def main():
"""Run DQN until the environment throws an exception."""
env = AllowBacktracking(make_env(stack=False, scale_rew=False))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 4)
optim, optimize = dqn.optimize(learning_rate=0.0001)
sess.run(tf.global_variables_initializer())
dqn.train(
num_steps=3000000, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=PrioritizedReplayBuffer(500000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=1024,
batch_size=16,
min_buffer_size=20000)
def main():
"""Run DQN until the environment throws an exception."""
parser = argparse.ArgumentParser()
parser.add_argument('--game', help='game: Use retro.data.list_games() to see a list of available games', type=str, default="SuperMarioBros-Nes")
parser.add_argument('--state', help='game state: Use retro.data.list_states(game) to see a list of available starting states', type=str, default="Level1-1")
parser.add_argument('--num_steps', help='The number of steps to train the model.', default=3000000, type=int)
parser.add_argument('--resume_training', help='Resume training the most recent model', action='store_true')
parser.add_argument('--show_gameplay', help='Display the agent playing the game in realtime', action='store_true')
parser.add_argument('--save_screens', help='save screen shots to $DATA_DIR/screenshots/[GAME]/[STATE]/[RUN_ID]/', action='store_true')
parser.add_argument('--generate_map', help='Generate a level map', action='store_true')
parser.add_argument('--show_map', help='Show the level map panorama as it is generated', action='store_true')
parser.add_argument('--show_map_matches', help='Show the keypoint matches in the level map. Implies --show_map=True', action='store_true')
parser.add_argument('-l', help='list games and states', action='store_true')
tensorboard_dir = path.join(environ.get('DATA_DIR', environ.get('HOME', '.')), 'tensorboard')
writer = tf.summary.FileWriter(tensorboard_dir)
args = parser.parse_args()
if args.l:
import retro
games = retro.data.list_games()
for game in games:
print('{}: {}'.format(game, retro.data.list_states(game)))
sys.exit(0)
game = args.game
state = args.state
resume = args.resume_training
env = AllowBacktracking(make_env(stack=False, scale_rew=False, game=game, state=state))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 4)
model_number = get_newest_model(game, state)
print('model_number: {}'.format(model_number))
print('resume: {}'.format(resume))
if resume and model_number is not None:
model_name = path.join(get_models_dir(game, state), '{}-{}'.format(state, model_number))
saver = tf.train.Saver()
saver.restore(sess, model_name)
print('loaded model {}'.format(model_name))
else:
model_number = 0
optim, optimize = dqn.optimize(learning_rate=0.0001)
#import pdb; pdb.set_trace()
#print(sess.graph.get_tensor_by_name('layer_1:0'))
#sys.exit(0)
if resume and model_number > 0:
print('resuming at model number {}'.format(model_number))
sess.run(tf.variables_initializer(optim.variables()))
else:
print('creating a new model')
sess.run(tf.global_variables_initializer())
writer.add_graph(sess.graph)
dqn.train(num_steps=args.num_steps, # Make sure an exception arrives before we stop.
initial_step=model_number,
player=player,
replay_buffer=PrioritizedReplayBuffer(500000, 0.5, 0.4, epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=1024,
batch_size=16,
min_buffer_size=20000,
save_iters=2048,
game=game,
state=state,
generate_map=args.generate_map,
show_gameplay=args.show_gameplay,
show_map=args.show_map,
show_map_matches=args.show_map_matches)
