COMPAT: Further Expand Compatibility with fromnumeric.py

[gfyoung]

Follow-on to #12810 by expanding compatibility with fromnumeric.py in the following modules:

1. tslib.pyx
2. window.py
3. groupby.py and resample.py (shared classes)

Closes #12811.

[jreback]

I would just make this a 1-liner. This is not something we care to advertise.

[gfyoung]

Fair enough. Done.

[codecov-io]

Current coverage is 84.14%

Merging #13148 into master will decrease coverage by <.01%

@@ master #13148 diff @@ ========================================== Files 138 138 Lines 50394 50409 +15 Methods 0 0 Messages 0 0 Branches 0 0 ========================================== + Hits 42409 42416 +7  - Misses 7985 7993 +8  Partials 0 0 

Powered by Codecov. Last updated by 2a120cf...e5465a6

[jreback]

@gfyoung let's maybe just disallow using numpy functions on window/groupby ops (the Timestamp one is prob ok).

this should just raise a nicer error message (say this is not supported)

In [6]: np.mean(Series(range(3)).groupby(range(3))) TypeError: mean() got an unexpected keyword argument 'axis' 

[gfyoung]

Fair enough. So instead of validating as I do now, I just check if the args or kwargs contain any numpy-specific arguments (not sure how else to check if the pandas function is being called from numpy otherwise)?

[jreback]

not really sure how to figure out if called via numpy w/o stack inspection (not a good idea!).

though you could simply just have it raise when it validates, e.g. its sort of duck like in the argument signature when called from numpy. IOW numpy is always passing dtype/out I think. So just raise a nicer error message on that (you prob need to pass a parameter to your validation routines to say hey if this is not validation, then print this nice error message; e.g. you want to say something like)

""" numpy operations are not valid with groupy, use .groupby(...).mean() instead

[jreback]

hmm not sure I like the name. maybe UnsupportedFunctionCall? and move to core.common. inherit from ValueError

[gfyoung]

That works too. Will change.

[jorisvandenbossche]

Can you give a easier explanation? Users don't know what fromnumeric.py is

[gfyoung]

Done.

[gfyoung]

@jreback, @jorisvandenbossche : Made the changes, and Travis is giving the green light. Ready to merge if there is nothing else.

[jreback]

thanks @gfyoung

[jreback]

@gfyoung breaking with numpy master on rounding: https://travis-ci.org/pydata/pandas/jobs/131735642

[gfyoung]

Okay, I think we might need to undo this compatibility change to tslib.pyx. If anyone complains about it, say that "round in numpy is for arrays and not timestamps, and you shouldn't be doing that" (currently, the function will throw an error because keyword arguments like decimals and out are not in the signature and shouldn't be). Here is why:

# signature in tslib.pyx def round(self, freq, *args, **kwargs): ... # signature in numpy master def around(a, decimals=0, out=None): return _wrapfunc(a, 'round', decimals=decimals, out=out)

When we call np.round(timestamp, freq), freq is being passed in as the decimals argument. However, decimals is being passed in with a **kwargs bundle, so it isn't being picked up as the positional argument needed for freq. I forgot that that change was not in v1.11.0 but was in master. My bad for not checking.

[jreback]

ok so just revert that part of the PR
compat for timestamps is not s big deal for this