WNJXYK/TTA-ImageNet-C

---
license: cc-by-4.0
task_categories: [image-classification]
tags: [imagenet, corruption, robustness, tta, test-time-adaptation]
pretty_name: "TTA ImageNet-C"
size_categories: [1M<n<10M]
configs:
- config_name: gaussian_noise
data_files:
- split: severity_1
path: data/gaussian_noise/severity_1/*.parquet
- split: severity_2
path: data/gaussian_noise/severity_2/*.parquet
- split: severity_3
path: data/gaussian_noise/severity_3/*.parquet
- split: severity_4
path: data/gaussian_noise/severity_4/*.parquet
- split: severity_5
path: data/gaussian_noise/severity_5/*.parquet
- config_name: shot_noise
data_files:
- split: severity_1
path: data/shot_noise/severity_1/*.parquet
- split: severity_2
path: data/shot_noise/severity_2/*.parquet
- split: severity_3
path: data/shot_noise/severity_3/*.parquet
- split: severity_4
path: data/shot_noise/severity_4/*.parquet
- split: severity_5
path: data/shot_noise/severity_5/*.parquet
- config_name: impulse_noise
data_files:
- split: severity_1
path: data/impulse_noise/severity_1/*.parquet
- split: severity_2
path: data/impulse_noise/severity_2/*.parquet
- split: severity_3
path: data/impulse_noise/severity_3/*.parquet
- split: severity_4
path: data/impulse_noise/severity_4/*.parquet
- split: severity_5
path: data/impulse_noise/severity_5/*.parquet
- config_name: defocus_blur
data_files:
- split: severity_1
path: data/defocus_blur/severity_1/*.parquet
- split: severity_2
path: data/defocus_blur/severity_2/*.parquet
- split: severity_3
path: data/defocus_blur/severity_3/*.parquet
- split: severity_4
path: data/defocus_blur/severity_4/*.parquet
- split: severity_5
path: data/defocus_blur/severity_5/*.parquet
- config_name: glass_blur
data_files:
- split: severity_1
path: data/glass_blur/severity_1/*.parquet
- split: severity_2
path: data/glass_blur/severity_2/*.parquet
- split: severity_3
path: data/glass_blur/severity_3/*.parquet
- split: severity_4
path: data/glass_blur/severity_4/*.parquet
- split: severity_5
path: data/glass_blur/severity_5/*.parquet
- config_name: motion_blur
data_files:
- split: severity_1
path: data/motion_blur/severity_1/*.parquet
- split: severity_2
path: data/motion_blur/severity_2/*.parquet
- split: severity_3
path: data/motion_blur/severity_3/*.parquet
- split: severity_4
path: data/motion_blur/severity_4/*.parquet
- split: severity_5
path: data/motion_blur/severity_5/*.parquet
- config_name: zoom_blur
data_files:
- split: severity_1
path: data/zoom_blur/severity_1/*.parquet
- split: severity_2
path: data/zoom_blur/severity_2/*.parquet
- split: severity_3
path: data/zoom_blur/severity_3/*.parquet
- split: severity_4
path: data/zoom_blur/severity_4/*.parquet
- split: severity_5
path: data/zoom_blur/severity_5/*.parquet
- config_name: snow
data_files:
- split: severity_1
path: data/snow/severity_1/*.parquet
- split: severity_2
path: data/snow/severity_2/*.parquet
- split: severity_3
path: data/snow/severity_3/*.parquet
- split: severity_4
path: data/snow/severity_4/*.parquet
- split: severity_5
path: data/snow/severity_5/*.parquet
- config_name: frost
data_files:
- split: severity_1
path: data/frost/severity_1/*.parquet
- split: severity_2
path: data/frost/severity_2/*.parquet
- split: severity_3
path: data/frost/severity_3/*.parquet
- split: severity_4
path: data/frost/severity_4/*.parquet
- split: severity_5
path: data/frost/severity_5/*.parquet
- config_name: fog
data_files:
- split: severity_1
path: data/fog/severity_1/*.parquet
- split: severity_2
path: data/fog/severity_2/*.parquet
- split: severity_3
path: data/fog/severity_3/*.parquet
- split: severity_4
path: data/fog/severity_4/*.parquet
- split: severity_5
path: data/fog/severity_5/*.parquet
- config_name: brightness
data_files:
- split: severity_1
path: data/brightness/severity_1/*.parquet
- split: severity_2
path: data/brightness/severity_2/*.parquet
- split: severity_3
path: data/brightness/severity_3/*.parquet
- split: severity_4
path: data/brightness/severity_4/*.parquet
- split: severity_5
path: data/brightness/severity_5/*.parquet
- config_name: contrast
data_files:
- split: severity_1
path: data/contrast/severity_1/*.parquet
- split: severity_2
path: data/contrast/severity_2/*.parquet
- split: severity_3
path: data/contrast/severity_3/*.parquet
- split: severity_4
path: data/contrast/severity_4/*.parquet
- split: severity_5
path: data/contrast/severity_5/*.parquet
- config_name: elastic_transform
data_files:
- split: severity_1
path: data/elastic_transform/severity_1/*.parquet
- split: severity_2
path: data/elastic_transform/severity_2/*.parquet
- split: severity_3
path: data/elastic_transform/severity_3/*.parquet
- split: severity_4
path: data/elastic_transform/severity_4/*.parquet
- split: severity_5
path: data/elastic_transform/severity_5/*.parquet
- config_name: pixelate
data_files:
- split: severity_1
path: data/pixelate/severity_1/*.parquet
- split: severity_2
path: data/pixelate/severity_2/*.parquet
- split: severity_3
path: data/pixelate/severity_3/*.parquet
- split: severity_4
path: data/pixelate/severity_4/*.parquet
- split: severity_5
path: data/pixelate/severity_5/*.parquet
- config_name: jpeg_compression
data_files:
- split: severity_1
path: data/jpeg_compression/severity_1/*.parquet
- split: severity_2
path: data/jpeg_compression/severity_2/*.parquet
- split: severity_3
path: data/jpeg_compression/severity_3/*.parquet
- split: severity_4
path: data/jpeg_compression/severity_4/*.parquet
- split: severity_5
path: data/jpeg_compression/severity_5/*.parquet
---

# TTA-ImageNet-C

Mirror of **ImageNet-C** (Hendrycks & Dietterich, ICLR 2019) with a revision
pin for reproducible test-time adaptation evaluation.

- **Upstream**: [Zenodo record 2235448](https://zenodo.org/records/2235448)
- **License**: CC BY 4.0 (matches upstream)
- **Maintained as part of**: [TTA-Evaluation-Harness](https://github.com/WNJXYK/TTA-Evaluation-Harness)

## Citation

```bibtex
@inproceedings{hendrycks2019benchmarking,
title={Benchmarking Neural Network Robustness to Common Corruptions and Perturbations},
author={Hendrycks, Dan and Dietterich, Thomas},
booktitle={ICLR},
year={2019}
}
```

## Structure

- **15 configs**: one per corruption type (`gaussian_noise`, `shot_noise`, ..., `jpeg_compression`).
- **5 splits per config**: `severity_1` through `severity_5`, 50 000 images each (1000 classes x 50).
- Labels are `ClassLabel` with 1000 WordNet-ID names in torchvision order
(lexicographic on wnid; `n01440764` = idx 0 = tench).

## Usage

```python
from datasets import load_dataset

ds = load_dataset("WNJXYK/TTA-ImageNet-C",
name="gaussian_noise",
split="severity_5",
revision="v1.0")
```

## Provenance

This mirror was built by `scripts/publish_imagenetc.py` in the
TTA-Evaluation-Harness repo. JPEG bytes are copied 1:1 from the upstream
files - no re-encoding, pixel-for-pixel identical to Hendrycks's release.