cifar10_n

  • Descriptif :

Une version réétiquetée de CIFAR-10 avec de vraies erreurs d'annotation humaine. Pour chaque paire (image, étiquette) du train CIFAR-10 original, il fournit plusieurs étiquettes supplémentaires données par de vrais annotateurs humains.

  • Page d' accueil : https://ucsc-real.soe.ucsc.edu:1995/Home.html/

  • Code source : tfds.image_classification.cifar10_n.Cifar10N

  • Versions :

    • 1.0.0 : Version initiale.
    • 1.0.1 : Correction d'une faute de frappe dans la clé worse_label .
    • 1.0.2 : Correction de la correspondance entre les annotations et les images.
    • 1.0.3 : Correction des fichiers dans MANUAL_DIR .
    • 1.0.4 (par défaut) : Correction du chargement des informations secondaires.
  • Taille du téléchargement : 162.17 MiB

  • Taille du jeu de données : 147.91 MiB

  • Instructions de téléchargement manuel : cet ensemble de données nécessite que vous téléchargiez manuellement les données sources dans download_config.manual_dir (par défaut ~/tensorflow_datasets/downloads/manual/ ) :
    Téléchargez 'side_info_cifar10N.csv', 'CIFAR-10_human_ordered.npy' et 'image_order_c10.npy' depuis https://github.com/UCSC-REAL/cifar-10-100n

Ensuite, convertissez 'CIFAR-10_human_ordered.npy' en un fichier CSV 'CIFAR-10_human_annotations.csv'. Cela peut être fait avec le code suivant :

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-10_human_ordered.npy'
human_labels_csv_path = '<local_path>/CIFAR-10_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)
  • Mise en cache automatique ( documentation ): Oui

  • Fractionnements :

Diviser Exemples
'test' 10 000
'train' 50 000
  • Structure des fonctionnalités :
FeaturesDict({
    'aggre_label': 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),
    'random_label1': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'random_label2': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'random_label3': ClassLabel(shape=(), dtype=int64, num_classes=10),
    'worker1_id': int64,
    'worker1_time': float32,
    'worker2_id': int64,
    'worker2_time': float32,
    'worker3_id': int64,
    'worker3_time': float32,
    'worse_label': ClassLabel(shape=(), dtype=int64, num_classes=10),
})
  • Documentation des fonctionnalités :
Caractéristique Classer Forme Dtype La description
FonctionnalitésDict
aggre_label Étiquette de classe int64
identifiant Texte chaîne de caractères
image Image (32, 32, 3) uint8
étiquette Étiquette de classe int64
random_label1 Étiquette de classe int64
random_label2 Étiquette de classe int64
random_label3 Étiquette de classe int64
worker1_id Tenseur int64
worker1_time Tenseur float32
worker2_id Tenseur int64
worker2_time Tenseur float32
worker3_id Tenseur int64
worker3_time Tenseur float32
pire_étiquette Étiquette de classe int64

Visualisation

  • Citation :
@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}
}