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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270

use std::num::NonZeroUsize;

use proptest::prelude::*;

use crate::diffusion::NoiseLevel;
use crate::eval::ClusterBootstrap;

use super::*;

fn level(gamma: f32) -> NoiseLevel {
    NoiseLevel::from_signal_variance(gamma).unwrap()
}

fn bootstrap() -> ClusterBootstrap {
    ClusterBootstrap::new(NonZeroUsize::new(256).unwrap(), 0.95, 7).unwrap()
}

/// A clean-target loss at `gamma` whose value is `loss` (a one-element
/// estimate/target pair with squared difference `loss`). Pass perfect squares for
/// an exact `loss`, since the estimate is `sqrt(loss)` and `f32` rounds otherwise.
fn loss_at(gamma: f32, loss: f32) -> CleanTargetLoss {
    let target = Tensor::vector(vec![0.0]);
    let estimate = Tensor::vector(vec![loss.sqrt()]);
    CleanTargetLoss::from_estimate(level(gamma), &estimate, &target).unwrap()
}

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

#[test]
fn estimate_equal_to_target_has_zero_loss() {
    let t = Tensor::new([2, 2], vec![1.0, -2.0, 3.0, 0.5]).unwrap();
    let l = CleanTargetLoss::from_estimate(level(0.5), &t, &t).unwrap();
    assert_eq!(l.loss(), 0.0);
    assert_eq!(l.level(), level(0.5));
}

#[test]
fn estimate_loss_matches_hand_computed_mse() {
    let estimate = Tensor::new([1, 2], vec![1.0, 4.0]).unwrap();
    let target = Tensor::new([1, 2], vec![0.0, 0.0]).unwrap();
    // mean of (1, 16) = 8.5
    let l = CleanTargetLoss::from_estimate(level(0.3), &estimate, &target).unwrap();
    assert!((l.loss() - 8.5).abs() < 1e-9);
}

#[test]
fn from_velocity_recovers_the_true_target_exactly() {
    // x_t = diffuse(x0, eps); v = the true velocity target. The recovery is exact,
    // so the clean-target loss is ~0 regardless of the noise level.
    let x0 = Tensor::new([2, 2], vec![0.7, -0.3, 0.4, 0.1]).unwrap();
    let eps = Tensor::new([2, 2], vec![-1.0, 0.5, 0.2, -0.8]).unwrap();
    for gamma in [0.05, 0.5, 0.95] {
        let lvl = level(gamma);
        let x_t = lvl.diffuse(&x0, &eps).unwrap();
        let v = lvl.velocity_target(&x0, &eps).unwrap();
        let l = CleanTargetLoss::from_velocity(lvl, &x_t, &v, &x0).unwrap();
        assert!(l.loss() < 1e-9, "gamma {gamma}: loss {}", l.loss());
    }
}

#[test]
fn high_snr_recovery_is_target_regardless_of_velocity() {
    // At gamma = 1 the noisy sample *is* the target, so any v_hat recovers it: the
    // structural reason the high-SNR bins are cheap and the unweighted objective
    // overspends there.
    let x0 = Tensor::new([1, 3], vec![0.2, -0.5, 0.9]).unwrap();
    let eps = Tensor::zeros([1, 3]).unwrap();
    let lvl = NoiseLevel::pure_signal();
    let x_t = lvl.diffuse(&x0, &eps).unwrap();
    let arbitrary_v = Tensor::new([1, 3], vec![5.0, -7.0, 3.0]).unwrap();
    let l = CleanTargetLoss::from_velocity(lvl, &x_t, &arbitrary_v, &x0).unwrap();
    assert!(l.loss() < 1e-12, "loss {}", l.loss());
}

#[test]
fn from_estimate_rejects_shape_mismatch() {
    let a = Tensor::new([1, 2], vec![0.0, 0.0]).unwrap();
    let b = Tensor::new([2, 1], vec![0.0, 0.0]).unwrap();
    assert!(matches!(
        CleanTargetLoss::from_estimate(level(0.5), &a, &b),
        Err(Error::Validation(_))
    ));
}

#[test]
fn from_estimate_rejects_non_finite_and_empty() {
    let target = Tensor::new([1, 2], vec![0.0, 0.0]).unwrap();
    let nan = Tensor::new([1, 2], vec![f32::NAN, 1.0]).unwrap();
    assert!(matches!(
        CleanTargetLoss::from_estimate(level(0.5), &nan, &target),
        Err(Error::Validation(_))
    ));
    let empty = Tensor::new([0, 2], vec![]).unwrap();
    assert!(matches!(
        CleanTargetLoss::from_estimate(level(0.5), &empty, &empty),
        Err(Error::Validation(_))
    ));
}

#[test]
fn curve_rejects_empty_losses() {
    assert!(matches!(
        PerSnrLossCurve::measure(&[], bins(4)),
        Err(Error::Validation(_))
    ));
}

#[test]
fn curve_assigns_losses_to_expected_bins() {
    // Perfect-square losses keep the `sqrt` round-trip exact (see `loss_at`).
    let losses = [
        loss_at(0.1, 1.0),  // bin 0: [0.00, 0.25)
        loss_at(0.3, 4.0),  // bin 1: [0.25, 0.50)
        loss_at(0.6, 9.0),  // bin 2: [0.50, 0.75)
        loss_at(0.9, 16.0), // bin 3: [0.75, 1.00)
        loss_at(1.0, 4.0),  // gamma == 1 folds into the top bin
    ];
    let curve = PerSnrLossCurve::measure(&losses, bins(4)).unwrap();
    assert_eq!(curve.observation_count(), 5);
    let b = curve.bins();
    assert_eq!(b.len(), 4);
    assert_eq!(b[0].count(), 1);
    assert_eq!(b[0].mean_loss(), Some(1.0));
    assert_eq!(b[1].mean_loss(), Some(4.0));
    assert_eq!(b[2].mean_loss(), Some(9.0));
    assert_eq!(b[3].count(), 2);
    assert_eq!(b[3].mean_loss(), Some(10.0)); // (16 + 4) / 2
}

#[test]
fn empty_bins_report_no_mean() {
    let curve = PerSnrLossCurve::measure(&[loss_at(0.6, 4.0)], bins(4)).unwrap();
    let b = curve.bins();
    assert_eq!(b[0].count(), 0);
    assert_eq!(b[0].mean_loss(), None);
    assert_eq!(b[2].mean_loss(), Some(4.0));
}

#[test]
fn bin_ranges_partition_the_unit_interval() {
    let curve = PerSnrLossCurve::measure(&[loss_at(0.5, 1.0)], bins(5)).unwrap();
    let b = curve.bins();
    assert_eq!(b[0].range().0, 0.0);
    assert!((b[4].range().1 - 1.0).abs() < 1e-6);
    for w in b.windows(2) {
        assert_eq!(w[0].range().1, w[1].range().0);
    }
}

#[test]
fn bin_intervals_center_on_bin_means() {
    // Two items, each scored at one gamma in bin 0 ([0, 0.25)) and one in bin 2
    // ([0.5, 0.75)). The interval is the pooled-mean bootstrap, so its point is the
    // bin's pooled `mean_loss`; with one loss per item per bin the clusters are
    // singletons, so the pooled and equal-weight statistics coincide either way.
    let per_item = vec![
        vec![loss_at(0.1, 1.0), loss_at(0.6, 4.0)],
        vec![loss_at(0.1, 9.0), loss_at(0.6, 16.0)],
    ];
    let curve = PerSnrLossCurve::measure(&per_item.concat(), bins(4)).unwrap();
    let intervals = PerSnrLossCurve::bin_intervals(&per_item, bins(4), &bootstrap()).unwrap();
    assert_eq!(intervals.len(), curve.bins().len());
    for (bin, ci) in curve.bins().iter().zip(&intervals) {
        match (bin.mean_loss(), ci) {
            (Some(mean), Some(ci)) => {
                assert!(
                    (ci.point() - mean).abs() < 1e-12,
                    "{} vs {mean}",
                    ci.point()
                );
                // Bounds stay inside the bin's own observation range.
                assert!(ci.lo() <= ci.point() && ci.point() <= ci.hi());
            }
            (None, None) => {}
            other => panic!("bin/interval populated-ness disagree: {other:?}"),
        }
    }
    // Bin 0 pools 1 and 9; bin 2 pools 4 and 16.
    assert_eq!(intervals[0].unwrap().point(), 5.0);
    assert_eq!(intervals[2].unwrap().point(), 10.0);
    assert!(intervals[1].is_none() && intervals[3].is_none());
}

#[test]
fn bin_intervals_center_on_pooled_mean_under_unequal_items() {
    // A conditioning item that segments into two windows contributes two losses to a
    // bin; another item contributes one. The interval must stay centered on the bin's
    // pooled `mean_loss` (observation-weighted), not the equal-weight mean of per-item
    // means, even though the clusters are now unequal in size. This is the centering
    // half of the EVAL-8 review fix: windows of one item cluster together, but the
    // headline weighting they bracket is pooled.
    // Perfect-square losses keep the `sqrt` round-trip exact (see `loss_at`).
    let per_item = vec![
        vec![loss_at(0.1, 1.0), loss_at(0.1, 1.0)], // one item, two windows -> bin 0
        vec![loss_at(0.1, 16.0)],                   // another item, one window -> bin 0
    ];
    let curve = PerSnrLossCurve::measure(&per_item.concat(), bins(4)).unwrap();
    let intervals = PerSnrLossCurve::bin_intervals(&per_item, bins(4), &bootstrap()).unwrap();
    // Pooled bin-0 mean is (1 + 1 + 16) / 3 = 6.0; the equal-weight mean of per-item
    // means would instead be ((1 + 1) / 2 + 16) / 2 = 8.5. The interval tracks the pooled.
    assert_eq!(curve.bins()[0].mean_loss(), Some(6.0));
    let ci = intervals[0].unwrap();
    assert!(
        (ci.point() - 6.0).abs() < 1e-12,
        "point {} should equal the pooled bin mean 6.0",
        ci.point()
    );
    assert!(ci.lo() <= ci.point() && ci.point() <= ci.hi());
    assert!(intervals[1].is_none() && intervals[2].is_none() && intervals[3].is_none());
}

#[test]
fn bin_intervals_reject_no_losses() {
    assert!(matches!(
        PerSnrLossCurve::bin_intervals(&[], bins(4), &bootstrap()),
        Err(Error::Validation(_))
    ));
    assert!(matches!(
        PerSnrLossCurve::bin_intervals(&[vec![], vec![]], bins(4), &bootstrap()),
        Err(Error::Validation(_))
    ));
}

proptest! {
    #[test]
    fn true_velocity_recovery_is_exact(
        x0 in proptest::collection::vec(-5.0f32..5.0, 4),
        eps in proptest::collection::vec(-5.0f32..5.0, 4),
        gamma in 0.0f32..=1.0,
    ) {
        let x0 = Tensor::new([2, 2], x0).unwrap();
        let eps = Tensor::new([2, 2], eps).unwrap();
        let lvl = level(gamma);
        let x_t = lvl.diffuse(&x0, &eps).unwrap();
        let v = lvl.velocity_target(&x0, &eps).unwrap();
        let l = CleanTargetLoss::from_velocity(lvl, &x_t, &v, &x0).unwrap();
        prop_assert!(l.loss() < 1e-4, "loss {} at gamma {gamma}", l.loss());
    }

    #[test]
    fn curve_conserves_observations(
        gammas in proptest::collection::vec(0.0f32..=1.0, 1..32),
        n in 1usize..8,
    ) {
        let losses: Vec<CleanTargetLoss> = gammas.iter().map(|&g| loss_at(g, 1.0)).collect();
        let curve = PerSnrLossCurve::measure(&losses, bins(n)).unwrap();
        prop_assert_eq!(curve.observation_count(), losses.len());
        prop_assert_eq!(curve.bins().len(), n);
    }

    #[test]
    fn bin_means_are_bounded_by_their_members(
        values in proptest::collection::vec((0.0f32..=1.0, 0.0f32..9.0), 1..24),
        n in 1usize..6,
    ) {
        let losses: Vec<CleanTargetLoss> =
            values.iter().map(|&(g, v)| loss_at(g, v)).collect();
        let curve = PerSnrLossCurve::measure(&losses, bins(n)).unwrap();
        for bin in curve.bins() {
            if let Some(mean) = bin.mean_loss() {
                // every observed loss is in [0, 9], so a bin mean must be too.
                prop_assert!((0.0..=9.0 + 1e-3).contains(&mean), "mean {mean}");
            } else {
                prop_assert_eq!(bin.count(), 0);
            }
        }
    }
}