scripts/benchmark_quant.sh

#!/bin/bash
set -euo pipefail

ENCODER="${ENCODER:?must be set}"
SWEEP_NAME="${1:?usage: $0 <sweep-name>}"
REPO_URL="${REPO_URL:-https://github.com/vedatonuryilmaz/dev_genopix.git}"
REPO_BRANCH="${REPO_BRANCH:-ml-intern}"
DATA_DATASET="${DATA_DATASET:-vedatonuryilmaz/te-quant-benchmark-v1}"
TOOLS="${TOOLS:-snk_em,snk_ot,vis_em,rsem,salmon}"
SAMPLES_PER_SETTING="${SAMPLES_PER_SETTING:-4}"
DEVICE="${DEVICE:-cpu}"

git clone --branch "$REPO_BRANCH" --depth 1 "$REPO_URL" deepgenopix
cd deepgenopix
uv sync --extra hf --extra quant

OUTPUT_DIR="data/output/quant_benchmark/$SWEEP_NAME"
mkdir -p "$OUTPUT_DIR"

uv run python -c "
import json
from pathlib import Path

from deepgenopix.quant import SyntheticConfig, load_quant_encoder
from deepgenopix.quant.synthetic import run_calibration_benchmark

encoder = load_quant_encoder('$ENCODER', device='$DEVICE')
tools = tuple(item.strip() for item in '$TOOLS'.split(',') if item.strip())
df = run_calibration_benchmark(
    encoder,
    SyntheticConfig(),
    tools=tools,
    samples_per_setting=int('$SAMPLES_PER_SETTING'),
)
out = Path('$OUTPUT_DIR')
df.to_parquet(out / 'results.parquet', index=False)
df.to_csv(out / 'results.csv', index=False)
(out / 'summary.json').write_text(json.dumps({
    'encoder': '$ENCODER',
    'tools': tools,
    'rows': int(len(df)),
    'sweep_name': '$SWEEP_NAME',
}, indent=2))
print(json.dumps({'rows': int(len(df)), 'output': str(out)}))
"

uv run python scripts/hf_artifacts.py push-dataset-folder \
  --repo "$DATA_DATASET" \
  --source "$OUTPUT_DIR" \
  --path-in-repo "$SWEEP_NAME" \
  --message "quant benchmark $SWEEP_NAME"