def run(args):
logging.basicConfig(filename=args.LOG_FILE, level=logging.DEBUG)
logging.getLogger().addHandler(logging.StreamHandler())
game_handler = GameStateHandler(random_seed=123, frame_skip=args.FRAME_SKIP, use_sdl=False,
image_processing=lambda x: crop_and_resize(x, args.IMAGE_HEIGHT, args.IMAGE_WIDTH))
game_handler.loadROM(args.ROM_FILE)
height, width = game_handler.getScreenDims()
logging.info('Screen resolution is %dx%d' % (height, width))
num_actions = game_handler.num_actions
net = theano_qnetwork.DeepQNetwork(args.IMAGE_HEIGHT, args.IMAGE_WIDTH, num_actions, args.STATE_FRAMES, args.DISCOUNT_FACTOR)
replay_memory = ReplayMemoryManager(args.IMAGE_HEIGHT, args.IMAGE_WIDTH, args.STATE_FRAMES, args.REPLAY_MEMORY_SIZE)
monitor = Monitoring(log_train_step_every=100, smooth_episode_scores_over=50)
agent = Agent(game_handler, net, replay_memory, None, monitor, args.TRAIN_FREQ, batch_size=args.BATCH_SIZE)
start_epsilon = args.START_EPSILON
exploring_duration = args.EXPLORING_DURATION
agent.populate_replay_memory(args.MIN_REPLAY_MEMORY)
agent.play(train_steps_limit=args.LEARNING_BEYOND_EXPLORING+args.EXPLORING_DURATION, start_eps=start_epsilon,
final_eps=args.FINAL_EPSILON, exploring_duration=exploring_duration)
def main(_=None):
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
d = os.path.dirname(args.log_file)
if not os.path.exists(d):
os.makedirs(d)
if not args.continue_training:
with open(args.log_file, 'w') as f:
f.write('')
logging.basicConfig(filename=args.log_file, level=logging.DEBUG)
logging.getLogger().addHandler(logging.StreamHandler())
game_handler = GameStateHandler(
args.rom_directory + args.rom_name,
random_seed=args.random_seed,
frame_skip=args.frame_skip,
use_sdl=args.use_sdl,
repeat_action_probability=args.repeat_action_probability,
minimum_actions=args.minimum_action_set,
test_mode=args.test_mode,
image_processing=lambda x: crop_and_resize(x, args.image_height, args.image_width, args.cut_top))
num_actions = game_handler.num_actions
if args.optimizer == 'rmsprop':
optimizer = tf.train.RMSPropOptimizer(
learning_rate=args.learning_rate,
decay=args.decay,
momentum=0.0,
epsilon=args.rmsprop_epsilon)
if not args.multi_gpu:
if args.double_dqn:
net = qnetwork.DualDeepQNetwork(args.image_height, args.image_width, sess, num_actions,
args.state_frames, args.discount_factor, args.target_net_refresh_rate,
net_type=args.net_type, optimizer=optimizer)
else:
net = qnetwork.DeepQNetwork(args.image_height, args.image_width, sess, num_actions, args.state_frames,
args.discount_factor, net_type=args.net_type, optimizer=optimizer)
else:
net = multi_gpu_qnetwork.MultiGPUDualDeepQNetwork(args.image_height, args.image_width, sess, num_actions,
args.state_frames, args.discount_factor,
optimizer=optimizer, gpus=[0, 1, 2, 3])
saver = Saver(sess, args.data_dir, args.continue_training)
if saver.replay_memory_found():
replay_memory = saver.get_replay_memory()
else:
if args.test_mode:
logging.error('NO SAVED NETWORKS IN TEST MODE!!!')
replay_memory = ReplayMemoryManager(args.image_height, args.image_width, args.state_frames,
args.replay_memory_size, reward_clip_min=args.reward_clip_min,
reward_clip_max=args.reward_clip_max)
# todo: add parameters to handle monitor
monitor = Monitoring(log_train_step_every=100, smooth_episode_scores_over=50)
agent = Agent(
game_handler=game_handler,
qnetwork=net,
replay_memory=replay_memory,
saver=saver,
monitor=monitor,
train_freq=args.train_freq,
test_mode=args.test_mode,
batch_size=args.batch_size,
save_every_x_episodes=args.saving_freq)
sess.run(tf.initialize_all_variables())
saver.restore(args.data_dir)
start_epsilon = max(args.final_epsilon,
args.start_epsilon - saver.get_start_frame() * (args.start_epsilon - args.final_epsilon) / args.exploration_duration)
exploring_duration = max(args.exploration_duration - saver.get_start_frame(), 1)
if args.test_mode:
agent.populate_replay_memory(args.state_frames, force_early_stop=True)
agent.play_in_test_mode(args.epsilon_in_test_mode)
else:
agent.populate_replay_memory(args.min_replay_memory)
agent.play(train_steps_limit=args.number_of_train_steps, start_eps=start_epsilon,
final_eps=args.final_epsilon, exploring_duration=exploring_duration)
