Pin numpy version at build to handle Numpy's API change in v1.20

[jpauwels]

Changes proposed in this pull request

The C-API change in Numpy v1.20 has the potential to cause incompatibility issues, as seen in #469 (comment).

This happens when a user hasn't upgraded yet to Numpy 1.20 (or can't upgrade, e.g. because of apt-get availability or pinned due to Tensorflow requirements). Because the build requirement in pyproject.toml is an unversioned numpy, it will pull in the latest >=1.20 version and build madmom against it, which gives problems when running with the <1.20 version because the latter is not forward compatible.

Numpy's C-API is forward compatible though, so the solution is to pin the build version at a low enough version, it doesn't matter if this version is lower than the version used at runtime. You could do this per Python version, as is currently done in pyproject.toml, but there even is a meta-package for this since it's a common problem. Using the meta-package also makes the solution robust against future Python versions, so that's what I did for this PR.

I also snuck in an unrelated warning fix that will become an error in Python 3.10 (due in October).

[superbock]

Thanks for the fix! I was not aware of this rather elegant fix :)

[mohkoma]

Is there any update yet? as you said Tensorflow require < 1.20 version.
Thanks!

[faroit]

@superbock this issue is blocking a JOSS review in progress openjournals/joss-reviews#3391 and prevents one of the reviewers (@keunwoochoi) to evaluate the performance of the software. It would be great if a new madmom release can be put out soon! Thanks for your great work

[keunwoochoi]

Thank you all!

[faroit]

@superbock could you issue a new madmom pypi release with this change included to help dependencies resolving? 🙏

[samuelbradshaw]

Hi! It looks like this was merged a couple of years ago, but there hasn't been a release since. Is there a madmom release planned soon?