diff options
Diffstat (limited to 'training')
| -rw-r--r-- | training/Makefile.am | 8 | ||||
| -rw-r--r-- | training/augment_grammar.cc | 4 | ||||
| -rw-r--r-- | training/collapse_weights.cc | 6 | ||||
| -rw-r--r-- | training/compute_cllh.cc | 23 | ||||
| -rw-r--r-- | training/grammar_convert.cc | 8 | ||||
| -rw-r--r-- | training/mpi_batch_optimize.cc | 127 | ||||
| -rw-r--r-- | training/mpi_online_optimize.cc | 69 | ||||
| -rw-r--r-- | training/mr_optimize_reduce.cc | 19 |
8 files changed, 72 insertions, 192 deletions
diff --git a/training/Makefile.am b/training/Makefile.am index e075e417..6e2c06f5 100644 --- a/training/Makefile.am +++ b/training/Makefile.am @@ -12,9 +12,7 @@ bin_PROGRAMS = \ cllh_filter_grammar \ mpi_online_optimize \ mpi_batch_optimize \ - mpi_em_optimize \ compute_cllh \ - feature_expectations \ augment_grammar noinst_PROGRAMS = \ @@ -29,12 +27,6 @@ mpi_online_optimize_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval mpi_batch_optimize_SOURCES = mpi_batch_optimize.cc optimize.cc mpi_batch_optimize_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a ../klm/lm/libklm.a ../klm/util/libklm_util.a -lz -feature_expectations_SOURCES = feature_expectations.cc -feature_expectations_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a ../klm/lm/libklm.a ../klm/util/libklm_util.a -lz - -mpi_em_optimize_SOURCES = mpi_em_optimize.cc optimize.cc -mpi_em_optimize_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a ../klm/lm/libklm.a ../klm/util/libklm_util.a -lz - compute_cllh_SOURCES = compute_cllh.cc compute_cllh_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a ../klm/lm/libklm.a ../klm/util/libklm_util.a -lz diff --git a/training/augment_grammar.cc b/training/augment_grammar.cc index df8d4ee8..e89a92d5 100644 --- a/training/augment_grammar.cc +++ b/training/augment_grammar.cc @@ -134,9 +134,7 @@ int main(int argc, char** argv) { } else { ngram = NULL; } extra_feature = conf.count("extra_lex_feature") > 0; if (conf.count("collapse_weights")) { - Weights w; - w.InitFromFile(conf["collapse_weights"].as<string>()); - w.InitVector(&col_weights); + Weights::InitFromFile(conf["collapse_weights"].as<string>(), &col_weights); } clear_features = conf.count("clear_features_after_collapse") > 0; gather_rules = false; diff --git a/training/collapse_weights.cc b/training/collapse_weights.cc index 4fb742fb..dc480f6c 100644 --- a/training/collapse_weights.cc +++ b/training/collapse_weights.cc @@ -59,10 +59,8 @@ int main(int argc, char** argv) { InitCommandLine(argc, argv, &conf); const string wfile = conf["weights"].as<string>(); const string gfile = conf["grammar"].as<string>(); - Weights wm; - wm.InitFromFile(wfile); - vector<double> w; - wm.InitVector(&w); + vector<weight_t> w; + Weights::InitFromFile(wfile, &w); MarginalMap e_tots; MarginalMap f_tots; prob_t tot; diff --git a/training/compute_cllh.cc b/training/compute_cllh.cc index 332f6d0c..b496d196 100644 --- a/training/compute_cllh.cc +++ b/training/compute_cllh.cc @@ -148,15 +148,6 @@ int main(int argc, char** argv) { if (!InitCommandLine(argc, argv, &conf)) return false; - // load initial weights - Weights weights; - if (conf.count("weights")) - weights.InitFromFile(conf["weights"].as<string>()); - - // freeze feature set - //const bool freeze_feature_set = conf.count("freeze_feature_set"); - //if (freeze_feature_set) FD::Freeze(); - // load cdec.ini and set up decoder ReadFile ini_rf(conf["decoder_config"].as<string>()); Decoder decoder(ini_rf.stream()); @@ -165,17 +156,22 @@ int main(int argc, char** argv) { abort(); } + // load weights + vector<weight_t>& weights = decoder.CurrentWeightVector(); + if (conf.count("weights")) + Weights::InitFromFile(conf["weights"].as<string>(), &weights); + + // freeze feature set + //const bool freeze_feature_set = conf.count("freeze_feature_set"); + //if (freeze_feature_set) FD::Freeze(); + vector<string> corpus; vector<int> ids; ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus, &ids); assert(corpus.size() > 0); assert(corpus.size() == ids.size()); - vector<double> wv; - weights.InitVector(&wv); - decoder.SetWeights(wv); TrainingObserver observer; double objective = 0; - bool converged = false; observer.Reset(); if (rank == 0) @@ -197,3 +193,4 @@ int main(int argc, char** argv) { return 0; } + diff --git a/training/grammar_convert.cc b/training/grammar_convert.cc index 8d292f8a..bf8abb26 100644 --- a/training/grammar_convert.cc +++ b/training/grammar_convert.cc @@ -251,12 +251,10 @@ int main(int argc, char **argv) { const bool is_split_input = (conf["format"].as<string>() == "split"); const bool is_json_input = is_split_input || (conf["format"].as<string>() == "json"); const bool collapse_weights = conf.count("collapse_weights"); - Weights wts; vector<double> w; - if (conf.count("weights")) { - wts.InitFromFile(conf["weights"].as<string>()); - wts.InitVector(&w); - } + if (conf.count("weights")) + Weights::InitFromFile(conf["weights"].as<string>(), &w); + if (collapse_weights && !w.size()) { cerr << "--collapse_weights requires a weights file to be specified!\n"; exit(1); diff --git a/training/mpi_batch_optimize.cc b/training/mpi_batch_optimize.cc index 39a8af7d..cc5953f6 100644 --- a/training/mpi_batch_optimize.cc +++ b/training/mpi_batch_optimize.cc @@ -31,42 +31,12 @@ using namespace std; using boost::shared_ptr; namespace po = boost::program_options; -void SanityCheck(const vector<double>& w) { - for (int i = 0; i < w.size(); ++i) { - assert(!isnan(w[i])); - assert(!isinf(w[i])); - } -} - -struct FComp { - const vector<double>& w_; - FComp(const vector<double>& w) : w_(w) {} - bool operator()(int a, int b) const { - return fabs(w_[a]) > fabs(w_[b]); - } -}; - -void ShowLargestFeatures(const vector<double>& w) { - vector<int> fnums(w.size()); - for (int i = 0; i < w.size(); ++i) - fnums[i] = i; - vector<int>::iterator mid = fnums.begin(); - mid += (w.size() > 10 ? 10 : w.size()); - partial_sort(fnums.begin(), mid, fnums.end(), FComp(w)); - cerr << "TOP FEATURES:"; - for (vector<int>::iterator i = fnums.begin(); i != mid; ++i) { - cerr << ' ' << FD::Convert(*i) << '=' << w[*i]; - } - cerr << endl; -} - bool InitCommandLine(int argc, char** argv, po::variables_map* conf) { po::options_description opts("Configuration options"); opts.add_options() ("input_weights,w",po::value<string>(),"Input feature weights file") ("training_data,t",po::value<string>(),"Training data") ("decoder_config,d",po::value<string>(),"Decoder configuration file") - ("sharded_input,s",po::value<string>(), "Corpus and grammar files are 'sharded' so each processor loads its own input and grammar file. Argument is the directory containing the shards.") ("output_weights,o",po::value<string>()->default_value("-"),"Output feature weights file") ("optimization_method,m", po::value<string>()->default_value("lbfgs"), "Optimization method (sgd, lbfgs, rprop)") ("correction_buffers,M", po::value<int>()->default_value(10), "Number of gradients for LBFGS to maintain in memory") @@ -88,14 +58,10 @@ bool InitCommandLine(int argc, char** argv, po::variables_map* conf) { } po::notify(*conf); - if (conf->count("help") || !conf->count("input_weights") || !(conf->count("training_data") | conf->count("sharded_input")) || !conf->count("decoder_config")) { + if (conf->count("help") || !conf->count("input_weights") || !(conf->count("training_data")) || !conf->count("decoder_config")) { cerr << dcmdline_options << endl; return false; } - if (conf->count("training_data") && conf->count("sharded_input")) { - cerr << "Cannot specify both --training_data and --sharded_input\n"; - return false; - } return true; } @@ -236,42 +202,9 @@ int main(int argc, char** argv) { po::variables_map conf; if (!InitCommandLine(argc, argv, &conf)) return 1; - string shard_dir; - if (conf.count("sharded_input")) { - shard_dir = conf["sharded_input"].as<string>(); - if (!DirectoryExists(shard_dir)) { - if (rank == 0) cerr << "Can't find shard directory: " << shard_dir << endl; - return 1; - } - if (rank == 0) - cerr << "Shard directory: " << shard_dir << endl; - } - - // load initial weights - Weights weights; - if (rank == 0) { cerr << "Loading weights...\n"; } - weights.InitFromFile(conf["input_weights"].as<string>()); - if (rank == 0) { cerr << "Done loading weights.\n"; } - - // freeze feature set (should be optional?) - const bool freeze_feature_set = true; - if (freeze_feature_set) FD::Freeze(); - // load cdec.ini and set up decoder vector<string> cdec_ini; ReadConfig(conf["decoder_config"].as<string>(), &cdec_ini); - if (shard_dir.size()) { - if (rank == 0) { - for (int i = 0; i < cdec_ini.size(); ++i) { - if (cdec_ini[i].find("grammar=") == 0) { - cerr << "!!! using sharded input and " << conf["decoder_config"].as<string>() << " contains a grammar specification:\n" << cdec_ini[i] << "\n VERIFY THAT THIS IS CORRECT!\n"; - } - } - } - ostringstream g; - g << "grammar=" << shard_dir << "/grammar." << rank << "_of_" << size << ".gz"; - cdec_ini.push_back(g.str()); - } istringstream ini; StoreConfig(cdec_ini, &ini); if (rank == 0) cerr << "Loading grammar...\n"; @@ -282,22 +215,28 @@ int main(int argc, char** argv) { } if (rank == 0) cerr << "Done loading grammar!\n"; + // load initial weights + if (rank == 0) { cerr << "Loading weights...\n"; } + vector<weight_t>& lambdas = decoder->CurrentWeightVector(); + Weights::InitFromFile(conf["input_weights"].as<string>(), &lambdas); + if (rank == 0) { cerr << "Done loading weights.\n"; } + + // freeze feature set (should be optional?) + const bool freeze_feature_set = true; + if (freeze_feature_set) FD::Freeze(); + const int num_feats = FD::NumFeats(); if (rank == 0) cerr << "Number of features: " << num_feats << endl; + lambdas.resize(num_feats); + const bool gaussian_prior = conf.count("gaussian_prior"); - vector<double> means(num_feats, 0); + vector<weight_t> means(num_feats, 0); if (conf.count("means")) { if (!gaussian_prior) { cerr << "Don't use --means without --gaussian_prior!\n"; exit(1); } - Weights wm; - wm.InitFromFile(conf["means"].as<string>()); - if (num_feats != FD::NumFeats()) { - cerr << "[ERROR] Means file had unexpected features!\n"; - exit(1); - } - wm.InitVector(&means); + Weights::InitFromFile(conf["means"].as<string>(), &means); } shared_ptr<BatchOptimizer> o; if (rank == 0) { @@ -309,26 +248,13 @@ int main(int argc, char** argv) { cerr << "Optimizer: " << o->Name() << endl; } double objective = 0; - vector<double> lambdas(num_feats, 0.0); - weights.InitVector(&lambdas); - if (lambdas.size() != num_feats) { - cerr << "Initial weights file did not have all features specified!\n feats=" - << num_feats << "\n weights file=" << lambdas.size() << endl; - lambdas.resize(num_feats, 0.0); - } vector<double> gradient(num_feats, 0.0); - vector<double> rcv_grad(num_feats, 0.0); + vector<double> rcv_grad; + rcv_grad.clear(); bool converged = false; vector<string> corpus; - if (shard_dir.size()) { - ostringstream os; os << shard_dir << "/corpus." << rank << "_of_" << size; - ReadTrainingCorpus(os.str(), 0, 1, &corpus); - cerr << os.str() << " has " << corpus.size() << " training examples. " << endl; - if (corpus.size() > 500) { corpus.resize(500); cerr << " TRUNCATING\n"; } - } else { - ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus); - } + ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus); assert(corpus.size() > 0); TrainingObserver observer; @@ -341,19 +267,20 @@ int main(int argc, char** argv) { if (rank == 0) { cerr << "Starting decoding... (~" << corpus.size() << " sentences / proc)\n"; } - decoder->SetWeights(lambdas); for (int i = 0; i < corpus.size(); ++i) decoder->Decode(corpus[i], &observer); cerr << " process " << rank << '/' << size << " done\n"; fill(gradient.begin(), gradient.end(), 0); - fill(rcv_grad.begin(), rcv_grad.end(), 0); observer.SetLocalGradientAndObjective(&gradient, &objective); double to = 0; #ifdef HAVE_MPI + rcv_grad.resize(num_feats, 0.0); mpi::reduce(world, &gradient[0], gradient.size(), &rcv_grad[0], plus<double>(), 0); - mpi::reduce(world, objective, to, plus<double>(), 0); swap(gradient, rcv_grad); + rcv_grad.clear(); + + mpi::reduce(world, objective, to, plus<double>(), 0); objective = to; #endif @@ -378,7 +305,7 @@ int main(int argc, char** argv) { for (int i = 0; i < gradient.size(); ++i) gnorm += gradient[i] * gradient[i]; cerr << " GNORM=" << sqrt(gnorm) << endl; - vector<double> old = lambdas; + vector<weight_t> old = lambdas; int c = 0; while (old == lambdas) { ++c; @@ -387,9 +314,8 @@ int main(int argc, char** argv) { assert(c < 5); } old.clear(); - SanityCheck(lambdas); - ShowLargestFeatures(lambdas); - weights.InitFromVector(lambdas); + Weights::SanityCheck(lambdas); + Weights::ShowLargestFeatures(lambdas); converged = o->HasConverged(); if (converged) { cerr << "OPTIMIZER REPORTS CONVERGENCE!\n"; } @@ -399,7 +325,7 @@ int main(int argc, char** argv) { ostringstream vv; vv << "Objective = " << objective << " (eval count=" << o->EvaluationCount() << ")"; const string svv = vv.str(); - weights.WriteToFile(fname, true, &svv); + Weights::WriteToFile(fname, lambdas, true, &svv); } // rank == 0 int cint = converged; #ifdef HAVE_MPI @@ -411,3 +337,4 @@ int main(int argc, char** argv) { } return 0; } + diff --git a/training/mpi_online_optimize.cc b/training/mpi_online_optimize.cc index 32033c19..2ef4a2e7 100644 --- a/training/mpi_online_optimize.cc +++ b/training/mpi_online_optimize.cc @@ -31,35 +31,6 @@ namespace mpi = boost::mpi; using namespace std; namespace po = boost::program_options; -void SanityCheck(const vector<double>& w) { - for (int i = 0; i < w.size(); ++i) { - assert(!isnan(w[i])); - assert(!isinf(w[i])); - } -} - -struct FComp { - const vector<double>& w_; - FComp(const vector<double>& w) : w_(w) {} - bool operator()(int a, int b) const { - return fabs(w_[a]) > fabs(w_[b]); - } -}; - -void ShowLargestFeatures(const vector<double>& w) { - vector<int> fnums(w.size()); - for (int i = 0; i < w.size(); ++i) - fnums[i] = i; - vector<int>::iterator mid = fnums.begin(); - mid += (w.size() > 10 ? 10 : w.size()); - partial_sort(fnums.begin(), mid, fnums.end(), FComp(w)); - cerr << "TOP FEATURES:"; - for (vector<int>::iterator i = fnums.begin(); i != mid; ++i) { - cerr << ' ' << FD::Convert(*i) << '=' << w[*i]; - } - cerr << endl; -} - bool InitCommandLine(int argc, char** argv, po::variables_map* conf) { po::options_description opts("Configuration options"); opts.add_options() @@ -250,10 +221,25 @@ int main(int argc, char** argv) { if (!InitCommandLine(argc, argv, &conf)) return 1; + vector<pair<string, int> > agenda; + if (!LoadAgenda(conf["training_agenda"].as<string>(), &agenda)) + return 1; + if (rank == 0) + cerr << "Loaded agenda defining " << agenda.size() << " training epochs\n"; + + assert(agenda.size() > 0); + + if (1) { // hack to load the feature hash functions -- TODO this should not be in cdec.ini + const string& cur_config = agenda[0].first; + const unsigned max_iteration = agenda[0].second; + ReadFile ini_rf(cur_config); + Decoder decoder(ini_rf.stream()); + } + // load initial weights - Weights weights; + vector<weight_t> init_weights; if (conf.count("input_weights")) - weights.InitFromFile(conf["input_weights"].as<string>()); + Weights::InitFromFile(conf["input_weights"].as<string>(), &init_weights); vector<int> frozen_fids; if (conf.count("frozen_features")) { @@ -310,19 +296,12 @@ int main(int argc, char** argv) { rng.reset(new MT19937); SparseVector<double> x; - weights.InitSparseVector(&x); + Weights::InitSparseVector(init_weights, &x); TrainingObserver observer; int write_weights_every_ith = 100; // TODO configure int titer = -1; - vector<pair<string, int> > agenda; - if (!LoadAgenda(conf["training_agenda"].as<string>(), &agenda)) - return 1; - if (rank == 0) - cerr << "Loaded agenda defining " << agenda.size() << " training epochs\n"; - - vector<double> lambdas; for (int ai = 0; ai < agenda.size(); ++ai) { const string& cur_config = agenda[ai].first; const unsigned max_iteration = agenda[ai].second; @@ -331,6 +310,8 @@ int main(int argc, char** argv) { // load cdec.ini and set up decoder ReadFile ini_rf(cur_config); Decoder decoder(ini_rf.stream()); + vector<weight_t>& lambdas = decoder.CurrentWeightVector(); + if (ai == 0) { lambdas.swap(init_weights); init_weights.clear(); } if (rank == 0) o->ResetEpoch(); // resets the learning rate-- TODO is this good? @@ -341,15 +322,13 @@ int main(int argc, char** argv) { #ifdef HAVE_MPI mpi::timer timer; #endif - weights.InitFromVector(x); - weights.InitVector(&lambdas); + x.init_vector(&lambdas); ++iter; ++titer; observer.Reset(); - decoder.SetWeights(lambdas); if (rank == 0) { converged = (iter == max_iteration); - SanityCheck(lambdas); - ShowLargestFeatures(lambdas); + Weights::SanityCheck(lambdas); + Weights::ShowLargestFeatures(lambdas); string fname = "weights.cur.gz"; if (iter % write_weights_every_ith == 0) { ostringstream o; o << "weights.epoch_" << (ai+1) << '.' << iter << ".gz"; @@ -360,7 +339,7 @@ int main(int argc, char** argv) { vv << "total iter=" << titer << " (of current config iter=" << iter << ") minibatch=" << size_per_proc << " sentences/proc x " << size << " procs. num_feats=" << x.size() << '/' << FD::NumFeats() << " passes_thru_data=" << (titer * size_per_proc / static_cast<double>(corpus.size())) << " eta=" << lr->eta(titer); const string svv = vv.str(); cerr << svv << endl; - weights.WriteToFile(fname, true, &svv); + Weights::WriteToFile(fname, lambdas, true, &svv); } for (int i = 0; i < size_per_proc; ++i) { diff --git a/training/mr_optimize_reduce.cc b/training/mr_optimize_reduce.cc index b931991d..15e28fa1 100644 --- a/training/mr_optimize_reduce.cc +++ b/training/mr_optimize_reduce.cc @@ -88,25 +88,19 @@ int main(int argc, char** argv) { const bool use_b64 = conf["input_format"].as<string>() == "b64"; - Weights weights; - weights.InitFromFile(conf["input_weights"].as<string>()); + vector<weight_t> lambdas; + Weights::InitFromFile(conf["input_weights"].as<string>(), &lambdas); const string s_obj = "**OBJ**"; int num_feats = FD::NumFeats(); cerr << "Number of features: " << num_feats << endl; const bool gaussian_prior = conf.count("gaussian_prior"); - vector<double> means(num_feats, 0); + vector<weight_t> means(num_feats, 0); if (conf.count("means")) { if (!gaussian_prior) { cerr << "Don't use --means without --gaussian_prior!\n"; exit(1); } - Weights wm; - wm.InitFromFile(conf["means"].as<string>()); - if (num_feats != FD::NumFeats()) { - cerr << "[ERROR] Means file had unexpected features!\n"; - exit(1); - } - wm.InitVector(&means); + Weights::InitFromFile(conf["means"].as<string>(), &means); } shared_ptr<BatchOptimizer> o; const string omethod = conf["optimization_method"].as<string>(); @@ -124,8 +118,6 @@ int main(int argc, char** argv) { cerr << "No state file found, assuming ITERATION 1\n"; } - vector<double> lambdas(num_feats, 0); - weights.InitVector(&lambdas); double objective = 0; vector<double> gradient(num_feats, 0); // 0<TAB>**OBJ**=12.2;Feat1=2.3;Feat2=-0.2; @@ -223,8 +215,7 @@ int main(int argc, char** argv) { old.clear(); SanityCheck(lambdas); ShowLargestFeatures(lambdas); - weights.InitFromVector(lambdas); - weights.WriteToFile(conf["output_weights"].as<string>(), false); + Weights::WriteToFile(conf["output_weights"].as<string>(), lambdas, false); const bool conv = o->HasConverged(); if (conv) { cerr << "OPTIMIZER REPORTS CONVERGENCE!\n"; } |