Fix numpy warnings

[mstimberg]

This fixes a number of issues that will raise warnings in the forthcoming numpy version (1.25).

[thesamovar]

Can you briefly summarise what's going on here? I have to say I don't really understand the changes.

[thesamovar]

Not sure I understand these changes, can you summarise?

[mstimberg]

This change is a slightly more complex issue related to what I wrote in my earlier comment: in future versions, we can no longer use float(x) as a check that x is either a scalar number or a length-1 array with a number (and not for example a string) – it will wail on length-1 arrays.

[thesamovar]

Any reason for change from sin to exp etc.? Not that it matters much, just curious.

[mstimberg]

I briefly explain this in the commit message: our docttest relied on the fact that np.sin(np.pi/2) == 1.0, which is reasonable but not guaranteed (results are only guaranteed to be < 4 epsilons away from the correct value). With the previous versions it matched 1.0 exactly, but with the newer, faster implementations in numpy 1.25 this is no longer the case (even with older versions, np.sin(0.0) or np.sin(np.pi) did not return 0.0). So I replaced it with exp(-inf) which I think is guaranteed to be exactly zero – at least it works with both numpy 1.24 and numpy 1.25 :)

[mstimberg]

Can you briefly summarise what's going on here? I have to say I don't really understand the changes.

Sure. The main issue that is adressed here is that in a few places, we used code like this:

int(self.variables["N"].get_value())

The result of self.variables["N"].get_value() is sometimes a scalar value (e.g. for NeuronGroup where N is a constant), and sometimes a length-1 array (e.g. for Synapses, where the value can change). By wrapping it in int, we were sure about always getting a scalar Python int. Starting with numpy 1.25 (currently a release candidate), this will raise a warning, and in future versions it will raise an error. To replace this code, I added an item() function to ArrayVariable, Constant, and Quantity (the latter already had it inherited from numpy, but discarded the units). This mirrors the built-in function that numpy offers:

>>> import numpy as np
>>> np.array(3).item()
3
>>> np.array([3]).item()
3
>>> np.array([3, 4]).item()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ValueError: can only convert an array of size 1 to a Python scalar

[mstimberg]

BTW: This was always wrong (we need a stable sort here), but it only now got triggered with numpy 1.25 (and only on GitHub, not on my local machine).

[thesamovar]

OK, all that makes sense - thanks!