cifar100_n

  • 説明:

CIFAR-100 の再ラベル付けされたバージョンで、実際の人間の注釈エラーがあります。元の CIFAR-100 トレイン セットのすべてのペア (画像、ラベル) に対して、実際の人間のアノテーターによって付与された追加のラベルが提供されます。

  • ホームページ: https://www.cs.toronto.edu/~kriz/cifar.html

  • ソース コード: tfds.image_classification.cifar100_n.Cifar100N

  • バージョン:

    • 1.0.0 : 初期リリース。
    • 1.0.1 (デフォルト): 注釈と画像の対応を修正しました。
  • ダウンロードサイズ: 160.71 MiB

  • データセットサイズ: 136.07 MiB

  • 手動ダウンロードの手順: このデータセットでは、ソース データを手動でdownload_config.manual_dir (デフォルトは~/tensorflow_datasets/downloads/manual/ ) にダウンロードする必要があります。
    https://github.com/UCSC-REAL/cifar-10-100nから「side_info_cifar100N.csv」、「CIFAR-100_human_ordered.npy」、「image_order_c100.npy」をダウンロードします

次に、「CIFAR-100_human_ordered.npy」を CSV ファイル「CIFAR-100_human_annotations.csv」に変換します。これは、次のコードで実行できます。

import numpy as np
from tensorflow_datasets.core.utils.lazy_imports_utils import pandas as pd
from tensorflow_datasets.core.utils.lazy_imports_utils import tensorflow as tf

human_labels_np_path = '<local_path>/CIFAR-100_human_ordered.npy'
human_labels_csv_path = '<local_path>/CIFAR-100_human_annotations.csv'

with tf.io.gfile.GFile(human_labels_np_path, "rb") as f:
  human_annotations = np.load(f, allow_pickle=True)

df = pd.DataFrame(human_annotations[()])

with tf.io.gfile.GFile(human_labels_csv_path, "w") as f:
  df.to_csv(f, index=False)
スプリット
'test' 10,000
'train' 50,000
  • 機能構造:
FeaturesDict({
    'coarse_label': ClassLabel(shape=(), dtype=int64, num_classes=20),
    'id': Text(shape=(), dtype=string),
    'image': Image(shape=(32, 32, 3), dtype=uint8),
    'label': ClassLabel(shape=(), dtype=int64, num_classes=100),
    'noise_label': ClassLabel(shape=(), dtype=int64, num_classes=100),
    'worker_id': int64,
    'worker_time': float32,
})
  • 機能のドキュメント:
特徴クラスDtype説明
特徴辞書
粗いラベルクラスラベルint64
ID文章ストリング
画像画像(32, 32, 3) uint8
ラベルクラスラベルint64
ノイズラベルクラスラベルint64
worker_idテンソルint64
worker_timeテンソルfloat32

視覚化

  • 引用
@inproceedings{wei2022learning,
  title={Learning with Noisy Labels Revisited: A Study Using Real-World Human
  Annotations},
  author={Jiaheng Wei and Zhaowei Zhu and Hao Cheng and Tongliang Liu and Gang
  Niu and Yang Liu},
  booktitle={International Conference on Learning Representations},
  year={2022},
  url={https://openreview.net/forum?id=TBWA6PLJZQm}
}