Struct SacTrainer 

pub struct SacTrainer<B: AutodiffBackend> { /* private fields */ }

Burn-backend Soft Actor-Critic trainer.

Generic only over the autodiff backend B; the network shapes are fixed (a SacActor plus four ContinuousQNetwork critics) so the trainer can own the three concrete Adam optimizers directly.

Implementations

impl<B: AutodiffBackend> SacTrainer<B>

pub fn new( config: SacConfig, obs_dim: usize, action_dim: usize, device: B::Device, ) -> Result<Self>

Build a new SAC trainer for an obs_dim-dimensional observation and action_dim-dimensional continuous action space.

All five networks are seeded off SacConfig::seed (the actor, each critic, and the targets get distinct derived seeds so the twin critics are decorrelated), and the targets are hard-copied from their online counterparts so they start byte-equal.

pub fn config(&self) -> &SacConfig

Borrow the configuration.

pub fn obs_dim(&self) -> usize

Observation dimension.

pub fn action_dim(&self) -> usize

Action dimension.

pub fn target_entropy(&self) -> f32

Effective target entropy (after the -action_dim heuristic).

pub fn actor(&self) -> &SacActor<B>

Borrow the actor. Panics if called mid-step.

pub fn q1(&self) -> &ContinuousQNetwork<B>

Borrow the first online critic. Panics if called mid-step.

pub fn q2(&self) -> &ContinuousQNetwork<B>

Borrow the second online critic. Panics if called mid-step.

pub fn buffer(&self) -> &ContinuousReplayBuffer

Borrow the replay buffer.

pub fn buffer_mut(&mut self) -> &mut ContinuousReplayBuffer

Mutably borrow the replay buffer.

pub fn buffer_len(&self) -> usize

Number of transitions currently in the buffer.

pub fn alpha(&self) -> f64

Current entropy temperature alpha = exp(log_alpha).

pub fn total_env_steps(&self) -> usize

Current env-step counter.

pub fn total_train_steps(&self) -> usize

Number of completed gradient updates.

pub fn total_episodes(&self) -> usize

Number of completed episodes.

pub fn increment_env_step(&mut self)

Caller invokes this once per environment step.

pub fn increment_episodes(&mut self, n: usize)

Caller invokes this when an episode terminates / truncates.

pub fn in_warmup(&self) -> bool

true while the trainer is still in the random-action warmup window (env steps < learning_starts).

pub fn select_action(&mut self, obs: &[f32]) -> Vec<f32>

Select an action for obs (shape [obs_dim]).

During the warmup window (Self::in_warmup) the action is drawn uniformly in (-1, 1) per dimension; afterwards it is a stochastic sample from the actor. The returned action is in the tanh-squashed (-1, 1) box — the env is responsible for rescaling to its own action range.

pub fn eval_action(&self, obs: &[f32]) -> Vec<f32>

Deterministic evaluation action tanh(mu) for obs (shape [obs_dim]), bypassing exploration noise.

pub fn train(&mut self) -> Result<Option<SacStepStats>>

Run gradient_steps_per_env_step SAC gradient updates if the buffer is ready.

Returns Ok(None) until the buffer holds min_buffer_size transitions; afterwards returns the stats from the last of the performed updates.

pub fn train_step(&mut self) -> Result<SacStepStats>

Run exactly one SAC gradient update (critics, actor, optional alpha, then a Polyak soft update of both targets).

Returns an error if any loss is non-finite. Assumes the buffer holds at least one transition; callers typically gate this on ContinuousReplayBuffer::is_ready via Self::train.