CreateDataset.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "pandas",
#     "pyarrow",
#     "pydantic",
# ]
# ///
import os
import argparse
from pathlib import Path
import pandas as pd
from pydantic import BaseModel, field_validator
from typing import Optional


IMAGE_EXT = {".jpg", ".jpeg", ".png", ".webp"}
VIDEO_EXT = {".mp4", ".webm", ".avi", ".mov"}


def clean_name(name: str) -> str:
    return name.strip().replace(" ", "_")


class MemeRecord(BaseModel):
    image_file_name: Optional[str] = None
    video_file_name: Optional[str] = None
    category: str
    caption: str = ""

    @field_validator("image_file_name", "video_file_name")
    @classmethod
    def normalize_path(cls, v: Optional[str]) -> Optional[str]:
        return v.replace("\\", "/") if v else v

    @field_validator("category")
    @classmethod
    def normalize_category(cls, v: str) -> str:
        return clean_name(v)


def main(root: Path):
    src = root / "meme"
    if not src.exists():
        raise SystemExit(f"❌ Missing folder: {src}")

    records: list[MemeRecord] = []

    for category_dir in src.iterdir():
        if not category_dir.is_dir():
            continue

        category = clean_name(category_dir.name)

        for root_dir, _, files in os.walk(category_dir):
            for file in files:
                path = Path(root_dir) / file
                ext = path.suffix.lower()

                rel = str(path.relative_to(root))

                if ext in IMAGE_EXT:
                    records.append(
                        MemeRecord(
                            image_file_name=rel,
                            category=category
                        )
                    )
                elif ext in VIDEO_EXT:
                    records.append(
                        MemeRecord(
                            video_file_name=rel,
                            category=category
                        )
                    )

    if not records:
        raise SystemExit("❌ No media found")

    df = pd.DataFrame([r.model_dump() for r in records])
    df.to_csv(root / "metadata.csv", index=False)
    df.to_parquet(root / "metadata.parquet", index=False)

    print(f"✔ Indexed {len(df)} files (HF-safe)")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--root", type=Path, default=Path.cwd())
    args = parser.parse_args()
    main(args.root)