script to filter reachable sentences, weight cleanup

6 files changed, 174 insertions, 211 deletions

diff --git a/decoder/apply_models.cc b/decoder/apply_models.cc
index 26cdb881..40fd27e4 100644
--- a/decoder/apply_models.cc
+++ b/decoder/apply_models.cc
@@ -276,8 +276,7 @@ public:
     make_heap(cand.begin(), cand.end(), HeapCandCompare());
     State2Node state2node;   // "buf" in Figure 2
     int pops = 0;
-    int pop_limit_eff=max(1,int(v.promise*pop_limit_));
-    while(!cand.empty() && pops < pop_limit_eff) {
+    while(!cand.empty() && pops < pop_limit_) {
       pop_heap(cand.begin(), cand.end(), HeapCandCompare());
       Candidate* item = cand.back();
       cand.pop_back();
diff --git a/decoder/hg.h b/decoder/hg.h
index e5ef05f8..f0ddbb76 100644
--- a/decoder/hg.h
+++ b/decoder/hg.h
@@ -49,16 +49,14 @@ public:
   // TODO get rid of cat_?
   // TODO keep cat_ and add span and/or state? :)
   struct Node {
-    Node() : id_(), cat_(), promise(1) {}
+    Node() : id_(), cat_() {}
     int id_; // equal to this object's position in the nodes_ vector
     WordID cat_;  // non-terminal category if <0, 0 if not set
     WordID NT() const { return -cat_; }
     EdgesVector in_edges_;   // an in edge is an edge with this node as its head.  (in edges come from the bottom up to us)  indices in edges_
     EdgesVector out_edges_;  // an out edge is an edge with this node as its tail.  (out edges leave us up toward the top/goal). indices in edges_
-    double promise; // set in global pruning; in [0,infty) so that mean is 1.  use: e.g. scale cube poplimit.  //TODO: appears to be useless, compile without this?  on the other hand, pretty cheap.
     void copy_fixed(Node const& o) { // nonstructural fields only - structural ones are managed by sorting/pruning/subsetting
       cat_=o.cat_;
-      promise=o.promise;
     }
     void copy_reindex(Node const& o,indices_after const& n2,indices_after const& e2) {
       copy_fixed(o);
@@ -81,7 +79,7 @@ public:
     int head_node_;               // refers to a position in nodes_
     TailNodeVector tail_nodes_;   // contents refer to positions in nodes_
     TRulePtr rule_;
-    FeatureVector feature_values_;
+    SparseVector<weight_t> feature_values_;
     prob_t edge_prob_;             // dot product of weights and feat_values
     int id_;   // equal to this object's position in the edges_ vector
 
@@ -468,7 +466,7 @@ public:
   /// drop edge i if edge_margin[i] < prune_below, unless preserve_mask[i]
   void MarginPrune(EdgeProbs const& edge_margin,prob_t prune_below,EdgeMask const* preserve_mask=0,bool safe_inside=false,bool verbose=false);
 
-  //TODO: in my opinion, looking at the ratio of logprobs (features \dot weights) rather than the absolute difference generalizes more nicely across sentence lengths and weight vectors that are constant multiples of one another.  at least make that an option.  i worked around this a little in cdec by making "beam alpha per source word" but that's not helping with different tuning runs.  this would also make me more comfortable about allocating Node.promise
+  //TODO: in my opinion, looking at the ratio of logprobs (features \dot weights) rather than the absolute difference generalizes more nicely across sentence lengths and weight vectors that are constant multiples of one another.  at least make that an option.  i worked around this a little in cdec by making "beam alpha per source word" but that's not helping with different tuning runs.
 
   // beam_alpha=0 means don't beam prune, otherwise drop things that are e^beam_alpha times worse than best -   // prunes any edge whose prob_t on the best path taking that edge is more than e^alpha times
   //density=0 means don't density prune:   // for density>=1.0, keep this many times the edges needed for the 1best derivation
diff --git a/training/Makefile.am b/training/Makefile.am
index 7ceeda34..5752859e 100644
--- a/training/Makefile.am
+++ b/training/Makefile.am
@@ -9,9 +9,9 @@ bin_PROGRAMS = \
   atools \
   plftools \
   collapse_weights \
-  cllh_filter_grammar \
-  mpi_online_optimize \
+  mpi_extract_reachable \
   mpi_extract_features \
+  mpi_online_optimize \
   mpi_batch_optimize \
   compute_cllh \
   augment_grammar
@@ -25,6 +25,9 @@ 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_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
+
 mpi_extract_features_SOURCES = mpi_extract_features.cc
 mpi_extract_features_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
 
@@ -34,9 +37,6 @@ mpi_batch_optimize_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/
 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
 
-cllh_filter_grammar_SOURCES = cllh_filter_grammar.cc
-cllh_filter_grammar_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
-
 augment_grammar_SOURCES = augment_grammar.cc
 augment_grammar_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/cllh_filter_grammar.cc b/training/cllh_filter_grammar.cc
deleted file mode 100644
index 6998ec2b..00000000
--- a/training/cllh_filter_grammar.cc
+++ /dev/null
@@ -1,197 +0,0 @@
-#include <iostream>
-#include <vector>
-#include <cassert>
-#include <unistd.h>   // fork
-#include <sys/wait.h> // waitpid
-
-#include <boost/program_options.hpp>
-#include <boost/program_options/variables_map.hpp>
-
-#include "tdict.h"
-#include "ff_register.h"
-#include "verbose.h"
-#include "hg.h"
-#include "decoder.h"
-#include "filelib.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()
-        ("training_data,t",po::value<string>(),"Training data corpus")
-        ("decoder_config,c",po::value<string>(),"Decoder configuration file")
-        ("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")
-        ("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;
-    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;
-  assert(size > 0);
-  assert(rank < size);
-  while(in) {
-    getline(in, line);
-    if (!in) break;
-    if (lc % size == rank) {
-      c->push_back(line);
-      ids->push_back(lc);
-    }
-    ++lc;
-  }
-}
-
-struct TrainingObserver : public DecoderObserver {
-  TrainingObserver() : s_lhs(-TD::Convert("S")), goal_lhs(-TD::Convert("Goal")) {}
-
-  void Reset() {
-    total_complete = 0;
-  } 
-
-  virtual void NotifyDecodingStart(const SentenceMetadata& smeta) {
-    state = 1;
-    used.clear();
-    failed = true;
-  }
-
-  virtual void NotifyTranslationForest(const SentenceMetadata& smeta, Hypergraph* hg) {
-    assert(state == 1);
-    for (int i = 0; i < hg->edges_.size(); ++i) {
-      const TRule* rule = hg->edges_[i].rule_.get();
-      if (rule->lhs_ == s_lhs || rule->lhs_ == goal_lhs)  // fragile hack to filter out glue rules
-        continue;
-      used.insert(rule);
-    }
-    state = 2;
-  }
-
-  virtual void NotifyAlignmentForest(const SentenceMetadata& smeta, Hypergraph* hg) {
-    assert(state == 2);
-    state = 3;
-  }
-
-  virtual void NotifyDecodingComplete(const SentenceMetadata& smeta) {
-    if (state == 3) {
-      failed = false;
-    } else {
-      failed = true;
-    }
-  }
-
-  set<const TRule*> used;
-
-  const int s_lhs;
-  const int goal_lhs;
-  bool failed;
-  int total_complete;
-  int state;
-};
-
-void work(const string& fname, int rank, int size, Decoder* decoder) {
-  cerr << "Worker " << rank << '/' << size << " starting.\n";
-  vector<string> corpus;
-  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());
-  ostringstream og; og << "grammar." << rank << "_of_" << size << ".gz";
-  WriteFile fog(og.str());
-
-  set<const TRule*> all_used;
-  TrainingObserver observer;
-  for (int i = 0; i < corpus.size(); ++i) {
-    const int sent_id = ids[i];
-    const string& input = corpus[i];
-    decoder->SetId(sent_id);
-    decoder->Decode(input, &observer);
-    if (observer.failed) {
-      // do nothing
-    } else {
-      (*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;
-      }
-    }
-  }
-}
-
-int main(int argc, char** argv) {
-  register_feature_functions();
-
-  po::variables_map conf;
-  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());
-  if (decoder.GetConf()["input"].as<string>() != "-") {
-    cerr << "cdec.ini must not set an input file\n";
-    abort();
-  }
-  SetSilent(true);  // turn off verbose decoder output
-  cerr << "Forking " << ncpus << " time(s)\n";
-  vector<pid_t> children;
-  for (int i = 0; i < ncpus; ++i) {
-    pid_t pid = fork();
-    if (pid < 0) {
-      cerr << "Fork failed!\n";
-      exit(1);
-    }
-    if (pid > 0) {
-      children.push_back(pid);
-    } else {
-      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);
-    }
-  }
-  for (int i = 0; i < children.size(); ++i) {
-    int status;
-    int w = waitpid(children[i], &status, 0);
-    if (w < 0) { cerr << "Error while waiting for children!"; return 1; }
-    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/mpi_extract_reachable.cc b/training/mpi_extract_reachable.cc
new file mode 100644
index 00000000..2a7c2b9d
--- /dev/null
+++ b/training/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/utils/feature_vector.h b/utils/feature_vector.h
index 733aa99e..a7b61a66 100755
--- a/utils/feature_vector.h
+++ b/utils/feature_vector.h
@@ -3,9 +3,9 @@
 
 #include <vector>
 #include "sparse_vector.h"
-#include "fdict.h"
+#include "weights.h"
 
-typedef double Featval;
+typedef weight_t Featval;
 typedef SparseVector<Featval> FeatureVector;
 typedef SparseVector<Featval> WeightVector;
 typedef std::vector<Featval> DenseWeightVector;