Support for numpy NaN functions: nan{min,max,mean,argmin,argmax,sum}

[mankoff]

No description provided.

[hgrecco]

It works fine in my computer (Pint develop branch)

>>> import numpy as np
>>> import pint
>>> ureg = pint.UnitRegistry()
>>> [1., 2., 3., float('NaN')] * ureg.meter
<Quantity([  1.   2.   3.  nan], 'meter')>
>>> q = [1., 2., 3., float('NaN')] * ureg.meter
>>> np.nanmean(q)
2.0
>>> np.nanmax(q)
3.0
>>> np.nanargmin(q)
0
>>> np.nanargmax(q)
2

What it does not work (and should) is assigning a NaN:

>>> q[0] = np.nan
Traceback (most recent call last):
...
ValueError
>>> q[0] = float('NaN')
Traceback (most recent call last):
...
ValueError

I am opening a new issue for this.

[hgrecco]

Sorry. My mistake. nanmean and nanmax does not work as the units are lost. I guess this is related to the fact that numpy does not recognize classes not derived from ndarray as numpy arrays.

[mankoff]

np.nanmean(q) works, but units a dropped. q.mean() works and retains units. q.nanmean() does not work. The latter is what I would like, so that units are retained. Note that when uncertainty propagation is fully implemented, it will (should?) work with q.mean(), but I'm not sure how it would work with np.mean(q). Similarly, uncertainties should also work with q.nanmean().

Also, q[0] = np.nan * ureg.m does work.

[hgrecco]

(Oops, it seems that we commented together) Also, please take a look at the docs and this numpy issue. NumPy devs don't think that this is a problem as the issue was closes right away. I am seriously considering having an additional _NPQuantity class for 0.6 to fix this issue and also, #39 and #37

[mankoff]

I'm fairly new to this level of Python development, but it seems we cannot / should not patch numpy so that np.mean(q) returns units based on this Q&A: https://stackoverflow.com/questions/22600365/can-i-decorate-an-explicit-function-call-such-as-np-sqrt However, we can write pint so that q.mean() returns units.

[hgrecco]

Pint supports NumPy functions without monkey patching NumPy. It already supports mean since 0.2 (see https://github.com/hgrecco/pint/blob/master/pint/testsuite/test_numpy.py#L146)
The problem with nanmean, etc is explained above.

[hgrecco]

Review once that #905 is merged