#!/usr/bin/env bash
# Fetch the third-party code / weights that embed.py needs for RAW inputs (images, text, counts, mIF tiles).
# Not needed for the feature-level API (E.*_from_feats), evaluate_saved.py or training (52_final_model.py).
#
#   PYTHON=python bash setup_encoders.sh
#
# Code is cloned at the exact commits the features were extracted with; nothing is pip-installed. The other frozen
# encoders (KEEP, BiomedCLIP, Qwen3-Embedding-0.6B, Tahoe-x1, SE-600M, AIDO.Cell, VirTues weights) are downloaded
# from the Hugging Face Hub on first use at pinned revisions (see the 4x_ / 14_ scripts).
# UNI2-h is gated (MahmoodLab/UNI2-h, accept the licence on its HF page): either log in with an account that has
# access, or point $UNI2_PATH at a local pytorch_model.bin.
set -euo pipefail
cd "$(dirname "$0")"
PY=${PYTHON:-python}
mkdir -p ext/medvtok

clone() {  # url dir commit
  [ -d "$2/.git" ] || git clone -q "$1" "$2"
  git -C "$2" checkout -q "$3"
  echo "$2 @ $(git -C "$2" rev-parse --short HEAD)"
}
clone https://github.com/bunnelab/virtues.git ext/virtues 2c9c6cb            # VirTues model code (MIT)
clone https://github.com/genbio-ai/ModelGenerator.git ext/modelgenerator bda6f2d  # AIDO.Cell gene list
clone https://github.com/Masaaki-75/medvtok.git ext/medvtok/repo a558c24     # MedVTok model code (MIT)

# MedVTok weights (3.3 GB; MIT)
"$PY" -I -c "from huggingface_hub import snapshot_download as s; print(s('massaki75/medvtok', revision='9e70a000b1040ce498450a877da91a8fe02dac41', local_dir='ext/medvtok/weights', allow_patterns=['ckpt.pth', 'config.json']))"

# ext/marker_embeddings/ (ESM-2 embeddings of the 16 marker proteins, VirTues input) ships with this repo
test -d ext/marker_embeddings/esm2_t30_150M_UR50D && echo "marker embeddings ok"
"$PY" -I -c "import os, sys; sys.path.insert(0, os.getcwd()); import embed; p = os.environ.get('UNI2_PATH', embed.UNI2_LOCAL); print('UNI2-h:', p if os.path.exists(p) else 'will be downloaded from the gated HF repo MahmoodLab/UNI2-h')"