Non-sparse x sparse

[lorenlugosch]

Is it possible to multiply a sparse matrix by a non-sparse matrix? What I'd like to do is multiply a (sparse) transition matrix in an HMM with a (non-sparse) vector of forward log probabilities.

[srush]

Great question. We don't have this yet, but I would like to add it. In the meantime you have two interesting options:

1. If you read and understand this code which implements HMMs using parallel prefix scan (also described in the paper) https://github.com/harvardnlp/pytorch-struct/blob/master/torch_struct/linearchain.py , then you will see that all the operations of sparse matrix x sparse matrix multiplies. Using this method you could have a really fast sparse HMM.

2. If 1 is still not fast enough, we have some unreleased code that implements HMM entirely in Cuda. If you email me we could help adapt this code for your task.

[lorenlugosch]

Thanks!