CUB-200.py

import os
import datasets
from datasets import DatasetInfo, DownloadManager, Image
import pickle
import random
from tqdm import tqdm
#from PIL import Image

_URLS = {
    "train": "train/filenames.pickle",
    "test": "test/filenames.pickle",
    "images": "images.zip",
    "text": "text.zip",
}

class CUB200Dataset(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        # add number before name for sorting
        datasets.BuilderConfig(name="full"),
    ]

    def _info(self) -> DatasetInfo:
        features = {
            "image": datasets.Image(),
            "text": datasets.Value("string")
        }
        info = datasets.DatasetInfo(
            features=datasets.Features(features),
            supervised_keys=None,
            citation="",
        )
        return info

    def _split_generators(self, dl_manager: DownloadManager):
        downloaded_files = dl_manager.download_and_extract(_URLS)
        return [datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"file_path": downloaded_files["train"],"images_path":downloaded_files["images"]+"/images","text_path":downloaded_files["text"]+"/text"}),
                datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={"file_path": downloaded_files["test"],"images_path":downloaded_files["images"]+"/images","text_path":downloaded_files["text"]+"/text"}),
                ]

    def _generate_examples(self, file_path, images_path, text_path):
        name_list = pickle.load(open(file_path, "rb"), encoding="bytes")
        sum = len(name_list)
        for index in tqdm(range(sum)):
            name = name_list[index]
            this_text_path = os.path.join(text_path, name+".txt")
            with open(this_text_path, "r") as f:
                caption = random.choice(f.readlines()).replace("\n", "").lower()
            image = Image().encode_example(os.path.join(images_path, name+".jpg"))
            yield index, {
                "image": image,
                "text": caption,
            }