rigl

End-to-end training of sparse deep neural networks with little-to-no performance loss.