One-Shot Data Preprocessing

The collection of raw and generated (with transformations) images from the original Omniglot Dataset.

Setup

Run the following command to initialize the following submodules.

git submodule update --init

Submodules

• brendenlake/omniglot: Omniglot data set for one-shot learning.

Usage

List of original Omniglot datasets.

• images_background_small1
• images_background_small2
• images_background
• images_evalution
• strokes_background_small1
• strokes_background_small2
• strokes_background
• strokes_evalution

To run preprocessing on a <dataset> from the list, issue the following command to generate the transformed data to <output_dataset>.

sh scripts/augment_wrapper.sh <dataset> <output_dataset>

Ensure that <dataset> and <output_dataset> do not have the same name.

References

Lake, B. M., Salakhutdinov, R., and Tenenbaum, J. B. (2015). Human-level concept learning through probabilistic program induction. Science, 350(6266), 1332-1338.

• Update: Lake, B. M., Salakhutdinov, R., and Tenenbaum, J. B. (2019). The Omniglot Challenge: A 3-Year Progress Report. Preprint available on arXiv:1902.03477.