helena 0.1.0

Core types and component interfaces for helena, a latent data-to-waveform generation platform.
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

use std::num::NonZeroUsize;

use proptest::prelude::*;

use super::*;

fn steps(n: usize) -> NonZeroUsize {
    NonZeroUsize::new(n).unwrap()
}

/// A `conds x seeds` grid of `elems`-wide cells filled row-major from `vals`.
fn grid(conds: usize, seeds: usize, elems: usize, vals: &[f32]) -> Vec<Vec<Latent<Continuous>>> {
    assert_eq!(vals.len(), conds * seeds * elems);
    let mut k = 0;
    (0..conds)
        .map(|_| {
            (0..seeds)
                .map(|_| {
                    let cell = latent([elems, 1], vals[k..k + elems].to_vec());
                    k += elems;
                    cell
                })
                .collect()
        })
        .collect()
}

#[test]
fn identical_grids_have_zero_distance() {
    let g = grid(2, 2, 2, &[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0]);
    assert_eq!(grid_distance(&g, &g).unwrap(), 0.0);
}

#[test]
fn distance_matches_hand_computed_mean() {
    // One condition, two seeds, one element. Cell distances: |1-2| and |3-7| = 1, 4.
    let reference = grid(1, 2, 1, &[1.0, 3.0]);
    let candidate = grid(1, 2, 1, &[2.0, 7.0]);
    // mean of (1, 4) = 2.5
    assert!((grid_distance(&reference, &candidate).unwrap() - 2.5).abs() < 1e-6);
}

#[test]
fn rejects_empty_reference() {
    let candidate = grid(1, 1, 1, &[0.0]);
    assert!(matches!(
        grid_distance(&[], &candidate),
        Err(Error::Validation(_))
    ));
}

#[test]
fn rejects_condition_count_mismatch() {
    let reference = grid(2, 1, 1, &[1.0, 2.0]);
    let candidate = grid(1, 1, 1, &[1.0]);
    assert!(matches!(
        grid_distance(&reference, &candidate),
        Err(Error::Validation(_))
    ));
}

#[test]
fn rejects_seed_budget_mismatch() {
    let reference = grid(1, 2, 1, &[1.0, 2.0]);
    let candidate = grid(1, 1, 1, &[1.0]);
    assert!(matches!(
        grid_distance(&reference, &candidate),
        Err(Error::Validation(_))
    ));
}

#[test]
fn rejects_cell_shape_mismatch() {
    let reference = grid(1, 1, 2, &[1.0, 2.0]);
    let candidate = grid(1, 1, 1, &[1.0]);
    assert!(matches!(
        grid_distance(&reference, &candidate),
        Err(Error::Validation(_))
    ));
}

#[test]
fn rejects_ragged_grids_even_when_pairwise_consistent() {
    // Cell (0,0) is [1,1] in both grids and cell (1,0) is [2,1] in both, so each pair
    // matches; but the grid-wide shape differs, which would average incomparable L2
    // scales. The grid-wide shape check rejects it.
    let reference = vec![
        vec![latent([1, 1], vec![0.0])],
        vec![latent([2, 1], vec![0.0, 0.0])],
    ];
    let candidate = vec![
        vec![latent([1, 1], vec![1.0])],
        vec![latent([2, 1], vec![1.0, 1.0])],
    ];
    assert!(matches!(
        grid_distance(&reference, &candidate),
        Err(Error::Validation(_))
    ));
}

// `rejects_tokens_and_non_finite` is gone: a grid of `Latent<Continuous>` cannot
// hold tokens, and a `FrameSeq` cell cannot hold a NaN, so both rejections are
// unrepresentable.

#[test]
fn curve_orders_gaps_and_scores_each_budget() {
    let reference = grid(1, 1, 1, &[0.0]);
    let near = grid(1, 1, 1, &[1.0]);
    let far = grid(1, 1, 1, &[4.0]);
    // Passed out of order; the curve sorts ascending by step budget.
    let curve = StepGapCurve::measure(
        steps(64),
        &reference,
        &[(steps(16), &near), (steps(4), &far)],
    )
    .unwrap();
    assert_eq!(curve.reference_steps(), steps(64));
    let g = curve.gaps();
    assert_eq!(g[0].steps(), steps(4));
    assert_eq!(g[0].gap(), 4.0);
    assert_eq!(g[1].steps(), steps(16));
    assert_eq!(g[1].gap(), 1.0);
    // Fewer steps -> larger gap -> the curve converges as the budget grows.
    assert!(curve.is_converging());
}

#[test]
fn curve_detects_non_convergence() {
    let reference = grid(1, 1, 1, &[0.0]);
    let far = grid(1, 1, 1, &[4.0]);
    let near = grid(1, 1, 1, &[1.0]);
    // Larger budget is *further* from the reference: not converging.
    let curve = StepGapCurve::measure(
        steps(64),
        &reference,
        &[(steps(4), &near), (steps(16), &far)],
    )
    .unwrap();
    assert!(!curve.is_converging());
}

#[test]
fn curve_rejects_empty_and_duplicate_budgets() {
    let reference = grid(1, 1, 1, &[0.0]);
    assert!(matches!(
        StepGapCurve::measure(steps(64), &reference, &[]),
        Err(Error::Validation(_))
    ));
    let a = grid(1, 1, 1, &[1.0]);
    let b = grid(1, 1, 1, &[2.0]);
    assert!(matches!(
        StepGapCurve::measure(steps(64), &reference, &[(steps(4), &a), (steps(4), &b)]),
        Err(Error::Validation(_))
    ));
}

#[test]
fn cell_distances_are_clustered_by_condition() {
    // Two conditions, two seeds. Per-cell |ref - cand|: cond 0 -> (1, 4), cond 1 -> (2, 2).
    let reference = grid(2, 2, 1, &[0.0, 0.0, 0.0, 0.0]);
    let candidate = grid(2, 2, 1, &[1.0, 4.0, 2.0, 2.0]);
    let clusters = grid_cell_distances(&reference, &candidate).unwrap();
    assert_eq!(clusters.len(), 2, "one cluster per condition");
    assert_eq!(clusters[0], vec![1.0, 4.0]);
    assert_eq!(clusters[1], vec![2.0, 2.0]);
    // `grid_distance` is exactly the pooled mean of these cells: (1 + 4 + 2 + 2) / 4.
    assert!((grid_distance(&reference, &candidate).unwrap() - 2.25).abs() < 1e-6);
}

#[test]
fn cell_distances_share_grid_distance_validation() {
    // The two share one core, so a misalignment rejected by `grid_distance` is rejected
    // here too (condition-count mismatch stands in for the whole boundary).
    let reference = grid(2, 1, 1, &[1.0, 2.0]);
    let candidate = grid(1, 1, 1, &[1.0]);
    assert!(matches!(
        grid_cell_distances(&reference, &candidate),
        Err(Error::Validation(_))
    ));
}

proptest! {
    #[test]
    fn self_distance_is_zero(
        conds in 1usize..3,
        seeds in 1usize..4,
        elems in 1usize..4,
        seed in -10.0f32..10.0,
    ) {
        let vals: Vec<f32> = (0..conds * seeds * elems)
            .map(|i| seed + i as f32)
            .collect();
        let g = grid(conds, seeds, elems, &vals);
        prop_assert_eq!(grid_distance(&g, &g).unwrap(), 0.0);
    }

    #[test]
    fn distance_is_symmetric_and_nonnegative(
        conds in 1usize..3,
        seeds in 1usize..3,
        elems in 1usize..3,
        a in proptest::collection::vec(-8.0f32..8.0, 18),
        b in proptest::collection::vec(-8.0f32..8.0, 18),
    ) {
        let n = conds * seeds * elems;
        let ga = grid(conds, seeds, elems, &a[..n]);
        let gb = grid(conds, seeds, elems, &b[..n]);
        let ab = grid_distance(&ga, &gb).unwrap();
        let ba = grid_distance(&gb, &ga).unwrap();
        prop_assert!(ab >= 0.0);
        prop_assert!((ab - ba).abs() <= 1e-4 * ab.max(1.0), "{ab} vs {ba}");
    }
}