Merge branch 'master' into master

.vscode/launch.json

@@ -0,0 +1,16 @@
+{
+    // Use IntelliSense to learn about possible attributes.
+    // Hover to view descriptions of existing attributes.
+    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "name": "(ctest) Launch",
+            "type": "cppdbg",
+            "cwd": "${workspaceFolder}",
+            "request": "launch",
+            "program": "${cmake.testProgram}",
+            "args": [ "${cmake.testArgs}" ]
+        }
+    ]
+}

ext_libs/ext_libs.cmake

@@ -38,7 +38,7 @@ if(RAPIDJSON_SYS_DEP)
   # Since EXACT is not specified, any version compatible with 1.1.0 is accepted (>= 1.1.0)
   find_package(RapidJSON 1.1.0 CONFIG REQUIRED)
   add_library(RapidJSON INTERFACE)
-  target_include_directories(RapidJSON INTERFACE ${RapidJSON_INCLUDE_DIRS})
+  target_include_directories(RapidJSON INTERFACE ${RapidJSON_INCLUDE_DIRS} ${RAPIDJSON_INCLUDE_DIRS})
 else()
   add_library(RapidJSON INTERFACE)
   target_include_directories(RapidJSON SYSTEM INTERFACE "${CMAKE_CURRENT_LIST_DIR}/rapidjson/include")
@@ -127,4 +127,4 @@ if(VW_FEAT_CB_GRAPH_FEEDBACK)
   target_include_directories(mlpack_ensmallen SYSTEM INTERFACE ${CMAKE_CURRENT_LIST_DIR}/armadillo-code/include)
 
   target_include_directories(mlpack_ensmallen SYSTEM INTERFACE ${CMAKE_CURRENT_LIST_DIR}/ensmallen/include)
-endif()
+endif()

python/docs/source/tutorials/DFtoVW_tutorial.ipynb

@@ -802,15 +802,17 @@
     "\n",
     "# Adding columns for easier visualization\n",
     "weights_df[\"feature_name\"] = weights_df.apply(\n",
-    "    lambda row: row.vw_feature_name.split(\"=\")[0]\n",
-    "    if row.is_cat\n",
-    "    else row.vw_feature_name,\n",
+    "    lambda row: (\n",
+    "        row.vw_feature_name.split(\"=\")[0] if row.is_cat else row.vw_feature_name\n",
+    "    ),\n",
     "    axis=1,\n",
     ")\n",
     "weights_df[\"feature_value\"] = weights_df.apply(\n",
-    "    lambda row: row.vw_feature_name.split(\"=\")[1].zfill(2)\n",
-    "    if row.is_cat\n",
-    "    else row.vw_feature_name,\n",
+    "    lambda row: (\n",
+    "        row.vw_feature_name.split(\"=\")[1].zfill(2)\n",
+    "        if row.is_cat\n",
+    "        else row.vw_feature_name\n",
+    "    ),\n",
     "    axis=1,\n",
     ")\n",
     "weights_df.sort_values([\"feature_name\", \"feature_value\"], inplace=True)"

python/docs/source/tutorials/cmd_first_steps.md

@@ -116,6 +116,6 @@ The model predicted a value of **0**. This result means our house will not need
 ## More to explore
 
 - See [Python tutorial](python_first_steps.ipynb) for a quick introduction to the basics of training and testing your model.
-- To learn more about how to approach a contextual bandits problem using  tVowpal Wabbit — including how to  work with different contextual bandits approaches, how to format data, and understand the results — see the [Contextual Bandit Reinforcement Learning Tutorial](python_Contextual_bandits_and_Vowpal_Wabbit.ipynb).
+- To learn more about how to approach a contextual bandits problem using Vowpal Wabbit — including how to work with different contextual bandits approaches, how to format data, and understand the results — see the [Contextual Bandit Reinforcement Learning Tutorial](python_Contextual_bandits_and_Vowpal_Wabbit.ipynb).
 - For more on the contextual bandits approach to reinforcement learning, including a content personalization scenario, see the [Contextual Bandit Simulation Tutorial](python_Simulating_a_news_personalization_scenario_using_Contextual_Bandits.ipynb).
 - See the [Linear Regression Tutorial](cmd_linear_regression.md) for a different look at the roof replacement problem and learn more about Vowpal Wabbit's format and understanding the results.

python/tests/confidence_sequence.py

@@ -189,6 +189,5 @@ def lblogwealth(self, *, t, sumXt, v, eta, s, alpha):
 
         return max(
             0,
-            (sumXt - sqrt(gamma1**2 * ll * v + gamma2**2 * ll**2) - gamma2 * ll)
-            / t,
+            (sumXt - sqrt(gamma1**2 * ll * v + gamma2**2 * ll**2) - gamma2 * ll) / t,
         )

python/tests/crminustwo.py

@@ -440,21 +440,23 @@ def intervaldiff(
                     candidates.append(
                         (
                             gstar,
-                            None
-                            if isclose(kappa, 0)
-                            else {
-                                "kappastar": kappa,
-                                "betastar": beta,
-                                "gammastar": gamma,
-                                "taustar": tau,
-                                "ufake": ufake,
-                                "wfake": wfake,
-                                "rfake": rex,
-                                "qfunc": lambda c, u, w, r, k=kappa, g=gamma, b=beta, t=tau, s=sign, num=n: -c
-                                * (b + g * u + t * w + s * (u - w) * r)
-                                / ((num + 1) * k),
-                                "mle": mle,
-                            },
+                            (
+                                None
+                                if isclose(kappa, 0)
+                                else {
+                                    "kappastar": kappa,
+                                    "betastar": beta,
+                                    "gammastar": gamma,
+                                    "taustar": tau,
+                                    "ufake": ufake,
+                                    "wfake": wfake,
+                                    "rfake": rex,
+                                    "qfunc": lambda c, u, w, r, k=kappa, g=gamma, b=beta, t=tau, s=sign, num=n: -c
+                                    * (b + g * u + t * w + s * (u - w) * r)
+                                    / ((num + 1) * k),
+                                    "mle": mle,
+                                }
+                            ),
                         )
                     )
 

python/vowpalwabbit/pyvw.py

@@ -532,9 +532,9 @@ def parse(
                     for ex in str_ex
                 ]
             ):
-                str_ex: List[
-                    Example
-                ] = str_ex  # pytype: disable=annotation-type-mismatch
+                str_ex: List[Example] = (
+                    str_ex  # pytype: disable=annotation-type-mismatch
+                )
                 return str_ex
 
         if not isinstance(str_ex, (list, str)):

test/core.vwtest.json

@@ -6073,5 +6073,34 @@
     "depends_on": [
       467
     ]
+  },
+  {
+    "id": 469,
+    "desc": "https://github.com/VowpalWabbit/vowpal_wabbit/issues/4669",
+    "vw_command": "--ccb_explore_adf --dsjson -d train-sets/issue4669.dsjson -f issue4669.model",
+    "diff_files": {
+      "stderr": "train-sets/ref/issue4669_train.stderr",
+      "stdout": "train-sets/ref/issue4669_train.stdout"
+    },
+    "input_files": [
+      "train-sets/issue4669.dsjson"
+    ]
+  },
+  {
+    "id": 470,
+    "desc": "https://github.com/VowpalWabbit/vowpal_wabbit/issues/4669",
+    "vw_command": "--ccb_explore_adf --dsjson --all_slots_loss --epsilon 0 -t -i issue4669.model -t -d train-sets/issue4669.dsjson -p issue4669_test_pred.txt",
+    "diff_files": {
+      "stderr": "train-sets/ref/issue4669_test.stderr",
+      "stdout": "train-sets/ref/issue4669_test.stdout",
+      "issue4669_test_pred.txt": "train-sets/ref/issue4669_test_pred.txt"
+    },
+    "input_files": [
+      "train-sets/issue4669.dsjson",
+      "issue4669.model"
+    ],
+    "depends_on": [
+      469
+    ]
   }
 ]

test/run_tests.py

@@ -68,17 +68,21 @@ def _are_same(expected: Any, actual: Any, key: str) -> Tuple[bool, str]:
         elif isinstance(expected, (int, bool, str)):
             return (
                 expected == actual,
-                f"Key '{key}' value mismatch. Expected: '{expected}', but found: '{actual}'"
-                if expected != actual
-                else "",
+                (
+                    f"Key '{key}' value mismatch. Expected: '{expected}', but found: '{actual}'"
+                    if expected != actual
+                    else ""
+                ),
             )
         elif isinstance(expected, (float)):
             delta = abs(expected - actual)
             return (
                 delta < epsilon,
-                f"Key '{key}' value mismatch. Expected: '{expected}', but found: '{actual}' (using epsilon: '{epsilon}')"
-                if delta >= epsilon
-                else "",
+                (
+                    f"Key '{key}' value mismatch. Expected: '{expected}', but found: '{actual}' (using epsilon: '{epsilon}')"
+                    if delta >= epsilon
+                    else ""
+                ),
             )
         elif isinstance(expected, dict):
             expected_keys = set(expected.keys())

test/save_resume_test.py

@@ -1,6 +1,7 @@
 """
 Test that the models generated with and without --predict_only_model produce the same predictions when loaded in test_mode.
 """
+
 import sys
 import os
 import optparse

test/train-sets/issue4669.dsjson

@@ -0,0 +1 @@
+{"c":{"_multi":[{"f":"1"},{"f":"2"}],"_slots":[{"_inc":[0,1]},{"_inc":[1]}]},"_outcomes":[{"_label_cost":1.0,"_a":[0,1],"_p":[0.5,0.5]},{"_label_cost":0.0,"_a":[1],"_p":[1]}]}

test/train-sets/ref/issue4669_test.stderr

@@ -0,0 +1,23 @@
+only testing
+predictions = issue4669_test_pred.txt
+using no cache
+Reading datafile = train-sets/issue4669.dsjson
+num sources = 1
+Num weight bits = 18
+learning rate = 0.5
+initial_t = 1
+power_t = 0.5
+cb_type = mtr
+Enabled learners: gd, generate_interactions, scorer-identity, csoaa_ldf-rank, cb_adf, cb_explore_adf_greedy, cb_sample, shared_feature_merger, ccb_explore_adf
+Input label = CCB
+Output pred = DECISION_PROBS
+average  since         example        example        current        current  current
+loss     last          counter         weight          label        predict features
+0.000000 0.000000            1            1.0        0:1,1:0         1,None        9
+
+finished run
+number of examples = 1
+weighted example sum = 1.000000
+weighted label sum = 0.000000
+average loss = 0.000000
+total feature number = 9

test/train-sets/ref/issue4669_test_pred.txt

@@ -0,0 +1,3 @@
+1:1,0:0
+
+

test/train-sets/ref/issue4669_train.stderr

@@ -0,0 +1,22 @@
+final_regressor = issue4669.model
+using no cache
+Reading datafile = train-sets/issue4669.dsjson
+num sources = 1
+Num weight bits = 18
+learning rate = 0.5
+initial_t = 0
+power_t = 0.5
+cb_type = mtr
+Enabled learners: gd, generate_interactions, scorer-identity, csoaa_ldf-rank, cb_adf, cb_explore_adf_greedy, cb_sample, shared_feature_merger, ccb_explore_adf
+Input label = CCB
+Output pred = DECISION_PROBS
+average  since         example        example        current        current  current
+loss     last          counter         weight          label        predict features
+1.000000 1.000000            1            1.0        0:1,1:0            0,1       12
+
+finished run
+number of examples = 1
+weighted example sum = 1.000000
+weighted label sum = 0.000000
+average loss = 1.000000
+total feature number = 12

utl/flatbuffer/vw_to_flat.cc

@@ -299,10 +299,10 @@ void to_flat::create_no_label(VW::example* v, ExampleBuilder& ex_builder)
   ex_builder.label = VW::parsers::flatbuffer::Createno_label(_builder, (uint8_t)'\000').Union();
 }
 
-flatbuffers::Offset<VW::parsers::flatbuffer::Namespace> to_flat::create_namespace(VW::features::audit_iterator begin,
-    VW::features::audit_iterator end, VW::namespace_index index, uint64_t hash, bool audit)
+// Create namespace when audit is true
+flatbuffers::Offset<VW::parsers::flatbuffer::Namespace> to_flat::create_namespace_audit(
+    VW::features::audit_iterator begin, VW::features::audit_iterator end, VW::namespace_index index, uint64_t hash)
 {
-  std::vector<flatbuffers::Offset<VW::parsers::flatbuffer::Feature>> fts;
   std::stringstream ss;
   ss << index;
 
@@ -316,26 +316,61 @@ flatbuffers::Offset<VW::parsers::flatbuffer::Namespace> to_flat::create_namespac
   if (find_ns_offset == _share_examples.end())
   {
     flatbuffers::Offset<VW::parsers::flatbuffer::Namespace> namespace_offset;
+    std::vector<flatbuffers::Offset<flatbuffers::String>> feature_names;
+    std::vector<float> feature_values;
+    std::vector<uint64_t> feature_hashes;
+
     // new namespace
-    if (audit)
+
+    std::string ns_name;
+    for (auto it = begin; it != end; ++it)
     {
-      std::string ns_name;
-      for (auto it = begin; it != end; ++it)
-      {
-        ns_name = it.audit()->ns;
-        fts.push_back(
-            VW::parsers::flatbuffer::CreateFeatureDirect(_builder, it.audit()->name.c_str(), it.value(), it.index()));
-      }
-      namespace_offset = VW::parsers::flatbuffer::CreateNamespaceDirect(_builder, ns_name.c_str(), index, &fts, hash);
+      if ((it.audit()->ns).c_str() != nullptr) ns_name = it.audit()->ns;
+
+      (feature_names).push_back(_builder.CreateString(it.audit()->name.c_str()));
+      (feature_values).push_back(it.value());
+      (feature_hashes).push_back(it.index());
     }
-    else
+    namespace_offset = VW::parsers::flatbuffer::CreateNamespaceDirect(
+        _builder, ns_name.c_str(), index, hash, &feature_names, &feature_values, &feature_hashes);
+
+    _share_examples[refid] = namespace_offset;
+  }
+
+  return _share_examples[refid];
+}
+
+// Create namespace when audit is false
+flatbuffers::Offset<VW::parsers::flatbuffer::Namespace> to_flat::create_namespace(
+    features::const_iterator begin, features::const_iterator end, VW::namespace_index index, uint64_t hash)
+{
+  std::stringstream ss;
+  ss << index;
+
+  for (auto it = begin; it != end; ++it) { ss << it.index() << it.value(); }
+  ss << ":" << hash;
+
+  std::string s = ss.str();
+  uint64_t refid = VW::uniform_hash(s.c_str(), s.size(), 0);
+  const auto find_ns_offset = _share_examples.find(refid);
+
+  if (find_ns_offset == _share_examples.end())
+  {
+    flatbuffers::Offset<VW::parsers::flatbuffer::Namespace> namespace_offset;
+    std::vector<float> feature_values;
+    std::vector<uint64_t> feature_hashes;
+
+    for (auto it = begin; it != end; ++it)
     {
-      for (auto it = begin; it != end; ++it)
+      if (it.value() != 0)  // store the feature data only if the value is non zero
       {
-        fts.push_back(VW::parsers::flatbuffer::CreateFeatureDirect(_builder, nullptr, it.value(), it.index()));
+        (feature_values).push_back(it.value());
+        (feature_hashes).push_back(it.index());
       }
-      namespace_offset = VW::parsers::flatbuffer::CreateNamespaceDirect(_builder, nullptr, index, &fts, hash);
     }
+    namespace_offset = VW::parsers::flatbuffer::CreateNamespaceDirect(
+        _builder, nullptr, index, hash, nullptr, &feature_values, &feature_hashes);
+
     _share_examples[refid] = namespace_offset;
   }
 
@@ -438,13 +473,25 @@ void to_flat::convert_txt_to_flat(VW::workspace& all)
           VW::details::flatten_namespace_extents(ae->feature_space[ns].namespace_extents, ae->feature_space[ns].size());
       auto unflattened_with_ranges_that_dont_have_extents = unflatten_namespace_extents_dont_skip(flattened_extents);
 
-      for (const auto& extent : unflattened_with_ranges_that_dont_have_extents)
+      if (all.output_config.audit || all.output_config.hash_inv)
+      {
+        for (const auto& extent : unflattened_with_ranges_that_dont_have_extents)
+        {
+          // The extent hash for a non-hash-extent will be 0, which is the same as the field no existing to flatbuffers.
+          auto created_ns = create_namespace_audit(ae->feature_space[ns].audit_begin() + extent.begin_index,
+              ae->feature_space[ns].audit_begin() + extent.end_index, ns, extent.hash);
+          namespaces.push_back(created_ns);
+        }
+      }
+      else
       {
-        // The extent hash for a non-hash-extent will be 0, which is the same as the field no existing to flatbuffers.
-        auto created_ns = create_namespace(ae->feature_space[ns].audit_begin() + extent.begin_index,
-            ae->feature_space[ns].audit_begin() + extent.end_index, ns, extent.hash,
-            all.output_config.audit || all.output_config.hash_inv);
-        namespaces.push_back(created_ns);
+        for (const auto& extent : unflattened_with_ranges_that_dont_have_extents)
+        {
+          // The extent hash for a non-hash-extent will be 0, which is the same as the field no existing to flatbuffers.
+          auto created_ns = create_namespace(ae->feature_space[ns].cbegin() + extent.begin_index,
+              ae->feature_space[ns].cbegin() + extent.end_index, ns, extent.hash);
+          namespaces.push_back(created_ns);
+        }
       }
     }
     std::string tag(ae->tag.begin(), ae->tag.size());

utl/flatbuffer/vw_to_flat.h

@@ -85,6 +85,8 @@ class to_flat
   void write_to_file(bool collection, bool is_multiline, MultiExampleBuilder& multi_ex_builder,
       ExampleBuilder& ex_builder, std::ofstream& outfile);
 
-  flatbuffers::Offset<VW::parsers::flatbuffer::Namespace> create_namespace(VW::features::audit_iterator begin,
-      VW::features::audit_iterator end, VW::namespace_index index, uint64_t hash, bool audit);
+  flatbuffers::Offset<VW::parsers::flatbuffer::Namespace> create_namespace(
+      VW::features::const_iterator begin, VW::features::const_iterator end, VW::namespace_index index, uint64_t hash);
+  flatbuffers::Offset<VW::parsers::flatbuffer::Namespace> create_namespace_audit(
+      VW::features::audit_iterator begin, VW::features::audit_iterator end, VW::namespace_index index, uint64_t hash);
 };