From 9007216a43c5572c2c343a1700ac79fb35b7d82f Mon Sep 17 00:00:00 2001 From: Chris Dyer Date: Sat, 25 Feb 2012 21:22:27 -0500 Subject: really slow hiero lm --- gi/pf/Makefile.am | 4 +- gi/pf/hierolm.cc | 309 +++++++++++++++++++++++++++++++++++++++++++++ phrasinator/ccrp.h | 294 ------------------------------------------- utils/ccrp.h | 340 ++++++++++++++++++++++++++++++++++++++++++++++++++ utils/ccrp_onetable.h | 12 ++ utils/sampler.h | 2 +- 6 files changed, 665 insertions(+), 296 deletions(-) create mode 100644 gi/pf/hierolm.cc delete mode 100644 phrasinator/ccrp.h create mode 100644 utils/ccrp.h diff --git a/gi/pf/Makefile.am b/gi/pf/Makefile.am index 8d43f36d..ed5b6fd3 100644 --- a/gi/pf/Makefile.am +++ b/gi/pf/Makefile.am @@ -1,4 +1,4 @@ -bin_PROGRAMS = cbgi brat dpnaive pfbrat pfdist itg pfnaive condnaive align-lexonly align-lexonly-pyp +bin_PROGRAMS = cbgi brat dpnaive pfbrat pfdist itg pfnaive condnaive align-lexonly align-lexonly-pyp hierolm noinst_LIBRARIES = libpf.a libpf_a_SOURCES = base_distributions.cc reachability.cc cfg_wfst_composer.cc corpus.cc unigrams.cc ngram_base.cc @@ -9,6 +9,8 @@ align_lexonly_pyp_SOURCES = align-lexonly-pyp.cc itg_SOURCES = itg.cc +hierolm_SOURCES = hierolm.cc + condnaive_SOURCES = condnaive.cc dpnaive_SOURCES = dpnaive.cc diff --git a/gi/pf/hierolm.cc b/gi/pf/hierolm.cc new file mode 100644 index 00000000..afb12fef --- /dev/null +++ b/gi/pf/hierolm.cc @@ -0,0 +1,309 @@ +#include +#include +#include + +#include +#include +#include + +#include "inside_outside.h" +#include "hg.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; + +shared_ptr prng; + +void InitCommandLine(int argc, char** argv, po::variables_map* conf) { + po::options_description opts("Configuration options"); + opts.add_options() + ("samples,s",po::value()->default_value(1000),"Number of samples") + ("input,i",po::value(),"Read parallel data from") + ("random_seed,S",po::value(), "Random seed"); + po::options_description clo("Command line options"); + clo.add_options() + ("config", po::value(), "Configuration file") + ("help,h", "Print this help message and exit"); + po::options_description dconfig_options, dcmdline_options; + dconfig_options.add(opts); + dcmdline_options.add(opts).add(clo); + + po::store(parse_command_line(argc, argv, dcmdline_options), *conf); + if (conf->count("config")) { + ifstream config((*conf)["config"].as().c_str()); + po::store(po::parse_config_file(config, dconfig_options), *conf); + } + po::notify(*conf); + + if (conf->count("help") || (conf->count("input") == 0)) { + cerr << dcmdline_options << endl; + exit(1); + } +} + +void ReadCorpus(const string& filename, + vector >* e, + set* vocab_e) { + e->clear(); + vocab_e->clear(); + istream* in; + if (filename == "-") + in = &cin; + else + in = new ifstream(filename.c_str()); + assert(*in); + string line; + while(*in) { + getline(*in, line); + if (line.empty() && !*in) break; + e->push_back(vector()); + vector& le = e->back(); + TD::ConvertSentence(line, &le); + for (unsigned i = 0; i < le.size(); ++i) + vocab_e->insert(le[i]); + } + if (in != &cin) delete in; +} + +struct Grid { + // a b c d e + // 0 - 0 - - + vector grid; +}; + +struct BaseRuleModel { + explicit BaseRuleModel(unsigned term_size, + unsigned nonterm_size = 1) : + unif_term(1.0 / term_size), + unif_nonterm(1.0 / nonterm_size) {} + prob_t operator()(const TRule& r) const { + prob_t p; p.logeq(Md::log_poisson(1.0, r.f_.size())); + const prob_t term_prob((2.0 + 0.01*r.f_.size()) / (r.f_.size() + 2)); + const prob_t nonterm_prob(1.0 - term_prob.as_float()); + for (unsigned i = 0; i < r.f_.size(); ++i) { + if (r.f_[i] <= 0) { // nonterminal + p *= nonterm_prob; + p *= unif_nonterm; + } else { // terminal + p *= term_prob; + p *= unif_term; + } + } + return p; + } + const prob_t unif_term, unif_nonterm; +}; + +struct HieroLMModel { + explicit HieroLMModel(unsigned vocab_size) : p0(vocab_size), x(1,1,1,1) {} + + prob_t Prob(const TRule& r) const { + return x.probT(r, p0(r)); + } + + int Increment(const TRule& r, MT19937* rng) { + return x.incrementT(r, p0(r), rng); + // return x.increment(r); + } + + int Decrement(const TRule& r, MT19937* rng) { + return x.decrement(r, rng); + //return x.decrement(r); + } + + prob_t Likelihood() const { + prob_t p; + p.logeq(x.log_crp_prob()); + for (CCRP::const_iterator it = x.begin(); it != x.end(); ++it) { + prob_t tp = p0(it->first); + tp.poweq(it->second.table_counts_.size()); + p *= tp; + } + //for (CCRP_OneTable::const_iterator it = x.begin(); it != x.end(); ++it) + // p *= p0(it->first); + return p; + } + + void ResampleHyperparameters(MT19937* rng) { + x.resample_hyperparameters(rng); + cerr << " d=" << x.discount() << ", alpha=" << x.concentration() << endl; + } + + const BaseRuleModel p0; + CCRP x; + //CCRP_OneTable x; +}; + +vector tofreelist; + +HieroLMModel* plm; + +struct NPGrammarIter : public GrammarIter, public RuleBin { + NPGrammarIter() : arity() { tofreelist.push_back(this); } + NPGrammarIter(const TRulePtr& inr, const int a, int symbol) : arity(a + (symbol < 0 ? 1 : 0)) { + if (inr) { + r.reset(new TRule(*inr)); + } else { + static const int kLHS = -TD::Convert("X"); + r.reset(new TRule); + r->lhs_ = kLHS; + } + TRule& rr = *r; + rr.f_.push_back(symbol); + rr.e_.push_back(symbol < 0 ? (1-int(arity)) : symbol); + tofreelist.push_back(this); + } + virtual int GetNumRules() const { + if (r) return 1; else return 0; + } + virtual TRulePtr GetIthRule(int) const { + return r; + } + virtual int Arity() const { + return arity; + } + virtual const RuleBin* GetRules() const { + if (!r) return NULL; else return this; + } + virtual const GrammarIter* Extend(int symbol) const { + return new NPGrammarIter(r, arity, symbol); + } + const unsigned char arity; + TRulePtr r; +}; + +struct NPGrammar : public Grammar { + virtual const GrammarIter* GetRoot() const { + return new NPGrammarIter; + } +}; + +void SampleDerivation(const Hypergraph& hg, MT19937* rng, vector* sampled_deriv, HieroLMModel* plm) { + HieroLMModel& lm = *plm; + vector node_probs; + const prob_t total_prob = Inside(hg, &node_probs); + queue q; + q.push(hg.nodes_.size() - 3); + 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 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& d, HieroLMModel* 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& d, HieroLMModel* plm, MT19937* rng) { + for (unsigned i = 0; i < d.size(); ++i) + plm->Decrement(*hg.edges_[d[i]].rule_, rng); +} + +int main(int argc, char** argv) { + po::variables_map conf; + vector grammars; + grammars.push_back(GrammarPtr(new NPGrammar)); + + InitCommandLine(argc, argv, &conf); + const unsigned samples = conf["samples"].as(); + + if (conf.count("random_seed")) + prng.reset(new MT19937(conf["random_seed"].as())); + else + prng.reset(new MT19937); + MT19937& rng = *prng; + + vector > corpuse; + set vocabe; + cerr << "Reading corpus...\n"; + ReadCorpus(conf["input"].as(), &corpuse, &vocabe); + cerr << "E-corpus size: " << corpuse.size() << " sentences\t (" << vocabe.size() << " word types)\n"; + HieroLMModel lm(vocabe.size()); + + plm = &lm; + ExhaustiveBottomUpParser parser("X", grammars); + + Hypergraph hg; + const int kX = -TD::Convert("X"); + const int kLP = FD::Convert("LogProb"); + SparseVector v; v.set_value(kLP, 1.0); + vector > derivs(corpuse.size()); + for (int SS=0; SS < samples; ++SS) { + for (int ci = 0; ci < corpuse.size(); ++ci) { + vector& src = corpuse[ci]; + Lattice lat(src.size()); + for (unsigned i = 0; i < src.size(); ++i) + lat[i].push_back(LatticeArc(src[i], 0.0, 1)); + cerr << TD::GetString(src) << endl; + hg.clear(); + parser.Parse(lat, &hg); // exhaustive parse + DecrementDerivation(hg, derivs[ci], &lm, &rng); + for (unsigned i = 0; i < hg.edges_.size(); ++i) { + TRule& r = *hg.edges_[i].rule_; + if (r.lhs_ == kX) + hg.edges_[i].edge_prob_ = lm.Prob(r); + } + vector d; + SampleDerivation(hg, &rng, &d, &lm); + derivs[ci] = d; + IncrementDerivation(hg, derivs[ci], &lm, &rng); + if (tofreelist.size() > 100000) { + cerr << "Freeing ... "; + for (unsigned i = 0; i < tofreelist.size(); ++i) + delete tofreelist[i]; + tofreelist.clear(); + cerr << "Freed.\n"; + } + } + cerr << "LLH=" << lm.Likelihood() << endl; + } + return 0; +} + diff --git a/phrasinator/ccrp.h b/phrasinator/ccrp.h deleted file mode 100644 index 9acf12ab..00000000 --- a/phrasinator/ccrp.h +++ /dev/null @@ -1,294 +0,0 @@ -#ifndef _CCRP_H_ -#define _CCRP_H_ - -#include -#include -#include -#include -#include -#include -#include -#include -#include "sampler.h" -#include "slice_sampler.h" - -// Chinese restaurant process (Pitman-Yor parameters) with table tracking. - -template > -class CCRP { - public: - CCRP(double disc, double conc) : - num_tables_(), - num_customers_(), - discount_(disc), - concentration_(conc), - discount_prior_alpha_(std::numeric_limits::quiet_NaN()), - discount_prior_beta_(std::numeric_limits::quiet_NaN()), - concentration_prior_shape_(std::numeric_limits::quiet_NaN()), - concentration_prior_rate_(std::numeric_limits::quiet_NaN()) {} - - CCRP(double d_alpha, double d_beta, double c_shape, double c_rate, double d = 0.1, double c = 10.0) : - num_tables_(), - num_customers_(), - discount_(d), - concentration_(c), - discount_prior_alpha_(d_alpha), - discount_prior_beta_(d_beta), - concentration_prior_shape_(c_shape), - concentration_prior_rate_(c_rate) {} - - double discount() const { return discount_; } - double concentration() const { return concentration_; } - - bool has_discount_prior() const { - return !std::isnan(discount_prior_alpha_); - } - - bool has_concentration_prior() const { - return !std::isnan(concentration_prior_shape_); - } - - void clear() { - num_tables_ = 0; - num_customers_ = 0; - dish_locs_.clear(); - } - - unsigned num_tables() const { - return num_tables_; - } - - unsigned num_tables(const Dish& dish) const { - const typename std::tr1::unordered_map::const_iterator it = dish_locs_.find(dish); - if (it == dish_locs_.end()) return 0; - return it->second.table_counts_.size(); - } - - unsigned num_customers() const { - return num_customers_; - } - - unsigned num_customers(const Dish& dish) const { - const typename std::tr1::unordered_map::const_iterator it = dish_locs_.find(dish); - if (it == dish_locs_.end()) return 0; - return it->total_dish_count_; - } - - // returns +1 or 0 indicating whether a new table was opened - int increment(const Dish& dish, const double& p0, MT19937* rng) { - DishLocations& loc = dish_locs_[dish]; - bool share_table = false; - if (loc.total_dish_count_) { - const double p_empty = (concentration_ + num_tables_ * discount_) * p0; - const double p_share = (loc.total_dish_count_ - loc.table_counts_.size() * discount_); - share_table = rng->SelectSample(p_empty, p_share); - } - if (share_table) { - double r = rng->next() * (loc.total_dish_count_ - loc.table_counts_.size() * discount_); - for (typename std::list::iterator ti = loc.table_counts_.begin(); - ti != loc.table_counts_.end(); ++ti) { - r -= (*ti - discount_); - if (r <= 0.0) { - ++(*ti); - break; - } - } - if (r > 0.0) { - std::cerr << "Serious error: r=" << r << std::endl; - Print(&std::cerr); - assert(r <= 0.0); - } - } else { - loc.table_counts_.push_back(1u); - ++num_tables_; - } - ++loc.total_dish_count_; - ++num_customers_; - return (share_table ? 0 : 1); - } - - // returns -1 or 0, indicating whether a table was closed - int decrement(const Dish& dish, MT19937* rng) { - DishLocations& loc = dish_locs_[dish]; - assert(loc.total_dish_count_); - if (loc.total_dish_count_ == 1) { - dish_locs_.erase(dish); - --num_tables_; - --num_customers_; - return -1; - } else { - int delta = 0; - // sample customer to remove UNIFORMLY. that is, do NOT use the discount - // here. if you do, it will introduce (unwanted) bias! - double r = rng->next() * loc.total_dish_count_; - --loc.total_dish_count_; - for (typename std::list::iterator ti = loc.table_counts_.begin(); - ti != loc.table_counts_.end(); ++ti) { - r -= *ti; - if (r <= 0.0) { - if ((--(*ti)) == 0) { - --num_tables_; - delta = -1; - loc.table_counts_.erase(ti); - } - break; - } - } - if (r > 0.0) { - std::cerr << "Serious error: r=" << r << std::endl; - Print(&std::cerr); - assert(r <= 0.0); - } - --num_customers_; - return delta; - } - } - - double prob(const Dish& dish, const double& p0) const { - const typename std::tr1::unordered_map::const_iterator it = dish_locs_.find(dish); - const double r = num_tables_ * discount_ + concentration_; - if (it == dish_locs_.end()) { - return r * p0 / (num_customers_ + concentration_); - } else { - return (it->second.total_dish_count_ - discount_ * it->second.table_counts_.size() + r * p0) / - (num_customers_ + concentration_); - } - } - - double log_crp_prob() const { - return log_crp_prob(discount_, concentration_); - } - - static double log_beta_density(const double& x, const double& alpha, const double& beta) { - assert(x > 0.0); - assert(x < 1.0); - assert(alpha > 0.0); - assert(beta > 0.0); - const double lp = (alpha-1)*log(x)+(beta-1)*log(1-x)+lgamma(alpha+beta)-lgamma(alpha)-lgamma(beta); - return lp; - } - - static double log_gamma_density(const double& x, const double& shape, const double& rate) { - assert(x >= 0.0); - assert(shape > 0.0); - assert(rate > 0.0); - const double lp = (shape-1)*log(x) - shape*log(rate) - x/rate - lgamma(shape); - return lp; - } - - // taken from http://en.wikipedia.org/wiki/Chinese_restaurant_process - // does not include P_0's - double log_crp_prob(const double& discount, const double& concentration) const { - double lp = 0.0; - if (has_discount_prior()) - lp = log_beta_density(discount, discount_prior_alpha_, discount_prior_beta_); - if (has_concentration_prior()) - lp += log_gamma_density(concentration, concentration_prior_shape_, concentration_prior_rate_); - assert(lp <= 0.0); - if (num_customers_) { - if (discount > 0.0) { - const double r = lgamma(1.0 - discount); - lp += lgamma(concentration) - lgamma(concentration + num_customers_) - + num_tables_ * log(discount) + lgamma(concentration / discount + num_tables_) - - lgamma(concentration / discount); - assert(std::isfinite(lp)); - for (typename std::tr1::unordered_map::const_iterator it = dish_locs_.begin(); - it != dish_locs_.end(); ++it) { - const DishLocations& cur = it->second; - for (std::list::const_iterator ti = cur.table_counts_.begin(); ti != cur.table_counts_.end(); ++ti) { - lp += lgamma(*ti - discount) - r; - } - } - } else { - assert(!"not implemented yet"); - } - } - assert(std::isfinite(lp)); - return lp; - } - - void resample_hyperparameters(MT19937* rng, const unsigned nloop = 5, const unsigned niterations = 10) { - assert(has_discount_prior() || has_concentration_prior()); - DiscountResampler dr(*this); - ConcentrationResampler cr(*this); - for (int iter = 0; iter < nloop; ++iter) { - if (has_concentration_prior()) { - concentration_ = slice_sampler1d(cr, concentration_, *rng, 0.0, - std::numeric_limits::infinity(), 0.0, niterations, 100*niterations); - } - if (has_discount_prior()) { - discount_ = slice_sampler1d(dr, discount_, *rng, std::numeric_limits::min(), - 1.0, 0.0, niterations, 100*niterations); - } - } - concentration_ = slice_sampler1d(cr, concentration_, *rng, 0.0, - std::numeric_limits::infinity(), 0.0, niterations, 100*niterations); - } - - struct DiscountResampler { - DiscountResampler(const CCRP& crp) : crp_(crp) {} - const CCRP& crp_; - double operator()(const double& proposed_discount) const { - return crp_.log_crp_prob(proposed_discount, crp_.concentration_); - } - }; - - struct ConcentrationResampler { - ConcentrationResampler(const CCRP& crp) : crp_(crp) {} - const CCRP& crp_; - double operator()(const double& proposed_concentration) const { - return crp_.log_crp_prob(crp_.discount_, proposed_concentration); - } - }; - - struct DishLocations { - DishLocations() : total_dish_count_() {} - unsigned total_dish_count_; // customers at all tables with this dish - std::list table_counts_; // list<> gives O(1) deletion and insertion, which we want - // .size() is the number of tables for this dish - }; - - void Print(std::ostream* out) const { - std::cerr << "PYP(d=" << discount_ << ",c=" << concentration_ << ") customers=" << num_customers_ << std::endl; - for (typename std::tr1::unordered_map::const_iterator it = dish_locs_.begin(); - it != dish_locs_.end(); ++it) { - (*out) << it->first << " (" << it->second.total_dish_count_ << " on " << it->second.table_counts_.size() << " tables): "; - for (typename std::list::const_iterator i = it->second.table_counts_.begin(); - i != it->second.table_counts_.end(); ++i) { - (*out) << " " << *i; - } - (*out) << std::endl; - } - } - - typedef typename std::tr1::unordered_map::const_iterator const_iterator; - const_iterator begin() const { - return dish_locs_.begin(); - } - const_iterator end() const { - return dish_locs_.end(); - } - - unsigned num_tables_; - unsigned num_customers_; - std::tr1::unordered_map dish_locs_; - - double discount_; - double concentration_; - - // optional beta prior on discount_ (NaN if no prior) - double discount_prior_alpha_; - double discount_prior_beta_; - - // optional gamma prior on concentration_ (NaN if no prior) - double concentration_prior_shape_; - double concentration_prior_rate_; -}; - -template -std::ostream& operator<<(std::ostream& o, const CCRP& c) { - c.Print(&o); - return o; -} - -#endif diff --git a/utils/ccrp.h b/utils/ccrp.h new file mode 100644 index 00000000..1a9e3ed5 --- /dev/null +++ b/utils/ccrp.h @@ -0,0 +1,340 @@ +#ifndef _CCRP_H_ +#define _CCRP_H_ + +#include +#include +#include +#include +#include +#include +#include +#include +#include "sampler.h" +#include "slice_sampler.h" + +// Chinese restaurant process (Pitman-Yor parameters) with table tracking. + +template > +class CCRP { + public: + CCRP(double disc, double conc) : + num_tables_(), + num_customers_(), + discount_(disc), + concentration_(conc), + discount_prior_alpha_(std::numeric_limits::quiet_NaN()), + discount_prior_beta_(std::numeric_limits::quiet_NaN()), + concentration_prior_shape_(std::numeric_limits::quiet_NaN()), + concentration_prior_rate_(std::numeric_limits::quiet_NaN()) {} + + CCRP(double d_alpha, double d_beta, double c_shape, double c_rate, double d = 0.9, double c = 1.0) : + num_tables_(), + num_customers_(), + discount_(d), + concentration_(c), + discount_prior_alpha_(d_alpha), + discount_prior_beta_(d_beta), + concentration_prior_shape_(c_shape), + concentration_prior_rate_(c_rate) {} + + double discount() const { return discount_; } + double concentration() const { return concentration_; } + + bool has_discount_prior() const { + return !std::isnan(discount_prior_alpha_); + } + + bool has_concentration_prior() const { + return !std::isnan(concentration_prior_shape_); + } + + void clear() { + num_tables_ = 0; + num_customers_ = 0; + dish_locs_.clear(); + } + + unsigned num_tables() const { + return num_tables_; + } + + unsigned num_tables(const Dish& dish) const { + const typename std::tr1::unordered_map::const_iterator it = dish_locs_.find(dish); + if (it == dish_locs_.end()) return 0; + return it->second.table_counts_.size(); + } + + unsigned num_customers() const { + return num_customers_; + } + + unsigned num_customers(const Dish& dish) const { + const typename std::tr1::unordered_map::const_iterator it = dish_locs_.find(dish); + if (it == dish_locs_.end()) return 0; + return it->total_dish_count_; + } + + // returns +1 or 0 indicating whether a new table was opened + int increment(const Dish& dish, const double& p0, MT19937* rng) { + DishLocations& loc = dish_locs_[dish]; + bool share_table = false; + if (loc.total_dish_count_) { + const double p_empty = (concentration_ + num_tables_ * discount_) * p0; + const double p_share = (loc.total_dish_count_ - loc.table_counts_.size() * discount_); + share_table = rng->SelectSample(p_empty, p_share); + } + if (share_table) { + double r = rng->next() * (loc.total_dish_count_ - loc.table_counts_.size() * discount_); + for (typename std::list::iterator ti = loc.table_counts_.begin(); + ti != loc.table_counts_.end(); ++ti) { + r -= (*ti - discount_); + if (r <= 0.0) { + ++(*ti); + break; + } + } + if (r > 0.0) { + std::cerr << "Serious error: r=" << r << std::endl; + Print(&std::cerr); + assert(r <= 0.0); + } + } else { + loc.table_counts_.push_back(1u); + ++num_tables_; + } + ++loc.total_dish_count_; + ++num_customers_; + return (share_table ? 0 : 1); + } + + // returns +1 or 0 indicating whether a new table was opened + template + int incrementT(const Dish& dish, const T& p0, MT19937* rng) { + DishLocations& loc = dish_locs_[dish]; + bool share_table = false; + if (loc.total_dish_count_) { + const T p_empty = T(concentration_ + num_tables_ * discount_) * p0; + const T p_share = T(loc.total_dish_count_ - loc.table_counts_.size() * discount_); + share_table = rng->SelectSample(p_empty, p_share); + } + if (share_table) { + double r = rng->next() * (loc.total_dish_count_ - loc.table_counts_.size() * discount_); + for (typename std::list::iterator ti = loc.table_counts_.begin(); + ti != loc.table_counts_.end(); ++ti) { + r -= (*ti - discount_); + if (r <= 0.0) { + ++(*ti); + break; + } + } + if (r > 0.0) { + std::cerr << "Serious error: r=" << r << std::endl; + Print(&std::cerr); + assert(r <= 0.0); + } + } else { + loc.table_counts_.push_back(1u); + ++num_tables_; + } + ++loc.total_dish_count_; + ++num_customers_; + return (share_table ? 0 : 1); + } + + // returns -1 or 0, indicating whether a table was closed + int decrement(const Dish& dish, MT19937* rng) { + DishLocations& loc = dish_locs_[dish]; + assert(loc.total_dish_count_); + if (loc.total_dish_count_ == 1) { + dish_locs_.erase(dish); + --num_tables_; + --num_customers_; + return -1; + } else { + int delta = 0; + // sample customer to remove UNIFORMLY. that is, do NOT use the discount + // here. if you do, it will introduce (unwanted) bias! + double r = rng->next() * loc.total_dish_count_; + --loc.total_dish_count_; + for (typename std::list::iterator ti = loc.table_counts_.begin(); + ti != loc.table_counts_.end(); ++ti) { + r -= *ti; + if (r <= 0.0) { + if ((--(*ti)) == 0) { + --num_tables_; + delta = -1; + loc.table_counts_.erase(ti); + } + break; + } + } + if (r > 0.0) { + std::cerr << "Serious error: r=" << r << std::endl; + Print(&std::cerr); + assert(r <= 0.0); + } + --num_customers_; + return delta; + } + } + + double prob(const Dish& dish, const double& p0) const { + const typename std::tr1::unordered_map::const_iterator it = dish_locs_.find(dish); + const double r = num_tables_ * discount_ + concentration_; + if (it == dish_locs_.end()) { + return r * p0 / (num_customers_ + concentration_); + } else { + return (it->second.total_dish_count_ - discount_ * it->second.table_counts_.size() + r * p0) / + (num_customers_ + concentration_); + } + } + + template + T probT(const Dish& dish, const T& p0) const { + const typename std::tr1::unordered_map::const_iterator it = dish_locs_.find(dish); + const T r = T(num_tables_ * discount_ + concentration_); + if (it == dish_locs_.end()) { + return r * p0 / T(num_customers_ + concentration_); + } else { + return (T(it->second.total_dish_count_ - discount_ * it->second.table_counts_.size()) + r * p0) / + T(num_customers_ + concentration_); + } + } + + double log_crp_prob() const { + return log_crp_prob(discount_, concentration_); + } + + static double log_beta_density(const double& x, const double& alpha, const double& beta) { + assert(x > 0.0); + assert(x < 1.0); + assert(alpha > 0.0); + assert(beta > 0.0); + const double lp = (alpha-1)*log(x)+(beta-1)*log(1-x)+lgamma(alpha+beta)-lgamma(alpha)-lgamma(beta); + return lp; + } + + static double log_gamma_density(const double& x, const double& shape, const double& rate) { + assert(x >= 0.0); + assert(shape > 0.0); + assert(rate > 0.0); + const double lp = (shape-1)*log(x) - shape*log(rate) - x/rate - lgamma(shape); + return lp; + } + + // taken from http://en.wikipedia.org/wiki/Chinese_restaurant_process + // does not include P_0's + double log_crp_prob(const double& discount, const double& concentration) const { + double lp = 0.0; + if (has_discount_prior()) + lp = log_beta_density(discount, discount_prior_alpha_, discount_prior_beta_); + if (has_concentration_prior()) + lp += log_gamma_density(concentration, concentration_prior_shape_, concentration_prior_rate_); + assert(lp <= 0.0); + if (num_customers_) { + if (discount > 0.0) { + const double r = lgamma(1.0 - discount); + lp += lgamma(concentration) - lgamma(concentration + num_customers_) + + num_tables_ * log(discount) + lgamma(concentration / discount + num_tables_) + - lgamma(concentration / discount); + assert(std::isfinite(lp)); + for (typename std::tr1::unordered_map::const_iterator it = dish_locs_.begin(); + it != dish_locs_.end(); ++it) { + const DishLocations& cur = it->second; + for (std::list::const_iterator ti = cur.table_counts_.begin(); ti != cur.table_counts_.end(); ++ti) { + lp += lgamma(*ti - discount) - r; + } + } + } else { + assert(!"not implemented yet"); + } + } + assert(std::isfinite(lp)); + return lp; + } + + void resample_hyperparameters(MT19937* rng, const unsigned nloop = 5, const unsigned niterations = 10) { + assert(has_discount_prior() || has_concentration_prior()); + DiscountResampler dr(*this); + ConcentrationResampler cr(*this); + for (int iter = 0; iter < nloop; ++iter) { + if (has_concentration_prior()) { + concentration_ = slice_sampler1d(cr, concentration_, *rng, 0.0, + std::numeric_limits::infinity(), 0.0, niterations, 100*niterations); + } + if (has_discount_prior()) { + discount_ = slice_sampler1d(dr, discount_, *rng, std::numeric_limits::min(), + 1.0, 0.0, niterations, 100*niterations); + } + } + concentration_ = slice_sampler1d(cr, concentration_, *rng, 0.0, + std::numeric_limits::infinity(), 0.0, niterations, 100*niterations); + } + + struct DiscountResampler { + DiscountResampler(const CCRP& crp) : crp_(crp) {} + const CCRP& crp_; + double operator()(const double& proposed_discount) const { + return crp_.log_crp_prob(proposed_discount, crp_.concentration_); + } + }; + + struct ConcentrationResampler { + ConcentrationResampler(const CCRP& crp) : crp_(crp) {} + const CCRP& crp_; + double operator()(const double& proposed_concentration) const { + return crp_.log_crp_prob(crp_.discount_, proposed_concentration); + } + }; + + struct DishLocations { + DishLocations() : total_dish_count_() {} + unsigned total_dish_count_; // customers at all tables with this dish + std::list table_counts_; // list<> gives O(1) deletion and insertion, which we want + // .size() is the number of tables for this dish + }; + + void Print(std::ostream* out) const { + std::cerr << "PYP(d=" << discount_ << ",c=" << concentration_ << ") customers=" << num_customers_ << std::endl; + for (typename std::tr1::unordered_map::const_iterator it = dish_locs_.begin(); + it != dish_locs_.end(); ++it) { + (*out) << it->first << " (" << it->second.total_dish_count_ << " on " << it->second.table_counts_.size() << " tables): "; + for (typename std::list::const_iterator i = it->second.table_counts_.begin(); + i != it->second.table_counts_.end(); ++i) { + (*out) << " " << *i; + } + (*out) << std::endl; + } + } + + typedef typename std::tr1::unordered_map::const_iterator const_iterator; + const_iterator begin() const { + return dish_locs_.begin(); + } + const_iterator end() const { + return dish_locs_.end(); + } + + unsigned num_tables_; + unsigned num_customers_; + std::tr1::unordered_map dish_locs_; + + double discount_; + double concentration_; + + // optional beta prior on discount_ (NaN if no prior) + double discount_prior_alpha_; + double discount_prior_beta_; + + // optional gamma prior on concentration_ (NaN if no prior) + double concentration_prior_shape_; + double concentration_prior_rate_; +}; + +template +std::ostream& operator<<(std::ostream& o, const CCRP& c) { + c.Print(&o); + return o; +} + +#endif diff --git a/utils/ccrp_onetable.h b/utils/ccrp_onetable.h index a868af9a..b63737d1 100644 --- a/utils/ccrp_onetable.h +++ b/utils/ccrp_onetable.h @@ -117,6 +117,18 @@ class CCRP_OneTable { } } + template + T probT(const Dish& dish, const T& p0) const { + const typename DishMapType::const_iterator it = dish_counts_.find(dish); + const T r(num_tables_ * discount_ + concentration_); + if (it == dish_counts_.end()) { + return r * p0 / T(num_customers_ + concentration_); + } else { + return (T(it->second - discount_) + r * p0) / + T(num_customers_ + concentration_); + } + } + double log_crp_prob() const { return log_crp_prob(discount_, concentration_); } diff --git a/utils/sampler.h b/utils/sampler.h index 153e7ef1..22c873d4 100644 --- a/utils/sampler.h +++ b/utils/sampler.h @@ -48,7 +48,7 @@ struct RandomNumberGenerator { template size_t SelectSample(const F& a, const F& b, double T = 1.0) { if (T == 1.0) { - if (this->next() > (a / (a + b))) return 1; else return 0; + if (F(this->next()) > (a / (a + b))) return 1; else return 0; } else { assert(!"not implemented"); } -- cgit v1.2.3