Sparse Matrix status


I’m trying to find any up to date information about the status on Pytorch’s sparse matrix capabilities, but I can’t seem to find anything recent.

Does anyone have any estimated timeline for when the sparse module would be operational?

More concretely for my particular issue:
I’m doing research in machine learning, and we are using pytorch, but unfortunately the lack of sparse matrix module is a huge hindrance for the networks we are producing.
Thus far I have managed to hack my own sparse matrix multiplication together when one of the matrices is a diagonal matrix. (It isn’t fast and neither is it pretty, but it does work), and combined with the matrix multiplication of (sparse,dense) I can actually compute all the operations I need for the network I’m making. However when I run my network and do .backward() I encounter “RuntimeError: calculating the gradient of a sparse Tensor argument to mm is not supported.”


With pytorch 1.6, now you can call, B) where A is the sparse matrix and B is a dense matrix (I think it also work with sparse B).