Struct AdaptationNetwork 

pub struct AdaptationNetwork<T: Float + Debug + ScalarOperand + Send + Sync + 'static> { /* private fields */ }

Task adaptation network

Implementations

impl<T: Float + Debug + ScalarOperand + Send + Sync + 'static> AdaptationNetwork<T>

pub fn new(model_dimension: usize, hidden_dimension: usize) -> Result<Self>

pub fn parameter_count(&self) -> usize

Encode a task descriptor into a context embedding.

The raw descriptor is three scalars (difficulty, complexity, data_characteristics). They are lifted to model_dimension features by a fixed random-Fourier-style basis: feature k is cos(2π·(k+1)·d_{k mod 3} / 3 + φ_k) with a deterministic per-index phase, so distinct descriptors map to distinct, bounded, non-collinear embeddings. Previously this ignored the task entirely and fed Array2::ones, which made every task produce the same context. Total learnable parameters across the three sub-networks.

pub fn encode_task_context(&mut self, task: &TaskBatch<T>) -> Result<Array2<T>>

pub fn predict_parameters(&mut self, context: &Array2<T>) -> Result<Vec<T>>

pub fn generate_parameter_update( &mut self, transformer_output: &Array2<T>, current_parameters: &Array1<T>, ) -> Result<Array1<T>>