#!/usr/bin/env bash # Quick start: download the released CausalBind checkpoints and reproduce the # DUD-E and LIT-PCBA results of Table 1 on a single GPU. # # Usage: bash scripts/quick_start.sh [variant ...] (default: sp lr emb atom) # Requires the test benchmarks in test_datasets/ (see DATA.md) and the # pretrained backbones in pretrain/ (bash scripts/download_data.sh pretrain). set -euo pipefail REPO_DIR="$(cd "$(dirname "$0")/.." && pwd)" cd "${REPO_DIR}" variants=("$@") [ ${#variants[@]} -eq 0 ] && variants=(sp lr emb atom) # Hugging Face; the GitHub mirror is https://github.com/lokali/CausalBind/releases/download/v1.0 CKPT_URL="${CKPT_URL:-https://huggingface.co/LokaLi/CausalBind/resolve/main}" RESULTS="${RESULTS:-results/quick_start}" mkdir -p checkpoints "${RESULTS}" for v in "${variants[@]}"; do ckpt="checkpoints/causalbind_${v}.pt" if [ ! -s "${ckpt}" ]; then echo ">>> downloading ${ckpt}" wget -nv -O "${ckpt}" "${CKPT_URL}/causalbind_${v}.pt" fi mkdir -p "${RESULTS}/${v}" for task in DUDE PCBA; do echo ">>> evaluating CausalBind-${v} on ${task}" bash scripts/test.sh "${v}" "${task}" "${ckpt}" "${RESULTS}/${v}" > "${RESULTS}/${v}/${task}.log" 2>&1 \ || { echo "evaluation failed, see ${RESULTS}/${v}/${task}.log"; exit 1; } done done python tools/summarize_results.py "${RESULTS}"