News for hackers. Stuff that matters.

A comment on Sparse matrix representations in SciPy (2017). Parent comment

Re:Sparse matrix representations in SciPy (2017) by timr on Thursday August 02, 2018 @11:59PM (#17677201)
That's true, but mostly you just want to be able to store massive sparse matrices without filling your RAM with zeros (or spending tons of time reading/writing zeros from disk). Any implementation will do, as long as it works.
[ Parent | on HN ]
Re:Sparse matrix representations in SciPy (2017) by srean on Friday August 03, 2018 @02:13AM (#17677613)
Storage is a minor worry. The main thing is whether the storage makes the frequent operations efficient. The tuple (index_position, value) is reasonably storage efficient but abysmal for linear algebraic operations. So one cannot decouple storage efficiency concerns from compute concerns too much.

I think this is one of those shibboleths that separate folks who are experts in ML/data-science at scale vs those who are coming more from a enterprise/DBA background. I often have to convince folks why its not a great idea to write sparse matrix multiplication in SQL.

[ Parent | on HN ]
Re:Sparse matrix representations in SciPy (2017) by timr on Friday August 03, 2018 @02:22PM (#17682243)
"Storage is a minor worry."

For many domains, it's a big worry. I routinely encounter large, sparse datasets where I/O and storage dominate other concerns. Big sparse tensors are worse than the "slowness" associated with a suboptimal sparse vector math implementation.

"I think this is one of those shibboleths that separate folks who are experts in ML/data-science at scale vs those who are coming more from a enterprise/DBA background."

I've been doing ML work for a long time. This is actually a shibboleth that separates folks who mostly do image classification from other kinds of ML people.

[ Parent | on HN ]
Re:Sparse matrix representations in SciPy (2017) by srean on Saturday August 04, 2018 @11:35AM (#17687227)
"Minor" in the sense relatively easy to solve if not constrained by the compute needs.

Interesting comment regarding image classification. I would have assumed they are the ones who don't have to worry about sparsity much. Physics, and image and video are some of the places you encounter large and dense matrices. Lapack's wet dream :)

[ Parent | on HN ]