Merge branch 'master' of https://github.com/redpony/cdec

17 files changed, 395 insertions, 233 deletions

diff --git a/training/crf/mpi_online_optimize.cc b/training/crf/mpi_online_optimize.cc
index 9e1ae34c..6b5b7d64 100644
--- a/training/crf/mpi_online_optimize.cc
+++ b/training/crf/mpi_online_optimize.cc
@@ -4,11 +4,11 @@
 #include <vector>
 #include <cassert>
 #include <cmath>
-#include <tr1/memory>
 #include <ctime>
 
 #include <boost/program_options.hpp>
 #include <boost/program_options/variables_map.hpp>
+#include <boost/shared_ptr.hpp>
 
 #include "stringlib.h"
 #include "verbose.h"
@@ -219,7 +219,7 @@ int main(int argc, char** argv) {
 #endif
   if (size > 1) SetSilent(true);  // turn off verbose decoder output
   register_feature_functions();
-  std::tr1::shared_ptr<MT19937> rng;
+  boost::shared_ptr<MT19937> rng;
 
   po::variables_map conf;
   if (!InitCommandLine(argc, argv, &conf))
@@ -264,8 +264,8 @@ int main(int argc, char** argv) {
   ReadTrainingCorpus(conf["training_data"].as<string>(), rank, size, &corpus, &ids);
   assert(corpus.size() > 0);
 
-  std::tr1::shared_ptr<OnlineOptimizer> o;
-  std::tr1::shared_ptr<LearningRateSchedule> lr;
+  boost::shared_ptr<OnlineOptimizer> o;
+  boost::shared_ptr<LearningRateSchedule> lr;
 
   const unsigned size_per_proc = conf["minibatch_size_per_proc"].as<unsigned>();
   if (size_per_proc > corpus.size()) {
diff --git a/training/dtrain/Makefile.am b/training/dtrain/Makefile.am
index 844c790d..ecb6c128 100644
--- a/training/dtrain/Makefile.am
+++ b/training/dtrain/Makefile.am
@@ -1,7 +1,7 @@
 bin_PROGRAMS = dtrain
 
 dtrain_SOURCES = dtrain.cc score.cc dtrain.h kbestget.h ksampler.h pairsampling.h score.h
-dtrain_LDADD   = ../../decoder/libcdec.a ../../klm/search/libksearch.a ../../mteval/libmteval.a ../../utils/libutils.a ../../klm/lm/libklm.a ../../klm/util/libklm_util.a ../../klm/util/double-conversion/libklm_util_double.a
+dtrain_LDADD   = ../../decoder/libcdec.a ../../klm/search/libksearch.a ../../mteval/libmteval.a ../../utils/libutils.a ../../klm/lm/libklm.a ../../klm/util/libklm_util.a ../../klm/util/double-conversion/libklm_util_double.a -lboost_regex
 
 AM_CPPFLAGS = -W -Wall -Wno-sign-compare -I$(top_srcdir)/utils -I$(top_srcdir)/decoder -I$(top_srcdir)/mteval
 
diff --git a/training/dtrain/README.md b/training/dtrain/README.md
index 2bae6b48..aa1ab3e7 100644
--- a/training/dtrain/README.md
+++ b/training/dtrain/README.md
@@ -1,10 +1,15 @@
 This is a simple (and parallelizable) tuning method for cdec
-which is able to train the weights of very many (sparse) features.
-It was used here:
-  "Joint Feature Selection in Distributed Stochastic
-   Learning for Large-Scale Discriminative Training in
-   SMT"
-(Simianer, Riezler, Dyer; ACL 2012)
+which is able to train the weights of very many (sparse) features
+on the training set.
+
+It was used in these papers:
+>  "Joint Feature Selection in Distributed Stochastic
+>   Learning for Large-Scale Discriminative Training in
+>   SMT" (Simianer, Riezler, Dyer; ACL 2012)
+>
+>  "Multi-Task Learning for Improved Discriminative
+>   Training in SMT" (Simianer, Riezler; WMT 2013)
+>
 
 
 Building
@@ -17,20 +22,9 @@ To build only parts needed for dtrain do
   cd training/dtrain/; make
 ```
 
-Ideas
------
- * get approx_bleu to work?
- * implement minibatches (Minibatch and Parallelization for Online Large Margin Structured Learning)
- * learning rate 1/T?
- * use an oracle? mira-like (model vs. BLEU), feature repr. of reference!? 
- * implement lc_bleu properly
- * merge kbest lists of previous epochs (as MERT does)
- * ``walk entire regularization path''
- * rerank after each update?
-
 Running
 -------
-See directories under test/ .
+See directories under examples/ .
 
 Legal
 -----
diff --git a/training/dtrain/dtrain.cc b/training/dtrain/dtrain.cc
index 0ee2f124..0a27a068 100644
--- a/training/dtrain/dtrain.cc
+++ b/training/dtrain/dtrain.cc
@@ -12,8 +12,9 @@ dtrain_init(int argc, char** argv, po::variables_map* cfg)
 {
   po::options_description ini("Configuration File Options");
   ini.add_options()
-    ("input",             po::value<string>()->default_value("-"),                                             "input file (src)")
+    ("input",             po::value<string>(),                                                                 "input file (src)")
     ("refs,r",            po::value<string>(),                                                                       "references")
+    ("bitext,b",          po::value<string>(),                                                            "bitext: 'src ||| tgt'")
     ("output",            po::value<string>()->default_value("-"),                          "output weights file, '-' for STDOUT")
     ("input_weights",     po::value<string>(),                                "input weights file (e.g. from previous iteration)")
     ("decoder_config",    po::value<string>(),                                                      "configuration file for cdec")
@@ -40,6 +41,10 @@ dtrain_init(int argc, char** argv, po::variables_map* cfg)
     ("scale_bleu_diff",   po::value<bool>()->zero_tokens(),                      "learning rate <- bleu diff of a misranked pair")
     ("loss_margin",       po::value<weight_t>()->default_value(0.),  "update if no error in pref pair but model scores this near")
     ("max_pairs",         po::value<unsigned>()->default_value(std::numeric_limits<unsigned>::max()), "max. # of pairs per Sent.")
+    ("pclr",              po::value<string>()->default_value("no"),         "use a (simple|adagrad) per-coordinate learning rate")
+    ("batch",             po::value<bool>()->zero_tokens(),                                               "do batch optimization")
+    ("repeat",            po::value<unsigned>()->default_value(1),          "repeat optimization over kbest list this number of times")
+    //("test-k-best",       po::value<bool>()->zero_tokens(),                       "check if optimization works (use repeat >= 2)")
     ("noup",              po::value<bool>()->zero_tokens(),                                               "do not update weights");
   po::options_description cl("Command Line Options");
   cl.add_options()
@@ -72,13 +77,17 @@ dtrain_init(int argc, char** argv, po::variables_map* cfg)
     cerr << "Wrong 'pair_sampling' param: '" << (*cfg)["pair_sampling"].as<string>() << "'." << endl;
     return false;
   }
-  if(cfg->count("hi_lo") && (*cfg)["pair_sampling"].as<string>() != "XYX") {
+  if (cfg->count("hi_lo") && (*cfg)["pair_sampling"].as<string>() != "XYX") {
     cerr << "Warning: hi_lo only works with pair_sampling XYX." << endl;
   }
-  if((*cfg)["hi_lo"].as<float>() > 0.5 || (*cfg)["hi_lo"].as<float>() < 0.01) {
+  if ((*cfg)["hi_lo"].as<float>() > 0.5 || (*cfg)["hi_lo"].as<float>() < 0.01) {
     cerr << "hi_lo must lie in [0.01, 0.5]" << endl;
     return false;
   }
+  if ((cfg->count("input")>0 || cfg->count("refs")>0) && cfg->count("bitext")>0) {
+    cerr << "Provide 'input' and 'refs' or 'bitext', not both." << endl;
+    return false;
+  }
   if ((*cfg)["pair_threshold"].as<score_t>() < 0) {
     cerr << "The threshold must be >= 0!" << endl;
     return false;
@@ -120,10 +129,16 @@ main(int argc, char** argv)
   const float hi_lo = cfg["hi_lo"].as<float>();
   const score_t approx_bleu_d = cfg["approx_bleu_d"].as<score_t>();
   const unsigned max_pairs = cfg["max_pairs"].as<unsigned>();
+  int repeat = cfg["repeat"].as<unsigned>();
+  //bool test_k_best = false;
+  //if (cfg.count("test-k-best")) test_k_best = true;
   weight_t loss_margin = cfg["loss_margin"].as<weight_t>();
+  bool batch = false;
+  if (cfg.count("batch")) batch = true;
   if (loss_margin > 9998.) loss_margin = std::numeric_limits<float>::max();
   bool scale_bleu_diff = false;
   if (cfg.count("scale_bleu_diff")) scale_bleu_diff = true;
+  const string pclr = cfg["pclr"].as<string>();
   bool average = false;
   if (select_weights == "avg")
     average = true;
@@ -131,7 +146,6 @@ main(int argc, char** argv)
   if (cfg.count("print_weights"))
     boost::split(print_weights, cfg["print_weights"].as<string>(), boost::is_any_of(" "));
 
-
   // setup decoder
   register_feature_functions();
   SetSilent(true);
@@ -178,17 +192,16 @@ main(int argc, char** argv)
   observer->SetScorer(scorer);
 
   // init weights
-  vector<weight_t>& dense_weights = decoder.CurrentWeightVector();
+  vector<weight_t>& decoder_weights = decoder.CurrentWeightVector();
   SparseVector<weight_t> lambdas, cumulative_penalties, w_average;
-  if (cfg.count("input_weights")) Weights::InitFromFile(cfg["input_weights"].as<string>(), &dense_weights);
-  Weights::InitSparseVector(dense_weights, &lambdas);
+  if (cfg.count("input_weights")) Weights::InitFromFile(cfg["input_weights"].as<string>(), &decoder_weights);
+  Weights::InitSparseVector(decoder_weights, &lambdas);
 
   // meta params for perceptron, SVM
   weight_t eta = cfg["learning_rate"].as<weight_t>();
   weight_t gamma = cfg["gamma"].as<weight_t>();
 
   // faster perceptron: consider only misranked pairs, see
-  // DO NOT ENABLE  WITH SVM (gamma > 0) OR loss_margin!
   bool faster_perceptron = false;
   if (gamma==0 && loss_margin==0) faster_perceptron = true;
 
@@ -208,13 +221,24 @@ main(int argc, char** argv)
   // output
   string output_fn = cfg["output"].as<string>();
   // input
-  string input_fn = cfg["input"].as<string>();
+  bool read_bitext = false;
+  string input_fn;
+  if (cfg.count("bitext")) {
+    read_bitext = true;
+    input_fn = cfg["bitext"].as<string>();
+  } else {
+    input_fn = cfg["input"].as<string>();
+  }
   ReadFile input(input_fn);
   // buffer input for t > 0
   vector<string> src_str_buf;          // source strings (decoder takes only strings)
   vector<vector<WordID> > ref_ids_buf; // references as WordID vecs
-  string refs_fn = cfg["refs"].as<string>();
-  ReadFile refs(refs_fn);
+  ReadFile refs;
+  string refs_fn;
+  if (!read_bitext) {
+    refs_fn = cfg["refs"].as<string>();
+    refs.Init(refs_fn);
+  }
 
   unsigned in_sz = std::numeric_limits<unsigned>::max(); // input index, input size
   vector<pair<score_t, score_t> > all_scores;
@@ -229,6 +253,7 @@ main(int argc, char** argv)
     cerr << setw(25) << "k " << k << endl;
     cerr << setw(25) << "N " << N << endl;
     cerr << setw(25) << "T " << T << endl;
+    cerr << setw(25) << "batch " << batch << endl;
     cerr << setw(26) << "scorer '" << scorer_str << "'" << endl;
     if (scorer_str == "approx_bleu")
       cerr << setw(25) << "approx. B discount " << approx_bleu_d << endl;
@@ -249,10 +274,14 @@ main(int argc, char** argv)
       cerr << setw(25) << "l1 reg " << l1_reg << " '" << cfg["l1_reg"].as<string>() << "'" << endl;
     if (rescale)
       cerr << setw(25) << "rescale " << rescale << endl;
+    cerr << setw(25) << "pclr " << pclr << endl;
     cerr << setw(25) << "max pairs " << max_pairs << endl;
+    cerr << setw(25) << "repeat " << repeat << endl;
+    //cerr << setw(25) << "test k-best " << test_k_best << endl;
     cerr << setw(25) << "cdec cfg " << "'" << cfg["decoder_config"].as<string>() << "'" << endl;
     cerr << setw(25) << "input " << "'" << input_fn << "'" << endl;
-    cerr << setw(25) << "refs " << "'" << refs_fn << "'" << endl;
+    if (!read_bitext)
+      cerr << setw(25) << "refs " << "'" << refs_fn << "'" << endl;
     cerr << setw(25) << "output " << "'" << output_fn << "'" << endl;
     if (cfg.count("input_weights"))
       cerr << setw(25) << "weights in " << "'" << cfg["input_weights"].as<string>() << "'" << endl;
@@ -261,6 +290,11 @@ main(int argc, char** argv)
     if (!verbose) cerr << "(a dot represents " << DTRAIN_DOTS << " inputs)" << endl;
   }
 
+  // pclr
+  SparseVector<weight_t> learning_rates;
+  // batch
+  SparseVector<weight_t> batch_updates;
+  score_t batch_loss;
 
   for (unsigned t = 0; t < T; t++) // T epochs
   {
@@ -269,16 +303,24 @@ main(int argc, char** argv)
   time(&start);
   score_t score_sum = 0.;
   score_t model_sum(0);
-  unsigned ii = 0, rank_errors = 0, margin_violations = 0, npairs = 0, f_count = 0, list_sz = 0;
+  unsigned ii = 0, rank_errors = 0, margin_violations = 0, npairs = 0, f_count = 0, list_sz = 0, kbest_loss_improve = 0;
+  batch_loss = 0.;
   if (!quiet) cerr << "Iteration #" << t+1 << " of " << T << "." << endl;
 
   while(true)
   {
 
     string in;
+    string ref;
     bool next = false, stop = false; // next iteration or premature stop
     if (t == 0) {
       if(!getline(*input, in)) next = true;
+      if(read_bitext) {
+        vector<string> strs;
+        boost::algorithm::split_regex(strs, in, boost::regex(" \\|\\|\\| "));
+        in = strs[0];
+        ref = strs[1];
+      }
     } else {
       if (ii == in_sz) next = true; // stop if we reach the end of our input
     }
@@ -310,15 +352,16 @@ main(int argc, char** argv)
     if (next || stop) break;
 
     // weights
-    lambdas.init_vector(&dense_weights);
+    lambdas.init_vector(&decoder_weights);
 
     // getting input
     vector<WordID> ref_ids; // reference as vector<WordID>
     if (t == 0) {
-      string r_;
-      getline(*refs, r_);
+      if (!read_bitext) {
+        getline(*refs, ref);
+      }
       vector<string> ref_tok;
-      boost::split(ref_tok, r_, boost::is_any_of(" "));
+      boost::split(ref_tok, ref, boost::is_any_of(" "));
       register_and_convert(ref_tok, ref_ids);
       ref_ids_buf.push_back(ref_ids);
       src_str_buf.push_back(in);
@@ -348,8 +391,10 @@ main(int argc, char** argv)
       }
     }
 
-    score_sum += (*samples)[0].score; // stats for 1best
-    model_sum += (*samples)[0].model;
+    if (repeat == 1) {
+      score_sum += (*samples)[0].score; // stats for 1best
+      model_sum += (*samples)[0].model;
+    }
 
     f_count += observer->get_f_count();
     list_sz += observer->get_sz();
@@ -364,30 +409,74 @@ main(int argc, char** argv)
         partXYX(samples, pairs, pair_threshold, max_pairs, faster_perceptron, hi_lo);
       if (pair_sampling == "PRO")
         PROsampling(samples, pairs, pair_threshold, max_pairs);
-      npairs += pairs.size();
+      int cur_npairs = pairs.size();
+      npairs += cur_npairs;
+
+      score_t kbest_loss_first, kbest_loss_last = 0.0;
 
-      SparseVector<weight_t> lambdas_copy;
+      for (vector<pair<ScoredHyp,ScoredHyp> >::iterator it = pairs.begin();
+           it != pairs.end(); it++) {
+        score_t model_diff = it->first.model - it->second.model;
+        kbest_loss_first += max(0.0, -1.0 * model_diff);
+      }
+
+      for (int ki=0; ki < repeat; ki++) {
+
+      score_t kbest_loss = 0.0; // test-k-best
+      SparseVector<weight_t> lambdas_copy; // for l1 regularization
+      SparseVector<weight_t> sum_up; // for pclr
       if (l1naive||l1clip||l1cumul) lambdas_copy = lambdas;
 
       for (vector<pair<ScoredHyp,ScoredHyp> >::iterator it = pairs.begin();
            it != pairs.end(); it++) {
-        bool rank_error;
+        score_t model_diff = it->first.model - it->second.model;
+        if (repeat > 1) {
+          model_diff = lambdas.dot(it->first.f) - lambdas.dot(it->second.f);
+          kbest_loss += max(0.0, -1.0 * model_diff);
+        }
+        bool rank_error = false;
         score_t margin;
         if (faster_perceptron) { // we only have considering misranked pairs
           rank_error = true; // pair sampling already did this for us
           margin = std::numeric_limits<float>::max();
         } else {
-          rank_error = it->first.model <= it->second.model;
-          margin = fabs(it->first.model - it->second.model);
+          rank_error = model_diff<=0.0;
+          margin = fabs(model_diff);
           if (!rank_error && margin < loss_margin) margin_violations++;
         }
-        if (rank_error) rank_errors++;
+        if (rank_error && ki==1) rank_errors++;
         if (scale_bleu_diff) eta = it->first.score - it->second.score;
         if (rank_error || margin < loss_margin) {
           SparseVector<weight_t> diff_vec = it->first.f - it->second.f;
-          lambdas.plus_eq_v_times_s(diff_vec, eta);
-          if (gamma)
-            lambdas.plus_eq_v_times_s(lambdas, -2*gamma*eta*(1./npairs));
+          if (batch) {
+            batch_loss += max(0., -1.0*model_diff);
+            batch_updates += diff_vec;
+            continue;
+          }
+          if (pclr != "no") {
+            sum_up += diff_vec;
+          } else {
+            lambdas.plus_eq_v_times_s(diff_vec, eta);
+            if (gamma) lambdas.plus_eq_v_times_s(lambdas, -2*gamma*eta*(1./cur_npairs));
+          }
+        }
+      }
+
+      // per-coordinate learning rate
+      if (pclr != "no") {
+        SparseVector<weight_t>::iterator it = sum_up.begin();
+        for (; it != sum_up.end(); ++it) {
+          if (pclr == "simple") {
+           lambdas[it->first] += it->second / max(1.0, learning_rates[it->first]);
+           learning_rates[it->first]++;
+          } else if (pclr == "adagrad") {
+            if (learning_rates[it->first] == 0) {
+             lambdas[it->first] +=  it->second * eta;
+            } else {
+             lambdas[it->first] +=  it->second * eta * learning_rates[it->first];
+            }
+            learning_rates[it->first] += pow(it->second, 2.0);
+          }
         }
       }
 
@@ -395,14 +484,16 @@ main(int argc, char** argv)
       // please note that this regularizations happen
       // after a _sentence_ -- not after each example/pair!
       if (l1naive) {
-        FastSparseVector<weight_t>::iterator it = lambdas.begin();
+        SparseVector<weight_t>::iterator it = lambdas.begin();
         for (; it != lambdas.end(); ++it) {
           if (!lambdas_copy.get(it->first) || lambdas_copy.get(it->first)!=it->second) {
+              it->second *= max(0.0000001, eta/(eta+learning_rates[it->first])); // FIXME
+              learning_rates[it->first]++;
             it->second -= sign(it->second) * l1_reg;
           }
         }
       } else if (l1clip) {
-        FastSparseVector<weight_t>::iterator it = lambdas.begin();
+        SparseVector<weight_t>::iterator it = lambdas.begin();
         for (; it != lambdas.end(); ++it) {
           if (!lambdas_copy.get(it->first) || lambdas_copy.get(it->first)!=it->second) {
             if (it->second != 0) {
@@ -417,7 +508,7 @@ main(int argc, char** argv)
         }
       } else if (l1cumul) {
         weight_t acc_penalty = (ii+1) * l1_reg; // ii is the index of the current input
-        FastSparseVector<weight_t>::iterator it = lambdas.begin();
+        SparseVector<weight_t>::iterator it = lambdas.begin();
         for (; it != lambdas.end(); ++it) {
           if (!lambdas_copy.get(it->first) || lambdas_copy.get(it->first)!=it->second) {
             if (it->second != 0) {
@@ -435,7 +526,28 @@ main(int argc, char** argv)
         }
       }
 
-    }
+      if (ki==repeat-1) { // done
+        kbest_loss_last = kbest_loss;
+        if (repeat > 1) {
+          score_t best_score = -1.;
+          score_t best_model = -std::numeric_limits<score_t>::max();
+          unsigned best_idx;
+          for (unsigned i=0; i < samples->size(); i++) {
+            score_t s = lambdas.dot((*samples)[i].f);
+            if (s > best_model) {
+              best_idx = i;
+              best_model = s;
+            }
+          }
+          score_sum += (*samples)[best_idx].score;
+          model_sum += best_model;
+        }
+      }
+    } // repeat
+
+    if ((kbest_loss_first - kbest_loss_last) >= 0) kbest_loss_improve++;
+
+    } // noup
 
     if (rescale) lambdas /= lambdas.l2norm();
 
@@ -443,14 +555,19 @@ main(int argc, char** argv)
 
   } // input loop
 
-  if (average) w_average += lambdas;
+  if (t == 0) in_sz = ii; // remember size of input (# lines)
 
-  if (scorer_str == "approx_bleu" || scorer_str == "lc_bleu") scorer->Reset();
 
-  if (t == 0) {
-    in_sz = ii; // remember size of input (# lines)
+  if (batch) {
+    lambdas.plus_eq_v_times_s(batch_updates, eta);
+    if (gamma) lambdas.plus_eq_v_times_s(lambdas, -2*gamma*eta*(1./npairs));
+    batch_updates.clear();
   }
 
+  if (average) w_average += lambdas;
+
+  if (scorer_str == "approx_bleu" || scorer_str == "lc_bleu") scorer->Reset();
+
   // print some stats
   score_t score_avg = score_sum/(score_t)in_sz;
   score_t model_avg = model_sum/(score_t)in_sz;
@@ -477,13 +594,15 @@ main(int argc, char** argv)
     cerr << _np << " 1best avg model score: " << model_avg;
     cerr << _p << " (" << model_diff << ")" << endl;
     cerr << "           avg # pairs: ";
-    cerr << _np << npairs/(float)in_sz;
+    cerr << _np << npairs/(float)in_sz << endl;
+    cerr << "        avg # rank err: ";
+    cerr << rank_errors/(float)in_sz;
     if (faster_perceptron) cerr << " (meaningless)";
     cerr << endl;
-    cerr << "        avg # rank err: ";
-    cerr << rank_errors/(float)in_sz << endl;
     cerr << "     avg # margin viol: ";
     cerr << margin_violations/(float)in_sz << endl;
+    if (batch) cerr << "            batch loss: " << batch_loss << endl;
+    cerr << "       k-best loss imp: " << ((float)kbest_loss_improve/in_sz)*100 << "%" << endl;
     cerr << "    non0 feature count: " <<  nonz << endl;
     cerr << "           avg list sz: " << list_sz/(float)in_sz << endl;
     cerr << "           avg f count: " << f_count/(float)list_sz << endl;
@@ -510,9 +629,9 @@ main(int argc, char** argv)
 
   // write weights to file
   if (select_weights == "best" || keep) {
-    lambdas.init_vector(&dense_weights);
+    lambdas.init_vector(&decoder_weights);
     string w_fn = "weights." + boost::lexical_cast<string>(t) + ".gz";
-    Weights::WriteToFile(w_fn, dense_weights, true);
+    Weights::WriteToFile(w_fn, decoder_weights, true);
   }
 
   } // outer loop
diff --git a/training/dtrain/dtrain.h b/training/dtrain/dtrain.h
index 3981fb39..ccb5ad4d 100644
--- a/training/dtrain/dtrain.h
+++ b/training/dtrain/dtrain.h
@@ -9,6 +9,8 @@
 #include <string.h>
 
 #include <boost/algorithm/string.hpp>
+#include <boost/regex.hpp>
+#include <boost/algorithm/string/regex.hpp>
 #include <boost/program_options.hpp>
 
 #include "decoder.h"
diff --git a/training/dtrain/examples/standard/dtrain.ini b/training/dtrain/examples/standard/dtrain.ini
index 23e94285..fc83f08e 100644
--- a/training/dtrain/examples/standard/dtrain.ini
+++ b/training/dtrain/examples/standard/dtrain.ini
@@ -1,5 +1,6 @@
-input=./nc-wmt11.de.gz
-refs=./nc-wmt11.en.gz
+#input=./nc-wmt11.de.gz
+#refs=./nc-wmt11.en.gz
+bitext=./nc-wmt11.gz
 output=-                  # a weights file (add .gz for gzip compression) or STDOUT '-'
 select_weights=VOID       # output average (over epochs) weight vector
 decoder_config=./cdec.ini # config for cdec
@@ -10,11 +11,11 @@ print_weights=Glue WordPenalty LanguageModel LanguageModel_OOV PhraseModel_0 Phr
 stop_after=10 # stop epoch after 10 inputs
 
 # interesting stuff
-epochs=2                 # run over input 2 times
+epochs=3                 # run over input 3 times
 k=100                    # use 100best lists
 N=4                      # optimize (approx) BLEU4
 scorer=fixed_stupid_bleu # use 'stupid' BLEU+1
-learning_rate=1.0        # learning rate, don't care if gamma=0 (perceptron)
+learning_rate=0.1        # learning rate, don't care if gamma=0 (perceptron) and loss_margin=0 (not margin perceptron)
 gamma=0                  # use SVM reg
 sample_from=kbest        # use kbest lists (as opposed to forest)
 filter=uniq              # only unique entries in kbest (surface form)
@@ -22,3 +23,5 @@ pair_sampling=XYX        #
 hi_lo=0.1                # 10 vs 80 vs 10 and 80 vs 10 here
 pair_threshold=0         # minimum distance in BLEU (here: > 0)
 loss_margin=0            # update if correctly ranked, but within this margin
+repeat=1                 # repeat training on a kbest list 1 times 
+#batch=true              # batch tuning, update after accumulating over all sentences and all kbest lists
diff --git a/training/dtrain/examples/standard/expected-output b/training/dtrain/examples/standard/expected-output
index 21f91244..75f47337 100644
--- a/training/dtrain/examples/standard/expected-output
+++ b/training/dtrain/examples/standard/expected-output
@@ -4,17 +4,18 @@ Reading ./nc-wmt11.en.srilm.gz
 ----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
 ****************************************************************************************************
   Example feature: Shape_S00000_T00000
-Seeding random number sequence to 970626287
+Seeding random number sequence to 3751911392
 
 dtrain
 Parameters:
                        k 100
                        N 4
-                       T 2
+                       T 3
+                   batch 0
                   scorer 'fixed_stupid_bleu'
              sample from 'kbest'
                   filter 'uniq'
-           learning rate 1
+           learning rate 0.1
                    gamma 0
              loss margin 0
        faster perceptron 1
@@ -23,69 +24,99 @@ Parameters:
           pair threshold 0
           select weights 'VOID'
                   l1 reg 0 'none'
+                    pclr no
                max pairs 4294967295
+                  repeat 1
                 cdec cfg './cdec.ini'
-                   input './nc-wmt11.de.gz'
-                    refs './nc-wmt11.en.gz'
+                   input './nc-wmt11.gz'
                   output '-'
               stop_after 10
 (a dot represents 10 inputs)
-Iteration #1 of 2.
+Iteration #1 of 3.
  . 10
 Stopping after 10 input sentences.
 WEIGHTS
-              Glue = -614
-       WordPenalty = +1256.8
-     LanguageModel = +5610.5
- LanguageModel_OOV = -1449
-     PhraseModel_0 = -2107
-     PhraseModel_1 = -4666.1
-     PhraseModel_2 = -2713.5
-     PhraseModel_3 = +4204.3
-     PhraseModel_4 = -1435.8
-     PhraseModel_5 = +916
-     PhraseModel_6 = +190
-       PassThrough = -2527
+              Glue = -110
+       WordPenalty = -8.2082
+     LanguageModel = -319.91
+ LanguageModel_OOV = -19.2
+     PhraseModel_0 = +312.82
+     PhraseModel_1 = -161.02
+     PhraseModel_2 = -433.65
+     PhraseModel_3 = +291.03
+     PhraseModel_4 = +252.32
+     PhraseModel_5 = +50.6
+     PhraseModel_6 = +146.7
+       PassThrough = -38.7
         ---
-       1best avg score: 0.17874 (+0.17874)
- 1best avg model score: 88399 (+88399)
-           avg # pairs: 798.2 (meaningless)
-        avg # rank err: 798.2
+       1best avg score: 0.16966 (+0.16966)
+ 1best avg model score: 29874 (+29874)
+           avg # pairs: 906.3
+        avg # rank err: 0 (meaningless)
      avg # margin viol: 0
-    non0 feature count: 887
+       k-best loss imp: 100%
+    non0 feature count: 832
            avg list sz: 91.3
-           avg f count: 126.85
-(time 0.33 min, 2 s/S)
+           avg f count: 139.77
+(time 0.35 min, 2.1 s/S)
 
-Iteration #2 of 2.
+Iteration #2 of 3.
  . 10
 WEIGHTS
-              Glue = -1025
-       WordPenalty = +1751.5
-     LanguageModel = +10059
- LanguageModel_OOV = -4490
-     PhraseModel_0 = -2640.7
-     PhraseModel_1 = -3757.4
-     PhraseModel_2 = -1133.1
-     PhraseModel_3 = +1837.3
-     PhraseModel_4 = -3534.3
-     PhraseModel_5 = +2308
-     PhraseModel_6 = +1677
-       PassThrough = -6222
+              Glue = -122.1
+       WordPenalty = +83.689
+     LanguageModel = +233.23
+ LanguageModel_OOV = -145.1
+     PhraseModel_0 = +150.72
+     PhraseModel_1 = -272.84
+     PhraseModel_2 = -418.36
+     PhraseModel_3 = +181.63
+     PhraseModel_4 = -289.47
+     PhraseModel_5 = +140.3
+     PhraseModel_6 = +3.5
+       PassThrough = -109.7
         ---
-       1best avg score: 0.30764 (+0.12891)
- 1best avg model score: -2.5042e+05 (-3.3882e+05)
-           avg # pairs: 725.9 (meaningless)
-        avg # rank err: 725.9
+       1best avg score: 0.17399 (+0.004325)
+ 1best avg model score: 4936.9 (-24937)
+           avg # pairs: 662.4
+        avg # rank err: 0 (meaningless)
      avg # margin viol: 0
-    non0 feature count: 1499
+       k-best loss imp: 100%
+    non0 feature count: 1240
            avg list sz: 91.3
-           avg f count: 114.34
-(time 0.32 min, 1.9 s/S)
+           avg f count: 125.11
+(time 0.27 min, 1.6 s/S)
+
+Iteration #3 of 3.
+ . 10
+WEIGHTS
+              Glue = -157.4
+       WordPenalty = -1.7372
+     LanguageModel = +686.18
+ LanguageModel_OOV = -399.7
+     PhraseModel_0 = -39.876
+     PhraseModel_1 = -341.96
+     PhraseModel_2 = -318.67
+     PhraseModel_3 = +105.08
+     PhraseModel_4 = -290.27
+     PhraseModel_5 = -48.6
+     PhraseModel_6 = -43.6
+       PassThrough = -298.5
+        ---
+       1best avg score: 0.30742 (+0.13343)
+ 1best avg model score: -15393 (-20329)
+           avg # pairs: 623.8
+        avg # rank err: 0 (meaningless)
+     avg # margin viol: 0
+       k-best loss imp: 100%
+    non0 feature count: 1776
+           avg list sz: 91.3
+           avg f count: 118.58
+(time 0.28 min, 1.7 s/S)
 
 Writing weights file to '-' ...
 done
 
 ---
-Best iteration: 2 [SCORE 'fixed_stupid_bleu'=0.30764].
-This took 0.65 min.
+Best iteration: 3 [SCORE 'fixed_stupid_bleu'=0.30742].
+This took 0.9 min.
diff --git a/training/dtrain/examples/standard/nc-wmt11.gz b/training/dtrain/examples/standard/nc-wmt11.gz
new file mode 100644
index 00000000..c39c5aef
--- /dev/null
+++ b/training/dtrain/examples/standard/nc-wmt11.gz
diff --git a/training/dtrain/parallelize.rb b/training/dtrain/parallelize.rb
index 285f3c9b..60ca9422 100755
--- a/training/dtrain/parallelize.rb
+++ b/training/dtrain/parallelize.rb
@@ -21,6 +21,8 @@ opts = Trollop::options do
   opt :qsub, "use qsub", :type => :bool, :default => false
   opt :dtrain_binary, "path to dtrain binary", :type => :string
   opt :extra_qsub, "extra qsub args", :type => :string, :default => ""
+  opt :per_shard_decoder_configs, "give special decoder config per shard", :type => :string, :short => '-o'
+  opt :first_input_weights, "input weights for first iter", :type => :string, :default => '', :short => '-w'
 end
 usage if not opts[:config]&&opts[:shards]&&opts[:input]&&opts[:references]
 
@@ -41,9 +43,11 @@ epochs     = opts[:epochs]
 rand       = opts[:randomize]
 reshard    = opts[:reshard]
 predefined_shards = false
+per_shard_decoder_configs = false
 if opts[:shards] == 0
   predefined_shards = true
   num_shards = 0
+  per_shard_decoder_configs = true if opts[:per_shard_decoder_configs]
 else
   num_shards = opts[:shards]
 end
@@ -51,6 +55,7 @@ input = opts[:input]
 refs  = opts[:references]
 use_qsub       = opts[:qsub]
 shards_at_once = opts[:processes_at_once]
+first_input_weights  = opts[:first_input_weights]
 
 `mkdir work`
 
@@ -101,6 +106,9 @@ refs_files = []
 if predefined_shards
   input_files = File.new(input).readlines.map {|i| i.strip }
   refs_files = File.new(refs).readlines.map {|i| i.strip }
+  if per_shard_decoder_configs
+    decoder_configs = File.new(opts[:per_shard_decoder_configs]).readlines.map {|i| i.strip}
+  end
   num_shards = input_files.size
 else
   input_files, refs_files = make_shards input, refs, num_shards, 0, rand
@@ -126,10 +134,18 @@ end
       else
         local_end = "2>work/out.#{shard}.#{epoch}"
       end
+      if per_shard_decoder_configs
+        cdec_cfg = "--decoder_config #{decoder_configs[shard]}"
+      else
+        cdec_cfg = ""
+      end
+      if first_input_weights!='' && epoch == 0
+        input_weights = "--input_weights #{first_input_weights}"
+      end
       pids << Kernel.fork {
-        `#{qsub_str_start}#{dtrain_bin} -c #{ini}\
+        `#{qsub_str_start}#{dtrain_bin} -c #{ini} #{cdec_cfg} #{input_weights}\
           --input #{input_files[shard]}\
-          --refs #{refs_files[shard]} #{input_weights}\
+          --refs #{refs_files[shard]}\
           --output work/weights.#{shard}.#{epoch}#{qsub_str_end} #{local_end}`
       }
       weights_files << "work/weights.#{shard}.#{epoch}"
diff --git a/training/latent_svm/latent_svm.cc b/training/latent_svm/latent_svm.cc
index ab9c1d5d..60e52550 100644
--- a/training/latent_svm/latent_svm.cc
+++ b/training/latent_svm/latent_svm.cc
@@ -32,7 +32,6 @@ total_loss and prev_loss actually refer not to loss, but the metric (usually BLE
 #include "sampler.h"
 
 using namespace std;
-using boost::shared_ptr;
 namespace po = boost::program_options;
 
 bool invert_score; 
@@ -128,7 +127,7 @@ struct HypothesisInfo {
 };
 
 struct GoodOracle {
-  shared_ptr<HypothesisInfo> good;
+  boost::shared_ptr<HypothesisInfo> good;
 };
 
 struct TrainingObserver : public DecoderObserver {
@@ -143,9 +142,9 @@ struct TrainingObserver : public DecoderObserver {
   const DocScorer& ds;
   const vector<weight_t>& feature_weights;
   vector<GoodOracle>& oracles;
-  shared_ptr<HypothesisInfo> cur_best;
-  shared_ptr<HypothesisInfo> cur_costaug_best;
-  shared_ptr<HypothesisInfo> cur_ref; 
+  boost::shared_ptr<HypothesisInfo> cur_best;
+  boost::shared_ptr<HypothesisInfo> cur_costaug_best;
+  boost::shared_ptr<HypothesisInfo> cur_ref; 
   const int kbest_size;
   const double mt_metric_scale;
   const double mu;
@@ -168,8 +167,8 @@ struct TrainingObserver : public DecoderObserver {
     UpdateOracles(smeta.GetSentenceID(), *hg);
   }
 
-  shared_ptr<HypothesisInfo> MakeHypothesisInfo(const SparseVector<double>& feats, const double metric) {
-    shared_ptr<HypothesisInfo> h(new HypothesisInfo);
+  boost::shared_ptr<HypothesisInfo> MakeHypothesisInfo(const SparseVector<double>& feats, const double metric) {
+    boost::shared_ptr<HypothesisInfo> h(new HypothesisInfo);
     h->features = feats;
     h->mt_metric_score = metric;
     return h;
diff --git a/training/mira/kbest_cut_mira.cc b/training/mira/kbest_cut_mira.cc
index 59fa860a..990609d7 100644
--- a/training/mira/kbest_cut_mira.cc
+++ b/training/mira/kbest_cut_mira.cc
@@ -30,7 +30,6 @@
 #include "sparse_vector.h"
 
 using namespace std;
-using boost::shared_ptr;
 namespace po = boost::program_options;
 
 bool invert_score;
@@ -50,13 +49,6 @@ bool sent_approx;
 bool checkloss;
 bool stream;
 
-void SanityCheck(const vector<double>& w) {
-  for (int i = 0; i < w.size(); ++i) {
-    assert(!isnan(w[i]));
-    assert(!isinf(w[i]));
-  }
-}
-
 struct FComp {
   const vector<double>& w_;
   FComp(const vector<double>& w) : w_(w) {}
@@ -149,7 +141,7 @@ struct HypothesisInfo {
   double alpha;
   double oracle_loss;
   SparseVector<double> oracle_feat_diff;
-  shared_ptr<HypothesisInfo> oracleN;
+  boost::shared_ptr<HypothesisInfo> oracleN;
 };
 
 bool ApproxEqual(double a, double b) {
@@ -157,7 +149,7 @@ bool ApproxEqual(double a, double b) {
   return (fabs(a-b)/fabs(b)) < EPSILON;
 }
 
-typedef shared_ptr<HypothesisInfo> HI;
+typedef boost::shared_ptr<HypothesisInfo> HI;
 bool HypothesisCompareB(const HI& h1, const HI& h2 ) 
 {
   return h1->mt_metric > h2->mt_metric;
@@ -185,11 +177,11 @@ bool HypothesisCompareG(const HI& h1, const HI& h2 )
 };
 
 
-void CuttingPlane(vector<shared_ptr<HypothesisInfo> >* cur_c, bool* again, vector<shared_ptr<HypothesisInfo> >& all_hyp, vector<weight_t> dense_weights)
+void CuttingPlane(vector<boost::shared_ptr<HypothesisInfo> >* cur_c, bool* again, vector<boost::shared_ptr<HypothesisInfo> >& all_hyp, vector<weight_t> dense_weights)
 {
   bool DEBUG_CUT = false;
-  shared_ptr<HypothesisInfo> max_fear, max_fear_in_set;
-  vector<shared_ptr<HypothesisInfo> >& cur_constraint = *cur_c;
+  boost::shared_ptr<HypothesisInfo> max_fear, max_fear_in_set;
+  vector<boost::shared_ptr<HypothesisInfo> >& cur_constraint = *cur_c;
 
   if(no_reweight)
     {
@@ -235,9 +227,9 @@ void CuttingPlane(vector<shared_ptr<HypothesisInfo> >* cur_c, bool* again, vecto
 }
 
 
-double ComputeDelta(vector<shared_ptr<HypothesisInfo> >* cur_p, double max_step_size,vector<weight_t> dense_weights )
+double ComputeDelta(vector<boost::shared_ptr<HypothesisInfo> >* cur_p, double max_step_size,vector<weight_t> dense_weights )
 {
-  vector<shared_ptr<HypothesisInfo> >& cur_pair = *cur_p;
+  vector<boost::shared_ptr<HypothesisInfo> >& cur_pair = *cur_p;
    double loss = cur_pair[0]->oracle_loss - cur_pair[1]->oracle_loss;
 
    double margin = -(cur_pair[0]->oracleN->features.dot(dense_weights)- cur_pair[0]->features.dot(dense_weights)) + (cur_pair[1]->oracleN->features.dot(dense_weights) - cur_pair[1]->features.dot(dense_weights));
@@ -261,12 +253,12 @@ double ComputeDelta(vector<shared_ptr<HypothesisInfo> >* cur_p, double max_step_
 }
 
 
-vector<shared_ptr<HypothesisInfo> > SelectPair(vector<shared_ptr<HypothesisInfo> >* cur_c)
+vector<boost::shared_ptr<HypothesisInfo> > SelectPair(vector<boost::shared_ptr<HypothesisInfo> >* cur_c)
 {
   bool DEBUG_SELECT= false;
-  vector<shared_ptr<HypothesisInfo> >& cur_constraint = *cur_c;
+  vector<boost::shared_ptr<HypothesisInfo> >& cur_constraint = *cur_c;
   
-  vector<shared_ptr<HypothesisInfo> > pair;
+  vector<boost::shared_ptr<HypothesisInfo> > pair;
 
   if (no_select || optimizer == 2){ //skip heuristic search and return oracle and fear for pa-mira
 
@@ -278,7 +270,7 @@ vector<shared_ptr<HypothesisInfo> > SelectPair(vector<shared_ptr<HypothesisInfo>
   
   for(int u=0;u != cur_constraint.size();u++)	
     {
-      shared_ptr<HypothesisInfo> max_fear;
+      boost::shared_ptr<HypothesisInfo> max_fear;
       
       if(DEBUG_SELECT) cerr<< "cur alpha " << u  << " " << cur_constraint[u]->alpha;
       for(int i=0; i < cur_constraint.size();i++) //select maximal violator
@@ -323,8 +315,8 @@ vector<shared_ptr<HypothesisInfo> > SelectPair(vector<shared_ptr<HypothesisInfo>
 }
 
 struct GoodBadOracle {
-  vector<shared_ptr<HypothesisInfo> > good;
-  vector<shared_ptr<HypothesisInfo> > bad;
+  vector<boost::shared_ptr<HypothesisInfo> > good;
+  vector<boost::shared_ptr<HypothesisInfo> > bad;
 };
 
 struct BasicObserver: public DecoderObserver {
@@ -367,8 +359,8 @@ struct TrainingObserver : public DecoderObserver {
   const DocScorer& ds;
   vector<ScoreP>& corpus_bleu_sent_stats;
   vector<GoodBadOracle>& oracles;
-  vector<shared_ptr<HypothesisInfo> > cur_best;
-  shared_ptr<HypothesisInfo> cur_oracle;
+  vector<boost::shared_ptr<HypothesisInfo> > cur_best;
+  boost::shared_ptr<HypothesisInfo> cur_oracle;
   const int kbest_size;
   Hypergraph forest;
   int cur_sent;
@@ -386,7 +378,7 @@ struct TrainingObserver : public DecoderObserver {
     return *cur_best[0];
   }
 
-  const vector<shared_ptr<HypothesisInfo> > GetCurrentBest() const {
+  const vector<boost::shared_ptr<HypothesisInfo> > GetCurrentBest() const {
     return cur_best;
   }
   
@@ -411,8 +403,8 @@ struct TrainingObserver : public DecoderObserver {
     
   }
 
-  shared_ptr<HypothesisInfo> MakeHypothesisInfo(const SparseVector<double>& feats, const double score, const vector<WordID>& hyp) {
-    shared_ptr<HypothesisInfo> h(new HypothesisInfo);
+  boost::shared_ptr<HypothesisInfo> MakeHypothesisInfo(const SparseVector<double>& feats, const double score, const vector<WordID>& hyp) {
+    boost::shared_ptr<HypothesisInfo> h(new HypothesisInfo);
     h->features = feats;
     h->mt_metric = score;
     h->hyp = hyp;
@@ -424,14 +416,14 @@ struct TrainingObserver : public DecoderObserver {
 
 	if (stream) sent_id = 0;
     bool PRINT_LIST= false;
-    vector<shared_ptr<HypothesisInfo> >& cur_good = oracles[sent_id].good;
-    vector<shared_ptr<HypothesisInfo> >& cur_bad = oracles[sent_id].bad;
+    vector<boost::shared_ptr<HypothesisInfo> >& cur_good = oracles[sent_id].good;
+    vector<boost::shared_ptr<HypothesisInfo> >& cur_bad = oracles[sent_id].bad;
     //TODO: look at keeping previous iterations hypothesis lists around
     cur_best.clear();
     cur_good.clear();
     cur_bad.clear();
 
-    vector<shared_ptr<HypothesisInfo> > all_hyp;
+    vector<boost::shared_ptr<HypothesisInfo> > all_hyp;
 
     typedef KBest::KBestDerivations<vector<WordID>, ESentenceTraversal,Filter> K;
     K kbest(forest,kbest_size);
@@ -527,7 +519,7 @@ struct TrainingObserver : public DecoderObserver {
     if(PRINT_LIST) { cerr << "GOOD" << endl;  for(int u=0;u!=cur_good.size();u++) cerr << cur_good[u]->mt_metric << " " << cur_good[u]->hope << endl;}     
 
     //use hope for fear selection
-    shared_ptr<HypothesisInfo>& oracleN = cur_good[0];
+    boost::shared_ptr<HypothesisInfo>& oracleN = cur_good[0];
 
     if(fear_select == 1){   //compute fear hyps with model - bleu
       if (PRINT_LIST) cerr << "FEAR " << endl;
@@ -663,13 +655,13 @@ int main(int argc, char** argv) {
     invert_score = false;
   }
 
-  shared_ptr<DocScorer> ds;
+  boost::shared_ptr<DocScorer> ds;
   //normal: load references, stream: start stream scorer
   if (stream) {
-	  ds = shared_ptr<DocScorer>(new DocStreamScorer(type, vector<string>(0), ""));
+	  ds = boost::shared_ptr<DocScorer>(new DocStreamScorer(type, vector<string>(0), ""));
 	  cerr << "Scoring doc stream with " << metric_name << endl;
   } else {
-      ds = shared_ptr<DocScorer>(new DocScorer(type, conf["reference"].as<vector<string> >(), ""));
+      ds = boost::shared_ptr<DocScorer>(new DocScorer(type, conf["reference"].as<vector<string> >(), ""));
       cerr << "Loaded " << ds->size() << " references for scoring with " << metric_name << endl;
   }
   vector<ScoreP> corpus_bleu_sent_stats;
@@ -774,9 +766,9 @@ int main(int argc, char** argv) {
       const HypothesisInfo& cur_good = *oracles[cur_sent].good[0];
       const HypothesisInfo& cur_bad = *oracles[cur_sent].bad[0];
 
-      vector<shared_ptr<HypothesisInfo> >& cur_good_v = oracles[cur_sent].good;
-      vector<shared_ptr<HypothesisInfo> >& cur_bad_v = oracles[cur_sent].bad;
-      vector<shared_ptr<HypothesisInfo> > cur_best_v = observer.GetCurrentBest();
+      vector<boost::shared_ptr<HypothesisInfo> >& cur_good_v = oracles[cur_sent].good;
+      vector<boost::shared_ptr<HypothesisInfo> >& cur_bad_v = oracles[cur_sent].bad;
+      vector<boost::shared_ptr<HypothesisInfo> > cur_best_v = observer.GetCurrentBest();
 
       tot_loss += cur_hyp.mt_metric;
       
@@ -824,13 +816,13 @@ int main(int argc, char** argv) {
 	}
       else if(optimizer == 5) //full mira with n-best list of constraints from hope, fear, model best
 	{
-	  vector<shared_ptr<HypothesisInfo> > cur_constraint;
+	  vector<boost::shared_ptr<HypothesisInfo> > cur_constraint;
 	  cur_constraint.insert(cur_constraint.begin(), cur_bad_v.begin(), cur_bad_v.end());
 	  cur_constraint.insert(cur_constraint.begin(), cur_best_v.begin(), cur_best_v.end());
 	  cur_constraint.insert(cur_constraint.begin(), cur_good_v.begin(), cur_good_v.end());
 
 	  bool optimize_again;
-	  vector<shared_ptr<HypothesisInfo> > cur_pair;
+	  vector<boost::shared_ptr<HypothesisInfo> > cur_pair;
 	  //SMO 
 	  for(int u=0;u!=cur_constraint.size();u++)	
 	    cur_constraint[u]->alpha =0;	      
@@ -879,7 +871,7 @@ int main(int argc, char** argv) {
       else if(optimizer == 2 || optimizer == 3) //PA and Cutting Plane MIRA update
 	  {
 	    bool DEBUG_SMO= true;
-	    vector<shared_ptr<HypothesisInfo> > cur_constraint;
+	    vector<boost::shared_ptr<HypothesisInfo> > cur_constraint;
 	    cur_constraint.push_back(cur_good_v[0]); //add oracle to constraint set
 	    bool optimize_again = true;
 	    int cut_plane_calls = 0;
@@ -919,7 +911,7 @@ int main(int argc, char** argv) {
 		    while (iter < smo_iter)
 		      {			
 			//select pair to optimize from constraint set
-			vector<shared_ptr<HypothesisInfo> > cur_pair = SelectPair(&cur_constraint);
+			vector<boost::shared_ptr<HypothesisInfo> > cur_pair = SelectPair(&cur_constraint);
 			
 			if(cur_pair.empty()){
 			  iter=MAX_SMO; 
diff --git a/training/mira/kbest_mira.cc b/training/mira/kbest_mira.cc
index d59b4224..2868de0c 100644
--- a/training/mira/kbest_mira.cc
+++ b/training/mira/kbest_mira.cc
@@ -3,10 +3,10 @@
 #include <vector>
 #include <cassert>
 #include <cmath>
-#include <tr1/memory>
 
 #include <boost/program_options.hpp>
 #include <boost/program_options/variables_map.hpp>
+#include <boost/shared_ptr.hpp>
 
 #include "stringlib.h"
 #include "hg_sampler.h"
@@ -30,7 +30,7 @@ using namespace std;
 namespace po = boost::program_options;
 
 bool invert_score;
-std::tr1::shared_ptr<MT19937> rng;
+boost::shared_ptr<MT19937> rng;
 
 void RandomPermutation(int len, vector<int>* p_ids) {
   vector<int>& ids = *p_ids;
@@ -88,8 +88,8 @@ struct HypothesisInfo {
 };
 
 struct GoodBadOracle {
-  std::tr1::shared_ptr<HypothesisInfo> good;
-  std::tr1::shared_ptr<HypothesisInfo> bad;
+  boost::shared_ptr<HypothesisInfo> good;
+  boost::shared_ptr<HypothesisInfo> bad;
 };
 
 struct TrainingObserver : public DecoderObserver {
@@ -97,7 +97,7 @@ struct TrainingObserver : public DecoderObserver {
   const DocumentScorer& ds;
   const EvaluationMetric& metric;
   vector<GoodBadOracle>& oracles;
-  std::tr1::shared_ptr<HypothesisInfo> cur_best;
+  boost::shared_ptr<HypothesisInfo> cur_best;
   const int kbest_size;
   const bool sample_forest;
 
@@ -109,16 +109,16 @@ struct TrainingObserver : public DecoderObserver {
     UpdateOracles(smeta.GetSentenceID(), *hg);
   }
 
-  std::tr1::shared_ptr<HypothesisInfo> MakeHypothesisInfo(const SparseVector<double>& feats, const double score) {
-    std::tr1::shared_ptr<HypothesisInfo> h(new HypothesisInfo);
+  boost::shared_ptr<HypothesisInfo> MakeHypothesisInfo(const SparseVector<double>& feats, const double score) {
+    boost::shared_ptr<HypothesisInfo> h(new HypothesisInfo);
     h->features = feats;
     h->mt_metric = score;
     return h;
   }
 
   void UpdateOracles(int sent_id, const Hypergraph& forest) {
-    std::tr1::shared_ptr<HypothesisInfo>& cur_good = oracles[sent_id].good;
-    std::tr1::shared_ptr<HypothesisInfo>& cur_bad = oracles[sent_id].bad;
+    boost::shared_ptr<HypothesisInfo>& cur_good = oracles[sent_id].good;
+    boost::shared_ptr<HypothesisInfo>& cur_bad = oracles[sent_id].bad;
     cur_bad.reset();  // TODO get rid of??
 
     if (sample_forest) {
diff --git a/training/mira/mira.py b/training/mira/mira.py
index 29c51e1d..d5a1d9f8 100755
--- a/training/mira/mira.py
+++ b/training/mira/mira.py
@@ -4,8 +4,19 @@ import subprocess, shlex, glob
 import argparse
 import logging
 import random, time
-import cdec.score
 import gzip, itertools
+try:
+  import cdec.score
+except ImportError:
+  sys.stderr.write('Could not import pycdec, see cdec/python/README.md for details\n')
+  sys.exit(1)
+have_mpl = True
+try: 
+  import matplotlib
+  matplotlib.use('Agg')
+  import matplotlib.pyplot as plt
+except ImportError:
+  have_mpl = False
 
 #mira run script
 #requires pycdec to be built, since it is used for scoring hypothesis
@@ -16,17 +27,17 @@ import gzip, itertools
 #scoring function using pycdec scoring
 def fast_score(hyps, refs, metric):
   scorer = cdec.score.Scorer(metric)
-  logging.info('loaded {0} references for scoring with {1}\n'.format(
+  logging.info('loaded {0} references for scoring with {1}'.format(
                 len(refs), metric))
   if metric=='BLEU':
     logging.warning('BLEU is ambiguous, assuming IBM_BLEU\n')
     metric = 'IBM_BLEU'
   elif metric=='COMBI':
     logging.warning('COMBI metric is no longer supported, switching to '
-                    'COMB:TER=-0.5;BLEU=0.5\n')
+                    'COMB:TER=-0.5;BLEU=0.5')
     metric = 'COMB:TER=-0.5;BLEU=0.5'
   stats = sum(scorer(r).evaluate(h) for h,r in itertools.izip(hyps,refs))
-  logging.info(stats.detail+'\n')
+  logging.info('Score={} ({})'.format(stats.score, stats.detail))
   return stats.score
 
 #create new parallel input file in output directory in sgml format
@@ -71,6 +82,8 @@ def main():
   #set logging to write all info messages to stderr
   logging.basicConfig(level=logging.INFO)
   script_dir = os.path.dirname(os.path.abspath(sys.argv[0]))
+  if not have_mpl:
+    logging.warning('Failed to import matplotlib, graphs will not be generated.')
 
   parser= argparse.ArgumentParser(
             formatter_class=argparse.ArgumentDefaultsHelpFormatter)
@@ -181,10 +194,11 @@ def main():
   dev_size = enseg(args.devset, newdev, args.grammar_prefix)
   args.devset = newdev
   
-  write_config(args)
+  log_config(args)
   args.weights, hope_best_fear = optimize(args, script_dir, dev_size)
   
-  graph_file = graph(args.output_dir, hope_best_fear, args.metric)
+  graph_file = ''
+  if have_mpl: graph_file = graph(args.output_dir, hope_best_fear, args.metric)
 
   dev_results, dev_bleu = evaluate(args.devset, args.weights, args.config, 
                          script_dir, args.output_dir)
@@ -205,17 +219,12 @@ def main():
 
   if graph_file:
     logging.info('A graph of the best/hope/fear scores over the iterations '
-                 'has been saved to {}\n'.format(graph_file))
+                 'has been saved to {}'.format(graph_file))
 
   print 'final weights:\n{}\n'.format(args.weights)
 
 #graph of hope/best/fear metric values across all iterations
 def graph(output_dir, hope_best_fear, metric):
-  try: 
-    import matplotlib.pyplot as plt
-  except ImportError:
-    logging.error('Error importing matplotlib. Graphing disabled.\n')
-    return ''
   max_y = float(max(hope_best_fear['best']))*1.5
   plt.plot(hope_best_fear['best'], label='best')
   plt.plot(hope_best_fear['hope'], label='hope')
@@ -308,6 +317,7 @@ def optimize(args, script_dir, dev_size):
   decoder = script_dir+'/kbest_cut_mira'
   (source, refs) = split_devset(args.devset, args.output_dir)
   port = random.randint(15000,50000)
+  logging.info('using port {}'.format(port))
   num_features = 0
   last_p_score = 0
   best_score_iter = -1
@@ -316,8 +326,8 @@ def optimize(args, script_dir, dev_size):
   hope_best_fear = {'hope':[],'best':[],'fear':[]}
   #main optimization loop
   while i<args.max_iterations:
-    logging.info('\n\nITERATION {}\n========\n'.format(i))
-    logging.info('using port {}\n'.format(port))
+    logging.info('======= STARTING ITERATION {} ======='.format(i))
+    logging.info('Starting at {}'.format(time.asctime()))
 
     #iteration specific files
     runfile = args.output_dir+'/run.raw.'+str(i)
@@ -327,10 +337,8 @@ def optimize(args, script_dir, dev_size):
     weightdir = args.output_dir+'/weights.pass'+str(i)
     os.mkdir(logdir)
     os.mkdir(weightdir)
-    
-    logging.info('RUNNING DECODER AT {}'.format(time.asctime()))
     weightsfile = args.output_dir+'/weights.'+str(i)
-    logging.info('ITER {}\n'.format(i))
+    logging.info('  log directory={}'.format(logdir))
     curr_pass = '0{}'.format(i)
     decoder_cmd = ('{0} -c {1} -w {2} -r{3} -m {4} -s {5} -b {6} -k {7} -o {8}'
                    ' -p {9} -O {10} -D {11} -h {12} -f {13} -C {14}').format(
@@ -350,7 +358,7 @@ def optimize(args, script_dir, dev_size):
                     parallelize, logdir, args.jobs)
     
     cmd = parallel_cmd + ' ' + decoder_cmd
-    logging.info('COMMAND: \n{}\n'.format(cmd))
+    logging.info('OPTIMIZATION COMMAND: {}'.format(cmd))
    
     dlog = open(decoderlog,'w')
     runf = open(runfile,'w')
@@ -365,27 +373,26 @@ def optimize(args, script_dir, dev_size):
       p1.stdout.close()
       
       if exit_code:
-        logging.error('Failed with exit code {}\n'.format(exit_code))
+        logging.error('Failed with exit code {}'.format(exit_code))
         sys.exit(exit_code)
 
       try:
         f = open(runfile)
       except IOError, msg:
-        logging.error('Unable to open {}\n'.format(runfile))
+        logging.error('Unable to open {}'.format(runfile))
         sys.exit()
       
       num_topbest = sum(1 for line in f)
       f.close()
       if num_topbest == dev_size: break
-      logging.warning('Incorrect number of top best. '
-                      'Waiting for distributed filesystem and retrying.')
+      logging.warning('Incorrect number of top best. Sleeping for 10 seconds and retrying...')
       time.sleep(10)
       retries += 1
     
     if dev_size != num_topbest:
       logging.error("Dev set contains "+dev_size+" sentences, but we don't "
                     "have topbest for all of these. Decoder failure? "
-                    " Check "+decoderlog+'\n')
+                    " Check "+decoderlog)
       sys.exit()
     dlog.close()
     runf.close()
@@ -427,7 +434,7 @@ def optimize(args, script_dir, dev_size):
     hope_best_fear['hope'].append(dec_score)
     hope_best_fear['best'].append(dec_score_h)
     hope_best_fear['fear'].append(dec_score_f)
-    logging.info('DECODER SCORE: {0} HOPE: {1} FEAR: {2}\n'.format(
+    logging.info('DECODER SCORE: {0} HOPE: {1} FEAR: {2}'.format(
                   dec_score, dec_score_h, dec_score_f))
     if dec_score > best_score:
       best_score_iter = i
@@ -436,12 +443,13 @@ def optimize(args, script_dir, dev_size):
     new_weights_file = '{}/weights.{}'.format(args.output_dir, i+1)
     last_weights_file = '{}/weights.{}'.format(args.output_dir, i)
     i += 1
-    weight_files = weightdir+'/weights.mira-pass*.*[0-9].gz'
+    weight_files = args.output_dir+'/weights.pass*/weights.mira-pass*[0-9].gz'
     average_weights(new_weights_file, weight_files)
 
-  logging.info('\nBEST ITER: {} :: {}\n\n'.format(
+  logging.info('BEST ITERATION: {} (SCORE={})'.format(
                best_score_iter, best_score))
   weights_final = args.output_dir+'/weights.final'
+  logging.info('WEIGHTS FILE: {}'.format(weights_final))
   shutil.copy(last_weights_file, weights_final)
   average_final_weights(args.output_dir)
   
@@ -481,15 +489,15 @@ def gzip_file(filename):
 
 #average the weights for a given pass
 def average_weights(new_weights, weight_files):
-  logging.info('AVERAGE {} {}\n'.format(new_weights, weight_files))
+  logging.info('AVERAGE {} {}'.format(new_weights, weight_files))
   feature_weights = {}
   total_mult = 0.0
   for path in glob.glob(weight_files):
     score = gzip.open(path)
     mult = 0
-    logging.info('FILE {}\n'.format(path))
+    logging.info('  FILE {}'.format(path))
     msg, ran, mult = score.readline().strip().split(' ||| ')
-    logging.info('Processing {} {}'.format(ran, mult))
+    logging.info('  Processing {} {}'.format(ran, mult))
     for line in score:
       f,w = line.split(' ',1)
       if f in feature_weights:
@@ -500,34 +508,30 @@ def average_weights(new_weights, weight_files):
     score.close()
   
   #write new weights to outfile
+  logging.info('Writing averaged weights to {}'.format(new_weights))
   out = open(new_weights, 'w')
   for f in iter(feature_weights):
     avg = feature_weights[f]/total_mult
-    logging.info('{} {} {} ||| Printing {} {}\n'.format(f,feature_weights[f], 
-                 total_mult, f, avg))
     out.write('{} {}\n'.format(f,avg))
 
-def write_config(args):
-  config = ('\n'
-            'DECODER: '
-            '/usr0/home/eschling/cdec/training/mira/kbest_cut_mira\n'
-            'INI FILE: '+args.config+'\n'
-            'WORKING DIRECTORY: '+args.output_dir+'\n'
-            'DEVSET: '+args.devset+'\n'
-            'EVAL METRIC: '+args.metric+'\n'
-            'MAX ITERATIONS: '+str(args.max_iterations)+'\n'
-            'DECODE NODES: '+str(args.jobs)+'\n'
-            'INITIAL WEIGHTS: '+args.weights+'\n')
+def log_config(args):
+  logging.info('WORKING DIRECTORY={}'.format(args.output_dir))
+  logging.info('INI FILE={}'.format(args.config))
+  logging.info('DEVSET={}'.format(args.devset))
+  logging.info('EVAL METRIC={}'.format(args.metric))
+  logging.info('MAX ITERATIONS={}'.format(args.max_iterations))
+  logging.info('PARALLEL JOBS={}'.format(args.jobs))
+  logging.info('INITIAL WEIGHTS={}'.format(args.weights))
   if args.grammar_prefix:
-    config += 'GRAMMAR PREFIX: '+str(args.grammar_prefix)+'\n'
+    logging.info('GRAMMAR PREFIX={}'.format(args.grammar_prefix))
   if args.test:
-    config += 'TEST SET: '+args.test+'\n'
+    logging.info('TEST SET={}'.format(args.test))
+  else:
+    logging.info('TEST SET=none specified')
   if args.test_config:
-    config += 'TEST CONFIG: '+args.test_config+'\n'
+    logging.info('TEST CONFIG={}'.format(args.test_config))
   if args.email:
-    config += 'EMAIL: '+args.email+'\n'
-           
-  logging.info(config)
+    logging.info('EMAIL={}'.format(args.email))
 
 if __name__=='__main__':
   main()
diff --git a/training/pro/mr_pro_map.cc b/training/pro/mr_pro_map.cc
index eef40b8a..a5e6e48f 100644
--- a/training/pro/mr_pro_map.cc
+++ b/training/pro/mr_pro_map.cc
@@ -2,7 +2,6 @@
 #include <iostream>
 #include <fstream>
 #include <vector>
-#include <tr1/unordered_map>
 
 #include <boost/functional/hash.hpp>
 #include <boost/shared_ptr.hpp>
diff --git a/training/utils/candidate_set.cc b/training/utils/candidate_set.cc
index 087efec3..33dae9a3 100644
--- a/training/utils/candidate_set.cc
+++ b/training/utils/candidate_set.cc
@@ -1,6 +1,11 @@
 #include "candidate_set.h"
 
-#include <tr1/unordered_set>
+#ifndef HAVE_OLD_CPP
+# include <unordered_set>
+#else
+# include <tr1/unordered_set>
+namespace std { using std::tr1::unordered_set; }
+#endif
 
 #include <boost/functional/hash.hpp>
 
@@ -139,12 +144,12 @@ void CandidateSet::ReadFromFile(const string& file) {
 
 void CandidateSet::Dedup() {
   if(!SILENT) cerr << "Dedup in=" << cs.size();
-  tr1::unordered_set<Candidate, CandidateHasher, CandidateCompare> u;
+  unordered_set<Candidate, CandidateHasher, CandidateCompare> u;
   while(cs.size() > 0) {
     u.insert(cs.back());
     cs.pop_back();
   }
-  tr1::unordered_set<Candidate, CandidateHasher, CandidateCompare>::iterator it = u.begin();
+  unordered_set<Candidate, CandidateHasher, CandidateCompare>::iterator it = u.begin();
   while (it != u.end()) {
     cs.push_back(*it);
     it = u.erase(it);
diff --git a/training/utils/online_optimizer.h b/training/utils/online_optimizer.h
index 28d89344..19223e9d 100644
--- a/training/utils/online_optimizer.h
+++ b/training/utils/online_optimizer.h
@@ -1,10 +1,10 @@
 #ifndef _ONL_OPTIMIZE_H_
 #define _ONL_OPTIMIZE_H_
 
-#include <tr1/memory>
 #include <set>
 #include <string>
 #include <cmath>
+#include <boost/shared_ptr.hpp>
 #include "sparse_vector.h"
 
 struct LearningRateSchedule {
@@ -56,7 +56,7 @@ struct ExponentialDecayLearningRate : public LearningRateSchedule {
 class OnlineOptimizer {
  public:
   virtual ~OnlineOptimizer();
-  OnlineOptimizer(const std::tr1::shared_ptr<LearningRateSchedule>& s,
+  OnlineOptimizer(const boost::shared_ptr<LearningRateSchedule>& s,
                   size_t batch_size,
                   const std::vector<int>& frozen_feats = std::vector<int>())
       : N_(batch_size),schedule_(s),k_() {
@@ -77,13 +77,13 @@ class OnlineOptimizer {
   std::set<int> frozen_;  // frozen (non-optimizing) features
 
  private:
-  std::tr1::shared_ptr<LearningRateSchedule> schedule_;
+  boost::shared_ptr<LearningRateSchedule> schedule_;
   int k_;  // iteration count
 };
 
 class CumulativeL1OnlineOptimizer : public OnlineOptimizer {
  public:
-  CumulativeL1OnlineOptimizer(const std::tr1::shared_ptr<LearningRateSchedule>& s,
+  CumulativeL1OnlineOptimizer(const boost::shared_ptr<LearningRateSchedule>& s,
                               size_t training_instances, double C,
                               const std::vector<int>& frozen) :
     OnlineOptimizer(s, training_instances, frozen), C_(C), u_() {}
diff --git a/training/utils/optimize_test.cc b/training/utils/optimize_test.cc
index bff2ca03..72fcef6d 100644
--- a/training/utils/optimize_test.cc
+++ b/training/utils/optimize_test.cc
@@ -2,6 +2,7 @@
 #include <iostream>
 #include <sstream>
 #include <boost/program_options/variables_map.hpp>
+#include <boost/shared_ptr.hpp>
 #include "optimize.h"
 #include "online_optimizer.h"
 #include "sparse_vector.h"
@@ -96,14 +97,11 @@ void TestOptimizerVariants(int num_vars) {
   cerr << oa.Name() << " SUCCESS\n";
 }
 
-using namespace std::tr1;
-
 void TestOnline() {
   size_t N = 20;
   double C = 1.0;
   double eta0 = 0.2;
-  std::tr1::shared_ptr<LearningRateSchedule> r(new ExponentialDecayLearningRate(N, eta0, 0.85));
-  //shared_ptr<LearningRateSchedule> r(new StandardLearningRate(N, eta0));
+  boost::shared_ptr<LearningRateSchedule> r(new ExponentialDecayLearningRate(N, eta0, 0.85));
   CumulativeL1OnlineOptimizer opt(r, N, C, std::vector<int>());
   assert(r->eta(10) < r->eta(1));
 }