#!/bin/bash
# VLAlert-X v2 Phase 3 — 5-seed Danger Head training.
#
# Each seed: 50 epochs cosine LR + early-stop (patience 10) on val per_frame AUC.
# Expected per-seed wall time: ~1 GPU-hr  (small head on cached features)
# 5 seeds total: ~5 GPU-hr.
set -euo pipefail
cd "$(dirname "$0")/../.."

OUT_ROOT="checkpoints/danger_v2"
mkdir -p logs "$OUT_ROOT"

for seed in 0 1 2 3 4; do
  echo "================================================================"
  echo "Danger Head seed=${seed}"
  echo "================================================================"
  python -m training.Policy.train_danger_head \
      --out_dir "${OUT_ROOT}/seed${seed}" \
      --epochs 50 \
      --batch_size 128 \
      --lr 3e-4 \
      --weight_decay 1e-4 \
      --hidden 512 \
      --k_queries 4 \
      --dropout 0.2 \
      --w_clip 0.5 \
      --patience 10 \
      --seed "${seed}" 2>&1 | tee "logs/phase3_danger_seed${seed}.log"
done

echo ""
echo "==============================================================="
echo "5-seed summary (val per_frame AUC):"
for seed in 0 1 2 3 4; do
  if [[ -f "${OUT_ROOT}/seed${seed}/best.pt" ]]; then
    python -c "
import torch
d = torch.load('${OUT_ROOT}/seed${seed}/best.pt', weights_only=False, map_location='cpu')
m = d['val_metrics']
print(f\"  seed${seed}: per_frame_auc={m.get('per_frame_auc',0):.4f}  \" +
      f\"clip_auc={m.get('clip_auc',0):.4f}  ep={d['epoch']}\")
"
  fi
done
echo "==============================================================="