This repository contains the revised split protocol for creating splits for few shot domain adaptation on the MNIST-USPS datasets.

Contrary to often seen splits, we define an independent test split here and only let the train-val split vary according to a user-defined random seed.

pip install mnist-usps

Getting the splits is a simple as:

from mnistusps import mnistusps

train, val, test = mnistusps(
    source_name = "mnist",
    target_name = "usps",
    seed=1,
    num_source_per_class=200,
    num_target_per_class=3,
    same_to_diff_class_ratio=3,
    image_resize=(240, 240),
    group_in_out=True, # groups data: ((img_s, img_t), (lbl_s, _lbl_t))
    framework_conversion="tensorflow",
    data_path = None, # downloads to "~/data" per default
)

The function automatically downloads and unpacks the data using Torchvision internally. It then creates the splits using the Dataset Ops library. Depending on your choice of machine learning library, the dataset can be converted to Tensorflow or PyTorch (assuming either is pre-installed) using Dataset Ops.