Merge branch 'master' into gd_get

vowpalwabbit/core/include/vw/core/reductions/active.h

@@ -16,10 +16,18 @@ namespace reductions
 class active
 {
 public:
-  active(float active_c0, VW::workspace* all) : active_c0(active_c0), _all(all) {}
+  active(float active_c0, std::shared_ptr<shared_data> shared_data, std::shared_ptr<rand_state> random_state,
+      VW::version_struct model_version)
+      : active_c0(active_c0)
+      , _shared_data(shared_data)
+      , _random_state(std::move(random_state))
+      , _model_version{std::move(model_version)}
+  {
+  }
 
   float active_c0;
-  VW::workspace* _all = nullptr;
+  std::shared_ptr<shared_data> _shared_data;  // statistics, loss
+  std::shared_ptr<rand_state> _random_state;
 
   float _min_seen_label = 0.f;
   float _max_seen_label = 1.f;

vowpalwabbit/core/include/vw/core/vw_versions.h

@@ -59,6 +59,9 @@ constexpr VW::version_struct VERSION_PASS_UINT64{8, 3, 3};
 
 /// Added serialized seen min and max labels in the --active reduction
 constexpr VW::version_struct VERSION_FILE_WITH_ACTIVE_SEEN_LABELS{9, 0, 0};
+/// Active seen labels was accidentally reverted out in 9.4.0
+constexpr VW::version_struct VERSION_FILE_WITH_ACTIVE_SEEN_LABELS_REVERTED{9, 4, 0};
+constexpr VW::version_struct VERSION_FILE_WITH_ACTIVE_SEEN_LABELS_FIXED{9, 10, 0};
 
 /// Moved option values from command line to model data
 constexpr VW::version_struct VERSION_FILE_WITH_L1_AND_L2_STATE_IN_MODEL_DATA{9, 0, 0};

vowpalwabbit/core/src/reductions/active.cc

@@ -50,13 +50,13 @@ float query_decision(const active& a, float ec_revert_weight, float k)
   if (k <= 1.f) { bias = 1.f; }
   else
   {
-    const auto weighted_queries = static_cast<float>(a._all->sd->weighted_labeled_examples);
-    const float avg_loss = (static_cast<float>(a._all->sd->sum_loss) / k) +
+    const auto weighted_queries = static_cast<float>(a._shared_data->weighted_labeled_examples);
+    const float avg_loss = (static_cast<float>(a._shared_data->sum_loss) / k) +
         std::sqrt((1.f + 0.5f * std::log(k)) / (weighted_queries + 0.0001f));
     bias = get_active_coin_bias(k, avg_loss, ec_revert_weight / k, a.active_c0);
   }
 
-  return (a._all->get_random_state()->get_and_update_random() < bias) ? 1.f / bias : -1.f;
+  return (a._random_state->get_and_update_random() < bias) ? 1.f / bias : -1.f;
 }
 
 template <bool is_learn>
@@ -66,15 +66,15 @@ void predict_or_learn_simulation(active& a, learner& base, VW::example& ec)
 
   if (is_learn)
   {
-    const auto k = static_cast<float>(a._all->sd->t);
+    const auto k = static_cast<float>(a._shared_data->t);
     constexpr float threshold = 0.f;
 
     ec.confidence = fabsf(ec.pred.scalar - threshold) / base.sensitivity(ec);
     const float importance = query_decision(a, ec.confidence, k);
 
     if (importance > 0.f)
     {
-      a._all->sd->queries += 1;
+      a._shared_data->queries += 1;
       ec.weight *= importance;
       base.learn(ec);
     }
@@ -94,7 +94,7 @@ void predict_or_learn_active(active& a, learner& base, VW::example& ec)
 
   if (ec.l.simple.label == FLT_MAX)
   {
-    const float threshold = (a._all->sd->max_label + a._all->sd->min_label) * 0.5f;
+    const float threshold = (a._shared_data->max_label + a._shared_data->min_label) * 0.5f;
     // We want to understand the change in prediction if the label were to be
     // the opposite of what was predicted. 0 and 1 are used for the expected min
     // and max labels to be coming in from the active interactor.
@@ -129,8 +129,14 @@ void active_print_result(
 
 void save_load(active& a, VW::io_buf& io, bool read, bool text)
 {
+  using namespace VW::version_definitions;
   if (io.num_files() == 0) { return; }
-  if (a._model_version >= VW::version_definitions::VERSION_FILE_WITH_ACTIVE_SEEN_LABELS)
+  // This code is valid if version is within
+  // [VERSION_FILE_WITH_ACTIVE_SEEN_LABELS, VERSION_FILE_WITH_ACTIVE_SEEN_LABELS_REVERTED)
+  // or >= VERSION_FILE_WITH_ACTIVE_SEEN_LABELS_FIXED
+  if ((a._model_version >= VERSION_FILE_WITH_ACTIVE_SEEN_LABELS &&
+          a._model_version < VERSION_FILE_WITH_ACTIVE_SEEN_LABELS_REVERTED) ||
+      a._model_version >= VERSION_FILE_WITH_ACTIVE_SEEN_LABELS_FIXED)
   {
     if (read)
     {
@@ -195,7 +201,7 @@ std::shared_ptr<VW::LEARNER::learner> VW::reductions::active_setup(VW::setup_bas
   if (!options.add_parse_and_check_necessary(new_options)) { return nullptr; }
 
   if (options.was_supplied("lda")) { THROW("lda cannot be combined with active learning") }
-  auto data = VW::make_unique<active>(active_c0, &all);
+  auto data = VW::make_unique<active>(active_c0, all.sd, all.get_random_state(), all.runtime_state.model_file_ver);
   auto base = require_singleline(stack_builder.setup_base_learner());
 
   using learn_pred_func_t = void (*)(active&, VW::LEARNER::learner&, VW::example&);