summaryrefslogtreecommitdiff
path: root/utils
diff options
context:
space:
mode:
authorPatrick Simianer <p@simianer.de>2012-04-23 21:44:02 +0200
committerPatrick Simianer <p@simianer.de>2012-04-23 21:44:02 +0200
commit2f427278616cbe3fa6f56d6b97c40b3894dbd950 (patch)
tree6998435e4677437c474cf0f835ce9f72d70d3945 /utils
parent6d0d0eb6bbfaee6b6998659a55e2195977ccd217 (diff)
parent217c4aaeba1c9f19b3420b526235bffd86c7a92b (diff)
Merge remote-tracking branch 'upstream/master'
Conflicts: Makefile.am configure.ac
Diffstat (limited to 'utils')
-rw-r--r--utils/alias_sampler.h50
-rw-r--r--utils/logval.h13
-rw-r--r--utils/unigram_pyp_lm.cc168
-rw-r--r--utils/weights.cc4
4 files changed, 231 insertions, 4 deletions
diff --git a/utils/alias_sampler.h b/utils/alias_sampler.h
new file mode 100644
index 00000000..81541f7a
--- /dev/null
+++ b/utils/alias_sampler.h
@@ -0,0 +1,50 @@
+#ifndef _ALIAS_SAMPLER_H_
+#define _ALIAS_SAMPLER_H_
+
+#include <vector>
+#include <limits>
+
+// R. A. Kronmal and A. V. Peterson, Jr. (1977) On the alias method for
+// generating random variables from a discrete distribution. In The American
+// Statistician, Vol. 33, No. 4. Pages 214--218.
+//
+// Intuition: a multinomial with N outcomes can be rewritten as a uniform
+// mixture of N Bernoulli distributions. The ith Bernoulli returns i with
+// probability F[i], otherwise it returns an "alias" value L[i]. The
+// constructor computes the F's and L's given an arbitrary multionimial p in
+// O(n) time and Draw returns samples in O(1) time.
+struct AliasSampler {
+ AliasSampler() {}
+ explicit AliasSampler(const std::vector<double>& p) { Init(p); }
+ void Init(const std::vector<double>& p) {
+ const unsigned N = p.size();
+ cutoffs_.resize(p.size());
+ aliases_.clear();
+ aliases_.resize(p.size(), std::numeric_limits<unsigned>::max());
+ std::vector<unsigned> s,g;
+ for (unsigned i = 0; i < N; ++i) {
+ const double cutoff = cutoffs_[i] = N * p[i];
+ if (cutoff >= 1.0) g.push_back(i); else s.push_back(i);
+ }
+ while(!s.empty() && !g.empty()) {
+ const unsigned k = g.back();
+ const unsigned j = s.back();
+ aliases_[j] = k;
+ cutoffs_[k] -= 1.0 - cutoffs_[j];
+ s.pop_back();
+ if (cutoffs_[k] < 1.0) {
+ g.pop_back();
+ s.push_back(k);
+ }
+ }
+ }
+ template <typename Uniform01Generator>
+ unsigned Draw(Uniform01Generator& u01) const {
+ const unsigned n = u01() * cutoffs_.size();
+ if (u01() > cutoffs_[n]) return aliases_[n]; else return n;
+ }
+ std::vector<double> cutoffs_; // F
+ std::vector<unsigned> aliases_; // L
+};
+
+#endif
diff --git a/utils/logval.h b/utils/logval.h
index 8a59d0b1..ec1f6acd 100644
--- a/utils/logval.h
+++ b/utils/logval.h
@@ -30,8 +30,6 @@ class LogVal {
LogVal(init_minus_1) : s_(true),v_(0) { }
LogVal(init_1) : s_(),v_(0) { }
LogVal(init_0) : s_(),v_(LOGVAL_LOG0) { }
- explicit LogVal(int x) : s_(x<0), v_(s_ ? std::log(-x) : std::log(x)) {}
- explicit LogVal(unsigned x) : s_(0), v_(std::log(x)) { }
LogVal(double lnx,bool sign) : s_(sign),v_(lnx) {}
LogVal(double lnx,init_lnx) : s_(),v_(lnx) {}
static Self exp(T lnx) { return Self(lnx,false); }
@@ -126,7 +124,7 @@ class LogVal {
}
Self operator-() const {
- return Self(v_,-s_);
+ return Self(v_,!s_);
}
void negate() { s_ = !s_; }
@@ -193,6 +191,15 @@ T log(const LogVal<T>& o) {
return o.v_;
}
+template<class T>
+LogVal<T> abs(const LogVal<T>& o) {
+ if (o.s_) {
+ LogVal<T> res = o;
+ res.s_ = false;
+ return res;
+ } else { return o; }
+}
+
template <class T>
LogVal<T> pow(const LogVal<T>& b, const T& e) {
return b.pow(e);
diff --git a/utils/unigram_pyp_lm.cc b/utils/unigram_pyp_lm.cc
new file mode 100644
index 00000000..510e8839
--- /dev/null
+++ b/utils/unigram_pyp_lm.cc
@@ -0,0 +1,168 @@
+#include <iostream>
+#include <tr1/memory>
+#include <queue>
+
+#include <boost/functional.hpp>
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#include "corpus_tools.h"
+#include "m.h"
+#include "tdict.h"
+#include "sampler.h"
+#include "ccrp.h"
+
+// A not very memory-efficient implementation of an 1-gram LM based on PYPs
+// as described in Y.-W. Teh. (2006) A Hierarchical Bayesian Language Model
+// based on Pitman-Yor Processes. In Proc. ACL.
+
+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,n",po::value<unsigned>()->default_value(50),"Number of samples")
+ ("train,i",po::value<string>(),"Training data file")
+ ("test,T",po::value<string>(),"Test data file")
+ ("discount_prior_a,a",po::value<double>()->default_value(1.0), "discount ~ Beta(a,b): a=this")
+ ("discount_prior_b,b",po::value<double>()->default_value(1.0), "discount ~ Beta(a,b): b=this")
+ ("strength_prior_s,s",po::value<double>()->default_value(1.0), "strength ~ Gamma(s,r): s=this")
+ ("strength_prior_r,r",po::value<double>()->default_value(1.0), "strength ~ Gamma(s,r): r=this")
+ ("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("train") == 0)) {
+ cerr << dcmdline_options << endl;
+ exit(1);
+ }
+}
+
+// uniform base distribution (0-gram model)
+struct UniformWordModel {
+ explicit UniformWordModel(unsigned vocab_size) : p0(1.0 / vocab_size), draws() {}
+ void increment() { ++draws; }
+ void decrement() { --draws; assert(draws >= 0); }
+ double prob(WordID) const { return p0; } // all words have equal prob
+ double log_likelihood() const { return draws * log(p0); }
+ const double p0;
+ int draws;
+};
+
+// represents an Unigram LM
+struct UnigramLM {
+ UnigramLM(unsigned vs, double da, double db, double ss, double sr) :
+ uniform_vocab(vs),
+ crp(da, db, ss, sr, 0.8, 1.0) {}
+ void increment(WordID w, MT19937* rng) {
+ const double backoff = uniform_vocab.prob(w);
+ if (crp.increment(w, backoff, rng))
+ uniform_vocab.increment();
+ }
+ void decrement(WordID w, MT19937* rng) {
+ if (crp.decrement(w, rng))
+ uniform_vocab.decrement();
+ }
+ double prob(WordID w) const {
+ const double backoff = uniform_vocab.prob(w);
+ return crp.prob(w, backoff);
+ }
+
+ double log_likelihood() const {
+ double llh = uniform_vocab.log_likelihood();
+ llh += crp.log_crp_prob();
+ return llh;
+ }
+
+ void resample_hyperparameters(MT19937* rng) {
+ crp.resample_hyperparameters(rng);
+ }
+
+ double discount_a, discount_b, strength_s, strength_r;
+ double d, strength;
+ UniformWordModel uniform_vocab;
+ CCRP<WordID> crp;
+};
+
+int main(int argc, char** argv) {
+ po::variables_map conf;
+
+ InitCommandLine(argc, argv, &conf);
+ 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<vector<WordID> > corpuse;
+ set<WordID> vocabe;
+ const WordID kEOS = TD::Convert("</s>");
+ cerr << "Reading corpus...\n";
+ CorpusTools::ReadFromFile(conf["train"].as<string>(), &corpuse, &vocabe);
+ cerr << "E-corpus size: " << corpuse.size() << " sentences\t (" << vocabe.size() << " word types)\n";
+ vector<vector<WordID> > test;
+ if (conf.count("test"))
+ CorpusTools::ReadFromFile(conf["test"].as<string>(), &test);
+ else
+ test = corpuse;
+ UnigramLM lm(vocabe.size(),
+ conf["discount_prior_a"].as<double>(),
+ conf["discount_prior_b"].as<double>(),
+ conf["strength_prior_s"].as<double>(),
+ conf["strength_prior_r"].as<double>());
+ for (int SS=0; SS < samples; ++SS) {
+ for (int ci = 0; ci < corpuse.size(); ++ci) {
+ const vector<WordID>& s = corpuse[ci];
+ for (int i = 0; i <= s.size(); ++i) {
+ WordID w = (i < s.size() ? s[i] : kEOS);
+ if (SS > 0) lm.decrement(w, &rng);
+ lm.increment(w, &rng);
+ }
+ if (SS > 0) lm.decrement(kEOS, &rng);
+ lm.increment(kEOS, &rng);
+ }
+ cerr << "LLH=" << lm.log_likelihood() << endl;
+ //if (SS % 10 == 9) lm.resample_hyperparameters(&rng);
+ }
+ double llh = 0;
+ unsigned cnt = 0;
+ unsigned oovs = 0;
+ for (int ci = 0; ci < test.size(); ++ci) {
+ const vector<WordID>& s = test[ci];
+ for (int i = 0; i <= s.size(); ++i) {
+ WordID w = (i < s.size() ? s[i] : kEOS);
+ double lp = log(lm.prob(w)) / log(2);
+ if (i < s.size() && vocabe.count(w) == 0) {
+ cerr << "**OOV ";
+ ++oovs;
+ lp = 0;
+ }
+ cerr << "p(" << TD::Convert(w) << ") = " << lp << endl;
+ llh -= lp;
+ cnt++;
+ }
+ }
+ cerr << " Log_10 prob: " << (-llh * log(2) / log(10)) << endl;
+ cerr << " Count: " << cnt << endl;
+ cerr << " OOVs: " << oovs << endl;
+ cerr << "Cross-entropy: " << (llh / cnt) << endl;
+ cerr << " Perplexity: " << pow(2, llh / cnt) << endl;
+ return 0;
+}
+
diff --git a/utils/weights.cc b/utils/weights.cc
index ac407dfb..39c18474 100644
--- a/utils/weights.cc
+++ b/utils/weights.cc
@@ -144,8 +144,10 @@ void Weights::ShowLargestFeatures(const vector<weight_t>& w) {
vector<int> fnums(w.size());
for (int i = 0; i < w.size(); ++i)
fnums[i] = i;
+ int nf = FD::NumFeats();
+ if (nf > 10) nf = 10;
vector<int>::iterator mid = fnums.begin();
- mid += (w.size() > 10 ? 10 : w.size());
+ mid += nf;
partial_sort(fnums.begin(), mid, fnums.end(), FComp(w));
cerr << "TOP FEATURES:";
for (vector<int>::iterator i = fnums.begin(); i != mid; ++i) {