Sparse matrix multiplication?

[huyng]

Hi There,

Is there currently support for sparse matrix multiplication in tensorflow?

I found this post:
http://stackoverflow.com/questions/34030140/is-sparse-tensor-multiplication-implemented-in-tensorflow

But it seems to indicate that one must first convert the sparse matrix into a dense matrix before running the multiplication.

[ebrevdo]

This was recently added. See sparse_ops.sparse_tensor_dense_matmul. The first input should be a 2-D SparseTensor and the second a regular dense tensor (matrix) of the same dtype.

Backprop coming soon (at least into the dense tensor, no backprop into the SparseTensor is currently planned).

[ebrevdo]

I'm closing this issue unless there's missing functionality in the op I mentioned.

[ebrevdo]

Use sparse_ops.sparse_tensor_dense_matmul. Does this do what you need?

[huyng]

Woah awesome! Thanks @ebrevdo . Had no idea this recently made it in. I think this will fit what I need, and will try this out tomorrow.

[tansey]

I don't think this is sufficient to close this issue. The stackoverflow post is talking about the general case where you want something like sparse_matmul(sp_a, sp_b).

I currently have that problem: I want to calculate a graph Laplacian given a sparse oriented edge matrix, D. Every row is an edge between two nodes, with -1 for node_i and 1 for node_j; the graph Laplacian is then (D'.D). My graph in this case has something like a million nodes but only about 3 million edges, so the overall sparse D representation has about 6 million entries. This is a piece of cake to calculate using something like scipy's sparse matrix multiplication, but converting it to tensorflow requires creating a dense 1M x 1M matrix.

[ebrevdo]

Feel free to open a new issue asking specifically for (SparseTensor, SparseTensor) matmul.

[theeluwin]

I also have similar problem to @tansey.
I know that there is a matrix solver using LU decomposition in Tensorflow, but it's supported to dense matrices only. I can calculate very sparse L, U, Linv, Uinv matrices (as a SparseTensor type) but there's no way to use it..
Using dense version for any of those blows up my memory (I'm working currently working on graphs with several million nodes).

[Corallus-Caninus]

I also need sparse@sparse matrix multiplication. Even with a_is_sparse and b_is_sparse set to true TF still load n^2 dense matrices into memory.
#3246 (comment)

it seems possible theoretically but I am unsure, are there any updates pertaining to this? Thanks.