---
license: mit
tags:
- image-classification
- computer-vision
- imagenet-c
---

# Nano ImageNet-C (Severity 5)

This is a randomly sampled subset of the ImageNet-C dataset, containing 5,000 images exclusively from corruption **severity level 5**. It is designed for efficient testing and validation of model robustness.

这是一个从 ImageNet-C 数据集中随机抽样的子集，包含 5000 张仅来自损坏等级为 **5** 的图像。它旨在用于高效地测试和验证模型的鲁棒性。

## How to Generate / 如何生成

This dataset was generated using the `create_nano_dataset.py` script included in this repository. To ensure reproducibility, the following parameters were used:

本数据集使用此仓库中包含的 `create_nano_dataset.py` 脚本生成。为确保可复现性，生成时使用了以下参数：

- **Source Dataset / 源数据集**: The full ImageNet-C dataset is required. / 需要完整的 ImageNet-C 数据集。
- **Random Seed / 随机种子**: `7600`
- **Python Version / Python 版本**: `3.10.14`

## Dataset Structure / 数据集结构

The dataset is provided as a single `.tar` file named `nano-imagenet-c.tar` in the `webdataset` format. The internal structure preserves the original ImageNet-C hierarchy: `corruption_type/class_name/image.jpg`.

数据集以 `webdataset` 格式打包在名为 `nano-imagenet-c.tar` 的单个 `.tar` 文件中。其内部结构保留了原始 ImageNet-C 的层次结构：`corruption_type/class_name/image.jpg`。

## Citation / 引用

If you use this dataset, please cite the original ImageNet-C paper:

如果您使用此数据集，请引用原始 ImageNet-C 的论文：

```bibtex
@inproceedings{danhendrycks2019robustness,
  title={Benchmarking Neural Network Robustness to Common Corruptions and Perturbations},
  author={Dan Hendrycks and Thomas Dietterich},
  booktitle={International Conference on Learning Representations},
  year={2019},
  url={https://openreview.net/forum?id=HJz6tiCqYm},
}
```