Alpha-Zero-Gomoku

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

Alpha-Zero-Gomoku

★ 10 Open GitHub ↗

Gomoku AI (Alpha Zero) implemented by pytorch and onnxruntime.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Alpha-Zero-Gomoku
Gomoku AI (Alpha Zero) implemented by pytorch and onnxruntime.

# Tutorial About Alpha Zero (Chinese):
bilibili (part 1-3):

https://www.bilibili.com/video/BV1a5411Q7en/

https://www.bilibili.com/video/BV1tL4y1K7UB/

https://www.bilibili.com/video/BV1Fa411L7ea/

PPT(Chinese):

Tutorial/AlphaZero-AI网络教程.pptx

# Supported Games
Currently Gomoku and Chinese Chess

Welcome other game implementions if you want to become the contributor!

# Supported OS System
linux/Windows (tested on Ubuntu 22 + GPU and Windows + GPU or CPU)

# Supported Enviroment
Both GPU and CPU (GPU test on GTX 4090 + Cuda 12 / CPU test on Intel i7)

# Language
C++ (for speed!) and python. The model is trained by pytorch (Python) and onnxruntime (C++,for selfplay), and inferenced by onnxruntime (C++).

# Dependence
gcc (linux) or visual studio 19+ (windows)

cmake 4.0.0+

pytorch (tested on onnxruntime-v1.22.0)

onnxruntime-gpu (tested on onnxruntime-v1.22.0)

# Installation
Download and install miniconda / python
and "pip install" all the dependent packages such as pytorch，onnx ...

train.sh: convert "python" to "/data/miniconda3/bin/python" or "python3" or your own python intepreter path

Download onnxruntime: https://github.com/microsoft/onnxruntime/releases/tag/v1.22.0

mv CMakeLists-Gomoku.txt CMakefiles.txt

convert onnxruntime path in CMakefiles.txt to your own path：
set(ONNXRUNTIME_ROOTDIR "/mnt/data/myc/projects/onnxruntime-linux-x64-gpu-1.22.0")

mkdir build

cp *.sh ./build

cd ./build

cmake ..    (or  win10: "cmake -A x64 ..")

cmake --build . --config Release   (or win10: open .sln file through visual Studio 19 and generate)

# Train (Linux)
cd ./build

chmod 777 train.sh

bash train.sh

If you want to train the model on windows 10/11, convert "train.sh" to "train.bat" and transfer commands.

# Human play with AI (inference)
run mcts_test, for example in linux:

./mcts_test ./weights/423.onnx 1   (now saved in "model" dir)

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.

# About This Project (Chinese)

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。

目前游戏支持五子棋和中国象棋，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相应的代码不会及时更新，需要自行修改一下才能用。

Alpha-Zero-Gomoku

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README