Dataset Card for FEMNIST

The FEMNIST dataset is a part of the LEAF benchmark. It represents image classification of handwritten digits, lower and uppercase letters, giving 62 unique labels.

Dataset Details

Dataset Description

Each sample is comprised of a (28x28) grayscale image, writer_id, hsf_id, and character.

• Curated by: LEAF
• License: BSD 2-Clause License

Dataset Sources

The FEMNIST is a preprocessed (in a way that resembles preprocessing for MNIST) version of NIST SD 19.

Uses

This dataset is intended to be used in Federated Learning settings.

Direct Use

We recommend using Flower Dataset (flwr-datasets) and Flower (flwr).

To partition the dataset, do the following.

1. Install the package.

pip install flwr-datasets[vision]

2. Use the HF Dataset under the hood in Flower Datasets.

from flwr_datasets import FederatedDataset
from flwr_datasets.partitioner import NaturalIdPartitioner

fds = FederatedDataset(
    dataset="flwrlabs/femnist",
    partitioners={"train": NaturalIdPartitioner(partition_by="writer_id")}
)
partition = fds.load_partition(partition_id=0)

Dataset Structure

The whole dataset is kept in the train split. If you want to leave out some part of the dataset for centralized evaluation, use Resplitter. (The full example is coming soon here) Dataset fields:

• image: grayscale of size (28, 28), PIL Image,
• writer_id: string, unique value per each writer,
• hsf_id: string, corresponds to the way that the data was collected (see more details here,
• character: ClassLabel (it means it's int if you access it in the dataset, but you can convert it to the original value by femnist["train"].features["character"].int2str(value).

Dataset Creation

Curation Rationale

This dataset was created as a part of the LEAF benchmark. We make it available in the HuggingFace Hub to facilitate its seamless use in FlowerDatasets.

Source Data

NIST SD 19

Data Collection and Processing

For the preprocessing details, please refer to the original paper, the source code and NIST SD 19

Who are the source data producers?

For the preprocessing details, please refer to the original paper, the source code and NIST SD 19

Citation

When working on the LEAF benchmark, please cite the original paper. If you're using this dataset with Flower Datasets, you can cite Flower.

BibTeX:

@article{DBLP:journals/corr/abs-1812-01097,
  author       = {Sebastian Caldas and
                  Peter Wu and
                  Tian Li and
                  Jakub Kone{\v{c}}n{\'y} and
                  H. Brendan McMahan and
                  Virginia Smith and
                  Ameet Talwalkar},
  title        = {{LEAF:} {A} Benchmark for Federated Settings},
  journal      = {CoRR},
  volume       = {abs/1812.01097},
  year         = {2018},
  url          = {http://arxiv.org/abs/1812.01097},
  eprinttype    = {arXiv},
  eprint       = {1812.01097},
  timestamp    = {Wed, 23 Dec 2020 09:35:18 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-1812-01097.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

@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

In case of any doubts, please contact Flower Labs.