cn_pose_dataset.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(
    {
        "image": datasets.Image(),
        "pose": datasets.Image(),
        "prompt": datasets.Value("string"),
    },
)

METADATA_URL = hf_hub_url(
    "vision-paper/cn_pose_dataset",
    filename="train.jsonl",
    repo_type="dataset",
)

IMAGE_URL = hf_hub_url(
    "vision-paper/cn_pose_dataset",
    filename="image.zip",
    repo_type="dataset",
)

POSE_URL = hf_hub_url(
    "vision-paper/cn_pose_dataset",
    filename="pose.zip",
    repo_type="dataset",
)

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


class cn_pose_dataset(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)
        image_dir = dl_manager.download_and_extract(
            IMAGE_URL
        )
        pose_dir = dl_manager.download_and_extract(
            POSE_URL
        )

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

    def _generate_examples(self, metadata_path, image_dir, pose_dir):
        metadata = pd.read_json(metadata_path, lines=True)

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

            image_path = row["image"]
            image_path = os.path.join(image_dir, image_path)
            image = open(image_path, "rb").read()

            pose_path = row["pose"]
            pose_path = os.path.join(
                pose_dir, row["pose"]
            )
            pose = open(pose_path, "rb").read()

            yield row["image"], {
                "prompt": prompt,
                "image": {
                    "path": image_path,
                    "bytes": image,
                },
                "pose": {
                    "path": pose_path,
                    "bytes": pose,
                },
            }