Struct ModuleReshaper 

pub struct ModuleReshaper<B: Backend, M: Module<B>> { /* private fields */ }

A ParamReshaper backed by a cloned template module.

Construction clones the supplied module once and counts its float leaves. Each unflatten call clones that template and maps the flat buffer into the clone’s leaves; each flatten call visits a module’s leaves and concatenates them.

Single-source width convention

A reshaper is the genome-width source of truth. Build one reshaper for the width, then hand reshaper.clone() (this type is Clone) to the fitness adapter — the strategy and its evaluator then agree on num_params by construction. Prefer this over building two ModuleReshaper::new(template.clone()) instances whose widths match only by convention: a silent divergence surfaces late as the documented width-mismatch panic in unflatten / evaluate_batch. Where a WeightOnly wrapper already owns a reshaper, clone its via strategy.reshaper().clone().

Example

See the module_eval_fn tests for a runnable end-to-end example of flatten/unflatten in the weight-only pipeline.

Implementations

impl<B: Backend, M: Module<B>> ModuleReshaper<B, M>

pub fn new(template: M) -> Self

Build a reshaper from a template module.

The template is cloned and retained; its float-leaf count is computed once and cached as num_params.

pub fn template(&self) -> &M

Borrow the retained template module.

pub fn num_params(&self) -> usize

Number of float parameters (flat-vector length).

Inherent mirror of ParamReshaper::num_params so callers can read the width without the M: Sync bound the trait requires.

Trait Implementations

impl<B: Backend, M: Module<B>> Clone for ModuleReshaper<B, M>

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<B: Backend, M: Module<B>> Debug for ModuleReshaper<B, M>

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

Formats the value using the given formatter.

impl<B, M> ParamReshaper<B> for ModuleReshaper<B, M>

where B: Backend, M: Module<B> + Sync,

type Module = M

The Burn module type this reshaper flattens and reconstructs.

fn num_params(&self) -> usize

Total number of trainable float parameters (the flat-vector length).

fn flatten(&self, module: &M, device: &B::Device) -> Tensor<B, 1>

Flatten all float Param leaves of module into a 1-D tensor.

fn unflatten(&self, flat: Tensor<B, 1>) -> M

Clone the template module and replace its float leaves with slices of flat, in the same order as flatten.