File size: 2,394 Bytes
e0db531 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 | """Split the judged pairs into train / unseen-validation.
The paper holds out a 500-pair validation set and reports on 75 of them. We hold
out the same *proportion* and, critically, make the split **disjoint by video**:
no clip that appears in a validation pair (as reference or target) may appear
anywhere in training. A naive random split would leak, because the same clip is
reused across many pairs.
"""
from __future__ import annotations
import argparse
import json
import random
from pathlib import Path
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--pairs", default="data/pairs.jsonl")
ap.add_argument("--train_out", default="data/pairs_train.jsonl")
ap.add_argument("--val_out", default="data/pairs_val.jsonl")
ap.add_argument("--val_frac", type=float, default=0.15)
ap.add_argument("--seed", type=int, default=0)
args = ap.parse_args()
rows = [json.loads(l) for l in open(args.pairs)]
rng = random.Random(args.seed)
order = list(range(len(rows)))
rng.shuffle(order)
# Select validation *pairs* first, then quarantine every clip they touch.
# (Selecting held-out *videos* first and keeping only pairs fully inside that
# set is far too restrictive at this dataset size -- it yielded 1 val pair.)
n_val = max(1, int(len(rows) * args.val_frac))
val_idx = set(order[:n_val])
val = [rows[i] for i in sorted(val_idx)]
val_vids = {r["ref_id"] for r in val} | {r["tgt_id"] for r in val}
# Train pairs must not touch any held-out clip at all, or physics/appearance
# from a validation clip leaks into training.
train = [r for i, r in enumerate(rows)
if i not in val_idx
and r["ref_id"] not in val_vids and r["tgt_id"] not in val_vids]
vids = sorted({r["ref_id"] for r in rows} | {r["tgt_id"] for r in rows})
for path, data in ((args.train_out, train), (args.val_out, val)):
Path(path).parent.mkdir(parents=True, exist_ok=True)
with open(path, "w") as f:
for r in data:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
dropped = len(rows) - len(train) - len(val)
print(f"videos: {len(vids)} ({len(val_vids)} held out)")
print(f"train pairs: {len(train)}")
print(f"val pairs: {len(val)}")
print(f"dropped (cross-split, would leak): {dropped}")
if __name__ == "__main__":
main()
|