datarust 0.6.5

Scikit-learn-style preprocessing and classical ML in Rust
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
206
207
208
209
210
211
212
213
214
215
216
217
218
219

use crate::error::{DatarustError, Result};
use crate::matrix::Matrix;
use crate::traits::{default_input_names, FeatureNames};
use crate::Transformer;

#[cfg(feature = "rayon")]
use rayon::prelude::*;

/// Norm kind for row-wise normalization, mirroring
/// `sklearn.preprocessing.Normalizer`.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
#[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
pub enum Norm {
    /// Normalize by the L1 (Manhattan) norm.
    L1,
    #[default]
    /// Normalize by the L2 (Euclidean) norm.
    L2,
    /// Normalize by the maximum absolute value.
    Max,
}

/// Normalize samples individually to unit norm (row-wise).
///
/// `Normalizer` is stateless — `fit` does nothing meaningful — matching
/// sklearn behavior.
#[derive(Debug, Clone)]
#[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
pub struct Normalizer {
    norm: Norm,
    n_features: usize,
    fitted: bool,
}

impl Normalizer {
    /// Creates a new normalizer using the given norm.
    pub fn new(norm: Norm) -> Self {
        Self {
            norm,
            n_features: 0,
            fitted: false,
        }
    }

    /// Returns the norm used for normalization.
    pub fn norm(&self) -> Norm {
        self.norm
    }

    fn scale_row(&self, row: &[f64]) -> Vec<f64> {
        match self.norm {
            Norm::L1 => {
                let s: f64 = row.iter().map(|v| v.abs()).sum();
                if s == 0.0 {
                    row.to_vec()
                } else {
                    row.iter().map(|v| v / s).collect()
                }
            }
            Norm::L2 => {
                let s: f64 = row.iter().map(|v| v * v).sum::<f64>().sqrt();
                if s == 0.0 {
                    row.to_vec()
                } else {
                    row.iter().map(|v| v / s).collect()
                }
            }
            Norm::Max => {
                let m = row.iter().fold(0.0_f64, |a, &v| a.max(v.abs()));
                if m == 0.0 {
                    row.to_vec()
                } else {
                    row.iter().map(|v| v / m).collect()
                }
            }
        }
    }
}

impl Default for Normalizer {
    fn default() -> Self {
        Self::new(Norm::default())
    }
}

impl FeatureNames for Normalizer {
    fn feature_names_out(&self, input_features: Option<&[String]>) -> Vec<String> {
        match input_features {
            Some(fs) => fs.to_vec(),
            None => default_input_names(self.n_features),
        }
    }
}

impl Transformer for Normalizer {
    fn name(&self) -> &'static str {
        "Normalizer"
    }

    fn fit(&mut self, x: &Matrix) -> Result<()> {
        x.validate_finite()?;
        self.n_features = x.ncols();
        self.fitted = true;
        Ok(())
    }

    fn transform(&self, x: &Matrix) -> Result<Matrix> {
        if !self.fitted {
            return Err(DatarustError::NotFitted("Normalizer".into()));
        }
        if x.ncols() != self.n_features {
            return Err(DatarustError::ShapeMismatch {
                expected: format!("{} features", self.n_features),
                actual: format!("{} features", x.ncols()),
            });
        }
        x.validate_finite()?;
        #[cfg(feature = "rayon")]
        {
            let out: Vec<Vec<f64>> = x.rows_ref().par_iter().map(|r| self.scale_row(r)).collect();
            Matrix::new(out)
        }
        #[cfg(not(feature = "rayon"))]
        {
            let out: Vec<Vec<f64>> = x.rows_ref().iter().map(|r| self.scale_row(r)).collect();
            Matrix::new(out)
        }
    }

    fn is_fitted(&self) -> bool {
        self.fitted
    }
}

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

    fn m1() -> Matrix {
        Matrix::new(vec![vec![3.0, 4.0], vec![1.0, 2.0], vec![0.0, 0.0]]).unwrap()
    }

    #[test]
    fn l2_row_norm_is_one() {
        let mut n = Normalizer::new(Norm::L2);
        let out = n.fit_transform(&m1()).unwrap();
        // 3,4 -> 0.6, 0.8 ; norm = 1
        assert!((out.get(0, 0) - 0.6).abs() < 1e-12);
        assert!((out.get(0, 1) - 0.8).abs() < 1e-12);
        for i in 0..2 {
            let r = out.row(i);
            let nrm: f64 = r.iter().map(|v| v * v).sum::<f64>().sqrt();
            assert!((nrm - 1.0).abs() < 1e-9);
        }
    }

    #[test]
    fn l1_sum_abs_one() {
        let mut n = Normalizer::new(Norm::L1);
        let out = n.fit_transform(&m1()).unwrap();
        // 3,4 -> 3/7, 4/7
        assert!((out.get(0, 0) - 3.0 / 7.0).abs() < 1e-12);
        assert!((out.get(0, 1) - 4.0 / 7.0).abs() < 1e-12);
        for i in 0..2 {
            let s: f64 = out.row(i).iter().map(|v| v.abs()).sum();
            assert!((s - 1.0).abs() < 1e-9);
        }
    }

    #[test]
    fn max_norm() {
        let mut n = Normalizer::new(Norm::Max);
        let out = n.fit_transform(&m1()).unwrap();
        // 3,4 -> 0.75, 1.0
        assert!((out.get(0, 0) - 0.75).abs() < 1e-12);
        assert!((out.get(0, 1) - 1.0).abs() < 1e-12);
        for i in 0..2 {
            let m = out.row(i).iter().fold(0.0_f64, |a, &v| a.max(v.abs()));
            assert!((m - 1.0).abs() < 1e-9);
        }
    }

    #[test]
    fn zero_row_passthrough() {
        let mut n = Normalizer::new(Norm::L2);
        let out = n.fit_transform(&m1()).unwrap();
        // zero row stays zero
        assert!((out.get(2, 0) - 0.0).abs() < 1e-12);
        assert!((out.get(2, 1) - 0.0).abs() < 1e-12);
    }

    #[test]
    fn transform_before_fit_errors() {
        let n = Normalizer::new(Norm::L2);
        assert!(matches!(
            n.transform(&m1()),
            Err(DatarustError::NotFitted(_))
        ));
    }

    #[test]
    fn negative_values_l2() {
        let x = Matrix::new(vec![vec![-3.0, 4.0]]).unwrap();
        let mut n = Normalizer::new(Norm::L2);
        let out = n.fit_transform(&x).unwrap();
        assert!((out.get(0, 0) - (-0.6)).abs() < 1e-12);
        assert!((out.get(0, 1) - 0.8).abs() < 1e-12);
    }

    #[test]
    fn independent_of_other_rows() {
        // Adding an extreme row does not change other rows' normalization.
        let x = Matrix::new(vec![vec![3.0, 4.0], vec![1000.0, 2000.0]]).unwrap();
        let mut n = Normalizer::new(Norm::L2);
        let out = n.fit_transform(&x).unwrap();
        assert!((out.get(0, 0) - 0.6).abs() < 1e-12);
        assert!((out.get(0, 1) - 0.8).abs() < 1e-12);
    }
}