Skip to content

sgrvinod/chess-transformers

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

72 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Chess Transformers

Teaching transformers to play chess

Version License


Chess Transformers is a library for training transformer models to play chess by learning from human games.

Contents

Install

Models

Datasets

Play

Train Models

Contribute

Install

To install Chess Transformers, clone this repository and install as a Python package locally.

gh repo clone sgrvinod/chess-transformers
cd chess-transformers
pip install .

If you are planning to develop or contribute or make changes to the codebase, install the package in editable mode, using the -e flag.

pip install -e .

OPTIONAL — If you want to train or evaluate a model, you may need to set some of the following environment variables on your computer:

  • Set CT_DATA_FOLDER to the folder on your computer where you have the training data. You do not need to set this if you do not plan to train any models.

  • Set CT_STOCKFISH_PATH to the executable of the Stockfish 16 chess engine. You do not need to set this if you do not plan to have a model play against this chess engine.

  • Set CT_FAIRY_STOCKFISH_PATH to the executable of the Fairy Stockfish chess engine. You do not need to set this if you do not plan to have a model play against this chess engine.

Models

There are currently four models available for use in Chess Transformers.

Model Name # Params Training Data Architecture Predictions
CT-E-20 20M LE22ct Transformer encoder only Best next half-move (or ply)
eg. f2e3
CT-EFT-20 20M LE22ct Transformer encoder only Best From and To squares corresponding to the next half-move eg. from f2 to e3
CT-ED-45 45M LE22ct Transformer encoder
and decoder
Sequence of half-moves (or plies)
eg. f2e3 -> b4b3 -> e3h6 -> b3b2 -> g4e6 -> g8f8 -> g3g7 -> f8e8 -> g7f7 -> loses
CT-EFT-85 85M LE22c Transformer encoder only Best From and To squares corresponding to the next half-move eg. from f2 to e3

All models are evaluated against the Fairy Stockfish chess engine at increasing strength levels 1 to 6, as predefined for use in the popular Stockfish chess bots on Lichess. The engine is run on an AMD Ryzen 7 3800X 8-Core Processor, with 8 CPU threads, and a hash table size of 8 GB. All other engine parameters are at their default values.

At each strength level of the chess engine, $n=1000$ games are played by the model, i.e. $500$ games each with black and white pieces.

Win ratios and the difference between the Elo rating of the model and the chess engine are calculated from these games' outcomes.

Detailed evaluation results for each model are provided below.

CT-E-20

Configuration File | Checkpoint | TensorBoard Logs

This is the encoder from the original transformer model in Vaswani et al. (2017) trained on the LE22ct dataset. A classification head at the turn token predicts the best half-move to be made (in UCI notation).

This is essentially a sequence (or image) classification task, where the sequence is the current state of the board, and the classes are the various moves that can be made on a chessboard in UCI notation.

CT-E-20 contains about 20 million parameters.

from chess_transformers.play import load_model
from chess_transformers.configs import import_config

CONFIG = import_config("CT-E-20")
model = load_model(CONFIG)

You do not need to download the model checkpoint manually. It will be downloaded automatically if required.

Model Strength

CT-E-20 was evaluated against the Fairy Stockfish chess engine at various strength levels as predefined for use in the popular Stockfish chess bots on Lichess. The engine is run on an AMD Ryzen 7 3800X 8-Core Processor, with 8 CPU threads, and a hash table size of 8 GB. All other engine parameters are at their default values.

These evaluation games can be viewed here.

Strength Level Games Wins Losses Draws Win Ratio Elo Difference Likelihood of Superiority
$LL$ $n$ $w$ $l$ $d$ $\frac{w + \frac{d}{2}}{n}$ $\Delta_{Elo}$ $LOS$
1 1000 989 0 11 99.45% 902.90
(± 117.67)
100.00%
2 1000 980 0 20 99.00% 798.25
(± 81.48)
100.00%
3 1000 872 61 67 90.55% 392.58
(± 33.31)
100.00%
4 1000 431 455 114 48.80% -8.34
(± 20.30)
21.00%
5 1000 205 685 110 26.00% -181.70
(± 22.78)
0.00%
6 1000 24 952 24 3.60% -571.11
(± 54.08)
0.00%

CT-EFT-20

Configuration File | Checkpoint | TensorBoard Logs

This is the encoder from the original transformer model in Vaswani et al. (2017) trained on the LE22ct dataset. Two classification heads operate upon the encoder outputs at all chessboard squares to predict the best candidates for the source (From) and destination (To) squares that correspond to the best half-move to be made.

This is essentially a sequence (or image) labeling task, where the sequence is the current state of the chessboard, and each square competes to be labeled as the From or To square.

CT-E-20 contains about 20 million parameters.

from chess_transformers.play import load_model
from chess_transformers.configs import import_config

CONFIG = import_config("CT-EFT-20")
model = load_model(CONFIG)

You do not need to download the model checkpoint manually. It will be downloaded automatically if required.

Model Strength

CT-EFT-20 was evaluated against the Fairy Stockfish chess engine at various strength levels as predefined for use in the popular Stockfish chess bots on Lichess. The engine is run on an AMD Ryzen 7 3800X 8-Core Processor, with 8 CPU threads, and a hash table size of 8 GB. All other engine parameters are at their default values.

These evaluation games can be viewed here.

Strength Level Games Wins Losses Draws Win Ratio Elo Difference Likelihood of Superiority
$LL$ $n$ $w$ $l$ $d$ $\frac{w + \frac{d}{2}}{n}$ $\Delta_{Elo}$ $LOS$
1 1000 994 0 6 99.70% 1008.63
(± 190.18)
100.00%
2 1000 988 0 12 99.40% 887.69
(± 111.13)
100.00%
3 1000 942 11 47 96.55% 578.77
(± 48.57)
100.00%
4 1000 697 192 111 75.25% 193.17
(± 23.08)
100.00%
5 1000 482 379 139 55.15% 35.91
(± 20.09)
99.98%
6 1000 61 872 67 9.45% -392.58
(± 33.31)
0.00%

CT-ED-45

Configuration File | Checkpoint | TensorBoard Logs

This is the original transformer model (encoder and decoder) in Vaswani et al. (2017) trained on the LE22ct dataset. A classification head after the last decoder layer predicts a sequence of half-moves, starting with the best half-move to be made next, followed by the likely course of the game an arbitrary number of half-moves into the future.

This is essentially a sequence-to-sequence (or image-to-sequence) task, where the input sequence is the current state of the board, and the output sequence is a string of half-moves that will likely occur on the board from that point onwards. Potentially, strategies applied to such tasks, such as beam search for decoding the best possible sequence of half-moves, can also be applied to this model. Training the model to predict not only the best half-move to make on the board right now, but also the sequence of half-moves that follow, can be viewed as a type of multitask training.

We are ultimately only interested in the very first half-move. Nevertheless, the full sequence of half-moves might help explain the model's decision for this important first half-move.

CT-ED-45 contains about 45 million parameters.

from chess_transformers.play import load_model
from chess_transformers.configs import import_config

CONFIG = import_config("CT-ED-45")
model = load_model(CONFIG)

You do not need to download the model checkpoint manually. It will be downloaded automatically if required.

Model Strength

CT-ED-45 was evaluated against the Fairy Stockfish chess engine at various strength levels as predefined for use in the popular Stockfish chess bots on Lichess. The engine is run on an AMD Ryzen 7 3800X 8-Core Processor, with 8 CPU threads, and a hash table size of 8 GB. All other engine parameters are at their default values.

Strength Level Games Wins Losses Draws Win Ratio Elo Difference Likelihood of Superiority
$LL$ $n$ $w$ $l$ $d$ $\frac{w + \frac{d}{2}}{n}$ $\Delta_{Elo}$ $LOS$
1 1000 976 0 24 98.80% 766.23
(± 73.45)
100.00%
2 1000 977 2 21 98.75% 759.05
(± 78.19)
100.00%
3 1000 676 244 80 71.60% 160.64
(± 22.72)
100.00%
4 1000 195 726 79 23.45% -205.52
(± 24.04)
0.00%
5 1000 67 895 38 8.60% -410.58
(± 36.41)
0.00%
6 1000 6 987 7 0.95% -807.25
(± 113.69)
0.00%

CT-EFT-85

Configuration File | Checkpoint | TensorBoard Logs

This is a larger version of the encoder from the original transformer model in Vaswani et al. (2017) trained on the LE22c dataset. Its size is analogous to BERTBASE in Devlin et al. (2018). Two classification heads operate upon the encoder outputs at all chessboard squares to predict the best candidates for the source (From) and destination (To) squares that correspond to the best half-move to be made.

This is essentially a sequence (or image) labeling task, where the sequence is the current state of the chessboard, and each square competes to be labeled as the From or To square.

CT-E-85 contains about 85 million parameters.

from chess_transformers.play import load_model
from chess_transformers.configs import import_config

CONFIG = import_config("CT-EFT-85")
model = load_model(CONFIG)

You do not need to download the model checkpoint manually. It will be downloaded automatically if required.

Model Strength

CT-EFT-85 was evaluated against the Fairy Stockfish chess engine at various strength levels as predefined for use in the popular Stockfish chess bots on Lichess. The engine is run on an AMD Ryzen 7 3800X 8-Core Processor, with 8 CPU threads, and a hash table size of 8 GB. All other engine parameters are at their default values.

These evaluation games can be viewed here.

Strength Level Games Wins Losses Draws Win Ratio Elo Difference Likelihood of Superiority
$LL$ $n$ $w$ $l$ $d$ $\frac{w + \frac{d}{2}}{n}$ $\Delta_{Elo}$ $LOS$
1 1000 999 0 1 99.95% 1320.33
(± 34.06)
100.00%
2 1000 997 0 3 99.85% 1129.30
(± 101.08)
100.00%
3 1000 979 0 21 98.95% 789.69
(± 79.22)
100.00%
4 1000 883 65 52 90.90% 399.81
(± 34.71)
100.00%
5 1000 712 183 105 76.45% 204.55
(± 23.52)
100.00%
6 1000 184 713 103 23.55% -204.55
(± 23.56)
0.00%

Datasets

There are currently five training datasets available in Chess Transformers.

Dataset Name Components # Datapoints
ML23c Board positions, turn, castling rights, next-move sequence (up to 10 half-moves) 10,797,366
LE22ct Board positions, turn, castling rights, next-move sequence (up to 10 half-moves) 13,287,522
GC22c Board positions, turn, castling rights, next-move sequence (up to 10 half-moves) 26,162,415
LE22c Board positions, turn, castling rights, next-move sequence (up to 10 half-moves) 127,684,720
ML23d Board positions, turn, castling rights, next-move sequence (up to 10 half-moves) 144,625,397

These datasets are sourced from groups of PGN files containing real games played by humans. There are currently three PGN filesets:

  • LE22 consists of games from the Lichess Elite Database put together by nikonoel, a collection of all standard chess games played on Lichess.org by players with a Lichess Elo rating of 2400+ against players with a Lichess Elo rating of 2200+ up to December 2021, and players rated 2500+ against players rated 2300+ from December 2021 up to December 2022

  • ML23 consists of Master-level games downloaded from PGN mentor, TWIC, and Caissabase in December 2023

  • GC22 consists of games from a Kaggle dataset, purported to contain chess games played by Grandmaster-titled players on chess.com up to July 2022

The lowercase letters at the end of every dataset denote specific filters that were applied to games from the corresponding PGN filesets:

  • "c for games that ended in a checkmate
  • "t" for games that used a specific time control
  • "d" for games that ended decisively

ML23c

This consists of Master-level games downloaded from PGN mentor, TWIC, and Caissabase in December 2023.

On this data (11,081,724 games), we apply the following filters to keep only those games that:

  • are unique (5,213,634 games)
  • and ended in a checkmate (250,694 games)

These 250,694 games consist of a total 10,797,366 half-moves made by the winners of the games, which alone constitute the dataset. For each such half-move, the chessboard, turn (white or black), and castling rights of both players before the move are calculated, as well as the sequence of half-moves beginning with this half-move up to 10 half-moves into the future. Draw potential is not calculated.

Download here. The data is zipped and will need to be extracted.

It consists of the following files:

  • ML23c.h5, an HDF5 file containing two tables, one with the raw data and the other encoded with indices (that will be used in the transformer model), containing the following fields:
    • board_position, the chessboard layout, or positions of pieces on the board
    • turn, the color of the pieces of the player to play
    • white_kingside_castling_rights, whether white can castle kingside
    • white_queenside_castling_rights, whether white can castle queenside
    • black_kingside_castling_rights, whether black can castle kingside
    • black_queenside_castling_rights, whether black can castle queenside
    • moves, 10 half-moves into the future made by both players
    • length, the number of half-moves in the sequence, as this will be less than 10 at the end of the game

LE22ct

This consists of games from the Lichess Elite Database put together by nikonoel, a collection of all standard chess games played on Lichess.org by players with a Lichess Elo rating of 2400+ against players with a Lichess Elo rating of 2200+ up to December 2021, and players rated 2500+ against players rated 2300+ from December 2021 up to December 2022.

On this data (20,241,368 games), we apply the following filters to keep only those games that:

  • used a time control of at least 5 minutes (2,073,780 games)
  • and ended in a checkmate (274,794 games)

These 274,794 games consist of a total 13,287,522 half-moves made by the winners of the games, which alone constitute the dataset. For each such half-move, the chessboard, turn (white or black), and castling rights of both players before the move are calculated, as well as the sequence of half-moves beginning with this half-move up to 10 half-moves into the future. Draw potential is not calculated.

Download here. The data is zipped and will need to be extracted.

It consists of the following files:

  • LE22ct.h5, an HDF5 file containing two tables, one with the raw data and the other encoded with indices (that will be used in the transformer model), containing the following fields:
    • board_position, the chessboard layout, or positions of pieces on the board
    • turn, the color of the pieces of the player to play
    • white_kingside_castling_rights, whether white can castle kingside
    • white_queenside_castling_rights, whether white can castle queenside
    • black_kingside_castling_rights, whether black can castle kingside
    • black_queenside_castling_rights, whether black can castle queenside
    • moves, 10 half-moves into the future made by both players
    • length, the number of half-moves in the sequence, as this will be less than 10 at the end of the game

GC22c

This consists of games from a Kaggle dataset, purported to contain chess games played by Grandmaster-titled players on chess.com up to July 2022.

On this data (4,811,076 games), we apply the following filters to keep only those games that:

  • are unique (4,178,495 games)
  • and ended in a checkmate (593,693 games)

These 593,693 games consist of a total 26,162,415 half-moves made by the winners of the games, which alone constitute the dataset. For each such half-move, the chessboard, turn (white or black), and castling rights of both players before the move are calculated, as well as the sequence of half-moves beginning with this half-move up to 10 half-moves into the future. Draw potential is not calculated.

Download here. The data is zipped and will need to be extracted.

It consists of the following files:

  • GC22c.h5, an HDF5 file containing two tables, one with the raw data and the other encoded with indices (that will be used in the transformer model), containing the following fields:
    • board_position, the chessboard layout, or positions of pieces on the board
    • turn, the color of the pieces of the player to play
    • white_kingside_castling_rights, w(hether white can castle kingside
    • white_queenside_castling_rights, whether white can castle queenside
    • black_kingside_castling_rights, whether black can castle kingside
    • black_queenside_castling_rights, whether black can castle queenside
    • moves, 10 half-moves into the future made by both players
    • length, the number of half-moves in the sequence, as this will be less than 10 at the end of the game

LE22c

This is an extended version of LE22ct, and consists of games from the Lichess Elite Database put together by nikonoel, a collection of all standard chess games played on Lichess.org by players with a Lichess Elo rating of 2400+ against players with a Lichess Elo rating of 2200+ up to December 2021, and players rated 2500+ against players rated 2300+ from December 2021 up to December 2022.

On this data (20,241,368 games), we apply the following filters to keep only those games that:

  • ended in a checkmate (2,751,394 games)

These 2,751,394 games consist of a total 127,684,720 half-moves made by the winners of the games, which alone constitute the dataset. For each such half-move, the chessboard, turn (white or black), and castling rights of both players before the move are calculated, as well as the sequence of half-moves beginning with this half-move up to 10 half-moves into the future. Draw potential is not calculated.

Download here. The data is zipped and will need to be extracted.

It consists of the following files:

  • LE22c.h5, an HDF5 file containing two tables, one with the raw data and the other encoded with indices (that will be used in the transformer model), containing the following fields:
    • board_position, the board layout or positions of pieces on the board
    • turn, the color of the pieces of the player to play
    • white_kingside_castling_rights, whether white can castle kingside
    • white_queenside_castling_rights, whether white can castle queenside
    • black_kingside_castling_rights, whether black can castle kingside
    • black_queenside_castling_rights, whether black can castle queenside
    • moves, 10 half-moves into the future made by both players
    • length, the number of half-moves in the sequence, as this will be less than 10 at the end of the game

ML23d

This consists of Master-level games downloaded from PGN mentor, TWIC, and Caissabase in December 2023.

On this data (11,081,724 games), we apply the following filters to keep only those games that:

  • are unique (5,213,634 games)
  • and are decisive, i.e. a player won (3,739,604 games)

These 3,739,604 games consist of a total 144,625,397 half-moves made by the winners of the games, which alone constitute the dataset. For each such half-move, the chessboard, turn (white or black), and castling rights of both players before the move are calculated, as well as the sequence of half-moves beginning with this half-move up to 10 half-moves into the future. Draw potential is not calculated.

Download here. The data is zipped and will need to be extracted.

It consists of the following files:

  • ML23d.h5, an HDF5 file containing two tables, one with the raw data and the other encoded with indices (that will be used in the transformer model), containing the following fields:
    • board_position, the chessboard layout, or positions of pieces on the board
    • turn, the color of the pieces of the player to play
    • white_kingside_castling_rights, whether white can castle kingside
    • white_queenside_castling_rights, whether white can castle queenside
    • black_kingside_castling_rights, whether black can castle kingside
    • black_queenside_castling_rights, whether black can castle queenside
    • moves, 10 half-moves into the future made by both players
    • length, the number of half-moves in the sequence, as this will be less than 10 at the end of the game

Play

After installing Chess Transformers, you can play games against an available model or have a model play against a chess engine.

You v. Model

You could either play in a Jupyter notebook (recommended for better UI) or in a Python shell.

import os
from chess_transformers.configs import import_config
from chess_transformers.play.utils import write_pgns
from chess_transformers.play import load_model, warm_up, human_v_model 

# Load configuration
config_name = "CT-EFT-85"
CONFIG = import_config(config_name)

# Load assets
model = load_model(CONFIG)

# Warmup model (triggers compilation)
warm_up(
    model=model
)

# Play
wins, losses, draws, pgns = human_v_model(
    human_color="b",  # color you want to play
    model=model,
    k=1,  # "k" in "top_k sampling", k=1 is best
    use_amp=True,
    rounds=1,  # number of rounds you want to play
    clock=None, 
    white_player_name=config_name,
    black_player_name="Me",
)

# Print games in Portable Game Notation (PGN) format
print(pgns)

# Save PGNs if you wish
write_pgns(
    pgns,
    pgn_file="somewhere/something.pgn",
)

You could also just make a copy of human_play.ipynb and play in that notebook.

Model v. Engine

The process is the same as above, except you must use a different set of functions:

from chess_transformers.play import model_v_engine
from chess_transformers.play.utils import load_engine

# Load engine
engine = load_engine(CONFIG.FAIRY_STOCKFISH_PATH)

# Play
LL = 1  # Try strength levels 1 to 8 (note: 7 and 8 may be slow)
model_color = "w"  # Try "w" and "b"
wins, losses, draws, pgns = model_v_engine(
    model=model,
    k=CONFIG.SAMPLING_K,
    use_amp=CONFIG.USE_AMP,
    model_color=model_color,
    engine=engine,
    time_limit=CONFIG.LICHESS_LEVELS[LL]["TIME_CONSTRAINT"],
    depth_limit=CONFIG.LICHESS_LEVELS[LL]["DEPTH"],
    uci_options={"Skill Level": CONFIG.LICHESS_LEVELS[LL]["SKILL"]},
    rounds=500,
    clock=None,
    white_player_name="Fairy Stockfish @ LL {}".format(LL)
    if model_color == "b"
    else config_name,
    black_player_name="Fairy Stockfish @ LL {}".format(LL)
    if model_color == "w"
    else config_name,
    event=config_name + " v. Fairy Stockfish @ LL {}".format(LL)
    if model_color == "w"
    else "Fairy Stockfish @ LL {} v. ".format(LL) + config_name,
)

See evaluate.py for an example.

Time Control

If you're using a Unix-type operating system — basically, not Windows — you can also set a time control for your games. Currently, only Fischer time control is available.

from chess_transformers.play.clocks import ChessClock

clock = ChessClock(base_time=60, 
                   increment=1)

Pass this clock to the functions above instead of clock=None.

Train Models

You're welcome to try to train your own models, but if you wish to contribute trained models, please discuss first.

Dataset

You can skip this step if you wish to use one of the existing datasets.

  • Collect PGN files containing games you wish to use for training the model.

  • Create a bash script for parsing these PGN files into a collection of FENs and moves using pgn-extract, like in LE22ct.sh, and execute it in the folder with the PGN files.

  • Create a configuration file for the dataset, like in LE22ct.py.

  • Run prep.py like python prep.py [config_name], or do it in your own Python script.

from chess_transformers.data import prepare_data
from chess_transformers.configs import import_config

# Load configuration
CONFIG = import_config("[config_name]")

# Prepare data
prepare_data(
    data_folder=CONFIG.DATA_FOLDER,
    h5_file=CONFIG.H5_FILE,
    max_move_sequence_length=CONFIG.MAX_MOVE_SEQUENCE_LENGTH,
    expected_rows=CONFIG.EXPECTED_ROWS,
    val_split_fraction=CONFIG.VAL_SPLIT_FRACTION,
)

Data files will be created in CONFIG.DATA_FOLDER.

Training

  • Create a configuration file for the model, like in CT-E-20.py.

  • Run train.py like python train.py [config_name], or do it in your own Python script.

from chess_transformers.train import train_model
from chess_transformers.configs import import_config

# Load configuration
CONFIG = import_config("[config_name]")

# Train model
train_model(CONFIG)
  • Monitor training with TensorBoard with tensorboard --logdir $CT_LOGS_DIR.

Evaluation

Run evaluate.py like python evaluate.py [config_name], or do it in your own Python notebook/script.

from chess_transformers.configs import import_config
from chess_transformers.evaluate import evaluate_model

# Load configuration
CONFIG = import_config("[config_name]")

# Evaluate model
evaluate_model(CONFIG)

Contribute

Contributions — and any discussion thereof — are welcome. As you may have noticed, Chess Transformers is in initial development and the public API is not to be considered stable.

If you are planning to contribute bug-fixes, please go ahead and do so. If you are planning to contribute in a way that extends Chess Transformers, or adds any new features, data, or models, please create a discussion thread to discuss it before you spend any time on it. Otherwise, your PR may be rejected due to lack of consensus or alignment with current goals.

Presently, the following types of contributions may be useful:

  • Better, more robust evaluation methods of models.
  • Evaluation of existing models against chess engines on different CPUs to study the effect of CPU specifications on engine strength and evaluation.
  • New models with:
    • the same transformer architectures but of a larger size, and trained on larger datasets.
    • or different transformer architectures or internal mechanisms.
    • or in general, improved evaluation scores.
  • Chess clocks for Windows OS, or for Unix-type OS but for time controls other than Fischer time control.
  • Refactoring of code that improves its ease of use.
  • Model visualization for explainable AI, such as visualizing positional and move embeddings, or attention patterns.

This list is not exhaustive. Please do not hesitate to discuss your ideas. Thank you!

License

Chess Transformers is licensed under the MIT license.