backgammon/main.py

import argparse
import sys
import os
import time

# Define helper functions
def log_train_outcome(outcome, trained_eps = 0):
    format_vars = { 'trained_eps': trained_eps,
                    'count': len(train_outcome),
                    'sum': sum(train_outcome),
                    'mean': sum(train_outcome) / len(train_outcome),
                    'time': int(time.time())
    }
    with open(os.path.join(config['model_path'], 'logs', "train.log"), 'a+') as f:
        f.write("{time};{trained_eps};{count};{sum};{mean}".format(**format_vars) + "\n")
    

def log_eval_outcomes(outcomes, trained_eps = 0):
    for outcome in outcomes:
        scores = outcome[1]
        format_vars = { 'trained_eps': trained_eps,
                        'method': outcome[0],
                        'count': len(scores),
                        'sum': sum(scores),
                        'mean': sum(scores) / len(scores),
                        'time': int(time.time())
        }
        with open(os.path.join(config['model_path'], 'logs', "eval.log"), 'a+') as f:
            f.write("{time};{method};{trained_eps};{count};{sum};{mean}".format(**format_vars) + "\n")

            
# Parse command line arguments
parser = argparse.ArgumentParser(description="Backgammon games")
parser.add_argument('--episodes', action='store', dest='episode_count',
                    type=int, default=1000,
                    help='number of episodes to train')
parser.add_argument('--model-path', action='store', dest='model_path',
                    default='./model',
                    help='path to Tensorflow model')
parser.add_argument('--eval-methods', action='store',
                    default=['random'], nargs='*',
                    help='specifies evaluation methods')
parser.add_argument('--eval', action='store_true',
                    help='whether to evaluate the neural network with a random choice bot')
parser.add_argument('--train', action='store_true',
                    help='whether to train the neural network')
parser.add_argument('--eval-after-train', action='store_true', dest='eval_after_train',
                    help='whether to evaluate after each training session')
parser.add_argument('--play', action='store_true',
                    help='whether to play with the neural network')
parser.add_argument('--start-episode', action='store', dest='start_episode',
                    type=int, default=0,
                    help='episode count to start at; purely for display purposes')

args = parser.parse_args()

config = {
    'model_path': args.model_path,
    'episode_count': args.episode_count,
    'eval_methods': args.eval_methods,
    'train': args.train,
    'play': args.play,
    'eval': args.eval,
    'eval_after_train': args.eval_after_train,
    'start_episode': args.start_episode
}

# Make sure directories exist
model_path = os.path.join(config['model_path'])
log_path   = os.path.join(model_path, 'logs')
if not os.path.isdir(model_path):
    os.mkdir(model_path)
if not os.path.isdir(log_path):
    os.mkdir(log_path)
    

# Set up game
import game
g = game.Game(config = config)
g.set_up_bots()


# Set up variables
episode_count = config['episode_count']


# Do actions specified by command-line
if args.train:
    eps = config['start_episode']
    while True:
        train_outcome = g.train_model(episodes = episode_count, trained_eps = eps)
        eps += episode_count
        log_train_outcome(train_outcome, trained_eps = eps)
        if config['eval_after_train']:
            eval_outcomes = g.eval(trained_eps = eps)
            log_eval_outcomes(eval_outcomes, trained_eps = eps)
elif args.eval:
    eps = config['start_episode']
    outcomes = g.eval()
    log_eval_outcomes(outcomes, trained_eps = eps)
#elif args.play:
#    g.play(episodes = episode_count)