conf.py

from math import sqrt
import os
conf = {
    'ROOT_DIR': os.path.dirname(__file__),
    'MODEL_DIR': 'models',
    'LOG_DIR': 'logs',
    'GAMES_DIR': 'games',
    'BEST_MODEL': 'best_model.h5',
    'MOVE_INDEX': 'move_index.csv',
    'GAME_FILE': 'game_%03d.h5',
    'SHOW_EACH_MOVE': False,
    'SHOW_END_GAME': False,

    ### MODEL ###
    'N_RESIDUAL_BLOCKS': 20, # Size of the tower of residual blocks, 20 for small model, 40 for alphagozero full size
    'L2_EPSILON': 1e-4,# The epsilon coefficient in the loss value function, 1e-4 in paper
    'MOMENTUM': 0.9,# 0.9 in paper
    'LEARNING_RATE': 1e-2,# 1e-2 in paper

    ### SELF-PLAY ###
    'N_GAMES': 2500,  # Number of games of self play generated by best_model, 25k in paper
    'MCTS_SIMULATIONS': 64, # 1.6k in paper
    'SIZE': 5, # board size 19 in paper
    'KOMI': 3.5, # The komi points given to white player
    'STOP_EXPLORATION': 4, # Number of plays after which temperature goes to 0 , 30 in paper
    'MCTS_BATCH_SIZE': 8, # Size of the prediction batch while exploring mcts
    'DIRICHLET_ALPHA': .03, # The value of dirichlet coefficient in the nois of root_node of mcts simulation
    'DIRICHLET_EPSILON': .25, # How much the noise is accounted for

    'RESIGNATION_PERCENT': .99, # % of the time we DO NOT use resignation to assess resignation value (10% in paper)
    'RESIGNATION_ALLOWED_ERROR': .05, # 5% of the time we resign a game we could have won

    ### TRAIN ###
    'TRAIN_BATCH_SIZE': 32,  # Batch size in the training phase, 32 in paper
    'EPOCHS_PER_SAVE': 1000, # A model will be saved to be evaluated this amount of epochs, 1000 in paper
    'NUM_WORKERS': 64,# We use this many GPU workers so split the task, 64 in paper
    'HISTOGRAM_FREQ': 0, # Shows the histograms in Tensorboard. For debugging
    'VALIDATION_SPLIT': 0, # Needed if you want histograms in Tensorboard.

    ### EVALUATOR ###
    'EVALUATE_N_GAMES': 100,# The number of games to test on to elect new best model, 400 in paper
    'EVALUATE_MARGIN': .5 + 1/sqrt(100),# Model has to win by that margin to be elected, 55% in paper
    'SGF_ENABLED': True,


    ### GTP ###
    'GTP_LOGFILE': 'gtp.log',
}