Struct BurnOptimizer 

pub struct BurnOptimizer<B, M, O>
where
    B: AutodiffBackend,
    M: AutodiffModule<B>,
    O: Optimizer<M, B>,
{ /* private fields */ }

Burn-side optimizer wrapper.

Wraps a Burn OptimizerAdaptor<O, M, B> and exposes it through the BackendOptimizer trait. The trainer bodies hold one of these and route their gradient step through BackendOptimizer::step_module.

Implementations

impl<B, M, O> BurnOptimizer<B, M, O>

where B: AutodiffBackend, M: AutodiffModule<B>, O: Optimizer<M, B>,

pub fn new(inner: O, learning_rate: f64) -> Self

Wrap a Burn optimizer (typically AdamConfig::new().init()).

pub fn inner(&self) -> &O

Borrow the wrapped Burn optimizer.

pub fn inner_mut(&mut self) -> &mut O

Mutably borrow the wrapped Burn optimizer. The trainer bodies call inner_mut().step(lr, module, grads) directly from their loss closures.

pub fn grad_clip_norm(&self) -> Option<f64>

The currently-staged gradient-norm cap, if any.

Trait Implementations

impl<B, M, O> BackendOptimizer for BurnOptimizer<B, M, O>

where B: AutodiffBackend, M: AutodiffModule<B>, O: Optimizer<M, B>,

type Module = M

The module type the optimizer steps.

fn clip_grad_norm(&mut self, max: f64)

Stage the maximum global gradient L2-norm.

fn step_module(&mut self, module: Self::Module) -> Self::Module

Burn-style move-through update.

fn learning_rate(&self) -> f64

Construction-time learning rate. Exposed for diagnostics.

impl<B, M, O> Debug for BurnOptimizer<B, M, O>

where B: AutodiffBackend, M: AutodiffModule<B>, O: Optimizer<M, B>,

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

Formats the value using the given formatter.