سيفار 100_n

  • الوصف :

نسخة معاد تسميتها من CIFAR-100 بها أخطاء حقيقية في التعليقات التوضيحية البشرية. لكل زوج (صورة ، ملصق) في مجموعة قطار CIFAR-100 الأصلية ، فإنه يوفر تسمية إضافية قدمها مضيف توضيحي بشري حقيقي.

  • الصفحة الرئيسية : https://www.cs.toronto.edu/~kriz/cifar.html

  • كود المصدر : tfds.image_classification.cifar100_n.Cifar100N

  • إصدارات :

    • 1.0.0 : الإصدار الأولي.
    • 1.0.1 (افتراضي): إصلاح المراسلات بين التعليقات التوضيحية والصور.
  • حجم التحميل : 160.71 MiB

  • حجم مجموعة البيانات : 136.07 MiB

  • إرشادات التنزيل اليدوي : تتطلب مجموعة البيانات هذه تنزيل بيانات المصدر يدويًا إلى download_config.manual_dir (الإعدادات الافتراضية على ~/tensorflow_datasets/downloads/manual/ ):
    قم بتنزيل "side_info_cifar100N.csv" و "CIFAR-100_human_ordered.npy" و "image_order_c100.npy" من https://github.com/UCSC-REAL/cifar-10-100n

ثم قم بتحويل "CIFAR-100_human_ordered.npy" إلى ملف CSV "CIFAR-100_human_annotations.csv". يمكن القيام بذلك باستخدام الكود التالي:

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)
  • التخزين المؤقت التلقائي ( الوثائق ): نعم

  • الانقسامات :

انشق، مزق أمثلة
'test' 10000
'train' 50000
  • هيكل الميزة :
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,
})
  • وثائق الميزة :
ميزة فصل شكل نوع وصف
الميزات
coarse_label ClassLabel int64
هوية شخصية نص سلسلة
صورة صورة (32 ، 32 ، 3) uint8
ضع الكلمة المناسبة ClassLabel int64
الضوضاء ClassLabel int64
worker_id موتر int64
عامل_وقت موتر تعويم 32

التصور

  • الاقتباس :
@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}
}