我不确定你到底想实现什么，但看起来你想要一个类似于Torch中的数据集类的东西。我将为我以前做的工作链接，希望它能为你工作：
https://github.com/Khaliladib11/INM705-CW-Khalil-Aziz/blob/main/Semnatic%20Segmentation/CityScapes.py

你通常会把这些东西分开，但它们应该是相同的顺序。让X是图像，y是标签，那么你可以这样做：

# allows to concatenate path fragments
from os.path import join
# allows to search for all files of a type
from glob import glob
# allows you to handle csv files
import pandas as pd

IMG_DIR = "my/img/dir/"

# your sorted img files
img_paths = list(sorted(glob(join(IMG_DIR, "*.png")))

def read_img(path):
  # your image load method of choice
  img = ...
  return img

# all loaded imgs
imgs = [read_img(p) for p in img_paths]

# labels
df = pd.read_csv("my.csv", sep=",")
labels = list(sorted(df.sort_values(by="identifier that is the same as the image name")["mylabelcolumn"].tolist()))

# this does not work if there is no 1:1 mapping from imgs to labels
len(labels) == len(img_paths)

# here, label[123] corresponds to img_paths[123] / imgs[123]

注意：有很多方法可以加载图像数据（OpenCV、tensorflow、pytorch、PIL等）。
注意：这将把所有img加载到内存中...这对大数据集不起作用。