summaryrefslogtreecommitdiff
path: root/dtrain/dtrain.cc
blob: 581c985a3b7e5824450de8780805e627df343540 (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
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
#include "dtrain.h"


bool
dtrain_init(int argc, char** argv, po::variables_map* cfg)
{
  po::options_description ini("Configuration File Options");
  ini.add_options()
    ("input",           po::value<string>()->default_value("-"),                                                "input file")
    ("output",          po::value<string>()->default_value("-"),                       "output weights file, '-' for STDOUT")
    ("input_weights",   po::value<string>(),                             "input weights file (e.g. from previous iteration)")
    ("decoder_config",  po::value<string>(),                                                   "configuration file for cdec")
    ("print_weights",   po::value<string>(),                                            "weights to print on each iteration")
    ("stop_after",      po::value<unsigned>()->default_value(0),                              "stop after X input sentences")
    ("tmp",             po::value<string>()->default_value("/tmp"),                                        "temp dir to use")
    ("keep",            po::value<bool>()->zero_tokens(),                            "keep weights files for each iteration")
    ("hstreaming",      po::value<string>(),                                "run in hadoop streaming mode, arg is a task id")
    ("epochs",          po::value<unsigned>()->default_value(10),                            "# of iterations T (per shard)") 
    ("k",               po::value<unsigned>()->default_value(100),                         "how many translations to sample")
    ("sample_from",     po::value<string>()->default_value("kbest"),  "where to sample translations from: 'kbest', 'forest'")
    ("filter",          po::value<string>()->default_value("uniq"),                       "filter kbest list: 'not', 'uniq'")
    ("pair_sampling",   po::value<string>()->default_value("108010"),        "how to sample pairs: 'all', '108010' or 'PRO'")
    ("pair_threshold",  po::value<score_t>()->default_value(0),                       "bleu [0,1] threshold to filter pairs")
    ("N",               po::value<unsigned>()->default_value(4),                                       "N for Ngrams (BLEU)")
    ("scorer",          po::value<string>()->default_value("stupid_bleu"),        "scoring: bleu, stupid_, smooth_, approx_")
    ("learning_rate",   po::value<weight_t>()->default_value(0.0001),                                        "learning rate")
    ("gamma",           po::value<weight_t>()->default_value(0),                          "gamma for SVM (0 for perceptron)")
    ("select_weights",  po::value<string>()->default_value("last"), "output 'best' or 'last' weights ('VOID' to throw away)")
    ("rescale",         po::value<bool>()->zero_tokens(),                           "rescale weight vector after each input")
    ("l1_reg",          po::value<string>()->default_value("none"),   "apply l1 regularization as in 'Tsuroka et al' (2010)")
    ("l1_reg_strength", po::value<weight_t>(),                                                  "l1 regularization strength")
    ("funny",           po::value<bool>()->zero_tokens(),                      "include correctly ranked pairs into updates")
#ifdef DTRAIN_LOCAL
    ("refs,r",         po::value<string>(),                                                       "references in local mode")
#endif
    ("noup",           po::value<bool>()->zero_tokens(),                                             "do not update weights");
  po::options_description cl("Command Line Options");
  cl.add_options()
    ("config,c",         po::value<string>(),              "dtrain config file")
    ("quiet,q",          po::value<bool>()->zero_tokens(),           "be quiet")
    ("verbose,v",        po::value<bool>()->zero_tokens(),         "be verbose");
  cl.add(ini);
  po::store(parse_command_line(argc, argv, cl), *cfg);
  if (cfg->count("config")) {
    ifstream ini_f((*cfg)["config"].as<string>().c_str());
    po::store(po::parse_config_file(ini_f, ini), *cfg);
  }
  po::notify(*cfg);
  if (!cfg->count("decoder_config")) { 
    cerr << cl << endl;
    return false;
  }
  if (cfg->count("hstreaming") && (*cfg)["output"].as<string>() != "-") {
    cerr << "When using 'hstreaming' the 'output' param should be '-'." << endl;
    return false;
  }
#ifdef DTRAIN_LOCAL
  if ((*cfg)["input"].as<string>() == "-") {
    cerr << "Can't use stdin as input with this binary. Recompile without DTRAIN_LOCAL" << endl;
    return false;
  }
#endif
  if ((*cfg)["sample_from"].as<string>() != "kbest"
       && (*cfg)["sample_from"].as<string>() != "forest") {
    cerr << "Wrong 'sample_from' param: '" << (*cfg)["sample_from"].as<string>() << "', use 'kbest' or 'forest'." << endl;
    return false;
  }
  if ((*cfg)["sample_from"].as<string>() == "kbest" && (*cfg)["filter"].as<string>() != "uniq" &&
        (*cfg)["filter"].as<string>() != "not") {
    cerr << "Wrong 'filter' param: '" << (*cfg)["filter"].as<string>() << "', use 'uniq' or 'not'." << endl;
    return false;
  }
  if ((*cfg)["pair_sampling"].as<string>() != "all" && (*cfg)["pair_sampling"].as<string>() != "108010" &&
        (*cfg)["pair_sampling"].as<string>() != "PRO") {
    cerr << "Wrong 'pair_sampling' param: '" << (*cfg)["pair_sampling"].as<string>() << "'." << endl;
    return false;
  }
  if ((*cfg)["pair_threshold"].as<score_t>() < 0) {
    cerr << "The threshold must be >= 0!" << endl;
    return false;
  }
  if ((*cfg)["select_weights"].as<string>() != "last" && (*cfg)["select_weights"].as<string>() != "best" &&
        (*cfg)["select_weights"].as<string>() != "VOID") {
    cerr << "Wrong 'select_weights' param: '" << (*cfg)["select_weights"].as<string>() << "', use 'last' or 'best'." << endl;
    return false;
  }
  return true;
}

int
main(int argc, char** argv)
{
  // handle most parameters
  po::variables_map cfg;
  if (!dtrain_init(argc, argv, &cfg)) exit(1); // something is wrong 
  bool quiet = false;
  if (cfg.count("quiet")) quiet = true;
  bool verbose = false;  
  if (cfg.count("verbose")) verbose = true;
  bool noup = false;
  if (cfg.count("noup")) noup = true;
  bool hstreaming = false;
  string task_id;
  if (cfg.count("hstreaming")) {
    hstreaming = true;
    quiet = true;
    task_id = cfg["hstreaming"].as<string>();
    cerr.precision(17);
  }
  bool rescale = false;
  if (cfg.count("rescale")) rescale = true;
  HSReporter rep(task_id);
  bool keep = false;
  if (cfg.count("keep")) keep = true;
  bool funny = false;
  if (cfg.count("funny"))
    funny = true;

  const unsigned k = cfg["k"].as<unsigned>();
  const unsigned N = cfg["N"].as<unsigned>(); 
  const unsigned T = cfg["epochs"].as<unsigned>();
  const unsigned stop_after = cfg["stop_after"].as<unsigned>();
  const string filter_type = cfg["filter"].as<string>();
  const string sample_from = cfg["sample_from"].as<string>();
  const string pair_sampling = cfg["pair_sampling"].as<string>();
  const score_t pair_threshold = cfg["pair_threshold"].as<score_t>();
  const string select_weights = cfg["select_weights"].as<string>();
  vector<string> print_weights;
  if (cfg.count("print_weights"))
    boost::split(print_weights, cfg["print_weights"].as<string>(), boost::is_any_of(" "));

  // setup decoder
  register_feature_functions();
  SetSilent(true);
  ReadFile ini_rf(cfg["decoder_config"].as<string>());
  if (!quiet)
    cerr << setw(25) << "cdec cfg " << "'" << cfg["decoder_config"].as<string>() << "'" << endl;
  Decoder decoder(ini_rf.stream());

  // scoring metric/scorer
  string scorer_str = cfg["scorer"].as<string>();
  LocalScorer* scorer;
  if (scorer_str == "bleu") {
    scorer = dynamic_cast<BleuScorer*>(new BleuScorer);
  } else if (scorer_str == "stupid_bleu") {
    scorer = dynamic_cast<StupidBleuScorer*>(new StupidBleuScorer);
  } else if (scorer_str == "smooth_bleu") {
    scorer = dynamic_cast<SmoothBleuScorer*>(new SmoothBleuScorer);
  } else if (scorer_str == "approx_bleu") {
    scorer = dynamic_cast<ApproxBleuScorer*>(new ApproxBleuScorer(N));
  } else {
    cerr << "Don't know scoring metric: '" << scorer_str << "', exiting." << endl;
    exit(1);
  }
  vector<score_t> bleu_weights;
  scorer->Init(N, bleu_weights);
  if (!quiet) cerr << setw(26) << "scorer '" << scorer_str << "'" << endl << endl;

  // setup decoder observer
  MT19937 rng; // random number generator
  HypSampler* observer;
  if (sample_from == "kbest")
    observer = dynamic_cast<KBestGetter*>(new KBestGetter(k, filter_type));
  else
    observer = dynamic_cast<KSampler*>(new KSampler(k, &rng));
  observer->SetScorer(scorer);

  // init weights
  vector<weight_t>& dense_weights = decoder.CurrentWeightVector();
  SparseVector<weight_t> lambdas, cumulative_penalties;
  if (cfg.count("input_weights")) Weights::InitFromFile(cfg["input_weights"].as<string>(), &dense_weights);
  Weights::InitSparseVector(dense_weights, &lambdas);

  // meta params for perceptron, SVM
  weight_t eta = cfg["learning_rate"].as<weight_t>();
  weight_t gamma = cfg["gamma"].as<weight_t>();

  // l1 regularization
  bool l1naive = false;
  bool l1clip = false;
  bool l1cumul = false;
  weight_t l1_reg = 0;
  if (cfg["l1_reg"].as<string>() != "none") {
    string s = cfg["l1_reg"].as<string>();
    if (s == "naive") l1naive = true;
    else if (s == "clip") l1clip = true;
    else if (s == "cumul") l1cumul = true;
    l1_reg = cfg["l1_reg_strength"].as<weight_t>();
  }

  // output
  string output_fn = cfg["output"].as<string>();
  // input
  string input_fn = cfg["input"].as<string>();
  ReadFile input(input_fn);
  // buffer input for t > 0
  vector<string> src_str_buf;          // source strings (decoder takes only strings)
  vector<vector<WordID> > ref_ids_buf; // references as WordID vecs
  // where temp files go
  string tmp_path = cfg["tmp"].as<string>();
  vector<string> w_tmp_files; // used for keep
#ifdef DTRAIN_LOCAL
  string refs_fn = cfg["refs"].as<string>();
  ReadFile refs(refs_fn);
#else
  string grammar_buf_fn = gettmpf(tmp_path, "dtrain-grammars");
  ogzstream grammar_buf_out;
  grammar_buf_out.open(grammar_buf_fn.c_str());
#endif

  unsigned in_sz = UINT_MAX; // input index, input size
  vector<pair<score_t, score_t> > all_scores;
  score_t max_score = 0.;
  unsigned best_it = 0;
  float overall_time = 0.;

  // output cfg
  if (!quiet) {
    cerr << _p5;
    cerr << endl << "dtrain" << endl << "Parameters:" << endl;
    cerr << setw(25) << "k " << k << endl;
    cerr << setw(25) << "N " << N << endl;
    cerr << setw(25) << "T " << T << endl;
    cerr << setw(25) << "sample from " << "'" << sample_from << "'" << endl;
    if (sample_from == "kbest")
      cerr << setw(25) << "filter " << "'" << filter_type << "'" << endl;
    cerr << setw(25) << "learning rate " << eta << endl;
    cerr << setw(25) << "gamma " << gamma << endl;
    cerr << setw(25) << "pairs " << "'" << pair_sampling << "'" << endl;
    cerr << setw(25) << "pair threshold " << pair_threshold << endl;
    cerr << setw(25) << "select weights " << "'" << select_weights << "'" << endl;
    if (cfg.count("l1_reg"))
      cerr << setw(25) << "l1 reg " << l1_reg << " '" << cfg["l1_reg"].as<string>() << "'" << endl;
    if (funny)
      cerr << setw(25) << "funny " << funny << endl;
    if (rescale)
      cerr << setw(25) << "rescale " << rescale << endl;
    cerr << setw(25) << "cdec cfg " << "'" << cfg["decoder_config"].as<string>() << "'" << endl;
    cerr << setw(25) << "input " << "'" << input_fn << "'" << endl;
#ifdef DTRAIN_LOCAL 
    cerr << setw(25) << "refs " << "'" << refs_fn << "'" << endl;
#endif
    cerr << setw(25) << "output " << "'" << output_fn << "'" << endl;
    if (cfg.count("input_weights"))
      cerr << setw(25) << "weights in" << cfg["input_weights"].as<string>() << endl;
    if (cfg.count("stop-after"))
      cerr << setw(25) << "stop_after " << stop_after << endl;
    if (!verbose) cerr << "(a dot represents " << DTRAIN_DOTS << " lines of input)" << endl;
  }


  for (unsigned t = 0; t < T; t++) // T epochs
  {

  if (hstreaming) cerr << "reporter:status:Iteration #" << t+1 << " of " << T << endl;

  time_t start, end;  
  time(&start);
#ifndef DTRAIN_LOCAL
  igzstream grammar_buf_in;
  if (t > 0) grammar_buf_in.open(grammar_buf_fn.c_str());
#endif
  score_t score_sum = 0.;
  score_t model_sum(0);
  unsigned ii = 0, rank_errors = 0, margin_violations = 0, npairs = 0;
  if (!quiet) cerr << "Iteration #" << t+1 << " of " << T << "." << endl;

  while(true)
  {

    string in;
    bool next = false, stop = false; // next iteration or premature stop
    if (t == 0) {
      if(!getline(*input, in)) next = true;
    } else {
      if (ii == in_sz) next = true; // stop if we reach the end of our input
    }
    // stop after X sentences (but still iterate for those)
    if (stop_after > 0 && stop_after == ii && !next) stop = true;
    
    // produce some pretty output
    if (!quiet && !verbose) {
      if (ii == 0) cerr << " ";
      if ((ii+1) % (DTRAIN_DOTS) == 0) {
        cerr << ".";
        cerr.flush();
      }
      if ((ii+1) % (20*DTRAIN_DOTS) == 0) {
        cerr << " " << ii+1 << endl;
        if (!next && !stop) cerr << " ";
      }
      if (stop) {
        if (ii % (20*DTRAIN_DOTS) != 0) cerr << " " << ii << endl;
        cerr << "Stopping after " << stop_after << " input sentences." << endl;
      } else {
        if (next) {
          if (ii % (20*DTRAIN_DOTS) != 0) cerr << " " << ii << endl;
        }
      }
    }
   
    // next iteration
    if (next || stop) break;

    // weights
    lambdas.init_vector(&dense_weights);

    // getting input
    vector<WordID> ref_ids; // reference as vector<WordID>
#ifndef DTRAIN_LOCAL
    vector<string> in_split; // input: sid\tsrc\tref\tpsg
    if (t == 0) {
      // handling input
      split_in(in, in_split); 
      if (hstreaming && ii == 0) cerr << "reporter:counter:" << task_id << ",First ID," << in_split[0] << endl;
      // getting reference
      vector<string> ref_tok;
      boost::split(ref_tok, in_split[2], boost::is_any_of(" "));
      register_and_convert(ref_tok, ref_ids);
      ref_ids_buf.push_back(ref_ids);
      // process and set grammar
      bool broken_grammar = true;
      for (string::iterator it = in.begin(); it != in.end(); it++) {
        if (!isspace(*it)) {
          broken_grammar = false;
          break;
        }
      }
      if (broken_grammar) continue;
      boost::replace_all(in, "\t", "\n");
      in += "\n";
      grammar_buf_out << in << DTRAIN_GRAMMAR_DELIM << " " << in_split[0] << endl;
      decoder.SetSentenceGrammarFromString(in);
      src_str_buf.push_back(in_split[1]);
      // decode
      observer->SetRef(ref_ids);
      decoder.Decode(in_split[1], observer);
    } else {
      // get buffered grammar
      string grammar_str;
      while (true) {
        string rule;
        getline(grammar_buf_in, rule);
        if (boost::starts_with(rule, DTRAIN_GRAMMAR_DELIM)) break;
        grammar_str += rule + "\n";
      }
      decoder.SetSentenceGrammarFromString(grammar_str);
      // decode
      observer->SetRef(ref_ids_buf[ii]);
      decoder.Decode(src_str_buf[ii], observer);
    }
#else
    if (t == 0) {
      string r_;
      getline(*refs, r_);
      vector<string> ref_tok;
      boost::split(ref_tok, r_, boost::is_any_of(" "));
      register_and_convert(ref_tok, ref_ids);
      ref_ids_buf.push_back(ref_ids);
      src_str_buf.push_back(in);
    } else {
      ref_ids = ref_ids_buf[ii];
    }
    observer->SetRef(ref_ids);
    if (t == 0) 
      decoder.Decode(in, observer);
    else
      decoder.Decode(src_str_buf[ii], observer);
#endif

    // get (scored) samples 
    vector<ScoredHyp>* samples = observer->GetSamples();

    if (verbose) {
      cerr << "--- ref for " << ii << " ";
      if (t > 0) printWordIDVec(ref_ids_buf[ii]);
      else printWordIDVec(ref_ids);
      for (unsigned u = 0; u < samples->size(); u++) {
        cerr << _p5 << _np << "[" << u << ". '";
        printWordIDVec((*samples)[u].w);
        cerr << "'" << endl;
        cerr << "SCORE=" << (*samples)[0].score << ",model="<< (*samples)[0].model << endl;
        cerr << "F{" << (*samples)[0].f << "} ]" << endl << endl;
      }
    }

    score_sum += (*samples)[0].score;
    model_sum += (*samples)[0].model;

    // weight updates
    if (!noup) {
      vector<pair<ScoredHyp,ScoredHyp> > pairs;
      if (pair_sampling == "all")
        all_pairs(samples, pairs, pair_threshold);
      if (pair_sampling == "108010")
        part108010(samples, pairs, pair_threshold);
      if (pair_sampling == "PRO")
        PROsampling(samples, pairs);
      npairs += pairs.size();

      for (vector<pair<ScoredHyp,ScoredHyp> >::iterator it = pairs.begin();
           it != pairs.end(); it++) {
        score_t rank_error = it->second.score - it->first.score;
        if (!gamma) {
          // perceptron
          if (rank_error > 0) {
            SparseVector<weight_t> diff_vec = it->second.f - it->first.f;
            lambdas.plus_eq_v_times_s(diff_vec, eta);
            rank_errors++;
          } else {
            if (funny) {
              SparseVector<weight_t> diff_vec = it->first.f - it->second.f;
              lambdas.plus_eq_v_times_s(diff_vec, eta);
            }
          }
          if (it->first.model - it->second.model < 1) margin_violations++;
        } else {
          // SVM
          score_t margin = it->first.model - it->second.model;
          if (rank_error > 0 || margin < 1) {
            SparseVector<weight_t> diff_vec = it->second.f - it->first.f;
            lambdas.plus_eq_v_times_s(diff_vec, eta);
            if (rank_error > 0) rank_errors++;
            if (margin < 1) margin_violations++;
          }
          // regularization
          lambdas.plus_eq_v_times_s(lambdas, -2*gamma*eta*(1./npairs));
        }
      }

      ////////
      // TEST THIS
      // reset cumulative_penalties after 1 iter? 
      // do this only once per INPUT (not per pair)
if (false) {
      if (l1naive) {
        for (unsigned d = 0; d < lambdas.size(); d++) {
          weight_t v = lambdas.get(d);
          lambdas.set_value(d, v - sign(v) * l1_reg);
        }
      } else if (l1clip) {
        for (unsigned d = 0; d < lambdas.size(); d++) {
          if (lambdas.nonzero(d)) {
            weight_t v = lambdas.get(d);
            if (v > 0) {
              lambdas.set_value(d, max(0., v - l1_reg));
            } else {
              lambdas.set_value(d, min(0., v + l1_reg));
            }
          }
        }
      } else if (l1cumul) {
        weight_t acc_penalty = (ii+1) * l1_reg; // ii is the index of the current input 
        for (unsigned d = 0; d < lambdas.size(); d++) {
          if (lambdas.nonzero(d)) {
            weight_t v = lambdas.get(d);
            weight_t penalty = 0;
            if (v > 0) {
              penalty = max(0., v-(acc_penalty + cumulative_penalties.get(d)));
            } else {
              penalty = min(0., v+(acc_penalty - cumulative_penalties.get(d)));
            }
            lambdas.set_value(d, penalty);
            cumulative_penalties.set_value(d, cumulative_penalties.get(d)+penalty);
          }
        }
      }
    }
}
    ////////

    if (rescale) lambdas /= lambdas.l2norm();
    
    ++ii;

    if (hstreaming) {
      rep.update_counter("Seen #"+boost::lexical_cast<string>(t+1), 1u);
      rep.update_counter("Seen", 1u);
    }

  } // input loop

  if (scorer_str == "approx_bleu") scorer->Reset();

  if (t == 0) {
    in_sz = ii; // remember size of input (# lines)
    if (hstreaming) {
      rep.update_counter("|Input|", ii);
      rep.update_gcounter("|Input|", ii);
      rep.update_gcounter("Shards", 1u);
    }
  }

#ifndef DTRAIN_LOCAL
  if (t == 0) {
    grammar_buf_out.close();
  } else {
    grammar_buf_in.close();
  }
#endif

  // print some stats
  score_t score_avg = score_sum/(score_t)in_sz;
  score_t model_avg = model_sum/(score_t)in_sz;
  score_t score_diff, model_diff;
  if (t > 0) {
    score_diff = score_avg - all_scores[t-1].first;
    model_diff = model_avg - all_scores[t-1].second;
  } else {
    score_diff = score_avg;
    model_diff = model_avg;
  }
  
  unsigned nonz;
  if (!quiet || hstreaming) nonz = (unsigned)lambdas.size_nonzero();

  if (!quiet) {
    cerr << _p5 << _p << "WEIGHTS" << endl;
    for (vector<string>::iterator it = print_weights.begin(); it != print_weights.end(); it++) {
      cerr << setw(18) << *it << " = " << lambdas.get(FD::Convert(*it)) << endl;
    }
    cerr << "        ---" << endl;
    cerr << _np << "      1best avg score: " << score_avg;
    cerr << _p << " (" << score_diff << ")" << endl;
    cerr << _np << "1best avg model score: " << model_avg;
    cerr << _p << " (" << model_diff << ")" << endl;
    cerr << "           avg #pairs: ";
    cerr << _np << npairs/(float)in_sz << endl;
    cerr << "        avg #rank err: ";
    cerr << rank_errors/(float)in_sz << endl;
    cerr << "     avg #margin viol: ";
    cerr << margin_violations/float(in_sz) << endl;
    cerr << "   non0 feature count: ";
    cerr << nonz << endl;
  }

  if (hstreaming) {
    rep.update_counter("Score 1best avg #"+boost::lexical_cast<string>(t+1), (unsigned)(score_avg*DTRAIN_SCALE)); 
    rep.update_counter("Model 1best avg #"+boost::lexical_cast<string>(t+1), (unsigned)(model_avg*DTRAIN_SCALE)); 
    rep.update_counter("Pairs avg #"+boost::lexical_cast<string>(t+1), (unsigned)((npairs/(weight_t)in_sz)*DTRAIN_SCALE)); 
    rep.update_counter("Rank errors avg #"+boost::lexical_cast<string>(t+1), (unsigned)((rank_errors/(weight_t)in_sz)*DTRAIN_SCALE)); 
    rep.update_counter("Margin violations avg #"+boost::lexical_cast<string>(t+1), (unsigned)((margin_violations/(weight_t)in_sz)*DTRAIN_SCALE)); 
    rep.update_counter("Non zero feature count #"+boost::lexical_cast<string>(t+1), nonz); 
    rep.update_gcounter("Non zero feature count #"+boost::lexical_cast<string>(t+1), nonz);
  }

  pair<score_t,score_t> remember;
  remember.first = score_avg;
  remember.second = model_avg;
  all_scores.push_back(remember);
  if (score_avg > max_score) {
    max_score = score_avg;
    best_it = t;
  }
  time (&end);
  float time_diff = difftime(end, start);
  overall_time += time_diff;
  if (!quiet) {
    cerr << _p2 << _np << "(time " << time_diff/60. << " min, ";
    cerr << time_diff/(float)in_sz<< " s/S)" << endl;
  }
  if (t+1 != T && !quiet) cerr << endl;

  if (noup) break;

  // write weights to file
  if (select_weights == "best" || keep) {
    lambdas.init_vector(&dense_weights);
    string w_fn = "weights." + boost::lexical_cast<string>(t) + ".gz";
    Weights::WriteToFile(w_fn, dense_weights, true); 
  }

  } // outer loop

#ifndef DTRAIN_LOCAL
  unlink(grammar_buf_fn.c_str());
#endif

  if (!noup) {
    if (!quiet) cerr << endl << "Writing weights file to '" << output_fn << "' ..." << endl;
    if (select_weights == "last") { // last
      WriteFile of(output_fn); // works with '-'
      ostream& o = *of.stream();
      o.precision(17);
      o << _np;
      for (SparseVector<weight_t>::const_iterator it = lambdas.begin(); it != lambdas.end(); ++it) {
	    if (it->second == 0) continue;
        o << FD::Convert(it->first) << '\t' << it->second << endl;
      }
    } else if (select_weights == "VOID") { // do nothing with the weights
    } else { // best
      if (output_fn != "-") {
        CopyFile("weights."+boost::lexical_cast<string>(best_it)+".gz", output_fn);
      } else {
        ReadFile bestw("weights."+boost::lexical_cast<string>(best_it)+".gz");
        string o;
        cout.precision(17);
        cout << _np;
        while(getline(*bestw, o)) cout << o << endl;
      }
      if (!keep) {
        for (unsigned i = 0; i < T; i++) {
          string s = "weights." + boost::lexical_cast<string>(i) + ".gz";
          unlink(s.c_str());
        }
      }
    }
    if (output_fn == "-" && hstreaming) cout << "__SHARD_COUNT__\t1" << endl;
    if (!quiet) cerr << "done" << endl;
  }
  
  if (!quiet) {
    cerr << _p5 << _np << endl << "---" << endl << "Best iteration: ";
    cerr << best_it+1 << " [SCORE '" << scorer_str << "'=" << max_score << "]." << endl;
    cerr << _p2 << "This took " << overall_time/60. << " min." << endl;
  }

  if (keep) {
    cout << endl << "Weight files per iteration:" << endl;
    for (unsigned i = 0; i < w_tmp_files.size(); i++) {
      cout << w_tmp_files[i] << endl;
    }
  }

  return 0;
}