Struct tangram_tree::TrainOptions

pub struct TrainOptions {

    pub binned_features_layout: BinnedFeaturesLayout,
    pub compute_losses: bool,
    pub early_stopping_options: Option<EarlyStoppingOptions>,
    pub l2_regularization_for_continuous_splits: f32,
    pub l2_regularization_for_discrete_splits: f32,
    pub learning_rate: f32,
    pub max_depth: Option<usize>,
    pub max_examples_for_computing_bin_thresholds: usize,
    pub max_leaf_nodes: usize,
    pub max_rounds: usize,
    pub max_valid_bins_for_number_features: u8,
    pub min_examples_per_node: usize,
    pub min_gain_to_split: f32,
    pub min_sum_hessians_per_node: f32,
    pub smoothing_factor_for_discrete_bin_sorting: f32,
}

These are the options passed to Regressor::train, BinaryClassifier::train, and MulticlassClassifier::train.

Fields

binned_features_layout: BinnedFeaturesLayout

This option controls whether binned features will be laid out in row major or column major order. Each will produce the same result, but row major will be faster for datasets with more rows and fewer columns, while column major will be faster for datasets with fewer rows and more columns.

compute_losses: bool

If true, the model will include the loss on the training data after each round.

early_stopping_options: Option<EarlyStoppingOptions>

This option controls early stopping. If it is Some, then early stopping will be enabled. If it is None, then early stopping will be disabled.

l2_regularization_for_continuous_splits: f32

This option sets the L2 regularization value for continuous splits, which helps avoid overfitting.

l2_regularization_for_discrete_splits: f32

This option sets the L2 regularization value for discrete splits, which helps avoid overfitting.

learning_rate: f32

The learning rate scales the leaf values to control the effect each tree has on the output.

max_depth: Option<usize>

This is the maximum depth of a single tree. If this value is None, the depth will not be limited.

max_examples_for_computing_bin_thresholds: usize

This is the maximum number of examples to consider when determining the bin thresholds for number features.

max_leaf_nodes: usize

This is the maximum number of leaf nodes in a single tree.

max_rounds: usize

This is the maximum number of rounds of training that will occur. Fewer rounds may be trained if early stopping is enabled.

max_valid_bins_for_number_features: u8

When computing the bin thresholds for number features, this is the maximum number of bins for valid values to create. If the number of unique values in the number feature is less than this value, the thresholds will be equal to the unique values, which can improve accuracy when number features have a small set of possible values.

min_examples_per_node: usize

A split will only be considered valid if the number of training examples sent to each of the resulting children is at least this value.

min_gain_to_split: f32

A node will only be split if the best split achieves at least this minimum gain.

min_sum_hessians_per_node: f32

A split will only be considered valid if the sum of hessians in each of the resulting children is at least this value.

smoothing_factor_for_discrete_bin_sorting: f32

When choosing which direction each enum variant should be sent in a discrete split, the enum variants are sorted by a score computed from the sum of gradients and hessians for examples with that enum variant. This smoothing factor is added to the denominator of that score.

Trait Implementations

impl Clone for TrainOptions

fn clone(&self) -> TrainOptions

Returns a copy of the value.

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

Performs copy-assignment from source.

impl Debug for TrainOptions

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

Formats the value using the given formatter.

impl Default for TrainOptions

fn default() -> TrainOptions

Returns the “default value” for a type.