module emote.nn.gaussian_policy
Classes
class BasePolicy(nn.Module):
Methods
def __init__(self) -> None
def post_process(self, actions) -> None
Post-process a pre-action into a post-action.
Arguments:
actions
def infer(self, x) -> None
Samples pre-actions and associated post-actions (actual decisions) from the policy given the encoder input.
Only for use at inference time; defaults to identity transformation. Crucial to reimplement for discrete reparametrized policies.
Arguments:
x(Tensor)
class GaussianPolicyHead(nn.Module):
Methods
def __init__(self, hidden_dim, action_dim) -> None
def forward(self, x, epsilon) -> Tensor | Tuple[Tensor]
Sample pre-actions and associated log-probabilities.
Arguments:
x(Tensor)epsilon(Tensor | None)
Returns:
- Direct samples (pre-actions) from the policy log- probabilities associated to those samples
class GaussianMlpPolicy(nn.Module):
Methods
def __init__(self, observation_dim, action_dim, hidden_dims) -> None
def forward(self, obs, epsilon) -> Tensor | Tuple[Tensor]