Pytorch_geometric: How can I implement the dot product (torch.mul()) of a dense matrix and a sparse matrix?

[RenShijie]

I have a mx1 dense matrix D and a mxm sparse matrix S. I implemented that by torch.mul(S.to_dense(), D), and transformed the result into a sparse tensor, but the GPU memory is not enough and running on CPU is too slow for large m. I wonder if there is a fast/sparse implementation of torch_geometric for this, without converting S to a dense matrix (converting D to sparse is acceptable \ (•◡•) /). Thanks for your help ≧◠◡◠≦✌

[rusty1s]

You can do the following:

row, col = edge_index  # Indices of S
edge_weight # Values of S
D_row = D[row]

out = D_row * edge_weight

[RenShijie]

Thanks for your kind and timely reply. By the way, I changed the last line 'out = D_row * edge_weight' into 'out = D_row * edge_weight.squeeze()', that really works ^_^

[RenShijie]

Hi, I have another question:
If I want to put a sparse mask M on a sparse matrix S, except for implementation via M * S, how can I operate on edges and indices to save time and memory?

[rusty1s]

I am not sure I understand, but what about removing the edges in S that also lie in M?

[RenShijie]

Sorry, maybe my description is not clear enough. What I want to do is to preserve the edges in S that the mask sparse matrix M also contains (or remove the edges that M doesn't have), so that the indices of target sparse matrix S are exactly the same as M.
So, how can I implement this? (Maybe how to finding out the common edges they have and delete the others?)

[rusty1s]

There is no PyTorch function for that, but can use numpys isin1d function for that. You can convert sparse indices to 1d via:

idx = N * row + col

[RenShijie]

Thanks for your kind and fast reply, I will have a try on that conversion. I thought there could be an implementation without using 'M * S' (sad...T_T).