Einsum?

[justanhduc]

Hey @rusty1s. Thanks for maintaining this great library. I am wondering if it is possible to support einsum for sparse matrix and tensor multiplication. Is there any difficulty? Thanks in advance!

[rusty1s]

We do not have any plans to integrate einsum support into torch-sparse. Which operation do you try to perform? Eventually, you might be able to do so with the tools already provided in torch-sparse.

[justanhduc]

My use case is that I wanna dot a sparse matrix m x n with another tensor b x n x d. With the current machinery I have to do a lot of transpose/untranspose/flatten/contiguous etc... Do you have any idea how to avoid all these?

[rusty1s]

This should be already supported in torch-sparse:

from torch_sparse import SparseTensor
sparse_mat = SparseTensor(row=row, col=col, sparse_sizes(m, n))
dense_mat = torch.randn(b, n, d)
out = sparse_mat @ dense_mat  # b x m x d

[justanhduc]

Ok I missed it. Perhaps you may want to make it more visible in the readme. Anw thanks a lot!

[yonatansverdlov]

This should be already supported in torch-sparse:

from torch_sparse import SparseTensor
sparse_mat = SparseTensor(row=row, col=col, sparse_sizes(m, n))
dense_mat = torch.randn(b, n, d)
out = sparse_mat @ dense_mat  # b x m x d

If I have a tensor ten1 of shape [b,n,dim,i] which is sparse, and tne2 tensor of shape [i,m] and I want to output ten1@ten2, how can I use what you did here?

[rusty1s]

This wouldn't be supported since sparse matrix multiplication cannot handle 4-dimensional tensors (needs to be 2 dimensions). Your best option is to reshape ten1 into [b * n * dim, i].

[yonatansverdlov]

Why not?
Will it be supported in the future?
Do you know something that helps with sparse tensor multiplication?

[rusty1s]

Likely not, SparseTensor is designed to have at most two sparse dimensions (while any other dimension afterwards can be dense).