scripts/verify_dataset.py

"""Verify the Sawit-Weight dataset package.

Two layers are checked. The as-captured layer under annotations/raw/ must still match
every SHA-256 its manifests recorded. The corrected layer (annotations/labels/,
annotations/trees.jsonl, data/train/metadata.jsonl) must be internally consistent and
must differ from the raw layer only where a correction is declared below.

    python scripts/verify_dataset.py [dataset_root]
"""

import hashlib
import json
import os
import sys

TOL = 2e-3  # bbox_pixel and bbox_yolo agree to within this fraction of the image side

# Corrections applied on top of the as-captured annotations, mirrored from the README.
ADDED_BOXES = {("TANERA_303_BW_0004", 1)}          # (tree, side_index)
FRACTION_OVERRIDES = {"TANERA_303_BW_0003", "TANERA_303_BW_0004", "TANERA_303_BW_0005"}


def sha256(path):
    h = hashlib.sha256()
    with open(path, "rb") as f:
        for chunk in iter(lambda: f.read(1 << 20), b""):
            h.update(chunk)
    return h.hexdigest()


def read_json(path):
    with open(path, encoding="utf-8") as f:
        return json.load(f)


def read_jsonl(path):
    with open(path, encoding="utf-8") as f:
        return [json.loads(line) for line in f if line.strip()]


def read_label(path):
    with open(path, encoding="utf-8") as f:
        return [line.split() for line in f.read().splitlines() if line.strip()]


def main(root):
    images = os.path.join(root, "data", "train")
    ann = os.path.join(root, "annotations")
    raw = os.path.join(ann, "raw")
    fail = []

    def check(ok, message):
        if not ok:
            fail.append(message)

    trees = sorted(f[:-5] for f in os.listdir(os.path.join(raw, "trees")) if f.endswith(".json"))
    check(len(trees) > 0, "no tree JSON found")

    seen_images, seen_labels, digests = set(), set(), {}

    for name in trees:
        tree = read_json(os.path.join(raw, "trees", f"{name}.json"))
        manifest = read_json(os.path.join(raw, "manifests", f"{name}.json"))
        device = os.path.join(raw, "device_metadata", f"{name}.json")

        check(tree["tree_name"] == name, f"{name}: tree_name does not match filename")

        # Layer 1: the as-captured artifacts still hash to what the device recorded.
        for side in manifest["sides"]:
            rgb = os.path.join(images, side["rgb"]["file"])
            raw_label = os.path.join(raw, "labels", side["label"]["file"])
            check(os.path.exists(rgb), f"{name}: missing image {side['rgb']['file']}")
            check(os.path.exists(raw_label), f"{name}: missing raw label {side['label']['file']}")
            if os.path.exists(rgb):
                digest = sha256(rgb)
                check(digest == side["rgb"]["sha256"], f"{name}: SHA-256 mismatch {side['rgb']['file']}")
                check(digest not in digests, f"{name}: duplicate image content, also in {digests.get(digest)}")
                digests[digest] = side["rgb"]["file"]
                seen_images.add(side["rgb"]["file"])
            if os.path.exists(raw_label):
                check(sha256(raw_label) == side["label"]["sha256"],
                      f"{name}: SHA-256 mismatch {side['label']['file']}")
                seen_labels.add(side["label"]["file"])

        check(sha256(os.path.join(raw, "trees", f"{name}.json")) == manifest["outputJson"]["sha256"],
              f"{name}: SHA-256 mismatch on tree JSON")
        check(sha256(device) == manifest["metadata"]["sha256"], f"{name}: SHA-256 mismatch on device metadata")

        # Layer 2: the corrected labels are the raw ones plus only the declared additions.
        for key, side in tree["images"].items():
            raw_rows = read_label(os.path.join(raw, "labels", side["label_file"]))
            fixed_rows = read_label(os.path.join(ann, "labels", side["label_file"]))
            expected = side["bbox_count"] + (1 if (name, side["side_index"]) in ADDED_BOXES else 0)
            check(len(raw_rows) == side["bbox_count"],
                  f"{name}/{key}: raw label has {len(raw_rows)} rows, bbox_count is {side['bbox_count']}")
            check(len(fixed_rows) == expected,
                  f"{name}/{key}: corrected label has {len(fixed_rows)} rows, expected {expected}")

            for parts in fixed_rows:
                check(parts[0] == "0", f"{name}/{key}: class id is not 0")
                check(all(0.0 <= float(v) <= 1.0 for v in parts[1:]),
                      f"{name}/{key}: bbox outside [0, 1]")
            for i, parts in enumerate(raw_rows):
                check([round(float(v), 5) for v in parts[1:]] == [round(float(v), 5) for v in fixed_rows[i][1:]],
                      f"{name}/{key}: corrected label altered the original box {i}")

    # Capture set coverage.
    capture_set = read_json(os.path.join(raw, "capture_set.json"))
    check({t["treeName"] for t in capture_set["trees"]} == set(trees),
          "capture_set.json does not list exactly the trees present")

    # No orphan files.
    on_disk_images = {f for f in os.listdir(images) if f.lower().endswith(".jpg")}
    on_disk_raw = {f for f in os.listdir(os.path.join(raw, "labels")) if f.endswith(".txt")}
    on_disk_fixed = {f for f in os.listdir(os.path.join(ann, "labels")) if f.endswith(".txt")}
    check(on_disk_images == seen_images, f"orphan images: {sorted(on_disk_images ^ seen_images)}")
    check(on_disk_raw == seen_labels, f"orphan raw labels: {sorted(on_disk_raw ^ seen_labels)}")
    check(on_disk_fixed == seen_labels, f"orphan corrected labels: {sorted(on_disk_fixed ^ seen_labels)}")

    # metadata.jsonl agrees with the corrected labels and with itself.
    rows = read_jsonl(os.path.join(images, "metadata.jsonl"))
    check({r["file_name"] for r in rows} == seen_images, "metadata.jsonl does not cover every image")
    check(len(rows) == len(seen_images), "metadata.jsonl has duplicate file_name rows")

    for row in rows:
        objects = row["objects"]
        lengths = {len(v) for v in objects.values()}
        check(lengths == {row["num_bunches"]}, f"{row['file_name']}: objects lists disagree with num_bunches")

        label_rows = read_label(os.path.join(ann, "labels", row["file_name"][:-4] + ".txt"))
        check(len(label_rows) == row["num_bunches"],
              f"{row['file_name']}: label row count disagrees with num_bunches")
        for i, parts in enumerate(label_rows):
            check([round(float(v), 6) for v in parts[1:]] == [round(v, 6) for v in objects["bbox_yolo"][i]],
                  f"{row['file_name']}: label row {i} disagrees with bbox_yolo")

        w, h = row["width"], row["height"]
        for i, (cx, cy, bw, bh) in enumerate(objects["bbox_yolo"]):
            x1, y1, x2, y2 = objects["bbox_pixel"][i]
            check(x1 < x2 and y1 < y2, f"{row['file_name']}: box {i} has no positive area")
            check(abs((x1 + x2) / 2 / w - cx) < TOL and abs((y1 + y2) / 2 / h - cy) < TOL
                  and abs((x2 - x1) / w - bw) < TOL and abs((y2 - y1) / h - bh) < TOL,
                  f"{row['file_name']}: box {i} pixel and yolo forms disagree")

        for i, frac in enumerate(objects["ripeness_fraction"]):
            source = objects["ripeness_fraction_source"][i]
            check(frac in (1, 2, 3, 4, 5), f"{row['file_name']}: ripeness_fraction {frac} outside 1-5")
            check(source in ("field_note", "visual_estimate"),
                  f"{row['file_name']}: unknown ripeness_fraction_source {source}")
            check(source == "field_note" or row["tree_name"] in FRACTION_OVERRIDES,
                  f"{row['file_name']}: undeclared ripeness override")
        for i, source in enumerate(objects["bbox_source"]):
            check(source in ("field", "visual_estimate"), f"{row['file_name']}: unknown bbox_source {source}")
            check(source == "field" or (row["tree_name"], row["side_index"]) in ADDED_BOXES,
                  f"{row['file_name']}: undeclared added box")
        for note in objects["notes"]:
            check(note is None or "FAKSI" not in note.upper().replace("FRAKSI", ""),
                  f"{row['file_name']}: FAKSI left unnormalised in notes")

    tree_rows = read_jsonl(os.path.join(ann, "trees.jsonl"))
    check({r["tree_name"] for r in tree_rows} == set(trees), "trees.jsonl does not cover every tree")
    boxes_by_tree = {}
    for row in rows:
        boxes_by_tree[row["tree_name"]] = boxes_by_tree.get(row["tree_name"], 0) + row["num_bunches"]
    for row in tree_rows:
        check(row["total_detections"] == boxes_by_tree[row["tree_name"]],
              f"{row['tree_name']}: total_detections disagrees with metadata.jsonl")
        check(len(row["bunches"]) == row["total_unique_bunches"],
              f"{row['tree_name']}: bunch count disagrees with total_unique_bunches")
        for bunch in row["bunches"]:
            check(len(bunch["appearances"]) == bunch["appearance_count"],
                  f"{row['tree_name']}: appearance_count disagrees with appearances")

    print(f"trees {len(trees)} | images {len(seen_images)} | labels {len(seen_labels)} | "
          f"boxes {sum(r['num_bunches'] for r in rows)}")
    if fail:
        print(f"FAIL: {len(fail)} problem(s)")
        for message in fail:
            print(f"  - {message}")
        return 1
    print("PASS: raw SHA-256 digests hold and the corrected layer is consistent")
    return 0


if __name__ == "__main__":
    sys.exit(main(sys.argv[1] if len(sys.argv) > 1 else "."))