diff-hf-datasets.py

# /// script
# requires-python = ">=3.11"
# dependencies = [
#     "datasets>=3.1.0",
#     "huggingface-hub",
#     "tqdm",
#     "Pillow", 
# ]
# ///

"""
Semantic diff between two object detection datasets on Hugging Face Hub.

Mirrors panlabel's diff command. Compares two dataset versions and reports:

- Images shared / only-in-A / only-in-B
- Categories shared / only-in-A / only-in-B
- Annotations added / removed / modified
- Bbox geometry changes (IoU-based matching)

Matching strategies:
  - ID-based: Match images by file_name or image_id column
  - For annotations within shared images, match by IoU threshold

Examples:
  uv run diff-hf-datasets.py merve/dataset-v1 merve/dataset-v2
  uv run diff-hf-datasets.py merve/old merve/new --iou-threshold 0.7 --detail
  uv run diff-hf-datasets.py merve/old merve/new --report json
"""

import argparse
import json
import logging
import math
import os
import sys
from collections import Counter, defaultdict
from datetime import datetime
from typing import Any

from datasets import load_dataset
from huggingface_hub import login
from tqdm.auto import tqdm

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

BBOX_FORMATS = ["coco_xywh", "xyxy", "voc", "yolo", "tfod", "label_studio"]


def to_xyxy(bbox: list[float], fmt: str, img_w: float = 1.0, img_h: float = 1.0) -> tuple[float, float, float, float]:
    if fmt == "coco_xywh":
        x, y, w, h = bbox
        return (x, y, x + w, y + h)
    elif fmt in ("xyxy", "voc"):
        return tuple(bbox[:4])
    elif fmt == "yolo":
        cx, cy, w, h = bbox
        return (cx - w / 2) * img_w, (cy - h / 2) * img_h, (cx + w / 2) * img_w, (cy + h / 2) * img_h
    elif fmt == "tfod":
        xmin_n, ymin_n, xmax_n, ymax_n = bbox
        return (xmin_n * img_w, ymin_n * img_h, xmax_n * img_w, ymax_n * img_h)
    elif fmt == "label_studio":
        x_pct, y_pct, w_pct, h_pct = bbox
        return (
            x_pct / 100.0 * img_w,
            y_pct / 100.0 * img_h,
            (x_pct + w_pct) / 100.0 * img_w,
            (y_pct + h_pct) / 100.0 * img_h,
        )
    else:
        raise ValueError(f"Unknown bbox format: {fmt}")


def compute_iou(box_a: tuple, box_b: tuple) -> float:
    """Compute IoU between two XYXY boxes."""
    xa = max(box_a[0], box_b[0])
    ya = max(box_a[1], box_b[1])
    xb = min(box_a[2], box_b[2])
    yb = min(box_a[3], box_b[3])

    inter = max(0, xb - xa) * max(0, yb - ya)
    area_a = (box_a[2] - box_a[0]) * (box_a[3] - box_a[1])
    area_b = (box_b[2] - box_b[0]) * (box_b[3] - box_b[1])
    union = area_a + area_b - inter

    if union <= 0:
        return 0.0
    return inter / union


def extract_annotations(
    example: dict[str, Any],
    bbox_column: str,
    category_column: str,
    bbox_format: str,
    width_column: str | None,
    height_column: str | None,
) -> list[dict]:
    """Extract annotations from example as list of {bbox_xyxy, category}."""
    objects = example.get("objects", example)
    bboxes = objects.get(bbox_column, []) or []
    categories = objects.get(category_column, []) or []

    img_w = 1.0
    img_h = 1.0
    if width_column:
        img_w = float(example.get(width_column, 1.0) or 1.0)
    if height_column:
        img_h = float(example.get(height_column, 1.0) or 1.0)

    anns = []
    for i, bbox in enumerate(bboxes):
        if bbox is None or len(bbox) < 4:
            continue
        if not all(math.isfinite(v) for v in bbox[:4]):
            continue
        xyxy = to_xyxy(bbox[:4], bbox_format, img_w, img_h)
        cat = str(categories[i]) if i < len(categories) else "<unknown>"
        anns.append({"bbox_xyxy": xyxy, "category": cat})

    return anns


def match_annotations_iou(
    anns_a: list[dict],
    anns_b: list[dict],
    iou_threshold: float,
) -> tuple[list[tuple[int, int, float]], list[int], list[int]]:
    """Greedy IoU matching. Returns (matched_pairs, unmatched_a, unmatched_b)."""
    if not anns_a or not anns_b:
        return [], list(range(len(anns_a))), list(range(len(anns_b)))

    # Compute all pairwise IoUs
    pairs = []
    for i, a in enumerate(anns_a):
        for j, b in enumerate(anns_b):
            iou = compute_iou(a["bbox_xyxy"], b["bbox_xyxy"])
            if iou >= iou_threshold:
                pairs.append((iou, i, j))

    pairs.sort(reverse=True)

    matched_a = set()
    matched_b = set()
    matches = []

    for iou, i, j in pairs:
        if i not in matched_a and j not in matched_b:
            matches.append((i, j, iou))
            matched_a.add(i)
            matched_b.add(j)

    unmatched_a = [i for i in range(len(anns_a)) if i not in matched_a]
    unmatched_b = [j for j in range(len(anns_b)) if j not in matched_b]

    return matches, unmatched_a, unmatched_b


def get_image_key(example: dict, id_column: str) -> str:
    """Get a unique key for an image example."""
    val = example.get(id_column)
    if val is not None:
        return str(val)
    return str(example.get("file_name", ""))


def main(
    dataset_a: str,
    dataset_b: str,
    bbox_column: str = "bbox",
    category_column: str = "category",
    bbox_format: str = "coco_xywh",
    id_column: str = "image_id",
    width_column: str | None = "width",
    height_column: str | None = "height",
    split: str = "train",
    max_samples: int | None = None,
    iou_threshold: float = 0.5,
    detail: bool = False,
    report_format: str = "text",
    hf_token: str | None = None,
):
    """Compare two object detection datasets semantically."""

    start_time = datetime.now()

    HF_TOKEN = hf_token or os.environ.get("HF_TOKEN")
    if HF_TOKEN:
        login(token=HF_TOKEN)

    logger.info(f"Loading dataset A: {dataset_a}")
    ds_a = load_dataset(dataset_a, split=split)
    logger.info(f"Loading dataset B: {dataset_b}")
    ds_b = load_dataset(dataset_b, split=split)

    if max_samples:
        ds_a = ds_a.select(range(min(max_samples, len(ds_a))))
        ds_b = ds_b.select(range(min(max_samples, len(ds_b))))

    # Index by image key
    logger.info("Indexing images...")
    index_a = {}
    for i in tqdm(range(len(ds_a)), desc="Indexing A"):
        ex = ds_a[i]
        key = get_image_key(ex, id_column)
        index_a[key] = ex

    index_b = {}
    for i in tqdm(range(len(ds_b)), desc="Indexing B"):
        ex = ds_b[i]
        key = get_image_key(ex, id_column)
        index_b[key] = ex

    keys_a = set(index_a.keys())
    keys_b = set(index_b.keys())

    shared_keys = keys_a & keys_b
    only_a_keys = keys_a - keys_b
    only_b_keys = keys_b - keys_a

    # Collect all categories
    cats_a = set()
    cats_b = set()
    total_added = 0
    total_removed = 0
    total_modified = 0
    total_matched = 0
    detail_records = []

    logger.info(f"Comparing {len(shared_keys)} shared images...")
    for key in tqdm(sorted(shared_keys), desc="Diffing"):
        ex_a = index_a[key]
        ex_b = index_b[key]

        anns_a = extract_annotations(ex_a, bbox_column, category_column, bbox_format, width_column, height_column)
        anns_b = extract_annotations(ex_b, bbox_column, category_column, bbox_format, width_column, height_column)

        for a in anns_a:
            cats_a.add(a["category"])
        for b in anns_b:
            cats_b.add(b["category"])

        matches, unmatched_a, unmatched_b = match_annotations_iou(anns_a, anns_b, iou_threshold)

        total_matched += len(matches)
        total_removed += len(unmatched_a)
        total_added += len(unmatched_b)

        # Check for category changes in matched pairs
        for i, j, iou in matches:
            if anns_a[i]["category"] != anns_b[j]["category"]:
                total_modified += 1
                if detail:
                    detail_records.append({
                        "image": key,
                        "type": "modified",
                        "from_category": anns_a[i]["category"],
                        "to_category": anns_b[j]["category"],
                        "iou": round(iou, 3),
                    })

        if detail:
            for idx in unmatched_a:
                detail_records.append({
                    "image": key,
                    "type": "removed",
                    "category": anns_a[idx]["category"],
                    "bbox": list(anns_a[idx]["bbox_xyxy"]),
                })
            for idx in unmatched_b:
                detail_records.append({
                    "image": key,
                    "type": "added",
                    "category": anns_b[idx]["category"],
                    "bbox": list(anns_b[idx]["bbox_xyxy"]),
                })

    # Count annotations in only-A and only-B images
    anns_only_a = 0
    for key in only_a_keys:
        anns = extract_annotations(index_a[key], bbox_column, category_column, bbox_format, width_column, height_column)
        anns_only_a += len(anns)
        for a in anns:
            cats_a.add(a["category"])

    anns_only_b = 0
    for key in only_b_keys:
        anns = extract_annotations(index_b[key], bbox_column, category_column, bbox_format, width_column, height_column)
        anns_only_b += len(anns)
        for b in anns:
            cats_b.add(b["category"])

    shared_cats = cats_a & cats_b
    only_a_cats = cats_a - cats_b
    only_b_cats = cats_b - cats_a

    processing_time = datetime.now() - start_time

    report = {
        "dataset_a": dataset_a,
        "dataset_b": dataset_b,
        "split": split,
        "iou_threshold": iou_threshold,
        "images": {
            "in_a": len(keys_a),
            "in_b": len(keys_b),
            "shared": len(shared_keys),
            "only_in_a": len(only_a_keys),
            "only_in_b": len(only_b_keys),
        },
        "categories": {
            "in_a": len(cats_a),
            "in_b": len(cats_b),
            "shared": len(shared_cats),
            "only_in_a": sorted(only_a_cats),
            "only_in_b": sorted(only_b_cats),
        },
        "annotations": {
            "matched": total_matched,
            "modified": total_modified,
            "added_in_shared_images": total_added,
            "removed_in_shared_images": total_removed,
            "in_only_a_images": anns_only_a,
            "in_only_b_images": anns_only_b,
        },
        "processing_time_seconds": processing_time.total_seconds(),
    }

    if detail:
        report["details"] = detail_records

    if report_format == "json":
        print(json.dumps(report, indent=2))
    else:
        print("\n" + "=" * 60)
        print(f"Dataset Diff")
        print(f"  A: {dataset_a}")
        print(f"  B: {dataset_b}")
        print("=" * 60)

        img = report["images"]
        print(f"\n  Images:")
        print(f"    A: {img['in_a']:,}  |  B: {img['in_b']:,}")
        print(f"    Shared: {img['shared']:,}")
        print(f"    Only in A: {img['only_in_a']:,}")
        print(f"    Only in B: {img['only_in_b']:,}")

        cat = report["categories"]
        print(f"\n  Categories:")
        print(f"    A: {cat['in_a']}  |  B: {cat['in_b']}  |  Shared: {cat['shared']}")
        if cat["only_in_a"]:
            print(f"    Only in A: {', '.join(cat['only_in_a'][:10])}")
        if cat["only_in_b"]:
            print(f"    Only in B: {', '.join(cat['only_in_b'][:10])}")

        ann = report["annotations"]
        print(f"\n  Annotations (IoU >= {iou_threshold}):")
        print(f"    Matched:  {ann['matched']:,}")
        print(f"    Modified: {ann['modified']:,} (category changed)")
        print(f"    Added:    {ann['added_in_shared_images']:,} (in shared images)")
        print(f"    Removed:  {ann['removed_in_shared_images']:,} (in shared images)")
        if ann["in_only_a_images"]:
            print(f"    In A-only images: {ann['in_only_a_images']:,}")
        if ann["in_only_b_images"]:
            print(f"    In B-only images: {ann['in_only_b_images']:,}")

        if detail and detail_records:
            print(f"\n  Detail ({len(detail_records)} changes):")
            for rec in detail_records[:20]:
                if rec["type"] == "modified":
                    print(f"    [{rec['image']}] {rec['from_category']} -> {rec['to_category']} (IoU={rec['iou']})")
                elif rec["type"] == "added":
                    print(f"    [{rec['image']}] + {rec['category']}")
                elif rec["type"] == "removed":
                    print(f"    [{rec['image']}] - {rec['category']}")
            if len(detail_records) > 20:
                print(f"    ... and {len(detail_records) - 20} more")

        print(f"\n  Processing time: {processing_time.total_seconds():.1f}s")
        print("=" * 60)


if __name__ == "__main__":
    parser = argparse.ArgumentParser(
        description="Semantic diff between two object detection datasets on HF Hub",
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog="""
Examples:
  uv run diff-hf-datasets.py merve/dataset-v1 merve/dataset-v2
  uv run diff-hf-datasets.py merve/old merve/new --iou-threshold 0.7 --detail
  uv run diff-hf-datasets.py merve/old merve/new --report json
        """,
    )

    parser.add_argument("dataset_a", help="First dataset ID (A)")
    parser.add_argument("dataset_b", help="Second dataset ID (B)")
    parser.add_argument("--bbox-column", default="bbox", help="Column containing bboxes (default: bbox)")
    parser.add_argument("--category-column", default="category", help="Column containing categories (default: category)")
    parser.add_argument("--bbox-format", choices=BBOX_FORMATS, default="coco_xywh", help="Bbox format (default: coco_xywh)")
    parser.add_argument("--id-column", default="image_id", help="Column to match images by (default: image_id)")
    parser.add_argument("--width-column", default="width", help="Column for image width (default: width)")
    parser.add_argument("--height-column", default="height", help="Column for image height (default: height)")
    parser.add_argument("--split", default="train", help="Dataset split (default: train)")
    parser.add_argument("--max-samples", type=int, help="Max samples per dataset")
    parser.add_argument("--iou-threshold", type=float, default=0.5, help="IoU threshold for matching (default: 0.5)")
    parser.add_argument("--detail", action="store_true", help="Show per-annotation changes")
    parser.add_argument("--report", choices=["text", "json"], default="text", help="Report format (default: text)")
    parser.add_argument("--hf-token", help="HF API token")

    args = parser.parse_args()

    main(
        dataset_a=args.dataset_a,
        dataset_b=args.dataset_b,
        bbox_column=args.bbox_column,
        category_column=args.category_column,
        bbox_format=args.bbox_format,
        id_column=args.id_column,
        width_column=args.width_column,
        height_column=args.height_column,
        split=args.split,
        max_samples=args.max_samples,
        iou_threshold=args.iou_threshold,
        detail=args.detail,
        report_format=args.report,
        hf_token=args.hf_token,
    )