updates to pro to support regularization to previous weight vectors, regualarization normalization, disable broken regularization tuning

2 files changed, 66 insertions, 38 deletions

diff --git a/pro-train/dist-pro.pl b/pro-train/dist-pro.pl
index dbfa329a..4bc9cfe3 100755
--- a/pro-train/dist-pro.pl
+++ b/pro-train/dist-pro.pl
@@ -41,6 +41,7 @@ my $lines_per_mapper = 30;
 my $iteration = 1;
 my $run_local = 0;
 my $best_weights;
+my $psi = 1;
 my $max_iterations = 30;
 my $decode_nodes = 15;   # number of decode nodes
 my $pmem = "4g";
@@ -62,7 +63,8 @@ my $cpbin=1;
 
 # regularization strength
 my $tune_regularizer = 0;
-my $reg = 1e-2;
+my $reg = 10;
+my $reg_previous = 0;
 
 # Process command-line options
 Getopt::Long::Configure("no_auto_abbrev");
@@ -73,10 +75,12 @@ if (GetOptions(
         "use-fork" => \$usefork,
 	"dry-run" => \$dryrun,
 	"epsilon=s" => \$epsilon,
+	"interpolate-with-weights=f" => \$psi,
 	"help" => \$help,
         "weights=s" => \$initial_weights,
 	"tune-regularizer" => \$tune_regularizer,
 	"reg=f" => \$reg,
+	"reg-previous=f" => \$reg_previous,
 	"local" => \$run_local,
 	"use-make=i" => \$use_make,
 	"max-iterations=i" => \$max_iterations,
@@ -91,6 +95,8 @@ if (GetOptions(
 	exit;
 }
 
+die "--tune-regularizer is no longer supported with --reg-previous and --reg. Please tune manually.\n" if $tune_regularizer;
+
 if ($usefork) { $usefork = "--use-fork"; } else { $usefork = ''; }
 
 if ($metric =~ /^(combi|ter)$/i) {
@@ -411,7 +417,7 @@ while (1){
 	}
 	print STDERR "\nRUNNING CLASSIFIER (REDUCER)\n";
 	print STDERR unchecked_output("date");
-	$cmd="cat @dev_outs | $REDUCER -w $dir/weights.$im1 -s $reg";
+	$cmd="cat @dev_outs | $REDUCER -w $dir/weights.$im1 -C $reg -y $reg_previous --interpolate_with_weights $psi";
 	if ($tune_regularizer) {
 		$cmd .= " -T -t $dev_test_file";
 	}
@@ -605,11 +611,21 @@ General options:
 
 Regularization options:
 
+	--interpolate-with-weights <F>
+		[deprecated] At each iteration the resulting weights are
+                interpolated with the weights from the previous iteration, with
+                this factor.
+
 	--tune-regularizer
 		Hold out one third of the tuning data and used this to tune the
-		regularization parameter.
+		regularization parameter. [this doesn't work well]
 
 	--reg <F>
+		l2 regularization strength
+
+	--reg-previous <F>
+		l2 penalty for moving away from the weights from the previous
+		iteration.
 
 Help
 }
diff --git a/pro-train/mr_pro_reduce.cc b/pro-train/mr_pro_reduce.cc
index aff410a0..98cddba2 100644
--- a/pro-train/mr_pro_reduce.cc
+++ b/pro-train/mr_pro_reduce.cc
@@ -23,11 +23,12 @@ void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   po::options_description opts("Configuration options");
   opts.add_options()
         ("weights,w", po::value<string>(), "Weights from previous iteration (used as initialization and interpolation")
-        ("interpolation,p",po::value<double>()->default_value(0.9), "Output weights are p*w + (1-p)*w_prev")
+        ("regularize_to_weights,y",po::value<double>()->default_value(0.0), "Differences in learned weights to previous weights are penalized with an l2 penalty with this strength; 0.0 = no effect")
+        ("interpolate_with_weights,p",po::value<double>()->default_value(1.0), "Output weights are p*w + (1-p)*w_prev; 1.0 = no effect")
         ("memory_buffers,m",po::value<unsigned>()->default_value(200), "Number of memory buffers (LBFGS)")
-        ("sigma_squared,s",po::value<double>()->default_value(0.1), "Sigma squared for Gaussian prior")
-        ("min_reg,r",po::value<double>()->default_value(1e-8), "When tuning (-T) regularization strength, minimum regularization strenght")
-        ("max_reg,R",po::value<double>()->default_value(10.0), "When tuning (-T) regularization strength, maximum regularization strenght")
+        ("regularization_strength,C",po::value<double>()->default_value(1.0), "l2 regularization strength")
+        ("min_reg,r",po::value<double>()->default_value(0.01), "When tuning (-T) regularization strength, minimum regularization strenght")
+        ("max_reg,R",po::value<double>()->default_value(1e6), "When tuning (-T) regularization strength, maximum regularization strenght")
         ("testset,t",po::value<string>(), "Optional held-out test set")
         ("tune_regularizer,T", "Use the held out test set (-t) to tune the regularization strength")
         ("help,h", "Help");
@@ -95,6 +96,27 @@ void GradAdd(const SparseVector<weight_t>& v, const double scale, vector<weight_
   }
 }
 
+double ApplyRegularizationTerms(const double C,
+                                const double T,
+                                const vector<weight_t>& weights,
+                                const vector<weight_t>& prev_weights,
+                                vector<weight_t>* g) {
+  assert(weights.size() == g->size());
+  double reg = 0;
+  for (size_t i = 0; i < weights.size(); ++i) {
+    const double prev_w_i = (i < prev_weights.size() ? prev_weights[i] : 0.0);
+    const double& w_i = weights[i];
+    double& g_i = (*g)[i];
+    reg += C * w_i * w_i;
+    g_i += 2 * C * w_i;
+
+    const double diff_i = w_i - prev_w_i;
+    reg += T * diff_i * diff_i;
+    g_i += 2 * T * diff_i;
+  }
+  return reg;
+}
+
 double TrainingInference(const vector<weight_t>& x,
                          const vector<pair<bool, SparseVector<weight_t> > >& corpus,
                          vector<weight_t>* g = NULL) {
@@ -134,8 +156,10 @@ double TrainingInference(const vector<weight_t>& x,
 // return held-out log likelihood
 double LearnParameters(const vector<pair<bool, SparseVector<weight_t> > >& training,
                        const vector<pair<bool, SparseVector<weight_t> > >& testing,
-                       const double sigsq,
+                       const double C,
+                       const double T,
                        const unsigned memory_buffers,
+                       const vector<weight_t>& prev_x,
                        vector<weight_t>* px) {
   vector<weight_t>& x = *px;
   vector<weight_t> vg(FD::NumFeats(), 0.0);
@@ -157,26 +181,12 @@ double LearnParameters(const vector<pair<bool, SparseVector<weight_t> > >& train
     }
 
     // handle regularizer
-#if 1
-    double norm = 0;
-    for (int i = 1; i < x.size(); ++i) {
-      const double mean_i = 0.0;
-      const double param = (x[i] - mean_i);
-      norm += param * param;
-      vg[i] += param / sigsq;
-    } 
-    const double reg = norm / (2.0 * sigsq);
-#else
-    double reg = 0;
-#endif
+    double reg = ApplyRegularizationTerms(C, T, x, prev_x, &vg);
     cll += reg;
-    cerr << cll << " (REG=" << reg << ")\tPPL=" << ppl << "\t TEST_PPL=" << tppl << "\t";
+    cerr << cll << " (REG=" << reg << ")\tPPL=" << ppl << "\t TEST_PPL=" << tppl << "\t" << endl;
     try {
-      vector<weight_t> old_x = x;
-      do {
-        opt.Optimize(cll, vg, &x);
-        converged = opt.HasConverged();
-      } while (!converged && x == old_x);
+      opt.Optimize(cll, vg, &x);
+      converged = opt.HasConverged();
     } catch (...) {
       cerr << "Exception caught, assuming convergence is close enough...\n";
       converged = true;
@@ -201,13 +211,14 @@ int main(int argc, char** argv) {
   }
   const double min_reg = conf["min_reg"].as<double>();
   const double max_reg = conf["max_reg"].as<double>();
-  double sigsq = conf["sigma_squared"].as<double>(); // will be overridden if parameter is tuned
-  assert(sigsq > 0.0);
+  double C = conf["regularization_strength"].as<double>(); // will be overridden if parameter is tuned
+  const double T = conf["regularize_to_weights"].as<double>();
+  assert(C > 0.0);
   assert(min_reg > 0.0);
   assert(max_reg > 0.0);
   assert(max_reg > min_reg);
-  const double psi = conf["interpolation"].as<double>();
-  if (psi < 0.0 || psi > 1.0) { cerr << "Invalid interpolation weight: " << psi << endl; }
+  const double psi = conf["interpolate_with_weights"].as<double>();
+  if (psi < 0.0 || psi > 1.0) { cerr << "Invalid interpolation weight: " << psi << endl; return 1; }
   ReadCorpus(&cin, &training);
   if (conf.count("testset")) {
     ReadFile rf(conf["testset"].as<string>());
@@ -231,14 +242,15 @@ int main(int argc, char** argv) {
   vector<pair<double,double> > sp;
   vector<double> smoothed;
   if (tune_regularizer) {
-    sigsq = min_reg;
+    C = min_reg;
     const double steps = 18;
     double sweep_factor = exp((log(max_reg) - log(min_reg)) / steps);
     cerr << "SWEEP FACTOR: " << sweep_factor << endl;
-    while(sigsq < max_reg) {
-      tppl = LearnParameters(training, testing, sigsq, conf["memory_buffers"].as<unsigned>(), &x);
-      sp.push_back(make_pair(sigsq, tppl));
-      sigsq *= sweep_factor;
+    while(C < max_reg) {
+      cerr << "C=" << C << "\tT=" <<T << endl;
+      tppl = LearnParameters(training, testing, C, T, conf["memory_buffers"].as<unsigned>(), prev_x, &x);
+      sp.push_back(make_pair(C, tppl));
+      C *= sweep_factor;
     }
     smoothed.resize(sp.size(), 0);
     smoothed[0] = sp[0].second;
@@ -257,16 +269,16 @@ int main(int argc, char** argv) {
         best_i = i;
       }
     }
-    sigsq = sp[best_i].first;
+    C = sp[best_i].first;
   }  // tune regularizer
-  tppl = LearnParameters(training, testing, sigsq, conf["memory_buffers"].as<unsigned>(), &x);
+  tppl = LearnParameters(training, testing, C, T, conf["memory_buffers"].as<unsigned>(), prev_x, &x);
   if (conf.count("weights")) {
     for (int i = 1; i < x.size(); ++i) {
       x[i] = (x[i] * psi) + prev_x[i] * (1.0 - psi);
     }
   }
   cout.precision(15);
-  cout << "# sigma^2=" << sigsq << "\theld out perplexity=";
+  cout << "# C=" << C << "\theld out perplexity=";
   if (tppl) { cout << tppl << endl; } else { cout << "N/A\n"; }
   if (sp.size()) {
     cout << "# Parameter sweep:\n";