gras 2.0.1

A Genetic Programming framework for Neural Architecture Search in Rust
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
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
//! Scoring helpers — convenience functions for fitness metrics.

use flodl::Variable;
use flodl::tensor::Result;

/// `1 − SS_res/SS_tot` — fraction of target variance explained (higher = better; `1.0` = perfect).
pub fn r2_score(pred: &Variable, y: &Variable) -> Result<f32> {
    let t = y.data().to_f32_vec()?;
    let p = pred.data().to_f32_vec()?;
    let n = t.len() as f32;
    if n == 0.0 {
        return Ok(0.0);
    }
    let mean_t = t.iter().sum::<f32>() / n;
    let ss_tot: f32 = t.iter().map(|&v| (v - mean_t).powi(2)).sum();
    let ss_res: f32 = t.iter().zip(&p).map(|(&a, &b)| (a - b).powi(2)).sum();
    if ss_tot.abs() < f32::EPSILON {
        return Ok(if ss_res.abs() < f32::EPSILON {
            1.0
        } else {
            0.0
        });
    }
    Ok(1.0 - ss_res / ss_tot)
}

/// Argmax class index per row, as plain `Vec<i64>`.
pub fn argmax_classes(pred: &Variable, y: &Variable) -> Result<(Vec<i64>, Vec<i64>)> {
    let pa = pred.data().argmax(1, false)?.to_i64_vec()?;
    let ta = y.data().argmax(1, false)?.to_i64_vec()?;
    Ok((pa, ta))
}

/// Fraction of rows whose argmax class matches the target (higher = better).
pub fn accuracy_score(pred: &Variable, y: &Variable) -> Result<f32> {
    let (pa, ta) = argmax_classes(pred, y)?;
    let n = pa.len() as f32;
    if n == 0.0 {
        return Ok(0.0);
    }
    Ok(pa.iter().zip(&ta).filter(|(a, b)| a == b).count() as f32 / n)
}

/// Macro F1 from argmax class vectors (per-class F1 averaged over classes).
pub fn f1_score(pred: &Variable, y: &Variable) -> Result<f32> {
    let (pa, ta) = argmax_classes(pred, y)?;
    let classes = y.data().shape()[1] as usize;
    // eli5: average the F1 score of each class, where F1 = 2 * precision * recall / (precision + recall)
    Ok(f1_from_vecs(&pa, &ta, classes))
}

/// Macro F1 from precomputed argmax vectors.
pub fn f1_from_vecs(pa: &[i64], ta: &[i64], classes: usize) -> f32 {
    if classes == 0 || pa.is_empty() {
        return 0.0;
    }
    let mut sum = 0.0f32;
    for c in 0..classes as i64 {
        let mut tp = 0usize;
        let mut fp = 0usize;
        let mut fn_ = 0usize;
        for (p, t) in pa.iter().zip(ta.iter()) {
            match (*p == c, *t == c) {
                (true, true) => tp += 1,
                (true, false) => fp += 1,
                (false, true) => fn_ += 1,
                (false, false) => {}
            }
        }
        let denom = (tp + fp + tp + fn_) as f32;
        if denom > 0.0 {
            sum += 2.0 * tp as f32 / denom;
        }
    }
    sum / classes as f32
}

/// Macro precision from argmax class vectors.
pub fn precision_score(pred: &Variable, y: &Variable) -> Result<f32> {
    let (pa, ta) = argmax_classes(pred, y)?;
    let classes = y.data().shape()[1] as usize;
    // eli5: average the precision of each class, where precision = TP / (TP + FP)
    Ok(precision_from_vecs(&pa, &ta, classes))
}

/// Macro RECALL from argmax class vectors (higher = better).
pub fn recall_from_vecs(pa: &[i64], ta: &[i64], classes: usize) -> f32 {
    if classes == 0 || pa.is_empty() {
        return 0.0;
    }
    let mut sum = 0.0f32;
    let mut present = 0usize;
    for c in 0..classes as i64 {
        let (mut tp, mut fn_) = (0usize, 0usize);
        for (p, t) in pa.iter().zip(ta.iter()) {
            if *t == c {
                if *p == c {
                    tp += 1;
                } else {
                    fn_ += 1;
                }
            }
        }
        if tp + fn_ > 0 {
            sum += tp as f32 / (tp + fn_) as f32;
            present += 1;
        }
    }
    if present == 0 {
        0.0
    } else {
        sum / present as f32
    }
}

/// Macro recall (higher = better) — [`recall_from_vecs`] over argmax vectors.
pub fn recall_score(pred: &Variable, y: &Variable) -> Result<f32> {
    let (pa, ta) = argmax_classes(pred, y)?;
    let classes = y.data().shape()[1] as usize;
    Ok(recall_from_vecs(&pa, &ta, classes))
}

/// BALANCED accuracy: the mean of per-class `(TP + TN) / n` over the classes the TARGET contains
/// (higher = better).
pub fn balanced_accuracy_score(pred: &Variable, y: &Variable) -> Result<f32> {
    let (pa, ta) = argmax_classes(pred, y)?;
    let classes = y.data().shape()[1] as usize;
    if classes == 0 || pa.is_empty() {
        return Ok(0.0);
    }
    let n = pa.len() as f32;
    let mut sum = 0.0f32;
    let mut present = 0usize;
    for c in 0..classes as i64 {
        let mut present_in_target = false;
        let mut correct = 0usize;
        for (p, t) in pa.iter().zip(ta.iter()) {
            // Per-class accuracy counts EVERY row: right for this class (TP) plus
            // right for another one while the label is not this class (TN).
            if (*p == c) == (*t == c) {
                correct += 1;
            }
            if *t == c {
                present_in_target = true;
            }
        }
        if present_in_target {
            sum += correct as f32 / n;
            present += 1;
        }
    }
    Ok(if present == 0 {
        0.0
    } else {
        sum / present as f32
    })
}

/// TOP-K accuracy: the share of rows whose TARGET class is among the net's `k` highest-scoring
/// classes (higher = better).
pub fn top_k_accuracy_score(pred: &Variable, y: &Variable, k: usize) -> Result<f32> {
    let shape = pred.data().shape();
    let rows = shape[0] as usize;
    let classes = shape[1] as usize;
    if rows == 0 || classes == 0 {
        return Ok(0.0);
    }
    let k = k.clamp(1, classes);
    let (_, ta) = argmax_classes(pred, y)?;
    let p = pred.data().to_f32_vec()?;
    let mut hits = 0usize;
    for r in 0..rows {
        let truth = ta.get(r).copied().unwrap_or(-1);
        if truth < 0 || truth as usize >= classes {
            continue;
        }
        let row = &p[r * classes..(r + 1) * classes];
        let target = row[truth as usize];
        let strictly_better = row.iter().filter(|v| **v > target).count();
        if strictly_better < k {
            hits += 1;
        }
    }
    Ok(hits as f32 / rows as f32)
}

/// AUROC — area under the ROC curve for BINARY classification (higher = better; `0.5` = a coin
/// flip).
pub fn auroc_score(pred: &Variable, y: &Variable) -> Result<f32> {
    let p_shape = pred.data().shape().to_vec();
    let t_shape = y.data().shape().to_vec();
    let rows = p_shape[0] as usize;
    if rows == 0 {
        return Ok(0.5);
    }
    let p_cols = *p_shape.get(1).unwrap_or(&1) as usize;
    let t_cols = *t_shape.get(1).unwrap_or(&1) as usize;
    let p = pred.data().to_f32_vec()?;
    let t = y.data().to_f32_vec()?;
    // Positive-class score per row: the class-1 column when there is one, else
    // only column. NaN sorts to the end under `total_cmp` — deterministic either way.
    let mut pairs: Vec<(f32, bool)> = (0..rows)
        .map(|r| {
            let score = if p_cols >= 2 {
                p[r * p_cols + 1]
            } else {
                p[r * p_cols]
            };
            let positive = if t_cols >= 2 {
                t[r * t_cols + 1] > 0.5
            } else {
                t[r * t_cols] > 0.5
            };
            (score, positive)
        })
        .collect();
    let n_pos = pairs.iter().filter(|(_, pos)| *pos).count();
    let n_neg = rows - n_pos;
    if n_pos == 0 || n_neg == 0 {
        return Ok(0.5);
    }
    pairs.sort_by(|a, b| a.0.total_cmp(&b.0));
    // Mann–Whitney U from average ranks: a block of equal scores shares its mean rank.
    let mut rank_sum_pos = 0.0f64;
    let mut i = 0usize;
    while i < pairs.len() {
        let mut j = i + 1;
        while j < pairs.len() && pairs[j].0.total_cmp(&pairs[i].0).is_eq() {
            j += 1;
        }
        let rank = ((i + 1) as f64 + j as f64) / 2.0;
        for (_, positive) in &pairs[i..j] {
            if *positive {
                rank_sum_pos += rank;
            }
        }
        i = j;
    }
    let u = rank_sum_pos - (n_pos as f64 * (n_pos as f64 + 1.0)) / 2.0;
    Ok((u / (n_pos as f64 * n_neg as f64)) as f32)
}

/// MAPE — mean absolute percentage error (lower = better).
pub fn mape_score(pred: &Variable, y: &Variable) -> Result<f32> {
    let t = y.data().to_f32_vec()?;
    let p = pred.data().to_f32_vec()?;
    let mut sum = 0.0f32;
    let mut n = 0usize;
    for (&a, &b) in t.iter().zip(&p) {
        if a.abs() > f32::EPSILON {
            sum += ((a - b) / a).abs();
            n += 1;
        }
    }
    Ok(if n == 0 { 0.0 } else { sum / n as f32 })
}

/// EXPLAINED variance — `1 − Var(target − pred) / Var(target)` (higher = better).
pub fn explained_variance_score(pred: &Variable, y: &Variable) -> Result<f32> {
    let t = y.data().to_f32_vec()?;
    let p = pred.data().to_f32_vec()?;
    let n = t.len() as f32;
    if n == 0.0 {
        return Ok(0.0);
    }
    let mean_t = t.iter().sum::<f32>() / n;
    let residual: Vec<f32> = t.iter().zip(&p).map(|(&a, &b)| a - b).collect();
    let mean_r = residual.iter().sum::<f32>() / n;
    let var_t: f32 = t.iter().map(|&v| (v - mean_t).powi(2)).sum::<f32>() / n;
    let var_r: f32 = residual.iter().map(|&v| (v - mean_r).powi(2)).sum::<f32>() / n;
    if var_t.abs() < f32::EPSILON {
        return Ok(if var_r.abs() < f32::EPSILON { 1.0 } else { 0.0 });
    }
    Ok(1.0 - var_r / var_t)
}

/// MSE score (lower = better) — `mse_loss(pred, y).item()`.
pub fn mse_loss_score(pred: &Variable, y: &Variable) -> Result<f32> {
    Ok(flodl::nn::loss::mse_loss(pred, y)?.item()? as f32)
}

/// L1/MAE score (lower = better) — `l1_loss(pred, y).item()`.
pub fn l1_loss_score(pred: &Variable, y: &Variable) -> Result<f32> {
    Ok(flodl::nn::loss::l1_loss(pred, y)?.item()? as f32)
}

/// RMSE score (lower = better) — `sqrt(mse_loss(pred, y).item())`.
pub fn rmse_score(pred: &Variable, y: &Variable) -> Result<f32> {
    Ok(mse_loss_score(pred, y)?.sqrt())
}

/// Macro precision from precomputed argmax vectors.
pub fn precision_from_vecs(pa: &[i64], ta: &[i64], classes: usize) -> f32 {
    if classes == 0 || pa.is_empty() {
        return 0.0;
    }
    let mut sum = 0.0f32;
    for c in 0..classes as i64 {
        let mut tp = 0usize;
        let mut fp = 0usize;
        for (p, t) in pa.iter().zip(ta.iter()) {
            if *p == c {
                if *t == c {
                    tp += 1;
                } else {
                    fp += 1;
                }
            }
        }
        let denom = tp + fp;
        if denom > 0 {
            sum += tp as f32 / denom as f32;
        }
    }
    sum / classes as f32
}

/// Softmax cross-entropy against a **one-hot** target, computed as `−mean(one_hot ·
/// log_softmax(pred))` (lower = better).
pub fn cross_entropy_onehot(pred: &Variable, y: &Variable) -> Result<f32> {
    let n = y.data().shape()[0] as f32;
    if n == 0.0 {
        return Ok(0.0);
    }
    let ls = pred.data().log_softmax(1)?;
    let masked = ls.mul(&y.data())?;
    Ok(-masked.sum()?.item()? as f32 / n)
}

/// Label-smoothed CE loss (for backward): the one-hot target is softened
/// `1−ε` on the true class and `ε/n_classes` elsewhere. `ε = 0` = plain CE.
pub fn label_smoothing_cross_entropy_loss(
    pred: &Variable,
    y: &Variable,
    smoothing: f32,
) -> Result<Variable> {
    let n_classes = y.data().shape()[1] as f32;
    let scale = flodl::Tensor::from_f32(&[1.0 - smoothing], &[1], y.data().device())?;
    let off = flodl::Tensor::from_f32(&[smoothing / n_classes], &[1], y.data().device())?;
    let smoothed_y = y.data().mul(&scale)?.add(&off)?;
    cross_entropy_onehot_loss(pred, &Variable::new(smoothed_y, false))
}

/// Focal loss against a ONE-HOT target: `−(1−p_t)^γ·log p_t`, averaged over rows.
pub fn focal_loss_onehot(pred: &Variable, y: &Variable, gamma: f32) -> Result<Variable> {
    let log_p = pred.data().log_softmax(1)?;
    // One term per row: the true class's log-probability, then its probability.
    let log_pt = log_p.mul(&y.data())?.sum_dims(&[1], false)?;
    let pt = log_pt.exp()?;
    // (1 − p_t)^γ — the modulating factor.
    let weight = pt.neg()?.add_scalar(1.0)?.pow_scalar(gamma as f64)?;
    // mean over rows of −(1−p_t)^γ log p_t, matching the CE reduction.
    let per_row = weight.mul(&log_pt)?.neg()?;
    Ok(Variable::new(per_row.mean()?, false))
}

/// Softmax cross-entropy loss tensor (for backward).
pub fn cross_entropy_onehot_loss(pred: &Variable, y: &Variable) -> Result<Variable> {
    let ls = pred.data().log_softmax(1)?;
    let masked = ls.mul(&y.data())?;
    let neg = flodl::Tensor::from_f32(&[-1.0], &[1], masked.device())?;
    let n = flodl::Tensor::from_f32(&[y.data().shape()[0] as f32], &[1], y.data().device())?;
    Ok(Variable::new(masked.sum()?.mul(&neg)?.div(&n)?, false))
}

#[cfg(test)]
mod tests {
    use super::*;
    use flodl::{Device, Tensor};

    fn var(xs: &[f32], shape: &[i64]) -> Variable {
        Variable::new(Tensor::from_f32(xs, shape, Device::CPU).unwrap(), false)
    }

    /// Three rows, one target class each; row 2 puts its truth SECOND.
    fn three_class_case() -> (Variable, Variable) {
        (
            var(
                &[
                    3.0, 0.0, 0.0, // truth 0 — top-1
                    0.0, 3.0, 0.5, // truth 1 — top-1
                    0.0, 3.0, 1.0, // truth 2 — ranked 3rd, so top-2 hit, top-1 miss
                ],
                &[3, 3],
            ),
            var(&[1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0], &[3, 3]),
        )
    }

    #[test]
    fn top_k_is_accuracy_at_one_and_inclusive_at_the_class_count() {
        let (pred, y) = three_class_case();
        let one = top_k_accuracy_score(&pred, &y, 1).unwrap();
        let acc = accuracy_score(&pred, &y).unwrap();
        assert!(
            (one - acc).abs() < 1e-6,
            "k=1 must BE accuracy: {one} vs {acc}"
        );
        assert!((one - 2.0 / 3.0).abs() < 1e-6, "k=1 = {one}");
        // Row 2's truth is second, so k=2 accepts every row.
        assert!((top_k_accuracy_score(&pred, &y, 2).unwrap() - 1.0).abs() < 1e-6);
        // k beyond the class count is clamped, not an error.
        assert!((top_k_accuracy_score(&pred, &y, 99).unwrap() - 1.0).abs() < 1e-6);
        // k=0 is meaningless and clamps up to k=1 rather than scoring 0 rows.
        assert!((top_k_accuracy_score(&pred, &y, 0).unwrap() - one).abs() < 1e-6);
    }

    fn binary_case(positive_scores: &[f32; 4]) -> (Variable, Variable) {
        // Class-1 scores: rows 0 and 1 are positive, rows 2 and 3 negative.
        let mut xs = Vec::new();
        for s in positive_scores {
            xs.push(1.0 - s);
            xs.push(*s);
        }
        (
            var(&xs, &[4, 2]),
            var(&[0.0, 1.0, 0.0, 1.0, 1.0, 0.0, 1.0, 0.0], &[4, 2]),
        )
    }

    #[test]
    fn auroc_is_one_perfect_zero_reversed_half_tied() {
        let perfect = auroc_score(
            &binary_case(&[0.9, 0.8, 0.2, 0.1]).0,
            &binary_case(&[0.9, 0.8, 0.2, 0.1]).1,
        )
        .unwrap();
        assert!(
            (perfect - 1.0).abs() < 1e-6,
            "perfect separation = {perfect}"
        );
        let reversed = auroc_score(
            &binary_case(&[0.1, 0.2, 0.8, 0.9]).0,
            &binary_case(&[0.1, 0.2, 0.8, 0.9]).1,
        )
        .unwrap();
        assert!(reversed.abs() < 1e-6, "reversed separation = {reversed}");
        let tied = auroc_score(
            &binary_case(&[0.5, 0.5, 0.5, 0.5]).0,
            &binary_case(&[0.5, 0.5, 0.5, 0.5]).1,
        )
        .unwrap();
        assert!(
            (tied - 0.5).abs() < 1e-6,
            "all tied = {tied}, expected a coin flip"
        );
    }

    #[test]
    fn auroc_on_one_class_says_the_exam_cannot_tell() {
        // Every target row positive: no negative pair exists, so 0.5 (never NaN).
        let pred = var(&[0.9, 0.8, 0.7, 0.6], &[4, 1]);
        let y = var(&[1.0, 1.0, 1.0, 1.0], &[4, 1]);
        let got = auroc_score(&pred, &y).unwrap();
        assert!((got - 0.5).abs() < 1e-6, "single-class AUROC = {got}");
        assert!(got.is_finite());
    }
}