Kuzushiji-49
Overview
The Kuzushiji-49 dataset is an image classification benchmark of 270,912 28×28 grayscale images of cursive Japanese (Kuzushiji) hiragana characters across 49 classes: the 46 modern hiragana, two obsolete hiragana (ゐ “wi” and ゑ “we”), and the iteration mark ゝ. It is a larger and harder companion to Kuzushiji-MNIST.
Train: 232,365 images
Test: 38,547 images
Unlike Kuzushiji-MNIST, the classes are highly imbalanced: the training split has between 392 and 6,000 images per class. Balanced (macro-averaged) accuracy is therefore more informative than plain accuracy.
Data Structure
When accessing an example using ds[i], you will receive a dictionary with the following keys:
Key |
Type |
Description |
|---|---|---|
|
|
28×28 grayscale image |
|
int |
Class label (0-48) |
Class names are romanized readings in the order of the official class map (k49_classmap.csv): a, i, u, …, wo, n, and iteration_mark for label 48.
Usage Example
Basic Usage
from stable_datasets.images.k_mnist49 import KMNIST49
# First run will download + prepare cache, then return the split as a HF Dataset
ds = KMNIST49(split="train")
# If you omit the split (split=None), you get a DatasetDict with all available splits
ds_all = KMNIST49(split=None)
sample = ds[0]
print(sample.keys()) # {"image", "label"}
# Map a label back to its class name
print(ds.info.features["label"].names[sample["label"]])
# Optional: make it PyTorch-friendly
ds_torch = ds.with_format("torch")
References
Official website: http://codh.rois.ac.jp/kmnist/
License: CC BY-SA 4.0
Attribution: “KMNIST Dataset” (created by CODH), adapted from “Kuzushiji Dataset” (created by NIJL and others), doi:10.20676/00000341
Citation
@online{clanuwat2018deep,
author = {Tarin Clanuwat and Mikel Bober-Irizar and Asanobu Kitamoto and Alex Lamb and Kazuaki Yamamoto and David Ha},
title = {Deep Learning for Classical Japanese Literature},
date = {2018-12-03},
year = {2018},
eprintclass = {cs.CV},
eprinttype = {arXiv},
eprint = {cs.CV/1812.01718}
}