scripts/MultiID/process_multiid.py

#!/usr/bin/env python3
"""Process MultiID-2M for MVEB: crop, build annotations from split JSON, pack parquet."""

from __future__ import annotations

import argparse
import json
import shutil
import sys
import threading
from concurrent.futures import ThreadPoolExecutor, as_completed
from pathlib import Path
from typing import Dict, List, Sequence, Tuple

from PIL import Image
from tqdm import tqdm

_SCRIPTS_ROOT = Path(__file__).resolve().parent.parent
if str(_SCRIPTS_ROOT) not in sys.path:
    sys.path.insert(0, str(_SCRIPTS_ROOT))
from pack_media_parquet import pack_dataset_with_media, resolve_split_output_dir

SCRIPT_DIR = Path(__file__).resolve().parent
ROOT_DIR = SCRIPT_DIR.parent.parent.parent

QUERY_INSTRUCTION = "You are a helpful assistant."
QUERY_TEXT = "Represent the face in the given image."
CANDIDATE_INSTRUCTION = "You are a helpful assistant."
CANDIDATE_TEXT = "Represent all faces in the given image."


def crop_faces(source_dir: Path, *, num_workers: int = 10, force: bool = False) -> None:
    """Crop faces from train_cp json/jpg pairs into cropped_grounding_cp_images."""
    train_cp_dir = source_dir / "train_cp"
    cropped_dir = source_dir / "cropped_grounding_cp_images"
    if not train_cp_dir.is_dir():
        raise FileNotFoundError(f"Missing train_cp dir: {train_cp_dir}")

    cropped_dir.mkdir(parents=True, exist_ok=True)
    subdirs = sorted(d.name for d in train_cp_dir.iterdir() if d.is_dir())
    if not subdirs:
        raise FileNotFoundError(f"No extracted shards under {train_cp_dir}")

    lock = threading.Lock()
    fail_count = 0

    def process_subdir(subdir: str) -> int:
        local_fail = 0
        anno_files = sorted((train_cp_dir / subdir).glob("*.json"))
        for anno_file in tqdm(anno_files, desc=f"crop {subdir}", leave=False):
            try:
                anno = json.loads(anno_file.read_text(encoding="utf-8"))
                img_path = anno_file.with_suffix(".jpg")
                if not img_path.is_file():
                    local_fail += 1
                    continue
                crop_box = anno["crop"]
                img = Image.open(img_path)
                for i, bbox in enumerate(anno["bboxes"]):
                    inst_id = str(anno["name"][i])
                    out_name = f"{img_path.stem}_{inst_id}.jpg"
                    out_path = cropped_dir / subdir / out_name
                    if out_path.is_file() and not force:
                        continue
                    face_bbox = [
                        bbox[0] - crop_box[0],
                        bbox[1] - crop_box[1],
                        bbox[2] - crop_box[0],
                        bbox[3] - crop_box[1],
                    ]
                    out_path.parent.mkdir(parents=True, exist_ok=True)
                    img.crop(face_bbox).save(out_path)
            except Exception:
                local_fail += 1
        return local_fail

    with ThreadPoolExecutor(max_workers=num_workers) as executor:
        futures = {executor.submit(process_subdir, subdir): subdir for subdir in subdirs}
        for future in tqdm(as_completed(futures), total=len(futures), desc="Cropping shards"):
            local_fail = future.result()
            with lock:
                fail_count += local_fail

    if fail_count:
        print(f"[warn] crop failed samples: {fail_count}")


def _parse_query_instance_id(query_rel_path: str) -> str:
    # cropped_grounding_cp_images/re_x/000010_000434.jpg -> 000434
    stem = Path(query_rel_path).stem
    if "_" not in stem:
        raise ValueError(f"Invalid query filename (expect *_<instance>): {query_rel_path}")
    return stem.rsplit("_", 1)[1]


def _candidate_json_rel_path(candidate_rel_path: str) -> str:
    # train_cp/re_x/000010.jpg -> train_cp/re_x/000010.json
    return str(Path(candidate_rel_path).with_suffix(".json")).replace("\\", "/")


def _load_split(split_json: Path) -> Dict[str, List[str]]:
    data = json.loads(split_json.read_text(encoding="utf-8"))
    for key in ("train", "test"):
        if key not in data or not isinstance(data[key], list):
            raise ValueError(f"{split_json} must contain {key!r} list")
    return data


def _build_annotations_for_split(
    split_name: str,
    split_paths: Sequence[str],
    source_dir: Path,
) -> Tuple[List[dict], List[dict]]:
    """Build annotations from split JSON per user-specified logic.

    - candidates: all `train_cp/...` images in split
    - query: all `cropped_grounding_cp_images/...` images in split
    - candidate instance mapping: parse corresponding train_cp/*.json `name` list
    - query pos_ids: mapping_dict[query_instance_id]
    """
    candidate_paths = sorted({p.replace("\\", "/") for p in split_paths if p.startswith("train_cp/")})
    query_paths = sorted(
        {p.replace("\\", "/") for p in split_paths if p.startswith("cropped_grounding_cp_images/")}
    )
    if not candidate_paths or not query_paths:
        raise ValueError(
            f"[{split_name}] split must contain both train_cp and cropped_grounding_cp_images paths"
        )

    candidate_rows: List[dict] = []
    inst_to_candidate_ids: Dict[str, List[str]] = {}

    for idx, rel_path in tqdm(enumerate(candidate_paths), total=len(candidate_paths), desc="Building candidate rows"):
        abs_path = source_dir / rel_path
        if not abs_path.is_file():
            raise FileNotFoundError(f"[{split_name}] missing candidate image: {abs_path}")

        anno_rel = _candidate_json_rel_path(rel_path)
        anno_abs = source_dir / anno_rel
        if not anno_abs.is_file():
            raise FileNotFoundError(f"[{split_name}] missing candidate json: {anno_abs}")

        anno = json.loads(anno_abs.read_text(encoding="utf-8"))
        instance_ids = [str(x) for x in anno.get("name", [])]
        if not instance_ids:
            continue

        candidate_rows.append(
            {
                "id": rel_path,
                # multi-person image: keep all ids in one field for audit
                "instance_id": "|".join(sorted(set(instance_ids))),
                "image_path": rel_path,
                "instruction": CANDIDATE_INSTRUCTION,
                "text": CANDIDATE_TEXT,
            }
        )
        for inst_id in set(instance_ids):
            inst_to_candidate_ids.setdefault(inst_id, []).append(rel_path)

    query_rows: List[dict] = []
    missing_pos = 0
    for rel_path in tqdm(query_paths, total=len(query_paths), desc="Building query rows"):
        abs_path = source_dir / rel_path
        if not abs_path.is_file():
            raise FileNotFoundError(f"[{split_name}] missing query image: {abs_path}")

        inst_id = _parse_query_instance_id(rel_path)
        pos_ids = inst_to_candidate_ids.get(inst_id, [])
        if not pos_ids:
            missing_pos += 1
            continue

        query_rows.append(
            {
                "id": rel_path,
                "instance_id": inst_id,
                "image_path": rel_path,
                "instruction": QUERY_INSTRUCTION,
                "text": QUERY_TEXT,
                "pos_ids": sorted(set(pos_ids)),
            }
        )

    if missing_pos:
        print(f"[{split_name}] warn: {missing_pos} query images have no matched candidate")
    if not query_rows or not candidate_rows:
        raise ValueError(f"[{split_name}] empty query/candidate after building")
    return query_rows, candidate_rows


def _process_one_split(
    split_name: str,
    split_paths: Sequence[str],
    source_dir: Path,
    output_root: Path,
    overwrite: bool,
    media_rows_per_shard: int,
    row_group_size: int,
    num_workers: int,
) -> None:
    out_dir = resolve_split_output_dir(output_root, split_name, "MultiID")
    if overwrite and out_dir.exists():
        shutil.rmtree(out_dir)
    out_dir.mkdir(parents=True, exist_ok=True)

    query_rows, candidate_rows = _build_annotations_for_split(split_name, split_paths, source_dir)
    stats = pack_dataset_with_media(
        query_annotations=query_rows,
        candidate_annotations=candidate_rows,
        image_dir=str(source_dir),
        output_dir=str(out_dir),
        media_rows_per_shard=media_rows_per_shard,
        row_group_size=row_group_size,
        num_workers=num_workers,
        dataset_name="MultiID",
        data_split=split_name,
        write_subset_readme=True,
        show_progress=True,
    )
    print(
        f"[{split_name}] done: media={stats['num_media']}, "
        f"query={stats['num_query']}, candidate={stats['num_candidate']}, "
        f"shards={stats['num_shards']} -> {out_dir}"
    )


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--source-dir",
        type=Path,
        default=ROOT_DIR / "source" / "MultiID-2M",
        help="Root containing train_cp/ and cropped_grounding_cp_images/.",
    )
    parser.add_argument(
        "--split-json",
        type=Path,
        default=SCRIPT_DIR / "train_test_split.json",
        help="JSON with split image paths (train_cp + cropped_grounding_cp_images).",
    )
    parser.add_argument(
        "--skip-crop",
        action="store_true",
        help="Skip face cropping; require existing cropped images + anno file.",
    )
    parser.add_argument(
        "--force-crop",
        action="store_true",
        help="Re-crop even if cropped jpg already exists.",
    )
    parser.add_argument("--crop-workers", type=int, default=10)
    parser.add_argument("--output-root", type=Path, default=ROOT_DIR)
    parser.add_argument("--overwrite", action="store_true")
    parser.add_argument("--media-rows-per-shard", type=int, default=5000)
    parser.add_argument("--row-group-size", type=int, default=100)
    parser.add_argument("--num-workers", type=int, default=1)
    parser.add_argument("--splits", nargs="+", default=["train", "test"], choices=["train", "test"])
    return parser.parse_args()


def main() -> None:
    args = parse_args()
    source_dir = args.source_dir
    if not source_dir.exists():
        raise FileNotFoundError(f"source dir not found: {source_dir}")

    if not args.skip_crop:
        print(f"==> Crop faces from {source_dir / 'train_cp'}")
        crop_faces(source_dir, num_workers=args.crop_workers, force=args.force_crop)

    split_data = _load_split(args.split_json)
    for split_name in args.splits:
        _process_one_split(
            split_name=split_name,
            split_paths=split_data[split_name],
            source_dir=source_dir,
            output_root=args.output_root,
            overwrite=args.overwrite,
            media_rows_per_shard=args.media_rows_per_shard,
            row_group_size=args.row_group_size,
            num_workers=args.num_workers,
        )


if __name__ == "__main__":
    main()