Add eval sample builder for valN
Browse files
benchmarks/edit/build_eval_samples.py
ADDED
|
@@ -0,0 +1,167 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Build val_N eval sample jsonl files from Ditto manifests.
|
| 3 |
+
|
| 4 |
+
The output schema matches the saved val20/val100 artifacts:
|
| 5 |
+
{"id", "target_video", "control_video", "prompt"}.
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
from __future__ import annotations
|
| 9 |
+
|
| 10 |
+
import argparse
|
| 11 |
+
import json
|
| 12 |
+
import random
|
| 13 |
+
import re
|
| 14 |
+
from collections import defaultdict
|
| 15 |
+
from pathlib import Path
|
| 16 |
+
from typing import Any
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
DEFAULT_MANIFESTS = (
|
| 20 |
+
"datas/ditto_face/manifest.json",
|
| 21 |
+
"datas/ditto_face2/manifest.json",
|
| 22 |
+
)
|
| 23 |
+
|
| 24 |
+
DEFAULT_BUCKET_PREFIXES = (
|
| 25 |
+
"global_freeform1",
|
| 26 |
+
"global_freeform1_filtered",
|
| 27 |
+
"global_freeform2",
|
| 28 |
+
"global_freeform2_filtered",
|
| 29 |
+
"global_freeform3",
|
| 30 |
+
"global_style1",
|
| 31 |
+
"global_style2",
|
| 32 |
+
)
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
def read_json(path: Path) -> list[dict[str, Any]]:
|
| 36 |
+
data = json.loads(path.read_text(encoding="utf-8"))
|
| 37 |
+
if not isinstance(data, list):
|
| 38 |
+
raise TypeError(f"{path} must contain a JSON list")
|
| 39 |
+
return data
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
def sanitize_flat_relpath(relpath: str) -> str:
|
| 43 |
+
parts = [re.sub(r"[^0-9A-Za-z._-]+", "_", piece) for piece in relpath.split("/")]
|
| 44 |
+
return "__".join(parts)
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
def bucket_root(bucket: str) -> str:
|
| 48 |
+
return bucket.split("/", 1)[0]
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
def category_from_relpath(relpath: str) -> str:
|
| 52 |
+
parts = Path(relpath).parts
|
| 53 |
+
return parts[1] if len(parts) >= 3 else ""
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
def video_path(dataset: str, side: str, relpath: str, materialized: str, path_style: str) -> str:
|
| 57 |
+
if path_style == "materialized" and materialized:
|
| 58 |
+
return materialized
|
| 59 |
+
return str(Path("datas") / dataset / side / sanitize_flat_relpath(relpath))
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
def row_to_sample(row: dict[str, Any], dataset: str, path_style: str) -> dict[str, str] | None:
|
| 63 |
+
prompt = str(row.get("prompt", "") or "").strip()
|
| 64 |
+
low_rel = str(row.get("low_rel", "") or row.get("low_video_relpath", "") or "")
|
| 65 |
+
high_rel = str(row.get("high_rel", "") or row.get("high_video_relpath", "") or "")
|
| 66 |
+
if not prompt or not low_rel or not high_rel:
|
| 67 |
+
return None
|
| 68 |
+
return {
|
| 69 |
+
"target_video": video_path(dataset, "high", high_rel, str(row.get("high_materialized", "") or ""), path_style),
|
| 70 |
+
"control_video": video_path(dataset, "low", low_rel, str(row.get("low_materialized", "") or ""), path_style),
|
| 71 |
+
"prompt": prompt,
|
| 72 |
+
"_dataset": dataset,
|
| 73 |
+
"_bucket": bucket_root(str(row.get("target_bucket", "") or high_rel.split("/", 1)[0])),
|
| 74 |
+
"_category": category_from_relpath(high_rel),
|
| 75 |
+
"_dedupe_key": f"{dataset}\n{low_rel}\n{prompt}",
|
| 76 |
+
}
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
def collect_candidates(
|
| 80 |
+
repo_root: Path,
|
| 81 |
+
manifests: list[Path],
|
| 82 |
+
bucket_prefixes: tuple[str, ...],
|
| 83 |
+
path_style: str,
|
| 84 |
+
) -> list[dict[str, str]]:
|
| 85 |
+
candidates = []
|
| 86 |
+
seen = set()
|
| 87 |
+
for manifest in manifests:
|
| 88 |
+
path = manifest if manifest.is_absolute() else repo_root / manifest
|
| 89 |
+
dataset = path.parent.name
|
| 90 |
+
for row in read_json(path):
|
| 91 |
+
bucket = bucket_root(str(row.get("target_bucket", "") or ""))
|
| 92 |
+
if bucket and bucket not in bucket_prefixes:
|
| 93 |
+
continue
|
| 94 |
+
sample = row_to_sample(row, dataset, path_style)
|
| 95 |
+
if sample is None:
|
| 96 |
+
continue
|
| 97 |
+
key = sample["_dedupe_key"]
|
| 98 |
+
if key in seen:
|
| 99 |
+
continue
|
| 100 |
+
seen.add(key)
|
| 101 |
+
candidates.append(sample)
|
| 102 |
+
return candidates
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
def stratified_sample(candidates: list[dict[str, str]], count: int, seed: int) -> list[dict[str, str]]:
|
| 106 |
+
rng = random.Random(seed)
|
| 107 |
+
buckets: dict[tuple[str, str, str], list[dict[str, str]]] = defaultdict(list)
|
| 108 |
+
for sample in candidates:
|
| 109 |
+
buckets[(sample["_dataset"], sample["_bucket"], sample["_category"])].append(sample)
|
| 110 |
+
for values in buckets.values():
|
| 111 |
+
rng.shuffle(values)
|
| 112 |
+
|
| 113 |
+
queues = list(buckets.values())
|
| 114 |
+
rng.shuffle(queues)
|
| 115 |
+
selected = []
|
| 116 |
+
while queues and len(selected) < count:
|
| 117 |
+
next_queues = []
|
| 118 |
+
for queue in queues:
|
| 119 |
+
if len(selected) >= count:
|
| 120 |
+
break
|
| 121 |
+
if queue:
|
| 122 |
+
selected.append(queue.pop())
|
| 123 |
+
if queue:
|
| 124 |
+
next_queues.append(queue)
|
| 125 |
+
queues = next_queues
|
| 126 |
+
rng.shuffle(queues)
|
| 127 |
+
if len(selected) < count:
|
| 128 |
+
raise RuntimeError(f"only selected {len(selected)} samples from {len(candidates)} candidates")
|
| 129 |
+
return selected[:count]
|
| 130 |
+
|
| 131 |
+
|
| 132 |
+
def write_samples(path: Path, samples: list[dict[str, str]]) -> None:
|
| 133 |
+
path.parent.mkdir(parents=True, exist_ok=True)
|
| 134 |
+
with path.open("w", encoding="utf-8") as handle:
|
| 135 |
+
for index, sample in enumerate(samples):
|
| 136 |
+
row = {
|
| 137 |
+
"id": f"val_{index:04d}",
|
| 138 |
+
"target_video": sample["target_video"],
|
| 139 |
+
"control_video": sample["control_video"],
|
| 140 |
+
"prompt": sample["prompt"],
|
| 141 |
+
}
|
| 142 |
+
handle.write(json.dumps(row, ensure_ascii=False) + "\n")
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
def main() -> None:
|
| 146 |
+
parser = argparse.ArgumentParser()
|
| 147 |
+
parser.add_argument("--repo-root", type=Path, default=Path.cwd())
|
| 148 |
+
parser.add_argument("--manifest", type=Path, action="append", default=[])
|
| 149 |
+
parser.add_argument("--output", type=Path, required=True)
|
| 150 |
+
parser.add_argument("--count", type=int, default=1000)
|
| 151 |
+
parser.add_argument("--seed", type=int, default=20260511)
|
| 152 |
+
parser.add_argument("--path-style", choices=("flat", "materialized"), default="flat")
|
| 153 |
+
parser.add_argument("--bucket-prefix", action="append", default=[])
|
| 154 |
+
args = parser.parse_args()
|
| 155 |
+
|
| 156 |
+
manifests = args.manifest or [Path(p) for p in DEFAULT_MANIFESTS]
|
| 157 |
+
bucket_prefixes = tuple(args.bucket_prefix or DEFAULT_BUCKET_PREFIXES)
|
| 158 |
+
candidates = collect_candidates(args.repo_root, manifests, bucket_prefixes, args.path_style)
|
| 159 |
+
selected = stratified_sample(candidates, args.count, args.seed)
|
| 160 |
+
write_samples(args.output, selected)
|
| 161 |
+
print(f"candidates={len(candidates)}")
|
| 162 |
+
print(f"selected={len(selected)} -> {args.output}")
|
| 163 |
+
print(f"path_style={args.path_style}, seed={args.seed}")
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
if __name__ == "__main__":
|
| 167 |
+
main()
|