Convert TopK reduction to be multiline example based (VowpalWabbit#1752)

* Make topk a multiline learner * Fix test for new format and rename B to K * revert destructor usage * remove all from data, move to output seq in finish * Revert "remove all from data, move to output seq in finish" This reverts commit 2c8f965.
lucjb · Mar 7, 2019 · 5edac80 · 5edac80
1 parent 1ef3cac
commit 5edac80
Show file tree

Hide file tree

Showing 2 changed files with 39 additions and 37 deletions.
diff --git a/test/train-sets/ref/topk-rec.stderr b/test/train-sets/ref/topk-rec.stderr
@@ -1,4 +1,4 @@
-creating quadratic features for pairs: MF 
+creating quadratic features for pairs: MF
 predictions = topk-rec.predict
 Num weight bits = 18
 learning rate = 0.5
@@ -12,20 +12,17 @@ loss     last          counter         weight    label  predict features
 0.000003 0.000003            1            1.0   3.0000   2.9983        4
 0.000004 0.000005            2            2.0   0.0000   0.0022        4
 0.000003 0.000000            3            3.0   2.0000   2.0000        4
-0.000003     n.a.            4            4.0  unknown   0.0000        1
-0.000003 0.000003            5            5.0   0.0000   0.0018        4
-0.000004 0.000011            6            6.0   3.0000   2.9968        4
-0.000004 0.000001            7            7.0   1.0000   1.0007        4
-0.000004     n.a.            8            8.0  unknown   0.0000        1
-0.000003 0.000000            9            9.0   2.0000   2.0004        4
-0.000003 0.000000           10           10.0   1.0000   1.0003        4
-0.000002 0.000000           11           11.0   3.0000   2.9995        4
-0.000002     n.a.           12           12.0  unknown   0.0000        1
+0.000003 0.000003            4            4.0   0.0000   0.0018        4
+0.000004 0.000011            5            5.0   3.0000   2.9968        4
+0.000004 0.000001            6            6.0   1.0000   1.0007        4
+0.000003 0.000000            7            7.0   2.0000   2.0004        4
+0.000003 0.000000            8            8.0   1.0000   1.0003        4
+0.000002 0.000000            9            9.0   3.0000   2.9995        4
 
 finished run
-number of examples = 12
-weighted example sum = 12.000000
+number of examples = 9
+weighted example sum = 9.000000
 weighted label sum = 15.000000
 average loss = 0.000002
 best constant = 1.666667
-total feature number = 39
+total feature number = 36
diff --git a/vowpalwabbit/topk.cc b/vowpalwabbit/topk.cc
@@ -13,7 +13,7 @@ license as described in the file LICENSE.
 using namespace std;
 using namespace VW::config;
 
-typedef pair<float, v_array<char> > scored_example;
+using scored_example = pair<float, v_array<char>>;
 
 struct compare_scored_examples
 {
@@ -22,8 +22,9 @@ struct compare_scored_examples
 
 struct topk
 {
-  uint32_t B;  // rec number
+  uint32_t K;  // rec number
   priority_queue<scored_example, vector<scored_example>, compare_scored_examples> pr_queue;
+  vw* all;
 };
 
 void print_result(int f, priority_queue<scored_example, vector<scored_example>, compare_scored_examples>& pr_queue)
@@ -56,45 +57,47 @@ void print_result(int f, priority_queue<scored_example, vector<scored_example>,
   }
 }
 
-void output_example(vw& all, topk& d, example& ec)
+void output_example(vw& all, example& ec)
 {
   label_data& ld = ec.l.simple;
 
   all.sd->update(ec.test_only, ld.label != FLT_MAX, ec.loss, ec.weight, ec.num_features);
   if (ld.label != FLT_MAX)
     all.sd->weighted_labels += ((double)ld.label) * ec.weight;
 
-  if (example_is_newline(ec))
-    for (int sink : all.final_prediction_sink) print_result(sink, d.pr_queue);
-
   print_update(all, ec);
 }
 
 template <bool is_learn>
-void predict_or_learn(topk& d, LEARNER::single_learner& base, example& ec)
+void predict_or_learn(topk& d, LEARNER::single_learner& base, multi_ex& ec_seq)
 {
-  if (example_is_newline(ec))
-    return;  // do not predict newline
+  for (auto example : ec_seq)
+  {
+    auto ec = *example;
 
-  if (is_learn)
-    base.learn(ec);
-  else
-    base.predict(ec);
+    if (is_learn)
+      base.learn(ec);
+    else
+      base.predict(ec);
 
-  if (d.pr_queue.size() < d.B)
-    d.pr_queue.push(make_pair(ec.pred.scalar, ec.tag));
+    if (d.pr_queue.size() < d.K)
+      d.pr_queue.push(make_pair(ec.pred.scalar, ec.tag));
+    else if (d.pr_queue.top().first < ec.pred.scalar)
+    {
+      d.pr_queue.pop();
+      d.pr_queue.push(make_pair(ec.pred.scalar, ec.tag));
+    }
 
-  else if (d.pr_queue.top().first < ec.pred.scalar)
-  {
-    d.pr_queue.pop();
-    d.pr_queue.push(make_pair(ec.pred.scalar, ec.tag));
+    output_example(*d.all, ec);
   }
 }
 
-void finish_example(vw& all, topk& d, example& ec)
+void finish_example(vw& all, topk& d, multi_ex& ec_seq)
 {
-  output_example(all, d, ec);
-  VW::finish_example(all, ec);
+  for (int sink : all.final_prediction_sink)
+    print_result(sink, d.pr_queue);
+
+  VW::clear_seq_and_finish_examples(all, ec_seq);
 }
 
 void finish(topk& d) { d.pr_queue = priority_queue<scored_example, vector<scored_example>, compare_scored_examples>(); }
@@ -104,13 +107,15 @@ LEARNER::base_learner* topk_setup(options_i& options, vw& all)
   auto data = scoped_calloc_or_throw<topk>();
 
   option_group_definition new_options("Top K");
-  new_options.add(make_option("top", data->B).keep().help("top k recommendation"));
+  new_options.add(make_option("top", data->K).keep().help("top k recommendation"));
   options.add_and_parse(new_options);
 
   if (!options.was_supplied("top"))
     return nullptr;
 
-  LEARNER::learner<topk, example>& l =
+  data->all = &all;
+
+  LEARNER::learner<topk, multi_ex>& l =
       init_learner(data, as_singleline(setup_base(options, all)), predict_or_learn<true>, predict_or_learn<false>);
   l.set_finish_example(finish_example);
   l.set_finish(finish);