From e4bc27b99d21ce4f3c6a5568070186015c655155 Mon Sep 17 00:00:00 2001 From: Chris Dyer Date: Fri, 14 Oct 2011 22:39:37 +0100 Subject: le optimizer --- training/Makefile.am | 4 + training/mpi_flex_optimize.cc | 346 ++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 350 insertions(+) create mode 100644 training/mpi_flex_optimize.cc diff --git a/training/Makefile.am b/training/Makefile.am index 0b598fd5..2a11ae52 100644 --- a/training/Makefile.am +++ b/training/Makefile.am @@ -12,6 +12,7 @@ bin_PROGRAMS = \ mpi_extract_reachable \ mpi_extract_features \ mpi_online_optimize \ + mpi_flex_optimize \ mpi_batch_optimize \ mpi_compute_cllh \ augment_grammar @@ -25,6 +26,9 @@ TESTS = lbfgs_test optimize_test mpi_online_optimize_SOURCES = mpi_online_optimize.cc online_optimizer.cc mpi_online_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 +mpi_flex_optimize_SOURCES = mpi_flex_optimize.cc online_optimizer.cc optimize.cc +mpi_flex_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 + mpi_extract_reachable_SOURCES = mpi_extract_reachable.cc mpi_extract_reachable_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/mpi_flex_optimize.cc b/training/mpi_flex_optimize.cc new file mode 100644 index 00000000..87c5f331 --- /dev/null +++ b/training/mpi_flex_optimize.cc @@ -0,0 +1,346 @@ +#include +#include +#include +#include +#include +#include + +#include +#include +#include + +#include "stringlib.h" +#include "verbose.h" +#include "hg.h" +#include "prob.h" +#include "inside_outside.h" +#include "ff_register.h" +#include "decoder.h" +#include "filelib.h" +#include "optimize.h" +#include "fdict.h" +#include "weights.h" +#include "sparse_vector.h" +#include "sampler.h" + +#ifdef HAVE_MPI +#include +#include +namespace mpi = boost::mpi; +#endif + +using namespace std; +namespace po = boost::program_options; + +bool InitCommandLine(int argc, char** argv, po::variables_map* conf) { + po::options_description opts("Configuration options"); + opts.add_options() + ("cdec_config,c",po::value(),"Decoder configuration file") + ("weights,w",po::value(),"Initial feature weights") + ("training_data,d",po::value(),"Training data") + ("minibatch_size_per_proc,s", po::value()->default_value(6), "Number of training instances evaluated per processor in each minibatch") + ("optimization_method,m", po::value()->default_value("lbfgs"), "Optimization method (options: lbfgs, sgd, rprop)") + ("minibatch_iterations,i", po::value()->default_value(10), "Number of optimization iterations per minibatch (1 = standard SGD)") + ("iterations,I", po::value()->default_value(50), "Number of passes through the training data before termination") + ("random_seed,S", po::value(), "Random seed (if not specified, /dev/random will be used)") + ("lbfgs_memory_buffers,M", po::value()->default_value(10), "Number of memory buffers for LBFGS history") + ("eta_0,e", po::value()->default_value(0.1), "Initial learning rate for SGD") + ("L1,1","Use L1 regularization") + ("L2,2","Use L2 regularization") + ("regularization_strength,C", po::value()->default_value(1.0), "Regularization strength (C)"); + po::options_description clo("Command line options"); + clo.add_options() + ("config", po::value(), "Configuration file") + ("help,h", "Print this help message and exit"); + po::options_description dconfig_options, dcmdline_options; + dconfig_options.add(opts); + dcmdline_options.add(opts).add(clo); + + po::store(parse_command_line(argc, argv, dcmdline_options), *conf); + if (conf->count("config")) { + ifstream config((*conf)["config"].as().c_str()); + po::store(po::parse_config_file(config, dconfig_options), *conf); + } + po::notify(*conf); + + if (conf->count("help") || !conf->count("training_data") || !conf->count("cdec_config")) { + cerr << "General-purpose minibatch online optimizer (MPI support " +#if HAVE_MPI + << "enabled" +#else + << "not enabled" +#endif + << ")\n" << dcmdline_options << endl; + return false; + } + return true; +} + +void ReadTrainingCorpus(const string& fname, int rank, int size, vector* c, vector* order) { + ReadFile rf(fname); + istream& in = *rf.stream(); + string line; + int id = 0; + while(in) { + getline(in, line); + if (!in) break; + if (id % size == rank) { + c->push_back(line); + order->push_back(id); + } + ++id; + } +} + +static const double kMINUS_EPSILON = -1e-6; + +struct CopyHGsObserver : public DecoderObserver { + Hypergraph* hg_; + Hypergraph* gold_hg_; + + // this can free up some memory + void RemoveRules(Hypergraph* h) { + for (unsigned i = 0; i < h->edges_.size(); ++i) + h->edges_[i].rule_.reset(); + } + + void SetCurrentHypergraphs(Hypergraph* h, Hypergraph* gold_h) { + hg_ = h; + gold_hg_ = gold_h; + } + + virtual void NotifyDecodingStart(const SentenceMetadata&) { + state = 1; + } + + // compute model expectations, denominator of objective + virtual void NotifyTranslationForest(const SentenceMetadata&, Hypergraph* hg) { + *hg_ = *hg; + RemoveRules(hg_); + assert(state == 1); + state = 2; + } + + // compute "empirical" expectations, numerator of objective + virtual void NotifyAlignmentForest(const SentenceMetadata&, Hypergraph* hg) { + assert(state == 2); + state = 3; + *gold_hg_ = *hg; + RemoveRules(gold_hg_); + } + + virtual void NotifyDecodingComplete(const SentenceMetadata&) { + if (state == 3) { + } else { + hg_->clear(); + gold_hg_->clear(); + } + } + + int state; +}; + +void ReadConfig(const string& ini, istringstream* out) { + ReadFile rf(ini); + istream& in = *rf.stream(); + ostringstream os; + while(in) { + string line; + getline(in, line); + if (!in) continue; + os << line << endl; + } + out->str(os.str()); +} + +#ifdef HAVE_MPI +namespace boost { namespace mpi { + template<> + struct is_commutative >, SparseVector > + : mpl::true_ { }; +} } // end namespace boost::mpi +#endif + +void AddGrad(const SparseVector x, double s, SparseVector* acc) { + for (SparseVector::const_iterator it = x.begin(); it != x.end(); ++it) + acc->add_value(it->first, it->second.as_float() * s); +} + +int main(int argc, char** argv) { +#ifdef HAVE_MPI + mpi::environment env(argc, argv); + mpi::communicator world; + const int size = world.size(); + const int rank = world.rank(); +#else + const int size = 1; + const int rank = 0; +#endif + if (size > 1) SetSilent(true); // turn off verbose decoder output + register_feature_functions(); + MT19937* rng = NULL; + + po::variables_map conf; + if (!InitCommandLine(argc, argv, &conf)) + return 1; + + boost::shared_ptr o; + const unsigned lbfgs_memory_buffers = conf["lbfgs_memory_buffers"].as(); + + istringstream ins; + ReadConfig(conf["cdec_config"].as(), &ins); + Decoder decoder(&ins); + + // load initial weights + vector init_weights; + if (conf.count("weights")) + Weights::InitFromFile(conf["weights"].as(), &init_weights); + + vector corpus; + vector ids; + ReadTrainingCorpus(conf["training_data"].as(), rank, size, &corpus, &ids); + assert(corpus.size() > 0); + + const unsigned size_per_proc = conf["minibatch_size_per_proc"].as(); + if (size_per_proc > corpus.size()) { + cerr << "Minibatch size must be smaller than corpus size!\n"; + return 1; + } + + size_t total_corpus_size = 0; +#ifdef HAVE_MPI + reduce(world, corpus.size(), total_corpus_size, std::plus(), 0); +#else + total_corpus_size = corpus.size(); +#endif + + if (conf.count("random_seed")) + rng = new MT19937(conf["random_seed"].as()); + else + rng = new MT19937; + + const unsigned minibatch_iterations = conf["minibatch_iterations"].as(); + + if (rank == 0) { + cerr << "Total corpus size: " << total_corpus_size << endl; + const unsigned batch_size = size_per_proc * size; + } + + SparseVector x; + Weights::InitSparseVector(init_weights, &x); + CopyHGsObserver observer; + + int write_weights_every_ith = 100; // TODO configure + int titer = -1; + + vector& lambdas = decoder.CurrentWeightVector(); + lambdas.swap(init_weights); + init_weights.clear(); + + int iter = -1; + bool converged = false; + while (!converged) { +#ifdef HAVE_MPI + mpi::timer timer; +#endif + x.init_vector(&lambdas); + ++iter; ++titer; +#if 0 + if (rank == 0) { + converged = (iter == max_iteration); + 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"; + fname = o.str(); + } + if (converged && ((ai+1)==agenda.size())) { fname = "weights.final.gz"; } + ostringstream vv; + 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(corpus.size())) << " eta=" << lr->eta(titer); + const string svv = vv.str(); + cerr << svv << endl; + Weights::WriteToFile(fname, lambdas, true, &svv); + } +#endif + + vector hgs(size_per_proc); + vector gold_hgs(size_per_proc); + for (int i = 0; i < size_per_proc; ++i) { + int ei = corpus.size() * rng->next(); + int id = ids[ei]; + observer.SetCurrentHypergraphs(&hgs[i], &gold_hgs[i]); + decoder.SetId(id); + decoder.Decode(corpus[ei], &observer); + } + + SparseVector local_grad, g; + double local_obj = 0; + o.reset(); + for (unsigned mi = 0; mi < minibatch_iterations; ++mi) { + local_grad.clear(); + g.clear(); + local_obj = 0; + + for (unsigned i = 0; i < size_per_proc; ++i) { + Hypergraph& hg = hgs[i]; + Hypergraph& hg_gold = gold_hgs[i]; + if (hg.edges_.size() < 2) continue; + + hg.Reweight(lambdas); + hg_gold.Reweight(lambdas); + SparseVector model_exp, gold_exp; + const prob_t z = InsideOutside, + EdgeFeaturesAndProbWeightFunction>(hg, &model_exp); + local_obj += log(z); + model_exp /= z; + AddGrad(model_exp, 1.0, &local_grad); + model_exp.clear(); + + const prob_t goldz = InsideOutside, + EdgeFeaturesAndProbWeightFunction>(hg_gold, &gold_exp); + local_obj -= log(goldz); + + if (log(z) - log(goldz) < kMINUS_EPSILON) { + cerr << "DIFF. ERR! log_model_z < log_gold_z: " << log(z) << " " << log(goldz) << endl; + return 1; + } + + gold_exp /= goldz; + AddGrad(gold_exp, -1.0, &local_grad); + } + + double obj = 0; +#ifdef HAVE_MPI + // TODO obj + reduce(world, local_grad, g, std::plus >(), 0); +#else + obj = local_obj; + g.swap(local_grad); +#endif + local_grad.clear(); + if (rank == 0) { + g /= (size_per_proc * size); + if (!o) + o.reset(new LBFGSOptimizer(FD::NumFeats(), lbfgs_memory_buffers)); + vector gg(FD::NumFeats()); + if (gg.size() != lambdas.size()) { lambdas.resize(gg.size()); } + for (SparseVector::const_iterator it = g.begin(); it != g.end(); ++it) + if (it->first) { gg[it->first] = it->second; } + cerr << "OBJ: " << obj << endl; + o->Optimize(obj, gg, &lambdas); + } +#ifdef HAVE_MPI + broadcast(world, x, 0); + broadcast(world, converged, 0); + world.barrier(); + if (rank == 0) { cerr << " ELAPSED TIME THIS ITERATION=" << timer.elapsed() << endl; } +#endif + } + } + return 0; +} -- cgit v1.2.3