Struct LearningSchedule 

pub struct LearningSchedule<T: Float + Debug + ScalarOperand + Send + Sync + 'static> {
    pub initial_rate: T,
    pub current_rate: T,
    pub warmup_steps: usize,
    pub decay_factor: T,
    /* private fields */
}

Linear-warmup-then-exponential-decay learning-rate schedule.

initial_rate, warmup_steps and decay_factor used to be stored and never consulted: the schedule had no way to advance, so current_rate stayed at initial_rate forever and LearningState held a schedule it never used.

Fields

initial_rate: T

current_rate: T

warmup_steps: usize

decay_factor: T

Implementations

impl<T: Float + Debug + ScalarOperand + Send + Sync + 'static> LearningSchedule<T>

pub fn new(initial_rate: T, warmup_steps: usize) -> Self

pub fn rate_at(&self, step: usize) -> T

Rate at step: initial_rate · step / warmup_steps while warming up, then initial_rate · decay_factor^(step - warmup_steps).

With warmup_steps == 0 the warmup phase is skipped entirely.

pub fn step(&mut self) -> T

Advance one step and return the new rate.

pub fn steps_taken(&self) -> usize

Steps taken so far.

Trait Implementations

impl<T: Clone + Float + Debug + ScalarOperand + Send + Sync + 'static> Clone for LearningSchedule<T>

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<T: Debug + Float + Debug + ScalarOperand + Send + Sync + 'static> Debug for LearningSchedule<T>

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

Formats the value using the given formatter.