Struct NesterovAcceleratedGradient 

pub struct NesterovAcceleratedGradient { /* private fields */ }

Nesterov Accelerated Gradient optimizer.

NAG uses lookahead to evaluate the gradient at the predicted next position, which can lead to faster convergence than standard momentum methods. Update rule: v_t = momentum * v_{t-1} + lr * grad(x_t + momentum * v_{t-1}) x_{t+1} = x_t - v_t

Implementations

impl NesterovAcceleratedGradient

pub fn new(config: NesterovAcceleratedGradientConfig) -> Self

Create a new NAG optimizer.

pub fn with_defaults(learning_rate: f32, momentum: f32) -> Self

Create NAG with default configuration.

pub fn get_config(&self) -> &NesterovAcceleratedGradientConfig

Get the configuration.

pub fn set_current_loss(&mut self, loss: f32)

Set a loss value for restart detection.

Trait Implementations

impl Debug for NesterovAcceleratedGradient

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

Formats the value using the given formatter.

impl OptimizerState for NesterovAcceleratedGradient

fn zero_grad(&mut self) -> Result<()>

Zero out gradients

fn step(&mut self, parameters: &mut [Tensor]) -> Result<()>

Perform optimization step

fn get_lr(&self) -> f32

Get current learning rate

fn set_lr(&mut self, lr: f32)

Set learning rate

fn state_dict(&self) -> Result<HashMap<String, Tensor>>

Save optimizer state to dictionary

fn load_state_dict(&mut self, state: HashMap<String, Tensor>) -> Result<()>

Load optimizer state from dictionary