Struct ModuleEvalFn 

pub struct ModuleEvalFn<B: Backend, R: ParamReshaper<B>, F> { /* private fields */ }

Bridges a flat population tensor to per-member module scoring.

Implements BatchFitnessFn<B, Tensor<B, 2>>: each row of the (pop_size, num_params) population is unflattened into a module via the ParamReshaper, passed to a host-side scorer, and the resulting scalar is collected into the fitness tensor in the scorer’s natural value space. Direction is declared once via ObjectiveSense (default ObjectiveSense::Maximize; pass with_sense for a cost scorer like MSE) and reconciled by the harness — no hand-negation in the scorer.

Gradient isolation

B: Backend, not AutodiffBackend. The reconstructed modules carry no gradient tracking — scorer performs forward-only work (loss, rollout return, …).

Type Parameters

• R: a ParamReshaper producing R::Module.
• F: a host-side Fn(&R::Module) -> f32 scorer (MSE, accuracy, negative episode return, …).

Device convention

Population rows are sliced on the supplied device; the reshaper splats them into modules on that same device. Callers must construct the population and the reshaper’s template on one device.

Implementations

impl<B, R, F> ModuleEvalFn<B, R, F>

where B: Backend, R: ParamReshaper<B>, F: Fn(&R::Module) -> f32 + Send,

pub fn new(reshaper: R, scorer: F) -> Self

Build an evaluator from a reshaper and a per-module scorer, defaulting the objective sense to ObjectiveSense::Maximize (accuracy, reward, episode return). Use with_sense for a cost scorer such as MSE.

pub fn with_sense(reshaper: R, scorer: F, sense: ObjectiveSense) -> Self

Build an evaluator with an explicit ObjectiveSense.

pub fn reshaper(&self) -> &R

Borrow the underlying reshaper.

Trait Implementations

impl<B, R, F> BatchFitnessFn<B, Tensor<B, 2>> for ModuleEvalFn<B, R, F>

where B: Backend, R: ParamReshaper<B>, F: Fn(&R::Module) -> f32 + Send,

fn evaluate_batch( &mut self, population: &Tensor<B, 2>, device: &B::Device, ) -> Tensor<B, 1>

Score every row of population by reconstructing one module per row.

Panics

Panics at batch entry if the population width (population.dims()[1]) differs from the reshaper’s num_params, failing fast before any per-row work rather than mid-loop inside ParamReshaper::unflatten.

fn sense(&self) -> ObjectiveSense

The optimisation direction of this objective.

impl<B: Backend, R: ParamReshaper<B>, F> Debug for ModuleEvalFn<B, R, F>

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

Formats the value using the given formatter.