3 files changed, 34 insertions, 8 deletions

diff --git a/training/mpi_online_optimize.cc b/training/mpi_online_optimize.cc
index 325ba030..32033c19 100644
--- a/training/mpi_online_optimize.cc
+++ b/training/mpi_online_optimize.cc
@@ -64,6 +64,7 @@ bool InitCommandLine(int argc, char** argv, po::variables_map* conf) {
   po::options_description opts("Configuration options");
   opts.add_options()
         ("input_weights,w",po::value<string>(),"Input feature weights file")
+        ("frozen_features,z",po::value<string>(), "List of features not to optimize")
         ("training_data,t",po::value<string>(),"Training data corpus")
         ("training_agenda,a",po::value<string>(), "Text file listing a series of configuration files and the number of iterations to train using each configuration successively")
         ("minibatch_size_per_proc,s", po::value<unsigned>()->default_value(5), "Number of training instances evaluated per processor in each minibatch")
@@ -254,6 +255,20 @@ int main(int argc, char** argv) {
   if (conf.count("input_weights"))
     weights.InitFromFile(conf["input_weights"].as<string>());
 
+  vector<int> frozen_fids;
+  if (conf.count("frozen_features")) {
+    ReadFile rf(conf["frozen_features"].as<string>());
+    istream& in = *rf.stream();
+    string line;
+    while(in) {
+      getline(in, line);
+      if (line.empty()) continue;
+      if (line[0] == ' ' || line[line.size() - 1] == ' ') { line = Trim(line); }
+      frozen_fids.push_back(FD::Convert(line));
+    }
+    if (rank == 0) cerr << "Freezing " << frozen_fids.size() << " features.\n";
+  }
+
   vector<string> corpus;
   vector<int> ids;
   ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus, &ids);
@@ -284,7 +299,7 @@ int main(int argc, char** argv) {
     const string omethod = conf["optimization_method"].as<string>();
     if (omethod == "sgd") {
       const double C = conf["regularization_strength"].as<double>();
-      o.reset(new CumulativeL1OnlineOptimizer(lr, total_corpus_size, C));
+      o.reset(new CumulativeL1OnlineOptimizer(lr, total_corpus_size, C, frozen_fids));
     } else {
       assert(!"fail");
     }
diff --git a/training/online_optimizer.h b/training/online_optimizer.h
index 312aabae..28d89344 100644
--- a/training/online_optimizer.h
+++ b/training/online_optimizer.h
@@ -2,6 +2,7 @@
 #define _ONL_OPTIMIZE_H_
 
 #include <tr1/memory>
+#include <set>
 #include <string>
 #include <cmath>
 #include "sparse_vector.h"
@@ -56,8 +57,12 @@ class OnlineOptimizer {
  public:
   virtual ~OnlineOptimizer();
   OnlineOptimizer(const std::tr1::shared_ptr<LearningRateSchedule>& s,
-                  size_t batch_size)
-    : N_(batch_size),schedule_(s),k_() {}
+                  size_t batch_size,
+                  const std::vector<int>& frozen_feats = std::vector<int>())
+      : N_(batch_size),schedule_(s),k_() {
+    for (int i = 0; i < frozen_feats.size(); ++i)
+      frozen_.insert(frozen_feats[i]);
+  }
   void ResetEpoch() { k_ = 0; ResetEpochImpl(); }
   void UpdateWeights(const SparseVector<double>& approx_g, int max_feat, SparseVector<double>* weights) {
     ++k_;
@@ -69,6 +74,7 @@ class OnlineOptimizer {
   virtual void ResetEpochImpl();
   virtual void UpdateWeightsImpl(const double& eta, const SparseVector<double>& approx_g, int max_feat, SparseVector<double>* weights) = 0;
   const size_t N_; // number of training instances per batch
+  std::set<int> frozen_;  // frozen (non-optimizing) features
 
  private:
   std::tr1::shared_ptr<LearningRateSchedule> schedule_;
@@ -78,16 +84,21 @@ class OnlineOptimizer {
 class CumulativeL1OnlineOptimizer : public OnlineOptimizer {
  public:
   CumulativeL1OnlineOptimizer(const std::tr1::shared_ptr<LearningRateSchedule>& s,
-                              size_t training_instances, double C) :
-    OnlineOptimizer(s, training_instances), C_(C), u_() {}
+                              size_t training_instances, double C,
+                              const std::vector<int>& frozen) :
+    OnlineOptimizer(s, training_instances, frozen), C_(C), u_() {}
 
  protected:
   void ResetEpochImpl() { u_ = 0; }
   void UpdateWeightsImpl(const double& eta, const SparseVector<double>& approx_g, int max_feat, SparseVector<double>* weights) {
     u_ += eta * C_ / N_;
-    (*weights) += eta * approx_g;
+    for (SparseVector<double>::const_iterator it = approx_g.begin(); 
+         it != approx_g.end(); ++it) {
+      if (frozen_.count(it->first) == 0)
+        weights->add_value(it->first, eta * it->second);
+    }
     for (int i = 1; i < max_feat; ++i)
-      ApplyPenalty(i, weights);
+      if (frozen_.count(i) == 0) ApplyPenalty(i, weights);
   }
 
  private:
diff --git a/training/optimize_test.cc b/training/optimize_test.cc
index 6fa5efd4..fe7ca70f 100644
--- a/training/optimize_test.cc
+++ b/training/optimize_test.cc
@@ -104,7 +104,7 @@ void TestOnline() {
   double eta0 = 0.2;
   shared_ptr<LearningRateSchedule> r(new ExponentialDecayLearningRate(N, eta0, 0.85));
   //shared_ptr<LearningRateSchedule> r(new StandardLearningRate(N, eta0));
-  CumulativeL1OnlineOptimizer opt(r, N, C);
+  CumulativeL1OnlineOptimizer opt(r, N, C, std::vector<int>());
   assert(r->eta(10) < r->eta(1));
 }