def test_vnc_monitoring():
with helpers.tempdir() as temp:
env = gym.make('gym-core.Pong-v3')
env = wrappers.GymCoreAction(env)
env = wrappers.Monitor(temp)(env)
env.configure(remotes=2)
env.reset()
for i in range(2):
env.step([0, 0])
env.close()
results = gym.monitoring.load_results(temp)
assert results['env_info']['env_id'] == 'gym-core.Pong-v3'
def test_multiprocessing_env_monitoring():
with helpers.tempdir() as temp:
env = wrappers.WrappedMultiprocessingEnv('Pong-v3')
env = wrappers.Monitor(temp)(env)
env.configure(n=2)
env.reset()
for i in range(2):
env.step([0, 0])
env.close()
manifests = glob.glob(os.path.join(temp, '*.video.*'))
assert len(manifests) == 2, 'There are {} manifests: {}'.format(len(manifests), manifests)
results = gym.monitoring.load_results(temp)
assert results['env_info']['env_id'] == 'Pong-v3'
env = wrappers.WrappedVNCEnv()
# env = wrappers.BlockingReset(env)
if not isinstance(env, wrappers.GymCoreAction):
# The GymCoreSyncEnv's try to mimic their core counterparts,
# and thus came pre-wrapped wth an action space
# translator. Everything else probably wants a SafeActionSpace
# wrapper to shield them from random-agent clicking around
# everywhere.
env = wrappers.experimental.SafeActionSpace(env)
else:
# Only gym-core are seedable
env.seed([0])
env = wrappers.Logger(env)
if args.monitor:
env = wrappers.Monitor(env, '/tmp/vnc_random_agent', force=True)
if args.actions == 'random':
action_space = env.action_space
elif args.actions == 'noop':
action_space = NoopSpace()
elif args.actions == 'forward':
action_space = ForwardSpace()
elif args.actions == 'click':
spec = universe.runtime_spec('flashgames').server_registry[args.env_id]
height = spec["height"]
width = spec["width"]
noclick_regions = [
r['coordinates'] for r in spec['regions'] if r['type'] == 'noclick'
] if spec.get('regions') else []
active_region = (CHROME_X_OFFSET, CHROME_Y_OFFSET,
def main():
# You can optionally set up the logger. Also fine to set the level
# to logging.DEBUG or logging.WARN if you want to change the
# amount of output.
logger.setLevel(logging.INFO)
universe.configure_logging()
parser = argparse.ArgumentParser(description=None)
parser.add_argument('-e', '--env_id', default='gtav.SaneDriving-v0', help='Which environment to run on.')
parser.add_argument('-m', '--monitor', action='store_false', help='Whether to activate the monitor.')
parser.add_argument('-r', '--remote', help='The number of environments to create (e.g. -r 20), or the address of pre-existing VNC servers and rewarders to use (e.g. -r vnc://localhost:5900+15900,localhost:5901+15901), or a query to the allocator (e.g. -r http://allocator.sci.openai-tech.com?n=2)')
parser.add_argument('-v', '--verbose', action='count', dest='verbosity', default=0, help='Set verbosity.')
parser.add_argument('-R', '--no-render', action='store_true', help='Do not render the environment locally.')
parser.add_argument('-f', '--fps', default=8., type=float, help='Desired frames per second')
parser.add_argument('-N', '--max-steps', type=int, default=10**7, help='Maximum number of steps to take')
parser.add_argument('-d', '--driver', default='DeepDriver', help='Choose your driver')
parser.add_argument('-c', '--custom_camera', action='store_true', help='Customize the GTA camera')
args = parser.parse_args()
logging.getLogger('gym').setLevel(logging.NOTSET)
logging.getLogger('universe').setLevel(logging.NOTSET)
if args.verbosity == 0:
logger.setLevel(logging.INFO)
elif args.verbosity >= 1:
logger.setLevel(logging.DEBUG)
if args.env_id is not None:
# N.B. This does not set the actual environment mode yet, which
# is currently driven by environment itself.
env = gym.make(args.env_id)
else:
env = wrappers.WrappedVNCEnv()
if not isinstance(env, wrappers.GymCoreAction):
# The GymCoreSyncEnv's try to mimic their core counterparts,
# and thus came pre-wrapped wth an action space
# translator. Everything else probably wants a SafeActionSpace
# wrapper to shield them from random-agent clicking around
# everywhere.
env = wrappers.SafeActionSpace(env)
else:
# Only gym-core are seedable
env.seed([0])
env = wrappers.Logger(env)
env.configure(
fps=args.fps,
# print_frequency=None,
# ignore_clock_skew=True,
remotes=args.remote,
vnc_driver='go', vnc_kwargs={
'encoding': 'tight', 'compress_level': 0, 'fine_quality_level': 50, 'subsample_level': 0, 'quality_level': 5,
},
)
if args.driver == 'DeepDriver':
driver = DeepDriver()
elif args.driver == 'DeepDriverTF':
driver = DeepDriverTF()
else:
raise Exception('That driver is not available')
driver.setup()
if args.monitor:
# env.monitor.start('/tmp/vnc_random_agent', force=True, video_callable=lambda i: True)
wrappers.Monitor(env, '/tmp/vnc_random_agent', video_callable=False, force=True)
render = not args.no_render
observation_n = env.reset()
reward_n = [0] * env.n
done_n = [False] * env.n
info = None
for i in range(args.max_steps):
if render:
# Note the first time you call render, it'll be relatively
# slow and you'll have some aggregated rewards. We could
# open the render() window before `reset()`, but that's
# confusing since it pops up a black window for the
# duration of the reset.
env.render()
action_n = driver.step(observation_n, reward_n, done_n, info)
try:
if info is not None:
distance = info['n'][0]['distance_from_destination']
logger.info('distance %s', distance)
except KeyError as e:
logger.debug('distance not available %s', str(e))
if args.custom_camera:
# Sending this every step is probably overkill
for action in action_n:
action.append(GTASetting('use_custom_camera', True))
# Take an action
with pyprofile.push('env.step'):
_step = env.step(action_n)
observation_n, reward_n, done_n, info = _step
if any(done_n) and info and not any(info_n.get('env_status.artificial.done', False) for info_n in info['n']):
print('done_n', done_n, 'i', i)
logger.info('end of episode')
env.reset()
# We're done! clean up
env.close()