cifar10_h

  • توضیحات :

یک نسخه برچسب‌گذاری‌شده مجدد از مجموعه آزمایشی CIFAR-10 با برچسب‌های نرم که از حاشیه‌نویس‌های واقعی انسان تهیه شده است. برای هر جفت (تصویر، برچسب) در مجموعه آزمایشی اصلی CIFAR-10، چندین برچسب اضافی ارائه شده توسط حاشیه نویسان واقعی انسان و همچنین برچسب نرم متوسط ​​ارائه می کند. مجموعه آموزشی با مجموعه داده اصلی یکسان است.

شکاف مثال ها
'test' 10000
'train' 50000
  • ساختار ویژگی :
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)),
})
  • مستندات ویژگی :
ویژگی کلاس شکل نوع D شرح
FeaturesDict
annotator_ids دنباله (اسکالر) (هیچ یک،) int32
برچسب های انسانی دنباله (ClassLabel) (هیچ یک،) int64
شناسه متن رشته
تصویر تصویر (32، 32، 3) uint8
برچسب ClassLabel int64
واکنش_زمان ها دنباله (اسکالر) (هیچ یک،) float32
soft_label تانسور (10،) float32
شاخص های آزمایشی دنباله (اسکالر) (هیچ یک،) int32

تجسم

  • نقل قول :
@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}
}