cifar100_n

  • Descripción :

Una versión reetiquetada de CIFAR-100 con errores de anotación humanos reales. Para cada par (imagen, etiqueta) en el conjunto de trenes CIFAR-100 original, proporciona una etiqueta adicional proporcionada por un anotador humano real.

  • Página de inicio: https://www.cs.toronto.edu/~kriz/cifar.html

  • Código fuente : tfds.image_classification.cifar100_n.Cifar100N

  • Versiones :

    • 1.0.0 : Versión inicial.
    • 1.0.1 (predeterminado): correspondencia fija entre anotaciones e imágenes.
  • Tamaño de la descarga : 160.71 MiB

  • Tamaño del conjunto de datos : 136.07 MiB

  • Instrucciones de descarga manual : este conjunto de datos requiere que descargue los datos de origen manualmente en download_config.manual_dir (el valor predeterminado es ~/tensorflow_datasets/downloads/manual/ ):
    Descargue 'side_info_cifar100N.csv', 'CIFAR-100_human_ordered.npy' y 'image_order_c100.npy' desde https://github.com/UCSC-REAL/cifar-10-100n

Luego convierta 'CIFAR-100_human_ordered.npy' en un archivo CSV 'CIFAR-100_human_annotations.csv'. Esto se puede hacer con el siguiente código:

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)
  • Almacenamiento automático en caché ( documentación ): Sí

  • Divisiones :

Separar Ejemplos
'test' 10,000
'train' 50,000
  • Estructura de características :
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,
})
  • Documentación de características :
Rasgo Clase Forma Tipo D Descripción
CaracterísticasDict
etiqueta_gruesa Etiqueta de clase int64
identificación Texto cuerda
imagen Imagen (32, 32, 3) uint8
etiqueta Etiqueta de clase int64
etiqueta_ruido Etiqueta de clase int64
trabajador_id Tensor int64
trabajador_tiempo Tensor flotar32

Visualización

  • Cita :
@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}
}