Make backend array functions callable from Triangle

[genedan]

Description

There are several calls to xp = self.get_array_module() throughout the library. xp results in numpy or sparse. xp then calls an array function to do something, e.g.,:

xp = self.get_array_module()
fill = (xp.nan_to_num(self.values) == 0) * (self * 0 + value)

This pattern appears frequently throughout the library. I think it may be possible to simplify the call to just:

fill = (self.nan_to_num(self.values) == 0) * (self * 0 + value)

Where the array function (in this case, nan_to_num()) is the one pertaining to the Triangle's backend, e.g., numpy.nan_to_num() or sparse.nan_to_num().

Calling Triangle.set_backend() will swap the array functions to be called from the appropriate backend module.

Is your feature request aligned with the scope of the package?

• Yes, absolutely!
• No, but it's still worth discussing.
• N/A (this request is not a codebase enhancement).

Describe the solution you'd like, or your current workaround.

No response

Do you have any additional supporting notes?

No response

Would you be willing to contribute this ticket?

• Yes, absolutely!
• Yes, but I would like some help.
• No.

[henrydingliu]

is this at odds with your ambitions from #1002, i.e. we should move away from passthrus? as we discourage the pattern of calling get_array_module(), i think we need to pick a lane between providing a passthru at the object level (which is essentially what this issue turns into), or by implementing those methods into the packages.

[genedan]

Yeah it is conflicting. I ran into further difficulty when actually trying to sketch out a solution. These are module level functions that manipulate arrays rather than Triangle methods. I wonder if there would instead be a way to import the correct version of something like nan_to_num at the top of the module instead of repeated executions of xp = self.get_array_module().

[henrydingliu]

we can probably go through the repo and try to refactor out quite a few instances of xp, like nan_to_num to fillna. some other ones probably deserve their own implementation, like ones. hopefully by that time we are left with very few instances of xp that we can just let it slide