Struct BcTrainer 

pub struct BcTrainer<B, P, O>
where
    B: AutodiffBackend,
    P: AutodiffModule<B>,
    O: Optimizer<P, B>,
{ /* private fields */ }

Burn-backend Behavioral Cloning trainer.

Generic over:

• B: AutodiffBackend — the Burn backend (e.g. Autodiff<NdArray<f32>>).
• P: AutodiffModule<B> — the policy module (only its logits head is trained; the value head is ignored).
• O: Optimizer<P, B> — the Burn optimizer (typically AdamConfig::new().init()).

The policy is held in Option<P> because Burn’s Optimizer::step consumes the module by value; each minibatch .take()s it and puts back the updated copy.

Implementations

impl<B, P, O> BcTrainer<B, P, O>

where B: AutodiffBackend, P: AutodiffModule<B> + Clone, O: Optimizer<P, B>,

pub fn new( config: BcConfig, policy: P, optimizer: BurnOptimizer<B, P, O>, ) -> Result<Self>

Build a new Burn behavioral-cloning trainer.

Validates the config and seeds the trainer’s minibatch-shuffle RNG from BcConfig::seed.

pub fn config(&self) -> &BcConfig

Borrow the configuration.

pub fn policy(&self) -> &P

Borrow the policy. Panics if the trainer is mid-step (the policy has been moved into the optimizer); only safe to call between train_epoch invocations.

pub fn total_steps(&self) -> usize

Total completed gradient updates (one per minibatch across all epochs).

pub fn total_epochs(&self) -> usize

Total completed epochs.

pub fn train_epoch<F>( &mut self, demos: &Demonstrations, forward_fn: F, ) -> Result<BcEpochStats>

where F: FnMut(&P, Tensor<B, 2>) -> Tensor<B, 2>,

Train for one full supervised pass over the demonstration dataset.

1. Draw a seeded shuffle of all example indices, partitioned into minibatches of BcConfig::batch_size (ceil(len / batch_size) minibatches).
2. For each minibatch: gather (obs, actions), produce logits via forward_fn, compute the cross-entropy compute_bc_loss, backprop, and step the optimizer once.
3. Aggregate the example-weighted mean loss and action-match accuracy over the epoch.

forward_fn lets the caller pick how logits are produced from the policy — e.g. |p, o| p.forward(o).0 to drop MlpBurnPolicy’s value head.

Returns finite loss and accuracy in [0, 1]. Returns an Err if the dataset is empty.