cifar10_h

  • 설명 :

실제 인간 주석자가 제공한 소프트 라벨을 사용하여 CIFAR-10 테스트 세트의 라벨이 다시 지정된 버전입니다. 원본 CIFAR-10 테스트 세트의 모든 쌍(이미지, 레이블)에 대해 실제 인간 주석자가 제공한 여러 추가 레이블과 평균 소프트 레이블을 제공합니다. 훈련 세트는 원본 데이터세트와 동일합니다.

나뉘다
'test' 10,000
'train' 50,000
  • 기능 구조 :
FeaturesDict({
    'annotator_ids': Sequence(Scalar(shape=(), dtype=int32)),
    'human_labels': Sequence(ClassLabel(shape=(), dtype=int64, num_classes=10)),
    'id': Text(shape=(), dtype=string),
    'image': Image(shape=(32, 32, 3), dtype=uint8),
    'label': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'reaction_times': Sequence(Scalar(shape=(), dtype=float32)),
    'soft_label': Tensor(shape=(10,), dtype=float32),
    'trial_indices': Sequence(Scalar(shape=(), dtype=int32)),
})
  • 기능 문서 :
특징 수업 모양 Dtype 설명
특징Dict
annotator_ids 시퀀스(스칼라) (없음,) 정수32
인간_라벨 시퀀스(클래스 라벨) (없음,) 정수64
ID 텍스트
영상 영상 (32, 32, 3) uint8
상표 클래스 라벨 정수64
반응_시간 시퀀스(스칼라) (없음,) float32
소프트 라벨 텐서 (10,) float32
trial_indices 시퀀스(스칼라) (없음,) 정수32

심상

  • 인용 :
@inproceedings{wei2022learning,
  title={Human uncertainty makes classification more robust},
  author={Joshua C. Peterson and Ruairidh M. Battleday and Thomas L. Griffiths
  and Olga Russakovsky},
  booktitle={IEEE International Conference on Computer Vision and Pattern
  Recognition (CVPR)},
  year={2019}
}