rill-ml 0.8.0

Lightweight, serializable online machine learning for Rust applications and streaming data.
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
220
221
222
223
224
225
226

//! Mean imputer for missing data.
//!
//! Replaces `NaN` values with the running mean of observed non-NaN
//! values for each feature. Uses Welford's algorithm for numerical
//! stability.

use crate::error::{RillError, checked_finite_add, checked_increment, ensure_finite};
use crate::traits::Transformer;

/// Replaces `NaN` values with the per-feature running mean.
///
/// When a feature has seen zero non-NaN values, `NaN` is replaced
/// with `0.0`. This transformer accepts `NaN` in its input.
#[derive(Debug, Clone)]
#[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
pub struct MeanImputer {
    feature_count: usize,
    counts: Vec<u64>,
    means: Vec<f64>,
    samples_seen: u64,
}

impl MeanImputer {
    /// Create a new imputer for `feature_count` features.
    ///
    /// # Errors
    /// Returns [`RillError::EmptyFeatures`] if `feature_count` is `0`.
    pub fn new(feature_count: usize) -> Result<Self, RillError> {
        if feature_count == 0 {
            return Err(RillError::EmptyFeatures);
        }
        Ok(Self {
            feature_count,
            counts: vec![0; feature_count],
            means: vec![0.0; feature_count],
            samples_seen: 0,
        })
    }

    /// The per-feature running means of observed non-NaN values.
    pub fn means(&self) -> &[f64] {
        &self.means
    }

    /// The per-feature counts of observed non-NaN values.
    pub fn counts(&self) -> &[u64] {
        &self.counts
    }

    /// Validate only the dimension, allowing NaN values.
    fn check_dimension(&self, features: &[f64]) -> Result<(), RillError> {
        if features.is_empty() {
            return Err(RillError::EmptyFeatures);
        }
        if features.len() != self.feature_count {
            return Err(RillError::DimensionMismatch {
                expected: self.feature_count,
                actual: features.len(),
            });
        }
        Ok(())
    }

    /// Update the running mean for feature `idx` using Welford's algorithm.
    fn update_mean(&mut self, idx: usize, value: f64) -> Result<(), RillError> {
        let n = checked_increment(self.counts[idx], "feature count")?;
        self.counts[idx] = n;
        let delta = value - self.means[idx];
        ensure_finite("mean delta", delta)?;
        self.means[idx] = checked_finite_add(self.means[idx], delta / n as f64, "mean")?;
        Ok(())
    }
}

impl Transformer for MeanImputer {
    fn input_dim(&self) -> usize {
        self.feature_count
    }

    fn output_dim(&self) -> usize {
        self.feature_count
    }

    fn transform(&self, features: &[f64]) -> Result<Vec<f64>, RillError> {
        self.check_dimension(features)?;
        let mut out = Vec::with_capacity(features.len());
        for (i, &x) in features.iter().enumerate() {
            if x.is_nan() {
                out.push(if self.counts[i] == 0 {
                    0.0
                } else {
                    self.means[i]
                });
            } else {
                ensure_finite("feature", x)?;
                out.push(x);
            }
        }
        Ok(out)
    }

    fn update(&mut self, features: &[f64]) -> Result<(), RillError> {
        self.check_dimension(features)?;
        for (i, &x) in features.iter().enumerate() {
            if x.is_nan() {
                continue;
            }
            ensure_finite("feature", x)?;
            self.update_mean(i, x)?;
        }
        self.samples_seen = checked_increment(self.samples_seen, "samples_seen")?;
        Ok(())
    }

    fn samples_seen(&self) -> u64 {
        self.samples_seen
    }

    fn reset(&mut self) {
        for c in &mut self.counts {
            *c = 0;
        }
        for m in &mut self.means {
            *m = 0.0;
        }
        self.samples_seen = 0;
    }
}

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

    #[test]
    fn nan_replaced_with_mean_after_update() {
        let mut imp = MeanImputer::new(2).unwrap();
        // feature 0: observed [2.0, 4.0] -> mean 3.0
        // feature 1: observed [10.0] -> mean 10.0
        imp.update(&[2.0, 10.0]).unwrap();
        imp.update(&[4.0, f64::NAN]).unwrap();
        let out = imp.transform(&[f64::NAN, f64::NAN]).unwrap();
        assert!((out[0] - 3.0).abs() < 1e-12);
        assert!((out[1] - 10.0).abs() < 1e-12);
    }

    #[test]
    fn nan_replaced_with_zero_when_no_data() {
        let imp = MeanImputer::new(2).unwrap();
        let out = imp.transform(&[f64::NAN, f64::NAN]).unwrap();
        assert_eq!(out, vec![0.0, 0.0]);
    }

    #[test]
    fn non_nan_passed_through() {
        let mut imp = MeanImputer::new(2).unwrap();
        imp.update(&[5.0, 6.0]).unwrap();
        let out = imp.transform(&[1.5, -2.0]).unwrap();
        assert_eq!(out, vec![1.5, -2.0]);
    }

    #[test]
    fn mean_updates_correctly() {
        let mut imp = MeanImputer::new(1).unwrap();
        imp.update(&[1.0]).unwrap();
        imp.update(&[2.0]).unwrap();
        imp.update(&[3.0]).unwrap();
        assert!((imp.means()[0] - 2.0).abs() < 1e-12);
        assert_eq!(imp.counts()[0], 3);
    }

    #[test]
    fn nan_skipped_in_update() {
        let mut imp = MeanImputer::new(2).unwrap();
        // feature 0: [1.0, NaN, 3.0] -> mean 2.0 (NaN skipped)
        // feature 1: [NaN, NaN, NaN] -> count 0, mean 0.0
        imp.update(&[1.0, f64::NAN]).unwrap();
        imp.update(&[f64::NAN, f64::NAN]).unwrap();
        imp.update(&[3.0, f64::NAN]).unwrap();
        assert!((imp.means()[0] - 2.0).abs() < 1e-12);
        assert_eq!(imp.counts()[0], 2);
        assert_eq!(imp.counts()[1], 0);
        assert!((imp.means()[1] - 0.0).abs() < 1e-12);
    }

    #[test]
    fn dimension_mismatch_rejected() {
        let imp = MeanImputer::new(3).unwrap();
        assert!(matches!(
            imp.transform(&[1.0, 2.0]),
            Err(RillError::DimensionMismatch { .. })
        ));
        let mut imp = imp;
        assert!(matches!(
            imp.update(&[1.0, 2.0, 3.0, 4.0]),
            Err(RillError::DimensionMismatch { .. })
        ));
    }

    #[test]
    fn reset_clears_state() {
        let mut imp = MeanImputer::new(2).unwrap();
        imp.update(&[1.0, 2.0]).unwrap();
        imp.update(&[3.0, 4.0]).unwrap();
        assert_eq!(imp.samples_seen(), 2);
        assert_eq!(imp.counts()[0], 2);
        imp.reset();
        assert_eq!(imp.samples_seen(), 0);
        assert_eq!(imp.counts()[0], 0);
        assert!((imp.means()[0] - 0.0).abs() < 1e-12);
    }

    #[test]
    #[cfg(feature = "serde")]
    fn serde_roundtrip() {
        let mut imp = MeanImputer::new(2).unwrap();
        imp.update(&[1.0, f64::NAN]).unwrap();
        imp.update(&[3.0, 5.0]).unwrap();
        let json = serde_json::to_string(&imp).unwrap();
        let restored: MeanImputer = serde_json::from_str(&json).unwrap();
        assert_eq!(restored.input_dim(), imp.input_dim());
        assert_eq!(restored.output_dim(), imp.output_dim());
        assert_eq!(restored.samples_seen(), imp.samples_seen());
        assert_eq!(restored.counts(), imp.counts());
        assert!((restored.means()[0] - imp.means()[0]).abs() < 1e-12);
    }
}