**MOVED TO https://github.com/LeelaChessZero/leela-chess ** A chess adaption of GCP's Leela Zero
**MOVED TO https://github.com/LeelaChessZero/leela-chess ** A chess adaption of GCP's Leela Zero
This is an adaptation of GCP's Leela Zero repository to chess, using Stockfish's position representation and move generation. (No heuristics or prior knowledge are carried over from Stockfish.)
The goal is to build a strong UCT chess AI following the same type of techniques as AlphaZero, as described in Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm.
We will need to do this with a distributed project, as it requires a huge amount of computations.
Please visit the LCZero forum to discuss: https://groups.google.com/forum/#!forum/lczero, or the github issues.
For precompiled binaries, see:
For live status: http://lczero.org
The rest of this page is for users who want to compile the code themselves. Of course, we also appreciate code reviews, pull requests and Windows testers!
# Install dependencies
sudo apt install cmake g++ git libboost-all-dev libopenblas-dev opencl-headers ocl-icd-libopencl1 ocl-icd-opencl-dev zlib1g-dev
# Test for OpenCL support & compatibility
sudo apt install clinfo && clinfo
# Clone git repo
git clone https://github.com/glinscott/leela-chess.git
cd leela-chess
git submodule update --init --recursive
mkdir build && cd build
# Configure
cmake ..
# Or configure without GPU support
cmake -DFEATURE_USE_CPU_ONLY=1 ..
# Build and run tests
make
./testsSee https://github.com/glinscott/leela-chess/tree/master/go/src/client/README.md. This client will produce self-play games and upload them to http://lczero.org. A central server uses these self-play game data as inputs for the training process.
The weights from the distributed training are downloadable from http://lczero.org/networks. The best one is the top network that has some games played on it.
Weights that we trained to prove the engine was solid are here https://github.com/glinscott/lczero-weights. The best weights obtained through supervised learning on a human dataset were with elo ratings > 2000.
The training pipeline resides in training/tf, this requires tensorflow running on linux (Ubuntu 16.04 in this case).
In order to start a training session you first need to download trainingdata from http://lczero.org/training_data. This data is packed in tar.gz balls each containing 10'000 games or chunks as we call them. Preparing data requires the following steps:
tar -xzf games11160000.tar.gz
ls training.* | parallel gzip {}This repacks each chunk into a gzipped file ready to be parsed by the training pipeline. Note that the parallel command uses all your cores and can be installed with apt-get install parallel.
Now that the data is in the right format one can configure a training pipeline. This configuration is achieved through a yaml file, see training/tf/configs/example.yaml:
…The configuration is pretty self explanatory, if you're new to training I suggest looking at the machine learning glossary by google. Now you can invoke training with the following command:
./train.py --cfg configs/example.yaml --output /tmp/mymodel.txtThis will initialize the pipeline and start training a new neural network. You can view progress by invoking tensorboard:
tensorboard --logdir leelalogsIf you now point your browser at localhost:6006 you'll see the trainingprogress as the trainingsteps pass by. Have fun!
The training pipeline will automatically restore from a previous model if it exists in your training:path as configured by your yaml config. For initializing from a raw weights.txt file you can use training/tf/net_to_model.py, this will create a checkpoint for you.
Generating trainingdata from pgn files is currently broken and has low priority, feel free to create a PR.
The code is released under the GPLv3 or later, except for ThreadPool.h, cl2.hpp and the clblast_level3 subdir, which have specific licenses (compatible with GPLv3) mentioned in those files.
No open issues yet, or sync has not completed.