File size: 70,945 Bytes
cfc7a54
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
# SIFQ — Giải thích toàn bộ Research & Mapping vào Code

> Tài liệu này đọc song song với `sifq_pdf.txt` và source code trong `src/`.
> Mỗi khái niệm trong paper đều có pointer đến file code tương ứng.

---

## 1. Vấn đề SIFQ giải quyết

### NFIQ2 có 3 nhược điểm lớn

**Nhược điểm 1 — Matcher cũ làm target:**  
NFIQ2 train supervised với target là match score của VeriFinger (minutiae-based, ~2010). Khi matcher hiện đại như MDGT/DINOv2 đánh giá ảnh khác, quality score không còn tương quan.

**Nhược điểm 2 — Feature thiên về optical sensor:**  
5 feature tay của NFIQ2 (OCL, LCS, FDA, RVU, OFL) thiết kế cho contact optical. Fingerprint contactless hoặc flat press sẽ có Q thấp dù matcher hiện đại match được.

**Nhược điểm 3 — Sensor bias:**  
Cùng một ngón tay chụp bằng 2 sensor khác nhau → NFIQ2 cho 2 điểm chênh nhau 20–30 đơn vị. Metric đang đo sensor signature thay vì biometric quality thực sự.

### SIFQ giải quyết bằng cách nào?

Thay vì dùng matcher score làm label, SIFQ dùng **3 tín hiệu tự giám sát bổ trợ nhau**:

| Signal | Vai trò | Ràng buộc |
|--------|---------|-----------|
| `L_mat` — Matcher-as-teacher | Q phải tương quan với embedding quality của MDGT | Task-relevant |
| `L_sens` — Cross-sensor invariance | Q phải giống nhau cho cùng ngón tay ở 2 sensor khác nhau | Sensor-invariant |
| `L_deg` — Controlled degradation | Q phải giảm khi ảnh bị degraded (blur, noise, occlusion...) | Ordinal grounding |

Ba tín hiệu tạo **equilibrium**: Q không thể chỉ distill MDGT (vì L_sens ngăn sensor bias), không thể chỉ đo sensor difference (vì L_deg buộc phải predict degradation order), không thể chỉ rank degradation synthetic (vì L_mat kéo về matcher reality).

**Điểm cân bằng duy nhất là Q đo đúng biometric quality thực sự.** Bảng dưới cho thấy nếu thiếu bất kỳ signal nào:

| Nếu chỉ có... | Thì model sẽ... |
|--------------|----------------|
| Chỉ `L_mat` | Distill MDGT → không sensor-invariant |
| Chỉ `L_sens` | Cho tất cả ảnh cùng score → L_pair = 0 dễ dàng |
| Chỉ `L_deg` | Học ranking degradation synthetic, không transfer sang ảnh real |
| `L_mat + L_sens` | Không có ordinal grounding → score không có ý nghĩa tuyệt đối |
| `L_sens + L_deg` | Không liên kết với matcher reality |

**⚠️ Bài học từ v15–v19 (score collapse):**
Ngoài việc cần đủ 3 signal, còn cần đảm bảo:
1. **`L_mat` cần per-image quality gradient thực sự**: Raw cosine với multi-sensor prototype
   (proto_max_batches=0) → cosine varies with quality, not sensor. Partial prototype
   (proto_max=150) → cosine bị sensor-bias → không phân biệt quality → collapse.
2. **`L_deg` phải áp dụng cho SD302 chứ không chỉ FVC**: FVC-only L_deg → model học
   quality function cho FVC nhưng SD302 không có per-image signal → SD302 collapse ở inference.
   L_spread_ds (per-dataset spread, T27) ngăn score anchoring → an toàn để bật L_deg full cho SD302.
3. **`L_spread_ds` là batch-level, không đủ**: Batch spread thỏa mãn bằng arbitrary ordering,
   không phải quality. Phải kết hợp với per-image signal từ L_mat + L_deg.

---

## 2. Kiến trúc — 4 Components

```
  Image [B,1,224,224]


  ┌────────────┐
  │  Backbone  │  TinyViT-5M, grayscale
  │  TinyViT   │  forward_spatial() → [B,196,320]
  └────────────┘

         ├── mean(dim=1) → [B, 320] features ──────────────────────────┐
         │                                                              │
         │ (SpatialConceptHead path: [B,196,320])                       ▼
         ▼ (ConceptHead legacy:      mean-pool → [B,320])    ┌──────────────────┐
  ┌──────────────────┐                                        │  SensorDisc      │
  │  ConceptHead /   │  [B,?]→ … →[B,6] ∈ [0,1]             │  GRL(λ) →        │
  │ SpatialConcept   │  see §2.2 for both variants           │  [B,64]→[B,Ns]   │
  └──────────────────┘                                        └──────────────────┘
         │ concepts [B, 6]                                          │ sensor_logits
         ▼                                                          ▼
  ┌──────────────────┐                                  adversarial sensor loss
  │ ScoreAggregator  │  [B,6]→[B,32]→[B,1]→Sigmoid→×100
  └──────────────────┘
         │ Q [B,1] ∈ [0, 100]

  outputs = { "score": Q, "concepts": C, "sensor_logits": S, "features": F }
```

---

### 2.1 Backbone — `src/models/backbone.py`

| | |
|---|---|
| **Model** | `timm: tiny_vit_5m_224.dist_in22k` |
| **Input** | `[B, 1, 224, 224]` — grayscale |
| **Output** | `[B, 320]` — global average of spatial tokens |

`forward_spatial()` extracts 14×14 token features → mean pool → `[B, 320]`.

- TinyViT-5M (not 2M): 2M insufficient capacity for 3 simultaneous losses
- Not shared with MDGT: avoids circular dependency; on-sensor deployment target (<50ms ARM Cortex-A72)

---

### 2.2 Concept Head — `src/models/concept_head.py`

Two implementations exist. The active version depends on the training flag:

**`ConceptHead` (legacy — v16–v26, default `--spatial-concept-head` not set):**
```
Input:  [B, 320]  — globally-pooled backbone output
Linear(320→256) → LayerNorm → GELU → Linear(256→6) → Sigmoid
Output: [B, 6],  each concept ∈ [0, 1]  (high = better quality)
```
Loses all spatial structure before concept prediction.

**`SpatialConceptHead` (v27+, enabled via `--spatial-concept-head`):**
```
Input:  [B, 196, 320]  — backbone.forward_spatial() (14×14 token map)
Shared trunk : Linear(320→128) → LayerNorm → GELU    → [B, 196, 128]
Per-concept  : 6 × Linear(128→1) → mean(dim=1)       → [B, 6]
Activation   : Sigmoid                                 → [B, 6] ∈ (0, 1)
```
Each concept projection learns which spatial regions matter (ridge breaks for
`continuity`, orientation edges for `orientation_coherence`, bifurcation patches
for `minutiae_reliability`). Separate weights per concept reduce entanglement.

`SIFQ.forward()` auto-dispatches via `concept_head.uses_spatial` attribute —
backward-compatible with all existing checkpoints.

**6 concepts (all decrease with degradation — T25):**

| Idx | Name | Meaning | Supervised by |
|-----|------|---------|---------------|
| 0 | `orientation_coherence` | Ridge flow consistency | dry_skin, wet_press (T39) |
| 1 | `ridge_valley_clarity` | Sharp ridge-valley boundaries | blur, jpeg, wet_press, **noise (T42)** |
| 2 | `continuity` | Unbroken ridge lines | blur, jpeg |
| 3 | `noise_level` | Low noise (↑ = cleaner — T25) | noise |
| 4 | `contrast_uniformity` | Even foreground contrast | dry_skin |
| 5 | `minutiae_reliability` | Reliable minutiae extraction | occlusion, wet_press |

**T42 changes (v32):**
- `continuity[2]` removed from `dry_skin`: dry_skin causes ridge fragmentation — at TinyViT patch scale (16×16 px) this looks identical to Gaussian noise texture (both = local high-frequency disruptions). When concept[2] is pulled down by dry_skin, concept[3] (noise_level) follows via shared feature paths → spurious crosstalk (v31 Spearman −0.758). Fix: dry_skin supervised by `contrast[4] + orientation[0]` only — these operate at multi-patch/regional scale, not shared with noise.
- `clarity[1]` added to `noise`: TinyViT 16×16 patch embed averages out pixel Gaussian noise (σ=5–30) → concept[3] receives almost no gradient signal from noise alone (v31: noise→noise_level ρ = −0.081). Gaussian noise blurs ridge-valley boundaries at patch scale — clarity[1] detects this → strong ViT signal → provides gradient that reinforces concept[3] in the correct direction.

Concepts are not required to be independent — blur reducing both clarity and continuity is physically correct. `L_ortho` only prevents extreme redundancy.

---

### 2.3 Score Aggregator — `src/models/aggregator.py`

```
Linear(6→32) → GELU → Linear(32→1) → Sigmoid → ×100
Output: [B, 1] ∈ [0, 100]  (NFIQ2-compatible scale)
```

`Q(x) = Aggregator(concepts(backbone(x)))`

> **Important:** `outputs["score"]` is already in [0, 100]. Do not multiply by 100 again. Old bug did this → loss thresholds were off by 10×.

---

### 2.4 Sensor Discriminator — `src/models/sensor_discriminator.py`

```
GRL(λ) → Linear(320→64) → ReLU → Linear(64→N_sensors)
```

**Gradient Reversal Layer** (`src/models/grad_reverse.py`):
- Forward: `f(x) = x`
- Backward: `∂L/∂x → −λ · ∂L/∂x`

The discriminator is trained to classify sensor identity. GRL forces the backbone to produce features the discriminator *cannot* distinguish → backbone sheds sensor signature.

> **TODO:** Paper specifies discriminator should receive intermediate concept-branch features, not raw backbone features. Currently uses backbone output. Not yet implemented.

---

## 3. Ba Loss Functions

### 3.1 `L_mat` — Matcher-as-Teacher

**File:** `src/losses/matcher_teacher.py`  
**Teacher model:** `src/training/mdgt_teacher.py` — load từ `pad/TRAM-downstream/checkpoint/checkpoints_dinov2_tram/best_eer.pt` (DINOv2 ViT-S/14, frozen hoàn toàn)

**Công thức:**

$$q_{mat}(x) = \frac{\cos(\text{MDGT}(x),\ c_y) - \mu_y}{\sigma_y}$$

$$L_{mat} = \text{Huber}(Q_{normalized}(x),\ q_{mat}(x))$$

Trong đó:
- $c_y$ = prototype của identity $y$ (trung bình L2-normalized embeddings)
- $\mu_y, \sigma_y$ = mean/std cosine similarity trong class $y$
- $Q_{normalized} = (Q/100 - 0.5) \times 2$ — đưa [0,100] về xấp xỉ [-1, 1]

**Code mapping:**

```python
# Compute prototype: trung bình embeddings per identity
proto = F.normalize(stack(emb_list).mean(0), dim=-1)

# v16 (T31): compute per-identity cosine stats (second pass, frozen teacher)
# stats[identity] = (mean_cos, std_cos)  — min_sigma = 0.02
stats = compute_identity_cos_stats(loss_mat, prototypes, proto_loader, device)

# q_mat: cosine similarity normalize per-identity, clamp bằng tanh
cos = dot(emb[i], proto[identity_id])
z   = (cos − mu_i) / sigma_i
q_mat[i] = tanh(z)              # ∈ (−1, 1), zero-centred

# Loss
pred_normalized = (score / 100.0 - 0.5) * 2.0
return HuberLoss(pred_normalized, q_mat)
```

**Vấn đề v15 (stats=None):** Raw cosine ≈ 0.85 cho mọi ảnh → teacher kéo tất cả score về 92.5, conflict với L_spread → score collapse. Xem TODO T31.

**v16 fix (T31):** Per-identity stats normalization + tanh → teacher target zero-centred, tương thích L_spread. Ảnh tốt hơn trung bình identity → target > 0 → Q > 50. Ảnh kém hơn → target < 0 → Q < 50.

**v18 revert (T33a — `--no-mat-stats`):** Skip per-identity cosine stats hoàn toàn. Tất cả images (FVC + SD302) đều dùng raw cosine làm L_mat target → không còn FVC/SD302 quality signal asymmetry. Root cause v17: FVC images có stable per-image tanh targets (từ FVC-only stats) nhưng SD302 images chỉ có raw cosine constant ≈ 0.85 → model học shortcut "FVC=variable quality, SD302=fixed quality ~52.7" → GRL + L_pair triệt tiêu SD302 quality features vì không có L_mat gradient để duy trì chúng. Revert về v14 design: tất cả images dùng raw cosine → L_mat chỉ còn tác dụng anchor mean (không gây asymmetry).

---

### 3.2 `L_sens` — Cross-Sensor Invariance

**File:** `src/losses/sensor_invariance.py`

**Công thức:**

$$L_{pair} = \max(0,\ |Q(x_{s1}) - Q(x_{s2})| - \delta)$$

$$L_{adv} = -\log D(\text{sensor} \mid f(x))$$

$$L_{sens} = L_{pair} + \lambda_{adv} \cdot \min(L_{adv},\ \ln N_{sensors})$$

Với $\delta = 0.05$, $\lambda_{adv} = 0.3$. Clamp $\min(L_{adv}, \ln N)$ ngăn GRL anti-correlation (T16).

**Code mapping (v11 — GRL anti-correlation fix):**

```python
# Pairs: cùng identity_id + finger_id, khác sensor_id (built trong batch)
l_pair = relu((score_s1 - score_s2).abs() - 0.05).mean()

# Adversarial: backbone phải fool sensor discriminator
l_adv = cross_entropy(sensor_logits, sensor_labels)

# GRL anti-correlation fix (T16): clamp l_adv tại random-guess baseline.
# Nếu CE > log(num_sensors): backbone đã đủ confuse discriminator, dừng
# gradient reversal — tránh việc backbone học anti-encode sensor identity.
rand_ce = log(sensor_logits.size(1))   # log(32) ≈ 3.47
l_adv_clamped = l_adv.clamp(max=rand_ce)

# Returns (total, l_pair_detached, l_adv_detached) for diagnostics
return l_pair + 0.3 * l_adv_clamped, l_pair.detach(), l_adv.detach()
```

**Quan sát v11 trước khi fix (T16):** `l_adv` dao động 7–16 trong S2–S4, gấp 2–4× random baseline log(32)≈3.47. Diễn giải: backbone không chỉ "mờ" sensor mà học predict sai sensor với high confidence (anti-correlation). Discriminator thích ngưỡc lại → cái cycle không dừng. Clamp ngăn vòng lặp này.

**Pair building — `train_sifq.py::build_pair_indices()`:**  
Group batch theo `(identity_id, finger_id)` → tất cả cặp khác sensor_id trong cùng group.

**Quan hệ giữa `L_pair` và `L_adv` — hai cơ chế bổ trợ:**

| | `L_pair` | `L_adv` (GRL) |
|--|---------|---------------|
| Ép ở tầng nào | **Output** — Q score phải bằng nhau | **Feature space** — backbone không được encode sensor info |
| Cách hoạt động | Hinge loss trực tiếp trên Q | Gradient reversal qua discriminator |
| Tại sao cần cả 2 | Chỉ ép output: backbone vẫn giữ sensor info ẩn trong features | Chỉ ép feature: Q vẫn có thể tái encode sensor qua concept head |

**Dataset novelty:** SIFQ là method đầu tiên đưa explicit invariance constraint vào quality learning. NFIQ2 không có mechanism này.

---

### 3.3 `L_deg` — Controlled Degradation Ranking

**File:** `src/losses/degradation_ranking.py`  
**Pipeline:** `src/data/degradation.py`

**Công thức:**

$$L_{rank} = \max(0,\ Q(x_{high}) - Q(x_{low}) + m) + \max(0,\ Q(x_{low}) - Q(x_{clean}) + m)$$

$$L_{concept\_deg} = \sum_{(deg, c_{target})} \text{Huber}(c_{target}(x_{low}),\ c_{target}(x_{high}) + 0.1)$$

$$L_{deg} = L_{rank} + \gamma \cdot L_{concept\_deg}$$

Với $m=0.1$ (ranking margin), $\gamma=0.5$.

**Degradation concept map (index) — T42 (v32) — current:**

```python
DEGRADATION_CONCEPT_MAP = {
    "blur":      [1, 2],   # clarity↓, continuity↓
    "noise":     [1, 3],   # clarity↓ (T42 co-target), noise_level↓
    "jpeg":      [2, 1],   # continuity↓, clarity↓
    "occlusion": [5],      # minutiae_reliability↓ only  ← T38: reverted from [5,2] to paper
    "dry_skin":  [4, 0],   # contrast_uniformity↓, orientation_coherence↓  ← T42: removed continuity
    "wet_press": [1, 5, 0], # clarity↓, minutiae_reliability↓, orientation_coherence↓  (T39)
}
```

**T42 physical reasoning (v32):**

| Change | Before | After | Physical reason |
|--------|--------|-------|-----------------|
| `noise` | `[3]` | `[1, 3]` | TinyViT patch embed (16×16 px) averages out pixel Gaussian noise σ=5–30 → concept[3] gradient ≈ 0 alone. But noise **also blurs ridge-valley edges** at patch scale (clarity[1]) → ViT detects this → provides training signal that anchors concept[3] in the correct direction. v31: noise→noise_level ρ = −0.081 (very weak). |
| `dry_skin` | `[4, 2, 0]` | `[4, 0]` | `continuity[2]` operates at **local patch scale** — same as Gaussian noise texture. dry_skin ridge fragmentation and pixel noise both create "high-frequency local disruptions" in 16×16 patches → shared backbone feature path → when concept[2] is pulled ↓ by dry_skin, concept[3] follows → spurious crosstalk (v31: dry_skin→noise_level ρ = −0.758). Removing continuity forces backbone to use `contrast[4]` (regional brightness gradient) + `orientation[0]` (multi-patch coherence field) — both are **regional-scale** features, not shared with noise. |

**T38 fix — wet_press `[1, 4]` → `[1, 5]` (reverted to paper design):**

Root cause: `MorphologicalDilator` used a fixed 3×3 kernel — at 500 DPI, 1 erode iteration = ~1 px ridge expansion, too subtle. Model read level 1 as "good ink coverage" → `minutiae_reliability(level_1) > level_0` → non-monotonic → ρ > 0 (wrong). Consistently wrong: v14 = +0.565, v20 = +0.246, v21 = +0.307.

Fix: Scale kernel + blur with severity in `degradation.py`: `ks = 3 + 2*iterations` (5×5/7×7/9×9), `sigma = 1.0 + 0.8*iterations` (1.8/2.6/3.4). Level 1 now visibly impairs bifurcations. Concept map KEPT at `[1, 5]` as designed.

**T38 fix — occlusion `[5, 2]``[5]` (reverted to paper design) + coverage 55% → 40%:**

T26 added `continuity[2]` because occlusion signal was near-zero, and pushed coverage to 55%. Root cause of weak signal was actually the **unseeded random block positions in eval** — different levels placed blocks in different spots → Spearman ρ incoherent. Now fixed by `np.random.seed(level)` in `compute_crosstalk_matrix` (T38b). With deterministic eval, `[5]` alone provides a clean signal. Coverage reverted to paper range: `0.133 × level` → level 3 = 40% max.

**Note**: Training occlusion still uses random positions (generalization). Only eval is seeded.

**T30 (v15): SD302 concept-only L_deg** — Root cause regression v14: concept head chỉ thấy FVC texture khi bị degraded; eval trên SD302 images thất bại (noise→noise_level: -0.019).
- **T30a**: `concept_deg_gamma 0.5 → 2.0` — L_concept 4× mạnh hơn so với L_rank
- **T30b**: Apply degradation cho SD302 batch items, chỉ compute L_concept (không L_rank → không score collapse). Lần đầu tiên concept head thấy degraded SD302.
- **T30c**: `deg-every-n-steps 2` (giảm từ default 4, giới hạn bởi OOM: 3 forward pass/step với batch=64)

**DegradationPipeline — 6 loại, 4 levels (0=clean, 1–3 tăng dần):**

| Type | Kỹ thuật | Severity range |
|------|---------|----------------|
| `blur` | GaussianBlur kernel $k = 2\lfloor 0.5 + level \times 0.83 \rfloor + 1$ | σ tăng theo level |
| `noise` | Additive Gaussian, $\sigma = 5 + level \times 8.3$ | 5→30 |
| `jpeg` | JPEG compression quality = $90 - level \times 25$ | 90→15 |
| `occlusion` | White block $= 13.3\% \times level$ coverage | 13%→40% |
| `dry_skin` | `DrySkinSimulator`: contrast reduce + crack lines | severity 1→3 |
| `wet_press` | `MorphologicalDilator`: kernel scales 5×5/7×7/9×9, sigma 1.8/2.6/3.4 | levels 1→3 |

**Code trong training loop (v6 fix):**

```python
# Random sample 1 trong 6 types mỗi deg step
deg_type = random.choice(_DEG_TYPES)
level_lo  = random.randint(1, 2)
level_hi  = 3

# Tensor [B,1,H,W] float[0,1] → numpy [B,H,W] uint8
imgs_np = (images[:,0].cpu().numpy() * 255).astype(np.uint8)

# Apply degradation per image
imgs_low  = stack([deg_pipeline.apply(img, deg_type, level_lo) for img in imgs_np])
imgs_high = stack([deg_pipeline.apply(img, deg_type, level_hi) for img in imgs_np])

# Forward pass qua model
out_low  = model(imgs_low_tensor)
out_high = model(imgs_high_tensor)

l_deg = loss_deg(score_clean, out_low["score"], out_high["score"],
                 out_low["concepts"], out_high["concepts"], deg_type)
```

**Tất cả concepts GIẢM khi degradation tăng (v13 — T25 fix):**

```python
# T25: Bỏ if c_idx == 3 special case.
# Tất cả concepts: high = better quality, low = worse quality.
# noise_level mới: high = ít nhiễu = tốt (ngược v12)
for c_idx in DEGRADATION_CONCEPT_MAP[deg_type]:
    l_concept += huber(concepts_low[:, c_idx], concepts_high[:, c_idx] + 0.1)
    # concepts_low[c] > concepts_high[c] + 0.1
    # → more degraded image has lower concept score
```

**Trước v13 (v12, BUG):** `noise_level` được push tăng theo degradation level. ScoreAggregator resolve conflict bằng cách invert direction → noise_level=-0.168 (Track 4, sai chiều).

---

### 3.4 `L_ortho` — Concept Decorrelation

**File:** `src/losses/orthogonality.py`

$$L_{ortho} = \| \text{off\_diag}(\text{corrcoef}(C)) \|_F^2$$

```python
# Standardize per concept (zero-mean, unit-std across batch)
x = concepts - concepts.mean(dim=0, keepdim=True)
x = x / (x.std(dim=0, keepdim=True) + 1e-6)
corr = (x.T @ x) / max(1, x.shape[0] - 1)   # [6, 6] Pearson corr
off_diag = corr - eye(6)
return (off_diag ** 2).sum()                   # Frobenius² of off-diagonal
```

Regularize để concepts không hoàn toàn redundant. Không có weight — luôn cộng thẳng vào total loss.

---

### 3.5 Spread Loss (Uniformity) — thêm vào trong training loop

**Không có file riêng** — implement trực tiếp trong `scripts/train_sifq.py`.

```python
# uniform mode (v8 default) — force sorted Q-scores → linspace(10, 90)
q_sorted, _ = q_batch.sort()
target_unif = torch.linspace(10.0, 90.0, n_q, device=device)
l_spread = F.mse_loss(q_sorted / 100.0, target_unif / 100.0)
# Gradient ép trực tiếp phân bố score: ảnh tốt nhất → 90, xấu nhất → 10
```

Weight lịch sử: v15 = 2.0 → v16/v17 = 4.0 → **v18 revert 2.0 (T33b)****v21 = 3.0** (`--spread-weight 3.0`). Default code = 4.0. W_SPREAD áp dụng cho **CẢ HAI** `L_spread_global` và `L_spread_sd302`.

**Thay thế variance mode của v7** (`relu(10/100 - q_std/100)^2`, max=0.01) vốn bị bão hòa ngay từ đầu (spread=0.010 stuck toàn bộ training).

**v18 (T33b):** W_SPREAD revert về 2.0. W_SPREAD=4.0 cùng với gamma=2.0 làm tệ hơn v14 trong v15–v17.

**v21:** W_SPREAD=3.0 — trung dung giữa v14 (2.0) và v16/v17 (4.0), kết hợp với `--proto-max-batches 0` (full prototypes) và FVC-only L_deg.

---

### 3.6 Total Loss

$$L(t) = \alpha(t) \cdot L_{mat} + \beta(t) \cdot L_{sens} + \gamma(t) \cdot L_{deg} + L_{ortho} + w_{spread} \cdot (L_{spread} + L_{spread\_sd302})$$

Trong đó:
- $L_{spread}$: MSE toàn batch (sorted Q → linspace(10,90))
- $L_{spread\_sd302}$: MSE chỉ SD302 subset trong batch (≥8 ảnh) — ép SD302 spread riêng, tránh SD302 dùng FVC variation để thỏa spread (T27)
- $w_{spread}$ = 3.0 (v21) — áp cho **cả hai** thành phần spread

---

## 4. Training Stage Scheduler

**File:** `src/training/stage_scheduler.py`

| Stage | Epochs | α (mat) | β (sens) | γ (deg) | β/α | Mục đích |
|-------|--------|---------|----------|---------|-----|----------|
| S1 | 0–9 | 0.10 | **0.00** | 1.00 | 0 | Bootstrap quality ordering từ degradation |
| S1→S2 ramp | 10–19 | **0.10→0.28** | **0.00→0.27** | **1.00→0.64** | 0→0.96 | β ramps NHANH HƠN α — β/α → 1.0 ở S2 entry |
| S2 | 20–34 | 0.30 | 0.30 | 0.60 | **1.0** | Balanced: GRL pressure = mat pressure |
| S2→S3 ramp | 35–39 | **0.30→0.29** | **0.30→0.34** | **0.60→0.52** | >1.0 | β vượt α smooth |
| S3 + S4 | 40+ | 0.25 | 0.35 | **0.50** | **1.4** | β > α, γ floor = 0.50 |

> **v21 chạy 60 epochs** → S3 kết thúc ở ep59. S4 (ep60+) không được sử dụng. Code scheduler dùng `(40, 999)` làm sentinel nên logic không thay đổi — chỉ số epoch thực tế bị cắt ngắn bởi `--epochs 60`.

**GRL λ warm-up (DANN schedule):**

$$\lambda(p) = \min\!\left(0.6,\ \frac{2}{1+e^{-10p}}-1\right), \quad p = \frac{\text{epoch}}{\text{total\_epochs}}$$

- ep0: λ=0.000
- ep10: λ=0.555
- ep12+: λ=0.600 (cạn max)

**Cosine LR decay (v8 mới):** LR decay từ 1e-4 → 5e-6 theo cosine. Trước với LR cố định → S4 oscillation (v7: loss tăng từ 0.53 lên 1.18 trong ep61-80).

**Vì sao β/α phải = 1.0 ngay từ S2 (v11 vs v9):**  
V9 bug: β/α = 0.50 trong S2 (α=0.40, β=0.20). PolyU contact vs contactless là cross-modality pairs cực mạnh — GRL với β quá thấp không đủ mạnh để override L_mat pressure, backbone vẫn encode sensor/modality signature. Kết quả: L_sens oscillate 2–3 suốt S2/S3. V11: β ramps nhanh hơn α (β Δ0.30 vs α Δ0.20 qua 10 epoch) → β/α = 1.0 NGAY KHI vào S2. GRL và L_mat có sức nặng bằng nhau từ đầu.  
**PolyU treatment (v11):** PolyU chỉ góp vào L_sens. L_mat và L_deg mask out PolyU records (không compute prototype, không apply degradation cho PolyU images).

**Vì sao β entry ramp quan trọng (v8 vs v7):**  
V7: β nhảy 0→0.20 đột ngột tại ep10 → L_sens explosion: 0.77 (ep10) → 11.32 (ep18), mất 20 epoch phục hồi.

**Vì sao γ floor = 0.50 (v8/v9) thay vì 0.40 (v7):**  
V7: L_deg plateau tại 0.104 từ ep43+ — chính xác bằng margin m=0.10. γ=0.50 giữ grounding mạnh hơn.

---

## 5. Dataset Pipeline

**Files:** `src/data/nist302_loader.py`, `src/data/fvc_loader.py`

```
dataset/
├── 302a/images/challengers/     → sensors: A, B, C, D, E, F, G, H (8 sensors)
│   └── A/roll/png/                  filename: {subject}_{sensor}_{captype}_{finger}.png  (4 tokens)
├── 302b/images/baseline/        → sensors: R_500, S_500, U_500, V_500 (4 sensors)
│   └── U/500/roll/png/              filename: {subject}_{sensor}_{dpi}_{captype}_{finger}.png  (5 tokens)
├── nist_302d/images/auxiliary/  → sensors: flat_K, flat_L, flat_M, flat_P (4 sensors)
└── FVC_Dataset/
    ├── FVC2002/Dbs/Db1_a|Db2_a...  filename: {subject}_{impression}.tif
    └── FVC2004/Dbs/DB1_A|DB2_A...
```

**Đầy đủ sau khi fix: 37,607 ảnh, 22 sensors**

| Dataset | Ảnh | Sensors | Role | Signal |
|---------|------|---------|------|--------|
| SD302-A (302a) | 13,630 | A–H (8) | Cross-sensor invariance | L_sens, L_mat |
| SD302-B (302b) | 11,796 | R,S,U,V (4) | Cross-sensor invariance | L_sens, L_mat |
| SD302-D (302d) | 5,141 | K,L,M,P (4) | Cross-sensor invariance | L_sens, L_mat |
| FVC2002 | 3,520 | 4 DBs | Matcher-teacher quality anchor | L_mat, L_deg |
| FVC2004 | 3,520 | 4 DBs | Matcher-teacher quality anchor | L_mat, L_deg |
| PolyU (contact) | 2,976 | 1 (polyu_contact) | Cross-modality hardest invariance | **L_sens ONLY** |
| PolyU (contactless) | 2,976 | 1 (polyu_contactless) | Cross-modality hardest invariance | **L_sens ONLY** |


> **† Lưu ý cột `noise_level`**: Giá trị theo hướng **model concept[3]**: `Cao` = ít nhiễu = ảnh sạch = tốt (ngược chiều với lượng noise vật lý trong ảnh). Bảng gốc dùng hướng image-domain (Thấp noise=sạch=tốt), nhưng đã được đổi sang hướng concept để nhất quán với kỳ vọng model output. Ví dụ: SD302a rất sạch → concept[3] dự kiến = `Cao`; FVC2004 DB3 nhiều noise → concept[3] dự kiến = `Thấp`.

**Tổng v11: 43,559 ảnh, 24 sensors, 2,612 eligible cross-sensor anchor groups**

**Tổng v14: 42,683 ảnh, 29 sensors** (sau khi loại `R_1000_slap`, `R_500_slap`, `S_500_slap` — ~1,700 non-segmented slap records). Anchor groups: 2,336 eligible.

**PolyU Cross-Fingerprint Database (v9+):**  
336 subjects × 6 impressions × 2 modalities = 5,952 ảnh. Cùng ngón tay được chụp bằng CMOS camera (contactless) và URU sensor (contact). Đây là cross-modality pairs mạnh nhất cho L_sens — harder invariance target so với cross-brand sensors trong SD302. File: `src/data/polyu_loader.py`.

**Tại sao SD302 không góp vào L_deg:**  
SD302 toàn ảnh chất lượng cao (protocol NIST ảnh cần đạt chuẩn). Degradation synthetic trên nó là giả tạo, không phản ánh quality variation thực. FVC có 8 lần chụp cùng ngón tay với **chất lượng tự nhiên dao động** → MDGT xếp hạng thực sự giữa các impression → L_deg có signal vật lý. Thiết kế tách biệt vai trò: SD302 = sensor role, FVC = quality role.

**Tại sao FVC cần thiết cho L_mat:**  
SD302 chỉ có ảnh chất lượng cao → q_mat ≈ constant → L_mat tầm thường. FVC có 8 lần chụp cùng ngón tay (chất lượng thực sự dao động) → MDGT có thể xếp hạng các lần chụp trong cùng subject → q_mat có signal thực sự.

**Metadata parsing:**

| Filename format | Tokens | identity_id | finger_id | sensor_id |
|----------------|--------|-------------|-----------|----------|
| SD302-A: `00002303_A_roll_05.png` | 4 | `00002303` (stem[0]) | `F05` (stem[-1]) | path-based `A_roll_png` |
| SD302-B/D: `00002401_U_500_roll_07.png` | 5 | `00002401` | `F07` | `U_500_roll` |
| FVC: `042_3.tif` | 2 | `2002_db1_0042` | `f01` | `fvc2002_db1` |

**Audit kết quả (kiểm tra thực tế):**

- **Finger ID nhất quán across sensors ✅**: 2,000 (subject, finger) pairs xuất hiện ở ≥2 sensors trong cả 302a và 302b
- Ví dụ: subject `00002303`, finger `05` → xuất hiện ở cả 8 sensors A–H → cặp L_sens cực tốt
- 200 subjects xuất hiện trong cả 302b và 302d → cross-subset L_sens pairs
- FVC: mỗi subject có 1 ngón tay duy nhất (finger_id = `f01`) → không có cross-sensor pairs

**Bug đã fix: 302a parser:**  
Parser cũ yêu cầu `len(tokens) >= 5` nhưng 302a chỉ có 4 tokens → **13,630 ảnh bị bỏ hoàn toàn**. Đã sửa thành `>= 4`. Sensor_id từ path vẫn đúng (`A_roll_png`, `B_roll_png`...).

**Risk R1 (cập nhật):** Finger_id đã được xác nhận là nhất quán across sensors trong NIST SD302. Protocol NIST yêu cầu cùng ngón tay trình cho tất cả sensor → finger number có ý nghĩa vật lý. ✅ Không còn là risk.

---

## 6. Evaluation 5 Tracks

### Track 1 — Error Rejection Curve (Primary)

**File:** `src/evaluation/erc.py`

Sắp xếp ảnh theo Q tăng dần → reject bottom x% → compute FNMR tại FMR=1e-4 trên phần còn lại → plot curve.

**AUC_ERC: thấp hơn = tốt hơn.**

Kết quả v5: SIFQ=0.8884 vs Random=0.8937 → hơn Random 0.5% — quá yếu.  
Kỳ vọng v6: L_deg hoạt động → ảnh degraded được score thấp hơn → reject đúng → FNMR giảm nhanh hơn.

---

### Track 2 — Sensor Invariance (Core novelty)

**File:** `src/evaluation/sensor_invariance.py`

**KS statistic** (Kolmogorov-Smirnov) giữa Q distributions của 2 sensors: `KS = 0` → giống hoàn toàn.

**Pearson correlation** giữa Q(x_s1) và Q(x_s2) cho paired samples: cao hơn = nhất quán hơn.

| Metric | v11 | v12 | v13 | v14 | v15 | v16 | v17 | v18 actual | v20 actual | **v21 target** |
|--------|-----|-----|-----|-----|-----|-----|-----|------------|------------|--------|
| mean_ks | 0.557 | 0.510 | 0.634 ❌ | **0.263 ✅** | 0.276 ❌ | 0.527 ❌ | 0.616 ❌ | **0.249 ⚠️** | **0.4936 ❌** | **≤0.27** |
| Pearson | — | -0.003 | +0.224 | **+0.294 ✅** | — | 0.046 ❌ | -0.002 ❌ | **0.007 ❌** | **0.0048 ❌** | **≥0.20** |
| q_std (inference) | — | — | — | ~15 | — | 0.33 ❌ | 0.60 ❌ | **0.01 ❌❌** | **0.007 ❌❌** | **>12** |
| n\_sensor\_pairs | 91 | 91 | 91 | 55 | 55 | 55 | **171** (302a added) | **171** | **171** | **≥100** (302a) |

> **v21 — 🚀 Active training** (T36 fix: full prototypes `--proto-max-batches 0` + FVC-only L_deg). Reasoning: v20 collapsed because `--deg-include-sd302` created a ~53.38 attractor; removing it restores the v14 mechanism where FVC genuine quality variation trains the backbone quality features that transfer to SD302 at inference.
**⚠️ v20 score collapse — KS misleading (same as v18 pattern)**: Training q_std≈18.3 (looks healthy) but inference q_std≈0.007. All sensors cluster near 53.38 (A_roll: std=0.007, range 53.364–53.403). Root cause: `--deg-include-sd302` applied L_rank to all clean SD302 images. L_rank only requires Q(clean)>Q(degraded), NOT that different clean images score differently → single ~53.38 attractor satisfies the constraint for all clean SD302. FVC-only signals (genuine quality variation) were blocked by this attractor. Fix: **v21 removes `--deg-include-sd302`**, keeps `--proto-max-batches 0`.
> **⚠️ v18 score collapse — KS misleading**: KS=0.249 vượt target "≤0.27" nhưng là **false positive** — tất cả sensors output score ≈54.64 (q_std≈0.01, range 54.62–54.65), phân bố trivially giống nhau vì model không phân biệt được images. Pearson=0.007 xác nhận: hoàn toàn không có cross-sensor ranking consistency. Root cause: T33 revert về raw cosine L_mat (constant ≈0.85 cho tất cả) + không có per-image quality signal → model collapse về trung bình ≈54.64. Concept head cũng collapse: orientation_coherence≈0.995 và noise_level≈0.982 stuck near 1.0; clarity≈0.032, continuity≈0.027, contrast≈0.04, minutiae≈0.027 stuck near 0.

**v13 regression trên KS**: Non-segmented slap sensors (`R_1000_slap`, `R_500_slap`, `S_500_slap`) tạo cụm score ~21.2 (std≈0) → kéo mean_KS lên 0.634. **Pearson cải thiện đáng kể** (+0.224 vs -0.003) — consistent relative ranking tốt hơn.

**v14 fix**: Loại non-segmented slap khỏi training và eval (`--exclude-sensor R_1000_slap,R_500_slap,S_500_slap`).

**v14 training actuals (60 epochs, from scratch):**
- `q_mean` = 52.6 (ep59) — centered tốt hơn v13 (~33 train-time)
- `q_std` = 15.5 (ep59) — plateau sau khi GRL equilibrium (peak S1 = 22.6 → drop về 14 khi β kick-in ep11)
- `l_pair` = 0.006 (ep59) — sensor invariance xuất sắc, gần bằng 0
- `l_mat` = 0.270 — converged, giảm từ 0.378 (ep0)
- 3-phase dynamics: S1 (spread explosion ep0-10) → Transition (GRL compression ep11-15) → Equilibrium (stable ep16-59)

**v14 eval results:** mean_KS = **0.263** ✅ (giảm từ 0.634), Pearson = **+0.294** ✅ (tăng từ 0.224).

**Đây là figure quan trọng nhất của paper.** Plot histogram Q per sensor — NFIQ2 expected các histogram tách biệt, SIFQ expected overlap cao.

---

### Track 4 — Concept Grounding (Interpretability claim)

**File:** `src/evaluation/concept_grounding.py`


Sweep degradation level 0→3, tính Spearman ρ giữa level và concept score.

**Cross-talk matrix lý tưởng:** diagonal-heavy — mỗi degradation chỉ ảnh hưởng đến target concept của nó (Spearman ρ mạnh âm trên diagonal, gần 0 off-diagonal).

**Kết quả v13 (Spearman ρ):**

| degradation | orient. | clarity | continuity | noise_lvl | contrast | minutiae |
|-------------|---------|---------|-----------|-----------|----------|----------|
| **blur** | +0.19 | **-0.49** ✅ | -0.10 | +0.10 | +0.22 | +0.11 |
| **noise** | +0.05 | +0.46 | -0.12 | **-0.53** ✅ | -0.06 | -0.20 |
| **jpeg** | -0.11 | +0.03 | +0.03 | -0.08 | -0.04 | -0.02 |
| **occlusion** | +0.10 | -0.13 | -0.15 | +0.10 | -0.19 | **-0.09** ⚠️ |
| **dry_skin** | **-0.35** ✅ | +0.06 | +0.01 | -0.13 | **-0.38** ✅ | +0.01 |
| **wet_press** | -0.06 | **-0.48** ✅ | +0.24 | -0.16 | +0.44 | +0.33 |

**Kết quả v14 (Spearman ρ):**

| degradation | orient. | clarity | continuity | noise_lvl | contrast | minutiae |
|-------------|---------|---------|-----------|-----------|----------|----------|
| **blur** | +0.11 | **-0.19** ✅ | -0.003 | +0.08 | +0.22 | **+0.50** ❌ |
| **noise** | +0.001 | -0.16 | +0.03 | **-0.02** ⚠️ | +0.15 | +0.05 |
| **jpeg** | -0.18 | -0.07 | -0.10 | +0.08 | -0.11 | -0.09 |
| **occlusion** | +0.06 | +0.04 | **-0.12** ✅ | +0.18 | +0.07 | **+0.18** ❌ |
| **dry_skin** | +0.19 | -0.28 | **-0.56** ✅⬆ | **-0.56** ✅⬆ | **-0.62** ✅⬆ | +0.36 |
| **wet_press** | **-0.28** ✅ | -0.13 | +0.37 | -0.16 | +0.20 | **+0.57** ❌ |

**Nhận xét v14 so với v13:**
-**dry_skin**: cải thiện mạnh — clarity(-0.28), continuity(-0.56), noise_level(-0.56), contrast(-0.62) đều đúng chiều và mạnh hơn
- ✅ **Track 2 KS**: 0.634 → 0.263 (cải thiện lớn nhờ loại slap sensors)
- ✅ **Pearson**: 0.224 → 0.294
- ⚠️ **noise→noise_level**: -0.533 (v13) → -0.019 (v14) — regression, gần bằng 0
- ⚠️ **blur→clarity**: -0.490 (v13) → -0.187 (v14) — yếu hơn đáng kể
- ❌ **wet_press→minutiae**: +0.326 (v13) → +0.565 (v14) — sai chiều, tệ hơn
- ❌ **occlusion→minutiae**: -0.093 (v13) → +0.183 (v14) — flip sang sai chiều

**Root cause v14 regression (T30):** Concept head chỉ thấy FVC texture khi degraded (FVC-only L_deg). Eval trên SD302 images — concept head không generalize. v15 fix bằng SD302 concept-only L_deg (T30b) + tăng gamma (T30a) + tăng tần suất (T30c). Eval v15: KS=0.2758 (tệ hơn v14), score range 24–56 (score collapse, không discrimination).

**Root cause v15 failure (T31 — v16 fix):** Score collapse toàn bộ 60 epochs (q_mean≈52.5, q_std≈16, score range 24–56). Chi tiết xem mục TODO T31.

---

**Kết quả v18 (Spearman ρ — CONCEPT COLLAPSE + INVERSION):**

| degradation | orient. | clarity | continuity | noise_lvl | contrast | minutiae |
|-------------|---------|---------|-----------|-----------|----------|----------|
| **blur** | -0.572 ❌ | **+0.611** ❌❌ | +0.575 ❌❌ | -0.583 ❌ | +0.552 ❌ | -0.487 ❌ |
| **noise** | -0.410 ❌ | +0.288 ❌ | +0.368 ❌ | **-0.296** ✓ | +0.298 ❌ | +0.630 ❌❌ |
| **jpeg** | -0.097 | **+0.222** ❌ | +0.194 ❌ | -0.268 | +0.140 | -0.279 |
| **occlusion** | -0.162 | -0.151 ✓ | +0.042 ❌ | -0.150 | +0.063 | **-0.133** ✓ |
| **dry_skin** | +0.183 | -0.066 ✓ | +0.091 ❌ | +0.086 | **-0.247** ✓ | +0.230 ❌ |
| **wet_press** | -0.118 | **+0.217** ❌ | +0.196 ❌ | -0.320 | +0.140 | **-0.245** ✓ |

**Chẩn đoán v18 concept collapse**: blur làm TĂNG clarity (+0.611) — đảo chiều hoàn toàn. Root cause: concept head bị stuck ở giá trị cực biên — orientation_coherence≈0.995 và noise_level≈0.982 (bão hòa gần 1.0); clarity≈0.032, continuity≈0.027, contrast≈0.04, minutiae≈0.027 (bão hòa gần 0.0). Khi degradation áp vào ảnh, không có dư địa để concept giảm thêm → bất kỳ thay đổi nhỏ nào đều là noise ngẫu nhiên, dẫn đến Spearman ρ không ổn định. Cùng cơ chế với score collapse (q_std≈0.01): model không phân biệt được input.

---

### Track 3 — Cross-Matcher Transfer (Reviewer defense)

Dùng Q của SIFQ (train với MDGT teacher) để tính ERC với **VeriFinger v12 match scores** (matcher chưa từng thấy).

Nếu AUC_ERC vẫn thấp → Q generalizable, không phải chỉ distill MDGT.  
Hiện tại: chưa có VeriFinger license → bỏ qua hoặc dùng matcher khác.

---

### Track 5 — Human Correlation (Optional)

Expert rate ảnh trên Likert 1–5, Spearman ρ với SIFQ score. Tốn công nhưng là evidence mạnh nhất.

---

## 7. File Map — Research Concept → Code

| Khái niệm trong paper | File code |
|----------------------|-----------|
| Backbone (TinyViT-5M) | `src/models/backbone.py` |
| Concept Head (6 concepts, legacy v16–v26) | `src/models/concept_head.py::ConceptHead` |
| Concept Head (spatial, v27+) | `src/models/concept_head.py::SpatialConceptHead` |
| Score Aggregator Q(x) | `src/models/aggregator.py` |
| Gradient Reversal Layer | `src/models/grad_reverse.py` |
| Sensor Discriminator D | `src/models/sensor_discriminator.py` |
| SIFQ full model wrapper | `src/models/sifq.py` |
| $L_{mat}$ (Huber + prototype) | `src/losses/matcher_teacher.py` |
| $L_{sens}$ (pair + adversarial) | `src/losses/sensor_invariance.py` |
| $L_{deg}$ (ranking + concept) | `src/losses/degradation_ranking.py` |
| $L_{ortho}$ (decorrelation) | `src/losses/orthogonality.py` |
| MDGT teacher loader | `src/training/mdgt_teacher.py` |
| Stage scheduler α/β/γ | `src/training/stage_scheduler.py` |
| DegradationPipeline (6 types) | `src/data/degradation.py` |
| DrySkinSimulator | `src/data/degradation.py::DrySkinSimulator` |
| MorphologicalDilator | `src/data/degradation.py::MorphologicalDilator` |
| Dataset loader SD302 A/B/D | `src/data/nist302_loader.py` |
| Dataset loader FVC 2002/2004 | `src/data/fvc_loader.py` |
| Dataset loader PolyU (contactless↔contact) | `src/data/polyu_loader.py` |
| Cross-sensor pair builder | `scripts/train_sifq.py::build_pair_indices()` |
| Spread/Uniformity loss | `scripts/train_sifq.py` (inline) |
| Prototype computation | `scripts/train_sifq.py::compute_teacher_prototypes()` |
| Training main loop | `scripts/train_sifq.py::main()` |
| Track 1 ERC evaluation | `src/evaluation/erc.py` |
| Track 2 Sensor invariance | `src/evaluation/sensor_invariance.py` |
| Track 4 Concept grounding | `src/evaluation/concept_grounding.py` |
| Run training v21 (T36: full-proto + FVC-only L_deg, 60ep) | `scripts/run_train_v21.sh` |
| Run training v14 (segmented only, 60ep from scratch) | `scripts/run_train_v14.sh` |
| Run training v13 (80ep, resume v12) | `scripts/run_train_v13.sh` |
| Run eval v13 (exclude non-seg slap) | `scripts/run_eval_v13.sh` → `eval_results/v13_seg/` |
| Run inference | `scripts/run_infer.py` |
| Run evaluation | `scripts/run_eval.py` |
| Visualize score milestones | `visualize_score_milestones.py --exclude-sensor ...` |
| **MDGTv2 local pipeline** (DINOv2+TRAM+GNN) | `src/models/mdgt/pipeline.py` |
| MDGT teacher wrapper (loads local pipeline, frozen) | `src/training/mdgt_teacher.py` |
| Pre-cache teacher embeddings (one-time, startup) | `scripts/train_sifq.py::precompute_teacher_embeddings()` |
| Image preloading (RAM cache, uint8) | `scripts/train_sifq.py::RecordDataset._preload_images()` |
| GPU-native degradation (blur/noise/occlusion) | `scripts/train_sifq.py::_degrade_gpu()` |
| CPU degradation with preloaded numpy (jpeg/dry_skin/wet_press) | `scripts/train_sifq.py::_degrade_from_np()` |
| Vectorized prototype cosine (1 GPU sync vs ~100) | `src/losses/matcher_teacher.py::_compute_q_mat()` |

---

## 8. Luồng dữ liệu trong 1 training step (v21 — current)

**Cấu hình v21:** batch=128, W_SPREAD=3.0, deg_every_n_steps=2, deg_max_images=32, no-mat-stats, proto-max-batches=0, gpus=0+1 (DataParallel)

**Startup (1 lần trước khi train):**
```
1. RecordDataset._preload_images() → np.empty([N, 224, 224], uint8) ~2 GB RAM
   → __getitem__ trở thành: arr.copy().float().unsqueeze(0) / 255.0 (không đọc disk)

2. precompute_teacher_embeddings(teacher, train_ds) → Tensor[N, 256] CPU (float32)
   → Frozen DINOv2+TRAM+GNN chạy 1 lần duy nhất trên toàn bộ N records
   → Không bao giờ chạy lại trong training loop

3. compute_teacher_prototypes() → dict{identity_id → Tensor[256]} CPU
   → MDGT prototype per identity (L2-normalized mean embeddings) — 1 lần
```

**Training step:**
```
Batch (B=128, CrossSensorBatchSampler: 16 anchor pairs + fill)

├─ images_np = batch["images_np"]  ← [B, 224, 224] uint8, đã trong RAM (không từ GPU)
├─ images = batch["images"].to(device)  ← [B, 1, 224, 224] float32 trên GPU

├─ Forward SIFQ: model(images) → {score [B,1], concepts [B,6], sensor_logits [B,N_sensors], features [B,320]}

├─ L_mat: (chỉ SD302+FVC, bỏ PolyU)
│   ├─ emb_cache = emb_cache[record_idxs[_non_polyu_idx]]  ← CPU lookup, O(1)
│   ├─ _compute_q_mat VECTORIZED (không per-sample .item() sync):
│   │   ├─ proto_mat = stack([prototypes[id] for id in ids]).to(device)  ← 1 GPU transfer
│   │   ├─ cos_values = (emb * proto_mat).sum(dim=-1)                   ← 1 GPU kernel
│   │   ├─ cos_cpu = cos_values.cpu().tolist()                          ← 1 GPU sync
│   │   └─ apply tanh(z) nếu có stats, else raw_cosine (tất cả v21 dùng raw — --no-mat-stats)
│   └─ Huber(pred_normalized, q_mat)   pred_normalized = (Q/100 - 0.5) × 2

├─ L_sens:
│   ├─ build_pair_indices() → pairs (cùng finger, khác sensor) từ SD302 + PolyU trong batch
│   ├─ L_pair = relu(|Q(s1) - Q(s2)| - 0.05).mean()
│   ├─ L_adv = cross_entropy(sensor_logits, sensor_labels)  ← GRL đảo gradient
│   ├─ L_adv_clamped = L_adv.clamp(max=log(num_sensors))   ← T16: ngăn anti-corr
│   └─ (L_pair + 0.3 × L_adv_clamped, L_pair, L_adv)

├─ L_deg (mỗi deg_every_n_steps=2 bước, chỉ FVC — tối đa deg_max_images=32):
│   ├─ deg_type = random.choice([blur, noise, jpeg, occlusion, dry_skin, wet_press])
│   ├─ level_lo ∈ {1,2} random, level_hi = 3
│   │
│   ├─ GPU path (blur / noise / occlusion) — không transfer nào:
│   │   ├─ imgs_low  = _degrade_gpu(images[_fvc_idx], deg_type, level_lo)
│   │   └─ imgs_high = _degrade_gpu(images[_fvc_idx], deg_type, level_hi)
│   │
│   └─ CPU path (jpeg / dry_skin / wet_press) — đọc từ preloaded numpy, không GPU→CPU:
│       ├─ imgs_np = images_np[_fvc_idx]   ← uint8, đã trong RAM
│       ├─ apply cv2 ops per image (GaussianBlur/imencode/erode/etc.)
│       ├─ imgs_low  = stack(low_list).float().to(device) / 255.0   ← 1 CPU→GPU
│       └─ imgs_high = stack(high_list).float().to(device) / 255.0  ← 1 CPU→GPU

│   model(imgs_low), model(imgs_high) → 2 SIFQ forward thêm
│   L_rank + gamma * L_concept_deg → L_deg

├─ L_ortho: standardize concepts → (x.T @ x) / (B-1) → off_diag Frobenius²

├─ L_spread_global: uniform MSE(sorted Q_batch / 100, linspace(10,90) / 100)
├─ L_spread_sd302:  uniform MSE(sorted Q_sd302 / 100, linspace(10,90) / 100) — nếu ≥8 SD302

│   [epoch-end: log train_l_pair, train_l_adv, train_q_mean, train_q_std]

└─ total = α × L_mat + β × L_sens + γ × L_deg + L_ortho + 3.0 × L_spread_global + 3.0 × L_spread_sd302
           └── scaler.backward() → optimizer.step() → scaler.update()  (AMP fp16)
```

**v21 config summary:**

| Param | Giá trị | Ghi chú |
|-------|---------|---------|
| batch_size | 128 | DataParallel → 64/GPU |
| W_SPREAD | 3.0 | --spread-weight 3.0 |
| deg_every_n_steps | 2 | Mỗi 2 bước có 1 deg step |
| deg_max_images | 32 | Tối đa 32 FVC images/step |
| concept_deg_gamma | 0.5 | L_concept weight trong L_deg |
| proto_max_batches | 0 | Full dataset (không cap) |
| no_mat_stats | True | Raw cosine target (không tanh) |
| sd302_concept_weight | 0.0 | T30b disabled |
| deg_include_sd302 | False | FVC-only L_deg (T27/T36) |

---

## 9. Điểm còn thiếu / TODO

| # | Vấn đề | Impact | Todo |
|---|--------|--------|------|
| T1 | `L_adv` input nên từ concept-intermediate, không phải backbone features | Trung bình | Sửa `SIFQ.forward()` + `SensorDiscriminator` input dim 320→256 |
| T2 | `L_mat` signal yếu vì q_mat ≈ constant với SD302 | ✅ Đã cải thiện (v8) | FVC2002+2004 đã được thêm → 8 impressions/subject → q_mat có real ranking signal |
| T3 | Finger_id heuristic không verified | ✅ Đã xác nhận | Audit thực tế: finger ID nhất quán across sensors trong SD302. Không còn là risk. |
| T4 | Không có FVC2002/2004 và PolyU CL2CB | ✅ Đã thêm (v9) | FVC2002+2004 (7,040) + PolyU (5,952 contact+contactless). PolyU = cross-modality L_sens mạnh nhất. |
| T5 | Track 3 chưa có VeriFinger | Thấp | Dùng matcher khác (FLaRE, AFR-Net open-source) |
| T6 | `L_mat` tốn compute nhưng gần như không contribute khi chỉ có SD302 | ✅ Đã xử lý (v8) | FVC thêm vào, L_mat già trị hơn |
| T7 | `L_spread` threshold 10/100 tùy ý — variance mode trivially saturates | ✅ Đã fix (v8) | Uniform mode: MSE(sorted Q, linspace(10,90)), weight=2.0 |
| T8 | `L_pair` trong `L_sens` sparse vì batch random ít khi có cross-sensor pair | ✅ Đã fix (v7) | `CrossSensorBatchSampler`: mỗi batch đảm bảo k_cross=16 nhóm cóp 2 samples từ khác sensor. |
| T9 | GRL λ cố định + stage transition nhảy bậc — shock tại ep10 (β: 0→0.20 đột ngột) | ✅ Đã fix (v8) | DANN warm-up từ ep0; β ramp linear ep10–14 + ep30–34 thay vì nhảy bậc |
| T13 | v8: chỉ ramp β, α/γ vẫn step change tại ep15 → sens oscillate 3–3.5 từ ep18 | ✅ Đã fix (v9) | Ramp ALL THREE (α, β, γ) cùng lúc ep10–19; không có step change nào trong toàn bộ schedule |
| T10 | Score collapse: spread=0.010 stuck (max penalty) trong toàn bộ v7 | ✅ Đã fix (v8) | Uniform spread mode: gradient trực tiếp ép phân bố score |
| T11 | L_deg plateau 0.104 (tại margin boundary m=0.10) từ ep43+ trong v7 | ✅ Đã fix (v8) | γ floor = 0.50 (was 0.40) giữ deg grounding mạnh hơn |
| T12 | S4 oscillation: loss tăng từ 0.53 lên 1.18 trong v7 ep61-80 | ✅ Đã fix (v8) | Cosine LR decay 1e-4 → 5e-6 ngăn oscillation giai đoạn finetune |
| T14 | `l_pair``l_adv` không được log riêng — chỉ thấy `train_l_sens` tổng hợp | ✅ Đã fix (v11) | `SensorInvarianceLoss.forward()` trả về `(total, l_pair, l_adv)` tuple; train script log `train_l_pair`, `train_l_adv` riêng |
| T15 | `q_mean``q_std` không được log ra metrics.jsonl — không thể monitor score distribution | ✅ Đã fix (v11) | Train script tích lũy `q_sum`/`q_sq_sum`/`q_count` qua từng step → tính epoch-level `train_q_mean`, `train_q_std` |
| T16 | GRL anti-correlation: `l_adv` >> `log(num_sensors)` suốt S2–S4 (ep20–47: 7–16 vs random=3.47) → backbone học predict sai sensor với high confidence → GRL game cycling không hội tụ | ✅ Đã fix (v11) | Clamp `l_adv` tại `log(num_sensors)` trước khi đưa vào total. Khi backbone đã đủ confuse discriminator (CE ≥ log N), gradient reversal dừng lại → equilibrium ổn định thay vì anti-correlation cycle |
| T17 | **Score collapse v11: 95% ảnh kẹt tại ~58.3** (inference SD302: mean=59.8, std=6.5, median=58.3). Root cause: L_deg chỉ apply cho FVC (16% batch) → L_spread được thỏa mãn batch-level nhờ FVC variation, nhưng SD302 images không có ordinal grounding → model học "SD302 = medium quality ~58" mà không có gradient để phân biệt. Slap sensors (R,S) cho ~88 do có ít samples và backbone capture type signal; roll/flat sensors cho ~58.3 uniform → sensor bias residual. | ⚠️ Fix sai (v12) → Đúng (v13) | **v12 (sai):** Apply L_deg cho ALL datasets. SD302 images anchored tại ~28 (near degraded floor), tệ hơn. **v13 (đúng):** Revert FVC-only L_deg + thêm per-dataset L_spread riêng cho SD302 subset. |
| T18 | **`MorphologicalDilator` sai chiều cho wet_press**: Dùng `cv2.dilate` thay vì `cv2.erode`. NIST fingerprints có ridges tối (dark ridges, light valleys). `cv2.dilate` expand vùng sáng → shrink ridges (mô phỏng ngón tay KHÔ, không phải ướt). Wet press cần ridges phình to, lấn vào valleys → phải dùng `cv2.erode` (expand dark regions). Bug giải thích tại sao Track 4 wet_press → clarity=+0.219 (sai chiều, phải âm): dilated image có ridges mảnh hơn → model thấy valleys rõ hơn → clarity tăng. | ✅ Fix (v12) | Đổi `cv2.dilate``cv2.erode` trong `MorphologicalDilator.__call__`. Erode expand ridges tối → ridges merge vào valleys → clarity giảm đúng chiều. |
| T19 | **`jpeg` và `occlusion` concept signal gần 0** (Track 4 eval v11: jpeg→clarity=-0.045, occlusion→minutiae=-0.032). L_concept_deg quá yếu so với các loss khác cho những type này. Occlusion max 30% coverage (level 3) không đủ khuất minutiae. | ✅ Fix (v12) | Tăng `--deg-every-n-steps` từ 2 → giữ 4 (compute balance sau T17 tăng batch size). Tăng L_concept_deg gamma từ 0.5 → 1.0 trong `DegradationRankingLoss`. Tăng occlusion level 3 từ 30% → 40% coverage. |
| T20 | **SD302 scores anchored tại ~28 trong v12** (target 10–90). T17 fix (L_deg on all datasets) dùng synthetic degradation làm ordinal anchor cho SD302. SD302 images trông như "level 1-2 degraded FVC" đối với model → anchor tại ~28 (gần degraded floor). Per-sensor: R/S slap non-segmented std=0.0 (fully collapsed), roll sensors mean=27-28 std=6-7, flat sensors mean=35-40 std=10-15. KS=0.51 vẫn fail vì sensor TYPE bias (slap vs roll vs flat). | ✅ Fix (v13) | Revert T17 (FVC-only L_deg) + Add per-dataset L_spread cho SD302 subset trong mỗi batch. SD302 images buộc phải compete với nhau để có spread [10,90] thay vì được anchored bởi synthetic degradation. |
| T21 | **Non-segmented slap collapse** (R_*_slap, S_*_slap std=0.0, score≈21.2 cố định). Full-hand slap images visually homogeneous → không có quality variation tự nhiên → model gán cùng score. v13 eval: sensors này kéo mean_KS từ 0.51 (v12) lên 0.634 (v13). | ✅ Fix (v14) | **T28**: `--exclude-sensor R_1000_slap,R_500_slap,S_500_slap` trong training và eval. Loại ~1,700 records. `run_infer.py`, `run_eval.py`, `train_sifq.py` đều có flag `--exclude-sensor`. Eval v13 chính xác: `run_eval_v13.sh``eval_results/v13_seg/`. |
| T22 | **Noise concept regression trong v12**: noise_level: +0.188 (v11) → **-0.168 (v12)** (sai chiều hoàn toàn). Root cause: `if c_idx == 3` special case trong `DegradationRankingLoss` push noise_level TĂNG theo degradation. ScoreAggregator (unconstrained MLP) có thể thỏa mãn L_rank bằng cách GIẢM noise_level + dùng positive weight → conflict được resolve bằng cách invert concept direction. Gamma=1.0 (T19) làm conflict mạnh hơn → model phải chọn một hướng nhất quán → chọn inverted direction. | ✅ Fix (v13) | T25: Remove `if c_idx == 3` special case trong `DegradationRankingLoss.forward()`. Tất cả concepts GIẢM theo degradation (high = better quality). Noise_level semantic: **low noise level = low noise = good** (trước: high noise_level = more noise). Consistent với tất cả concepts khác → ScoreAggregator dùng positive weights cho tất cả. |
| T23 | **Occlusion concept vẫn dead sau T19** (v12 eval: contin=+0.023, minutiae=+0.075 — sai chiều cả hai). Hai nguyên nhân: (1) `DEGRADATION_CONCEPT_MAP` chỉ supervise minutiae_reliability, bỏ qua continuity; (2) coverage 40% chưa đủ để affect minutiae feature. | ✅ Fix (v13) | T26: Thêm continuity (c_idx=2) vào `DEGRADATION_CONCEPT_MAP["occlusion"] = [5, 2]`. Tăng occlusion level 3 coverage từ 40% → 55% (`0.183 * level`). v13 kết quả: occlusion→minutiae=-0.09 (yếu, chưa đủ). v14 cần monitor. |
| T28 | **Non-segmented slap gây nhiễu training và eval** (v13: R_1000_slap score≈21.2 std=0, kéo mean_KS lên 0.634). Full-hand slap ảnh 4 ngón tay → backbone extract feature gần như giống nhau → score collapse về 21.2 cho mọi người. SensorDiscriminator phải waste GRL capacity để align cụm ~21 với segmented counterparts ~54. | ✅ Fix (v14) | T28: `--exclude-sensor R_1000_slap,R_500_slap,S_500_slap`. Loại ~1,700 records từ SD302-B/D. Thêm flag vào `train_sifq.py`, `run_infer.py`, `run_eval.py`, `visualize_score_milestones.py`. Script: `run_train_v14.sh`, `run_eval_v13.sh`. |
| T29 | **Epoch plateau quá sớm — 80 epochs lãng phí** (v13: loss converge từ epoch 20; ep20–79 flat). | ✅ Fix (v14) | T29: v14 train từ đầu (không resume) với **60 epochs**. LR cosine schedule 1e-4→0 trên 60 epochs. Nếu resume từ checkpoint: 30 epochs đủ. |
| T31 | **Score collapse v15 — tất cả ảnh score gần như giống nhau** (eval range 24–56, q_mean≈52.5 flat suốt 60 epochs). Root cause: `stats=None` trong mọi lần gọi `loss_mat()`. Khi stats=None: `q_mat = raw cosine ≈ 0.85` cho mọi ảnh (MDGT quality-robust → cosine ổn định ở mức cao cho mọi ảnh của cùng identity). L_mat kéo pred_score về `(0.85/2 + 0.5)×100 ≈ 92.5`; L_spread kéo về [10,90]; equilibrium tại q_mean≈52 không thay đổi. Không có gradient nào chỉ ra ảnh nào nên cao/thấp hơn ảnh khác → model chỉ đáp ứng **rank thứ tự** trong batch (L_spread) mà không phân biệt ảnh tốt/xấu thực sự. | ⚠️ Partial fix (v16 — tạo ra bug mới T32) | **T31a (per-identity cos stats):** Sau khi tính prototype, chạy second pass `compute_identity_cos_stats()` để tính `(mean_cos, std_cos)` per identity. Pass `stats=cos_stats` vào `loss_mat()`. **T31b (tanh scaling trong `_compute_q_mat`):** `q_mat = tanh((cos − μᵢ) / σᵢ)` → teacher target zero-centred ∈ (−1,1). **T31c (W_SPREAD 2.0→4.0):** Tăng spread weight. **Nhưng T31 tạo ra conflict mới: xem T32.** |
| T32 | **Score collapse v16 — NGHIÊM TRỌNG HƠN v15** (eval range 46.9–50.5, q_std=0.33! Training q_std=15.6 nhưng là false positive — driven bởi FVC trong training batch). Root cause: T31 per-identity stats conflict trực tiếp với L_pair. L_mat muốn **within-identity variation** (`q_mat=tanh((cos−μᵢ)/σᵢ)` → ảnh cùng identity phải score khác nhau). L_pair muốn **within-identity uniformity** (`|Q(s1)−Q(s2)|≤0.05` → cùng identity, khác sensor, score bằng nhau). Conflict không thể resolve → model output ~50 cho tất cả SD302. FVC không có L_pair → FVC spread OK → training q_std=15.6 driven by FVC only. Inference trên SD302-only → q_std=0.33. Sensor bias: roll sensors cluster ≈49.67 vs slap/flat ≈50.2 → KS=0.90 trong range 3.6pt → mean_KS=0.527 (tệ hơn cả v13). | ✅ Fix (v17) | **T32a (`compute_identity_cos_stats(fvc_only=True)`):** Chỉ tính stats cho FVC identities. FVC không có cross-sensor pairs → L_pair=0 → không conflict với L_mat within-identity signal. **T32b (`_compute_q_mat` fallback to raw cosine):** SD302 identities không có trong FVC-only stats → fallback về `raw_cosine ≈ 0.85` (v14 behaviour — constant target, chỉ anchor mean, không conflict với L_pair). FVC images giữ tanh normalization với FVC-specific stats. Files: `src/losses/matcher_teacher.py::_compute_q_mat`, `scripts/train_sifq.py::compute_identity_cos_stats`. |
| T33 | **Score collapse v17 — WORSE THAN v16** (eval SD302 range 18.1–53.0, tất cả đổ về ~52.7, q_std=0.60! Training q_std=16.59 OK nhưng false positive driven by FVC). Root cause (sâu hơn T32 đã fix): T32 loại bỏ L_mat vs L_pair CONFLICT nhưng không fix SCORE COLLAPSE. (1) SD302 raw cosine ≈ 0.85 là CONSTANT → L_mat không tạo per-image quality gradient cho SD302 → không có signal để phân biệt SD302 images. (2) GRL + L_pair triệt tiêu discriminative features từ SD302 backbone: GRL xóa sensor-correlated info (bao gồm quality-correlated-with-sensor); L_pair ép same-identity same-score → SD302 feature space bị flatten. (3) FVC có stable per-image tanh targets → model học quality function cho FVC. SD302 chỉ có inconsistent batch-relative spread signals → feature collapse về ~52.7. (4) FVC tanh targets (variable) vs SD302 raw cosine targets (constant) tạo asymmetry → model học shortcut "FVC=variable, SD302=52.7". W_SPREAD=4.0 + gamma=2.0 + sd302_concept_weight=1.0 làm tệ thêm so với v14. KS=0.616 ❌, Pearson=-0.002 ❌. | ✅ Fix (v18) | **T33: Revert tất cả additions từ v15–v17 về v14 baseline + thêm SD302-A.** T33a: `--no-mat-stats` — skip per-identity cosine stats hoàn toàn; tất cả images (FVC + SD302) dùng raw cosine làm L_mat target → không còn FVC/SD302 asymmetry. T33b: W_SPREAD=2.0 (revert từ 4.0). T33c: concept_deg_gamma=0.5 (revert từ 2.0). T33d: sd302_concept_weight=0.0 (revert T30b). T33e: deg-every-n-steps=4 (revert từ 2). NEW: SD302-A included (13,630 images, 8 sensors A-H) — v14 chỉ dùng B+D. Targets: q_std>12, score range 10–90, mean_KS≤0.27, Pearson≥0.25. |
| T35 | **Score collapse v15–v19 — two root causes (T33 fixed but v19 still collapsed).** v18 actual: q_std=0.01 ❌❌ (score range ~54.62–54.65). Training showed healthy q_std but inference collapse. Root cause 1 (dominant): Truncated prototypes `--proto-max-batches 150` → only 45% of dataset → SD302 gets 2–3 partial sensor prototypes → cosine is sensor-biased (not quality-correlated) → per-image L_mat gradient near-zero for SD302. Root cause 2: FVC-only L_deg leaves SD302 without ordinal grounding → L_spread_ds is batch-level only → collapses at inference. T35a fix: `--proto-max-batches 0` (full dataset, 15–25 min). T35b fix (proposed): `--deg-include-sd302` → apply L_rank to SD302 too. Note: T35b was WRONG — see T36. | ⚠️ Partial (v20: T35a ✅, T35b ❌) | T35a: Set `--proto-max-batches 0` (no cap). T35b: `--deg-include-sd302` flag (applies L_rank to FVC+SD302 instead of FVC-only). **T35b introduced new collapse — removed in v21.** |
| T36 | **Score collapse v20 — `--deg-include-sd302` creates ~53.38 attractor.** v20 training q_std=18.3 (healthy-looking) but ALL SD302 inference scores collapse to ~53.38 (q_std=0.007, range 53.20–53.41). Root cause: L_rank applied to SD302 clean images. L_rank only requires Q(clean)>Q(degraded)+m, NOT that different clean images score differently. Model satisfies L_rank by assigning same ~53.38 to ALL clean SD302 images + lower to degraded → gradient has no reason to differentiate clean images → stable ~53.38 attractor. L_spread_ds (batch-level) + L_pair reinforce attractor. Training q_std=18.3 is FALSE POSITIVE driven entirely by L_spread gradient — at inference (no gradient) model outputs its "natural" ~53.38 for all clean SD302. **Why v14 worked without --deg-include-sd302:** FVC genuine quality variation → MDGT cosine genuinely varies (0.75→0.93 per FVC impression) → L_mat gradient shapes backbone to be quality-discriminative → features TRANSFER to SD302 at inference. L_rank on SD302 blocks this transfer. eval_results/v20: mean_KS=0.4936 ❌, Pearson=0.0048 ❌. | ✅ Fix (v21 — Training) | **T36: Remove `--deg-include-sd302`.** Keep `--proto-max-batches 0` from T35a. FVC-only L_deg: genuine quality variation → L_mat gradient → quality-discriminative backbone → intrinsic SD302 quality scoring. One-line change in run_train_v21.sh vs run_train_v20.sh. |
| T37 | **CrossSensorBatchSampler k_cross=16 + full prototypes destroys quality discrimination.** v21: Pearson=0.092, dry_skin orientation=+0.87 ❌. Root cause: `CrossSensorBatchSampler` (added for T8) default k_cross=16 guarantees 16 cross-sensor pairs/batch. With full prototypes (nearly-constant L_mat targets ≈0.85–0.90 for all SD302), backbone over-optimises sensor invariance at expense of quality ordering → Pearson collapses. v14 used random batching (~1–2 pairs by chance). Additionally: DataParallel (--gpus 0,1) causes 4.3× slowdown for TinyViT-5M (v21: ~970s/epoch vs v20 single GPU: ~226s). Redundant double teacher pass (emb_cache + compute_teacher_prototypes both ran teacher on all 42K images). | ✅ Fix (v22 — Training) | **T37: (1) `--k-cross 0`** (disable CrossSensorBatchSampler, random batching like v14); **(2) `--gpus 0`** (single GPU, no DataParallel); **(3) `build_prototypes_from_cache()`** computes prototypes from emb_cache in O(N) CPU — no second teacher pass. All other params same as v21 (proto-max-batches=0, no-mat-stats, spread-weight=3.0, concept-deg-gamma=0.5, FVC-only L_deg). Expected: Pearson≥0.28, KS≤0.28, ~220s/epoch, correct concept grounding. |
| T38 (v22) | **minutiae_reliability wrong direction for wet_press; occlusion deviates from paper.** (1) `wet_press → minutiae_reliability[5]` consistently POSITIVE across all versions (v14=+0.565) — simulation 3×3 kernel too subtle at level 1 (~1 px expansion) so model reads it as "good ink". (2) `occlusion` concept map had `[5, 2]` (continuity added in T26) and coverage 55%, both deviating from paper design `[minutiae_reliability only, 10–40%]`. T26 added continuity as a crutch for weak signal, but root cause was unseeded eval (incoherent Spearman ρ across random block positions per level). | ✅ Fix (v22 — Code) | **T38a** (wet_press simulation): Scale `MorphologicalDilator` kernel `ks = 3 + 2*iterations` (5×5/7×7/9×9) + `sigma = 1.0 + 0.8*iterations`. Level 1 visibly impairs bifurcations. Concept map KEPT at `[1, 5]` (paper). **T38b** (occlusion): Revert concept map `[5, 2]` → `[5]`; coverage `0.183×level` → `0.133×level` (max 40%). **T38c** (occlusion eval): `np.random.seed(level)` in `compute_crosstalk_matrix` — deterministic block position per severity level. Files: `src/data/degradation.py`, `src/losses/degradation_ranking.py`, `src/evaluation/concept_grounding.py`. |
| T_old38 | **MDGTCheckpointTeacher phụ thuộc vào external `fingerprint_pad` package** (dynamic sys.path injection) — brittle, không versioned, khó deploy. | ✅ Fix (v21, cùng phiên) | Import từ `src/models/mdgt/pipeline.py::MDGTv2` (local). `_ensure_on_path()` và `fingerprint_pad` import đã bị xóa. `MDGTv2` constructor nhận thêm `gnn_heads` và `pool_heads` (absent trong external API cũ). File: `src/training/mdgt_teacher.py`. |


---

## 10. Performance Engineering (v21)

### Bối cảnh

Training ban đầu chạy ~3.5–5.7s/it. Sau 3 vòng tối ưu, root causes được xác định và fix:

| Iteration | Thời gian | Root cause |
|-----------|-----------|------------|
| Before | 3.49 s/it (step 39) | Teacher DINOv2 forward mỗi step |
| After teacher cache | 4.05 s/it (step 6) | DataLoader disk IO bottleneck bị lộ |
| After image preload | 4.96 s/it (step 4) | CUDA warmup + _compute_q_mat GPU syncs |
| After vectorize q_mat | ~0.4 s/it (dự kiến, sau warmup) | — |

> **Lưu ý:** Các measurement ở step 2–6 bao gồm CUDA kernel JIT compilation (one-time, 2–5s). Cần đến step 20–30 để đo steady-state.

### Fix 1: Teacher Embedding Cache

**Vấn đề:** DINOv2 ViT-S/14 + TRAM + GNN forward chạy trên ~100 images mỗi step, mỗi epoch. Teacher là frozen → cùng image luôn cho cùng embedding. ~20,000 DINOv2 forward qua 60 epochs = pure waste.

**Thêm nữa:** Attention monkey-patch trong `DINOv2Backbone` vô hiệu hóa xformers, buộc dùng standard O(N²) attention + lưu 12 × `(B, 6, 257, 257)` attention maps per forward.

**Fix:**
```python
# scripts/train_sifq.py
emb_cache = precompute_teacher_embeddings(teacher, train_ds, device, batch_size=128)
# → Tensor[N, 256] CPU float32, ~43 MB cho 42K records

# Per step (thay vì teacher forward):
_cached_emb = emb_cache[record_idxs[_non_polyu_idx]]  # O(1) lookup
loss_mat(..., cached_emb=_cached_emb)
```

**Tiết kiệm:** ~2–3s/it (DINOv2 forward) → 0 (lookup). Startup cost: ~2–3 min (1 lần duy nhất).

---

### Fix 2: Image Preloading

**Vấn đề:** Mỗi `RecordDataset.__getitem__` gọi `Image.open(path).convert("L")` + resize. Với 8 DataLoader workers và B=128, mỗi batch cần ~128 disk reads + PIL decode + resize. Đây là CPU/IO bottleneck ẩn sau teacher GPU time.

**Fix:**
```python
# RecordDataset.__init__
def _preload_images(self, image_size):
    cache = np.empty((N, image_size, image_size), dtype=np.uint8)  # ~2 GB
    for i, rec in enumerate(self.records):
        sample = next(loader.iter_samples([rec]))
        cache[i] = (sample["image"].squeeze(0) * 255).byte().numpy()
    self._image_cache = cache

# __getitem__: O(1) numpy slice + dtype cast, no disk IO
image = torch.from_numpy(self._image_cache[idx].copy()).float().unsqueeze(0) / 255.0
```

**Lợi ích kép:** `batch["images_np"]` (uint8 numpy trong batch) cho phép degradation CPU path đọc trực tiếp từ RAM, không cần `images[idx].cpu().numpy()`.

---

### Fix 3: Vectorized `_compute_q_mat`

**Vấn đề:** `_compute_q_mat` gọi `torch.dot(emb[i], proto).item()` trong Python loop cho mỗi trong ~100 non-polyu samples. Mỗi `.item()` là một GPU synchronization barrier — CPU block cho đến khi GPU hoàn thành. ~100 syncs/step × ~10–100ms/sync = 1–10s overhead.

**Fix:**
```python
# Trước: 100 syncs
for i, identity in enumerate(identity_ids):
    cos = float(torch.dot(emb[i], proto).item())  # ← GPU sync!

# Sau: 1 sync
proto_mat = torch.stack([prototypes[id] for id in identity_ids]).to(emb.device)  # 1 transfer
proto_mat = F.normalize(proto_mat, dim=-1)
cos_values = (emb * proto_mat).sum(dim=-1)   # 1 GPU kernel
cos_cpu = cos_values.float().cpu().tolist()   # 1 sync
```

**Tiết kiệm:** ~100 GPU syncs → 1. Đây là fix quan trọng nhất sau khi teacher cache đã giải phóng bottleneck ẩn.

---

### Fix 4: GPU-Native Degradation

**Vấn đề:** Degradation loop cũ làm `images[_fvc_idx].detach().float().cpu().numpy()` → cv2 ops → `tensor.to(device)`. GPU→CPU là synchronous (stall pipeline).

**Fix:**
```python
# _degrade_gpu: blur/noise/occlusion — full GPU, không transfer
if deg_type == "blur":
    return TF.gaussian_blur(images, kernel_size)
if deg_type == "noise":
    return (images + torch.randn_like(images) * sigma).clamp(0, 1)
if deg_type == "occlusion":
    out = images.clone(); out[:, :, y:y+block, x:x+block] = 1.0; return out

# _degrade_from_np: jpeg/dry_skin/wet_press — CPU nhưng đọc từ preloaded numpy
imgs = images_np[deg_idx]           # [K, H, W] uint8, đã trong RAM
low  = [deg_pipeline.apply(img, t, lo) for img in imgs]
...  → single CPU→GPU transfer (không GPU→CPU)
```

**Breakdown degradation types:**

| Type | Đường đi | Transfer cost |
|------|---------|---------------|
| blur | GPU (`TF.gaussian_blur`) | 0 |
| noise | GPU (`torch.randn_like`) | 0 |
| occlusion | GPU (tensor indexing) | 0 |
| jpeg | CPU+cv2 → GPU (1 transfer) | 1 CPU→GPU |
| dry_skin | CPU+cv2 → GPU (1 transfer) | 1 CPU→GPU |
| wet_press | CPU+cv2 → GPU (1 transfer) | 1 CPU→GPU |

---

### Tổng quan pipeline v21

```
Startup (1 lần, ~3–5 phút):
  _preload_images()            ~2 min  → 2 GB RAM
  precompute_teacher_emb()     ~2 min  → 43 MB CPU tensor
  compute_teacher_prototypes() ~1 min  → dict[id → 256-D]

Per step (~0.4 s steady-state):
  DataLoader:   ~1 ms  (numpy slice + cast, no disk)
  H2D transfer: ~5 ms  (pin_memory)
  SIFQ forward: ~100 ms (TinyViT-5M B=128, DataParallel 2×GPU)
  L_mat:        ~2 ms  (emb_cache lookup + 1 GPU kernel + 1 sync)
  L_sens:       ~5 ms  (pair indices + hinge loss)
  L_deg (50%):  ~150 ms (2 extra SIFQ forwards B=32 + GPU/CPU deg ops)
  Backward:     ~100 ms
  Optimizer:    ~20 ms
  ─────────────────────
  Total:        ~0.35–0.55 s/it (estimate, post CUDA warmup)
```