KataGo/Linux

KataGo/Linux说明怎么在Linux系统上使用KataGo。

基本环境

这边是用AWS上的Ubuntu 18.04，开g4dn.xlarge测试的。

编译

先安装需要的套件：

sudo apt install -y build-essential cmake clinfo git libboost-filesystem-dev libgoogle-perftools-dev libzip-dev nvidia-compute-utils-435 nvidia-driver-435 nvidia-headless-435 nvidia-utils-435 ocl-icd-opencl-dev unzip wget zlib1g-dev

云端上面可能已经有显示卡的驱动了，这时候不要装Nvidia驱动：

sudo apt install -y build-essential clinfo cmake git libboost-filesystem-dev libgoogle-perftools-dev libzip-dev ocl-icd-opencl-dev unzip zlib1g-dev

再来编译：

cd ~; git clone https://github.com/lightvector/KataGo.git; cd ~/KataGo/cpp; cmake . -DUSE_BACKEND=OPENCL -DUSE_TCMALLOC=1; renice 20 $$; make -j16

然后把KataGo的执行档丢到Lizzie目录下：

cd ~/KataGo/cpp; cp katago ~/Lizzie

另外一种方式是编Eigen的版本：

cd ~; git clone https://github.com/lightvector/KataGo.git; cd ~/KataGo/cpp; cmake . -DUSE_BACKEND=EIGEN -DUSE_AVX2=1 -DUSE_TCMALLOC=1; renice 20 $$; make -j16

设定

等下会使用Lizzie当作前端，所以这边会把KataGo的设定档丢到Lizzie的目录下：

cd ~/KataGo/cpp; cp configs/gtp_example.cfg ~/Lizzie/

下载最新的训练资料

最新的训练资料可以在KataGo Distributed Training取得。

这边会下载先前最强的20 blocks、30 blocks与40 blocks版本到Lizzie的目录下并且解开：

cd ~/Lizzie/; wget https://d3dndmfyhecmj0.cloudfront.net/g170/neuralnets/g170e-b20c256x2-s3761649408-d809581368.zip https://d3dndmfyhecmj0.cloudfront.net/g170/neuralnets/g170-b30c320x2-s2846858752-d829865719.zip https://d3dndmfyhecmj0.cloudfront.net/g170/neuralnets/g170-b40c256x2-s2990766336-d830712531.zip; echo *.zip | xargs -n1 unzip

测试

先测试Nvidia显卡有没有生效（要记得先重开机让驱动程式生效）：

nvidia-smi

接下来可以跑KataGo的测试：

cd ~/KataGo/; ./katago benchmark -model g170e-b20c256x2-s3761649408-d809581368/model.bin.gz -config gtp_example.cfg

跑完测试后可以依照建议修改gtp_example.cfg里的numSearchThreads以加快速度。

快速安装

这边包括了KataGo、Lizzie以及VNC相关的套件。

sudo apt update; sudo apt install -y build-essential clinfo cmake default-jdk git icewm libboost-filesystem-dev libgoogle-perftools-dev libzip-dev nvidia-compute-utils-435 nvidia-driver-435 nvidia-headless-435 nvidia-utils-435 ocl-icd-opencl-dev tightvncserver unzip wget zlib1g-dev; cd /tmp; wget https://github.com/featurecat/lizzie/releases/download/0.7.2/Lizzie.0.7.2.Mac-Linux.zip; cd ~; unzip /tmp/Lizzie.0.7.2.Mac-Linux.zip; git clone https://github.com/lightvector/KataGo.git; cd KataGo/cpp; cmake . -DUSE_BACKEND=OPENCL -DUSE_TCMALLOC=1; make -j8; cp katago ~/Lizzie; cp configs/gtp_example.cfg ~/Lizzie; cd ~/Lizzie; wget https://d3dndmfyhecmj0.cloudfront.net/g170/neuralnets/g170e-b20c256x2-s3761649408-d809581368.zip https://d3dndmfyhecmj0.cloudfront.net/g170/neuralnets/g170-b30c320x2-s2846858752-d829865719.zip https://d3dndmfyhecmj0.cloudfront.net/g170/neuralnets/g170-b40c256x2-s2990766336-d830712531.zip; echo *.zip | xargs -n1 unzip; sudo reboot

云端上面不装显示卡的驱动：

sudo apt update; sudo apt install -y build-essential clinfo cmake default-jdk git icewm libboost-filesystem-dev libgoogle-perftools-dev libzip-dev ocl-icd-opencl-dev tightvncserver unzip wget zlib1g-dev; cd /tmp; wget https://github.com/featurecat/lizzie/releases/download/0.7.2/Lizzie.0.7.2.Mac-Linux.zip; cd ~; unzip /tmp/Lizzie.0.7.2.Mac-Linux.zip; git clone https://github.com/lightvector/KataGo.git; cd KataGo/cpp; cmake . -DUSE_BACKEND=OPENCL -DUSE_TCMALLOC=1; make -j8; cp katago ~/Lizzie; cp configs/gtp_example.cfg ~/Lizzie; cd ~/Lizzie; wget https://d3dndmfyhecmj0.cloudfront.net/g170/neuralnets/g170e-b20c256x2-s3761649408-d809581368.zip https://d3dndmfyhecmj0.cloudfront.net/g170/neuralnets/g170-b30c320x2-s2846858752-d829865719.zip https://d3dndmfyhecmj0.cloudfront.net/g170/neuralnets/g170-b40c256x2-s2990766336-d830712531.zip; echo *.zip | xargs -n1 unzip; sudo reboot

重开机后连进去再跑：

cd ~/Lizzie; ./katago benchmark -model g170e-b20c256x2-s3761649408-d809581368/model.bin.gz -config gtp_example.cfg

执行

先把VNC跑起来：

tightvncserver -geometry 1680x1050 -depth 24

透过VNC连进去后，用终端机（Terminal，预设是xterm）后执行Lizzie：

cd ~/Lizzie; java -jar lizzie.jar

另外也可以考虑设定sun.java2d.opengl让画面比较不卡：

cd ~/Lizzie; java -Dsun.java2d.opengl=true -jar lizzie.jar

相关连结

• KataGo
• Lizzie

外部连结

• 官方网站 （英文）