Struct Collector 

pub struct Collector {
    pub tensors: Vec<TensorSnapshot>,
    /* private fields */
}

Collects tensor views from modules without copying data.

This collector traverses a module hierarchy and creates lightweight views of tensors that can be materialized to TensorData on demand.

Examples

Collect all tensors

let collector = Collector::new(None, None, false);
// Use with module.visit(&mut collector);
let all_tensors = collector.tensors;

Filter with single pattern

let filter = PathFilter::new().with_regex(r"^encoder\..*");
let collector = Collector::new(Some(filter), None, false);
// Use with module.visit(&mut collector);
// Only collects tensors starting with "encoder."

Filter with multiple patterns (OR union)

let filter = PathFilter::new()
    .with_regex(r"^encoder\..*")  // Match all encoder tensors
    .with_regex(r".*\.bias$");    // OR match any bias tensors
let collector = Collector::new(Some(filter), None, false);
// Use with module.visit(&mut collector);
// Collects tensors matching ANY of the patterns

Fields

tensors: Vec<TensorSnapshot>

Collection of tensor views

Implementations

impl Collector

pub fn new( filter: Option<PathFilter>, adapter: Option<Box<dyn ModuleAdapter>>, skip_enum_variants: bool, ) -> Self

Create a new tensor view collector with an optional filter and adapter.

Arguments

• filter - An optional PathFilter to determine which tensors to collect. When None, all tensors are collected.
• adapter - Optional adapter to transform tensors based on container types. Applied to all collected tensors before returning.
• skip_enum_variants - Skip enum variant names when building paths. When true, paths will not include enum variant names (e.g., “feature.weight” instead of “feature.BaseConv.weight”). Useful when exporting to formats like PyTorch that don’t use enum variants.

Examples

// Collect all tensors without adapter
let collector = Collector::new(None, None, false);

// Use PathFilter builder
let filter = PathFilter::new()
    .with_regex(r"^encoder\..*")
    .with_full_path("decoder.weight");
let collector = Collector::new(Some(filter), None, false);

// Skip enum variants for PyTorch export
let collector = Collector::new(None, None, true);

pub fn into_tensors(self) -> Vec<TensorSnapshot>

Apply the adapter to collected tensors and return the result.

Trait Implementations

impl Default for Collector

fn default() -> Self

Returns the “default value” for a type.

impl<B: Backend> ModuleVisitor<B> for Collector

fn enter_module(&mut self, name: &str, container_type: &str)

Called when entering a submodule.

fn exit_module(&mut self, _name: &str, container_type: &str)

Called when exiting a submodule.

fn visit_float<const D: usize>(&mut self, param: &Param<Tensor<B, D>>)

Visit a float parameter in the module.

fn visit_int<const D: usize>(&mut self, param: &Param<Tensor<B, D, Int>>)

Visit an int parameter in the module.

fn visit_bool<const D: usize>(&mut self, param: &Param<Tensor<B, D, Bool>>)

Visit a bool parameter in the module.

fn visit_float_with_path<const D: usize>( &mut self, path: &[String], id: ParamId, tensor: &Tensor<B, D>, )

Visit a float tensor with its full module path.

fn visit_int_with_path<const D: usize>( &mut self, path: &[String], id: ParamId, tensor: &Tensor<B, D, Int>, )

Visit an int tensor with its full module path.

fn visit_bool_with_path<const D: usize>( &mut self, path: &[String], id: ParamId, tensor: &Tensor<B, D, Bool>, )

Visit a bool tensor with its full module path.