import numpy as np
from PIL import Image as PILImage
from stable_datasets.schema import (
ClassLabel,
DatasetInfo,
DatasetSource,
DownloadInfo,
Features,
Image,
Version,
)
from stable_datasets.utils import BaseDatasetBuilder
[docs]
class Optdigits(BaseDatasetBuilder):
"""Build the UCI Optical Recognition of Handwritten Digits dataset.
The dataset contains 5,620 grayscale 8x8 handwritten digit samples
across 10 classes.
It preserves the official UCI split of 3,823 training samples and
1,797 test samples. Original pixel intensities range from 0 to 16
and are converted to 8-bit grayscale images.
"""
VERSION = Version("1.0.0")
SOURCE = DatasetSource(
license="CC BY 4.0",
homepage="https://archive.ics.uci.edu/dataset/80/optical+recognition+of+handwritten+digits",
citation="""@misc{uci_optdigits,
title={Optical Recognition of Handwritten Digits},
author={Alpaydin, E. and Kaynak, C.},
year={1998},
publisher={UCI Machine Learning Repository}
}""",
assets={
"train": DownloadInfo(
url=("https://archive.ics.uci.edu/ml/machine-learning-databases/optdigits/optdigits.tra")
),
"test": DownloadInfo(
url=("https://archive.ics.uci.edu/ml/machine-learning-databases/optdigits/optdigits.tes")
),
},
)
def _info(self):
return DatasetInfo(
description=(
"The UCI Optical Recognition of Handwritten Digits "
"dataset contains 5,620 samples across 10 classes. "
"Each sample consists of 64 intensity features "
"representing an 8x8 grayscale image. "
"The original training and test splits contain "
"3,823 and 1,797 samples, respectively."
),
features=Features(
{
"image": Image(),
"label": ClassLabel(num_classes=10),
}
),
supervised_keys=("image", "label"),
homepage=self.SOURCE["homepage"],
citation=self.SOURCE["citation"],
)
def _generate_examples(self, data_path, split):
"""Generate 8x8 grayscale images and digit labels."""
data = np.loadtxt(
data_path,
delimiter=",",
dtype=np.int32,
)
if data.ndim != 2 or data.shape[1] != 65:
raise ValueError(f"Unexpected Optdigits data shape: {data.shape}")
features = data[:, :-1]
labels = data[:, -1]
if np.any((features < 0) | (features > 16)):
raise ValueError("Pixel intensities must be in [0, 16].")
if np.any((labels < 0) | (labels > 9)):
raise ValueError("Digit labels must be in [0, 9].")
for idx, (pixels, label) in enumerate(zip(features, labels)):
# Convert the original intensities to 8-bit grayscale.
image = np.rint(pixels.reshape(8, 8) * (255.0 / 16.0)).astype(np.uint8)
pil_image = PILImage.fromarray(image)
yield (
idx,
{
"image": pil_image,
"label": int(label),
},
)