def close(self):
if self._closed:
return
docker_closer.unregister(self._docker_closer_id)
# Make sure 1. we were the onse who started it, 2. it's
# actually been started, and 3. we're meant to kill it.
if self._container_id and not self.assigner.reuse:
self._remove()
self._closed = True
@property
def client(self):
return self.assigner.client
if __name__ == '__main__':
logging.getLogger().setLevel(logging.INFO)
from universe.runtimes import registration
# docker run --name test --rm -ti -p 5900:5900 -p 15900:15900 quay.io/openai/universe.gym-core
instance = DockerManager(
runtime=registration.runtime_spec('gym-core'),
n=2,
)
instance.start()
import ipdb
ipdb.set_trace()
def configure_with_latest_docker_runtime_tag(env):
original_image = registration.runtime_spec(env.spec.tags['runtime']).image
latest_image = re.sub(r':.*', ':latest', original_image)
logger.info("Using latest image: {}".format(latest_image))
env.configure(remotes=1, docker_image=latest_image)
def _configure(
self,
remotes=None,
client_id=None,
start_timeout=None,
docker_image=None,
ignore_clock_skew=False,
disable_action_probes=False,
vnc_driver=None,
vnc_kwargs={},
rewarder_driver=None,
replace_on_crash=False,
allocate_sync=True,
observer=False,
api_key=None,
record=False,
):
"""Standard Gym hook to configure the environment.
Args:
ignore_clock_skew (bool): Assume remotes are on the same machine as us,
for the purposes of diagnostics measurement.
If true, we skip measuring the clock skew over the network,
and skip generating diagnostics which rely on it.
True when used by the rewarder to measure latency between
the VNC frame and its calculation of reward for that
frame. In this case we share a common clock with the env
generating the VNC frame, so we don't need to send/receive
probes. Clock skew is zero in this case.
False when remotes are potentially different machines
(such as an agent, or a demonstrator), and we will be
sending probe keys and measuring network ping rountrip
times to calculate clock skew.
"""
if self._started:
raise error.Error(
'{} has already been started; cannot change configuration now.'
.format(self))
universe.configure_logging()
twisty.start_once()
if self.spec is not None:
runtime = registration.runtime_spec(self.spec.tags['runtime'])
# Let the user manually set the docker_image version
if docker_image:
# TODO: don't support this option?
runtime.image = docker_image
else:
runtime = None
if remotes is None:
remotes = os.environ.get('GYM_VNC_REMOTES', '1')
if client_id is None:
client_id = default_client_id()
self.remote_manager, self.n = remotes_module.build(
client_id=client_id,
remotes=remotes,
runtime=runtime,
start_timeout=start_timeout,
api_key=api_key,
use_recorder_ports=record,
)
self.connection_names = [None] * self.n
self.connection_labels = [None] * self.n
self.crashed = {}
self.allow_reconnect = replace_on_crash and self.remote_manager.supports_reconnect
if self.remote_manager.connect_vnc:
cls = vnc_session(vnc_driver)
vnc_kwargs.setdefault('start_timeout',
self.remote_manager.start_timeout)
if runtime == 'gym-core':
vnc_kwargs.setdefault('encoding', 'zrle')
else:
vnc_kwargs.setdefault('encoding', 'tight')
vnc_kwargs.setdefault('fine_quality_level', 50)
vnc_kwargs.setdefault('subsample_level', 2)
# Filter out None values, since some drivers may not handle them correctly
vnc_kwargs = {k: v for k, v in vnc_kwargs.items() if v is not None}
logger.info(
'Using VNCSession arguments: %s. (Customize by running "env.configure(vnc_kwargs={...})"',
vnc_kwargs)
self.vnc_kwargs = vnc_kwargs
self.vnc_session = cls()
else:
self.vnc_session = None
self._observer = observer
if self.remote_manager.connect_rewarder:
cls = rewarder_session(rewarder_driver)
self.rewarder_session = cls()
else:
self.rewarder_session = None
if ignore_clock_skew:
logger.info(
'Printed stats will ignore clock skew. (This usually makes sense only when the environment and agent are on the same machine.)'
)
if self.rewarder_session or ignore_clock_skew:
# Don't need rewarder session if we're ignoring clock skew
if self.spec is not None:
metadata_encoding = self.spec.tags.get('metadata_encoding')
else:
metadata_encoding = None
self.diagnostics = diagnostics.Diagnostics(
self.n,
self._probe_key,
ignore_clock_skew,
metadata_encoding=metadata_encoding,
disable_action_probes=disable_action_probes)
else:
self.diagnostics = None
self._reset_mask()
self._started = True
self.remote_manager.allocate([str(i) for i in range(self.n)],
initial=True)
if allocate_sync:
# Block until we've fulfilled n environments
self._handle_connect(n=self.n)
else:
# Handle any backends which synchronously fufill their
# allocation.
self._handle_connect()
def _configure(self, remotes=None,
client_id=None,
start_timeout=None, docker_image=None,
ignore_clock_skew=False, disable_action_probes=False,
vnc_driver=None, vnc_kwargs=None,
rewarder_driver=None,
replace_on_crash=False, allocate_sync=True,
observer=False, api_key=None,
record=False,
):
"""Standard Gym hook to configure the environment.
Args:
ignore_clock_skew (bool): Assume remotes are on the same machine as us,
for the purposes of diagnostics measurement.
If true, we skip measuring the clock skew over the network,
and skip generating diagnostics which rely on it.
True when used by the rewarder to measure latency between
the VNC frame and its calculation of reward for that
frame. In this case we share a common clock with the env
generating the VNC frame, so we don't need to send/receive
probes. Clock skew is zero in this case.
False when remotes are potentially different machines
(such as an agent, or a demonstrator), and we will be
sending probe keys and measuring network ping rountrip
times to calculate clock skew.
"""
if self._started:
raise error.Error('{} has already been started; cannot change configuration now.'.format(self))
universe.configure_logging()
twisty.start_once()
if self.spec is not None:
runtime = registration.runtime_spec(self.spec.tags['runtime'])
# Let the user manually set the docker_image version
if docker_image:
# TODO: don't support this option?
runtime.image = docker_image
else:
runtime = None
if remotes is None:
remotes = os.environ.get('GYM_VNC_REMOTES', '1')
if client_id is None:
client_id = default_client_id()
if vnc_kwargs is None:
vnc_kwargs = {}
self.remote_manager, self.n = remotes_module.build(
client_id=client_id,
remotes=remotes, runtime=runtime, start_timeout=start_timeout,
api_key=api_key,
use_recorder_ports=record,
)
self.connection_names = [None] * self.n
self.connection_labels = [None] * self.n
self.crashed = {}
self.allow_reconnect = replace_on_crash and self.remote_manager.supports_reconnect
if self.remote_manager.connect_vnc:
cls = vnc_session(vnc_driver)
vnc_kwargs.setdefault('start_timeout', self.remote_manager.start_timeout)
if runtime == 'gym-core':
vnc_kwargs.setdefault('encoding', 'zrle')
else:
vnc_kwargs.setdefault('encoding', 'tight')
vnc_kwargs.setdefault('fine_quality_level', 50)
vnc_kwargs.setdefault('subsample_level', 2)
# Filter out None values, since some drivers may not handle them correctly
vnc_kwargs = {k: v for k, v in vnc_kwargs.items() if v is not None}
logger.info('Using VNCSession arguments: %s. (Customize by running "env.configure(vnc_kwargs={...})"', vnc_kwargs)
self.vnc_kwargs = vnc_kwargs
self.vnc_session = cls()
else:
self.vnc_session = None
self._observer = observer
if self.remote_manager.connect_rewarder:
cls = rewarder_session(rewarder_driver)
self.rewarder_session = cls()
else:
self.rewarder_session = None
if ignore_clock_skew:
logger.info('Printed stats will ignore clock skew. (This usually makes sense only when the environment and agent are on the same machine.)')
if self.rewarder_session or ignore_clock_skew:
# Don't need rewarder session if we're ignoring clock skew
if self.spec is not None:
metadata_encoding = self.spec.tags.get('metadata_encoding')
else:
metadata_encoding = None
self.diagnostics = diagnostics.Diagnostics(self.n, self._probe_key, ignore_clock_skew, metadata_encoding=metadata_encoding, disable_action_probes=disable_action_probes)
else:
self.diagnostics = None
self._reset_mask()
self._started = True
self.remote_manager.allocate([str(i) for i in range(self.n)], initial=True)
if allocate_sync:
# Block until we've fulfilled n environments
self._handle_connect(n=self.n)
else:
# Handle any backends which synchronously fufill their
# allocation.
self._handle_connect()
def configure_with_latest_docker_runtime_tag(env):
original_image = registration.runtime_spec(env.spec.tags['runtime']).image
latest_image = re.sub(r':.*', ':latest', original_image)
logger.info("Using latest image: {}".format(latest_image))
env.configure(remotes=1, docker_image=latest_image)
self.close()
def close(self):
if self._closed:
return
docker_closer.unregister(self._docker_closer_id)
# Make sure 1. we were the onse who started it, 2. it's
# actually been started, and 3. we're meant to kill it.
if self._container_id and not self.assigner.reuse:
self._remove()
self._closed = True
@property
def client(self):
return self.assigner.client
if __name__ == '__main__':
logging.getLogger().setLevel(logging.INFO)
from universe.runtimes import registration
# docker run --name test --rm -ti -p 5900:5900 -p 15900:15900 quay.io/openai/universe.gym-core
instance = DockerManager(
runtime=registration.runtime_spec('gym-core'),
n=2,
)
instance.start()
import ipdb;ipdb.set_trace()