Distorted_Human_Images.py

import pandas as pd
from huggingface_hub import hf_hub_url
import datasets
import os

_VERSION = datasets.Version("0.0.1")

_DESCRIPTION = "TODO"
_HOMEPAGE = "TODO"
_LICENSE = "TODO"
_CITATION = "TODO"

_FEATURES = datasets.Features(
    {
        "flawless": datasets.Image(),
        "distorted": datasets.Image(),
        "mask": datasets.Image(),
        "reference": datasets.Image(),
        "prompt": datasets.Value("string"),
    },
)

METADATA_URL = hf_hub_url(
    "NegarMov/Distorted_Human_Images",
    filename="train.jsonl",
    repo_type="dataset",
)

FLAWLESS_URL = hf_hub_url(
    "NegarMov/Distorted_Human_Images",
    filename="flawless.zip",
    repo_type="dataset",
)

DISTORTED_URL = hf_hub_url(
    "NegarMov/Distorted_Human_Images",
    filename="distorted.zip",
    repo_type="dataset",
)

MASK_URL = hf_hub_url(
    "NegarMov/Distorted_Human_Images",
    filename="mask.zip",
    repo_type="dataset",
)

REFERENCE_URL = hf_hub_url(
    "NegarMov/Distorted_Human_Images",
    filename="reference.zip",
    repo_type="dataset",
)

_DEFAULT_CONFIG = datasets.BuilderConfig(name="default", version=_VERSION)


class Distorted_Human_Images(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [_DEFAULT_CONFIG]
    DEFAULT_CONFIG_NAME = "default"

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=_FEATURES,
            supervised_keys=None,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        metadata_path = dl_manager.download(METADATA_URL)
        flawless_dir = dl_manager.download_and_extract(
            FLAWLESS_URL
        )
        distorted_dir = dl_manager.download_and_extract(
            DISTORTED_URL
        )
        mask_dir = dl_manager.download_and_extract(
            MASK_URL
        )
        reference_dir = dl_manager.download_and_extract(
            REFERENCE_URL
        )

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                # These kwargs will be passed to _generate_examples
                gen_kwargs={
                    "metadata_path": metadata_path,
                    "flawless_dir": flawless_dir,
                    "distorted_dir": distorted_dir,
                    "mask_dir": mask_dir,
                    "reference_dir": reference_dir,
                },
            ),
        ]

    def _generate_examples(self, metadata_path, flawless_dir, distorted_dir, mask_dir, reference_dir):
        metadata = pd.read_json(metadata_path, lines=True)

        for _, row in metadata.iterrows():
            prompt = row["prompt"]

            flawless_path = row["flawless"]
            flawless_path = os.path.join(flawless_dir, flawless_path)
            flawless = open(flawless_path, "rb").read()

            distorted_path = row["distorted"]
            distorted_path = os.path.join(
                distorted_dir, row["distorted"]
            )
            distorted = open(distorted_path, "rb").read()

            mask_path = row["mask"]
            mask_path = os.path.join(
                mask_dir, row["mask"]
            )
            mask = open(mask_path, "rb").read()

            reference_path = row["reference"]
            reference_path = os.path.join(
                reference_dir, row["reference"]
            )
            reference = open(reference_path, "rb").read()

            yield row["flawless"], {
                "prompt": prompt,
                "flawless": {
                    "path": flawless_path,
                    "bytes": flawless,
                },
                "distorted": {
                    "path": distorted_path,
                    "bytes": distorted,
                },
                "mask": {
                    "path": mask_path,
                    "bytes": mask,
                },
                "reference": {
                    "path": reference_path,
                    "bytes": reference,
                },
            }