A little experiment at using transformers to play go. The basic concept is to feed a sequence of moves to the transformer and predict the next move.

A GTP interface can be started from the command-line, or if you prefer you can build a docker image that runs the bot in an isolated environment:

python3 -m hugging_go

make build-image
docker run -it `docker build -q .`

pip install -r requirements.txt

You can train a tokenizer from scratch on a given dataset, which should consist of one-line SGF files. The resulting tokenizer will be saved to model/tokenizer.json:

python3 -m hugging_go --train_token=True data/kgs_big.sgf

After training the tokenizer the main model can be trained using the same one-line SGF files. This will probably take a long time:

python3 -m hugging_go --train-model=True data/kgs_big.sgf