Struct GradientAccumulator 

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

Gradient accumulator for memory-efficient training

Accumulates gradients over multiple micro-batches before applying updates, reducing memory requirements for large batch training.

Examples

use scirs2_core::ndarray::Array1;
use optirs_core::memory_efficient_optimizer::GradientAccumulator;

let mut accumulator = GradientAccumulator::<f32>::new(1000);

// Accumulate gradients from 4 micro-batches
for _ in 0..4 {
    let micro_batch_grads = Array1::from_elem(1000, 0.1);
    accumulator.accumulate(&micro_batch_grads.view()).unwrap();
}

// Get averaged gradients
let avg_grads = accumulator.average().unwrap();

Implementations

impl<A: Float + ScalarOperand + Debug + Zero> GradientAccumulator<A>

pub fn new(size: usize) -> Self

Creates a new gradient accumulator

Arguments

• size - Size of gradient vectors

pub fn accumulate(&mut self, gradients: &ArrayView1<'_, A>) -> Result<()>

Accumulate a gradient vector

Arguments

• gradients - Gradients to accumulate

pub fn count(&self) -> usize

Get the number of accumulated gradients

pub fn average(&mut self) -> Result<Array1<A>>

Compute the average of accumulated gradients

Returns the averaged gradients and resets the accumulator.

pub fn reset(&mut self)

Reset the accumulator

pub fn is_ready(&self, target: usize) -> bool

Check if accumulator has reached target count

Arguments

• target - Target number of accumulations