#!/bin/bash
#SBATCH --job-name=al2_test_loss_72k
#SBATCH --mail-type=FAIL,END
#SBATCH --mail-user=dingqy@umich.edu
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=4
#SBATCH --mem-per-cpu=11GB
#SBATCH --time=1:30:00
#SBATCH --account=ahowens1
#SBATCH --partition=spgpu
#SBATCH --qos=foe
#SBATCH --reservation=spgpu_foe_nodes
#SBATCH --gres=gpu:a40:1
#SBATCH --output=/nfs/turbo/coe-ahowens-nobackup/dingqy/al2_test_loss_72k-%j.log
#
# Sweep new ckpts from the 70k→120k resume run (49235422). Training is still
# in progress (~step 99k as of submission); these 14 ckpts are 72k-98k.
# All on local turbo only — HF uploader hasn't run since training started.
set -eu

source /nfs/turbo/coe-ahowens-nobackup/dingqy/miniforge3/etc/profile.d/conda.sh
source /home/dingqy/.hf_token
export HF_CACHE=/nfs/turbo/coe-ahowens-nobackup/dingqy/.cache/huggingface

conda activate new_audioldm

OUT=/nfs/turbo/coe-ahowens-nobackup/dingqy/al2_test_loss_shard_72k.json

LOCAL_BASE=/nfs/turbo/coe-ahowens-nobackup/dingqy/AudioLDM-training-finetuning/log/audioldm2_v2/audioldm2_bg2fg_controlnet_rebalance
CKPTS=""
for s in 72000 74000 76000 78000 80000 82000 84000 86000 88000 90000 92000 94000 96000 98000; do
    CKPTS="$CKPTS $LOCAL_BASE/step_$(printf %08d $s)"
done

echo "ckpts: $CKPTS"
echo "out:   $OUT"

python /nfs/turbo/coe-ahowens-nobackup/dingqy/eval_audioldm2_test_loss.py \
    --ckpts $CKPTS \
    --out "$OUT" \
    --batch-size 8 --num-workers 4

echo ""
echo "=== sorted by step ==="
python -c "
import json, re
r = json.load(open('$OUT'))
def step(k): return int(re.search(r'step_0*(\d+)', k).group(1))
for k in sorted(r.keys(), key=step):
    print(f'  step={step(k):>5}  loss={r[k][\"test_loss\"]:.5f}')
"