cifar100_n

  • Descrizione :

Una versione rietichettata di CIFAR-100 con veri errori di annotazione umana. Per ogni coppia (immagine, etichetta) nel treno CIFAR-100 originale, fornisce un'etichetta aggiuntiva fornita da un vero annotatore umano.

Quindi converti "CIFAR-100_human_ordered.npy" in un file CSV "CIFAR-100_human_annotations.csv". Questo può essere fatto con il seguente codice:

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)
Diviso Esempi
'test' 10.000
'train' 50.000
  • Struttura delle caratteristiche :
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,
})
  • Documentazione delle funzionalità :
Caratteristica Classe Forma Tipo D Descrizione
CaratteristicheDict
etichetta_grossolana ClassLabel int64
id Testo corda
Immagine Immagine (32, 32, 3) uint8
etichetta ClassLabel int64
etichetta_rumore ClassLabel int64
lavoratore_id Tensore int64
lavoratore_tempo Tensore galleggiante32

Visualizzazione

  • Citazione :
@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}
}