Blame for load_example.py

"""split_3fold.json 으로 fold k 의 학습/inner_val/시험 목록을 만들고, 사진 1장과 마스크를 tar 에서 바로 읽는 예시.



사용법: python load_example.py [k]      (필요: numpy, pillow)

"""
import io
import json
import sys
import tarfile

import numpy as np
from PIL import Image

k = int(sys.argv[1]) if len(sys.argv) > 1 else 0
split = json.load(open("split_3fold.json", encoding="utf-8"))
classes = [c["name"] for c in json.load(open("classes.json", encoding="utf-8"))["classes"]]
inner = set(split["inner_val"][str(k)])
test = sorted(n for n, v in split["images"].items() if v["fold"] == k)
train = sorted(n for n, v in split["images"].items() if v["fold"] != k and n not in inner)
print(f"fold {k}: train {len(train)}, inner_val {len(inner)}, test {len(test)}")

name = test[0]
with tarfile.open(f"masks/fold{k}.tar") as t:
    mask = np.load(io.BytesIO(t.extractfile(f"masks/{name}.npz").read()))["mask"]   # (19, H, W) bool
with tarfile.open(f"images/fold{k}.tar") as t:
    img = Image.open(io.BytesIO(t.extractfile(f"images/{name}.jpg").read()))
print(f"{name}: image {img.size[0]}x{img.size[1]}, mask {mask.shape} {mask.dtype}")
for i, c in enumerate(classes):
    if mask[i].any():
        print(f"  {i:2d} {c:14s} {int(mask[i].sum()):>10d} px")
print(f"  multi-label pixels (2+ classes): {int((mask.sum(0) > 1).sum())}")