File size: 2,589 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 | #!/usr/bin/env bash
# v28 training — Matcher-free: L_mat disabled (--no-mat).
#
# Goal: Validate that L_sens + L_deg + L_ortho + L_spread are sufficient
# for sensor-invariant quality ordering without any external matcher teacher.
#
# Key change from v26:
# --no-mat : skip MDGT entirely (no loading, no emb_cache, no prototypes)
# l_mat = 0.0 for all steps/epochs; stage scheduler still runs
# normally for beta/gamma (sensor + degradation weights unaffected)
# --spread-weight 4.0 : slightly stronger spread to compensate for lost L_mat anchor
# NOTE: do NOT add --fixed-alpha 0.0 — that overrides beta/gamma too (-1.0 default),
#
# Paper argument if this works:
# "SIFQ quality emerges from degradation ordering and sensor consistency alone,
# without any external matcher supervision. This scorer-free quality generalises
# across matchers (Track 1 ERC) with no matcher-specific bias."
#
# Expected dynamics:
# S1 (ep 0-9): q_std grows from L_deg + L_spread (no L_mat anchor — may be slower)
# S2+ (ep 20+): L_sens stabilises sensor gap; L_deg maintains ordinal grounding
# Risk: without L_mat, collapse is possible if L_deg/L_spread insufficient.
# Watch q_std and l_spread in logs. If q_std < 5 at ep15, training failed.
set -euo pipefail
REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
export PATH="/home/aiserver/miniconda3/bin:$PATH"
VERSION="v28"
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" \
--no-mat \
--epochs 60 \
--batch-size 96 \
--image-size 224 \
--lr 1e-4 \
--spread-mode uniform \
--spread-weight 4.0 \
--concept-deg-gamma 2.0 \
--sd302-concept-weight 0.0 \
--deg-every-n-steps 2 \
--k-cross 0 \
--max-train-samples -1 \
--num-workers 8 \
--gpus 0 \
--save-dir "${SAVE_DIR}" \
2>&1 | tee "${LOG_FILE}"
echo "[auto-eval] Training done. Starting eval ${VERSION}..."
bash "${EVAL_SCRIPT}" >> "${LOG_FILE}" 2>&1
|