File size: 2,146 Bytes
b33e5eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env bash
# v11 training — PolyU back in L_sens (contact vs contactless pairs).
#
# Changes vs v10:
#   1. PolyU RE-INCLUDED in CrossSensorBatchSampler + build_pair_indices.
#      PolyU contact vs contactless = strongest cross-modal sensor signal → L_sens.
#   2. PolyU EXCLUDED from L_mat (no prototype / teacher loss for PolyU images).
#   3. PolyU EXCLUDED from L_deg (degradation ranking not applicable to PolyU).
#      → PolyU contributes ONLY to L_sens per original design.
#   4. New β/α schedule: β ramps FASTER than α so β/α → 1.0 at S2 entry (ep20).
#      α: 0.10→0.30  β: 0.00→0.30  γ: 1.00→0.60  over ep10-19.
#      S2 full: α=0.30 β=0.30 γ=0.60 (β/α=1.0).
#      S3:      α=0.25 β=0.35 γ=0.50 (β/α=1.4).
#
# Root cause of v9 failure: β/α=0.50 in S2 (α=0.40, β=0.20) → GRL too weak
#   to override L_mat's sensor-correlated gradients → L_sens stuck at 2-3 all S2/S3.
#
# Total dataset: SD302-A/B/D (30,567) + FVC2002+2004 (7,040) + PolyU (5,952)
#              = 43,559 images
#              L_sens anchors: SD302 (~2,276 groups) + PolyU (336 cross-modal groups)
#              L_mat:          SD302 + FVC only (PolyU masked)
#              L_deg:          FVC only (SD302 + PolyU masked)
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" \
  --epochs 80 \
  --batch-size 96 \
  --image-size 224 \
  --lr 1e-4 \
  --spread-mode uniform \
  --spread-weight 2.0 \
  --deg-every-n-steps 2 \
  --max-train-samples -1 \
  --num-workers 8 \
  --gpus 0,1 \
  --save-dir "${REPO_ROOT}/sifq/checkpoints_full_v11"