Struct ReduceLROnPlateau 

pub struct ReduceLROnPlateau { /* private fields */ }

Reduce the learning rate when a monitored metric plateaus.

Unlike the other schedulers which step automatically, this scheduler requires you to report the metric value (e.g., validation loss) and it decides whether to reduce the LR.

Arguments (builder pattern)

• factor: Factor to multiply LR by when reducing (default: 0.1)
• patience: Number of steps with no improvement before reducing (default: 10)
• min_lr: Lower bound on the learning rate (default: 1e-6)
• threshold: Minimum improvement to qualify as improvement (default: 1e-4)

Example

let mut sched = ReduceLROnPlateau::new(0.01);
// After each epoch:
let new_lr = sched.step_metric(val_loss);
optimizer.set_learning_rate(new_lr);

Implementations

impl ReduceLROnPlateau

pub fn new(initial_lr: f64) -> ReduceLROnPlateau

Create a new ReduceLROnPlateau with sensible defaults.

Default: factor=0.1, patience=10, min_lr=1e-6, threshold=1e-4, mode=min

pub fn factor(self, factor: f64) -> ReduceLROnPlateau

Set the factor by which to reduce LR (default: 0.1).

pub fn patience(self, patience: u64) -> ReduceLROnPlateau

Set patience (steps without improvement before reducing, default: 10).

pub fn min_lr(self, min_lr: f64) -> ReduceLROnPlateau

Set the minimum learning rate (default: 1e-6).

pub fn threshold(self, threshold: f64) -> ReduceLROnPlateau

Set the improvement threshold (default: 1e-4).

pub fn mode_max(self) -> ReduceLROnPlateau

Set mode to maximize (higher metric = better). Default is minimize (lower metric = better).

pub fn step_metric(&mut self, metric: f64) -> f64

Report a metric value and return the (possibly updated) learning rate.

Call this once per epoch/evaluation with the metric value (e.g., val loss).

pub fn lr(&self) -> f64

Get the current learning rate.

pub fn best_metric(&self) -> f64

Get the best metric value seen so far.

pub fn bad_steps(&self) -> u64

Get number of steps without improvement.

pub fn reset(&mut self)

Reset state.