ferrolearn-linear 0.5.0

Linear models for the ferrolearn ML framework
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
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
//! Linear Support Vector Regressor.
//!
//! This module provides [`LinearSVR`], a liblinear-faithful linear SVR that
//! operates directly in the primal space without kernel overhead. The fit
//! minimizes the L2-regularized epsilon-insensitive (or squared
//! epsilon-insensitive) objective
//!
//! ```text
//!   min_w  0.5 * ||w||^2  +  C * sum_i  L_eps(y_i - w . x_i)
//! ```
//!
//! (NO `1/n` averaging — the summed loss is scaled by `C`, matching
//! `sklearn/svm/_base.py` `_fit_liblinear`). The solver is liblinear's dual
//! coordinate descent for SVR (`solve_l2r_l1l2_svr` in
//! `sklearn/svm/src/liblinear/linear.cpp:1051`), which converges to the unique
//! minimizer of the strongly convex objective.
//!
//! # Examples
//!
//! ```
//! use ferrolearn_linear::linear_svr::LinearSVR;
//! use ferrolearn_core::{Fit, Predict};
//! use ndarray::{array, Array2};
//!
//! let x = Array2::from_shape_vec((5, 1), vec![1.0, 2.0, 3.0, 4.0, 5.0]).unwrap();
//! let y = array![2.0, 4.0, 6.0, 8.0, 10.0];
//!
//! let model = LinearSVR::<f64>::new();
//! let fitted = model.fit(&x, &y).unwrap();
//! let preds = fitted.predict(&x).unwrap();
//! assert_eq!(preds.len(), 5);
//! ```
//!
//! ## REQ status
//!
//! Binary (R-DEFER-2): SHIPPED = impl + non-test production consumer + tests +
//! green oracle verification; NOT-STARTED = open blocker `#`. `LinearSVR`/
//! `FittedLinearSVR`/`LinearSVRLoss` are boundary estimator types re-exported at
//! the crate root (`pub use linear_svr::{…} in lib.rs`); under S5/R-DEFER-1 the
//! public estimator type + the `PipelineEstimator` impl ARE the consumer surface
//! (no `ferrolearn-python` LinearSVR binding yet). See `.design/linear/linear_svr.md`.
//!
//! | REQ | Status | Evidence |
//! |---|---|---|
//! | REQ-1 (fit parity — coef_/intercept_ vs liblinear oracle) | SHIPPED | `fn fit` minimizes liblinear's `0.5·‖w‖² + C·Σ L_ε` via dual CD (`solve_l2r_l1l2_svr`, `linear.cpp:1051`). Pinned by `tests/divergence_linear_svr_fit.rs::linear_svr_coef_parity`. |
//! | REQ-2 (predict = X·coef + intercept) | SHIPPED | `fn predict` computes `x.dot(&coefficients) + intercept` (mirrors `LinearModel.predict`, `_base.py`). Pinned by `tests/divergence_linear_svr_fit.rs::linear_svr_predict` (epsilon=0.1, C=1.0 fit → `predict([[2.5]]) ≈ 5.0251` live oracle). Consumer: `predict_pipeline` (`FittedPipelineEstimator` impl). |
//! | REQ-3 (epsilon default = 0.0) | SHIPPED | `LinearSVR::new` sets `epsilon = F::zero()` (`_classes.py:522`). Pinned by `linear_svr_default_epsilon`. |
//! | REQ-4 (loss param + squared objective) | SHIPPED | the `SquaredEpsilonInsensitive` branch sets `lambda = 0.5/C`, `upper_bound = +inf` (the `L2R_L2LOSS_SVR_DUAL` path, `linear.cpp:1078-1081`), no `1/n`. Pinned by `tests/divergence_linear_svr_fit.rs::linear_svr_squared_loss` (`loss='squared_epsilon_insensitive'`, epsilon=0.1, C=1.0 → live oracle `coef [1.8913]`, `intercept [0.2821]`, `predict([[1.5]]) ≈ 3.119`). |
//! | REQ-5 (fit_intercept + intercept_scaling) | SHIPPED | augmented synthetic column = `intercept_scaling`, penalized in ‖w‖² (`_base.py:1189-1198`), `intercept_ = intercept_scaling·w_last`. Pinned by `linear_svr_coef_parity`, module `test_fit_intercept_false_zero_intercept`/`test_invalid_intercept_scaling`. |
//! | REQ-6 (dual param) | SHIPPED | `LinearSVR` exposes `pub dual: DualMode` (default `Auto`) + `#[must_use] with_dual`, modeling sklearn's `"dual": ["boolean", StrOptions({"auto"})]` (`_classes.py:513`, default `"auto"` `:528`). `fn fit` resolves `Auto`→dual solver and rejects `dual=False` + `EpsilonInsensitive` with `FerroError::InvalidParameter` (sklearn's `ValueError "Unsupported set of arguments"`, `_get_liblinear_solver_type`, `_base.py:1015,:1047`); `dual=False` + `SquaredEpsilonInsensitive` reuses the dual CD (same strongly-convex minimizer, R-DEV-7). Pinned by `tests/divergence_linear_svr_fit.rs::{linear_svr_dual_auto_true, linear_svr_dual_false_eps_rejected, linear_svr_dual_false_squared}`. Consumer: `pub use linear_svr::{…}` (`lib.rs`) + `PipelineEstimator` impl. |
//! | REQ-7 (C-scaling convention) | SHIPPED | the `/n` division is removed; dual CD uses `upper_bound = C` (L1) / `lambda = 0.5/C` (L2). Pinned by `linear_svr_coef_c_dependence`. |
//! | REQ-8 (tol/max_iter + n_iter_) | SHIPPED | `fn fit` counts dual-CD outer iterations into `FittedLinearSVR::n_iter`, exposed via `#[must_use] pub fn n_iter` (mirrors `n_iter_ = n_iter_.max().item()`, `_classes.py:603`); emits the `ConvergenceWarning`-equivalent via `eprintln!` at `max_iter` (`_base.py:1234-1238`, crate qda/lda warning channel). Pinned by `tests/divergence_linear_svr_fit.rs::linear_svr_n_iter` (`1 <= n_iter <= max_iter`). |
//! | REQ-9 (param validation + n_features_in_) | SHIPPED | `fn fit` validates `tol > 0` → `FerroError::InvalidParameter` (sklearn `"tol": [Interval(Real, 0.0, None, closed="neither")]`, `_classes.py:508`); `max_iter` is `usize` so `>= 0` always (sklearn `Interval(Integral,0,None,closed="left")`, `:516`) — documented, no check. Keeps `C>0`/`epsilon>=0` (both match sklearn's empirical fit-time behavior — negative epsilon raises ValueError at fit). `FittedLinearSVR` stores `n_features_in` (= `X.ncols()`), exposed via `#[must_use] pub fn n_features_in` mirroring `n_features_in_` (`_validate_data`, `_classes.py:569-576`). Pinned by `tests/divergence_linear_svr_fit.rs::{linear_svr_tol_validation, linear_svr_n_features_in}`. The `intercept_`-shape (length-1 ndarray vs scalar) sub-item is a binding-ABI concern deferred to the ferrolearn-python layer (cf. #600); `intercept()` keeps returning the scalar. Consumer: `pub use linear_svr::{…}` (`lib.rs`) + `PipelineEstimator` impl. |
//! | REQ-10 (ferray substrate) | NOT-STARTED | open prereq blocker #615. Imports `ndarray`, not `ferray-core`/`ferray::linalg` (R-SUBSTRATE). |
//! | REQ-11 (non-finite input rejected) | SHIPPED | `fn fit` rejects any NaN/+/-inf in X OR y BEFORE the dual coordinate-descent solve with `FerroError::InvalidParameter`, mirroring sklearn's `_validate_data(force_all_finite=True)` (`svm/_classes.py:302`, the liblinear `_fit_liblinear` path; cf. `svm/_base.py:190`) → `ValueError("Input X contains NaN.")` / `"Input y contains NaN."` / `"... contains infinity ..."`. `y` is `Array1<F>` (float regression targets), so X AND y are checked; ferrolearn's `LinearSVR::fit` takes no `sample_weight` argument. `.iter().any(|v| !v.is_finite())` catches NaN and Inf; the finite path is byte-identical. Verified vs the live sklearn 1.5.2 oracle (R-CHAR-3): `LinearSVR().fit` raises `ValueError` for NaN/+inf/-inf in X and NaN/inf in y (`tests/divergence_linear_nonfinite_batch4.rs::linsvr_*`). Non-test consumer: `pub use linear_svr::{…}` (`lib.rs`) + `PipelineEstimator` impl. (#2263) |

use ferrolearn_core::error::FerroError;
use ferrolearn_core::introspection::HasCoefficients;
use ferrolearn_core::pipeline::{FittedPipelineEstimator, PipelineEstimator};
use ferrolearn_core::traits::{Fit, Predict};
use ndarray::{Array1, Array2, ScalarOperand};
use num_traits::Float;

/// Loss function for [`LinearSVR`].
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum LinearSVRLoss {
    /// Epsilon-insensitive loss: `max(0, |y - f(x)| - epsilon)` (L1, the
    /// standard SVR loss). liblinear solver `L2R_L1LOSS_SVR_DUAL` (type 13).
    EpsilonInsensitive,
    /// Squared epsilon-insensitive loss: `max(0, |y - f(x)| - epsilon)^2` (L2).
    /// liblinear solver `L2R_L2LOSS_SVR_DUAL` (type 12).
    SquaredEpsilonInsensitive,
}

/// Solver selection for [`LinearSVR`], mirroring sklearn's `dual` parameter
/// (`sklearn/svm/_classes.py:506-516` `_parameter_constraints`:
/// `"dual": ["boolean", StrOptions({"auto"})]`, default `"auto"`).
///
/// sklearn's `dual` selects the liblinear dual (`True`) or primal (`False`)
/// solver; `"auto"` resolves to one of them based on `n_samples`/`n_features`
/// and whether the loss supports that solver
/// (`_validate_dual_parameter`, `_classes.py:13-29`). For SVR the
/// `epsilon_insensitive` loss is **dual-only** (solver type 13 — no `{False:…}`
/// entry, `_base.py:1015`), so `dual=False` with that loss raises a
/// `ValueError`; `squared_epsilon_insensitive` supports both
/// (primal type 11 / dual type 12, `_base.py:1016`).
///
/// The minimizer of the strongly-convex objective is solver-invariant (R-DEV-7),
/// so ferrolearn always uses its dual coordinate descent for the resolved
/// solver and produces the same observable `coef_`/`intercept_`.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum DualMode {
    /// `"auto"`: resolve automatically. For SVR this resolves to the dual
    /// solver (the supported solver for both losses), matching sklearn's
    /// `_validate_dual_parameter` (`_classes.py:13-29`).
    Auto,
    /// `True`: use the dual liblinear solver.
    True,
    /// `False`: use the primal solver. For `EpsilonInsensitive` this is
    /// unsupported (sklearn raises `ValueError`); for
    /// `SquaredEpsilonInsensitive` it reaches the same optimum as the dual
    /// solver (R-DEV-7).
    False,
}

/// Linear Support Vector Regressor (primal objective, liblinear dual CD).
///
/// Solves the L2-regularized epsilon-insensitive or squared
/// epsilon-insensitive objective `0.5*||w||^2 + C * sum_i L_eps(y_i - w.x_i)`
/// via liblinear's dual coordinate descent. Mirrors `sklearn.svm.LinearSVR`.
///
/// # Type Parameters
///
/// - `F`: The floating-point type (`f32` or `f64`).
#[derive(Debug, Clone)]
pub struct LinearSVR<F> {
    /// Regularization parameter. The strength of the regularization is
    /// inversely proportional to `C`. Must be strictly positive.
    pub c: F,
    /// Width of the epsilon-insensitive tube (sklearn default `0.0`).
    pub epsilon: F,
    /// Maximum number of dual coordinate descent iterations.
    pub max_iter: usize,
    /// Convergence tolerance on the projected-gradient L1 norm.
    pub tol: F,
    /// Loss function to use.
    pub loss: LinearSVRLoss,
    /// Whether to fit an intercept. When `true`, the design matrix is augmented
    /// with a synthetic constant column equal to [`intercept_scaling`]; the
    /// augmented weight is penalized like any feature (liblinear convention).
    ///
    /// [`intercept_scaling`]: Self::intercept_scaling
    pub fit_intercept: bool,
    /// Value of the synthetic intercept feature when [`fit_intercept`] is
    /// `true`. Must be strictly positive. `intercept_ = intercept_scaling *
    /// w_last`.
    ///
    /// [`fit_intercept`]: Self::fit_intercept
    pub intercept_scaling: F,
    /// Solver selection (sklearn `dual`, default [`DualMode::Auto`]). See
    /// [`DualMode`] for resolution semantics.
    pub dual: DualMode,
}

impl<F: Float> LinearSVR<F> {
    /// Create a new `LinearSVR` with scikit-learn's default settings.
    ///
    /// Defaults (matching `sklearn.svm.LinearSVR`, `_classes.py:519-532`):
    /// `C = 1.0`, `epsilon = 0.0`, `max_iter = 1000`, `tol = 1e-4`,
    /// `loss = EpsilonInsensitive`, `fit_intercept = true`,
    /// `intercept_scaling = 1.0`, `dual = Auto`.
    ///
    /// # Panics
    ///
    /// Never panics: the literal constants `0.0`, `1e-4`, `1.0` are exactly
    /// representable in every supported float type, so the `F::from` conversions
    /// cannot fail.
    #[must_use]
    pub fn new() -> Self {
        // These literals are exactly representable, so `or(F::zero())` /
        // `or(F::one())` fallbacks are never taken (no `.unwrap()` in lib code).
        let zero = F::zero();
        let one = F::one();
        Self {
            c: one,
            epsilon: zero,
            max_iter: 1000,
            tol: F::from(1e-4).unwrap_or_else(|| {
                // 1e-4 is representable in f32/f64; fall back defensively.
                let ten = F::from(10).unwrap_or(one);
                one / (ten * ten * ten * ten)
            }),
            loss: LinearSVRLoss::EpsilonInsensitive,
            fit_intercept: true,
            intercept_scaling: one,
            dual: DualMode::Auto,
        }
    }

    /// Set the regularization parameter C.
    #[must_use]
    pub fn with_c(mut self, c: F) -> Self {
        self.c = c;
        self
    }

    /// Set the epsilon tube width.
    #[must_use]
    pub fn with_epsilon(mut self, epsilon: F) -> Self {
        self.epsilon = epsilon;
        self
    }

    /// Set the maximum number of iterations.
    #[must_use]
    pub fn with_max_iter(mut self, max_iter: usize) -> Self {
        self.max_iter = max_iter;
        self
    }

    /// Set the convergence tolerance.
    #[must_use]
    pub fn with_tol(mut self, tol: F) -> Self {
        self.tol = tol;
        self
    }

    /// Set the loss function.
    #[must_use]
    pub fn with_loss(mut self, loss: LinearSVRLoss) -> Self {
        self.loss = loss;
        self
    }

    /// Set whether to fit an intercept (sklearn `fit_intercept`).
    #[must_use]
    pub fn with_fit_intercept(mut self, fit_intercept: bool) -> Self {
        self.fit_intercept = fit_intercept;
        self
    }

    /// Set the intercept scaling (sklearn `intercept_scaling`). Must be
    /// strictly positive when `fit_intercept` is `true`.
    #[must_use]
    pub fn with_intercept_scaling(mut self, intercept_scaling: F) -> Self {
        self.intercept_scaling = intercept_scaling;
        self
    }

    /// Set the solver selection (sklearn `dual`). See [`DualMode`].
    #[must_use]
    pub fn with_dual(mut self, dual: DualMode) -> Self {
        self.dual = dual;
        self
    }
}

impl<F: Float> Default for LinearSVR<F> {
    fn default() -> Self {
        Self::new()
    }
}

/// Fitted Linear Support Vector Regressor.
///
/// Stores the learned coefficients and intercept. `coef_` are the weights on
/// the original features; `intercept_ = intercept_scaling * w_last` when an
/// intercept was fit, otherwise `0`.
#[derive(Debug, Clone)]
pub struct FittedLinearSVR<F> {
    /// Learned coefficient vector (one per original feature).
    coefficients: Array1<F>,
    /// Learned intercept (bias) term.
    intercept: F,
    /// Number of dual coordinate-descent outer iterations actually run.
    ///
    /// Mirrors `sklearn.svm.LinearSVR.n_iter_` (`_classes.py:467-468, :603`,
    /// `n_iter_ = n_iter_.max().item()`). Because ferrolearn's dual CD is a
    /// distinct implementation from liblinear's, the exact count need not match
    /// liblinear's bookkeeping; it is a structural attribute satisfying
    /// `1 <= n_iter <= max_iter`.
    n_iter: usize,
    /// Number of features seen during `fit` (the number of columns of `X`).
    ///
    /// Mirrors sklearn's standard `n_features_in_` fitted attribute
    /// (set by `_validate_data` in `LinearSVR.fit`, `_classes.py:569-576`).
    n_features_in: usize,
}

impl<F> FittedLinearSVR<F> {
    /// Number of dual coordinate-descent outer iterations run during `fit`.
    ///
    /// Mirrors `sklearn.svm.LinearSVR.n_iter_` (`sklearn/svm/_classes.py:467`,
    /// set to `n_iter_.max().item()` at `_classes.py:603`). A value equal to
    /// `max_iter` indicates the solver did not converge (a
    /// `ConvergenceWarning` is emitted on the crate's warning channel in that
    /// case, mirroring `_base.py:1234-1238`).
    #[must_use]
    pub fn n_iter(&self) -> usize {
        self.n_iter
    }

    /// Number of features seen during `fit`.
    ///
    /// Mirrors `sklearn.svm.LinearSVR.n_features_in_` (the standard scikit-learn
    /// fitted attribute set by `_validate_data`, `_classes.py:569-576`): the
    /// number of columns of the `X` passed to `fit`.
    #[must_use]
    pub fn n_features_in(&self) -> usize {
        self.n_features_in
    }
}

impl<F: Float + Send + Sync + ScalarOperand + 'static> Fit<Array2<F>, Array1<F>> for LinearSVR<F> {
    type Fitted = FittedLinearSVR<F>;
    type Error = FerroError;

    /// Fit the linear SVR model using liblinear's dual coordinate descent.
    ///
    /// Minimizes `0.5*||w||^2 + C * sum_i L_eps(y_i - w.x_i)` (no `1/n`). When
    /// `fit_intercept` is set, the design matrix is augmented with a constant
    /// column equal to `intercept_scaling`, the augmented weight is penalized
    /// like any feature, and `intercept_ = intercept_scaling * w_last`.
    ///
    /// # Errors
    ///
    /// - [`FerroError::ShapeMismatch`] — sample count mismatch.
    /// - [`FerroError::InvalidParameter`] — `C` not positive, `epsilon`
    ///   negative, `tol` not positive, `intercept_scaling` not positive (when
    ///   fitting an intercept), or `dual = False` with `EpsilonInsensitive`
    ///   loss (sklearn's unsupported `(loss, dual)` combination).
    /// - [`FerroError::InsufficientSamples`] — no samples provided.
    fn fit(&self, x: &Array2<F>, y: &Array1<F>) -> Result<FittedLinearSVR<F>, FerroError> {
        let (n_samples, n_features) = x.dim();

        if n_samples != y.len() {
            return Err(FerroError::ShapeMismatch {
                expected: vec![n_samples],
                actual: vec![y.len()],
                context: "y length must match number of samples in X".into(),
            });
        }

        if n_samples == 0 {
            return Err(FerroError::InsufficientSamples {
                required: 1,
                actual: 0,
                context: "LinearSVR requires at least one sample".into(),
            });
        }

        if self.c <= F::zero() {
            return Err(FerroError::InvalidParameter {
                name: "C".into(),
                reason: "must be positive".into(),
            });
        }

        if self.epsilon < F::zero() {
            return Err(FerroError::InvalidParameter {
                name: "epsilon".into(),
                reason: "must be non-negative".into(),
            });
        }

        // `tol` must be strictly positive (sklearn
        // `"tol": [Interval(Real, 0.0, None, closed="neither")]`,
        // `_classes.py:508`). `max_iter` is `usize` so `>= 0` always holds
        // (sklearn `Interval(Integral, 0, None, closed="left")`,
        // `_classes.py:516`) — no runtime check is needed.
        if self.tol <= F::zero() {
            return Err(FerroError::InvalidParameter {
                name: "tol".into(),
                reason: "must be greater than 0".into(),
            });
        }

        // Resolve `dual` and reject unsupported (loss, dual) combinations,
        // mirroring sklearn's `_validate_dual_parameter` (`_classes.py:13-29`)
        // + `_get_liblinear_solver_type` (`_base.py:995-1047`). `"auto"`
        // resolves to the dual solver (the supported solver for both SVR
        // losses); `dual=False` with `epsilon_insensitive` is unsupported
        // (solver dict has no `{False:…}` entry, `_base.py:1015`) → ValueError.
        // The minimizer is solver-invariant (strongly convex; R-DEV-7), so for
        // the supported combinations ferrolearn keeps using its dual CD and
        // produces the oracle-matching `coef_`/`intercept_`.
        if matches!(self.dual, DualMode::False)
            && matches!(self.loss, LinearSVRLoss::EpsilonInsensitive)
        {
            return Err(FerroError::InvalidParameter {
                name: "dual".into(),
                reason: "dual=False is not supported for epsilon_insensitive loss".into(),
            });
        }

        // liblinear raises when intercept_scaling <= 0 with fit_intercept
        // (`_base.py:1190-1196`).
        if self.fit_intercept && self.intercept_scaling <= F::zero() {
            return Err(FerroError::InvalidParameter {
                name: "intercept_scaling".into(),
                reason: "must be greater than 0 when fit_intercept is true".into(),
            });
        }

        // Non-finite input validation (#2263). sklearn `LinearSVR.fit`
        // -> `self._validate_data(X, y, ...)` (`svm/_classes.py:302`, the shared
        // liblinear `_fit_liblinear` path; cf. `svm/_base.py:190`) keeps the
        // default `force_all_finite=True`, so `check_array` rejects any NaN or
        // +/-inf in X OR y with a `ValueError("Input X contains NaN.")` /
        // `"Input y contains NaN."` / `"... contains infinity ..."` BEFORE the
        // dual coordinate-descent solve. `y` is `Array1<F>` (float regression
        // targets), so X AND y need the runtime check; ferrolearn's `LinearSVR::
        // fit` takes no `sample_weight` argument.
        // `.iter().any(|v| !v.is_finite())` rejects both NaN and Inf (bounds-safe,
        // no panic, R-CODE-2); the finite path is byte-identical.
        if x.iter().any(|v| !v.is_finite()) {
            return Err(FerroError::InvalidParameter {
                name: "X".into(),
                reason: "Input X contains NaN or infinity.".into(),
            });
        }
        if y.iter().any(|v| !v.is_finite()) {
            return Err(FerroError::InvalidParameter {
                name: "y".into(),
                reason: "Input y contains NaN or infinity.".into(),
            });
        }

        // Build the (possibly augmented) design matrix. The synthetic bias
        // column holds `intercept_scaling`; its weight is penalized in ||w||^2,
        // exactly as liblinear treats it (`_base.py:1189-1198`).
        let w_size = if self.fit_intercept {
            n_features + 1
        } else {
            n_features
        };

        // Per-sample squared norm QD[i] = ||x_i||^2 (including the bias column).
        let mut qd = vec![F::zero(); n_samples];
        for (i, qd_i) in qd.iter_mut().enumerate() {
            let mut acc = F::zero();
            let row = x.row(i);
            for &v in row.iter() {
                acc = acc + v * v;
            }
            if self.fit_intercept {
                acc = acc + self.intercept_scaling * self.intercept_scaling;
            }
            *qd_i = acc;
        }

        // liblinear `solve_l2r_l1l2_svr` (`linear.cpp:1051`).
        //   p   = epsilon, C = C, eps = tol
        //   L1 loss: lambda[i] = 0,        upper_bound[i] = C
        //   L2 loss: lambda[i] = 0.5 / C,  upper_bound[i] = +inf
        let p = self.epsilon;
        let inf = F::infinity();
        let (lambda, upper_bound) = match self.loss {
            LinearSVRLoss::EpsilonInsensitive => (F::zero(), self.c),
            LinearSVRLoss::SquaredEpsilonInsensitive => {
                let half = F::from(0.5).unwrap_or_else(|| F::one() / (F::one() + F::one()));
                (half / self.c, inf)
            }
        };

        let mut beta = vec![F::zero(); n_samples];
        let mut w = vec![F::zero(); w_size];
        // beta starts at 0 so w starts at 0; no contribution to accumulate.

        let mut index: Vec<usize> = (0..n_samples).collect();
        let mut active_size = n_samples;
        let mut gmax_old = inf;
        let mut gnorm1_init = -F::one();

        let tiny = F::from(1.0e-12).unwrap_or_else(|| F::epsilon());

        // Closure-free helper to compute w . x_i (augmented).
        let dot_w_xi = |w: &[F], i: usize| -> F {
            let mut acc = F::zero();
            let row = x.row(i);
            for (j, &v) in row.iter().enumerate() {
                acc = acc + v * w[j];
            }
            if self.fit_intercept {
                acc = acc + self.intercept_scaling * w[n_features];
            }
            acc
        };

        // Count of dual-CD outer iterations actually run (for `n_iter_`) and
        // whether the convergence criterion was met before `max_iter`.
        let mut n_iter: usize = 0;
        let mut converged = false;

        for iter in 0..self.max_iter {
            n_iter = iter + 1;
            let mut gmax_new = F::zero();
            let mut gnorm1_new = F::zero();

            // liblinear shuffles `index` each sweep; the minimizer is unique so
            // order only affects the path, not the limit. We sweep in natural
            // order for determinism (no RNG), matching the converged optimum.

            let mut s = 0;
            while s < active_size {
                let i = index[s];

                // G = -y_i + lambda*beta_i + w . x_i
                let g = -y[i] + lambda * beta[i] + dot_w_xi(&w, i);
                let gp = g + p;
                let gn = g - p;

                let mut violation = F::zero();
                let mut shrink = false;

                if beta[i] == F::zero() {
                    if gp < F::zero() {
                        violation = -gp;
                    } else if gn > F::zero() {
                        violation = gn;
                    } else if gp > gmax_old && gn < -gmax_old {
                        shrink = true;
                    }
                } else if beta[i] >= upper_bound {
                    if gp > F::zero() {
                        violation = gp;
                    } else if gp < -gmax_old {
                        shrink = true;
                    }
                } else if beta[i] <= -upper_bound {
                    if gn < F::zero() {
                        violation = -gn;
                    } else if gn > gmax_old {
                        shrink = true;
                    }
                } else if beta[i] > F::zero() {
                    violation = gp.abs();
                } else {
                    violation = gn.abs();
                }

                if shrink {
                    active_size -= 1;
                    index.swap(s, active_size);
                    continue; // re-process the swapped-in element at `s`
                }

                gmax_new = if violation > gmax_new {
                    violation
                } else {
                    gmax_new
                };
                gnorm1_new = gnorm1_new + violation;

                // Newton direction d (1-D minimization of the dual along beta_i).
                let h = qd[i] + lambda;
                let d = if gp < h * beta[i] {
                    -gp / h
                } else if gn > h * beta[i] {
                    -gn / h
                } else {
                    -beta[i]
                };

                if d.abs() < tiny {
                    s += 1;
                    continue;
                }

                let beta_old = beta[i];
                let mut new_beta = beta[i] + d;
                // Clamp to [-upper_bound, upper_bound].
                if new_beta > upper_bound {
                    new_beta = upper_bound;
                } else if new_beta < -upper_bound {
                    new_beta = -upper_bound;
                }
                beta[i] = new_beta;
                let d_eff = beta[i] - beta_old;

                if d_eff != F::zero() {
                    let row = x.row(i);
                    for (j, &v) in row.iter().enumerate() {
                        w[j] = w[j] + d_eff * v;
                    }
                    if self.fit_intercept {
                        w[n_features] = w[n_features] + d_eff * self.intercept_scaling;
                    }
                }

                s += 1;
            }

            if iter == 0 {
                gnorm1_init = gnorm1_new;
            }

            // Convergence: Gnorm1_new <= tol * Gnorm1_init. If converged on the
            // active set, reactivate all and re-check once (liblinear pattern).
            if gnorm1_new <= self.tol * gnorm1_init {
                if active_size == n_samples {
                    converged = true;
                    break;
                }
                active_size = n_samples;
                gmax_old = inf;
                continue;
            }

            gmax_old = gmax_new;
        }

        // liblinear warns when the iteration count reaches `max_iter` without
        // satisfying the stopping criterion (`_base.py:1234-1238`). The crate's
        // warning channel is `eprintln!` (cf. qda.rs collinearity, lda.rs
        // priors renormalization).
        if !converged && n_iter >= self.max_iter {
            eprintln!("Liblinear failed to converge, increase the number of iterations.");
        }

        // Extract coef_ / intercept_ (`_base.py:598`, `_classes.py:581-598`).
        let coefficients = Array1::from_iter(w.iter().take(n_features).copied());
        let intercept = if self.fit_intercept {
            self.intercept_scaling * w[n_features]
        } else {
            F::zero()
        };

        // `n_iter` is at least 1 (the loop body always runs once for
        // `max_iter >= 1`; a zero `max_iter` leaves it at 0, the degenerate
        // sklearn case where no iteration runs).
        Ok(FittedLinearSVR {
            coefficients,
            intercept,
            n_iter,
            n_features_in: n_features,
        })
    }
}

impl<F: Float + Send + Sync + ScalarOperand + 'static> Predict<Array2<F>> for FittedLinearSVR<F> {
    type Output = Array1<F>;
    type Error = FerroError;

    /// Predict target values for the given feature matrix.
    ///
    /// Computes `X @ coefficients + intercept`.
    ///
    /// # Errors
    ///
    /// Returns [`FerroError::ShapeMismatch`] if the number of features
    /// does not match the fitted model.
    fn predict(&self, x: &Array2<F>) -> Result<Array1<F>, FerroError> {
        let n_features = x.ncols();
        if n_features != self.coefficients.len() {
            return Err(FerroError::ShapeMismatch {
                expected: vec![self.coefficients.len()],
                actual: vec![n_features],
                context: "number of features must match fitted model".into(),
            });
        }

        let preds = x.dot(&self.coefficients) + self.intercept;
        Ok(preds)
    }
}

impl<F: Float + Send + Sync + ScalarOperand + 'static> HasCoefficients<F> for FittedLinearSVR<F> {
    fn coefficients(&self) -> &Array1<F> {
        &self.coefficients
    }

    fn intercept(&self) -> F {
        self.intercept
    }
}

// Pipeline integration.
impl<F> PipelineEstimator<F> for LinearSVR<F>
where
    F: Float + ScalarOperand + Send + Sync + 'static,
{
    fn fit_pipeline(
        &self,
        x: &Array2<F>,
        y: &Array1<F>,
    ) -> Result<Box<dyn FittedPipelineEstimator<F>>, FerroError> {
        let fitted = self.fit(x, y)?;
        Ok(Box::new(fitted))
    }
}

impl<F> FittedPipelineEstimator<F> for FittedLinearSVR<F>
where
    F: Float + ScalarOperand + Send + Sync + 'static,
{
    fn predict_pipeline(&self, x: &Array2<F>) -> Result<Array1<F>, FerroError> {
        self.predict(x)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use approx::assert_relative_eq;
    use ndarray::array;

    #[test]
    fn test_default_constructor() {
        let m = LinearSVR::<f64>::new();
        assert_eq!(m.max_iter, 1000);
        assert!(m.c == 1.0);
        // sklearn default epsilon=0.0 (`_classes.py:522`).
        assert_relative_eq!(m.epsilon, 0.0);
        assert_eq!(m.loss, LinearSVRLoss::EpsilonInsensitive);
        assert!(m.fit_intercept);
        assert_relative_eq!(m.intercept_scaling, 1.0);
    }

    #[test]
    fn test_builder_setters() {
        let m = LinearSVR::<f64>::new()
            .with_c(10.0)
            .with_epsilon(0.5)
            .with_max_iter(500)
            .with_tol(1e-6)
            .with_loss(LinearSVRLoss::SquaredEpsilonInsensitive)
            .with_fit_intercept(false)
            .with_intercept_scaling(2.0);
        assert!(m.c == 10.0);
        assert_relative_eq!(m.epsilon, 0.5);
        assert_eq!(m.max_iter, 500);
        assert_eq!(m.loss, LinearSVRLoss::SquaredEpsilonInsensitive);
        assert!(!m.fit_intercept);
        assert_relative_eq!(m.intercept_scaling, 2.0);
    }

    #[test]
    fn test_fits_linear_data() {
        let x = Array2::from_shape_vec((5, 1), vec![1.0, 2.0, 3.0, 4.0, 5.0]).unwrap();
        let y = array![2.0, 4.0, 6.0, 8.0, 10.0];

        let model = LinearSVR::<f64>::new()
            .with_c(10.0)
            .with_epsilon(0.0)
            .with_max_iter(10000);
        let fitted = model.fit(&x, &y).unwrap();
        let preds = fitted.predict(&x).unwrap();

        // Should roughly recover y = 2x.
        for (p, &t) in preds.iter().zip(y.iter()) {
            assert!(
                (p - t).abs() < 3.0,
                "prediction {p} too far from target {t}"
            );
        }
    }

    #[test]
    fn test_squared_epsilon_insensitive() {
        let x = Array2::from_shape_vec((5, 1), vec![1.0, 2.0, 3.0, 4.0, 5.0]).unwrap();
        let y = array![2.0, 4.0, 6.0, 8.0, 10.0];

        let model = LinearSVR::<f64>::new()
            .with_c(10.0)
            .with_epsilon(0.0)
            .with_loss(LinearSVRLoss::SquaredEpsilonInsensitive)
            .with_max_iter(10000);
        let fitted = model.fit(&x, &y).unwrap();
        let preds = fitted.predict(&x).unwrap();
        assert_eq!(preds.len(), 5);
    }

    #[test]
    fn test_shape_mismatch() {
        let x = Array2::from_shape_vec((3, 1), vec![1.0, 2.0, 3.0]).unwrap();
        let y = array![1.0, 2.0];

        let model = LinearSVR::<f64>::new();
        assert!(model.fit(&x, &y).is_err());
    }

    #[test]
    fn test_invalid_c() {
        let x = Array2::from_shape_vec((3, 1), vec![1.0, 2.0, 3.0]).unwrap();
        let y = array![1.0, 2.0, 3.0];

        let model = LinearSVR::<f64>::new().with_c(0.0);
        assert!(model.fit(&x, &y).is_err());
    }

    #[test]
    fn test_negative_epsilon() {
        let x = Array2::from_shape_vec((3, 1), vec![1.0, 2.0, 3.0]).unwrap();
        let y = array![1.0, 2.0, 3.0];

        let model = LinearSVR::<f64>::new().with_epsilon(-0.1);
        assert!(model.fit(&x, &y).is_err());
    }

    #[test]
    fn test_invalid_intercept_scaling() {
        let x = Array2::from_shape_vec((3, 1), vec![1.0, 2.0, 3.0]).unwrap();
        let y = array![1.0, 2.0, 3.0];

        let model = LinearSVR::<f64>::new().with_intercept_scaling(0.0);
        assert!(model.fit(&x, &y).is_err());

        // But with fit_intercept=false, intercept_scaling is ignored.
        let model = LinearSVR::<f64>::new()
            .with_fit_intercept(false)
            .with_intercept_scaling(0.0);
        assert!(model.fit(&x, &y).is_ok());
    }

    #[test]
    fn test_predict_feature_mismatch() {
        let x = Array2::from_shape_vec((3, 2), vec![1.0, 0.0, 2.0, 0.0, 3.0, 0.0]).unwrap();
        let y = array![1.0, 2.0, 3.0];

        let fitted = LinearSVR::<f64>::new()
            .with_max_iter(5000)
            .fit(&x, &y)
            .unwrap();

        let x_bad = Array2::from_shape_vec((3, 1), vec![1.0, 2.0, 3.0]).unwrap();
        assert!(fitted.predict(&x_bad).is_err());
    }

    #[test]
    fn test_has_coefficients() {
        let x =
            Array2::from_shape_vec((4, 2), vec![1.0, 0.0, 2.0, 0.0, 3.0, 0.0, 4.0, 0.0]).unwrap();
        let y = array![1.0, 2.0, 3.0, 4.0];

        let fitted = LinearSVR::<f64>::new()
            .with_max_iter(5000)
            .fit(&x, &y)
            .unwrap();
        assert_eq!(fitted.coefficients().len(), 2);
    }

    #[test]
    fn test_fit_intercept_false_zero_intercept() {
        let x = Array2::from_shape_vec((5, 1), vec![1.0, 2.0, 3.0, 4.0, 5.0]).unwrap();
        let y = array![2.0, 4.0, 6.0, 8.0, 10.0];

        let fitted = LinearSVR::<f64>::new()
            .with_fit_intercept(false)
            .with_max_iter(5000)
            .fit(&x, &y)
            .unwrap();
        assert_relative_eq!(fitted.intercept(), 0.0);
    }

    #[test]
    fn test_pipeline_integration() {
        let x = Array2::from_shape_vec((4, 1), vec![1.0, 2.0, 3.0, 4.0]).unwrap();
        let y = array![3.0, 5.0, 7.0, 9.0];

        let model = LinearSVR::<f64>::new().with_max_iter(5000);
        let fitted_pipe = model.fit_pipeline(&x, &y).unwrap();
        let preds = fitted_pipe.predict_pipeline(&x).unwrap();
        assert_eq!(preds.len(), 4);
    }
}