Struct GradientAccumulator 

pub struct GradientAccumulator { /* private fields */ }

Gradient accumulation

Accumulates gradients over multiple micro-batches before updating. Useful for training with large effective batch sizes on limited memory.

Implementations

impl GradientAccumulator

pub fn new(accumulation_steps: usize) -> Self

pub fn accumulate(&mut self, gradients: &HashMap<String, Tensor>)

Accumulate gradients from a micro-batch

pub fn should_update(&self) -> bool

Check if ready to update (accumulated enough steps)

pub fn get_and_reset(&mut self) -> HashMap<String, Tensor>

Get accumulated gradients and reset

pub fn reset(&mut self)

Reset accumulator