Hugging Face's logo

Datasets:
anbinh39
/
UFR-Fing

Dataset card Files
xet
 Community
UFR-Fing
File size: 3,357 Bytes
cfc7a54
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
#!/usr/bin/env bash
# v22 training β€” First correct run: full prototypes + random batching + single GPU.
#
# Root cause analysis β€” why v15–v21 all failed:
#
#   BUG 1 β€” CrossSensorBatchSampler default k_cross=16:
#     k_cross=16 forces 16 guaranteed cross-sensor pairs per batch.
#     With full prototypes (nearly-constant L_mat targets for SD302), backbone
#     over-optimises sensor invariance β†’ loses quality discrimination β†’
#     Pearson collapses (~0.09 in v21). Fix: --k-cross 0 (random batching).
#
#   BUG 2 β€” proto-max-batches default=150:
#     Only 45% of dataset used for prototypes β†’ 2–3 partial sensor prototypes
#     per SD302 identity β†’ cosine target is sensor-biased, not quality-correlated
#     β†’ no per-image quality gradient β†’ score collapse in v15–v19.
#     Fix: --proto-max-batches 0 (full dataset).
#
#   BUG 3 β€” DataParallel (––gpus 0,1) 4.3Γ— slower for TinyViT-5M:
#     Fix: --gpus 0 (single GPU).
#
#   BUG 4 β€” Redundant teacher double-pass at startup:
#     Fix: build_prototypes_from_cache() β€” O(N) CPU, no second teacher forward.
#
cd /home/aiserver/works/fingerprint
nohup bash sifq/scripts/run_train_v22.sh > sifq/logs/train_v22.log 2>&1 &
echo "PID: $!"
tail -f sifq/logs/train_v22.log# v22 design goals (per research design Β§4–7):
#   Quality-discriminative  : q_std β‰₯ 18 (model ranks fingerprints by utility)
#   Sensor-invariant        : pair loss β†’ 0 (L_sens equilibrium), adv β†’ rand_ce
#   Concept-grounded        : all Track 4 target concepts negative Spearman ρ
#   Training speed          : ~220s/epoch (single GPU, no DataParallel overhead)
#
# Note on L_sens / pair loss:
#   v22 includes SD302-A (8 diverse roll sensors per finger) which creates genuine
#   cross-sensor quality variation in L_mat targets. L_pair starts high (~12 at
#   S2 onset) and decreases as L_sens equilibrium is found. This is expected
#   with a richer, more diverse dataset β€” it is NOT a training failure.

set -euo pipefail

REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
export PATH="/home/aiserver/miniconda3/bin:$PATH"

VERSION="v22"
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 96 \
  --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 \
  --k-cross 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