- 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。
バージョン:
-
1.0.0(デフォルト): 初期リリース。
-
ダウンロードサイズ:
172.92 MiBデータセットのサイズ:
144.85 MiB自動キャッシュ(ドキュメント): はい
分割:
| スプリット | 例 |
|---|---|
'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 | 形 | 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 | 弦 | ||
| 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 |
監視キー(
as_superviseddocを参照):None図( tfds.show_examples ):

- 例( tfds.as_dataframe ):
- 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}
}