Blame for scripts/verify.py

"""Integrity checks for the release. Run from anywhere:  python scripts/verify.py   (needs Pillow)
Checks: ids, boxes inside images, no duplicate names, no photo in two splits, tile/photo consistency,
counts vs stats.json and README.md, and checksums.sha256."""
import hashlib, json, re, sys
from pathlib import Path
from PIL import Image

root = Path(__file__).resolve().parent.parent
errs = []
def err(m): errs.append(m)
J = lambda p: json.loads((root / p).read_text())

classes = J("classes.json"); names = [c["name"] for c in classes]
assert [c["id"] for c in classes] == list(range(6)) and names == ["corn", "peanuts", "popcorn_corn", "pumpkin", "sunflower", "wheat"]
splits = J("splits.json"); stats = J("stats.json")

def check_coco(path, imgdir, tiles):
    d = J(path)
    if [c["name"] for c in d["categories"]] != names or [c["id"] for c in d["categories"]] != list(range(6)): err(f"{path}: categories")
    ids = [i["id"] for i in d["images"]]; fns = [i["file_name"] for i in d["images"]]
    if len(set(ids)) != len(ids): err(f"{path}: duplicate image ids")
    if len(set(fns)) != len(fns): err(f"{path}: duplicate file names")
    aids = [a["id"] for a in d["annotations"]]
    if len(set(aids)) != len(aids): err(f"{path}: duplicate annotation ids")
    im = {i["id"]: i for i in d["images"]}
    for i in d["images"]:
        p = root / imgdir / i["file_name"]
        if not p.exists(): err(f"{path}: missing {p}"); continue
        with Image.open(p) as x:
            if x.size != (i["width"], i["height"]): err(f"{path}: size mismatch {i['file_name']}")
    for a in d["annotations"]:
        i = im.get(a["image_id"])
        if i is None: err(f"{path}: ann {a['id']} bad image_id"); continue
        x, y, w, h = a["bbox"]
        if not (0 <= a["category_id"] < 6): err(f"{path}: ann {a['id']} bad category")
        if w <= 0 or h <= 0 or x < -0.5 or y < -0.5 or x + w > i["width"] + 0.5 or y + h > i["height"] + 0.5: err(f"{path}: ann {a['id']} box outside {i['file_name']}")
    on_disk = {p.name for p in (root / imgdir).glob("*.jpg")}
    if on_disk != set(fns): err(f"{path}: files on disk differ from labels ({len(on_disk)} vs {len(fns)})")
    return d

D = {s: check_coco(f"coco/{s}/labels.json", f"coco/{s}/images", True) for s in ("train", "val", "test")}
P = {s: check_coco(f"annotations/{s}_photos.json", f"photos/{s}", False) for s in ("val", "test")}

# no photo in two splits; tiles belong to the photo's split; photo files match the split map
photo_of = lambda fn: re.sub(r"_x\d+_y\d+\.jpg$", ".jpg", fn)
seen = {}
for s, d in D.items():
    for i in d["images"]:
        ph = photo_of(i["file_name"])
        if splits.get(ph) != s: err(f"tile {i['file_name']} in {s} but photo is in {splits.get(ph)}")
        seen.setdefault(ph, set()).add(s)
for ph, ss in seen.items():
    if len(ss) > 1: err(f"photo {ph} in several splits: {ss}")
for s in ("val", "test"):
    for i in P[s]["images"]:
        if splits.get(i["file_name"]) != s: err(f"photo {i['file_name']} not in split {s}")
if len(set(splits)) != len(splits): err("duplicate photo names in splits.json")
for ph in splits:
    if not re.fullmatch(r"(corn|peanuts|popcorn_corn|pumpkin|sunflower|wheat)_\d{4}\.jpg", ph): err(f"bad photo name {ph}")

# counts vs stats.json and README
for s in ("train", "val", "test"):
    t = stats["totals"][s]
    n_ph = sum(1 for v in splits.values() if v == s)
    if (t["tiles"], t["boxes"], t["photos"]) != (len(D[s]["images"]), len(D[s]["annotations"]), n_ph): err(f"stats.json mismatch for {s}")
readme = (root / "README.md").read_text()
for s in ("train", "val", "test"):
    t = stats["totals"][s]
    row = f"| **all** | **{s}** | {t['photos']} | {t['tiles']} |"
    if row not in readme: err(f"README table row for {s} does not match the data")
if f"{len(splits)} photos" not in readme: err("README photo total")

# checksums
cs = root / "checksums.sha256"
if cs.exists():
    n = 0
    for line in cs.read_text().splitlines():
        h, _, rel = line.partition("  ")
        if hashlib.sha256((root / rel).read_bytes()).hexdigest() != h: err(f"checksum mismatch {rel}")
        n += 1
    listed = {l.partition("  ")[2] for l in cs.read_text().splitlines()}
    for p in root.rglob("*"):
        if ".cache" in p.relative_to(root).parts or p.name == ".gitattributes":      # added by the Hub / `hf download`
            continue
        if p.is_file() and p.name != "checksums.sha256" and str(p.relative_to(root)) not in listed: err(f"not in checksums: {p.relative_to(root)}")
else:
    err("checksums.sha256 missing")

print("tiles:", {s: len(D[s]["images"]) for s in D}, "boxes:", {s: len(D[s]["annotations"]) for s in D}, "photos:", {s: sum(1 for v in splits.values() if v == s) for s in D})
if errs:
    print(len(errs), "PROBLEMS"); [print(" -", e) for e in errs[:30]]; sys.exit(1)
print("OK: all checks passed")