scripts/eval_vbench8_extended_subset.py

#!/usr/bin/env python3
"""Run the standard VBench metrics for one generated strategy.

The official VBench metric modules are used unchanged.  Only the metadata
builder is replaced so that the evaluator reads the exact extended prompt used
for generation and the auditable mapping can use suite-indexed filenames.
"""

from __future__ import annotations

import argparse
import hashlib
import json
import os
import sys
import tempfile
from pathlib import Path
from typing import Any


def preparse_gpu() -> str:
    parser = argparse.ArgumentParser(add_help=False)
    parser.add_argument("--gpu", default="0")
    args, _ = parser.parse_known_args()
    os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
    os.environ.setdefault("MPLCONFIGDIR", tempfile.mkdtemp(prefix="vbench_mpl_"))
    return str(args.gpu)


PHYSICAL_GPU = preparse_gpu()

_vbench_cache = Path(
    os.environ.get("VBENCH_CACHE_DIR", "/data3/chenzhuo/.cache/vbench")
).expanduser()
os.environ.setdefault(
    "VBENCH_BERT_MODEL_DIR", str(_vbench_cache / "bert-base-uncased")
)

from vbench import VBench

REPO_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
    sys.path.insert(0, str(REPO_ROOT))

from scripts.vbench8_protocol import (
    DIMENSIONS,
    PROTOCOL_NAME,
    SUITE_COUNTS,
    aggregate_selected_score,
)


def sha256(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for block in iter(lambda: handle.read(1024 * 1024), b""):
            digest.update(block)
    return digest.hexdigest()


def read_mapping(path: Path) -> list[dict[str, Any]]:
    value = json.loads(path.read_text(encoding="utf-8"))
    if not isinstance(value, list) or len(value) != 251:
        raise ValueError(f"Expected a 251-row mapping list: {path}")
    counts = {suite: 0 for suite in SUITE_COUNTS}
    globals_seen: set[int] = set()
    suite_seen: dict[str, set[int]] = {suite: set() for suite in SUITE_COUNTS}
    for row in value:
        suite = str(row["prompt_suite"])
        global_index = int(row["global_index"])
        suite_index = int(row["suite_index"])
        if suite not in counts:
            raise ValueError(f"Unknown suite {suite}")
        if global_index in globals_seen:
            raise ValueError(f"Duplicate global index {global_index}")
        if suite_index in suite_seen[suite]:
            raise ValueError(f"Duplicate suite index {suite}/{suite_index}")
        globals_seen.add(global_index)
        suite_seen[suite].add(suite_index)
        counts[suite] += 1
    if counts != SUITE_COUNTS:
        raise ValueError(f"Unexpected suite counts: {counts}")
    return sorted(value, key=lambda row: int(row["global_index"]))


def build_metadata(
    *,
    mapping: list[dict[str, Any]],
    videos_root: Path,
    strategy: str,
    metadata_path: Path,
) -> None:
    records: list[dict[str, Any]] = []
    for row in mapping:
        suite = str(row["prompt_suite"])
        suite_index = int(row["suite_index"])
        video_path = (
            videos_root / strategy / suite / f"{suite_index:03d}.mp4"
        ).resolve()
        if not video_path.is_file():
            raise FileNotFoundError(video_path)
        official_dimensions = list(row["official_dimensions"])
        if not set(official_dimensions).issubset(set(DIMENSIONS)):
            raise ValueError(
                f"Unsupported dimensions at global index {row['global_index']}: "
                f"{official_dimensions}"
            )
        record: dict[str, Any] = {
            "prompt_en": str(row["extended_prompt"]),
            "dimension": official_dimensions,
            "video_list": [str(video_path)],
        }
        # The standard scene implementation in vbench==0.1.5 needs the
        # official scene keyword in auxiliary_info.  It is not a replacement
        # for prompt_en: overall_consistency still receives the extended text.
        if "auxiliary_info" in row:
            record["auxiliary_info"] = row["auxiliary_info"]
        records.append(record)
    if len(records) != 251:
        raise ValueError(f"Expected 251 metadata records, got {len(records)}")
    metadata_path.parent.mkdir(parents=True, exist_ok=True)
    metadata_path.write_text(
        json.dumps(records, ensure_ascii=False, indent=2) + "\n", encoding="utf-8"
    )


class FixedMetadataVBench(VBench):
    """Use prepared metadata while retaining VBench's official evaluator."""

    def __init__(self, *, device: str, metadata_path: Path, output_path: Path):
        super().__init__(
            device=device,
            full_info_dir=str(metadata_path),
            output_path=str(output_path),
        )
        self.prepared_metadata_path = metadata_path

    def build_full_info_json(self, *args: Any, **kwargs: Any) -> str:
        return str(self.prepared_metadata_path)


def parse_result(value: Any) -> tuple[float, Any]:
    if isinstance(value, (list, tuple)) and len(value) == 2:
        score, details = value
    else:
        score, details = value, None
    score = float(score)
    if not 0.0 <= score <= 1.0:
        raise ValueError(f"VBench raw score is outside [0, 1]: {score}")
    return score, details


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--gpu", default=PHYSICAL_GPU)
    parser.add_argument("--strategy", required=True)
    parser.add_argument("--mapping", type=Path, required=True)
    parser.add_argument("--videos-root", type=Path, required=True)
    parser.add_argument("--output-root", type=Path, required=True)
    parser.add_argument("--vbench-info", type=Path, required=True)
    parser.add_argument("--skip-existing", action="store_true")
    return parser.parse_args()


def main() -> None:
    args = parse_args()
    mapping_path = args.mapping.resolve()
    videos_root = args.videos_root.resolve()
    output_root = args.output_root.resolve()
    mapping = read_mapping(mapping_path)
    metadata_path = output_root / "vbench/metadata" / f"{args.strategy}_full_info.json"
    raw_dir = output_root / "vbench/raw_results" / args.strategy
    raw_name = f"vbench8_{args.strategy}"
    raw_result_path = raw_dir / f"{raw_name}_eval_results.json"
    score_path = output_root / "vbench/scores" / f"{args.strategy}.json"
    if args.skip_existing and raw_result_path.is_file() and score_path.is_file():
        print(f"[cached] strategy={args.strategy}", flush=True)
        return

    build_metadata(
        mapping=mapping,
        videos_root=videos_root,
        strategy=args.strategy,
        metadata_path=metadata_path,
    )
    raw_dir.mkdir(parents=True, exist_ok=True)
    print(
        f"[vbench] gpu={args.gpu} strategy={args.strategy} videos=251 "
        f"dimensions={','.join(DIMENSIONS)}",
        flush=True,
    )
    bench = FixedMetadataVBench(
        device="cuda", metadata_path=metadata_path, output_path=raw_dir
    )
    bench.evaluate(
        videos_path=str(videos_root / args.strategy),
        name=raw_name,
        dimension_list=list(DIMENSIONS),
        local=True,
        mode="vbench_standard",
    )
    if not raw_result_path.is_file():
        raise FileNotFoundError(raw_result_path)
    raw_json = json.loads(raw_result_path.read_text(encoding="utf-8"))
    raw_scores: dict[str, float] = {}
    details: dict[str, Any] = {}
    for dimension in DIMENSIONS:
        if dimension not in raw_json:
            raise KeyError(f"Missing {dimension} in {raw_result_path}")
        score, dimension_details = parse_result(raw_json[dimension])
        raw_scores[dimension] = score
        if dimension_details is not None:
            details[dimension] = dimension_details
    aggregate = aggregate_selected_score(raw_scores)
    result = {
        "protocol": PROTOCOL_NAME,
        "strategy": args.strategy,
        "physical_gpu": str(args.gpu),
        "benchmark": "VBench",
        "vbench_version": "0.1.5",
        "vbench_long": False,
        "vbench_info": str(args.vbench_info.resolve()),
        "vbench_info_sha256": sha256(args.vbench_info.resolve()),
        "mapping": str(mapping_path),
        "mapping_sha256": sha256(mapping_path),
        "videos_root": str(videos_root / args.strategy),
        "num_videos": 251,
        "dimensions": list(DIMENSIONS),
        "raw_scores": raw_scores,
        **aggregate,
        "scene_prompt_note": (
            "prompt_en is the Self-Forcing extended prompt; vbench==0.1.5 "
            "scene uses the official auxiliary scene keyword by design."
        ),
    }
    if details:
        details_path = output_root / "vbench/details" / f"{args.strategy}.json"
        details_path.parent.mkdir(parents=True, exist_ok=True)
        details_path.write_text(
            json.dumps(details, ensure_ascii=False, indent=2) + "\n",
            encoding="utf-8",
        )
        result["details"] = str(details_path)
    score_path.parent.mkdir(parents=True, exist_ok=True)
    score_path.write_text(
        json.dumps(result, ensure_ascii=False, indent=2) + "\n", encoding="utf-8"
    )
    print(
        f"[complete] strategy={args.strategy} "
        f"selected={result['selected_vbench_percent']:.4f}%",
        flush=True,
    )


if __name__ == "__main__":
    main()