wlasl.py

# Copyright 2023 Thinh T. Duong
import os
import datasets
import pandas as pd
from glob import glob


logger = datasets.logging.get_logger(__name__)


_CITATION = """
"""
_DESCRIPTION = """
"""
_HOMEPAGE = "https://github.com/dxli94/WLASL"
_REPO_URL = "https://huggingface.co/datasets/VieSignLang/wlasl/resolve/main"
_URLS = {
    "meta": f"{_REPO_URL}/WLASL_v0.3.json",
    "videos": f"{_REPO_URL}/start_kit/videos/*.zip",
}


class WLASLConfig(datasets.BuilderConfig):
    """WLASL configuration."""

    def __init__(self, name, **kwargs):
        """
        Parameters
        ----------
        name : str
            Name of subset.
        kwargs : dict
            Keyword arguments.
        """
        super(WLASLConfig, self).__init__(
            name=name,
            version=datasets.Version("1.0.0"),
            description=_DESCRIPTION,
            **kwargs,
        )


class WLASL(datasets.GeneratorBasedBuilder):
    """WLASL dataset."""
    BUILDER_CONFIGS = [
        WLASLConfig(name="rgb_videos"),
    ]
    DEFAULT_CONFIG_NAME = "rgb_videos"

    def _info(self) -> datasets.DatasetInfo:
        features = datasets.Features({
            "gloss": datasets.Value("string"),
            "bbox": datasets.Sequence(datasets.Value("int16")),
            "fps": datasets.Value("int8"),
            "frame_end": datasets.Value("int32"),
            "frame_start": datasets.Value("int32"),
            "instance_id": datasets.Value("int32"),
            "signer_id": datasets.Value("int32"),
            "source": datasets.Value("string"),
            "url": datasets.Value("string"),
            "variation_id": datasets.Value("int8"),
            "video_id": datasets.Value("int32"),
            "video": datasets.Value("string"),
        })

        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            homepage=_HOMEPAGE,
            citation=_CITATION,
        )

    def _split_generators(
        self, dl_manager: datasets.DownloadManager
    ) -> list[datasets.SplitGenerator]:
        """
        Get splits.

        Parameters
        ----------
        dl_manager : datasets.DownloadManager
            Download manager.

        Returns
        -------
        list[datasets.SplitGenerator]
            Split generators.
        """
        metadata_path = dl_manager.download(_URLS["meta"])
        raw_df = pd.read_json(metadata_path)
        exploded_df = raw_df.explode("instances")
        df = pd.concat(
            [
                exploded_df[["gloss"]].reset_index(drop=True),
                pd.json_normalize(exploded_df.instances)
            ],
            axis=1,
        )
        split_dict = {
            datasets.Split.TRAIN: df[df.split == "train"].drop(columns=["split"]),
            datasets.Split.VALIDATION: df[df.split == "val"].drop(columns=["split"]),
            datasets.Split.TEST: df[df.split == "test"].drop(columns=["split"]),
        }

        video_dirs = dl_manager.download_and_extract(glob(_URLS["videos"]))

        return [
            datasets.SplitGenerator(
                name=name,
                gen_kwargs={
                    "split_df": split_df,
                    "video_dirs": video_dirs,
                },
            )
            for name, split_df in split_dict.items()
        ]

    def _generate_examples(
        self, split_df: str,
        video_dirs: list[str],
    ) -> tuple[int, dict]:
        """
        Generate examples from metadata.

        Parameters
        ----------
        split_df : str
            Split dataframe.
        video_dirs : list[str]
            List of video directories.

        Yields
        ------
        tuple[int, dict]
            Sample.
        """
        split = datasets.Dataset.from_pandas(split_df)
        for i, sample in enumerate(split):
            for video_dir in video_dirs:
                video_path = os.path.join(video_dir, sample["video_id"] + ".mp4")
                if os.path.exists(video_path):
                    yield i, {
                        "gloss": sample["gloss"],
                        "bbox": sample["bbox"],
                        "fps": sample["fps"],
                        "frame_end": sample["frame_end"],
                        "frame_start": sample["frame_start"],
                        "instance_id": sample["instance_id"],
                        "signer_id": sample["signer_id"],
                        "source": sample["source"],
                        "url": sample["url"],
                        "variation_id": sample["variation_id"],
                        "video_id": sample["video_id"],
                        "video": video_path,
                    }