ferrolearn-preprocess 0.5.0

Preprocessing transformers for the ferrolearn ML framework
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
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205

//! Divergence audit: `FunctionTransformer` (forward element-wise transform,
//! REQ-1) vs scikit-learn 1.5.2 `sklearn/preprocessing/_function_transformer.py`
//! `class FunctionTransformer` `_transform` (`:375-379`:
//! `return func(X, **(kw_args if kw_args else {}))`).
//!
//! ferrolearn ships a thin scalar wrapper applying `Fn(F) -> F` element-wise
//! via `Array2::mapv` (`function_transformer.rs:77-80`). For element-wise
//! ufuncs this is numerically identical to sklearn's whole-array `func(X)`.
//!
//! EVERY expected value below is taken from a LIVE sklearn 1.5.2 oracle call
//! (run from /tmp), pasted in the `// oracle:` comment above the constant, NOT
//! copied from the ferrolearn side (R-CHAR-3). The scalar-vs-array signature
//! mismatch (REQ-2/REQ-3) is a structural NOT-STARTED missing feature, not a
//! runtime value divergence, and is intentionally NOT pinned here (there is no
//! whole-array `func` callable through the `Fn(F) -> F` API).

use ferrolearn_core::traits::Transform;
use ferrolearn_preprocess::FunctionTransformer;
use ndarray::{Array2, array};

/// Green guard (REQ-1 / AC-1): element-wise `np.log1p` ufunc.
///
/// Mirrors `sklearn/preprocessing/_function_transformer.py:375-379`
/// (`func(X)` with `func = np.log1p`).
///
/// Oracle (live sklearn 1.5.2, run from /tmp):
/// ```text
/// python3 -c "import numpy as np; from sklearn.preprocessing import \
///   FunctionTransformer; import json; \
///   print(json.dumps(FunctionTransformer(np.log1p).transform(\
///   np.array([[0.,1.],[2.,3.]])).tolist()))"
/// -> [[0.0, 0.6931471805599453], [1.0986122886681096, 1.3862943611198906]]
/// ```
/// np.log1p(x) = ln(1+x); Rust analog `f64::ln_1p`.
#[test]
#[allow(
    clippy::approx_constant,
    reason = "exact live-sklearn-1.5.2 oracle value ln(2)=log1p(1); R-CHAR-3, not the std constant"
)]
fn guard_log1p_matches_sklearn_oracle() {
    // oracle: FunctionTransformer(np.log1p).transform([[0,1],[2,3]])
    let expected: [[f64; 2]; 2] = [
        [0.0, 0.693_147_180_559_945_3],
        [1.098_612_288_668_109_6, 1.386_294_361_119_890_6],
    ];

    let ft = FunctionTransformer::<f64>::new(|v: f64| v.ln_1p());
    let x = array![[0.0, 1.0], [2.0, 3.0]];
    let out = ft.transform(&x).unwrap();

    assert_eq!(out.shape(), x.shape(), "shape must be preserved");
    for ((r, c), &want) in expected
        .iter()
        .enumerate()
        .flat_map(|(i, row)| row.iter().enumerate().map(move |(j, v)| ((i, j), v)))
    {
        // Bit-exact: numpy and Rust both delegate log1p to libm.
        assert_eq!(
            out[[r, c]].to_bits(),
            want.to_bits(),
            "log1p mismatch at [{r},{c}]: ferro={} sklearn={}",
            out[[r, c]],
            want
        );
    }
}

/// Green guard (REQ-1): element-wise `np.expm1` ufunc cross-check.
///
/// Oracle (live sklearn 1.5.2, run from /tmp):
/// ```text
/// python3 -c "import numpy as np; from sklearn.preprocessing import \
///   FunctionTransformer; import json; \
///   print(json.dumps(FunctionTransformer(np.expm1).transform(\
///   np.array([[0.,1.],[2.,3.]])).tolist()))"
/// -> [[0.0, 1.718281828459045], [6.38905609893065, 19.085536923187668]]
/// ```
/// np.expm1(x) = exp(x) - 1; Rust analog `f64::exp_m1`.
#[test]
fn guard_expm1_matches_sklearn_oracle() {
    // oracle: FunctionTransformer(np.expm1).transform([[0,1],[2,3]])
    let expected: [[f64; 2]; 2] = [
        [0.0, 1.718_281_828_459_045],
        [6.389_056_098_930_65, 19.085_536_923_187_668],
    ];

    let ft = FunctionTransformer::<f64>::new(|v: f64| v.exp_m1());
    let x = array![[0.0, 1.0], [2.0, 3.0]];
    let out = ft.transform(&x).unwrap();

    for ((r, c), &want) in expected
        .iter()
        .enumerate()
        .flat_map(|(i, row)| row.iter().enumerate().map(move |(j, v)| ((i, j), v)))
    {
        assert_eq!(
            out[[r, c]].to_bits(),
            want.to_bits(),
            "expm1 mismatch at [{r},{c}]: ferro={} sklearn={}",
            out[[r, c]],
            want
        );
    }
}

/// Green guard (REQ-1): element-wise `np.sqrt` ufunc cross-check.
///
/// Oracle (live sklearn 1.5.2, run from /tmp):
/// ```text
/// python3 -c "import numpy as np; from sklearn.preprocessing import \
///   FunctionTransformer; import json; \
///   print(json.dumps(FunctionTransformer(np.sqrt).transform(\
///   np.array([[1.,4.],[9.,16.]])).tolist()))"
/// -> [[1.0, 2.0], [3.0, 4.0]]
/// ```
#[test]
fn guard_sqrt_matches_sklearn_oracle() {
    // oracle: FunctionTransformer(np.sqrt).transform([[1,4],[9,16]])
    let expected: [[f64; 2]; 2] = [[1.0, 2.0], [3.0, 4.0]];

    let ft = FunctionTransformer::<f64>::new(|v: f64| v.sqrt());
    let x = array![[1.0, 4.0], [9.0, 16.0]];
    let out = ft.transform(&x).unwrap();

    for ((r, c), &want) in expected
        .iter()
        .enumerate()
        .flat_map(|(i, row)| row.iter().enumerate().map(move |(j, v)| ((i, j), v)))
    {
        assert_eq!(
            out[[r, c]].to_bits(),
            want.to_bits(),
            "sqrt mismatch at [{r},{c}]: ferro={} sklearn={}",
            out[[r, c]],
            want
        );
    }
}

/// Green guard (REQ-1): NaN / -inf propagation through `np.log` on a 0.0 and a
/// negative input matches sklearn (sklearn forwards the whole array to np.log;
/// `_transform`, `:375-379`).
///
/// Oracle (live sklearn 1.5.2, run from /tmp, np.seterr(all='ignore')):
/// ```text
/// python3 -c "import numpy as np; from sklearn.preprocessing import \
///   FunctionTransformer; import json; np.seterr(all='ignore'); \
///   print(json.dumps(FunctionTransformer(np.log).transform(\
///   np.array([[0.,1.],[-1.,2.]])).tolist()))"
/// -> [[-Infinity, 0.0], [NaN, 0.6931471805599453]]
/// ```
#[test]
#[allow(
    clippy::approx_constant,
    reason = "exact live-sklearn-1.5.2 oracle value ln(2)=log(2); R-CHAR-3, not the std constant"
)]
fn guard_log_nan_inf_propagation_matches_sklearn_oracle() {
    let ft = FunctionTransformer::<f64>::new(|v: f64| v.ln());
    let x = array![[0.0, 1.0], [-1.0, 2.0]];
    let out = ft.transform(&x).unwrap();

    // oracle: np.log(0.0) -> -inf
    assert!(
        out[[0, 0]].is_infinite() && out[[0, 0]].is_sign_negative(),
        "expected -inf at [0,0], got {}",
        out[[0, 0]]
    );
    // oracle: np.log(1.0) -> 0.0
    assert_eq!(out[[0, 1]].to_bits(), 0.0_f64.to_bits());
    // oracle: np.log(-1.0) -> NaN
    assert!(
        out[[1, 0]].is_nan(),
        "expected NaN at [1,0], got {}",
        out[[1, 0]]
    );
    // oracle: np.log(2.0) -> 0.6931471805599453
    assert_eq!(
        out[[1, 1]].to_bits(),
        0.693_147_180_559_945_3_f64.to_bits(),
        "ln(2) mismatch: {}",
        out[[1, 1]]
    );
}

/// Green guard (REQ-1): empty-matrix `(0, n)` shape preservation matches
/// sklearn (whole-array func on an empty array returns the same shape).
///
/// Oracle (live sklearn 1.5.2, run from /tmp):
/// ```text
/// python3 -c "import numpy as np; from sklearn.preprocessing import \
///   FunctionTransformer; \
///   print(FunctionTransformer(np.log1p).transform(np.zeros((0,3))).shape)"
/// -> (0, 3)
/// ```
#[test]
fn guard_empty_matrix_shape_matches_sklearn_oracle() {
    // oracle: transform(zeros((0,3))).shape == (0, 3)
    let expected_shape = [0_usize, 3];

    let ft = FunctionTransformer::<f64>::new(|v: f64| v.ln_1p());
    let x: Array2<f64> = Array2::zeros((0, 3));
    let out = ft.transform(&x).unwrap();

    assert_eq!(out.shape(), &expected_shape);
}