Is this implemented in any of the deep learning libraries? It looks very useful for debugging, but it also looks like a lot of work to implement, so it would make sense to have it directly integrated in the libraries.