def run_align_norm_obs(raw_env, train_env, test_env, action_list): eps = np.finfo(np.float32).eps.item() raw_obs, train_obs = [raw_env.reset()], [train_env.reset()] for action in action_list: obs, rew, done, info = raw_env.step(action) raw_obs.append(obs) if np.any(done): raw_obs.append(raw_env.reset(np.where(done)[0])) obs, rew, done, info = train_env.step(action) train_obs.append(obs) if np.any(done): train_obs.append(train_env.reset(np.where(done)[0])) ref_rms = RunningMeanStd() for ro, to in zip(raw_obs, train_obs): ref_rms.update(ro) no = (ro - ref_rms.mean) / np.sqrt(ref_rms.var + eps) assert np.allclose(no, to) assert np.allclose(ref_rms.mean, train_env.get_obs_rms().mean) assert np.allclose(ref_rms.var, train_env.get_obs_rms().var) assert np.allclose(ref_rms.mean, test_env.get_obs_rms().mean) assert np.allclose(ref_rms.var, test_env.get_obs_rms().var) test_obs = [test_env.reset()] for action in action_list: obs, rew, done, info = test_env.step(action) test_obs.append(obs) if np.any(done): test_obs.append(test_env.reset(np.where(done)[0])) for ro, to in zip(raw_obs, test_obs): no = (ro - ref_rms.mean) / np.sqrt(ref_rms.var + eps) assert np.allclose(no, to)
def __init__( self, model: torch.nn.Module, optim: torch.optim.Optimizer, dist_fn: Type[torch.distributions.Distribution], discount_factor: float = 0.99, reward_normalization: bool = False, action_scaling: bool = True, action_bound_method: str = "clip", deterministic_eval: bool = False, **kwargs: Any, ) -> None: super().__init__( action_scaling=action_scaling, action_bound_method=action_bound_method, **kwargs ) self.actor = model self.optim = optim self.dist_fn = dist_fn assert 0.0 <= discount_factor <= 1.0, "discount factor should be in [0, 1]" self._gamma = discount_factor self._rew_norm = reward_normalization self.ret_rms = RunningMeanStd() self._eps = 1e-8 self._deterministic_eval = deterministic_eval
def __init__( self, venv: BaseVectorEnv, update_obs_rms: bool = True, ) -> None: super().__init__(venv) # initialize observation running mean/std self.update_obs_rms = update_obs_rms self.obs_rms = RunningMeanStd()
class VectorEnvNormObs(VectorEnvWrapper): """An observation normalization wrapper for vectorized environments. :param bool update_obs_rms: whether to update obs_rms. Default to True. :param float clip_obs: the maximum absolute value for observation. Default to 10.0. :param float epsilon: To avoid division by zero. """ def __init__( self, venv: BaseVectorEnv, update_obs_rms: bool = True, clip_obs: float = 10.0, epsilon: float = np.finfo(np.float32).eps.item(), ) -> None: super().__init__(venv) # initialize observation running mean/std self.update_obs_rms = update_obs_rms self.obs_rms = RunningMeanStd() self.clip_max = clip_obs self.eps = epsilon # TODO: compatible issue with reset -> (obs, info) def reset( self, id: Optional[Union[int, List[int], np.ndarray]] = None) -> np.ndarray: obs = self.venv.reset(id) if self.obs_rms and self.update_obs_rms: self.obs_rms.update(obs) return self._norm_obs(obs) def step( self, action: np.ndarray, id: Optional[Union[int, List[int], np.ndarray]] = None, ) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]: obs, rew, done, info = self.venv.step(action, id) if self.obs_rms and self.update_obs_rms: self.obs_rms.update(obs) return self._norm_obs(obs), rew, done, info def _norm_obs(self, obs: np.ndarray) -> np.ndarray: if self.obs_rms: obs = (obs - self.obs_rms.mean) / np.sqrt(self.obs_rms.var + self.eps) obs = np.clip(obs, -self.clip_max, self.clip_max) return obs def set_obs_rms(self, obs_rms: RunningMeanStd) -> None: """Set with given observation running mean/std.""" self.obs_rms = obs_rms def get_obs_rms(self) -> RunningMeanStd: """Return observation running mean/std.""" return self.obs_rms
def normalize_all_obs_in_replay_buffer( replay_buffer: ReplayBuffer) -> Tuple[ReplayBuffer, RunningMeanStd]: # compute obs mean and var obs_rms = RunningMeanStd() obs_rms.update(replay_buffer.obs) _eps = np.finfo(np.float32).eps.item() # normalize obs replay_buffer._meta["obs"] = (replay_buffer.obs - obs_rms.mean) / np.sqrt(obs_rms.var + _eps) replay_buffer._meta["obs_next"] = ( replay_buffer.obs_next - obs_rms.mean) / np.sqrt(obs_rms.var + _eps) return replay_buffer, obs_rms
def __init__( self, venv: BaseVectorEnv, update_obs_rms: bool = True, clip_obs: float = 10.0, epsilon: float = np.finfo(np.float32).eps.item(), ) -> None: super().__init__(venv) # initialize observation running mean/std self.update_obs_rms = update_obs_rms self.obs_rms = RunningMeanStd() self.clip_max = clip_obs self.eps = epsilon
def __init__( self, env_fns: List[Callable[[], gym.Env]], worker_fn: Callable[[Callable[[], gym.Env]], EnvWorker], wait_num: Optional[int] = None, timeout: Optional[float] = None, norm_obs: bool = False, obs_rms: Optional[RunningMeanStd] = None, update_obs_rms: bool = True, ) -> None: self._env_fns = env_fns # A VectorEnv contains a pool of EnvWorkers, which corresponds to # interact with the given envs (one worker <-> one env). self.workers = [worker_fn(fn) for fn in env_fns] self.worker_class = type(self.workers[0]) assert issubclass(self.worker_class, EnvWorker) assert all([isinstance(w, self.worker_class) for w in self.workers]) self.env_num = len(env_fns) self.wait_num = wait_num or len(env_fns) assert 1 <= self.wait_num <= len(env_fns), \ f"wait_num should be in [1, {len(env_fns)}], but got {wait_num}" self.timeout = timeout assert self.timeout is None or self.timeout > 0, \ f"timeout is {timeout}, it should be positive if provided!" self.is_async = self.wait_num != len(env_fns) or timeout is not None self.waiting_conn: List[EnvWorker] = [] # environments in self.ready_id is actually ready # but environments in self.waiting_id are just waiting when checked, # and they may be ready now, but this is not known until we check it # in the step() function self.waiting_id: List[int] = [] # all environments are ready in the beginning self.ready_id = list(range(self.env_num)) self.is_closed = False # initialize observation running mean/std self.norm_obs = norm_obs self.update_obs_rms = update_obs_rms self.obs_rms = RunningMeanStd( ) if obs_rms is None and norm_obs else obs_rms self.__eps = np.finfo(np.float32).eps.item()
def test_rms(): rms = RunningMeanStd() assert np.allclose(rms.mean, 0) assert np.allclose(rms.var, 1) rms.update(np.array([[[1, 2], [3, 5]]])) rms.update(np.array([[[1, 2], [3, 4]], [[1, 2], [0, 0]]])) assert np.allclose(rms.mean, np.array([[1, 2], [2, 3]]), atol=1e-3) assert np.allclose(rms.var, np.array([[0, 0], [2, 14 / 3.]]), atol=1e-3)
class PGPolicy(BasePolicy): """Implementation of REINFORCE algorithm. :param torch.nn.Module model: a model following the rules in :class:`~tianshou.policy.BasePolicy`. (s -> logits) :param torch.optim.Optimizer optim: a torch.optim for optimizing the model. :param dist_fn: distribution class for computing the action. :type dist_fn: Type[torch.distributions.Distribution] :param float discount_factor: in [0, 1]. Default to 0.99. :param bool action_scaling: whether to map actions from range [-1, 1] to range [action_spaces.low, action_spaces.high]. Default to True. :param str action_bound_method: method to bound action to range [-1, 1], can be either "clip" (for simply clipping the action), "tanh" (for applying tanh squashing) for now, or empty string for no bounding. Default to "clip". :param Optional[gym.Space] action_space: env's action space, mandatory if you want to use option "action_scaling" or "action_bound_method". Default to None. :param lr_scheduler: a learning rate scheduler that adjusts the learning rate in optimizer in each policy.update(). Default to None (no lr_scheduler). :param bool deterministic_eval: whether to use deterministic action instead of stochastic action sampled by the policy. Default to False. .. seealso:: Please refer to :class:`~tianshou.policy.BasePolicy` for more detailed explanation. """ def __init__( self, model: torch.nn.Module, optim: torch.optim.Optimizer, dist_fn: Type[torch.distributions.Distribution], discount_factor: float = 0.99, reward_normalization: bool = False, action_scaling: bool = True, action_bound_method: str = "clip", deterministic_eval: bool = False, **kwargs: Any, ) -> None: super().__init__( action_scaling=action_scaling, action_bound_method=action_bound_method, **kwargs ) self.actor = model self.optim = optim self.dist_fn = dist_fn assert 0.0 <= discount_factor <= 1.0, "discount factor should be in [0, 1]" self._gamma = discount_factor self._rew_norm = reward_normalization self.ret_rms = RunningMeanStd() self._eps = 1e-8 self._deterministic_eval = deterministic_eval def process_fn( self, batch: Batch, buffer: ReplayBuffer, indices: np.ndarray ) -> Batch: r"""Compute the discounted returns for each transition. .. math:: G_t = \sum_{i=t}^T \gamma^{i-t}r_i where :math:`T` is the terminal time step, :math:`\gamma` is the discount factor, :math:`\gamma \in [0, 1]`. """ v_s_ = np.full(indices.shape, self.ret_rms.mean) unnormalized_returns, _ = self.compute_episodic_return( batch, buffer, indices, v_s_=v_s_, gamma=self._gamma, gae_lambda=1.0 ) if self._rew_norm: batch.returns = (unnormalized_returns - self.ret_rms.mean) / \ np.sqrt(self.ret_rms.var + self._eps) self.ret_rms.update(unnormalized_returns) else: batch.returns = unnormalized_returns return batch def forward( self, batch: Batch, state: Optional[Union[dict, Batch, np.ndarray]] = None, **kwargs: Any, ) -> Batch: """Compute action over the given batch data. :return: A :class:`~tianshou.data.Batch` which has 4 keys: * ``act`` the action. * ``logits`` the network's raw output. * ``dist`` the action distribution. * ``state`` the hidden state. .. seealso:: Please refer to :meth:`~tianshou.policy.BasePolicy.forward` for more detailed explanation. """ logits, hidden = self.actor(batch.obs, state=state) if isinstance(logits, tuple): dist = self.dist_fn(*logits) else: dist = self.dist_fn(logits) if self._deterministic_eval and not self.training: if self.action_type == "discrete": act = logits.argmax(-1) elif self.action_type == "continuous": act = logits[0] else: act = dist.sample() return Batch(logits=logits, act=act, state=hidden, dist=dist) def learn( # type: ignore self, batch: Batch, batch_size: int, repeat: int, **kwargs: Any ) -> Dict[str, List[float]]: losses = [] for _ in range(repeat): for minibatch in batch.split(batch_size, merge_last=True): self.optim.zero_grad() result = self(minibatch) dist = result.dist act = to_torch_as(minibatch.act, result.act) ret = to_torch(minibatch.returns, torch.float, result.act.device) log_prob = dist.log_prob(act).reshape(len(ret), -1).transpose(0, 1) loss = -(log_prob * ret).mean() loss.backward() self.optim.step() losses.append(loss.item()) return {"loss": losses}
class BaseVectorEnv(object): """Base class for vectorized environments wrapper. Usage: :: env_num = 8 envs = DummyVectorEnv([lambda: gym.make(task) for _ in range(env_num)]) assert len(envs) == env_num It accepts a list of environment generators. In other words, an environment generator ``efn`` of a specific task means that ``efn()`` returns the environment of the given task, for example, ``gym.make(task)``. All of the VectorEnv must inherit :class:`~tianshou.env.BaseVectorEnv`. Here are some other usages: :: envs.seed(2) # which is equal to the next line envs.seed([2, 3, 4, 5, 6, 7, 8, 9]) # set specific seed for each env obs = envs.reset() # reset all environments obs = envs.reset([0, 5, 7]) # reset 3 specific environments obs, rew, done, info = envs.step([1] * 8) # step synchronously envs.render() # render all environments envs.close() # close all environments .. warning:: If you use your own environment, please make sure the ``seed`` method is set up properly, e.g., :: def seed(self, seed): np.random.seed(seed) Otherwise, the outputs of these envs may be the same with each other. :param env_fns: a list of callable envs, ``env_fns[i]()`` generates the i-th env. :param worker_fn: a callable worker, ``worker_fn(env_fns[i])`` generates a worker which contains the i-th env. :param int wait_num: use in asynchronous simulation if the time cost of ``env.step`` varies with time and synchronously waiting for all environments to finish a step is time-wasting. In that case, we can return when ``wait_num`` environments finish a step and keep on simulation in these environments. If ``None``, asynchronous simulation is disabled; else, ``1 <= wait_num <= env_num``. :param float timeout: use in asynchronous simulation same as above, in each vectorized step it only deal with those environments spending time within ``timeout`` seconds. :param bool norm_obs: Whether to track mean/std of data and normalize observation on return. For now, observation normalization only support observation of type np.ndarray. :param obs_rms: class to track mean&std of observation. If not given, it will initialize a new one. Usually in envs that is used to evaluate algorithm, obs_rms should be passed in. Default to None. :param bool update_obs_rms: Whether to update obs_rms. Default to True. """ def __init__( self, env_fns: List[Callable[[], gym.Env]], worker_fn: Callable[[Callable[[], gym.Env]], EnvWorker], wait_num: Optional[int] = None, timeout: Optional[float] = None, norm_obs: bool = False, obs_rms: Optional[RunningMeanStd] = None, update_obs_rms: bool = True, ) -> None: self._env_fns = env_fns # A VectorEnv contains a pool of EnvWorkers, which corresponds to # interact with the given envs (one worker <-> one env). self.workers = [worker_fn(fn) for fn in env_fns] self.worker_class = type(self.workers[0]) assert issubclass(self.worker_class, EnvWorker) assert all([isinstance(w, self.worker_class) for w in self.workers]) self.env_num = len(env_fns) self.wait_num = wait_num or len(env_fns) assert 1 <= self.wait_num <= len(env_fns), \ f"wait_num should be in [1, {len(env_fns)}], but got {wait_num}" self.timeout = timeout assert self.timeout is None or self.timeout > 0, \ f"timeout is {timeout}, it should be positive if provided!" self.is_async = self.wait_num != len(env_fns) or timeout is not None self.waiting_conn: List[EnvWorker] = [] # environments in self.ready_id is actually ready # but environments in self.waiting_id are just waiting when checked, # and they may be ready now, but this is not known until we check it # in the step() function self.waiting_id: List[int] = [] # all environments are ready in the beginning self.ready_id = list(range(self.env_num)) self.is_closed = False # initialize observation running mean/std self.norm_obs = norm_obs self.update_obs_rms = update_obs_rms self.obs_rms = RunningMeanStd( ) if obs_rms is None and norm_obs else obs_rms self.__eps = np.finfo(np.float32).eps.item() def _assert_is_not_closed(self) -> None: assert not self.is_closed, \ f"Methods of {self.__class__.__name__} cannot be called after close." def __len__(self) -> int: """Return len(self), which is the number of environments.""" return self.env_num def __getattribute__(self, key: str) -> Any: """Switch the attribute getter depending on the key. Any class who inherits ``gym.Env`` will inherit some attributes, like ``action_space``. However, we would like the attribute lookup to go straight into the worker (in fact, this vector env's action_space is always None). """ if key in [ 'metadata', 'reward_range', 'spec', 'action_space', 'observation_space' ]: # reserved keys in gym.Env return self.get_env_attr(key) else: return super().__getattribute__(key) def get_env_attr(self, key: str, id: Optional[Union[int, List[int], np.ndarray]] = None) -> List[Any]: """Get an attribute from the underlying environments. If id is an int, retrieve the attribute denoted by key from the environment underlying the worker at index id. The result is returned as a list with one element. Otherwise, retrieve the attribute for all workers at indices id and return a list that is ordered correspondingly to id. :param str key: The key of the desired attribute. :param id: Indice(s) of the desired worker(s). Default to None for all env_id. :return list: The list of environment attributes. """ self._assert_is_not_closed() id = self._wrap_id(id) if self.is_async: self._assert_id(id) return [self.workers[j].get_env_attr(key) for j in id] def set_env_attr( self, key: str, value: Any, id: Optional[Union[int, List[int], np.ndarray]] = None) -> None: """Set an attribute in the underlying environments. If id is an int, set the attribute denoted by key from the environment underlying the worker at index id to value. Otherwise, set the attribute for all workers at indices id. :param str key: The key of the desired attribute. :param Any value: The new value of the attribute. :param id: Indice(s) of the desired worker(s). Default to None for all env_id. """ self._assert_is_not_closed() id = self._wrap_id(id) if self.is_async: self._assert_id(id) for j in id: self.workers[j].set_env_attr(key, value) def _wrap_id( self, id: Optional[Union[int, List[int], np.ndarray]] = None ) -> Union[List[int], np.ndarray]: if id is None: return list(range(self.env_num)) return [id] if np.isscalar(id) else id # type: ignore def _assert_id(self, id: Union[List[int], np.ndarray]) -> None: for i in id: assert i not in self.waiting_id, \ f"Cannot interact with environment {i} which is stepping now." assert i in self.ready_id, \ f"Can only interact with ready environments {self.ready_id}." # TODO: compatible issue with reset -> (obs, info) def reset( self, id: Optional[Union[int, List[int], np.ndarray]] = None) -> np.ndarray: """Reset the state of some envs and return initial observations. If id is None, reset the state of all the environments and return initial observations, otherwise reset the specific environments with the given id, either an int or a list. """ self._assert_is_not_closed() id = self._wrap_id(id) if self.is_async: self._assert_id(id) # send(None) == reset() in worker for i in id: self.workers[i].send(None) obs_list = [self.workers[i].recv() for i in id] try: obs = np.stack(obs_list) except ValueError: # different len(obs) obs = np.array(obs_list, dtype=object) if self.obs_rms and self.update_obs_rms: self.obs_rms.update(obs) return self.normalize_obs(obs) def step( self, action: np.ndarray, id: Optional[Union[int, List[int], np.ndarray]] = None ) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]: """Run one timestep of some environments' dynamics. If id is None, run one timestep of all the environments’ dynamics; otherwise run one timestep for some environments with given id, either an int or a list. When the end of episode is reached, you are responsible for calling reset(id) to reset this environment’s state. Accept a batch of action and return a tuple (batch_obs, batch_rew, batch_done, batch_info) in numpy format. :param numpy.ndarray action: a batch of action provided by the agent. :return: A tuple including four items: * ``obs`` a numpy.ndarray, the agent's observation of current environments * ``rew`` a numpy.ndarray, the amount of rewards returned after \ previous actions * ``done`` a numpy.ndarray, whether these episodes have ended, in \ which case further step() calls will return undefined results * ``info`` a numpy.ndarray, contains auxiliary diagnostic \ information (helpful for debugging, and sometimes learning) For the async simulation: Provide the given action to the environments. The action sequence should correspond to the ``id`` argument, and the ``id`` argument should be a subset of the ``env_id`` in the last returned ``info`` (initially they are env_ids of all the environments). If action is None, fetch unfinished step() calls instead. """ self._assert_is_not_closed() id = self._wrap_id(id) if not self.is_async: assert len(action) == len(id) for i, j in enumerate(id): self.workers[j].send(action[i]) result = [] for j in id: obs, rew, done, info = self.workers[j].recv() info["env_id"] = j result.append((obs, rew, done, info)) else: if action is not None: self._assert_id(id) assert len(action) == len(id) for act, env_id in zip(action, id): self.workers[env_id].send(act) self.waiting_conn.append(self.workers[env_id]) self.waiting_id.append(env_id) self.ready_id = [x for x in self.ready_id if x not in id] ready_conns: List[EnvWorker] = [] while not ready_conns: ready_conns = self.worker_class.wait(self.waiting_conn, self.wait_num, self.timeout) result = [] for conn in ready_conns: waiting_index = self.waiting_conn.index(conn) self.waiting_conn.pop(waiting_index) env_id = self.waiting_id.pop(waiting_index) obs, rew, done, info = conn.recv() info["env_id"] = env_id result.append((obs, rew, done, info)) self.ready_id.append(env_id) obs_list, rew_list, done_list, info_list = zip(*result) try: obs_stack = np.stack(obs_list) except ValueError: # different len(obs) obs_stack = np.array(obs_list, dtype=object) rew_stack, done_stack, info_stack = map( np.stack, [rew_list, done_list, info_list]) if self.obs_rms and self.update_obs_rms: self.obs_rms.update(obs_stack) return self.normalize_obs(obs_stack), rew_stack, done_stack, info_stack def seed( self, seed: Optional[Union[int, List[int]]] = None) -> List[Optional[List[int]]]: """Set the seed for all environments. Accept ``None``, an int (which will extend ``i`` to ``[i, i + 1, i + 2, ...]``) or a list. :return: The list of seeds used in this env's random number generators. The first value in the list should be the "main" seed, or the value which a reproducer pass to "seed". """ self._assert_is_not_closed() seed_list: Union[List[None], List[int]] if seed is None: seed_list = [seed] * self.env_num elif isinstance(seed, int): seed_list = [seed + i for i in range(self.env_num)] else: seed_list = seed return [w.seed(s) for w, s in zip(self.workers, seed_list)] def render(self, **kwargs: Any) -> List[Any]: """Render all of the environments.""" self._assert_is_not_closed() if self.is_async and len(self.waiting_id) > 0: raise RuntimeError( f"Environments {self.waiting_id} are still stepping, cannot " "render them now.") return [w.render(**kwargs) for w in self.workers] def close(self) -> None: """Close all of the environments. This function will be called only once (if not, it will be called during garbage collected). This way, ``close`` of all workers can be assured. """ self._assert_is_not_closed() for w in self.workers: w.close() self.is_closed = True def normalize_obs(self, obs: np.ndarray) -> np.ndarray: """Normalize observations by statistics in obs_rms.""" if self.obs_rms and self.norm_obs: clip_max = 10.0 # this magic number is from openai baselines # see baselines/common/vec_env/vec_normalize.py#L10 obs = (obs - self.obs_rms.mean) / np.sqrt(self.obs_rms.var + self.__eps) obs = np.clip(obs, -clip_max, clip_max) return obs
class VectorEnvNormObs(VectorEnvWrapper): """An observation normalization wrapper for vectorized environments. :param bool update_obs_rms: whether to update obs_rms. Default to True. """ def __init__( self, venv: BaseVectorEnv, update_obs_rms: bool = True, ) -> None: super().__init__(venv) # initialize observation running mean/std self.update_obs_rms = update_obs_rms self.obs_rms = RunningMeanStd() def reset( self, id: Optional[Union[int, List[int], np.ndarray]] = None, **kwargs: Any, ) -> Union[np.ndarray, Tuple[np.ndarray, List[dict]]]: retval = self.venv.reset(id, **kwargs) reset_returns_info = isinstance( retval, (tuple, list)) and len(retval) == 2 and isinstance( retval[1], dict) if reset_returns_info: obs, info = retval else: obs = retval if isinstance(obs, tuple): raise TypeError( "Tuple observation space is not supported. ", "Please change it to array or dict space", ) if self.obs_rms and self.update_obs_rms: self.obs_rms.update(obs) obs = self._norm_obs(obs) if reset_returns_info: return obs, info else: return obs def step( self, action: np.ndarray, id: Optional[Union[int, List[int], np.ndarray]] = None, ) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]: obs, rew, done, info = self.venv.step(action, id) if self.obs_rms and self.update_obs_rms: self.obs_rms.update(obs) return self._norm_obs(obs), rew, done, info def _norm_obs(self, obs: np.ndarray) -> np.ndarray: if self.obs_rms: return self.obs_rms.norm(obs) # type: ignore return obs def set_obs_rms(self, obs_rms: RunningMeanStd) -> None: """Set with given observation running mean/std.""" self.obs_rms = obs_rms def get_obs_rms(self) -> RunningMeanStd: """Return observation running mean/std.""" return self.obs_rms