在强化学习中（on-line）的算法如果是on-policy的算法都是需要较大的采样样本的，因此采样的效率往往对整个算法运行效率有着自关重要的影响，在deepmind（Google）公司的强化学习的并行采样设计中往往使用带有timeout的带有时间过期的队列（queue）来进行并行采样的多进程间的同步和通信，但是这种设计往往十分复杂，编程难度也极大，不过也正因如此也比较吸引人们的关注，对此我也是如此，不过我一直在考虑这种timeout的同步通信方式是否真的可以提高运行效率，如果不能的话或者提高的效率有限的话，那么搞这么复杂的设计是不是自找苦吃，也或者只是一种炫技的操作呢。
为了分析强化学习的并行采样到底应该如何设计，或者说不同种类的on-line的on-policy的同步并行采样应该如何设计，为此在项目：
https://openi.pcl.ac.cn/devilmaycry812839668/pytorch-maml-rl
基础上修改出了下面的实验项目：
https://openi.pcl.ac.cn/devilmaycry812839668/rl-sampling-experiments
其实，这个新改出的项目其实就是改了两个文件，即，multi_task_sampler.py 和 sync_vector_env_steps.py
给出 multi_task_sampler.py 的代码：
点击查看代码
########################################
import os
import time
import signal
import sys
import copy
import numpy as np
from queue import Empty
from multiprocessing import Queue
from multiprocessing.sharedctypes import RawArray
from typing import List, Callable, Tuple, Dict
from gymnasium import Env ##################################
from ctypes import c_uint, c_float, c_double, c_int64
import torch
from maml_rl.envs.utils.sync_vector_env_steps import ChildEnv
from maml_rl.samplers.sampler import make_env
NUMPY_TO_C_DTYPE = {np.float32: c_float,
np.float64: c_double,
np.uint8: c_uint,
np.int64: c_int64,}
class MultiTaskSampler():
def __init__(self,
env_name,
env_kwargs,
batch_size,
policy,
baseline,
seed=None,
num_workers=8):
self.seed = seed
self.batch_size = batch_size
self.num_workers = num_workers
self.policy = policy
self.baseline = baseline
env_fns = [make_env(env_name, env_kwargs=env_kwargs) for _ in range(batch_size*num_workers)]
self.vector_env = SynVectorEnv(env_fns, num_workers)
self.env = make_env(env_name, env_kwargs=env_kwargs)()
self.first_init = True
def close(self):
self.vector_env.close()
def sample(self, max_length, params=None):
self.done_times = np.array([max_length+1, ]*self.num_workers)
if self.first_init:
self.first_init = False
seed = self.seed
else:
seed = None
# self.task_results = [{"obs":[[],], "action":[[],], "reward":[[],], "terminated":[[],], \
# "truncated":[[],], "length":0, "episodes":0} \
# for _ in range(self.batch_size*self.num_workers)]
self.task_results = [[] for _ in range(self.num_workers)]
####################################### reset_task
tasks_profile = self.env.unwrapped.sample_tasks(self.batch_size*self.num_workers)
self.vector_env.reset_task(tasks_profile)
####################################### reset
# obss, infos = self.vector_env.reset(seed=seed)
data_dict = self.vector_env.reset(seed=seed)
while data_dict:
print(len(data_dict))
remove_list = []
for i, (index, data) in enumerate(data_dict.items()):
# data[1] ### infos
self.task_results[index].append(data[0])
self.done_times[index] -= 1
if self.done_times[index] == 0:
remove_list.append(index)
for it in remove_list:
data_dict.pop(it)
if len(data_dict) == 0:
data_dict = self.vector_env.step_recv()
continue
obs_numpy = np.concatenate([data_dict[index][0]["next_obs"] for index in data_dict.keys()], dtype=np.float32)
observations_tensor = torch.from_numpy(obs_numpy)
pi = self.policy(observations_tensor, params=params)
actions_tensor = pi.sample()
actions = actions_tensor.cpu().numpy()
for i, (index, data) in enumerate(data_dict.items()):
# data[1] ### infos
# self.task_results[index][-1]["next_actions"] = actions[i*self.batch_size:(i+1)*self.batch_size]
self.vector_env.step_send(index, actions[i*self.batch_size:(i+1)*self.batch_size])
data_dict = self.vector_env.step_recv()
"""
next_obs
next_obs_truncated
next_actions
rewards
terminated
truncated
self.shared_data["actions"] = _actions
self.shared_data["rewards"] = _rewards
self.shared_data["next_obs"] = _next_obs
self.shared_data["next_obs_truncated"] = _next_obs_truncated
self.shared_data["terminated"] = _terminated
self.shared_data["truncated"] = _truncated
"""
def collect(self):
for i, t_result in enumerate(self.task_results):
task_obs = []
task_actions = []
task_rewards = []
for j, step_dict in enumerate(t_result):
if j == 0:
obs = [step_dict["next_obs"],]
actions = []
rewards = []
continue
elif step_dict["terminated"]:
actions.append(step_dict["actions"])
rewards.append(step_dict["rewards"])
# obs.append(step_dict["next_obs"])
task_obs.append(obs)
task_actions.append(actions)
task_rewards.append(rewards)
obs = [step_dict["next_obs"],]
actions = []
rewards = []
elif step_dict["truncated"]:
actions.append(step_dict["actions"])
rewards.append(step_dict["rewards"])
obs.append(step_dict["next_obs_truncated"])
task_obs.append(obs)
task_actions.append(actions)
task_rewards.append(rewards)
obs = [step_dict["next_obs"],]
actions = []
rewards = []
else:
actions.append(step_dict["actions"])
rewards.append(step_dict["rewards"])
obs.append(step_dict["next_obs"])
task_episodes = {}
for i, result in enumerate(self.task_results):
for j in range(self.batch_size):
# result[j] self.batch_size*i+j
x = 0
task_episodes[self.batch_size*i+j]={x:{"obs":[],}, }
obs = [[], ]
actions = [[], ]
rewards = [[], ]
terminated = [[], ]
truncated = [[], ]
#########################################1
class SynVectorEnv(object):
def __init__(self, env_fns:List[Callable[[], Env]], n_process:int):
self.n_process = n_process
self.env_fns = env_fns
assert len(env_fns)%n_process == 0, "env_fns must be divided by n_process"
self.notify_queues = [Queue() for _ in range(n_process)]
self.barrier = Queue()
# self.shared_data = {"actions":None,
# "rewards":None,
# "next_obs":None,
# "next_obs_truncated":None,
# "terminated":None,
# "truncated":None,}
self.shared_data = dict()
self.init_shared_data(env_fns)
self.sub_shared_data = []
self.runners = []
for i, _s_env_fns in enumerate(np.split(np.asarray(env_fns), n_process)):
_s_shared_data = dict()
for _name, _data in self.shared_data.items():
_s_shared_data[_name] = np.split(_data, n_process)[i]
self.runners.append(ChildEnv(i, list(_s_env_fns), _s_shared_data, self.notify_queues[i], self.barrier))
self.sub_shared_data.append(_s_shared_data)
for r in self.runners: ### 启动子进程
r.start()
self._running_flag = True
self.clean()
def init_shared_data(self, env_fns:List[Callable[[], Env]]):
N = len(env_fns)
_env:Env = env_fns[0]()
_obs, _ = _env.reset()
if _env.action_space.sample().size == 1:
_actions = np.zeros(N, dtype=_env.action_space.dtype)
else:
_actions = np.asarray([np.zeros(_env.action_space.sample().size, dtype=_env.action_space.dtype) for _ in range(N)])
_rewards = np.zeros(N, dtype=np.float32)
_next_obs = np.asarray([ _obs for env in env_fns ])
_next_obs_truncated = np.asarray([ _obs for env in env_fns ])
_terminated = np.zeros(N, dtype=np.float32)
_truncated = np.zeros(N, dtype=np.float32)
_actions = self._get_shared(_actions)
_rewards = self._get_shared(_rewards)
_next_obs = self._get_shared(_next_obs)
_next_obs_truncated = self._get_shared(_next_obs_truncated)
_terminated = self._get_shared(_terminated)
_truncated = self._get_shared(_truncated)
self.shared_data["actions"] = _actions
self.shared_data["rewards"] = _rewards
self.shared_data["next_obs"] = _next_obs
self.shared_data["next_obs_truncated"] = _next_obs_truncated
self.shared_data["terminated"] = _terminated
self.shared_data["truncated"] = _truncated
self.running_num = 0
def _get_shared(self, array:np.ndarray):
"""
Returns a RawArray backed numpy array that can be shared between processes.
:param array: the array to be shared
:return: the RawArray backed numpy array
"""
dtype = NUMPY_TO_C_DTYPE[array.dtype.type]
shared_array = RawArray(dtype, array.size)
return np.frombuffer(shared_array, dtype).reshape(array.shape)
"""
instruction:
None: 结束进程
0 -> "reset": 重置环境, "seed": 设置随机种子
1 -> "reset_task": 重置环境任务
2 -> "step": 执行step操作
"""
def reset(self, seed:int|None=None):
results = {}
for i, _queue in enumerate(self.notify_queues):
if seed is None:
_queue.put((0, seed))
else:
_queue.put((0, seed+i*len(self.env_fns)//self.n_process))
for _ in range(self.n_process):
_index, _instruction, _info = self.barrier.get()
if _instruction != 0:
raise ValueError("instruction must be 0, reset operation")
# results[_index] = [np.copy(self.sub_shared_data[i]["next_obs"]), _info]
results[_index] = [copy.deepcopy(self.sub_shared_data[i]), _info]
return results
def reset_task(self, tasks):
for i, _queue in enumerate(self.notify_queues):
_queue.put((1, tasks[i*(self.n_process):(i+1)*(self.n_process)]))
def step_send(self, index:int, actions:np.ndarray):
self.sub_shared_data[index]["actions"][:] = actions
self.notify_queues[index].put((2, None)) ### None 只是为了占位用
self.running_num += 1
def step_recv(self):
if self.running_num == 0:
return None
_index, _instruction, _info = self.barrier.get()
assert _instruction == 2, "instruction must be 2, step operation"
results ={_index: [copy.deepcopy(self.sub_shared_data[_index]), _info]}
for _ in range(self.running_num-1):
try:
_index, _instruction, _info = self.barrier.get(timeout=0.0001)
# _index, _instruction, _info = self.barrier.get()
assert _instruction == 2, "instruction must be 2, step operation"
results[_index] = [copy.deepcopy(self.sub_shared_data[_index]), _info]
except Empty as e:
# print(e)
pass
self.running_num -= len(results)
return results
def close(self):
for queue in self.notify_queues:
queue.put((None, None))
def get_shared_variables(self) -> Dict[str, np.ndarray]:
return self.shared_data
def clean(self):
main_process_pid = os.getpid()
def signal_handler(signal, frame):
if os.getpid() == main_process_pid:
print('Signal ' + str(signal) + ' detected, cleaning up.')
if self._running_flag:
self._running_flag = False
self.close()
print('Cleanup completed, shutting down...')
sys.exit(0)
signal.signal(signal.SIGTERM, signal_handler)
signal.signal(signal.SIGINT, signal_handler)以及 sync_vector_env_steps.py 的代码实现：
点击查看代码
from multiprocessing import Process
from multiprocessing import Queue
import numpy as np
from numpy import ndarray
from typing import Callable, Dict, List, Tuple
"""目前主要支持gymnasium环境，gym环境未成功支持"""
from gymnasium import Env
# try:
# import gymnasium as gym
# except ImportError:
# import gym
# from gym import Env
"""
多进程实现：
以子进程的方式运行游戏环境并与主进程交互数据
"""
class ChildEnv(Process):
def __init__(self, index:int, env_fns:List[Callable[[], Env]], shared_data:Dict[str, ndarray],
notify_queue:Queue, barrier:Queue):
super(ChildEnv, self).__init__()
self.index = index # 子进程的编号
self.envs = [env_fn() for env_fn in env_fns]
self.actions = shared_data["actions"]
self.rewards = shared_data["rewards"]
self.next_obs = shared_data["next_obs"]
self.next_obs_truncated = shared_data["next_obs_truncated"]
self.terminated = shared_data["terminated"]
self.truncated = shared_data["truncated"]
self.notify_queue = notify_queue
self.barrier = barrier
# Check that all sub-environments are meta-learning environments
for env in self.envs:
if not hasattr(env.unwrapped, 'reset_task'):
raise ValueError('The environment provided is not a '
'meta-learning environment. It does not have '
'the method `reset_task` implemented.')
def run(self):
super(ChildEnv, self).run()
"""
instruction:
None: 结束进程
0 -> "reset": 重置环境, "seed": 设置随机种子
1 -> "reset_task": 重置环境任务
2 -> "step": 执行step操作
"""
while True:
instruction, data = self.notify_queue.get()
if instruction == 2: # step操作
# _actions = data # return: (self.index, instruction, _infos)
# _actions = self.actions
pass
elif instruction == 0: # reset操作
_seed = data if data else np.random.randint(2 ** 16 - 1)
_infos = []
for i, env in enumerate(self.envs):
self.next_obs[i], _info = env.reset(seed=_seed+i)
_infos.append(_info)
self.barrier.put((self.index, instruction, _infos)) # 通知主进程, reset, observations初始化
# self.barrier.put((self.index, instruction, None)) # 通知主进程, reset, observations初始化
continue
elif instruction == 1: # reset_task操作
_tasks = data
for env, _task in zip(self.envs, _tasks):
env.unwrapped.reset_task(_task)
continue
elif instruction is None:
break ### exit sub_process
else:
raise ValueError('Unknown instruction: {}'.format(instruction))
_infos = []
for i, (env, action) in enumerate(zip(self.envs, self.actions)):
next_obs, reward, terminated, truncated, info = env.step(action)
if terminated:
self.next_obs[i], _sub_info = env.reset()
info["terminated"] = {"info":_sub_info, }
elif truncated:
self.next_obs_truncated[i] = next_obs
self.next_obs[i], _sub_info = env.reset()
info["truncated"] = {"info":_sub_info,
# "next_obs":self.next_obs_truncated[i]
}
self.rewards[i] = reward
self.terminated[i] = terminated
self.truncated[i] = truncated
_infos.append(info)
# self.barrier.put((self.index, instruction, _infos))
self.barrier.put((self.index, instruction, None)) # 减少发送内容大小
"""
# def step_wait(self):
def step(self, actions):
self._actions = actions
observations_list, infos = [], []
batch_ids, j = [], 0
num_actions = len(self._actions)
rewards = np.zeros((num_actions,), dtype=np.float32)
for i, env in enumerate(self.envs):
if self._dones[i]:
continue
action = self._actions[j]
observation, rewards[j], self._done, _truncation, info = env.step(action)
batch_ids.append(i)
self._dones[i] = self._done or _truncation
if not self._dones[i]:
observations_list.append(observation)
infos.append(info)
j += 1
assert num_actions == j
if observations_list:
observations = create_empty_array(self.single_observation_space,
n=len(observations_list),
fn=np.zeros)
concatenate(self.single_observation_space,
observations_list,
observations)
observations = np.asarray(observations, dtype=np.float32)
else:
observations = None
return (observations, rewards, np.copy(self._dones),
{'batch_ids': batch_ids, 'infos': infos})
"""其中，sync_vector_env_steps.py 文件的实现是比较常见的，就是多进程同步并行，然后每个进程下环境串行，以此来增加当次的通信数据量并提高吞吐效率，最终实现提高并行运算的效率。
该实现中主进程和子进程间的通信使用共享内存的方式进行，同步的信号通过队列的方式进行传递，该中方式被认为是on-policy并行中效率最高的方式，具体可以参见gymnasium中的更详细的实现方式。
其实，最为关键的实现代码是在 文件 multi_task_sampler.py 中。
在这里设计 step_send 函数，该函数接受的是主进程发送给子进程的 actions数据，其中，index 代表发送给的目标子进程的ID编号。
def step_send(self, index:int, actions:np.ndarray):
self.sub_shared_data[index]["actions"][:] = actions
self.notify_queues[index].put((2, None)) ### None 只是为了占位用
self.running_num += 1step_recv 函数代表着从子进程中接受传递给主进程的消息信号，其中，_index 代表着发送方的子进程的ID号，收到这个信号就说明这个子进程对应的共享内存的数据已经被子进程写好了，子进程只需要从这块共享数据中将数据取走即可。因为这块共享内存未来还是要被下次的存放操作所覆盖的，因此这里需要对这块共享内存中的数据就行deepcopy，以保证这部分内存的数据在后期不会被改变。
def step_recv(self):
if self.running_num == 0:
return None
_index, _instruction, _info = self.barrier.get()
assert _instruction == 2, "instruction must be 2, step operation"
results ={_index: [copy.deepcopy(self.sub_shared_data[_index]), _info]}
for _ in range(self.running_num-1):
try:
_index, _instruction, _info = self.barrier.get(timeout=0.0001)
# _index, _instruction, _info = self.barrier.get()
assert _instruction == 2, "instruction must be 2, step operation"
results[_index] = [copy.deepcopy(self.sub_shared_data[_index]), _info]
except Empty as e:
# print(e)
pass
self.running_num -= len(results)
return results而其中体现出同步通信的队列的timeout的超时操作的实现如下：
_index, _instruction, _info = self.barrier.get(timeout=0.0001)
可以看到，如果timeout时间过小，那么主进程就不会等待所有子进程都完成一个step的操作，而是将部分进程的单步step数据传递给上一层的操作；同理，如果timeout的时间过大，那么也就意味着主进程会等待所有子进程的单步step的操作数据，然后一并上传递给上一层操作。
也正是这一点的不同才是Google的实现与其他实现的不同所在，在Google实现中是要把这个timeout设置为一个比较小的数值，以保证主进程不会一直堵塞在等待子进程数据的操作上，这样来提高算法的采样效率，而我一直疑问的也就是这个地方，这个timeout是否应该存在，存在的意义在哪，或者更应该说到底哪类应用中该设置这个timeout.
