dataset.py

import os
import numpy as np
import datasets

# Mô tả dataset
_DESCRIPTION = """
VOYA Vietnamese Sign Language (VSL) dataset.
Dataset gồm các chuỗi keypoints đã trích xuất bằng MediaPipe cho nhận dạng ngôn ngữ ký hiệu.
Mỗi sample có shape (60, 1605), lưu trong 'sequences', với nhãn tương ứng trong 'labels'.
"""

_HOMEPAGE = "https://huggingface.co/datasets/Kateht/VOYA_VSL"
_LICENSE = "MIT"
_CITATION = """
@misc{voya_vsl_2025,
  author       = {Kateht et al.},
  title        = {VOYA Vietnamese Sign Language Dataset},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\\url{https://huggingface.co/datasets/Kateht/VOYA_VSL}}
}
"""


class VOYAVSLConfig(datasets.BuilderConfig):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)


class VOYAVSL(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        VOYAVSLConfig(
            name="default",
            version=datasets.Version("1.0.0"),
            description="VOYA Vietnamese Sign Language dataset",
        )
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "sequences": datasets.Array2D(
                        shape=(60, 1605), dtype="float32"
                    ),
                    "labels": datasets.Value("int32"),
                }
            ),
            supervised_keys=("sequences", "labels"),
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        # Tải cả folder Merged thay vì 1 file
        data_dir = dl_manager.download_and_extract(
            "https://huggingface.co/datasets/Kateht/VOYA_VSL/resolve/main/Merged"
        )
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,  # bạn sẽ tự chia train/val/test sau
                gen_kwargs={"data_dir": data_dir},
            ),
        ]

    def _generate_examples(self, data_dir):
        idx = 0
        for fname in sorted(os.listdir(data_dir)):
            if not fname.endswith(".npz"):
                continue
            fpath = os.path.join(data_dir, fname)
            data = np.load(fpath)

            sequences, labels = data["sequences"], data["labels"]
            for seq, label in zip(sequences, labels):
                yield idx, {
                    "sequences": seq.astype("float32"),
                    "labels": int(label),
                }
                idx += 1