Skip to content

Latest commit

 

History

History
44 lines (34 loc) · 3.72 KB

zeroshot_dataset_en.md

File metadata and controls

44 lines (34 loc) · 3.72 KB
Raw

中文说明 | English

Zero-shot Image Classification Datasets

The collection of dataset is the Chinese version of the Image Classification in the Wild in the ELEVATER Benchmark. It consists of 20 datasets, including Caltech-101, CIFAR-10, CIFAR-100, MNIST, etc. We provide our organized datasets, which enable direct usage of our codes on the datasets.

Download link: Click here

For the ImageNet data, please visit the official website (link). You can refer to this doc to prepare the validation set into ImageFolder format. This project only provides the label names in Chinese and English.

Notes

We have organized 20 datasets into 20 directories, and zipped and uploaded them. Users can click the link above to download all the datasets. After unzipping the ELEVATER_all.zip, you will get the zipped files of each dataset in ELEVATER. Once again after unzipping the dataset zipfile, you will get the dataset directory with the following folder structure:

${dataset_name}
├── index.json  # Some datasets contain this file，which only serves for the submission to the ELEVATER benchmark
├── label_cn.txt  # File of Chinese labels，where the text in each line refers to the label name
├── label.txt  # File of English labels，where the text in each line refers to the label name
├── test/
│   ├── 000/
│   ├── 001/
│   └── 002/
└── train/
    ├── 000/
    ├── 001/
    └── 002/

${dataset_name} refers to the directory of each dataset, where there are two directories named train and test. Each directory contains sub-directories named with id, which refers to a category. Additionally, there are 2 files, namely the Chinese label filelabel_cn.txt, the English label file label.txt. Note that:

  • When the number of labels is no larger than 10, e,g., 10, the ids are [0-9]
  • When the number of labels is larger than 10, e.g., 100, the ids are [000-099]，which are left padded with 0 to 3-digit numbers. This serves for alphabetic order.
  • Each id refers to the label name in the ${id}-th line of the label file (0-index). For example, 0 refers to the label name in the 0-th line, and 099 refers to the label name in the 99-th line

The sub-directories of training and test sets are alphatically ordered for the reason that we use torchvision.dataset, which requires to organize data to sub-directories based on their labels and alphabetically ordered.

There are two label files for Chinese and English. We only use label_cn.txt in our experiments, and label.txt is for reference only. Each line contains a label name. The example is shown below:

飞机
汽车
……

index.json only serves for the submission of the ELEVATER benchmark, and not every dataset contains this file. The existence of this file is due to the specified order of test data. To make your submission available, you need to add index.json in your command.

Similarly, if you prepare the ImageNet data, please put the label files mentioned above in the directory ${dataset_name}, and create directories like train and test, and file the images according to their categories into directories, which should be ordered by the alphabetic order, like 000-999. The organization should be consistent with the abovementioned example.