Any thoughts of ways to improve this?

Aside, I wonder if the generator needs to be in a list?

List is definitely unnecessary.

The ideal solution is probably to make isscalar work correctly for this case.

Hmm, making gradient work correctly here could get tricky, especially with unevenly spaced samples as proposed in #8446. For scalars, it is suggested that custom scalar types inherit from np.generic. What does 1 * ureg('m') look like?

It's just a thin wrapper around a Python integer:

In [1]: import pint
In [2]: import numpy as np
In [3]: ureg = pint.UnitRegistry()

In [4]: 1 * ureg('m')
Out[4]: <Quantity(1, 'meter')>

In [5]: type((1 * ureg('m')).magnitude)
Out[5]: int

Seems like isscalar should be fixed to accept instances of int and long (Python2). I assume that `isinstance(ureg, int) works, is that correct?

It'd be more like

s = 1 * ureg('m')
isinstance(s, int)

But regardless, that last check returns False. I think only duck typing would solve that problem--and I understand if that's too much to bite off.

Not sure, but isscalar actually checks for scalar objects, not even 0-d arrays would fit that for example. Don't we have some other similar way to achieve this?

The former example (numpy v.1.10):
x = np.array([0, 1, 2, 3, 4])
dx = np.gradient(x)
y = x**2
np.gradient(y, dx, edge_order=2)
also fails, because dx is always an array of arrays and never an array of scalars if varargs is supplied.

@seberg

Not sure, but isscalar actually checks for scalar objects, not even 0-d arrays

Based on this, is it ever appropriate to use isscalar in its current form? isscalar was the cause of the bug behind #8441. Shouldn't the test just be np.ndim(x) == 0, always?

bump...I'm using another package pynbody which implements SimArrays, wrappers around numpy arrays that carry units (meters, seconds, etc...). Allowing zero dimensional arrays in np.gradient would be a big help. Older versions of numpy worked just fine with this.

@ibackus: This looks like an easy fix to me, making the change I suggest in the comment above. Do you want to have a crack at implementing it?

@eric-wieser: For now I'll have to pass...I've not contributed to numpy before and don't really have the time to getting things set up so that I can do that properly. all that is required is changing isscalar to np.ndim(x)==0 as you noted.

@eric-wieser, I would fix this but I do not grasp your solution fully. I see that np.isinstance is not working properly. Am I right that you are suggesting to get rid of that check and instead check the dimension of x?

Assuming you mean np.isscalar and not isinstance, then yes - checking dimension == 0 should suffice

Looking into it I found this might already been resolved by #8446?

@maluethi: Good find, but I think the fix is still needed on this line.

You should be able to test this with something like np.gradient(np.ones((3, 3)), np.array(1))