BUG: linalg.norm(): Don't convert object arrays to float

[mgeier]

This fixes #7575 (Regression in linalg.norm() using dtype=object).

Please tell me if I shall rebase this onto another branch.

[charris]

It would be good to test more of the norms, or at least all that work with this fix.

[mgeier]

I've added more tests in bfa3ed9d29314d9d4a06eb5172ae855d19ae7c8d.

Everything that was broken before the regression is still broken.

[charris]

I've added a comment to the issue #7575.

[charris]

I might put this in 1.11.1, but not 1.12, to give you more time to make your own version of norm and get that out there. Deal?

[mgeier]

If it's definitely going to be dropped for 1.12, I don't care about what's happening in 1.11.1.
I could very quickly switch to the work-around given by @pv in #7575.

But I don't think this would be a good decision for the linalg submodule to demote object arrays from being first-class citizens.

Let me offer a counter-deal:
You add my fix to 1.11.1, in the meantime I'll try to come up with a PR that adds support for object arrays in numpy.linalg.norm() for all possible values of the ord keyword.
If you are satisfied with my PR, it goes into 1.12. If not (or if I can't come up with one until the deadline for 1.12), you can drop the whole thing (and we are back to your deal from above).

What about that?

[charris]

Let's put this in with the understanding that it is something of an implementation detail. If you want a guarantee that your work will not be affected by future changes you should pursue your own implementation of a norm function that works as you need it to for your application.

[charris]

Thanks @mgeier .