dtrain_net_interface: features, more detailed update debug output

2 files changed, 5 insertions, 2 deletions

diff --git a/training/dtrain/dtrain_net_interface.cc b/training/dtrain/dtrain_net_interface.cc
index f16b9304..77ccde55 100644
--- a/training/dtrain/dtrain_net_interface.cc
+++ b/training/dtrain/dtrain_net_interface.cc
@@ -143,7 +143,7 @@ main(int argc, char** argv)
           cerr << "[dtrain] learning ..." << endl;
           source = parts[0];
           // debug --
-          debug_output << "\"source\":\"" <<  source.substr(source.find_first_of(">")+1, source.find_last_of("<")-3) <<  "\"," << endl;
+          debug_output << "\"source\":\"" <<  source.substr(source.find_first_of(">")+2, source.find_last_of(">")-6) <<  "\"," << endl;
           debug_output << "\"target\":\"" << parts[1] <<  "\"," << endl;
           // -- debug
           parts.erase(parts.begin());
@@ -198,11 +198,14 @@ main(int argc, char** argv)
     // get pairs and update
     SparseVector<weight_t> updates;
     size_t num_up = CollectUpdates(samples, updates, margin);
+    debug_output << "\"1best_features\":\"" << (*samples)[0].f << "\"," << endl;
+    debug_output << "\"update_raw\":\"" << updates << "\"," << endl;
     updates *= eta_sparse; // apply learning rate for sparse features
     for (auto feat: dense_features) { // apply learning rate for dense features
       updates[FD::Convert(feat)] /= eta_sparse;
       updates[FD::Convert(feat)] *= eta;
     }
+    debug_output << "\"update\":\"" << updates << "\"," << endl;
     // debug --
     debug_output << "\"num_up\":" << num_up << "," << endl;
     debug_output << "\"updated_features\":" << updates.size() << "," << endl;
diff --git a/training/dtrain/dtrain_net_interface.h b/training/dtrain/dtrain_net_interface.h
index debe8800..b201c7a3 100644
--- a/training/dtrain/dtrain_net_interface.h
+++ b/training/dtrain/dtrain_net_interface.h
@@ -66,7 +66,7 @@ dtrain_net_init(int argc, char** argv, po::variables_map* conf)
     ("learning_rate_sparse,l", po::value<weight_t>()->default_value(0.00001), "learning rate for sparse features")
     ("output_derivation,E",    po::bool_switch()->default_value(false),  "output derivation, not viterbi str")
     ("output_rules,R",    po::bool_switch()->default_value(false),                        "also output rules")
-    ("dense_features,D",       po::value<string>()->default_value("EgivenFCoherent SampleCountF CountEF MaxLexFgivenE MaxLexEgivenF IsSingletonF IsSingletonFE Glue WordPenalty PassThrough LanguageModel LanguageModel_OOV Shape_S01111_T11011 Shape_S11110_T11011 Shape_S11100_T11000 Shape_S01110_T01110 Shape_S01111_T01111 Shape_S01100_T11000 Shape_S10000_T10000 Shape_S11100_T11100 Shape_S11110_T11110 Shape_S11110_T11010 Shape_S01100_T11100 Shape_S01000_T01000 Shape_S01010_T01010 Shape_S01111_T01011 Shape_S01100_T01100 Shape_S01110_T11010 Shape_S11000_T11000 Shape_S11000_T01100 IsSupportedOnline ForceRule"),
+    ("dense_features,D",       po::value<string>()->default_value("EgivenFCoherent SampleCountF CountEF MaxLexFgivenE MaxLexEgivenF IsSingletonF IsSingletonFE Glue WordPenalty PassThrough LanguageModel LanguageModel_OOV Shape_S01111_T11011 Shape_S11110_T11011 Shape_S11100_T11000 Shape_S01110_T01110 Shape_S01111_T01111 Shape_S01100_T11000 Shape_S10000_T10000 Shape_S11100_T11100 Shape_S11110_T11110 Shape_S11110_T11010 Shape_S01100_T11100 Shape_S01000_T01000 Shape_S01010_T01010 Shape_S01111_T01011 Shape_S01100_T01100 Shape_S01110_T11010 Shape_S11000_T11000 Shape_S11000_T01100 IsSupportedOnline NewRule KnownRule OOVFix"),
                                                                                              "dense features")
     ("debug_output,d",   po::value<string>()->default_value(""),                      "file for debug output");
   po::options_description cl("Command Line Options");