Should we stop overloading NumPy altogether?

[ricardoV94]

Overloading NumPy functions when the input are TensorVariables is cute, but I think confusing as soon as a user tries to use a function that is not implemented or requires a different API (like set_subtensor)

From what I see around it seems that knowledgeable users don't rely on this at all and it's beginners who use it, usually by accident. I think it's more harmful than helpful.

import numpy as np
import aesara.tensor as at

x = at.scalar('x')
y = np.exp(x)
y  # Elemwise{exp,no_inplace}.0

[twiecki]

I do some benefit, it mostly works with just simple functions with np.exp() and new-comers seem to do this pretty automatically and are not confused when it works.

[brandonwillard]

We need to do a lot more to make it actually work. There's a very narrow space where it's usable, and, due to that, the feature is more confusing than not.

A good example is np.dot; I've accidentally used that a few times instead of at.dot and it led to extremely confusing errors—even by my standards.

[dhirschfeld]

It can be convenient for an array implementation to Just Work with the numpy methods but if it sometimes does the wrong thing that would be very bad.

IIUC you can implement __array_function__ to specify what you support, or return NotImplemented to disable the dispatching altogether.

[brandonwillard]

It can be convenient for an array implementation to Just Work with the numpy methods but if it sometimes does the wrong thing that would be very bad.

That's exactly the case, so I'm all for us removing the partial support until someone does it correctly with __array_function__ or whatnot.