Struct DynamicLossScaler 

pub struct DynamicLossScaler { /* private fields */ }

State for dynamic loss scaling during mixed-precision training.

Implementations

impl DynamicLossScaler

pub fn new(initial_scale: f32) -> Self

pub fn scale(&self) -> f32

pub fn scale_loss(&self, loss: &Tensor) -> Result<Tensor, ModelError>

Scale a loss tensor by the current loss scale factor.

pub fn unscale_gradients(&self, gradients: &[Tensor]) -> Vec<Tensor>

Unscale gradients by dividing by the current scale factor.

pub fn check_overflow(gradients: &[Tensor]) -> bool

Check if any gradient contains inf/nan (overflow indicator).

pub fn update(&mut self, overflow: bool) -> bool

Update the scaler state after a training step. Returns true if the step should be applied (no overflow), false to skip.

Trait Implementations

impl Clone for DynamicLossScaler

fn clone(&self) -> DynamicLossScaler

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for DynamicLossScaler

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

Formats the value using the given formatter.