Skip to content

README.md

Latest commit

 

History

History
57 lines (56 loc) · 4.39 KB

README.md

File metadata and controls

57 lines (56 loc) · 4.39 KB

How to train your own odds network

Training data generation

1. Option: Fastchess (recommended)

  1. Download Fastchess from here.
  2. start generating games with:
./fastchess-ubuntu-22.04 -engine cmd=lc0/build/release/lc0 name=lqo_v2_2000 nodes=2000 args="--weights=lqo_v2.pb.gz --temperature=0.8 --tempdecay-moves=8 --temp-cutoff-move=8 --temp-endgame=0.05" -engine cmd=lc0/build/release/lc0 name=EliteLeela_22 nodes=22 args="--weights=Elite-Leela.pb.gz --temperature=1.5 --tempdecay-moves=20 --temp-cutoff-move=20 --temp-endgame=0.05" -openings file=queen500.epd format=epd -rounds 500 -games 1 -concurrency 2 -pgnout file=lqo-training-data.pgn -noswap

make sure to adjust these parameters to reach a win-rate of >50% <60%. It is also highly recommended to use an opening book for generations 2+, to avoid overfitting on the training data. The -noswap parameter is also very important here, as this prevents the engines from switching the sides, which would defeat the purpose of the data generation for odds-play.

2. Option: Using my training data generation script

  1. Install libraries with pip install chess dropbox (dropbox is optional)
  2. Generate training data using the generate_games.py script in the training folder. Depending on your exact odds, you will have to make some adjustments to the config_v0.json file. The v0 file contains the exact configuration I used for the first version of this network. You will have to adjust the fen position or use an opening book, and adjust the settings such that you end up getting games with good variability and a win-rate of slightly more than 50%. I recommend generating 130k games for each of the iterations. Once you have the right configuration, start the script with:
python generate_games.py -c config_v0.json

Converting training data

  1. Now convert the pgn into training-data using the training-data tool.

Installing training environment

  1. Install tensorflow for executing training code, for me it was tf-2.10. Make sure you have miniconda installed and run:
conda create -n tf_210 python=3.10
conda activate tf_210
conda install tensorflow-gpu==2.10.0
pip install tensorflow-addons==0.20.0
pip install pyyaml

this should install all the libraries including the right cuda and cudnn libraries. Then bind these libraries with:

conda install -c nvidia cuda-nvcc --yes
# Configure the XLA cuda directory
mkdir -p $CONDA_PREFIX/etc/conda/activate.d
printf 'export XLA_FLAGS=--xla_gpu_cuda_data_dir=$CONDA_PREFIX/lib/\n' >> $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh
source $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh
# Copy libdevice file to the required path
mkdir -p $CONDA_PREFIX/lib/nvvm/libdevice
cp $CONDA_PREFIX/lib/libdevice.10.bc $CONDA_PREFIX/lib/nvvm/libdevice/

then clone the training code and compile protobuf files:

git clone --recurse-submodules https://github.com/LeelaChessZero/lczero-training.git
cd lczero-training/
./init.sh

Training the network

  1. Download the base-network T82 and convert the model for training (if generating second+ generations skip this step, or provide the older generations net to convert, if you dont have the training setup anymore used for these):
python net_to_model.py --ignore-errors --cfg=training/768x15x24h-t80_lqo.yaml net/768x15x24h-t82-swa-7464000.pb.gz

edit the input-path of the yaml config, so that it points to your training-data generated earlier with the trainingdata-tool. and edit path to point to where the converted base model is. Make sure you see in console output that it gets loaded, otherwise it will train from scratch. then start training with:

python train.py --cfg training/768x15x24h-t80_lqo.yaml

Once its done you should have QUEEN_ODDS-swa-10000.pb.gz file in your path. You should also have a non-swa version, but I recommend using that version.

Then repeat this process once but this time use the newly created network for generating games. I used the config_v1, but again, depending on your odds you might want to generate games against another opponent or with other settings. For the second training run, skip converting the net_to_model.py and, only change the data in the input-path, as it will then automatically resume training from your latest checkpoint.