Struct LinearDecayLR 

pub struct LinearDecayLR { /* private fields */ }

Available on crate feature alloc only.

Linearly interpolates the learning rate from initial_lr to final_lr over decay_steps, then holds final_lr forever.

The formula is:

lr = initial_lr - (initial_lr - final_lr) * min(step / decay_steps, 1.0)

This gives a smooth ramp-down that reaches final_lr at exactly step == decay_steps and clamps there for all subsequent steps.

Example

use irithyll::ensemble::lr_schedule::{LRScheduler, LinearDecayLR};

let mut sched = LinearDecayLR::new(0.1, 0.01, 100);
// At step 0 we get the initial rate.
assert!(crate::math::abs((sched.learning_rate(0, 0.0) - 0.1)) < 1e-12);
// At step 50 we're halfway.
assert!(crate::math::abs((sched.learning_rate(50, 0.0) - 0.055)) < 1e-12);
// At step 100 we've reached the final rate.
assert!(crate::math::abs((sched.learning_rate(100, 0.0) - 0.01)) < 1e-12);
// Beyond decay_steps the rate stays clamped.
assert!(crate::math::abs((sched.learning_rate(200, 0.0) - 0.01)) < 1e-12);

Implementations

impl LinearDecayLR

pub fn new(initial_lr: f64, final_lr: f64, decay_steps: u64) -> Self

Create a linear-decay scheduler.

Arguments

• initial_lr – Rate at step 0.
• final_lr – Rate from step decay_steps onward.
• decay_steps – Length of the linear ramp in steps.

Trait Implementations

impl Clone for LinearDecayLR

fn clone(&self) -> LinearDecayLR

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LinearDecayLR

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

Formats the value using the given formatter.

impl LRScheduler for LinearDecayLR

fn learning_rate(&mut self, step: u64, _current_loss: f64) -> f64

Return the learning rate for the given step and current_loss.

fn reset(&mut self)

Reset the scheduler to its initial state.