Struct GradAccumulator 

pub struct GradAccumulator { /* private fields */ }

Accumulates parameter gradients across multiple micro-batches.

Supports mean-reduction normalisation and global-norm gradient clipping.

Implementations

impl GradAccumulator

pub fn new() -> Self

Create an empty accumulator.

pub fn accumulate(&mut self, name: &str, grad: &Array1<f32>) -> ModelResult<()>

Accumulate grad into the named parameter slot.

If no gradient for name exists yet, it is initialised to zeros of the same length before accumulation.

Errors

Returns ModelError::numerical_instability if grad contains NaN or Inf, or ModelError::dimension_mismatch if the lengths differ between calls for the same name.

pub fn get(&self, name: &str) -> Option<&Array1<f32>>

Return a reference to the accumulated gradient for name, if any.

pub fn normalize(&mut self)

Divide each accumulated gradient by its accumulation count (mean reduction).

pub fn zero_grad(&mut self)

Zero all accumulated gradients and reset counts.

pub fn apply_clip(&mut self, max_norm: f32) -> f32

Clip gradients by global L2 norm.

Computes the L2 norm across all parameter gradients; if it exceeds max_norm, scales all gradients by max_norm / norm.

Returns the L2 norm before clipping.

pub fn param_names(&self) -> Vec<&str>

Return the names of all parameters with accumulated gradients.

Trait Implementations

impl Debug for GradAccumulator

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

Formats the value using the given formatter.

impl Default for GradAccumulator

fn default() -> GradAccumulator

Returns the “default value” for a type.