training/Policy/train_m10_v3.sh

#!/usr/bin/env bash
# M10-v3: MultiQueryPolicyHead on Qwen3-VL-4B per-frame belief cache.
# Reuses the existing M10 trainer (warm_start_trainer_m10.py); only the
# belief cache path differs (Qwen3 instead of Qwen2.5).
#
# Output: checkpoints/Policy/m10_qwen3vl4b_seed{0..4}/best.pt
set -euo pipefail
cd "$(dirname "$0")/../.."

CACHE_DIR=data/belief_cache_perframe_qwen3vl4b
LABEL_DIR=data/policy_labels
OUT_BASE=checkpoints/Policy
SEEDS=(0 1 2 3 4)

[[ -f $CACHE_DIR/multisrc_train.pt ]] || { echo "MISSING $CACHE_DIR/multisrc_train.pt"; exit 1; }
[[ -f $CACHE_DIR/multisrc_val.pt   ]] || { echo "MISSING $CACHE_DIR/multisrc_val.pt"; exit 1; }

for SEED in "${SEEDS[@]}"; do
    EXP=m10_qwen3vl4b_seed${SEED}
    echo "=== [m10_v3] seed=${SEED}  → $OUT_BASE/$EXP ==="
    python -m training.Policy.warm_start_trainer_m10 \
        --label_dir          $LABEL_DIR \
        --train_cache_path   $CACHE_DIR/multisrc_train.pt \
        --val_cache_path     $CACHE_DIR/multisrc_val.pt \
        --output_dir         $OUT_BASE \
        --experiment_name    $EXP \
        --K 4 --d_out 512 --n_heads 4 \
        --focal_alpha 0.75 --focal_gamma 2.0 \
        --label_smoothing 0.05 \
        --ortho_lambda 0.01 \
        --cost_lambda 0.3 \
        --ordinal_lambda 0.2 --ordinal_margin 0.2 \
        --belief_noise_std 0.01 \
        --num_epochs 6 \
        --batch_size 256 \
        --learning_rate 2e-4 \
        --warmup_steps 200 \
        --val_every_n_steps 200 \
        --early_stop_patience 7 \
        --use_balanced_sampler \
        --seed ${SEED}
done

echo
echo "=== [m10_v3] all seeds done ==="
ls -d $OUT_BASE/m10_qwen3vl4b_seed*