rok commented on issue #47520: URL: https://github.com/apache/arrow/issues/47520#issuecomment-3298308358
> [@rok](https://github.com/rok) I've investigated all tensor sparse formats and I want to give my final opinion: Thanks for checking! > 1. Regarding [this](https://github.com/apache/arrow/issues/47520#issuecomment-3271622837), my assumption about case 2 was wrong; the problem only exists in the CooTensor format. Great to hear it's a COO only issue. > 2. All types of sparse tensor format creation from dense tensors have the problem of handling negative zero. This is not great, but not sure how and if we should be fixing it. It's definitely is less urgent than the COO issue. Let's open a new issue for it? > 3. HalfFloatType is not handled during the creation of any format at all (negative zero does not cause a memory problem). I'd say this is a new feature and we'd want to handle it separately in a new issue. > > This is quite unfortunate, and we should address it ASAP. pitrou, it would probably be best to fix this before the release. > > My initial solution is ready; however, since I've changed 12 files, it will take several days to clean it up. I think the release is frozen on October 6th. Do you have an open PR? > [@rok](https://github.com/rok) [@pitrou](https://github.com/pitrou) One more point regarding the sparse tensor specification: the current format treats all negative zeros as positive zeros. So, converting a dense tensor to a sparse tensor and back will turn every negative zero into a positive zero. Is this intended behavior? This is the same as 2. ? It's not intended behavior if I remember correctly. I think this didn't come up when we originally implemented these. Perhaps we should document it. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
