rust_trainer 0.1.4

CPU-first pure-Rust supervised trainer for Selective State Space Models with Hyperspherical Prototype Networks.
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

use ndarray::{Array1, Array2, Array3, ArrayView2, ArrayView3};

const LN_EPS: f32 = 1e-5;
const NORM_EPS: f32 = 1e-8;

#[derive(Debug, Clone)]
pub struct LnCache {
    pub y_hat: Array3<f32>,
    pub inv_std: Array2<f32>,
}

pub fn layer_norm_forward(x: ArrayView3<f32>) -> (Array3<f32>, LnCache) {
    let (b, t, d) = (x.shape()[0], x.shape()[1], x.shape()[2]);
    let mut y = Array3::<f32>::zeros((b, t, d));
    let mut inv_std = Array2::<f32>::zeros((b, t));
    let dn = d as f32;
    for bi in 0..b {
        for ti in 0..t {
            let mut mean = 0.0;
            for di in 0..d {
                mean += x[(bi, ti, di)];
            }
            mean /= dn;
            let mut var = 0.0;
            for di in 0..d {
                let v = x[(bi, ti, di)] - mean;
                var += v * v;
            }
            var /= dn;
            let inv = 1.0 / (var + LN_EPS).sqrt();
            inv_std[(bi, ti)] = inv;
            for di in 0..d {
                y[(bi, ti, di)] = (x[(bi, ti, di)] - mean) * inv;
            }
        }
    }
    (y.clone(), LnCache { y_hat: y, inv_std })
}

pub fn layer_norm_backward(dy: ArrayView3<f32>, cache: &LnCache) -> Array3<f32> {
    let (b, t, d) = (dy.shape()[0], dy.shape()[1], dy.shape()[2]);
    let mut dx = Array3::<f32>::zeros((b, t, d));
    let dn = d as f32;
    for bi in 0..b {
        for ti in 0..t {
            let inv = cache.inv_std[(bi, ti)];
            let mut mean_dy = 0.0;
            let mut mean_dy_y = 0.0;
            for di in 0..d {
                let dyi = dy[(bi, ti, di)];
                mean_dy += dyi;
                mean_dy_y += dyi * cache.y_hat[(bi, ti, di)];
            }
            mean_dy /= dn;
            mean_dy_y /= dn;
            for di in 0..d {
                let dyi = dy[(bi, ti, di)];
                dx[(bi, ti, di)] = inv * (dyi - mean_dy - cache.y_hat[(bi, ti, di)] * mean_dy_y);
            }
        }
    }
    dx
}

pub fn hpn_loss_and_grad_z(
    z_flat: ArrayView2<f32>,
    targets: &[i64],
    prototypes: &Array2<f32>,
) -> (f32, Array2<f32>) {
    let (loss, dz, _d_proto) = hpn_loss_and_grads(z_flat, targets, prototypes);
    (loss, dz)
}

pub fn hpn_loss_and_grads(
    z_flat: ArrayView2<f32>,
    targets: &[i64],
    prototypes: &Array2<f32>,
) -> (f32, Array2<f32>, Array2<f32>) {
    let (n, d) = (z_flat.shape()[0], z_flat.shape()[1]);
    assert_eq!(targets.len(), n);
    let k = prototypes.shape()[0];

    let mut z_norm = Array2::<f32>::zeros((n, d));
    let mut z_invnorm = Array1::<f32>::zeros(n);
    let mut cos = Array1::<f32>::zeros(n);

    for i in 0..n {
        let mut sq = 0.0;
        for di in 0..d {
            let v = z_flat[(i, di)];
            sq += v * v;
        }
        let nrm = sq.sqrt().max(NORM_EPS);
        let inv = 1.0 / nrm;
        z_invnorm[i] = inv;
        let yi = targets[i].rem_euclid(k as i64) as usize;
        let mut c = 0.0;
        for di in 0..d {
            z_norm[(i, di)] = z_flat[(i, di)] * inv;
            c += z_norm[(i, di)] * prototypes[(yi, di)];
        }
        cos[i] = c;
    }

    let mut loss = 0.0;
    for &c in cos.iter() {
        let r = 1.0 - c;
        loss += r * r;
    }
    loss /= n as f32;

    let mut dz_norm = Array2::<f32>::zeros((n, d));
    let mut d_prototypes = Array2::<f32>::zeros((k, d));
    let coeff = -2.0 / (n as f32);
    for i in 0..n {
        let dc = coeff * (1.0 - cos[i]);
        let yi = targets[i].rem_euclid(k as i64) as usize;
        for di in 0..d {
            dz_norm[(i, di)] = dc * prototypes[(yi, di)];
            d_prototypes[(yi, di)] += dc * z_norm[(i, di)];
        }
    }

    let mut dz = Array2::<f32>::zeros((n, d));
    for i in 0..n {
        let mut dot = 0.0;
        for di in 0..d {
            dot += dz_norm[(i, di)] * z_norm[(i, di)];
        }
        let inv = z_invnorm[i];
        for di in 0..d {
            dz[(i, di)] = inv * (dz_norm[(i, di)] - dot * z_norm[(i, di)]);
        }
    }

    (loss, dz, d_prototypes)
}

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

    #[test]
    fn layer_norm_backward_preserves_shape_and_finiteness() {
        let x = Array3::from_shape_fn((2, 3, 4), |(b, t, d)| 0.1 * (1 + b + t + d) as f32);
        let dy = Array3::from_shape_fn((2, 3, 4), |(b, t, d)| 0.05 * (1 + b + t + d) as f32);
        let (_y, cache) = layer_norm_forward(x.view());
        let dx = layer_norm_backward(dy.view(), &cache);
        assert_eq!(dx.dim(), (2, 3, 4));
        assert!(dx.iter().all(|v| v.is_finite()));
    }

    #[test]
    fn hpn_loss_and_grad_returns_finite_outputs() {
        let z = Array2::from_shape_fn((6, 8), |(n, d)| 0.03 * (1 + n + d) as f32);
        let prototypes = Array2::from_shape_fn((10, 8), |(k, d)| 0.02 * (1 + k + d) as f32);
        let targets = vec![0, 1, 2, 3, 4, 5];
        let (loss, dz) = hpn_loss_and_grad_z(z.view(), &targets, &prototypes);
        assert!(loss.is_finite());
        assert_eq!(dz.dim(), (6, 8));
        assert!(dz.iter().all(|v| v.is_finite()));

        let (loss2, dz2, dproto) = hpn_loss_and_grads(z.view(), &targets, &prototypes);
        assert!((loss - loss2).abs() <= 1e-8);
        assert_eq!(dz2.dim(), (6, 8));
        assert_eq!(dproto.dim(), (10, 8));
        assert!(dproto.iter().all(|v| v.is_finite()));
    }
}