#!/bin/bash
#SBATCH --job-name=sa_pretrained_baseline
#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=1:00:00
#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/sa_pretrained_baseline-%j.log
#
# Sanity check: is the pretrained sa_open_1_0_bg_expanded.ckpt (training step
# 0) better than ANY of our 37 trained ckpts? If yes, the bg2fg fine-tune is
# net negative on test_loss — a strong signal training is hurting, not helping.
#
# Same eval as the 4-shard sweep: deterministic per-batch t/noise, same test
# split, same loss formula. Only difference: pretrained has no EMA so we use
# raw weights (the eval script auto-detects this).
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 stable-audio

OUT=/nfs/turbo/coe-ahowens-nobackup/dingqy/sa_test_loss_pretrained.json
PRETRAINED=/nfs/turbo/coe-ahowens-nobackup/dingqy/friendly-stable-audio-tools/data/ckpt/sa_open_1_0_bg_expanded.ckpt

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

echo ""
echo "=== pretrained vs best trained ==="
python -c "
import json
pre = json.load(open('$OUT'))
pre_loss = list(pre.values())[0]['test_loss']
best_trained = 0.68191  # lbkb1h5z step=1000
print(f'  pretrained (sa_open_1_0_bg_expanded):  {pre_loss:.5f}')
print(f'  best trained (lbkb1h5z step=1000):     {best_trained:.5f}')
delta = pre_loss - best_trained
print(f'  delta (pretrained - best_trained):     {delta:+.5f}')
if delta > 0:
    print('  → pretrained is WORSE → training did help (at least for the first 1k steps)')
else:
    print('  → pretrained is BETTER → fine-tune is net negative; training is hurting')
"