Struct OptimizationParams 

pub struct OptimizationParams {

    pub n_epochs: usize,
    pub batch_size: usize,
    pub learning_rate: f64,
    pub beta1: f64,
    pub beta2: f64,
    pub penalty: f32,
    pub repulsion_strength: f32,
    pub patience: Option<i32>,
    pub loss_reduction: LossReduction,
    pub min_desired_loss: Option<f64>,
    pub timeout: Option<u64>,
    pub verbose: bool,
    pub neg_sample_rate: usize,
    pub cooldown_ms: u64,
    pub figures_dir: Option<PathBuf>,
}

Configuration for stochastic gradient descent optimization.

These parameters control the embedding optimization process.

Fields

n_epochs: usize

Number of optimization epochs.

Default: 100

batch_size: usize

The number of samples to process in each training batch.

Default: 1000

learning_rate: f64

Initial learning rate for the Adam optimizer.

Default: 0.001

beta1: f64

Beta1 parameter for the Adam optimizer.

Default: 0.9

beta2: f64

Beta2 parameter for the Adam optimizer.

Default: 0.999

penalty: f32

L2 regularization (weight decay) penalty.

Default: 1e-5

repulsion_strength: f32

Weight applied to the repulsion term of the UMAP cross-entropy loss.

Default: 1.0

patience: Option<i32>

Number of epochs to wait for improvement before triggering early stopping. None disables early stopping.

Default: None

loss_reduction: LossReduction

The method used to reduce the loss (mean or sum).

Default: Sum

min_desired_loss: Option<f64>

Minimum desired loss to achieve before stopping early.

Default: None

timeout: Option<u64>

Maximum training time in seconds. None means no limit.

Default: None

verbose: bool

Whether to show detailed progress information during training.

Default: false

neg_sample_rate: usize

Number of negative (repulsion) samples drawn per positive (attraction) edge each epoch.

Higher values produce stronger repulsion and better cluster separation at the cost of more computation per epoch.

Default: 5

cooldown_ms: u64

Milliseconds to sleep at the end of every training epoch.

Inserting a small pause between epochs lets the GPU scheduler breathe, preventing the device from being pinned at 100 % utilisation for the entire run. Typical values:

cooldown_ms	Effect
0 (default)	No pause — maximum throughput
1–5	Barely perceptible pause, ~10–20 % GPU headroom
10–50	Noticeable slowdown, significant GPU headroom

Default: 0 (disabled)

figures_dir: Option<PathBuf>

Directory where loss-curve and embedding snapshot plots are written when verbose is true (or the verbose feature flag is enabled).

Defaults to "figures" (relative to the current working directory). Set this to an absolute path — or any writable location — if the process runs on a read-only filesystem.

When None the default "figures" directory is used.

Trait Implementations

impl Clone for OptimizationParams

fn clone(&self) -> OptimizationParams

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for OptimizationParams

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

Formats the value using the given formatter.

impl Default for OptimizationParams

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for OptimizationParams

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for OptimizationParams

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.