ActorValueOperator

class torchrl.modules.tensordict_module.ActorValueOperator(*args, **kwargs)

Actor-value 運算子。

此類別將一個 actor 和一個 value 模型包裝在一起，它們共享一個共同的觀測嵌入網路。

注意

對於一個返回動作和品質值 \(Q(s, a)\) 的類似類別，請參閱 ActorCriticOperator。對於一個沒有共同嵌入的版本，請參閱 ActorCriticWrapper。

為了方便工作流程，此類別配備了 get_policy_operator() 和 get_value_operator() 方法，它們都將返回一個具有專用功能的獨立 TDModule。

參數:

• common_operator (TensorDictModule) – 一個讀取觀測值並產生隱藏變數的通用運算子

• policy_operator (TensorDictModule) – 一個讀取隱藏變數並返回動作的策略運算子

• value_operator (TensorDictModule) – 一個讀取隱藏變數並返回值的 value 運算子

範例

>>> import torch
>>> from tensordict import TensorDict
>>> from torchrl.modules import ProbabilisticActor, SafeModule
>>> from torchrl.modules import ValueOperator, TanhNormal, ActorValueOperator, NormalParamExtractor
>>> module_hidden = torch.nn.Linear(4, 4)
>>> td_module_hidden = SafeModule(
...    module=module_hidden,
...    in_keys=["observation"],
...    out_keys=["hidden"],
...    )
>>> module_action = TensorDictModule(
...     nn.Sequential(torch.nn.Linear(4, 8), NormalParamExtractor()),
...     in_keys=["hidden"],
...     out_keys=["loc", "scale"],
...     )
>>> td_module_action = ProbabilisticActor(
...    module=module_action,
...    in_keys=["loc", "scale"],
...    out_keys=["action"],
...    distribution_class=TanhNormal,
...    return_log_prob=True,
...    )
>>> module_value = torch.nn.Linear(4, 1)
>>> td_module_value = ValueOperator(
...    module=module_value,
...    in_keys=["hidden"],
...    )
>>> td_module = ActorValueOperator(td_module_hidden, td_module_action, td_module_value)
>>> td = TensorDict({"observation": torch.randn(3, 4)}, [3,])
>>> td_clone = td_module(td.clone())
>>> print(td_clone)
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        hidden: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        loc: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        sample_log_prob: Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, is_shared=False),
        scale: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        state_value: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([3]),
    device=None,
    is_shared=False)
>>> td_clone = td_module.get_policy_operator()(td.clone())
>>> print(td_clone)  # no value
TensorDict(
    fields={
        action: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        hidden: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        loc: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        sample_log_prob: Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, is_shared=False),
        scale: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([3]),
    device=None,
    is_shared=False)
>>> td_clone = td_module.get_value_operator()(td.clone())
>>> print(td_clone)  # no action
TensorDict(
    fields={
        hidden: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        observation: Tensor(shape=torch.Size([3, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        state_value: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([3]),
    device=None,
    is_shared=False)

get_policy_head() → SafeSequential

返回策略 head。

get_policy_operator() → SafeSequential

返回一個獨立的策略運算子，它將觀測值對應到一個動作。

get_value_head() → SafeSequential

傳回 value head。

get_value_operator() → SafeSequential

傳回一個獨立的 value network operator，它將觀察值映射到價值估算。