[Debian-med-packaging] Bug#1030907: [Help] Possible scikit-learn issue (Was: Bug#1030907: umap-learn: FTBFS (failing tests))

[Drew Parsons]

On 2023-02-09 15:02, Andreas Tille wrote:
> Control: tags -1 help
> 
> Am Thu, Feb 09, 2023 at 12:50:45AM +0100 schrieb Santiago Vila:
>> Package: src:umap-learn
>> Version: 0.4.5+dfsg-3
>> Severity: serious
>> Tags: ftbfs
>> 
>> Dear maintainer:
>> 
>> During a rebuild of all packages in bookworm, your package failed to 
>> build:
> 
> In Salsa CI you can find a more informative log:
> 
>    https://salsa.debian.org/med-team/umap-learn/-/jobs/3923570
> 
> Looking at line 1307 and following I'm wondering, whether this issue is
> rather caused by scikit-learn which is not fully converted to numpy
> 1.24.
> 
> Any ideas how to fix this?

tldr;  https://github.com/lmcinnes/umap/pull/946

There's a couple of strange aspects to this problem.  The np.matrix 
seems to be generated by scipy.sparse.  np.matrix is "not recommended" 
by numpy, but not actually deprecated (at least not removed) yet.  
scikit-learn is jumping the gun by declaring it's not supported.

On the other hand, scitkit-learn has been "not supporting" it for two 
years now.  This corner of the debian archive is quite a bit out of 
date.  The current version of umap (i.e. umap-learn) is 0.5.3, released 
April last year.

That said, the np.matrix discrepancy was only just fixed recently, in 
December, by
https://github.com/lmcinnes/umap/pull/946

Drew