Merge remote-tracking branch 'upstream/master'

25 files changed, 899 insertions, 73 deletions

diff --git a/gi/pf/Makefile.am b/gi/pf/Makefile.am
index f9c979d0..86f8e07b 100644
--- a/gi/pf/Makefile.am
+++ b/gi/pf/Makefile.am
@@ -1,8 +1,14 @@
-bin_PROGRAMS = cbgi brat dpnaive pfbrat pfdist itg pfnaive condnaive align-lexonly-pyp learn_cfg pyp_lm nuisance_test align-tl
+bin_PROGRAMS = cbgi brat dpnaive pfbrat pfdist itg pfnaive condnaive align-lexonly-pyp learn_cfg pyp_lm nuisance_test align-tl pf_test bayes_lattice_score
 
 noinst_LIBRARIES = libpf.a
 
-libpf_a_SOURCES = base_distributions.cc reachability.cc cfg_wfst_composer.cc corpus.cc unigrams.cc ngram_base.cc transliterations.cc backward.cc pyp_word_model.cc pyp_tm.cc
+libpf_a_SOURCES = base_distributions.cc reachability.cc cfg_wfst_composer.cc corpus.cc unigrams.cc ngram_base.cc transliterations.cc backward.cc hpyp_tm.cc pyp_tm.cc
+
+bayes_lattice_score_SOURCES = bayes_lattice_score.cc
+bayes_lattice_score_LDADD = libpf.a $(top_srcdir)/decoder/libcdec.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
+
+pf_test_SOURCES = pf_test.cc
+pf_test_LDADD = libpf.a $(top_srcdir)/decoder/libcdec.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
 
 nuisance_test_SOURCES = nuisance_test.cc
 nuisance_test_LDADD = libpf.a $(top_srcdir)/decoder/libcdec.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
diff --git a/gi/pf/align-lexonly-pyp.cc b/gi/pf/align-lexonly-pyp.cc
index 942dcf51..e7509f57 100644
--- a/gi/pf/align-lexonly-pyp.cc
+++ b/gi/pf/align-lexonly-pyp.cc
@@ -11,6 +11,7 @@
 #include "sampler.h"
 #include "corpus.h"
 #include "pyp_tm.h"
+#include "hpyp_tm.h"
 #include "quasi_model2.h"
 
 using namespace std;
@@ -61,15 +62,17 @@ struct AlignedSentencePair {
   Array2D<short> posterior;
 };
 
+template <class LexicalTranslationModel>
 struct Aligner {
   Aligner(const vector<vector<WordID> >& lets,
+          int vocab_size,
           int num_letters,
           const po::variables_map& conf,
           vector<AlignedSentencePair>* c) :
       corpus(*c),
       paj_model(conf["align_alpha"].as<double>(), conf["p_null"].as<double>()),
       infer_paj(conf.count("infer_alignment_hyperparameters") > 0),
-      model(lets, num_letters),
+      model(lets, vocab_size, num_letters),
       kNULL(TD::Convert("NULL")) {
     assert(lets[kNULL].size() == 0);
   }
@@ -77,7 +80,7 @@ struct Aligner {
   vector<AlignedSentencePair>& corpus;
   QuasiModel2 paj_model;
   const bool infer_paj;
-  PYPLexicalTranslation model;
+  LexicalTranslationModel model;
   const WordID kNULL;
 
   void ResampleHyperparameters() {
@@ -217,7 +220,8 @@ int main(int argc, char** argv) {
   ExtractLetters(vocabf, &letters, NULL);
   letters[TD::Convert("NULL")].clear();
 
-  Aligner aligner(letters, letset.size(), conf, &corpus);
+  //Aligner<PYPLexicalTranslation> aligner(letters, vocabe.size(), letset.size(), conf, &corpus);
+  Aligner<HPYPLexicalTranslation> aligner(letters, vocabe.size(), letset.size(), conf, &corpus);
   aligner.InitializeRandom();
 
   const unsigned samples = conf["samples"].as<unsigned>();
diff --git a/gi/pf/align-tl.cc b/gi/pf/align-tl.cc
index cbe8c6c8..f6608f1d 100644
--- a/gi/pf/align-tl.cc
+++ b/gi/pf/align-tl.cc
@@ -58,7 +58,7 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   }
 }
 
-shared_ptr<MT19937> prng;
+boost::shared_ptr<MT19937> prng;
 
 struct LexicalAlignment {
   unsigned char src_index;
diff --git a/gi/pf/bayes_lattice_score.cc b/gi/pf/bayes_lattice_score.cc
new file mode 100644
index 00000000..70cb8dc2
--- /dev/null
+++ b/gi/pf/bayes_lattice_score.cc
@@ -0,0 +1,309 @@
+#include <iostream>
+#include <queue>
+
+#include <boost/functional.hpp>
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#include "inside_outside.h"
+#include "hg.h"
+#include "hg_io.h"
+#include "bottom_up_parser.h"
+#include "fdict.h"
+#include "grammar.h"
+#include "m.h"
+#include "trule.h"
+#include "tdict.h"
+#include "filelib.h"
+#include "dict.h"
+#include "sampler.h"
+#include "ccrp.h"
+#include "ccrp_onetable.h"
+
+using namespace std;
+using namespace tr1;
+namespace po = boost::program_options;
+
+boost::shared_ptr<MT19937> prng;
+
+void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
+  po::options_description opts("Configuration options");
+  opts.add_options()
+        ("samples,s",po::value<unsigned>()->default_value(1000),"Number of samples")
+        ("input,i",po::value<string>(),"Read parallel data from")
+        ("random_seed,S",po::value<uint32_t>(), "Random seed");
+  po::options_description clo("Command line options");
+  clo.add_options()
+        ("config", po::value<string>(), "Configuration file")
+        ("help", "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") == 0)) {
+    cerr << dcmdline_options << endl;
+    exit(1);
+  }
+}
+
+unsigned ReadCorpus(const string& filename,
+                    vector<Lattice>* e,
+                    set<WordID>* vocab_e) {
+  e->clear();
+  vocab_e->clear();
+  ReadFile rf(filename);
+  istream* in = rf.stream();
+  assert(*in);
+  string line;
+  unsigned toks = 0;
+  while(*in) {
+    getline(*in, line);
+    if (line.empty() && !*in) break;
+    e->push_back(Lattice());
+    Lattice& le = e->back();
+    LatticeTools::ConvertTextOrPLF(line, & le);
+    for (unsigned i = 0; i < le.size(); ++i)
+      for (unsigned j = 0; j < le[i].size(); ++j)
+        vocab_e->insert(le[i][j].label);
+    toks += le.size();
+  }
+  return toks;
+}
+
+struct BaseModel {
+  explicit BaseModel(unsigned tc) :
+      unif(1.0 / tc), p(prob_t::One()) {}
+  prob_t prob(const TRule& r) const {
+    return unif;
+  }
+  void increment(const TRule& r, MT19937* rng) {
+    p *= prob(r);
+  }
+  void decrement(const TRule& r, MT19937* rng) {
+    p /= prob(r);
+  }
+  prob_t Likelihood() const {
+    return p;
+  }
+  const prob_t unif;
+  prob_t p;
+};
+
+struct UnigramModel {
+  explicit UnigramModel(unsigned tc) : base(tc), crp(1,1,1,1), glue(1,1,1,1) {}
+  BaseModel base;
+  CCRP<TRule> crp;
+  CCRP<TRule> glue;
+
+  prob_t Prob(const TRule& r) const {
+    if (r.Arity() != 0) {
+      return glue.prob(r, prob_t(0.5));
+    }
+    return crp.prob(r, base.prob(r));
+  }
+
+  int Increment(const TRule& r, MT19937* rng) {
+    if (r.Arity() != 0) {
+      glue.increment(r, 0.5, rng);
+      return 0;
+    } else {
+      if (crp.increment(r, base.prob(r), rng)) {
+        base.increment(r, rng);
+        return 1;
+      }
+      return 0;
+    }
+  }
+
+  int Decrement(const TRule& r, MT19937* rng) {
+    if (r.Arity() != 0) {
+      glue.decrement(r, rng);
+      return 0;
+    } else {
+      if (crp.decrement(r, rng)) {
+        base.decrement(r, rng);
+        return -1;
+      }
+      return 0;
+    }
+  }
+
+  prob_t Likelihood() const {
+    prob_t p;
+    p.logeq(crp.log_crp_prob() + glue.log_crp_prob());
+    p *= base.Likelihood();
+    return p;
+  }
+
+  void ResampleHyperparameters(MT19937* rng) {
+    crp.resample_hyperparameters(rng);
+    glue.resample_hyperparameters(rng);
+    cerr << " d=" << crp.discount() << ", s=" << crp.strength() << "\t STOP d=" << glue.discount() << ", s=" << glue.strength() << endl;
+  }
+};
+
+UnigramModel* plm;
+
+void SampleDerivation(const Hypergraph& hg, MT19937* rng, vector<unsigned>* sampled_deriv) {
+  vector<prob_t> node_probs;
+  Inside<prob_t, EdgeProb>(hg, &node_probs);
+  queue<unsigned> q;
+  q.push(hg.nodes_.size() - 2);
+  while(!q.empty()) {
+    unsigned cur_node_id = q.front();
+//    cerr << "NODE=" << cur_node_id << endl;
+    q.pop();
+    const Hypergraph::Node& node = hg.nodes_[cur_node_id];
+    const unsigned num_in_edges = node.in_edges_.size();
+    unsigned sampled_edge = 0;
+    if (num_in_edges == 1) {
+      sampled_edge = node.in_edges_[0];
+    } else {
+      //prob_t z;
+      assert(num_in_edges > 1);
+      SampleSet<prob_t> ss;
+      for (unsigned j = 0; j < num_in_edges; ++j) {
+        const Hypergraph::Edge& edge = hg.edges_[node.in_edges_[j]];
+        prob_t p = edge.edge_prob_;
+        for (unsigned k = 0; k < edge.tail_nodes_.size(); ++k)
+          p *= node_probs[edge.tail_nodes_[k]];
+        ss.add(p);
+//        cerr << log(ss[j]) << " ||| " << edge.rule_->AsString() << endl;
+        //z += p;
+      }
+//      for (unsigned j = 0; j < num_in_edges; ++j) {
+//        const Hypergraph::Edge& edge = hg.edges_[node.in_edges_[j]];
+//        cerr << exp(log(ss[j] / z)) << " ||| " << edge.rule_->AsString() << endl;
+//      }
+//      cerr << " --- \n";
+      sampled_edge = node.in_edges_[rng->SelectSample(ss)];
+    }
+    sampled_deriv->push_back(sampled_edge);
+    const Hypergraph::Edge& edge = hg.edges_[sampled_edge];
+    for (unsigned j = 0; j < edge.tail_nodes_.size(); ++j) {
+      q.push(edge.tail_nodes_[j]);
+    }
+  }
+//  for (unsigned i = 0; i < sampled_deriv->size(); ++i) {
+//    cerr << *hg.edges_[(*sampled_deriv)[i]].rule_ << endl;
+//  }
+}
+
+void IncrementDerivation(const Hypergraph& hg, const vector<unsigned>& d, UnigramModel* plm, MT19937* rng) {
+  for (unsigned i = 0; i < d.size(); ++i)
+    plm->Increment(*hg.edges_[d[i]].rule_, rng);
+}
+
+void DecrementDerivation(const Hypergraph& hg, const vector<unsigned>& d, UnigramModel* plm, MT19937* rng) {
+  for (unsigned i = 0; i < d.size(); ++i)
+    plm->Decrement(*hg.edges_[d[i]].rule_, rng);
+}
+
+prob_t TotalProb(const Hypergraph& hg) {
+  return Inside<prob_t, EdgeProb>(hg);
+}
+
+void IncrementLatticePath(const Hypergraph& hg, const vector<unsigned>& d, Lattice* pl) {
+  Lattice& lat = *pl;
+  for (int i = 0; i < d.size(); ++i) {
+    const Hypergraph::Edge& edge = hg.edges_[d[i]];
+    if (edge.rule_->Arity() != 0) continue;
+    WordID sym = edge.rule_->e_[0];
+    vector<LatticeArc>& las = lat[edge.i_];
+    int dist = edge.j_ - edge.i_;
+    assert(dist > 0);
+    for (int j = 0; j < las.size(); ++j) {
+      if (las[j].dist2next == dist &&
+          las[j].label == sym) {
+        las[j].cost += 1;
+      }
+    }
+  }
+}
+
+int main(int argc, char** argv) {
+  po::variables_map conf;
+
+  InitCommandLine(argc, argv, &conf);
+  vector<GrammarPtr> grammars(2);
+  grammars[0].reset(new GlueGrammar("S","X"));
+  const unsigned samples = conf["samples"].as<unsigned>();
+
+  if (conf.count("random_seed"))
+    prng.reset(new MT19937(conf["random_seed"].as<uint32_t>()));
+  else
+    prng.reset(new MT19937);
+  MT19937& rng = *prng;
+  vector<Lattice> corpuse;
+  set<WordID> vocabe;
+  cerr << "Reading corpus...\n";
+  const unsigned toks = ReadCorpus(conf["input"].as<string>(), &corpuse, &vocabe);
+  cerr << "E-corpus size: " << corpuse.size() << " lattices\t (" << vocabe.size() << " word types)\n";
+  UnigramModel lm(vocabe.size());
+  vector<Hypergraph> hgs(corpuse.size());
+  vector<vector<unsigned> > derivs(corpuse.size());
+  for (int i = 0; i < corpuse.size(); ++i) {
+    grammars[1].reset(new PassThroughGrammar(corpuse[i], "X"));
+    ExhaustiveBottomUpParser parser("S", grammars);
+    bool res = parser.Parse(corpuse[i], &hgs[i]);  // exhaustive parse
+    assert(res);
+  }
+
+  double csamples = 0;
+  for (int SS=0; SS < samples; ++SS) {
+    const bool is_last = ((samples - 1) == SS);
+    prob_t dlh = prob_t::One();
+    bool record_sample = (SS > (samples * 1 / 3) && (SS % 5 == 3));
+    if (record_sample) csamples++;
+    for (int ci = 0; ci < corpuse.size(); ++ci) {
+      Lattice& lat = corpuse[ci];
+      Hypergraph& hg = hgs[ci];
+      vector<unsigned>& d = derivs[ci];
+      if (!is_last) DecrementDerivation(hg, d, &lm, &rng);
+      for (unsigned i = 0; i < hg.edges_.size(); ++i) {
+        TRule& r = *hg.edges_[i].rule_;
+        if (r.Arity() != 0)
+          hg.edges_[i].edge_prob_ = prob_t::One();
+        else
+          hg.edges_[i].edge_prob_ = lm.Prob(r);
+      }
+      if (!is_last) {
+        d.clear();
+        SampleDerivation(hg, &rng, &d);
+        IncrementDerivation(hg, derivs[ci], &lm, &rng);
+      } else {
+        prob_t p = TotalProb(hg);
+        dlh *= p;
+        cerr << " p(sentence) = " << log(p) << "\t" << log(dlh) << endl;
+      }
+      if (record_sample) IncrementLatticePath(hg, derivs[ci], &lat);
+    }
+    double llh = log(lm.Likelihood());
+    cerr << "LLH=" << llh << "\tENTROPY=" << (-llh / log(2) / toks) << "\tPPL=" << pow(2, -llh / log(2) / toks) << endl;
+    if (SS % 10 == 9) lm.ResampleHyperparameters(&rng);
+    if (is_last) {
+      double z = log(dlh);
+      cerr << "TOTAL_PROB=" << z << "\tENTROPY=" << (-z / log(2) / toks) << "\tPPL=" << pow(2, -z / log(2) / toks) << endl;
+    }
+  }
+  cerr << lm.crp << endl;
+  cerr << lm.glue << endl;
+  for (int i = 0; i < corpuse.size(); ++i) {
+    for (int j = 0; j < corpuse[i].size(); ++j)
+      for (int k = 0; k < corpuse[i][j].size(); ++k) {
+        corpuse[i][j][k].cost /= csamples;
+        corpuse[i][j][k].cost += 1e-3;
+        corpuse[i][j][k].cost = log(corpuse[i][j][k].cost);
+      }
+    cout << HypergraphIO::AsPLF(corpuse[i]) << endl;
+  }
+  return 0;
+}
+
diff --git a/gi/pf/brat.cc b/gi/pf/brat.cc
index c2c52760..832f22cf 100644
--- a/gi/pf/brat.cc
+++ b/gi/pf/brat.cc
@@ -489,7 +489,7 @@ int main(int argc, char** argv) {
     cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n";
     return 1;
   }
-  shared_ptr<MT19937> prng;
+  boost::shared_ptr<MT19937> prng;
   if (conf.count("random_seed"))
     prng.reset(new MT19937(conf["random_seed"].as<uint32_t>()));
   else
diff --git a/gi/pf/cfg_wfst_composer.cc b/gi/pf/cfg_wfst_composer.cc
index a31b5be8..20520c81 100644
--- a/gi/pf/cfg_wfst_composer.cc
+++ b/gi/pf/cfg_wfst_composer.cc
@@ -16,7 +16,6 @@
 #include "tdict.h"
 #include "hg.h"
 
-using boost::shared_ptr;
 namespace po = boost::program_options;
 using namespace std;
 using namespace std::tr1;
@@ -114,7 +113,7 @@ struct Edge {
   const Edge* const active_parent;    // back pointer, NULL for PREDICT items
   const Edge* const passive_parent;   // back pointer, NULL for SCAN and PREDICT items
   TRulePtr tps;   // translations
-  shared_ptr<SparseVector<double> > features; // features from CFG rule
+  boost::shared_ptr<SparseVector<double> > features; // features from CFG rule
 
   bool IsPassive() const {
     // when a rule is completed, this value will be set
diff --git a/gi/pf/condnaive.cc b/gi/pf/condnaive.cc
index 3ea88016..419731ac 100644
--- a/gi/pf/condnaive.cc
+++ b/gi/pf/condnaive.cc
@@ -55,7 +55,7 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   }
 }
 
-shared_ptr<MT19937> prng;
+boost::shared_ptr<MT19937> prng;
 
 struct ModelAndData {
   explicit ModelAndData(ConditionalParallelSegementationModel<PhraseConditionalBase>& m, const vector<vector<int> >& ce, const vector<vector<int> >& cf, const set<int>& ve, const set<int>& vf) :
diff --git a/gi/pf/dpnaive.cc b/gi/pf/dpnaive.cc
index 469dff5c..75ccad72 100644
--- a/gi/pf/dpnaive.cc
+++ b/gi/pf/dpnaive.cc
@@ -55,7 +55,7 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   }
 }
 
-shared_ptr<MT19937> prng;
+boost::shared_ptr<MT19937> prng;
 
 template <typename Base>
 struct ModelAndData {
diff --git a/gi/pf/hpyp_tm.cc b/gi/pf/hpyp_tm.cc
new file mode 100644
index 00000000..784f9958
--- /dev/null
+++ b/gi/pf/hpyp_tm.cc
@@ -0,0 +1,133 @@
+#include "hpyp_tm.h"
+
+#include <tr1/unordered_map>
+#include <iostream>
+#include <queue>
+
+#include "tdict.h"
+#include "ccrp.h"
+#include "pyp_word_model.h"
+#include "tied_resampler.h"
+
+using namespace std;
+using namespace std::tr1;
+
+struct FreqBinner {
+  FreqBinner(const std::string& fname) { fd_.Load(fname); }
+  unsigned NumberOfBins() const { return fd_.Max() + 1; }
+  unsigned Bin(const WordID& w) const { return fd_.LookUp(w); }
+  FreqDict<unsigned> fd_;
+};
+
+template <typename Base, class Binner = FreqBinner>
+struct ConditionalPYPWordModel {
+  ConditionalPYPWordModel(Base* b, const Binner* bnr = NULL) :
+      base(*b),
+      binner(bnr),
+      btr(binner ? binner->NumberOfBins() + 1u : 2u) {}
+
+  void Summary() const {
+    cerr << "Number of conditioning contexts: " << r.size() << endl;
+    for (RuleModelHash::const_iterator it = r.begin(); it != r.end(); ++it) {
+      cerr << TD::Convert(it->first) << "   \tPYP(d=" << it->second.discount() << ",s=" << it->second.strength() << ") --------------------------" << endl;
+      for (CCRP<vector<WordID> >::const_iterator i2 = it->second.begin(); i2 != it->second.end(); ++i2)
+        cerr << "   " << i2->second.total_dish_count_ << '\t' << TD::GetString(i2->first) << endl;
+    }
+  }
+
+  void ResampleHyperparameters(MT19937* rng) {
+    btr.ResampleHyperparameters(rng);
+  } 
+
+  prob_t Prob(const WordID src, const vector<WordID>& trglets) const {
+    RuleModelHash::const_iterator it = r.find(src);
+    if (it == r.end()) {
+      return base(trglets);
+    } else {
+      return it->second.prob(trglets, base(trglets));
+    }
+  }
+
+  void Increment(const WordID src, const vector<WordID>& trglets, MT19937* rng) {
+    RuleModelHash::iterator it = r.find(src);
+    if (it == r.end()) {
+      it = r.insert(make_pair(src, CCRP<vector<WordID> >(0.5,1.0))).first;
+      static const WordID kNULL = TD::Convert("NULL");
+      unsigned bin = (src == kNULL ? 0 : 1);
+      if (binner && bin) { bin = binner->Bin(src) + 1; }
+      btr.Add(bin, &it->second);
+    }
+    if (it->second.increment(trglets, base(trglets), rng))
+      base.Increment(trglets, rng);
+  }
+
+  void Decrement(const WordID src, const vector<WordID>& trglets, MT19937* rng) {
+    RuleModelHash::iterator it = r.find(src);
+    assert(it != r.end());
+    if (it->second.decrement(trglets, rng)) {
+      base.Decrement(trglets, rng);
+    }
+  }
+
+  prob_t Likelihood() const {
+    prob_t p = prob_t::One();
+    for (RuleModelHash::const_iterator it = r.begin(); it != r.end(); ++it) {
+      prob_t q; q.logeq(it->second.log_crp_prob());
+      p *= q;
+    }
+    return p;
+  }
+
+  unsigned UniqueConditioningContexts() const {
+    return r.size();
+  }
+
+  // TODO tie PYP hyperparameters based on source word frequency bins
+  Base& base;
+  const Binner* binner;
+  BinTiedResampler<CCRP<vector<WordID> > > btr;
+  typedef unordered_map<WordID, CCRP<vector<WordID> > > RuleModelHash;
+  RuleModelHash r;
+};
+
+HPYPLexicalTranslation::HPYPLexicalTranslation(const vector<vector<WordID> >& lets,
+                                               const unsigned vocab_size,
+                                               const unsigned num_letters) :
+    letters(lets),
+    base(vocab_size, num_letters, 5),
+    up0(new PYPWordModel<PoissonUniformWordModel>(&base)),
+    tmodel(new ConditionalPYPWordModel<PYPWordModel<PoissonUniformWordModel> >(up0, new FreqBinner("10k.freq"))),
+    kX(-TD::Convert("X")) {}
+
+void HPYPLexicalTranslation::Summary() const {
+  tmodel->Summary();
+  up0->Summary();
+}
+
+prob_t HPYPLexicalTranslation::Likelihood() const {
+  prob_t p = up0->Likelihood();
+  p *= tmodel->Likelihood();
+  return p;
+}
+
+void HPYPLexicalTranslation::ResampleHyperparameters(MT19937* rng) {
+  tmodel->ResampleHyperparameters(rng);
+  up0->ResampleHyperparameters(rng);
+}
+
+unsigned HPYPLexicalTranslation::UniqueConditioningContexts() const {
+  return tmodel->UniqueConditioningContexts();
+}
+
+prob_t HPYPLexicalTranslation::Prob(WordID src, WordID trg) const {
+  return tmodel->Prob(src, letters[trg]);
+}
+
+void HPYPLexicalTranslation::Increment(WordID src, WordID trg, MT19937* rng) {
+  tmodel->Increment(src, letters[trg], rng);
+}
+
+void HPYPLexicalTranslation::Decrement(WordID src, WordID trg, MT19937* rng) {
+  tmodel->Decrement(src, letters[trg], rng);
+}
+
diff --git a/gi/pf/hpyp_tm.h b/gi/pf/hpyp_tm.h
new file mode 100644
index 00000000..af3215ba
--- /dev/null
+++ b/gi/pf/hpyp_tm.h
@@ -0,0 +1,38 @@
+#ifndef HPYP_LEX_TRANS
+#define HPYP_LEX_TRANS
+
+#include <vector>
+#include "wordid.h"
+#include "prob.h"
+#include "sampler.h"
+#include "freqdict.h"
+#include "poisson_uniform_word_model.h"
+
+struct FreqBinner;
+template <class B> struct PYPWordModel;
+template <typename T, class B> struct ConditionalPYPWordModel;
+
+struct HPYPLexicalTranslation {
+  explicit HPYPLexicalTranslation(const std::vector<std::vector<WordID> >& lets,
+                                 const unsigned vocab_size,
+                                 const unsigned num_letters);
+
+  prob_t Likelihood() const;
+
+  void ResampleHyperparameters(MT19937* rng);
+  prob_t Prob(WordID src, WordID trg) const;  // return p(trg | src)
+  void Summary() const;
+  void Increment(WordID src, WordID trg, MT19937* rng);
+  void Decrement(WordID src, WordID trg, MT19937* rng);
+  unsigned UniqueConditioningContexts() const;
+
+ private:
+  const std::vector<std::vector<WordID> >& letters;   // spelling dictionary
+  PoissonUniformWordModel base;  // "generator" of English types
+  PYPWordModel<PoissonUniformWordModel>* up0;  // model English lexicon
+  ConditionalPYPWordModel<PYPWordModel<PoissonUniformWordModel>, FreqBinner>* tmodel;  // translation distributions
+                      // (model English word | French word)
+  const WordID kX;
+};
+
+#endif
diff --git a/gi/pf/itg.cc b/gi/pf/itg.cc
index a38fe672..29ec3860 100644
--- a/gi/pf/itg.cc
+++ b/gi/pf/itg.cc
@@ -231,7 +231,7 @@ int main(int argc, char** argv) {
     cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n";
     return 1;
   }
-  shared_ptr<MT19937> prng;
+  boost::shared_ptr<MT19937> prng;
   if (conf.count("random_seed"))
     prng.reset(new MT19937(conf["random_seed"].as<uint32_t>()));
   else
diff --git a/gi/pf/learn_cfg.cc b/gi/pf/learn_cfg.cc
index ed1772bf..44eaa162 100644
--- a/gi/pf/learn_cfg.cc
+++ b/gi/pf/learn_cfg.cc
@@ -24,7 +24,7 @@ using namespace std;
 using namespace tr1;
 namespace po = boost::program_options;
 
-shared_ptr<MT19937> prng;
+boost::shared_ptr<MT19937> prng;
 vector<int> nt_vocab;
 vector<int> nt_id_to_index;
 static unsigned kMAX_RULE_SIZE = 0;
diff --git a/gi/pf/mh_test.cc b/gi/pf/mh_test.cc
new file mode 100644
index 00000000..296e7285
--- /dev/null
+++ b/gi/pf/mh_test.cc
@@ -0,0 +1,148 @@
+#include "ccrp.h"
+
+#include <vector>
+#include <iostream>
+
+#include "tdict.h"
+#include "transliterations.h"
+
+using namespace std;
+
+MT19937 rng;
+
+static bool verbose = false;
+
+struct Model {
+
+  Model() : bp(), base(0.2, 0.6) , ccrps(5, CCRP<int>(0.8, 0.5)) {}
+
+  double p0(int x) const {
+    assert(x > 0);
+    assert(x < 5);
+    return 1.0/4.0;
+  }
+
+  double llh() const {
+    double lh = bp + base.log_crp_prob();
+    for (int ctx = 1; ctx < 5; ++ctx)
+      lh += ccrps[ctx].log_crp_prob();
+    return lh;
+  }
+
+  double prob(int ctx, int x) const {
+    assert(ctx > 0 && ctx < 5);
+    return ccrps[ctx].prob(x, base.prob(x, p0(x)));
+  }
+
+  void increment(int ctx, int x) {
+    assert(ctx > 0 && ctx < 5);
+    if (ccrps[ctx].increment(x, base.prob(x, p0(x)), &rng)) {
+      if (base.increment(x, p0(x), &rng)) {
+        bp += log(1.0 / 4.0);
+      }
+    }
+  }
+
+  // this is just a biased estimate
+  double est_base_prob(int x) {
+    return (x + 1) * x / 40.0;
+  }
+
+  void increment_is(int ctx, int x) {
+    assert(ctx > 0 && ctx < 5);
+    SampleSet<double> ss;
+    const int PARTICLES = 25;
+    vector<CCRP<int> > s1s(PARTICLES, CCRP<int>(0.5,0.5));
+    vector<CCRP<int> > sbs(PARTICLES, CCRP<int>(0.5,0.5));
+    vector<double> sp0s(PARTICLES);
+
+    CCRP<int> s1 = ccrps[ctx];
+    CCRP<int> sb = base;
+    double sp0 = bp;
+    for (int pp = 0; pp < PARTICLES; ++pp) {
+      if (pp > 0) {
+        ccrps[ctx] = s1;
+        base = sb;
+        bp = sp0;
+      }
+
+      double q = 1;
+      double gamma = 1;
+      double est_p = est_base_prob(x);
+      //base.prob(x, p0(x)) + rng.next() * 0.1;
+      if (ccrps[ctx].increment(x, est_p, &rng, &q)) {
+        gamma = q * base.prob(x, p0(x));
+        q *= est_p;
+        if (verbose) cerr << "(DP-base draw) ";
+        double qq = -1;
+        if (base.increment(x, p0(x), &rng, &qq)) {
+          if (verbose) cerr << "(G0 draw) ";
+          bp += log(p0(x));
+          qq *= p0(x);
+        }
+      } else { gamma = q; }
+      double w = gamma / q;
+      if (verbose)
+        cerr << "gamma=" << gamma << " q=" << q << "\tw=" << w << endl;
+      ss.add(w);
+      s1s[pp] = ccrps[ctx];
+      sbs[pp] = base;
+      sp0s[pp] = bp;
+    }
+    int ps = rng.SelectSample(ss);
+    ccrps[ctx] = s1s[ps];
+    base = sbs[ps];
+    bp = sp0s[ps];
+    if (verbose) {
+      cerr << "SELECTED: " << ps << endl;
+      static int cc = 0; cc++; if (cc ==10) exit(1);
+    }
+  }
+
+  void decrement(int ctx, int x) {
+    assert(ctx > 0 && ctx < 5);
+    if (ccrps[ctx].decrement(x, &rng)) {
+      if (base.decrement(x, &rng)) {
+        bp -= log(p0(x));
+      }
+    }
+  }
+
+  double bp;
+  CCRP<int> base;
+  vector<CCRP<int> > ccrps;
+
+};
+
+int main(int argc, char** argv) {
+  if (argc > 1) { verbose = true; }
+  vector<int> counts(15, 0);
+  vector<int> tcounts(15, 0);
+  int points[] = {1,2, 2,2, 3,2, 4,1, 3, 4, 3, 3, 2, 3, 4, 1, 4, 1, 3, 2, 1, 3, 1, 4, 0, 0};
+  double tlh = 0;
+  double tt = 0;
+  for (int n = 0; n < 1000; ++n) {
+    if (n % 10 == 0) cerr << '.';
+    if ((n+1) % 400 == 0) cerr << " [" << (n+1) << "]\n";
+    Model m;
+    for (int *x = points; *x; x += 2)
+      m.increment(x[0], x[1]);
+
+    for (int j = 0; j < 24; ++j) {
+      for (int *x = points; *x; x += 2) {
+        if (rng.next() < 0.8) {
+          m.decrement(x[0], x[1]);
+          m.increment_is(x[0], x[1]);
+        }
+      }
+    }
+    counts[m.base.num_customers()]++;
+    tcounts[m.base.num_tables()]++;
+    tlh += m.llh();
+    tt += 1.0;
+  }
+  cerr << "mean LLH = " << (tlh / tt) << endl;
+  for (int i = 0; i < 15; ++i)
+    cerr << i << ": " << (counts[i] / tt) << "\t" << (tcounts[i] / tt) << endl;
+}
+
diff --git a/gi/pf/pf_test.cc b/gi/pf/pf_test.cc
new file mode 100644
index 00000000..296e7285
--- /dev/null
+++ b/gi/pf/pf_test.cc
@@ -0,0 +1,148 @@
+#include "ccrp.h"
+
+#include <vector>
+#include <iostream>
+
+#include "tdict.h"
+#include "transliterations.h"
+
+using namespace std;
+
+MT19937 rng;
+
+static bool verbose = false;
+
+struct Model {
+
+  Model() : bp(), base(0.2, 0.6) , ccrps(5, CCRP<int>(0.8, 0.5)) {}
+
+  double p0(int x) const {
+    assert(x > 0);
+    assert(x < 5);
+    return 1.0/4.0;
+  }
+
+  double llh() const {
+    double lh = bp + base.log_crp_prob();
+    for (int ctx = 1; ctx < 5; ++ctx)
+      lh += ccrps[ctx].log_crp_prob();
+    return lh;
+  }
+
+  double prob(int ctx, int x) const {
+    assert(ctx > 0 && ctx < 5);
+    return ccrps[ctx].prob(x, base.prob(x, p0(x)));
+  }
+
+  void increment(int ctx, int x) {
+    assert(ctx > 0 && ctx < 5);
+    if (ccrps[ctx].increment(x, base.prob(x, p0(x)), &rng)) {
+      if (base.increment(x, p0(x), &rng)) {
+        bp += log(1.0 / 4.0);
+      }
+    }
+  }
+
+  // this is just a biased estimate
+  double est_base_prob(int x) {
+    return (x + 1) * x / 40.0;
+  }
+
+  void increment_is(int ctx, int x) {
+    assert(ctx > 0 && ctx < 5);
+    SampleSet<double> ss;
+    const int PARTICLES = 25;
+    vector<CCRP<int> > s1s(PARTICLES, CCRP<int>(0.5,0.5));
+    vector<CCRP<int> > sbs(PARTICLES, CCRP<int>(0.5,0.5));
+    vector<double> sp0s(PARTICLES);
+
+    CCRP<int> s1 = ccrps[ctx];
+    CCRP<int> sb = base;
+    double sp0 = bp;
+    for (int pp = 0; pp < PARTICLES; ++pp) {
+      if (pp > 0) {
+        ccrps[ctx] = s1;
+        base = sb;
+        bp = sp0;
+      }
+
+      double q = 1;
+      double gamma = 1;
+      double est_p = est_base_prob(x);
+      //base.prob(x, p0(x)) + rng.next() * 0.1;
+      if (ccrps[ctx].increment(x, est_p, &rng, &q)) {
+        gamma = q * base.prob(x, p0(x));
+        q *= est_p;
+        if (verbose) cerr << "(DP-base draw) ";
+        double qq = -1;
+        if (base.increment(x, p0(x), &rng, &qq)) {
+          if (verbose) cerr << "(G0 draw) ";
+          bp += log(p0(x));
+          qq *= p0(x);
+        }
+      } else { gamma = q; }
+      double w = gamma / q;
+      if (verbose)
+        cerr << "gamma=" << gamma << " q=" << q << "\tw=" << w << endl;
+      ss.add(w);
+      s1s[pp] = ccrps[ctx];
+      sbs[pp] = base;
+      sp0s[pp] = bp;
+    }
+    int ps = rng.SelectSample(ss);
+    ccrps[ctx] = s1s[ps];
+    base = sbs[ps];
+    bp = sp0s[ps];
+    if (verbose) {
+      cerr << "SELECTED: " << ps << endl;
+      static int cc = 0; cc++; if (cc ==10) exit(1);
+    }
+  }
+
+  void decrement(int ctx, int x) {
+    assert(ctx > 0 && ctx < 5);
+    if (ccrps[ctx].decrement(x, &rng)) {
+      if (base.decrement(x, &rng)) {
+        bp -= log(p0(x));
+      }
+    }
+  }
+
+  double bp;
+  CCRP<int> base;
+  vector<CCRP<int> > ccrps;
+
+};
+
+int main(int argc, char** argv) {
+  if (argc > 1) { verbose = true; }
+  vector<int> counts(15, 0);
+  vector<int> tcounts(15, 0);
+  int points[] = {1,2, 2,2, 3,2, 4,1, 3, 4, 3, 3, 2, 3, 4, 1, 4, 1, 3, 2, 1, 3, 1, 4, 0, 0};
+  double tlh = 0;
+  double tt = 0;
+  for (int n = 0; n < 1000; ++n) {
+    if (n % 10 == 0) cerr << '.';
+    if ((n+1) % 400 == 0) cerr << " [" << (n+1) << "]\n";
+    Model m;
+    for (int *x = points; *x; x += 2)
+      m.increment(x[0], x[1]);
+
+    for (int j = 0; j < 24; ++j) {
+      for (int *x = points; *x; x += 2) {
+        if (rng.next() < 0.8) {
+          m.decrement(x[0], x[1]);
+          m.increment_is(x[0], x[1]);
+        }
+      }
+    }
+    counts[m.base.num_customers()]++;
+    tcounts[m.base.num_tables()]++;
+    tlh += m.llh();
+    tt += 1.0;
+  }
+  cerr << "mean LLH = " << (tlh / tt) << endl;
+  for (int i = 0; i < 15; ++i)
+    cerr << i << ": " << (counts[i] / tt) << "\t" << (tcounts[i] / tt) << endl;
+}
+
diff --git a/gi/pf/pfbrat.cc b/gi/pf/pfbrat.cc
index c2c52760..832f22cf 100644
--- a/gi/pf/pfbrat.cc
+++ b/gi/pf/pfbrat.cc
@@ -489,7 +489,7 @@ int main(int argc, char** argv) {
     cerr << argv[0] << "Please use --model1 to specify model 1 parameters\n";
     return 1;
   }
-  shared_ptr<MT19937> prng;
+  boost::shared_ptr<MT19937> prng;
   if (conf.count("random_seed"))
     prng.reset(new MT19937(conf["random_seed"].as<uint32_t>()));
   else
diff --git a/gi/pf/pfdist.cc b/gi/pf/pfdist.cc
index 3d578db2..a3e46064 100644
--- a/gi/pf/pfdist.cc
+++ b/gi/pf/pfdist.cc
@@ -23,7 +23,7 @@ using namespace std;
 using namespace tr1;
 namespace po = boost::program_options;
 
-shared_ptr<MT19937> prng;
+boost::shared_ptr<MT19937> prng;
 
 void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   po::options_description opts("Configuration options");
diff --git a/gi/pf/pfnaive.cc b/gi/pf/pfnaive.cc
index e1a53f5c..958ec4e2 100644
--- a/gi/pf/pfnaive.cc
+++ b/gi/pf/pfnaive.cc
@@ -25,7 +25,7 @@ using namespace std;
 using namespace tr1;
 namespace po = boost::program_options;
 
-shared_ptr<MT19937> prng;
+boost::shared_ptr<MT19937> prng;
 
 void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   po::options_description opts("Configuration options");
diff --git a/gi/pf/poisson_uniform_word_model.h b/gi/pf/poisson_uniform_word_model.h
new file mode 100644
index 00000000..76204a0e
--- /dev/null
+++ b/gi/pf/poisson_uniform_word_model.h
@@ -0,0 +1,50 @@
+#ifndef _POISSON_UNIFORM_WORD_MODEL_H_
+#define _POISSON_UNIFORM_WORD_MODEL_H_
+
+#include <cmath>
+#include <vector>
+#include "prob.h"
+#include "m.h"
+
+// len ~ Poisson(lambda)
+//   for (1..len)
+//     e_i ~ Uniform({Vocabulary})
+struct PoissonUniformWordModel {
+  explicit PoissonUniformWordModel(const unsigned vocab_size,
+                                   const unsigned alphabet_size,
+                                   const double mean_len = 5) :
+    lh(prob_t::One()),
+    v0(-std::log(vocab_size)),
+    u0(-std::log(alphabet_size)),
+    mean_length(mean_len) {}
+
+  void ResampleHyperparameters(MT19937*) {}
+
+  inline prob_t operator()(const std::vector<WordID>& s) const {
+    prob_t p;
+    p.logeq(Md::log_poisson(s.size(), mean_length) + s.size() * u0);
+    //p.logeq(v0);
+    return p;
+  }
+
+  inline void Increment(const std::vector<WordID>& w, MT19937*) {
+    lh *= (*this)(w);
+  }
+
+  inline void Decrement(const std::vector<WordID>& w, MT19937 *) {
+    lh /= (*this)(w);
+  }
+
+  inline prob_t Likelihood() const { return lh; }
+
+  void Summary() const {}
+
+ private:
+
+  prob_t lh;  // keeps track of the draws from the base distribution
+  const double v0;  // uniform log prob of generating a word
+  const double u0;  // uniform log prob of generating a letter
+  const double mean_length;  // mean length of a word in the base distribution
+};
+
+#endif
diff --git a/gi/pf/pyp_lm.cc b/gi/pf/pyp_lm.cc
index 91029688..e2b67e17 100644
--- a/gi/pf/pyp_lm.cc
+++ b/gi/pf/pyp_lm.cc
@@ -25,7 +25,7 @@ using namespace std;
 using namespace tr1;
 namespace po = boost::program_options;
 
-shared_ptr<MT19937> prng;
+boost::shared_ptr<MT19937> prng;
 
 void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   po::options_description opts("Configuration options");
diff --git a/gi/pf/pyp_tm.cc b/gi/pf/pyp_tm.cc
index e21f0267..6bc8a5bf 100644
--- a/gi/pf/pyp_tm.cc
+++ b/gi/pf/pyp_tm.cc
@@ -91,26 +91,23 @@ struct ConditionalPYPWordModel {
 };
 
 PYPLexicalTranslation::PYPLexicalTranslation(const vector<vector<WordID> >& lets,
+                                             const unsigned vocab_size,
                                              const unsigned num_letters) :
     letters(lets),
-    up0(new PYPWordModel(num_letters)),
-    tmodel(new ConditionalPYPWordModel<PYPWordModel>(up0, new FreqBinner("10k.freq"))),
+    base(vocab_size, num_letters, 5),
+    tmodel(new ConditionalPYPWordModel<PoissonUniformWordModel>(&base, new FreqBinner("10k.freq"))),
     kX(-TD::Convert("X")) {}
 
 void PYPLexicalTranslation::Summary() const {
   tmodel->Summary();
-  up0->Summary();
 }
 
 prob_t PYPLexicalTranslation::Likelihood() const {
-  prob_t p = up0->Likelihood();
-  p *= tmodel->Likelihood();
-  return p;
+  return tmodel->Likelihood() * base.Likelihood();
 }
 
 void PYPLexicalTranslation::ResampleHyperparameters(MT19937* rng) {
   tmodel->ResampleHyperparameters(rng);
-  up0->ResampleHyperparameters(rng);
 }
 
 unsigned PYPLexicalTranslation::UniqueConditioningContexts() const {
diff --git a/gi/pf/pyp_tm.h b/gi/pf/pyp_tm.h
index 63e7c96d..2b076a25 100644
--- a/gi/pf/pyp_tm.h
+++ b/gi/pf/pyp_tm.h
@@ -6,13 +6,14 @@
 #include "prob.h"
 #include "sampler.h"
 #include "freqdict.h"
+#include "poisson_uniform_word_model.h"
 
 struct FreqBinner;
-struct PYPWordModel;
 template <typename T, class B> struct ConditionalPYPWordModel;
 
 struct PYPLexicalTranslation {
   explicit PYPLexicalTranslation(const std::vector<std::vector<WordID> >& lets,
+                                 const unsigned vocab_size,
                                  const unsigned num_letters);
 
   prob_t Likelihood() const;
@@ -26,8 +27,8 @@ struct PYPLexicalTranslation {
 
  private:
   const std::vector<std::vector<WordID> >& letters;   // spelling dictionary
-  PYPWordModel* up0;  // base distribuction (model English word)
-  ConditionalPYPWordModel<PYPWordModel, FreqBinner>* tmodel;  // translation distributions
+  PoissonUniformWordModel base;  // "generator" of English types
+  ConditionalPYPWordModel<PoissonUniformWordModel, FreqBinner>* tmodel;  // translation distributions
                       // (model English word | French word)
   const WordID kX;
 };
diff --git a/gi/pf/pyp_word_model.cc b/gi/pf/pyp_word_model.cc
deleted file mode 100644
index 12df4abf..00000000
--- a/gi/pf/pyp_word_model.cc
+++ /dev/null
@@ -1,20 +0,0 @@
-#include "pyp_word_model.h"
-
-#include <iostream>
-
-using namespace std;
-
-void PYPWordModel::ResampleHyperparameters(MT19937* rng) {
-  r.resample_hyperparameters(rng);
-  cerr << " PYPWordModel(d=" << r.discount() << ",s=" << r.strength() << ")\n";
-}
-
-void PYPWordModel::Summary() const {
-  cerr << "PYPWordModel: generations=" << r.num_customers()
-       << " PYP(d=" << r.discount() << ",s=" << r.strength() << ')' << endl;
-  for (CCRP<vector<WordID> >::const_iterator it = r.begin(); it != r.end(); ++it)
-    cerr << "   " << it->second.total_dish_count_
-              << " (on " << it->second.table_counts_.size() << " tables) "
-              << TD::GetString(it->first) << endl;
-}
-
diff --git a/gi/pf/pyp_word_model.h b/gi/pf/pyp_word_model.h
index ff366865..224a9034 100644
--- a/gi/pf/pyp_word_model.h
+++ b/gi/pf/pyp_word_model.h
@@ -11,48 +11,52 @@
 #include "os_phrase.h"
 
 // PYP(d,s,poisson-uniform) represented as a CRP
+template <class Base>
 struct PYPWordModel {
-  explicit PYPWordModel(const unsigned vocab_e_size, const double mean_len = 5) :
-      base(prob_t::One()), r(1,1,1,1,0.66,50.0), u0(-std::log(vocab_e_size)), mean_length(mean_len) {}
-
-  void ResampleHyperparameters(MT19937* rng);
+  explicit PYPWordModel(Base* b) :
+      base(*b),
+      r(1,1,1,1,0.66,50.0)
+    {}
+
+  void ResampleHyperparameters(MT19937* rng) {
+    r.resample_hyperparameters(rng);
+    std::cerr << " PYPWordModel(d=" << r.discount() << ",s=" << r.strength() << ")\n";
+  }
 
   inline prob_t operator()(const std::vector<WordID>& s) const {
-    return r.prob(s, p0(s));
+    return r.prob(s, base(s));
   }
 
   inline void Increment(const std::vector<WordID>& s, MT19937* rng) {
-    if (r.increment(s, p0(s), rng))
-      base *= p0(s);
+    if (r.increment(s, base(s), rng))
+      base.Increment(s, rng);
   }
 
   inline void Decrement(const std::vector<WordID>& s, MT19937 *rng) {
     if (r.decrement(s, rng))
-      base /= p0(s);
+      base.Decrement(s, rng);
   }
 
   inline prob_t Likelihood() const {
     prob_t p; p.logeq(r.log_crp_prob());
-    p *= base;
+    p *= base.Likelihood();
     return p;
   }
 
-  void Summary() const;
-
- private:
-  inline double logp0(const std::vector<WordID>& s) const {
-    return Md::log_poisson(s.size(), mean_length) + s.size() * u0;
+  void Summary() const {
+    std::cerr << "PYPWordModel: generations=" << r.num_customers()
+         << " PYP(d=" << r.discount() << ",s=" << r.strength() << ')' << std::endl;
+    for (typename CCRP<std::vector<WordID> >::const_iterator it = r.begin(); it != r.end(); ++it) {
+      std::cerr << "   " << it->second.total_dish_count_
+                << " (on " << it->second.table_counts_.size() << " tables) "
+                << TD::GetString(it->first) << std::endl;
+    }
   }
 
-  inline prob_t p0(const std::vector<WordID>& s) const {
-    prob_t p; p.logeq(logp0(s));
-    return p;
-  }
+ private:
 
-  prob_t base;  // keeps track of the draws from the base distribution
+  Base& base;  // keeps track of the draws from the base distribution
   CCRP<std::vector<WordID> > r;
-  const double u0;  // uniform log prob of generating a letter
-  const double mean_length;  // mean length of a word in the base distribution
 };
 
 #endif
diff --git a/gi/pf/quasi_model2.h b/gi/pf/quasi_model2.h
index 588c8f84..4075affe 100644
--- a/gi/pf/quasi_model2.h
+++ b/gi/pf/quasi_model2.h
@@ -9,6 +9,7 @@
 #include "array2d.h"
 #include "slice_sampler.h"
 #include "m.h"
+#include "have_64_bits.h"
 
 struct AlignmentObservation {
   AlignmentObservation() : src_len(), trg_len(), j(), a_j() {}
@@ -20,13 +21,23 @@ struct AlignmentObservation {
   unsigned short a_j;
 };
 
+#ifdef HAVE_64_BITS
 inline size_t hash_value(const AlignmentObservation& o) {
   return reinterpret_cast<const size_t&>(o);
 }
-
 inline bool operator==(const AlignmentObservation& a, const AlignmentObservation& b) {
   return hash_value(a) == hash_value(b);
 }
+#else
+inline size_t hash_value(const AlignmentObservation& o) {
+  size_t h = 1;
+  boost::hash_combine(h, o.src_len);
+  boost::hash_combine(h, o.trg_len);
+  boost::hash_combine(h, o.j);
+  boost::hash_combine(h, o.a_j);
+  return h;
+}
+#endif
 
 struct QuasiModel2 {
   explicit QuasiModel2(double alpha, double pnull = 0.1) :
diff --git a/gi/pf/tied_resampler.h b/gi/pf/tied_resampler.h
index 6f45fbce..a4f4af36 100644
--- a/gi/pf/tied_resampler.h
+++ b/gi/pf/tied_resampler.h
@@ -78,10 +78,8 @@ struct TiedResampler {
                             std::numeric_limits<double>::infinity(), 0.0, niterations, 100*niterations);
     std::cerr << "TiedCRPs(d=" << discount << ",s="
               << strength << ") = " << LogLikelihood(discount, strength) << std::endl;
-    for (typename std::set<CRP*>::iterator it = crps.begin(); it != crps.end(); ++it) {
-      (*it)->set_discount(discount);
-      (*it)->set_strength(strength);
-    }
+    for (typename std::set<CRP*>::iterator it = crps.begin(); it != crps.end(); ++it)
+      (*it)->set_hyperparameters(discount, strength);
   }
  private:
   std::set<CRP*> crps;