How to compute a large sparse matrix in a single GPU efficiently

Question

I have a very large dataframe named diagnoses. I am trying to do matrix multiplication as below

scores = torch.sparse.mm(diagnoses * freq_adjustment.unsqueeze(0), diagnoses.permute(1, 0))

I want to store the scores in a.pt file using pytorch. But keep getting cuda out of memory error. How to do this efficiently?

Answer 1

In order to benefit from torch.sparse you need your tensors diagnoses and freq_adjustment to be of type sparse tensor . Moreover, if the matrix is not sparse -- it would actually take more space in memory.