File size: 7,634 Bytes
5f0e48f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 | """Same-task / different-domain power-grid detection splits (YOLO format, images hardlinked).
A_weather (clear -> fog), classes insulator/defect, from SFID (UPID images + synthetic fog):
SFID id i comes from UPID source image i//10 (i%10==0 is the original, the rest are its augmentations;
checked against UPID augmented_labels_train.json: all 686 originals sit at 10g). SFID's own split puts copies
of one source in both train and test, so we re-split by source: 70/10/20. Train sources are halved so
clear-train and fog-train never share a scene; val/test use the same sources in both domains (paired).
B_region (China -> Brazil), class insulator:
cn = SFID clear images (CPLID-derived), all train sources, same val/test sources as A, defect boxes dropped.
br = InsPLAD-det (UAV, Brazil; Voxel51/InsPLAD FiftyOne export), glass + polymer insulator -> insulator.
Official det val -> test; official det train split by tower/flight prefix 90/10 -> train/val.
python prepare_data.py sfid --raw raw/sfid/SFID --out A_weather
python prepare_data.py region --raw raw/sfid/SFID --insplad raw/insplad_hf --out B_region
"""
import argparse, json, os, random
from pathlib import Path
def put(root, split, stem, img, lines):
dst = root / "images" / split / f"{stem}{img.suffix}"
dst.parent.mkdir(parents=True, exist_ok=True)
if not dst.exists(): os.link(img, dst)
lab = root / "labels" / split / f"{stem}.txt"
lab.parent.mkdir(parents=True, exist_ok=True)
lab.write_text("".join(l + "\n" for l in lines))
def write_yaml(path, root, names, train, val, test):
fmt = lambda v: "[" + ", ".join(v) + "]" if isinstance(v, list) else v
path.write_text(f"path: {root}\ntrain: {fmt(train)}\nval: {fmt(val)}\ntest: {fmt(test)}\nnames: {dict(enumerate(names))}\n")
def write_yamls(out, doms, names):
for d in doms:
write_yaml(out / f"{d}.yaml", out, names, *(f"{d}/images/{s}" for s in ("train", "val", "test")))
write_yaml(out / "multi.yaml", out, names, *([f"{d}/images/{s}" for d in doms] for s in ("train", "val", "test")))
def sfid_groups(raw, seed=0):
"""SFID images {stem: path} and UPID source group -> 'train_clear' | 'train_fog' | 'val' | 'test'."""
imgs = {p.stem: p for p in Path(raw).glob("images/*/*.jpg")}
groups = sorted({int(s.removeprefix("fogged_")) // 10 for s in imgs})
random.Random(seed).shuffle(groups)
n = len(groups); a, b = int(.7 * n), int(.8 * n)
split = {g: "val" for g in groups[a:b]} | {g: "test" for g in groups[b:]}
split |= {g: ("train_clear", "train_fog")[i % 2] for i, g in enumerate(groups[:a])}
return imgs, split
def sfid_labels(raw, img, stem):
p = Path(raw) / "labels" / img.parent.name / f"{stem}.txt"
return [l.strip() for l in p.read_text().splitlines() if l.strip()] if p.exists() else []
def sfid(raw, out, seed=0):
out = Path(out).resolve(); imgs, split = sfid_groups(raw, seed); count = {}
for stem, img in imgs.items():
dom, i = ("fog" if stem.startswith("fogged_") else "clear"), int(stem.removeprefix("fogged_"))
s = split[i // 10]
if s.startswith("train_"):
if s != f"train_{dom}": continue
s = "train"
put(out / dom, s, f"{i:06d}", img, sfid_labels(raw, img, stem))
count[dom, s] = count.get((dom, s), 0) + 1
write_yamls(out, ("clear", "fog"), ["insulator", "defect"])
print("A_weather images", dict(sorted(count.items())))
INSPLAD_INSULATOR = {"glass insulator", "polymer insulator"}
def region(raw, insplad, out, seed=0):
out, ins = Path(out).resolve(), Path(insplad).resolve(); imgs, split = sfid_groups(raw, seed); count = {}
for stem, img in imgs.items():
if stem.startswith("fogged_"): continue
s = split[int(stem) // 10].split("_")[0] # both train halves
put(out / "cn", s, stem, img, [l for l in sfid_labels(raw, img, stem) if l.split()[0] == "0"])
count["cn", s] = count.get(("cn", s), 0) + 1
det = [x for x in json.load(open(ins / "samples.json"))["samples"] if x["task"] == "detection"]
pre = lambda x: Path(x["filepath"]).stem.split("_")[0] # tower / flight id
assert len({Path(x["filepath"]).stem for x in det}) == len(det), "InsPLAD stems collide"
tr = sorted({pre(x) for x in det if "train" in x["tags"]}); random.Random(seed).shuffle(tr)
val_pre, missing = set(tr[: len(tr) // 10]), 0
for x in det:
img = ins / x["filepath"]
if not img.exists(): missing += 1; continue
s = "test" if "val" in x["tags"] else "val" if pre(x) in val_pre else "train"
boxes = [d["bounding_box"] for d in (x["ground_truth"] or {}).get("detections", []) if d["label"] in INSPLAD_INSULATOR]
put(out / "br", s, img.stem, img, [f"0 {b[0] + b[2] / 2:.6f} {b[1] + b[3] / 2:.6f} {b[2]:.6f} {b[3]:.6f}" for b in boxes])
count["br", s] = count.get(("br", s), 0) + 1
write_yamls(out, ("cn", "br"), ["insulator"])
print("B_region images", dict(sorted(count.items())), "| InsPLAD missing", missing)
def cnfog(raw, out, seed=0):
"""Third B site: the cn scenes under SFID's synthetic fog (same source groups and splits as cn), insulator only."""
out = Path(out).resolve(); imgs, split = sfid_groups(raw, seed); count = {}
for stem, img in imgs.items():
if not stem.startswith("fogged_"): continue
i = int(stem.removeprefix("fogged_")); s = split[i // 10].split("_")[0] # both train halves, as cn
put(out / "cnfog", s, f"{i:06d}", img, [l for l in sfid_labels(raw, img, stem) if l.split()[0] == "0"])
count[s] = count.get(s, 0) + 1
write_yaml(out / "cnfog.yaml", out, ["insulator"], *(f"cnfog/images/{s}" for s in ("train", "val", "test")))
write_yaml(out / "multi3.yaml", out, ["insulator"], *([f"{d}/images/{s}" for d in ("cn", "cnfog", "br")] for s in ("train", "val", "test")))
print("B_region cnfog images", dict(sorted(count.items())))
def coco_ref(zip_path, out):
"""Generic images with empty labels: the local-support switch's negative density (gridkd.py lsl --neg)."""
import zipfile
root = Path(out).resolve()
img, lab = root / "images" / "val", root / "labels" / "val"
img.mkdir(parents=True, exist_ok=True); lab.mkdir(parents=True, exist_ok=True)
with zipfile.ZipFile(zip_path) as z:
for n in z.namelist():
if n.lower().endswith(".jpg"):
(img / Path(n).name).write_bytes(z.read(n))
(lab / f"{Path(n).stem}.txt").write_text("")
for tag, names in (("1c", ["insulator"]), ("2c", ["insulator", "defect"])): # class names must match the model
write_yaml(root / f"ref_{tag}.yaml", root, names, "images/val", "images/val", "images/val")
print(len(list(img.glob("*.jpg"))), "reference images ->", root)
if __name__ == "__main__":
ap = argparse.ArgumentParser(); sub = ap.add_subparsers(dest="cmd", required=True)
for c in ("sfid", "region"):
p = sub.add_parser(c); p.add_argument("--raw", required=True); p.add_argument("--out", required=True); p.add_argument("--seed", type=int, default=0)
if c == "region": p.add_argument("--insplad", required=True)
p = sub.add_parser("coco_ref"); p.add_argument("--zip", required=True); p.add_argument("--out", required=True)
p = sub.add_parser("cnfog"); p.add_argument("--raw", required=True); p.add_argument("--out", required=True)
a = ap.parse_args()
if a.cmd == "coco_ref": coco_ref(a.zip, a.out)
elif a.cmd == "cnfog": cnfog(a.raw, a.out)
elif a.cmd == "sfid": sfid(a.raw, a.out, a.seed)
else: region(a.raw, a.insplad, a.out, a.seed)
|