Struct GradientAccumulator

pub struct GradientAccumulator<A: Float, D: Dimension> { /* private fields */ }

Gradient accumulation utility

Implementations

impl<A: Float + ScalarOperand + Debug, D: Dimension> GradientAccumulator<A, D>

pub fn new(accumulation_steps: usize, average_gradients: bool) -> Self

Create a new gradient accumulator

Arguments

• accumulation_steps - Number of micro-batches to accumulate before stepping
• average_gradients - Whether to average gradients (true) or sum them (false)

pub fn accumulate(&mut self, gradients: &Array<A, D>) -> bool

Add gradients from a micro-batch

Arguments

• gradients - Gradients from the current micro-batch

Returns

true if enough gradients have been accumulated and it’s time to step

pub fn get_and_reset(&mut self) -> Option<Array<A, D>>

Get the accumulated gradients and reset the accumulator

Returns

The accumulated gradients, ready for optimization step

pub fn progress(&self) -> (usize, usize)

Get current accumulation progress

pub fn is_ready(&self) -> bool

Check if ready for optimization step

pub fn reset(&mut self)

Reset the accumulator

pub fn set_accumulation_steps(&mut self, steps: usize)

Change accumulation steps

Trait Implementations

impl<A: Clone + Float, D: Clone + Dimension> Clone for GradientAccumulator<A, D>

fn clone(&self) -> GradientAccumulator<A, D>

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<A: Debug + Float, D: Debug + Dimension> Debug for GradientAccumulator<A, D>

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

Formats the value using the given formatter.