scripts/dedup/v2_write.py

"""v2 step 3: write the deduplicated, split release by copying stored rows byte for byte.

Every kept triplet (fold from <split_dir>/<subset>/{train,val,test}.parquet) is written to
    <out>/<subset>/<fold>/<subset>_<fold>_<NNNN>.aselmdb
in ascending ms_id order (R, S, P rows of a triplet adjacent), at most <per_shard> triplets per file. The files are
cut on the full split list, and then the BROKEN triplets are left out, so lemat_train_0213 and lemat_train_0479
hold per_shard - 1 triplets and the other files do not change. Rows are NOT
re-serialised: the zlib-compressed JSON value stored in the SciLM shard is copied unchanged under a new LMDB key
("1".."3n"; the id is not part of the stored value), then "nextid" = 3n+1 is written exactly as ase_db_backends does.
Each file is written as <name>.tmp and renamed only after its transaction committed.
  usage: v2_write.py <scilm_root> <scilm_index_dir> <split_dir> <out_root> [per_shard=50000] [workers=48] [subsets]
"""
import glob, os, sys, zlib, json
import numpy as np, pyarrow.parquet as pq
from ase.io.jsonio import encode
from multiprocessing import Pool

FOLDS = ["train", "val", "test"]
src_root, idx_dir, split_dir, out = sys.argv[1:5]
per = int(sys.argv[5]) if len(sys.argv) > 5 else 50000; nw = int(sys.argv[6]) if len(sys.argv) > 6 else 48
SUBS = sys.argv[7].split(",") if len(sys.argv) > 7 else ["lemat", "mp20bat", "oc20", "oc22"]
# saddle ms_ids of the 2 lemat triplets with broken forces (max |F| at the saddle 4.78e11 and 1.02e8 eV/A), found
# after the split was drawn. They stay in the split lists, so that the file cuts do not move, and are not written.
BROKEN = [35584423, 79800988]


def load_index(s):
    Z = [np.load(f) for f in sorted(glob.glob(f"{idx_dir}/{s}__*.npz"))]
    shards = [str(z["shard"]) for z in Z]
    ms = np.concatenate([z["ms"] for z in Z]); key = np.concatenate([z["key"] for z in Z])
    fi = np.concatenate([np.full(len(z["ms"]), i, np.int32) for i, z in enumerate(Z)])
    o = np.argsort(ms); ms, key, fi = ms[o], key[o], fi[o]
    assert (np.diff(ms) > 0).all(), f"{s}: repeated ms_id in the source"
    return ms, key, fi, shards


def plan():
    jobs = []
    for s in SUBS:
        ms, key, fi, shards = load_index(s)
        for f in FOLDS:
            fr = np.sort(np.asarray(pq.read_table(f"{split_dir}/{s}/{f}.parquet").column("ms_id")))
            sad = fr[fr % 3 == 1]
            assert np.array_equal(fr, np.sort(np.concatenate([sad - 1, sad, sad + 1]))), f"{s}/{f}: incomplete triplets"
            pos = np.searchsorted(ms, fr); assert (pos < len(ms)).all() and np.array_equal(ms[pos], fr), f"{s}/{f}: ms_id missing in source"
            n = (len(sad) + per - 1) // per
            for k in range(n):
                rows = pos[3 * k * per: 3 * (k + 1) * per]
                rows = rows[~np.isin(ms[rows] // 3, np.array(BROKEN) // 3)]      # leave out the 3 rows of a BROKEN triplet
                name = f"{s}/{f}/{s}_{f}_{k:04d}.aselmdb"
                jobs.append((name, [shards[i] for i in range(len(shards))], fi[rows], key[rows], ms[rows]))
    return jobs


def work(job):
    import lmdb
    name, shards, fi, key, ms = job
    dst = f"{out}/{name}"
    if os.path.exists(dst): return name, len(ms), "exists"
    os.makedirs(os.path.dirname(dst), exist_ok=True)
    vals = [None] * len(ms)
    for f in np.unique(fi):                                   # read grouped by source file, in key order
        env = lmdb.open(f"{src_root}/{shards[f]}", subdir=False, readonly=True, lock=False, readahead=False, meminit=False)
        sel = np.where(fi == f)[0]; sel = sel[np.argsort(key[sel])]
        with env.begin(buffers=False) as t:
            for i in sel:
                v = t.get(str(int(key[i])).encode("ascii")); assert v is not None, (name, shards[f], key[i]); vals[i] = v
        env.close()
    tmp = dst + ".tmp"
    for p in (tmp, tmp + "-lock"):
        if os.path.exists(p): os.remove(p)
    env = lmdb.open(tmp, subdir=False, map_size=2 ** 41, meminit=False, lock=False, sync=True)
    with env.begin(write=True) as t:
        for i, v in enumerate(vals):
            assert t.put(str(i + 1).encode("ascii"), v, overwrite=False)
        t.put(b"nextid", zlib.compress(bytes(encode(len(vals) + 1), "utf-8")), overwrite=False)
    env.sync(True); env.close()
    os.rename(tmp, dst)
    return name, len(ms), "written"


if __name__ == "__main__":
    jobs = plan(); jobs.sort(key=lambda j: -len(j[4]))
    print(f"{len(jobs)} output shards, {sum(len(j[4]) for j in jobs)} rows", flush=True)
    man = [(j[0], len(j[4]) // 3, int(j[4][0]), int(j[4][-1])) for j in sorted(jobs)]
    os.makedirs(out, exist_ok=True)
    if len(SUBS) == 4:
        with open(f"{out}/.manifest.tsv", "w") as fh:
            fh.write("file\ttriplets\tfirst_ms_id\tlast_ms_id\n")
            for m in man: fh.write("\t".join(map(str, m)) + "\n")
    with Pool(nw) as pool:
        for n, (name, r, st) in enumerate(pool.imap_unordered(work, jobs), 1):
            print(f"  {n}/{len(jobs)} {name}: {r} rows {st}", flush=True)