EMNIST-DA: A dataset for studying measurement shift

emnist_large.tar.gz contains the EMNIST-DA dataset consisting of 13 shifted versions of the 47-class extended-MNIST (EMNIST) dataset. As many methods achieve very good performance on MNIST datasets, this dataset was created to be more challenging with 47-classes and some difficult (measurement) shifts. Source code to generate the dataset is available at https://github.com/cianeastwood/bufr/blob/main/data/emnist.py.