ready to push

python/pylibvw.cc

@@ -359,24 +359,13 @@ BOOST_PYTHON_MODULE(pylibvw) {
       .def("get_topic_prediction", &VW::get_topic_prediction, "For LDA models, returns the topic prediction for the topic id given")
       .def("get_feature_number", &VW::get_feature_number, "Returns the total number of features for this example")
 
-      // the following four are redundant with get_simplelabel_...
-      //.def("get_label", &VW::get_label, "Returns the label (simple_label only!) for this example")
-      //.def("get_importance", &VW::get_importance, "Returns the importance (simple_label only!) for this example")
-      //.def("get_initial", &VW::get_initial, "TODO")
-      //.def("get_prediction", &VW::get_prediction, "TODO")
-
       .def("get_example_counter", &get_example_counter, "Returns the counter of total number of examples seen up to and including this one")
       .def("get_ft_offset", &get_ft_offset, "Returns the feature offset for this example (used, eg, by multiclass classification to bulk offset all features)")
       .def("get_partial_prediction", &get_partial_prediction, "Returns the partial prediction associated with this example")
       .def("get_updated_prediction", &get_updated_prediction, "Returns the partial prediction as if we had updated it after learning")
       .def("get_loss", &get_loss, "Returns the loss associated with this example")
       .def("get_example_t", &get_example_t, "The total sum of importance weights up to and including this example")
       .def("get_total_sum_feat_sq", &get_total_sum_feat_sq, "The total sum of feature-value squared for this example")
-      //      .def_readwrite("revert_weight", &example::revert_weight)
-      //      .def_readwrite("test_only", &example::test_only)
-      //      .def_readwrite("end_pass", &example::end_pass)
-      //      .def_readwrite("sorted", &example::sorted)
-      //      .def_readwrite("in_use", &example::in_use)
 
       .def("num_namespaces", &ex_num_namespaces, "The total number of namespaces associated with this example")
       .def("namespace", &ex_namespace, "Get the namespace id for namespace i (for i = 0.. num_namespaces); specifically returns the ord() of the corresponding character id")
@@ -415,23 +404,23 @@ BOOST_PYTHON_MODULE(pylibvw) {
       ;
 
   py::class_<Search::predictor, predictor_ptr>("predictor", py::no_init)
-      .def("set_input", &my_set_input, "TODO")
-      .def("set_input_ldf", &my_set_input_LDF, "TODO")
-      .def("add_oracle", &my_add_oracle, "TODO")
-      .def("add_oracles", &my_add_oracles, "TODO")
-      .def("add_allowed", &my_add_allowed, "TODO")
-      .def("add_alloweds", &my_add_alloweds, "TODO")
-      .def("add_condition", &my_add_condition, "TODO")
-      .def("add_condition_range", &my_add_condition_range, "TODO")
-      .def("set_oracle", &my_set_oracle, "TODO")
-      .def("set_oracles", &my_set_oracles, "TODO")
-      .def("set_allowed", &my_set_allowed, "TODO")
-      .def("set_alloweds", &my_set_alloweds, "TODO")
-      .def("set_condition", &my_set_condition, "TODO")
-      .def("set_condition_range", &my_set_condition_range, "TODO")
-      .def("set_learner_id", &my_set_learner_id, "TODO")
-      .def("set_tag", &my_set_tag, "TODO")
-      .def("predict", &Search::predictor::predict, "TODO")
+      .def("set_input", &my_set_input, "set the input (an example) for this predictor (non-LDF mode only)")
+      .def("set_input_ldf", &my_set_input_LDF, "set the inputs (a list of examples) for this predictor (LDF mode only)")
+      .def("add_oracle", &my_add_oracle, "add an action to the current list of oracle actions")
+      .def("add_oracles", &my_add_oracles, "add a list of actions to the current list of oracle actions")
+      .def("add_allowed", &my_add_allowed, "add an action to the current list of allowed actions")
+      .def("add_alloweds", &my_add_alloweds, "add a list of actions to the current list of allowed actions")
+      .def("add_condition", &my_add_condition, "add a (tag,char) pair to the list of variables on which to condition")
+      .def("add_condition_range", &my_add_condition_range, "given (tag,len,char), add (tag,char), (tag-1,char+1), ..., (tag-len,char+len) to the list of conditionings")
+      .def("set_oracle", &my_set_oracle, "set an action as the current list of oracle actions")
+      .def("set_oracles", &my_set_oracles, "set a list of actions as the current list of oracle actions")
+      .def("set_allowed", &my_set_allowed, "set an action as the current list of allowed actions")
+      .def("set_alloweds", &my_set_alloweds, "set a list of actions as the current list of allowed actions")
+      .def("set_condition", &my_set_condition, "set a (tag,char) pair as the list of variables on which to condition")
+      .def("set_condition_range", &my_set_condition_range, "given (tag,len,char), set (tag,char), (tag-1,char+1), ..., (tag-len,char+len) as the list of conditionings")
+      .def("set_learner_id", &my_set_learner_id, "select the learner with which to make this prediction")
+      .def("set_tag", &my_set_tag, "change the tag of this prediction")
+      .def("predict", &Search::predictor::predict, "make a prediction")
       ;
 
   py::class_<Search::search, search_ptr>("search")

test/RunTests

@@ -954,8 +954,8 @@ __DATA__
     train-sets/ref/sequence_data.ldf.test.stderr
     train-sets/ref/sequence_data.ldf.test.predict
 
-# Test 48: search sequence SPAN labeling BIO, non-ldf train
-{VW} -k -c -d train-sets/sequencespan_data --passes 20 --invariant --search_rollout oracle --search_alpha 1e-8 --search_task sequencespan --search 7 --holdout_off -f models/sequencespan_data.model
+# Test 48: search sequence SPAN labeling BIO, non-ldf train, no rollouts
+{VW} -k -c -d train-sets/sequencespan_data --passes 20 --invariant --search_rollout none --search_task sequencespan --search 7 --holdout_off -f models/sequencespan_data.model
     train-sets/ref/sequencespan_data.nonldf.train.stderr
 
 # Test 49: search sequence SPAN labeling BIO, non-ldf test

test/train-sets/ref/sequencespan_data.nonldf.train.stderr

@@ -4,16 +4,17 @@ learning rate = 10
 initial_t = 1
 power_t = 0.5
 decay_learning_rate = 1
+no rollout!
 creating cache_file = train-sets/sequencespan_data.cache
 Reading datafile = train-sets/sequencespan_data
 num sources = 1
 average    since      instance            current true      current predicted   cur   cur   predic    cache  examples          
 loss       last        counter           output prefix          output prefix  pass   pol     made     hits    gener  beta    
 6.000000   6.000000          1  [2 1 1 2 2 1 6 7 7 ..] [1 1 1 1 1 1 1 1 1 ..]     0     0       15        0       15  0.000000
 8.500000   11.000000         2  [2 1 1 2 2 1 6 7 7 ..] [2 1 6 4 1 6 4 1 6 ..]     1     0       30        0       30  0.000000
-6.000000   3.500000          4  [2 1 1 2 2 1 6 7 7 ..] [2 1 1 2 2 1 6 7 7 ..]     3     0       60        0       60  0.000001
-3.000000   0.000000          8  [2 1 1 2 2 1 6 7 7 ..] [2 1 1 2 2 1 6 7 7 ..]     7     0      120        0      120  0.000001
-1.500000   0.000000         16  [2 1 1 2 2 1 6 7 7 ..] [2 1 1 2 2 1 6 7 7 ..]    15     0      240        0      240  0.000002
+6.000000   3.500000          4  [2 1 1 2 2 1 6 7 7 ..] [2 1 1 2 2 1 6 7 7 ..]     3     0       60        0       60  0.000000
+3.000000   0.000000          8  [2 1 1 2 2 1 6 7 7 ..] [2 1 1 2 2 1 6 7 7 ..]     7     0      120        0      120  0.000000
+1.500000   0.000000         16  [2 1 1 2 2 1 6 7 7 ..] [2 1 1 2 2 1 6 7 7 ..]    15     0      240        0      240  0.000000
 
 finished run
 number of examples per pass = 1

vowpalwabbit/main.cc

@@ -85,7 +85,7 @@ int main(int argc, char *argv[])
         }
         cerr << endl << "weighted example sum = " << all->sd->weighted_examples;
         cerr << endl << "weighted label sum = " << all->sd->weighted_labels;
-        if(all->holdout_set_off)
+        if(all->holdout_set_off || (all->sd->holdout_best_loss == FLT_MAX))
         {
             cerr << endl << "average loss = " << all->sd->sum_loss / all->sd->weighted_examples;
         }