summaryrefslogtreecommitdiff
path: root/rst_parser/mst_train.cc
blob: a78df600982249a0758a71a97de6ff7874f41451 (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
#include "arc_factored.h"

#include <vector>
#include <iostream>
#include <boost/program_options.hpp>
#include <boost/program_options/variables_map.hpp>
// #define HAVE_THREAD 1
#if HAVE_THREAD
#include <boost/thread.hpp>
#endif

#include "arc_ff.h"
#include "stringlib.h"
#include "filelib.h"
#include "tdict.h"
#include "dep_training.h"
#include "optimize.h"
#include "weights.h"

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

void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
  po::options_description opts("Configuration options");
  string cfg_file;
  opts.add_options()
        ("training_data,t",po::value<string>()->default_value("-"), "File containing training data (jsent format)")
        ("weights,w",po::value<string>(), "Optional starting weights")
        ("output_every_i_iterations,I",po::value<unsigned>()->default_value(1), "Write weights every I iterations")
        ("regularization_strength,C",po::value<double>()->default_value(1.0), "Regularization strength")
#ifdef HAVE_CMPH
        ("cmph_perfect_feature_hash,h", po::value<string>(), "Load perfect hash function for features")
#endif
#if HAVE_THREAD
        ("threads,T",po::value<unsigned>()->default_value(1), "Number of threads")
#endif
        ("correction_buffers,m", po::value<int>()->default_value(10), "LBFGS correction buffers");
  po::options_description clo("Command line options");
  clo.add_options()
        ("config,c", po::value<string>(&cfg_file), "Configuration file")
        ("help,?", "Print this help message and exit");

  po::options_description dconfig_options, dcmdline_options;
  dconfig_options.add(opts);
  dcmdline_options.add(dconfig_options).add(clo);
  po::store(parse_command_line(argc, argv, dcmdline_options), *conf);
  if (cfg_file.size() > 0) {
    ReadFile rf(cfg_file);
    po::store(po::parse_config_file(*rf.stream(), dconfig_options), *conf);
  }
  if (conf->count("help")) {
    cerr << dcmdline_options << endl;
    exit(1);
  }
}

void AddFeatures(double prob, const SparseVector<double>& fmap, vector<double>* g) {
  for (SparseVector<double>::const_iterator it = fmap.begin(); it != fmap.end(); ++it)
    (*g)[it->first] += it->second * prob;
}

double ApplyRegularizationTerms(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 prev_w_i = (i < prev_weights.size() ? prev_weights[i] : 0.0);
    const double& w_i = weights[i];
    double& g_i = (*g)[i];
    reg += C * w_i * w_i;
    g_i += 2 * C * w_i;

//    reg += T * (w_i - prev_w_i) * (w_i - prev_w_i);
//    g_i += 2 * T * (w_i - prev_w_i);
  }
  return reg;
}

struct GradientWorker {
  GradientWorker(int f,
                 int t,
                 vector<double>* w,
                 vector<TrainingInstance>* c,
                 vector<ArcFactoredForest>* fs) : obj(), weights(*w), from(f), to(t), corpus(*c), forests(*fs), g(w->size()) {}
  void operator()() {
    int every = (to - from) / 20;
    if (!every) every++;
    for (int i = from; i < to; ++i) {
      if ((from == 0) && (i + 1) % every == 0) cerr << '.' << flush;
      const int num_words = corpus[i].ts.words.size();
      forests[i].Reweight(weights);
      prob_t z;
      forests[i].EdgeMarginals(&z);
      obj -= log(z);
      //cerr << " O = " << (-corpus[i].features.dot(weights)) << " D=" << -lz << "  OO= " << (-corpus[i].features.dot(weights) - lz) << endl;
      //cerr << " ZZ = " << zz << endl;
      for (int h = -1; h < num_words; ++h) {
        for (int m = 0; m < num_words; ++m) {
          if (h == m) continue;
          const ArcFactoredForest::Edge& edge = forests[i](h,m);
          const SparseVector<weight_t>& fmap = edge.features;
          double prob = edge.edge_prob.as_float();
          if (prob < -0.000001) { cerr << "Prob < 0: " << prob << endl; prob = 0; }
          if (prob > 1.000001) { cerr << "Prob > 1: " << prob << endl; prob = 1; }
          AddFeatures(prob, fmap, &g);
          //mfm += fmap * prob;  // DE
        }
      }
    }
  }
  double obj;
  vector<double>& weights;
  const int from, to;
  vector<TrainingInstance>& corpus;
  vector<ArcFactoredForest>& forests;
  vector<double> g; // local gradient
};

int main(int argc, char** argv) {
  int rank = 0;
  int size = 1;
  po::variables_map conf;
  InitCommandLine(argc, argv, &conf);
  if (conf.count("cmph_perfect_feature_hash")) {
    cerr << "Loading perfect hash function from " << conf["cmph_perfect_feature_hash"].as<string>() << " ...\n";
    FD::EnableHash(conf["cmph_perfect_feature_hash"].as<string>());
    cerr << "  " << FD::NumFeats() << " features in map\n";
  }
  ArcFeatureFunctions ffs;
  vector<TrainingInstance> corpus;
  TrainingInstance::ReadTrainingCorpus(conf["training_data"].as<string>(), &corpus, rank, size);
  vector<weight_t> weights;
    Weights::InitFromFile(conf["weights"].as<string>(), &weights);
  vector<ArcFactoredForest> forests(corpus.size());
  SparseVector<double> empirical;
  cerr << "Extracting features...\n";
  bool flag = false;
  for (int i = 0; i < corpus.size(); ++i) {
    TrainingInstance& cur = corpus[i];
    if (rank == 0 && (i+1) % 10 == 0) { cerr << '.' << flush; flag = true; }
    if (rank == 0 && (i+1) % 400 == 0) { cerr << " [" << (i+1) << "]\n"; flag = false; }
    ffs.PrepareForInput(cur.ts);
    SparseVector<weight_t> efmap;
    for (int j = 0; j < cur.tree.h_m_pairs.size(); ++j) {
      efmap.clear();
      ffs.EdgeFeatures(cur.ts, cur.tree.h_m_pairs[j].first,
                       cur.tree.h_m_pairs[j].second,
                       &efmap);
      cur.features += efmap;
    }
    for (int j = 0; j < cur.tree.roots.size(); ++j) {
      efmap.clear();
      ffs.EdgeFeatures(cur.ts, -1, cur.tree.roots[j], &efmap);
      cur.features += efmap;
    }
    empirical += cur.features;
    forests[i].resize(cur.ts.words.size());
    forests[i].ExtractFeatures(cur.ts, ffs);
  }
  if (flag) cerr << endl;
  //cerr << "EMP: " << empirical << endl; //DE
  weights.resize(FD::NumFeats(), 0.0);
  vector<weight_t> g(FD::NumFeats(), 0.0);
  cerr << "features initialized\noptimizing...\n";
  boost::shared_ptr<BatchOptimizer> o;
#if HAVE_THREAD
  unsigned threads = conf["threads"].as<unsigned>();
  if (threads > corpus.size()) threads = corpus.size();
#else
  const unsigned threads = 1;
#endif
  int chunk = corpus.size() / threads;
  o.reset(new LBFGSOptimizer(g.size(), conf["correction_buffers"].as<int>()));
  int iterations = 1000;
  for (int iter = 0; iter < iterations; ++iter) {
    cerr << "ITERATION " << iter << " " << flush;
    fill(g.begin(), g.end(), 0.0);
    for (SparseVector<double>::iterator it = empirical.begin(); it != empirical.end(); ++it)
      g[it->first] = -it->second;
    double obj = -empirical.dot(weights);
    vector<boost::shared_ptr<GradientWorker> > jobs;
    for (int from = 0; from < corpus.size(); from += chunk) {
      int to = from + chunk;
      if (to > corpus.size()) to = corpus.size();
      jobs.push_back(boost::shared_ptr<GradientWorker>(new GradientWorker(from, to, &weights, &corpus, &forests)));
    }
#if HAVE_THREAD
    boost::thread_group tg;
    for (int i = 0; i < threads; ++i)
      tg.create_thread(boost::ref(*jobs[i]));
    tg.join_all();
#else
    (*jobs[0])();
#endif
    for (int i = 0; i < threads; ++i) {
      obj += jobs[i]->obj;
      vector<double>& tg = jobs[i]->g;
      for (unsigned j = 0; j < g.size(); ++j)
        g[j] += tg[j];
    }
    // SparseVector<double> mfm;  //DE
    //cerr << endl << "E: " << empirical << endl;  // DE
    //cerr << "M: " << mfm << endl;  // DE
    double r = ApplyRegularizationTerms(conf["regularization_strength"].as<double>(), weights, &g);
    double gnorm = 0;
    for (int i = 0; i < g.size(); ++i)
      gnorm += g[i]*g[i];
    ostringstream ll;
    ll << "ITER=" << (iter+1) << "\tOBJ=" << (obj+r) << "\t[F=" << obj << " R=" << r << "]\tGnorm=" << sqrt(gnorm);
    cerr << ' ' << ll.str().substr(ll.str().find('\t')+1) << endl;
    obj += r;
    assert(obj >= 0);
    o->Optimize(obj, g, &weights);
    Weights::ShowLargestFeatures(weights);
    const bool converged = o->HasConverged();
    const char* ofname = converged ? "weights.final.gz" : "weights.cur.gz";
    if (converged || ((iter+1) % conf["output_every_i_iterations"].as<unsigned>()) == 0) {
      cerr << "writing..." << flush;
      const string sl = ll.str();
      Weights::WriteToFile(ofname, weights, true, &sl);
      cerr << "done" << endl;
    }
    if (converged) { cerr << "CONVERGED\n"; break; }
  }
  return 0;
}