summaryrefslogtreecommitdiff
path: root/vest/mr_vest_generate_mapper_input.cc
blob: 062d57b5bc1ece7ee051498a92cd6009469ea695 (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
//TODO: debug segfault when references supplied, null shared_ptr when oracle
#include <iostream>
#include <vector>
#include <sstream>

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

#include "sampler.h"
#include "filelib.h"
#include "weights.h"
#include "line_optimizer.h"
#include "hg.h"
#include "hg_io.h"
#include "scorer.h"
#include "oracle_bleu.h"
#include "ff_bleu.h"

const bool DEBUG_ORACLE=true;

boost::shared_ptr<FFRegistry> global_ff_registry;
namespace {
struct init_ff {
  init_ff() {
    global_ff_registry.reset(new FFRegistry);
    global_ff_registry->Register(new FFFactory<BLEUModel>);
  }
};
init_ff reg;
}

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

typedef SparseVector<double> Dir;
typedef Dir Point;

void compress_similar(vector<Dir> &dirs,double min_dist,ostream *log=&cerr,bool avg=true,bool verbose=true) {
  //  return; //TODO: debug
  if (min_dist<=0) return;
  double max_s=1.-min_dist;
  if (log&&verbose) *log<<"max allowed S="<<max_s<<endl;
  unsigned N=dirs.size();
  for (int i=0;i<N;++i) {
    for (int j=i+1;j<N;++j) {
      double s=dirs[i].tanimoto_coef(dirs[j]);
      if (log&&verbose) *log<<"S["<<i<<","<<j<<"]="<<s<<' ';
      if (s>max_s) {
        if (log) *log << "Collapsing similar directions (T="<<s<<" > "<<max_s<<").  dirs["<<i<<"]="<<dirs[i]<<" dirs["<<j<<"]"<<endl;
        if (avg) {
          dirs[i]+=dirs[j];
          dirs[i]/=2.;
          if (log) *log<<" averaged="<<dirs[i];
        }
        if (log) *log<<endl;
        swap(dirs[j],dirs[--N]);
      }
    }
    if (log&&verbose) *log<<endl;

  }
  dirs.resize(N);
}

struct oracle_directions {
  MT19937 rng;
  OracleBleu oracle;
  vector<Dir> directions;

  bool start_random;
  bool include_primary;
  bool old_to_hope;
  bool fear_to_hope;
  unsigned n_random;
  void AddPrimaryAndRandomDirections() {
    LineOptimizer::CreateOptimizationDirections(
      fids,n_random,&rng,&directions,include_primary);
  }

  void Print() {
    for (int i = 0; i < dev_set_size; ++i)
      for (int j = 0; j < directions.size(); ++j)
        cout << forest_file(i) <<" " << i << ' ' << origin << ' ' << directions[j] << endl;
  }

  void AddOptions(po::options_description *opts) {
    oracle.AddOptions(opts);
    opts->add_options()
      ("dev_set_size,s",po::value<unsigned>(&dev_set_size),"[REQD] Development set size (# of parallel sentences)")
      ("forest_repository,r",po::value<string>(&forest_repository),"[REQD] Path to forest repository")
      ("weights,w",po::value<string>(&weights_file),"[REQD] Current feature weights file")
      ("optimize_feature,o",po::value<vector<string> >(), "Feature to optimize (if none specified, all weights listed in the weights file will be optimized)")
      ("random_directions,d",po::value<unsigned>(&n_random)->default_value(10),"Number of random directions to run the line optimizer in")
      ("no_primary,n","don't use the primary (orthogonal each feature alone) directions")
      ("oracle_directions,O",po::value<unsigned>(&n_oracle)->default_value(0),"read the forests and choose this many directions based on heading toward a hope max (bleu+modelscore) translation.")
      ("oracle_start_random",po::bool_switch(&start_random),"sample random subsets of dev set for ALL oracle directions, not just those after a sequential run through it")
      ("oracle_batch,b",po::value<unsigned>(&oracle_batch)->default_value(10),"to produce each oracle direction, sum the 'gradient' over this many sentences")
      ("max_similarity,m",po::value<double>(&max_similarity)->default_value(0),"remove directions that are too similar (Tanimoto coeff. less than (1-this)).  0 means don't filter, 1 means only 1 direction allowed?")
      ("fear_to_hope,f",po::bool_switch(&fear_to_hope),"for each of the oracle_directions, also include a direction from fear to hope (as well as origin to hope)")
      ("no_old_to_hope,n","don't emit the usual old -> hope oracle")
      ("decoder_translations",po::value<string>(&decoder_translations_file)->default_value(""),"one per line decoder 1best translations for computing document BLEU vs. sentences-seen-so-far BLEU")
      ("verbose",po::bool_switch(&verbose),"detailed logs")
      ;
  }
  void InitCommandLine(int argc, char *argv[], po::variables_map *conf) {
    po::options_description opts("Configuration options");
    AddOptions(&opts);
    opts.add_options()("help,h", "Help");

    po::options_description dcmdline_options;
    dcmdline_options.add(opts);
    po::store(parse_command_line(argc, argv, dcmdline_options), *conf);
    bool flag = false;
    if (conf->count("dev_set_size") == 0) {
      cerr << "Please specify the size of the development set using -d N\n";
      flag = true;
    }
    if (conf->count("weights") == 0) {
      cerr << "Please specify the starting-point weights using -w <weightfile.txt>\n";
      flag = true;
    }
    if (conf->count("forest_repository") == 0) {
      cerr << "Please specify the forest repository location using -r <DIR>\n";
      flag = true;
    }
    if (flag || conf->count("help")) {
      cerr << dcmdline_options << endl;
      exit(1);
    }
    po::notify(*conf);

    if (0) {
    dev_set_size = (*conf)["dev_set_size"].as<unsigned>();
    forest_repository = (*conf)["forest_repository"].as<string>();
    weights_file = (*conf)["weights"].as<string>();
    n_random = (*conf)["random_directions"].as<unsigned>();
    }
  }

  int main(int argc, char *argv[]) {
    po::variables_map conf;
    InitCommandLine(argc,argv,&conf);
    UseConf(conf);
    Run();
    return 0;
  }
  bool verbose;
  void Run() {
    AddPrimaryAndRandomDirections();
    AddOracleDirections();
    compress_similar(directions,max_similarity,&cerr,true,verbose);
    Print();
  }


  Point origin; // old weights that gave model 1best.
  vector<string> optimize_features;
  void UseConf(po::variables_map const& conf) {
    oracle.UseConf(conf);
      // po::value<X>(&var) takes care of below:
      //    fear_to_hope=conf.count("fear_to_hope");
      //    n_random=conf["random_directions"].as<unsigned int>();
      //    forest_repository=conf["forest_repository"].as<string>();
      //    dev_set_size=conf["dev_set_size"].as<unsigned int>();
      //    n_oracle=conf["oracle_directions"].as<unsigned>();
      //    oracle_batch=conf["oracle_batch"].as<unsigned>();
      //    max_similarity=conf["max_similarity"].as<double>();
      //    weights_file=conf["weights"].as<string>();
    include_primary=!conf.count("no_primary");
    old_to_hope=!conf.count("no_old_to_hope");

    if (conf.count("optimize_feature") > 0)
      optimize_features=conf["optimize_feature"].as<vector<string> >();
    Init();
  }

  string weights_file;
  double max_similarity;
  unsigned n_oracle, oracle_batch;
  string forest_repository;
  unsigned dev_set_size;
  vector<Oracle> oracles;
  vector<int> fids;
  string forest_file(unsigned i) const {
    ostringstream o;
    o << forest_repository << '/' << i << ".json.gz";
    return o.str();
  }

  oracle_directions() { }

  Sentences model_hyps;

  vector<OracleBleu::ScoreP> model_scores;
  bool have_doc;
  void Init() {
    have_doc=!decoder_translations_file.empty();
    if (have_doc) {
      model_hyps.Load(decoder_translations_file);
      model_scores.resize(model_hyps.size());
      for (int i=0;i<model_hyps.size();++i) {
        //FIXME: what is scoreccand? with / without clipping? do without for consistency w/ oracle
        Score *s=oracle.ds[i]->ScoreCandidate(model_hyps[i]);
        model_scores[i].reset(s);
        oracle.doc_score->PlusEquals(*s);
      }
      //TODO: compute doc bleu stats for each sentence, then when getting oracle temporarily exclude stats for that sentence (skip regular score updating)
    }
    start_random=false;
    cerr << "Forest repo: " << forest_repository << endl;
    assert(DirectoryExists(forest_repository));
    vector<string> features;
    weights.InitFromFile(weights_file, &features);
    if (optimize_features.size())
      features=optimize_features;
    weights.InitSparseVector(&origin);
    fids.clear();
    AddFeatureIds(features);
    oracles.resize(dev_set_size);
  }

  Weights weights;
  void AddFeatureIds(vector<string> const& features) {
    int i = fids.size();
    fids.resize(fids.size()+features.size());
    for (; i < features.size(); ++i)
      fids[i] = FD::Convert(features[i]);
 }


  std::string decoder_translations_file; // one per line
  //TODO: is it worthwhile to get a complete document bleu first?  would take a list of 1best translations one per line from the decoders, rather than loading all the forests (expensive).  translations are in run.raw.N.gz - new arg
  void adjust_doc(unsigned i,double scale=1.) {
    oracle.doc_score->PlusEquals(*model_scores[i],scale);
  }

  Score &ds() {
    return *oracle.doc_score;
  }

  Oracle const& ComputeOracle(unsigned i) {
    Oracle &o=oracles[i];
    if (o.is_null()) {
      if (have_doc) {
        if (verbose) cerr<<"Before removing i="<<i<<" "<<ds().ScoreDetails()<<"\n";
        adjust_doc(i,-1);
      }
      ReadFile rf(forest_file(i));
      Hypergraph hg;
      {
        Timer t("Loading forest from JSON "+forest_file(i));
        HypergraphIO::ReadFromJSON(rf.stream(), &hg);
      }
      if (verbose) cerr<<"Before oracle["<<i<<"]: "<<ds().ScoreDetails();
      o=oracle.ComputeOracle(oracle.MakeMetadata(hg,i),&hg,origin,&cerr);
      if (verbose) {
        cerr << o;
        cerr<<" ; after: "<<ds().ScoreDetails()
            <<" oracle="<<oracle.GetScore(o.hope.sentence,i)->ScoreDetails()
            <<" model="<<oracle.GetScore(o.model.sentence,i)->ScoreDetails()<<endl;
        if (have_doc)
          cerr<<" doc (should = model): "<<model_scores[i]->ScoreDetails()<<endl;
      }
      if (have_doc) {
        adjust_doc(i,1);
      } else
        oracle.IncludeLastScore();
    }
    return o;
  }

  // if start_random is true, immediately sample w/ replacement from src sentences; otherwise, consume them sequentially until exhausted, then random.  oracle vectors are summed
  void AddOracleDirections() {
    MT19937::IntRNG rsg=rng.inclusive(0,dev_set_size-1);
    unsigned b=0;
    for(unsigned i=0;i<n_oracle;++i) {
      Dir o2hope;
      Dir fear2hope;
      for (unsigned j=0;j<oracle_batch;++j,++b) {
        Oracle const& o=ComputeOracle((start_random||b>=dev_set_size) ? rsg() : b);

        if (old_to_hope)
          o2hope+=o.ModelHopeGradient();
        if (fear_to_hope)
          fear2hope+=o.FearHopeGradient();
      }
      double N=(double)oracle_batch;
      if (old_to_hope) {
        o2hope/=N;
        directions.push_back(o2hope);
      }
      if (fear_to_hope) {
        fear2hope/=N;
        directions.push_back(fear2hope);
      }
    }
  }
};

int main(int argc, char** argv) {
  oracle_directions od;
  return od.main(argc,argv);
}