Big refactor from FAIR team

This commit includes a *ton* of refactoring goodness and various stability fixes.

Author: jma127

.gitmodules

@@ -1,21 +1,22 @@
 [submodule "third_party/concurrentqueue"]
 path = third_party/concurrentqueue
 url = https://github.com/cameron314/concurrentqueue.git
-[submodule "third_party/cppzmq"]
-path = third_party/cppzmq
-url = https://github.com/zeromq/cppzmq.git
-[submodule "third_party/googletest"]
-path = third_party/googletest
-url = https://github.com/google/googletest.git
-[submodule "third_party/json"]
-path = third_party/json
-url = https://github.com/nlohmann/json.git
 [submodule "third_party/pybind11"]
 path = third_party/pybind11
 url = https://github.com/pybind/pybind11.git
 [submodule "third_party/spdlog"]
 path = third_party/spdlog
 url = https://github.com/gabime/spdlog.git
+[submodule "third_party/json"]
+path = third_party/json
+url = https://github.com/nlohmann/json.git
+[submodule "third_party/googletest"]
+path = third_party/googletest
+url = https://github.com/google/googletest.git
 [submodule "third_party/tbb"]
 path = third_party/tbb
 url = https://github.com/01org/tbb.git
+ignore = untracked
+[submodule "third_party/cppzmq"]
+path = third_party/cppzmq
+url = https://github.com/zeromq/cppzmq.git

scripts/elfgames/go/analysis.sh

@@ -14,10 +14,10 @@ set(ELF_SOURCES
  options/Pybind.cc
 )
 
-set(ELF_TEST_SOURCES
- options/OptionMapTest.cc
- options/OptionSpecTest.cc
-)
+# set(ELF_TEST_SOURCES
+#  options/OptionMapTest.cc
+#  options/OptionSpecTest.cc
+# )
 
 # Main ELF library
 
@@ -26,10 +26,10 @@ target_compile_definitions(elf PUBLIC
  GIT_COMMIT_HASH=${GIT_COMMIT_HASH}
  GIT_STAGED=${GIT_STAGED_STRING}
 )
+
 target_link_libraries(elf PUBLIC
  #${Boost_LIBRARIES}
  concurrentqueue
- cppzmq
  nlohmann_json
  pybind11
  $<BUILD_INTERFACE:${PYTHON_LIBRARIES}>
@@ -40,7 +40,7 @@ target_link_libraries(elf PUBLIC
 # Tests
 
 enable_testing()
-add_cpp_tests(test_cpp_elf_ elf ${ELF_TEST_SOURCES})
+# add_cpp_tests(test_cpp_elf_ elf ${ELF_TEST_SOURCES})
 
 # Python bindings
 

src_cpp/elf/CMakeLists.txt

@@ -88,9 +88,9 @@ class AIClientT : public AI_T<S, A> {
  status == comm::ReplyStatus::UNKNOWN;
  }
 
- virtual bool act_batch(
+ bool act_batch(
  const std::vector<const S*>& batch_s,
- const std::vector<A*>& batch_a) {
+ const std::vector<A*>& batch_a) override {
  std::vector<elf::FuncsWithState> funcs_s =
  client_->BindStateToFunctions(targets_, batch_s);
  std::vector<elf::FuncsWithState> funcs_a =

src_cpp/elf/ai/ai.h

@@ -70,7 +70,7 @@ class TreeSearchSingleThreadT {
  : threadId_(thread_id), options_(options) {
  if (options_.verbose) {
  std::string log_file =
- "tree_search_" + std::to_string(thread_id) + ".txt";
+ options_.log_prefix + std::to_string(thread_id) + ".txt";
  output_.reset(new std::ofstream(log_file));
  }
  }
@@ -374,10 +374,7 @@ class TreeSearchT {
  return searchTree_.printTree();
  }
 
- MCTSResult runPolicyOnly(const State& /*root_state*/) {
- // TODO Policy only doesn't work.
- assert(false);
- /*
+ MCTSResult runPolicyOnly(const State& root_state) {
  if (actors_.empty() || treeSearches_.empty()) {
  throw std::range_error(
  "TreeSearch::runPolicyOnly works when there is at least one thread");
@@ -386,15 +383,17 @@ class TreeSearchT {
 
  // Some hack here.
  Node* root = searchTree_.getRootNode();
- treeSearches_[0]->visit(*actors_[0], root);
 
- // return StrongestPrior(root->getStateActions());
- */
+ if (!root->isVisited()) {
+ NodeResponseT<Action> resp;
+ actors_[0]->evaluate(*root->getStatePtr(), &resp);
+ root->setEvaluation(resp);
+ }
 
  MCTSResult result;
- // result.action_rank_method = MCTSResult::PRIOR;
- // result.addActions(root->getStateActions());
-
+ result.action_rank_method = MCTSResult::PRIOR;
+ result.addActions(root->getStateActions());
+ result.root_value = root->getValue();
  return result;
  }
 
@@ -490,6 +489,8 @@ class TreeSearchT {
 
  // Pick the best solution.
  MCTSResult result;
+ result.root_value = root->getValue();
+
  // MCTSResult result2;
  if (options_.pick_method == "strongest_prior") {
  result.action_rank_method = MCTSResult::PRIOR;

src_cpp/elf/ai/tree_search/tree_search.h

@@ -61,8 +61,10 @@ struct StateTrait {
  return s1 == s2;
  }
 
- static bool
- moves_since(const S& s, size_t* next_move_number, std::vector<A>* moves) {
+ static bool moves_since(
+ const S& /*s*/,
+ size_t* /*next_move_number*/,
+ std::vector<A>* /*moves*/) {
  // By default it is not provided.
  return false;
  }
@@ -84,7 +86,7 @@ struct ActionTrait {
 template <typename Actor>
 struct ActorTrait {
  public:
- static std::string to_string(const Actor& a) {
+ static std::string to_string(const Actor&) {
  return "";
  }
 };
@@ -213,6 +215,7 @@ struct MCTSResultT {
  enum RankCriterion { MOST_VISITED = 0, PRIOR = 1, UNIFORM_RANDOM };
 
  Action best_action;
+ float root_value;
  float max_score;
  EdgeInfo best_edge_info;
  MCTSPolicy<Action> mcts_policy;
@@ -224,6 +227,7 @@ struct MCTSResultT {
  // action_edges ssengupta@fb.com
  MCTSResultT()
  : best_action(ActionTrait<Action>::default_value()),
+ root_value(0.0),
  max_score(std::numeric_limits<float>::lowest()),
  best_edge_info(0),
  total_visits(0),

src_cpp/elf/ai/tree_search/tree_search_base.h

@@ -85,6 +85,7 @@ struct TSOptions {
  bool persistent_tree = false;
  float root_epsilon = 0.0;
  float root_alpha = 0.0;
+ std::string log_prefix = "";
 
  // [TODO] Not a good design.
  // string pick_method = "strongest_prior";
@@ -102,6 +103,7 @@ struct TSOptions {
  ss << "Maximal #moves (0 = no constraint): " << max_num_moves
  << std::endl;
  ss << "Seed: " << seed << std::endl;
+ ss << "Log Prefix: " << log_prefix << std::endl;
  ss << "#Threads: " << num_threads << std::endl;
  ss << "#Rollout per thread: " << num_rollouts_per_thread
  << ", #rollouts per batch: " << num_rollouts_per_batch << std::endl;
@@ -156,6 +158,9 @@ struct TSOptions {
  if (t1.pick_method != t2.pick_method) {
  return false;
  }
+ if (t1.log_prefix != t2.log_prefix) {
+ return false;
+ }
  if (t1.root_epsilon != t2.root_epsilon) {
  return false;
  }
@@ -181,6 +186,7 @@ struct TSOptions {
  JSON_SAVE(j, seed);
  JSON_SAVE(j, persistent_tree);
  JSON_SAVE(j, pick_method);
+ JSON_SAVE(j, log_prefix);
  JSON_SAVE(j, root_epsilon);
  JSON_SAVE(j, root_alpha);
  JSON_SAVE(j, virtual_loss);
@@ -198,6 +204,7 @@ struct TSOptions {
  JSON_LOAD(opt, j, seed);
  JSON_LOAD(opt, j, persistent_tree);
  JSON_LOAD(opt, j, pick_method);
+ JSON_LOAD(opt, j, log_prefix);
  JSON_LOAD(opt, j, root_epsilon);
  JSON_LOAD(opt, j, root_alpha);
  JSON_LOAD(opt, j, virtual_loss);
@@ -213,6 +220,7 @@ struct TSOptions {
  verbose,
  persistent_tree,
  pick_method,
+ log_prefix,
  virtual_loss,
  verbose_time,
  alg_opt,

src_cpp/elf/ai/tree_search/tree_search_options.h

@@ -124,7 +124,7 @@ class Context {
 
  void start() {
  th_.reset(new std::thread([&]() {
- assert(nice(10) == 10);
+ // assert(nice(10) == 10);
  collectAndSendBatch();
  }));
  }
@@ -183,11 +183,17 @@ class Context {
  }
  }
  smem_->waitBatchFillMem(server_);
- // LOG(INFO) << "Receiver: Batch received. #batch = "
- // << batch.size() << std::endl;
+ // std::cout << "Receiver[" << smem_opts.getLabel() << "] Batch
+ // received. #batch = "
+ // << smem_->getEffectiveBatchSize() << std::endl;
+
  comm::ReplyStatus batch_status =
  batchClient_->sendWait(smem_.get(), {""});
 
+ // std::cout << "Receiver[" << smem_opts.getLabel() << "] Batch
+ // releasing. #batch = "
+ // << smem_->getEffectiveBatchSize() << std::endl;
+
  // LOG(INFO) << "Receiver: Release batch" << std::endl;
  smem_->waitReplyReleaseBatch(server_, batch_status);
  }
@@ -280,7 +286,7 @@ class Context {
  auto* client = getClient();
  for (int i = 0; i < num_games_; ++i) {
  game_threads_.emplace_back([i, client, this]() {
- assert(nice(19) == 19);
+ // assert(nice(19) == 19);
  client->start();
  game_cb_(i, client);
  client->End();
@@ -329,7 +335,7 @@ class Context {
  std::atomic<bool> tmp_thread_done(false);
 
  std::thread tmp_thread([&]() {
- assert(nice(10) == 10);
+ // assert(nice(10) == 10);
 
  std::cout << "Prepare to stop ..." << std::endl;
  client_->prepareToStop();