Skip to content

Latest commit

 

History

History
93 lines (78 loc) · 4.33 KB

README.md

File metadata and controls

93 lines (78 loc) · 4.33 KB
Raw

nn-train

Chess (and variants) neural network (NN) and Efficiently Updatable Neural Network (NNUE) training program. This program takes labelled epd positions and trains a neural network to predict the outcome and/or move choice for a position.

  usage: train.py [-h] [--epd EPD] [--dir DIR] [--id ID]
                  [--global-steps GLOBAL_STEPS] [--batch-size BATCH_SIZE]
                  [--learning-rate LR] [--validation-split VALD_SPLIT]
                  [--cores CORES] [--gpus GPUS] [--gzip] [--net NET]
                  [--rsav RSAV] [--rsavo RSAVO] [--rand] [--opt OPT]
                  [--policy-channels POL_CHANNELS] [--policy-weight POL_W]
                  [--value-weight VAL_W] [--score-weight SCORE_W]
                  [--policy-gradient POL_GRAD] [--no-auxinp]
                  [--channels CHANNELS] [--boardx BOARDX] [--boardy BOARDY]
                  [--frac-z FRAC_Z] [--frac-pi FRAC_PI] [--piece-map PCMAP]
                  [--mixed] [--head-type HEAD_TYPE] [--max-steps MAX_STEPS]

  optional arguments:
    -h, --help            show this help message and exit
    --epd EPD, -e EPD     Path to labeled EPD file for training
    --dir DIR             Path to network files
    --id ID, -i ID        ID of neural networks to load.
    --global-steps GLOBAL_STEPS
                          Global number of steps trained so far.
    --batch-size BATCH_SIZE, -b BATCH_SIZE
                          Training batch size.
    --learning-rate LR, -l LR
                          Training learning rate.
    --validation-split VALD_SPLIT
                          Fraction of sample to use for validation.
    --cores CORES         Number of cores to use.
    --gpus GPUS           Number of gpus to use.
    --gzip, -z            Process zipped file.
    --net NET             Net to train from
                          0=2x32,6x64,12x128,20x256,4=30x384,5=NNUE.
    --rsav RSAV           Save graph every RSAV steps.
    --rsavo RSAVO         Save optimization state every RSAVO steps.
    --rand                Generate random network.
    --opt OPT             Optimizer 0=SGD 1=Adam.
    --policy-channels POL_CHANNELS
                          Number of policy channels
    --policy-weight POL_W
                          Policy loss weight.
    --value-weight VAL_W  Value loss weight.
    --score-weight SCORE_W
                          Score loss weight.
    --policy-gradient POL_GRAD
                          0=standard 1=multiply policy by score.
    --no-auxinp, -u       Don't use auxillary input
    --channels CHANNELS, -c CHANNELS
                          number of input channels of network.
    --boardx BOARDX, -x BOARDX
                          board x-dimension.
    --boardy BOARDY, -y BOARDY
                          board y-dimension.
    --frac-z FRAC_Z       Fraction of ouctome(Z) relative to MCTS value(Q).
    --frac-pi FRAC_PI     Fraction of MCTS policy (PI) relative to one-hot
                          policy(P).
    --piece-map PCMAP     Map pieces to planes
    --mixed               Use mixed precision training
    --head-type HEAD_TYPE
                          Heads of neural network, 0=value/policy,
                          1=value/score, 2=all three, 3=value only.
    --max-steps MAX_STEPS
                          Maximum number of steps to train for.

To train 2x32 networks from a gzipped labelled epd with result and best moves using 16 cpu cores and 1 gpu

python src/train.py --dir nets --gzip --epd quiet.epd.gz --net 0 --cores 16 --gpus 1

Then to convert your keras model to protobuf tensorflow format:

./scripts/convert-to-pb.sh nets/ID-1-model-0

To also convert to UFF format use

./scripts/prepare.sh nets 1 0

To restart interrupted training from specific ID e.g. 120

python src/train.py --epd quiet.epd --id 120

You can build your own network (different number of blocks and filters) by modifying resnet.py.

Self play training

To train networks by reinforcement learning issue command

./train.sh 3

This will train networks 20x256 resnet using selfplay games produced by the 20x256 network. The net used for producing selfplay games is mentioned first