Struct MAMLLearner 

pub struct MAMLLearner<T: Float + Debug + Send + Sync + 'static, D: Dimension> { /* private fields */ }

MAML implementation over the shared linear task model (super::linear_model).

The inner loop descends the support-set loss of that model with analytic gradients; the outer loop applies FOMAML by default and the exact second-order MAML meta-gradient when config.second_order is set (exact because the model is linear in its parameters, so its MSE Hessian is constant and the Hessian-vector product is closed-form).

Implementations

impl<T: Float + Default + Clone + Send + Sync + ScalarOperand + Debug, D: Dimension> MAMLLearner<T, D>

pub fn new(config: MAMLConfig<T>) -> Result<Self>

pub fn adaptations_recorded(&self) -> usize

Number of task adaptations recorded so far.

Trait Implementations

impl<T: Float + Debug + 'static + Default + Clone + Send + Sync + Sum + ScalarOperand, D: Dimension> MetaLearner<T> for MAMLLearner<T, D>

fn meta_train_step( &mut self, task_batch: &[MetaTask<T>], meta_parameters: &mut HashMap<String, Array1<T>>, ) -> Result<MetaTrainingResult<T>>

One meta-training step over a batch of tasks.

The meta-update is applied here (see the module docs): meta_parameters comes back already advanced by -outer_lr * meta_gradient.

fn adapt_to_task( &mut self, task: &MetaTask<T>, meta_parameters: &HashMap<String, Array1<T>>, adaptation_steps: usize, ) -> Result<TaskAdaptationResult<T>>

Inner loop: descend the support-set loss from the meta-parameters.

fn evaluate_query_set( &self, task: &MetaTask<T>, adapted_parameters: &HashMap<String, Array1<T>>, ) -> Result<QueryEvaluationResult<T>>

Evaluate the query set at the adapted parameters.

fn get_algorithm(&self) -> MetaLearningAlgorithm

Get meta-learner type