From 68b0969b1f41eacb4b336a66625894995b2f1e74 Mon Sep 17 00:00:00 2001 From: Patrick Simianer Date: Tue, 12 Nov 2013 20:07:47 +0100 Subject: impl repeat param --- training/dtrain/dtrain.cc | 78 ++++++++++++++++------------ training/dtrain/examples/standard/dtrain.ini | 6 ++- 2 files changed, 49 insertions(+), 35 deletions(-) diff --git a/training/dtrain/dtrain.cc b/training/dtrain/dtrain.cc index 23131810..441e2cd7 100644 --- a/training/dtrain/dtrain.cc +++ b/training/dtrain/dtrain.cc @@ -43,7 +43,7 @@ dtrain_init(int argc, char** argv, po::variables_map* cfg) ("max_pairs", po::value()->default_value(std::numeric_limits::max()), "max. # of pairs per Sent.") ("pclr", po::value()->default_value("no"), "use a (simple|adagrad) per-coordinate learning rate") ("batch", po::value()->zero_tokens(), "do batch optimization") - //("repeat", po::value()->default_value(1), "repeat optimization over kbest list this number of times") + ("repeat", po::value()->default_value(1), "repeat optimization over kbest list this number of times") //("test-k-best", po::value()->zero_tokens(), "check if optimization works (use repeat >= 2)") ("noup", po::value()->zero_tokens(), "do not update weights"); po::options_description cl("Command Line Options"); @@ -129,7 +129,7 @@ main(int argc, char** argv) const float hi_lo = cfg["hi_lo"].as(); const score_t approx_bleu_d = cfg["approx_bleu_d"].as(); const unsigned max_pairs = cfg["max_pairs"].as(); - //int repeat = cfg["repeat"].as(); + int repeat = cfg["repeat"].as(); //bool test_k_best = false; //if (cfg.count("test-k-best")) test_k_best = true; weight_t loss_margin = cfg["loss_margin"].as(); @@ -276,7 +276,7 @@ main(int argc, char** argv) cerr << setw(25) << "rescale " << rescale << endl; cerr << setw(25) << "pclr " << pclr << endl; cerr << setw(25) << "max pairs " << max_pairs << endl; - //cerr << setw(25) << "repeat " << repeat << endl; + cerr << setw(25) << "repeat " << repeat << endl; //cerr << setw(25) << "test k-best " << test_k_best << endl; cerr << setw(25) << "cdec cfg " << "'" << cfg["decoder_config"].as() << "'" << endl; cerr << setw(25) << "input " << "'" << input_fn << "'" << endl; @@ -294,23 +294,19 @@ main(int argc, char** argv) SparseVector learning_rates; // batch SparseVector batch_updates; - weight_t batch_loss; - - //int did_improve; // FIXME for test-k-best + score_t batch_loss; for (unsigned t = 0; t < T; t++) // T epochs { - + time_t start, end; time(&start); score_t score_sum = 0.; score_t model_sum(0); - unsigned ii = 0, rank_errors = 0, margin_violations = 0, npairs = 0, f_count = 0, list_sz = 0; + unsigned ii = 0, rank_errors = 0, margin_violations = 0, npairs = 0, f_count = 0, list_sz = 0, kbest_loss_improve = 0; batch_loss = 0.; if (!quiet) cerr << "Iteration #" << t+1 << " of " << T << "." << endl; - //did_improve = 0; - while(true) { @@ -395,8 +391,10 @@ main(int argc, char** argv) } } - score_sum += (*samples)[0].score; // stats for 1best - model_sum += (*samples)[0].model; + if (repeat == 1) { + score_sum += (*samples)[0].score; // stats for 1best + model_sum += (*samples)[0].model; + } f_count += observer->get_f_count(); list_sz += observer->get_sz(); @@ -414,24 +412,22 @@ main(int argc, char** argv) int cur_npairs = pairs.size(); npairs += cur_npairs; - weight_t kbest_loss_first, kbest_loss_last = 0.0; + score_t kbest_loss_first, kbest_loss_last = 0.0; -//for (int q=0; q < repeat; q++) { // repeat + for (int ki=0; ki < repeat; ki++) { - weight_t kbest_loss = 0.0; // test-k-best + score_t kbest_loss = 0.0; // test-k-best SparseVector lambdas_copy; // for l1 regularization SparseVector sum_up; // for pclr if (l1naive||l1clip||l1cumul) lambdas_copy = lambdas; for (vector >::iterator it = pairs.begin(); it != pairs.end(); it++) { - - /*if (repeat > 1) { - double x = max(0.0, -1.0 * (lambdas.dot(it->first.f) - lambdas.dot(it->second.f))); - kbest_loss += x; - }*/ - score_t model_diff = it->first.model - it->second.model; + if (repeat > 1) { + model_diff = lambdas.dot(it->first.f) - lambdas.dot(it->second.f); + kbest_loss += max(0.0, -1.0 * model_diff); + } bool rank_error = false; score_t margin; if (faster_perceptron) { // we only have considering misranked pairs @@ -442,7 +438,7 @@ main(int argc, char** argv) margin = fabs(model_diff); if (!rank_error && margin < loss_margin) margin_violations++; } - if (rank_error) rank_errors++; + if (rank_error && ki==1) rank_errors++; if (scale_bleu_diff) eta = it->first.score - it->second.score; if (rank_error || margin < loss_margin) { SparseVector diff_vec = it->first.f - it->second.f; @@ -524,12 +520,27 @@ main(int argc, char** argv) } } - //if (q==0) { kbest_loss_first = kbest_loss; } - //if (q==repeat-1) { kbest_loss_last = kbest_loss; } -//}//repeat -//if((kbest_loss_first - kbest_loss_last) > 0) did_improve++; + if (ki==0) kbest_loss_first = kbest_loss; + if (ki==repeat-1) { // done + kbest_loss_last = kbest_loss; + score_t best_score = -1.; + score_t best_model = -std::numeric_limits::max(); + unsigned best_idx; + for (unsigned i=0; i < samples->size(); i++) { + score_t s = lambdas.dot((*samples)[i].f); + if (s > best_model) { + best_idx = i; + best_model = s; + } + } + score_sum += (*samples)[best_idx].score; + model_sum += best_model; + } + } // repeat - } + if ((kbest_loss_first - kbest_loss_last) >= 0) kbest_loss_improve++; + + } // noup if (rescale) lambdas /= lambdas.l2norm(); @@ -539,7 +550,6 @@ main(int argc, char** argv) if (t == 0) in_sz = ii; // remember size of input (# lines) - //if (repeat > 1) cout << "did improve? " << did_improve << " out of " << in_sz << endl; if (batch) { lambdas.plus_eq_v_times_s(batch_updates, eta); @@ -577,14 +587,16 @@ main(int argc, char** argv) cerr << _np << " 1best avg model score: " << model_avg; cerr << _p << " (" << model_diff << ")" << endl; cerr << " avg # pairs: "; - cerr << _np << npairs/(float)in_sz; + cerr << _np << npairs/(float)in_sz << endl; + cerr << " avg # margin viol: "; + cerr << margin_violations/(float)in_sz << endl; + cerr << " avg # rank err: "; + cerr << rank_errors/(float)in_sz; if (faster_perceptron) cerr << " (meaningless)"; cerr << endl; - cerr << " avg # rank err: "; - cerr << rank_errors/(float)in_sz << endl; if (batch) cerr << " batch loss: " << batch_loss << endl; - cerr << " avg # margin viol: "; - cerr << margin_violations/(float)in_sz << endl; + if (repeat > 1) cerr << " k-best loss imp: " << ((float)kbest_loss_improve/in_sz)*100 << "%" << endl; + cerr << " non0 feature count: " << nonz << endl; cerr << " avg list sz: " << list_sz/(float)in_sz << endl; cerr << " avg f count: " << f_count/(float)list_sz << endl; diff --git a/training/dtrain/examples/standard/dtrain.ini b/training/dtrain/examples/standard/dtrain.ini index 4d096dfb..ef022469 100644 --- a/training/dtrain/examples/standard/dtrain.ini +++ b/training/dtrain/examples/standard/dtrain.ini @@ -11,11 +11,11 @@ print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 Phr stop_after=10 # stop epoch after 10 inputs # interesting stuff -epochs=100 # run over input 3 times +epochs=3 # run over input 3 times k=100 # use 100best lists N=4 # optimize (approx) BLEU4 scorer=fixed_stupid_bleu # use 'stupid' BLEU+1 -learning_rate=0.0001 # learning rate, don't care if gamma=0 (perceptron) and loss_margin=0 (not margin perceptron) +learning_rate=0.0001 # learning rate, don't care if gamma=0 (perceptron) and loss_margin=0 (not margin perceptron) gamma=0 # use SVM reg sample_from=kbest # use kbest lists (as opposed to forest) filter=uniq # only unique entries in kbest (surface form) @@ -23,3 +23,5 @@ pair_sampling=XYX # hi_lo=0.1 # 10 vs 80 vs 10 and 80 vs 10 here pair_threshold=0 # minimum distance in BLEU (here: > 0) loss_margin=0 # update if correctly ranked, but within this margin +repeat=1 # repeat training on a kbest list 1 times +#batch=true # batch tuning, update after accumulating over all sentences and all kbest lists -- cgit v1.2.3