| # v12 training — Fix score collapse + wet_press concept grounding. | |
| # | |
| # Changes vs v11: | |
| # T17 — L_deg applied to ALL datasets (SD302 + FVC + PolyU), not FVC-only. | |
| # Root cause of v11 score collapse: 95% of SD302 images stuck at ~58.3 | |
| # because they had no ordinal grounding from L_deg. Synthetic degradation | |
| # on SD302 is artificial but necessary to prevent collapse on majority data. | |
| # T18 — MorphologicalDilator fixed: cv2.erode instead of cv2.dilate. | |
| # NIST fingerprints have DARK ridges. Erode expands dark ridges (wet smear). | |
| # Previous dilate shrank ridges → paradoxically increased clarity score. | |
| # Fixes wet_press → clarity = +0.219 (wrong) → should be negative. | |
| # T19 — DegradationRankingLoss gamma: 0.5 → 1.0 (stronger concept supervision). | |
| # Occlusion block coverage: 30% → 40% at level 3. | |
| # Fixes jpeg/occlusion concept signal near zero (rho ≈ -0.05, -0.03 in v11). | |
| # | |
| # Training: resume from v11 last.pt (model weights only, not optimizer). | |
| # --spread-weight 5.0 (up from 2.0) to force wider score distribution. | |
| # --deg-every-n-steps 4 (up from 2) to keep compute balanced since L_deg | |
| # now processes full batch (96 images) instead of ~15 FVC images. | |
| # | |
| # Expected improvements: | |
| # Track 2 mean_KS: 0.5566 → < 0.15 (SD302 images should spread, less sensor clustering) | |
| # Track 4 wet_press → clarity: +0.219 (❌) → negative (✓) | |
| # Track 4 occlusion → minutiae: -0.032 (weak) → < -0.2 (strong) | |
| set -euo pipefail | |
| REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)" | |
| source "${REPO_ROOT}/.venv/bin/activate" | |
| 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 "${REPO_ROOT}/dataset/PolyU" \ | |
| --mdgt-checkpoint "${REPO_ROOT}/pad/TRAM-downstream/checkpoint/checkpoints_dinov2_tram/best_eer.pt" \ | |
| --resume "${REPO_ROOT}/sifq/checkpoints_full_v11/last.pt" \ | |
| --epochs 80 \ | |
| --batch-size 96 \ | |
| --image-size 224 \ | |
| --lr 1e-4 \ | |
| --spread-mode uniform \ | |
| --spread-weight 5.0 \ | |
| --deg-every-n-steps 4 \ | |
| --deg-max-images 32 \ | |
| --max-train-samples -1 \ | |
| --num-workers 8 \ | |
| --gpus 0,1 \ | |
| --save-dir "${REPO_ROOT}/sifq/checkpoints_full_v12" | |