Struct TrialOutcome 

pub struct TrialOutcome {
    pub final_loss: f32,
    pub convergence_epoch: usize,
    pub stability_score: f32,
    pub throughput: f32,
    pub gradient_norm_variance: f32,
    pub peak_memory_bytes: Option<usize>,
}

What a caller-supplied objective reports after really training a configuration.

The tuner cannot know any of these numbers: only the caller runs the model. Every field is therefore a measurement handed back by the objective, and the tuner adds only the wall-clock time it timed itself.

Fields

final_loss: f32

Validation loss at the end of the trial. Lower is better.

convergence_epoch: usize

Epoch at which the run converged (or the number of epochs actually run).

stability_score: f32

Training stability in [0, 1]; 1.0 means no divergence was observed.

throughput: f32

Observed throughput in samples per second.

gradient_norm_variance: f32

Variance of the gradient norm observed during the trial.

peak_memory_bytes: Option<usize>

Peak memory in bytes, if the caller measured it.

Implementations

impl TrialOutcome

pub fn from_loss(final_loss: f32, convergence_epoch: usize) -> Self

Minimal outcome for objectives that only produce a loss.

Unmeasured quantities stay neutral (stability_score = 1.0, zero throughput and gradient variance, no memory figure) rather than being invented.

Trait Implementations

impl Clone for TrialOutcome

fn clone(&self) -> TrialOutcome

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TrialOutcome

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

Formats the value using the given formatter.