Struct NaturalGradient 

pub struct NaturalGradient { /* private fields */ }

Natural Gradient optimizer

Natural gradients use the Fisher Information Matrix to precondition gradients, providing better-conditioned updates especially for neural networks and probabilistic models.

Reference: “Natural Gradient Works Efficiently in Learning” (Amari, 1998)

Implementations

impl NaturalGradient

pub fn new( params: Vec<Arc<RwLock<Tensor>>>, lr: Option<f32>, momentum: Option<f32>, damping: Option<f32>, fisher_update_freq: Option<usize>, use_empirical_fisher: Option<bool>, fisher_ema_decay: Option<f32>, ) -> Self

Create a new Natural Gradient optimizer

pub fn builder() -> NaturalGradientBuilder

Builder pattern for Natural Gradient optimizer

Trait Implementations

impl Optimizer for NaturalGradient

fn step(&mut self) -> OptimizerResult<()>

Perform a single optimization step

fn zero_grad(&mut self)

Zero all gradients

fn get_lr(&self) -> Vec<f32>

Get the current learning rate

fn set_lr(&mut self, lr: f32)

Set one learning rate for every parameter group (broadcast).

fn set_lrs(&mut self, lrs: &[f32])

Set the learning rate of each parameter group individually.

fn add_param_group( &mut self, params: Vec<Arc<RwLock<Tensor>>>, options: HashMap<String, f32>, )

Add a parameter group

fn parameters(&self) -> Vec<Arc<RwLock<Tensor>>>

Get the parameter tensors managed by this optimizer.

fn state_dict(&self) -> OptimizerResult<OptimizerState>

Get state dict for serialization

fn load_state_dict(&mut self, state: OptimizerState) -> OptimizerResult<()>

Load state dict