scripts/hf_push_dataset.py

from __future__ import annotations

import argparse
import os
from pathlib import Path
from typing import Dict, List, Tuple

# Best-effort load .env if present so users don't need the dotenv CLI.
try:
    from dotenv import load_dotenv  # type: ignore
    load_dotenv()  # loads .env from current working directory if it exists
except Exception:
    pass

import pandas as pd
from datasets import Dataset, DatasetDict, Image, Features, Value
from huggingface_hub import HfApi
from huggingface_hub.errors import HfHubHTTPError


def _find_splits(data_dir: Path) -> List[Tuple[str, Path]]:
    out = []
    for split in ("train", "validation", "test"):
        sd = data_dir / split
        if (sd / "metadata.csv").exists():
            out.append((split, sd))
    return out


def _normalize_columns(df: pd.DataFrame) -> pd.DataFrame:
    # Create a lowercase map but keep original for selection
    lower_map = {c.lower(): c for c in df.columns}

    # Image path
    img_col_candidates = [
        "image", "file_name", "filename", "path", "filepath", "file"
    ]
    img_col = next((lower_map[c] for c in img_col_candidates if c in lower_map), None)
    if img_col is None:
        raise ValueError(f"Could not find an image path column among: {df.columns.tolist()}")

    # Latitude
    lat_candidates = ["latitude", "lat", "Latitude", "LAT"]
    lat_col = None
    for c in lat_candidates:
        if c.lower() in lower_map:
            lat_col = lower_map[c.lower()]
            break
    if lat_col is None:
        raise ValueError("Latitude column not found (expected one of latitude/lat)")

    # Longitude
    lon_candidates = ["longitude", "lon", "Longitude", "LON", "long"]
    lon_col = None
    for c in lon_candidates:
        if c.lower() in lower_map:
            lon_col = lower_map[c.lower()]
            break
    if lon_col is None:
        raise ValueError("Longitude column not found (expected one of longitude/lon/long)")

    out_df = pd.DataFrame({
        "image": df[img_col].astype(str),
        "latitude": pd.to_numeric(df[lat_col], errors="coerce"),
        "longitude": pd.to_numeric(df[lon_col], errors="coerce"),
    })
    out_df = out_df.dropna(subset=["latitude", "longitude"]).reset_index(drop=True)
    return out_df


def _resolve_paths(df: pd.DataFrame, split_dir: Path) -> pd.DataFrame:
    paths = []
    for p in df["image"].tolist():
        pth = Path(p)
        if pth.is_absolute() and pth.exists():
            paths.append(str(pth))
            continue
        # try relative to split dir
        pth2 = (split_dir / p).resolve()
        if pth2.exists():
            paths.append(str(pth2))
            continue
        # try relative to data dir
        # keep as is if not found (datasets will error later, but we surface here)
        paths.append(str(p))
    df = df.copy()
    df["image"] = paths
    return df


def build_datasetdict(data_dir: Path) -> DatasetDict:
    splits = _find_splits(data_dir)
    if not splits:
        raise SystemExit(f"No splits found under {data_dir}. Expected metadata.csv in train/validation/test.")

    feats = Features({
        "image": Image(),
        "latitude": Value("float64"),
        "longitude": Value("float64"),
    })

    dd: Dict[str, Dataset] = {}
    for split, sd in splits:
        csv_path = sd / "metadata.csv"
        df = pd.read_csv(csv_path)
        df = _normalize_columns(df)
        df = _resolve_paths(df, sd)

        ds = Dataset.from_dict(df.to_dict(orient="list"), features=feats)
        dd[split] = ds
        print(f"Split {split}: {len(ds)} rows")

    return DatasetDict(dd)


def push_to_hub(ds: DatasetDict, repo_id: str, private: bool, max_shard_size: str) -> None:
    # Prefer explicit token if provided
    token = os.environ.get("HUGGINGFACE_HUB_TOKEN") or os.environ.get("HF_TOKEN")
    if not token:
        print("[auth] 未检测到 Token。请在 .env 设置 HUGGINGFACE_HUB_TOKEN=hf_xxx，或设置环境变量 HF_TOKEN/HUGGINGFACE_HUB_TOKEN。")
        print("[auth] 也可以先运行: python -c \"from huggingface_hub import login; login('hf_xxx')\"")
    try:
        api = HfApi(token=token)
        api.create_repo(repo_id=repo_id, repo_type="dataset", exist_ok=True, private=private)
        ds.push_to_hub(repo_id, private=private, max_shard_size=max_shard_size, token=token)
        print(f"Pushed to https://huggingface.co/datasets/{repo_id}")
    except HfHubHTTPError as e:
        if hasattr(e, "response") and getattr(e.response, "status_code", None) == 401:
            print("[auth] 401 Unauthorized：请检查 Token 是否有效、是否具备 write 权限、是否属于 LarryD123 账号。")
            print("[auth] 建议：\n - 在 https://huggingface.co/settings/tokens 重新生成 write Token\n - 将其写入项目根目录 .env (HUGGINGFACE_HUB_TOKEN=hf_xxx)\n - 重新运行上传命令")
        raise


def main():
    ap = argparse.ArgumentParser(description="Build and push a 3-column Image+GPS dataset to Hugging Face.")
    ap.add_argument("--data-dir", type=Path, required=True, help="Folder containing split subfolders (train/validation/test)")
    ap.add_argument("--repo-id", type=str, required=False, help="<user>/<dataset_name> on Hugging Face")
    ap.add_argument("--private", type=str, default="false", help="true/false for private dataset")
    ap.add_argument("--max-shard-size", type=str, default="500MB", help="Shard size for HF push")
    ap.add_argument("--dry-run", action="store_true", help="Build locally without pushing to Hub")
    args = ap.parse_args()

    ds = build_datasetdict(args.data_dir)
    print(ds)

    if args.dry_run:
        print("Dry run: not pushing to hub.")
        return

    if not args.repo_id:
        raise SystemExit("--repo-id is required unless --dry-run is set")

    private = str(args.private).lower() in ("1", "true", "yes", "y")
    push_to_hub(ds, args.repo_id, private=private, max_shard_size=args.max_shard_size)


if __name__ == "__main__":
    main()