cifar10_h

  • 説明:

CIFAR-10 のテスト セットのラベルを変更したバージョンで、実際の人間のアノテーターによるソフトラベルが付けられています。元の CIFAR-10 テスト セット内のすべてのペア (画像、ラベル) に対して、平均的なソフト ラベルに加えて、実際の人間のアノテーターによって与えられたいくつかの追加ラベルが提供されます。トレーニング セットは元のデータセットと同一です。

スプリット
'test' 10,000
'train' 50,000
  • 機能の構造:
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)),
})
  • 機能ドキュメント:
特徴クラスDタイプ説明
特徴辞書
annotator_idsシーケンス(スカラー) (なし、) int32
人間のラベルシーケンス(クラスラベル) (なし、) int64
ID文章
画像画像(32、32、3) uint8
ラベルクラスラベルint64
反応時間シーケンス(スカラー) (なし、) float32
ソフトラベルテンソル(10,) float32
トライアルインデックスシーケンス(スカラー) (なし、) int32

視覚化

  • 引用
@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}
}