Merge remote-tracking branch 'upstream/master' into aml_mm

python/docs/source/examples/epsilon_decay.ipynb

@@ -5,7 +5,7 @@
    "id": "c06fbf85",
    "metadata": {},
    "source": [
-    "### [Experimental] Testing Basic Models with Varying Epsilon Values and Model Counts for Non-Stationary Epsilon Decay"
+    "# [Experimental] Testing Basic Models with Varying Epsilon Values and Model Counts for Non-Stationary Epsilon Decay"
    ]
   },
   {

python/docs/source/examples/index.md

@@ -5,6 +5,7 @@
 
 basics
 contextual_bandit
+epsilon_decay
 mini_vw
 poisson_regression
 predict

python/docs/source/tutorials/python_Simulating_a_news_personalization_scenario_using_Contextual_Bandits.ipynb

@@ -318,7 +318,7 @@
    "metadata": {},
    "source": [
     "\n",
-    "We want to be able to visualize what is occurring, so we are going to plot the click through rate over each iteration of the simulation. If VW is showing actions the get rewards the ctr will be higher. Below is a little utility function to make showing the plot easier.\n"
+    "We want to be able to visualize what is occurring, so we are going to plot the click through rate over each iteration of the simulation. If VW is showing actions that get rewards the ctr will be higher. Below is a little utility function to make showing the plot easier.\n"
    ]
   },
   {

python/tests/test_confidence_sequence_robust.py

@@ -92,13 +92,14 @@ def test_cs_robust_ci():
     csr.addobs((0.879389936, 1))
     csr.addobs((0.880083546, 1))
     csr.addobs((0.8807676, 1))
+    csr.addobs((43.91880613, 0.9))
 
     alpha = 0.05 / 16.0
     lb, ub = csr.getci(alpha)
 
     # Compare bounds to confidence_sequence_robust_test.cc
-    np.testing.assert_almost_equal(lb, 0.19224909776696902, 6)
-    np.testing.assert_almost_equal(ub, 0.93637288977188682, 6)
+    np.testing.assert_almost_equal(lb, 0.30209151281846858, 4)
+    np.testing.assert_almost_equal(ub, 0.90219143188334106, 4)
 
     # Test brentq separately
     s = 139.8326745448
@@ -114,8 +115,8 @@ def test_cs_robust_ci():
     )
 
     # Compare to brentq in confidence_sequence_robust_test.cc
-    np.testing.assert_almost_equal(res.root, 0.31672263211621371, 6)
+    np.testing.assert_almost_equal(res.root, 0.30209143008131789, 4)
 
     # Test that root of objective function is 0
     test_root = csr.lower.logwealthmix(mu=res.root, s=s, thres=thres, memo=memo) - thres
-    np.testing.assert_almost_equal(test_root, 0.0)
+    np.testing.assert_almost_equal(test_root, 0.0, 2)

test/train-sets/ref/coin_model_overflow.invert.txt

@@ -27,6 +27,6 @@ sd::oec.weighted_labeled_examples 1
 current_pass 1
 l1_state 0
 l2_state 1
-34875:0 2e+10 2e+10 1e+10 0 2
-116060:0 2 2 1 0 2
-195802:0 2e+20 2e+20 1e+20 0 2
+f:34875:0 2e+10 2e+10 1e+10 0 2
+Constant:116060:0 2 2 1 0 2
+f*f:195802:0 2e+20 2e+20 1e+20 0 2

test/train-sets/ref/help.stdout

@@ -53,8 +53,8 @@ Feature Options:
                                             only (type: list[str], keep)
     --ignore_features_dsjson_experimental args...
                                             Ignore specified features from namespace. To ignore a feature
-                                            arg should be namespace|feature To ignore a feature in the default
-                                            namespace, arg should be |feature (type: list[str], keep, experimental)
+                                            arg should be <namespace>|<feature>. <namespace> should be empty
+                                            for default (type: list[str], keep, experimental)
     --keep args...                          Keep namespaces beginning with character <arg> (type: list[str],
                                             keep)
     --redefine args...                      Redefine namespaces beginning with characters of std::string

test/train-sets/ref/help_cbadf.stdout

@@ -53,8 +53,8 @@ Feature Options:
                                             only (type: list[str], keep)
     --ignore_features_dsjson_experimental args...
                                             Ignore specified features from namespace. To ignore a feature
-                                            arg should be namespace|feature To ignore a feature in the default
-                                            namespace, arg should be |feature (type: list[str], keep, experimental)
+                                            arg should be <namespace>|<feature>. <namespace> should be empty
+                                            for default (type: list[str], keep, experimental)
     --keep args...                          Keep namespaces beginning with character <arg> (type: list[str],
                                             keep)
     --redefine args...                      Redefine namespaces beginning with characters of std::string

vowpalwabbit/core/include/vw/core/confidence_sequence_robust.h

@@ -45,16 +45,15 @@ class g_tilde
 class countable_discrete_base
 {
 public:
-  countable_discrete_base(double eta = 0.95f, double r = 2.0, double k = 1.5, double lambda_max = 0.5, double xi = 1.6);
+  countable_discrete_base(double eta = 0.95f, double k = 1.5, double lambda_max = 0.5, double xi = 1.6);
   double get_ci(double alpha) const;
   double get_lam_sqrt_tp1(double j) const;
   double get_v_impl(std::map<uint64_t, double>& memo, uint64_t j) const;
   double log_wealth_mix(double mu, double s, double thres, std::map<uint64_t, double>& memo) const;
   double root_brentq(double s, double thres, std::map<uint64_t, double>& memo, double min_mu, double max_mu,
-      double toll_x = 1e-10, double toll_f = 1e-12) const;
+      double toll_x = 1e-4, double toll_f = 1e-20) const;
   double log_sum_exp(const std::vector<double>& combined) const;
   double lb_log_wealth(double alpha) const;
-  double polylog(double r, double eta) const;
   double get_log_weight(double j) const;
   double get_log_remaining_weight(double j) const;
   double get_s() const;

vowpalwabbit/core/src/confidence_sequence_robust.cc

@@ -86,18 +86,17 @@ void g_tilde::reset_stats()
   t = 0;
 }
 
-countable_discrete_base::countable_discrete_base(double eta, double r, double k, double lambda_max, double xi)
+countable_discrete_base::countable_discrete_base(double eta, double k, double lambda_max, double xi)
     : log_xi(std::log1p(xi - 1))
     , log_xi_m1(std::log1p(xi - 2.0))
     , lambda_max(lambda_max)
-    , zeta_r(1.6449340668482264)  // std::riemann_zeta(r) -- Assuming r=2.0 is constant
-    , scale_fac(0.5 * (1.0 + polylog(r, eta) / (eta * zeta_r)))
+    , zeta_r(1.6449340668482264)                                    // std::riemann_zeta(r) -- Assuming r=2.0
+    , scale_fac(0.5 * (1.0 + 1.4406337969700393 / (eta * zeta_r)))  // polylog(r, eta) -- Assuming eta=0.95, r=2.0
     , log_scale_fac(std::log1p(scale_fac - 1.0))
     , t(0)
     , gt(k)
 {
   assert(0.0 < eta && eta < 1.0);
-  assert(r > 1.0);
   assert(0.0 < scale_fac && scale_fac < 1.0);
   assert(0.0 < lambda_max && lambda_max <= 1.0 + (-0.15859433956303937));  // sc.lambertw(-exp(-2)) in Python
   assert(1.0 < xi);
@@ -213,15 +212,7 @@ double countable_discrete_base::root_brentq(
     }
     else { mflag = false; }
 
-    size_t memo_size = memo.size();
     fs = f(s);
-    while (memo.size() > memo_size)
-    {
-      memo_size = memo.size();
-      fa = f(a);
-      fb = f(b);
-      fs = f(s);
-    }
     d = c;
     c = b;
     fc = fb;
@@ -244,7 +235,8 @@ double countable_discrete_base::root_brentq(
     }
   }
 
-  return s;
+  // Returning lower estimate of location of root
+  return std::min(a, b);
 }
 
 double countable_discrete_base::lb_log_wealth(double alpha) const
@@ -265,21 +257,6 @@ double countable_discrete_base::lb_log_wealth(double alpha) const
   return root_brentq(s, thres, memo, min_mu, max_mu);
 }
 
-double countable_discrete_base::polylog(double r, double eta) const
-{
-  double ret_val = 0.0;
-  double min_thres = 1e-10;
-  double curr_val = std::numeric_limits<int>::max();
-  uint64_t k = 0;
-  while (curr_val > min_thres)
-  {
-    k += 1;
-    curr_val = std::pow(eta, k) / std::pow(k, r);
-    ret_val += curr_val;
-  }
-  return ret_val;
-}
-
 double countable_discrete_base::get_log_weight(double j) const { return log_scale_fac + log_xi_m1 - (1 + j) * log_xi; }
 
 double countable_discrete_base::get_log_remaining_weight(double j) const { return log_scale_fac - j * log_xi; }

vowpalwabbit/core/src/parse_args.cc

@@ -560,8 +560,8 @@ void parse_feature_tweaks(options_i& options, VW::workspace& all, bool interacti
                .help("Ignore namespaces beginning with character <arg> for linear terms only"))
       .add(make_option("ignore_features_dsjson_experimental", ignore_features_dsjson)
                .keep()
-               .help("Ignore specified features from namespace. To ignore a feature arg should be namespace|feature "
-                     "To ignore a feature in the default namespace, arg should be |feature")
+               .help("Ignore specified features from namespace. To ignore a feature arg should be "
+                     "<namespace>|<feature>. <namespace> should be empty for default")
                .experimental())
       .add(make_option("keep", keeps).keep().help("Keep namespaces beginning with character <arg>"))
       .add(make_option("redefine", redefines)

vowpalwabbit/core/src/reductions/gd.cc

@@ -994,48 +994,39 @@ void save_load_online_state_weights(VW::workspace& all, VW::io_buf& model_file,
     for (typename T::iterator v = weights.begin(); v != weights.end(); ++v)
     {
       i = v.index() >> weights.stride_shift();
-
+      bool gd_write = *v != 0.f;
+      bool ftrl3_write = ftrl_size == 3 && (*v != 0.f || (&(*v))[1] != 0.f || (&(*v))[2] != 0.f);
+      bool ftrl4_write = ftrl_size == 4 && (*v != 0.f || (&(*v))[1] != 0.f || (&(*v))[2] != 0.f || (&(*v))[3] != 0.f);
+      bool ftrl6_write = ftrl_size == 6 &&
+          (*v != 0.f || (&(*v))[1] != 0.f || (&(*v))[2] != 0.f || (&(*v))[3] != 0.f || (&(*v))[4] != 0.f ||
+              (&(*v))[5] != 0.f);
       if (all.print_invert)  // write readable model with feature names
       {
-        if (*v != 0.f)
+        if (gd_write || ftrl3_write || ftrl4_write || ftrl6_write)
         {
           const auto map_it = all.index_name_map.find(i);
-          if (map_it != all.index_name_map.end())
-          {
-            msg << to_string(map_it->second) << ":";
-            VW::details::bin_text_write_fixed(model_file, nullptr /*unused*/, 0 /*unused*/, msg, true);
-          }
+          if (map_it != all.index_name_map.end()) { msg << to_string(map_it->second) << ":"; }
         }
       }
 
-      if (ftrl_size == 3)
+      if (ftrl3_write)
       {
-        if (*v != 0. || (&(*v))[1] != 0. || (&(*v))[2] != 0.)
-        {
-          brw = write_index(model_file, msg, text, all.num_bits, i);
-          msg << ":" << *v << " " << (&(*v))[1] << " " << (&(*v))[2] << "\n";
-          brw += VW::details::bin_text_write_fixed(model_file, (char*)&(*v), 3 * sizeof(*v), msg, text);
-        }
+        brw = write_index(model_file, msg, text, all.num_bits, i);
+        msg << ":" << *v << " " << (&(*v))[1] << " " << (&(*v))[2] << "\n";
+        brw += VW::details::bin_text_write_fixed(model_file, (char*)&(*v), 3 * sizeof(*v), msg, text);
       }
-      else if (ftrl_size == 4)
+      else if (ftrl4_write)
       {
-        if (*v != 0. || (&(*v))[1] != 0. || (&(*v))[2] != 0. || (&(*v))[3] != 0.)
-        {
-          brw = write_index(model_file, msg, text, all.num_bits, i);
-          msg << ":" << *v << " " << (&(*v))[1] << " " << (&(*v))[2] << " " << (&(*v))[3] << "\n";
-          brw += VW::details::bin_text_write_fixed(model_file, (char*)&(*v), 4 * sizeof(*v), msg, text);
-        }
+        brw = write_index(model_file, msg, text, all.num_bits, i);
+        msg << ":" << *v << " " << (&(*v))[1] << " " << (&(*v))[2] << " " << (&(*v))[3] << "\n";
+        brw += VW::details::bin_text_write_fixed(model_file, (char*)&(*v), 4 * sizeof(*v), msg, text);
       }
-      else if (ftrl_size == 6)
+      else if (ftrl6_write)
       {
-        if (*v != 0. || (&(*v))[1] != 0. || (&(*v))[2] != 0. || (&(*v))[3] != 0. || (&(*v))[4] != 0. ||
-            (&(*v))[5] != 0.)
-        {
-          brw = write_index(model_file, msg, text, all.num_bits, i);
-          msg << ":" << *v << " " << (&(*v))[1] << " " << (&(*v))[2] << " " << (&(*v))[3] << " " << (&(*v))[4] << " "
-              << (&(*v))[5] << "\n";
-          brw += VW::details::bin_text_write_fixed(model_file, (char*)&(*v), 6 * sizeof(*v), msg, text);
-        }
+        brw = write_index(model_file, msg, text, all.num_bits, i);
+        msg << ":" << *v << " " << (&(*v))[1] << " " << (&(*v))[2] << " " << (&(*v))[3] << " " << (&(*v))[4] << " "
+            << (&(*v))[5] << "\n";
+        brw += VW::details::bin_text_write_fixed(model_file, (char*)&(*v), 6 * sizeof(*v), msg, text);
       }
       else if (g == nullptr || (!all.weights.adaptive && !all.weights.normalized))
       {

vowpalwabbit/core/tests/automl_test.cc

@@ -328,11 +328,9 @@ TEST(Automl, NamespaceSwitchWIterations)
 
         auto champ_exclusions =
             aml->cm->_config_oracle.configs[aml->cm->estimators[aml->cm->current_champ].first.config_index].elements;
-        EXPECT_EQ(champ_exclusions.size(), 1);
-        std::vector<VW::namespace_index> ans{'U', 'U'};
-        EXPECT_NE(champ_exclusions.find(ans), champ_exclusions.end());
+        EXPECT_EQ(champ_exclusions.size(), 0);
         auto champ_interactions = aml->cm->estimators[aml->cm->current_champ].first.live_interactions;
-        EXPECT_EQ(champ_interactions.size(), 5);
+        EXPECT_EQ(champ_interactions.size(), 6);
         return true;
       });
 
@@ -361,9 +359,9 @@ TEST(Automl, ClearConfigsWIterations)
         EXPECT_EQ(aml->cm->current_champ, 0);
         EXPECT_EQ(aml->cm->_config_oracle.valid_config_size, 4);
         EXPECT_EQ(clear_champ_switch - 1, aml->cm->total_learn_count);
-        EXPECT_EQ(aml->cm->estimators[0].first.live_interactions.size(), 2);
-        EXPECT_EQ(aml->cm->estimators[1].first.live_interactions.size(), 3);
-        EXPECT_EQ(aml->cm->estimators[2].first.live_interactions.size(), 1);
+        EXPECT_EQ(aml->cm->estimators[0].first.live_interactions.size(), 3);
+        EXPECT_EQ(aml->cm->estimators[1].first.live_interactions.size(), 2);
+        EXPECT_EQ(aml->cm->estimators[2].first.live_interactions.size(), 2);
         EXPECT_EQ(aml->current_state, VW::reductions::automl::automl_state::Experimenting);
         return true;
       });
@@ -434,9 +432,9 @@ TEST(Automl, ClearConfigsOneDiffWIterations)
       {
         aml_test::aml_onediff* aml = aml_test::get_automl_data<VW::reductions::automl::one_diff_impl>(all);
         EXPECT_EQ(aml->cm->estimators.size(), 3);
-        EXPECT_EQ(aml->cm->estimators[0].first.live_interactions.size(), 2);
-        EXPECT_EQ(aml->cm->estimators[1].first.live_interactions.size(), 3);
-        EXPECT_EQ(aml->cm->estimators[2].first.live_interactions.size(), 1);
+        EXPECT_EQ(aml->cm->estimators[0].first.live_interactions.size(), 3);
+        EXPECT_EQ(aml->cm->estimators[1].first.live_interactions.size(), 2);
+        EXPECT_EQ(aml->cm->estimators[2].first.live_interactions.size(), 2);
         return true;
       });
 

vowpalwabbit/core/tests/confidence_sequence_robust_test.cc

@@ -96,13 +96,14 @@ TEST(ConfidenceSequenceRobust, PythonEquivalenceCI)
   csr.update(0.879389936, 1);
   csr.update(0.880083546, 1);
   csr.update(0.8807676, 1);
+  csr.update(43.91880613, 0.9);
 
   double lb = csr.lower_bound();
   double ub = csr.upper_bound();
 
   // Compare to test_confidence_sequence_robust.py
-  EXPECT_NEAR(lb, 0.19224909776696902, 1e-6);
-  EXPECT_NEAR(ub, 0.93637288977188682, 1e-6);
+  EXPECT_NEAR(lb, 0.30209151281846858, 1e-6);
+  EXPECT_NEAR(ub, 0.90219143188334106, 1e-6);
 
   // Test brentq separately
   double s = 139.8326745448;
@@ -118,9 +119,9 @@ TEST(ConfidenceSequenceRobust, PythonEquivalenceCI)
   double root = csr.lower.root_brentq(s, thres, memo, min_mu, max_mu);
 
   // Compare root to test_confidence_sequence_robust.py
-  EXPECT_NEAR(root, 0.31672263211621371, 1e-6);
+  EXPECT_NEAR(root, 0.30209143008131789, 1e-6);
 
   // Test that root of objective function is 0
   auto test_root = csr.lower.log_wealth_mix(root, s, thres, memo) - thres;
-  EXPECT_NEAR(test_root, 0.0, 1e-6);
+  EXPECT_NEAR(test_root, 0.0, 1e-2);
 }

vowpalwabbit/core/tests/epsilon_decay_test.cc

@@ -60,29 +60,28 @@ TEST(EpsilonDecay, InitWIterations)
       "--epsilon_decay", "--model_count=3", "--cb_explore_adf", "--quiet", "--epsilon=0.2", "--random_seed=5"});
 }
 
-#if !defined(__APPLE__) && !defined(_WIN32)
 TEST(EpsilonDecay, ChampChangeWIterations)
 {
   const size_t num_iterations = 610;
-  const size_t seed = 4;
+  const size_t seed = 36;
   const std::vector<uint64_t> swap_after = {500};
-  const size_t deterministic_champ_switch = 601;
+  const size_t deterministic_champ_switch = 600;
   callback_map test_hooks;
 
   test_hooks.emplace(deterministic_champ_switch - 1,
       [&](cb_sim&, VW::workspace& all, VW::multi_ex&)
       {
         epsilon_decay_data* epsilon_decay = epsilon_decay_test::get_epsilon_decay_data(all);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[0][0].update_count, 88);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[1][0].update_count, 88);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][0].update_count, 88);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][0].update_count, 88);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[1][1].update_count, 93);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][1].update_count, 93);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][1].update_count, 93);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][2].update_count, 94);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][2].update_count, 94);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][3].update_count, 600);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[0][0].update_count, 87);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[1][0].update_count, 87);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][0].update_count, 87);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][0].update_count, 87);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[1][1].update_count, 92);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][1].update_count, 92);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][1].update_count, 92);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][2].update_count, 93);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][2].update_count, 93);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][3].update_count, 599);
         return true;
       });
 
@@ -94,12 +93,12 @@ TEST(EpsilonDecay, ChampChangeWIterations)
         EXPECT_EQ(epsilon_decay->conf_seq_estimators[1][0].update_count, 0);
         EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][0].update_count, 0);
         EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][0].update_count, 0);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[1][1].update_count, 0);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][1].update_count, 0);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][1].update_count, 0);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][2].update_count, 0);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][2].update_count, 0);
-        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][3].update_count, 89);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[1][1].update_count, 88);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][1].update_count, 88);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][1].update_count, 88);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[2][2].update_count, 93);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][2].update_count, 93);
+        EXPECT_EQ(epsilon_decay->conf_seq_estimators[3][3].update_count, 94);
         return true;
       });
 
@@ -110,7 +109,6 @@ TEST(EpsilonDecay, ChampChangeWIterations)
 
   EXPECT_GT(ctr.back(), 0.4f);
 }
-#endif
 
 TEST(EpsilonDecay, UpdateCountWIterations)
 {