README.md

metadata

dataset_info:
  features:
    - name: qid
      dtype: uint32
    - name: image
      dtype: image
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: q_lang
      dtype:
        class_label:
          names:
            '0': en
            '1': zh
    - name: img_id
      dtype: uint32
    - name: location
      dtype:
        class_label:
          names:
            '0': Abdomen
            '1': Lung
            '2': Chest_heart
            '3': Chest_lung
            '4': Brain_Tissue
            '5': Brain_Face
            '6': Brain
            '7': Neck
            '8': Chest_mediastinal
            '9': Pelvic Cavity
    - name: modality
      dtype:
        class_label:
          names:
            '0': MRI
            '1': CT
            '2': X-Ray
    - name: base_type
      dtype:
        class_label:
          names:
            '0': vqa
            '1': kvqa
    - name: answer_type
      dtype:
        class_label:
          names:
            '0': OPEN
            '1': CLOSED
    - name: content_type
      dtype:
        class_label:
          names:
            '0': Modality
            '1': Position
            '2': Organ
            '3': Size
            '4': Abnormality
            '5': Quantity
            '6': Plane
            '7': Shape
            '8': Color
            '9': KG
    - name: triple
      list: string
  splits:
    - name: test
      num_bytes: 215442470
      num_examples: 2094
    - name: train
      num_bytes: 1331237033
      num_examples: 9835
    - name: validation
      num_bytes: 195808761
      num_examples: 2099
  download_size: 1324949790
  dataset_size: 1742488264
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
license: cc-by-4.0
task_categories:
  - visual-question-answering
language:
  - en
  - zh
tags:
  - medical

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.")