Layrub is a runtime data placement strategy for extreme-scale net training. It is developed on BVLC Caffe, and achieves memory savings more than 50% over BVLC Caffe. - View it on GitHub