chore: [LAS] remove compile time flag and its own custom CI

CMakeLists.txt

@@ -129,9 +129,9 @@ option(SPDLOG_SYS_DEP "Override using the submodule for spdlog dependency. Inste
 option(VW_BOOST_MATH_SYS_DEP "Override using the submodule for boost math dependency. Instead will use find_package" OFF)
 option(VW_ZLIB_SYS_DEP "Override using the submodule for zlib dependency. Instead will use find_package" ON)
 option(VW_GTEST_SYS_DEP "Override using fetch package for gtest dependency. Instead will use find_package" OFF)
+option(VW_EIGEN_SYS_DEP "Override using the submodule for Eigen3 dependency. Instead will use find_package" OFF)
 option(VW_BUILD_VW_C_WRAPPER "Enable building the c_wrapper project" ON)
 option(VW_BUILD_CSV "Build csv parser" OFF)
-option(VW_BUILD_LARGE_ACTION_SPACE "Enable large action space reduction" OFF)
 option(vw_BUILD_NET_CORE "Build .NET Core targets" OFF)
 option(vw_BUILD_NET_FRAMEWORK "Build .NET Framework targets" OFF)
 option(VW_USE_ASAN "Compile with AddressSanitizer" OFF)

CMakePresets.json

@@ -42,6 +42,10 @@
         "VW_GTEST_SYS_DEP": {
           "type": "BOOL",
           "value": "ON"
+        },
+        "VW_EIGEN_SYS_DEP": {
+          "type": "BOOL",
+          "value": "ON"
         }
       }
     },

ext_libs/ext_libs.cmake

@@ -91,10 +91,12 @@ if(BUILD_FLATBUFFERS)
   include(FlatbufferUtils)
 endif()
 
-if(VW_BUILD_LARGE_ACTION_SPACE)
+if(VW_EIGEN_SYS_DEP)
+  # Since EXACT is not specified, any version compatible with 3.4.0 is accepted (>= 3.4.0)
+  find_package(Eigen3 3.4.0 CONFIG REQUIRED)
   add_library(eigen INTERFACE)
-  target_include_directories(eigen SYSTEM INTERFACE
-    $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}/eigen>
-    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
-  )
+  target_include_directories(eigen INTERFACE ${EIGEN3_INCLUDE_DIR})
+else()
+  add_library(eigen INTERFACE)
+  target_include_directories(eigen SYSTEM INTERFACE $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}/eigen>)
 endif()

test/benchmarks/standalone/benchmark_text_input.cc

@@ -278,7 +278,6 @@ BENCHMARK_CAPTURE(benchmark_multi, ccb_adf_same_char_interactions, gen_ccb_examp
     "--ccb_explore_adf --quiet -q ::")
     ->MinTime(15.0);
 
-#ifdef BUILD_LARGE_ACTION_SPACE
 BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_small_300_onestep,
     gen_cb_examples(1, 50, 10, 300, 5, 5, 20, 10, false),
     "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet")
@@ -320,151 +319,3 @@ BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_small_500_plaincb,
     ->MinTime(15.0)
     ->UseRealTime()
     ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_small_1k_onestep,
-    gen_cb_examples(1, 50, 10, 1000, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_small_1k_onestep_max_threads,
-    gen_cb_examples(1, 50, 10, 1000, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet --thread_pool_size " +
-        std::to_string(std::thread::hardware_concurrency()))
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_small_1k_plaincb,
-    gen_cb_examples(1, 50, 10, 1000, 5, 5, 20, 10, false), "--cb_explore_adf -q :: --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_medium_300_onestep,
-    gen_cb_examples(1, 50, 20, 300, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_medium_300_onestep_max_threads,
-    gen_cb_examples(1, 50, 20, 300, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet --thread_pool_size " +
-        std::to_string(std::thread::hardware_concurrency()))
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_medium_300_plaincb,
-    gen_cb_examples(1, 50, 20, 300, 5, 5, 20, 10, false), "--cb_explore_adf -q :: --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_medium_500_onestep,
-    gen_cb_examples(1, 50, 20, 500, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_medium_500_onestep_max_threads,
-    gen_cb_examples(1, 50, 20, 500, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet --thread_pool_size " +
-        std::to_string(std::thread::hardware_concurrency()))
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_medium_500_plaincb,
-    gen_cb_examples(1, 50, 20, 500, 5, 5, 20, 10, false), "--cb_explore_adf -q :: --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_medium_1k_onestep,
-    gen_cb_examples(1, 50, 20, 1000, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_medium_1k_onestep_max_threads,
-    gen_cb_examples(1, 50, 20, 1000, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet --thread_pool_size " +
-        std::to_string(std::thread::hardware_concurrency()))
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_medium_1k_plaincb,
-    gen_cb_examples(1, 50, 20, 1000, 5, 5, 20, 10, false), "--cb_explore_adf -q :: --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_large_300_onestep,
-    gen_cb_examples(1, 50, 50, 300, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_large_300_onestep_max_threads,
-    gen_cb_examples(1, 50, 50, 300, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet --thread_pool_size " +
-        std::to_string(std::thread::hardware_concurrency()))
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_large_300_plaincb,
-    gen_cb_examples(1, 50, 50, 300, 5, 5, 20, 10, false), "--cb_explore_adf -q :: --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_large_500_onestep,
-    gen_cb_examples(1, 50, 50, 500, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_large_500_onestep_max_threads,
-    gen_cb_examples(1, 50, 50, 500, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet --thread_pool_size " +
-        std::to_string(std::thread::hardware_concurrency()))
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_large_500_plaincb,
-    gen_cb_examples(1, 50, 50, 500, 5, 5, 20, 10, false), "--cb_explore_adf -q :: --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_large_1k_onestep,
-    gen_cb_examples(1, 50, 50, 1000, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_large_1k_onestep_max_threads,
-    gen_cb_examples(1, 50, 50, 1000, 5, 5, 20, 10, false),
-    "--cb_explore_adf --large_action_space -q :: --max_actions 20 --quiet --thread_pool_size " +
-        std::to_string(std::thread::hardware_concurrency()))
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-
-BENCHMARK_CAPTURE(benchmark_multi_predict, cb_las_large_1k_plaincb,
-    gen_cb_examples(1, 50, 50, 1000, 5, 5, 20, 10, false), "--cb_explore_adf -q :: --quiet")
-    ->MinTime(15.0)
-    ->UseRealTime()
-    ->Unit(benchmark::kMillisecond);
-#endif

test/train-sets/ref/help.stdout

@@ -599,6 +599,27 @@ Weight Options:
                                             uint, keep, necessary)
     --replay_m_count arg                    How many times (in expectation) should each example be played
                                             (default: 1 = permuting) (type: uint, default: 1)
+[Reduction] Experimental: Contextual Bandit Exploration with ADF with large action space filtering Options:
+    --cb_explore_adf                        Online explore-exploit for a contextual bandit problem with multiline
+                                            action dependent features (type: bool, keep, necessary)
+    --large_action_space                    Large action space filtering (type: bool, keep, necessary, experimental)
+    --max_actions arg                       Max number of actions to hold (type: uint, default: 20, keep,
+                                            experimental)
+    --spanner_c arg                         Parameter for computing c-approximate spanner (type: float, default:
+                                            2, keep, experimental)
+    --full_predictions                      Full representation of the prediction's action probabilities,
+                                            if not specified filtered out actions will not appear in the
+                                            final predictions. If specified, filtered out actions will appear
+                                            with a probability of zero (type: bool, experimental)
+    --thread_pool_size arg                  Number of threads in the thread pool that will be used when running
+                                            with one pass svd implementation (default svd implementation
+                                            option). Default thread pool size will be half of the available
+                                            hardware threads (type: uint)
+    --block_size arg                        Number of actions in a block to be scheduled for multithreading
+                                            when using one pass svd implementation (by default, block_size
+                                            = num_actions / thread_pool_size) (type: uint, default: 0)
+    --two_pass_svd                          A more accurate svd that is much slower than the default (one
+                                            pass svd) (type: bool, experimental)
 [Reduction] Explore Evaluation Options:
     --explore_eval                          Evaluate explore_eval adf policies (type: bool, keep, necessary)
     --multiplier arg                        Multiplier used to make all rejection sample probabilities <=

test/unit_test/CMakeLists.txt

@@ -11,6 +11,9 @@ add_executable(vw-unit-test.out
   cats_tree_test.cc
   cats_user_provided_pdf.cc
   cb_explore_adf_test.cc
+  cb_large_actions_test.cc
+  cb_las_one_pass_svd_test.cc
+  cb_las_spanner_test.cc
   ccb_parser_test.cc
   ccb_test.cc
   chain_hashing.cc
@@ -59,10 +62,6 @@ add_executable(vw-unit-test.out
   weights_test.cc
 )
 
-if(VW_BUILD_LARGE_ACTION_SPACE)
-  target_sources(vw-unit-test.out PRIVATE cb_large_actions_test.cc cb_las_one_pass_svd_test.cc cb_las_spanner_test.cc)
-endif()
-
 # Add the include directories from vw target for testing
 target_link_libraries(vw-unit-test.out PRIVATE vw_core Boost::unit_test_framework)
 target_include_directories(vw-unit-test.out PRIVATE $<TARGET_PROPERTY:vw_core,INCLUDE_DIRECTORIES>)

test/unit_test/cb_las_spanner_test.cc

@@ -464,7 +464,7 @@ static std::vector<std::string> gen_cb_examples(
 
     action_ss << "| ";
     for (int action_feat = 0; action_feat < coordinates; ++action_feat)
-    { action_ss << "x" << action_feat << ":" << (drand48() * scale) << " "; }
+    { action_ss << "x" << action_feat << ":" << ((static_cast<double>(std::rand()) / RAND_MAX) * scale) << " "; }
 
     examples.push_back(action_ss.str());
   }

vcpkg.json

@@ -4,6 +4,7 @@
   "version": "9.5.0",
   "dependencies": [
     "boost-math",
+    "eigen3",
     "flatbuffers",
     "fmt",
     "rapidjson",

vowpalwabbit/core/CMakeLists.txt

@@ -75,6 +75,7 @@ set(vw_core_headers
   include/vw/core/prediction_type.h
   include/vw/core/print_utils.h
   include/vw/core/prob_dist_cont.h
+   include/vw/core/qr_decomposition.h
   include/vw/core/queue.h
   include/vw/core/rand_state.h
   include/vw/core/rand48.h
@@ -94,6 +95,7 @@ set(vw_core_headers
   include/vw/core/reductions/cats_pdf.h
   include/vw/core/reductions/cats_tree.h
   include/vw/core/reductions/cats.h
+  include/vw/core/reductions/cb/cb_actions_mask.h
   include/vw/core/reductions/cb/cb_adf.h
   include/vw/core/reductions/cb/cb_algs.h
   include/vw/core/reductions/cb/cb_dro.h
@@ -102,6 +104,7 @@ set(vw_core_headers
   include/vw/core/reductions/cb/cb_explore_adf_cover.h
   include/vw/core/reductions/cb/cb_explore_adf_first.h
   include/vw/core/reductions/cb/cb_explore_adf_greedy.h
+  include/vw/core/reductions/cb/cb_explore_adf_large_action_space.h
   include/vw/core/reductions/cb/cb_explore_adf_regcb.h
   include/vw/core/reductions/cb/cb_explore_adf_rnd.h
   include/vw/core/reductions/cb/cb_explore_adf_softmax.h
@@ -239,6 +242,7 @@ set(vw_core_sources
   src/prediction_type.cc
   src/print_utils.cc
   src/prob_dist_cont.cc
+  src/qr_decomposition.cc
   src/rand48.cc
   src/reduction_stack.cc
   src/reductions/active_cover.cc
@@ -255,13 +259,18 @@ set(vw_core_sources
   src/reductions/cats_pdf.cc
   src/reductions/cats_tree.cc
   src/reductions/cats.cc
+  src/reductions/cb/details/large_action/one_pass_svd_impl.cc
+  src/reductions/cb/details/large_action/one_rank_spanner_impl.cc
+  src/reductions/cb/details/large_action/two_pass_svd_impl.cc
+  src/reductions/cb/cb_actions_mask.cc
   src/reductions/cb/cb_adf.cc
   src/reductions/cb/cb_algs.cc
   src/reductions/cb/cb_dro.cc
   src/reductions/cb/cb_explore_adf_bag.cc
   src/reductions/cb/cb_explore_adf_cover.cc
   src/reductions/cb/cb_explore_adf_first.cc
   src/reductions/cb/cb_explore_adf_greedy.cc
+  src/reductions/cb/cb_explore_adf_large_action_space.cc
   src/reductions/cb/cb_explore_adf_regcb.cc
   src/reductions/cb/cb_explore_adf_rnd.cc
   src/reductions/cb/cb_explore_adf_softmax.cc
@@ -343,23 +352,6 @@ set(vw_core_sources
   src/vw_validate.cc
 )
 
-if(VW_BUILD_LARGE_ACTION_SPACE)
-  set(vw_core_headers
-    ${vw_core_headers}
-    include/vw/core/reductions/cb/cb_explore_adf_large_action_space.h
-    include/vw/core/reductions/cb/cb_actions_mask.h include/vw/core/qr_decomposition.h
-  )
-
-  set(vw_core_sources
-    ${vw_core_sources}
-    src/reductions/cb/cb_explore_adf_large_action_space.cc
-    src/reductions/cb/details/large_action/one_pass_svd_impl.cc
-    src/reductions/cb/details/large_action/two_pass_svd_impl.cc
-    src/reductions/cb/details/large_action/one_rank_spanner_impl.cc
-    src/reductions/cb/cb_actions_mask.cc src/qr_decomposition.cc
-  )
-endif()
-
 vw_add_library(
     NAME "core"
     TYPE "STATIC_ONLY"
@@ -370,6 +362,7 @@ vw_add_library(
         # Use BUILD_INTERFACE to prevent them from being exported, i.e. treat them as PRIVATE
         # https://gitlab.kitware.com/cmake/cmake/issues/15415
         ${CMAKE_DL_LIBS} ${LINK_THREADS} vw_io $<BUILD_INTERFACE:${boost_math_target}> $<BUILD_INTERFACE:RapidJSON>
+        $<BUILD_INTERFACE:eigen>
     DESCRIPTION "This contains all remaining VW code, all reduction implementations, driver, option handling"
     EXCEPTION_DESCRIPTION "Yes"
     ENABLE_INSTALL
@@ -387,11 +380,6 @@ if(BUILD_FLATBUFFERS)
   target_compile_definitions(vw_core PUBLIC BUILD_FLATBUFFERS)
 endif()
 
-if(VW_BUILD_LARGE_ACTION_SPACE)
-  target_link_libraries(vw_core PUBLIC $<BUILD_INTERFACE:eigen>)
-  target_compile_definitions(vw_core PUBLIC BUILD_LARGE_ACTION_SPACE)
-endif()
-
 # Handle generated header
 configure_file(config.h.in ${CMAKE_CURRENT_BINARY_DIR}/include/vw/core/config.h)
 target_include_directories(vw_core PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_BINARY_DIR}/include/>)