def close(self, name=None, reason=u'closed by RewarderSession.close'):
if name is None:
names = list(self.names_by_id.values())
else:
logger.info('[%s] Closing rewarder connection', name)
names = [name]
self.ids_by_name = {name: id for id, name in self.names_by_id.items()}
for name in names:
with self.lock:
id = self.ids_by_name.pop(name, None)
if id is None:
# already closed
continue
del self.names_by_id[id]
del self.reward_buffers[id]
del self.env_statuses[id]
self.errors.pop(id, None)
network = self.networks.pop(id)
network.close()
client = self.clients.pop(id, None)
if client is not None:
reactor.callFromThread(client.close, reason=reason)
def send_env_reward(self, reward, done, info, episode_id):
pyprofile.incr('agent_conn.reward', reward)
if done:
pyprofile.incr('agent_conn.done')
reactor.callFromThread(self._send_env_reward, reward, done, info,
episode_id)
def connect(self, name, address, label, password, env_id=None, seed=None, fps=60,
start_timeout=None, observer=False, skip_network_calibration=False):
if name in self.reward_buffers:
self.close(name)
network = Network()
self.names_by_id[self.i] = name
self.reward_buffers[self.i] = reward_buffer.RewardBuffer(label)
self.env_statuses[self.i] = env_status.EnvStatus(label=label, primary=False)
self.networks[self.i] = network
reactor.callFromThread(self._connect,
name=name,
address=address,
env_id=env_id,
seed=seed,
fps=fps,
i=self.i,
network=network,
env_status=self.env_statuses[self.i],
reward_buffer=self.reward_buffers[self.i],
label=label,
start_timeout=start_timeout,
password=password,
observer=observer,
skip_network_calibration=skip_network_calibration,
)
self.i += 1
return network
def listen(self, port=15900):
logger.info('Starting Rewarder on port=%s', port)
factory = websocket.WebSocketServerFactory()
factory.agent_conn = self
factory.protocol = RewarderProtocol
reactor.callFromThread(reactor.listenTCP, port, factory)
def connect(self, name, address, label, password, env_id=None, seed=None, fps=60,
start_timeout=None, observer=False, skip_network_calibration=False):
if name in self.reward_buffers:
self.close(name, reason='closing previous connection to reconnect with the same name')
network = Network()
self.names_by_id[self.i] = name
self.reward_buffers[self.i] = reward_buffer.RewardBuffer(label)
self.env_statuses[self.i] = env_status.EnvStatus(label=label, primary=False)
self.networks[self.i] = network
reactor.callFromThread(self._connect,
name=name,
address=address,
env_id=env_id,
seed=seed,
fps=fps,
i=self.i,
network=network,
env_status=self.env_statuses[self.i],
reward_buffer=self.reward_buffers[self.i],
label=label,
start_timeout=start_timeout,
password=password,
observer=observer,
skip_network_calibration=skip_network_calibration,
)
self.i += 1
return network
def close(self, name=None):
if name is None:
names = list(self.names_by_id.values())
else:
logger.info('[%s] Closing rewarder connection', name)
names = [name]
self.ids_by_name = {name: id for id, name in self.names_by_id.items()}
for name in names:
with self.lock:
id = self.ids_by_name.pop(name, None)
if id is None:
# already closed
continue
del self.names_by_id[id]
del self.reward_buffers[id]
del self.env_statuses[id]
self.errors.pop(id, None)
network = self.networks.pop(id)
network.close()
client = self.clients.pop(id, None)
if client is not None:
reactor.callFromThread(client.close)
def send_env_describe(self,
env_id,
env_state,
episode_id,
fps,
headers=None,
parent_message_id=None,
parent_context=None):
reactor.callFromThread(self._send_env_describe, env_id, env_state,
episode_id, fps, headers, parent_message_id,
parent_context)
def blockingCallFromThread(f, *a, **kw):
local_queue = queue.Queue()
def _callFromThread():
result = defer.maybeDeferred(f, *a, **kw)
result.addBoth(local_queue.put)
reactor.callFromThread(_callFromThread)
result = queue_get(local_queue)
if isinstance(result, failure.Failure):
if result.frames:
e = error.Error(str(result))
else:
e = result.value
raise e
return result
def blockingCallFromThread(f, *a, **kw):
local_queue = queue.Queue()
def _callFromThread():
result = defer.maybeDeferred(f, *a, **kw)
result.addBoth(local_queue.put)
reactor.callFromThread(_callFromThread)
result = queue_get(local_queue)
if isinstance(result, failure.Failure):
if result.frames:
e = error.Error(str(result))
else:
e = result.value
raise e
return result
def send_action(self, action_n, env_id):
reactor.callFromThread(self._send_action, env_id, action_n)
return self.pop_errors()
def reset(self, seed=None):
with self.lock:
for i, reward_buffer in self.reward_buffers.items():
reward_buffer.mask()
reactor.callFromThread(self._reset, seed=seed)
def send_reply_env_reset(self, *args, **kwargs):
reactor.callFromThread(self._send_reply_env_reset, *args, **kwargs)
def send_reply_error(self, *args, **kwargs):
reactor.callFromThread(self._send_reply_error, *args, **kwargs)
def send_env_observation(self, observation, episode_id):
reactor.callFromThread(self._send_env_observation, observation, episode_id)
def send_env_text(self, text, episode_id):
''' text channel to communicate with the agent '''
reactor.callFromThread(self._send_env_text, text, episode_id)
def step(self, action):
reactor.callFromThread(self._step, action)
return self.flip()
def reset(self, seed=None):
with self.lock:
for i, reward_buffer in self.reward_buffers.items():
reward_buffer.mask()
reactor.callFromThread(self._reset, seed=seed)
def send_action(self, action_n, env_id):
reactor.callFromThread(self._send_action, env_id, action_n)
return self.pop_errors()
