Research done at Johns Hopkins University by Michael DeLeo

Contact: [email protected]

Github

HuggingFace

BertHarmon is a BERT model trained for the task of Chess.

from transformers import pipeline
task = pipeline('fill-mask', model='BertHarmon')
task("rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1 White <MOVE_SEP> [MASK]")

The base string consists of the FEN_position followed by the player color and a move seperator. Finally with the [MASK] token. The mask token is the algebraic notation for a chess move to be taken givent the current board state in FEN Notation

Build Stockfish

cd stockfish_build
. ./build.sh

Then install the requirements for either a CPU machine or CUDA.

CPU Machine

pip install -r requirements.txt

CUDA Machine

pip install -r requirements-cuda.txt

All our training scripts utilize jupyter notebooks.

jupyter lab

Training requires CUDA 11.3

Reference the Train BERT NIM notebook

Reference the Train BERT Chess notebook

Using docker is optional, it can help to standardize your environment. The reqs here are not that complicated though.

docker-compose -- 1.27 docker -- 20.10.12

1. Make sure docker and docker compose are installed.
2. For a GPU enabled machine, install nvidia runtime
   1. LINUX: sudo apt-get install nvidia-docker2 -y
   2. Windows: TODO

CPU Machine

docker compose up --build

OR

GPU Machine

docker compose up --build -f docker-compose-gpu.yml

docker compose up -f docker-compose-remote.yml