code/infer_audioldm2_test_set.sbat

#!/bin/bash
#SBATCH --job-name=al2_test_set_infer
#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=4:00:00
#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_set_infer-%j.log
#
# AudioLDM2 ControlNet bulk inference on full 3919-pair test set.
# Default: best ckpt by test_loss = step=110000 (test_loss=0.08349). Override
# AL2_CKPT_PATH on sbatch CLI for A/B comparisons.
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

# step=110000 is on HF (sweeper has uploaded it).
: "${AL2_CKPT_PATH:=audioldm2_bg2fg_controlnet_rebalance/step_00110000}"
: "${OUT_DIR:=/nfs/turbo/coe-ahowens-nobackup/dingqy/inference_demo/audioldm2_test_set_step110k}"

echo "ckpt:  $AL2_CKPT_PATH"
echo "out:   $OUT_DIR"

python /nfs/turbo/coe-ahowens-nobackup/dingqy/infer_audioldm2_test_set.py \
    --ckpt "$AL2_CKPT_PATH" \
    --out "$OUT_DIR" \
    --steps 50 --batch-size 4

echo ""
echo "=== sample subfolder ==="
SAMPLE=$(find "$OUT_DIR" -name "mix.wav" 2>/dev/null | head -1 | xargs dirname 2>/dev/null)
echo "first subfolder: $SAMPLE"
[ -n "$SAMPLE" ] && ls -la "$SAMPLE"
echo ""
echo "=== count ==="
find "$OUT_DIR" -name "mix.wav" 2>/dev/null | wc -l