fix: some adjustments. 🚀

CMakeLists.txt

@@ -5,7 +5,7 @@ if(CMAKE_COMPILER_IS_GNUCC)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++17 -pthread -Wl,-rpath=./:./lib")
     set(ONNXRUNTIME_ROOTDIR "/mnt/d/Linker/onnxruntime-linux-x64-1.11.1")
 else()
-    set(ONNXRUNTIME_ROOTDIR "D:/Linker/onnxruntime-win-x86-1.11.1")
+    set(ONNXRUNTIME_ROOTDIR "D:/Linker/onnxruntime-win-x64-gpu-1.11.1")
 endif()
 
 # tensorrt_provider_factory.h contains old APIs of the tensorrt execution provider

README.md

@@ -51,8 +51,6 @@ cmake ..    # (or "cmake -A x64 ..")
 cmake --build . --config Release   # (or open .sln file through visual Studio 19 and generate for win10)
 ```
 
-add `-DUSE_CUDA=ON` to enable gpu.
-
 ## Train (Linux)
 
 ```shell

src/mcts.cpp

@@ -463,9 +463,6 @@ void MCTS::simulate(std::shared_ptr<Gomoku> game)
       }
     }
 
-    // std::cout << "value = " << value << std::endl;
-    // std::cout << "action_priors = " << action_priors << std::endl;
-
     double sum = 0;
     for (unsigned int i = 0; i < action_priors.size(); i++)
     {

src/onnx.cpp

@@ -58,11 +58,10 @@ NeuralNetwork::NeuralNetwork(const std::string model_path, const unsigned int ba
 
   session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_EXTENDED);
 
+//#define USE_CUDA
 #ifdef USE_CUDA
-  void enable_cuda(OrtSessionOptions * session_options)
-  {
-    ORT_ABORT_ON_ERROR(OrtSessionOptionsAppendExecutionProvider_CUDA(session_options, 0));
-  }
+    OrtSessionOptionsAppendExecutionProvider_CUDA(session_options, 0);
+    //std::cout << "DEBUG ......Enable CUDA......" << std::endl;
 #endif
 
 #ifdef _WIN32

src/pbrain-Z2I/pbrain-Z2I.cpp

@@ -116,11 +116,11 @@ int main(int argc, char *argv[])
     module = new NeuralNetwork(s_model_path, NUM_MCT_SIMS);
 
     unsigned int u_timeout_turn = 30000;
-    double per_sims = (double)(u_timeout_turn)/(double)(4*60000);
+    double per_sims = (double)(u_timeout_turn)/(double)(4*30000);
     cout << "DEBUG per_sims: " << per_sims << endl;
 
     MCTS *m;
-    m = new MCTS(module, NUM_MCT_THREADS, C_PUCT, (unsigned int)(NUM_MCT_SIMS * NUM_MCT_THREADS * per_sims), C_VIRTUAL_LOSS, BORAD_SIZE * BORAD_SIZE);
+    m = new MCTS(module, NUM_MCT_THREADS, C_PUCT, (unsigned int)(NUM_MCT_SIMS * log(NUM_MCT_THREADS) * per_sims), C_VIRTUAL_LOSS, BORAD_SIZE * BORAD_SIZE);
 
     string command;
     unsigned int size;
@@ -485,7 +485,7 @@ int main(int argc, char *argv[])
                             delete m;
                             m = nullptr;
                         }
-                        m = new MCTS(module, NUM_MCT_THREADS, C_PUCT, (unsigned int)(NUM_MCT_SIMS * NUM_MCT_THREADS * per_sims * (value * 0.1 / u_timeout_turn) + 1), C_VIRTUAL_LOSS, BORAD_SIZE * BORAD_SIZE);
+                        m = new MCTS(module, NUM_MCT_THREADS, C_PUCT, (unsigned int)(NUM_MCT_SIMS * log(NUM_MCT_THREADS) * per_sims * (value * 0.1 / u_timeout_turn) + 1), C_VIRTUAL_LOSS, BORAD_SIZE * BORAD_SIZE);
                     }
                 }
             }

src/play.cpp

@@ -61,12 +61,11 @@ void SelfPlay::play(unsigned int saved_id)
 
     while (game_state.first == 0)
     {
-        // g->render();
-        double temp = step < EXPLORE_STEP ? 1 : 0;
+        // cout << "game id: " << saved_id << endl;
+        double temp = step < EXPLORE_STEP ? 1.0 : 1e-3;
         auto action_probs = mcts->get_action_probs(g.get(), temp);
         // auto action_probs = m->get_action_probs(g.get(), 1);
         // int best_action = m->get_best_action_from_prob(action_probs);
-
         board_type board = g->get_board();
         for (int i = 0; i < BORAD_SIZE * BORAD_SIZE; i++)
         {
@@ -115,6 +114,7 @@ void SelfPlay::play(unsigned int saved_id)
         mcts->update_with_move(res);
         g->execute_move(res);
         game_state = g->get_game_status();
+        g->render();
         step++;
     }
     cout << "Self play: total step num = " << step << " winner = " << game_state.second << endl;