cifar100_n

  • Mô tả :

Phiên bản CIFAR-100 được dán nhãn lại với các lỗi chú thích thực của con người. Đối với mỗi cặp (hình ảnh, nhãn) trong tập hợp CIFAR-100 ban đầu, nó cung cấp một nhãn bổ sung do người chú thích thực sự đưa ra.

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

  • Mã nguồn : tfds.image_classification.cifar100_n.Cifar100N

  • Phiên bản :

    • 1.0.0 : Bản phát hành đầu tiên.
    • 1.0.1 (mặc định): Đã sửa lỗi tương ứng giữa chú thích và hình ảnh.
  • Kích thước tải xuống : 160.71 MiB

  • Kích thước tập dữ liệu : 136.07 MiB

  • Hướng dẫn tải xuống thủ công : Bộ dữ liệu này yêu cầu bạn tải xuống dữ liệu nguồn theo cách thủ công vào download_config.manual_dir (mặc định là ~/tensorflow_datasets/downloads/manual/ ):
    Tải xuống 'side_info_cifar100N.csv', 'CIFAR-100_human_ordered.npy' và 'image_order_c100.npy' từ https://github.com/UCSC-REAL/cifar-10-100n

Sau đó, chuyển đổi 'CIFAR-100_human_ordered.npy' thành tệp CSV 'CIFAR-100_human_annotations.csv'. Điều này có thể được thực hiện với đoạn mã sau:

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)
  • Tự động lưu vào bộ đệm ( tài liệu ): Có

  • Chia tách :

Tách ra ví dụ
'test' 10.000
'train' 50.000
  • Cấu trúc tính năng :
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,
})
  • Tài liệu tính năng :
Tính năng Lớp Hình dạng Dtype Sự mô tả
Tính năngDict
thô_nhãn LớpNhãn int64
Tôi Chữ sợi dây
hình ảnh Hình ảnh (32, 32, 3) uint8
nhãn mác LớpNhãn int64
noise_label LớpNhãn int64
worker_id tenxơ int64
worker_time tenxơ phao32

Hình dung

  • trích dẫn :
@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}
}