diff options
author | Patrick Simianer <p@simianer.de> | 2011-09-25 20:23:09 +0200 |
---|---|---|
committer | Patrick Simianer <p@simianer.de> | 2011-09-25 20:23:09 +0200 |
commit | fe471bb707226052551d75b043295ca5f57261c0 (patch) | |
tree | 73ba37bf8d5c1de6de50f63888a49e918e4a8cd4 /dtrain | |
parent | 5e1ab3481551607f1c2a10027049044cd41f78ab (diff) |
removed some quirks, less boost, prettier code, score_t
Diffstat (limited to 'dtrain')
19 files changed, 298 insertions, 313 deletions
diff --git a/dtrain/README b/dtrain/README index 137c1b48..42b91b9b 100644 --- a/dtrain/README +++ b/dtrain/README @@ -1,13 +1,4 @@ -NOTES - learner gets all used features (binary! and dense (logprob is sum of logprobs!)) - weights: see decoder/decoder.cc line 548 - (40k sents, k=100 = ~400M mem, 1 iteration 45min)? - utils/weights.cc: why wv_? - FD, Weights::wv_ grow too large, see utils/weights.cc; - decoder/hg.h; decoder/scfg_translator.cc; utils/fdict.cc - TODO - enable kbest FILTERING (nofiler vs unique) MULTIPARTITE ranking (108010, 1 vs all, cluster modelscore;score) what about RESCORING? REMEMBER kbest (merge) weights? @@ -30,7 +21,7 @@ TODO non deterministic, high variance, RANDOM RESTARTS use separate TEST SET -KNOWN BUGS PROBLEMS +KNOWN BUGS, PROBLEMS doesn't select best iteration for weigts if size of candidate < N => 0 score cdec kbest vs 1best (no -k param), rescoring? => ok(?) diff --git a/dtrain/dtrain.cc b/dtrain/dtrain.cc index 76fdb49c..a70ca2f1 100644 --- a/dtrain/dtrain.cc +++ b/dtrain/dtrain.cc @@ -4,64 +4,66 @@ bool dtrain_init(int argc, char** argv, po::variables_map* cfg) { - po::options_description conff("Configuration File Options"); - conff.add_options() - ("decoder_config", po::value<string>(), "configuration file for cdec") - ("kbest", po::value<size_t>()->default_value(100), "k for kbest") - ("ngrams", po::value<size_t>()->default_value(3), "N for Ngrams") - ("filter", po::value<string>()->default_value("unique"), "filter kbest list") - ("epochs", po::value<size_t>()->default_value(2), "# of iterations T") - ("input", po::value<string>()->default_value("-"), "input file") - ("output", po::value<string>()->default_value("-"), "output weights file") - ("scorer", po::value<string>()->default_value("stupid_bleu"), "scoring metric") - ("stop_after", po::value<size_t>()->default_value(0), "stop after X input sentences") - ("input_weights", po::value<string>(), "input weights file (e.g. from previous iteration)") - ("wprint", po::value<string>(), "weights to print on each iteration") - ("hstreaming", po::value<bool>()->zero_tokens(), "run in hadoop streaming mode") - ("noup", po::value<bool>()->zero_tokens(), "do not update weights"); - - po::options_description clo("Command Line Options"); - clo.add_options() + 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 (or VOID)") + ("input_weights", po::value<string>(), "input weights file (e.g. from previous iteration)") + ("decoder_config", po::value<string>(), "configuration file for cdec") + ("ksamples", po::value<size_t>()->default_value(100), "size of kbest or sample from forest") + ("sample_from", po::value<string>()->default_value("kbest"), "where to get translations from") + ("filter", po::value<string>()->default_value("unique"), "filter kbest list") + ("pair_sampling", po::value<string>()->default_value("all"), "how to sample pairs: all, rand") + ("ngrams", po::value<size_t>()->default_value(3), "N for Ngrams") + ("epochs", po::value<size_t>()->default_value(2), "# of iterations T") + ("scorer", po::value<string>()->default_value("stupid_bleu"), "scoring metric") + ("stop_after", po::value<size_t>()->default_value(0), "stop after X input sentences") + ("print_weights", po::value<string>(), "weights to print on each iteration") + ("hstreaming", po::value<bool>()->zero_tokens(), "run in hadoop streaming mode") + ("learning_rate", po::value<double>()->default_value(0.0005), "learning rate") + ("gamma", po::value<double>()->default_value(0.), "gamma for SVM (0 for perceptron)") + ("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"); - po::options_description config_options, cmdline_options; - - config_options.add(conff); - cmdline_options.add(clo); - cmdline_options.add(conff); - - po::store(parse_command_line(argc, argv, cmdline_options), *cfg); + cl.add(ini); + po::store(parse_command_line(argc, argv, cl), *cfg); if (cfg->count("config")) { - ifstream config((*cfg)["config"].as<string>().c_str()); - po::store(po::parse_config_file(config, config_options), *cfg); + 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 << cmdline_options << endl; + cerr << cl << endl; return false; } if (cfg->count("hstreaming") && (*cfg)["output"].as<string>() != "-") { cerr << "When using 'hstreaming' the 'output' param should be '-'."; return false; } - if (cfg->count("filter") && (*cfg)["filter"].as<string>() != "unique" + if ((*cfg)["filter"].as<string>() != "unique" && (*cfg)["filter"].as<string>() != "no") { - cerr << "Wrong 'filter' type: '" << (*cfg)["filter"].as<string>() << "'." << endl; + cerr << "Wrong 'filter' param: '" << (*cfg)["filter"].as<string>() << "', use 'unique' or 'no'." << endl; + } + if ((*cfg)["pair_sampling"].as<string>() != "all" + && (*cfg)["pair_sampling"].as<string>() != "rand") { + cerr << "Wrong 'pair_sampling' param: '" << (*cfg)["pair_sampling"].as<string>() << "', use 'all' or 'rand'." << endl; + } + 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 true; } -#include "filelib.h" - int main(int argc, char** argv) { - cout << _p5; // handle most parameters po::variables_map cfg; - if (! dtrain_init(argc, argv, &cfg)) exit(1); // something is wrong + if (!dtrain_init(argc, argv, &cfg)) exit(1); // something is wrong bool quiet = false; if (cfg.count("quiet")) quiet = true; bool verbose = false; @@ -73,43 +75,37 @@ main(int argc, char** argv) hstreaming = true; quiet = true; } - const size_t k = cfg["kbest"].as<size_t>(); + const size_t k = cfg["ksamples"].as<size_t>(); const size_t N = cfg["ngrams"].as<size_t>(); const size_t T = cfg["epochs"].as<size_t>(); const size_t stop_after = cfg["stop_after"].as<size_t>(); const string filter_type = cfg["filter"].as<string>(); - if (!quiet) { - cout << endl << "dtrain" << endl << "Parameters:" << endl; - cout << setw(25) << "k " << k << endl; - cout << setw(25) << "N " << N << endl; - cout << setw(25) << "T " << T << endl; - if (cfg.count("stop-after")) - cout << setw(25) << "stop_after " << stop_after << endl; - if (cfg.count("input_weights")) - cout << setw(25) << "weights " << cfg["weights"].as<string>() << endl; - cout << setw(25) << "input " << "'" << cfg["input"].as<string>() << "'" << endl; - cout << setw(25) << "filter " << "'" << filter_type << "'" << endl; - } - - vector<string> wprint; - if (cfg.count("wprint")) { - boost::split(wprint, cfg["wprint"].as<string>(), boost::is_any_of(" ")); - } + const string sample_from = cfg["sample_from"].as<string>(); + const string pair_sampling = cfg["pair_sampling"].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, observer + // setup decoder register_feature_functions(); SetSilent(true); ReadFile ini_rf(cfg["decoder_config"].as<string>()); if (!quiet) cout << setw(25) << "cdec cfg " << "'" << cfg["decoder_config"].as<string>() << "'" << endl; Decoder decoder(ini_rf.stream()); - KBestGetter observer(k, filter_type); - MT19937 rng; - //KSampler observer(k, &rng); + + MT19937 rng; // random number generator + // setup decoder observer + HypoSampler* observer; + if (sample_from == "kbest") { + observer = dynamic_cast<KBestGetter*>(new KBestGetter(k, filter_type)); + } else { + observer = dynamic_cast<KSampler*>(new KSampler(k, &rng)); + } // scoring metric/scorer string scorer_str = cfg["scorer"].as<string>(); - double (*scorer)(NgramCounts&, const size_t, const size_t, size_t, vector<float>); + score_t (*scorer)(NgramCounts&, const size_t, const size_t, size_t, vector<score_t>); if (scorer_str == "bleu") { scorer = &bleu; } else if (scorer_str == "stupid_bleu") { @@ -122,58 +118,64 @@ main(int argc, char** argv) cerr << "Don't know scoring metric: '" << scorer_str << "', exiting." << endl; exit(1); } - // for approx_bleu NgramCounts global_counts(N); // counts for 1 best translations - size_t global_hyp_len = 0; // sum hypothesis lengths - size_t global_ref_len = 0; // sum reference lengths - // this is all BLEU implmentations - vector<float> bleu_weights; // we leave this empty -> 1/N; TODO? + size_t global_hyp_len = 0; // sum hypothesis lengths + size_t global_ref_len = 0; // sum reference lengths + // ^^^ global_* for approx_bleu + vector<score_t> bleu_weights; // we leave this empty -> 1/N if (!quiet) cout << setw(26) << "scorer '" << scorer_str << "'" << endl << endl; // init weights Weights weights; - if (cfg.count("weights")) weights.InitFromFile(cfg["weights"].as<string>()); + if (cfg.count("input_weights")) weights.InitFromFile(cfg["input_weights"].as<string>()); SparseVector<double> lambdas; weights.InitSparseVector(&lambdas); vector<double> dense_weights; + // meta params for perceptron, SVM + double eta = cfg["learning_rate"].as<double>(); + double gamma = cfg["gamma"].as<double>(); + lambdas.add_value(FD::Convert("__bias"), 0); + // input - if (!quiet && !verbose) - cout << "(a dot represents " << DTRAIN_DOTS << " lines of input)" << endl; string input_fn = cfg["input"].as<string>(); - ifstream input; - if (input_fn != "-") input.open(input_fn.c_str()); - string in; - vector<string> in_split; // input: src\tref\tpsg - vector<string> ref_tok; // tokenized reference - vector<WordID> ref_ids; // reference as vector of WordID - - // buffer input for t > 0 - vector<string> src_str_buf; // source strings, TODO? memory - vector<vector<WordID> > ref_ids_buf; // references as WordID vecs + ReadFile input(input_fn); + // buffer input for t > 0 + vector<string> src_str_buf; // source strings + vector<vector<WordID> > ref_ids_buf; // references as WordID vecs // this is for writing the grammar buffer file char grammar_buf_fn[] = DTRAIN_TMP_DIR"/dtrain-grammars-XXXXXX"; mkstemp(grammar_buf_fn); ogzstream grammar_buf_out; grammar_buf_out.open(grammar_buf_fn); - size_t sid = 0, in_sz = 99999999; // sentence id, input size - double acc_1best_score = 0., acc_1best_model = 0.; - vector<vector<double> > scores_per_iter; - double max_score = 0.; - size_t best_t = 0; - bool next = false, stop = false; - double score = 0.; - size_t cand_len = 0; - double overall_time = 0.; - - // for the perceptron/SVM; TODO as params - double eta = 0.0005; - double gamma = 0.;//01; // -> SVM - lambdas.add_value(FD::Convert("__bias"), 0); - - // for random sampling - srand (time(NULL)); + size_t in_sz = 999999999; // input index, input size + vector<pair<score_t,score_t> > all_scores; + score_t max_score = 0.; + size_t best_it = 0; + float overall_time = 0.; + + // output cfg + if (!quiet) { + cout << _p5; + cout << endl << "dtrain" << endl << "Parameters:" << endl; + cout << setw(25) << "k " << k << endl; + cout << setw(25) << "N " << N << endl; + cout << setw(25) << "T " << T << endl; + if (cfg.count("stop-after")) + cout << setw(25) << "stop_after " << stop_after << endl; + if (cfg.count("input_weights")) + cout << setw(25) << "weights in" << cfg["input_weights"].as<string>() << endl; + cout << setw(25) << "input " << "'" << cfg["input"].as<string>() << "'" << endl; + cout << setw(25) << "output " << "'" << cfg["output"].as<string>() << "'" << endl; + if (sample_from == "kbest") + cout << setw(25) << "filter " << "'" << filter_type << "'" << endl; + cout << setw(25) << "learning rate " << eta << endl; + cout << setw(25) << "gamma " << gamma << endl; + cout << setw(25) << "sample from " << "'" << sample_from << "'" << endl; + cout << setw(25) << "pairs " << "'" << pair_sampling << "'" << endl; + if (!verbose) cout << "(a dot represents " << DTRAIN_DOTS << " lines of input)" << endl; + } for (size_t t = 0; t < T; t++) // T epochs @@ -181,58 +183,44 @@ main(int argc, char** argv) time_t start, end; time(&start); - - // actually, we need only need this if t > 0 FIXME igzstream grammar_buf_in; if (t > 0) grammar_buf_in.open(grammar_buf_fn); - - // reset average scores - acc_1best_score = acc_1best_model = 0.; - - // reset sentence counter - sid = 0; - + score_t score_sum = 0., model_sum = 0.; + size_t ii = 0; if (!quiet) cout << "Iteration #" << t+1 << " of " << T << "." << endl; while(true) { - // get input from stdin or file - in.clear(); - next = stop = false; // next iteration, premature stop - if (t == 0) { - if (input_fn == "-") { - if (!getline(cin, in)) next = true; - } else { - if (!getline(input, in)) next = true; - } + string in; + bool next = false, stop = false; // next iteration or premature stop + if (t == 0) { + if(!getline(*input, in)) next = true; } else { - if (sid == in_sz) next = true; // stop if we reach the end of our input + 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 == sid && !next) stop = true; + if (stop_after > 0 && stop_after == ii && !next) stop = true; // produce some pretty output if (!quiet && !verbose) { - if (sid == 0) cout << " "; - if ((sid+1) % (DTRAIN_DOTS) == 0) { - cout << "."; - cout.flush(); - } - if ((sid+1) % (20*DTRAIN_DOTS) == 0) { - cout << " " << sid+1 << endl; - if (!next && !stop) cout << " "; - } - if (stop) { - if (sid % (20*DTRAIN_DOTS) != 0) cout << " " << sid << endl; - cout << "Stopping after " << stop_after << " input sentences." << endl; - } else { - if (next) { - if (sid % (20*DTRAIN_DOTS) != 0) { - cout << " " << sid << endl; - } - } + if (ii == 0) cout << " "; + if ((ii+1) % (DTRAIN_DOTS) == 0) { + cout << "."; + cout.flush(); + } + if ((ii+1) % (20*DTRAIN_DOTS) == 0) { + cout << " " << ii+1 << endl; + if (!next && !stop) cout << " "; + } + if (stop) { + if (ii % (20*DTRAIN_DOTS) != 0) cout << " " << ii << endl; + cout << "Stopping after " << stop_after << " input sentences." << endl; + } else { + if (next) { + if (ii % (20*DTRAIN_DOTS) != 0) cout << " " << ii << endl; } + } } // next iteration @@ -244,12 +232,15 @@ main(int argc, char** argv) weights.InitVector(&dense_weights); decoder.SetWeights(dense_weights); + // getting input + vector<string> in_split; // input: sid\tsrc\tref\tpsg + vector<WordID> ref_ids; // reference as vector<WordID> if (t == 0) { // handling input - in_split.clear(); strsplit(in, in_split, '\t', 4); // getting reference - ref_tok.clear(); ref_ids.clear(); + ref_ids.clear(); + vector<string> ref_tok; strsplit(in_split[2], ref_tok, ' '); register_and_convert(ref_tok, ref_ids); ref_ids_buf.push_back(ref_ids); @@ -268,7 +259,7 @@ main(int argc, char** argv) decoder.SetSentenceGrammarFromString(in_split[3]); // decode src_str_buf.push_back(in_split[1]); - decoder.Decode(in_split[1], &observer); + decoder.Decode(in_split[1], observer); } else { // get buffered grammar string grammar_str; @@ -280,73 +271,67 @@ main(int argc, char** argv) } decoder.SetSentenceGrammarFromString(grammar_str); // decode - decoder.Decode(src_str_buf[sid], &observer); + decoder.Decode(src_str_buf[ii], observer); } - // get kbest list - KBestList* kb; - //if () { // TODO get from forest - kb = observer.GetKBest(); - //} + Samples* samples = observer->GetSamples(); // (local) scoring - if (t > 0) ref_ids = ref_ids_buf[sid]; - for (size_t i = 0; i < kb->GetSize(); i++) { - NgramCounts counts = make_ngram_counts(ref_ids, kb->sents[i], N); + if (t > 0) ref_ids = ref_ids_buf[ii]; + score_t score = 0.; + for (size_t i = 0; i < samples->GetSize(); i++) { + NgramCounts counts = make_ngram_counts(ref_ids, samples->sents[i], N); if (scorer_str == "approx_bleu") { + size_t hyp_len = 0; if (i == 0) { // 'context of 1best translations' global_counts += counts; - global_hyp_len += kb->sents[i].size(); + global_hyp_len += samples->sents[i].size(); global_ref_len += ref_ids.size(); counts.reset(); - cand_len = 0; } else { - cand_len = kb->sents[i].size(); + hyp_len = samples->sents[i].size(); } NgramCounts counts_tmp = global_counts + counts; - score = .9*scorer(counts_tmp, - global_ref_len, - global_hyp_len + cand_len, N, bleu_weights); + score = .9 * scorer(counts_tmp, + global_ref_len, + global_hyp_len + hyp_len, N, bleu_weights); } else { - cand_len = kb->sents[i].size(); score = scorer(counts, - ref_ids.size(), - kb->sents[i].size(), N, bleu_weights); + ref_ids.size(), + samples->sents[i].size(), N, bleu_weights); } - kb->scores.push_back(score); + samples->scores.push_back(score); if (i == 0) { - acc_1best_score += score; - acc_1best_model += kb->model_scores[i]; + score_sum += score; + model_sum += samples->model_scores[i]; } if (verbose) { if (i == 0) cout << "'" << TD::GetString(ref_ids) << "' [ref]" << endl; - cout << _p5 << _np << "[hyp " << i << "] " << "'" << TD::GetString(kb->sents[i]) << "'"; - cout << " [SCORE=" << score << ",model="<< kb->model_scores[i] << "]" << endl; - //cout << kb->feats[i] << endl; // too verbose + cout << _p5 << _np << "[hyp " << i << "] " << "'" << TD::GetString(samples->sents[i]) << "'"; + cout << " [SCORE=" << score << ",model="<< samples->model_scores[i] << "]" << endl; + cout << samples->feats[i] << endl; } - } // Nbest loop + } // sample/scoring loop if (verbose) cout << endl; ////////////////////////////////////////////////////////// // UPDATE WEIGHTS if (!noup) { - - int up = 0; - - TrainingInstances pairs; - sample_all_pairs(kb, pairs); - //sample_rand_pairs(kb, pairs, &rng); + vector<Pair> pairs; + if (pair_sampling == "all") + sample_all_pairs(samples, pairs); + if (pair_sampling == "rand") + sample_rand_pairs(samples, pairs, &rng); - for (TrainingInstances::iterator ti = pairs.begin(); + for (vector<Pair>::iterator ti = pairs.begin(); ti != pairs.end(); ti++) { SparseVector<double> dv; if (ti->first_score - ti->second_score < 0) { - up++; dv = ti->second - ti->first; //} else { //dv = ti->first - ti->second; @@ -360,10 +345,10 @@ main(int argc, char** argv) if (verbose) { cout << "{{ f("<< ti->first_rank <<") > f(" << ti->second_rank << ") but g(i)="<< ti->first_score <<" < g(j)="<< ti->second_score << " so update" << endl; - cout << " i " << TD::GetString(kb->sents[ti->first_rank]) << endl; - cout << " " << kb->feats[ti->first_rank] << endl; - cout << " j " << TD::GetString(kb->sents[ti->second_rank]) << endl; - cout << " " << kb->feats[ti->second_rank] << endl; + cout << " i " << TD::GetString(samples->sents[ti->first_rank]) << endl; + cout << " " << samples->feats[ti->first_rank] << endl; + cout << " j " << TD::GetString(samples->sents[ti->second_rank]) << endl; + cout << " " << samples->feats[ti->second_rank] << endl; cout << " diff vec: " << dv << endl; cout << " lambdas after update: " << lambdas << endl; cout << "}}" << endl; @@ -378,69 +363,66 @@ main(int argc, char** argv) //double l2 = lambdas.l2norm(); //if (l2) lambdas /= lambdas.l2norm(); - //cout << up << endl; } ////////////////////////////////////////////////////////// - ++sid; + ++ii; - if (hstreaming) cerr << "reporter:counter:dtrain,sid," << sid << endl; + if (hstreaming) cerr << "reporter:counter:dtrain,sid," << in_split[0] << endl; } // input loop if (t == 0) { - in_sz = sid; // remember size (lines) of input + in_sz = ii; // remember size of input (# lines) grammar_buf_out.close(); - if (input_fn != "-") input.close(); } else { grammar_buf_in.close(); } // print some stats - double avg_1best_score = acc_1best_score/(double)in_sz; - double avg_1best_model = acc_1best_model/(double)in_sz; - double avg_1best_score_diff, avg_1best_model_diff; + 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) { - avg_1best_score_diff = avg_1best_score - scores_per_iter[t-1][0]; - avg_1best_model_diff = avg_1best_model - scores_per_iter[t-1][1]; + score_diff = score_avg - all_scores[t-1].first; + model_diff = model_avg - all_scores[t-1].second; } else { - avg_1best_score_diff = avg_1best_score; - avg_1best_model_diff = avg_1best_model; + score_diff = score_avg; + model_diff = model_avg; } if (!quiet) { cout << _p5 << _p << "WEIGHTS" << endl; - for (vector<string>::iterator it = wprint.begin(); it != wprint.end(); it++) { - cout << setw(16) << *it << " = " << dense_weights[FD::Convert(*it)] << endl; + for (vector<string>::iterator it = print_weights.begin(); it != print_weights.end(); it++) { + cout << setw(16) << *it << " = " << lambdas.get(FD::Convert(*it)) << endl; } cout << " ---" << endl; - cout << _np << " avg score: " << avg_1best_score; - cout << _p << " (" << avg_1best_score_diff << ")" << endl; - cout << _np << "avg model score: " << avg_1best_model; - cout << _p << " (" << avg_1best_model_diff << ")" << endl; + cout << _np << " 1best avg score: " << score_avg; + cout << _p << " (" << score_diff << ")" << endl; + cout << _np << "1best avg model score: " << model_avg; + cout << _p << " (" << model_diff << ")" << endl; } - vector<double> remember_scores; - remember_scores.push_back(avg_1best_score); - remember_scores.push_back(avg_1best_model); - scores_per_iter.push_back(remember_scores); - if (avg_1best_score > max_score) { - max_score = avg_1best_score; - best_t = t; + 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); - double time_dif = difftime(end, start); - overall_time += time_dif; + float time_diff = difftime(end, start); + overall_time += time_diff; if (!quiet) { - cout << _p2 << _np << "(time " << time_dif/60. << " min, "; - cout << time_dif/(double)in_sz<< " s/S)" << endl; + cout << _p2 << _np << "(time " << time_diff/60. << " min, "; + cout << time_diff/(float)in_sz<< " s/S)" << endl; } - if (t+1 != T && !quiet) cout << endl; if (noup) break; } // outer loop - //unlink(grammar_buf_fn); + unlink(grammar_buf_fn); if (!noup) { if (!quiet) cout << endl << "writing weights file '" << cfg["output"].as<string>() << "' ..."; @@ -452,7 +434,7 @@ main(int argc, char** argv) cout << _np << FD::Convert(ti->first) << "\t" << ti->second << endl; } if (hstreaming) cout << "__SHARD_COUNT__\t1" << endl; - } else { + } else if (cfg["output"].as<string>() != "VOID") { weights.InitFromVector(lambdas); weights.WriteToFile(cfg["output"].as<string>(), true); } @@ -461,7 +443,7 @@ main(int argc, char** argv) if (!quiet) { cout << _p5 << _np << endl << "---" << endl << "Best iteration: "; - cout << best_t+1 << " [SCORE '" << scorer_str << "'=" << max_score << "]." << endl; + cout << best_it+1 << " [SCORE '" << scorer_str << "'=" << max_score << "]." << endl; cout << _p2 << "This took " << overall_time/60. << " min." << endl; } diff --git a/dtrain/kbestget.h b/dtrain/kbestget.h index cf466fe4..79201182 100644 --- a/dtrain/kbestget.h +++ b/dtrain/kbestget.h @@ -1,19 +1,14 @@ #ifndef _DTRAIN_KBESTGET_H_ #define _DTRAIN_KBESTGET_H_ - #include "kbest.h" - namespace dtrain { -/* - * KBestList - * - */ -struct KBestList { +struct Samples +{ vector<SparseVector<double> > feats; vector<vector<WordID> > sents; vector<double> model_scores; @@ -21,71 +16,71 @@ struct KBestList { size_t GetSize() { return sents.size(); } }; +struct HypoSampler : public DecoderObserver +{ + virtual Samples* GetSamples() {} +}; -/* - * KBestGetter - * - */ -struct KBestGetter : public DecoderObserver +struct KBestGetter : public HypoSampler { const size_t k_; const string filter_type; - KBestList kb; + Samples s; - KBestGetter( const size_t k, const string filter_type ) : + KBestGetter(const size_t k, const string filter_type) : k_(k), filter_type(filter_type) {} virtual void - NotifyTranslationForest( const SentenceMetadata& smeta, Hypergraph* hg ) + NotifyTranslationForest(const SentenceMetadata& smeta, Hypergraph* hg) { - KBest( *hg ); + KBest(*hg); } - KBestList* GetKBest() { return &kb; } + Samples* GetSamples() { return &s; } void - KBest( const Hypergraph& forest ) + KBest(const Hypergraph& forest) { - if ( filter_type == "unique" ) { - KBestUnique( forest ); - } else if ( filter_type == "no" ) { - KBestNoFilter( forest ); + if (filter_type == "unique") { + KBestUnique(forest); + } else if (filter_type == "no") { + KBestNoFilter(forest); } } void - KBestUnique( const Hypergraph& forest ) + KBestUnique(const Hypergraph& forest) { - kb.sents.clear(); - kb.feats.clear(); - kb.model_scores.clear(); - kb.scores.clear(); - KBest::KBestDerivations<vector<WordID>, ESentenceTraversal, KBest::FilterUnique, prob_t, EdgeProb> kbest( forest, k_ ); - for ( size_t i = 0; i < k_; ++i ) { + s.sents.clear(); + s.feats.clear(); + s.model_scores.clear(); + s.scores.clear(); + KBest::KBestDerivations<vector<WordID>, ESentenceTraversal, KBest::FilterUnique, prob_t, EdgeProb> kbest(forest, k_); + for (size_t i = 0; i < k_; ++i) { const KBest::KBestDerivations<vector<WordID>, ESentenceTraversal, KBest::FilterUnique, prob_t, EdgeProb>::Derivation* d = - kbest.LazyKthBest( forest.nodes_.size() - 1, i ); + kbest.LazyKthBest(forest.nodes_.size() - 1, i); if (!d) break; - kb.sents.push_back( d->yield); - kb.feats.push_back( d->feature_values ); - kb.model_scores.push_back( log(d->score) ); + s.sents.push_back(d->yield); + s.feats.push_back(d->feature_values); + s.model_scores.push_back(log(d->score)); } } void - KBestNoFilter( const Hypergraph& forest ) + KBestNoFilter(const Hypergraph& forest) { - kb.sents.clear(); - kb.feats.clear(); - kb.model_scores.clear(); - kb.scores.clear(); - KBest::KBestDerivations<vector<WordID>, ESentenceTraversal> kbest( forest, k_ ); - for ( size_t i = 0; i < k_; ++i ) { + s.sents.clear(); + s.feats.clear(); + s.model_scores.clear(); + s.scores.clear(); + KBest::KBestDerivations<vector<WordID>, ESentenceTraversal> kbest(forest, k_); + for (size_t i = 0; i < k_; ++i) { const KBest::KBestDerivations<vector<WordID>, ESentenceTraversal>::Derivation* d = - kbest.LazyKthBest( forest.nodes_.size() - 1, i ); + kbest.LazyKthBest(forest.nodes_.size() - 1, i); if (!d) break; - kb.sents.push_back( d->yield); - kb.feats.push_back( d->feature_values ); - kb.model_scores.push_back( log(d->score) ); + s.sents.push_back(d->yield); + s.feats.push_back(d->feature_values); + s.model_scores.push_back(log(d->score)); } } }; @@ -93,6 +88,5 @@ struct KBestGetter : public DecoderObserver } // namespace - #endif diff --git a/dtrain/ksampler.h b/dtrain/ksampler.h index 914e9723..ac88b643 100644 --- a/dtrain/ksampler.h +++ b/dtrain/ksampler.h @@ -1,21 +1,22 @@ #ifndef _DTRAIN_KSAMPLER_H_ #define _DTRAIN_KSAMPLER_H_ -#include "kbest.h" #include "hgsampler.h" +#include "kbest.h" // cdec #include "sampler.h" namespace dtrain { + /* * KSampler * */ -struct KSampler : public DecoderObserver +struct KSampler : public HypoSampler { const size_t k_; - KBestList kb; + Samples s; MT19937* rng; explicit KSampler( const size_t k, MT19937* prng ) : @@ -27,19 +28,19 @@ struct KSampler : public DecoderObserver Sample( *hg ); } - KBestList* GetKBest() { return &kb; } + Samples* GetSamples() { return &s; } void Sample( const Hypergraph& forest ) { - kb.sents.clear(); - kb.feats.clear(); - kb.model_scores.clear(); - kb.scores.clear(); + s.sents.clear(); + s.feats.clear(); + s.model_scores.clear(); + s.scores.clear(); std::vector<HypergraphSampler::Hypothesis> samples; HypergraphSampler::sample_hypotheses(forest, k_, rng, &samples); for ( size_t i = 0; i < k_; ++i ) { - kb.sents.push_back( samples[i].words ); - kb.feats.push_back( samples[i].fmap ); - kb.model_scores.push_back( log(samples[i].model_score) ); + s.sents.push_back( samples[i].words ); + s.feats.push_back( samples[i].fmap ); + s.model_scores.push_back( log(samples[i].model_score) ); } } }; @@ -47,6 +48,5 @@ struct KSampler : public DecoderObserver } // namespace - #endif diff --git a/dtrain/pairsampling.h b/dtrain/pairsampling.h index e06036ca..a8521485 100644 --- a/dtrain/pairsampling.h +++ b/dtrain/pairsampling.h @@ -8,21 +8,19 @@ namespace dtrain { -struct TPair +struct Pair { SparseVector<double> first, second; size_t first_rank, second_rank; double first_score, second_score; }; -typedef vector<TPair> TrainingInstances; - inline void -sample_all_pairs(KBestList* kb, TrainingInstances &training) +sample_all_pairs(Samples* kb, vector<Pair> &training) { for (size_t i = 0; i < kb->GetSize()-1; i++) { for (size_t j = i+1; j < kb->GetSize(); j++) { - TPair p; + Pair p; p.first = kb->feats[i]; p.second = kb->feats[j]; p.first_rank = i; @@ -35,12 +33,12 @@ sample_all_pairs(KBestList* kb, TrainingInstances &training) } inline void -sample_rand_pairs(KBestList* kb, TrainingInstances &training, MT19937* prng) +sample_rand_pairs(Samples* kb, vector<Pair> &training, MT19937* prng) { for (size_t i = 0; i < kb->GetSize()-1; i++) { for (size_t j = i+1; j < kb->GetSize(); j++) { if (prng->next() < .5) { - TPair p; + Pair p; p.first = kb->feats[i]; p.second = kb->feats[j]; p.first_rank = i; diff --git a/dtrain/score.cc b/dtrain/score.cc index d08e87f3..c6d3a05f 100644 --- a/dtrain/score.cc +++ b/dtrain/score.cc @@ -47,27 +47,27 @@ make_ngram_counts(vector<WordID> hyp, vector<WordID> ref, size_t N) * * NOTE: 0 if one n in {1..N} has 0 count */ -double +score_t brevity_penaly(const size_t hyp_len, const size_t ref_len) { if (hyp_len > ref_len) return 1; - return exp(1 - (double)ref_len/(double)hyp_len); + return exp(1 - (score_t)ref_len/(score_t)hyp_len); } -double +score_t bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, - size_t N, vector<float> weights ) + size_t N, vector<score_t> weights ) { if (hyp_len == 0 || ref_len == 0) return 0; if (ref_len < N) N = ref_len; - float N_ = (float)N; + score_t N_ = (score_t)N; if (weights.empty()) { for (size_t i = 0; i < N; i++) weights.push_back(1/N_); } - double sum = 0; + score_t sum = 0; for (size_t i = 0; i < N; i++) { if (counts.clipped[i] == 0 || counts.sum[i] == 0) return 0; - sum += weights[i] * log((double)counts.clipped[i] / (double)counts.sum[i]); + sum += weights[i] * log((score_t)counts.clipped[i] / (score_t)counts.sum[i]); } return brevity_penaly(hyp_len, ref_len) * exp(sum); } @@ -82,22 +82,22 @@ bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, * * NOTE: 0 iff no 1gram match */ -double +score_t stupid_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, - size_t N, vector<float> weights ) + size_t N, vector<score_t> weights ) { if (hyp_len == 0 || ref_len == 0) return 0; if (ref_len < N) N = ref_len; - float N_ = (float)N; + score_t N_ = (score_t)N; if (weights.empty()) { for (size_t i = 0; i < N; i++) weights.push_back(1/N_); } - double sum = 0; - float add = 0; + score_t sum = 0; + score_t add = 0; for (size_t i = 0; i < N; i++) { if (i == 1) add = 1; - sum += weights[i] * log(((double)counts.clipped[i] + add) / ((double)counts.sum[i] + add)); + sum += weights[i] * log(((score_t)counts.clipped[i] + add) / ((score_t)counts.sum[i] + add)); } return brevity_penaly(hyp_len, ref_len) * exp(sum); } @@ -111,21 +111,21 @@ stupid_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, * * NOTE: max is 0.9375 */ -double +score_t smooth_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, - const size_t N, vector<float> weights ) + const size_t N, vector<score_t> weights ) { if (hyp_len == 0 || ref_len == 0) return 0; - float N_ = (float)N; + score_t N_ = (score_t)N; if (weights.empty()) { for (size_t i = 0; i < N; i++) weights.push_back(1/N_); } - double sum = 0; - float j = 1; + score_t sum = 0; + score_t j = 1; for (size_t i = 0; i < N; i++) { if (counts.clipped[i] == 0 || counts.sum[i] == 0) continue; - sum += exp((weights[i] * log((double)counts.clipped[i]/(double)counts.sum[i]))) / pow(2, N_-j+1); + sum += exp((weights[i] * log((score_t)counts.clipped[i]/(score_t)counts.sum[i]))) / pow(2, N_-j+1); j++; } return brevity_penaly(hyp_len, ref_len) * sum; @@ -138,9 +138,9 @@ smooth_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, * and Structural Translation Features" * (Chiang et al. '08) */ -double +score_t approx_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, - const size_t N, vector<float> weights) + const size_t N, vector<score_t> weights) { return brevity_penaly(hyp_len, ref_len) * 0.9 * bleu(counts, hyp_len, ref_len, N, weights); diff --git a/dtrain/score.h b/dtrain/score.h index 0afb6237..bff0b10c 100644 --- a/dtrain/score.h +++ b/dtrain/score.h @@ -15,15 +15,16 @@ namespace dtrain { +typedef double score_t; // float + struct NgramCounts { - NgramCounts(const size_t N) : N_(N) { - reset(); - } size_t N_; map<size_t, size_t> clipped; map<size_t, size_t> sum; + NgramCounts(const size_t N) : N_(N) { reset(); } + void operator+=(const NgramCounts& rhs) { @@ -76,22 +77,22 @@ struct NgramCounts }; typedef map<vector<WordID>, size_t> Ngrams; + Ngrams make_ngrams(vector<WordID>& s, size_t N); NgramCounts make_ngram_counts(vector<WordID> hyp, vector<WordID> ref, size_t N); -double brevity_penaly(const size_t hyp_len, const size_t ref_len); -double bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N, - vector<float> weights = vector<float>()); -double stupid_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, size_t N, - vector<float> weights = vector<float>()); -double smooth_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N, - vector<float> weights = vector<float>()); -double approx_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N, - vector<float> weights = vector<float>()); +score_t brevity_penaly(const size_t hyp_len, const size_t ref_len); +score_t bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N, + vector<score_t> weights = vector<score_t>()); +score_t stupid_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, size_t N, + vector<score_t> weights = vector<score_t>()); +score_t smooth_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N, + vector<score_t> weights = vector<score_t>()); +score_t approx_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len, const size_t N, + vector<score_t> weights = vector<score_t>()); } // namespace - #endif diff --git a/dtrain/test/cdec_toy/cdec.ini b/dtrain/test/cdec_toy/cdec.ini index 3a6bab68..9eb34512 100644 --- a/dtrain/test/cdec_toy/cdec.ini +++ b/dtrain/test/cdec_toy/cdec.ini @@ -1,3 +1,4 @@ formalism=scfg grammar=../dtrain/test/toy_cdec/grammar add_pass_through_rules=true +weights=../dtrain/test/toy_cdec/weights diff --git a/dtrain/test/example/cdec.ini b/dtrain/test/example/cdec.ini new file mode 100644 index 00000000..cdc8a8bb --- /dev/null +++ b/dtrain/test/example/cdec.ini @@ -0,0 +1,7 @@ +formalism=scfg +add_pass_through_rules=true +cubepruning_pop_limit=30 +scfg_max_span_limit=15 +feature_function=WordPenalty +feature_function=KLanguageModel /home/pks/z/X/x/dtrain/test/example/nc-wmt11.en.srilm.gz +#feature_function=RuleIdentityFeatures diff --git a/dtrain/test/example/dtrain.ini b/dtrain/test/example/dtrain.ini new file mode 100644 index 00000000..aee3c89e --- /dev/null +++ b/dtrain/test/example/dtrain.ini @@ -0,0 +1,11 @@ +decoder_config=test/example/cdec.ini +ksamples=100 +ngrams=3 +epochs=1000 +input=test/example/nc-1k.gz +scorer=stupid_bleu +output=test/example/weights.gz +stop_after=10 +sample_from=kbest +pair_sampling=all +print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 PhraseModel_1 PhraseModel_2 PhraseModel_3 PhraseModel_4 PassThrough diff --git a/dtrain/test/example/nc-1k-tabs.gz b/dtrain/test/example/nc-1k-tabs.gz Binary files differnew file mode 100644 index 00000000..45496cd8 --- /dev/null +++ b/dtrain/test/example/nc-1k-tabs.gz diff --git a/dtrain/test/example/nc-1k.gz b/dtrain/test/example/nc-1k.gz Binary files differnew file mode 100644 index 00000000..f638a166 --- /dev/null +++ b/dtrain/test/example/nc-1k.gz diff --git a/dtrain/test/example/nc-wmt11.en.srilm.gz b/dtrain/test/example/nc-wmt11.en.srilm.gz Binary files differnew file mode 100644 index 00000000..7ce81057 --- /dev/null +++ b/dtrain/test/example/nc-wmt11.en.srilm.gz diff --git a/dtrain/test/example/weights.gz b/dtrain/test/example/weights.gz Binary files differnew file mode 100644 index 00000000..21157427 --- /dev/null +++ b/dtrain/test/example/weights.gz diff --git a/dtrain/test/logreg/bin_class.cc b/dtrain/test/log_reg_dyer/bin_class.cc index 19bcde25..19bcde25 100644 --- a/dtrain/test/logreg/bin_class.cc +++ b/dtrain/test/log_reg_dyer/bin_class.cc diff --git a/dtrain/test/logreg/bin_class.h b/dtrain/test/log_reg_dyer/bin_class.h index 3466109a..3466109a 100644 --- a/dtrain/test/logreg/bin_class.h +++ b/dtrain/test/log_reg_dyer/bin_class.h diff --git a/dtrain/test/logreg/log_reg.cc b/dtrain/test/log_reg_dyer/log_reg.cc index ec2331fe..ec2331fe 100644 --- a/dtrain/test/logreg/log_reg.cc +++ b/dtrain/test/log_reg_dyer/log_reg.cc diff --git a/dtrain/test/logreg/log_reg.h b/dtrain/test/log_reg_dyer/log_reg.h index ecc560b8..ecc560b8 100644 --- a/dtrain/test/logreg/log_reg.h +++ b/dtrain/test/log_reg_dyer/log_reg.h diff --git a/dtrain/test/toy_example/dtrain.ini b/dtrain/test/toy_example/dtrain.ini index 0cc222e1..3ab4f8d4 100644 --- a/dtrain/test/toy_example/dtrain.ini +++ b/dtrain/test/toy_example/dtrain.ini @@ -1,5 +1,5 @@ decoder_config=test/toy_example/cdec.ini -kbest=4 +ksamples=4 ngrams=3 epochs=2 input=test/toy_example/toy.in |