dtrain: added pclr variants and new expected-output; fixed bug in soft syntax features

4 files changed, 100 insertions, 63 deletions

diff --git a/training/dtrain/dtrain.cc b/training/dtrain/dtrain.cc
index 9d60a903..38a9b69a 100644
--- a/training/dtrain/dtrain.cc
+++ b/training/dtrain/dtrain.cc
@@ -40,7 +40,7 @@ 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<bool>()->zero_tokens(),                         "use a (simple) per-coordinate learning rate")
+    ("pclr",              po::value<string>()->default_value("no"),         "use a (simple|adagrad) per-coordinate learning rate")
     ("noup",              po::value<bool>()->zero_tokens(),                                               "do not update weights");
   po::options_description cl("Command Line Options");
   cl.add_options()
@@ -125,8 +125,7 @@ main(int argc, char** argv)
   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;
-  bool pclr = false;
-  if (cfg.count("pclr")) pclr = true;
+  const string pclr = cfg["pclr"].as<string>();
   bool average = false;
   if (select_weights == "avg")
     average = true;
@@ -190,7 +189,6 @@ main(int argc, char** argv)
   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;
 
@@ -251,8 +249,7 @@ 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;
-    if (pclr)
-      cerr << setw(25) << "pclr " << pclr << endl;
+    cerr << setw(25) << "pclr " << pclr << endl;
     cerr << setw(25) << "max pairs " << max_pairs << endl;
     cerr << setw(25) << "cdec cfg " << "'" << cfg["decoder_config"].as<string>() << "'" << endl;
     cerr << setw(25) << "input " << "'" << input_fn << "'" << endl;
@@ -392,22 +389,30 @@ main(int argc, char** argv)
         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;
-          if (pclr) {
+          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./npairs)); // FIXME
           }
-          if (gamma)
-            lambdas.plus_eq_v_times_s(lambdas, -2*gamma*eta*(1./npairs));
         }
       }
 
       // per-coordinate learning rate
-      if (pclr) {
+      if (pclr != "no") {
         SparseVector<weight_t>::iterator it = sum_up.begin();
-        for (; it != lambdas.end(); ++it) {
-          learning_rates[it->first]++;
-          lambdas[it->first] += it->second / learning_rates[it->first]; //* max(0.00000001, eta/(eta+learning_rates[it->first]));
+        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);
+          }
         }
       }
 
diff --git a/training/dtrain/examples/standard/dtrain.ini b/training/dtrain/examples/standard/dtrain.ini
index c0912a62..e6d6382e 100644
--- a/training/dtrain/examples/standard/dtrain.ini
+++ b/training/dtrain/examples/standard/dtrain.ini
@@ -1,6 +1,6 @@
 input=./nc-wmt11.de.gz
 refs=./nc-wmt11.en.gz
-output=asdf                  # a weights file (add .gz for gzip compression) or STDOUT '-'
+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
 # weights for these features will be printed on each iteration
@@ -10,11 +10,11 @@ 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=1.0        # 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)
diff --git a/training/dtrain/examples/standard/expected-output b/training/dtrain/examples/standard/expected-output
index 21f91244..a35bbe6f 100644
--- a/training/dtrain/examples/standard/expected-output
+++ b/training/dtrain/examples/standard/expected-output
@@ -4,13 +4,13 @@ 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 4049211323
 
 dtrain
 Parameters:
                        k 100
                        N 4
-                       T 2
+                       T 3
                   scorer 'fixed_stupid_bleu'
              sample from 'kbest'
                   filter 'uniq'
@@ -23,6 +23,7 @@ Parameters:
           pair threshold 0
           select weights 'VOID'
                   l1 reg 0 'none'
+                    pclr no
                max pairs 4294967295
                 cdec cfg './cdec.ini'
                    input './nc-wmt11.de.gz'
@@ -30,62 +31,88 @@ Parameters:
                   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 = -1100
+       WordPenalty = -82.082
+     LanguageModel = -3199.1
+ LanguageModel_OOV = -192
+     PhraseModel_0 = +3128.2
+     PhraseModel_1 = -1610.2
+     PhraseModel_2 = -4336.5
+     PhraseModel_3 = +2910.3
+     PhraseModel_4 = +2523.2
+     PhraseModel_5 = +506
+     PhraseModel_6 = +1467
+       PassThrough = -387
         ---
-       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: 2.9874e+05 (+2.9874e+05)
+           avg # pairs: 906.3 (meaningless)
+        avg # rank err: 906.3
      avg # margin viol: 0
-    non0 feature count: 887
+    non0 feature count: 825
            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 = -1221
+       WordPenalty = +836.89
+     LanguageModel = +2332.3
+ LanguageModel_OOV = -1451
+     PhraseModel_0 = +1507.2
+     PhraseModel_1 = -2728.4
+     PhraseModel_2 = -4183.6
+     PhraseModel_3 = +1816.3
+     PhraseModel_4 = -2894.7
+     PhraseModel_5 = +1403
+     PhraseModel_6 = +35
+       PassThrough = -1097
         ---
-       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: 49369 (-2.4937e+05)
+           avg # pairs: 662.4 (meaningless)
+        avg # rank err: 662.4
      avg # margin viol: 0
-    non0 feature count: 1499
+    non0 feature count: 1235
            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 = -1574
+       WordPenalty = -17.372
+     LanguageModel = +6861.8
+ LanguageModel_OOV = -3997
+     PhraseModel_0 = -398.76
+     PhraseModel_1 = -3419.6
+     PhraseModel_2 = -3186.7
+     PhraseModel_3 = +1050.8
+     PhraseModel_4 = -2902.7
+     PhraseModel_5 = -486
+     PhraseModel_6 = -436
+       PassThrough = -2985
+        ---
+       1best avg score: 0.30742 (+0.13343)
+ 1best avg model score: -1.5393e+05 (-2.0329e+05)
+           avg # pairs: 623.8 (meaningless)
+        avg # rank err: 623.8
+     avg # margin viol: 0
+    non0 feature count: 1770
+           avg list sz: 91.3
+           avg f count: 118.58
+(time 0.25 min, 1.5 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 0.86667 min.
diff --git a/training/dtrain/parallelize.rb b/training/dtrain/parallelize.rb
index 2fc66cab..60ca9422 100755
--- a/training/dtrain/parallelize.rb
+++ b/training/dtrain/parallelize.rb
@@ -21,7 +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 :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]
 
@@ -54,6 +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`
 
@@ -137,10 +139,13 @@ end
       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} #{cdec_cfg}\
+        `#{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}"