ModelBasedEnvBase

torchrl.envs.ModelBasedEnvBase(*args, **kwargs)

MBRL sota-implementations 的基础环境。

MBRL 算法模型封装。它旨在为世界模型（包括但不限于观察、奖励、完成状态和安全约束模型）提供一个环境框架，并像经典环境一样运行。

这是一个基础类，不应直接使用。

示例

>>> import torch
>>> from tensordict import TensorDict
>>> from torchrl.data import Composite, Unbounded
>>> class MyMBEnv(ModelBasedEnvBase):
...     def __init__(self, world_model, device="cpu", dtype=None, batch_size=None):
...         super().__init__(world_model, device=device, dtype=dtype, batch_size=batch_size)
...         self.observation_spec = Composite(
...             hidden_observation=Unbounded((4,))
...         )
...         self.state_spec = Composite(
...             hidden_observation=Unbounded((4,)),
...         )
...         self.action_spec = Unbounded((1,))
...         self.reward_spec = Unbounded((1,))
...
...     def _reset(self, tensordict: TensorDict) -> TensorDict:
...         tensordict = TensorDict(
...             batch_size=self.batch_size,
...             device=self.device,
...         )
...         tensordict = tensordict.update(self.state_spec.rand())
...         tensordict = tensordict.update(self.observation_spec.rand())
...         return tensordict
>>> # This environment is used as follows:
>>> import torch.nn as nn
>>> from torchrl.modules import MLP, WorldModelWrapper
>>> world_model = WorldModelWrapper(
...     TensorDictModule(
...         MLP(out_features=4, activation_class=nn.ReLU, activate_last_layer=True, depth=0),
...         in_keys=["hidden_observation", "action"],
...         out_keys=["hidden_observation"],
...     ),
...     TensorDictModule(
...         nn.Linear(4, 1),
...         in_keys=["hidden_observation"],
...         out_keys=["reward"],
...     ),
... )
>>> env = MyMBEnv(world_model)
>>> tensordict = env.rollout(max_steps=10)
>>> print(tensordict)
TensorDict(
    fields={
        action: Tensor(torch.Size([10, 1]), dtype=torch.float32),
        done: Tensor(torch.Size([10, 1]), dtype=torch.bool),
        hidden_observation: Tensor(torch.Size([10, 4]), dtype=torch.float32),
        next: LazyStackedTensorDict(
            fields={
                hidden_observation: Tensor(torch.Size([10, 4]), dtype=torch.float32)},
            batch_size=torch.Size([10]),
            device=cpu,
            is_shared=False),
        reward: Tensor(torch.Size([10, 1]), dtype=torch.float32)},
    batch_size=torch.Size([10]),
    device=cpu,
    is_shared=False)

属性

observation_spec (Composite): 观测的采样规范；action_spec (TensorSpec): 操作的采样规范；reward_spec (TensorSpec): 奖励的采样规范；input_spec (Composite):输入的采样规范；batch_size (torch.Size): 环境使用的批次大小。如果未设置，环境将接受所有批次大小的 tensordicts。device (torch.device): 环境输入和输出的预期所在设备

参数:

• world_model (nn.Module) – 生成世界状态及其相应奖励的模型；

• params (List[torch.Tensor], optional) – 世界模型的参数列表；

• buffers (List[torch.Tensor], optional) – 世界模型的缓冲区列表；

• device (torch.device, optional) – 环境输入和输出的预期所在设备

• dtype (torch.dtype, optional) – 环境输入和输出的数据类型

• batch_size (torch.Size, optional) – 实例中包含的环境数量

• run_type_check (bool, optional) – 是否在环境的步骤中运行类型检查

torchrl.envs.step(TensorDict -> TensorDict)

环境中的一步

torchrl.envs.reset(TensorDict, optional -> TensorDict)

重置环境

torchrl.envs.set_seed(int -> int)

设置环境的种子

torchrl.envs.rand_step(TensorDict, optional -> TensorDict)

根据动作规范进行随机步进

torchrl.envs.rollout(Callable, ... -> TensorDict)

使用给定的策略（或在未提供策略时随机步进）在环境中执行 rollout