barbakan_zero

In this repo I will try to adapt KataGo's zero knowledge training pipeline to gomoku. This repo is a truncated version of repo: https://github.com/lightvector/KataGo
As this is an experimental project derived form KataGo, I definitely recommend to investigate the KataGo repository for best practices. This repo is not matured, but helped me produce Barbakan engine, a strong gomoku program, competing successfully on gomoku ai tournament: https://gomocup.org/

Setup repo

Supported os: Linux

Known dependencies:
cmake
ninja
python3.10-venv (python version is most likely flexible)
cuda toolkit (https://developer.nvidia.com/cuda-toolkit)
TCMalloc (sudo apt-get install google-perftools && sudo apt -y install libgoogle-perftools-dev)
zlib (sudo apt-get install zlib1g-dev && sudo apt-get install libzip-dev)

To check that you have a working repo for development run the followings from the root of the repo:
./source/scripts/configure.sh release
./source/scripts/build.sh release
./source/scripts/test.sh release
(last message should be 'All tests passed')

Alternatively, it is possible to setup the work environment of the repo with Docker. For more info, please check part "Docker (misc)".

Analyze with the engine

Next to the executable (in ./build/release/bin/barbakan_zero) one need to copy a config file (default_gtp.cfg can be found in ./data/configs/gtp) and a model file (model.bin.gz can be found in ./data/model, or one can train a new model with the training pipeline).

For a graphical interface please set up the q5gomoku project. Clone and follow the Readme.md instructions of:
https://github.com/jano-wol/q5gomoku

gtp engine can be also used without the graphical interface, but this is very inconvenient. Some example commands:
./build/release/bin/barbakan_zero gtp
boardsize 19
showboard
play B A19
play W C1
showboard

Train NN model

After the build step is ready nn model train can be started by:
./source/scripts/train.sh release train_id

If train_id is a new id, then a new training process will be started. Otherwise, the training process corresponding to the already existing id will be continued. The output folder of the training process is ./data/train/train_id. (This is not cleaned by the configure step.)

Currently train.sh will always train a hard coded NN model corresponding to model id b6c96 (see ./source/python/modelconfigs.py for all the possibilities). To change this hard coded value, adjust line
NN_ID=b6c96
in
./source/scripts/train/loop_call.sh

The training process hyperparameters are determined by the following files:
./data/configs/selfplay/gatekeeper_train.cfg.in
./data/configs/selfplay/selfplay_train.cfg.in
./source/scripts/train/synchronous_loop.sh
It is possible that as the training process evolves, some changes are needed in the constants. (e.g. one can try to experiment with NUM_GAMES_PER_CYCLE or surewinDepth etc.)

Docker (misc)

Prerequisite: Docker. From the root of the repo run (sudo might be needed):
./source/scripts/misc/docker/build.sh
./source/scripts/misc/docker/run.sh

In case run.sh produce the following error:
Error response from daemon: could not select device driver "" with capabilities: [[gpu]].
try the following commands on the host machine (sudo might be needed):
apt-get install -y nvidia-container-toolkit
apt install -y nvidia-docker2
systemctl daemon-reload
systemctl restart docker
and rerun
./source/scripts/misc/docker/run.sh

In case run.sh is successful, a new command port is starting in the image. Now git has to be configured in the image by:
git config --global --add safe.directory /barbakan_zero

Finally, similarly to the conventional setup, it is possible to test that the repository is functioning correctly by calling:
./source/scripts/configure.sh release
./source/scripts/build.sh release
./source/scripts/test.sh release