Conversion from CSR to BCSR

[jselvam11]

Hi,

This library seems really innovative. I would like to convert the CSR result from https://www.open3d.org/docs/release/python_api/open3d.ml.torch.ops.fixed_radius_search.html.

returns (for example)

neighbors_index = [1, 4, 4]
neighbors_row_splits = [0, 1, 2, 3]
neighbors_distance = []

as torch Tensors

to the Matrix BCSR format introduced in this library. Is there an intuitive way of doing this transformation?

[tgale96]

Hi, thanks! Is the matrix that you have encoded in CSR block-sparse? It will need to be to exploit BSCR and block-sparse computation.

If it is, the conversion should be fairly simple? A simple way to get started would be to convert your matrix to dense and then to BSCR with stk.ops.to_sparse(...). Then we could optimized that transition to avoid going through dense.

[jselvam11]

Took a further look and visualized the sparsity by plotting the csr matrix using pyplot.spy, seems like there are banded regions with block sparse patterns.

Some context, I am trying to learn the association between points and their neighbors which are sparsely represented through an MLP, but my current approach requires me to create a dense tensor from the CSR result which results in me running out of GPU memory. I saw the SparseMLP implementation in megablocks and thought about giving it a shot.

[tgale96]

Oh interesting! Is there a reason you can use the built-in types/ops for unstructured sparsity in PyTorch?