segovia 0.4.0

A fast, chunked, memory-bounded Rust engine for electrophysiology (Neuropixels) signal processing, with Python bindings.
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

use nalgebra::{DMatrix, SymmetricEigen};
use numpy::ndarray::{Array1, Array2, Axis};

pub struct Whitener {
    w: Array2<f32>,
    mean: Option<Array1<f32>>,
}

impl Whitener {
    pub fn estimate(calib: &Array2<f32>, eps: f64, apply_mean: bool) -> Self {
        let n = calib.nrows();
        let c = calib.ncols();

        let calib64 = calib.mapv(|v| v as f64);
        let mean64 = if apply_mean {
            Some(calib64.mean_axis(Axis(0)).expect("non-empty calibration"))
        } else {
            None
        };

        let mut centered = calib64.clone();
        if let Some(m) = &mean64 {
            for mut row in centered.rows_mut() {
                row -= m;
            }
        }

        let cov = centered.t().dot(&centered) / (n as f64);

        let mut dm = DMatrix::<f64>::zeros(c, c);
        for i in 0..c {
            for j in 0..c {
                dm[(i, j)] = cov[[i, j]];
            }
        }

        let eig = SymmetricEigen::new(dm);
        let vals = &eig.eigenvalues;
        let vecs = &eig.eigenvectors;

        let mut d = DMatrix::<f64>::zeros(c, c);
        for i in 0..c {
            d[(i, i)] = 1.0 / (vals[i] + eps).sqrt();
        }
        let w_dm = vecs * d * vecs.transpose();

        let mut w = Array2::<f32>::zeros((c, c));
        for i in 0..c {
            for j in 0..c {
                w[[i, j]] = w_dm[(i, j)] as f32;
            }
        }
        let mean = mean64.map(|m| m.mapv(|v| v as f32));

        Self { w, mean }
    }

    pub fn apply(&self, chunk: &Array2<f32>) -> Array2<f32> {
        let s = chunk.nrows();
        let c = chunk.ncols();

        let mut centered = chunk.to_owned();
        if let Some(m) = &self.mean {
            for mut row in centered.rows_mut() {
                row -= m;
            }
        }

        let mut out = Array2::<f32>::zeros((s, c));
        let lhs = centered
            .as_slice()
            .expect("centered is row-major contiguous");
        let rhs = self.w.as_slice().expect("w is row-major contiguous");
        let dst = out.as_slice_mut().expect("out is row-major contiguous");
        let cs = c as isize;
        unsafe {
            gemm::gemm(
                s,
                c,
                c,
                dst.as_mut_ptr(),
                1,
                cs,
                false,
                lhs.as_ptr(),
                1,
                cs,
                rhs.as_ptr(),
                1,
                cs,
                0.0f32,
                1.0f32,
                false,
                false,
                false,
                gemm::Parallelism::None,
            );
        }
        out
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use numpy::ndarray::Array2;

    #[test]
    fn whitened_data_has_identity_covariance() {
        let n = 4000;
        let c = 3;
        let mut data = Array2::<f32>::zeros((n, c));
        let mut s: u64 = 1;
        let mut rng = || {
            s = s
                .wrapping_mul(6364136223846793005)
                .wrapping_add(1442695040888963407);
            ((s >> 33) as f64) / (1u64 << 31) as f64 - 1.0
        };
        for i in 0..n {
            let a = rng();
            let b = rng();
            let d = rng();
            data[[i, 0]] = (2.0 * a) as f32;
            data[[i, 1]] = (a + 0.5 * b) as f32;
            data[[i, 2]] = (a - b + 0.3 * d) as f32;
        }

        let w = Whitener::estimate(&data, 0.0, true);
        let out = w.apply(&data).mapv(|v| v as f64);

        let mean = out.mean_axis(Axis(0)).unwrap();
        let mut centered = out.clone();
        for mut row in centered.rows_mut() {
            row -= &mean;
        }
        let cov = centered.t().dot(&centered) / (n as f64);
        for i in 0..c {
            for j in 0..c {
                let expected = if i == j { 1.0 } else { 0.0 };
                assert!(
                    (cov[[i, j]] - expected).abs() < 1e-4,
                    "cov[{i},{j}] = {} expected {expected}",
                    cov[[i, j]]
                );
            }
        }
    }
}