ActionMask

class torchrl.envs.transforms.ActionMask(action_key: NestedKey = 'action', mask_key: NestedKey = 'action_mask')

一種自適應動作遮罩器。

此轉換在步執行後從輸入 tensordict 讀取遮罩，並調整 one-hot / 分類動作規範的遮罩。

注意

在沒有環境的情況下使用此轉換將失敗。

引數：

• action_key (NestedKey, optional) – 動作張量所在的鍵。預設為 "action"。

• mask_key (NestedKey, optional) – 動作遮罩所在的鍵。預設為 "action_mask"。

示例

>>> import torch
>>> from torchrl.data.tensor_specs import Categorical, Binary, Unbounded, Composite
>>> from torchrl.envs.transforms import ActionMask, TransformedEnv
>>> from torchrl.envs.common import EnvBase
>>> class MaskedEnv(EnvBase):
...     def __init__(self, *args, **kwargs):
...         super().__init__(*args, **kwargs)
...         self.action_spec = Categorical(4)
...         self.state_spec = Composite(action_mask=Binary(4, dtype=torch.bool))
...         self.observation_spec = Composite(obs=Unbounded(3))
...         self.reward_spec = Unbounded(1)
...
...     def _reset(self, tensordict=None):
...         td = self.observation_spec.rand()
...         td.update(torch.ones_like(self.state_spec.rand()))
...         return td
...
...     def _step(self, data):
...         td = self.observation_spec.rand()
...         mask = data.get("action_mask")
...         action = data.get("action")
...         mask = mask.scatter(-1, action.unsqueeze(-1), 0)
...
...         td.set("action_mask", mask)
...         td.set("reward", self.reward_spec.rand())
...         td.set("done", ~mask.any().view(1))
...         return td
...
...     def _set_seed(self, seed):
...         return seed
...
>>> torch.manual_seed(0)
>>> base_env = MaskedEnv()
>>> env = TransformedEnv(base_env, ActionMask())
>>> r = env.rollout(10)
>>> env = TransformedEnv(base_env, ActionMask())
>>> r = env.rollout(10)
>>> r["action_mask"]
tensor([[ True,  True,  True,  True],
        [ True,  True, False,  True],
        [ True,  True, False, False],
        [ True, False, False, False]])

forward(tensordict: TensorDictBase) → TensorDictBase

讀取輸入的 tensordict，並對選定的鍵應用轉換。