Struct border_tch_agent::sac::SAC

pub struct SAC<E, Q, P, R> where
    E: Env,
    Q: SubModel2<Output = Tensor>,
    P: SubModel<Output = (Tensor, Tensor)>,
    R: ReplayBufferBase,
    E::Obs: Into<Q::Input1> + Into<P::Input>,
    E::Act: Into<Q::Input2>,
    Q::Input2: From<Tensor>,
    Q::Config: DeserializeOwned + Serialize + OutDim + Debug + PartialEq + Clone,
    P::Config: DeserializeOwned + Serialize + OutDim + Debug + PartialEq + Clone,
    <R::Batch as Batch>::ObsBatch: Into<Q::Input1> + Into<P::Input> + Clone,
    <R::Batch as Batch>::ActBatch: Into<Q::Input2> + Into<Tensor>,  { /* private fields */ }

Soft actor critic (SAC) agent.

Trait Implementations

impl<E, Q, P, R> Agent<E, R> for SAC<E, Q, P, R> where
    E: Env,
    Q: SubModel2<Output = Tensor>,
    P: SubModel<Output = (Tensor, Tensor)>,
    R: ReplayBufferBase,
    E::Obs: Into<Q::Input1> + Into<P::Input>,
    E::Act: Into<Q::Input2> + From<Tensor>,
    Q::Input2: From<Tensor>,
    Q::Config: DeserializeOwned + Serialize + OutDim + Debug + PartialEq + Clone,
    P::Config: DeserializeOwned + Serialize + OutDim + Debug + PartialEq + Clone,
    <R::Batch as Batch>::ObsBatch: Into<Q::Input1> + Into<P::Input> + Clone,
    <R::Batch as Batch>::ActBatch: Into<Q::Input2> + Into<Tensor>, 

fn train(&mut self)

Set the policy to training mode.

fn eval(&mut self)

Set the policy to evaluation mode.

fn is_train(&self) -> bool

Return if it is in training mode.

fn opt(&mut self, buffer: &mut R) -> Option<Record>

Do an optimization step.

fn save<T: AsRef<Path>>(&self, path: T) -> Result<()>

Save the agent in the given directory. This method commonly creates a number of files consisting the agent in the directory. For example, the DQN agent in border_tch_agent crate saves two Q-networks corresponding to the original and target networks.

fn load<T: AsRef<Path>>(&mut self, path: T) -> Result<()>

Load the agent from the given directory.

impl<E, Q, P, R> Policy<E> for SAC<E, Q, P, R> where
    E: Env,
    Q: SubModel2<Output = Tensor>,
    P: SubModel<Output = (Tensor, Tensor)>,
    R: ReplayBufferBase,
    E::Obs: Into<Q::Input1> + Into<P::Input>,
    E::Act: Into<Q::Input2> + From<Tensor>,
    Q::Input2: From<Tensor>,
    Q::Config: DeserializeOwned + Serialize + OutDim + Debug + PartialEq + Clone,
    P::Config: DeserializeOwned + Serialize + OutDim + Debug + PartialEq + Clone,
    <R::Batch as Batch>::ObsBatch: Into<Q::Input1> + Into<P::Input> + Clone,
    <R::Batch as Batch>::ActBatch: Into<Q::Input2> + Into<Tensor>, 

fn build(config: Self::Config) -> Self

Constructs SAC agent.

type Config = SACConfig<Q, P>

Configuration of the policy.

fn sample(&mut self, obs: &E::Obs) -> E::Act

Sample an action given an observation.