cifar100_n

  • คำอธิบาย :

CIFAR-100 เวอร์ชันที่ติดฉลากใหม่ซึ่งมีข้อผิดพลาดในการอธิบายประกอบโดยมนุษย์จริง สำหรับทุกคู่ (ภาพ, ป้ายกำกับ) ในชุดรถไฟ CIFAR-100 ดั้งเดิม จะมีป้ายกำกับเพิ่มเติมที่ได้รับจากคำอธิบายประกอบของมนุษย์จริงๆ

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

  • รหัสที่มา : tfds.image_classification.cifar100_n.Cifar100N

  • รุ่น :

    • 1.0.0 : การเปิดตัวครั้งแรก
    • 1.0.1 (ค่าเริ่มต้น): แก้ไขความสอดคล้องกันระหว่างคำอธิบายประกอบและรูปภาพ
  • ขนาดการดาวน์โหลด : 160.71 MiB

  • ขนาดชุดข้อมูล : 136.07 MiB

  • คำแนะนำในการดาวน์โหลดด้วยตนเอง : ชุดข้อมูลนี้กำหนดให้คุณต้องดาวน์โหลดแหล่งข้อมูลด้วยตนเองลงใน download_config.manual_dir (ค่าเริ่มต้นเป็น ~/tensorflow_datasets/downloads/manual/ ):
    ดาวน์โหลด 'side_info_cifar100N.csv', 'CIFAR-100_human_ordered.npy' และ 'image_order_c100.npy' จาก https://github.com/UCSC-REAL/cifar-10-100n

จากนั้นแปลง 'CIFAR-100_human_ordered.npy' เป็นไฟล์ CSV 'CIFAR-100_human_annotations.csv' สามารถทำได้ด้วยรหัสต่อไปนี้:

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)
แยก ตัวอย่าง
'test' 10,000
'train' 50,000
  • โครงสร้างคุณลักษณะ :
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,
})
  • เอกสารคุณสมบัติ :
ลักษณะเฉพาะ ระดับ รูปร่าง Dประเภท คำอธิบาย
คุณสมบัติDict
coarse_label ป้ายกำกับคลาส int64
รหัส ข้อความ สตริง
ภาพ ภาพ (32, 32, 3) uint8
ฉลาก ป้ายกำกับคลาส int64
noise_label ป้ายกำกับคลาส int64
รหัสผู้ปฏิบัติงาน เทนเซอร์ int64
คนทำงาน_เวลา เทนเซอร์ ลอย32

การสร้างภาพ

  • การอ้างอิง :
@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}
}