Gomoku AI (Alpha Zero) implemented by pytorch and onnxruntime.
bilibili (part 1-3):
https://www.bilibili.com/video/BV1a5411Q7en/
https://www.bilibili.com/video/BV1tL4y1K7UB/
https://www.bilibili.com/video/BV1Fa411L7ea/
PPT(Chinese):
AlphaZero-AI网络教程.pptx
Currently only Gomoku and similar games such as Tic-Tac-Toe.
Welcome other game implementions if you want to become the contributor!
linux/Windows (tested on Ubuntu 20 + GPU and Windows 10 + GPU or CPU)
Both GPU and CPU (GPU test on Tesla V100 + Cuda 11 / CPU test on Intel i7)
C++ (for speed!) and python. The model is trained by pytorch (Python) and onnxruntime (C++,for selfplay), and inferenced by onnxruntime (C++).
gcc (linux) or visual studio 19 (windows)
cmake 3.13+
pytorch (tested on 1.11)
onnxruntime-gpu (tested on 1.11)
Download and install miniconda / python and "pip install" all the dependent packages such as pytorch
train.sh: convert "/data/miniconda3/bin/python" to "python" or "python3" or your own python intepreter path
Download onnxruntime: https://github.com/microsoft/onnxruntime/releases/tag/v1.11.1
CMakefiles.txt: convert the onnxruntime path to your own path
mkdir build
copy *.sh to ./build
cd ./build
cmake .. (or "cmake -A x64 ..")
cmake --build . --config Release (or open .sln file through visual Studio 19 and generate for win10)
cd ./build
bash train_net.sh
If you want to train the model on windows 10, convert "train.sh" to "train.bat" and change corresponding commands.
run mcts_test, for example in linux:
./mcts_test ./weights/1000.onnx 1
Here 1(or 0) = AI play with black(or white) pieces.
The newest trained model will be updated in "model" directory.
Increase or decrease "NUM_MCT_SIMS" in include/common.h (default 1600) to increase the power or speed of AI.
一个使用pytorch + onnxruntime训练的Alpha Zero训练框架。 onnxruntime主要负责“左右互搏”的部分，pytorch负责模型参数优化。
目前游戏仅支持五子棋和井字棋，如果有其它小伙伴愿意提供其它棋类游戏的源码，这里非常欢迎。
支持多线程蒙特卡洛树搜索,该部分和模型推理（用于左右互搏）部分均由c++完成，之所以不用python主要是为了加快速度，并且避开python GIL的坑。
本项目支持windows和linux平台。由于我主要在linux上训练，windows相应的代码可能未来不会及时更新，需要修改一下才能用。