| """Step 6: package the benchmark as a Hugging Face dataset and push it. |
| |
| python -m fsbench.package_hf --bench data/bench --processed data/processed \ |
| --sources configs/fundusshift_open.yaml --card hf/README.md --repo le-dang1haha/FundusShift [--public] |
| Image splits go to parquet shards via `datasets`; metadata (manifest incl. unused rows, duplicate pairs, audit, |
| EDA, configs, pipeline code) is uploaded under metadata/ and code/. |
| """ |
| import argparse |
| import shutil |
| import tempfile |
| from pathlib import Path |
|
|
| import pandas as pd |
| import yaml |
| from datasets import ClassLabel, Dataset, DatasetDict, Features, Image, Value |
| from huggingface_hub import HfApi, hf_hub_download |
|
|
| from .common import CLASSES |
| SPLIT_NAMES = ["train", "validation", "calibration", "test", "test_ood", "test_openset"] |
| STR_COLS = ["image_id", "label_name", "source", "family", "region", "country", "device", "licence", "native_label", |
| "normal_scope", "patient_id", "eye", "group_id", "cal_subset", "orig_sha256", "quality_flags"] |
| FLOAT_COLS = ["fov_fraction", "brightness", "contrast", "sharpness"] |
|
|
|
|
| def features(core): |
| f = {"image": Image(), "label": ClassLabel(names=core)} |
| f.update({c: Value("string") for c in STR_COLS}) |
| f.update({"dr_grade": Value("int8"), "orig_width": Value("int32"), "orig_height": Value("int32")}) |
| f.update({c: Value("float32") for c in FLOAT_COLS}) |
| return Features(f) |
|
|
|
|
| def build(bench, processed, sources): |
| meta = {s["id"]: s for s in sources} |
| d = pd.read_csv(bench / "splits_final.csv", dtype=str, keep_default_na=False) |
| d = d[d.split.isin(SPLIT_NAMES)].copy() |
| core = [c for c in CLASSES if c in set(d.label[d.split == "train"])] |
| d["image"] = [str(processed / s / f"{i}.jpg") for s, i in zip(d.source, d.image_id)] |
| d["label_name"] = d.label |
| d["label"] = [core.index(x) if x in core else -1 for x in d.label] |
| d["device"] = [str(meta[s].get("device", "")) for s in d.source] |
| d["licence"] = [str(meta[s].get("licence", "")) for s in d.source] |
| d["orig_sha256"] = d.sha256 |
| d["dr_grade"] = pd.to_numeric(d.dr_grade, errors="coerce").fillna(-1).astype(int) |
| for c in ["orig_width", "orig_height"]: |
| d[c] = pd.to_numeric(d[c], errors="coerce").fillna(0).astype(float).astype(int) |
| for c in FLOAT_COLS: |
| d[c] = pd.to_numeric(d[c], errors="coerce").astype(float) |
| cols = list(features(core)) |
| dd = {} |
| for sp in SPLIT_NAMES: |
| sub = d[d.split == sp].sort_values(["source", "image_id"]) |
| if len(sub): |
| dd[sp] = Dataset.from_pandas(sub[cols].reset_index(drop=True), features=features(core), preserve_index=False) |
| return DatasetDict(dd) |
|
|
|
|
| def merge_card(repo, card_path, api): |
| """Keep the auto-generated dataset_info/configs YAML from push_to_hub, add our front matter and body.""" |
| auto = Path(hf_hub_download(repo, "README.md", repo_type="dataset")).read_text() |
| auto_yaml = yaml.safe_load(auto.split("---")[1]) if auto.startswith("---") else {} |
| text = Path(card_path).read_text() |
| _, ours_yaml, body = text.split("---", 2) |
| front = {**yaml.safe_load(ours_yaml), **{k: v for k, v in auto_yaml.items() if k in ("dataset_info", "configs")}} |
| out = "---\n" + yaml.safe_dump(front, sort_keys=False, allow_unicode=True) + "---" + body |
| api.upload_file(path_or_fileobj=out.encode(), path_in_repo="README.md", repo_id=repo, repo_type="dataset", |
| commit_message="Dataset card") |
|
|
|
|
| def main(argv=None): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--bench", required=True) |
| ap.add_argument("--processed", required=True) |
| ap.add_argument("--sources", required=True) |
| ap.add_argument("--card", required=True) |
| ap.add_argument("--repo", required=True) |
| ap.add_argument("--public", action="store_true") |
| ap.add_argument("--skip-images", action="store_true", help="only refresh card + metadata") |
| a = ap.parse_args(argv) |
| bench, processed = Path(a.bench), Path(a.processed) |
| sources = yaml.safe_load(open(a.sources))["sources"] |
| api = HfApi() |
| api.create_repo(a.repo, repo_type="dataset", private=not a.public, exist_ok=True) |
|
|
| if not a.skip_images: |
| ds = build(bench, processed, sources) |
| print(ds) |
| ds.push_to_hub(a.repo, private=not a.public, max_shard_size="500MB", |
| commit_message="FundusShift image splits") |
|
|
| root = Path(__file__).resolve().parents[1] |
| with tempfile.TemporaryDirectory() as tmp: |
| t = Path(tmp) |
| (t / "metadata/eda").mkdir(parents=True) |
| for f in ["manifest.csv", "near_duplicates.csv", "dup_candidates.csv", "cross_source_reuse.csv", |
| "audit_report.md", "splits_final.csv", "quality.csv"]: |
| if (bench / f).exists(): |
| shutil.copy(bench / f, t / "metadata" / f) |
| for f in (bench / "eda").glob("*"): |
| shutil.copy(f, t / "metadata/eda" / f.name) |
| shutil.copytree(root / "configs", t / "code/configs") |
| shutil.copytree(root / "fsbench", t / "code/fsbench", ignore=shutil.ignore_patterns("__pycache__")) |
| shutil.copytree(root / "scripts", t / "code/scripts") |
| shutil.copytree(root / "baselines", t / "code/baselines", ignore=shutil.ignore_patterns("__pycache__")) |
| shutil.copytree(root / "tests", t / "code/tests", ignore=shutil.ignore_patterns("__pycache__")) |
| for f in ["requirements.txt", "run_all.sh"]: |
| shutil.copy(root / f, t / "code" / f) |
| api.upload_folder(folder_path=str(t), repo_id=a.repo, repo_type="dataset", |
| commit_message="Metadata, audit, EDA and pipeline code") |
| merge_card(a.repo, a.card, api) |
| print(f"https://huggingface.co/datasets/{a.repo}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
