BatchSizeTransform

class torchrl.envs.transforms.BatchSizeTransform(*, batch_size: torch.Size | None = None, reshape_fn: Callable[[TensorDictBase], TensorDictBase] | None = None, reset_func: Callable[[TensorDictBase, TensorDictBase], TensorDictBase] | None =None, env_kwarg: bool =False)

一個用於修改環境批大小的 transform。

此 transform 有兩種不同的用法：它可以用於為非批鎖定的（例如無狀態的）環境設定批大小，以便使用資料收集器進行資料收集。它也可以用於修改環境的批大小（例如，squeeze、unsqueeze 或 reshape 操作）。

此 transform 修改環境批大小以匹配提供的批大小。它要求父環境的批大小可以擴充套件到提供的批大小。

關鍵字引數：

• batch_size (torch.Size 或 等效型別, 可選) – 環境的新批大小。與 reshape_fn 互斥。

• reshape_fn (可呼叫物件, 可選) –

  一個用於修改環境批大小的可呼叫物件。與 batch_size 互斥。

  注意

  目前，支援涉及 reshape、flatten、unflatten、squeeze 和 unsqueeze 的轉換。如果需要其他 reshape 操作，請在 TorchRL github 上提交功能請求。

• reset_func (可呼叫物件, 可選) – 一個生成 reset tensordict 的函式。簽名必須匹配 Callable[[TensorDictBase, TensorDictBase], TensorDictBase]，其中第一個輸入引數是在呼叫 reset() 時傳遞給環境的可選 tensordict，第二個引數是 TransformedEnv.base_env.reset 的輸出。如果 env_kwarg=True，它也可以支援一個可選的 env 關鍵字引數。

• env_kwarg (布林值, 可選) – 如果為 True，則 reset_func 必須支援一個 env 關鍵字引數。預設為 False。傳遞的環境將是帶有其 transform 的環境。

示例

>>> # Changing the batch-size with a function
>>> from torchrl.envs import GymEnv
>>> base_env = GymEnv("CartPole-v1")
>>> env = TransformedEnv(base_env, BatchSizeTransform(reshape_fn=lambda data: data.reshape(1, 1)))
>>> env.rollout(4)
>>> # Setting the shape of a stateless environment
>>> class MyEnv(EnvBase):
...     batch_locked = False
...     def __init__(self):
...         super().__init__()
...         self.observation_spec = Composite(observation=Unbounded(3))
...         self.reward_spec = Unbounded(1)
...         self.action_spec = Unbounded(1)
...
...     def _reset(self, tensordict: TensorDictBase, **kwargs) -> TensorDictBase:
...         tensordict_batch_size = tensordict.batch_size if tensordict is not None else torch.Size([])
...         result = self.observation_spec.rand(tensordict_batch_size)
...         result.update(self.full_done_spec.zero(tensordict_batch_size))
...         return result
...
...     def _step(
...         self,
...         tensordict: TensorDictBase,
...     ) -> TensorDictBase:
...         result = self.observation_spec.rand(tensordict.batch_size)
...         result.update(self.full_done_spec.zero(tensordict.batch_size))
...         result.update(self.full_reward_spec.zero(tensordict.batch_size))
...         return result
...
...     def _set_seed(self, seed: Optional[int]):
...         pass
...
>>> env = TransformedEnv(MyEnv(), BatchSizeTransform([5]))
>>> assert env.batch_size == torch.Size([5])
>>> assert env.rollout(10).shape == torch.Size([5, 10])

reset_func 可以建立一個具有所需批大小的 tensordict，從而實現細粒度的 reset 呼叫

>>> def reset_func(tensordict, tensordict_reset, env):
...     result = env.observation_spec.rand()
...     result.update(env.full_done_spec.zero())
...     assert result.batch_size != torch.Size([])
...     return result
>>> env = TransformedEnv(MyEnv(), BatchSizeTransform([5], reset_func=reset_func, env_kwarg=True))
>>> print(env.rollout(2))
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.float32, is_shared=False),
        done: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([5, 2, 3]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                terminated: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([5, 2]),
            device=None,
            is_shared=False),
        observation: Tensor(shape=torch.Size([5, 2, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([5, 2]),
    device=None,
    is_shared=False)

此 transform 可用於在資料收集器內部部署非批鎖定的環境

>>> from torchrl.collectors import SyncDataCollector
>>> collector = SyncDataCollector(env, lambda td: env.rand_action(td), frames_per_batch=10, total_frames=-1)
>>> for data in collector:
...     print(data)
...     break
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.float32, is_shared=False),
        collector: TensorDict(
            fields={
                traj_ids: Tensor(shape=torch.Size([5, 2]), device=cpu, dtype=torch.int64, is_shared=False)},
            batch_size=torch.Size([5, 2]),
            device=None,
            is_shared=False),
        done: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
        next: TensorDict(
            fields={
                done: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                observation: Tensor(shape=torch.Size([5, 2, 3]), device=cpu, dtype=torch.float32, is_shared=False),
                reward: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.float32, is_shared=False),
                terminated: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([5, 2]),
            device=None,
            is_shared=False),
        observation: Tensor(shape=torch.Size([5, 2, 3]), device=cpu, dtype=torch.float32, is_shared=False),
        terminated: Tensor(shape=torch.Size([5, 2, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
    batch_size=torch.Size([5, 2]),
    device=None,
    is_shared=False)
>>> collector.shutdown()

forward(tensordict: TensorDictBase) → TensorDictBase

讀取輸入的 tensordict，並對選定的鍵應用 transform。

transform_env_batch_size(batch_size: Size)

轉換父環境的批大小。

transform_input_spec(input_spec: Composite) → Composite

轉換輸入 spec，使結果 spec 與 transform 對映匹配。

引數：

input_spec (TensorSpec) – transform 之前的 spec

返回：

transform 之後的預期 spec

transform_output_spec(output_spec: Composite) → Composite

轉換輸出 spec，使結果 spec 與 transform 對映匹配。

此方法通常不應修改。更改應使用 transform_observation_spec()、transform_reward_spec() 和 transform_full_done_spec() 實現。 :param output_spec: transform 之前的 spec :type output_spec: TensorSpec

返回：

transform 之後的預期 spec