MNIST-M

Overview

MNIST-M was introduced by Ganin & Lempitsky (2015) as the target domain for unsupervised domain adaptation from MNIST. Each image is an MNIST digit blended over a random color patch cropped from the BSDS500 photo collection, using the absolute difference between the digit and the patch. The result keeps MNIST’s labels but has natural-image color and texture statistics, so models trained only on MNIST transfer poorly to it (57.5% accuracy in the original paper, against 98.9% for a model trained on MNIST-M itself).

Train: 59001 images
Test: 9001 images

This builder uses the original release by the authors, linked from Yaroslav Ganin’s homepage. Other copies of MNIST-M circulating online (for example the 60000/10000-image, 28x28 versions regenerated with keras_mnistm) are different data, so their sizes and results do not match this release.

Data Structure

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

Key	Type	Description
`image`	`PIL.Image.Image`	32×32×3 RGB image
`label`	int	Digit label (0-9)

Usage Example

Basic Usage

from stable_datasets.images.mnist_m import MNISTM

# First run will download + prepare cache, then return the split
ds = MNISTM(split="train")

# If you omit the split (split=None), you get a StableDatasetDict with all available splits
ds_all = MNISTM(split=None)

sample = ds[0]
print(sample.keys())  # {"image", "label"}

# Optional: make it PyTorch-friendly
ds_torch = ds.with_format("torch")

References

Official website: https://yaroslav.ganin.net/ (the “MNIST-M” link under Unsupervised Domain Adaptation by Backpropagation)
Download: mnist_m.tar.gz from Google Drive (MD5 859df31c91afe82e80e5012ba928f279; no checksum is published with the release)
License: Unknown. The release declares no license. The data is derived from MNIST and BSDS500.

Citation

@inproceedings{ganin2015unsupervised,
    title     = {Unsupervised Domain Adaptation by Backpropagation},
    author    = {Ganin, Yaroslav and Lempitsky, Victor},
    booktitle = {Proceedings of the 32nd International Conference on Machine Learning},
    pages     = {1180--1189},
    year      = {2015}}

@article{ganin2016domain,
    title     = {Domain-Adversarial Training of Neural Networks},
    author    = {Ganin, Yaroslav and Ustinova, Evgeniya and Ajakan, Hana and Germain, Pascal and Larochelle, Hugo and Laviolette, Fran{\c{c}}ois and Marchand, Mario and Lempitsky, Victor},
    journal   = {Journal of Machine Learning Research},
    volume    = {17},
    number    = {59},
    pages     = {1--35},
    year      = {2016}}