Struct AdamW 

pub struct AdamW {
    pub lr: f64,
    pub beta1: f64,
    pub beta2: f64,
    pub eps: f64,
    pub weight_decay: f64,
    /* private fields */
}

AdamW optimizer (Loshchilov & Hutter 2017) — Adam with decoupled weight decay.

Unlike Adam (which couples weight decay into the gradient), AdamW applies weight decay directly to the parameters before the gradient update: params ← params - lr · weight_decay · params (L2 shrinkage) then Adam update on the raw gradient

Fields

lr: f64

Learning rate.

beta1: f64

First moment decay.

beta2: f64

Second moment decay.

eps: f64

Numerical stability constant.

weight_decay: f64

Decoupled L2 weight-decay coefficient.

Implementations

impl AdamW

pub fn new(lr: f64) -> Self

Create a new AdamW optimizer with default β₁=0.9, β₂=0.999, ε=1e-8, weight_decay=0.01.

pub fn step( &mut self, params: &mut Vec<f64>, grad: &[f64], ) -> OptimizeResult<()>

Perform one AdamW update step.

pub fn reset_state(&mut self)

Reset all state.

Trait Implementations

impl Clone for AdamW

fn clone(&self) -> AdamW

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for AdamW

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

Formats the value using the given formatter.