compute obj, fixes for grammar filter

git-svn-id: https://ws10smt.googlecode.com/svn/trunk@668 ec762483-ff6d-05da-a07a-a48fb63a330f

4 files changed, 225 insertions, 13 deletions

diff --git a/training/Makefile.am b/training/Makefile.am
index 83c15ecc..b3f93529 100644
--- a/training/Makefile.am
+++ b/training/Makefile.am
@@ -20,13 +20,17 @@ TESTS = lbfgs_test optimize_test
 
 if MPI
 bin_PROGRAMS += mpi_batch_optimize \
-             mpi_online_optimize
+             mpi_online_optimize \
+             compute_cllh
 
 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 -lz
 
 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 -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 -lz
 endif
 
 cllh_filter_grammar_SOURCES = cllh_filter_grammar.cc
diff --git a/training/cllh_filter_grammar.cc b/training/cllh_filter_grammar.cc
index 90fe9fba..6998ec2b 100644
--- a/training/cllh_filter_grammar.cc
+++ b/training/cllh_filter_grammar.cc
@@ -22,7 +22,10 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   opts.add_options()
         ("training_data,t",po::value<string>(),"Training data corpus")
         ("decoder_config,c",po::value<string>(),"Decoder configuration file")
-        ("ncpus,n",po::value<unsigned>()->default_value(1),"Number of CPUs to use");
+        ("shards,s",po::value<unsigned>()->default_value(1),"Number of shards")
+        ("starting_shard,S",po::value<unsigned>()->default_value(0), "In this invocation only process shards >= S")
+        ("work_limit,l",po::value<unsigned>()->default_value(9999), "Process maximially this many shards")
+        ("ncpus,C",po::value<unsigned>()->default_value(1),"Number of CPUs to use");
   po::options_description clo("Command line options");
   clo.add_options()
         ("config", po::value<string>(), "Configuration file")
@@ -49,6 +52,8 @@ void ReadTrainingCorpus(const string& fname, int rank, int size, vector<string>*
   istream& in = *rf.stream();
   string line;
   int lc = 0;
+  assert(size > 0);
+  assert(rank < size);
   while(in) {
     getline(in, line);
     if (!in) break;
@@ -112,6 +117,7 @@ void work(const string& fname, int rank, int size, Decoder* decoder) {
   vector<int> ids;
   ReadTrainingCorpus(fname, rank, size, &corpus, &ids);
   assert(corpus.size() > 0);
+  assert(corpus.size() == ids.size());
   cerr << "  " << rank << '/' << size << ": has " << corpus.size() << " sentences to process\n";
   ostringstream oc; oc << "corpus." << rank << "_of_" << size;
   WriteFile foc(oc.str());
@@ -121,13 +127,14 @@ void work(const string& fname, int rank, int size, Decoder* decoder) {
   set<const TRule*> all_used;
   TrainingObserver observer;
   for (int i = 0; i < corpus.size(); ++i) {
-    int ex_num = ids[i];
-    decoder->SetId(ex_num);
-    decoder->Decode(corpus[ex_num], &observer);
+    const int sent_id = ids[i];
+    const string& input = corpus[i];
+    decoder->SetId(sent_id);
+    decoder->Decode(input, &observer);
     if (observer.failed) {
-      (*foc.stream()) << "*** id=" << ex_num << " is unreachable\n";
+      // do nothing
     } else {
-      (*foc.stream()) << corpus[ex_num] << endl;
+      (*foc.stream()) << input << endl;
       for (set<const TRule*>::iterator it = observer.used.begin(); it != observer.used.end(); ++it) {
         if (all_used.insert(*it).second)
           (*fog.stream()) << **it << endl;
@@ -143,6 +150,11 @@ int main(int argc, char** argv) {
   InitCommandLine(argc, argv, &conf);
   const string fname = conf["training_data"].as<string>();
   const unsigned ncpus = conf["ncpus"].as<unsigned>();
+  const unsigned shards = conf["shards"].as<unsigned>();
+  const unsigned start = conf["starting_shard"].as<unsigned>();
+  const unsigned work_limit = conf["work_limit"].as<unsigned>();
+  const unsigned eff_shards = min(start + work_limit, shards);
+  cerr << "Processing shards " << start << "/" << shards << " to " << eff_shards << "/" << shards << endl;
   assert(ncpus > 0);
   ReadFile ini_rf(conf["decoder_config"].as<string>());
   Decoder decoder(ini_rf.stream());
@@ -162,8 +174,13 @@ int main(int argc, char** argv) {
     if (pid > 0) {
       children.push_back(pid);
     } else {
-      work(fname, i, ncpus, &decoder);
-      cerr << "  " << i << "/" << ncpus << " finished.\n";
+      for (int j = start; j < eff_shards; ++j) {
+        if (j % ncpus == i) {
+          cerr << "  CPU " << i << " processing shard " << j << endl;
+          work(fname, j, shards, &decoder);
+          cerr << "  Shard " << j << "/" << shards << " finished.\n";
+        }
+      }
       _exit(0);
     }
   }
@@ -171,7 +188,10 @@ int main(int argc, char** argv) {
     int status;
     int w = waitpid(children[i], &status, 0);
     if (w < 0) { cerr << "Error while waiting for children!"; return 1; }
-    cerr << "Child " << i << ": status=" << status << " sig?=" << WIFSIGNALED(status) << " sig=" << WTERMSIG(status) << endl;
+    if (WIFSIGNALED(status)) {
+      cerr << "Child " << i << " received signal " << WTERMSIG(status) << endl;
+      if (WTERMSIG(status) == 11) { cerr << " this is a SEGV- you may be trying to print temporarily created rules\n"; }
+    }
   }
   return 0;
 }
diff --git a/training/compute_cllh.cc b/training/compute_cllh.cc
new file mode 100644
index 00000000..f25e17c3
--- /dev/null
+++ b/training/compute_cllh.cc
@@ -0,0 +1,185 @@
+#include <sstream>
+#include <iostream>
+#include <fstream>
+#include <vector>
+#include <cassert>
+#include <cmath>
+
+#include <mpi.h>
+#include <boost/mpi.hpp>
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#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;
+
+void 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;
+    MPI::Finalize();
+    exit(1);
+  }
+}
+
+void ReadTrainingCorpus(const string& fname, int rank, int size, vector<string>* c, vector<int>* ids) {
+  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);
+      ids->push_back(lc);
+    }
+    ++lc;
+  }
+}
+
+static const double kMINUS_EPSILON = -1e-6;
+
+struct TrainingObserver : public DecoderObserver {
+  void Reset() {
+    acc_obj = 0;
+  } 
+
+  virtual void NotifyDecodingStart(const SentenceMetadata&) {
+    cur_obj = 0;
+    state = 1;
+  }
+
+  // compute model expectations, denominator of objective
+  virtual void 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);
+  }
+
+  // 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);
+
+    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(!isnan(log_ref_z));
+    acc_obj += (cur_obj - log_ref_z);
+  }
+
+  double acc_obj;
+  double cur_obj;
+  int state;
+};
+
+namespace mpi = boost::mpi;
+
+int main(int argc, char** argv) {
+  mpi::environment env(argc, argv);
+  mpi::communicator world;
+  const int size = world.size(); 
+  const int rank = world.rank();
+  if (size > 1) SetSilent(true);  // turn off verbose decoder output
+  register_feature_functions();
+
+  po::variables_map conf;
+  InitCommandLine(argc, argv, &conf);
+
+  // 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());
+  if (decoder.GetConf()["input"].as<string>() != "-") {
+    cerr << "cdec.ini must not set an input file\n";
+    abort();
+  }
+
+  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)
+    cerr << "Each processor is decoding " << corpus.size() << " training examples...\n";
+
+  for (int i = 0; i < corpus.size(); ++i) {
+    decoder.SetId(ids[i]);
+    decoder.Decode(corpus[i], &observer);
+  }
+
+  reduce(world, observer.acc_obj, objective, std::plus<double>(), 0);
+
+  if (rank == 0)
+    cout << "OBJECTIVE: " << objective << endl;
+
+  return 0;
+}
diff --git a/training/model1.cc b/training/model1.cc
index 92a70985..3e27689f 100644
--- a/training/model1.cc
+++ b/training/model1.cc
@@ -29,8 +29,8 @@ int main(int argc, char** argv) {
     double denom = 0.0;
     int lc = 0;
     bool flag = false;
+    string line;
     while(true) {
-      string line;
       getline(in, line);
       if (!in) break;
       ++lc;
@@ -41,8 +41,11 @@ int main(int argc, char** argv) {
       Lattice src, trg;
       LatticeTools::ConvertTextToLattice(ssrc, &src);
       LatticeTools::ConvertTextToLattice(strg, &trg);
-      assert(src.size() > 0);
-      assert(trg.size() > 0);
+      if (src.size() == 0 || trg.size() == 0) {
+        cerr << "Error: " << lc << "\n" << line << endl;
+        assert(src.size() > 0);
+        assert(trg.size() > 0);
+      }
       denom += 1.0;
       vector<double> probs(src.size() + 1);
       for (int j = 0; j < trg.size(); ++j) {