Struct OneCycle 

pub struct OneCycle {
    pub max_lr: f64,
    pub pct_start: f64,
    pub anneal_strategy: AnnealStrategy,
    pub total_epochs: usize,
    pub div_factor: f64,
    pub final_div_factor: f64,
}

One-cycle learning rate policy.

Implements Smith & Touvron’s 1cycle policy: the learning rate rises from base_lr to max_lr over the first pct_start fraction of training, then anneals back down to a minimum learning rate over the remainder.

Reference: Smith (2018), “A disciplined approach to neural network hyper-parameters”.

Fields

max_lr: f64

Maximum learning rate (peak of the cycle)

pct_start: f64

Fraction of total epochs for the increasing phase (0 < pct_start < 1)

anneal_strategy: AnnealStrategy

Annealing strategy for the decreasing phase

total_epochs: usize

Total number of training epochs

div_factor: f64

Minimum (final) learning rate as a fraction of base_lr

final_div_factor: f64

Final learning rate divisor (final_lr = base_lr / final_div_factor)

Implementations

impl OneCycle

pub fn new( max_lr: f64, pct_start: f64, anneal_strategy: AnnealStrategy, total_epochs: usize, ) -> Self

Create a new one-cycle schedule.

Arguments

• max_lr - Peak learning rate
• pct_start - Fraction of epochs for the warmup/increase phase
• anneal_strategy - How to anneal during the decrease phase
• total_epochs - Total training epochs

Trait Implementations

impl Clone for OneCycle

fn clone(&self) -> OneCycle

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for OneCycle

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

Formats the value using the given formatter.

impl LrSchedule for OneCycle

fn get_lr(&self, epoch: usize, base_lr: f64) -> f64

Compute the learning rate for the given epoch/step.