def test_run(self):
archivefile = tempfile.mktemp(suffix="nethack_test", prefix=".zip")
game = nethack.NetHack(archivefile=archivefile)
response = game.reset()
actions = [
nethack.MiscAction.MORE,
nethack.MiscAction.MORE,
nethack.MiscAction.MORE,
nethack.MiscAction.MORE,
nethack.MiscAction.MORE,
nethack.MiscAction.MORE,
]
for action in actions:
while not response.ProgramState().InMoveloop():
response, done, info = game.step(nethack.MiscAction.MORE)
response, done, info = game.step(action)
if done:
# Only the good die young.
response = game.reset()
obs = response.Observation()
chars = _fb_ndarray_to_np(obs.Chars())
glyphs = _fb_ndarray_to_np(obs.Glyphs())
status = response.Blstats()
x, y = status.CursX(), status.CursY()
self.assertEqual(np.count_nonzero(chars == ord("@")), 1)
self.assertEqual(chars[y, x], ord("@"))
mon = nethack.permonst(nethack.glyph_to_mon(glyphs[y][x]))
self.assertEqual(mon.mname, "monk")
self.assertEqual(mon.mlevel, 10)
class_sym = nethack.class_sym.from_mlet(mon.mlet)
self.assertEqual(class_sym.sym, "@")
self.assertEqual(class_sym.explain, "human or elf")
self.assertEqual(os.waitpid(info["pid"], os.WNOHANG), (0, 0))
del game # Should kill process.
with self.assertRaisesRegex(OSError, "No (child|such)? process"):
os.waitpid(info["pid"], 0)
def main():
episodes = 0
steps = 0
game = nethack.NetHack(archivefile="random.zip")
start = timeit.default_timer()
while True:
steps_delta = play(game, False)
time_delta = timeit.default_timer() - start
episodes += 1
steps += steps_delta
print("Episde: %i. Steps: %i. SPS: %f" %
(episodes, steps, steps_delta / time_delta))
start = timeit.default_timer()
def play(should_stop, queue):
game = nethack.NetHack(archivefile=None)
done = True
steps = 0
while not should_stop.is_set():
if done or steps >= 1000:
queue.put(steps)
steps = 0
observation = game.reset()
if observation.Internal() and observation.Internal().Xwaitforspace():
ch = nethack.MiscAction.MORE
else:
ch = random.choice(ACTIONS)
observation, done, _ = game.step(ch)
steps += 1
def __init__(
self,
savedir=None,
archivefile="nethack.%(pid)i.%(time)s.zip",
character="mon-hum-neu-mal",
max_episode_steps=5000,
observation_keys=("glyphs", "status", "message", "inventory"),
actions=None,
options=None,
):
"""Constructs a new NLE environment.
Args:
savedir (str or None): path to save archives into. Defaults to None.
archivefile (str or None): Template for the zip archive filename of
NetHack ttyrec files. Use "%(pid)i" for the process id of the
NetHack process, "%(time)s" for the creation time. Use None to
disable writing archivefiles.
character (str): name of character. Defaults to "mon-hum-neu-mal".
max_episode_steps (int): maximum amount of steps allowed before the
game is forcefully quit. In such cases, ``info["end_status"]``
will be equal to ``StepStatus.ABORTED``. Defaults to 5000.
observation_keys (list): keys to use when creating the observation.
Defaults to all.
actions (list): list of actions. If None, the full action space will
be used, i.e. ``nle.nethack.ACTIONS``. Defaults to None.
options (list): list of game options to initialize NetHack. If None,
NetHack will be initialized with the options found in
``nle.nethack.NETHACKOPTIONS`. Defaults to None.
"""
self.character = character
self._max_episode_steps = max_episode_steps
if actions is None:
actions = FULL_ACTIONS
self._actions = actions
self.response = None
if archivefile is not None:
try:
if savedir is None:
parent_dir = os.path.join(os.getcwd(), "nle_data")
os.makedirs(parent_dir, exist_ok=True)
# Create a unique subdirectory for us.
self.savedir = tempfile.mkdtemp(
prefix=time.strftime("%Y%m%d-%H%M%S_"), dir=parent_dir)
else:
self.savedir = savedir
os.makedirs(self.savedir)
except FileExistsError:
logger.info("Using existing savedir: %s", self.savedir)
else:
logger.info("Created savedir: %s", self.savedir)
self.archivefile = os.path.join(self.savedir, archivefile)
else:
self.savedir = None
self.archivefile = None
self._stats_file = None
self._stats_logger = None
self._setup_statsfile = archivefile is not None
self.env = nethack.NetHack(
archivefile=self.archivefile,
options=options,
playername="Agent%(pid)i-" + self.character,
)
self._random = random.SystemRandom()
# -1 so that it's 0-based on first reset
self._episode = -1
space_dict = {
"glyphs":
gym.spaces.Box(
low=np.iinfo(np.int16).min,
high=np.iinfo(np.int16).max,
shape=DUNGEON_SHAPE,
dtype=np.int16,
),
"status":
gym.spaces.Box(
low=np.iinfo(np.int32).min,
high=np.iinfo(np.int32).max,
shape=(23, ),
dtype=np.int32,
),
"message":
gym.spaces.Box(
low=np.iinfo(np.uint8).min,
high=np.iinfo(np.uint8).max,
shape=(DEFAULT_MSG_PAD, ),
dtype=np.uint8,
),
"inventory":
gym.spaces.Tuple((
gym.spaces.Box(
low=np.iinfo(np.int16).min,
high=np.iinfo(np.int16).max,
shape=(DEFAULT_INV_PAD, ),
dtype=np.int16,
),
gym.spaces.Box(
low=np.iinfo(np.uint8).min,
high=np.iinfo(np.uint8).max,
shape=(DEFAULT_INV_PAD, DEFAULT_INVSTR_PAD),
dtype=np.uint8,
),
gym.spaces.Box(
low=np.iinfo(np.uint8).min,
high=np.iinfo(np.uint8).max,
shape=(DEFAULT_INV_PAD, ),
dtype=np.uint8,
),
gym.spaces.Box(
low=np.iinfo(np.uint8).min,
high=np.iinfo(np.uint8).max,
shape=(DEFAULT_INV_PAD, ),
dtype=np.uint8,
),
)),
}
self.observation_space = gym.spaces.Dict(
{key: space_dict[key]
for key in observation_keys})
self._key_functions = {
"glyphs": _get_glyphs,
"status": _get_status_fast,
"message": _get_padded_message,
"inventory": _get_padded_inv,
}
for key in list(self._key_functions.keys()):
if key not in observation_keys:
del self._key_functions[key]
self.action_space = gym.spaces.Discrete(len(self._actions))
def test_forking_with_nethack_in_parent(self, num_procs=NUM_SUBPROCESSES):
# Breaks for pyzmq <= 18.0.0, fixed in
# https://github.com/zeromq/pyzmq/commit/28c2a36836fc45c09ede4d9962498db449b642d1 # noqa
env = nethack.NetHack(archivefile=None) # noqa: F841
self.assertEqual(_run_nethack_in_subprocesses(num_procs), [0] * num_procs)
def test_forking_with_nethack_in_parent_new_context(
self, num_procs=NUM_SUBPROCESSES
):
env = nethack.NetHack(archivefile=None, context=zmq.Context()) # noqa: F841
self.assertEqual(_run_nethack_in_subprocesses(num_procs), [0] * num_procs)
def _run_nethack():
env = nethack.NetHack(archivefile=None)
env.reset()
env.step(0)
env.step(0)
env.step(0)
def play(env_name, play_mode, ngames, max_steps, seeds, no_clear, no_render):
is_raw_env = env_name == "nethack"
if is_raw_env:
# TODO save data somewhere reasonable
env = nethack.NetHack(archivefile="./nle_data/play_data.zip")
else:
env = gym.make(env_name)
if seeds is not None:
env.seed(seeds)
obs = env.reset()
last_obs = None # needed for "nethack" env
steps = 0
episodes = 0
reward = 0.0
action = None
ch = None
start_time = timeit.default_timer()
while True:
if not no_render:
if not no_clear:
os.system("cls" if os.name == "nt" else "clear")
if not is_raw_env:
print("Previous reward:", reward)
print("Available actions:", env._actions)
print("Previous action: {}{!r})".format(
"{} --".format(chr(ch)) if ch is not None else "",
env._actions[action] if action is not None else None,
))
env.render()
else:
print("Available actions:", nle.env.base.FULL_ACTIONS)
print("Previous actions:", action)
print_message.print_message(obs)
action = get_action(env, play_mode, is_raw_env)
if action is None:
break
if is_raw_env:
last_obs = obs
obs, done, info = env.step(action)
else:
obs, reward, done, info = env.step(action)
steps += 1
done = done or steps >= max_steps
if not done:
continue
if not is_raw_env:
print("Final reward:", reward)
print("End status:", info["end_status"].name)
print("Env stats:")
pprint.pprint(info["stats"])
else:
if last_obs.ProgramState().Gameover():
# Print tombstone.
if last_obs.WindowsLength() < 1:
return steps
window = last_obs.Windows(1)
if not no_render:
for i in range(window.StringsLength()):
print(window.Strings(i).decode("ascii"))
time_delta = timeit.default_timer() - start_time
print("Episode: {}. Steps: {}. SPS: {:f}".format(
episodes, steps, steps / time_delta))
start_time = timeit.default_timer()
episodes += 1
steps = 0
if episodes == ngames:
break
env.reset()