"""Data augmentation for the training set: each training structure gets K extra text variants from a random site permutation plus a random rigid translation (physically the same crystal). Validation and test rows are left untouched.""" import os, sys, csv, signal, json, time os.environ.setdefault("CUDA_VISIBLE_DEVICES", ""); os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "3") for v in ("OMP_NUM_THREADS", "MKL_NUM_THREADS", "OPENBLAS_NUM_THREADS"): os.environ.setdefault(v, "1") import warnings; warnings.filterwarnings("ignore") from multiprocessing import Pool ROOT, BASE, OUT = sys.argv[1], sys.argv[2], sys.argv[3] class TO(Exception): pass def _alarm(*a): raise TO() def resolve(rel): d, f = rel.split("/", 1) for pre in ("P-", "N-", ""): p = os.path.join(ROOT, pre + d, f) if os.path.exists(p): return p return None NPROC = int(sys.argv[4]); K = int(sys.argv[5]) REPS = ["cif_p1", "cif_symmetrized", "slices", "zmatrix", "atom_sequences_plusplus"] def work(row): import numpy as np from pymatgen.core import Structure from xtal2txt.core import TextRep out = [] try: s0 = Structure.from_file(resolve(row["cif_path"])) except Exception: return out signal.signal(signal.SIGALRM, _alarm) rng = np.random.default_rng(abs(hash(row["id"])) % (2**32)) for k in range(1, K + 1): perm = rng.permutation(len(s0)); shift = rng.random(3) s = Structure(s0.lattice, [s0[i].species for i in perm], [(s0[i].frac_coords + shift) % 1.0 for i in perm]) r = dict(id=f"{row['id']}#a{k}", label=int(row["label"]), set="train", source_dir=row["source_dir"]) try: tr = TextRep.from_input(s) except Exception: continue for rep in REPS: signal.alarm(180 if rep == "slices" else 60) try: v = tr.get_requested_text_reps([rep])[rep]; r[rep] = v if isinstance(v, str) and v else None except Exception: r[rep] = None finally: signal.alarm(0) out.append(r) return out if __name__ == "__main__": import pandas as pd rows = [r for r in csv.DictReader(open(os.path.join(ROOT, "dataset_划分.csv"), encoding="utf-8-sig")) if r["set"] == "train"] t0 = time.time(); res = [] with Pool(NPROC, maxtasksperchild=200) as pool: for i, r in enumerate(pool.imap_unordered(work, rows, chunksize=4)): res.extend(r) if (i + 1) % 1000 == 0: print(f"{i+1}/{len(rows)} {time.time()-t0:.0f}s", flush=True) aug = pd.DataFrame(res); base = pd.read_parquet(BASE) df = pd.concat([base, aug], ignore_index=True); df.to_parquet(OUT) print("aug rows", len(aug), "total", len(df), f"{time.time()-t0:.0f}s") for rep in REPS: print(rep, "aug ok", int(aug[rep].notna().sum()), "changed vs original:", float((aug[rep].values != aug.id.str.split("#").str[0].map(base.set_index("id")[rep]).values).mean().round(3)))