Function quantile_transform 

pub fn quantile_transform<F: Float + Send + Sync + 'static>(
    x: &Array2<F>,
    axis: usize,
    n_quantiles: usize,
    output_distribution: OutputDistribution,
    subsample: usize,
) -> Result<Array2<F>, FerroError>

Transform features using quantiles information — the standalone, estimator-less API mirroring scikit-learn’s quantile_transform free function (sklearn/preprocessing/_data.py:2978,:3107-3119).

This is a thin wrapper that reuses the fitted estimator: it constructs a QuantileTransformer with the given parameters, then runs the SHIPPED Fit::fit → Transform::transform (i.e. fit_transform) — it does NOT reimplement any quantile math. Each call fits on the supplied data and immediately transforms it (sklearn n = QuantileTransformer(...); X = n.fit_transform(X), :3107-3116).

Parameters

• x — the data to transform, shape (n_samples, n_features).
• axis — axis along which to transform (sklearn axis=0 default, :3013-3015). axis == 0 transforms each feature (column) independently — fit_transform(X) (:3115-3116). axis == 1 transforms each sample (row) independently — fit_transform(X.T).T (:3117-3118): the matrix is transposed (so rows become columns / the per-feature path operates on each original row), fit-transformed, then transposed back to the original orientation.
• n_quantiles — number of quantile reference landmarks (sklearn default 1000, clamped to n_samples, :3017-3023).
• output_distribution — OutputDistribution::Uniform (sklearn default, :3025-3027) or OutputDistribution::Normal.
• subsample — maximum samples used to estimate the quantiles (0 = use all). See the scope note below.

Scope

Unlike sklearn’s quantile_transform, the ignore_implicit_zeros, random_state, and copy keyword arguments are not surfaced: this path is dense, deterministic, and always returns a freshly allocated array (copy is implicitly True). The subsample parameter is threaded to the estimator, but the actual random subsampling is REQ-6 (#1324, NOT-STARTED): for n_samples > subsample the deterministic strided pick is used instead of sklearn’s RNG draw, so this function matches sklearn’s output only when n_samples <= subsample (use subsample=None or a large value on the sklearn side to compare). NaN is disregarded when fitting and preserved in the output (sklearn Notes, :3078-3079).

Errors

Returns FerroError::InvalidParameter if axis is neither 0 nor 1 (mirroring sklearn’s ValueError("axis should be either equal to 0 or 1. Got axis={axis}")). Also propagates any error from the underlying Fit::fit / Transform::transform (e.g. FerroError::InsufficientSamples for fewer than 2 rows along the chosen axis, FerroError::InvalidParameter for n_quantiles < 2 or non-finite ±inf input).

Examples

use ferrolearn_preprocess::quantile_transformer::{quantile_transform, OutputDistribution};
use ndarray::array;

let x = array![[1.0_f64], [2.0], [3.0], [4.0], [5.0]];
let out = quantile_transform(&x, 0, 5, OutputDistribution::Uniform, 0).unwrap();
// First maps to 0.0, last to 1.0.
assert!((out[[0, 0]] - 0.0_f64).abs() < 1e-9);
assert!((out[[4, 0]] - 1.0_f64).abs() < 1e-9);