CAIR-M3LLM/SLAKE

Fork of BoKelvin/SLAKE converted to:

1. Wrap images as binary object
2. Classify categorical information into class labels

Metadata

Name	#train	#val	#test	img#train	img#val	img#test
SLAKE	9,835	2,099	2,094	586	174	180

Conversion script

from pathlib import Path

from datasets import ClassLabel, Dataset, Features, Image, Sequence, Value

SLAKE_FEAT = {
    "qid": Value("uint32"),
    "image": Image(decode=True),
    "question": Value("string"),
    "answer": Value("string"),
    "q_lang": ClassLabel(names=["en", "zh"]),
    "img_id": Value("uint32"),
    "location": ClassLabel(
        names=[
            "Abdomen",
            "Lung",
            "Chest_heart",
            "Chest_lung",
            "Brain_Tissue",
            "Brain_Face",
            "Brain",
            "Neck",
            "Chest_mediastinal",
            "Pelvic Cavity",
        ]
    ),
    "modality": ClassLabel(names=["MRI", "CT", "X-Ray"]),
    "base_type": ClassLabel(names=["vqa", "kvqa"]),
    "answer_type": ClassLabel(names=["OPEN", "CLOSED"]),
    "content_type": ClassLabel(
        names=[
            "Modality",
            "Position",
            "Organ",
            "Size",
            "Abnormality",
            "Quantity",
            "Plane",
            "Shape",
            "Color",
            "KG",
        ]
    ),
    "triple": Sequence(Value("string")),
}


def reformat(
    jsonl_path: Path,
    upload_to: str | None = None,
    image_dir: str = "images",
):
    split = jsonl_path.stem
    d = Dataset.from_json(jsonl_path.as_posix())
    d = d.map(
        lambda e: {
            "image": {"path": f"{image_dir}/{e['img_name']}"},
        },
        num_proc=16,
        features=Features(SLAKE_FEAT),
        remove_columns=["img_name"]
    )

    if upload_to:
        d.push_to_hub(upload_to, split=split)
    else:
        print(d)
        print(d[0])
        print(f"Would upload to split={split} on the hub.")