major restructure of the training code

9 files changed, 1685 insertions, 0 deletions

diff --git a/training/crf/Makefile.am b/training/crf/Makefile.am
new file mode 100644
index 00000000..d203df25
--- /dev/null
+++ b/training/crf/Makefile.am
@@ -0,0 +1,27 @@
+bin_PROGRAMS = \
+  mpi_batch_optimize \
+  mpi_compute_cllh \
+  mpi_extract_features \
+  mpi_extract_reachable \
+  mpi_flex_optimize \
+  mpi_online_optimize
+
+mpi_online_optimize_SOURCES = mpi_online_optimize.cc
+mpi_online_optimize_LDADD = $(top_srcdir)/training/utils/libtraining_utils.a $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/klm/search/libksearch.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a -lz
+
+mpi_flex_optimize_SOURCES = mpi_flex_optimize.cc
+mpi_flex_optimize_LDADD = $(top_srcdir)/training/utils/libtraining_utils.a $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/klm/search/libksearch.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/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)/klm/search/libksearch.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a -lz
+
+mpi_extract_features_SOURCES = mpi_extract_features.cc
+mpi_extract_features_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/klm/search/libksearch.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a -lz
+
+mpi_batch_optimize_SOURCES = mpi_batch_optimize.cc cllh_observer.cc
+mpi_batch_optimize_LDADD = $(top_srcdir)/training/utils/libtraining_utils.a $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/klm/search/libksearch.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a -lz
+
+mpi_compute_cllh_SOURCES = mpi_compute_cllh.cc cllh_observer.cc
+mpi_compute_cllh_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/klm/search/libksearch.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a $(top_srcdir)/klm/lm/libklm.a $(top_srcdir)/klm/util/libklm_util.a -lz
+
+AM_CPPFLAGS = -DBOOST_TEST_DYN_LINK -W -Wall -Wno-sign-compare -I$(top_srcdir)/training -I$(top_srcdir)/training/utils -I$(top_srcdir)/utils -I$(top_srcdir)/decoder -I$(top_srcdir)/mteval
diff --git a/training/crf/cllh_observer.cc b/training/crf/cllh_observer.cc
new file mode 100644
index 00000000..4ec2fa65
--- /dev/null
+++ b/training/crf/cllh_observer.cc
@@ -0,0 +1,52 @@
+#include "cllh_observer.h"
+
+#include <cmath>
+#include <cassert>
+
+#include "inside_outside.h"
+#include "hg.h"
+#include "sentence_metadata.h"
+
+using namespace std;
+
+static const double kMINUS_EPSILON = -1e-6;
+
+ConditionalLikelihoodObserver::~ConditionalLikelihoodObserver() {}
+
+void ConditionalLikelihoodObserver::NotifyDecodingStart(const SentenceMetadata&) {
+  cur_obj = 0;
+  state = 1;
+}
+
+void ConditionalLikelihoodObserver::NotifyTranslationForest(const SentenceMetadata&, Hypergraph* hg) {
+  assert(state == 1);
+  state = 2;
+  SparseVector<prob_t> cur_model_exp;
+  const prob_t z = InsideOutside<prob_t,
+                                 EdgeProb,
+                                 SparseVector<prob_t>,
+                                 EdgeFeaturesAndProbWeightFunction>(*hg, &cur_model_exp);
+  cur_obj = log(z);
+}
+
+void ConditionalLikelihoodObserver::NotifyAlignmentForest(const SentenceMetadata& smeta, Hypergraph* hg) {
+  assert(state == 2);
+  state = 3;
+  SparseVector<prob_t> ref_exp;
+  const prob_t ref_z = InsideOutside<prob_t,
+                                     EdgeProb,
+                                     SparseVector<prob_t>,
+                                     EdgeFeaturesAndProbWeightFunction>(*hg, &ref_exp);
+
+  double log_ref_z = log(ref_z);
+
+  // rounding errors means that <0 is too strict
+  if ((cur_obj - log_ref_z) < kMINUS_EPSILON) {
+    cerr << "DIFF. ERR! log_model_z < log_ref_z: " << cur_obj << " " << log_ref_z << endl;
+    exit(1);
+  }
+  assert(!std::isnan(log_ref_z));
+  acc_obj += (cur_obj - log_ref_z);
+  trg_words += smeta.GetReference().size();
+}
+
diff --git a/training/crf/cllh_observer.h b/training/crf/cllh_observer.h
new file mode 100644
index 00000000..0de47331
--- /dev/null
+++ b/training/crf/cllh_observer.h
@@ -0,0 +1,26 @@
+#ifndef _CLLH_OBSERVER_H_
+#define _CLLH_OBSERVER_H_
+
+#include "decoder.h"
+
+struct ConditionalLikelihoodObserver : public DecoderObserver {
+
+  ConditionalLikelihoodObserver() : trg_words(), acc_obj(), cur_obj() {}
+  ~ConditionalLikelihoodObserver();
+
+  void Reset() {
+    acc_obj = 0;
+    trg_words = 0;
+  }
+ 
+  virtual void NotifyDecodingStart(const SentenceMetadata&);
+  virtual void NotifyTranslationForest(const SentenceMetadata&, Hypergraph* hg);
+  virtual void NotifyAlignmentForest(const SentenceMetadata& smeta, Hypergraph* hg);
+
+  unsigned trg_words;
+  double acc_obj;
+  double cur_obj;
+  int state;
+};
+
+#endif
diff --git a/training/crf/mpi_batch_optimize.cc b/training/crf/mpi_batch_optimize.cc
new file mode 100644
index 00000000..2eff07e4
--- /dev/null
+++ b/training/crf/mpi_batch_optimize.cc
@@ -0,0 +1,372 @@
+#include <sstream>
+#include <iostream>
+#include <vector>
+#include <cassert>
+#include <cmath>
+
+#include "config.h"
+#ifdef HAVE_MPI
+#include <boost/mpi/timer.hpp>
+#include <boost/mpi.hpp>
+namespace mpi = boost::mpi;
+#endif
+
+#include <boost/shared_ptr.hpp>
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#include "sentence_metadata.h"
+#include "cllh_observer.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 "stringlib.h"
+#include "optimize.h"
+#include "fdict.h"
+#include "weights.h"
+#include "sparse_vector.h"
+
+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()
+        ("input_weights,w",po::value<string>(),"Input feature weights file")
+        ("training_data,t",po::value<string>(),"Training data")
+        ("test_data,T",po::value<string>(),"(optional) test data")
+        ("decoder_config,c",po::value<string>(),"Decoder configuration file")
+        ("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")
+        ("gaussian_prior,p","Use a Gaussian prior on the weights")
+        ("sigma_squared", po::value<double>()->default_value(1.0), "Sigma squared term for spherical Gaussian prior")
+        ("means,u", po::value<string>(), "(optional) file containing the means for Gaussian prior");
+  po::options_description clo("Command line options");
+  clo.add_options()
+        ("config", po::value<string>(), "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<string>().c_str());
+    po::store(po::parse_config_file(config, dconfig_options), *conf);
+  }
+  po::notify(*conf);
+
+  if (conf->count("help") || !conf->count("input_weights") || !(conf->count("training_data")) || !conf->count("decoder_config")) {
+    cerr << dcmdline_options << endl;
+    return false;
+  }
+  return true;
+}
+
+void ReadTrainingCorpus(const string& fname, int rank, int size, vector<string>* c) {
+  ReadFile rf(fname);
+  istream& in = *rf.stream();
+  string line;
+  int lc = 0;
+  while(in) {
+    getline(in, line);
+    if (!in) break;
+    if (lc % size == rank) c->push_back(line);
+    ++lc;
+  }
+}
+
+static const double kMINUS_EPSILON = -1e-6;
+
+struct TrainingObserver : public DecoderObserver {
+  void Reset() {
+    acc_grad.clear();
+    acc_obj = 0;
+    total_complete = 0;
+    trg_words = 0;
+  } 
+
+  void SetLocalGradientAndObjective(vector<double>* g, double* o) const {
+    *o = acc_obj;
+    for (SparseVector<prob_t>::const_iterator it = acc_grad.begin(); it != acc_grad.end(); ++it)
+      (*g)[it->first] = it->second.as_float();
+  }
+
+  virtual void NotifyDecodingStart(const SentenceMetadata& smeta) {
+    cur_model_exp.clear();
+    cur_obj = 0;
+    state = 1;
+  }
+
+  // compute model expectations, denominator of objective
+  virtual void NotifyTranslationForest(const SentenceMetadata& smeta, Hypergraph* hg) {
+    assert(state == 1);
+    state = 2;
+    const prob_t z = InsideOutside<prob_t,
+                                   EdgeProb,
+                                   SparseVector<prob_t>,
+                                   EdgeFeaturesAndProbWeightFunction>(*hg, &cur_model_exp);
+    cur_obj = log(z);
+    cur_model_exp /= z;
+  }
+
+  // compute "empirical" expectations, numerator of objective
+  virtual void NotifyAlignmentForest(const SentenceMetadata& smeta, Hypergraph* hg) {
+    assert(state == 2);
+    state = 3;
+    SparseVector<prob_t> ref_exp;
+    const prob_t ref_z = InsideOutside<prob_t,
+                                       EdgeProb,
+                                       SparseVector<prob_t>,
+                                       EdgeFeaturesAndProbWeightFunction>(*hg, &ref_exp);
+    ref_exp /= ref_z;
+
+    double log_ref_z;
+#if 0
+    if (crf_uniform_empirical) {
+      log_ref_z = ref_exp.dot(feature_weights);
+    } else {
+      log_ref_z = log(ref_z);
+    }
+#else
+    log_ref_z = log(ref_z);
+#endif
+
+    // rounding errors means that <0 is too strict
+    if ((cur_obj - log_ref_z) < kMINUS_EPSILON) {
+      cerr << "DIFF. ERR! log_model_z < log_ref_z: " << cur_obj << " " << log_ref_z << endl;
+      exit(1);
+    }
+    assert(!std::isnan(log_ref_z));
+    ref_exp -= cur_model_exp;
+    acc_grad -= ref_exp;
+    acc_obj += (cur_obj - log_ref_z);
+    trg_words += smeta.GetReference().size();
+  }
+
+  virtual void NotifyDecodingComplete(const SentenceMetadata& smeta) {
+    if (state == 3) {
+      ++total_complete;
+    } else {
+    }
+  }
+
+  int total_complete;
+  SparseVector<prob_t> cur_model_exp;
+  SparseVector<prob_t> acc_grad;
+  double acc_obj;
+  double cur_obj;
+  unsigned trg_words;
+  int state;
+};
+
+void ReadConfig(const string& ini, vector<string>* out) {
+  ReadFile rf(ini);
+  istream& in = *rf.stream();
+  while(in) {
+    string line;
+    getline(in, line);
+    if (!in) continue;
+    out->push_back(line);
+  }
+}
+
+void StoreConfig(const vector<string>& cfg, istringstream* o) {
+  ostringstream os;
+  for (int i = 0; i < cfg.size(); ++i) { os << cfg[i] << endl; }
+  o->str(os.str());
+}
+
+template <typename T>
+struct VectorPlus : public binary_function<vector<T>, vector<T>, vector<T> >  {
+  vector<T> operator()(const vector<int>& a, const vector<int>& b) const {
+    assert(a.size() == b.size());
+    vector<T> v(a.size());
+    transform(a.begin(), a.end(), b.begin(), v.begin(), plus<T>()); 
+    return v;
+  } 
+}; 
+
+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
+  SetSilent(true);  // turn off verbose decoder output
+  register_feature_functions();
+
+  po::variables_map conf;
+  if (!InitCommandLine(argc, argv, &conf)) return 1;
+
+  // load cdec.ini and set up decoder
+  vector<string> cdec_ini;
+  ReadConfig(conf["decoder_config"].as<string>(), &cdec_ini);
+  istringstream ini;
+  StoreConfig(cdec_ini, &ini);
+  if (rank == 0) cerr << "Loading grammar...\n";
+  Decoder* decoder = new Decoder(&ini);
+  if (decoder->GetConf()["input"].as<string>() != "-") {
+    cerr << "cdec.ini must not set an input file\n";
+    return 1;
+  }
+  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<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::InitFromFile(conf["means"].as<string>(), &means);
+  }
+  boost::shared_ptr<BatchOptimizer> o;
+  if (rank == 0) {
+    const string omethod = conf["optimization_method"].as<string>();
+    if (omethod == "rprop")
+      o.reset(new RPropOptimizer(num_feats));  // TODO add configuration
+    else
+      o.reset(new LBFGSOptimizer(num_feats, conf["correction_buffers"].as<int>()));
+    cerr << "Optimizer: " << o->Name() << endl;
+  }
+  double objective = 0;
+  vector<double> gradient(num_feats, 0.0);
+  vector<double> rcv_grad;
+  rcv_grad.clear();
+  bool converged = false;
+
+  vector<string> corpus, test_corpus;
+  ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus);
+  assert(corpus.size() > 0);
+  if (conf.count("test_data"))
+    ReadTrainingCorpus(conf["test_data"].as<string>(), rank, size, &test_corpus);
+
+  TrainingObserver observer;
+  ConditionalLikelihoodObserver cllh_observer;
+  while (!converged) {
+    observer.Reset();
+    cllh_observer.Reset();
+#ifdef HAVE_MPI
+    mpi::timer timer;
+    world.barrier();
+#endif
+    if (rank == 0) {
+      cerr << "Starting decoding... (~" << corpus.size() << " sentences / proc)\n";
+      cerr << "  Testset size: " << test_corpus.size() << " sentences / proc)\n";
+    }
+    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);
+    observer.SetLocalGradientAndObjective(&gradient, &objective);
+
+    unsigned total_words = 0;
+#ifdef HAVE_MPI
+    double to = 0;
+    rcv_grad.resize(num_feats, 0.0);
+    mpi::reduce(world, &gradient[0], gradient.size(), &rcv_grad[0], plus<double>(), 0);
+    swap(gradient, rcv_grad);
+    rcv_grad.clear();
+
+    reduce(world, observer.trg_words, total_words, std::plus<unsigned>(), 0);
+    mpi::reduce(world, objective, to, plus<double>(), 0);
+    objective = to;
+#else
+    total_words = observer.trg_words;
+#endif
+    if (rank == 0)
+      cerr << "TRAINING CORPUS: ln p(f|e)=" << objective << "\t log_2 p(f|e) = " << (objective/log(2)) << "\t cond. entropy = " << (objective/log(2) / total_words) << "\t ppl = " << pow(2, (objective/log(2) / total_words)) << endl;
+
+    for (int i = 0; i < test_corpus.size(); ++i)
+      decoder->Decode(test_corpus[i], &cllh_observer);
+
+    double test_objective = 0;
+    unsigned test_total_words = 0;
+#ifdef HAVE_MPI
+    reduce(world, cllh_observer.acc_obj, test_objective, std::plus<double>(), 0);
+    reduce(world, cllh_observer.trg_words, test_total_words, std::plus<unsigned>(), 0);
+#else
+    test_objective = cllh_observer.acc_obj;
+    test_total_words = cllh_observer.trg_words;
+#endif
+
+    if (rank == 0) {  // run optimizer only on rank=0 node
+      if (test_corpus.size())
+        cerr << "    TEST CORPUS: ln p(f|e)=" << test_objective << "\t log_2 p(f|e) = " << (test_objective/log(2)) << "\t cond. entropy = " << (test_objective/log(2) / test_total_words) << "\t ppl = " << pow(2, (test_objective/log(2) / test_total_words)) << endl;
+      if (gaussian_prior) {
+        const double sigsq = conf["sigma_squared"].as<double>();
+        double norm = 0;
+        for (int k = 1; k < lambdas.size(); ++k) {
+          const double& lambda_k = lambdas[k];
+          if (lambda_k) {
+            const double param = (lambda_k - means[k]);
+            norm += param * param;
+            gradient[k] += param / sigsq;
+          }
+        }
+        const double reg = norm / (2.0 * sigsq);
+        cerr << "REGULARIZATION TERM: " << reg << endl;
+        objective += reg;
+      }
+      cerr << "EVALUATION #" << o->EvaluationCount() << " OBJECTIVE: " << objective << endl;
+      double gnorm = 0;
+      for (int i = 0; i < gradient.size(); ++i)
+        gnorm += gradient[i] * gradient[i];
+      cerr << "  GNORM=" << sqrt(gnorm) << endl;
+      vector<weight_t> old = lambdas;
+      int c = 0;
+      while (old == lambdas) {
+        ++c;
+        if (c > 1) { cerr << "Same lambdas, repeating optimization\n"; }
+        o->Optimize(objective, gradient, &lambdas);
+        assert(c < 5);
+      }
+      old.clear();
+      Weights::SanityCheck(lambdas);
+      Weights::ShowLargestFeatures(lambdas);
+
+      converged = o->HasConverged();
+      if (converged) { cerr << "OPTIMIZER REPORTS CONVERGENCE!\n"; }
+
+      string fname = "weights.cur.gz";
+      if (converged) { fname = "weights.final.gz"; }
+      ostringstream vv;
+      vv << "Objective = " << objective << "  (eval count=" << o->EvaluationCount() << ")";
+      const string svv = vv.str();
+      Weights::WriteToFile(fname, lambdas, true, &svv);
+    }  // rank == 0
+    int cint = converged;
+#ifdef HAVE_MPI
+    mpi::broadcast(world, &lambdas[0], lambdas.size(), 0);
+    mpi::broadcast(world, cint, 0);
+    if (rank == 0) { cerr << "  ELAPSED TIME THIS ITERATION=" << timer.elapsed() << endl; }
+#endif
+    converged = cint;
+  }
+  return 0;
+}
+
diff --git a/training/crf/mpi_compute_cllh.cc b/training/crf/mpi_compute_cllh.cc
new file mode 100644
index 00000000..066389d0
--- /dev/null
+++ b/training/crf/mpi_compute_cllh.cc
@@ -0,0 +1,134 @@
+#include <iostream>
+#include <vector>
+#include <cassert>
+#include <cmath>
+
+#include "config.h"
+#ifdef HAVE_MPI
+#include <boost/mpi.hpp>
+#endif
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#include "cllh_observer.h"
+#include "sentence_metadata.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 "weights.h"
+
+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()
+        ("weights,w",po::value<string>(),"Input feature weights file")
+        ("training_data,t",po::value<string>(),"Training data corpus")
+        ("decoder_config,c",po::value<string>(),"Decoder configuration file");
+  po::options_description clo("Command line options");
+  clo.add_options()
+        ("config", po::value<string>(), "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<string>().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("decoder_config")) {
+    cerr << dcmdline_options << endl;
+    return false;
+  }
+  return true;
+}
+
+void ReadInstances(const string& fname, int rank, int size, vector<string>* c) {
+  assert(fname != "-");
+  ReadFile rf(fname);
+  istream& in = *rf.stream();
+  string line;
+  int lc = 0;
+  while(in) {
+    getline(in, line);
+    if (!in) break;
+    if (lc % size == rank) c->push_back(line);
+    ++lc;
+  }
+}
+
+static const double kMINUS_EPSILON = -1e-6;
+
+#ifdef HAVE_MPI
+namespace mpi = boost::mpi;
+#endif
+
+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();
+
+  po::variables_map conf;
+  if (!InitCommandLine(argc, argv, &conf))
+    return false;
+
+  // load cdec.ini and set up decoder
+  ReadFile ini_rf(conf["decoder_config"].as<string>());
+  Decoder decoder(ini_rf.stream());
+  if (decoder.GetConf()["input"].as<string>() != "-") {
+    cerr << "cdec.ini must not set an input file\n";
+    abort();
+  }
+
+  // load weights
+  vector<weight_t>& weights = decoder.CurrentWeightVector();
+  if (conf.count("weights"))
+    Weights::InitFromFile(conf["weights"].as<string>(), &weights);
+
+  vector<string> corpus;
+  ReadInstances(conf["training_data"].as<string>(), rank, size, &corpus);
+  assert(corpus.size() > 0);
+
+  if (rank == 0)
+    cerr << "Each processor is decoding ~" << corpus.size() << " training examples...\n";
+
+  ConditionalLikelihoodObserver observer;
+  for (int i = 0; i < corpus.size(); ++i)
+    decoder.Decode(corpus[i], &observer);
+
+  double objective = 0;
+  unsigned total_words = 0;
+#ifdef HAVE_MPI
+  reduce(world, observer.acc_obj, objective, std::plus<double>(), 0);
+  reduce(world, observer.trg_words, total_words, std::plus<unsigned>(), 0);
+#else
+  objective = observer.acc_obj;
+#endif
+
+  if (rank == 0) {
+    cout << "CONDITIONAL LOG_e LIKELIHOOD: " << objective << endl;
+    cout << "CONDITIONAL LOG_2 LIKELIHOOD: " << (objective/log(2)) << endl;
+    cout << "         CONDITIONAL ENTROPY: " << (objective/log(2) / total_words) << endl;
+    cout << "                  PERPLEXITY: " << pow(2, (objective/log(2) / total_words)) << endl;
+  }
+
+  return 0;
+}
+
diff --git a/training/crf/mpi_extract_features.cc b/training/crf/mpi_extract_features.cc
new file mode 100644
index 00000000..6750aa15
--- /dev/null
+++ b/training/crf/mpi_extract_features.cc
@@ -0,0 +1,151 @@
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <cassert>
+
+#include "config.h"
+#ifdef HAVE_MPI
+#include <boost/mpi.hpp>
+#endif
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#include "ff_register.h"
+#include "verbose.h"
+#include "filelib.h"
+#include "fdict.h"
+#include "decoder.h"
+#include "weights.h"
+
+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()
+        ("training_data,t",po::value<string>(),"Training data corpus")
+        ("decoder_config,c",po::value<string>(),"Decoder configuration file")
+        ("weights,w", po::value<string>(), "(Optional) weights file; weights may affect what features are encountered in pruning configurations")
+        ("output_prefix,o",po::value<string>()->default_value("features"),"Output path prefix");
+  po::options_description clo("Command line options");
+  clo.add_options()
+        ("config", po::value<string>(), "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<string>().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("decoder_config")) {
+    cerr << "Decode an input set (optionally in parallel using MPI) and write\nout the feature strings encountered.\n";
+    cerr << dcmdline_options << endl;
+    return false;
+  }
+  return true;
+}
+
+void ReadTrainingCorpus(const string& fname, int rank, int size, vector<string>* c) {
+  ReadFile rf(fname);
+  istream& in = *rf.stream();
+  string line;
+  int lc = 0;
+  while(in) {
+    getline(in, line);
+    if (!in) break;
+    if (lc % size == rank) c->push_back(line);
+    ++lc;
+  }
+}
+
+static const double kMINUS_EPSILON = -1e-6;
+
+struct TrainingObserver : public DecoderObserver {
+
+  virtual void NotifyDecodingStart(const SentenceMetadata&) {
+  }
+
+  // compute model expectations, denominator of objective
+  virtual void NotifyTranslationForest(const SentenceMetadata&, Hypergraph* hg) {
+  }
+
+  // compute "empirical" expectations, numerator of objective
+  virtual void NotifyAlignmentForest(const SentenceMetadata& smeta, Hypergraph* hg) {
+  }
+};
+
+#ifdef HAVE_MPI
+namespace mpi = boost::mpi;
+#endif
+
+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();
+
+  po::variables_map conf;
+  if (!InitCommandLine(argc, argv, &conf))
+    return false;
+
+  // load cdec.ini and set up decoder
+  ReadFile ini_rf(conf["decoder_config"].as<string>());
+  Decoder decoder(ini_rf.stream());
+  if (decoder.GetConf()["input"].as<string>() != "-") {
+    cerr << "cdec.ini must not set an input file\n";
+    abort();
+  }
+
+  if (FD::UsingPerfectHashFunction()) {
+    cerr << "Your configuration file has enabled a cmph hash function. Please disable.\n";
+    return 1;
+  }
+
+  // load optional weights
+  if (conf.count("weights"))
+    Weights::InitFromFile(conf["weights"].as<string>(), &decoder.CurrentWeightVector());
+
+  vector<string> corpus;
+  ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus);
+  assert(corpus.size() > 0);
+
+  TrainingObserver observer;
+
+  if (rank == 0)
+    cerr << "Each processor is decoding ~" << corpus.size() << " training examples...\n";
+
+  for (int i = 0; i < corpus.size(); ++i)
+    decoder.Decode(corpus[i], &observer);
+
+  {
+    ostringstream os;
+    os << conf["output_prefix"].as<string>() << '.' << rank << "_of_" << size;
+    WriteFile wf(os.str());
+    ostream& out = *wf.stream();
+    const unsigned num_feats = FD::NumFeats();
+    for (unsigned i = 1; i < num_feats; ++i) {
+      out << FD::Convert(i) << endl;
+    }
+    cerr << "Wrote " << os.str() << endl;
+  }
+
+#ifdef HAVE_MPI
+  world.barrier();
+#else
+#endif
+
+  return 0;
+}
+
diff --git a/training/crf/mpi_extract_reachable.cc b/training/crf/mpi_extract_reachable.cc
new file mode 100644
index 00000000..2a7c2b9d
--- /dev/null
+++ b/training/crf/mpi_extract_reachable.cc
@@ -0,0 +1,163 @@
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <cassert>
+
+#include "config.h"
+#ifdef HAVE_MPI
+#include <boost/mpi.hpp>
+#endif
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#include "ff_register.h"
+#include "verbose.h"
+#include "filelib.h"
+#include "fdict.h"
+#include "decoder.h"
+#include "weights.h"
+
+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()
+        ("training_data,t",po::value<string>(),"Training data corpus")
+        ("decoder_config,c",po::value<string>(),"Decoder configuration file")
+        ("weights,w", po::value<string>(), "(Optional) weights file; weights may affect what features are encountered in pruning configurations")
+        ("output_prefix,o",po::value<string>()->default_value("reachable"),"Output path prefix");
+  po::options_description clo("Command line options");
+  clo.add_options()
+        ("config", po::value<string>(), "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<string>().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("decoder_config")) {
+    cerr << "Decode an input set (optionally in parallel using MPI) and write\nout the inputs that produce reachable parallel parses.\n";
+    cerr << dcmdline_options << endl;
+    return false;
+  }
+  return true;
+}
+
+void ReadTrainingCorpus(const string& fname, int rank, int size, vector<string>* c) {
+  ReadFile rf(fname);
+  istream& in = *rf.stream();
+  string line;
+  int lc = 0;
+  while(in) {
+    getline(in, line);
+    if (!in) break;
+    if (lc % size == rank) c->push_back(line);
+    ++lc;
+  }
+}
+
+static const double kMINUS_EPSILON = -1e-6;
+
+struct ReachabilityObserver : public DecoderObserver {
+
+  virtual void NotifyDecodingStart(const SentenceMetadata&) {
+    reachable = false;
+  }
+
+  // compute model expectations, denominator of objective
+  virtual void NotifyTranslationForest(const SentenceMetadata&, Hypergraph* hg) {
+  }
+
+  // compute "empirical" expectations, numerator of objective
+  virtual void NotifyAlignmentForest(const SentenceMetadata& smeta, Hypergraph* hg) {
+    reachable = true;
+  }
+
+  bool reachable;
+};
+
+#ifdef HAVE_MPI
+namespace mpi = boost::mpi;
+#endif
+
+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();
+
+  po::variables_map conf;
+  if (!InitCommandLine(argc, argv, &conf))
+    return false;
+
+  // load cdec.ini and set up decoder
+  ReadFile ini_rf(conf["decoder_config"].as<string>());
+  Decoder decoder(ini_rf.stream());
+  if (decoder.GetConf()["input"].as<string>() != "-") {
+    cerr << "cdec.ini must not set an input file\n";
+    abort();
+  }
+
+  if (FD::UsingPerfectHashFunction()) {
+    cerr << "Your configuration file has enabled a cmph hash function. Please disable.\n";
+    return 1;
+  }
+
+  // load optional weights
+  if (conf.count("weights"))
+    Weights::InitFromFile(conf["weights"].as<string>(), &decoder.CurrentWeightVector());
+
+  vector<string> corpus;
+  ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus);
+  assert(corpus.size() > 0);
+
+
+  if (rank == 0)
+    cerr << "Each processor is decoding ~" << corpus.size() << " training examples...\n";
+
+  size_t num_reached = 0;
+  {
+    ostringstream os;
+    os << conf["output_prefix"].as<string>() << '.' << rank << "_of_" << size;
+    WriteFile wf(os.str());
+    ostream& out = *wf.stream();
+    ReachabilityObserver observer;
+    for (int i = 0; i < corpus.size(); ++i) {
+      decoder.Decode(corpus[i], &observer);
+      if (observer.reachable) {
+         out << corpus[i] << endl;
+         ++num_reached;
+      }
+      corpus[i].clear();
+    }
+    cerr << "Shard " << rank << '/' << size << " finished, wrote "
+         << num_reached << " instances to " << os.str() << endl;
+  }
+
+  size_t total = 0;
+#ifdef HAVE_MPI
+  reduce(world, num_reached, total, std::plus<double>(), 0);
+#else
+  total = num_reached;
+#endif
+  if (rank == 0) {
+    cerr << "-----------------------------------------\n";
+    cerr << "TOTAL = " << total << " instances\n";
+  }
+  return 0;
+}
+
diff --git a/training/crf/mpi_flex_optimize.cc b/training/crf/mpi_flex_optimize.cc
new file mode 100644
index 00000000..b52decdc
--- /dev/null
+++ b/training/crf/mpi_flex_optimize.cc
@@ -0,0 +1,386 @@
+#include <sstream>
+#include <iostream>
+#include <fstream>
+#include <vector>
+#include <cassert>
+#include <cmath>
+
+#include <boost/shared_ptr.hpp>
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#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 <boost/mpi/timer.hpp>
+#include <boost/mpi.hpp>
+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<string>(),"Decoder configuration file")
+        ("weights,w",po::value<string>(),"Initial feature weights")
+        ("training_data,d",po::value<string>(),"Training data")
+        ("minibatch_size_per_proc,s", po::value<unsigned>()->default_value(6), "Number of training instances evaluated per processor in each minibatch")
+        ("minibatch_iterations,i", po::value<unsigned>()->default_value(10), "Number of optimization iterations per minibatch")
+        ("iterations,I", po::value<unsigned>()->default_value(50), "Number of passes through the training data before termination")
+        ("regularization_strength,C", po::value<double>()->default_value(0.2), "Regularization strength")
+        ("time_series_strength,T", po::value<double>()->default_value(0.0), "Time series regularization strength")
+        ("random_seed,S", po::value<uint32_t>(), "Random seed (if not specified, /dev/random will be used)")
+        ("lbfgs_memory_buffers,M", po::value<unsigned>()->default_value(10), "Number of memory buffers for LBFGS history");
+  po::options_description clo("Command line options");
+  clo.add_options()
+        ("config", po::value<string>(), "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<string>().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 << "LBFGS 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<string>* c, vector<int>* 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<std::plus<SparseVector<double> >, SparseVector<double> > 
+    : mpl::true_ { };
+} } // end namespace boost::mpi
+#endif
+
+void AddGrad(const SparseVector<prob_t> x, double s, SparseVector<double>* acc) {
+  for (SparseVector<prob_t>::const_iterator it = x.begin(); it != x.end(); ++it)
+    acc->add_value(it->first, it->second.as_float() * s);
+}
+
+double PNorm(const vector<double>& v, const double p) {
+  double acc = 0;
+  for (int i = 0; i < v.size(); ++i)
+    acc += pow(v[i], p);
+  return pow(acc, 1.0 / p);
+}
+
+void VV(ostream&os, const vector<double>& v) {
+  for (int i = 1; i < v.size(); ++i)
+    if (v[i]) os << FD::Convert(i) << "=" << v[i] << " ";
+}
+
+double ApplyRegularizationTerms(const double C,
+                                const double T,
+                                const vector<double>& weights,
+                                const vector<double>& prev_weights,
+                                double* g) {
+  double reg = 0;
+  for (size_t i = 0; i < weights.size(); ++i) {
+    const double prev_w_i = (i < prev_weights.size() ? prev_weights[i] : 0.0);
+    const double& w_i = weights[i];
+    reg += C * w_i * w_i;
+    g[i] += 2 * C * w_i;
+
+    reg += T * (w_i - prev_w_i) * (w_i - prev_w_i);
+    g[i] += 2 * T * (w_i - prev_w_i);
+  }
+  return reg;
+}
+
+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<BatchOptimizer> o;
+  const unsigned lbfgs_memory_buffers = conf["lbfgs_memory_buffers"].as<unsigned>();
+  const unsigned size_per_proc = conf["minibatch_size_per_proc"].as<unsigned>();
+  const unsigned minibatch_iterations = conf["minibatch_iterations"].as<unsigned>();
+  const double regularization_strength = conf["regularization_strength"].as<double>();
+  const double time_series_strength = conf["time_series_strength"].as<double>();
+  const bool use_time_series_reg = time_series_strength > 0.0;
+  const unsigned max_iteration = conf["iterations"].as<unsigned>();
+
+  vector<string> corpus;
+  vector<int> ids;
+  ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus, &ids);
+  assert(corpus.size() > 0);
+
+  if (size_per_proc > corpus.size()) {
+    cerr << "Minibatch size (per processor) must be smaller or equal to the local corpus size!\n";
+    return 1;
+  }
+
+  // initialize decoder (loads hash functions if necessary)
+  istringstream ins;
+  ReadConfig(conf["cdec_config"].as<string>(), &ins);
+  Decoder decoder(&ins);
+
+  // load initial weights
+  vector<weight_t> prev_weights;
+  if (conf.count("weights"))
+    Weights::InitFromFile(conf["weights"].as<string>(), &prev_weights);
+
+  if (conf.count("random_seed"))
+    rng = new MT19937(conf["random_seed"].as<uint32_t>());
+  else
+    rng = new MT19937;
+
+  size_t total_corpus_size = 0;
+#ifdef HAVE_MPI
+  reduce(world, corpus.size(), total_corpus_size, std::plus<size_t>(), 0);
+#else
+  total_corpus_size = corpus.size();
+#endif
+
+  if (rank == 0)
+    cerr << "Total corpus size: " << total_corpus_size << endl;
+
+  CopyHGsObserver observer;
+
+  int write_weights_every_ith = 100; // TODO configure
+  int titer = -1;
+
+  vector<weight_t>& cur_weights = decoder.CurrentWeightVector();
+  if (use_time_series_reg) {
+    cur_weights = prev_weights;
+  } else {
+    cur_weights.swap(prev_weights);
+    prev_weights.clear();
+  }
+
+  int iter = -1;
+  bool converged = false;
+  vector<double> gg;
+  while (!converged) {
+#ifdef HAVE_MPI
+    mpi::timer timer;
+#endif
+    ++iter; ++titer;
+    if (rank == 0) {
+      converged = (iter == max_iteration);
+        string fname = "weights.cur.gz";
+        if (iter % write_weights_every_ith == 0) {
+          ostringstream o; o << "weights.epoch_" << iter << ".gz";
+          fname = o.str();
+        }
+        if (converged) { 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=" << FD::NumFeats() << "   passes_thru_data=" << (titer * size_per_proc / static_cast<double>(corpus.size()));
+        const string svv = vv.str();
+        Weights::WriteToFile(fname, cur_weights, true, &svv);
+      }
+
+      vector<Hypergraph> hgs(size_per_proc);
+      vector<Hypergraph> 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<double> 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(cur_weights);
+          hg_gold.Reweight(cur_weights);
+          SparseVector<prob_t> model_exp, gold_exp;
+          const prob_t z = InsideOutside<prob_t,
+                                         EdgeProb,
+                                         SparseVector<prob_t>,
+                                         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<prob_t,
+                                         EdgeProb,
+                                         SparseVector<prob_t>,
+                                         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
+        reduce(world, local_obj, obj, std::plus<double>(), 0);
+        reduce(world, local_grad, g, std::plus<SparseVector<double> >(), 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));
+          gg.clear();
+          gg.resize(FD::NumFeats());
+          if (gg.size() != cur_weights.size()) { cur_weights.resize(gg.size()); }
+          for (SparseVector<double>::iterator it = g.begin(); it != g.end(); ++it)
+            if (it->first) { gg[it->first] = it->second; }
+          g.clear();
+          double r = ApplyRegularizationTerms(regularization_strength,
+                                time_series_strength, // * (iter == 0 ? 0.0 : 1.0),
+                                cur_weights,
+                                prev_weights,
+                                &gg[0]);
+          obj += r;
+          if (mi == 0 || mi == (minibatch_iterations - 1)) {
+            if (!mi) cerr << iter << ' '; else cerr << ' ';
+            cerr << "OBJ=" << obj << " (REG=" << r << ")" << " |g|=" << PNorm(gg, 2) << " |w|=" << PNorm(cur_weights, 2); 
+            if (mi > 0) cerr << endl << flush; else cerr << ' ';
+          } else { cerr << '.' << flush; }
+          // cerr << "w = "; VV(cerr, cur_weights); cerr << endl;
+          // cerr << "g = "; VV(cerr, gg); cerr << endl;
+          o->Optimize(obj, gg, &cur_weights);
+        }
+#ifdef HAVE_MPI
+        broadcast(world, cur_weights, 0);
+        broadcast(world, converged, 0);
+        world.barrier();
+#endif
+    }
+    prev_weights = cur_weights;
+  }
+  return 0;
+}
diff --git a/training/crf/mpi_online_optimize.cc b/training/crf/mpi_online_optimize.cc
new file mode 100644
index 00000000..d6968848
--- /dev/null
+++ b/training/crf/mpi_online_optimize.cc
@@ -0,0 +1,374 @@
+#include <sstream>
+#include <iostream>
+#include <fstream>
+#include <vector>
+#include <cassert>
+#include <cmath>
+#include <tr1/memory>
+
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#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 "online_optimizer.h"
+#include "fdict.h"
+#include "weights.h"
+#include "sparse_vector.h"
+#include "sampler.h"
+
+#ifdef HAVE_MPI
+#include <boost/mpi/timer.hpp>
+#include <boost/mpi.hpp>
+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()
+        ("input_weights,w",po::value<string>(),"Input feature weights file")
+        ("frozen_features,z",po::value<string>(), "List of features not to optimize")
+        ("training_data,t",po::value<string>(),"Training data corpus")
+        ("training_agenda,a",po::value<string>(), "Text file listing a series of configuration files and the number of iterations to train using each configuration successively")
+        ("minibatch_size_per_proc,s", po::value<unsigned>()->default_value(5), "Number of training instances evaluated per processor in each minibatch")
+        ("optimization_method,m", po::value<string>()->default_value("sgd"), "Optimization method (sgd)")
+        ("random_seed,S", po::value<uint32_t>(), "Random seed (if not specified, /dev/random will be used)")
+        ("eta_0,e", po::value<double>()->default_value(0.2), "Initial learning rate for SGD (eta_0)")
+        ("L1,1","Use L1 regularization")
+        ("regularization_strength,C", po::value<double>()->default_value(1.0), "Regularization strength (C)");
+  po::options_description clo("Command line options");
+  clo.add_options()
+        ("config", po::value<string>(), "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<string>().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("training_agenda")) {
+    cerr << dcmdline_options << endl;
+    return false;
+  }
+  return true;
+}
+
+void ReadTrainingCorpus(const string& fname, int rank, int size, vector<string>* c, vector<int>* 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 TrainingObserver : public DecoderObserver {
+  void Reset() {
+    acc_grad.clear();
+    acc_obj = 0;
+    total_complete = 0;
+  } 
+
+  void SetLocalGradientAndObjective(vector<double>* g, double* o) const {
+    *o = acc_obj;
+    for (SparseVector<prob_t>::const_iterator it = acc_grad.begin(); it != acc_grad.end(); ++it)
+      (*g)[it->first] = it->second.as_float();
+  }
+
+  virtual void NotifyDecodingStart(const SentenceMetadata& smeta) {
+    cur_model_exp.clear();
+    cur_obj = 0;
+    state = 1;
+  }
+
+  // compute model expectations, denominator of objective
+  virtual void NotifyTranslationForest(const SentenceMetadata& smeta, Hypergraph* hg) {
+    assert(state == 1);
+    state = 2;
+    const prob_t z = InsideOutside<prob_t,
+                                   EdgeProb,
+                                   SparseVector<prob_t>,
+                                   EdgeFeaturesAndProbWeightFunction>(*hg, &cur_model_exp);
+    cur_obj = log(z);
+    cur_model_exp /= z;
+  }
+
+  // compute "empirical" expectations, numerator of objective
+  virtual void NotifyAlignmentForest(const SentenceMetadata& smeta, Hypergraph* hg) {
+    assert(state == 2);
+    state = 3;
+    SparseVector<prob_t> ref_exp;
+    const prob_t ref_z = InsideOutside<prob_t,
+                                       EdgeProb,
+                                       SparseVector<prob_t>,
+                                       EdgeFeaturesAndProbWeightFunction>(*hg, &ref_exp);
+    ref_exp /= ref_z;
+
+    double log_ref_z;
+#if 0
+    if (crf_uniform_empirical) {
+      log_ref_z = ref_exp.dot(feature_weights);
+    } else {
+      log_ref_z = log(ref_z);
+    }
+#else
+    log_ref_z = log(ref_z);
+#endif
+
+    // rounding errors means that <0 is too strict
+    if ((cur_obj - log_ref_z) < kMINUS_EPSILON) {
+      cerr << "DIFF. ERR! log_model_z < log_ref_z: " << cur_obj << " " << log_ref_z << endl;
+      exit(1);
+    }
+    assert(!std::isnan(log_ref_z));
+    ref_exp -= cur_model_exp;
+    acc_grad += ref_exp;
+    acc_obj += (cur_obj - log_ref_z);
+  }
+
+  virtual void NotifyDecodingComplete(const SentenceMetadata& smeta) {
+    if (state == 3) {
+      ++total_complete;
+    } else {
+    }
+  }
+
+  void GetGradient(SparseVector<double>* g) const {
+    g->clear();
+    for (SparseVector<prob_t>::const_iterator it = acc_grad.begin(); it != acc_grad.end(); ++it)
+      g->set_value(it->first, it->second.as_float());
+  }
+
+  int total_complete;
+  SparseVector<prob_t> cur_model_exp;
+  SparseVector<prob_t> acc_grad;
+  double acc_obj;
+  double cur_obj;
+  int state;
+};
+
+#ifdef HAVE_MPI
+namespace boost { namespace mpi {
+  template<>
+  struct is_commutative<std::plus<SparseVector<double> >, SparseVector<double> > 
+    : mpl::true_ { };
+} } // end namespace boost::mpi
+#endif
+
+bool LoadAgenda(const string& file, vector<pair<string, int> >* a) {
+  ReadFile rf(file);
+  istream& in = *rf.stream();
+  string line;
+  while(in) {
+    getline(in, line);
+    if (!in) break;
+    if (line.empty()) continue;
+    if (line[0] == '#') continue;
+    int sc = 0;
+    if (line.size() < 3) return false;
+    for (int i = 0; i < line.size(); ++i) { if (line[i] == ' ') ++sc; }
+    if (sc != 1) { cerr << "Too many spaces in line: " << line << endl; return false; }
+    size_t d = line.find(" ");
+    pair<string, int> x;
+    x.first = line.substr(0,d);
+    x.second = atoi(line.substr(d+1).c_str());
+    a->push_back(x);
+    if (!FileExists(x.first)) {
+      cerr << "Can't find file " << x.first << endl;
+      return false;
+    }
+  }
+  return true;
+}
+
+int main(int argc, char** argv) {
+  cerr << "THIS SOFTWARE IS DEPRECATED YOU SHOULD USE mpi_flex_optimize\n";
+#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();
+  std::tr1::shared_ptr<MT19937> rng;
+
+  po::variables_map conf;
+  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
+  vector<weight_t> init_weights;
+  if (conf.count("input_weights"))
+    Weights::InitFromFile(conf["input_weights"].as<string>(), &init_weights);
+
+  vector<int> frozen_fids;
+  if (conf.count("frozen_features")) {
+    ReadFile rf(conf["frozen_features"].as<string>());
+    istream& in = *rf.stream();
+    string line;
+    while(in) {
+      getline(in, line);
+      if (line.empty()) continue;
+      if (line[0] == ' ' || line[line.size() - 1] == ' ') { line = Trim(line); }
+      frozen_fids.push_back(FD::Convert(line));
+    }
+    if (rank == 0) cerr << "Freezing " << frozen_fids.size() << " features.\n";
+  }
+
+  vector<string> corpus;
+  vector<int> ids;
+  ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus, &ids);
+  assert(corpus.size() > 0);
+
+  std::tr1::shared_ptr<OnlineOptimizer> o;
+  std::tr1::shared_ptr<LearningRateSchedule> lr;
+
+  const unsigned size_per_proc = conf["minibatch_size_per_proc"].as<unsigned>();
+  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<size_t>(), 0);
+#else
+  total_corpus_size = corpus.size();
+#endif
+
+  if (rank == 0) {
+    cerr << "Total corpus size: " << total_corpus_size << endl;
+    const unsigned batch_size = size_per_proc * size;
+    // TODO config
+    lr.reset(new ExponentialDecayLearningRate(batch_size, conf["eta_0"].as<double>()));
+
+    const string omethod = conf["optimization_method"].as<string>();
+    if (omethod == "sgd") {
+      const double C = conf["regularization_strength"].as<double>();
+      o.reset(new CumulativeL1OnlineOptimizer(lr, total_corpus_size, C, frozen_fids));
+    } else {
+      assert(!"fail");
+    }
+  }
+  if (conf.count("random_seed"))
+    rng.reset(new MT19937(conf["random_seed"].as<uint32_t>()));
+  else
+    rng.reset(new MT19937);
+
+  SparseVector<double> x;
+  Weights::InitSparseVector(init_weights, &x);
+  TrainingObserver observer;
+
+  int write_weights_every_ith = 100; // TODO configure
+  int titer = -1;
+
+  for (int ai = 0; ai < agenda.size(); ++ai) {
+    const string& cur_config = agenda[ai].first;
+    const unsigned max_iteration = agenda[ai].second;
+    if (rank == 0)
+      cerr << "STARTING TRAINING EPOCH " << (ai+1) << ". CONFIG=" << cur_config << endl;
+    // 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?
+
+    int iter = -1;
+    bool converged = false;
+    while (!converged) {
+#ifdef HAVE_MPI
+      mpi::timer timer;
+#endif
+      x.init_vector(&lambdas);
+      ++iter; ++titer;
+      observer.Reset();
+      if (rank == 0) {
+        converged = (iter == max_iteration);
+        Weights::SanityCheck(lambdas);
+        static int cc = 0; ++cc; if (cc > 1) { 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<double>(corpus.size())) << "   eta=" << lr->eta(titer);
+        const string svv = vv.str();
+        cerr << svv << endl;
+        Weights::WriteToFile(fname, lambdas, true, &svv);
+      }
+
+      for (int i = 0; i < size_per_proc; ++i) {
+        int ei = corpus.size() * rng->next();
+        int id = ids[ei];
+        decoder.SetId(id);
+        decoder.Decode(corpus[ei], &observer);
+      }
+      SparseVector<double> local_grad, g;
+      observer.GetGradient(&local_grad);
+#ifdef HAVE_MPI
+      reduce(world, local_grad, g, std::plus<SparseVector<double> >(), 0);
+#else
+      g.swap(local_grad);
+#endif
+      local_grad.clear();
+      if (rank == 0) {
+        g /= (size_per_proc * size);
+        o->UpdateWeights(g, FD::NumFeats(), &x);
+      }
+#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;
+}