DI-engine/ding/envs/env_manager/subprocess_env_manager.py

from typing import Any, Union, List, Tuple, Dict, Callable, Optional
from multiprocessing import connection, get_context
from collections import namedtuple
from ditk import logging
import platform
import time
import copy
import gymnasium
import gym
import traceback
import torch
import pickle
import numpy as np
import treetensor.numpy as tnp
from easydict import EasyDict
from types import MethodType
from ding.data import ShmBufferContainer, ShmBuffer

from ding.envs.env import BaseEnvTimestep
from ding.utils import PropagatingThread, LockContextType, LockContext, ENV_MANAGER_REGISTRY, make_key_as_identifier, \
    remove_illegal_item, CloudPickleWrapper
from .base_env_manager import BaseEnvManager, EnvState, timeout_wrapper


def is_abnormal_timestep(timestep: namedtuple) -> bool:
    if isinstance(timestep.info, dict):
        return timestep.info.get('abnormal', False)
    elif isinstance(timestep.info, list) or isinstance(timestep.info, tuple):
        return timestep.info[0].get('abnormal', False) or timestep.info[1].get('abnormal', False)
    else:
        raise TypeError("invalid env timestep type: {}".format(type(timestep.info)))


@ENV_MANAGER_REGISTRY.register('async_subprocess')
class AsyncSubprocessEnvManager(BaseEnvManager):
    """
    Overview:
        Create an AsyncSubprocessEnvManager to manage multiple environments.
        Each Environment is run by a respective subprocess.
    Interfaces:
        seed, launch, ready_obs, step, reset, active_env
    """

    config = dict(
        episode_num=float("inf"),
        max_retry=1,
        step_timeout=None,
        auto_reset=True,
        retry_type='reset',
        reset_timeout=None,
        retry_waiting_time=0.1,
        # subprocess specified args
        shared_memory=True,
        copy_on_get=True,
        context='spawn' if platform.system().lower() == 'windows' else 'fork',
        wait_num=2,
        step_wait_timeout=0.01,
        connect_timeout=60,
        reset_inplace=False,
    )

    def __init__(
            self,
            env_fn: List[Callable],
            cfg: EasyDict = EasyDict({}),
    ) -> None:
        """
        Overview:
            Initialize the AsyncSubprocessEnvManager.
        Arguments:
            - env_fn (:obj:`List[Callable]`): The function to create environment
            - cfg (:obj:`EasyDict`): Config

        .. note::

            - wait_num: for each time the minimum number of env return to gather
            - step_wait_timeout: for each time the minimum number of env return to gather
        """
        super().__init__(env_fn, cfg)
        self._shared_memory = self._cfg.shared_memory
        self._copy_on_get = self._cfg.copy_on_get
        self._context = self._cfg.context
        self._wait_num = self._cfg.wait_num
        self._step_wait_timeout = self._cfg.step_wait_timeout

        self._lock = LockContext(LockContextType.THREAD_LOCK)
        self._connect_timeout = self._cfg.connect_timeout
        self._async_args = {
            'step': {
                'wait_num': min(self._wait_num, self._env_num),
                'timeout': self._step_wait_timeout
            }
        }
        self._reset_inplace = self._cfg.reset_inplace
        if not self._auto_reset:
            assert not self._reset_inplace, "reset_inplace is unavailable when auto_reset=False."

    def _create_state(self) -> None:
        r"""
        Overview:
            Fork/spawn sub-processes(Call ``_create_env_subprocess``) and create pipes to transfer the data.
        """
        self._env_episode_count = {env_id: 0 for env_id in range(self.env_num)}
        self._ready_obs = {env_id: None for env_id in range(self.env_num)}
        self._reset_param = {i: {} for i in range(self.env_num)}
        if self._shared_memory:
            obs_space = self._observation_space
            if isinstance(obs_space, (gym.spaces.Dict, gymnasium.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(self.env_num)
            }
        else:
            self._obs_buffers = {env_id: None for env_id in range(self.env_num)}
        self._pipe_parents, self._pipe_children = {}, {}
        self._subprocesses = {}
        for env_id in range(self.env_num):
            self._create_env_subprocess(env_id)
        self._waiting_env = {'step': set()}
        self._closed = False

    def _create_env_subprocess(self, env_id):
        # start a new one
        ctx = get_context(self._context)
        self._pipe_parents[env_id], self._pipe_children[env_id] = ctx.Pipe()
        self._subprocesses[env_id] = ctx.Process(
            # target=self.worker_fn,
            target=self.worker_fn_robust,
            args=(
                self._pipe_parents[env_id],
                self._pipe_children[env_id],
                CloudPickleWrapper(self._env_fn[env_id]),
                self._obs_buffers[env_id],
                self.method_name_list,
                self._reset_timeout,
                self._step_timeout,
                self._reset_inplace,
            ),
            daemon=True,
            name='subprocess_env_manager{}_{}'.format(env_id, time.time())
        )
        self._subprocesses[env_id].start()
        self._pipe_children[env_id].close()
        self._env_states[env_id] = EnvState.INIT

        if self._env_replay_path is not None:
            self._pipe_parents[env_id].send(['enable_save_replay', [self._env_replay_path[env_id]], {}])
            self._pipe_parents[env_id].recv()

    @property
    def ready_env(self) -> List[int]:
        active_env = [i for i, s in self._env_states.items() if s == EnvState.RUN]
        return [i for i in active_env if i not in self._waiting_env['step']]

    @property
    def ready_obs(self) -> Dict[int, Any]:
        """
        Overview:
            Get the next observations.
        Return:
            A dictionary with observations and their environment IDs.
        Note:
            The observations are returned in np.ndarray.
        Example:
            >>>     obs_dict = env_manager.ready_obs
            >>>     actions_dict = {env_id: model.forward(obs) for env_id, obs in obs_dict.items())}
        """
        no_done_env_idx = [i for i, s in self._env_states.items() if s != EnvState.DONE]
        sleep_count = 0
        while not any([self._env_states[i] == EnvState.RUN for i in no_done_env_idx]):
            if sleep_count != 0 and sleep_count % 10000 == 0:
                logging.warning(
                    'VEC_ENV_MANAGER: all the not done envs are resetting, sleep {} times'.format(sleep_count)
                )
            time.sleep(0.001)
            sleep_count += 1
        return {i: self._ready_obs[i] for i in self.ready_env}

    @property
    def ready_imgs(self, render_mode: Optional[str] = 'rgb_array') -> Dict[int, Any]:
        """
        Overview:
            Get the next renderd frames.
        Return:
            A dictionary with rendered frames and their environment IDs.
        Note:
            The rendered frames are returned in np.ndarray.
        """
        for i in self.ready_env:
            self._pipe_parents[i].send(['render', None, {'render_mode': render_mode}])
        data = {i: self._pipe_parents[i].recv() for i in self.ready_env}
        self._check_data(data)
        return data

    def launch(self, reset_param: Optional[Dict] = None) -> 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.
        """
        assert self._closed, "please first close the env manager"
        if reset_param is not None:
            assert len(reset_param) == len(self._env_fn)
        self._create_state()
        self.reset(reset_param)

    def reset(self, reset_param: Optional[Dict] = None) -> None:
        """
        Overview:
            Reset the environments their parameters.
        Arguments:
            - reset_param (:obj:`List`): Dict of reset parameters for each environment, key is the env_id, \
                value is the cooresponding reset parameters.
        """
        self._check_closed()

        if reset_param is None:
            reset_env_list = [env_id for env_id in range(self._env_num)]
        else:
            reset_env_list = reset_param.keys()
            for env_id in reset_param:
                self._reset_param[env_id] = reset_param[env_id]

        # clear previous info
        for env_id in reset_env_list:
            if env_id in self._waiting_env['step']:
                self._pipe_parents[env_id].recv()
                self._waiting_env['step'].remove(env_id)

        sleep_count = 0
        while any([self._env_states[i] == EnvState.RESET for i in reset_env_list]):
            if sleep_count != 0 and sleep_count % 10000 == 0:
                logging.warning(
                    'VEC_ENV_MANAGER: not all the envs finish resetting, sleep {} times'.format(sleep_count)
                )
            time.sleep(0.001)
            sleep_count += 1

        # reset env
        reset_thread_list = []
        for i, env_id in enumerate(reset_env_list):
            # set seed
            if self._env_seed[env_id] is not None:
                try:
                    if self._env_dynamic_seed is not None:
                        self._pipe_parents[env_id].send(['seed', [self._env_seed[env_id], self._env_dynamic_seed], {}])
                    else:
                        self._pipe_parents[env_id].send(['seed', [self._env_seed[env_id]], {}])
                    ret = self._pipe_parents[env_id].recv()
                    self._check_data({env_id: ret})
                    self._env_seed[env_id] = None  # seed only use once
                except BaseException as e:
                    logging.warning(
                        "subprocess reset set seed failed, ignore and continue... \n subprocess exception traceback: \n"
                        + traceback.format_exc()
                    )
            self._env_states[env_id] = EnvState.RESET
            reset_thread = PropagatingThread(target=self._reset, args=(env_id, ))
            reset_thread.daemon = True
            reset_thread_list.append(reset_thread)

        for t in reset_thread_list:
            t.start()
        for t in reset_thread_list:
            t.join()

    def _reset(self, env_id: int) -> None:

        def reset_fn():
            if self._pipe_parents[env_id].poll():
                recv_data = self._pipe_parents[env_id].recv()
                raise RuntimeError("unread data left before sending to the pipe: {}".format(repr(recv_data)))
            # if self._reset_param[env_id] is None, just reset specific env, not pass reset param
            if self._reset_param[env_id] is not None:
                assert isinstance(self._reset_param[env_id], dict), type(self._reset_param[env_id])
                self._pipe_parents[env_id].send(['reset', [], self._reset_param[env_id]])
            else:
                self._pipe_parents[env_id].send(['reset', [], None])

            if not self._pipe_parents[env_id].poll(self._connect_timeout):
                raise ConnectionError("env reset connection timeout")  # Leave it to try again

            obs = self._pipe_parents[env_id].recv()
            self._check_data({env_id: obs}, close=False)
            if self._shared_memory:
                obs = self._obs_buffers[env_id].get()
            # it is necessary to add lock for the updates of env_state
            with self._lock:
                self._env_states[env_id] = EnvState.RUN
                self._ready_obs[env_id] = obs

        exceptions = []
        for _ in range(self._max_retry):
            try:
                reset_fn()
                return
            except BaseException as e:
                logging.info("subprocess exception traceback: \n" + traceback.format_exc())
                if self._retry_type == 'renew' or isinstance(e, pickle.UnpicklingError):
                    self._pipe_parents[env_id].close()
                    if self._subprocesses[env_id].is_alive():
                        self._subprocesses[env_id].terminate()
                    self._create_env_subprocess(env_id)
                exceptions.append(e)
                time.sleep(self._retry_waiting_time)

        logging.error("Env {} reset has exceeded max retries({})".format(env_id, self._max_retry))
        runtime_error = RuntimeError(
            "Env {} reset has exceeded max retries({}), and the latest exception is: {}".format(
                env_id, self._max_retry, str(exceptions[-1])
            )
        )
        runtime_error.__traceback__ = exceptions[-1].__traceback__
        if self._closed:  # exception cased by main thread closing parent_remote
            return
        else:
            self.close()
            raise runtime_error

    def step(self, actions: Dict[int, Any]) -> Dict[int, namedtuple]:
        """
        Overview:
            Step all environments. Reset an env if done.
        Arguments:
            - actions (:obj:`Dict[int, Any]`): {env_id: action}
        Returns:
            - timesteps (:obj:`Dict[int, namedtuple]`): {env_id: timestep}. Timestep is a \
                ``BaseEnvTimestep`` tuple with observation, reward, done, env_info.
        Example:
            >>>     actions_dict = {env_id: model.forward(obs) for env_id, obs in obs_dict.items())}
            >>>     timesteps = env_manager.step(actions_dict):
            >>>     for env_id, timestep in timesteps.items():
            >>>         pass

        .. note:

            - The env_id that appears in ``actions`` will also be returned in ``timesteps``.
            - Each environment is run by a subprocess separately. Once an environment is done, it is reset immediately.
            - Async subprocess env manager use ``connection.wait`` to poll.
        """
        self._check_closed()
        env_ids = list(actions.keys())
        assert all([self._env_states[env_id] == EnvState.RUN for env_id in env_ids]
                   ), 'current env state are: {}, please check whether the requested env is in reset or done'.format(
                       {env_id: self._env_states[env_id]
                        for env_id in env_ids}
                   )

        for env_id, act in actions.items():
            self._pipe_parents[env_id].send(['step', [act], None])

        timesteps = {}
        step_args = self._async_args['step']
        wait_num, timeout = min(step_args['wait_num'], len(env_ids)), step_args['timeout']
        rest_env_ids = list(set(env_ids).union(self._waiting_env['step']))
        ready_env_ids = []
        cur_rest_env_ids = copy.deepcopy(rest_env_ids)
        while True:
            rest_conn = [self._pipe_parents[env_id] for env_id in cur_rest_env_ids]
            ready_conn, ready_ids = AsyncSubprocessEnvManager.wait(rest_conn, min(wait_num, len(rest_conn)), timeout)
            cur_ready_env_ids = [cur_rest_env_ids[env_id] for env_id in ready_ids]
            assert len(cur_ready_env_ids) == len(ready_conn)
            # timesteps.update({env_id: p.recv() for env_id, p in zip(cur_ready_env_ids, ready_conn)})
            for env_id, p in zip(cur_ready_env_ids, ready_conn):
                try:
                    timesteps.update({env_id: p.recv()})
                except pickle.UnpicklingError as e:
                    timestep = BaseEnvTimestep(None, None, None, {'abnormal': True})
                    timesteps.update({env_id: timestep})
                    self._pipe_parents[env_id].close()
                    if self._subprocesses[env_id].is_alive():
                        self._subprocesses[env_id].terminate()
                    self._create_env_subprocess(env_id)
            self._check_data(timesteps)
            ready_env_ids += cur_ready_env_ids
            cur_rest_env_ids = list(set(cur_rest_env_ids).difference(set(cur_ready_env_ids)))
            # At least one not done env timestep, or all envs' steps are finished
            if any([not t.done for t in timesteps.values()]) or len(ready_conn) == len(rest_conn):
                break
        self._waiting_env['step']: set
        for env_id in rest_env_ids:
            if env_id in ready_env_ids:
                if env_id in self._waiting_env['step']:
                    self._waiting_env['step'].remove(env_id)
            else:
                self._waiting_env['step'].add(env_id)

        if self._shared_memory:
            for i, (env_id, timestep) in enumerate(timesteps.items()):
                timesteps[env_id] = timestep._replace(obs=self._obs_buffers[env_id].get())

        for env_id, timestep in timesteps.items():
            if is_abnormal_timestep(timestep):
                self._env_states[env_id] = EnvState.ERROR
                continue
            if timestep.done:
                self._env_episode_count[env_id] += 1
                if self._env_episode_count[env_id] < self._episode_num:
                    if self._auto_reset:
                        if self._reset_inplace:  # reset in subprocess at once
                            self._env_states[env_id] = EnvState.RUN
                            self._ready_obs[env_id] = timestep.obs
                        else:
                            # in this case, ready_obs is updated in ``self._reset``
                            self._env_states[env_id] = EnvState.RESET
                            reset_thread = PropagatingThread(target=self._reset, args=(env_id, ), name='regular_reset')
                            reset_thread.daemon = True
                            reset_thread.start()
                    else:
                        # in the case that auto_reset=False, caller should call ``env_manager.reset`` manually
                        self._env_states[env_id] = EnvState.NEED_RESET
                else:
                    self._env_states[env_id] = EnvState.DONE
            else:
                self._ready_obs[env_id] = timestep.obs
        return timesteps

    # This method must be staticmethod, otherwise there will be some resource conflicts(e.g. port or file)
    # Env must be created in worker, which is a trick of avoiding env pickle errors.
    # A more robust version is used by default. But this one is also preserved.
    @staticmethod
    def worker_fn(
            p: connection.Connection,
            c: connection.Connection,
            env_fn_wrapper: 'CloudPickleWrapper',
            obs_buffer: ShmBuffer,
            method_name_list: list,
            reset_inplace: bool = False,
    ) -> None:  # noqa
        """
        Overview:
            Subprocess's target function to run.
        """
        torch.set_num_threads(1)
        env_fn = env_fn_wrapper.data
        env = env_fn()
        p.close()
        try:
            while True:
                try:
                    cmd, args, kwargs = c.recv()
                except EOFError:  # for the case when the pipe has been closed
                    c.close()
                    break
                try:
                    if cmd == 'getattr':
                        ret = getattr(env, args[0])
                    elif cmd in method_name_list:
                        if cmd == 'step':
                            timestep = env.step(*args, **kwargs)
                            if is_abnormal_timestep(timestep):
                                ret = timestep
                            else:
                                if reset_inplace and timestep.done:
                                    obs = env.reset()
                                    timestep = timestep._replace(obs=obs)
                                if obs_buffer is not None:
                                    obs_buffer.fill(timestep.obs)
                                    timestep = timestep._replace(obs=None)
                                ret = timestep
                        elif cmd == 'reset':
                            ret = env.reset(*args, **kwargs)  # obs
                            if obs_buffer is not None:
                                obs_buffer.fill(ret)
                                ret = None
                        elif args is None and kwargs is None:
                            ret = getattr(env, cmd)()
                        else:
                            ret = getattr(env, cmd)(*args, **kwargs)
                    else:
                        raise KeyError("not support env cmd: {}".format(cmd))
                    c.send(ret)
                except Exception as e:
                    # when there are some errors in env, worker_fn will send the errors to env manager
                    # directly send error to another process will lose the stack trace, so we create a new Exception
                    logging.warning("subprocess exception traceback: \n" + traceback.format_exc())
                    c.send(
                        e.__class__(
                            '\nEnv Process Exception:\n' + ''.join(traceback.format_tb(e.__traceback__)) + repr(e)
                        )
                    )
                if cmd == 'close':
                    c.close()
                    break
        except KeyboardInterrupt:
            c.close()

    @staticmethod
    def worker_fn_robust(
            parent,
            child,
            env_fn_wrapper,
            obs_buffer,
            method_name_list,
            reset_timeout=None,
            step_timeout=None,
            reset_inplace=False,
    ) -> None:
        """
        Overview:
            A more robust version of subprocess's target function to run. Used by default.
        """
        torch.set_num_threads(1)
        env_fn = env_fn_wrapper.data
        env = env_fn()
        parent.close()

        @timeout_wrapper(timeout=step_timeout)
        def step_fn(*args, **kwargs):
            timestep = env.step(*args, **kwargs)
            if is_abnormal_timestep(timestep):
                ret = timestep
            else:
                if reset_inplace and timestep.done:
                    obs = env.reset()
                    timestep = timestep._replace(obs=obs)
                if obs_buffer is not None:
                    obs_buffer.fill(timestep.obs)
                    timestep = timestep._replace(obs=None)
                ret = timestep
            return ret

        @timeout_wrapper(timeout=reset_timeout)
        def reset_fn(*args, **kwargs):
            try:
                ret = env.reset(*args, **kwargs)
                if obs_buffer is not None:
                    obs_buffer.fill(ret)
                    ret = None
                return ret
            except BaseException as e:
                logging.warning("subprocess exception traceback: \n" + traceback.format_exc())
                env.close()
                raise e

        while True:
            try:
                cmd, args, kwargs = child.recv()
            except EOFError:  # for the case when the pipe has been closed
                child.close()
                break
            try:
                if cmd == 'getattr':
                    ret = getattr(env, args[0])
                elif cmd in method_name_list:
                    if cmd == 'step':
                        ret = step_fn(*args)
                    elif cmd == 'reset':
                        if kwargs is None:
                            kwargs = {}
                        ret = reset_fn(*args, **kwargs)
                    elif cmd == 'render':
                        from ding.utils import render
                        ret = render(env, **kwargs)
                    elif args is None and kwargs is None:
                        ret = getattr(env, cmd)()
                    else:
                        ret = getattr(env, cmd)(*args, **kwargs)
                else:
                    raise KeyError("not support env cmd: {}".format(cmd))
                child.send(ret)
            except BaseException as e:
                logging.debug("Sub env '{}' error when executing {}".format(str(env), cmd))
                # when there are some errors in env, worker_fn will send the errors to env manager
                # directly send error to another process will lose the stack trace, so we create a new Exception
                logging.warning("subprocess exception traceback: \n" + traceback.format_exc())
                child.send(
                    e.__class__('\nEnv Process Exception:\n' + ''.join(traceback.format_tb(e.__traceback__)) + repr(e))
                )
            if cmd == 'close':
                child.close()
                break

    def _check_data(self, data: Dict, close: bool = True) -> None:
        exceptions = []
        for i, d in data.items():
            if isinstance(d, BaseException):
                self._env_states[i] = EnvState.ERROR
                exceptions.append(d)
        # when receiving env Exception, env manager will safely close and raise this Exception to caller
        if len(exceptions) > 0:
            if close:
                self.close()
            raise exceptions[0]

    # override
    def __getattr__(self, key: str) -> Any:
        self._check_closed()
        # we suppose that all the envs has the same attributes, if you need different envs, please
        # create different env managers.
        if not hasattr(self._env_ref, key):
            raise AttributeError("env `{}` doesn't have the attribute `{}`".format(type(self._env_ref), key))
        if isinstance(getattr(self._env_ref, key), MethodType) and key not in self.method_name_list:
            raise RuntimeError("env getattr doesn't supports method({}), please override method_name_list".format(key))
        for _, p in self._pipe_parents.items():
            p.send(['getattr', [key], {}])
        data = {i: p.recv() for i, p in self._pipe_parents.items()}
        self._check_data(data)
        ret = [data[i] for i in self._pipe_parents.keys()]
        return ret

    # override
    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

    # override
    def close(self) -> None:
        """
        Overview:
            CLose the env manager and release all related resources.
        """
        if self._closed:
            return
        self._closed = True
        for _, p in self._pipe_parents.items():
            p.send(['close', None, None])
        for env_id, p in self._pipe_parents.items():
            if not p.poll(5):
                continue
            p.recv()
        for i in range(self._env_num):
            self._env_states[i] = EnvState.VOID
        # disable process join for avoiding hang
        # for p in self._subprocesses:
        #     p.join()
        for _, p in self._subprocesses.items():
            p.terminate()
        for _, p in self._pipe_parents.items():
            p.close()

    @staticmethod
    def wait(rest_conn: list, wait_num: int, timeout: Optional[float] = None) -> Tuple[list, list]:
        """
        Overview:
            Wait at least enough(len(ready_conn) >= wait_num) connections within timeout constraint.
            If timeout is None and wait_num == len(ready_conn), means sync mode;
            If timeout is not None, will return when len(ready_conn) >= wait_num and
            this method takes more than timeout seconds.
        """
        assert 1 <= wait_num <= len(rest_conn
                                    ), 'please indicate proper wait_num: <wait_num: {}, rest_conn_num: {}>'.format(
                                        wait_num, len(rest_conn)
                                    )
        rest_conn_set = set(rest_conn)
        ready_conn = set()
        start_time = time.time()
        while len(rest_conn_set) > 0:
            if len(ready_conn) >= wait_num and timeout:
                if (time.time() - start_time) >= timeout:
                    break
            finish_conn = set(connection.wait(rest_conn_set, timeout=timeout))
            ready_conn = ready_conn.union(finish_conn)
            rest_conn_set = rest_conn_set.difference(finish_conn)
        ready_ids = [rest_conn.index(c) for c in ready_conn]
        return list(ready_conn), ready_ids


@ENV_MANAGER_REGISTRY.register('subprocess')
class SyncSubprocessEnvManager(AsyncSubprocessEnvManager):
    config = dict(
        episode_num=float("inf"),
        max_retry=1,
        step_timeout=None,
        auto_reset=True,
        reset_timeout=None,
        retry_type='reset',
        retry_waiting_time=0.1,
        # subprocess specified args
        shared_memory=True,
        copy_on_get=True,
        context='spawn' if platform.system().lower() == 'windows' else 'fork',
        wait_num=float("inf"),  # inf mean all the environments
        step_wait_timeout=None,
        connect_timeout=60,
        reset_inplace=False,  # if reset_inplace=True in SyncSubprocessEnvManager, the interaction can be reproducible.
    )

    def step(self, actions: Dict[int, Any]) -> Dict[int, namedtuple]:
        """
        Overview:
            Step all environments. Reset an env if done.
        Arguments:
            - actions (:obj:`Dict[int, Any]`): {env_id: action}
        Returns:
            - timesteps (:obj:`Dict[int, namedtuple]`): {env_id: timestep}. Timestep is a \
                ``BaseEnvTimestep`` tuple with observation, reward, done, env_info.
        Example:
            >>>     actions_dict = {env_id: model.forward(obs) for env_id, obs in obs_dict.items())}
            >>>     timesteps = env_manager.step(actions_dict):
            >>>     for env_id, timestep in timesteps.items():
            >>>         pass

        .. note::

            - The env_id that appears in ``actions`` will also be returned in ``timesteps``.
            - Each environment is run by a subprocess separately. Once an environment is done, it is reset immediately.
        """
        self._check_closed()
        env_ids = list(actions.keys())
        assert all([self._env_states[env_id] == EnvState.RUN for env_id in env_ids]
                   ), 'current env state are: {}, please check whether the requested env is in reset or done'.format(
                       {env_id: self._env_states[env_id]
                        for env_id in env_ids}
                   )
        for env_id, act in actions.items():
            # it is necessary to set kwargs as None for saving cost of serialization in some env like cartpole,
            # and step method never uses kwargs in known envs.
            self._pipe_parents[env_id].send(['step', [act], None])

        # ===     This part is different from async one.     ===
        # === Because operate in this way is more efficient. ===
        timesteps = {}
        ready_conn = [self._pipe_parents[env_id] for env_id in env_ids]
        # timesteps.update({env_id: p.recv() for env_id, p in zip(env_ids, ready_conn)})
        for env_id, p in zip(env_ids, ready_conn):
            try:
                timesteps.update({env_id: p.recv()})
            except pickle.UnpicklingError as e:
                timestep = BaseEnvTimestep(None, None, None, {'abnormal': True})
                timesteps.update({env_id: timestep})
                self._pipe_parents[env_id].close()
                if self._subprocesses[env_id].is_alive():
                    self._subprocesses[env_id].terminate()
                self._create_env_subprocess(env_id)
        self._check_data(timesteps)
        # ======================================================

        if self._shared_memory:
            # TODO(nyz) optimize sync shm
            for i, (env_id, timestep) in enumerate(timesteps.items()):
                timesteps[env_id] = timestep._replace(obs=self._obs_buffers[env_id].get())
        for env_id, timestep in timesteps.items():
            if is_abnormal_timestep(timestep):
                self._env_states[env_id] = EnvState.ERROR
                continue
            if timestep.done:
                self._env_episode_count[env_id] += 1
                if self._env_episode_count[env_id] < self._episode_num:
                    if self._auto_reset:
                        if self._reset_inplace:  # reset in subprocess at once
                            self._env_states[env_id] = EnvState.RUN
                            self._ready_obs[env_id] = timestep.obs
                        else:
                            # in this case, ready_obs is updated in ``self._reset``
                            self._env_states[env_id] = EnvState.RESET
                            reset_thread = PropagatingThread(target=self._reset, args=(env_id, ), name='regular_reset')
                            reset_thread.daemon = True
                            reset_thread.start()
                    else:
                        # in the case that auto_reset=False, caller should call ``env_manager.reset`` manually
                        self._env_states[env_id] = EnvState.NEED_RESET
                else:
                    self._env_states[env_id] = EnvState.DONE
            else:
                self._ready_obs[env_id] = timestep.obs
        return timesteps


@ENV_MANAGER_REGISTRY.register('subprocess_v2')
class SubprocessEnvManagerV2(SyncSubprocessEnvManager):
    """
    Overview:
        SyncSubprocessEnvManager for new task pipeline and interfaces coupled with treetensor.
    """

    @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)
        """
        no_done_env_idx = [i for i, s in self._env_states.items() if s != EnvState.DONE]
        sleep_count = 0
        while not any([self._env_states[i] == EnvState.RUN for i in no_done_env_idx]):
            if sleep_count != 0 and sleep_count % 10000 == 0:
                logging.warning(
                    'VEC_ENV_MANAGER: all the not done envs are resetting, sleep {} times'.format(sleep_count)
                )
            time.sleep(0.001)
            sleep_count += 1
        return tnp.stack([tnp.array(self._ready_obs[i]) for i in self.ready_env])

    def step(self, actions: Union[List[tnp.ndarray], tnp.ndarray]) -> List[tnp.ndarray]:
        """
        Overview:
            Execute env step according to input actions. And reset an env if done.
        Arguments:
            - actions (:obj:`Union[List[tnp.ndarray], tnp.ndarray]`): actions came from outer caller like policy.
        Returns:
            - timesteps (:obj:`List[tnp.ndarray]`): Each timestep is a tnp.array with observation, reward, done, \
                info, env_id.
        """
        if isinstance(actions, tnp.ndarray):
            # zip operation will lead to wrong behaviour if not split data
            split_action = tnp.split(actions, actions.shape[0])
            split_action = [s.squeeze(0) for s in split_action]
        else:
            split_action = actions
        actions = {env_id: a for env_id, a in zip(self.ready_obs_id, split_action)}
        timesteps = super().step(actions)
        new_data = []
        for env_id, timestep in timesteps.items():
            obs, reward, done, info = timestep
            # 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)
            info = remove_illegal_item(info)
            new_data.append(tnp.array({'obs': obs, 'reward': reward, 'done': done, 'info': info, 'env_id': env_id}))
        return new_data