frozen_lake.py

import datasets
import json
import tarfile
from os import listdir, makedirs
from os.path import isfile, join

import numpy as np

from datasets import DatasetInfo, GeneratorBasedBuilder, Split, SplitGenerator


_CITATION = ""

_DESCRIPTION = "Expert Dataset for the Frozen Lake Custom Environment. See paper for more details."

_HOMEPAGE = "https://huggingface.co/datasets/Ptisni/frozen_lake"

_LICENSE = ""

_REPO = "https://huggingface.co/datasets/Ptisni/frozen_lake"


class ImageSet(GeneratorBasedBuilder):

    def _info(self):
        return DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features({
                "obs": datasets.Value("string"),
                "actions": datasets.Value("int32"),
                "rewards": datasets.Value("float32"),
                "episode_starts": datasets.Value("bool"),
            }),
            supervised_keys=None,
            homepage=_HOMEPAGE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        info_path = dl_manager.download_and_extract(f"{_REPO}/resolve/main/fl_dataset.tar.gz")
        image_path = dl_manager.download_and_extract(f"{_REPO}/resolve/main/images.tar.gz")

        split = []
        for value in np.arange(0.0, 0.55, 0.05):
            value = round(value, 2)
            split.append(SplitGenerator(
                name="%.2f" % value,
                gen_kwargs={
                    "images_paths": f"{image_path}/images_{value}",
                    "infos": f"{info_path}/teacher_{value}.jsonl"
                }
            ))
        return split

    def _generate_examples(self, images_paths, infos):
        images = [join(images_paths, f) for f in listdir(images_paths) if isfile(join(images_paths, f))]

        images_dict = {}
        for image in images:
            images_dict[image.split("/")[-1].split(".")[0]] = image

        with open(infos, encoding="utf-8") as data:
            for idx, line in enumerate(data):
                record = json.loads(line)
                index = record["states"].split(".")[0]
                yield idx, {
                    "obs": images_dict[index],
                    "actions": record["actions"],
                    "rewards": record["rewards"],
                    "episode_starts": record["episode_starts"],
                }