator world;
  const int size = world.size(); 
  const int rank = world.rank();
#else
  const int size = 1;
  const int rank = 0;
#endif
  if (size > 1) SetSilent(true);  // turn off verbose decoder output
  register_feature_functions();
  std::tr1::shared_ptr<MT19937> rng;

  po::variables_map conf;
  if (!InitCommandLine(argc, argv, &conf))
    return 1;

  // load initial weights
  Weights weights;
  if (conf.count("input_weights"))
    weights.InitFromFile(conf["input_weights"].as<string>());

  vector<string> corpus;
  vector<int> ids;
  ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus, &ids);
  assert(corpus.size() > 0);

  std::tr1::shared_ptr<OnlineOptimizer> o;
  std::tr1::shared_ptr<LearningRateSchedule> lr;

  const unsigned size_per_proc = conf["minibatch_size_per_proc"].as<unsigned>();
  if (size_per_proc > corpus.size()) {
    cerr << "Minibatch size must be smaller than corpus size!\n";
    return 1;
  }

  size_t total_corpus_size = 0;
#ifdef HAVE_MPI
  reduce(world, corpus.size(), total_corpus_size, std::plus<size_t>(), 0);
#else
  total_corpus_size = corpus.size();
#endif

  if (rank == 0) {
    cerr << "Total corpus size: " << total_corpus_size << endl;
    const unsigned batch_size = size_per_proc * size;
    // TODO config
    lr.reset(new ExponentialDecayLearningRate(batch_size, conf["eta_0"].as<double>()));

    const string omethod = conf["optimization_method"].as<string>();
    if (omethod == "sgd") {
      const double C = conf["regularization_strength"].as<double>();
      o.reset(new CumulativeL1OnlineOptimizer(lr, total_corpus_size, C));
    } else {
      assert(!"fail");
    }
  }
  if (conf.count("random_seed"))
    rng.reset(new MT19937(conf["random_seed"].as<uint32_t>()));
  else
    rng.reset(new MT19937);

  SparseVector<double> x;
  weights.InitSparseVector(&x);
  TrainingObserver observer;

  int write_weights_every_ith = 100; // TODO configure
  int titer = -1;

  vector<pair<string, int> > agenda;
  if (!LoadAgenda(conf["training_agenda"].as<string>(), &agenda))
    return 1;
  if (rank == 0)
    cerr << "Loaded agenda defining " << agenda.size() << " training epochs\n";

  vector<double> lambdas;
  for (int ai = 0; ai < agenda.size(); ++ai) {
    const string& cur_config = agenda[ai].first;
    const unsigned max_iteration = agenda[ai].second;
    if (rank == 0)
      cerr << "STARTING TRAINING EPOCH " << (ai+1) << ". CONFIG=" << cur_config << endl;
    // load cdec.ini and set up decoder
    ReadFile ini_rf(cur_config);
    Decoder decoder(ini_rf.stream());

    if (rank == 0)
      o->ResetEpoch(); // resets the learning rate-- TODO is this good?

    int iter = -1;
    bool converged = false;
    while (!converged) {
#ifdef HAVE_MPI
      mpi::timer timer;
#endif
      weights.InitFromVector(x);
      weights.InitVector(&lambdas);
      ++iter; ++titer;
      observer.Reset();
      decoder.SetWeights(lambdas);
      if (rank == 0) {
        converged = (iter == max_iteration);
        SanityCheck(lambdas);
        ShowLargestFeatures(lambdas);
        string fname = "weights.cur.gz";
        if (iter % write_weights_every_ith == 0) {
          ostringstream o; o << "weights.epoch_" << (ai+1) << '.' << iter << ".gz";
          fname = o.str();
        }
        if (converged && ((ai+1)==agenda.size())) { fname = "weights.final.gz"; }
        ostringstream vv;
        vv << "total iter=" << titer << " (of current config iter=" << iter << ")  minibatch=" << size_per_proc << " sentences/proc x " << size << " procs.   num_feats=" << x.size() << '/' << FD::NumFeats() << "   passes_thru_data=" << (titer * size_per_proc / static_cast<double>(corpus.size())) << "   eta=" << lr->eta(titer);
        const string svv = vv.str();
        cerr << svv << endl;
        weights.WriteToFile(fname, true, &svv);
      }

      for (int i = 0; i < size_per_proc; ++i) {
        int ei = corpus.size() * rng->next();
        int id = ids[ei];
        decoder.SetId(id);
        decoder.Decode(corpus[ei], &observer);
      }
      SparseVector<double> local_grad, g;
      observer.GetGradient(&local_grad);
#ifdef HAVE_MPI
      reduce(world, local_grad, g, std::plus<SparseVector<double> >(), 0);
#else
      g.swap(local_grad);
#endif
      local_grad.clear();
      if (rank == 0) {
        g /= (size_per_proc * size);
        o->UpdateWeights(g, FD::NumFeats(), &x);
      }
#ifdef HAVE_MPI
      broadcast(world, x, 0);
      broadcast(world, converged, 0);
      world.barrier();
      if (rank == 0) { cerr << "  ELAPSED TIME THIS ITERATION=" << timer.elapsed() << endl; }
#endif
    }
  }
  return 0;
}