Enum SchedulerStrategy 

pub enum SchedulerStrategy {
    Constant {
        lr: f64,
    },
    StepDecay {
        initial_lr: f64,
        decay_factor: f64,
        step_size: u64,
    },
    ExponentialDecay {
        initial_lr: f64,
        decay_rate: f64,
    },
    CosineAnnealing {
        initial_lr: f64,
        min_lr: f64,
        t_max: u64,
    },
    WarmupCosine {
        warmup_epochs: u64,
        initial_lr: f64,
        peak_lr: f64,
        min_lr: f64,
        t_max: u64,
    },
    CyclicLR {
        base_lr: f64,
        max_lr: f64,
        step_size: u64,
    },
    ReduceOnPlateau {
        initial_lr: f64,
        factor: f64,
        patience: u64,
        min_lr: f64,
        threshold: f64,
    },
}

Multi-strategy learning rate scheduling for distributed ML training.

Supports seven scheduling strategies including plateau detection.

Variants

Constant

Fixed learning rate that never changes.

Fields

lr: f64

The fixed learning rate.

StepDecay

Multiply learning rate by decay_factor every step_size epochs.

Fields

initial_lr: f64

Starting learning rate.

decay_factor: f64

Multiplicative factor applied every step_size epochs.

step_size: u64

Number of epochs between each decay application.

ExponentialDecay

Exponential decay: lr = initial * exp(-decay_rate * epoch).

Fields

initial_lr: f64

Starting learning rate.

decay_rate: f64

Rate of exponential decay.

CosineAnnealing

Cosine annealing from initial_lr to min_lr over t_max epochs.

Fields

initial_lr: f64

Starting learning rate.

min_lr: f64

Floor learning rate.

t_max: u64

Period (number of epochs for one half-cycle).

WarmupCosine

Linear warmup over warmup_epochs then cosine annealing to min_lr.

Fields

warmup_epochs: u64

Number of epochs for linear warm-up phase.

initial_lr: f64

Starting LR for warm-up (typically a small value or 0).

peak_lr: f64

Peak learning rate reached after warm-up.

min_lr: f64

Floor learning rate after the cosine phase.

t_max: u64

Length of the cosine phase in epochs (not including warmup).

CyclicLR

Triangular (cyclic) learning rate oscillating between base_lr and max_lr.

Fields

base_lr: f64

Lower bound of the LR range.

max_lr: f64

Upper bound of the LR range.

step_size: u64

Half the cycle length (each triangle leg = step_size epochs).

ReduceOnPlateau

Reduce learning rate when validation loss stops improving.

Fields

initial_lr: f64

Starting learning rate.

factor: f64

Multiplicative reduction factor (< 1.0).

patience: u64

Number of epochs with no improvement before reduction.

min_lr: f64

Minimum allowed learning rate.

threshold: f64

Minimum change in loss to qualify as improvement.

Trait Implementations

impl Clone for SchedulerStrategy

fn clone(&self) -> SchedulerStrategy

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for SchedulerStrategy

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

Formats the value using the given formatter.