zjowowen's picture
init space
079c32c
raw
history blame
24.1 kB
from collections import defaultdict
import math
import queue
from time import sleep, time
import gym
from ding.framework import Supervisor
from typing import TYPE_CHECKING, Any, List, Union, Dict, Optional, Callable
from ding.framework.supervisor import ChildType, RecvPayload, SendPayload
from ding.utils import make_key_as_identifier
from ditk import logging
from ding.data import ShmBufferContainer
import enum
import treetensor.numpy as tnp
import numbers
if TYPE_CHECKING:
from gym.spaces import Space
class EnvState(enum.IntEnum):
"""
VOID -> RUN -> DONE
"""
VOID = 0
INIT = 1
RUN = 2
RESET = 3
DONE = 4
ERROR = 5
NEED_RESET = 6
class EnvRetryType(str, enum.Enum):
RESET = "reset"
RENEW = "renew"
class EnvSupervisor(Supervisor):
"""
Manage multiple envs with supervisor.
New features (compared to env manager):
- Consistent interface in multi-process and multi-threaded mode.
- Add asynchronous features and recommend using asynchronous methods.
- Reset is performed after an error is encountered in the step method.
Breaking changes (compared to env manager):
- Without some states.
"""
def __init__(
self,
type_: ChildType = ChildType.PROCESS,
env_fn: List[Callable] = None,
retry_type: EnvRetryType = EnvRetryType.RESET,
max_try: Optional[int] = None,
max_retry: Optional[int] = None,
auto_reset: bool = True,
reset_timeout: Optional[int] = None,
step_timeout: Optional[int] = None,
retry_waiting_time: Optional[int] = None,
episode_num: int = float("inf"),
shared_memory: bool = True,
copy_on_get: bool = True,
**kwargs
) -> None:
"""
Overview:
Supervisor that manage a group of envs.
Arguments:
- type_ (:obj:`ChildType`): Type of child process.
- env_fn (:obj:`List[Callable]`): The function to create environment
- retry_type (:obj:`EnvRetryType`): Retry reset or renew env.
- max_try (:obj:`EasyDict`): Max try times for reset or step action.
- max_retry (:obj:`Optional[int]`): Alias of max_try.
- auto_reset (:obj:`bool`): Auto reset env if reach done.
- reset_timeout (:obj:`Optional[int]`): Timeout in seconds for reset.
- step_timeout (:obj:`Optional[int]`): Timeout in seconds for step.
- retry_waiting_time (:obj:`Optional[float]`): Wait time on each retry.
- shared_memory (:obj:`bool`): Use shared memory in multiprocessing.
- copy_on_get (:obj:`bool`): Use copy on get in multiprocessing.
"""
if kwargs:
logging.warning("Unknown parameters on env supervisor: {}".format(kwargs))
super().__init__(type_=type_)
if type_ is not ChildType.PROCESS and (shared_memory or copy_on_get):
logging.warning("shared_memory and copy_on_get only works in process mode.")
self._shared_memory = type_ is ChildType.PROCESS and shared_memory
self._copy_on_get = type_ is ChildType.PROCESS and copy_on_get
self._env_fn = env_fn
self._create_env_ref()
self._obs_buffers = None
if env_fn:
if self._shared_memory:
obs_space = self._observation_space
if isinstance(obs_space, gym.spaces.Dict):
# For multi_agent case, such as multiagent_mujoco and petting_zoo mpe.
# Now only for the case that each agent in the team have the same obs structure
# and corresponding shape.
shape = {k: v.shape for k, v in obs_space.spaces.items()}
dtype = {k: v.dtype for k, v in obs_space.spaces.items()}
else:
shape = obs_space.shape
dtype = obs_space.dtype
self._obs_buffers = {
env_id: ShmBufferContainer(dtype, shape, copy_on_get=self._copy_on_get)
for env_id in range(len(self._env_fn))
}
for env_init in env_fn:
self.register(env_init, shm_buffer=self._obs_buffers, shm_callback=self._shm_callback)
else:
for env_init in env_fn:
self.register(env_init)
self._retry_type = retry_type
self._auto_reset = auto_reset
if max_retry:
logging.warning("The `max_retry` is going to be deprecated, use `max_try` instead!")
self._max_try = max_try or max_retry or 1
self._reset_timeout = reset_timeout
self._step_timeout = step_timeout
self._retry_waiting_time = retry_waiting_time
self._env_replay_path = None
self._episode_num = episode_num
self._init_states()
def _init_states(self):
self._env_seed = {}
self._env_dynamic_seed = None
self._env_replay_path = None
self._env_states = {}
self._reset_param = {}
self._ready_obs = {}
self._env_episode_count = {i: 0 for i in range(self.env_num)}
self._retry_times = defaultdict(lambda: 0)
self._last_called = defaultdict(lambda: {"step": math.inf, "reset": math.inf})
def _shm_callback(self, payload: RecvPayload, obs_buffers: Any):
"""
Overview:
This method will be called in child worker, so we can put large data into shared memory
and replace the original payload data to none, then reduce the serialization/deserialization cost.
"""
if payload.method == "reset" and payload.data is not None:
obs_buffers[payload.proc_id].fill(payload.data)
payload.data = None
elif payload.method == "step" and payload.data is not None:
obs_buffers[payload.proc_id].fill(payload.data.obs)
payload.data._replace(obs=None)
def _create_env_ref(self):
# env_ref is used to acquire some common attributes of env, like obs_shape and act_shape
self._env_ref = self._env_fn[0]()
self._env_ref.reset()
self._observation_space = self._env_ref.observation_space
self._action_space = self._env_ref.action_space
self._reward_space = self._env_ref.reward_space
self._env_ref.close()
def step(self, actions: Union[Dict[int, List[Any]], List[Any]], block: bool = True) -> Optional[List[tnp.ndarray]]:
"""
Overview:
Execute env step according to input actions. And reset an env if done.
Arguments:
- actions (:obj:`List[tnp.ndarray]`): Actions came from outer caller like policy, \
in structure of {env_id: actions}.
- block (:obj:`bool`): If block, return timesteps, else return none.
Returns:
- timesteps (:obj:`List[tnp.ndarray]`): Each timestep is a tnp.array with observation, reward, done, \
info, env_id.
"""
assert not self.closed, "Env supervisor has closed."
if isinstance(actions, List):
actions = {i: p for i, p in enumerate(actions)}
assert actions, "Action is empty!"
send_payloads = []
for env_id, act in actions.items():
payload = SendPayload(proc_id=env_id, method="step", args=[act])
send_payloads.append(payload)
self.send(payload)
if not block:
# Retrieve the data for these steps from the recv method
return
# Wait for all steps returns
recv_payloads = self.recv_all(
send_payloads, ignore_err=True, callback=self._recv_callback, timeout=self._step_timeout
)
return [payload.data for payload in recv_payloads]
def recv(self, ignore_err: bool = False) -> RecvPayload:
"""
Overview:
Wait for recv payload, this function will block the thread.
Arguments:
- ignore_err (:obj:`bool`): If ignore_err is true, payload with error object will be discarded.\
This option will not catch the exception.
Returns:
- recv_payload (:obj:`RecvPayload`): Recv payload.
"""
self._detect_timeout()
try:
payload = super().recv(ignore_err=True, timeout=0.1)
payload = self._recv_callback(payload=payload)
if payload.err:
return self.recv(ignore_err=ignore_err)
else:
return payload
except queue.Empty:
return self.recv(ignore_err=ignore_err)
def _detect_timeout(self):
"""
Overview:
Try to restart all timeout environments if detected timeout.
"""
for env_id in self._last_called:
if self._step_timeout and time() - self._last_called[env_id]["step"] > self._step_timeout:
payload = RecvPayload(
proc_id=env_id, method="step", err=TimeoutError("Step timeout on env {}".format(env_id))
)
self._recv_queue.put(payload)
continue
if self._reset_timeout and time() - self._last_called[env_id]["reset"] > self._reset_timeout:
payload = RecvPayload(
proc_id=env_id, method="reset", err=TimeoutError("Step timeout on env {}".format(env_id))
)
self._recv_queue.put(payload)
continue
@property
def env_num(self) -> int:
return len(self._children)
@property
def observation_space(self) -> 'Space':
return self._observation_space
@property
def action_space(self) -> 'Space':
return self._action_space
@property
def reward_space(self) -> 'Space':
return self._reward_space
@property
def ready_obs(self) -> tnp.array:
"""
Overview:
Get the ready (next) observation in ``tnp.array`` type, which is uniform for both async/sync scenarios.
Return:
- ready_obs (:obj:`tnp.array`): A stacked treenumpy-type observation data.
Example:
>>> obs = env_manager.ready_obs
>>> action = model(obs) # model input np obs and output np action
>>> timesteps = env_manager.step(action)
"""
active_env = [i for i, s in self._env_states.items() if s == EnvState.RUN]
active_env.sort()
obs = [self._ready_obs.get(i) for i in active_env]
if len(obs) == 0:
return tnp.array([])
return tnp.stack(obs)
@property
def ready_obs_id(self) -> List[int]:
return [i for i, s in self.env_states.items() if s == EnvState.RUN]
@property
def done(self) -> bool:
return all([s == EnvState.DONE for s in self.env_states.values()])
@property
def method_name_list(self) -> List[str]:
return ['reset', 'step', 'seed', 'close', 'enable_save_replay']
@property
def env_states(self) -> Dict[int, EnvState]:
return {env_id: self._env_states.get(env_id) or EnvState.VOID for env_id in range(self.env_num)}
def env_state_done(self, env_id: int) -> bool:
return self.env_states[env_id] == EnvState.DONE
def launch(self, reset_param: Optional[Dict] = None, block: bool = True) -> None:
"""
Overview:
Set up the environments and their parameters.
Arguments:
- reset_param (:obj:`Optional[Dict]`): Dict of reset parameters for each environment, key is the env_id, \
value is the cooresponding reset parameters.
- block (:obj:`block`): Whether will block the process and wait for reset states.
"""
assert self.closed, "Please first close the env supervisor before launch it"
if reset_param is not None:
assert len(reset_param) == self.env_num
self.start_link()
self._send_seed(self._env_seed, self._env_dynamic_seed, block=block)
self.reset(reset_param, block=block)
self._enable_env_replay()
def reset(self, reset_param: Optional[Dict[int, List[Any]]] = None, block: bool = True) -> None:
"""
Overview:
Reset an environment.
Arguments:
- reset_param (:obj:`Optional[Dict[int, List[Any]]]`): Dict of reset parameters for each environment, \
key is the env_id, value is the cooresponding reset parameters.
- block (:obj:`block`): Whether will block the process and wait for reset states.
"""
if not reset_param:
reset_param = {i: {} for i in range(self.env_num)}
elif isinstance(reset_param, List):
reset_param = {i: p for i, p in enumerate(reset_param)}
send_payloads = []
for env_id, kw_param in reset_param.items():
self._reset_param[env_id] = kw_param # For auto reset
send_payloads += self._reset(env_id, kw_param=kw_param)
if not block:
return
self.recv_all(send_payloads, ignore_err=True, callback=self._recv_callback, timeout=self._reset_timeout)
def _recv_callback(
self, payload: RecvPayload, remain_payloads: Optional[Dict[str, SendPayload]] = None
) -> RecvPayload:
"""
Overview:
The callback function for each received payload, within this method will modify the state of \
each environment, replace objects in shared memory, and determine if a retry is needed due to an error.
Arguments:
- payload (:obj:`RecvPayload`): The received payload.
- remain_payloads (:obj:`Optional[Dict[str, SendPayload]]`): The callback may be called many times \
until remain_payloads be cleared, you can append new payload into remain_payloads to call this \
callback recursively.
"""
self._set_shared_obs(payload=payload)
self.change_state(payload=payload)
if payload.method == "reset":
return self._recv_reset_callback(payload=payload, remain_payloads=remain_payloads)
elif payload.method == "step":
return self._recv_step_callback(payload=payload, remain_payloads=remain_payloads)
return payload
def _set_shared_obs(self, payload: RecvPayload):
if self._obs_buffers is None:
return
if payload.method == "reset" and payload.err is None:
payload.data = self._obs_buffers[payload.proc_id].get()
elif payload.method == "step" and payload.err is None:
payload.data._replace(obs=self._obs_buffers[payload.proc_id].get())
def _recv_reset_callback(
self, payload: RecvPayload, remain_payloads: Optional[Dict[str, SendPayload]] = None
) -> RecvPayload:
assert payload.method == "reset", "Recv error callback({}) in reset callback!".format(payload.method)
if remain_payloads is None:
remain_payloads = {}
env_id = payload.proc_id
if payload.err:
self._retry_times[env_id] += 1
if self._retry_times[env_id] > self._max_try - 1:
self.shutdown(5)
raise RuntimeError(
"Env {} reset has exceeded max_try({}), and the latest exception is: {}".format(
env_id, self._max_try, payload.err
)
)
if self._retry_waiting_time:
sleep(self._retry_waiting_time)
if self._retry_type == EnvRetryType.RENEW:
self._children[env_id].restart()
send_payloads = self._reset(env_id)
for p in send_payloads:
remain_payloads[p.req_id] = p
else:
self._retry_times[env_id] = 0
self._ready_obs[env_id] = payload.data
return payload
def _recv_step_callback(
self, payload: RecvPayload, remain_payloads: Optional[Dict[str, SendPayload]] = None
) -> RecvPayload:
assert payload.method == "step", "Recv error callback({}) in step callback!".format(payload.method)
if remain_payloads is None:
remain_payloads = {}
if payload.err:
send_payloads = self._reset(payload.proc_id)
for p in send_payloads:
remain_payloads[p.req_id] = p
info = {"abnormal": True, "err": payload.err}
payload.data = tnp.array(
{
'obs': None,
'reward': None,
'done': None,
'info': info,
'env_id': payload.proc_id
}
)
else:
obs, reward, done, info, *_ = payload.data
if done:
self._env_episode_count[payload.proc_id] += 1
if self._env_episode_count[payload.proc_id] < self._episode_num and self._auto_reset:
send_payloads = self._reset(payload.proc_id)
for p in send_payloads:
remain_payloads[p.req_id] = p
# make the type and content of key as similar as identifier,
# in order to call them as attribute (e.g. timestep.xxx), such as ``TimeLimit.truncated`` in cartpole info
info = make_key_as_identifier(info)
payload.data = tnp.array(
{
'obs': obs,
'reward': reward,
'done': done,
'info': info,
'env_id': payload.proc_id
}
)
self._ready_obs[payload.proc_id] = obs
return payload
def _reset(self, env_id: int, kw_param: Optional[Dict[str, Any]] = None) -> List[SendPayload]:
"""
Overview:
Reset an environment. This method does not wait for the result to be returned.
Arguments:
- env_id (:obj:`int`): Environment id.
- kw_param (:obj:`Optional[Dict[str, Any]]`): Reset parameters for the environment.
Returns:
- send_payloads (:obj:`List[SendPayload]`): The request payloads for seed and reset actions.
"""
assert not self.closed, "Env supervisor has closed."
send_payloads = []
kw_param = kw_param or self._reset_param[env_id]
if self._env_replay_path is not None and self.env_states[env_id] == EnvState.RUN:
logging.warning("Please don't reset an unfinished env when you enable save replay, we just skip it")
return send_payloads
# Reset env
payload = SendPayload(proc_id=env_id, method="reset", kwargs=kw_param)
send_payloads.append(payload)
self.send(payload)
return send_payloads
def _send_seed(self, env_seed: Dict[int, int], env_dynamic_seed: Optional[bool] = None, block: bool = True) -> None:
send_payloads = []
for env_id, seed in env_seed.items():
if seed is None:
continue
args = [seed]
if env_dynamic_seed is not None:
args.append(env_dynamic_seed)
payload = SendPayload(proc_id=env_id, method="seed", args=args)
send_payloads.append(payload)
self.send(payload)
if not block or not send_payloads:
return
self.recv_all(send_payloads, ignore_err=True, callback=self._recv_callback, timeout=self._reset_timeout)
def change_state(self, payload: RecvPayload):
self._last_called[payload.proc_id][payload.method] = math.inf # Have recevied
if payload.err:
self._env_states[payload.proc_id] = EnvState.ERROR
elif payload.method == "reset":
self._env_states[payload.proc_id] = EnvState.RUN
elif payload.method == "step":
if payload.data[2]:
self._env_states[payload.proc_id] = EnvState.DONE
def send(self, payload: SendPayload) -> None:
self._last_called[payload.proc_id][payload.method] = time()
return super().send(payload)
def seed(self, seed: Union[Dict[int, int], List[int], int], dynamic_seed: Optional[bool] = None) -> None:
"""
Overview:
Set the seed for each environment. The seed function will not be called until supervisor.launch \
was called.
Arguments:
- seed (:obj:`Union[Dict[int, int], List[int], int]`): List of seeds for each environment; \
Or one seed for the first environment and other seeds are generated automatically. \
Note that in threading mode, no matter how many seeds are given, only the last one will take effect. \
Because the execution in the thread is asynchronous, the results of each experiment \
are different even if a fixed seed is used.
- dynamic_seed (:obj:`Optional[bool]`): Dynamic seed is used in the training environment, \
trying to make the random seed of each episode different, they are all generated in the reset \
method by a random generator 100 * np.random.randint(1 , 1000) (but the seed of this random \
number generator is fixed by the environmental seed method, guranteeing the reproducibility \
of the experiment). You need not pass the dynamic_seed parameter in the seed method, or pass \
the parameter as True.
"""
self._env_seed = {}
if isinstance(seed, numbers.Integral):
self._env_seed = {i: seed + i for i in range(self.env_num)}
elif isinstance(seed, list):
assert len(seed) == self.env_num, "len(seed) {:d} != env_num {:d}".format(len(seed), self.env_num)
self._env_seed = {i: _seed for i, _seed in enumerate(seed)}
elif isinstance(seed, dict):
self._env_seed = {env_id: s for env_id, s in seed.items()}
else:
raise TypeError("Invalid seed arguments type: {}".format(type(seed)))
self._env_dynamic_seed = dynamic_seed
def enable_save_replay(self, replay_path: Union[List[str], str]) -> None:
"""
Overview:
Set each env's replay save path.
Arguments:
- replay_path (:obj:`Union[List[str], str]`): List of paths for each environment; \
Or one path for all environments.
"""
if isinstance(replay_path, str):
replay_path = [replay_path] * self.env_num
self._env_replay_path = replay_path
def _enable_env_replay(self):
if self._env_replay_path is None:
return
send_payloads = []
for env_id, s in enumerate(self._env_replay_path):
payload = SendPayload(proc_id=env_id, method="enable_save_replay", args=[s])
send_payloads.append(payload)
self.send(payload)
self.recv_all(send_payloads=send_payloads)
def __getattr__(self, key: str) -> List[Any]:
if not hasattr(self._env_ref, key):
raise AttributeError("env `{}` doesn't have the attribute `{}`".format(type(self._env_ref), key))
return super().__getattr__(key)
def close(self, timeout: Optional[float] = None) -> None:
"""
In order to be compatible with BaseEnvManager, the new version can use `shutdown` directly.
"""
self.shutdown(timeout=timeout)
def shutdown(self, timeout: Optional[float] = None) -> None:
if self._running:
send_payloads = []
for env_id in range(self.env_num):
payload = SendPayload(proc_id=env_id, method="close")
send_payloads.append(payload)
self.send(payload)
self.recv_all(send_payloads=send_payloads, ignore_err=True, timeout=timeout)
super().shutdown(timeout=timeout)
self._init_states()
@property
def closed(self) -> bool:
return not self._running