English.md

Simple Usage

Requirements

1. PyTorch (1.x version)
2. NumPy
3. Tkinter
4. Matplotlib

Open With Built-in GUI

You may download the pre-training model from release section named 預先訓練好的小型權重. Then put the pt file to the dlgo directory and enter following command.

$ python3 dlgo.py --weights nn_2x64.pt --gui

The dlgo will use your GPU automatically. If you want to disable GPU, set the value USE_GPU False.

Open With GTP GUI

The dlgo support the GTP GUI. Sabaki is recommanded. Some helpful optional arguments are here.

optional arguments:
  -p <integer>, --playouts <integer>
                        The number of playouts.
  -w <string>, --weights <string>
                        The weights file name.
  -r <float>, --resign-threshold <float>
                        Resign when winrate is less than x.

The sample command is here.

$ dlgo.py --weights nn_2x64.pt -p 1600 -r 0.25

Training

Following above simple steps to train a new weights

1. Preparing the sgf files. You may just use the sgf.zip. The zip including around 35000 9x9 games.

2. Set the network parametes in the config.py. Including BOARD_SIZE, BLOCK_SIZE, FILTER_SIZE.

3. Start training.

    $ python3 train.py --dir sgf-directory-name --steps 128000 --batch-size 512 --learning-rate 0.005
   

Some helpful arguments are here.

optional arguments:
  -h, --help            show this help message and exit
  -d <string>, --dir <string>
                        The input SGF files directory. Will use data cache if set None.
  -s <integer>, --steps <integer>
                        Terminate after these steps.
  -v <integer>, --verbose-steps <integer>
                        Dump verbose on every X steps.
  -b <integer>, --batch-size <integer>
                        The batch size number.
  -l <float>, --learning-rate <float>
                        The learning rate.
  --noplot              Disable plotting.