DTActor

class torchrl.modules.DTActor(state_dim: int, action_dim: int, transformer_config: Dict | DecisionTransformer.DTConfig = None, device: DEVICE_TYPING | None = None)

Decision Transformer Actor class.

Actor class for the Decision Transformer to output deterministic action as presented in “Decision Transformer” <https://arxiv.org/abs/2202.05607.pdf>. Returns the deterministic actions.

Parameters:

• state_dim (int) – state dimension.

• action_dim (int) – action dimension.

• transformer_config (Dict or DecisionTransformer.DTConfig, optional) – config for the GPT2 transformer. Defaults to default_config().

• device (torch.device, optional) – device to use. Defaults to None.

Examples

>>> model = DTActor(state_dim=4, action_dim=2,
...     transformer_config=DTActor.default_config())
>>> observation = torch.randn(32, 10, 4)
>>> action = torch.randn(32, 10, 2)
>>> return_to_go = torch.randn(32, 10, 1)
>>> output = model(observation, action, return_to_go)
>>> output.shape
torch.Size([32, 10, 2])

classmethod default_config()

Default configuration for DTActor.

forward(observation: Tensor, action: Tensor, return_to_go: Tensor) → Tensor

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.