Struct border::agent::tch::sac::base::SAC

pub struct SAC<E, Q, P, O, A> where
    E: Env,
    O: TchBuffer<Item = E::Obs>,
    A: TchBuffer<Item = E::Act>,  { /* fields omitted */ }

Soft actor critic agent.

Trait Implementations

impl<E, Q, P, O, A> Agent<E> for SAC<E, Q, P, O, A> where
    E: Env,
    Q: Model2<Input1 = O::SubBatch, Input2 = A::SubBatch, Output = Tensor> + Clone,
    P: Model1<Input = Tensor, Output = (Tensor, Tensor)> + Clone,
    E::Obs: Into<O::SubBatch>,
    E::Act: From<Tensor>,
    O: TchBuffer<Item = E::Obs, SubBatch = P::Input>,
    A: TchBuffer<Item = E::Act, SubBatch = 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 push_obs(&self, obs: &E::Obs)

Push observation to the agent. This method is used when resetting the environment.

fn observe(&mut self, step: Step<E>) -> Option<Record>

Update model parameters.

When the return value is Some(Record), it includes:

• loss_critic: Loss of critic
• loss_actor: Loss of actor

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

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

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

Load the agent from the given directory.

impl<E, Q, P, O, A> Policy<E> for SAC<E, Q, P, O, A> where
    E: Env,
    Q: Model2<Input1 = O::SubBatch, Input2 = A::SubBatch, Output = Tensor> + Clone,
    P: Model1<Input = Tensor, Output = (Tensor, Tensor)> + Clone,
    E::Obs: Into<O::SubBatch>,
    E::Act: From<Tensor>,
    O: TchBuffer<Item = E::Obs, SubBatch = P::Input>,
    A: TchBuffer<Item = E::Act, SubBatch = Tensor>, 

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

Sample an action given an observation.