Tf-Agents ParallelPyEnvironment以静默方式失败

Question

我已经编写了一个自定义环境，这样我就可以使用强化学习(PPO)和tf-agent。如果我将我的env (继承自py_environment.PyEnvironment)包装在一个TfPyEnvironment中，这可以很好地工作，但是如果我试图将它包装到一个ParallelPyEnvironment中，就会失败。我尝试过使用ParallelPyEnvironment的所有关键字参数，但是代码一直运行到行，然后什么也没有发生--没有例外，程序不会终止等等。

下面是初始化环境并展示eval_env的工作变体的代码

train_env = tf_py_environment.TFPyEnvironment(
    ParallelPyEnvironment(
        [CardGameEnv()] * hparams['parallel_environments']
    )
)
# this works perfectly:
eval_env = tf_py_environment.TFPyEnvironment(CardGameEnv(debug=True))

如果我通过CTRL+C终止脚本，则会输出以下内容：

Traceback (most recent call last):
Traceback (most recent call last):
  File "E:\Users\tmp\Documents\Programming\Neural Nets\Poker_AI\poker_logic\train.py", line 229, in <module>
  File "<string>", line 1, in <module>
    train(model_num=3)
  File "C:\Python37\lib\multiprocessing\spawn.py", line 105, in spawn_main
  File "E:\Users\tmp\Documents\Programming\Neural Nets\Poker_AI\poker_logic\train.py", line 64, in train
    [CardGameEnv()] * hparams['parallel_environments']
    exitcode = _main(fd)
  File "E:\Users\tmp\AppData\Roaming\Python\Python37\site-packages\gin\config.py", line 1009, in wrapper
  File "C:\Python37\lib\multiprocessing\spawn.py", line 113, in _main
    preparation_data = reduction.pickle.load(from_parent)
KeyboardInterrupt
    return fn(*new_args, **new_kwargs)
  File "C:\Python37\lib\site-packages\tf_agents\environments\parallel_py_environment.py", line 70, in __init__
    self.start()
  File "C:\Python37\lib\site-packages\tf_agents\environments\parallel_py_environment.py", line 83, in start
    env.start(wait_to_start=self._start_serially)
  File "C:\Python37\lib\site-packages\tf_agents\environments\parallel_py_environment.py", line 223, in start
    self._process.start()
  File "C:\Python37\lib\multiprocessing\process.py", line 112, in start
    self._popen = self._Popen(self)
  File "C:\Python37\lib\multiprocessing\context.py", line 223, in _Popen
    return _default_context.get_context().Process._Popen(process_obj)
  File "C:\Python37\lib\multiprocessing\context.py", line 322, in _Popen
    return Popen(process_obj)
  File "C:\Python37\lib\multiprocessing\popen_spawn_win32.py", line 65, in __init__
    reduction.dump(process_obj, to_child)
  File "C:\Python37\lib\multiprocessing\reduction.py", line 60, in dump
    ForkingPickler(file, protocol).dump(obj)
  File "C:\Python37\lib\site-packages\tf_agents\environments\parallel_py_environment.py", line 264, in __getattr__
    return self._receive()
  File "C:\Python37\lib\site-packages\tf_agents\environments\parallel_py_environment.py", line 333, in _receive
    message, payload = self._conn.recv()
  File "C:\Python37\lib\multiprocessing\connection.py", line 250, in recv
    buf = self._recv_bytes()
  File "C:\Python37\lib\multiprocessing\connection.py", line 306, in _recv_bytes
    [ov.event], False, INFINITE)
KeyboardInterrupt
Error in atexit._run_exitfuncs:
Traceback (most recent call last):
  File "C:\Python37\lib\site-packages\tf_agents\environments\parallel_py_environment.py", line 289, in close
    self._process.join(5)
  File "C:\Python37\lib\multiprocessing\process.py", line 139, in join
    assert self._popen is not None, 'can only join a started process'
AssertionError: can only join a started process

由此我得出结论，ParallelPyEnvironment正在尝试启动的线程不能做到这一点，但由于我对Python中的线程没有太多的经验，我不知道从哪里开始，特别是如何修复这个问题。目前的培训需要很长的时间，而且根本不使用我的PC的功能(使用了3 3GB的32 3GB内存，处理器为3%，GPU几乎不能工作，但VRAM已满)，因此这应该会大大加快培训时间。

Answer 1

解决方案是传入可调用对象，而不是环境，这样ParallelPyEnvironment就可以自己构造它们：

train_env = tf_py_environment.TFPyEnvironment(
    ParallelPyEnvironment(
        [CardGameEnv] * hparams['parallel_environments']
    )
)