Struct ExponentialMovingAverage 

pub struct ExponentialMovingAverage { /* private fields */ }

Exponential Moving Average of model parameters.

Maintains shadow copies of parameters that are updated as a weighted running average: shadow = decay * shadow + (1 - decay) * param. This is commonly used to produce a smoothed version of model weights that often generalises better at inference time.

Implementations

impl ExponentialMovingAverage

pub fn new(decay: f32) -> Self

Creates a new EMA tracker with the given decay factor (e.g. 0.999).

pub fn register(&mut self, params: &[Tensor])

Registers initial parameter values as shadow copies.

pub fn update(&mut self, params: &[Tensor])

Updates shadow parameters: shadow = decay * shadow + (1 - decay) * param.

Panics if the number of tensors does not match the registered count or if any tensor length differs from its shadow counterpart.

pub fn shadow_params(&self) -> &[Tensor]

Returns a reference to the shadow parameters.

pub fn apply_shadow(&self, params: &mut [Tensor])

Copies shadow parameter values into the provided mutable slice.

Panics if the slice length does not match the shadow parameter count or if any tensor length differs.

pub fn num_updates(&self) -> usize

Returns the number of update steps performed so far.