This is the chess variant NNUE training code for Fairy-Stockfish. See the documentation in the wiki for more details on the training process and join our discord to ask questions. This project is derived from the trainer for standard chess used for official Stockfish.
Requires a CUDA capable GPU.
python3 -m venv env
source env/bin/activate
pip install python-chess==0.31.4 "pytorch-lightning<1.5.0" torch matplotlib
First check what version of cuda is being used by pytorch.
import torch
torch.version.cuda
Then install CuPy with the matching CUDA version.
pip install cupy-cudaXXX
where XXX corresponds to the first 3 digits of the CUDA version. For example cupy-cuda112
for CUDA 11.2.
CuPy might use the PyTorch's private installation of CUDA, but it is better to install the matching version of CUDA separately. CUDA Downloads
This requires a C++17 compiler and cmake.
Windows:
compile_data_loader.bat
Linux/Mac:
sh compile_data_loader.bat
source env/bin/activate
python train.py train_data.bin val_data.bin
python train.py --resume_from_checkpoint <path> ...
python train.py --gpus 1 ...
By default the trainer uses a factorized HalfKAv2 feature set (named "HalfKAv2^")
If you wish to change the feature set used then you can use the --features=NAME
option. For the list of available features see --help
The default is:
python train.py ... --features="HalfKAv2^"
--smart-fen-skipping
currently skips over moves where the king is in check, or where the bestMove is a capture (typical of non-quiet positions).
--random-fen-skipping N
skip N fens on average before using one. Uses fewer fens per game, useful with large data sets.
python train.py --smart-fen-skipping --random-fen-skipping 3 --batch-size 16384 --threads 8 --num-workers 8 --gpus 1 trainingdata validationdata
best nets have been trained with 16B d9-scored nets, training runs >200 epochs
Using either a checkpoint (.ckpt
) or serialized model (.pt
),
you can export to SF NNUE format. This will convert last.ckpt
to nn.nnue
, which you can load directly in SF.
python serialize.py last.ckpt nn.nnue
Import an existing SF NNUE network to the pytorch network format.
python serialize.py nn.nnue converted.pt
Visualize a network from either a checkpoint (.ckpt
), a serialized model (.pt
)
or a SF NNUE file (.nnue
).
python visualize.py nn.nnue --features="HalfKAv2"
Visualize the difference between two networks from either a checkpoint (.ckpt
), a serialized model (.pt
)
or a SF NNUE file (.nnue
).
python visualize.py nn.nnue --features="HalfKAv2" --ref-model nn.cpkt --ref-features="HalfKAv2^"
pip install tensorboard
tensorboard --logdir=logs
Then, go to https://localhost:6006/
python run_games.py --concurrency 16 --stockfish_exe ./stockfish.master --c_chess_exe ./c-chess-cli --ordo_exe ./ordo --book_file_name ./noob_3moves.epd run96
Automatically converts all .ckpt
found under run96
to .nnue
and runs games to find the best net. Games are played using c-chess-cli
and nets are ranked using ordo
.
This script runs in a loop, and will monitor the directory for new checkpoints. Can be run in parallel with the training, if idle cores are available.
- Sopel - for the amazing fast sparse data loader
- connormcmonigle - https://github.com/connormcmonigle/seer-nnue, and loss function advice.
- syzygy - https://www.talkchess.com/forum3/viewtopic.php?f=7&t=75506
- https://github.com/DanielUranga/TensorFlowNNUE
- https://hxim.github.io/Stockfish-Evaluation-Guide/
- dkappe - Suggesting ranger (https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer)