ซิฟาร์10_h

  • คำอธิบาย :

ชุดการทดสอบของ CIFAR-10 ที่ได้รับการเปลี่ยนชื่อใหม่ โดยมีฉลากแบบอ่อนที่มาจากผู้อธิบายประกอบที่เป็นมนุษย์จริงๆ สำหรับทุกคู่ (รูปภาพ, ป้ายกำกับ) ในชุดการทดสอบ CIFAR-10 ดั้งเดิม จะมีป้ายกำกับเพิ่มเติมหลายฉลากที่มอบให้โดยผู้อธิบายประกอบที่เป็นมนุษย์จริง เช่นเดียวกับฉลากแบบอ่อนโดยเฉลี่ย ชุดการฝึกจะเหมือนกับชุดข้อมูลต้นฉบับชุดหนึ่ง

แยก ตัวอย่าง
'test' 10,000
'train' 50,000
  • โครงสร้างคุณสมบัติ :
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)),
})
  • เอกสารคุณสมบัติ :
คุณสมบัติ ระดับ รูปร่าง ประเภทD คำอธิบาย
คุณสมบัติDict
annotator_ids ลำดับ(สเกลาร์) (ไม่มี,) int32
human_labels ลำดับ(ClassLabel) (ไม่มี,) int64
รหัส ข้อความ เชือก
ภาพ ภาพ (32, 32, 3) uint8
ฉลาก ClassLabel int64
ปฏิกิริยา_ครั้ง ลำดับ(สเกลาร์) (ไม่มี,) ลอย32
soft_label เทนเซอร์ (10,) ลอย32
ดัชนีทดลอง ลำดับ(สเกลาร์) (ไม่มี,) int32

การแสดงภาพ

  • การอ้างอิง :
@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}
}