Add files using upload-large-folder tool

.claude/settings.local.json

@@ -0,0 +1,10 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(find /home/aiserver/works/fingerprint/sifq -name \"train_v16.log\" 2>/dev/null; find /home/aiserver/works/fingerprint -name \"train_v16.log\" 2>/dev/null)",
+      "Read(//home/aiserver/works/fingerprint/**)",
+      "Bash(grep -E \"^\\\\[Ep|^\\\\[Val|^Ep [0-9]+.*\\\\]$|val_loss|val_eer|EER|AUC|Rank|✓|✗|saved|checkpoint|best\")",
+      "Bash(grep -v \"^\\\\[S[12]\\\\]$\")"
+    ]
+  }
+}

.venv/.gitignore

@@ -0,0 +1 @@
+*

.venv/CACHEDIR.TAG

@@ -0,0 +1 @@
+Signature: 8a477f597d28d172789f06886806bc55

.venv/pyvenv.cfg

@@ -0,0 +1,5 @@
+home = /home/aiserver/.local/share/uv/python/cpython-3.11.12-linux-x86_64-gnu/bin
+implementation = CPython
+uv = 0.7.2
+version_info = 3.11.12
+include-system-site-packages = false

archive/research_fingerjet.md

@@ -0,0 +1,113 @@
+Read [](file:///home/aiserver/works/fingerprint/sifq/rules/SIFQ_explained.md#1-1), lines 1 to 100
+
+Read [](file:///home/aiserver/works/fingerprint/sifq/rules/plan.md#1-1), lines 1 to 80
+
+Read [](file:///home/aiserver/works/fingerprint/sifq/rules/SIFQ_explained.md#100-100), lines 100 to 200
+
+Read [](file:///home/aiserver/works/fingerprint/sifq/src/evaluation/concept_grounding.py)
+
+Read [](file:///home/aiserver/works/fingerprint/sifq/src/losses/degradation_ranking.py#1-1), lines 1 to 80
+
+Read [](file:///home/aiserver/works/fingerprint/sifq/rules/sifq_pdf.txt#1-1), lines 1 to 100
+
+Hãy để tôi đọc thêm phần đánh giá và design của concept 5 để phân tích kỹ hơn.  
+
+Read [](file:///home/aiserver/works/fingerprint/sifq/rules/sifq_pdf.txt#100-100), lines 100 to 200
+
+Read [](file:///home/aiserver/works/fingerprint/sifq/src/losses/degradation_ranking.py#80-80), lines 80 to 120
+
+## Phân tích lý thuyết: SIFQ concept 5 vs FingerJetFX OSE
+
+### Tóm tắt: Phù hợp có điều kiện — nhưng có 3 cảnh báo quan trọng
+
+---
+
+### 1. Alignment vật lý: ✅ Cùng phenomenon
+
+| | SIFQ concept 5 | FingerJetFX OSE (NFIQ2) |
+|---|---|---|
+| **Đo gì** | Latent proxy: "điều kiện để minutiae extraction thành công" | Trực tiếp: trích xuất minutiae + gán confidence |
+| **Output** | Scalar ∈ [0,1] — toàn cục (global image) | Minutiae count + per-minutiae quality + OFL/LCS features |
+| **Supervision** | Occlusion, wet_press degradation | Hand-engineered, VeriFinger-based |
+
+Hai approach đều hướng về cùng một câu hỏi: *"Liệu có thể trích xuất minutiae đáng tin cậy từ ảnh này không?"* → Alignment vật lý hợp lý.
+
+---
+
+### 2. Vấn đề coverage gap: ⚠️ Concept 5 quá hẹp
+
+FingerJetFX minutiae reliability bị ảnh hưởng bởi **nhiều degradation**:
+
+| Degradation | FingerJetFX | SIFQ concept 5 hiện tại |
+|---|---|---|
+| Blur | Ridge endings mờ → không xác định minutiae position | ❌ Không supervise concept 5 |
+| Noise | False minutiae tăng | ❌ Không supervise concept 5 |
+| JPEG | Blocking artifacts tạo false minutiae | ❌ Không supervise concept 5 |
+| Occlusion | Mất minutiae vùng bị che | ✅ Supervise concept 5 |
+| Wet press | Ridge merging → false terminations | ✅ Supervise concept 5 |
+| Dry skin | Ridge gaps → false bifurcations | ❌ Không supervise concept 5 |
+
+**Hệ quả:** SIFQ concept 5 sẽ cho kết quả đúng với occlusion/wet_press nhưng có thể không phản ứng đúng với blur/noise/dry_skin theo nghĩa "minutiae reliability" đầy đủ — trong khi FingerJetFX sẽ bị ảnh hưởng bởi tất cả.
+
+---
+
+### 3. Sensor bias của FingerJetFX: ❌ Mâu thuẫn với novelty
+
+Đây là vấn đề lý thuyết **nghiêm trọng nhất**:
+
+```
+FingerJetFX được thiết kế cho optical contact fingerprint.
+Trên contactless hoặc cross-sensor → minutiae count systematic shift.
+```
+
+SIFQ concept 5 **được thiết kế để sensor-invariant**. Nếu validate concept 5 cross-sensor bằng FingerJetFX:
+- FingerJetFX sẽ cho minutiae count **khác nhau** cho cùng ngón tay ở 2 sensor
+- SIFQ concept 5 (nếu hoạt động đúng) sẽ cho giá trị **tương tự** → hai metric **không đồng ý**
+- Kết luận sai: "SIFQ concept 5 không tương quan với FingerJetFX" — nhưng thực ra đây là **bằng chứng novelty**, không phải failure
+
+**Dùng FingerJetFX làm oracle cross-sensor sẽ penalize đúng cái mà SIFQ cố gắng làm tốt hơn.**
+
+---
+
+### 4. Cách dùng FingerJetFX đúng cho novelty của bạn
+
+**Hợp lệ — within-sensor correlation:**
+```
+Cùng sensor, nhiều ảnh khác nhau:
+  Spearman(SIFQ concept 5, FingerJetFX minutiae count) → nên > 0.30
+```
+Chứng minh concept 5 capture được signal vật lý thực sự.
+
+**Novelty proof — cross-sensor stability comparison:**
+```
+Same finger, sensor A vs sensor B:
+  KS(FingerJetFX minutiae count, across sensors) >> KS(SIFQ concept 5, across sensors)
+```
+Đây là bằng chứng mạnh nhất: FingerJetFX bị sensor bias nghiêm trọng, SIFQ concept 5 ổn định hơn → đây là contribution chính của bài báo.
+
+**Rất thú vị — divergence case:**
+Tìm ảnh contactless mà FingerJetFX cho minutiae count thấp (vì đặc trưng texture khác) nhưng MDGT vẫn match tốt và SIFQ Q cao → cho thấy SIFQ captures "modern matcher utility" không phải "VeriFinger-era minutiae utility".
+
+---
+
+### 5. Vấn đề thiết kế cần xem xét nếu muốn validate nghiêm túc
+
+Concept 5 trong paper được định nghĩa là *"synthesis of previous 3 concepts + own signal from occlusion"*. Nhưng implementation hiện tại **chỉ có occlusion + wet_press**. Nếu muốn concept 5 thực sự đại diện cho "minutiae reliability" đầy đủ như FingerJetFX:
+
+**Option A (nhẹ):** Thêm `dry_skin: [5]` vào DEGRADATION_CONCEPT_MAP — dry skin tạo ridge gaps → false bifurcations → minutiae unreliable. Về mặt vật lý có cơ sở.
+
+**Option B (mạnh hơn):** Thêm blur vào concept 5 supervision — blur không xác định được minutiae position chính xác là failure mode của mọi minutiae extractor. Nhưng cần cẩn thận: concept 1 (clarity) và 2 (continuity) đã được blur supervise → thêm concept 5 vào blur có thể tạo entanglement không cần thiết.
+
+---
+
+### Kết luận lý thuyết
+
+| Câu hỏi | Kết luận |
+|---|---|
+| FingerJetFX có đo cùng phenomenon? | ✅ Có, nhưng minutiae trực tiếp thay vì latent proxy |
+| Validation within-sensor có hợp lệ? | ✅ Hợp lệ, expected Spearman > 0.30 |
+| Validation cross-sensor có hợp lệ? | ❌ Không — FingerJetFX có sensor bias, sẽ mislead |
+| Dùng FingerJetFX làm oracle cho novelty? | ❌ Mâu thuẫn — novelty SIFQ là vượt qua giới hạn này |
+| FingerJetFX có giúp chứng minh novelty? | ✅ Nếu so sánh KS stability của FingerJetFX vs SIFQ cross-sensor |
+
+Thiết kế evaluation phù hợp: FingerJetFX dùng làm **baseline comparison** ("SIFQ ổn định hơn FingerJetFX trong điều kiện cross-sensor"), không phải **ground truth oracle** cho concept 5.

eval_results/sifq_report_v24.txt

@@ -0,0 +1,151 @@
+======================================================================
+SIFQ -- Concept Quality Model: Ly thuyet & Ket qua (v24)
+Generated: 2026-06-01 17:17
+======================================================================
+
+1. DINH NGHIA 6 CONCEPTS
+----------------------------------------------------------------------
+Tat ca concepts: cao = chat luong tot hon, dau ra trong [0, 1].
+
+  #  Concept                    Y nghia vat ly
+  -- -------------------------- -----------------------------------------------
+  0  orientation_coherence      Ridge flow nhat quan, local orientation field smooth
+  1  ridge_valley_clarity       Bien ridge-valley sac net, contrast cuc bo cao
+  2  continuity                 Ridge lines lien tuc, khong bi dut gay
+  3  noise_level                It nhieu ngau nhien (cao = it noise = tot)
+  4  contrast_uniformity        Contrast deu tren toan foreground
+  5  minutiae_reliability       Minutiae co the trich xuat chinh xac
+
+  Degradation Concept Map (v25 -- T40):
+  blur      -> [clarity[1], continuity[2], orient_coh[0]]
+  noise     -> [noise_level[3], contrast_u[4]]
+  jpeg      -> [continuity[2], clarity[1], contrast_u[4]]
+  occlusion -> [minutiae_reliability[5]]
+  dry_skin  -> [contrast_u[4], continuity[2], orient_coh[0]]
+  wet_press -> [clarity[1], minutiae_rel[5], orient_coh[0]]
+
+  Ly do thiet ke map nhu vay:
+  - Moi concept phai duoc giam sat boi >= 2 loai degradation (tranh single-
+    point-of-failure: 1 degradation target 1 concept -> signal yeu, de bi
+    gradient interference invert chieu).
+  - Chon degradation phu hop vat ly: jpeg blocking -> contrast bands (khong
+    chi ridge artifacts), blur -> orientation blur (khong chi clarity loss).
+
+
+2. CO CHE TRAINING CONCEPT (L_concept)
+----------------------------------------------------------------------
+Voi moi cap anh (mild degradation vs severe degradation cung loai):
+
+  L_concept = SUM Huber( c_mild[c], c_severe[c] + 0.1 )
+              c in targets(degradation_type)
+
+  -> Anh degradation nhe phai co concept cao hon anh degradation nang >= 0.1
+  -> Model hoc tung concept phan ung dung chieu voi loai hu hong tuong ung
+
+Ket hop ranking loss:
+  L_rank = relu(Q_severe - Q_mild + m) + relu(Q_mild - Q_clean + m)
+  -> Q giam theo thu tu: clean > mild_deg > severe_deg
+
+  L_deg = L_rank + gamma * L_concept
+  gamma = 2.0 (v24) -> 1.5 (v25)
+    - gamma qua cao (2.0): gradient conflict qua manh -> noise_level inversion
+    - gamma qua thap (0.5, v22): blur->continuity FAIL (+0.261)
+    - gamma = 1.5: compromise, du manh cho blur/jpeg, khong gay inversion
+
+
+3. VAN DE CONCEPT SATURATION
+----------------------------------------------------------------------
+Nguyen nhan goc: L_concept chi train tren synthetic degradation pairs.
+Voi real fingerprint images, KHONG co gradient dinh huong concept.
+-> Concept troi ve gia tri mac dinh cua backbone features.
+
+Hau qua trong v24 (42,683 real fingerprint images):
+
+  Concept                      mean    std   rho_Q  Tinh trang
+  ---------------------------- ------  -----  ------  ---------------------------
+  orientation_coherence         0.111  0.048  -0.480  !! Saturated LOW  -- T39 overcorrected
+  ridge_valley_clarity          0.050  0.074  +0.085  !! Near-dead -- low variance
+  continuity                    0.043  0.029  +0.138  !! Near-dead -- low variance
+  noise_level                   0.543  0.340  -0.989  !! DOMINATES Q (rho=-0.989)
+  contrast_uniformity           0.049  0.081  -0.131  !! Near-dead -- low variance
+  minutiae_reliability          0.868  0.033  -0.672  !! Saturated HIGH -- dead (range 0.84-0.93)
+
+  Vong lap nguy hiem (self-reinforcing collapse):
+    ScoreAggregator chon noise_level (std=0.340, cao nhat)
+    -> gradient tap trung update noise pathway
+    -> cac concept khac it duoc update -> variance thap hon
+    -> cang bi bo qua -> variance cang thap (vong lap)
+
+  He qua: Q ≈ f(noise_level) -- model thuc chat la "noise detector",
+  khong phai "quality estimator" da khai niem.
+
+
+4. KET QUA v24
+----------------------------------------------------------------------
+  Track 2 -- Sensor Invariance:
+    Mean KS (cross-sensor) = 0.1263   [goal <= 0.30]  PASS
+    Cross-sensor Pearson   = 0.8004   [goal >= 0.20]  PASS
+
+    Y nghia: cung mot ngon tay chup bang cac sensor khac nhau -> SIFQ cho
+    score nhat quan. KS thap = phan bo giong nhau, Pearson cao = ranking on dinh.
+    NFIQ2 thuong co KS > 0.5 cho cross-sensor pairs.
+
+  Track 4 -- Concept Grounding (rho < 0 = PASS):
+    Degradation  Concept                          rho  Status
+    ------------ ---------------------------- -------  ------
+    blur         ridge_valley_clarity          -0.483  PASS
+    blur         continuity                    -0.459  PASS
+    noise        noise_level                   +0.365  FAIL <--
+    jpeg         continuity                    -0.412  PASS
+    jpeg         ridge_valley_clarity          -0.345  PASS
+    occlusion    minutiae_reliability          -0.055  PASS
+    dry_skin     contrast_uniformity           +0.051  FAIL <--
+    dry_skin     continuity                    -0.114  PASS
+    dry_skin     orientation_coherence         +0.008  FAIL <--
+    wet_press    ridge_valley_clarity          -0.042  PASS
+    wet_press    minutiae_reliability          -0.081  PASS
+    wet_press    orientation_coherence         -0.448  PASS
+
+    Result: 9/12 pairs PASS
+
+  3 failures phan tich:
+
+  [1] noise -> noise_level (rho=+0.365):
+      Regression tu v22 (-0.052, OK) -> v24 (+0.365, FAIL).
+      gamma=2.0 + T39 (orient_coh added to dry_skin/wet_press) thay doi
+      gradient landscape cua ScoreAggregator. Noise la degradation duy nhat
+      target noise_level -> single-point pressure -> de bi invert.
+
+  [2] dry_skin -> contrast_u (rho=+0.051):
+      Persistent qua cac phien ban (v22: +0.504). Dry_skin la degradation
+      DUY NHAT target contrast_uniformity -> signal yeu, de bi gradient
+      tu cac loss khac at di.
+
+  [3] dry_skin -> orient_coh (rho=+0.008):
+      T39 moi them orient_coh vao dry_skin, nhung wet_press->orient_coh
+      hoat dong tot (-0.448). Ly do: wet_press + blur deu co orientation
+      disruption manh, nhung dry_skin tao ra orientation noise khong nhat
+      quan -> khong du signal de orient_coh feature phan biet.
+
+
+5. v25 -- T40 FIXES (dang chay)
+----------------------------------------------------------------------
+  Thay doi so voi v24:
+
+  gamma: 2.0 -> 1.5
+
+  blur:  [1, 2]    -> [1, 2, 0]  (them orient_coh[0])
+  noise: [3]       -> [3, 4]     (them contrast_u[4])
+  jpeg:  [2, 1]    -> [2, 1, 4]  (them contrast_u[4])
+  dry_skin / wet_press / occlusion: khong doi
+
+  Ket qua ky vong:
+  - noise->noise_lv:    +0.365 -> negative  (gamma thap + noise 2-concept)
+  - dry_skin->contrast_u: +0.051 -> negative  (3 degs co-supervise contrast_u)
+  - dry_skin->orient_coh: +0.008 -> negative  (blur gives orient_coh strong signal)
+  - Cac pairs da PASS: giu nguyen (blur/jpeg/occlusion/wet_press)
+  - Track 2: giu PASS (KS, Pearson khong thay doi co ban)
+
+======================================================================
+END
+======================================================================

rules/HOWTO_RUN.md

@@ -0,0 +1,775 @@
+# SIFQ — Hướng dẫn chạy
+
+Tất cả lệnh chạy từ thư mục gốc: `/home/aiserver/works/fingerprint`
+
+| Version | Status | Epochs | Highlights |
+|---------|--------|--------|------------|
+| v11 | ✅ Done | 80 | PolyU back in L_sens |
+| v12 | ✅ Done | 80 | T17–T19 concept fixes |
+| v13 | ✅ Done | 80 | Revert T17, per-dataset L_spread, occlusion fix |
+| v14 | ✅ Done | 60 | Exclude non-segmented slap; KS=0.263 ✅, Pearson=0.294 ✅; dry_skin ↑↑ |
+| v15 | ✅ Done | 60 | T30: SD302 concept-only L_deg + gamma=2.0; KS=0.2758 (tệ), score collapse (24–56) ❌ |
+| v16 | ✅ Done | 60 | T31: per-identity cos stats + tanh; score collapse 46.9–50.5 ❌ (worse!); KS=0.527 ❌ |
+| v17 | ✅ Done | 60 | T32: FVC-only cos stats; score collapse ALL SD302 at 52.7, q_std=0.60 ❌; KS=0.616 ❌ |
+| v18 | ✅ Done ❌ | 60 | T33: revert v14 loss + SD302-A; **score collapse** q_std≈0.01, KS=0.249⚠️(false positive), Pearson=0.007❌; concept head stuck near bounds |
+| v19 | ✅ Done ❌ | 60 | T34: true revert to v14 — spread-weight=3.0, deg-every-n-steps=2; score collapse ALL ~53.1x, q_std≈0.01 ❌ |
+| v20 | ✅ Done ❌ | 60 | T35: full prototypes (proto-max=0) + L_deg on SD302; score collapse q_std≈0.007 ❌ — --deg-include-sd302 was wrong fix |
+| v21 | ✅ Done ❌ | 60 | T36: full prototypes + FVC-only L_deg; KS=0.312, Pearson=0.092 ❌ — CrossSensorBatchSampler k_cross=16 destroyed quality discrimination |
+| v22 | ✅ Done | 60 | T37: fix k_cross=0 + single GPU; KS=0.278, Pearson=0.546 |
+| v24 | ✅ Done | 60 | Best: KS=0.126 ✅, Pearson=0.800 ✅, q_std~22 |
+| v25 | ❌ Done | 60 | gamma=1.5 too weak; KS=0.213, Pearson=0.620 |
+| v26 | ✅ Done | 60 | gamma=2.0 restored |
+| v27 | ✅ Done | 60 | SpatialConceptHead |
+| v28 | ✅ Done | 60 | No L_mat (--no-mat); KS=0.1346, Pearson=0.7645 |
+| v29 | ❌ Done | 60 | T41 concept map + ortho-weight 3.0 + concept-spread-weight 1.0; KS=0.2985, Pearson=0.1448 |
+| v31 | ✅ Done | 60 | DINOv2-ViTS/14 teacher (public); KS=0.1152, Pearson=0.8858 — best so far |
+| v32 | 🚀 Next | 60 | T42: dry_skin[4,0] + noise[1,3] — fix noise_level crosstalk + weak noise signal |
+
+---
+
+## 1. Smoke test (kiểm tra pipeline ~2 phút)
+
+```bash
+cd /home/aiserver/works/fingerprint
+bash sifq/scripts/run_smoke.sh
+```
+
+Kết quả ở `sifq/checkpoints_smoke/last.pt` và `metrics.jsonl`.
+
+---
+
+## 2. Full training v21 — Active 🚀 (60 epochs, từ đầu)
+
+> **v21 — T36: Full prototypes + FVC-only L_deg (remove --deg-include-sd302).**
+>
+> **Root cause of v20 failure (new diagnosis):**
+>
+> v20 training showed q_std≈18.3 (healthy) but ALL SD302 inference scores collapsed to ~53.38
+> (q_std≈0.007). The `--proto-max-batches 0` fix (T35a) was correct. But `--deg-include-sd302`
+> (T35b) was WRONG and caused the collapse:
+>
+> - L_rank on SD302 says: Q(clean) > Q(low_deg) + m > Q(high_deg) + m
+> - The model satisfies this with a single ~53.38 for ALL clean SD302 images + lower for degraded
+> - L_rank does NOT require different clean images to score differently → attractor at ~53.38
+> - L_spread_ds (batch-level) + L_pair reinforce the collapse
+>
+> **Why v14 worked without --deg-include-sd302:**
+> FVC genuine quality variation → MDGT raw cosine genuinely varies (0.75→0.93) per impression
+> → L_mat gradient shapes backbone to be quality-discriminative → transfers to SD302 at inference.
+> L_rank on SD302 BLOCKS this transfer by creating the stable ~53.38 attractor.
+>
+> **T36 fix:** Remove `--deg-include-sd302`. Keep `--proto-max-batches 0`.
+> FVC-only L_deg: genuine quality signal trains backbone → SD302 scored by intrinsic quality.
+
+> **v20 — T35: Full prototypes + L_deg on SD302. Fix root causes of ALL v15–v19 failures.**
+>
+> **Root cause analysis (v19 failure = same collapse as v15–v18):**
+>
+> v19 was supposed to reproduce v14 exactly but STILL collapsed (all scores ~53.1x, q_std≈0.01).
+> The "spread-weight/deg-every-n-steps mismatch" explanation was wrong. Two deeper root causes:
+>
+> **Root cause 1 — Truncated prototypes (T35a):**
+> - v14's code had NO `--proto-max-batches` cap → prototypes computed on full 42,683 images
+> - When this param was added (default=150 batches = ~19,200 images = 45% of data), SD302 identities
+>   went from full 19-sensor multi-sensor prototypes to 2–3 sensor partial prototypes
+> - Full prototype: raw cosine varies with image quality across sensors (quality-discriminative)
+> - Partial prototype: raw cosine correlates with WHICH sensors are in the prototype window
+>   → sensor-biased, NOT quality-correlated → no per-image quality gradient for SD302 → collapse
+>
+> **Root cause 2 — FVC-only L_deg leaves SD302 without ordinal grounding (T35b):**
+> - T27 reverted T17 (L_deg on all images) because clean SD302 anchored at ~28
+> - But T27 ITSELF introduced per-dataset L_spread_ds — which prevents anchoring
+> - With L_spread_ds forcing SD302 to span [10,90], re-enabling full L_rank for SD302 is now safe
+> - L_deg (L_rank + L_concept) on SD302 provides the only stable per-image quality signal
+>   for SD302 at inference (synthetic degradation response ≈ ridge quality proxy)
+>
+> **T35 fixes:**
+> 1. **T35a — `--proto-max-batches 0`**: Full dataset prototype computation. Cost: ~10–15 min
+>    overhead at epoch 0 (334 batches vs 150). Restores genuine per-image L_mat quality signal.
+> 2. **T35b — `--deg-include-sd302`**: Apply full L_deg (L_rank + L_concept) to SD302 images.
+>    L_spread_ds prevents the T17 score anchoring issue. Per-image ordinal grounding for SD302.
+>
+> **All other settings from v19 kept unchanged:**
+> - `--no-mat-stats`, `--spread-weight 3.0`, `--deg-every-n-steps 2`
+> - `--concept-deg-gamma 0.5`, `--sd302-concept-weight 0.0`
+> - SD302-A included, `--batch-size 128`, `--gpus 0`
+
+## 2. Full training v22 — 🚀 Next run (60 epochs, từ đầu)
+
+> **v22 — T37: Fix ALL bugs from v15–v21. First correct run with full dataset.**
+>
+> **Three bugs caused v15–v21 to fail:**
+>
+> **BUG 1 — CrossSensorBatchSampler k_cross=16:**
+> - k_cross=16 forces 16 guaranteed pairs/batch = 8–16× more L_sens pressure
+> - With full prototypes (nearly-constant L_mat targets), backbone over-optimises sensor
+>   invariance and loses quality discrimination → Pearson collapses to ~0.09 (v21)
+> - **Fix: `--k-cross 0` (random batching)**
+>
+> **BUG 2 — proto-max-batches default=150:**
+> - Only 45% of dataset used for prototypes → partial sensor prototypes → sensor-biased cosine
+> - **Fix: `--proto-max-batches 0` (full dataset)**
+>
+> **BUG 3 — DataParallel (--gpus 0,1) causes 4.3× slowdown for TinyViT-5M:**
+> - **Fix: `--gpus 0` (single GPU)**
+>
+> **BONUS — Redundant teacher double-pass eliminated:**
+> - Now `build_prototypes_from_cache()` computes prototypes from emb_cache (O(N) CPU)
+
+> **Note on high L_sens / pair loss in v22:**
+> v22 includes SD302-A (8 diverse roll sensors per finger). L_mat targets vary by sensor
+> for the same finger → L_pair starts high (~12 at S2 onset) and decreases gradually.
+> This is EXPECTED with a richer dataset — not a failure. q_std=21–22 (best quality
+> discrimination so far). L_sens equilibrium is found by ep60.
+
+```bash
+cd /home/aiserver/works/fingerprint
+version=v32
+nohup bash sifq/scripts/run_train_$version.sh > sifq/logs/train_$version.log 2>&1 &
+echo "PID: $!"
+tail -f sifq/logs/train_$version.log
+
+pkill -f train_sifq.py
+```
+
+**Expected results (research design goals, 171 sensor pairs):**
+
+| Metric | v21 ❌ | v22 (ep51) | v22 goal |
+|--------|--------|--------|----------|
+| q_std (train) | ~17.7 | **~22** ✅ | ≥18 |
+| epoch time | ~970s ❌ | **~220s** ✅ | ≤250s |
+| Track 2 mean_KS | 0.312 | — (training) | **≤0.30** |
+| Track 2 Pearson | +0.092 ❌ | — (training) | **≥0.20** |
+| Concept grounding | dry_skin+0.83 ❌ | — (training) | **all targets negative ρ** |
+
+**Dấu hiệu v22 đang train đúng:**
+
+| Epoch | Dấu hiệu tốt |
+|-------|-------------|
+| ep 0 | Startup nhanh (~teacher 1 pass + cache); epoch ~220s |
+| ep 0–10 | q_std tăng nhanh 1→10→18; l_mat ~0.18–0.20; l_deg > 0 |
+| ep 12+ | l_pair cao (~12) khi S2 bắt đầu — BÌNH THƯỜNG với SD302-A 8 sensors |
+| ep 20–50 | l_pair giảm dần (12→4→2); adv tăng về rand_ce; q_std ~20–22 ✅ |
+| ep 40–60 | Stable: l_spread ≈0.002, q_mean ~52, q_std ~21–22 |
+
+---
+
+## 3. Full training v21 — Done ❌ (60 epochs)
+| ep 20+ | `train_l_pair` → 0.01–0.05, q_std ổn định 12–18 |
+
+> **Key difference from v20:** `train_q_std` in v20 was also 18.3 (false positive driven by L_spread).
+> In v21, watch for **INFERENCE q_std > 5** at end — run a quick spot-check:
+> `python sifq/scripts/run_infer.py --checkpoint sifq/checkpoints/v21/last.pt --max-samples 200 --output /tmp/spot.jsonl && python -c "import json,statistics; s=[json.loads(l)['q_score'] for l in open('/tmp/spot.jsonl')]; print(f'std={statistics.stdev(s):.2f} range=[{min(s):.1f},{max(s):.1f}]')"`
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_train_v21.sh > sifq/logs/train_v21.log 2>&1 &
+echo "PID: $!"
+tail -f sifq/logs/train_v21.log
+```
+
+- Checkpoint: `sifq/checkpoints/v21/last.pt`
+- Train từ đầu, LR: `1e-4` cosine → 5e-6, Batch: `128`, GPU: `0`
+- **Auto-eval** sau training: kết quả ở `sifq/eval_results/v21/`
+
+---
+
+## 2b. Eval v21 — Chạy thủ công nếu cần
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_eval_v32.sh > sifq/eval_v32.log 2>&1 &
+echo "PID: $!"
+```
+
+Kết quả ở `sifq/eval_results/v21/`.
+
+---
+
+## 2c. Full training v20 — Completed ❌ (score collapse)
+
+> **v20 actual results:** q_std≈0.007 (collapsed), mean_KS=0.4936 ❌, Pearson=0.0048 ❌.
+> Root cause: `--deg-include-sd302` applied L_rank to all clean SD302 → stable ~53.38 attractor.
+> See v21 for fix.
+
+---
+
+## 2d. Eval v20 — Chạy thủ công nếu cần
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_eval_v20.sh > sifq/eval_v20.log 2>&1 &
+echo "PID: $!"
+```
+
+Kết quả ở `sifq/eval_results/v20/`.
+
+---
+
+## 2c. Full training v19 — Completed ❌ (score collapse — 60 epochs)
+
+> **v19 — T34: True revert to v14 hyperparameters (FAILED — same collapse as v18).**
+> Root cause was NOT the spread/deg values (those were already correct in v19).
+> Real root cause: truncated prototypes + FVC-only L_deg. See v20 analysis above.
+
+---
+
+## 2. Full training v18 — Completed ❌ (score collapse — 60 epochs)
+
+> **v18 — T33: revert toward v14 + include SD302-A.**
+>
+> Root cause từ v17 eval:
+> - T32 chỉ remove explicit L_mat vs L_pair conflict, nhưng SD302 VẪNN collapse → q_std=0.60 at inference.
+> - 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.
+>
+> **T33 fixes: revert tất cả additions từ v15–v17:**
+> 1. **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.
+> 2. **T33b — W_SPREAD = 2.0** (revert từ 4.0 — v14 level)
+> 3. **T33c — concept_deg_gamma = 0.5** (revert từ 2.0 — v14 default)
+> 4. **T33d — sd302_concept_weight = 0.0** (revert T30b — v14 default)
+> 5. **T33e — deg-every-n-steps = 4** (revert từ 2 — v14 default)
+> 6. **NEW — SD302-A included**: 13,630 images, 8 sensors (A-H) trong training (v14 chỉ dùng B+D)
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_train_v18.sh > sifq/logs/train_v18.log 2>&1 &
+echo "PID: $!"
+tail -f sifq/logs/train_v18.log
+```
+
+- Checkpoint: `sifq/checkpoints/v18/last.pt`
+- Metrics: `sifq/checkpoints/v18/metrics.jsonl`
+- Train từ đầu (không resume)
+- LR: `1e-4` cosine → 5e-6
+- Batch: `128`
+- **Auto-eval** sau training: kết quả ở `sifq/eval_results/v18/`
+
+**Kết quả v18 thực tế (FAILED — score collapse):**
+
+| Metric | v14 ✅ | v17 ❌ | v18 actual ❌ | Target |
+|--------|--------|--------|--------------|--------|
+| q_std (inference) | ~15 | 0.60 | **0.01 ❌❌** | **>12** |
+| Score range | 10–90 | 18–53 | **54.62–54.65 ❌** | **10–90** |
+| Track 2 mean_KS | 0.263 | 0.616 | **0.249 ⚠️** | **≤0.27** |
+| Track 2 Pearson | +0.294 | -0.002 | **0.007 ❌** | **≥0.25** |
+| n_sensor_pairs | 55 | 171 | **171** | **≥100** |
+
+> **⚠️ v18 KS=0.249 là false positive**: Model không phân biệt được images — tất cả sensors cho cùng score ≈54.64. KS thấp vì các phân bố trivially giống nhau (đều là Dirac delta tại 54.64). Root cause mới: raw cosine L_mat hằng số (~0.85) không tạo per-image quality gradient → model collapse về mean. Cần T34: tìm cách tạo per-image quality signal mà không gây FVC/SD302 asymmetry.
+
+**Dấu hiệu v18 đang train đúng:**
+
+| Epoch | Dấu hiệu tốt |
+|-------|-------------|
+| ep 0–10 | `train_q_std` tăng > 15, `train_l_mat` giảm (~0.25–0.35) |
+| ep 11–20 | `train_l_pair` giảm (GRL kick-in), q_std giảm tạm (~12-15) |
+| ep 20+ | `train_l_pair` → 0.01–0.05, q_std ổn định 12–18 |
+
+---
+
+## 2a. Eval v18 — Chạy thủ công nếu cần
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_eval_v18.sh > sifq/eval_v18.log 2>&1 &
+echo "PID: $!"
+```
+
+Kết quả ở `sifq/eval_results/v18/`.
+
+---
+
+## 2. Full training v19 — Active 🚀 (60 epochs, từ đầu)
+
+> **v19 — T34: True revert to v14 hyperparameters.**
+>
+> Root cause từ v18 failure:
+> - v18 comments nói "revert to v14" nhưng dùng sai giá trị:
+>   - `--spread-weight 2.0` (v14 thực tế: **3.0** — sai 50%)
+>   - `--deg-every-n-steps 4` (v14 thực tế: **2** — sai 2×)
+> - Hai sai số này làm L_spread + L_deg yếu hơn v14 → backbone không học được quality-discriminative features → inference collapse (q_std≈0.01).
+> - Xác nhận bằng `diff run_train_v14.sh run_train_v18.sh`.
+>
+> **T34 fixes:**
+> 1. **T34a — `--spread-weight 3.0`**: Restore v14 actual value. Gradient L_spread mạnh hơn buộc backbone phân biệt quality.
+> 2. **T34b — `--deg-every-n-steps 2`**: Restore v14 actual value. 2× tần suất L_deg → ordinal grounding mạnh hơn.
+>
+> **Giữ nguyên từ v18:**
+> - `--no-mat-stats`: cần thiết để reproduce v14 pre-T31 behavior với code hiện tại
+> - `--concept-deg-gamma 0.5`, `--sd302-concept-weight 0.0` (v14 defaults)
+> - SD302-A included (v18 addition, giữ lại)
+> - `--batch-size 128`, `--gpus 0`
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_train_v19.sh > sifq/logs/train_v19.log 2>&1 &
+echo "PID: $!"
+tail -f sifq/logs/train_v19.log
+```
+
+- Checkpoint: `sifq/checkpoints/v19/last.pt`
+- Metrics: `sifq/checkpoints/v19/metrics.jsonl`
+- Train từ đầu (không resume)
+- LR: `1e-4` cosine → 5e-6
+- Batch: `128`, GPU: `0`
+- **Auto-eval** sau training: kết quả ở `sifq/eval_results/v19/`
+
+**Expected results (target = v14 baseline):**
+
+| Metric | v14 ✅ | v18 ❌ | v19 target |
+|--------|--------|--------|-----------|
+| q_std (inference) | ~15.5 | 0.01 | **>12** |
+| Score range | 10–90 | 54.62–54.65 | **10–90** |
+| Track 2 mean_KS | 0.263 | 0.249⚠️ | **≤0.27 (real)** |
+| Track 2 Pearson | +0.294 | 0.007 | **≥0.25** |
+
+**Dấu hiệu v19 đang train đúng:**
+
+| Epoch | Dấu hiệu tốt |
+|-------|-------------|
+| ep 0–10 | `train_q_std` > 15, `train_l_mat` giảm (~0.25–0.35) |
+| ep 11–20 | `train_l_pair` giảm (GRL kick-in), q_std giảm tạm (~12–15) |
+| ep 20+ | `train_l_pair` → 0.01–0.05, q_std ổn định 12–18 |
+
+---
+
+## 2b. Eval v19 — Chạy thủ công nếu cần
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_eval_v19.sh > sifq/eval_v19.log 2>&1 &
+echo "PID: $!"
+```
+
+Kết quả ở `sifq/eval_results/v19/`.
+
+---
+
+## 3. Full training v17 — Completed (60 epochs) ✅ (score collapse ❌)
+
+> **v17 thất bại**: Score collapse nghiêm trọng — tất cả SD302 images output ~52.7 (q_std=0.60 at inference). Training q_std=16.59 vẫn ổn (driven by FVC). KS=0.616 ❌.
+>
+> **Root cause v17 (sâu hơn T32 đã fix)**:
+> - T32 fix (FVC-only stats, SD302 raw cosine fallback) đã loại bỏ L_mat vs L_pair CONFLICT nhưng không fix SCORE COLLAPSE.
+> - SD302 raw cosine ≈ 0.85 là CONSTANT cho tất cả images → L_mat không phân biệt được SD302 identities với nhau → không có per-image quality signal cho SD302.
+> - 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.
+> - L_spread_ds tạo relative constraints per-batch nhưng KHÔNG CONSISTENT qua các batches → average score cho mỗi SD302 image settle về ~52.7.
+> - FVC images: có stable per-image tanh targets → model học được quality function. SD302: chỉ có inconsistent batch-relative spread signals → feature collapse.
+>
+> **v16 root cause (T32 mục tiêu)**:
+
+> **v17 là training hiện tại.** Root cause từ v16 eval:
+> - **Score collapse nghiêm trọng**: inference q_std=0.33, range 46.9–50.5 (tệ hơn v15!). Training q_std=15.6 trông ổn nhưng driven bởi FVC images. Inference chỉ chạy SD302 → collapse.
+> - **T32 — Root cause**: T31's per-identity stats conflict trực tiếp với L_pair.
+>   - L_mat (T31b): `q_mat = tanh((cos−μᵢ)/σᵢ)` → muốn **within-identity variation** (ảnh khác nhau của cùng ngón tay nên score khác nhau)
+>   - L_pair: `|Q(s1)−Q(s2)| ≤ 0.05` → muốn **within-identity uniformity** (cùng ngón tay, khác sensor, phải score bằng nhau)
+>   - Hai loss conflict → model giải quyết bằng output ~50 cho tất cả SD302. FVC không có L_pair → spread được → training q_std=15.6 driven by FVC only.
+>
+> **v17 fixes (T32): FVC-only per-identity stats**
+> 1. **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.
+> 2. **T32b — `_compute_q_mat` fallback to raw cosine**: SD302 identities không có trong stats → dùng raw cosine (~0.85, hằng số) giống v14. Không còn conflict với L_pair.
+> 3. **Giữ nguyên**: W_SPREAD=4.0, concept_deg_gamma=2.0, sd302_concept_weight=1.0, deg-every=2.
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_train_v17.sh > sifq/logs/train_v17.log 2>&1 &
+echo "PID: $!"
+tail -f sifq/logs/train_v17.log
+```
+
+- Checkpoint: `sifq/checkpoints/v17/last.pt`
+- Metrics: `sifq/checkpoints/v17/metrics.jsonl`
+- Train từ đầu (không resume)
+- LR: `1e-4` cosine → 5e-6
+- Batch: `128`
+- **Auto-eval** sau training: kết quả ở `sifq/eval_results/v17/`
+
+**Kết quả v17 (thực tế so với target):**
+
+| Metric | v14 | v16 | Target v17 | Thực tế v17 |
+|--------|-----|-----|------------|-------------|
+| Score range (eval, SD302) | — | 46.9–50.5 ❌ | **0–100** | 18.1–53.0 (all ~52.7) ❌ |
+| q_std (inference) | — | 0.33 ❌ | **>10** | **0.60** ❌ (worse collapse) |
+| q_std (train, final) | ~15.5 | 15.6 | ~15–20 | 16.59 ✅ |
+| Track 2 mean_KS | 0.263 ✅ | 0.527 ❌ | **≤0.27** | **0.616** ❌ |
+| Track 2 Pearson | 0.294 ✅ | 0.046 ❌ | **≥0.25** | **-0.002** ❌ |
+
+**Dấu hiệu v17 đang train đúng (watch trong log):**
+
+| Epoch | Dấu hiệu tốt |
+|-------|-------------|
+| ep 0–5 | `train_l_mat` bắt đầu giảm (~0.15–0.25) — FVC quality signal hoạt động |
+| ep 0–10 | `train_q_std` tăng lên > 15 (L_spread có tác dụng trên mixed batch) |
+| ep 11–15 | `train_l_pair` giảm nhanh (GRL kick-in), q_std giảm tạm thời |
+| ep 20+ | `train_l_pair` → 0.01–0.05, `train_q_std` ≥ 12, ổn định |
+
+---
+
+## 2a. Eval v17 — Chạy thủ công nếu cần
+
+Auto-eval đã được tích hợp vào `run_train_v17.sh`. Nếu cần chạy lại riêng:
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_eval_v17.sh > sifq/eval_v17.log 2>&1 &
+echo "PID: $!"
+```
+
+Kết quả ở `sifq/eval_results/v17/`.
+
+---
+
+## 2b. Full training v16 — Completed (60 epochs) ✅ (score collapse ❌)
+
+> **v16 thất bại**: score collapse 46.9–50.5 (tệ hơn v15!), KS=0.527. Root cause: T31 conflict với L_pair → T32 fix.
+
+```bash
+# Xem kết quả v16 eval
+cat sifq/eval_results/v16/eval_summary.json
+```
+
+- Checkpoint: `sifq/checkpoints/v16/last.pt`
+- Eval kết quả: `sifq/eval_results/v16/`
+
+---
+
+## 2c. Full training v15 — Completed (60 epochs) ✅ (score collapse ❌)
+
+> **v15 thất bại**: score range 24–56 (score collapse), KS=0.2758 (tệ hơn v14). Root cause: T31.
+
+```bash
+# Chỉ xem logs — không cần chạy lại
+cat sifq/eval_v15.log
+cat sifq/checkpoints_full_v15/metrics.jsonl | tail -5
+```
+
+---
+
+## 2c. Full training v14 — Completed (60 epochs, từ đầu) ✅
+
+```bash
+nohup bash sifq/scripts/run_eval_v15.sh > sifq/eval_v15.log 2>&1 &
+echo "PID: $!"
+```
+
+Kết quả ở `sifq/eval_results/v15/`.
+
+---
+
+## 2c. Full training v14 — Completed (60 epochs, từ đầu) ✅
+
+> **v14 đã train xong.** Nguyên nhân gốc từ v13 eval:
+> 1. **T28 — Non-segmented slap std≈0**: `R_1000_slap`, `R_500_slap`, `S_500_slap` là ảnh 4 ngón tay → backbone cho score ~21.2 cho mọi người (std=0). Kéo mean_KS từ 0.51 (v12) lên 0.634 (v13 — tệ hơn). Loại khỏi training và eval.
+> 2. **T29 — Train từ đầu (không resume)**: backbone TinyViT từ IN22k pretrained, concept head + aggregator ngẫu nhiên → cần 60 epochs để converge đầy đủ.
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_train_v14.sh > sifq/train_v14.log 2>&1 &
+echo "PID: $!"
+tail -f sifq/train_v14.log
+```
+
+- Checkpoint: `sifq/checkpoints_full_v14/last.pt`
+- Metrics: `sifq/checkpoints_full_v14/metrics.jsonl`
+- Train từ đầu (không resume)
+- **Sensors bị loại**: `R_1000_slap`, `R_500_slap`, `S_500_slap` (~1,700 records)
+- **Dataset**: 42,683 records, 29 sensors
+
+**Kết quả training v14 (60 epochs):**
+
+| Metric | v13 | v14 thực tế | Trend |
+|--------|-----|-------------|-------|
+| `q_mean` (train) | ~33 | **52.6** | ✅ Centered tốt hơn hẳn |
+| `q_std` (train) | 12.5 | **~15.5** | ✅ Spread ổn định |
+| `l_pair` (ep59) | — | **0.006** | ✅ Sensor invariance xuất sắc |
+| `l_mat` (ep59) | — | **0.270** | ✅ Converged tốt |
+| `l_spread` (ep59) | — | **0.025** | ✅ Ổn định |
+| Eval (KS/Pearson) | 0.634 / 0.224 | **0.263 / 0.294** ✅ | Target ✅ |
+
+**3-phase training dynamics (quan sát được lần đầu ở v14 do train from scratch):**
+
+| Phase | Epochs | q_std | l_pair | Mô tả |
+|-------|--------|-------|--------|-------|
+| S1 — Spread | 0–10 | 1 → 22.6 | 0.47 → 25 | GRL λ warm-up, L_spread mở rộng phân bố |
+| Transition | 11–15 | 22.6 → 14 | 24 → 0.13 | β kick-in, GRL triệt tiêu sensor signature |
+| Equilibrium | 16–59 | ~15.5 | → 0.006 | l_pair/l_mat converged, l_spread stable |
+
+**Kết quả eval v14 (Track 2 + Track 4):**
+
+| Track | Metric | v12 | v13 | v14 |
+|-------|--------|-----|-----|-----|
+| Track 2 | mean_KS | 0.510 | 0.634 ❌ | **0.263** ✅ |
+| Track 2 | Pearson | -0.003 | +0.224 | **+0.294** ✅ |
+| Track 2 | n\_sensor\_pairs | 91 | 91 | 55 (slap loại) |
+| Track 4 | dry_skin → contrast | — | -0.376 | **-0.623** ✅⬆ |
+| Track 4 | dry_skin → continuity | — | +0.009 | **-0.563** ✅⬆ |
+| Track 4 | blur → clarity | — | -0.490 ✅ | -0.187 ⚠️ (yếu hơn) |
+| Track 4 | noise → noise_level | — | -0.533 ✅ | -0.019 ⚠️ (regression) |
+| Track 4 | wet_press → minutiae | — | +0.326 ❌ | +0.565 ❌ (tệ hơn) |
+
+**Expected improvements v14 — đã đạt:**
+
+| Track | Metric | v13 | Target v14 |
+|-------|--------|-----|------------|
+| Score | Sensor clusters | 3 cụm (~21, ~50, ~55) | **2 cụm hoặc phẳng** (slap đã loại) |
+| Track 2 | mean_KS | 0.634 ❌ | **<0.35** ✅ |
+| Track 2 | Pearson | 0.224 | **≥0.25** |
+| Train | Epochs | 80 | **60** (from scratch) |
+
+---
+
+## 2d. Eval v14 — Chạy sau khi training xong ⭐
+
+> **Bước tiếp theo bắt buộc** sau khi v14 training completed.
+
+```bash
+# Bước 1: Inference — sinh Q scores
+python sifq/scripts/run_infer.py \
+  --checkpoint sifq/checkpoints_full_v14/last.pt \
+  --exclude-sensor "R_1000_slap,R_500_slap,S_500_slap" \
+  --output     sifq/eval_results/sifq_scores_v14.jsonl
+
+# Bước 2: Eval — Track 2 + Track 4
+python sifq/scripts/run_eval.py \
+  --sifq-scores sifq/eval_results/sifq_scores_v14.jsonl \
+  --checkpoint  sifq/checkpoints_full_v14/last.pt \
+  --out-dir     sifq/eval_results/v14 \
+  --skip-track1
+
+# Bước 3: Visual score milestones
+mkdir -p sifq/eval_results/v14
+python sifq/visualize_score_milestones.py \
+  --scores  sifq/eval_results/sifq_scores_v14.jsonl \
+  --output  sifq/eval_results/v14/milestone_samples.png \
+  --n_buckets 8
+```
+
+---
+
+## 2b. Eval v13 (chính xác, exclude non-segmented slap)
+
+```bash
+bash sifq/scripts/run_eval_v13.sh
+```
+
+Kết quả ở `sifq/eval_results/v13_seg/` (exclude `R/S_*_slap`).
+
+---
+
+## 2c. Full training v13 — Completed (80 epochs)
+
+> **v13 là training hiện tại.** Phân tích sau v12 eval cho thấy T17 fix (L_deg on all datasets) là sai về nguyên tắc:
+> 1. **T20 — SD302 anchored tại ~28**: Synthetic degradation trên SD302 kéo clean images về gần degraded floor (~28). SD302 toàn ảnh chất lượng cao → synthetic degradation không phản ánh quality variation thực → anchoring sai.
+> 2. **T21 — Non-segmented slap std=0**: Full-hand slap images visually homogeneous → model gán cùng score (behavior đúng, không cần fix).
+> 3. **T22 — Noise concept regression**: `noise_level: +0.188 (v11) → -0.168 (v12)`. Old `c_idx==3` special case gây conflict với L_rank → model invert noise_level direction.
+> 4. **T23 — Occlusion vẫn dead**: 40% coverage chưa đủ + continuity concept không được supervise.
+>
+> **v13 fixes:**
+> - **T27**: Revert T17 (FVC-only L_deg) + **per-dataset L_spread** cho SD302 subset riêng (prevents SD302 collapse mà không cần synthetic degradation anchoring).
+> - **T25**: Remove `c_idx==3` special case. Tất cả concepts giảm với degradation (high=better).
+> - **T26**: Occlusion: add continuity (c_idx=2) to DEGRADATION_CONCEPT_MAP + coverage 40%→55%.
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_train_v13.sh > sifq/train_v13.log 2>&1 &
+echo "PID: $!"
+tail -f sifq/train_v13.log
+```
+
+- Checkpoint: `sifq/checkpoints_full_v13/last.pt`
+- Metrics: `sifq/checkpoints_full_v13/metrics.jsonl`
+- Resume từ: `sifq/checkpoints_full_v12/last.pt`
+
+**Expected improvements v13:**
+
+| Track | Metric | v11 | v12 | Target v13 |
+|-------|--------|-----|-----|----------|
+| Score | q_std (inference) | 6.47 | 12.53 | **>18** (spread đều, bao gồm SD302 roll/flat) |
+| Score | mean | 59.81 | 33.63 | **45–65** (SD302 anchored đúng chỗ) |
+| Track 2 | mean_KS | 0.5566 | 0.5100 | **<0.20** (sensor type bias vẫn còn nhưng giảm) |
+| Track 4 | wet_press → clarity | +0.219 ❌ | -0.647 ✅ | giữ ✅ |
+| Track 4 | noise → noise_level | +0.188 ✅ | -0.168 ❌ | **positive ✅** (T25 fix) |
+| Track 4 | occlusion → continuity | -0.029 ❌ | +0.023 ❌ | **negative ✅** (T26 fix) |
+| Track 4 | occlusion → minutiae | -0.032 ❌ | +0.075 ❌ | **negative ✅** (T26 fix) |
+
+---
+
+## 2b. Full training v11 — Completed (80 epochs)
+
+> v11 đã train xong. Checkpoint tại `sifq/checkpoints_full_v11/last.pt`.
+> 1. **PolyU trở lại L_sens** — contact vs contactless pairs là tín hiệu L_sens mạnh nhất (2,612 anchor groups thay vì 2,276)
+> 2. **PolyU bị mask khỏi L_mat** — không compute prototype / teacher loss cho PolyU
+> 3. **PolyU bị mask khỏi L_deg** — không apply degradation cho PolyU images
+> 4. **Schedule mới: β/α = 1.0 ngay từ S2 (ep20)** — β ramp nhanh hơn α trong ep10–19
+>    - S2: α=0.30 β=0.30 γ=0.60 (β/α=1.0) thay vì α=0.40 β=0.20 của v9
+>    - S3: α=0.25 β=0.35 γ=0.50 (β/α=1.4)
+> + Root cause của v9 oscillation: β/α=0.50 trong S2 quá yếu — GRL không đủ override L_mat với PolyU cross-modality pairs
+
+```bash
+cd /home/aiserver/works/fingerprint
+nohup bash sifq/scripts/run_train_v14.sh > sifq/train_v14.log 2>&1 &
+echo "PID: $!"
+
+# Theo dõi log
+tail -f sifq/train_v11.log
+```
+
+- Dữ liệu: SD302-A/B/D + FVC2002 + FVC2004 + PolyU = **43,559 ảnh, 24 sensors**
+- L_sens anchors: SD302 (~2,276 groups) + PolyU (336 cross-modal groups) = 2,612 total
+- L_mat: SD302 + FVC (PolyU masked)
+- L_deg: FVC only (SD302 + PolyU masked)
+- Checkpoint: `sifq/checkpoints_full_v11/last.pt`
+- Metrics: `sifq/checkpoints_full_v11/metrics.jsonl`
+- Thời gian ước tính: **~9 giờ** (2× RTX A4000, DataParallel, 453 steps/epoch)
+
+Kiểm tra nhanh training đang chạy:
+
+```bash
+# Xem loss mới nhất
+tail -5 sifq/checkpoints_full_v11/metrics.jsonl | python3 -c "
+import sys, json
+for l in sys.stdin: r=json.loads(l); print(f'ep{r[\"epoch\"]:02d} deg={r[\"train_l_deg\"]:.4f} sens={r[\"train_l_sens\"]:.4f} spread={r[\"train_l_spread\"]:.4f} lr={r[\"lr\"]:.2e} total={r[\"train_total\"]:.4f}')
+"
+```
+
+**Expected trends v11:**
+
+| Loss | S1 (ep0-9) | S2 (ep20-34) | S3 (ep40+) |
+|------|-----------|-------------|------------|
+| `L_deg` | 0.10 → **<0.05** | giữ <0.05 | giữ <0.05 (γ=0.50) |
+| `L_sens` | tracked only | **<1.5** (β/α=1.0 → GRL cân bằng) | <1.0 |
+| `spread` | **giảm rõ** (uniform mode) | tiếp tục giảm | → ~0 |
+| `L_mat` | flat ~0.3 | giảm (có FVC signal) | flat ~0.01 |
+
+---
+
+## 3. Inference — sinh Q scores từ model đã train
+
+```bash
+cd /home/aiserver/works/fingerprint
+source .venv/bin/activate
+
+python sifq/scripts/run_infer.py \
+  --checkpoint sifq/checkpoints_full_v11/last.pt \
+  --output     sifq/eval_results/sifq_scores_v11.jsonl
+```
+
+Output: `sifq/eval_results/sifq_scores_v11.jsonl` — mỗi dòng là 1 ảnh với Q score + 6 concepts.
+
+---
+
+## 4. Evaluation — verify SOTA
+
+### 4a. Nhanh: chỉ Track 2 (sensor invariance) + Track 4 (concept grounding)
+
+```bash
+python sifq/scripts/run_eval.py \
+  --sifq-scores sifq/eval_results/sifq_scores_v11.jsonl \
+  --checkpoint  sifq/checkpoints_full_v11/last.pt \
+  --out-dir     sifq/eval_results \
+  --skip-track1
+```
+
+### 4b. Đầy đủ: Track 1 (ERC) + Track 2 + Track 4
+
+```bash
+python sifq/scripts/run_eval.py \
+  --sifq-scores sifq/eval_results/sifq_scores_v11.jsonl \
+  --checkpoint  sifq/checkpoints_full_v11/last.pt \
+  --out-dir     sifq/eval_results \
+  --max-pairs   3000
+```
+
+### 4c. Có NFIQ2 baseline (so sánh trực tiếp với SOTA)
+
+```bash
+# NFIQ2 is not on PyPI — generate proxy scores using image quality heuristics:
+python sifq/scripts/gen_nfiq2_proxy_scores.py \
+  --sifq-scores sifq/eval_results/sifq_scores_v11.jsonl \
+  --output      /tmp/nfiq2_scores.jsonl
+python sifq/scripts/run_eval.py \
+  --sifq-scores  sifq/eval_results/sifq_scores_v11.jsonl \
+  --checkpoint   sifq/checkpoints_full_v11/last.pt \
+  --out-dir      sifq/eval_results \
+  --nfiq2-scores /tmp/nfiq2_scores.jsonl
+```
+
+---
+
+## 5. Xem kết quả
+
+```bash
+# Bảng số
+cat sifq/eval_results/eval_summary.json
+
+# Plots (mở file)
+ls sifq/eval_results/*.png
+```
+
+| File | Nội dung |
+|------|----------|
+| `eval_summary.json` | Tất cả metrics (AUC_ERC, KS statistic, Pearson) |
+| `plot_erc.png` | Track 1 — ERC curve, AUC thấp hơn là tốt hơn |
+| `plot_sensor_hist.png` | Track 2 — Q distribution per sensor |
+| `plot_sensor_scatter.png` | Track 2 — scatter Q_s1 vs Q_s2 |
+| `plot_crosstalk.png` | Track 4 — concept grounding heatmap |
+
+---
+
+## 6. Tiêu chí SOTA rank 1
+
+| Track | Metric | v5 | v7 nofvc | Target v11 |
+|-------|--------|-----|---------|----------|
+| Track 1 (ERC) | AUC_ERC | 0.8884 | — | < NFIQ2 proxy |
+| Track 2 (Sensor) | mean_KS | 0.326 | 0.326 (collapse) | < 0.10 |
+| Track 2 (Sensor) | Pearson cross-sensor | 0.179 | — | > 0.70 |
+| Track 4 (Concepts) | blur → clarity rho | +0.483 ❌ | — | negative |
+| Track 4 (Concepts) | occlusion → minutiae rho | -0.013 ❌ | — | negative |
+
+---
+
+## Checkpoint paths (hiện tại)
+
+| Thành phần | Path |
+|-----------|------|
+| MDGT teacher | `pad/TRAM-downstream/checkpoint/checkpoints_dinov2_tram/best_eer.pt` |
+| NIST SD302-A | `dataset/302a/images/challengers` |
+| NIST SD302-B | `dataset/302b/images/baseline` |
+| NIST SD302-D | `dataset/nist_302d/images/auxiliary` |
+| FVC2002 | `dataset/FVC_Dataset/FVC2002` |
+| FVC2004 | `dataset/FVC_Dataset/FVC2004` |
+| PolyU | `dataset/PolyU` |
+| SIFQ v18 (done, score collapse q_std≈0.01, Pearson=0.007 ❌) | `sifq/checkpoints/v18/last.pt` |
+| SIFQ v17 (done, score collapse 18.1–53.0, q_std=0.60 ❌) | `sifq/checkpoints/v17/last.pt` |
+| SIFQ v16 (done, score collapse 46.9–50.5 ❌) | `sifq/checkpoints/v16/last.pt` |
+| SIFQ v15 (done, score collapse 24–56 ❌) | `sifq/checkpoints_full_v15/last.pt` |
+| SIFQ v14 (eval done, KS=0.263 ✅) | `sifq/checkpoints_full_v14/last.pt` |
+| SIFQ v13 (completed) | `sifq/checkpoints_full_v13/last.pt` |
+| SIFQ v12 (completed) | `sifq/checkpoints_full_v12/last.pt` |
+| SIFQ v11 (completed, score collapse ~58.3) | `sifq/checkpoints_full_v11/last.pt` |
+| SIFQ v10 (stopped, PolyU excluded from L_sens) | `sifq/checkpoints_full_v10/last.pt` |
+| SIFQ v9 (stopped ep63, β/α=0.50 too weak) | `sifq/checkpoints_full_v9/last.pt` |
+| SIFQ v8 (stopped ep22, sens stuck 3-3.5) | `sifq/checkpoints_full_v8/last.pt` |
+| SIFQ v7 nofvc (ref) | `sifq/checkpoints_full_v7_nofvc/last.pt` |
+
+## Lịch sử checkpoint
+
+| Version | Epochs | Vấn đề | Trạng thái |
+|---------|--------|--------|----------|
+| `checkpoints_full/` | 21 | beta=0 bug → GRL tắt | Lỗi |
+| `checkpoints_full_v2/` | 50 | spread loss bimodal | Lỗi |
+| `checkpoints_full_v3/` | 30 | Q collapse, deg dead | Lỗi |
+| `checkpoints_full_v5/` | 25 | Chỉ noise degradation | l_deg≈0 |
+| `checkpoints_full_v6/` | 80 | Full deg pipeline | Ref |
+| `checkpoints_full_v7_nofvc/` | 80 | Score collapse (spread=0.010), L_sens explosion ep17-18, γ=0.40 làm L_deg plateau 0.104 | Ref |
+| `checkpoints_full_v8/` | 22 | Uniform spread, β ramp, γ=0.50, cosine LR, FVC — dừng ep22 vì α/γ step change tại ep15 gây sens oscillate 3–3.5 | Stopped |
+| `checkpoints_full_v9/` | 63 | Ramp ALL THREE (α/β/γ) ep10–19 + PolyU — dừng ep63 vì β/α=0.50 trong S2 quá yếu, sens oscillate 2–3 suốt S2/S3 | Stopped |
+| `checkpoints_full_v10/` | early | PolyU excluded từ L_sens — dừng sớm, sai thiết kế | Stopped |
+| `checkpoints_full_v11/` | 80 | Score collapse: 95% ảnh tại ~58.3, wet_press concept sai chiều | Completed |
+| `checkpoints_full_v12/` | 80 | T17+T18+T19: L_deg all datasets, fix wet_press erode, gamma=1.0 | Completed |
+| `checkpoints_full_v13/` | 80 | Revert T17, per-dataset L_spread, T25 noise fix, T26 occlusion fix | Completed |
+| `checkpoints_full_v14/` | 60 | Train from scratch; exclude non-seg slap; KS=0.263; Pearson=0.294 | Completed |
+| `checkpoints_full_v15/` | 60 | T30: SD302 concept-only L_deg, gamma=2.0; score collapse 24–56 ❌ | Completed |
+| `checkpoints/v16/` | 60 | T31: per-identity stats ALL datasets; score collapse 46.9–50.5 ❌; L_mat vs L_pair conflict | Completed |
+| `checkpoints/v17/` | 60 | T32: FVC-only stats; score collapse 18.1–53.0 (q_std=0.60 ❌); GRL suppresses SD302 quality | Completed |
+| `checkpoints/v18/` | 60 | T33: --no-mat-stats + W_SPREAD=2.0 + gamma=0.5 + SD302-A; revert to v14 design; score collapse q_std≈0.01, KS=0.249⚠️(false positive) ❌ | Completed ❌ |

rules/SIFQ_explained.md

@@ -0,0 +1,960 @@
+# 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` và `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` và `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)
+```

rules/plan.md

@@ -0,0 +1,320 @@
+# SIFQ — Plan & Experiment Tracker
+
+Dùng file này để: ghi kế hoạch trước khi thay đổi code, track tiến độ, và revert khi cần.
+
+---
+
+## Active Experiments
+
+### v32 — Done ❌
+**Goal:** Fix two concept grounding failures in v31 via T42 concept map
+**Key changes from v31:** T42 concept map (2 changes):
+- `dry_skin: [4, 2, 0] → [4, 0]` — remove continuity
+- `noise: [3] → [1, 3]` — add clarity as co-target
+**Script:** `scripts/run_train_v32.sh`
+**Result:** KS=0.1375, Pearson=0.7892 — Track 2 regression vs v31 (KS +0.022, Pearson -0.097)
+**Track 4 regressions vs v31:**
+- blur→clarity: +0.046 ✗ (was -0.305 ✓)
+- noise→clarity: +0.114 ✗ (was -0.236 ✓)
+- noise→noise_level: +0.121 ✗ (was -0.081 ✓)
+- occlusion→minutiae: +0.121 ✗ (was -0.123 ✓)
+- wet_press→minutiae: +0.676 ✗ (was -0.695 ✓)
+**Root cause:** Adding clarity[1] to noise map caused clarity concept to absorb noise-degradation gradient, destroying its response to blur. Removing continuity[2] from dry_skin destabilised concept interactions across other degradation types.
+**Reverted:** concept map → T39 (restored `noise:[3]`, `dry_skin:[4,2,0]`)
+**Status:** Failed — reverted to T39 concept map
+
+---
+
+### v31 — Done ✅
+**Goal:** Benchmark DINOv2-ViTS/14 as public teacher replacing unpublished MDGT
+**Key change from v28:** `--teacher dinov2` (frozen DINOv2-ViTS/14 CLS token [B,384])
+**Script:** `scripts/run_train_v31.sh`
+**Result:** KS=0.1152, Pearson=0.8858, q_std≥15 ✅ — **best Track 2 so far**
+**Concept issues:** dry_skin→noise_level Spearman -0.758 (spurious, noise_level NOT in target);
+blur→continuity +0.214 (wrong direction); noise→noise_level -0.081 (weak signal)
+**Status:** Complete
+
+---
+
+### v29 — Done ❌
+**Goal:** Fix concept grounding failures in v28 — T41 concept map + anti-saturation losses
+**Key changes from v28:**
+- T41 concept map: jpeg→[1], dry_skin→[4,0], noise→[1,3] (fix gradient conflicts + noise inversion)
+- `--ortho-weight 3.0` (vs 1.0 default) — push saturated concepts apart
+- `--concept-spread-weight 1.0` — penalise concepts with batch std < 0.20
+**Script:** `scripts/run_train_v29.sh`
+**Result:** KS=0.2985, Pearson=0.1448 — catastrophic sensor invariance regression
+**Root cause of failure:** `--concept-spread-weight 1.0` forces per-batch concept diversity
+which amplifies sensor-specific texture features, destroying cross-sensor score alignment.
+G_roll_png and H_roll_png particularly affected (KS 0.86 and 0.77 vs other sensors).
+**Status:** Failed — reverted to v28 code base (train_sifq.py, degradation_ranking.py T39)
+
+---
+
+### v28 — Done ✅
+**Goal:** Validate matcher-free quality learning — L_mat disabled entirely (`--no-mat`)
+**Key change from v26:** No MDGT teacher. Loss = L_sens + L_deg + L_ortho + L_spread only.
+**Script:** `scripts/run_train_v28.sh`
+**Result:** KS=0.1346, Pearson=0.7645, q_std~23.2 (60 epochs)
+**Concept issues found:** continuity collapsed (mean=0.068), noise_level inverted (+0.56), orient_coh saturated (0.94), contrast_uni saturated (0.85), clarity flat (std=0.043). Only minutiae_rel discriminating.
+**Paper claim:** ✅ "SIFQ quality is self-supervised — no external matcher needed" validated by KS close to v24
+**Status:** Complete → concept grounding needs fix (next experiment TBD)
+
+---
+
+### v26 — Đang train 🚀
+**Goal:** Verify v25 regression root cause = gamma=1.5 (not T40 concept map)  
+**Key change from v25:** `--concept-deg-gamma 2.0` (restored), DEGRADATION_CONCEPT_MAP reverted to T39  
+**Script:** `scripts/run_train_v26.sh`  
+**Log:** `logs/train_v26.log`  
+**Expected:** KS ≈ 0.126, Pearson ≈ 0.80 (match v24)  
+**Eval:** `scripts/run_eval_v26.sh` (auto-runs after training)  
+**Status:** Training in progress
+
+### v27 — Pending ⏳
+**Goal:** Validate SpatialConceptHead (14×14 spatial tokens) improves concept grounding  
+**Key change from v26:** `--spatial-concept-head` flag — `SpatialConceptHead` replaces `ConceptHead`  
+**Script:** `scripts/run_train_v27.sh`  
+**Expected:** KS ≈ v26, Track 4 diagonal stronger (especially orientation_coherence, continuity, minutiae_reliability)  
+**Blocker:** Wait for v26 to confirm KS/Pearson target first  
+**Status:** Code ready, not launched
+
+---
+
+## Code Change Log
+
+### 2026-06-04 — T42 concept map: 2 changes for v32
+
+**Files changed:**
+- `src/losses/degradation_ranking.py` — T42 DEGRADATION_CONCEPT_MAP:
+  - `dry_skin: [4,2,0] → [4,0]` (removed continuity [2])
+  - `noise: [3] → [1,3]` (added clarity [1] as co-target)
+
+**Root cause (dry_skin):** v31 shows noise_level (c=3) Spearman -0.758 with dry_skin (NOT in target).
+Continuity [2] shares backbone patch-scale features (~16×16) with noise texture → cross-activation.
+Fix: use contrast[4] + orientation[0] (regional features at multi-patch scale) only.
+
+**Root cause (noise):** v31 shows noise→noise_level Spearman only -0.081 (nearly no signal).
+TinyViT 16×16 patch embed averages out pixel Gaussian noise (σ=5–30) → concept[3] gradient ≈ 0.
+Adding clarity[1] anchors noise degradation to ridge-valley blur (detectable at patch scale).
+
+**Backward compatibility:** ⚠️ Modifies shared DEGRADATION_CONCEPT_MAP.
+
+**Revert T42 → T39:**
+```python
+"noise":     [3],
+"dry_skin":  [4, 2, 0],
+```
+
+---
+
+### 2026-06-03 — DINOv2Teacher: replace MDGT with public DINOv2-ViTS/14
+
+**Files changed:**
+- `src/training/mdgt_teacher.py` — Added `DINOv2Teacher` class: frozen DINOv2-ViTS/14 (via `torch.hub`), handles grayscale→RGB channel repeat + ImageNet normalization internally, returns L2-normalized [B, 384] CLS embeddings. `MDGTCheckpointTeacher` unchanged.
+- `src/train.py` — Added `--teacher {dinov2,mdgt}` arg (default=`dinov2`). Import `DINOv2Teacher`. Instantiation in `main()` dispatches on `args.teacher`. `--mdgt-checkpoint` arg remains but only used when `--teacher=mdgt`.
+
+**Root cause / motivation:** MDGT is unpublished work → not reproducible by reviewers → academic integrity risk. DINOv2-ViTS/14 is public (Meta, ICLR 2024), cite-able, and `torch.hub` reproducible.
+
+**Expected Pearson impact:** DINOv2 raw → expect Pearson ~0.55–0.70 vs 0.80 with MDGT. Run v31 to benchmark.
+
+**Backward compatibility:** ✅ All existing scripts using `--teacher mdgt --mdgt-checkpoint <path>` unaffected.
+
+**Revert:**
+```bash
+# train.py: change --teacher default back to "mdgt"
+# or pass --teacher mdgt --mdgt-checkpoint <ckpt_path> explicitly
+```
+
+---
+
+### 2026-06-02 — T41 concept map + anti-saturation losses (v29)
+
+**Files changed:**
+- `src/losses/degradation_ranking.py` — T41 DEGRADATION_CONCEPT_MAP:
+  - `jpeg: [2,1] → [1]` (remove continuity — JPEG artifacts wrong gradient direction in v28)
+  - `dry_skin: [4,2,0] → [4,0]` (remove continuity — gradient conflict, Spearman+0.10 wrong)
+  - `noise: [3] → [1,3]` (add clarity co-target — noise blurs ridges, anchors concept 3)
+- `scripts/train_sifq.py` — Added `--ortho-weight` (default 1.0) and `--concept-spread-weight` (default 0.0) args; l_orth now weighted; per-concept spread loss (std < 0.20 → penalty) added to total loss; logged as `l_cspread` in running dict and epoch print
+
+**Root cause:** v28 concept collapse/inversion diagnosed via inference stats:
+  - continuity std=0.076, mean=0.068 → dead concept (3 conflicting grad sources: blur✓, jpeg✗, dry_skin✗)
+  - noise_level Spearman +0.56 → inverted (Gaussian noise increases texture energy in TinyViT)
+  - orient_coh mean=0.942, contrast_uni mean=0.851 → L_ortho=1.0 too weak to break saturation
+
+**Backward compatibility:** ✅ `--ortho-weight` default=1.0, `--concept-spread-weight` default=0.0 → all v16-v28 unaffected.
+
+**Revert T41 → T39:**
+```python
+# degradation_ranking.py DEGRADATION_CONCEPT_MAP:
+"noise":     [3],
+"jpeg":      [2, 1],
+"dry_skin":  [4, 2, 0],
+```
+
+---
+
+### 2026-06-02 — --no-mat flag (v28 teacher-free experiment)
+
+**Files changed:**
+- `scripts/train_sifq.py` — Added `--no-mat` flag; MDGT teacher/emb_cache/prototypes wrapped in `if not args.no_mat`; training loop forces `l_mat = 0.0` when flag is set
+
+**Backward compatibility:** ✅ Default `--no-mat=False` — all existing versions unaffected.
+
+**Revert:** Remove `--no-mat` block in `parse_args()` and restore unconditional MDGT instantiation in `main()`.
+
+---
+
+### 2026-06-01 — SpatialConceptHead
+
+**Files changed:**
+- `src/models/concept_head.py` — Added `SpatialConceptHead` class
+- `src/models/sifq.py` — `SIFQ.forward()` dispatches via `concept_head.uses_spatial`
+- `src/models/__init__.py` — Export `SpatialConceptHead`
+- `scripts/train_sifq.py` — `--spatial-concept-head` flag (default=False)
+- `src/train.py` — same flag
+- `scripts/run_eval.py` — auto-detect from `ckpt["config"]["spatial_concept_head"]`
+- `scripts/run_infer.py` — same auto-detect
+- `rules/SIFQ_explained.md` — Section 2.2 updated, diagram updated, File Map updated
+
+
+**Backward compatibility:** ✅ All v16–v26 checkpoints load cleanly without flag.
+
+---
+
+### 2026-06-01 — T39 concept map revert + gamma restore (v26)
+
+**Problem:** v25 used gamma=1.5 AND T40 concept map. Regression KS 0.126→0.213.  
+**Root cause identified:** gamma=1.5 too weak during S1→S2 ramp.  
+  - v24 (gamma=2.0): spread stable ~0.003, q_std grows 18→22 monotonically
+  - v25 (gamma=1.5): spread spikes to 0.028, q_std collapses 22→10 (epochs 13–16)
+
+**Files changed:**
+- `src/losses/degradation_ranking.py` — T40 reverted to T39 DEGRADATION_CONCEPT_MAP
+- `scripts/run_train_v26.sh` — `--concept-deg-gamma 2.0`
+
+**DEGRADATION_CONCEPT_MAP T39 (current, correct):**
+```python
+"blur":      [1, 2]
+"noise":     [3]
+"jpeg":      [2, 1]
+"occlusion": [5]
+"dry_skin":  [4, 2, 0]
+"wet_press": [1, 5, 0]
+```
+
+**Revert target:** T40 map (v25):
+```python
+"blur":      [1, 2, 0]    # + orient_coh
+"noise":     [3, 4]       # + contrast_u
+"jpeg":      [2, 1, 4]    # + contrast_u
+"occlusion": [5]
+"dry_skin":  [4, 2, 0]
+"wet_press": [1, 5, 0]
+```
+**Do NOT revert to T40** unless v26 confirms gamma=2.0 alone is insufficient and T39 concept grounding is weaker than expected.
+
+---
+
+## Revert Cookbook
+
+### Revert concept map to previous version
+```bash
+# Check what the map looked like in a specific git commit:
+git log --oneline src/losses/degradation_ranking.py
+git show <commit>:sifq/src/losses/degradation_ranking.py | grep -A 30 "DEGRADATION_CONCEPT_MAP"
+
+# Edit directly:
+# src/losses/degradation_ranking.py — DEGRADATION_CONCEPT_MAP dict
+```
+
+### Revert to v24 hyperparameters (known good baseline)
+```bash
+# Key v24 flags (from scripts/run_train_v24.sh):
+--concept-deg-gamma 2.0
+--spread-weight 3.0
+--spread-mode uniform
+--deg-every-n-steps 2
+--no-mat-stats
+--proto-max-batches 0
+--k-cross 0
+--batch-size 96
+```
+
+### Load and inspect a checkpoint
+```python
+import torch
+ckpt = torch.load("checkpoints/v24/last.pt", map_location="cpu", weights_only=False)
+print(ckpt["metrics"])     # KS, Pearson, q_std, etc.
+print(ckpt["config"])      # all argparse flags used
+print(ckpt["epoch"])       # which epoch
+```
+
+### Compare two checkpoints' configs
+```python
+import torch, json
+c24 = torch.load("checkpoints/v24/last.pt", map_location="cpu", weights_only=False)["config"]
+c25 = torch.load("checkpoints/v25/last.pt", map_location="cpu", weights_only=False)["config"]
+for k in c24:
+    if c24.get(k) != c25.get(k):
+        print(f"{k}: v24={c24.get(k)}  v25={c25.get(k)}")
+```
+
+### Run eval manually on any checkpoint
+```bash
+cd /home/aiserver/works/fingerprint
+
+# Step 1: Inference
+python sifq/scripts/run_infer.py \
+    --checkpoint sifq/checkpoints/vXX/last.pt \
+    --root-302a  dataset/302a/images/challengers \
+    --root-302b  dataset/302b/images/baseline \
+    --root-302d  dataset/nist_302d/images/auxiliary \
+    --exclude-sensor "R_1000_slap,R_500_slap,S_500_slap" \
+    --output sifq/eval_results/sifq_scores_vXX.jsonl
+
+# Step 2: Track 2 + Track 4
+python sifq/scripts/run_eval.py \
+    --sifq-scores sifq/eval_results/sifq_scores_vXX.jsonl \
+    --checkpoint  sifq/checkpoints/vXX/last.pt \
+    --out-dir     sifq/eval_results/vXX \
+    --exclude-sensor "R_1000_slap,R_500_slap,S_500_slap" \
+    --skip-track1
+```
+
+### Smoke test (~2 min)
+```bash
+cd /home/aiserver/works/fingerprint
+bash sifq/scripts/run_smoke.sh
+```
+
+---
+
+## Decision Log
+
+### Why not T40 concept map?
+T40 added `orient_coh` to blur/noise/jpeg. Reverted in v26 because root cause of v25 regression was **gamma=1.5**, not concept map. T40 change had no confirmed benefit. Defer until v26 evaluation is done.
+
+### Why FVC-only L_deg (not SD302)?
+SD302 images are all high quality by protocol → clean images satisfy L_rank at a single ~53 attractor → blocks quality discrimination signal from L_mat. FVC has genuine quality variation (8 impressions/subject) → MDGT cosine varies → L_mat gradient meaningful → transfers to SD302 at inference. (v20 root cause analysis)
+
+### Why --no-mat-stats?
+Per-identity cosine stats (T31) created asymmetry: FVC has stable tanh targets, SD302 gets raw cosine ~0.85 constant. Backbone learned "FVC=quality-variable, SD302=fixed" → no L_mat gradient for SD302 → GRL destroyed SD302 features. (v17 root cause)
+
+### Why --k-cross 0?
+k_cross>0 forces k guaranteed cross-sensor pairs per batch → over-constrains L_sens → backbone over-optimises sensor invariance → loses quality discrimination → Pearson collapses to ~0.09. (v21 root cause)
+
+### Why --proto-max-batches 0?
+Partial prototypes (150 batches = 45% data) → SD302 identities have 2–3 sensor prototypes instead of full 19-sensor → cosine correlates with WHICH sensors in prototype window (sensor-biased) → no quality gradient for SD302. (v19/v20 root cause)
+
+### Why gamma must be ≥ 2.0?
+During S1→S2 ramp, mat loss is introduced alongside existing deg loss. If deg loss too weak (gamma=1.5), mat loss dominates momentarily → spread spikes → model partially collapses and never fully recovers. gamma=2.0 keeps deg strong enough to maintain ordinal grounding through the ramp. (v25 root cause)
+
+---
+
+## Pending Research Questions
+
+- [ ] **v26 eval:** Does restoring gamma=2.0 fully recover v24 KS/Pearson? (expected: yes)
+- [ ] **v27 eval:** Does SpatialConceptHead improve Track 4 diagonal ρ vs v26?
+- [ ] **noise→noise_level:** v24 Track 4 shows +0.365 (wrong direction). Defer to v28.
+- [ ] **T40 revisit:** Once v26/v27 stable, evaluate if blur/noise/jpeg → orient_coh improves orientation_coherence grounding.

rules/sifq_pdf.txt

@@ -0,0 +1,488 @@
+SIFQ: Sensor-Invariant Fingerprint Quality Assessment
+A Self-Supervised, Concept-Grounded Alternative to NFIQ2
+Research Design Document
+Author: Bình An
+Collaborator / Drafting: Claude
+Version 0.1 · Research Design
+
+1. Tóm tắt ý tưởng
+SIFQ (Sensor-Invariant Fingerprint Quality) là một framework đánh giá chất lượng vân tay được huấn luyện
+hoàn toàn theo hướng self-supervised và weakly-supervised, không phụ thuộc vào nhãn chất lượng do
+con người đánh. Khác với NFIQ2 (SOTA) vốn dùng match score từ VeriFinger-era matcher làm supervision
+gián tiếp và chịu sensor-bias rõ rệt khi đánh giá dữ liệu cross-modality, SIFQ kết hợp ba tín hiệu huấn luyện
+bổ trợ nhau: matcher embedding concentration từ MDGT làm teacher, cross-sensor invariance enforced
+trên paired data, và controlled degradation ranking trên synthetic perturbations.
+Đầu ra của SIFQ gồm một vector concepts có thể diễn giải (ridge clarity, orientation coherence, continuity,
+noise level, contrast uniformity, minutiae reliability) và một scalar utility score tương thích thang đo
+NFIQ2. Đóng góp cốt lõi về mặt phương pháp luận là việc attack trực diện vấn đề sensor-bias mà literature
+hiện tại chưa giải quyết, và lần đầu tiên đưa concept bottleneck vào fingerprint quality assessment.
+Tài liệu trình bày kiến trúc, ba loss functions, training protocol bốn stage, evaluation protocol năm track,
+các rủi ro chính cùng mitigation tương ứng, và mối quan hệ với các dự án MDGT/OMFR hiện đang triển
+khai song song.
+
+2. Động lực và tuyên bố vấn đề
+2.1. Hạn chế của NFIQ2 trong bối cảnh matcher hiện đại
+NFIQ2 được phát triển cuối thập niên 2010 và trở thành chuẩn ISO/IEC 29794-4. Tuy nhiên đang gặp 3 vấn
+đề lớn trong bối cảnh biometric hiện đại.
+Thứ nhất, NFIQ2 học supervised với target là match score của VeriFinger. Điều này đóng khung quality
+như một utility predictor cho một matcher cụ thể, không nhất thiết chuyển giao được sang matcher khác,
+đặc biệt là deep matcher dựa trên learned embeddings như FLaRE, AFR-Net. Khi matcher thay đổi
+paradigm từ minutiae-based sang representation-based, quality signal cũng cần được recalibrate.
+Thứ hai, các feature tay trong NFIQ2 (OCL, LCS, FDA, RVU, OFL) được thiết kế trên giả định sensor optical
+contact đồng nhất. Khi áp dụng lên contactless fingerprint hoặc latent prints, nhiều feature trả về giá trị
+cực thấp vì đặc tính thống kê của ảnh khác biệt căn bản, dẫn đến Q(x) thấp mặc dù matcher hiện đại vẫn
+có thể match tốt những ảnh này.
+Thứ ba, NFIQ2 không có cơ chế bảo đảm sensor invariance. Trong quan sát thực nghiệm trên NIST SD302
+mà chúng tôi ghi nhận trong quá trình phát triển MDGT, cùng một ngón tay capture trên hai sensor khác
+nhau thường nhận được NFIQ2 score khác biệt 20-30 điểm, nghĩa là metric đang capture sensor signature
+nhiều hơn biometric content.
+
+Tóm lại, th NFIQ2 chỉ đánh giá tốt với Minuatie, còn với các model trên fix-length hoặc feature khác thì
+giở. Đặc biệt lại khi kiểm định cùng 2 sensors, 1 ngón tay sẽ trả về 2 điểm NFIQ2 khác nhau
+
+2.2. Vấn đề sensor-bias từ kinh nghiệm MDGT
+Trong quá trình phát triển MDGT trên NIST SD302, chúng tôi quan sát thấy hiện tượng shortcut learning:
+impostor similarity trong cùng một sensor cao hơn genuine similarity cross-sensor. Điều này cho thấy
+không chỉ matcher mà mọi feature extractor có khả năng capture sensor identity thay vì biometric identity
+khi không có cơ chế debiasing rõ ràng. Quality predictor nếu huấn luyện ngây thơ cũng sẽ mắc cùng failure
+mode.
+Một quality metric đáng tin cậy cho biometric system cần bất biến với sensor, chỉ phản ánh thông tin
+ridge-valley thực chất. Đây là một gap mà literature hiện tại chưa lấp và là cơ hội về mặt nghiên cứu.
+
+2.3. Nhu cầu về task-aware và interpretable quality
+Trong một fingerprint biometric pipeline hiện đại, một ảnh có thể có chất lượng khác nhau tuỳ task. Ảnh
+contactless có thể tốt cho identity matching nhưng tệ cho liveness/PAD vì thiếu texture bề mặt. Một
+quality score duy nhất không thể phục vụ mọi downstream task. SIFQ đưa ra giải pháp thông qua concept
+decomposition, cho phép downstream system quyết định concept nào quan trọng cho task của nó.
+Ngoài ra, khi operator ở sensor nhận thông báo quality thấp, việc biết được vấn đề nằm ở concept nào
+(noise quá cao, ridge clarity kém, occlusion) cho phép hướng dẫn re-capture hiệu quả hơn, thay vì chỉ báo
+chung một con số.
+
+3. Related works
+3.1. Fingerprint quality assessment
+NFIQ (2004) và phiên bản kế nhiệm NFIQ2 (2016, sau này trở thành ISO/IEC 29794-4) là hai baseline chính.
+NFIQ2 dùng tập hợp 14 features bao gồm orientation certainty, minutiae count and quality, ridge-valley
+uniformity, frequency domain analysis, tổng hợp qua một random forest hoặc gradient boosting classifier
+regress về match score. Hạn chế đã phân tích ở Mục 2.1.
+MiDeCon (Tapia et al., 2022) thử nghiệm deep quality estimation nhưng vẫn huấn luyện supervised với
+NFIQ2 labels làm pseudo-ground-truth, vô hình chung kế thừa các bias của NFIQ2. Một số nghiên cứu sử
+dụng CNN để classify quality tốt/xấu dưới dạng binary task, nhưng mất đi tính ordinal của quality score.
+
+3.2. Face quality methods áp dụng cho fingerprint
+SER-FIQ (Terhörst et al., 2020) giới thiệu stochastic embedding robustness như quality signal: với một ảnh
+x, chạy multiple forward passes với dropout kích hoạt, quality là độ ổn định của embeddings. Không cần
+labels nhưng đòi hỏi dropout trong architecture và nhiều forward passes lúc inference.
+SDD-FIQA (Ou et al., 2021) sử dụng similarity distribution distance: với mỗi sample, so sánh genuine và
+impostor similarity distributions, quality là mức độ tách biệt. Yêu cầu identity labels.
+
+CR-FIQA (Boutros et al., 2023) là closest conceptual analog: quality được định nghĩa là class-center ratio,
+hiệu giữa intra-class distance và nearest inter-class distance trong embedding space. Được train như
+regression head trên matcher backbone, không cần labels ngoài identity.
+Điểm chung của ba method này là chúng được phát triển cho face, nơi mà cross-sensor/cross-modality
+không phải vấn đề chính. Khi chuyển sang fingerprint, đặc biệt với các dataset cross-modality như PolyU
+CL2CB, chúng sẽ kế thừa sensor-bias của backbone matcher.
+
+3.3. Concept bottleneck models
+Koh et al. (2020) đề xuất Concept Bottleneck Models: thay vì map input trực tiếp đến target, model học
+trước một tập concepts có tên, rồi concepts predict target. Được chứng minh cho tác vụ classification trên
+medical imaging và birds. Chưa có công bố nào đưa concept bottleneck vào fingerprint quality assessment.
+Áp dụng concept bottleneck vào fingerprint quality có ba lợi ích: interpretability (operator biết vấn đề cụ
+thể), controllability (có thể intervene vào concept values để test what-if), và regularization (concepts
+grounded ngăn model overfit vào spurious correlations).
+
+3.4. So sánh tóm tắt
+Method
+
+Supervision
+
+Teacher
+
+Sensorinvariant
+
+Interpretable
+concepts
+
+Domain
+
+NFIQ2
+
+Supervised
+
+VeriFinger
+scores
+
+No
+
+Partial (handcrafted)
+
+Fingerprint
+
+SER-FIQ
+
+Self-supervised
+
+Dropoutensemble
+
+No
+
+No
+
+Face
+
+SDD-FIQA
+
+Weaklysupervised
+
+Similarity
+distribution
+
+No
+
+No
+
+Face
+
+4. Phương pháp đề xuất
+4.1. Tổng quan kiến trúc
+SIFQ gồm ba thành phần chính. Backbone là một ViT nhẹ, đề xuất TinyViT-2M hoặc MobileViT tương
+đương, nhận input 1-channel fingerprint image (hoặc 3-channel sau Gabor stem nếu muốn share
+preprocessing với MDGT) và trả về feature map spatial tại resolution 14x14. Concept head là một MLP với
+k output nodes, mỗi node là một concept scalar trong khoảng [0,1] thông qua sigmoid activation.
+Aggregator là một MLP nhỏ nhận concept vector và predict scalar Q(x) trong thang [0,100].
+Chọn backbone không share weights với MDGT có hai lý do. Thứ nhất, SIFQ cần deploy được on-sensor
+với latency thấp (mục tiêu dưới 50ms trên ARM Cortex-A72), MDGT quá nặng. Thứ hai, giữ hai model tách
+
+biệt tránh circular dependency giữa quality và matching, nơi mà improvement ở một bên ngẫu nhiên dẫn
+đến degradation ở bên kia do shared representation drift.
+
+4.2. Concept set
+Chọn k=6 concepts đầu tiên, có thể mở rộng trong follow-up. Lựa chọn concepts dựa trên các yếu tố đã
+được fingerprint expert community công nhận và có thể induce được bằng controlled degradation.
+1. orientation_coherence: độ nhất quán của orientation field trong các local patches. Cao khi ridge
+flow smooth, thấp khi có đứt gãy hoặc chỗ chuyển hướng đột ngột.
+2. ridge_valley_clarity: độ sắc nét của biên ridge-valley. Thấp khi blur, compression, hoặc dry skin
+làm mờ boundary.
+3. continuity: mức độ liên tục của ridge lines. Giảm khi có gap, fragmentation, hoặc overbinarization.
+4. noise_level: mức nhiễu background và foreground. Tăng với sensor noise, JPEG artifacts, scan
+noise.
+5. contrast_uniformity: độ đồng đều contrast across foreground. Thấp khi dry skin, uneven
+pressure, partial contact.
+6. minutiae_reliability: expected reliability của minutiae extraction. Tổng hợp từ ba concepts
+trước nhưng có signal riêng khi occlusion hoặc edge-cropping xảy ra.
+Lưu ý rằng concepts không cần độc lập hoàn toàn về mặt toán học. Blur affect cả ridge_valley_clarity và
+continuity đồng thời, điều này chấp nhận được và được reflect trong degradation-concept mapping ở
+Mục 4.4.
+
+4.3. Ba tín hiệu huấn luyện
+4.3.1. Signal 1: Matcher-as-teacher (L_mat)
+Cho mỗi identity y có đủ samples (tối thiểu 5 samples per identity), compute class prototype c_y là trung
+bình của MDGT embeddings đã L2-normalize. Với mỗi sample x có label y, quality signal từ matcher là:
+q_mat(x) = ( cos(MDGT(x), c_y) − mu_y ) / sigma_y
+
+trong đó mu_y và sigma_y là mean và standard deviation của cosine similarity trong class y. Việc chuẩn
+hoá per-identity giúp tránh bias toward identities có ít intra-variance (tức là identities có nhiều sample rất
+giống nhau thì không được ưu tiên hơn identities có variance lớn hơn một cách tự nhiên).
+SIFQ được huấn luyện regress q_mat(x) bằng Huber loss, robust hơn MSE khi có outliers:
+L_mat = Huber( Q_normalized(x), q_mat(x) )
+
+MDGT được frozen trong suốt training. L_mat một mình không đủ vì nó kế thừa toàn bộ bias của MDGT
+bao gồm sensor-bias.
+
+4.3.2. Signal 2: Cross-sensor invariance (L_sens)
+Cho mỗi paired sample (x_s1, x_s2) cùng ngón tay khác sensor, enforce ràng buộc Q bất biến qua margin:
+L_pair = max( 0, | Q(x_s1) − Q(x_s2) | − delta )
+
+Margin delta = 0.05 cho phép fluctuation nhỏ nhưng penalize drift hệ thống. Ngoài pairwise constraint,
+một adversarial term được thêm vào để enforce global invariance:
+L_adv = − log D(sensor | f_intermediate(x))
+
+trong đó D là một sensor classifier nhận intermediate features từ concept head và cố gắng predict sensor
+type. Backbone và concept head được train ngược lại để fool D thông qua gradient reversal layer
+(lambda_adv = 0.3 mặc định). Tổng loss:
+L_sens = L_pair + lambda_adv * L_adv
+
+Đây là contribution cốt lõi về mặt novelty. NFIQ2 hoàn toàn không có mechanism tương tự, và các face
+quality method cũng không vì face thường capture trong single modality. SIFQ là method đầu tiên đưa
+invariance constraint trực tiếp vào quality learning.
+
+4.3.3. Signal 3: Controlled degradation ranking (L_deg)
+Build một degradation pipeline sáu loại, mỗi loại có bốn levels cường độ. Với mỗi clean image x, sample
+hai levels i < j trong cùng degradation type, và enforce ordinal ranking:
+L_rank = max( 0, Q(x_j) − Q(x_i) + m ) + max( 0, Q(x_i) − Q(x) + m )
+
+với margin m = 0.1. Ngoài ranking loss trên scalar Q, thêm một supervision concept-level quan trọng: mỗi
+degradation type được gán một hoặc nhiều target concepts, và concepts tương ứng phải phản ứng
+monotonically theo level. Ví dụ, với blur, ridge_valley_clarity và continuity đều phải giảm khi level tăng.
+Mapping cụ thể được trình bày trong Mục 4.4. Concept supervision này transform concepts từ latent
+variables vô nghĩa thành grounded representations. Nếu không có bước này, concepts chỉ là 6 output
+neurons tình cờ được đặt tên, không có quan hệ nào với tên của chúng.
+L_concept_deg = sum over (degradation, target_concept) of Huber(
+c_target(x_level), expected_response(level) )
+
+Tổng loss cho degradation signal:
+L_deg = L_rank + gamma * L_concept_deg
+
+với gamma = 0.5.
+
+4.3.4. Tổng loss và trade-off có chủ ý
+Tổng loss tại epoch t:
+L(t) = alpha(t) * L_mat + beta(t) * L_sens + gamma_stage(t) * L_deg + L_ortho
+
+trong đó alpha và beta ramp theo stage (xem Mục 5), L_ortho là một decorrelation loss nhẹ trên concept
+activations để hạn chế entanglement:
+
+L_ortho = || off_diag( corr(concept_activations) ) ||_F^2
+
+Balance giữa 3 loại signal. Đây là điểm conceptual quan trọng nhất. L_mat kéo Q theo hướng behavior
+của MDGT (kể cả bias). L_sens đẩy Q rời khỏi bất cứ signature nào của sensor. L_deg cung cấp ordinal
+anchor độc lập với cả hai. Khi huấn luyện đồng thời, ba lực tạo ra một equilibrium mà Q phải học quality
+nghĩa là gì không phụ thuộc sensor, nhưng vẫn predictive cho matcher hiện đại, và có grounding ordinal
+thực tế. Nếu chỉ dùng L_mat, Q clone bias của MDGT. Nếu chỉ L_sens, Q không có task relevance. Nếu chỉ
+L_deg, Q chỉ predict mức độ xuống cấp synthetic.
+
+4.4 Các loại concept-degradation
+Bảng sau định nghĩa cách mỗi loại degradation ảnh hưởng đến concepts nào. Đây là core của concept
+grounding mechanism.
+Degradation type
+
+Physical simulation
+
+Target concept(s)
+
+Expected response
+
+Blur
+
+Gaussian kernel, sigma 0.5
+to 3.0
+
+ridge_valley_clarity,
+continuity
+
+Both decrease
+monotonically
+
+Noise
+
+Additive Gaussian, sigma
+5 to 30
+
+noise_level
+
+Increases
+monotonically
+
+JPEG compression
+
+Quality factor 90 to 15
+
+continuity,
+ridge_valley_clarity
+
+Both decrease
+
+Partial occlusion
+
+Random mask, coverage
+10% to 40%
+
+minutiae_reliability
+
+Decreases with
+coverage
+
+Dry skin
+
+Local contrast reduction +
+fragmentation
+
+contrast_uniformity,
+continuity
+
+Both decrease
+
+Pressure (wet)
+
+Morphological dilation,
+ridge thickening
+
+ridge_valley_clarity,
+minutiae_reliability
+
+Both decrease
+
+Lưu ý thiết kế. Một degradation có thể ảnh hưởng nhiều concepts, điều này phản ánh thực tế ngoài đời.
+Blur thực tế giảm cả clarity và continuity. Mapping không cần hoàn hảo, chỉ cần đủ tín hiệu để concepts
+học được hướng đúng. Cross-talk không mong muốn giữa concepts được regulate bởi L_ortho và được
+measure trong Track 4 của evaluation.
+
+5. Datasets
+SIFQ dựa vào bốn dataset chính. Cấu hình yêu cầu paired data cho L_sens, identity labels cho L_mat, và
+clean images cho L_deg.
+
+Dataset
+
+Paired?
+
+Identity
+labels?
+
+Role in SIFQ
+
+Signal used
+
+Single-sensor
+
+Yes
+
+Matcher-teacher training,
+degradation base images
+
+L_mat, L_deg
+
+NIST SD302 a/b/d
+
+Yes (multisensor)
+
+Yes
+
+Cross-sensor invariance
+training, matcher-teacher
+
+L_sens, L_mat
+
+PolyU CL2CB
+
+Yes (crossmodality)
+
+Yes
+
+Hardest invariance case:
+contactless vs contact
+
+L_sens
+
+NIST SD300
+
+No
+
+Partial
+
+Held-out evaluation only
+
+Evaluation, no training
+
+FVC2002/2004
+
+Điểm quan trọng cần xác nhận sớm là paired structure của NIST SD302. Cần metadata đảm bảo same
+finger ID persistent across sensors, không chỉ same subject ID. Nếu chỉ có subject ID, việc pair dựa trên
+finger position dễ sai và cần manual audit cho một subset.
+PolyU CL2CB là trường hợp invariance khó nhất (contactless vs contact) và cũng là case thuyết phục nhất
+nếu SIFQ xử lý tốt. Dataset này nhỏ, chủ yếu dùng cho L_sens và evaluation Track 2, không đủ cho general
+training.
+
+7. Evaluation Protocol
+Năm evaluation tracks được thiết kế để address từng claim cụ thể của paper, kèm theo vai trò rhetorical
+trong narrative. Các tracks không đồng quyền, Track 1 và 2 là primary, Track 3-5 là supporting.
+#
+
+Track
+
+Metric
+
+Role in narrative
+
+1
+
+Error Rejection Curve
+
+FNMR@FMR=1e-4 vs rejection
+ratio; AUC_ERC
+
+Primary utility metric, directly
+comparable to NFIQ2
+
+2
+
+Sensor invariance
+
+KS statistic across sensor
+populations; cross-sensor Pearson
+of Q
+
+Core novelty claim figure
+
+3
+
+Cross-matcher transfer
+
+Q correlation with VeriFinger and
+AFR-Net error rates
+
+Rebuts the 'only distills MDGT'
+critique
+
+4
+
+Concept grounding
+
+Spearman rho between degradation
+level and concept score; cross-talk
+matrix
+
+Validates interpretability claim
+
+5
+
+Human correlation
+(optional)
+
+Spearman rho with expert Likert
+ratings on 300 to 500 images
+
+Strongest rhetorical evidence if
+available
+
+7.1. Track 1: Error Rejection Curve
+
+Đây là metric chuẩn trong quality literature, cho biết quality score hữu ích thế nào khi dùng để reject
+samples. Sắp xếp samples theo Q tăng dần, reject bottom x%, compute FNMR tại FMR fixed trên phần còn
+lại. Plot curve FNMR vs rejection ratio. Q tốt giảm FNMR nhanh khi tăng rejection.
+Baseline so sánh: NFIQ2 (industry standard), SER-FIQ và CR-FIQA (state-of-the-art self-supervised từ face
+domain adapted cho fingerprint). Matcher dùng để tính FNMR: cả VeriFinger v12 (classical) lẫn MDGT
+(deep). Thử nghiệm trên FVC2004, NIST SD302 single-sensor subset, PolyU CL2CB.
+Metric tổng hợp: AUC_ERC (area under ERC curve). Thấp hơn là tốt hơn.
+
+7.2. Track 2: Sensor-invariance quantification
+Đây là figure KEY của paper, cần plot rõ ràng và ấn tượng. Trên NIST SD302, tách samples theo sensor,
+plot histogram của Q cho mỗi sensor trên cùng một axis. NFIQ2 expected cho histograms rất khác nhau
+giữa optical và capacitive. SIFQ expected cho histograms overlap cao.
+Metric định lượng: Kolmogorov-Smirnov statistic giữa Q distributions cross-sensor. Thấp hơn là tốt hơn.
+Thêm một bảng với từng cặp sensor và KS của từng method. Thêm một scatter plot Q_sensor1 vs
+Q_sensor2 cho paired samples, Pearson correlation. Cao hơn là tốt hơn.
+Đây là track mà SIFQ phải win rõ ràng, nếu không paper mất contribution cốt lõi.
+
+7.3. Track 3: Cross-matcher transfer
+Sau khi train SIFQ với MDGT teacher, evaluate khả năng generalize sang matcher khác. Compute ERC của
+SIFQ với VeriFinger v12 làm matcher (metric tính FNMR dựa trên VeriFinger scores). Nếu Q vẫn giảm FNMR
+effectively, chứng tỏ SIFQ không chỉ distill MDGT mà học được general quality signal.
+Track này address reviewer concern phổ biến: 'ông chỉ dạy SIFQ bắt chước MDGT'. Nếu Q giảm FNMR trên
+matcher chưa từng thấy, critique đó bị refute.
+
+7.4. Track 4: Concept grounding
+Validate claim interpretability. Với mỗi degradation type, sweep level từ 0 (clean) đến max, plot response
+của mỗi concept. Tính Spearman correlation giữa degradation level và concept response. Target concepts
+phải có rho cao (gần 1 hoặc -1), non-target concepts phải có rho thấp.
+Plot cross-talk matrix: hàng là degradation type, cột là concept, cell là Spearman rho. Matrix lý tưởng có
+diagonal-heavy pattern (mỗi degradation ảnh hưởng chủ yếu đến target concept của nó). Matrix dày đặc
+cho thấy concept entanglement, cần mitigation.
+
+7.5. Track 5: Human correlation (tuỳ chọn)
+
+Mời 2-3 fingerprint examiners (ví dụ từ phòng forensic hoặc academia partner) rate 300 đến 500 ảnh trên
+Likert scale 1-5. Compute Spearman rho giữa expert rating trung bình và SIFQ score. Bonus: concept-level
+rating, expert rate từng concept riêng, so với SIFQ concept scores.
+Track này tốn công (2-3 ngày expert time, cần ethics clearance) nhưng nếu có kết quả tích cực thì là
+evidence rhetorical mạnh nhất. Optional và có thể để cho follow-up paper.
+
+

scripts/_gen_report_v24.py

@@ -0,0 +1,207 @@
+import json, numpy as np
+from scipy.stats import spearmanr
+from datetime import datetime
+
+records = [json.loads(l) for l in open('eval_results/sifq_scores_v24.jsonl')]
+CNAMES = ['orientation_coherence','ridge_valley_clarity','continuity','noise_level','contrast_uniformity','minutiae_reliability']
+concepts = np.array([r['concepts'] for r in records])
+qs = np.array([r['q_score'] for r in records])
+
+d = json.load(open('eval_results/v24/eval_summary.json'))
+t2 = d['track2_sensor_invariance']['SIFQ']
+t4 = d['track4_concept_grounding']
+
+TARGETS = {
+    'blur':      ['ridge_valley_clarity', 'continuity'],
+    'noise':     ['noise_level'],
+    'jpeg':      ['continuity', 'ridge_valley_clarity'],
+    'occlusion': ['minutiae_reliability'],
+    'dry_skin':  ['contrast_uniformity', 'continuity', 'orientation_coherence'],
+    'wet_press': ['ridge_valley_clarity', 'minutiae_reliability', 'orientation_coherence'],
+}
+
+SEP  = '=' * 70
+SEP2 = '-' * 70
+
+lines = [
+    SEP,
+    'SIFQ -- Concept Quality Model: Ly thuyet & Ket qua (v24)',
+    f'Generated: {datetime.now().strftime("%Y-%m-%d %H:%M")}',
+    SEP,
+    '',
+    '1. DINH NGHIA 6 CONCEPTS',
+    SEP2,
+    'Tat ca concepts: cao = chat luong tot hon, dau ra trong [0, 1].',
+    '',
+    '  #  Concept                    Y nghia vat ly',
+    '  -- -------------------------- -----------------------------------------------',
+    '  0  orientation_coherence      Ridge flow nhat quan, local orientation field smooth',
+    '  1  ridge_valley_clarity       Bien ridge-valley sac net, contrast cuc bo cao',
+    '  2  continuity                 Ridge lines lien tuc, khong bi dut gay',
+    '  3  noise_level                It nhieu ngau nhien (cao = it noise = tot)',
+    '  4  contrast_uniformity        Contrast deu tren toan foreground',
+    '  5  minutiae_reliability       Minutiae co the trich xuat chinh xac',
+    '',
+    '  Degradation Concept Map (v25 -- T40):',
+    '  blur      -> [clarity[1], continuity[2], orient_coh[0]]',
+    '  noise     -> [noise_level[3], contrast_u[4]]',
+    '  jpeg      -> [continuity[2], clarity[1], contrast_u[4]]',
+    '  occlusion -> [minutiae_reliability[5]]',
+    '  dry_skin  -> [contrast_u[4], continuity[2], orient_coh[0]]',
+    '  wet_press -> [clarity[1], minutiae_rel[5], orient_coh[0]]',
+    '',
+    '  Ly do thiet ke map nhu vay:',
+    '  - Moi concept phai duoc giam sat boi >= 2 loai degradation (tranh single-',
+    '    point-of-failure: 1 degradation target 1 concept -> signal yeu, de bi',
+    '    gradient interference invert chieu).',
+    '  - Chon degradation phu hop vat ly: jpeg blocking -> contrast bands (khong',
+    '    chi ridge artifacts), blur -> orientation blur (khong chi clarity loss).',
+    '',
+    '',
+    '2. CO CHE TRAINING CONCEPT (L_concept)',
+    SEP2,
+    'Voi moi cap anh (mild degradation vs severe degradation cung loai):',
+    '',
+    '  L_concept = SUM Huber( c_mild[c], c_severe[c] + 0.1 )',
+    '              c in targets(degradation_type)',
+    '',
+    '  -> Anh degradation nhe phai co concept cao hon anh degradation nang >= 0.1',
+    '  -> Model hoc tung concept phan ung dung chieu voi loai hu hong tuong ung',
+    '',
+    'Ket hop ranking loss:',
+    '  L_rank = relu(Q_severe - Q_mild + m) + relu(Q_mild - Q_clean + m)',
+    '  -> Q giam theo thu tu: clean > mild_deg > severe_deg',
+    '',
+    '  L_deg = L_rank + gamma * L_concept',
+    '  gamma = 2.0 (v24) -> 1.5 (v25)',
+    '    - gamma qua cao (2.0): gradient conflict qua manh -> noise_level inversion',
+    '    - gamma qua thap (0.5, v22): blur->continuity FAIL (+0.261)',
+    '    - gamma = 1.5: compromise, du manh cho blur/jpeg, khong gay inversion',
+    '',
+    '',
+    '3. VAN DE CONCEPT SATURATION',
+    SEP2,
+    'Nguyen nhan goc: L_concept chi train tren synthetic degradation pairs.',
+    'Voi real fingerprint images, KHONG co gradient dinh huong concept.',
+    '-> Concept troi ve gia tri mac dinh cua backbone features.',
+    '',
+    'Hau qua trong v24 (42,683 real fingerprint images):',
+    '',
+    '  Concept                      mean    std   rho_Q  Tinh trang',
+    '  ---------------------------- ------  -----  ------  ---------------------------',
+]
+
+STATUS = {
+    'noise_level':          '!! DOMINATES Q (rho=-0.989)',
+    'minutiae_reliability': '!! Saturated HIGH -- dead (range 0.84-0.93)',
+    'orientation_coherence':'!! Saturated LOW  -- T39 overcorrected',
+    'ridge_valley_clarity': '!! Near-dead -- low variance',
+    'continuity':           '!! Near-dead -- low variance',
+    'contrast_uniformity':  '!! Near-dead -- low variance',
+}
+
+for i, name in enumerate(CNAMES):
+    c = concepts[:, i]
+    rho, _ = spearmanr(c, qs)
+    status = STATUS.get(name, 'OK')
+    lines.append(f'  {name:<28} {c.mean():>6.3f} {c.std():>6.3f} {rho:>+7.3f}  {status}')
+
+lines += [
+    '',
+    '  Vong lap nguy hiem (self-reinforcing collapse):',
+    '    ScoreAggregator chon noise_level (std=0.340, cao nhat)',
+    '    -> gradient tap trung update noise pathway',
+    '    -> cac concept khac it duoc update -> variance thap hon',
+    '    -> cang bi bo qua -> variance cang thap (vong lap)',
+    '',
+    '  He qua: Q ≈ f(noise_level) -- model thuc chat la "noise detector",',
+    '  khong phai "quality estimator" da khai niem.',
+    '',
+    '',
+    '4. KET QUA v24',
+    SEP2,
+]
+
+ks      = t2['mean_ks_across_sensors']
+pearson = t2['cross_sensor_pearson']
+lines += [
+    '  Track 2 -- Sensor Invariance:',
+    f'    Mean KS (cross-sensor) = {ks:.4f}   [goal <= 0.30]  {"PASS" if ks <= 0.30 else "FAIL"}',
+    f'    Cross-sensor Pearson   = {pearson:.4f}   [goal >= 0.20]  {"PASS" if pearson >= 0.20 else "FAIL"}',
+    '',
+    '    Y nghia: cung mot ngon tay chup bang cac sensor khac nhau -> SIFQ cho',
+    '    score nhat quan. KS thap = phan bo giong nhau, Pearson cao = ranking on dinh.',
+    '    NFIQ2 thuong co KS > 0.5 cho cross-sensor pairs.',
+    '',
+    '  Track 4 -- Concept Grounding (rho < 0 = PASS):',
+    f'    {"Degradation":<12} {"Concept":<28} {"rho":>7}  Status',
+    f'    {"------------":<12} {"----------------------------":<28} {"-------":>7}  ------',
+]
+
+pass_count = total = 0
+for row in t4:
+    deg = row['degradation']
+    for concept in TARGETS.get(deg, []):
+        rho = row.get(concept)
+        if rho is None:
+            continue
+        total += 1
+        ok = rho < 0
+        if ok:
+            pass_count += 1
+        status = 'PASS' if ok else 'FAIL <--'
+        lines.append(f'    {deg:<12} {concept:<28} {rho:>+7.3f}  {status}')
+
+lines += [
+    '',
+    f'    Result: {pass_count}/{total} pairs PASS',
+    '',
+    '  3 failures phan tich:',
+    '',
+    '  [1] noise -> noise_level (rho=+0.365):',
+    '      Regression tu v22 (-0.052, OK) -> v24 (+0.365, FAIL).',
+    '      gamma=2.0 + T39 (orient_coh added to dry_skin/wet_press) thay doi',
+    '      gradient landscape cua ScoreAggregator. Noise la degradation duy nhat',
+    '      target noise_level -> single-point pressure -> de bi invert.',
+    '',
+    '  [2] dry_skin -> contrast_u (rho=+0.051):',
+    '      Persistent qua cac phien ban (v22: +0.504). Dry_skin la degradation',
+    '      DUY NHAT target contrast_uniformity -> signal yeu, de bi gradient',
+    '      tu cac loss khac at di.',
+    '',
+    '  [3] dry_skin -> orient_coh (rho=+0.008):',
+    '      T39 moi them orient_coh vao dry_skin, nhung wet_press->orient_coh',
+    '      hoat dong tot (-0.448). Ly do: wet_press + blur deu co orientation',
+    '      disruption manh, nhung dry_skin tao ra orientation noise khong nhat',
+    '      quan -> khong du signal de orient_coh feature phan biet.',
+    '',
+    '',
+    '5. v25 -- T40 FIXES (dang chay)',
+    SEP2,
+    '  Thay doi so voi v24:',
+    '',
+    '  gamma: 2.0 -> 1.5',
+    '',
+    '  blur:  [1, 2]    -> [1, 2, 0]  (them orient_coh[0])',
+    '  noise: [3]       -> [3, 4]     (them contrast_u[4])',
+    '  jpeg:  [2, 1]    -> [2, 1, 4]  (them contrast_u[4])',
+    '  dry_skin / wet_press / occlusion: khong doi',
+    '',
+    '  Ket qua ky vong:',
+    '  - noise->noise_lv:    +0.365 -> negative  (gamma thap + noise 2-concept)',
+    '  - dry_skin->contrast_u: +0.051 -> negative  (3 degs co-supervise contrast_u)',
+    '  - dry_skin->orient_coh: +0.008 -> negative  (blur gives orient_coh strong signal)',
+    '  - Cac pairs da PASS: giu nguyen (blur/jpeg/occlusion/wet_press)',
+    '  - Track 2: giu PASS (KS, Pearson khong thay doi co ban)',
+    '',
+    SEP,
+    'END',
+    SEP,
+]
+
+txt = '\n'.join(lines)
+with open('eval_results/sifq_report_v24.txt', 'w', encoding='utf-8') as f:
+    f.write(txt)
+print(txt)
+print()
+print('>>> Saved: eval_results/sifq_report_v24.txt')

scripts/gen_nfiq2_proxy_scores.py

@@ -0,0 +1,162 @@
+"""Generate NFIQ2-proxy quality scores using image heuristics.
+
+NFIQ2 is not available as a pip package and requires building from source.
+This script approximates it using:
+  - Gabor filter energy at fingerprint ridge frequencies (8–16 px/cycle)
+  - Block-wise local contrast (RMS of intensity)
+  - Laplacian variance (sharpness)
+
+The three signals are fused and rescaled to [0, 100], matching NFIQ2 convention
+(higher = better quality).  Scores are written as JSONL with the same schema
+expected by run_eval.py:
+  {"image_path": "...", "q_score": 73.2}
+
+Usage (reads image paths from an existing scores file, e.g. sifq_scores.jsonl):
+  python sifq/scripts/gen_nfiq2_proxy_scores.py \
+      --sifq-scores sifq/eval_results/sifq_scores.jsonl \
+      --output      /tmp/nfiq2_scores.jsonl
+
+Or scan a directory directly:
+  python sifq/scripts/gen_nfiq2_proxy_scores.py \
+      --image-dir dataset/302b/images/baseline \
+      --output    /tmp/nfiq2_scores.jsonl
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+
+# ---------------------------------------------------------------------------
+# Quality features
+# ---------------------------------------------------------------------------
+
+def _gabor_energy(gray: np.ndarray, image_size: int = 224) -> float:
+    """Mean Gabor filter energy at typical fingerprint ridge frequencies."""
+    img = cv2.resize(gray, (image_size, image_size)).astype(np.float32) / 255.0
+    total = 0.0
+    count = 0
+    for theta_deg in range(0, 180, 30):
+        theta = np.deg2rad(theta_deg)
+        # wavelength in pixels: 8–16 px for 500 dpi prints at 224px crops
+        for wavelength in (8, 12, 16):
+            sigma = wavelength * 0.56
+            kern = cv2.getGaborKernel(
+                (31, 31), sigma, theta, wavelength, gamma=0.5, psi=0,
+                ktype=cv2.CV_32F,
+            )
+            filtered = cv2.filter2D(img, cv2.CV_32F, kern)
+            total += float(np.mean(filtered ** 2))
+            count += 1
+    return total / count if count else 0.0
+
+
+def _local_contrast(gray: np.ndarray, block: int = 16) -> float:
+    """Mean RMS contrast over non-overlapping blocks."""
+    img = cv2.resize(gray, (224, 224)).astype(np.float32)
+    h, w = img.shape
+    vals = []
+    for r in range(0, h - block + 1, block):
+        for c in range(0, w - block + 1, block):
+            patch = img[r : r + block, c : c + block]
+            vals.append(float(np.std(patch)))
+    return float(np.mean(vals)) if vals else 0.0
+
+
+def _laplacian_var(gray: np.ndarray) -> float:
+    """Variance of Laplacian — sharpness metric."""
+    img = cv2.resize(gray, (224, 224))
+    lap = cv2.Laplacian(img, cv2.CV_64F)
+    return float(np.var(lap))
+
+
+def _quality_score(image_path: str) -> float | None:
+    img = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
+    if img is None:
+        return None
+    g_energy = _gabor_energy(img)
+    contrast  = _local_contrast(img)
+    lap_var   = _laplacian_var(img)
+    # Combine: each signal contributes roughly equally after normalisation.
+    # Raw typical ranges (empirical, 500-dpi fingerprints at 224 px):
+    #   g_energy  ~[0.0001, 0.006]
+    #   contrast  ~[5, 60]
+    #   lap_var   ~[50, 3000]
+    score = (
+        np.clip(g_energy / 0.006, 0.0, 1.0) * 0.40 +
+        np.clip(contrast  / 60.0,  0.0, 1.0) * 0.35 +
+        np.clip(lap_var   / 3000.0, 0.0, 1.0) * 0.25
+    )
+    return round(float(score) * 100.0, 2)
+
+
+# ---------------------------------------------------------------------------
+# CLI
+# ---------------------------------------------------------------------------
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description="Generate NFIQ2-proxy quality scores")
+    src = p.add_mutually_exclusive_group(required=True)
+    src.add_argument(
+        "--sifq-scores", type=str,
+        help="Path to SIFQ scores JSONL; image paths are read from 'image_path' field",
+    )
+    src.add_argument(
+        "--image-dir", type=str,
+        help="Root directory to scan recursively for .png/.bmp/.wsq images",
+    )
+    p.add_argument("--output", type=str, default="/tmp/nfiq2_scores.jsonl")
+    p.add_argument(
+        "--extensions", type=str, default="png,bmp,wsq,jpg,jpeg",
+        help="Comma-separated file extensions to scan (only with --image-dir)",
+    )
+    return p.parse_args()
+
+
+def collect_paths(args: argparse.Namespace) -> list[str]:
+    if args.sifq_scores:
+        paths: list[str] = []
+        with open(args.sifq_scores, encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if line:
+                    paths.append(json.loads(line)["image_path"])
+        return paths
+    # --image-dir
+    exts = {f".{e.lstrip('.')}" for e in args.extensions.split(",")}
+    return [
+        str(p) for p in Path(args.image_dir).rglob("*")
+        if p.suffix.lower() in exts
+    ]
+
+
+def main() -> None:
+    args = parse_args()
+    paths = collect_paths(args)
+    print(f"Scoring {len(paths)} images → {args.output}")
+
+    out = Path(args.output)
+    out.parent.mkdir(parents=True, exist_ok=True)
+
+    skipped = 0
+    with out.open("w", encoding="utf-8") as fout:
+        for i, p in enumerate(paths):
+            score = _quality_score(p)
+            if score is None:
+                skipped += 1
+                continue
+            fout.write(json.dumps({"image_path": p, "q_score": score}) + "\n")
+            if (i + 1) % 1000 == 0:
+                print(f"  {i + 1}/{len(paths)}  (skipped={skipped})")
+
+    print(f"Done. Written {len(paths) - skipped} records, skipped {skipped}.")
+    print(f"Output: {out}")
+
+
+if __name__ == "__main__":
+    main()

scripts/run_eval_v16.sh

@@ -0,0 +1,53 @@
+#!/usr/bin/env bash
+# Eval v16: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+VENV="$REPO_ROOT/.venv"
+VERSION="v16"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+source "$VENV/bin/activate"
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores  "$SCORES" \
+    --output  "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8
+
+echo "[eval_${VERSION}] Milestone image: $OUTDIR/milestone_samples.png"
+
+echo ""
+echo "=== SUMMARY ==="
+cat "$OUTDIR/eval_summary.json"

scripts/run_eval_v17.sh

@@ -0,0 +1,57 @@
+#!/usr/bin/env bash
+# Eval v17: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+VENV="/home/aiserver/miniconda3"
+VERSION="v17"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+export PATH="$VENV/bin:$PATH"
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"
+
+echo "[eval_${VERSION}] Milestone image: $OUTDIR/milestone_samples.png"
+
+echo ""
+echo "=== SUMMARY ==="
+cat "$OUTDIR/eval_summary.json"

scripts/run_eval_v18.sh

@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+# Eval v18: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v18"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"
+
+echo "[eval_${VERSION}] Milestone image: $OUTDIR/milestone_samples.png"
+
+echo ""
+echo "=== SUMMARY ==="
+cat "$OUTDIR/eval_summary.json"

scripts/run_eval_v19.sh

@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+# Eval v19: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v19"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"
+
+echo "[eval_${VERSION}] Milestone image: $OUTDIR/milestone_samples.png"
+
+echo ""
+echo "=== SUMMARY ==="
+cat "$OUTDIR/eval_summary.json"

scripts/run_eval_v20.sh

@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+# Eval v20: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v20"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"
+
+echo "[eval_${VERSION}] Milestone image: $OUTDIR/milestone_samples.png"
+
+echo ""
+echo "=== SUMMARY ==="
+cat "$OUTDIR/eval_summary.json"

scripts/run_eval_v21.sh

@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+# Eval v21: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v21"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"
+
+echo "[eval_${VERSION}] Milestone image: $OUTDIR/milestone_samples.png"
+
+echo ""
+echo "=== SUMMARY ==="
+cat "$OUTDIR/eval_summary.json"

scripts/run_eval_v23.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Eval v23: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v23"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"

scripts/run_eval_v26.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Eval v26: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v26"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"

scripts/run_eval_v27.sh

@@ -0,0 +1,59 @@
+#!/usr/bin/env bash
+# Eval v27: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v27"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+echo "[eval_${VERSION}] Running inference on $CKPT ..."
+python sifq/scripts/run_infer.py \
+    --checkpoint "$CKPT" \
+    --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+    --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+    --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+    --exclude-sensor "$EXCLUDE" \
+    --output "$SCORES"
+    
+# ── Inference ──────────────────────────────────────────────────────────────
+# if [[ ! -f "$SCORES" ]]; then
+#     echo "[eval_${VERSION}] Running inference on $CKPT ..."
+#     python sifq/scripts/run_infer.py \
+#         --checkpoint "$CKPT" \
+#         --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+#         --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+#         --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+#         --exclude-sensor "$EXCLUDE" \
+#         --output "$SCORES"
+#     echo "[eval_${VERSION}] Inference done: $SCORES"
+# else
+#     echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+# fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"

scripts/run_eval_v28.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Eval v28: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v28"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"

scripts/run_eval_v29.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Eval v29: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v29"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"

scripts/run_eval_v31.sh

@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# Eval v31: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v31"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"
+
+echo "[eval_${VERSION}] All done."

scripts/run_eval_v32.sh

@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# Eval v32: inference + Track 2 + Track 4 + visual score milestones.
+
+set -euo pipefail
+
+REPO_ROOT="/home/aiserver/works/fingerprint"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+VERSION="v32"
+CKPT="$REPO_ROOT/sifq/checkpoints/${VERSION}/last.pt"
+SCORES="$REPO_ROOT/sifq/eval_results/sifq_scores_${VERSION}.jsonl"
+OUTDIR="$REPO_ROOT/sifq/eval_results/${VERSION}"
+EXCLUDE="R_1000_slap,R_500_slap,S_500_slap"
+
+cd "$REPO_ROOT"
+
+mkdir -p "$OUTDIR"
+
+# ── Inference ──────────────────────────────────────────────────────────────
+if [[ ! -f "$SCORES" ]]; then
+    echo "[eval_${VERSION}] Running inference on $CKPT ..."
+    python sifq/scripts/run_infer.py \
+        --checkpoint "$CKPT" \
+        --root-302a  "$REPO_ROOT/dataset/302a/images/challengers" \
+        --root-302b  "$REPO_ROOT/dataset/302b/images/baseline" \
+        --root-302d  "$REPO_ROOT/dataset/nist_302d/images/auxiliary" \
+        --exclude-sensor "$EXCLUDE" \
+        --output "$SCORES"
+    echo "[eval_${VERSION}] Inference done: $SCORES"
+else
+    echo "[eval_${VERSION}] Scores file already exists: $SCORES (skipping inference)"
+fi
+
+# ── Evaluation (Track 2 + Track 4) ────────────────────────────────────────
+echo "[eval_${VERSION}] Running evaluation (Track 2 + Track 4, exclude: $EXCLUDE)..."
+python sifq/scripts/run_eval.py \
+    --sifq-scores "$SCORES" \
+    --checkpoint  "$CKPT" \
+    --out-dir     "$OUTDIR" \
+    --exclude-sensor "$EXCLUDE" \
+    --skip-track1
+
+echo "[eval_${VERSION}] Eval done. Results in $OUTDIR"
+
+# ── Visual score milestones ────────────────────────────────────────────────
+echo "[eval_${VERSION}] Generating score milestone samples..."
+python sifq/scripts/visualize_score_milestones.py \
+    --scores   "$SCORES" \
+    --output   "$OUTDIR/milestone_samples.png" \
+    --n_buckets 8 \
+    --version  "$VERSION"
+
+echo "[eval_${VERSION}] All done."

scripts/run_infer.py

@@ -0,0 +1,158 @@
+"""Inference script: run trained SIFQ on a dataset and save per-image scores.
+
+Output JSON format (one line per image):
+  {"image_path": ..., "q_score": 73.2, "concepts": [0.8, 0.6, ...],
+   "identity_id": "00002401", "finger_id": "F07", "sensor_id": "U_500_roll"}
+
+Usage:
+  python scripts/run_infer.py --checkpoint checkpoints/v16/last.pt \\
+      --root-302b dataset/302b/images/baseline \\
+      --output /tmp/sifq_scores.jsonl
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+import torch
+
+ROOT = Path(__file__).resolve().parents[1]
+SRC_ROOT = ROOT / "src"
+if str(SRC_ROOT) not in sys.path:
+    sys.path.insert(0, str(SRC_ROOT))
+
+from data.nist302_loader import NIST302Loader, NIST302Paths
+from models.aggregator import ScoreAggregator
+from models.backbone import SIFQBackbone
+from models.concept_head import ConceptHead, SpatialConceptHead
+from models.sensor_discriminator import SensorDiscriminator
+from models.sifq import SIFQ
+
+
+def load_model(checkpoint_path: str, device: torch.device, num_sensors: int = 10) -> SIFQ:
+    ckpt = torch.load(checkpoint_path, map_location=device, weights_only=False)
+    # Infer num_sensors from checkpoint metrics ("n_sensors" key saved by train_sifq.py)
+    num_sensors = ckpt.get("metrics", {}).get("n_sensors", num_sensors)
+
+    backbone = SIFQBackbone(model_name="tiny_vit_5m_224.dist_in22k", pretrained=False)
+    # Detect architecture from saved config — spatial head was introduced in v27
+    _use_spatial = ckpt.get("config", {}).get("spatial_concept_head", False)
+    if _use_spatial:
+        concept_head = SpatialConceptHead(in_dim=backbone.feature_dim)
+    else:
+        concept_head = ConceptHead(in_dim=backbone.feature_dim)
+    aggregator = ScoreAggregator(k=6)
+    sensor_disc = SensorDiscriminator(in_dim=backbone.feature_dim, num_sensors=num_sensors)
+    model = SIFQ(backbone, concept_head, aggregator, sensor_disc)
+    model.load_state_dict(ckpt["model"], strict=True)
+    model.to(device).eval()
+    return model
+
+
+def parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description="SIFQ inference — generate quality scores")
+    p.add_argument("--checkpoint", type=str, required=True,
+                   help="Path to trained SIFQ checkpoint (last.pt or best.pt)")
+    p.add_argument("--root-302a", type=str, default="",
+                   help="Root for NIST SD302-A challengers (optional)")
+    p.add_argument("--root-302b", type=str,
+                   default="/home/aiserver/works/fingerprint/dataset/302b/images/baseline",
+                   help="Root for NIST SD302-B baseline")
+    p.add_argument("--root-302d", type=str,
+                   default="/home/aiserver/works/fingerprint/dataset/nist_302d/images/auxiliary",
+                   help="Root for NIST SD302-D auxiliary")
+    p.add_argument("--image-size", type=int, default=224)
+    p.add_argument("--batch-size", type=int, default=32)
+    p.add_argument("--num-workers", type=int, default=2)
+    p.add_argument("--output", type=str, default="/tmp/sifq_scores.jsonl",
+                   help="Output JSONL file path")
+    p.add_argument("--max-samples", type=int, default=-1,
+                   help="Cap number of images for quick eval; -1 means all")
+    p.add_argument("--exclude-sensor", type=str, default="",
+                   help="Comma-separated sensor_ids to skip inference on. "
+                        "E.g. 'R_1000_slap,R_500_slap,S_500_slap'")
+    return p.parse_args()
+
+
+@torch.no_grad()
+def main() -> None:
+    args = parse_args()
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    # --- Load model ---
+    print(f"Loading checkpoint: {args.checkpoint}")
+    model = load_model(args.checkpoint, device)
+    print(f"Model loaded. Device: {device}")
+
+    # --- Discover records ---
+    paths = NIST302Paths(
+        root_302a=args.root_302a or "",
+        root_302b=args.root_302b,
+        root_302d=args.root_302d,
+    )
+    loader = NIST302Loader(image_size=args.image_size)
+    records = loader.discover(paths)
+    if args.exclude_sensor:
+        excluded = {s.strip() for s in args.exclude_sensor.split(",") if s.strip()}
+        records = [r for r in records if r["sensor_id"] not in excluded]
+        print(f"After excluding sensors {excluded}: {len(records)} records remain")
+    if args.max_samples > 0:
+        records = records[:args.max_samples]
+    print(f"Discovered {len(records)} records")
+
+    # --- Batch inference ---
+    output_path = Path(args.output)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    batch_records: list[dict] = []
+    batch_tensors: list[torch.Tensor] = []
+
+    def flush_batch() -> None:
+        if not batch_tensors:
+            return
+        images = torch.stack(batch_tensors, dim=0).to(device)
+        outputs = model(images)
+        scores = outputs["score"].squeeze(-1).cpu().tolist()
+        concepts_batch = outputs["concepts"].cpu().tolist()
+        for rec, q, conc in zip(batch_records, scores, concepts_batch):
+            row = {
+                "image_path": rec["image_path"],
+                "identity_id": rec["identity_id"],
+                "finger_id": rec["finger_id"],
+                "sensor_id": rec["sensor_id"],
+                "dataset": rec["dataset"],
+                "q_score": round(float(q), 3),
+                "concepts": [round(float(c), 4) for c in conc],
+            }
+            with open(output_path, "a", encoding="utf-8") as f:
+                f.write(json.dumps(row) + "\n")
+        batch_records.clear()
+        batch_tensors.clear()
+
+    # Clear output file
+    output_path.write_text("")
+
+    n_done = 0
+    for sample in loader.iter_samples(records):
+        batch_records.append({
+            "image_path": sample["image_path"],
+            "identity_id": sample["identity_id"],
+            "finger_id": sample["finger_id"],
+            "sensor_id": sample["sensor_id"],
+            "dataset": sample["dataset"],
+        })
+        batch_tensors.append(sample["image"])
+        if len(batch_tensors) >= args.batch_size:
+            flush_batch()
+            n_done += args.batch_size
+            if n_done % 500 == 0:
+                print(f"  {n_done}/{len(records)} done")
+
+    flush_batch()
+    print(f"Done. Scores saved to: {output_path}")
+
+
+if __name__ == "__main__":
+    main()

scripts/run_smoke.sh

@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+# Smoke test: 1 epoch, 50 samples, no multiprocessing
+# Verifies the full pipeline runs end-to-end in ~1-2 minutes
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
+VENV="$ROOT/../.venv"
+
+[[ -f "$VENV/bin/activate" ]] && source "$VENV/bin/activate"
+
+python "$SCRIPT_DIR/train_sifq.py" \
+  --epochs 1 \
+  --batch-size 8 \
+  --max-train-samples 80 \
+  --num-workers 0 \
+  --save-dir "$ROOT/checkpoints/smoke" \
+  "$@"

scripts/run_train_v16.sh

@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+# v16 training — Fix score collapse / no discrimination từ v15.
+#
+# Root cause analysis (v15 failure):
+#
+#   Tất cả ảnh được score gần như giống nhau (range 24–56, thay vì 0–100).
+#   Model không phân biệt được ảnh tốt/xấu.
+#
+#   Nguyên nhân chính: stats=None trong mọi lần gọi loss_mat().
+#
+#   Khi stats=None:
+#     q_mat = raw cosine ≈ 0.85 cho mọi ảnh (MDGT quality-robust)
+#     → model tối ưu pred_score → (0.85/2 + 0.5)×100 ≈ 92.5 cho tất cả
+#     → L_spread kéo về [10,90], L_mat kéo về 92.5
+#     → equilibrium tại q_mean ≈ 52 không thay đổi suốt 60 epoch
+#
+# v16 fixes:
+#   T31a — Per-identity cosine stats (compute_identity_cos_stats):
+#           Sau khi tính prototype, chạy second pass để tính (mean_cos, std_cos)
+#           per identity. Pass stats vào loss_mat() → target được normalize:
+#             q_mat = tanh((cos - mu_i) / sigma_i)
+#           Ảnh tốt hơn trung bình của identity → target > 0 → score > 50.
+#           Ảnh kém hơn → target < 0 → score < 50.
+#           L_mat và L_spread giờ CÙNG chiều thay vì conflict.
+#
+#   T31b — tanh scaling trong _compute_q_mat:
+#           tanh maps z ∈ (-∞,+∞) về (-1,1), tương thích với
+#           pred_normalized = (score/100 - 0.5) × 2 ∈ [-1,1].
+#           Không cần clamp thủ công, không có extreme target.
+#
+#   T31c — Tăng W_SPREAD từ 2.0 → 4.0:
+#           Với teacher target đã zero-centred, tăng spread weight
+#           giúp model học spread nhanh hơn mà không conflict.
+#
+#   T31d — deg-every-n-steps 1 (thay vì 2):
+#           L_deg mỗi step = 2× signal so với v15. FVC-only ranking.
+#
+# Train từ đầu, 60 epochs, LR=1e-4 cosine.
+
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+source "${REPO_ROOT}/.venv/bin/activate"
+
+VERSION="v16"
+SAVE_DIR="${REPO_ROOT}/sifq/checkpoints/${VERSION}"
+LOG_FILE="${REPO_ROOT}/sifq/logs/train_${VERSION}.log"
+EVAL_SCRIPT="${REPO_ROOT}/sifq/scripts/run_eval_${VERSION}.sh"
+
+python "${REPO_ROOT}/sifq/scripts/train_sifq.py" \
+  --root-302a  "${REPO_ROOT}/dataset/302a/images/challengers" \
+  --root-302b  "${REPO_ROOT}/dataset/302b/images/baseline" \
+  --root-302d  "${REPO_ROOT}/dataset/nist_302d/images/auxiliary" \
+  --root-fvc2002 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2002" \
+  --root-fvc2004 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2004" \
+  --root-polyu "${REPO_ROOT}/dataset/PolyU" \
+  --exclude-sensor "R_1000_slap,R_500_slap,S_500_slap" \
+  --mdgt-checkpoint "${REPO_ROOT}/pad/TRAM-downstream/checkpoint/checkpoints_dinov2_tram/best_eer.pt" \
+  --epochs 60 \
+  --batch-size 128 \
+  --image-size 224 \
+  --lr 1e-4 \
+  --spread-mode uniform \
+  --spread-weight 4.0 \
+  --concept-deg-gamma 2.0 \
+  --sd302-concept-weight 1.0 \
+  --deg-every-n-steps 2 \
+  --min-sigma 0.02 \
+  --max-train-samples -1 \
+  --num-workers 8 \
+  --gpus 0 \
+  --save-dir "${SAVE_DIR}"
+
+echo "[auto-eval] Training done. Starting eval ${VERSION}..."
+bash "${EVAL_SCRIPT}" >> "${LOG_FILE}" 2>&1

scripts/run_train_v17.sh

@@ -0,0 +1,87 @@
+#!/usr/bin/env bash
+# v17 training — T32: fix score collapse from v16 (L_mat vs L_pair conflict).
+#
+# Root cause analysis (v16 failure):
+#
+#   Score collapse nghiêm trọng hơn v15: q_std=0.33 ở inference (range 46.9–50.5).
+#   Training q_std=15.6 bề ngoài ổn nhưng KHÔNG phản ánh thực tế:
+#     - q_std=15.6 được tạo bởi FVC images trong training batch (có quality variation)
+#     - SD302 images trong training cũng bị collapse về ~50 nhưng bị che bởi FVC
+#     - Inference chỉ chạy trên SD302 → q_std=0.33 (hoàn toàn sụp đổ)
+#
+#   Root cause (T32): Per-identity stats (T31) tạo L_mat targets conflict trực tiếp
+#   với L_pair:
+#     - L_mat (T31b): q_mat = tanh((cos−μᵢ)/σᵢ) → muốn score khác nhau cho ảnh
+#       khác nhau của cùng identity (within-identity variation)
+#     - L_pair: |Q(s1)−Q(s2)| ≤ 0.05 → ép tất cả ảnh cùng identity có score
+#       giống nhau (within-identity uniformity)
+#     - Hai loss xung đột → model giải quyết bằng cách output ~50 cho tất cả
+#       SD302 images. Training q_std=15.6 driven by FVC (no L_pair conflict).
+#     - Inference (SD302 only): q_std=0.33, range 46.9–50.5.
+#     - KS=0.527 (tệ hơn v14=0.263): residual sensor bias giữa roll/flat sensors
+#       trong range tiny (0.5pt) với intra-sensor std ~0.1 → CDF shapes khác nhau.
+#
+# v17 fix (T32): FVC-only per-identity stats trong L_mat
+#
+#   T32a — compute_identity_cos_stats(fvc_only=True):
+#           Chỉ tính (mean_cos, std_cos) cho FVC identities.
+#           FVC không có cross-sensor pairs → L_pair=0 → không conflict với L_mat.
+#           FVC q_mat = tanh((cos−μ_fvc)/σ_fvc) → proper quality ordering ✅
+#
+#   T32b — _compute_q_mat fallback to raw cosine:
+#           SD302 identities KHÔNG có trong stats → fallback về raw cosine (~0.85).
+#           Đây là v14 behaviour (stats=None) đã cho KS=0.263 ✅.
+#           Không còn conflict với L_pair → SD302 scores có thể spread.
+#
+#   Giữ nguyên từ v16:
+#     - W_SPREAD=4.0 (T31c)
+#     - concept_deg_gamma=2.0 (T30a)
+#     - sd302_concept_weight=1.0 (T30b)
+#     - deg-every-n-steps=2
+#
+#   Expected improvements vs v16:
+#     - Score range: 0–100 (thay vì 46.9–50.5)
+#     - KS: ≤0.27 (target v14 = 0.263)
+#     - Pearson: ≥0.25 (target v14 = 0.294)
+#     - FVC quality ordering: tốt hơn v14 (per-identity tanh targets)
+#
+# Train từ đầu, 60 epochs, LR=1e-4 cosine.
+
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+source "${REPO_ROOT}/sifq/.venv/bin/activate"
+
+VERSION="v17"
+SAVE_DIR="${REPO_ROOT}/sifq/checkpoints/${VERSION}"
+LOG_FILE="${REPO_ROOT}/sifq/logs/train_${VERSION}.log"
+EVAL_SCRIPT="${REPO_ROOT}/sifq/scripts/run_eval_${VERSION}.sh"
+
+mkdir -p "${REPO_ROOT}/sifq/logs"
+
+python "${REPO_ROOT}/sifq/scripts/train_sifq.py" \
+  --root-302a  "${REPO_ROOT}/dataset/302a/images/challengers" \
+  --root-302b  "${REPO_ROOT}/dataset/302b/images/baseline" \
+  --root-302d  "${REPO_ROOT}/dataset/nist_302d/images/auxiliary" \
+  --root-fvc2002 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2002" \
+  --root-fvc2004 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2004" \
+  --root-polyu "${REPO_ROOT}/dataset/PolyU" \
+  --exclude-sensor "R_1000_slap,R_500_slap,S_500_slap" \
+  --mdgt-checkpoint "${REPO_ROOT}/pad/TRAM-downstream/checkpoint/checkpoints_dinov2_tram/best_eer.pt" \
+  --epochs 60 \
+  --batch-size 128 \
+  --image-size 224 \
+  --lr 1e-4 \
+  --spread-mode uniform \
+  --spread-weight 4.0 \
+  --concept-deg-gamma 2.0 \
+  --sd302-concept-weight 1.0 \
+  --deg-every-n-steps 2 \
+  --min-sigma 0.02 \
+  --max-train-samples -1 \
+  --num-workers 8 \
+  --gpus 0 \
+  --save-dir "${SAVE_DIR}"
+
+echo "[auto-eval] Training done. Starting eval ${VERSION}..."
+bash "${EVAL_SCRIPT}" >> "${LOG_FILE}" 2>&1

scripts/run_train_v18.sh

@@ -0,0 +1,85 @@
+#!/usr/bin/env bash
+# v18 training — T33: revert to v14 loss design + include SD302-A.
+#
+# Root cause analysis (v17 failure):
+#
+#   Score collapse persists despite T32 fix: all SD302 images → ~52.7, q_std=0.60.
+#   Training q_std=16.59 looks healthy but driven by FVC only (same pattern as v15/v16).
+#
+#   Deeper root cause (T33):
+#     T32 removed explicit L_mat vs L_pair conflict (SD302 gets raw cosine fallback),
+#     but FVC still gets tanh-normalised targets (high variation) while SD302 gets
+#     constant raw cosine targets. Model learns to distinguish "FVC mode" (variable)
+#     from "SD302 mode" (constant) — dataset shortcut. At inference (SD302 only): collapse.
+#
+#     Additional factors in v15–v17 that didn't exist in v14 (which worked):
+#       - W_SPREAD 2.0 → 4.0 (stronger spread pressure amplifies FVC/SD302 asymmetry)
+#       - concept_deg_gamma 0.5/1.0 → 2.0 (stronger concept loss may destabilise SD302 features)
+#       - sd302_concept_weight 0 → 1.0 (T30b concept deg on SD302 adds noise to SD302 features)
+#       - deg-every-n-steps 4 → 2 (more frequent deg steps = more FVC-specific gradients)
+#
+# v18 fix (T33): revert all v15–v17 additions, add SD302-A as new data
+#
+#   T33a — --no-mat-stats: Disable per-identity cosine stats entirely.
+#           ALL identities (FVC + SD302) use raw cosine as L_mat target (v14 behaviour).
+#           Eliminates FVC/SD302 quality signal asymmetry. Model must treat both datasets
+#           similarly — SD302 cannot shortcut to constant output.
+#
+#   T33b — W_SPREAD = 2.0: Revert from 4.0. Lower spread pressure → backbone retains
+#           more discriminative features for SD302. v14 used ~2.0.
+#
+#   T33c — concept_deg_gamma = 0.5: Revert from 2.0 to v14 default.
+#
+#   T33d — sd302_concept_weight = 0.0: Disable SD302 concept degradation (default, v14).
+#
+#   T33e — deg-every-n-steps = 4: Revert to original (v14). Fewer FVC-only deg steps.
+#
+#   NEW vs v14 — SD302-A: Include dataset/302a/images/challengers (13,630 images, sensors A-H).
+#           More cross-sensor pairs for L_sens. More GRL training signal. Better coverage.
+#
+# Expected improvements vs v14:
+#   - Score spread: q_std > 10 at inference (like v14 ~15)
+#   - KS: ~0.26 (v14 baseline), potentially better with 8 new SD302-A sensors
+#   - Pearson: ~0.29 (v14 baseline)
+#   - More sensor-type coverage (8 SD302-A sensors vs 4 SD302-B + 4 SD302-D in v14)
+#
+# Train from scratch, 60 epochs, LR=1e-4 cosine.
+
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+
+VERSION="v18"
+SAVE_DIR="${REPO_ROOT}/sifq/checkpoints/${VERSION}"
+LOG_FILE="${REPO_ROOT}/sifq/logs/train_${VERSION}.log"
+EVAL_SCRIPT="${REPO_ROOT}/sifq/scripts/run_eval_${VERSION}.sh"
+
+mkdir -p "${REPO_ROOT}/sifq/logs"
+
+python "${REPO_ROOT}/sifq/scripts/train_sifq.py" \
+  --root-302a  "${REPO_ROOT}/dataset/302a/images/challengers" \
+  --root-302b  "${REPO_ROOT}/dataset/302b/images/baseline" \
+  --root-302d  "${REPO_ROOT}/dataset/nist_302d/images/auxiliary" \
+  --root-fvc2002 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2002" \
+  --root-fvc2004 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2004" \
+  --root-polyu "${REPO_ROOT}/dataset/PolyU" \
+  --exclude-sensor "R_1000_slap,R_500_slap,S_500_slap" \
+  --mdgt-checkpoint "${REPO_ROOT}/pad/TRAM-downstream/checkpoint/checkpoints_dinov2_tram/best_eer.pt" \
+  --epochs 60 \
+  --batch-size 128 \
+  --image-size 224 \
+  --lr 1e-4 \
+  --spread-mode uniform \
+  --spread-weight 2.0 \
+  --concept-deg-gamma 0.5 \
+  --sd302-concept-weight 0.0 \
+  --deg-every-n-steps 4 \
+  --no-mat-stats \
+  --max-train-samples -1 \
+  --num-workers 8 \
+  --gpus 0 \
+  --save-dir "${SAVE_DIR}"
+
+echo "[auto-eval] Training done. Starting eval ${VERSION}..."
+bash "${EVAL_SCRIPT}" >> "${LOG_FILE}" 2>&1

scripts/run_train_v19.sh

@@ -0,0 +1,79 @@
+#!/usr/bin/env bash
+# v19 training — T34: true revert to v14 hyperparameters (fix v18 misconfiguration).
+#
+# Root cause analysis (v18 failure):
+#
+#   v18 claimed to "revert to v14 design" but used WRONG hyperparameters:
+#     v18: --spread-weight 2.0  (v14 actual: 3.0  ← wrong)
+#     v18: --deg-every-n-steps 4  (v14 actual: 2  ← wrong)
+#   v18 comments were incorrect — v14 actually used spread-weight=3.0 and
+#   deg-every-n-steps=2 (confirmed via diff of original run_train_v14.sh).
+#
+#   These weaker settings were the root cause of v18 score collapse (q_std≈0.01):
+#   - L_spread at 2.0 was insufficient to force per-image quality differentiation
+#   - L_deg at every 4 steps provided half the degradation grounding vs v14
+#   - Combined result: backbone learned no quality-discriminative features → inference collapse
+#
+# v19 fix (T34): use v14 ACTUAL hyperparameters
+#
+#   T34a — --spread-weight 3.0: Restore v14 value (v18 used 2.0, docs erroneously said "v14=~2.0").
+#           Stronger L_spread gradient forces backbone to differentiate quality across images.
+#
+#   T34b — --deg-every-n-steps 2: Restore v14 value (v18 used 4).
+#           2× more frequent L_deg provides stronger ordinal quality signal.
+#
+#   KEPT from v18:
+#   - --no-mat-stats: Required to reproduce v14 pre-T31 behavior with current code.
+#     v14 predated per-identity cosine stats (T31). Current code default tries to compute
+#     stats; --no-mat-stats disables this, giving same raw cosine L_mat as original v14.
+#   - --sd302-concept-weight 0.0: disable SD302 concept-only L_deg (v14 default)
+#   - --concept-deg-gamma 0.5: v14 default
+#   - SD302-A included (302a challengers): added in v18, kept here for better sensor coverage
+#   - --batch-size 128 on --gpus 0 (v14 used 96×2 GPU, effective per-GPU is similar)
+#
+# Expected improvements vs v18:
+#   - Score spread: q_std > 10 at inference (like v14 ~15.5)
+#   - Track 2 mean_KS: ~0.26 (v14 baseline, not false positive like v18)
+#   - Track 2 Pearson: ~0.29 (v14 baseline)
+#   - Score range: 10–90 (vs v18: 54.62–54.65 collapsed)
+#
+# Train from scratch, 60 epochs, LR=1e-4 cosine.
+
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+
+VERSION="v19"
+SAVE_DIR="${REPO_ROOT}/sifq/checkpoints/${VERSION}"
+LOG_FILE="${REPO_ROOT}/sifq/logs/train_${VERSION}.log"
+EVAL_SCRIPT="${REPO_ROOT}/sifq/scripts/run_eval_${VERSION}.sh"
+
+mkdir -p "${REPO_ROOT}/sifq/logs"
+
+python "${REPO_ROOT}/sifq/scripts/train_sifq.py" \
+  --root-302a  "${REPO_ROOT}/dataset/302a/images/challengers" \
+  --root-302b  "${REPO_ROOT}/dataset/302b/images/baseline" \
+  --root-302d  "${REPO_ROOT}/dataset/nist_302d/images/auxiliary" \
+  --root-fvc2002 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2002" \
+  --root-fvc2004 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2004" \
+  --root-polyu "${REPO_ROOT}/dataset/PolyU" \
+  --exclude-sensor "R_1000_slap,R_500_slap,S_500_slap" \
+  --mdgt-checkpoint "${REPO_ROOT}/pad/TRAM-downstream/checkpoint/checkpoints_dinov2_tram/best_eer.pt" \
+  --epochs 60 \
+  --batch-size 128 \
+  --image-size 224 \
+  --lr 1e-4 \
+  --spread-mode uniform \
+  --spread-weight 3.0 \
+  --concept-deg-gamma 0.5 \
+  --sd302-concept-weight 0.0 \
+  --deg-every-n-steps 2 \
+  --no-mat-stats \
+  --max-train-samples -1 \
+  --num-workers 8 \
+  --gpus 0 \
+  --save-dir "${SAVE_DIR}"
+
+echo "[auto-eval] Training done. Starting eval ${VERSION}..."
+bash "${EVAL_SCRIPT}" >> "${LOG_FILE}" 2>&1

scripts/run_train_v20.sh

@@ -0,0 +1,93 @@
+#!/usr/bin/env bash
+# v20 training — T35: full prototypes + L_deg on SD302.
+#
+# Root cause analysis (v19 failure — same collapse as v15–v18):
+#
+#   v19 was supposed to reproduce v14 exactly but STILL collapsed (q_std≈0.01
+#   at inference, all scores 53.1x). This means the "v14 hyperparameter diff"
+#   explanation (spread-weight and deg-every-n-steps) was wrong. Two deeper root
+#   causes identified by analysing every code change since v14:
+#
+#   Root cause 1 — Truncated prototypes (T35a):
+#     v14's code had NO --proto-max-batches cap; prototypes were computed on the
+#     full dataset (42,683 images, ~334 batches). When this parameter was added
+#     (default 150 batches = ~19,200 images, 45% of data), SD302 identities went
+#     from having full 19-sensor multi-sensor prototypes to 2–3 sensor partial
+#     prototypes.
+#       Full prototype  → raw cosine varies with image quality across all sensors
+#                        (quality-discriminative, sensor-invariant).
+#       Partial proto   → raw cosine correlated with which sensors are in the
+#                        prototype window (sensor-biased, quality-agnostic) →
+#                        no per-image quality gradient for SD302 → collapse.
+#
+#   Root cause 2 — FVC-only L_deg (T27, unchanged since v13):
+#     T27 reverted T17 (L_deg on all images) because clean SD302 images anchored
+#     at ~28 when there was no per-dataset spread. T27 itself introduced
+#     per-dataset L_spread_ds. Now that L_spread_ds forces SD302 to span [10,90],
+#     re-enabling full L_rank for SD302 is safe: L_spread_ds prevents the
+#     anchoring, and L_rank orders SD302 images by their synthetic degradation
+#     response (a proxy for ridge clarity / quality). This gives the ONLY stable
+#     per-image quality signal for SD302 at inference.
+#
+# v20 fixes (T35):
+#   T35a — --proto-max-batches 0: Compute prototypes on the full training set.
+#           Full multi-sensor prototypes restore per-image L_mat quality signal
+#           for SD302 (raw cosine varies with quality, not sensor).
+#           Cost: ~10–15 extra minutes at epoch 0 for prototype computation.
+#
+#   T35b — --deg-include-sd302: Apply full L_deg (L_rank + L_concept) to SD302
+#           images in addition to FVC. L_spread_ds (present since v13) prevents
+#           score anchoring at ~28. Provides robust per-image ordinal quality
+#           grounding for SD302 independent of root cause 1.
+#
+# All other settings kept from v19 (which correctly reproduced v14 hyperparams):
+#   --spread-weight 3.0  --deg-every-n-steps 2  --no-mat-stats
+#   --concept-deg-gamma 0.5  --sd302-concept-weight 0.0
+#   --batch-size 128  --gpus 0  --epochs 60
+#
+# Expected improvements vs v19:
+#   q_std (inference)  : >12  (v19: ~0.01)
+#   Score range        : 10–90 (v19: 53.10–53.16)
+#   Track 2 mean_KS    : ≤0.27 real  (v19: 0.3706 with collapsed distributions)
+#   Track 2 Pearson    : ≥0.25  (v19: -0.0123)
+
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+
+VERSION="v20"
+SAVE_DIR="${REPO_ROOT}/sifq/checkpoints/${VERSION}"
+LOG_FILE="${REPO_ROOT}/sifq/logs/train_${VERSION}.log"
+EVAL_SCRIPT="${REPO_ROOT}/sifq/scripts/run_eval_${VERSION}.sh"
+
+mkdir -p "${REPO_ROOT}/sifq/logs"
+
+python "${REPO_ROOT}/sifq/scripts/train_sifq.py" \
+  --root-302a  "${REPO_ROOT}/dataset/302a/images/challengers" \
+  --root-302b  "${REPO_ROOT}/dataset/302b/images/baseline" \
+  --root-302d  "${REPO_ROOT}/dataset/nist_302d/images/auxiliary" \
+  --root-fvc2002 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2002" \
+  --root-fvc2004 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2004" \
+  --root-polyu "${REPO_ROOT}/dataset/PolyU" \
+  --exclude-sensor "R_1000_slap,R_500_slap,S_500_slap" \
+  --mdgt-checkpoint "${REPO_ROOT}/pad/TRAM-downstream/checkpoint/checkpoints_dinov2_tram/best_eer.pt" \
+  --epochs 60 \
+  --batch-size 128 \
+  --image-size 224 \
+  --lr 1e-4 \
+  --spread-mode uniform \
+  --spread-weight 3.0 \
+  --concept-deg-gamma 0.5 \
+  --sd302-concept-weight 0.0 \
+  --deg-every-n-steps 2 \
+  --deg-include-sd302 \
+  --no-mat-stats \
+  --proto-max-batches 0 \
+  --max-train-samples -1 \
+  --num-workers 8 \
+  --gpus 0 \
+  --save-dir "${SAVE_DIR}"
+
+echo "[auto-eval] Training done. Starting eval ${VERSION}..."
+bash "${EVAL_SCRIPT}" >> "${LOG_FILE}" 2>&1

scripts/run_train_v21.sh

@@ -0,0 +1,93 @@
+#!/usr/bin/env bash
+# v21 training — T36: full prototypes + FVC-only L_deg (remove --deg-include-sd302).
+#
+# Root cause analysis (v20 failure — same collapse as v15–v19):
+#
+#   v20 training showed healthy q_std≈18.3 across all 60 epochs but INFERENCE
+#   collapsed (all SD302 sensors score ~53.38, q_std≈0.007, concepts stuck at
+#   [0.994, 0.008, ...]).  The T35 diagnosis was right about truncated
+#   prototypes (root cause 1) but the second fix (--deg-include-sd302) was
+#   WRONG and caused the collapse:
+#
+#   Root cause — L_rank on SD302 creates a stable attractor at ~53.38 (T35b was wrong):
+#     SD302 images are ALL high quality (controlled NIST acquisition protocol).
+#     With --deg-include-sd302, L_rank says:
+#         Q(clean_SD302) > Q(low_deg_SD302) + m
+#         Q(low_deg_SD302) > Q(high_deg_SD302) + m
+#     The model satisfies this constraint by assigning a single "clean score" (~53.38)
+#     to ALL clean SD302 images and lower scores to degraded variants.  There is
+#     NO gradient pushing different clean SD302 images apart — L_rank only requires
+#     clean > degraded, not that clean images differ from each other.  Combined with
+#     L_spread_ds (batch-level, arbitrary ordering) and L_pair (same-identity
+#     same-score pressure), the model collapses all clean SD302 to ~53.38 at inference.
+#
+#   Why v14 worked without --deg-include-sd302:
+#     FVC has GENUINE quality variation (8 impressions per subject with naturally
+#     different quality — blur, dry, wet, pressure).  MDGT raw cosine to prototype
+#     GENUINELY varies (~0.75 worst to ~0.93 best) for FVC subjects.  L_mat gradient
+#     on FVC shapes the backbone to be quality-discriminative.  At inference on SD302,
+#     the backbone's learned quality features transfer (ridge clarity, noise level,
+#     ridge continuity) → SD302 images scored by intrinsic quality, not batch rank.
+#     This transfer mechanism is BLOCKED when L_rank on SD302 creates the ~53.38
+#     attractor that overrides the learned quality representation.
+#
+# v21 fix (T36):
+#   T36 — Remove --deg-include-sd302. Revert to FVC-only L_deg (T27 original design).
+#   Keep --proto-max-batches 0 (T35a, the correct fix from v20).
+#   L_deg applied only to FVC images: genuine quality variation → genuine L_mat
+#   gradient → backbone learns quality features that transfer to SD302 at inference.
+#   L_spread_ds still applied to SD302 to force batch-level spread (prevents batch-
+#   collapse without creating the attractor problem).
+#
+# All other settings kept from v20:
+#   --no-mat-stats  --spread-weight 3.0  --deg-every-n-steps 2
+#   --concept-deg-gamma 0.5  --sd302-concept-weight 0.0
+#   --batch-size 128  --gpus 0  --epochs 60
+#   SD302-A+B+D included (full 30K sensor-invariant images, 19 sensors)
+#
+# Expected improvements vs v20:
+#   q_std (inference)  : >12   (v20: ~0.007)
+#   Score range        : 10–90 (v20: 53.20–53.41)
+#   Track 2 mean_KS    : ≤0.30 (v20: 0.4936 — driven by sensor clusters, not quality)
+#   Track 2 Pearson    : ≥0.20 (v20: 0.0048)
+#   Concept grounding  : blur→clarity<0, noise→noise_level<0, etc.
+
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+export PATH="/home/aiserver/miniconda3/bin:$PATH"
+
+VERSION="v21"
+SAVE_DIR="${REPO_ROOT}/sifq/checkpoints/${VERSION}"
+LOG_FILE="${REPO_ROOT}/sifq/logs/train_${VERSION}.log"
+EVAL_SCRIPT="${REPO_ROOT}/sifq/scripts/run_eval_${VERSION}.sh"
+
+mkdir -p "${REPO_ROOT}/sifq/logs"
+
+python "${REPO_ROOT}/sifq/scripts/train_sifq.py" \
+  --root-302a  "${REPO_ROOT}/dataset/302a/images/challengers" \
+  --root-302b  "${REPO_ROOT}/dataset/302b/images/baseline" \
+  --root-302d  "${REPO_ROOT}/dataset/nist_302d/images/auxiliary" \
+  --root-fvc2002 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2002" \
+  --root-fvc2004 "${REPO_ROOT}/dataset/FVC_Dataset/FVC2004" \
+  --root-polyu "${REPO_ROOT}/dataset/PolyU" \
+  --exclude-sensor "R_1000_slap,R_500_slap,S_500_slap" \
+  --mdgt-checkpoint "${REPO_ROOT}/pad/TRAM-downstream/checkpoint/checkpoints_dinov2_tram/best_eer.pt" \
+  --epochs 60 \
+  --batch-size 128 \
+  --image-size 224 \
+  --lr 1e-4 \
+  --spread-mode uniform \
+  --spread-weight 3.0 \
+  --concept-deg-gamma 0.5 \
+  --sd302-concept-weight 0.0 \
+  --deg-every-n-steps 2 \
+  --no-mat-stats \
+  --proto-max-batches 0 \
+  --max-train-samples -1 \
+  --num-workers 8 \
+  --gpus 0,1 \
+  --save-dir "${SAVE_DIR}"
+
+echo "[auto-eval] Training done. Starting eval ${VERSION}..."
+bash "${EVAL_SCRIPT}" >> "${LOG_FILE}" 2>&1