USPS

Overview

USPS is a handwritten digit classification dataset containing 9,298 16 x 16 grayscale images across 10 classes. This loader uses the LIBSVM distribution and preserves its provided splits:

• Train: 7,291 images

• Test: 2,007 images

The source files store each image as 256 pixel features in LIBSVM format. Pixel values are converted from [-1, 1] to [0, 255] and cast to uint8 to construct grayscale images. No separate validation split is provided.

Label Convention

The LIBSVM files encode each digit d as label d + 1 (1 to 10). Labels are shifted down by 1, following torchvision’s USPS convention, so the label field equals the digit shown in the image (0-9).

Data Structure

When accessing an example using ds[i], you receive a dictionary with the following keys:

Key	Type	Description
image	PIL.Image.Image	16 x 16 grayscale image with uint8 pixel values.
label	int	Digit shown in the image (0-9).

Usage Example

Basic Usage

Save this example as a Python file and run it. The main guard supports platforms that use multiprocessing with the spawn start method.

from stable_datasets.images.usps import USPS

def main():
    # First load downloads the data and prepares the processed cache.
    # Later loads reuse the cache and return a StableDataset.
    train = USPS(split="train")
    test = USPS(split="test")

    sample = train[0]
    print(sample.keys())  # image, label
    print(sample["image"].size)  # (16, 16)
    print(sample["image"].mode)  # L
    print(sample["label"])
    print(len(train), len(test))  # 7291 2007

    # Load all available splits as a StableDatasetDict.
    all_splits = USPS(split=None)
    print(all_splits.keys())  # train, test

    # Optionally request PyTorch-formatted samples.
    train_torch = train.with_format("torch")
    print(train_torch[0]["image"].shape)

if __name__ == "__main__":
    main()

Related Datasets

• MNIST: Handwritten digit recognition with 28 x 28 grayscale images.

• EMNIST: Handwritten character recognition, including letters and digits.

• SVHN: Digit recognition using natural images of house numbers.

References

• LIBSVM dataset page

• Training data

• Test data

• torchvision USPS implementation

• License: Unknown

Neither the LIBSVM dataset page nor the original release from CEDAR (SUNY Buffalo) publishes a license or terms of use. Check with the data provider for applicable use terms, and cite the reference below when using the dataset.

Citation

The LIBSVM dataset page cites the following reference for USPS:

@article{hull1994database,
  author = {Hull, J. J.},
  title = {A Database for Handwritten Text Recognition Research},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume = {16},
  number = {5},
  pages = {550--554},
  year = {1994}
}