stackmix_hkr.py

import os
import pandas as pd
import zipfile

import datasets
from datasets import load_dataset


class StackmixHKR(datasets.GeneratorBasedBuilder):

    def _info(self):
        return datasets.DatasetInfo(
            features=datasets.Features(
                {
                    "path": datasets.Value("string"),
                    "name": datasets.Value("string"),
                    "text": datasets.Value("string"),
                    "image": datasets.Image()
                }
            )
        )

    def _split_generators(self, dl_manager):
        _URLS = {f"img{i:02d}": f"img/img-{i:02d}.zip" for i in range(10)}
        _URLS["labels"] = "gt.txt"
        
        data_paths = dl_manager.download(_URLS)
        archive_paths = [data_paths[key] for key in data_paths if key.startswith("img")]
        img_path = os.path.split(data_paths["img00"])[0]

        for archive_path in archive_paths:
            with zipfile.ZipFile(archive_path) as archive:
                archive.extractall(img_path)

        image_paths = [os.path.join(img_path, img_name) for img_name in os.listdir(img_path)]
        return [datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,
                    gen_kwargs={
                        "image_paths": image_paths,
                        "labels_path": data_paths["labels"]
                    })]

    def _generate_examples(self, image_paths, labels_path):
        df = pd.read_csv(labels_path, sep=",", names=["path", "text"])
        df["path"] = df.path.str[4:]
        df.set_index("path", inplace=True)

        for image_path in image_paths:
            image_name = os.path.basename(image_path)
 
            if image_name in df.index:

                example = {
                    "path": image_path,
                    "name": image_name,
                    "text": df["text"][image_name],
                    "image": image_path
                }
                yield image_name, example