rsomics-quantile-transform 0.1.0

Quantile transformer — value-exact scikit-learn QuantileTransformer port mapping features to uniform or normal distributions
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155

//! Per-column forward transform — sklearn `QuantileTransformer._transform_col`.
//!
//! Forward map:
//!   1. Double-interp averaging (handles ties):
//!      `0.5 * (interp(x, quantiles, refs) - interp(-x, -quantiles[::-1], -refs[::-1]))`
//!   2. Force boundary values to exactly 0/1.
//!   3. For normal output: apply `ndtri` then clip to `[CLIP_MIN, CLIP_MAX]`.
//!
//! Uniform output is purely linear interpolation → BIT-EXACT vs sklearn (0 ULP).
//! Normal output adds the Cephes `ndtri` transcendental; cross-arch last bits can
//! differ ≤1 ULP, so compat tolerance is ≤1e-12 relative.

use crate::ndtri::{CLIP_MAX, CLIP_MIN, ndtri};

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum OutputDistribution {
    Uniform,
    Normal,
}

/// Linear interpolation matching `np.interp(x, xp, fp)`.
///
/// numpy's C implementation computes `slope = (fp[j+1]-fp[j])/(xp[j+1]-xp[j])` then
/// returns `fp[j] + slope * (x - xp[j])` — which is `fp[j].mul_add(slope, dx)` in FP.
/// Using the same `fp[i].mul_add(slope, dx)` → `fp[i] + slope*(x-xp[i])` order
/// via `f64::mul_add` (FMA) matches numpy bit-for-bit including in the subsample regime.
fn np_interp(x: f64, xp: &[f64], fp: &[f64]) -> f64 {
    debug_assert_eq!(xp.len(), fp.len());
    let n = xp.len();
    if x <= xp[0] {
        return fp[0];
    }
    if x >= xp[n - 1] {
        return fp[n - 1];
    }
    let idx = xp.partition_point(|&v| v <= x);
    let i = idx - 1;
    let slope = (fp[i + 1] - fp[i]) / (xp[i + 1] - xp[i]);
    // FMA: slope * (x - xp[i]) + fp[i]
    slope.mul_add(x - xp[i], fp[i])
}

/// Transform a single column in place, matching `_transform_col(inverse=False)`.
pub fn transform_col(
    col: &mut [f64],
    quantiles: &[f64],
    references: &[f64],
    dist: OutputDistribution,
) {
    let lower_bound_x = quantiles[0];
    let upper_bound_x = quantiles[quantiles.len() - 1];

    // Build reversed views for the second interp.
    let q_rev: Vec<f64> = quantiles.iter().rev().map(|&v| -v).collect();
    let r_rev: Vec<f64> = references.iter().rev().map(|&v| -v).collect();

    for v in col.iter_mut() {
        if v.is_nan() {
            continue;
        }
        let x = *v;

        // sklearn uses `== lower/upper_bound_x` for uniform boundary detection.
        let at_lower = x == lower_bound_x;
        let at_upper = x == upper_bound_x;

        // Double-interp averaging (ties handling, per sklearn comment).
        let fwd = np_interp(x, quantiles, references);
        let rev = -np_interp(-x, &q_rev, &r_rev);
        let mut y = 0.5 * (fwd + rev);

        // Exact boundary values must land at exactly 0 / 1 (sklearn sets them after interp).
        if at_upper {
            y = 1.0;
        }
        if at_lower {
            y = 0.0;
        }

        if dist == OutputDistribution::Normal {
            y = ndtri(y).clamp(CLIP_MIN, CLIP_MAX);
        }

        *v = y;
    }
}

/// Transform every column of the matrix (row-major `data`, `n_rows × n_cols`).
pub fn transform_matrix(
    data: &mut [f64],
    n_rows: usize,
    n_cols: usize,
    quantiles_per_col: &[Vec<f64>],
    references: &[f64],
    dist: OutputDistribution,
) {
    // Work column by column; extract → transform → scatter back.
    for j in 0..n_cols {
        let mut col: Vec<f64> = (0..n_rows).map(|i| data[i * n_cols + j]).collect();
        transform_col(&mut col, &quantiles_per_col[j], references, dist);
        for (i, v) in col.into_iter().enumerate() {
            data[i * n_cols + j] = v;
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    fn close(a: f64, b: f64) {
        assert!(
            (a - b).abs() < 1e-12,
            "got={a} want={b} diff={}",
            (a - b).abs()
        );
    }

    #[test]
    fn np_interp_basic() {
        let xp = [0.0, 1.0, 2.0];
        let fp = [0.0, 0.5, 1.0];
        close(np_interp(0.5, &xp, &fp), 0.25);
        close(np_interp(0.0, &xp, &fp), 0.0);
        close(np_interp(2.0, &xp, &fp), 1.0);
        close(np_interp(-1.0, &xp, &fp), 0.0); // below → fp[0]
        close(np_interp(3.0, &xp, &fp), 1.0); // above → fp[-1]
    }

    #[test]
    fn uniform_ties_average() {
        // quantiles = [1,2,2,2,3], refs = [0,.25,.5,.75,1]
        // x=2 → fwd=interp(2,[1,2,2,2,3],[0,.25,.5,.75,1])=0.25
        //        rev=-interp(-2,[-3,-2,-2,-2,-1],[−1,−.75,−.5,−.25,0])
        //        fwd for rev interp: x=-2 in [-3,-2,-2,-2,-1] → 0.75
        //        rev term = -(−0.75) but wait — refs_rev = [−1,−.75,−.5,−.25,0]
        //        -interp(-2,[-3,-2,-2,-2,-1],[-1,-.75,-.5,-.25,0]) = -(-0.75) = 0.75
        //        average = 0.5*(0.25+0.75) = 0.5
        let quantiles = [1.0, 2.0, 2.0, 2.0, 3.0];
        let refs = [0.0, 0.25, 0.5, 0.75, 1.0];
        let mut col = [2.0];
        transform_col(&mut col, &quantiles, &refs, OutputDistribution::Uniform);
        close(col[0], 0.5);
    }

    #[test]
    fn boundary_forced_to_exact() {
        let quantiles = [1.0, 2.0, 3.0];
        let refs = [0.0, 0.5, 1.0];
        let mut col = [1.0, 3.0];
        transform_col(&mut col, &quantiles, &refs, OutputDistribution::Uniform);
        assert_eq!(col[0], 0.0);
        assert_eq!(col[1], 1.0);
    }
}