Struct TrainingLoop 

pub struct TrainingLoop { /* private fields */ }

The main training orchestrator.

Coordinates data batching, forward/backward passes, optimizer steps, LR scheduling, early stopping, distributed gradient synchronisation, and callback dispatch.

Implementations

impl TrainingLoop

pub fn new(config: TrainingConfig) -> Self

Create a new TrainingLoop with the given configuration.

Uses LocalGradientSync by default (no distributed sync).

pub fn add_callback(&mut self, cb: Box<dyn TrainingCallback>)

Register a training callback.

pub fn with_gradient_sync(self, sync: Box<dyn GradientSync>) -> Self

Override the gradient synchronization strategy.

pub fn run( &mut self, data: &dyn DataProvider, optimizer: &mut dyn Optimizer, lr_scheduler: &mut dyn LrScheduler, early_stopping: Option<&mut EarlyStopping>, model_weights: &mut Array1<f32>, model_bias: &mut f32, ) -> ModelResult<TrainingResult>

Run training.

Parameters

• data: dataset providing batched samples.
• optimizer: mutable reference to any Optimizer implementation.
• lr_scheduler: mutable reference to any LrScheduler implementation.
• early_stopping: optional early-stopping guard from crate::checkpoint.
• model_weights: mutable weight vector; updated in-place each epoch.
• model_bias: mutable bias scalar; updated in-place each epoch.

Returns

A TrainingResult summarising the completed run.