Does torch.bmm support batch sparsexdense->dense?

Yiru_Shen · January 15, 2019, 10:44pm

I have found several topics related to sparse x dense -> dense batch-mul batch-mul-2 and also found the to-do list.
However, it’s not clear about the most up to date status of this question.

So my question is, does torch.bmm support batch version of sparse x dense -> dense?
For example, the expected operation is GxM = Y, where G is sparse of (5, 100, 100), M is dense of (5, 100, 3) and Y is a dense matrix of (5, 100, 3).
Or does torch.spmm support this batch operation?

Sorry that I cannot find the detailed documentation and forgive me if I repost this question.

kevinj22 · January 22, 2019, 12:18pm

I don’t know to be honest. There seems to be a promising package here: https://github.com/cornellius-gp/gpytorch

Check out BDSMM on line 83:

github.com

cornellius-gp/gpytorch/blob/master/gpytorch/utils/sparse.py

#!/usr/bin/env python3

from operator import mul

import torch


def make_sparse_from_indices_and_values(interp_indices, interp_values, num_rows):
    """
    This produces a sparse tensor with a fixed number of non-zero entries in each column.

    Args:
        - interp_indices - Tensor (batch_size) x num_cols x n_nonzero_entries
            A matrix which has the indices of the nonzero_entries for each column
        - interp_values - Tensor (batch_size) x num_cols x n_nonzero_entries
            The corresponding values
        - num_rows - the number of rows in the result matrix

    Returns:
        - SparseTensor - (batch_size) x num_cols x num_rows

This file has been truncated. show original

Seems like they change the indices of what would be the 3d sparse matrix to a 2d by adding the number of rows of the each matrix to the next set of indices.