code/featurization/featurize_chemeleon.py

"""Extract CheMeleon embeddings (2048-dim frozen) for train + test.

Reads:  data/raw/pxr-challenge_TRAIN.csv, pxr-challenge_TEST_BLINDED.csv
Writes: data/processed/features/chemeleon_train.npy, chemeleon_test.npy
"""
from __future__ import annotations
import sys, time
from pathlib import Path
import numpy as np
import pandas as pd

EXP_ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(EXP_ROOT / "code"))
OUT_DIR = EXP_ROOT / "data" / "processed" / "features"
OUT_DIR.mkdir(parents=True, exist_ok=True)


def main():
    from chemeleon_fingerprint import CheMeleonFingerprint

    p_tr = OUT_DIR / "chemeleon_train.npy"
    p_te = OUT_DIR / "chemeleon_test.npy"
    if p_tr.exists() and p_te.exists():
        print(f"  cache hit: {p_tr.name}, {p_te.name}")
        a = np.load(p_tr); b = np.load(p_te)
        print(f"  shapes: train={a.shape}  test={b.shape}")
        return 0

    train = pd.read_csv(EXP_ROOT / "data" / "raw" / "pxr-challenge_TRAIN.csv")
    test = pd.read_csv(EXP_ROOT / "data" / "raw" / "pxr-challenge_TEST_BLINDED.csv")
    print(f"loading CheMeleon...")
    fp = CheMeleonFingerprint()
    print(f"computing train embeddings ({len(train)})...")
    t0 = time.time()
    # batch in chunks of 256 to avoid OOM
    BATCH = 256
    train_embs = []
    for i in range(0, len(train), BATCH):
        chunk = train["SMILES"].iloc[i:i+BATCH].tolist()
        train_embs.append(fp(chunk))
        print(f"  {i+len(chunk)}/{len(train)} ({(time.time()-t0):.1f}s)")
    train_emb = np.vstack(train_embs)
    np.save(p_tr, train_emb)

    print(f"computing test embeddings ({len(test)})...")
    test_embs = []
    for i in range(0, len(test), BATCH):
        chunk = test["SMILES"].iloc[i:i+BATCH].tolist()
        test_embs.append(fp(chunk))
    test_emb = np.vstack(test_embs)
    np.save(p_te, test_emb)
    print(f"saved train={train_emb.shape} test={test_emb.shape} in {time.time()-t0:.1f}s")
    return 0


if __name__ == "__main__":
    sys.exit(main())