fast creg training code for univariate linear and logistic regression

4 files changed, 358 insertions, 13 deletions

diff --git a/training/Makefile.am b/training/Makefile.am
index 991ac210..4b69ea94 100644
--- a/training/Makefile.am
+++ b/training/Makefile.am
@@ -1,5 +1,6 @@
 bin_PROGRAMS = \
   model1 \
+  creg \
   lbl_model \
   test_ngram \
   mr_em_map_adapter \
@@ -23,6 +24,9 @@ noinst_PROGRAMS = \
 
 TESTS = lbfgs_test optimize_test
 
+creg_SOURCES = creg.cc
+creg_LDADD = ./liblbfgs/liblbfgs.a $(top_srcdir)/utils/libutils.a -lz
+
 mpi_online_optimize_SOURCES = mpi_online_optimize.cc online_optimizer.cc
 mpi_online_optimize_LDADD = $(top_srcdir)/decoder/libcdec.a $(top_srcdir)/mteval/libmteval.a $(top_srcdir)/utils/libutils.a ../klm/lm/libklm.a ../klm/util/libklm_util.a -lz
 
diff --git a/training/creg.cc b/training/creg.cc
new file mode 100644
index 00000000..58adea00
--- /dev/null
+++ b/training/creg.cc
@@ -0,0 +1,334 @@
+#include <cstdlib>
+#include <iostream>
+#include <vector>
+#include <tr1/unordered_map>
+
+#include <boost/program_options.hpp>
+#include <boost/program_options/variables_map.hpp>
+
+#include "json_feature_map_lexer.h"
+#include "prob.h"
+#include "filelib.h"
+#include "weights.h"
+#include "sparse_vector.h"
+#include "liblbfgs/lbfgs++.h"
+
+using namespace std;
+using namespace std::tr1;
+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_features,x", po::value<string>(), "File containing training instance features (ARKRegression format)")
+        ("training_responses,y", po::value<string>(), "File containing training response features (ARKRegression format)")
+        ("linear,n", "Linear (rather than logistic) regression")
+        ("l1",po::value<double>()->default_value(0.0), "l_1 regularization strength")
+        ("l2",po::value<double>()->default_value(0.0), "l_2 regularization strength")
+        ("weights,w", po::value<string>(), "Initial weights")
+        ("epsilon,e", po::value<double>()->default_value(1e-4), "Epsilon for convergence test. Terminates when ||g|| < epsilon * max(1, ||x||)")
+        ("memory_buffers,m",po::value<unsigned>()->default_value(40), "Number of memory buffers for LBFGS")
+        ("help,h", "Help");
+  po::options_description dcmdline_options;
+  dcmdline_options.add(opts);
+  po::store(parse_command_line(argc, argv, dcmdline_options), *conf);
+  if (conf->count("help") || !conf->count("training_features") || !conf->count("training_responses")) {
+    cerr << dcmdline_options << endl;
+    exit(1);
+  }
+}
+
+struct TrainingInstance {
+  SparseVector<float> x;
+  union {
+    unsigned label;  // for categorical predictions
+    float value;     // for continuous predictions
+  } y;
+};
+
+struct ReaderHelper {
+  explicit ReaderHelper(vector<TrainingInstance>* xyp) : xy_pairs(xyp), lc(), flag() {}
+  unordered_map<string, unsigned> id2ind;
+  vector<TrainingInstance>* xy_pairs;
+  int lc;
+  bool flag;
+};
+
+void ReaderCB(const string& id, const SparseVector<float>& fmap, void* extra) {
+  ReaderHelper& rh = *reinterpret_cast<ReaderHelper*>(extra);
+  ++rh.lc;
+  if (rh.lc % 1000 == 0) { cerr << '.'; rh.flag = true; }
+  if (rh.lc % 40000 == 0) { cerr << " [" << rh.lc << "]\n"; rh.flag = false; }
+  const unordered_map<string, unsigned>::iterator it = rh.id2ind.find(id);
+  if (it == rh.id2ind.end()) {
+    cerr << "Unlabeled example in line " << rh.lc << endl;
+    abort();
+  }
+  (*rh.xy_pairs)[it->second - 1].x = fmap;
+}
+
+void ReadLabeledInstances(const string& ffeats,
+                 const string& fresp,
+                 const bool is_continuous,
+                 vector<TrainingInstance>* xy_pairs,
+                 vector<string>* labels) {
+  bool flag = false;
+  xy_pairs->clear();
+  int lc = 0;
+  ReaderHelper rh(xy_pairs);
+  unordered_map<string, unsigned> label2id;
+  cerr << "Reading training responses from " << fresp << " ..." << endl;
+  ReadFile fr(fresp);
+  for (unsigned i = 0; i < labels->size(); ++i)
+    label2id[(*labels)[i]] = i;
+  istream& in = *fr.stream();
+  string line;
+  while(getline(in, line)) {
+    ++lc;
+    if (lc % 1000 == 0) { cerr << '.'; flag = true; }
+    if (lc % 40000 == 0) { cerr << " [" << lc << "]\n"; flag = false; }
+    if (line.size() == 0) continue;
+    if (line[0] == '#') continue;
+    unsigned p = 0;
+    while (p < line.size() && line[p] != ' ' && line[p] != '\t') { ++p; }
+    unsigned& ind = rh.id2ind[line.substr(0, p)];
+    if (ind != 0) { cerr << "ID " << line.substr(0, p) << " duplicated in line " << lc << endl; abort(); }
+    while (p < line.size() && (line[p] == ' ' || line[p] == '\t')) { ++p; }
+    assert(p < line.size());
+    xy_pairs->push_back(TrainingInstance());
+    ind = xy_pairs->size();
+    if (is_continuous) {
+      xy_pairs->back().y.value = strtof(&line[p], 0);
+    } else { // categorical predictions
+      unordered_map<string, unsigned>::iterator it = label2id.find(line.substr(p));
+      if (it == label2id.end()) {
+        const string label = line.substr(p);
+        it = label2id.insert(make_pair(label, labels->size())).first;
+        labels->push_back(label);
+      }
+      xy_pairs->back().y.label = it->second;  // label id
+    }
+  }
+  if (flag) cerr << endl;
+  if (!is_continuous) {
+    cerr << "LABELS:";
+    for (unsigned j = 0; j < labels->size(); ++j)
+      cerr << " " << (*labels)[j];
+    cerr << endl;
+  }
+  cerr << "Reading training features from " << ffeats << " ..." << endl;
+  ReadFile ff(ffeats);
+  JSONFeatureMapLexer::ReadRules(ff.stream(), ReaderCB, &rh);
+  if (rh.flag) cerr << endl;
+}
+
+// helper base class (not polymorphic- just a container and some helper functions) for loss functions
+// real loss functions should implement double operator()(const vector<double>& x, double* g),
+// which should evaluate f(x) and g = f'(x)
+struct BaseLoss {
+  // dimp1 = number of categorial outputs possible for logistic regression
+  // for linear regression, it should be 1 more than the dimension of the response variable
+  BaseLoss(
+      const vector<TrainingInstance>& tr,
+      unsigned dimp1,
+      unsigned numfeats,
+      unsigned ll2) : training(tr), K(dimp1), p(numfeats), l2(ll2) {}
+
+  // weight vector layout for K classes, with p features
+  //   w[0 : K-1] = bias weights
+  //   w[y*p + K : y*p + K + p - 1] = feature weights for y^th class
+  // this representation is used in ComputeDotProducts and GradAdd
+  void ComputeDotProducts(const SparseVector<float>& fx,  // feature vector of x
+                          const vector<double>& w,         // full weight vector
+                          vector<double>* pdotprods) const {
+    vector<double>& dotprods = *pdotprods;
+    const unsigned km1 = K - 1;
+    dotprods.resize(km1);
+    for (unsigned y = 0; y < km1; ++y)
+      dotprods[y] = w[y];  // bias terms
+    for (SparseVector<float>::const_iterator it = fx.begin(); it != fx.end(); ++it) {
+      const float fval = it->second;
+      const unsigned fid = it->first;
+      for (unsigned y = 0; y < km1; ++y)
+        dotprods[y] += w[fid + y * p + km1] * fval;
+    }
+  }
+
+  double ApplyRegularizationTerms(const vector<double>& weights,
+                                  double* g) const {
+    double reg = 0;
+    for (size_t i = K - 1; i < weights.size(); ++i) {
+      const double& w_i = weights[i];
+      reg += l2 * w_i * w_i;
+      g[i] += 2 * l2 * w_i;
+    }
+    return reg;
+  }
+
+  void GradAdd(const SparseVector<float>& fx,
+               const unsigned y,
+               const double scale,
+               double* acc) const {
+    acc[y] += scale; // class bias
+    for (SparseVector<float>::const_iterator it = fx.begin();
+         it != fx.end(); ++it)
+      acc[it->first + y * p + K - 1] += it->second * scale;
+  }
+
+  const vector<TrainingInstance>& training;
+  const unsigned K, p;
+  const double l2;
+};
+
+struct UnivariateSquaredLoss : public BaseLoss {
+  UnivariateSquaredLoss(
+          const vector<TrainingInstance>& tr,
+          unsigned numfeats,
+          const double l2) : BaseLoss(tr, 2, numfeats, l2) {}
+
+  // evaluate squared loss and gradient
+  double operator()(const vector<double>& x, double* g) const {
+    fill(g, g + x.size(), 0.0);
+    double cll = 0;
+    vector<double> dotprods(1);  // univariate prediction
+    for (int i = 0; i < training.size(); ++i) {
+      const SparseVector<float>& fmapx = training[i].x;
+      const double refy = training[i].y.value;
+      ComputeDotProducts(fmapx, x, &dotprods);
+      double diff = dotprods[0] - refy;
+      cll += diff * diff;
+
+      double scale = 2 * diff;
+      GradAdd(fmapx, 0, scale, g);
+    }
+    double reg = ApplyRegularizationTerms(x, g);
+    return cll + reg;
+  }
+};
+
+struct MulticlassLogLoss : public BaseLoss {
+  MulticlassLogLoss(
+          const vector<TrainingInstance>& tr,
+          unsigned k,
+          unsigned numfeats,
+          const double l2) : BaseLoss(tr, k, numfeats, l2) {}
+
+  // evaluate log loss and gradient
+  double operator()(const vector<double>& x, double* g) const {
+    fill(g, g + x.size(), 0.0);
+    vector<double> dotprods(K - 1);  // K-1 degrees of freedom
+    vector<prob_t> probs(K);
+    double cll = 0;
+    for (int i = 0; i < training.size(); ++i) {
+      const SparseVector<float>& fmapx = training[i].x;
+      const unsigned refy = training[i].y.label;
+      //cerr << "FMAP: " << fmapx << endl;
+      ComputeDotProducts(fmapx, x, &dotprods);
+      prob_t z;
+      for (unsigned j = 0; j < dotprods.size(); ++j)
+        z += (probs[j] = prob_t(dotprods[j], init_lnx()));
+      z += (probs.back() = prob_t::One());
+      for (unsigned y = 0; y < probs.size(); ++y) {
+        probs[y] /= z;
+        //cerr << "  p(y=" << y << ")=" << probs[y].as_float() << "\tz=" << z << endl;
+      }
+      cll -= log(probs[refy]);  // log p(y | x)
+
+      for (unsigned y = 0; y < dotprods.size(); ++y) {
+        double scale = probs[y].as_float();
+        if (y == refy) { scale -= 1.0; }
+        GradAdd(fmapx, y, scale, g);
+      }
+    }
+    double reg = ApplyRegularizationTerms(x, g);
+    return cll + reg;
+  }
+};
+
+template <class LossFunction>
+double LearnParameters(LossFunction& loss,
+                       const double l1,
+                       const unsigned l1_start,
+                       const unsigned memory_buffers,
+                       const double eps,
+                       vector<double>* px) {
+  LBFGS<LossFunction> lbfgs(px, loss, memory_buffers, l1, l1_start, eps);
+  lbfgs.MinimizeFunction();
+  return 0;
+}
+
+int main(int argc, char** argv) {
+  po::variables_map conf;
+  InitCommandLine(argc, argv, &conf);
+  string line;
+  vector<TrainingInstance> training;
+  const string xfile = conf["training_features"].as<string>();
+  const string yfile = conf["training_responses"].as<string>();
+  double l1 = conf["l1"].as<double>();
+  double l2 = conf["l2"].as<double>();
+  const unsigned memory_buffers = conf["memory_buffers"].as<unsigned>();
+  const double epsilon = conf["epsilon"].as<double>();
+  if (l1 < 0.0) {
+    cerr << "L1 strength must be >= 0\n";
+    return 1;
+  }
+  if (l2 < 0.0) {
+    cerr << "L2 strength must be >= 0\n";
+    return 2;
+  }
+
+  const bool is_continuous = conf.count("linear");
+  vector<string> labels; // only populated for non-continuous models
+  ReadLabeledInstances(xfile, yfile, is_continuous, &training, &labels);
+
+  if (conf.count("weights")) {
+    cerr << "Initial weights are not implemented, please implement." << endl;
+    // TODO read weights for categorical and continuous predictions
+    // can't use normal cdec weight framework
+    abort();
+  }
+
+  cerr << "         Number of features: " << FD::NumFeats() << endl;
+  cerr << "Number of training examples: " << training.size() << endl;
+  const unsigned p = FD::NumFeats();
+  cout.precision(15);
+
+  if (conf.count("linear")) {  // linear regression
+    vector<double> weights(1 + FD::NumFeats(), 0.0);
+    cerr << "       Number of parameters: " << weights.size() << endl;
+    UnivariateSquaredLoss loss(training, p, l2);
+    LearnParameters(loss, l1, 1, memory_buffers, epsilon, &weights);
+    cout << p << "\t***CONTINUOUS***" << endl;
+    cout << "***BIAS***\t" << weights[0] << endl;
+    for (unsigned f = 0; f < p; ++f) {
+      const double w = weights[1 + f];
+      if (w)
+        cout << FD::Convert(f) << "\t" << w << endl;
+    }
+  } else {                     // logistic regression
+    vector<double> weights((1 + FD::NumFeats()) * (labels.size() - 1), 0.0);
+    cerr << "       Number of parameters: " << weights.size() << endl;
+    cerr << "           Number of labels: " << labels.size() << endl;
+    const unsigned K = labels.size();
+    const unsigned km1 = K - 1;
+    MulticlassLogLoss loss(training, K, p, l2);
+    LearnParameters(loss, l1, km1, memory_buffers, epsilon, &weights);
+
+    cout << p << "\t***CATEGORICAL***";
+    for (unsigned y = 0; y < K; ++y)
+      cout << '\t' << labels[y];
+    cout << endl;
+    for (unsigned y = 0; y < km1; ++y)
+      cout << labels[y] << "\t***BIAS***\t" << weights[y] << endl;
+    for (unsigned y = 0; y < km1; ++y) {
+      for (unsigned f = 0; f < p; ++f) {
+        const double w = weights[km1 + y * p + f];
+        if (w)
+          cout << labels[y] << "\t" << FD::Convert(f) << "\t" << w << endl;
+      }
+    }
+  }
+
+  return 0;
+}
+
diff --git a/training/liblbfgs/lbfgs++.h b/training/liblbfgs/lbfgs++.h
index 342f9b0e..92ead955 100644
--- a/training/liblbfgs/lbfgs++.h
+++ b/training/liblbfgs/lbfgs++.h
@@ -16,28 +16,33 @@
 template <typename Function>
 class LBFGS {
  public:
-  LBFGS(size_t n,            // number of variables
-        const Function& f,   // function to optimize
-        double l1_c = 0.0,   // l1 penalty strength
-        size_t m = 10        // number of memory buffers
-                             // TODO should use custom allocator here:
+  LBFGS(size_t n,              // number of variables
+        const Function& f,     // function to optimize
+        size_t m = 10,         // number of memory buffers
+        double l1_c = 0.0,     // l1 penalty strength
+        unsigned l1_start = 0, // l1 penalty starting index
+        double eps = 1e-5      // convergence epsilon
+                               // TODO should use custom allocator here:
         ) : p_x(new std::vector<lbfgsfloatval_t>(n, 0.0)),
                              owned(true),
                              m_x(*p_x),
                              func(f) {
-    Init(m, l1_c);
+    Init(m, l1_c, l1_start, eps);
   }
 
   // constructor where external vector storage for variables is used
   LBFGS(std::vector<lbfgsfloatval_t>* px,
         const Function& f,
-        double l1_c = 0.0,   // l1 penalty strength
-        size_t m = 10
+        size_t m = 10,         // number of memory buffers
+        double l1_c = 0.0,     // l1 penalty strength
+        unsigned l1_start = 0, // l1 penalty starting index
+        double eps = 1e-5      // convergence epsilon
+                               // TODO should use custom allocator here:
         ) : p_x(px),
                              owned(false),
                              m_x(*p_x),
                              func(f) {
-    Init(m, l1_c);
+    Init(m, l1_c, l1_start, eps);
   }
 
   ~LBFGS() {
@@ -60,12 +65,14 @@ class LBFGS {
   }
 
  private:
-  void Init(size_t m, double l1_c) {
+  void Init(size_t m, double l1_c, unsigned l1_start, double eps) {
     lbfgs_parameter_init(&param);
     param.m = m;
+    param.epsilon = eps;
     if (l1_c > 0.0) {
       param.linesearch = LBFGS_LINESEARCH_BACKTRACKING;
-      param.orthantwise_c = 1.0;
+      param.orthantwise_c = l1_c;
+      param.orthantwise_start = l1_start;
     }
     silence = false;
   }
diff --git a/training/liblbfgs/ll_test.cc b/training/liblbfgs/ll_test.cc
index 43c0f214..48bc0366 100644
--- a/training/liblbfgs/ll_test.cc
+++ b/training/liblbfgs/ll_test.cc
@@ -5,7 +5,7 @@ using namespace std;
 
 // Function must be lbfgsfloatval_t f(x.begin, x.end, g.begin)
 lbfgsfloatval_t func(const vector<lbfgsfloatval_t>& x, lbfgsfloatval_t* g) {
-    int i;
+    unsigned i;
     lbfgsfloatval_t fx = 0.0;
 
     for (i = 0;i < x.size();i += 2) {
@@ -24,7 +24,7 @@ void Opt(F& f) {
   lbfgs.MinimizeFunction();
 }
 
-int main(int argc, char** argv) {
+int main() {
   Opt(func);
   return 0;
 }