Struct DeterministicPredictor 

pub struct DeterministicPredictor { /* private fields */ }

Deterministic gradient predictor.

Maintains gradient history and fits a linear model to predict future gradients deterministically.

Implementations

impl DeterministicPredictor

pub fn new( param_shapes: &[(String, Vec<usize>)], config: DeterministicPredictionConfig, device: &Device, ) -> Result<DeterministicPredictor, OptimError>

Create a new deterministic predictor.

Arguments

• param_shapes - List of (name, shape) tuples for parameters
• config - Prediction configuration
• device - Device for tensor storage

pub fn in_warmup(&self) -> bool

Check if still in warmup phase (must compute full gradients).

pub fn needs_correction(&self) -> bool

Check if correction is needed based on residual magnitude.

pub fn record_gradient( &mut self, gradients: &HashMap<String, Tensor>, is_correction: bool, ) -> Result<(), OptimError>

Record a gradient from full computation.

Updates history and potentially refits the prediction model.

Arguments

• gradients - Map of parameter names to gradient tensors
• is_correction - Whether this is a correction step (vs. regular full step)

pub fn predict_gradient( &mut self, ) -> Result<HashMap<String, Tensor>, OptimError>

Predict gradient for current step.

Uses the fitted linear model to extrapolate from history. Prediction is fully deterministic.

Returns

Map of parameter names to predicted gradient tensors.

pub fn get_stats(&self) -> &PredictorStatistics

Get prediction statistics.

pub fn reset(&mut self) -> Result<(), OptimError>

Reset predictor state.

pub const fn global_step(&self) -> usize

Get current global step.

pub const fn is_ready(&self) -> bool

Check if warmup is complete.