Struct DartsSearch 

pub struct DartsSearch {
    pub cells: Vec<DartsCell>,
    pub config: DartsConfig,
    /* private fields */
}

Top-level DARTS search controller.

Manages a stack of DartsCells and implements the bi-level optimisation loop.

Fields

cells: Vec<DartsCell>

Stack of cells forming the super-network.

config: DartsConfig

Configuration.

Implementations

impl DartsSearch

pub fn new(config: DartsConfig) -> Self

Construct a DartsSearch from the given config.

pub fn arch_parameters(&self) -> Vec<f64>

Return all architecture parameters across all cells, flattened.

pub fn n_arch_params(&self) -> usize

Total number of architecture parameters.

pub fn update_arch_params( &mut self, grads: &[f64], lr: f64, ) -> OptimizeResult<()>

Apply a gradient step to architecture parameters.

grads must be the same length as arch_parameters().

pub fn derive_discrete_arch_indices(&self) -> Vec<Vec<Vec<usize>>>

Derive the discrete architecture: for each cell, argmax op per edge.

Returns Vec<Vec<Vec<usize>>> — [cell][intermediate_node][predecessor].

pub fn derive_discrete_arch(&self) -> Vec<Vec<Operation>>

Derive the discrete architecture as a vec of Operation vectors.

Uses Operation::all() to map operation index to Operation. If n_operations exceeds the canonical list length, Operation::Identity is used as a fallback.

pub fn bilevel_step( &mut self, train_x: &[Vec<f64>], train_y: &[f64], val_x: &[Vec<f64>], val_y: &[f64], ) -> (f64, f64)

One bilevel optimisation step (approximate first-order DARTS).

Inner step: update network weights on train_x/train_y. Outer step: update architecture params on val_x/val_y.

Returns (train_loss, val_loss) before the update.

Trait Implementations

impl Clone for DartsSearch

fn clone(&self) -> DartsSearch

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for DartsSearch

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.