flwrlabs/cinic10

---
license: cc-by-4.0
dataset_info:
features:
- name: image
dtype: image
- name: label
dtype:
class_label:
names:
'0': airplane
'1': automobile
'2': bird
'3': cat
'4': deer
'5': dog
'6': frog
'7': horse
'8': ship
'9': truck
splits:
- name: train
num_bytes: 178662714
num_examples: 90000
- name: validation
num_bytes: 180126542
num_examples: 90000
- name: test
num_bytes: 178913694
num_examples: 90000
download_size: 771149160
dataset_size: 537702950
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
task_categories:
- image-classification
size_categories:
- 100K<n<1M
---
# Dataset Card for CINIC-10

CINIC-10 has a total of 270,000 images equally split amongst three subsets: train, validate, and test. This means that CINIC-10 has 4.5 times as many samples than CIFAR-10.

## Dataset Details

In each subset (90,000 images), there are ten classes (identical to [CIFAR-10](https://www.cs.toronto.edu/~kriz/cifar.html) classes). There are 9000 images per class per subset. Using the suggested data split (an equal three-way split), CINIC-10 has 1.8 times as many training samples as in CIFAR-10. CINIC-10 is designed to be directly swappable with CIFAR-10.
To understand the motivation behind the dataset creation please visit the [GitHub repository](https://github.com/BayesWatch/cinic-10 ).

### Dataset Sources

- **Repository:** https://github.com/BayesWatch/cinic-10
- **Paper:** https://arxiv.org/abs/1810.03505
- **Dataset:** http://dx.doi.org/10.7488/ds/2448
- **Benchmarking, Papers with code:** https://paperswithcode.com/sota/image-classification-on-cinic-10

## Use in FL

In order to prepare the dataset for the FL settings, we recommend using [Flower Dataset](https://flower.ai/docs/datasets/) (flwr-datasets) for the dataset download and partitioning and [Flower](https://flower.ai/docs/framework/) (flwr) for conducting FL experiments.

To partition the dataset, do the following.
1. Install the package.
```bash
pip install flwr-datasets[vision]
```
2. Use the HF Dataset under the hood in Flower Datasets.
```python
from flwr_datasets import FederatedDataset
from flwr_datasets.partitioner import IidPartitioner

fds = FederatedDataset(
dataset="flwrlabs/cinic10",
partitioners={"train": IidPartitioner(num_partitions=10)}
)
partition = fds.load_partition(partition_id=0)
```

## Dataset Structure

### Data Instances
The first instance of the train split is presented below:
```
{
'image': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=32x32>,
'label': 0
}
```
### Data Split

```
DatasetDict({
train: Dataset({
features: ['image', 'label'],
num_rows: 90000
})
validation: Dataset({
features: ['image', 'label'],
num_rows: 90000
})
test: Dataset({
features: ['image', 'label'],
num_rows: 90000
})
})
```

## Citation

When working with the CINIC-10 dataset, please cite the original paper.
If you're using this dataset with Flower Datasets and Flower, cite Flower.

**BibTeX:**

Original paper:
```
@misc{darlow2018cinic10imagenetcifar10,
title={CINIC-10 is not ImageNet or CIFAR-10},
author={Luke N. Darlow and Elliot J. Crowley and Antreas Antoniou and Amos J. Storkey},
year={2018},
eprint={1810.03505},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/1810.03505},
}
````

Flower:

```
@article{DBLP:journals/corr/abs-2007-14390,
author = {Daniel J. Beutel and
Taner Topal and
Akhil Mathur and
Xinchi Qiu and
Titouan Parcollet and
Nicholas D. Lane},
title = {Flower: {A} Friendly Federated Learning Research Framework},
journal = {CoRR},
volume = {abs/2007.14390},
year = {2020},
url = {https://arxiv.org/abs/2007.14390},
eprinttype = {arXiv},
eprint = {2007.14390},
timestamp = {Mon, 03 Aug 2020 14:32:13 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2007-14390.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
```

## Dataset Card Contact

If you have any questions about the dataset preprocessing and preparation, please contact [Flower Labs](https://flower.ai/).