- Descrizione :
Una versione rietichettata del set di test CIFAR-10 con etichette soft provenienti da veri annotatori umani. Per ogni coppia (immagine, etichetta) nel set di test CIFAR-10 originale, fornisce diverse etichette aggiuntive fornite da annotatori umani reali, oltre alla soft-label media. Il set di addestramento è identico a quello del set di dati originale.
Pagina iniziale : https://github.com/jcpeterson/cifar-10h
Codice sorgente :
tfds.image_classification.cifar10_h.Cifar10HVersioni :
-
1.0.0(impostazione predefinita): versione iniziale.
-
Dimensione download :
172.92 MiBDimensione del set di dati :
144.85 MiBMemorizzazione nella cache automatica ( documentazione ): sì
Divide :
| Diviso | Esempi |
|---|---|
'test' | 10.000 |
'train' | 50.000 |
- Struttura delle caratteristiche :
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)),
})
- Documentazione delle funzionalità :
| Caratteristica | Classe | Forma | Tipo D | Descrizione |
|---|---|---|---|---|
| CaratteristicheDict | ||||
| annotator_ids | Sequenza (scalare) | (Nessuno,) | int32 | |
| etichette_umane | Sequenza(ClassLabel) | (Nessuno,) | int64 | |
| id | Testo | corda | ||
| Immagine | Immagine | (32, 32, 3) | uint8 | |
| etichetta | ClassLabel | int64 | ||
| tempi_di_reazione | Sequenza (scalare) | (Nessuno,) | float32 | |
| soft_label | Tensore | (10,) | float32 | |
| prova_indices | Sequenza (scalare) | (Nessuno,) | int32 |
Chiavi supervisionate (vedi il documento
as_supervised):NoneFigura ( tfds.show_examples ):

- Esempi ( tfds.as_dataframe ):
- Citazione :
@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}
}