l1 regularzation

1 files changed, 50 insertions, 1 deletions

diff --git a/dtrain/dtrain.cc b/dtrain/dtrain.cc
index ca5f0c5e..448e639c 100644
--- a/dtrain/dtrain.cc
+++ b/dtrain/dtrain.cc
@@ -26,6 +26,8 @@ dtrain_init(int argc, char** argv, po::variables_map* cfg)
     ("select_weights", po::value<string>()->default_value("last"), "output 'best' or 'last' weights ('VOID' to throw away)")
     ("keep_w",         po::value<bool>()->zero_tokens(),                              "protocol weights for each iteration")
     ("unit_weight_vector", po::value<bool>()->zero_tokens(),                       "Rescale weight vector after each input")
+    ("l1_reg",         po::value<string>()->default_value("no"),         "apply l1 regularization as in Tsuroka et al 2010")
+    ("l1_reg_strength", po::value<weight_t>(),                                                 "l1 regularization strength")
 #ifdef DTRAIN_LOCAL
     ("refs,r",         po::value<string>(),                                                      "references in local mode")
 #endif
@@ -155,13 +157,25 @@ main(int argc, char** argv)
 
   // init weights
   vector<weight_t>& dense_weights = decoder.CurrentWeightVector();
-  SparseVector<weight_t> lambdas;
+  SparseVector<weight_t> lambdas, cumulative_penalties;
   if (cfg.count("input_weights")) Weights::InitFromFile(cfg["input_weights"].as<string>(), &dense_weights);
   Weights::InitSparseVector(dense_weights, &lambdas);
 
   // meta params for perceptron, SVM
   weight_t eta = cfg["learning_rate"].as<weight_t>();
   weight_t gamma = cfg["gamma"].as<weight_t>();
+  // l1 regularization
+  bool l1naive = false;
+  bool l1clip = false;
+  bool l1cumul = false;
+  weight_t l1_reg = 0;
+  if (cfg["l1_reg"].as<string>() != "no") {
+    string s = cfg["l1_reg"].as<string>();
+    if (s == "naive") l1naive = true;
+    else if (s == "clip") l1clip = true;
+    else if (s == "cumul") l1cumul = true;
+    l1_reg = cfg["l1_reg_strength"].as<weight_t>();
+  }
 
   // output
   string output_fn = cfg["output"].as<string>();
@@ -388,6 +402,41 @@ main(int argc, char** argv)
           lambdas.plus_eq_v_times_s(lambdas, -2*gamma*eta*(1./npairs));
         }
       }
+
+      // reset cumulative_penalties after 1 iter? 
+      // do this only once per INPUT (not per pair)
+      if (l1naive) {
+        for (unsigned d = 0; d < lambdas.size(); d++) {
+          weight_t v = lambdas.get(d);
+          lambdas.set_value(d, v - sign(v) * l1_reg);
+        }
+      } else if (l1clip) {
+        for (unsigned d = 0; d < lambdas.size(); d++) {
+          if (lambdas.nonzero(d)) {
+            weight_t v = lambdas.get(d);
+            if (v > 0) {
+              lambdas.set_value(d, max(0., v - l1_reg));
+            } else {
+              lambdas.set_value(d, min(0., v + l1_reg));
+            }
+          }
+        }
+      } else if (l1cumul) {
+        weight_t acc_penalty = (ii+1) * l1_reg; // ii is the index of the current input 
+        for (unsigned d = 0; d < lambdas.size(); d++) {
+          if (lambdas.nonzero(d)) {
+            weight_t v = lambdas.get(d);
+            weight_t penalty = 0;
+            if (v > 0) {
+              penalty = max(0., v-(acc_penalty + cumulative_penalties.get(d)));
+            } else {
+              penalty = min(0., v+(acc_penalty - cumulative_penalties.get(d)));
+            }
+            lambdas.set_value(d, penalty);
+            cumulative_penalties.set_value(d, cumulative_penalties.get(d)+penalty);
+          }
+        }
+      }
     }
 
     if (unit_weight_vector && sample_from == "forest") lambdas /= lambdas.l2norm();